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;
761 hdl = &exp->exp_imp_reverse->imp_remote_handle;
762 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
763 conn->cookie = exp->exp_handle.h_cookie;
764 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
765 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
769 target = exp->exp_obd;
771 /* Might be a re-connect after a partition. */
772 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
773 LCONSOLE_WARN("%s: already connected client %s (at %s) "
774 "with handle %#llx. Rejecting client "
775 "with the same UUID trying to reconnect "
776 "with handle %#llx\n", target->obd_name,
777 obd_uuid2str(&exp->exp_client_uuid),
778 obd_export_nid2str(exp),
779 hdl->cookie, conn->cookie);
780 memset(conn, 0, sizeof *conn);
781 /* target_handle_connect() treats EALREADY and
782 * -EALREADY differently. -EALREADY is an error
783 * (same UUID, different handle). */
787 if (!target->obd_recovering) {
788 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
789 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
790 obd_export_nid2str(exp));
791 GOTO(out_already, rc);
794 now = ktime_get_seconds();
795 deadline = jiffies_to_msecs(target->obd_recovery_timer.expires) /
797 if (now < deadline) {
798 struct target_distribute_txn_data *tdtd;
803 timeout = deadline - now;
804 tdtd = class_exp2tgt(exp)->lut_tdtd;
805 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
806 buf = tdtd->tdtd_show_update_logs_retrievers(
807 tdtd->tdtd_show_retrievers_cbdata,
811 LCONSOLE_WARN("%s: Recovery already passed deadline "
812 "%lld:%.02lld. It is due to DNE recovery "
813 "failed/stuck on the %d MDT(s):%s. "
814 "Please wait until all MDTs recovered "
815 "or abort the recovery by force.\n",
816 target->obd_name, timeout / 60,
818 buf ? buf : "unknown (not enough RAM)");
820 LCONSOLE_WARN("%s: Recovery already passed deadline "
821 "%lld:%.02lld. If you do not want to wait "
822 "more, please abort the recovery by "
823 "force.\n", target->obd_name,
824 timeout / 60, timeout % 60);
829 timeout = now - deadline;
830 LCONSOLE_WARN("%s: Recovery already passed deadline"
831 " %lld:%.02lld, It is most likely due to DNE"
832 " recovery is failed or stuck, please wait a"
833 " few more minutes or abort the recovery.\n",
834 target->obd_name, timeout / 60, timeout % 60);
838 conn->cookie = exp->exp_handle.h_cookie;
839 /* target_handle_connect() treats EALREADY and
840 * -EALREADY differently. EALREADY means we are
841 * doing a valid reconnect from the same client. */
846 check_and_start_recovery_timer(struct obd_device *obd,
847 struct ptlrpc_request *req, int new_client);
850 * update flags for import during reconnect process
852 static int rev_import_flags_update(struct obd_import *revimp,
853 struct ptlrpc_request *req)
856 struct obd_connect_data *data;
858 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
860 if (data->ocd_connect_flags & OBD_CONNECT_AT)
861 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
863 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
865 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
867 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
869 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
870 revimp->imp_client->cli_name,
871 libcfs_id2str(req->rq_peer), rc);
879 * Allocate a new reverse import for an export.
881 * \retval -errno in case error hit
882 * \retval 0 if reverse import correctly init
884 int rev_import_init(struct obd_export *export)
886 struct obd_device *obd = export->exp_obd;
887 struct obd_import *revimp;
889 LASSERT(export->exp_imp_reverse == NULL);
891 revimp = class_new_import(obd);
895 revimp->imp_remote_handle.cookie = 0ULL;
896 revimp->imp_client = &obd->obd_ldlm_client;
897 revimp->imp_dlm_fake = 1;
899 /* it is safe to connect import in new state as no sends possible */
900 spin_lock(&export->exp_lock);
901 export->exp_imp_reverse = revimp;
902 spin_unlock(&export->exp_lock);
903 class_import_put(revimp);
907 EXPORT_SYMBOL(rev_import_init);
910 * Handle reconnect for an export.
