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 if (connect_op == MDS_CONNECT) {
440 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
441 OBD_ALLOC(cli->cl_mod_tag_bitmap,
442 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
443 if (cli->cl_mod_tag_bitmap == NULL)
444 GOTO(err, rc = -ENOMEM);
449 CERROR("ldlm_get_ref failed: %d\n", rc);
453 ptlrpc_init_client(rq_portal, rp_portal, name,
454 &obddev->obd_ldlm_client);
456 imp = class_new_import(obddev);
458 GOTO(err_ldlm, rc = -ENOENT);
459 imp->imp_client = &obddev->obd_ldlm_client;
460 imp->imp_connect_op = connect_op;
461 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
462 LUSTRE_CFG_BUFLEN(lcfg, 1));
463 class_import_put(imp);
465 if (lustre_cfg_buf(lcfg, 4)) {
466 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
468 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
470 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
471 obddev->obd_name, lustre_cfg_string(lcfg, 4),
473 GOTO(err_import, rc);
475 fake_conn.c_self = LNET_MKNID(refnet, 0);
476 imp->imp_connection = &fake_conn;
479 rc = client_import_add_conn(imp, &server_uuid, 1);
481 CERROR("can't add initial connection\n");
482 GOTO(err_import, rc);
484 imp->imp_connection = NULL;
486 cli->cl_import = imp;
487 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
488 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
490 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
491 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
492 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
493 name, obddev->obd_name,
494 cli->cl_target_uuid.uuid);
495 spin_lock(&imp->imp_lock);
496 imp->imp_deactive = 1;
497 spin_unlock(&imp->imp_lock);
501 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
502 LDLM_NAMESPACE_CLIENT,
503 LDLM_NAMESPACE_GREEDY,
505 if (obddev->obd_namespace == NULL) {
506 CERROR("Unable to create client namespace - %s\n",
508 GOTO(err_import, rc = -ENOMEM);
514 class_destroy_import(imp);
518 if (cli->cl_mod_tag_bitmap != NULL)
519 OBD_FREE(cli->cl_mod_tag_bitmap,
520 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
521 cli->cl_mod_tag_bitmap = NULL;
525 EXPORT_SYMBOL(client_obd_setup);
527 int client_obd_cleanup(struct obd_device *obddev)
529 struct client_obd *cli = &obddev->u.cli;
532 ldlm_namespace_free_post(obddev->obd_namespace);
533 obddev->obd_namespace = NULL;
535 obd_cleanup_client_import(obddev);
536 LASSERT(obddev->u.cli.cl_import == NULL);
540 if (cli->cl_mod_tag_bitmap != NULL)
541 OBD_FREE(cli->cl_mod_tag_bitmap,
542 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
543 cli->cl_mod_tag_bitmap = NULL;
547 EXPORT_SYMBOL(client_obd_cleanup);
549 /* ->o_connect() method for client side (OSC and MDC and MGC) */
550 int client_connect_import(const struct lu_env *env,
551 struct obd_export **exp,
552 struct obd_device *obd, struct obd_uuid *cluuid,
553 struct obd_connect_data *data, void *localdata)
555 struct client_obd *cli = &obd->u.cli;
556 struct obd_import *imp = cli->cl_import;
557 struct obd_connect_data *ocd;
558 struct lustre_handle conn = { 0 };
563 down_write(&cli->cl_sem);
564 if (cli->cl_conn_count > 0)
565 GOTO(out_sem, rc = -EALREADY);
567 rc = class_connect(&conn, obd, cluuid);
571 cli->cl_conn_count++;
572 *exp = class_conn2export(&conn);
574 LASSERT(obd->obd_namespace);
576 imp->imp_dlm_handle = conn;
577 rc = ptlrpc_init_import(imp);
581 ocd = &imp->imp_connect_data;
584 imp->imp_connect_flags_orig = data->ocd_connect_flags;
585 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
588 rc = ptlrpc_connect_import(imp);
590 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
593 LASSERT(*exp != NULL && (*exp)->exp_connection);
596 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
597 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
598 data->ocd_connect_flags, ocd->ocd_connect_flags);
599 data->ocd_connect_flags = ocd->ocd_connect_flags;
602 ptlrpc_pinger_add_import(imp);
608 cli->cl_conn_count--;
609 class_disconnect(*exp);
613 up_write(&cli->cl_sem);
617 EXPORT_SYMBOL(client_connect_import);
619 int client_disconnect_export(struct obd_export *exp)
621 struct obd_device *obd = class_exp2obd(exp);
622 struct client_obd *cli;
623 struct obd_import *imp;
628 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
629 exp, exp ? exp->exp_handle.h_cookie : -1);
634 imp = cli->cl_import;
636 down_write(&cli->cl_sem);
637 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
640 if (cli->cl_conn_count == 0) {
641 CERROR("disconnecting disconnected device (%s)\n",
643 GOTO(out_disconnect, rc = -EINVAL);
646 cli->cl_conn_count--;
647 if (cli->cl_conn_count != 0)
648 GOTO(out_disconnect, rc = 0);
650 /* Mark import deactivated now, so we don't try to reconnect if any
651 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
652 * fully deactivate the import, or that would drop all requests. */
653 spin_lock(&imp->imp_lock);
654 imp->imp_deactive = 1;
655 spin_unlock(&imp->imp_lock);
657 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
658 * delete it regardless. (It's safe to delete an import that was
660 (void)ptlrpc_pinger_del_import(imp);
662 if (obd->obd_namespace != NULL) {
663 /* obd_force == local only */
664 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
665 obd->obd_force ? LCF_LOCAL : 0, NULL);
666 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
669 /* There's no need to hold sem while disconnecting an import,
670 * and it may actually cause deadlock in GSS. */
671 up_write(&cli->cl_sem);
672 rc = ptlrpc_disconnect_import(imp, 0);
673 down_write(&cli->cl_sem);
675 ptlrpc_invalidate_import(imp);
680 /* Use server style - class_disconnect should be always called for
682 err = class_disconnect(exp);
686 up_write(&cli->cl_sem);
690 EXPORT_SYMBOL(client_disconnect_export);
692 #ifdef HAVE_SERVER_SUPPORT
693 int server_disconnect_export(struct obd_export *exp)
698 /* Disconnect early so that clients can't keep using export. */
699 rc = class_disconnect(exp);
700 /* Close import to avoid sending any requests. */
701 if (exp->exp_imp_reverse)
702 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
704 ldlm_bl_thread_wakeup();
706 /* complete all outstanding replies */
707 spin_lock(&exp->exp_lock);
708 while (!list_empty(&exp->exp_outstanding_replies)) {
709 struct ptlrpc_reply_state *rs =
710 list_entry(exp->exp_outstanding_replies.next,
711 struct ptlrpc_reply_state, rs_exp_list);
712 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
714 spin_lock(&svcpt->scp_rep_lock);
716 list_del_init(&rs->rs_exp_list);
717 spin_lock(&rs->rs_lock);
718 ptlrpc_schedule_difficult_reply(rs);
719 spin_unlock(&rs->rs_lock);
721 spin_unlock(&svcpt->scp_rep_lock);
723 spin_unlock(&exp->exp_lock);
727 EXPORT_SYMBOL(server_disconnect_export);
729 /* --------------------------------------------------------------------------
730 * from old lib/target.c
731 * -------------------------------------------------------------------------- */
733 static int target_handle_reconnect(struct lustre_handle *conn,
734 struct obd_export *exp,
735 struct obd_uuid *cluuid)
737 struct obd_device *target;
738 struct lustre_handle *hdl;
745 hdl = &exp->exp_imp_reverse->imp_remote_handle;
746 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
747 conn->cookie = exp->exp_handle.h_cookie;
748 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
749 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
753 target = exp->exp_obd;
755 /* Might be a re-connect after a partition. */
756 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
757 LCONSOLE_WARN("%s: already connected client %s (at %s) "
758 "with handle %#llx. Rejecting client "
759 "with the same UUID trying to reconnect "
760 "with handle %#llx\n", target->obd_name,
761 obd_uuid2str(&exp->exp_client_uuid),
762 obd_export_nid2str(exp),
763 hdl->cookie, conn->cookie);
764 memset(conn, 0, sizeof *conn);
765 /* target_handle_connect() treats EALREADY and
766 * -EALREADY differently. -EALREADY is an error
767 * (same UUID, different handle). */
771 if (!target->obd_recovering) {
772 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
773 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
774 obd_export_nid2str(exp));
775 GOTO(out_already, rc);
778 now = cfs_time_current();
779 deadline = target->obd_recovery_timer.expires;
780 if (cfs_time_before(now, deadline)) {
781 struct target_distribute_txn_data *tdtd =
782 class_exp2tgt(exp)->lut_tdtd;
787 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
788 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
789 buf = tdtd->tdtd_show_update_logs_retrievers(
790 tdtd->tdtd_show_retrievers_cbdata,
794 LCONSOLE_WARN("%s: Recovery already passed deadline "
795 "%d:%.02d. It is due to DNE recovery "
796 "failed/stuck on the %d MDT(s):%s. "
797 "Please wait until all MDTs recovered "
798 "or abort the recovery by force.\n",
799 target->obd_name, timeout / 60,
801 buf ? buf : "unknown (not enough RAM)");
803 LCONSOLE_WARN("%s: Recovery already passed deadline "
804 "%d:%.02d. If you do not want to wait "
805 "more, please abort the recovery by "
806 "force.\n", target->obd_name,
807 timeout / 60, timeout % 60);
812 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
813 LCONSOLE_WARN("%s: Recovery already passed deadline"
814 " %d:%.02d, It is most likely due to DNE"
815 " recovery is failed or stuck, please wait a"
816 " few more minutes or abort the recovery.\n",
817 target->obd_name, timeout / 60, timeout % 60);
821 conn->cookie = exp->exp_handle.h_cookie;
822 /* target_handle_connect() treats EALREADY and
823 * -EALREADY differently. EALREADY means we are
824 * doing a valid reconnect from the same client. */
829 check_and_start_recovery_timer(struct obd_device *obd,
830 struct ptlrpc_request *req, int new_client);
833 * update flags for import during reconnect process
835 static int rev_import_flags_update(struct obd_import *revimp,
836 struct ptlrpc_request *req)
839 struct obd_connect_data *data;
841 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
843 if (data->ocd_connect_flags & OBD_CONNECT_AT)
844 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
846 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
848 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
850 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
852 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
853 revimp->imp_client->cli_name,
854 libcfs_id2str(req->rq_peer), rc);
862 * Allocate a new reverse import for an export.
