4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
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/jiffies.h>
43 #include <linux/kthread.h>
44 #include <libcfs/libcfs.h>
46 #include <obd_class.h>
47 #include <lustre_dlm.h>
48 #include <lustre_net.h>
49 #include <lustre_sec.h>
50 #include "ldlm_internal.h"
52 /* @priority: If non-zero, move the selected connection to the list head.
53 * @create: If zero, only search in existing connections.
55 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
56 int priority, int create)
58 struct ptlrpc_connection *ptlrpc_conn;
59 struct obd_import_conn *imp_conn = NULL, *item;
60 lnet_nid_t nid4refnet = LNET_NID_ANY;
64 if (!create && !priority) {
65 CDEBUG(D_HA, "Nothing to do\n");
69 if (imp->imp_connection &&
70 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
71 /* nid4refnet is used to restrict network connections */
72 nid4refnet = imp->imp_connection->c_self;
73 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
75 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
80 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
82 GOTO(out_put, rc = -ENOMEM);
85 spin_lock(&imp->imp_lock);
86 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
87 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
89 list_del(&item->oic_item);
90 list_add(&item->oic_item,
92 item->oic_last_attempt = 0;
94 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
95 imp, imp->imp_obd->obd_name, uuid->uuid,
96 (priority ? ", moved to head" : ""));
97 spin_unlock(&imp->imp_lock);
98 GOTO(out_free, rc = 0);
101 /* No existing import connection found for \a uuid. */
103 imp_conn->oic_conn = ptlrpc_conn;
104 imp_conn->oic_uuid = *uuid;
105 imp_conn->oic_last_attempt = 0;
107 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
109 list_add_tail(&imp_conn->oic_item,
110 &imp->imp_conn_list);
111 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
112 imp, imp->imp_obd->obd_name, uuid->uuid,
113 (priority ? "head" : "tail"));
115 spin_unlock(&imp->imp_lock);
116 GOTO(out_free, rc = -ENOENT);
119 spin_unlock(&imp->imp_lock);
123 OBD_FREE(imp_conn, sizeof(*imp_conn));
125 ptlrpc_connection_put(ptlrpc_conn);
129 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
131 return import_set_conn(imp, uuid, 1, 0);
134 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
137 return import_set_conn(imp, uuid, priority, 1);
139 EXPORT_SYMBOL(client_import_add_conn);
141 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
143 struct obd_import_conn *imp_conn;
144 struct obd_export *dlmexp;
148 spin_lock(&imp->imp_lock);
149 if (list_empty(&imp->imp_conn_list)) {
150 LASSERT(!imp->imp_connection);
154 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
155 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
157 LASSERT(imp_conn->oic_conn);
159 if (imp_conn == imp->imp_conn_current) {
160 LASSERT(imp_conn->oic_conn == imp->imp_connection);
162 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
163 imp->imp_state != LUSTRE_IMP_DISCON) {
164 CERROR("can't remove current connection\n");
165 GOTO(out, rc = -EBUSY);
168 ptlrpc_connection_put(imp->imp_connection);
169 imp->imp_connection = NULL;
171 dlmexp = class_conn2export(&imp->imp_dlm_handle);
172 if (dlmexp && dlmexp->exp_connection) {
173 LASSERT(dlmexp->exp_connection ==
175 ptlrpc_connection_put(dlmexp->exp_connection);
176 dlmexp->exp_connection = NULL;
180 class_export_put(dlmexp);
183 list_del(&imp_conn->oic_item);
184 ptlrpc_connection_put(imp_conn->oic_conn);
185 OBD_FREE(imp_conn, sizeof(*imp_conn));
186 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
187 imp, imp->imp_obd->obd_name, uuid->uuid);
192 spin_unlock(&imp->imp_lock);
194 CERROR("connection %s not found\n", uuid->uuid);
197 EXPORT_SYMBOL(client_import_del_conn);
200 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
201 * to find a conn uuid of \a imp which can reach \a peer.
203 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
204 struct obd_uuid *uuid)
206 struct obd_import_conn *conn;
210 spin_lock(&imp->imp_lock);
211 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
212 /* Check if conn UUID does have this peer NID. */
213 if (class_check_uuid(&conn->oic_uuid, peer)) {
214 *uuid = conn->oic_uuid;
219 spin_unlock(&imp->imp_lock);
222 EXPORT_SYMBOL(client_import_find_conn);
224 void client_destroy_import(struct obd_import *imp)
226 /* Drop security policy instance after all RPCs have finished/aborted
227 * to let all busy contexts be released. */
228 class_import_get(imp);
229 class_destroy_import(imp);
230 sptlrpc_import_sec_put(imp);
231 class_import_put(imp);
233 EXPORT_SYMBOL(client_destroy_import);
236 * Check whether or not the OSC is on MDT.
239 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
241 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
244 static int osc_on_mdt(char *obdname)
248 ptr = strrchr(obdname, '-');
252 if (strncmp(ptr + 1, "MDT", 3) == 0)
258 /* Configure an RPC client OBD device.
263 * 3 - inactive-on-startup
264 * 4 - restrictive net
266 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
268 struct client_obd *cli = &obddev->u.cli;
269 struct obd_import *imp;
270 struct obd_uuid server_uuid;
271 int rq_portal, rp_portal, connect_op;
272 char *name = obddev->obd_type->typ_name;
273 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
274 char *cli_name = lustre_cfg_buf(lcfg, 0);
275 struct ptlrpc_connection fake_conn = { .c_self = 0,
276 .c_remote_uuid.uuid[0] = 0 };
280 /* In a more perfect world, we would hang a ptlrpc_client off of
281 * obd_type and just use the values from there. */
282 if (!strcmp(name, LUSTRE_OSC_NAME)) {
283 rq_portal = OST_REQUEST_PORTAL;
284 rp_portal = OSC_REPLY_PORTAL;
285 connect_op = OST_CONNECT;
286 cli->cl_sp_me = LUSTRE_SP_CLI;
287 cli->cl_sp_to = LUSTRE_SP_OST;
288 ns_type = LDLM_NS_TYPE_OSC;
289 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
290 !strcmp(name, LUSTRE_LWP_NAME)) {
291 rq_portal = MDS_REQUEST_PORTAL;
292 rp_portal = MDC_REPLY_PORTAL;
293 connect_op = MDS_CONNECT;
294 if (is_lwp_on_ost(cli_name))
295 cli->cl_sp_me = LUSTRE_SP_OST;
296 else if (is_lwp_on_mdt(cli_name))
297 cli->cl_sp_me = LUSTRE_SP_MDT;
299 cli->cl_sp_me = LUSTRE_SP_CLI;
300 cli->cl_sp_to = LUSTRE_SP_MDT;
301 ns_type = LDLM_NS_TYPE_MDC;
302 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
303 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
304 /* OSP_on_MDT for other MDTs */
305 connect_op = MDS_CONNECT;
306 cli->cl_sp_to = LUSTRE_SP_MDT;
307 ns_type = LDLM_NS_TYPE_MDC;
308 rq_portal = OUT_PORTAL;
310 /* OSP on MDT for OST */
311 connect_op = OST_CONNECT;
312 cli->cl_sp_to = LUSTRE_SP_OST;
313 ns_type = LDLM_NS_TYPE_OSC;
314 rq_portal = OST_REQUEST_PORTAL;
316 rp_portal = OSC_REPLY_PORTAL;
317 cli->cl_sp_me = LUSTRE_SP_MDT;
318 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
319 rq_portal = MGS_REQUEST_PORTAL;
320 rp_portal = MGC_REPLY_PORTAL;
321 connect_op = MGS_CONNECT;
322 cli->cl_sp_me = LUSTRE_SP_MGC;
323 cli->cl_sp_to = LUSTRE_SP_MGS;
324 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
325 ns_type = LDLM_NS_TYPE_MGC;
327 CERROR("unknown client OBD type \"%s\", can't setup\n",
332 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
333 CERROR("requires a TARGET UUID\n");
337 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
338 CERROR("client UUID must be less than 38 characters\n");
342 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
343 CERROR("setup requires a SERVER UUID\n");
347 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
348 CERROR("target UUID must be less than 38 characters\n");
352 init_rwsem(&cli->cl_sem);
353 mutex_init(&cli->cl_mgc_mutex);
355 init_rwsem(&cli->cl_seq_rwsem);
356 cli->cl_conn_count = 0;
357 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
358 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
359 sizeof(server_uuid)));
361 cli->cl_dirty_pages = 0;
362 cli->cl_dirty_max_pages = 0;
363 cli->cl_avail_grant = 0;
364 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
365 /* cl_dirty_max_pages may be changed at connect time in
366 * ptlrpc_connect_interpret(). */
367 client_adjust_max_dirty(cli);
368 INIT_LIST_HEAD(&cli->cl_cache_waiters);
369 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
370 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
371 INIT_LIST_HEAD(&cli->cl_loi_write_list);
372 INIT_LIST_HEAD(&cli->cl_loi_read_list);
373 spin_lock_init(&cli->cl_loi_list_lock);
374 atomic_set(&cli->cl_pending_w_pages, 0);
375 atomic_set(&cli->cl_pending_r_pages, 0);
376 cli->cl_r_in_flight = 0;
377 cli->cl_w_in_flight = 0;
379 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
380 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
381 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
382 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
383 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
384 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
387 INIT_LIST_HEAD(&cli->cl_lru_osc);
388 atomic_set(&cli->cl_lru_shrinkers, 0);
389 atomic_long_set(&cli->cl_lru_busy, 0);
390 atomic_long_set(&cli->cl_lru_in_list, 0);
391 INIT_LIST_HEAD(&cli->cl_lru_list);
392 spin_lock_init(&cli->cl_lru_list_lock);
393 atomic_long_set(&cli->cl_unstable_count, 0);