912 * \param exp export to handle reconnect process
913 * \param req client reconnect request
915 * \retval -rc in case securitfy flavor can't be changed
916 * \retval 0 in case none problems
918 static int rev_import_reconnect(struct obd_export *exp,
919 struct ptlrpc_request *req)
921 struct obd_import *revimp = exp->exp_imp_reverse;
922 struct lustre_handle *lh;
925 /* avoid sending a request until import flags are changed */
926 ptlrpc_import_enter_resend(revimp);
928 if (revimp->imp_connection != NULL)
929 ptlrpc_connection_put(revimp->imp_connection);
932 * client from recovery don't have a handle so we need to take from
933 * request. it may produce situation when wrong client connected
934 * to recovery as we trust a client uuid
936 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
937 revimp->imp_remote_handle = *lh;
939 /* unknown versions will be caught in
940 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
941 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
943 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
945 rc = rev_import_flags_update(revimp, req);
947 /* it is safe to still be in RECOVERY phase as we are not able
948 * to setup correct security flavor so requests are not able to
949 * be delivered correctly */
953 /* resend all rpc's via new connection */
954 return ptlrpc_import_recovery_state_machine(revimp);
957 int target_handle_connect(struct ptlrpc_request *req)
959 struct obd_device *target = NULL;
960 struct obd_export *export = NULL;
961 /* connect handle - filled from target_handle_reconnect in
963 struct lustre_handle conn;
964 struct lustre_handle *tmp;
965 struct obd_uuid cluuid;
970 bool mds_conn = false, lw_client = false, initial_conn = false;
971 bool mds_mds_conn = false;
972 bool new_mds_mds_conn = false;
973 struct obd_connect_data *data, *tmpdata;
975 lnet_nid_t *client_nid = NULL;
978 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
980 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
982 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
983 GOTO(out, rc = -EINVAL);
986 target = class_dev_by_str(str);
988 deuuidify(str, NULL, &target_start, &target_len);
989 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
990 "from %s (no target). If you are running "
991 "an HA pair check that the target is "
992 "mounted on the other server.\n", str,
993 libcfs_nid2str(req->rq_peer.nid));
994 GOTO(out, rc = -ENODEV);
997 spin_lock(&target->obd_dev_lock);
999 target->obd_conn_inprogress++;
1001 if (target->obd_stopping || !target->obd_set_up) {
1002 spin_unlock(&target->obd_dev_lock);
1004 deuuidify(str, NULL, &target_start, &target_len);
1005 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1006 target_len, target_start,
1007 libcfs_nid2str(req->rq_peer.nid),
1008 (target->obd_stopping ?
1009 "stopping" : "not set up"));
1010 GOTO(out, rc = -ENODEV);
1013 if (target->obd_no_conn) {
1014 spin_unlock(&target->obd_dev_lock);
1016 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1017 "from %s\n", target->obd_name,
1018 libcfs_nid2str(req->rq_peer.nid));
1019 GOTO(out, rc = -EAGAIN);
1022 spin_unlock(&target->obd_dev_lock);
1024 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1026 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1027 GOTO(out, rc = -EINVAL);
1030 obd_str2uuid(&cluuid, str);
1032 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1034 GOTO(out, rc = -EPROTO);
1038 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1040 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1042 GOTO(out, rc = -EPROTO);
1044 rc = req_capsule_server_pack(&req->rq_pill);
1048 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1049 /* Don't allow clients to connect that are using old 1.8 format
1050 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1051 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1052 * FULL20 flag should be set on all connections since 2.0, but no
1053 * longer affects behaviour.
1055 * Later this check will be disabled and the flag can be retired
1056 * completely once interop with 3.0 is no longer needed.
1058 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1059 GOTO(out, rc = -EPROTO);
1061 /* Don't allow liblustre clients to connect.