864 * \retval -errno in case error hit
865 * \retval 0 if reverse import correctly init
867 int rev_import_init(struct obd_export *export)
869 struct obd_device *obd = export->exp_obd;
870 struct obd_import *revimp;
872 LASSERT(export->exp_imp_reverse == NULL);
874 revimp = class_new_import(obd);
878 revimp->imp_remote_handle.cookie = 0ULL;
879 revimp->imp_client = &obd->obd_ldlm_client;
880 revimp->imp_dlm_fake = 1;
882 /* it is safe to connect import in new state as no sends possible */
883 spin_lock(&export->exp_lock);
884 export->exp_imp_reverse = revimp;
885 spin_unlock(&export->exp_lock);
886 class_import_put(revimp);
890 EXPORT_SYMBOL(rev_import_init);
893 * Handle reconnect for an export.
895 * \param exp export to handle reconnect process
896 * \param req client reconnect request
898 * \retval -rc in case securitfy flavor can't be changed
899 * \retval 0 in case none problems
901 static int rev_import_reconnect(struct obd_export *exp,
902 struct ptlrpc_request *req)
904 struct obd_import *revimp = exp->exp_imp_reverse;
905 struct lustre_handle *lh;
908 /* avoid sending a request until import flags are changed */
909 ptlrpc_import_enter_resend(revimp);
911 if (revimp->imp_connection != NULL)
912 ptlrpc_connection_put(revimp->imp_connection);
915 * client from recovery don't have a handle so we need to take from
916 * request. it may produce situation when wrong client connected
917 * to recovery as we trust a client uuid
919 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
920 revimp->imp_remote_handle = *lh;
922 /* unknown versions will be caught in
923 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
924 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
926 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
928 rc = rev_import_flags_update(revimp, req);
930 /* it is safe to still be in RECOVERY phase as we are not able
931 * to setup correct security flavor so requests are not able to
932 * be delivered correctly */
936 /* resend all rpc's via new connection */
937 return ptlrpc_import_recovery_state_machine(revimp);
940 int target_handle_connect(struct ptlrpc_request *req)
942 struct obd_device *target = NULL;
943 struct obd_export *export = NULL;
944 /* connect handle - filled from target_handle_reconnect in
946 struct lustre_handle conn;
947 struct lustre_handle *tmp;
948 struct obd_uuid tgtuuid;
949 struct obd_uuid cluuid;
954 bool mds_conn = false, lw_client = false, initial_conn = false;
955 bool mds_mds_conn = false;
956 bool new_mds_mds_conn = false;
957 bool target_referenced = false;
958 struct obd_connect_data *data, *tmpdata;
960 lnet_nid_t *client_nid = NULL;
963 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
965 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
967 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
968 GOTO(out, rc = -EINVAL);
971 obd_str2uuid(&tgtuuid, str);
972 target = class_uuid2obd(&tgtuuid);
974 target = class_name2obd(str);
977 deuuidify(str, NULL, &target_start, &target_len);
978 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
979 "from %s (no target). If you are running "
980 "an HA pair check that the target is "
981 "mounted on the other server.\n", str,
982 libcfs_nid2str(req->rq_peer.nid));
983 GOTO(out, rc = -ENODEV);
986 spin_lock(&target->obd_dev_lock);
987 if (target->obd_stopping || !target->obd_set_up) {
988 spin_unlock(&target->obd_dev_lock);
990 deuuidify(str, NULL, &target_start, &target_len);
991 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
992 target_len, target_start,
993 libcfs_nid2str(req->rq_peer.nid),
994 (target->obd_stopping ?
995 "stopping" : "not set up"));
996 GOTO(out, rc = -ENODEV);
999 if (target->obd_no_conn) {
1000 spin_unlock(&target->obd_dev_lock);
1002 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1003 "from %s\n", target->obd_name,
1004 libcfs_nid2str(req->rq_peer.nid));
1005 GOTO(out, rc = -EAGAIN);
1008 /* Make sure the target isn't cleaned up while we're here. Yes,
1009 * there's still a race between the above check and our incref here.
1010 * Really, class_uuid2obd should take the ref. */
1011 class_incref(target, __func__, current);
1012 target_referenced = true;
1014 target->obd_conn_inprogress++;
1015 spin_unlock(&target->obd_dev_lock);
1017 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1019 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1020 GOTO(out, rc = -EINVAL);
1023 obd_str2uuid(&cluuid, str);
1025 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1027 GOTO(out, rc = -EPROTO);
1031 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1033 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1035 GOTO(out, rc = -EPROTO);
1037 rc = req_capsule_server_pack(&req->rq_pill);
1041 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1042 /* Don't allow clients to connect that are using old 1.8 format
1043 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1044 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1045 * FULL20 flag should be set on all connections since 2.0, but no
1046 * longer affects behaviour.
1048 * Later this check will be disabled and the flag can be retired
1049 * completely once interop with 3.0 is no longer needed.
1051 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1052 GOTO(out, rc = -EPROTO);
1055 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1056 if (data->ocd_version < LUSTRE_VERSION_CODE -
1057 LUSTRE_VERSION_ALLOWED_OFFSET ||
1058 data->ocd_version > LUSTRE_VERSION_CODE +
1059 LUSTRE_VERSION_ALLOWED_OFFSET) {
1060 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1061 "libclient connection attempt",
1062 data->ocd_version < LUSTRE_VERSION_CODE ?
1064 OBD_OCD_VERSION_MAJOR(data->ocd_version),
1065 OBD_OCD_VERSION_MINOR(data->ocd_version),
1066 OBD_OCD_VERSION_PATCH(data->ocd_version),
1067 OBD_OCD_VERSION_FIX(data->ocd_version));
1068 data = req_capsule_server_sized_get(&req->rq_pill,
1070 offsetof(typeof(*data), ocd_version) +
1071 sizeof(data->ocd_version));
1073 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1074 data->ocd_version = LUSTRE_VERSION_CODE;
1076 GOTO(out, rc = -EPROTO);
1080 /* Note: lw_client is needed in MDS-MDS failover during update log
1081 * processing, so we needs to allow lw_client to be connected at
1082 * anytime, instead of only the initial connection */
1083 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1085 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1086 initial_conn = true;
1087 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1088 mds_mds_conn = (data->ocd_connect_flags &
1089 OBD_CONNECT_MDS_MDS) != 0;
1091 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1092 * for Imperative Recovery connection from MGC to MGS.