394 INIT_LIST_HEAD(&cli->cl_shrink_list);
395 INIT_LIST_HEAD(&cli->cl_grant_chain);
397 INIT_LIST_HEAD(&cli->cl_flight_waiters);
398 cli->cl_rpcs_in_flight = 0;
400 init_waitqueue_head(&cli->cl_destroy_waitq);
401 atomic_set(&cli->cl_destroy_in_flight, 0);
403 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
404 #ifdef ENABLE_CHECKSUM
405 /* Turn on checksumming by default. */
406 cli->cl_checksum = 1;
408 * The supported checksum types will be worked out at connect time
409 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
412 cli->cl_cksum_type = cli->cl_supp_cksum_types;
414 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
416 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
417 * from OFD after connecting. */
418 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
420 cli->cl_max_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
422 /* set cl_chunkbits default value to PAGE_SHIFT,
423 * it will be updated at OSC connection time. */
424 cli->cl_chunkbits = PAGE_SHIFT;
426 if (!strcmp(name, LUSTRE_MDC_NAME)) {
427 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
428 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
429 cli->cl_max_rpcs_in_flight = 2;
430 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
431 cli->cl_max_rpcs_in_flight = 3;
432 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
433 cli->cl_max_rpcs_in_flight = 4;
435 if (osc_on_mdt(obddev->obd_name))
436 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
438 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
441 spin_lock_init(&cli->cl_mod_rpcs_lock);
442 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
443 cli->cl_max_mod_rpcs_in_flight = 0;
444 cli->cl_mod_rpcs_in_flight = 0;
445 cli->cl_close_rpcs_in_flight = 0;
446 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
447 cli->cl_mod_tag_bitmap = NULL;
449 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
451 if (connect_op == MDS_CONNECT) {
452 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
453 OBD_ALLOC(cli->cl_mod_tag_bitmap,
454 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
455 if (cli->cl_mod_tag_bitmap == NULL)
456 GOTO(err, rc = -ENOMEM);
461 CERROR("ldlm_get_ref failed: %d\n", rc);
465 ptlrpc_init_client(rq_portal, rp_portal, name,
466 &obddev->obd_ldlm_client);
468 imp = class_new_import(obddev);
470 GOTO(err_ldlm, rc = -ENOENT);
471 imp->imp_client = &obddev->obd_ldlm_client;
472 imp->imp_connect_op = connect_op;
473 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
474 LUSTRE_CFG_BUFLEN(lcfg, 1));
475 class_import_put(imp);
477 if (lustre_cfg_buf(lcfg, 4)) {
478 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
480 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
482 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
483 obddev->obd_name, lustre_cfg_string(lcfg, 4),
485 GOTO(err_import, rc);
487 fake_conn.c_self = LNET_MKNID(refnet, 0);
488 imp->imp_connection = &fake_conn;
491 rc = client_import_add_conn(imp, &server_uuid, 1);
493 CERROR("can't add initial connection\n");
494 GOTO(err_import, rc);
496 imp->imp_connection = NULL;
498 cli->cl_import = imp;
499 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
500 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
502 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
503 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
504 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
505 name, obddev->obd_name,
506 cli->cl_target_uuid.uuid);
507 spin_lock(&imp->imp_lock);
508 imp->imp_deactive = 1;
509 spin_unlock(&imp->imp_lock);
513 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
514 LDLM_NAMESPACE_CLIENT,
515 LDLM_NAMESPACE_GREEDY,
517 if (obddev->obd_namespace == NULL) {
518 CERROR("Unable to create client namespace - %s\n",
520 GOTO(err_import, rc = -ENOMEM);
526 class_destroy_import(imp);
530 if (cli->cl_mod_tag_bitmap != NULL)
531 OBD_FREE(cli->cl_mod_tag_bitmap,
532 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
533 cli->cl_mod_tag_bitmap = NULL;
537 EXPORT_SYMBOL(client_obd_setup);
539 int client_obd_cleanup(struct obd_device *obddev)
541 struct client_obd *cli = &obddev->u.cli;
544 ldlm_namespace_free_post(obddev->obd_namespace);
545 obddev->obd_namespace = NULL;
547 obd_cleanup_client_import(obddev);
548 LASSERT(obddev->u.cli.cl_import == NULL);
552 if (cli->cl_mod_tag_bitmap != NULL)
553 OBD_FREE(cli->cl_mod_tag_bitmap,
554 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
555 cli->cl_mod_tag_bitmap = NULL;
559 EXPORT_SYMBOL(client_obd_cleanup);
561 /* ->o_connect() method for client side (OSC and MDC and MGC) */
562 int client_connect_import(const struct lu_env *env,
563 struct obd_export **exp,
564 struct obd_device *obd, struct obd_uuid *cluuid,
565 struct obd_connect_data *data, void *localdata)
567 struct client_obd *cli = &obd->u.cli;
568 struct obd_import *imp = cli->cl_import;
569 struct obd_connect_data *ocd;
570 struct lustre_handle conn = { 0 };
575 down_write(&cli->cl_sem);
576 if (cli->cl_conn_count > 0)
577 GOTO(out_sem, rc = -EALREADY);
579 rc = class_connect(&conn, obd, cluuid);
583 cli->cl_conn_count++;
584 *exp = class_conn2export(&conn);
586 LASSERT(obd->obd_namespace);
588 imp->imp_dlm_handle = conn;
589 rc = ptlrpc_init_import(imp);
593 ocd = &imp->imp_connect_data;
596 imp->imp_connect_flags_orig = data->ocd_connect_flags;
597 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
600 rc = ptlrpc_connect_import(imp);
602 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
605 LASSERT(*exp != NULL && (*exp)->exp_connection);
608 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
609 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
610 data->ocd_connect_flags, ocd->ocd_connect_flags);
611 data->ocd_connect_flags = ocd->ocd_connect_flags;
612 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
615 ptlrpc_pinger_add_import(imp);
621 cli->cl_conn_count--;
622 class_disconnect(*exp);
626 up_write(&cli->cl_sem);
630 EXPORT_SYMBOL(client_connect_import);
632 int client_disconnect_export(struct obd_export *exp)
634 struct obd_device *obd = class_exp2obd(exp);
635 struct client_obd *cli;
636 struct obd_import *imp;
641 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
642 exp, exp ? exp->exp_handle.h_cookie : -1);
647 imp = cli->cl_import;
649 down_write(&cli->cl_sem);
650 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
653 if (cli->cl_conn_count == 0) {
654 CERROR("disconnecting disconnected device (%s)\n",
656 GOTO(out_disconnect, rc = -EINVAL);
659 cli->cl_conn_count--;
660 if (cli->cl_conn_count != 0)
661 GOTO(out_disconnect, rc = 0);
663 /* Mark import deactivated now, so we don't try to reconnect if any
664 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
665 * fully deactivate the import, or that would drop all requests. */
666 spin_lock(&imp->imp_lock);
667 imp->imp_deactive = 1;
668 spin_unlock(&imp->imp_lock);
670 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
671 * delete it regardless. (It's safe to delete an import that was
673 (void)ptlrpc_pinger_del_import(imp);
675 if (obd->obd_namespace != NULL) {
676 /* obd_force == local only */
677 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
678 obd->obd_force ? LCF_LOCAL : 0, NULL);
679 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
682 /* There's no need to hold sem while disconnecting an import,
683 * and it may actually cause deadlock in GSS. */
684 up_write(&cli->cl_sem);
685 rc = ptlrpc_disconnect_import(imp, 0);
686 down_write(&cli->cl_sem);
688 ptlrpc_invalidate_import(imp);
693 /* Use server style - class_disconnect should be always called for
695 err = class_disconnect(exp);
699 up_write(&cli->cl_sem);
703 EXPORT_SYMBOL(client_disconnect_export);
705 #ifdef HAVE_SERVER_SUPPORT
706 int server_disconnect_export(struct obd_export *exp)
711 /* Disconnect early so that clients can't keep using export. */
712 rc = class_disconnect(exp);
713 /* Close import to avoid sending any requests. */
714 if (exp->exp_imp_reverse)
715 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
717 ldlm_bl_thread_wakeup();
719 /* complete all outstanding replies */
720 spin_lock(&exp->exp_lock);
721 while (!list_empty(&exp->exp_outstanding_replies)) {
722 struct ptlrpc_reply_state *rs =
723 list_entry(exp->exp_outstanding_replies.next,
724 struct ptlrpc_reply_state, rs_exp_list);
725 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
727 spin_lock(&svcpt->scp_rep_lock);
729 list_del_init(&rs->rs_exp_list);
731 spin_lock(&rs->rs_lock);
732 /* clear rs_convert_lock to make sure rs is handled and put */
733 rs->rs_convert_lock = 0;
734 ptlrpc_schedule_difficult_reply(rs);
735 spin_unlock(&rs->rs_lock);
737 spin_unlock(&svcpt->scp_rep_lock);
739 spin_unlock(&exp->exp_lock);
743 EXPORT_SYMBOL(server_disconnect_export);
745 /* --------------------------------------------------------------------------
746 * from old lib/target.c
747 * -------------------------------------------------------------------------- */
749 static int target_handle_reconnect(struct lustre_handle *conn,
750 struct obd_export *exp,
751 struct obd_uuid *cluuid)
753 struct obd_device *target;
754 struct lustre_handle *hdl;
760 hdl = &exp->exp_imp_reverse->imp_remote_handle;
761 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
762 conn->cookie = exp->exp_handle.h_cookie;
763 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
764 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
768 target = exp->exp_obd;
770 /* Might be a re-connect after a partition. */
771 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
772 LCONSOLE_WARN("%s: already connected client %s (at %s) "
773 "with handle %#llx. Rejecting client "
774 "with the same UUID trying to reconnect "
775 "with handle %#llx\n", target->obd_name,
776 obd_uuid2str(&exp->exp_client_uuid),
777 obd_export_nid2str(exp),
778 hdl->cookie, conn->cookie);
779 memset(conn, 0, sizeof *conn);
780 /* target_handle_connect() treats EALREADY and
781 * -EALREADY differently. -EALREADY is an error
782 * (same UUID, different handle). */
786 if (!target->obd_recovering) {
787 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
788 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
789 obd_export_nid2str(exp));
790 GOTO(out_already, rc);
793 remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
794 timeout = ktime_divns(remaining, NSEC_PER_SEC);
796 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
798 obd_uuid2str(&exp->exp_client_uuid),
799 obd_export_nid2str(exp),
800 target->obd_max_recoverable_clients,
801 timeout / 60, timeout % 60);
803 struct target_distribute_txn_data *tdtd;
808 tdtd = class_exp2tgt(exp)->lut_tdtd;
809 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
810 buf = tdtd->tdtd_show_update_logs_retrievers(
811 tdtd->tdtd_show_retrievers_cbdata,
815 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting, waiting for %d MDTs (%s) in recovery for %lld:%.02lld. Please wait until all MDTs recovered or you may force MDT evicition via 'lctl --device %s abort_recovery.\n",
817 obd_uuid2str(&exp->exp_client_uuid),
818 obd_export_nid2str(exp), count,
819 buf ? buf : "unknown (not enough RAM)",
820 (abs(timeout) + target->obd_recovery_timeout) / 60,
821 (abs(timeout) + target->obd_recovery_timeout) % 60,
824 LCONSOLE_WARN("%s: Recovery already passed deadline %lld:%.02lld. If you do not want to wait more, you may force taget eviction via 'lctl --device %s abort_recovery.\n",
825 target->obd_name, abs(timeout) / 60,
826 abs(timeout) % 60, target->obd_name);
833 conn->cookie = exp->exp_handle.h_cookie;
834 /* target_handle_connect() treats EALREADY and
835 * -EALREADY differently. EALREADY means we are
836 * doing a valid reconnect from the same client. */
841 check_and_start_recovery_timer(struct obd_device *obd,
842 struct ptlrpc_request *req, int new_client);
845 * update flags for import during reconnect process
847 static int rev_import_flags_update(struct obd_import *revimp,
848 struct ptlrpc_request *req)
851 struct obd_connect_data *data;
853 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
855 if (data->ocd_connect_flags & OBD_CONNECT_AT)
856 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
858 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
860 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
862 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
864 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
865 revimp->imp_client->cli_name,
866 libcfs_id2str(req->rq_peer), rc);
874 * Allocate a new reverse import for an export.
876 * \retval -errno in case error hit
877 * \retval 0 if reverse import correctly init
879 int rev_import_init(struct obd_export *export)
881 struct obd_device *obd = export->exp_obd;
882 struct obd_import *revimp;
884 LASSERT(export->exp_imp_reverse == NULL);
886 revimp = class_new_import(obd);
890 revimp->imp_remote_handle.cookie = 0ULL;
891 revimp->imp_client = &obd->obd_ldlm_client;
892 revimp->imp_dlm_fake = 1;
894 /* it is safe to connect import in new state as no sends possible */
895 spin_lock(&export->exp_lock);
896 export->exp_imp_reverse = revimp;
897 spin_unlock(&export->exp_lock);
898 class_import_put(revimp);
902 EXPORT_SYMBOL(rev_import_init);
905 * Handle reconnect for an export.
907 * \param exp export to handle reconnect process
908 * \param req client reconnect request
910 * \retval -rc in case securitfy flavor can't be changed
911 * \retval 0 in case none problems
913 static int rev_import_reconnect(struct obd_export *exp,
914 struct ptlrpc_request *req)
916 struct obd_import *revimp = exp->exp_imp_reverse;
917 struct lustre_handle *lh;
920 /* avoid sending a request until import flags are changed */
921 ptlrpc_import_enter_resend(revimp);
923 if (revimp->imp_connection != NULL)
924 ptlrpc_connection_put(revimp->imp_connection);
927 * client from recovery don't have a handle so we need to take from
928 * request. it may produce situation when wrong client connected
929 * to recovery as we trust a client uuid
931 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
932 revimp->imp_remote_handle = *lh;
934 /* unknown versions will be caught in
935 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
936 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
938 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
940 rc = rev_import_flags_update(revimp, req);
942 /* it is safe to still be in RECOVERY phase as we are not able
943 * to setup correct security flavor so requests are not able to
944 * be delivered correctly */
948 /* resend all rpc's via new connection */
949 return ptlrpc_import_recovery_state_machine(revimp);
952 int target_handle_connect(struct ptlrpc_request *req)
954 struct obd_device *target = NULL;
955 struct obd_export *export = NULL;
956 /* connect handle - filled from target_handle_reconnect in
958 struct lustre_handle conn;
959 struct lustre_handle *tmp;
960 struct obd_uuid cluuid;
965 bool mds_conn = false, lw_client = false, initial_conn = false;
966 bool mds_mds_conn = false;
967 bool new_mds_mds_conn = false;
968 struct obd_connect_data *data, *tmpdata;
970 lnet_nid_t *client_nid = NULL;
973 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
975 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
977 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
978 GOTO(out, rc = -EINVAL);
981 target = class_dev_by_str(str);
983 deuuidify(str, NULL, &target_start, &target_len);
984 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
985 "from %s (no target). If you are running "
986 "an HA pair check that the target is "
987 "mounted on the other server.\n", str,
988 libcfs_nid2str(req->rq_peer.nid));
989 GOTO(out, rc = -ENODEV);
992 spin_lock(&target->obd_dev_lock);
994 target->obd_conn_inprogress++;
996 if (target->obd_stopping || !target->obd_set_up) {
997 spin_unlock(&target->obd_dev_lock);
999 deuuidify(str, NULL, &target_start, &target_len);
1000 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1001 target_len, target_start,
1002 libcfs_nid2str(req->rq_peer.nid),
1003 (target->obd_stopping ?
1004 "stopping" : "not set up"));
1005 GOTO(out, rc = -ENODEV);
1008 if (target->obd_no_conn) {
1009 spin_unlock(&target->obd_dev_lock);
1011 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1012 "from %s\n", target->obd_name,
1013 libcfs_nid2str(req->rq_peer.nid));
1014 GOTO(out, rc = -EAGAIN);
1017 spin_unlock(&target->obd_dev_lock);
1019 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1021 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1022 GOTO(out, rc = -EINVAL);
1025 obd_str2uuid(&cluuid, str);
1027 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1029 GOTO(out, rc = -EPROTO);
1033 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1035 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1037 GOTO(out, rc = -EPROTO);
1039 rc = req_capsule_server_pack(&req->rq_pill);
1043 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1044 /* Don't allow clients to connect that are using old 1.8 format
1045 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1046 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1047 * FULL20 flag should be set on all connections since 2.0, but no
1048 * longer affects behaviour.
1050 * Later this check will be disabled and the flag can be retired
1051 * completely once interop with 3.0 is no longer needed.
1053 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1054 GOTO(out, rc = -EPROTO);
1056 /* Don't allow liblustre clients to connect.
1057 * - testing was disabled in v2_2_50_0-61-g6a75d65
1058 * - building was disabled in v2_5_58_0-28-g7277179
1059 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1060 * - clients were refused connect for version difference > 0.0.1.32 */
1061 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1062 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1063 GOTO(out, rc = -EPROTO);
1067 /* Note: lw_client is needed in MDS-MDS failover during update log
1068 * processing, so we needs to allow lw_client to be connected at
1069 * anytime, instead of only the initial connection */
1070 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1072 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1073 initial_conn = true;
1074 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1075 mds_mds_conn = (data->ocd_connect_flags &
1076 OBD_CONNECT_MDS_MDS) != 0;
1078 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1079 * for Imperative Recovery connection from MGC to MGS.