1094 * Via check OBD_CONNECT_FID, we can distinguish whether
1095 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1098 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1099 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1100 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1101 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1102 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1103 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1105 /* We do not support the MDT-MDT interoperations with
1106 * different version MDT because of protocol changes. */
1107 if (unlikely(major != LUSTRE_MAJOR ||
1108 minor != LUSTRE_MINOR ||
1109 abs(patch - LUSTRE_PATCH) > 3)) {
1110 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1111 "connection from different version MDT "
1112 "(%d.%d.%d.%d) %s %s\n",
1113 target->obd_name, LUSTRE_MAJOR,
1114 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1115 major, minor, patch,
1116 OBD_OCD_VERSION_FIX(data->ocd_version),
1117 libcfs_nid2str(req->rq_peer.nid), str);
1119 GOTO(out, rc = -EPROTO);
1124 /* lctl gets a backstage, all-access pass. */
1125 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1126 goto dont_check_exports;
1128 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1132 /* We've found an export in the hash. */
1134 spin_lock(&export->exp_lock);
1136 if (export->exp_connecting) { /* bug 9635, et. al. */
1137 spin_unlock(&export->exp_lock);
1138 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1139 export->exp_obd->obd_name, export,
1140 libcfs_nid2str(req->rq_peer.nid));
1141 class_export_put(export);
1144 } else if ((mds_conn || (lw_client && initial_conn) ||
1145 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1146 export->exp_connection != NULL) {
1147 spin_unlock(&export->exp_lock);
1148 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1149 /* MDS or LWP reconnected after failover. */
1150 LCONSOLE_WARN("%s: Received %s connection from "
1151 "%s, removing former export from %s\n",
1152 target->obd_name, mds_conn ? "MDS" : "LWP",
1153 libcfs_nid2str(req->rq_peer.nid),
1154 libcfs_nid2str(export->exp_connection->c_peer.nid));
1156 /* New MDS connection from the same NID. */
1157 LCONSOLE_WARN("%s: Received new %s connection from "
1158 "%s, removing former export from same NID\n",
1159 target->obd_name, mds_conn ? "MDS" : "LWP",
1160 libcfs_nid2str(req->rq_peer.nid));
1163 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1164 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1165 /* Because exports between MDTs will always be
1166 * kept, let's do not fail such export if they
1167 * come from the same NID, otherwise it might
1168 * cause eviction between MDTs, which might
1169 * cause namespace inconsistency */
1170 spin_lock(&export->exp_lock);
1171 export->exp_connecting = 1;
1172 spin_unlock(&export->exp_lock);
1173 conn.cookie = export->exp_handle.h_cookie;
1176 class_fail_export(export);
1177 class_export_put(export);
1181 } else if (export->exp_connection != NULL && initial_conn &&
1182 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1183 spin_unlock(&export->exp_lock);
1184 /* In MDS failover we have static UUID but NID can change. */
1185 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1186 "existing nid %s is already connected\n",
1187 target->obd_name, cluuid.uuid,
1188 libcfs_nid2str(req->rq_peer.nid),
1190 export->exp_connection->c_peer.nid));
1192 class_export_put(export);
1195 export->exp_connecting = 1;
1196 spin_unlock(&export->exp_lock);
1197 LASSERT(export->exp_obd == target);
1199 rc = target_handle_reconnect(&conn, export, &cluuid);
1202 /* If we found an export, we already unlocked. */
1205 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1206 } else if (req->rq_export == NULL &&
1207 atomic_read(&export->exp_rpc_count) > 0) {
1208 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1209 "still busy with %d references\n",
1210 target->obd_name, cluuid.uuid,
1211 libcfs_nid2str(req->rq_peer.nid),
1212 atomic_read(&export->exp_refcount));
1213 GOTO(out, rc = -EBUSY);
1214 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1215 if (!strstr(cluuid.uuid, "mdt"))
1216 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1217 "known client %s (at %s) because it "
1218 "is indicating it is a new client",
1219 target->obd_name, cluuid.uuid,
1220 libcfs_nid2str(req->rq_peer.nid));
1221 GOTO(out, rc = -EALREADY);
1223 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1230 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1231 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1232 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1233 export, (long)cfs_time_current_sec(),
1234 export ? (long)export->exp_last_request_time : 0);
1236 /* If this is the first time a client connects, reset the recovery
1237 * timer. Discard lightweight connections which might be local. */
1238 if (!lw_client && rc == 0 && target->obd_recovering)
1239 check_and_start_recovery_timer(target, req, export == NULL);
1241 /* We want to handle EALREADY but *not* -EALREADY from
1242 * target_handle_reconnect(), return reconnection state in a flag. */
1243 if (rc == EALREADY) {
1244 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1250 /* Tell the client if we support replayable requests. */
1251 if (target->obd_replayable)
1252 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1253 client_nid = &req->rq_peer.nid;
1255 if (export == NULL) {
1256 /* allow lightweight connections during recovery */
1257 /* allow "new" MDT to be connected during recovery, since we
1258 * need retrieve recovery update records from it */
1259 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1261 int c; /* connected */
1262 int i; /* in progress */
1264 int s; /* stale/evicted */
1266 c = atomic_read(&target->obd_connected_clients);
1267 i = atomic_read(&target->obd_lock_replay_clients);
1268 k = target->obd_max_recoverable_clients;
1269 s = target->obd_stale_clients;
1270 t = target->obd_recovery_timer.expires;
1271 t = cfs_time_sub(t, cfs_time_current());
1272 t = cfs_duration_sec(t);
1273 LCONSOLE_WARN("%s: Denying connection for new client %s"
1274 "(at %s), waiting for %d known clients "
1275 "(%d recovered, %d in progress, and %d "
1276 "evicted) to recover in %d:%.02d\n",
1277 target->obd_name, cluuid.uuid,
1278 libcfs_nid2str(req->rq_peer.nid), k,
1279 c - i, i, s, (int)t / 60,
1284 rc = obd_connect(req->rq_svc_thread->t_env,
1285 &export, target, &cluuid, data,
1287 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1288 lustre_msg_add_op_flags(req->rq_repmsg,
1289 MSG_CONNECT_RECOVERING);
1291 conn.cookie = export->exp_handle.h_cookie;
1292 rc = rev_import_init(export);
1296 new_mds_mds_conn = true;
1299 rc = obd_reconnect(req->rq_svc_thread->t_env,
1300 export, target, &cluuid, data, client_nid);
1305 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1306 data->ocd_instance = target->u.obt.obt_instance;
1308 /* Return only the parts of obd_connect_data that we understand, so the
1309 * client knows that we don't understand the rest. */
1311 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1313 tmpdata = req_capsule_server_get(&req->rq_pill,
1315 /* Don't use struct assignment here, because the client reply
1316 * buffer may be smaller/larger than the local struct
1317 * obd_connect_data. */
1318 memcpy(tmpdata, data, min(tmpsize, size));
1321 /* If the client and the server are the same node, we will already
1322 * have an export that really points to the client's DLM export,
1323 * because we have a shared handles table.
1325 * XXX this will go away when shaver stops sending the "connect" handle
1326 * in the real "remote handle" field of the request --phik 24 Apr 2003
1328 ptlrpc_request_change_export(req, export);
1330 spin_lock(&export->exp_lock);
1331 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1332 spin_unlock(&export->exp_lock);
1333 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1334 "or equal conn_cnt: %d >= %d\n",
1335 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1336 export->exp_conn_cnt,
1337 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1339 GOTO(out, rc = -EALREADY);
1341 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1342 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1344 /* Don't evict liblustre clients for not pinging. */
1345 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1346 export->exp_libclient = 1;
1347 spin_unlock(&export->exp_lock);
1349 spin_lock(&target->obd_dev_lock);
1350 list_del_init(&export->exp_obd_chain_timed);
1351 spin_unlock(&target->obd_dev_lock);
1353 spin_unlock(&export->exp_lock);
1356 if (export->exp_connection != NULL) {
1357 /* Check to see if connection came from another NID. */
1358 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1359 !hlist_unhashed(&export->exp_nid_hash))
1360 cfs_hash_del(export->exp_obd->obd_nid_hash,
1361 &export->exp_connection->c_peer.nid,
1362 &export->exp_nid_hash);
1364 ptlrpc_connection_put(export->exp_connection);
1367 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1370 if (hlist_unhashed(&export->exp_nid_hash))
1371 cfs_hash_add(export->exp_obd->obd_nid_hash,
1372 &export->exp_connection->c_peer.nid,
1373 &export->exp_nid_hash);
1375 lustre_msg_set_handle(req->rq_repmsg, &conn);
1377 rc = rev_import_reconnect(export, req);
1381 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1383 __u64 transno = data->ocd_transno;
1385 spin_lock(&export->exp_lock);
1386 /* possible race with class_disconnect_stale_exports,
1387 * export may be already in the eviction process */
1388 if (export->exp_failed) {
1389 spin_unlock(&export->exp_lock);
1390 GOTO(out, rc = -ENODEV);
1392 export->exp_in_recovery = 1;
1393 export->exp_req_replay_needed = 1;
1394 export->exp_lock_replay_needed = 1;
1395 spin_unlock(&export->exp_lock);
1397 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1398 MSG_CONNECT_TRANSNO);
1399 if (has_transno && transno == 0)
1400 CWARN("Connect with zero transno!\n");
1402 if (has_transno && transno > 0 &&
1403 transno < target->obd_next_recovery_transno &&
1404 transno > target->obd_last_committed) {
1405 /* Another way is to use cmpxchg() to be lock-free. */
1406 spin_lock(&target->obd_recovery_task_lock);
1407 if (transno < target->obd_next_recovery_transno)
1408 target->obd_next_recovery_transno = transno;
1409 spin_unlock(&target->obd_recovery_task_lock);
1412 atomic_inc(&target->obd_req_replay_clients);
1413 atomic_inc(&target->obd_lock_replay_clients);
1414 /* Note: MDS-MDS connection is allowed to be connected during
1415 * recovery, no matter if the exports needs to be recoveried.
1416 * Because we need retrieve updates logs from all other MDTs.