1081 * Via check OBD_CONNECT_FID, we can distinguish whether
1082 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1083 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1086 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1087 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1088 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1089 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1090 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1091 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1093 /* We do not support the MDT-MDT interoperations with
1094 * different version MDT because of protocol changes. */
1095 if (unlikely(major != LUSTRE_MAJOR ||
1096 minor != LUSTRE_MINOR ||
1097 abs(patch - LUSTRE_PATCH) > 3)) {
1098 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1099 "connection from different version MDT "
1100 "(%d.%d.%d.%d) %s %s\n",
1101 target->obd_name, LUSTRE_MAJOR,
1102 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1103 major, minor, patch,
1104 OBD_OCD_VERSION_FIX(data->ocd_version),
1105 libcfs_nid2str(req->rq_peer.nid), str);
1107 GOTO(out, rc = -EPROTO);
1112 /* lctl gets a backstage, all-access pass. */
1113 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1114 goto dont_check_exports;
1116 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1120 /* We've found an export in the hash. */
1122 spin_lock(&export->exp_lock);
1124 if (export->exp_connecting) { /* bug 9635, et. al. */
1125 spin_unlock(&export->exp_lock);
1126 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1127 export->exp_obd->obd_name, export,
1128 libcfs_nid2str(req->rq_peer.nid));
1129 class_export_put(export);
1132 } else if ((mds_conn || (lw_client && initial_conn) ||
1133 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1134 export->exp_connection != NULL) {
1135 spin_unlock(&export->exp_lock);
1136 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1137 /* MDS or LWP reconnected after failover. */
1138 LCONSOLE_WARN("%s: Received %s connection from "
1139 "%s, removing former export from %s\n",
1140 target->obd_name, mds_conn ? "MDS" : "LWP",
1141 libcfs_nid2str(req->rq_peer.nid),
1142 libcfs_nid2str(export->exp_connection->c_peer.nid));
1144 /* New MDS connection from the same NID. */
1145 LCONSOLE_WARN("%s: Received new %s connection from "
1146 "%s, removing former export from same NID\n",
1147 target->obd_name, mds_conn ? "MDS" : "LWP",
1148 libcfs_nid2str(req->rq_peer.nid));
1151 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1152 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1153 /* Because exports between MDTs will always be
1154 * kept, let's do not fail such export if they
1155 * come from the same NID, otherwise it might
1156 * cause eviction between MDTs, which might
1157 * cause namespace inconsistency */
1158 spin_lock(&export->exp_lock);
1159 export->exp_connecting = 1;
1160 export->exp_conn_cnt = 0;
1161 spin_unlock(&export->exp_lock);
1162 conn.cookie = export->exp_handle.h_cookie;
1165 class_fail_export(export);
1166 class_export_put(export);
1170 } else if (export->exp_connection != NULL && initial_conn &&
1171 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1172 spin_unlock(&export->exp_lock);
1173 /* In MDS failover we have static UUID but NID can change. */
1174 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1175 "existing nid %s is already connected\n",
1176 target->obd_name, cluuid.uuid,
1177 libcfs_nid2str(req->rq_peer.nid),
1179 export->exp_connection->c_peer.nid));
1181 class_export_put(export);
1184 export->exp_connecting = 1;
1185 spin_unlock(&export->exp_lock);
1186 LASSERT(export->exp_obd == target);
1188 rc = target_handle_reconnect(&conn, export, &cluuid);
1191 /* If we found an export, we already unlocked. */
1194 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1195 } else if (req->rq_export == NULL &&
1196 atomic_read(&export->exp_rpc_count) > 0) {
1197 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1198 "still busy with %d references\n",
1199 target->obd_name, cluuid.uuid,
1200 libcfs_nid2str(req->rq_peer.nid),
1201 atomic_read(&export->exp_refcount));
1202 GOTO(out, rc = -EBUSY);
1203 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1205 if (!strstr(cluuid.uuid, "mdt"))
1206 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1207 "known client %s (at %s) because it "
1208 "is indicating it is a new client",
1209 target->obd_name, cluuid.uuid,
1210 libcfs_nid2str(req->rq_peer.nid));
1211 GOTO(out, rc = -EALREADY);
1213 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1220 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1221 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1222 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1223 export, ktime_get_real_seconds(),
1224 export ? export->exp_last_request_time : 0);
1226 /* If this is the first time a client connects, reset the recovery
1227 * timer. Discard lightweight connections which might be local. */
1228 if (!lw_client && rc == 0 && target->obd_recovering)
1229 check_and_start_recovery_timer(target, req, export == NULL);
1231 /* We want to handle EALREADY but *not* -EALREADY from
1232 * target_handle_reconnect(), return reconnection state in a flag. */
1233 if (rc == EALREADY) {
1234 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1240 /* Tell the client if we support replayable requests. */
1241 if (target->obd_replayable)
1242 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1243 client_nid = &req->rq_peer.nid;
1245 if (export == NULL) {
1246 /* allow lightweight connections during recovery */
1247 /* allow "new" MDT to be connected during recovery, since we
1248 * need retrieve recovery update records from it */
1249 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1250 struct hrtimer *timer = &target->obd_recovery_timer;
1259 connected = atomic_read(&target->obd_connected_clients);
1260 in_progress = atomic_read(&target->obd_lock_replay_clients);
1261 known = target->obd_max_recoverable_clients;
1262 stale = target->obd_stale_clients;
1263 remaining = hrtimer_expires_remaining(timer);
1264 left = ktime_divns(remaining, NSEC_PER_SEC);
1265 if (ktime_to_ns(remaining) > 0) {
1266 msg = "to recover in";
1269 msg = "already passed deadline";
1273 LCONSOLE_WARN("%s: Denying connection for new client %s (at %s), waiting for %d known clients (%d recovered, %d in progress, and %d evicted) %s %lld:%.02lld\n",
1274 target->obd_name, cluuid.uuid,
1275 libcfs_nid2str(req->rq_peer.nid), known,
1276 connected - in_progress, in_progress,
1277 stale, msg, timeout / 60, timeout % 60);
1281 rc = obd_connect(req->rq_svc_thread->t_env,
1282 &export, target, &cluuid, data,
1284 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1285 lustre_msg_add_op_flags(req->rq_repmsg,
1286 MSG_CONNECT_RECOVERING);
1288 conn.cookie = export->exp_handle.h_cookie;
1289 rc = rev_import_init(export);
1293 new_mds_mds_conn = true;
1296 rc = obd_reconnect(req->rq_svc_thread->t_env,
1297 export, target, &cluuid, data, client_nid);
1302 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1303 data->ocd_instance = target->u.obt.obt_instance;
1305 /* Return only the parts of obd_connect_data that we understand, so the
1306 * client knows that we don't understand the rest. */
1308 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1310 tmpdata = req_capsule_server_get(&req->rq_pill,
1312 /* Don't use struct assignment here, because the client reply
1313 * buffer may be smaller/larger than the local struct
1314 * obd_connect_data. */
1315 memcpy(tmpdata, data, min(tmpsize, size));
1318 /* If the client and the server are the same node, we will already
1319 * have an export that really points to the client's DLM export,
1320 * because we have a shared handles table.
1322 * XXX this will go away when shaver stops sending the "connect" handle
1323 * in the real "remote handle" field of the request --phik 24 Apr 2003
1325 ptlrpc_request_change_export(req, export);
1327 spin_lock(&export->exp_lock);
1328 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1329 spin_unlock(&export->exp_lock);
1330 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1331 "or equal conn_cnt: %d >= %d\n",
1332 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1333 export->exp_conn_cnt,
1334 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1336 GOTO(out, rc = -EALREADY);
1338 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1339 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1340 spin_unlock(&export->exp_lock);
1342 if (export->exp_connection != NULL) {
1343 /* Check to see if connection came from another NID. */
1344 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1345 !hlist_unhashed(&export->exp_nid_hash))
1346 cfs_hash_del(export->exp_obd->obd_nid_hash,
1347 &export->exp_connection->c_peer.nid,
1348 &export->exp_nid_hash);
1350 ptlrpc_connection_put(export->exp_connection);
1353 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1356 if (hlist_unhashed(&export->exp_nid_hash))
1357 cfs_hash_add(export->exp_obd->obd_nid_hash,
1358 &export->exp_connection->c_peer.nid,
1359 &export->exp_nid_hash);
1361 lustre_msg_set_handle(req->rq_repmsg, &conn);
1363 rc = rev_import_reconnect(export, req);
1367 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1369 __u64 transno = data->ocd_transno;
1371 spin_lock(&export->exp_lock);
1372 /* possible race with class_disconnect_stale_exports,
1373 * export may be already in the eviction process */
1374 if (export->exp_failed) {
1375 spin_unlock(&export->exp_lock);
1376 GOTO(out, rc = -ENODEV);
1378 export->exp_in_recovery = 1;
1379 export->exp_req_replay_needed = 1;
1380 export->exp_lock_replay_needed = 1;
1381 spin_unlock(&export->exp_lock);
1383 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1384 MSG_CONNECT_TRANSNO);
1385 if (has_transno && transno == 0)
1386 CWARN("Connect with zero transno!\n");
1388 if (has_transno && transno > 0 &&
1389 transno < target->obd_next_recovery_transno &&
1390 transno > target->obd_last_committed) {
1391 /* Another way is to use cmpxchg() to be lock-free. */
1392 spin_lock(&target->obd_recovery_task_lock);
1393 if (transno < target->obd_next_recovery_transno)
1394 target->obd_next_recovery_transno = transno;
1395 spin_unlock(&target->obd_recovery_task_lock);
1398 atomic_inc(&target->obd_req_replay_clients);
1399 atomic_inc(&target->obd_lock_replay_clients);
1400 /* Note: MDS-MDS connection is allowed to be connected during
1401 * recovery, no matter if the exports needs to be recoveried.
1402 * Because we need retrieve updates logs from all other MDTs.
1403 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1404 * also needs to be increased to match other recovery checking
1406 if (new_mds_mds_conn)
1407 target->obd_max_recoverable_clients++;
1408 if (atomic_inc_return(&target->obd_connected_clients) ==
1409 target->obd_max_recoverable_clients)
1410 wake_up(&target->obd_next_transno_waitq);
1413 /* Tell the client we're in recovery, when client is involved in it. */
1414 if (target->obd_recovering && !lw_client)
1415 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1419 spin_lock(&export->exp_lock);
1420 export->exp_connecting = 0;
1421 spin_unlock(&export->exp_lock);
1423 class_export_put(export);
1425 if (target != NULL) {
1426 spin_lock(&target->obd_dev_lock);
1427 target->obd_conn_inprogress--;
1428 spin_unlock(&target->obd_dev_lock);
1429 class_decref(target, "find", current);
1431 req->rq_status = rc;
1435 int target_handle_disconnect(struct ptlrpc_request *req)
1440 rc = req_capsule_server_pack(&req->rq_pill);
1444 /* Keep the rq_export around so we can send the reply. */
1445 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1450 void target_destroy_export(struct obd_export *exp)
1452 struct obd_import *imp = NULL;
1453 /* exports created from last_rcvd data, and "fake"
1454 exports created by lctl don't have an import */
1455 spin_lock(&exp->exp_lock);
1456 if (exp->exp_imp_reverse != NULL) {
1457 imp = exp->exp_imp_reverse;
1458 exp->exp_imp_reverse = NULL;
1460 spin_unlock(&exp->exp_lock);
1462 client_destroy_import(imp);
1464 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1465 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1466 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1467 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1469 EXPORT_SYMBOL(target_destroy_export);
1472 * Recovery functions
1474 static void target_request_copy_get(struct ptlrpc_request *req)
1476 class_export_rpc_inc(req->rq_export);
1477 LASSERT(list_empty(&req->rq_list));
1478 INIT_LIST_HEAD(&req->rq_replay_list);
1480 /* Increase refcount to keep request in queue. */
1481 atomic_inc(&req->rq_refcount);
1482 /* Let export know it has replays to be handled. */
1483 atomic_inc(&req->rq_export->exp_replay_count);
1486 static void target_request_copy_put(struct ptlrpc_request *req)
1488 LASSERT(list_empty(&req->rq_replay_list));
1489 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1491 atomic_dec(&req->rq_export->exp_replay_count);
1492 class_export_rpc_dec(req->rq_export);
1493 ptlrpc_server_drop_request(req);
1496 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1498 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1499 struct obd_export *exp = req->rq_export;
1500 struct ptlrpc_request *reqiter;
1501 struct ptlrpc_request *dup_req = NULL;
1506 spin_lock(&exp->exp_lock);
1507 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1509 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1517 /* We expect it with RESENT and REPLAY flags. */
1518 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1519 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1520 CERROR("invalid flags %x of resent replay\n",
1521 lustre_msg_get_flags(req->rq_reqmsg));
1523 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1526 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1528 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1529 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1533 list_add_tail(&req->rq_replay_list,
1534 &exp->exp_req_replay_queue);
1537 spin_unlock(&exp->exp_lock);
1541 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1543 LASSERT(!list_empty(&req->rq_replay_list));
1544 LASSERT(req->rq_export);
1546 spin_lock(&req->rq_export->exp_lock);
1547 list_del_init(&req->rq_replay_list);
1548 spin_unlock(&req->rq_export->exp_lock);
1551 static void target_finish_recovery(struct lu_target *lut)
1553 struct obd_device *obd = lut->lut_obd;
1556 /* Only log a recovery message when recovery has occurred. */
1557 if (obd->obd_recovery_start) {
1558 time64_t now = ktime_get_real_seconds();
1559 time64_t elapsed_time;
1561 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1562 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1563 "%d recovered and %d %s evicted.\n", obd->obd_name,
1564 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1565 obd->obd_max_recoverable_clients,
1566 atomic_read(&obd->obd_connected_clients),
1567 obd->obd_stale_clients,
1568 obd->obd_stale_clients == 1 ? "was" : "were");
1571 ldlm_reprocess_recovery_done(obd->obd_namespace);
1572 spin_lock(&obd->obd_recovery_task_lock);
1573 if (!list_empty(&obd->obd_req_replay_queue) ||
1574 !list_empty(&obd->obd_lock_replay_queue) ||
1575 !list_empty(&obd->obd_final_req_queue)) {
1576 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1578 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1579 list_empty(&obd->obd_lock_replay_queue) ? \
1581 list_empty(&obd->obd_final_req_queue) ? \
1583 spin_unlock(&obd->obd_recovery_task_lock);
1586 spin_unlock(&obd->obd_recovery_task_lock);
1588 obd->obd_recovery_end = ktime_get_real_seconds();
1590 /* When recovery finished, cleanup orphans on MDS and OST. */
1591 if (obd->obd_type && OBP(obd, postrecov)) {
1592 int rc = OBP(obd, postrecov)(obd);
1595 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1601 static void abort_req_replay_queue(struct obd_device *obd)
1603 struct ptlrpc_request *req, *n;
1604 struct list_head abort_list;
1606 INIT_LIST_HEAD(&abort_list);
1607 spin_lock(&obd->obd_recovery_task_lock);
1608 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1609 spin_unlock(&obd->obd_recovery_task_lock);
1610 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1611 DEBUG_REQ(D_WARNING, req, "aborted:");
1612 req->rq_status = -ENOTCONN;
1613 if (ptlrpc_error(req)) {
1614 DEBUG_REQ(D_ERROR, req,
1615 "failed abort_req_reply; skipping");
1617 target_exp_dequeue_req_replay(req);
1618 target_request_copy_put(req);
1622 static void abort_lock_replay_queue(struct obd_device *obd)
1624 struct ptlrpc_request *req, *n;
1625 struct list_head abort_list;
1627 INIT_LIST_HEAD(&abort_list);
1628 spin_lock(&obd->obd_recovery_task_lock);
1629 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1630 spin_unlock(&obd->obd_recovery_task_lock);
1631 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1632 DEBUG_REQ(D_ERROR, req, "aborted:");
1633 req->rq_status = -ENOTCONN;
1634 if (ptlrpc_error(req)) {
1635 DEBUG_REQ(D_ERROR, req,
1636 "failed abort_lock_reply; skipping");
1638 target_request_copy_put(req);
1642 /* Called from a cleanup function if the device is being cleaned up
1643 forcefully. The exports should all have been disconnected already,
1644 the only thing left to do is
1645 - clear the recovery flags
1647 - free queued requests and replies, but don't send replies
1648 Because the obd_stopping flag is set, no new requests should be received.
1651 void target_cleanup_recovery(struct obd_device *obd)
1653 struct ptlrpc_request *req, *n;
1654 struct list_head clean_list;
1656 INIT_LIST_HEAD(&clean_list);
1657 spin_lock(&obd->obd_dev_lock);
1658 if (!obd->obd_recovering) {
1659 spin_unlock(&obd->obd_dev_lock);
1663 obd->obd_recovering = obd->obd_abort_recovery = 0;
1664 spin_unlock(&obd->obd_dev_lock);
1666 spin_lock(&obd->obd_recovery_task_lock);
1667 target_cancel_recovery_timer(obd);
1668 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1669 spin_unlock(&obd->obd_recovery_task_lock);
1671 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1672 LASSERT(req->rq_reply_state == NULL);
1673 target_exp_dequeue_req_replay(req);
1674 target_request_copy_put(req);
1677 spin_lock(&obd->obd_recovery_task_lock);
1678 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1679 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1680 spin_unlock(&obd->obd_recovery_task_lock);
1682 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1683 LASSERT(req->rq_reply_state == NULL);
1684 target_request_copy_put(req);
1689 EXPORT_SYMBOL(target_cleanup_recovery);
1691 /* obd_recovery_task_lock should be held */
1692 void target_cancel_recovery_timer(struct obd_device *obd)
1694 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1695 hrtimer_cancel(&obd->obd_recovery_timer);
1698 static void target_start_recovery_timer(struct obd_device *obd)
1702 if (obd->obd_recovery_start != 0)
1705 spin_lock(&obd->obd_dev_lock);
1706 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1707 spin_unlock(&obd->obd_dev_lock);
1711 LASSERT(obd->obd_recovery_timeout != 0);
1713 if (obd->obd_recovery_start != 0) {
1714 spin_unlock(&obd->obd_dev_lock);
1718 delay = ktime_set(obd->obd_recovery_timeout, 0);
1719 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_REL);
1720 obd->obd_recovery_start = ktime_get_real_seconds();
1721 spin_unlock(&obd->obd_dev_lock);
1723 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1725 obd->obd_recovery_timeout / 60,
1726 obd->obd_recovery_timeout % 60,
1727 obd->obd_max_recoverable_clients,
1728 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1729 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1733 * extend recovery window.
1735 * if @extend is true, extend recovery window to have @drt remaining at least;
1736 * otherwise, make sure the recovery timeout value is not less than @drt.
1738 static void extend_recovery_timer(struct obd_device *obd, time_t drt,
1745 spin_lock(&obd->obd_dev_lock);
1746 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1747 spin_unlock(&obd->obd_dev_lock);
1750 LASSERT(obd->obd_recovery_start != 0);
1752 to = obd->obd_recovery_timeout;
1753 left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1754 left = ktime_divns(left_ns, NSEC_PER_SEC);
1755 if (extend && (drt > left))
1757 else if (!extend && (drt > to))
1760 if (to > obd->obd_recovery_time_hard) {
1761 to = obd->obd_recovery_time_hard;
1762 CWARN("%s: extended recovery timer reaching hard limit: %ld, extend: %d\n",
1763 obd->obd_name, to, extend);
1766 if (obd->obd_recovery_timeout < to) {
1767 ktime_t now = ktime_get_real();
1770 obd->obd_recovery_timeout = to;
1771 end = ktime_set(obd->obd_recovery_start + to, 0);
1772 left_ns = ktime_sub(end, now);
1773 hrtimer_forward_now(&obd->obd_recovery_timer, left_ns);
1774 left = ktime_divns(left_ns, NSEC_PER_MSEC);
1776 spin_unlock(&obd->obd_dev_lock);
1778 CDEBUG(D_HA, "%s: recovery timer will expire in %ld seconds\n",
1779 obd->obd_name, left);
1782 /* Reset the timer with each new client connection */
1784 * This timer is actually reconnect_timer, which is for making sure
1785 * the total recovery window is at least as big as my reconnect
1786 * attempt timing. So the initial recovery time_out will be set to
1787 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1788 * from client is bigger than this, then the recovery time_out will
1789 * be extended to make sure the client could be reconnected, in the
1790 * process, the timeout from the new client should be ignored.