1417 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1418 * also needs to be increased to match other recovery checking
1420 if (new_mds_mds_conn)
1421 target->obd_max_recoverable_clients++;
1422 if (atomic_inc_return(&target->obd_connected_clients) ==
1423 target->obd_max_recoverable_clients)
1424 wake_up(&target->obd_next_transno_waitq);
1427 /* Tell the client we're in recovery, when client is involved in it. */
1428 if (target->obd_recovering && !lw_client)
1429 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1433 spin_lock(&export->exp_lock);
1434 export->exp_connecting = 0;
1435 spin_unlock(&export->exp_lock);
1437 class_export_put(export);
1439 if (target_referenced == true && target != NULL) {
1440 spin_lock(&target->obd_dev_lock);
1441 target->obd_conn_inprogress--;
1442 spin_unlock(&target->obd_dev_lock);
1444 class_decref(target, __func__, current);
1446 req->rq_status = rc;
1450 int target_handle_disconnect(struct ptlrpc_request *req)
1455 rc = req_capsule_server_pack(&req->rq_pill);
1459 /* Keep the rq_export around so we can send the reply. */
1460 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1465 void target_destroy_export(struct obd_export *exp)
1467 struct obd_import *imp = NULL;
1468 /* exports created from last_rcvd data, and "fake"
1469 exports created by lctl don't have an import */
1470 spin_lock(&exp->exp_lock);
1471 if (exp->exp_imp_reverse != NULL) {
1472 imp = exp->exp_imp_reverse;
1473 exp->exp_imp_reverse = NULL;
1475 spin_unlock(&exp->exp_lock);
1477 client_destroy_import(imp);
1479 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1480 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1481 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1482 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1484 EXPORT_SYMBOL(target_destroy_export);
1487 * Recovery functions
1489 static void target_request_copy_get(struct ptlrpc_request *req)
1491 class_export_rpc_inc(req->rq_export);
1492 LASSERT(list_empty(&req->rq_list));
1493 INIT_LIST_HEAD(&req->rq_replay_list);
1495 /* Increase refcount to keep request in queue. */
1496 atomic_inc(&req->rq_refcount);
1497 /* Let export know it has replays to be handled. */
1498 atomic_inc(&req->rq_export->exp_replay_count);
1501 static void target_request_copy_put(struct ptlrpc_request *req)
1503 LASSERT(list_empty(&req->rq_replay_list));
1504 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1506 atomic_dec(&req->rq_export->exp_replay_count);
1507 class_export_rpc_dec(req->rq_export);
1508 ptlrpc_server_drop_request(req);
1511 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1513 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1514 struct obd_export *exp = req->rq_export;
1515 struct ptlrpc_request *reqiter;
1516 struct ptlrpc_request *dup_req = NULL;
1521 spin_lock(&exp->exp_lock);
1522 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1524 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1532 /* We expect it with RESENT and REPLAY flags. */
1533 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1534 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1535 CERROR("invalid flags %x of resent replay\n",
1536 lustre_msg_get_flags(req->rq_reqmsg));
1538 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1541 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1543 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1544 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1548 list_add_tail(&req->rq_replay_list,
1549 &exp->exp_req_replay_queue);
1552 spin_unlock(&exp->exp_lock);
1556 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1558 LASSERT(!list_empty(&req->rq_replay_list));
1559 LASSERT(req->rq_export);
1561 spin_lock(&req->rq_export->exp_lock);
1562 list_del_init(&req->rq_replay_list);
1563 spin_unlock(&req->rq_export->exp_lock);
1566 static void target_finish_recovery(struct lu_target *lut)
1568 struct obd_device *obd = lut->lut_obd;
1571 /* Only log a recovery message when recovery has occurred. */
1572 if (obd->obd_recovery_start) {
1573 time64_t now = ktime_get_real_seconds();
1574 time64_t elapsed_time;
1576 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1577 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1578 "%d recovered and %d %s evicted.\n", obd->obd_name,
1579 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1580 obd->obd_max_recoverable_clients,
1581 atomic_read(&obd->obd_connected_clients),
1582 obd->obd_stale_clients,
1583 obd->obd_stale_clients == 1 ? "was" : "were");
1586 ldlm_reprocess_recovery_done(obd->obd_namespace);
1587 spin_lock(&obd->obd_recovery_task_lock);
1588 if (!list_empty(&obd->obd_req_replay_queue) ||
1589 !list_empty(&obd->obd_lock_replay_queue) ||
1590 !list_empty(&obd->obd_final_req_queue)) {
1591 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1593 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1594 list_empty(&obd->obd_lock_replay_queue) ? \
1596 list_empty(&obd->obd_final_req_queue) ? \
1598 spin_unlock(&obd->obd_recovery_task_lock);
1601 spin_unlock(&obd->obd_recovery_task_lock);
1603 obd->obd_recovery_end = ktime_get_real_seconds();
1605 /* When recovery finished, cleanup orphans on MDS and OST. */
1606 if (OBT(obd) && OBP(obd, postrecov)) {
1607 int rc = OBP(obd, postrecov)(obd);
1609 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1615 static void abort_req_replay_queue(struct obd_device *obd)
1617 struct ptlrpc_request *req, *n;
1618 struct list_head abort_list;
1620 INIT_LIST_HEAD(&abort_list);
1621 spin_lock(&obd->obd_recovery_task_lock);
1622 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1623 spin_unlock(&obd->obd_recovery_task_lock);
1624 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1625 DEBUG_REQ(D_WARNING, req, "aborted:");
1626 req->rq_status = -ENOTCONN;
1627 if (ptlrpc_error(req)) {
1628 DEBUG_REQ(D_ERROR, req,
1629 "failed abort_req_reply; skipping");
1631 target_exp_dequeue_req_replay(req);
1632 target_request_copy_put(req);
1636 static void abort_lock_replay_queue(struct obd_device *obd)
1638 struct ptlrpc_request *req, *n;
1639 struct list_head abort_list;
1641 INIT_LIST_HEAD(&abort_list);
1642 spin_lock(&obd->obd_recovery_task_lock);
1643 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1644 spin_unlock(&obd->obd_recovery_task_lock);
1645 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1646 DEBUG_REQ(D_ERROR, req, "aborted:");
1647 req->rq_status = -ENOTCONN;
1648 if (ptlrpc_error(req)) {
1649 DEBUG_REQ(D_ERROR, req,
1650 "failed abort_lock_reply; skipping");
1652 target_request_copy_put(req);
1656 /* Called from a cleanup function if the device is being cleaned up
1657 forcefully. The exports should all have been disconnected already,
1658 the only thing left to do is
1659 - clear the recovery flags
1661 - free queued requests and replies, but don't send replies
1662 Because the obd_stopping flag is set, no new requests should be received.
1665 void target_cleanup_recovery(struct obd_device *obd)
1667 struct ptlrpc_request *req, *n;
1668 struct list_head clean_list;
1670 INIT_LIST_HEAD(&clean_list);
1671 spin_lock(&obd->obd_dev_lock);
1672 if (!obd->obd_recovering) {
1673 spin_unlock(&obd->obd_dev_lock);
1677 obd->obd_recovering = obd->obd_abort_recovery = 0;
1678 spin_unlock(&obd->obd_dev_lock);
1680 spin_lock(&obd->obd_recovery_task_lock);
1681 target_cancel_recovery_timer(obd);
1682 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1683 spin_unlock(&obd->obd_recovery_task_lock);
1685 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1686 LASSERT(req->rq_reply_state == NULL);
1687 target_exp_dequeue_req_replay(req);
1688 target_request_copy_put(req);
1691 spin_lock(&obd->obd_recovery_task_lock);
1692 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1693 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1694 spin_unlock(&obd->obd_recovery_task_lock);
1696 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1697 LASSERT(req->rq_reply_state == NULL);
1698 target_request_copy_put(req);
1703 EXPORT_SYMBOL(target_cleanup_recovery);
1705 /* obd_recovery_task_lock should be held */
1706 void target_cancel_recovery_timer(struct obd_device *obd)
1708 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1709 del_timer(&obd->obd_recovery_timer);
1712 static void target_start_recovery_timer(struct obd_device *obd)
1714 if (obd->obd_recovery_start != 0)
1717 spin_lock(&obd->obd_dev_lock);
1718 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1719 spin_unlock(&obd->obd_dev_lock);
1723 LASSERT(obd->obd_recovery_timeout != 0);
1725 if (obd->obd_recovery_start != 0) {
1726 spin_unlock(&obd->obd_dev_lock);
1730 mod_timer(&obd->obd_recovery_timer,
1731 cfs_time_shift(obd->obd_recovery_timeout));
1732 obd->obd_recovery_start = ktime_get_real_seconds();
1733 spin_unlock(&obd->obd_dev_lock);
1735 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1737 obd->obd_recovery_timeout / 60,
1738 obd->obd_recovery_timeout % 60,
1739 obd->obd_max_recoverable_clients,
1740 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1741 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1745 * extend recovery window.
1747 * if @extend is true, extend recovery window to have @drt remaining at least;
1748 * otherwise, make sure the recovery timeout value is not less than @drt.
1750 static void extend_recovery_timer(struct obd_device *obd, int drt,
1758 spin_lock(&obd->obd_dev_lock);
1759 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1760 spin_unlock(&obd->obd_dev_lock);
1763 LASSERT(obd->obd_recovery_start != 0);
1765 now = ktime_get_real_seconds();
1766 to = obd->obd_recovery_timeout;
1767 end = obd->obd_recovery_start + to;
1770 if (extend && (drt > left)) {
1772 } else if (!extend && (drt > to)) {
1776 if (to > obd->obd_recovery_time_hard) {
1777 to = obd->obd_recovery_time_hard;
1778 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1779 obd->obd_name, to, extend);
1782 if (obd->obd_recovery_timeout < to) {
1783 obd->obd_recovery_timeout = to;
1784 end = obd->obd_recovery_start + to;
1785 mod_timer(&obd->obd_recovery_timer,
1786 cfs_time_shift(end - now));
1788 spin_unlock(&obd->obd_dev_lock);
1790 CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1791 obd->obd_name, (s64)(end - now));
1794 /* Reset the timer with each new client connection */
1796 * This timer is actually reconnect_timer, which is for making sure
1797 * the total recovery window is at least as big as my reconnect
1798 * attempt timing. So the initial recovery time_out will be set to
1799 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1800 * from client is bigger than this, then the recovery time_out will
1801 * be extended to make sure the client could be reconnected, in the
1802 * process, the timeout from the new client should be ignored.