1793 check_and_start_recovery_timer(struct obd_device *obd,
1794 struct ptlrpc_request *req,
1797 time_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1798 struct obd_device_target *obt = &obd->u.obt;
1800 if (!new_client && service_time)
1801 /* Teach server about old server's estimates, as first guess
1802 * at how long new requests will take.
1804 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1807 target_start_recovery_timer(obd);
1809 /* Convert the service time to RPC timeout,
1810 * and reuse service_time to limit stack usage.
1812 service_time = at_est2timeout(service_time);
1814 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1815 service_time < at_extra)
1816 service_time = at_extra;
1818 /* We expect other clients to timeout within service_time, then try
1819 * to reconnect, then try the failover server. The max delay between
1820 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1821 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1823 LASSERT(obt->obt_magic == OBT_MAGIC);
1824 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1825 if (service_time > obd->obd_recovery_timeout && !new_client)
1826 extend_recovery_timer(obd, service_time, false);
1829 /** Health checking routines */
1830 static inline int exp_connect_healthy(struct obd_export *exp)
1832 return (exp->exp_in_recovery);
1835 /** if export done req_replay or has replay in queue */
1836 static inline int exp_req_replay_healthy(struct obd_export *exp)
1838 return (!exp->exp_req_replay_needed ||
1839 atomic_read(&exp->exp_replay_count) > 0);
1843 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1845 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1846 exp_req_replay_healthy(exp);
1849 /** if export done lock_replay or has replay in queue */
1850 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1852 return (!exp->exp_lock_replay_needed ||
1853 atomic_read(&exp->exp_replay_count) > 0);
1856 static inline int exp_vbr_healthy(struct obd_export *exp)
1858 return (!exp->exp_vbr_failed);
1861 static inline int exp_finished(struct obd_export *exp)
1863 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1866 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1868 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1872 static int check_for_next_transno(struct lu_target *lut)
1874 struct ptlrpc_request *req = NULL;
1875 struct obd_device *obd = lut->lut_obd;
1876 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1877 int wake_up = 0, connected, completed, queue_len;
1878 __u64 req_transno = 0;
1879 __u64 update_transno = 0;
1880 __u64 next_transno = 0;
1883 spin_lock(&obd->obd_recovery_task_lock);
1884 if (!list_empty(&obd->obd_req_replay_queue)) {
1885 req = list_entry(obd->obd_req_replay_queue.next,
1886 struct ptlrpc_request, rq_list);
1887 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1891 update_transno = distribute_txn_get_next_transno(tdtd);
1893 connected = atomic_read(&obd->obd_connected_clients);
1894 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1895 queue_len = obd->obd_requests_queued_for_recovery;
1896 next_transno = obd->obd_next_recovery_transno;
1898 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1899 "req_transno: %llu, next_transno: %llu\n",
1900 obd->obd_max_recoverable_clients, connected, completed,
1901 queue_len, req_transno, next_transno);
1903 if (obd->obd_abort_recovery) {
1904 CDEBUG(D_HA, "waking for aborted recovery\n");
1906 } else if (obd->obd_recovery_expired) {
1907 CDEBUG(D_HA, "waking for expired recovery\n");
1909 } else if (tdtd != NULL && req != NULL &&
1910 is_req_replayed_by_update(req)) {
1911 LASSERTF(req_transno < next_transno, "req_transno %llu"
1912 "next_transno%llu\n", req_transno, next_transno);
1913 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1916 } else if (req_transno == next_transno ||
1917 (update_transno != 0 && update_transno <= next_transno)) {
1918 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1920 } else if (queue_len > 0 &&
1921 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1922 /** handle gaps occured due to lost reply or VBR */
1923 LASSERTF(req_transno >= next_transno,
1924 "req_transno: %llu, next_transno: %llu\n",
1925 req_transno, next_transno);
1927 "%s: waking for gap in transno, VBR is %s (skip: "
1928 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1929 ", next_update %lld last_committed: %lld)\n",
1930 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1931 next_transno, queue_len, completed, connected,
1932 req_transno, update_transno, obd->obd_last_committed);
1933 obd->obd_next_recovery_transno = req_transno;
1935 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1936 CDEBUG(D_HA, "waking for completed recovery\n");
1938 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1939 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1940 " by fail_lock, waking up (%lld)\n", next_transno);
1941 obd->obd_next_recovery_transno = req_transno;
1944 spin_unlock(&obd->obd_recovery_task_lock);
1948 static int check_for_next_lock(struct lu_target *lut)
1950 struct obd_device *obd = lut->lut_obd;
1953 spin_lock(&obd->obd_recovery_task_lock);
1954 if (!list_empty(&obd->obd_lock_replay_queue)) {
1955 CDEBUG(D_HA, "waking for next lock\n");
1957 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1958 CDEBUG(D_HA, "waking for completed lock replay\n");
1960 } else if (obd->obd_abort_recovery) {
1961 CDEBUG(D_HA, "waking for aborted recovery\n");
1963 } else if (obd->obd_recovery_expired) {
1964 CDEBUG(D_HA, "waking for expired recovery\n");
1967 spin_unlock(&obd->obd_recovery_task_lock);
1973 * wait for recovery events,
1974 * check its status with help of check_routine
1975 * evict dead clients via health_check
1977 static int target_recovery_overseer(struct lu_target *lut,
1978 int (*check_routine)(struct lu_target *),
1979 int (*health_check)(struct obd_export *))
1981 struct obd_device *obd = lut->lut_obd;
1982 struct target_distribute_txn_data *tdtd;
1986 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
1987 now = ktime_get_seconds();
1988 if (now - last > 600) {
1989 LCONSOLE_INFO("%s: in recovery but waiting for "
1990 "the first client to connect\n",
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 time_t 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
2181 to = max_t(time_t, at_est2timeout(at_get(&svcpt->scp_at_estimate)),
2182 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 */
2476 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2478 trd->trd_processing_task = current_pid();
2480 spin_lock(&obd->obd_dev_lock);
2481 obd->obd_recovering = 1;
2482 spin_unlock(&obd->obd_dev_lock);
2483 complete(&trd->trd_starting);
2485 /* first of all, we have to know the first transno to replay */
2486 if (target_recovery_overseer(lut, check_for_recovery_ready,
2487 exp_connect_healthy)) {
2488 abort_req_replay_queue(obd);
2489 abort_lock_replay_queue(obd);
2490 if (lut->lut_tdtd != NULL)
2491 dtrq_list_destroy(lut->lut_tdtd);
2494 /* next stage: replay requests or update */
2496 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2497 atomic_read(&obd->obd_req_replay_clients),
2498 obd->obd_next_recovery_transno);
2499 replay_request_or_update(env, lut, trd, thread);
2502 * The second stage: replay locks
2504 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2505 atomic_read(&obd->obd_lock_replay_clients));
2506 while ((req = target_next_replay_lock(lut))) {
2507 LASSERT(trd->trd_processing_task == current_pid());
2508 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2509 libcfs_nid2str(req->rq_peer.nid));
2510 handle_recovery_req(thread, req,
2511 trd->trd_recovery_handler);
2512 target_request_copy_put(req);
2513 obd->obd_replayed_locks++;
2517 * The third stage: reply on final pings, at this moment all clients
2518 * must have request in final queue
2520 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2521 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2522 /** Update server last boot epoch */
2523 tgt_boot_epoch_update(lut);
2524 /* We drop recoverying flag to forward all new requests
2525 * to regular mds_handle() since now */
2526 spin_lock(&obd->obd_dev_lock);
2527 obd->obd_recovering = obd->obd_abort_recovery = 0;
2528 spin_unlock(&obd->obd_dev_lock);
2529 spin_lock(&obd->obd_recovery_task_lock);
2530 target_cancel_recovery_timer(obd);
2531 spin_unlock(&obd->obd_recovery_task_lock);
2532 while ((req = target_next_final_ping(obd))) {
2533 LASSERT(trd->trd_processing_task == current_pid());
2534 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2535 libcfs_nid2str(req->rq_peer.nid));
2536 handle_recovery_req(thread, req,
2537 trd->trd_recovery_handler);
2538 /* Because the waiting client can not send ping to server,
2539 * so we need refresh the last_request_time, to avoid the
2540 * export is being evicted */
2541 ptlrpc_update_export_timer(req->rq_export, 0);
2542 target_request_copy_put(req);
2545 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2546 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2547 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2548 if (delta > OBD_RECOVERY_TIME_SOFT) {
2549 CWARN("too long recovery - read logs\n");
2550 libcfs_debug_dumplog();
2553 target_finish_recovery(lut);
2555 lu_context_fini(&env->le_ctx);
2556 trd->trd_processing_task = 0;
2557 complete(&trd->trd_finishing);