1806 check_and_start_recovery_timer(struct obd_device *obd,
1807 struct ptlrpc_request *req,
1810 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1811 struct obd_device_target *obt = &obd->u.obt;
1813 if (!new_client && service_time)
1814 /* Teach server about old server's estimates, as first guess
1815 * at how long new requests will take. */
1816 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1819 target_start_recovery_timer(obd);
1821 /* Convert the service time to RPC timeout,
1822 * and reuse service_time to limit stack usage. */
1823 service_time = at_est2timeout(service_time);
1825 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1826 service_time < at_extra)
1827 service_time = at_extra;
1829 /* We expect other clients to timeout within service_time, then try
1830 * to reconnect, then try the failover server. The max delay between
1831 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1832 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1834 LASSERT(obt->obt_magic == OBT_MAGIC);
1835 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1836 if (service_time > obd->obd_recovery_timeout && !new_client)
1837 extend_recovery_timer(obd, service_time, false);
1840 /** Health checking routines */
1841 static inline int exp_connect_healthy(struct obd_export *exp)
1843 return (exp->exp_in_recovery);
1846 /** if export done req_replay or has replay in queue */
1847 static inline int exp_req_replay_healthy(struct obd_export *exp)
1849 return (!exp->exp_req_replay_needed ||
1850 atomic_read(&exp->exp_replay_count) > 0);
1854 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1856 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1857 exp_req_replay_healthy(exp);
1860 /** if export done lock_replay or has replay in queue */
1861 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1863 return (!exp->exp_lock_replay_needed ||
1864 atomic_read(&exp->exp_replay_count) > 0);
1867 static inline int exp_vbr_healthy(struct obd_export *exp)
1869 return (!exp->exp_vbr_failed);
1872 static inline int exp_finished(struct obd_export *exp)
1874 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1877 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1879 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1883 static int check_for_next_transno(struct lu_target *lut)
1885 struct ptlrpc_request *req = NULL;
1886 struct obd_device *obd = lut->lut_obd;
1887 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1888 int wake_up = 0, connected, completed, queue_len;
1889 __u64 req_transno = 0;
1890 __u64 update_transno = 0;
1891 __u64 next_transno = 0;
1894 spin_lock(&obd->obd_recovery_task_lock);
1895 if (!list_empty(&obd->obd_req_replay_queue)) {
1896 req = list_entry(obd->obd_req_replay_queue.next,
1897 struct ptlrpc_request, rq_list);
1898 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1902 update_transno = distribute_txn_get_next_transno(tdtd);
1904 connected = atomic_read(&obd->obd_connected_clients);
1905 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1906 queue_len = obd->obd_requests_queued_for_recovery;
1907 next_transno = obd->obd_next_recovery_transno;
1909 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1910 "req_transno: %llu, next_transno: %llu\n",
1911 obd->obd_max_recoverable_clients, connected, completed,
1912 queue_len, req_transno, next_transno);
1914 if (obd->obd_abort_recovery) {
1915 CDEBUG(D_HA, "waking for aborted recovery\n");
1917 } else if (obd->obd_recovery_expired) {
1918 CDEBUG(D_HA, "waking for expired recovery\n");
1920 } else if (tdtd != NULL && req != NULL &&
1921 is_req_replayed_by_update(req)) {
1922 LASSERTF(req_transno < next_transno, "req_transno %llu"
1923 "next_transno%llu\n", req_transno, next_transno);
1924 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1927 } else if (req_transno == next_transno ||
1928 (update_transno != 0 && update_transno <= next_transno)) {
1929 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1931 } else if (queue_len > 0 &&
1932 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1933 /** handle gaps occured due to lost reply or VBR */
1934 LASSERTF(req_transno >= next_transno,
1935 "req_transno: %llu, next_transno: %llu\n",
1936 req_transno, next_transno);
1938 "%s: waking for gap in transno, VBR is %s (skip: "
1939 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1940 ", next_update %lld last_committed: %lld)\n",
1941 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1942 next_transno, queue_len, completed, connected,
1943 req_transno, update_transno, obd->obd_last_committed);
1944 obd->obd_next_recovery_transno = req_transno;
1946 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1947 CDEBUG(D_HA, "waking for completed recovery\n");
1949 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1950 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1951 " by fail_lock, waking up (%lld)\n", next_transno);
1952 obd->obd_next_recovery_transno = req_transno;
1955 spin_unlock(&obd->obd_recovery_task_lock);
1959 static int check_for_next_lock(struct lu_target *lut)
1961 struct obd_device *obd = lut->lut_obd;
1964 spin_lock(&obd->obd_recovery_task_lock);
1965 if (!list_empty(&obd->obd_lock_replay_queue)) {
1966 CDEBUG(D_HA, "waking for next lock\n");
1968 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1969 CDEBUG(D_HA, "waking for completed lock replay\n");
1971 } else if (obd->obd_abort_recovery) {
1972 CDEBUG(D_HA, "waking for aborted recovery\n");
1974 } else if (obd->obd_recovery_expired) {
1975 CDEBUG(D_HA, "waking for expired recovery\n");
1978 spin_unlock(&obd->obd_recovery_task_lock);
1984 * wait for recovery events,
1985 * check its status with help of check_routine
1986 * evict dead clients via health_check
1988 static int target_recovery_overseer(struct lu_target *lut,
1989 int (*check_routine)(struct lu_target *),
1990 int (*health_check)(struct obd_export *))
1992 struct obd_device *obd = lut->lut_obd;
1993 struct target_distribute_txn_data *tdtd;
1995 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
1996 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
1997 __u64 next_update_transno = 0;
1999 /* Only abort the recovery if there are no update recovery
2000 * left in the queue */
2001 spin_lock(&obd->obd_recovery_task_lock);
2002 if (lut->lut_tdtd != NULL) {
2003 next_update_transno =
2004 distribute_txn_get_next_transno(lut->lut_tdtd);
2006 tdtd = lut->lut_tdtd;
2007 /* If next_update_transno == 0, it probably because
2008 * updatelog retrieve threads did not get any records
2009 * yet, let's wait those threads stopped */
2010 if (next_update_transno == 0) {
2011 struct l_wait_info lwi = { 0 };
2013 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2015 &tdtd->tdtd_recovery_threads_count) == 0,
2018 next_update_transno =
2019 distribute_txn_get_next_transno(
2024 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2025 obd->obd_next_recovery_transno = next_update_transno;
2026 spin_unlock(&obd->obd_recovery_task_lock);
2027 /* Disconnect unfinished exports from clients, and
2028 * keep connection from MDT to make sure the update
2029 * recovery will still keep trying until some one
2030 * manually abort the recovery */
2031 class_disconnect_stale_exports(obd,
2032 exp_finished_or_from_mdt);
2033 /* Abort all of replay and replay lock req from
2035 abort_req_replay_queue(obd);
2036 abort_lock_replay_queue(obd);
2037 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2038 ")in the queue.\n", obd->obd_name,
2039 next_update_transno);
2041 obd->obd_abort_recovery = 1;
2042 spin_unlock(&obd->obd_recovery_task_lock);
2043 CWARN("%s recovery is aborted by hard timeout\n",
2048 while (wait_event_timeout(obd->obd_next_transno_waitq,
2050 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2051 /* wait indefinitely for event, but don't trigger watchdog */;
2053 if (obd->obd_abort_recovery) {
2054 CWARN("recovery is aborted, evict exports in recovery\n");
2055 if (lut->lut_tdtd != NULL) {
2056 struct l_wait_info lwi = { 0 };
2058 tdtd = lut->lut_tdtd;
2059 /* Let's wait all of the update log recovery thread
2061 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2062 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2064 /* Then abort the update recovery list */
2065 dtrq_list_destroy(lut->lut_tdtd);
2068 /** evict exports which didn't finish recovery yet */
2069 class_disconnect_stale_exports(obd, exp_finished);
2071 } else if (obd->obd_recovery_expired) {
2072 obd->obd_recovery_expired = 0;
2073 /** If some clients died being recovered, evict them */
2074 LCONSOLE_WARN("%s: recovery is timed out, "
2075 "evict stale exports\n", obd->obd_name);
2076 /** evict cexports with no replay in queue, they are stalled */
2077 class_disconnect_stale_exports(obd, health_check);
2079 /** continue with VBR */
2080 spin_lock(&obd->obd_dev_lock);
2081 obd->obd_version_recov = 1;
2082 spin_unlock(&obd->obd_dev_lock);
2084 * reset timer, recovery will proceed with versions now,
2085 * timeout is set just to handle reconnection delays
2087 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2088 /** Wait for recovery events again, after evicting bad clients */
2094 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2096 struct obd_device *obd = lut->lut_obd;
2097 struct ptlrpc_request *req = NULL;
2099 CDEBUG(D_HA, "Waiting for lock\n");
2100 if (target_recovery_overseer(lut, check_for_next_lock,
2101 exp_lock_replay_healthy))
2102 abort_lock_replay_queue(obd);
2104 spin_lock(&obd->obd_recovery_task_lock);
2105 if (!list_empty(&obd->obd_lock_replay_queue)) {
2106 req = list_entry(obd->obd_lock_replay_queue.next,
2107 struct ptlrpc_request, rq_list);
2108 list_del_init(&req->rq_list);
2109 spin_unlock(&obd->obd_recovery_task_lock);
2111 spin_unlock(&obd->obd_recovery_task_lock);
2112 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2113 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2114 /** evict exports failed VBR */
2115 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2120 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2122 struct ptlrpc_request *req = NULL;
2124 spin_lock(&obd->obd_recovery_task_lock);
2125 if (!list_empty(&obd->obd_final_req_queue)) {
2126 req = list_entry(obd->obd_final_req_queue.next,
2127 struct ptlrpc_request, rq_list);
2128 list_del_init(&req->rq_list);
2129 spin_unlock(&obd->obd_recovery_task_lock);
2130 if (req->rq_export->exp_in_recovery) {
2131 spin_lock(&req->rq_export->exp_lock);
2132 req->rq_export->exp_in_recovery = 0;
2133 spin_unlock(&req->rq_export->exp_lock);
2136 spin_unlock(&obd->obd_recovery_task_lock);
2141 static void handle_recovery_req(struct ptlrpc_thread *thread,
2142 struct ptlrpc_request *req,
2143 svc_handler_t handler)
2148 * export can be evicted during recovery, no need to handle replays for
2149 * it after that, discard such request silently
2151 if (req->rq_export->exp_disconnected)
2154 req->rq_session.lc_thread = thread;
2155 req->rq_svc_thread = thread;
2156 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2158 /* thread context */
2159 lu_context_enter(&thread->t_env->le_ctx);
2161 lu_context_exit(&thread->t_env->le_ctx);
2163 /* don't reset timer for final stage */
2164 if (!exp_finished(req->rq_export)) {
2165 int to = obd_timeout;
2168 * Add request timeout to the recovery time so next request from
2169 * this client may come in recovery time
2172 struct ptlrpc_service_part *svcpt;
2174 svcpt = req->rq_rqbd->rqbd_svcpt;
2175 /* If the server sent early reply for this request,
2176 * the client will recalculate the timeout according to
2177 * current server estimate service time, so we will
2178 * use the maxium timeout here for waiting the client
2179 * sending the next req */
2180 to = max((int)at_est2timeout(
2181 at_get(&svcpt->scp_at_estimate)),
2182 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2183 /* Add 2 net_latency, one for balance rq_deadline
2184 * (see ptl_send_rpc), one for resend the req to server,
2185 * Note: client will pack net_latency in replay req
2186 * (see ptlrpc_replay_req) */
2187 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2189 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2194 /** Checking routines for recovery */
2195 static int check_for_recovery_ready(struct lu_target *lut)
2197 struct obd_device *obd = lut->lut_obd;
2198 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2200 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2201 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2202 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2203 obd->obd_recovery_expired);
2205 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2206 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2207 if (clnts + obd->obd_stale_clients <
2208 obd->obd_max_recoverable_clients)
2212 if (lut->lut_tdtd != NULL) {
2213 if (!lut->lut_tdtd->tdtd_replay_ready &&
2214 !obd->obd_abort_recovery) {
2215 /* Let's extend recovery timer, in case the recovery
2216 * timer expired, and some clients got evicted */
2217 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2219 CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2220 obd->obd_name, obd->obd_recovery_timeout);
2229 REQUEST_RECOVERY = 1,
2230 UPDATE_RECOVERY = 2,
2233 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2237 if (!list_empty(&obd->obd_req_replay_queue)) {
2238 struct ptlrpc_request *req;
2240 req = list_entry(obd->obd_req_replay_queue.next,
2241 struct ptlrpc_request, rq_list);
2242 transno = lustre_msg_get_transno(req->rq_reqmsg);
2248 static __u64 get_next_transno(struct lu_target *lut, int *type)
2250 struct obd_device *obd = lut->lut_obd;
2251 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2253 __u64 update_transno;
2256 transno = get_next_replay_req_transno(obd);
2258 *type = REQUEST_RECOVERY;
2263 update_transno = distribute_txn_get_next_transno(tdtd);
2264 if (transno == 0 || (transno >= update_transno &&
2265 update_transno != 0)) {
2266 transno = update_transno;
2268 *type = UPDATE_RECOVERY;
2275 * drop duplicate replay request
2277 * Because the operation has been replayed by update recovery, the request
2278 * with the same transno will be dropped and also notify the client to send
2279 * next replay request.
2281 * \param[in] env execution environment
2282 * \param[in] obd failover obd device
2283 * \param[in] req request to be dropped
2285 static void drop_duplicate_replay_req(struct lu_env *env,
2286 struct obd_device *obd,
2287 struct ptlrpc_request *req)
2289 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2290 " update records are found.\n",
2291 lustre_msg_get_transno(req->rq_reqmsg),
2292 libcfs_nid2str(req->rq_peer.nid));
2294 /* Right now, only for MDS reint operation update replay and
2295 * normal request replay can have the same transno */
2296 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2297 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2298 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2299 if (likely(req->rq_export))
2300 target_committed_to_req(req);
2301 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2302 target_send_reply(req, req->rq_status, 0);
2304 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2305 libcfs_nid2str(req->rq_peer.nid));
2307 target_exp_dequeue_req_replay(req);
2308 target_request_copy_put(req);
2309 obd->obd_replayed_requests++;
2312 static void replay_request_or_update(struct lu_env *env,
2313 struct lu_target *lut,
2314 struct target_recovery_data *trd,
2315 struct ptlrpc_thread *thread)
2317 struct obd_device *obd = lut->lut_obd;
2318 struct ptlrpc_request *req = NULL;
2323 CDEBUG(D_HA, "Waiting for transno %lld\n",
2324 obd->obd_next_recovery_transno);
2326 /* Replay all of request and update by transno */
2328 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2330 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2332 /** It is needed to extend recovery window above
2333 * recovery_time_soft. Extending is possible only in the
2334 * end of recovery window (see more details in
2335 * handle_recovery_req()).
2337 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2339 if (target_recovery_overseer(lut, check_for_next_transno,
2340 exp_req_replay_healthy_or_from_mdt)) {
2341 abort_req_replay_queue(obd);
2342 abort_lock_replay_queue(obd);
2346 spin_lock(&obd->obd_recovery_task_lock);
2347 transno = get_next_transno(lut, &type);
2348 if (type == REQUEST_RECOVERY && transno != 0) {
2349 /* Drop replay request from client side, if the
2350 * replay has been executed by update with the
2352 req = list_entry(obd->obd_req_replay_queue.next,
2353 struct ptlrpc_request, rq_list);
2355 list_del_init(&req->rq_list);
2356 obd->obd_requests_queued_for_recovery--;
2357 spin_unlock(&obd->obd_recovery_task_lock);
2359 /* Let's check if the request has been redone by
2361 if (is_req_replayed_by_update(req)) {
2362 struct distribute_txn_replay_req *dtrq;
2364 dtrq = distribute_txn_lookup_finish_list(tdtd,
2366 LASSERT(dtrq != NULL);
2367 spin_lock(&tdtd->tdtd_replay_list_lock);
2368 list_del_init(&dtrq->dtrq_list);
2369 spin_unlock(&tdtd->tdtd_replay_list_lock);
2372 drop_duplicate_replay_req(env, obd, req);
2377 LASSERT(trd->trd_processing_task == current_pid());
2378 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2379 lustre_msg_get_transno(req->rq_reqmsg),
2380 libcfs_nid2str(req->rq_peer.nid));
2382 handle_recovery_req(thread, req,
2383 trd->trd_recovery_handler);
2385 * bz18031: increase next_recovery_transno before
2386 * target_request_copy_put() will drop exp_rpc reference
2388 spin_lock(&obd->obd_recovery_task_lock);
2389 obd->obd_next_recovery_transno++;
2390 spin_unlock(&obd->obd_recovery_task_lock);
2391 target_exp_dequeue_req_replay(req);
2392 target_request_copy_put(req);
2393 obd->obd_replayed_requests++;
2394 } else if (type == UPDATE_RECOVERY && transno != 0) {
2395 struct distribute_txn_replay_req *dtrq;
2398 spin_unlock(&obd->obd_recovery_task_lock);
2400 LASSERT(tdtd != NULL);
2401 dtrq = distribute_txn_get_next_req(tdtd);
2402 lu_context_enter(&thread->t_env->le_ctx);
2403 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2404 lu_context_exit(&thread->t_env->le_ctx);
2405 extend_recovery_timer(obd, obd_timeout, true);
2407 if (rc == 0 && dtrq->dtrq_xid != 0) {
2408 CDEBUG(D_HA, "Move x%llu t%llu"
2409 " to finish list\n", dtrq->dtrq_xid,
2410 dtrq->dtrq_master_transno);
2412 /* Add it to the replay finish list */
2413 spin_lock(&tdtd->tdtd_replay_list_lock);
2414 list_add(&dtrq->dtrq_list,
2415 &tdtd->tdtd_replay_finish_list);
2416 spin_unlock(&tdtd->tdtd_replay_list_lock);
2418 spin_lock(&obd->obd_recovery_task_lock);
2419 if (transno == obd->obd_next_recovery_transno)
2420 obd->obd_next_recovery_transno++;
2422 obd->obd_next_recovery_transno)
2423 obd->obd_next_recovery_transno =
2425 spin_unlock(&obd->obd_recovery_task_lock);
2430 spin_unlock(&obd->obd_recovery_task_lock);
2432 LASSERT(list_empty(&obd->obd_req_replay_queue));
2433 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2434 /** evict exports failed VBR */
2435 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2441 static int target_recovery_thread(void *arg)
2443 struct lu_target *lut = arg;
2444 struct obd_device *obd = lut->lut_obd;
2445 struct ptlrpc_request *req;
2446 struct target_recovery_data *trd = &obd->obd_recovery_data;
2447 unsigned long delta;
2449 struct ptlrpc_thread *thread = NULL;
2453 unshare_fs_struct();
2454 OBD_ALLOC_PTR(thread);
2460 OBD_FREE_PTR(thread);
2464 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2466 OBD_FREE_PTR(thread);
2471 thread->t_env = env;
2472 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2473 env->le_ctx.lc_thread = thread;
2474 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2475 thread->t_watchdog = NULL;
2477 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2479 trd->trd_processing_task = current_pid();
2481 spin_lock(&obd->obd_dev_lock);
2482 obd->obd_recovering = 1;
2483 spin_unlock(&obd->obd_dev_lock);
2484 complete(&trd->trd_starting);
2486 /* first of all, we have to know the first transno to replay */
2487 if (target_recovery_overseer(lut, check_for_recovery_ready,
2488 exp_connect_healthy)) {
2489 abort_req_replay_queue(obd);
2490 abort_lock_replay_queue(obd);
2491 if (lut->lut_tdtd != NULL)
2492 dtrq_list_destroy(lut->lut_tdtd);
2495 /* next stage: replay requests or update */
2497 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2498 atomic_read(&obd->obd_req_replay_clients),
2499 obd->obd_next_recovery_transno);
2500 replay_request_or_update(env, lut, trd, thread);
2503 * The second stage: replay locks
2505 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2506 atomic_read(&obd->obd_lock_replay_clients));
2507 while ((req = target_next_replay_lock(lut))) {
2508 LASSERT(trd->trd_processing_task == current_pid());
2509 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2510 libcfs_nid2str(req->rq_peer.nid));
2511 handle_recovery_req(thread, req,
2512 trd->trd_recovery_handler);
2513 target_request_copy_put(req);
2514 obd->obd_replayed_locks++;
2518 * The third stage: reply on final pings, at this moment all clients
2519 * must have request in final queue
2521 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2522 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2523 /** Update server last boot epoch */
2524 tgt_boot_epoch_update(lut);
2525 /* We drop recoverying flag to forward all new requests
2526 * to regular mds_handle() since now */
2527 spin_lock(&obd->obd_dev_lock);
2528 obd->obd_recovering = obd->obd_abort_recovery = 0;
2529 spin_unlock(&obd->obd_dev_lock);
2530 spin_lock(&obd->obd_recovery_task_lock);
2531 target_cancel_recovery_timer(obd);
2532 spin_unlock(&obd->obd_recovery_task_lock);
2533 while ((req = target_next_final_ping(obd))) {
2534 LASSERT(trd->trd_processing_task == current_pid());
2535 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2536 libcfs_nid2str(req->rq_peer.nid));
2537 handle_recovery_req(thread, req,
2538 trd->trd_recovery_handler);
2539 /* Because the waiting client can not send ping to server,
2540 * so we need refresh the last_request_time, to avoid the
2541 * export is being evicted */
2542 ptlrpc_update_export_timer(req->rq_export, 0);
2543 target_request_copy_put(req);
2546 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2547 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2548 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2549 if (delta > OBD_RECOVERY_TIME_SOFT) {
2550 CWARN("too long recovery - read logs\n");
2551 libcfs_debug_dumplog();
2554 target_finish_recovery(lut);
2556 lu_context_fini(&env->le_ctx);
2557 trd->trd_processing_task = 0;
2558 complete(&trd->trd_finishing);
2560 tgt_io_thread_done(thread);
2561 OBD_FREE_PTR(thread);
2566 static int target_start_recovery_thread(struct lu_target *lut,
2567 svc_handler_t handler)
2569 struct obd_device *obd = lut->lut_obd;
2571 struct target_recovery_data *trd = &obd->obd_recovery_data;
2574 memset(trd, 0, sizeof(*trd));
2575 init_completion(&trd->trd_starting);
2576 init_completion(&trd->trd_finishing);
2577 trd->trd_recovery_handler = handler;