2559 tgt_io_thread_done(thread);
2560 OBD_FREE_PTR(thread);
2565 static int target_start_recovery_thread(struct lu_target *lut,
2566 svc_handler_t handler)
2568 struct obd_device *obd = lut->lut_obd;
2570 struct target_recovery_data *trd = &obd->obd_recovery_data;
2573 memset(trd, 0, sizeof(*trd));
2574 init_completion(&trd->trd_starting);
2575 init_completion(&trd->trd_finishing);
2576 trd->trd_recovery_handler = handler;
2578 rc = server_name2index(obd->obd_name, &index, NULL);
2582 if (!IS_ERR(kthread_run(target_recovery_thread,
2583 lut, "tgt_recover_%d", index))) {
2584 wait_for_completion(&trd->trd_starting);
2585 LASSERT(obd->obd_recovering != 0);
2593 void target_stop_recovery_thread(struct obd_device *obd)
2595 if (obd->obd_recovery_data.trd_processing_task > 0) {
2596 struct target_recovery_data *trd = &obd->obd_recovery_data;
2597 /** recovery can be done but postrecovery is not yet */
2598 spin_lock(&obd->obd_dev_lock);
2599 if (obd->obd_recovering) {
2600 CERROR("%s: Aborting recovery\n", obd->obd_name);
2601 obd->obd_abort_recovery = 1;
2602 wake_up(&obd->obd_next_transno_waitq);
2604 spin_unlock(&obd->obd_dev_lock);
2605 wait_for_completion(&trd->trd_finishing);
2608 EXPORT_SYMBOL(target_stop_recovery_thread);
2610 void target_recovery_fini(struct obd_device *obd)
2612 class_disconnect_exports(obd);
2613 target_stop_recovery_thread(obd);
2614 target_cleanup_recovery(obd);
2616 EXPORT_SYMBOL(target_recovery_fini);
2618 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2620 struct obd_device *obd = container_of(timer, struct obd_device,
2621 obd_recovery_timer);
2624 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2625 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2626 ktime_get_real_seconds() - obd->obd_recovery_start,
2627 atomic_read(&obd->obd_connected_clients));
2629 obd->obd_recovery_expired = 1;
2630 wake_up(&obd->obd_next_transno_waitq);
2631 return HRTIMER_NORESTART;
2634 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2636 struct obd_device *obd = lut->lut_obd;
2638 if (lut->lut_bottom->dd_rdonly)
2641 if (obd->obd_max_recoverable_clients == 0) {
2642 /** Update server last boot epoch */
2643 tgt_boot_epoch_update(lut);
2647 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2648 "last_transno %llu\n", obd->obd_name,
2649 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2650 LASSERT(obd->obd_stopping == 0);
2651 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2652 obd->obd_recovery_start = 0;
2653 obd->obd_recovery_end = 0;
2655 hrtimer_init(&obd->obd_recovery_timer, CLOCK_REALTIME,
2657 obd->obd_recovery_timer.function = &target_recovery_expired;
2658 target_start_recovery_thread(lut, handler);
2660 EXPORT_SYMBOL(target_recovery_init);
2662 static int target_process_req_flags(struct obd_device *obd,
2663 struct ptlrpc_request *req)
2665 struct obd_export *exp = req->rq_export;
2666 LASSERT(exp != NULL);
2667 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2668 /* client declares he's ready to replay locks */
2669 spin_lock(&exp->exp_lock);
2670 if (exp->exp_req_replay_needed) {
2671 exp->exp_req_replay_needed = 0;
2672 spin_unlock(&exp->exp_lock);
2674 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2675 atomic_dec(&obd->obd_req_replay_clients);
2677 spin_unlock(&exp->exp_lock);
2680 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2681 /* client declares he's ready to complete recovery
2682 * so, we put the request on th final queue */
2683 spin_lock(&exp->exp_lock);
2684 if (exp->exp_lock_replay_needed) {
2685 exp->exp_lock_replay_needed = 0;
2686 spin_unlock(&exp->exp_lock);
2688 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2689 atomic_dec(&obd->obd_lock_replay_clients);
2691 spin_unlock(&exp->exp_lock);
2697 int target_queue_recovery_request(struct ptlrpc_request *req,
2698 struct obd_device *obd)
2700 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2701 struct ptlrpc_request *reqiter;
2705 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2706 /* Processing the queue right now, don't re-add. */
2710 target_process_req_flags(obd, req);
2712 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2713 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2714 if (cfs_fail_val == 1) {
2717 wake_up(&cfs_race_waitq);
2719 set_current_state(TASK_INTERRUPTIBLE);
2720 schedule_timeout(cfs_time_seconds(1));
2724 /* client declares he's ready to complete recovery
2725 * so, we put the request on th final queue */
2726 target_request_copy_get(req);
2727 DEBUG_REQ(D_HA, req, "queue final req");
2728 wake_up(&obd->obd_next_transno_waitq);
2729 spin_lock(&obd->obd_recovery_task_lock);
2730 if (obd->obd_recovering) {
2731 struct ptlrpc_request *tmp;
2732 struct ptlrpc_request *duplicate = NULL;
2734 if (likely(!req->rq_export->exp_replay_done)) {
2735 req->rq_export->exp_replay_done = 1;
2736 list_add_tail(&req->rq_list,
2737 &obd->obd_final_req_queue);
2738 spin_unlock(&obd->obd_recovery_task_lock);
2742 /* XXX O(n), but only happens if final ping is
2743 * timed out, probably reorganize the list as
2744 * a hash list later */
2745 list_for_each_entry_safe(reqiter, tmp,
2746 &obd->obd_final_req_queue,
2748 if (reqiter->rq_export == req->rq_export) {
2749 list_del_init(&reqiter->rq_list);
2750 duplicate = reqiter;
2755 list_add_tail(&req->rq_list,
2756 &obd->obd_final_req_queue);
2757 req->rq_export->exp_replay_done = 1;
2758 spin_unlock(&obd->obd_recovery_task_lock);
2760 if (duplicate != NULL) {
2761 DEBUG_REQ(D_HA, duplicate,
2762 "put prev final req\n");
2763 target_request_copy_put(duplicate);
2767 spin_unlock(&obd->obd_recovery_task_lock);
2768 target_request_copy_put(req);
2769 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2772 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2773 /* client declares he's ready to replay locks */
2774 target_request_copy_get(req);
2775 DEBUG_REQ(D_HA, req, "queue lock replay req");
2776 wake_up(&obd->obd_next_transno_waitq);
2777 spin_lock(&obd->obd_recovery_task_lock);
2778 LASSERT(obd->obd_recovering);
2779 /* usually due to recovery abort */
2780 if (!req->rq_export->exp_in_recovery) {
2781 spin_unlock(&obd->obd_recovery_task_lock);
2782 target_request_copy_put(req);
2785 LASSERT(req->rq_export->exp_lock_replay_needed);
2786 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2787 spin_unlock(&obd->obd_recovery_task_lock);
2791 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2792 * (i.e. buflens etc are in my own byte order), but type-dependent
2793 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2796 INIT_LIST_HEAD(&req->rq_list);
2797 DEBUG_REQ(D_HA, req, "not queueing");
2801 /* If we're processing the queue, we want don't want to queue this
2804 * Also, if this request has a transno less than the one we're waiting
2805 * for, we should process it now. It could (and currently always will)
2806 * be an open request for a descriptor that was opened some time ago.
2808 * Also, a resent, replayed request that has already been
2809 * handled will pass through here and be processed immediately.
2811 CDEBUG(D_HA, "Next recovery transno: %llu"
2812 ", current: %llu, replaying\n",
2813 obd->obd_next_recovery_transno, transno);
2815 /* If the request has been replayed by update replay, then sends this
2816 * request to the recovery thread (replay_request_or_update()), where
2817 * it will be handled */
2818 spin_lock(&obd->obd_recovery_task_lock);
2819 if (transno < obd->obd_next_recovery_transno &&
2820 !is_req_replayed_by_update(req)) {
2821 /* Processing the queue right now, don't re-add. */
2822 LASSERT(list_empty(&req->rq_list));
2823 spin_unlock(&obd->obd_recovery_task_lock);
2826 spin_unlock(&obd->obd_recovery_task_lock);
2828 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2831 target_request_copy_get(req);
2832 if (!req->rq_export->exp_in_recovery) {
2833 target_request_copy_put(req);
2836 LASSERT(req->rq_export->exp_req_replay_needed);
2838 if (target_exp_enqueue_req_replay(req)) {
2839 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2840 target_request_copy_put(req);
2845 spin_lock(&obd->obd_recovery_task_lock);
2846 LASSERT(obd->obd_recovering);
2847 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2848 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2849 list_add_tail(&req->rq_list, &reqiter->rq_list);
2854 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2856 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2857 "has been claimed by another client");
2858 spin_unlock(&obd->obd_recovery_task_lock);
2859 target_exp_dequeue_req_replay(req);
2860 target_request_copy_put(req);
2866 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2868 obd->obd_requests_queued_for_recovery++;
2869 spin_unlock(&obd->obd_recovery_task_lock);
2870 wake_up(&obd->obd_next_transno_waitq);
2874 void target_committed_to_req(struct ptlrpc_request *req)
2876 struct obd_export *exp = req->rq_export;
2878 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2879 lustre_msg_set_last_committed(req->rq_repmsg,
2880 exp->exp_last_committed);
2882 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2883 "%d)", exp->exp_obd->obd_no_transno,
2884 req->rq_repmsg == NULL);
2886 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2887 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2890 #endif /* HAVE_SERVER_SUPPORT */
2893 * Packs current SLV and Limit into \a req.
2895 int target_pack_pool_reply(struct ptlrpc_request *req)
2897 struct obd_device *obd;
2900 /* Check that we still have all structures alive as this may
2901 * be some late RPC at shutdown time. */
2902 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2903 !exp_connect_lru_resize(req->rq_export))) {
2904 lustre_msg_set_slv(req->rq_repmsg, 0);
2905 lustre_msg_set_limit(req->rq_repmsg, 0);
2909 /* OBD is alive here as export is alive, which we checked above. */
2910 obd = req->rq_export->exp_obd;
2912 read_lock(&obd->obd_pool_lock);
2913 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2914 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2915 read_unlock(&obd->obd_pool_lock);
2920 static int target_send_reply_msg(struct ptlrpc_request *req,
2921 int rc, int fail_id)
2923 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2924 DEBUG_REQ(D_ERROR, req, "dropping reply");
2927 /* We can have a null rq_reqmsg in the event of bad signature or
2928 * no context when unwrapping */
2929 if (req->rq_reqmsg &&
2930 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2931 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2935 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2936 req->rq_status = rc;
2937 return ptlrpc_send_error(req, 1);
2939 DEBUG_REQ(D_NET, req, "sending reply");
2942 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2945 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2947 struct ptlrpc_service_part *svcpt;
2949 struct ptlrpc_reply_state *rs;
2950 struct obd_export *exp;
2953 if (req->rq_no_reply) {
2958 svcpt = req->rq_rqbd->rqbd_svcpt;
2959 rs = req->rq_reply_state;
2960 if (rs == NULL || !rs->rs_difficult) {
2962 target_send_reply_msg (req, rc, fail_id);
2967 /* must be an export if locks saved */
2968 LASSERT(req->rq_export != NULL);
2969 /* req/reply consistent */
2970 LASSERT(rs->rs_svcpt == svcpt);
2973 LASSERT(!rs->rs_scheduled);
2974 LASSERT(!rs->rs_scheduled_ever);
2975 LASSERT(!rs->rs_handled);
2976 LASSERT(!rs->rs_on_net);
2977 LASSERT(rs->rs_export == NULL);
2978 LASSERT(list_empty(&rs->rs_obd_list));
2979 LASSERT(list_empty(&rs->rs_exp_list));
2981 exp = class_export_get(req->rq_export);
2983 /* disable reply scheduling while I'm setting up */
2984 rs->rs_scheduled = 1;
2986 rs->rs_xid = req->rq_xid;
2987 rs->rs_transno = req->rq_transno;
2988 rs->rs_export = exp;
2989 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2991 spin_lock(&exp->exp_uncommitted_replies_lock);
2992 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2993 rs->rs_transno, exp->exp_last_committed);
2994 if (rs->rs_transno > exp->exp_last_committed) {
2995 /* not committed already */
2996 list_add_tail(&rs->rs_obd_list,
2997 &exp->exp_uncommitted_replies);
2999 spin_unlock(&exp->exp_uncommitted_replies_lock);
3001 spin_lock(&exp->exp_lock);
3002 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3003 spin_unlock(&exp->exp_lock);
3005 netrc = target_send_reply_msg(req, rc, fail_id);
3007 spin_lock(&svcpt->scp_rep_lock);
3009 atomic_inc(&svcpt->scp_nreps_difficult);
3012 /* error sending: reply is off the net. Also we need +1
3013 * reply ref until ptlrpc_handle_rs() is done
3014 * with the reply state (if the send was successful, there
3015 * would have been +1 ref for the net, which
3016 * reply_out_callback leaves alone) */
3018 ptlrpc_rs_addref(rs);
3021 spin_lock(&rs->rs_lock);
3022 if (rs->rs_transno <= exp->exp_last_committed ||
3023 (!rs->rs_on_net && !rs->rs_no_ack) ||
3024 list_empty(&rs->rs_exp_list) || /* completed already */
3025 list_empty(&rs->rs_obd_list)) {
3026 CDEBUG(D_HA, "Schedule reply immediately\n");
3027 ptlrpc_dispatch_difficult_reply(rs);
3029 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3030 rs->rs_scheduled = 0; /* allow notifier to schedule */
3032 spin_unlock(&rs->rs_lock);
3033 spin_unlock(&svcpt->scp_rep_lock);
3037 enum ldlm_mode lck_compat_array[] = {
3038 [LCK_EX] = LCK_COMPAT_EX,
3039 [LCK_PW] = LCK_COMPAT_PW,
3040 [LCK_PR] = LCK_COMPAT_PR,
3041 [LCK_CW] = LCK_COMPAT_CW,
3042 [LCK_CR] = LCK_COMPAT_CR,
3043 [LCK_NL] = LCK_COMPAT_NL,
3044 [LCK_GROUP] = LCK_COMPAT_GROUP,
3045 [LCK_COS] = LCK_COMPAT_COS,
3049 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3050 * not escape to the user level.
3052 int ldlm_error2errno(enum ldlm_error error)
3058 case ELDLM_LOCK_MATCHED:
3061 case ELDLM_LOCK_CHANGED:
3064 case ELDLM_LOCK_ABORTED:
3067 case ELDLM_LOCK_REPLACED:
3070 case ELDLM_NO_LOCK_DATA:
3073 case ELDLM_NAMESPACE_EXISTS:
3076 case ELDLM_BAD_NAMESPACE:
3080 if (((int)error) < 0) { /* cast to signed type */
3081 result = error; /* as ldlm_error can be unsigned */
3083 CERROR("Invalid DLM result code: %d\n", error);
3089 EXPORT_SYMBOL(ldlm_error2errno);
3092 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3094 enum ldlm_error ldlm_errno2error(int err_no)
3103 error = ELDLM_LOCK_CHANGED;
3106 error = ELDLM_LOCK_ABORTED;
3109 error = ELDLM_LOCK_REPLACED;
3112 error = ELDLM_NO_LOCK_DATA;
3115 error = ELDLM_NAMESPACE_EXISTS;
3118 error = ELDLM_BAD_NAMESPACE;
3126 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3127 void ldlm_dump_export_locks(struct obd_export *exp)
3129 spin_lock(&exp->exp_locks_list_guard);
3130 if (!list_empty(&exp->exp_locks_list)) {
3131 struct ldlm_lock *lock;
3133 CERROR("dumping locks for export %p,"
3134 "ignore if the unmount doesn't hang\n", exp);
3135 list_for_each_entry(lock, &exp->exp_locks_list,
3137 LDLM_ERROR(lock, "lock:");
3139 spin_unlock(&exp->exp_locks_list_guard);
3143 #ifdef HAVE_SERVER_SUPPORT
3144 static int target_bulk_timeout(void *data)
3147 /* We don't fail the connection here, because having the export
3148 * killed makes the (vital) call to commitrw very sad.
3153 static inline const char *bulk2type(struct ptlrpc_request *req)
3155 if (req->rq_bulk_read)
3157 if (req->rq_bulk_write)
3162 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3163 struct l_wait_info *lwi)
3165 struct ptlrpc_request *req = desc->bd_req;
3166 time64_t start = ktime_get_real_seconds();
3172 /* If there is eviction in progress, wait for it to finish. */
3173 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3174 *lwi = LWI_INTR(NULL, NULL);
3175 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3176 !atomic_read(&exp->exp_obd->
3177 obd_evict_inprogress),
3181 /* Check if client was evicted or reconnected already. */
3182 if (exp->exp_failed ||
3183 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3186 if (req->rq_bulk_read)
3187 rc = sptlrpc_svc_wrap_bulk(req, desc);
3189 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3190 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3191 else /* old version, bulk matchbits is rq_xid */
3192 req->rq_mbits = req->rq_xid;
3195 rc = ptlrpc_start_bulk_transfer(desc);
3199 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3200 bulk2type(req), rc);
3204 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3205 ptlrpc_abort_bulk(desc);
3209 /* limit actual bulk transfer to bulk_timeout seconds */
3210 deadline = start + bulk_timeout;
3211 if (deadline > req->rq_deadline)
3212 deadline = req->rq_deadline;
3215 time64_t timeoutl = deadline - ktime_get_real_seconds();
3216 long timeout_jiffies = timeoutl <= 0 ?
3217 1 : cfs_time_seconds(timeoutl);
3218 time64_t rq_deadline;
3220 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3221 cfs_time_seconds(1),
3222 target_bulk_timeout, desc);
3223 rc = l_wait_event(desc->bd_waitq,
3224 !ptlrpc_server_bulk_active(desc) ||
3227 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3229 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3230 /* Wait again if we changed rq_deadline. */
3231 rq_deadline = READ_ONCE(req->rq_deadline);
3232 deadline = start + bulk_timeout;
3233 if (deadline > rq_deadline)
3234 deadline = rq_deadline;
3235 } while (rc == -ETIMEDOUT &&
3236 deadline > ktime_get_real_seconds());
3238 if (rc == -ETIMEDOUT) {
3239 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3240 bulk2type(req), deadline - start,
3241 ktime_get_real_seconds() - deadline);
3242 ptlrpc_abort_bulk(desc);
3243 } else if (exp->exp_failed) {
3244 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3247 ptlrpc_abort_bulk(desc);
3248 } else if (exp->exp_conn_cnt >
3249 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3250 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3252 /* We don't reply anyway. */
3254 ptlrpc_abort_bulk(desc);
3255 } else if (desc->bd_failure) {
3256 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3258 /* XXX should this be a different errno? */
3261 if (req->rq_bulk_write)
3262 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3263 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3264 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3265 bulk2type(req), desc->bd_nob_transferred,
3267 /* XXX should this be a different errno? */
3274 EXPORT_SYMBOL(target_bulk_io);
3276 #endif /* HAVE_SERVER_SUPPORT */