2579 rc = server_name2index(obd->obd_name, &index, NULL);
2583 if (!IS_ERR(kthread_run(target_recovery_thread,
2584 lut, "tgt_recover_%d", index))) {
2585 wait_for_completion(&trd->trd_starting);
2586 LASSERT(obd->obd_recovering != 0);
2594 void target_stop_recovery_thread(struct obd_device *obd)
2596 if (obd->obd_recovery_data.trd_processing_task > 0) {
2597 struct target_recovery_data *trd = &obd->obd_recovery_data;
2598 /** recovery can be done but postrecovery is not yet */
2599 spin_lock(&obd->obd_dev_lock);
2600 if (obd->obd_recovering) {
2601 CERROR("%s: Aborting recovery\n", obd->obd_name);
2602 obd->obd_abort_recovery = 1;
2603 wake_up(&obd->obd_next_transno_waitq);
2605 spin_unlock(&obd->obd_dev_lock);
2606 wait_for_completion(&trd->trd_finishing);
2609 EXPORT_SYMBOL(target_stop_recovery_thread);
2611 void target_recovery_fini(struct obd_device *obd)
2613 class_disconnect_exports(obd);
2614 target_stop_recovery_thread(obd);
2615 target_cleanup_recovery(obd);
2617 EXPORT_SYMBOL(target_recovery_fini);
2619 static void target_recovery_expired(unsigned long castmeharder)
2621 struct obd_device *obd = (struct obd_device *)castmeharder;
2622 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2623 " after %llus (%d clients connected)\n",
2624 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2625 (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2626 atomic_read(&obd->obd_connected_clients));
2628 obd->obd_recovery_expired = 1;
2629 wake_up(&obd->obd_next_transno_waitq);
2632 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2634 struct obd_device *obd = lut->lut_obd;
2636 if (lut->lut_bottom->dd_rdonly)
2639 if (obd->obd_max_recoverable_clients == 0) {
2640 /** Update server last boot epoch */
2641 tgt_boot_epoch_update(lut);
2645 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2646 "last_transno %llu\n", obd->obd_name,
2647 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2648 LASSERT(obd->obd_stopping == 0);
2649 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2650 obd->obd_recovery_start = 0;
2651 obd->obd_recovery_end = 0;
2653 setup_timer(&obd->obd_recovery_timer, target_recovery_expired,
2654 (unsigned long)obd);
2655 target_start_recovery_thread(lut, handler);
2657 EXPORT_SYMBOL(target_recovery_init);
2659 static int target_process_req_flags(struct obd_device *obd,
2660 struct ptlrpc_request *req)
2662 struct obd_export *exp = req->rq_export;
2663 LASSERT(exp != NULL);
2664 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2665 /* client declares he's ready to replay locks */
2666 spin_lock(&exp->exp_lock);
2667 if (exp->exp_req_replay_needed) {
2668 exp->exp_req_replay_needed = 0;
2669 spin_unlock(&exp->exp_lock);
2671 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2672 atomic_dec(&obd->obd_req_replay_clients);
2674 spin_unlock(&exp->exp_lock);
2677 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2678 /* client declares he's ready to complete recovery
2679 * so, we put the request on th final queue */
2680 spin_lock(&exp->exp_lock);
2681 if (exp->exp_lock_replay_needed) {
2682 exp->exp_lock_replay_needed = 0;
2683 spin_unlock(&exp->exp_lock);
2685 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2686 atomic_dec(&obd->obd_lock_replay_clients);
2688 spin_unlock(&exp->exp_lock);
2694 int target_queue_recovery_request(struct ptlrpc_request *req,
2695 struct obd_device *obd)
2697 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2698 struct ptlrpc_request *reqiter;
2702 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2703 /* Processing the queue right now, don't re-add. */
2707 target_process_req_flags(obd, req);
2709 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2710 /* client declares he's ready to complete recovery
2711 * so, we put the request on th final queue */
2712 target_request_copy_get(req);
2713 DEBUG_REQ(D_HA, req, "queue final req");
2714 wake_up(&obd->obd_next_transno_waitq);
2715 spin_lock(&obd->obd_recovery_task_lock);
2716 if (obd->obd_recovering) {
2717 struct ptlrpc_request *tmp;
2718 struct ptlrpc_request *duplicate = NULL;
2720 if (likely(!req->rq_export->exp_replay_done)) {
2721 req->rq_export->exp_replay_done = 1;
2722 list_add_tail(&req->rq_list,
2723 &obd->obd_final_req_queue);
2724 spin_unlock(&obd->obd_recovery_task_lock);
2728 /* XXX O(n), but only happens if final ping is
2729 * timed out, probably reorganize the list as
2730 * a hash list later */
2731 list_for_each_entry_safe(reqiter, tmp,
2732 &obd->obd_final_req_queue,
2734 if (reqiter->rq_export == req->rq_export) {
2735 list_del_init(&reqiter->rq_list);
2736 duplicate = reqiter;
2741 list_add_tail(&req->rq_list,
2742 &obd->obd_final_req_queue);
2743 req->rq_export->exp_replay_done = 1;
2744 spin_unlock(&obd->obd_recovery_task_lock);
2746 if (duplicate != NULL) {
2747 DEBUG_REQ(D_HA, duplicate,
2748 "put prev final req\n");
2749 target_request_copy_put(duplicate);
2753 spin_unlock(&obd->obd_recovery_task_lock);
2754 target_request_copy_put(req);
2755 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2758 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2759 /* client declares he's ready to replay locks */
2760 target_request_copy_get(req);
2761 DEBUG_REQ(D_HA, req, "queue lock replay req");
2762 wake_up(&obd->obd_next_transno_waitq);
2763 spin_lock(&obd->obd_recovery_task_lock);
2764 LASSERT(obd->obd_recovering);
2765 /* usually due to recovery abort */
2766 if (!req->rq_export->exp_in_recovery) {
2767 spin_unlock(&obd->obd_recovery_task_lock);
2768 target_request_copy_put(req);
2771 LASSERT(req->rq_export->exp_lock_replay_needed);
2772 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2773 spin_unlock(&obd->obd_recovery_task_lock);
2777 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2778 * (i.e. buflens etc are in my own byte order), but type-dependent
2779 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2782 INIT_LIST_HEAD(&req->rq_list);
2783 DEBUG_REQ(D_HA, req, "not queueing");
2787 /* If we're processing the queue, we want don't want to queue this
2790 * Also, if this request has a transno less than the one we're waiting
2791 * for, we should process it now. It could (and currently always will)
2792 * be an open request for a descriptor that was opened some time ago.
2794 * Also, a resent, replayed request that has already been
2795 * handled will pass through here and be processed immediately.
2797 CDEBUG(D_HA, "Next recovery transno: %llu"
2798 ", current: %llu, replaying\n",
2799 obd->obd_next_recovery_transno, transno);
2801 /* If the request has been replayed by update replay, then sends this
2802 * request to the recovery thread (replay_request_or_update()), where
2803 * it will be handled */
2804 spin_lock(&obd->obd_recovery_task_lock);
2805 if (transno < obd->obd_next_recovery_transno &&
2806 !is_req_replayed_by_update(req)) {
2807 /* Processing the queue right now, don't re-add. */
2808 LASSERT(list_empty(&req->rq_list));
2809 spin_unlock(&obd->obd_recovery_task_lock);
2812 spin_unlock(&obd->obd_recovery_task_lock);
2814 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2817 target_request_copy_get(req);
2818 if (!req->rq_export->exp_in_recovery) {
2819 target_request_copy_put(req);
2822 LASSERT(req->rq_export->exp_req_replay_needed);
2824 if (target_exp_enqueue_req_replay(req)) {
2825 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2826 target_request_copy_put(req);
2831 spin_lock(&obd->obd_recovery_task_lock);
2832 LASSERT(obd->obd_recovering);
2833 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2834 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2835 list_add_tail(&req->rq_list, &reqiter->rq_list);
2840 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2842 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2843 "has been claimed by another client");
2844 spin_unlock(&obd->obd_recovery_task_lock);
2845 target_exp_dequeue_req_replay(req);
2846 target_request_copy_put(req);
2852 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2854 obd->obd_requests_queued_for_recovery++;
2855 spin_unlock(&obd->obd_recovery_task_lock);
2856 wake_up(&obd->obd_next_transno_waitq);
2860 int target_handle_ping(struct ptlrpc_request *req)
2862 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2863 return req_capsule_server_pack(&req->rq_pill);
2866 void target_committed_to_req(struct ptlrpc_request *req)
2868 struct obd_export *exp = req->rq_export;
2870 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2871 lustre_msg_set_last_committed(req->rq_repmsg,
2872 exp->exp_last_committed);
2874 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2875 "%d)", exp->exp_obd->obd_no_transno,
2876 req->rq_repmsg == NULL);
2878 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2879 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2882 #endif /* HAVE_SERVER_SUPPORT */
2885 * Packs current SLV and Limit into \a req.
2887 int target_pack_pool_reply(struct ptlrpc_request *req)
2889 struct obd_device *obd;
2892 /* Check that we still have all structures alive as this may
2893 * be some late RPC at shutdown time. */
2894 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2895 !exp_connect_lru_resize(req->rq_export))) {
2896 lustre_msg_set_slv(req->rq_repmsg, 0);
2897 lustre_msg_set_limit(req->rq_repmsg, 0);
2901 /* OBD is alive here as export is alive, which we checked above. */
2902 obd = req->rq_export->exp_obd;
2904 read_lock(&obd->obd_pool_lock);
2905 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2906 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2907 read_unlock(&obd->obd_pool_lock);
2912 static int target_send_reply_msg(struct ptlrpc_request *req,
2913 int rc, int fail_id)
2915 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2916 DEBUG_REQ(D_ERROR, req, "dropping reply");
2919 /* We can have a null rq_reqmsg in the event of bad signature or
2920 * no context when unwrapping */
2921 if (req->rq_reqmsg &&
2922 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2923 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2927 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2928 req->rq_status = rc;
2929 return ptlrpc_send_error(req, 1);
2931 DEBUG_REQ(D_NET, req, "sending reply");
2934 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2937 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2939 struct ptlrpc_service_part *svcpt;
2941 struct ptlrpc_reply_state *rs;
2942 struct obd_export *exp;
2945 if (req->rq_no_reply) {
2950 svcpt = req->rq_rqbd->rqbd_svcpt;
2951 rs = req->rq_reply_state;
2952 if (rs == NULL || !rs->rs_difficult) {
2954 target_send_reply_msg (req, rc, fail_id);
2959 /* must be an export if locks saved */
2960 LASSERT(req->rq_export != NULL);
2961 /* req/reply consistent */
2962 LASSERT(rs->rs_svcpt == svcpt);
2965 LASSERT(!rs->rs_scheduled);
2966 LASSERT(!rs->rs_scheduled_ever);
2967 LASSERT(!rs->rs_handled);
2968 LASSERT(!rs->rs_on_net);
2969 LASSERT(rs->rs_export == NULL);
2970 LASSERT(list_empty(&rs->rs_obd_list));
2971 LASSERT(list_empty(&rs->rs_exp_list));
2973 exp = class_export_get(req->rq_export);
2975 /* disable reply scheduling while I'm setting up */
2976 rs->rs_scheduled = 1;
2978 rs->rs_xid = req->rq_xid;
2979 rs->rs_transno = req->rq_transno;
2980 rs->rs_export = exp;
2981 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2983 spin_lock(&exp->exp_uncommitted_replies_lock);
2984 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2985 rs->rs_transno, exp->exp_last_committed);
2986 if (rs->rs_transno > exp->exp_last_committed) {
2987 /* not committed already */
2988 list_add_tail(&rs->rs_obd_list,
2989 &exp->exp_uncommitted_replies);
2991 spin_unlock(&exp->exp_uncommitted_replies_lock);
2993 spin_lock(&exp->exp_lock);
2994 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2995 spin_unlock(&exp->exp_lock);
2997 netrc = target_send_reply_msg(req, rc, fail_id);
2999 spin_lock(&svcpt->scp_rep_lock);
3001 atomic_inc(&svcpt->scp_nreps_difficult);
3004 /* error sending: reply is off the net. Also we need +1
3005 * reply ref until ptlrpc_handle_rs() is done
3006 * with the reply state (if the send was successful, there
3007 * would have been +1 ref for the net, which
3008 * reply_out_callback leaves alone) */
3010 ptlrpc_rs_addref(rs);
3013 spin_lock(&rs->rs_lock);
3014 if (rs->rs_transno <= exp->exp_last_committed ||
3015 (!rs->rs_on_net && !rs->rs_no_ack) ||
3016 list_empty(&rs->rs_exp_list) || /* completed already */
3017 list_empty(&rs->rs_obd_list)) {
3018 CDEBUG(D_HA, "Schedule reply immediately\n");
3019 ptlrpc_dispatch_difficult_reply(rs);
3021 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3022 rs->rs_scheduled = 0; /* allow notifier to schedule */
3024 spin_unlock(&rs->rs_lock);
3025 spin_unlock(&svcpt->scp_rep_lock);
3029 enum ldlm_mode lck_compat_array[] = {
3030 [LCK_EX] = LCK_COMPAT_EX,
3031 [LCK_PW] = LCK_COMPAT_PW,
3032 [LCK_PR] = LCK_COMPAT_PR,
3033 [LCK_CW] = LCK_COMPAT_CW,
3034 [LCK_CR] = LCK_COMPAT_CR,
3035 [LCK_NL] = LCK_COMPAT_NL,
3036 [LCK_GROUP] = LCK_COMPAT_GROUP,
3037 [LCK_COS] = LCK_COMPAT_COS,
3041 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3042 * not escape to the user level.
3044 int ldlm_error2errno(enum ldlm_error error)
3050 case ELDLM_LOCK_MATCHED:
3053 case ELDLM_LOCK_CHANGED:
3056 case ELDLM_LOCK_ABORTED:
3059 case ELDLM_LOCK_REPLACED:
3062 case ELDLM_NO_LOCK_DATA:
3065 case ELDLM_NAMESPACE_EXISTS:
3068 case ELDLM_BAD_NAMESPACE:
3072 if (((int)error) < 0) { /* cast to signed type */
3073 result = error; /* as ldlm_error can be unsigned */
3075 CERROR("Invalid DLM result code: %d\n", error);
3081 EXPORT_SYMBOL(ldlm_error2errno);
3084 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3086 enum ldlm_error ldlm_errno2error(int err_no)
3095 error = ELDLM_LOCK_CHANGED;
3098 error = ELDLM_LOCK_ABORTED;
3101 error = ELDLM_LOCK_REPLACED;
3104 error = ELDLM_NO_LOCK_DATA;
3107 error = ELDLM_NAMESPACE_EXISTS;
3110 error = ELDLM_BAD_NAMESPACE;
3118 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3119 void ldlm_dump_export_locks(struct obd_export *exp)
3121 spin_lock(&exp->exp_locks_list_guard);
3122 if (!list_empty(&exp->exp_locks_list)) {
3123 struct ldlm_lock *lock;
3125 CERROR("dumping locks for export %p,"
3126 "ignore if the unmount doesn't hang\n", exp);
3127 list_for_each_entry(lock, &exp->exp_locks_list,
3129 LDLM_ERROR(lock, "lock:");
3131 spin_unlock(&exp->exp_locks_list_guard);
3135 #ifdef HAVE_SERVER_SUPPORT
3136 static int target_bulk_timeout(void *data)
3139 /* We don't fail the connection here, because having the export
3140 * killed makes the (vital) call to commitrw very sad.
3145 static inline const char *bulk2type(struct ptlrpc_request *req)
3147 if (req->rq_bulk_read)
3149 if (req->rq_bulk_write)
3154 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3155 struct l_wait_info *lwi)
3157 struct ptlrpc_request *req = desc->bd_req;
3158 time_t start = cfs_time_current_sec();
3164 /* If there is eviction in progress, wait for it to finish. */
3165 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3166 *lwi = LWI_INTR(NULL, NULL);
3167 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3168 !atomic_read(&exp->exp_obd->
3169 obd_evict_inprogress),
3173 /* Check if client was evicted or reconnected already. */
3174 if (exp->exp_failed ||
3175 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3178 if (req->rq_bulk_read)
3179 rc = sptlrpc_svc_wrap_bulk(req, desc);
3181 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3182 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3183 else /* old version, bulk matchbits is rq_xid */
3184 req->rq_mbits = req->rq_xid;
3187 rc = ptlrpc_start_bulk_transfer(desc);
3191 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3192 bulk2type(req), rc);
3196 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3197 ptlrpc_abort_bulk(desc);
3201 /* limit actual bulk transfer to bulk_timeout seconds */
3202 deadline = start + bulk_timeout;
3203 if (deadline > req->rq_deadline)
3204 deadline = req->rq_deadline;
3207 long timeoutl = deadline - cfs_time_current_sec();
3208 cfs_duration_t timeout = timeoutl <= 0 ?
3209 CFS_TICK : cfs_time_seconds(timeoutl);
3212 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3213 target_bulk_timeout, desc);
3214 rc = l_wait_event(desc->bd_waitq,
3215 !ptlrpc_server_bulk_active(desc) ||
3218 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3220 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3221 /* Wait again if we changed rq_deadline. */
3222 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3223 deadline = start + bulk_timeout;
3224 if (deadline > rq_deadline)
3225 deadline = rq_deadline;
3226 } while ((rc == -ETIMEDOUT) &&
3227 (deadline > cfs_time_current_sec()));
3229 if (rc == -ETIMEDOUT) {
3230 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3231 bulk2type(req), deadline - start,
3232 cfs_time_current_sec() - deadline);
3233 ptlrpc_abort_bulk(desc);
3234 } else if (exp->exp_failed) {
3235 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3238 ptlrpc_abort_bulk(desc);
3239 } else if (exp->exp_conn_cnt >
3240 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3241 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3243 /* We don't reply anyway. */
3245 ptlrpc_abort_bulk(desc);
3246 } else if (desc->bd_failure) {
3247 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3249 /* XXX should this be a different errno? */
3252 if (req->rq_bulk_write)
3253 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3254 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3255 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3256 bulk2type(req), desc->bd_nob_transferred,
3258 /* XXX should this be a different errno? */
3265 EXPORT_SYMBOL(target_bulk_io);
3267 #endif /* HAVE_SERVER_SUPPORT */