1062 * - testing was disabled in v2_2_50_0-61-g6a75d65
1063 * - building was disabled in v2_5_58_0-28-g7277179
1064 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1065 * - clients were refused connect for version difference > 0.0.1.32 */
1066 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1067 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1068 GOTO(out, rc = -EPROTO);
1072 /* Note: lw_client is needed in MDS-MDS failover during update log
1073 * processing, so we needs to allow lw_client to be connected at
1074 * anytime, instead of only the initial connection */
1075 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1077 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1078 initial_conn = true;
1079 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1080 mds_mds_conn = (data->ocd_connect_flags &
1081 OBD_CONNECT_MDS_MDS) != 0;
1083 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1084 * for Imperative Recovery connection from MGC to MGS.
1086 * Via check OBD_CONNECT_FID, we can distinguish whether
1087 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1088 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1091 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1092 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1093 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1094 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1095 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1096 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1098 /* We do not support the MDT-MDT interoperations with
1099 * different version MDT because of protocol changes. */
1100 if (unlikely(major != LUSTRE_MAJOR ||
1101 minor != LUSTRE_MINOR ||
1102 abs(patch - LUSTRE_PATCH) > 3)) {
1103 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1104 "connection from different version MDT "
1105 "(%d.%d.%d.%d) %s %s\n",
1106 target->obd_name, LUSTRE_MAJOR,
1107 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1108 major, minor, patch,
1109 OBD_OCD_VERSION_FIX(data->ocd_version),
1110 libcfs_nid2str(req->rq_peer.nid), str);
1112 GOTO(out, rc = -EPROTO);
1117 /* lctl gets a backstage, all-access pass. */
1118 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1119 goto dont_check_exports;
1121 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1125 /* We've found an export in the hash. */
1127 spin_lock(&export->exp_lock);
1129 if (export->exp_connecting) { /* bug 9635, et. al. */
1130 spin_unlock(&export->exp_lock);
1131 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1132 export->exp_obd->obd_name, export,
1133 libcfs_nid2str(req->rq_peer.nid));
1134 class_export_put(export);
1137 } else if ((mds_conn || (lw_client && initial_conn) ||
1138 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1139 export->exp_connection != NULL) {
1140 spin_unlock(&export->exp_lock);
1141 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1142 /* MDS or LWP reconnected after failover. */
1143 LCONSOLE_WARN("%s: Received %s connection from "
1144 "%s, removing former export from %s\n",
1145 target->obd_name, mds_conn ? "MDS" : "LWP",
1146 libcfs_nid2str(req->rq_peer.nid),
1147 libcfs_nid2str(export->exp_connection->c_peer.nid));
1149 /* New MDS connection from the same NID. */
1150 LCONSOLE_WARN("%s: Received new %s connection from "
1151 "%s, removing former export from same NID\n",
1152 target->obd_name, mds_conn ? "MDS" : "LWP",
1153 libcfs_nid2str(req->rq_peer.nid));
1156 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1157 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1158 /* Because exports between MDTs will always be
1159 * kept, let's do not fail such export if they
1160 * come from the same NID, otherwise it might
1161 * cause eviction between MDTs, which might
1162 * cause namespace inconsistency */
1163 spin_lock(&export->exp_lock);
1164 export->exp_connecting = 1;
1165 export->exp_conn_cnt = 0;
1166 spin_unlock(&export->exp_lock);
1167 conn.cookie = export->exp_handle.h_cookie;
1170 class_fail_export(export);
1171 class_export_put(export);
1175 } else if (export->exp_connection != NULL && initial_conn &&
1176 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1177 spin_unlock(&export->exp_lock);
1178 /* In MDS failover we have static UUID but NID can change. */
1179 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1180 "existing nid %s is already connected\n",
1181 target->obd_name, cluuid.uuid,
1182 libcfs_nid2str(req->rq_peer.nid),
1184 export->exp_connection->c_peer.nid));
1186 class_export_put(export);
1189 export->exp_connecting = 1;
1190 spin_unlock(&export->exp_lock);
1191 LASSERT(export->exp_obd == target);
1193 rc = target_handle_reconnect(&conn, export, &cluuid);
1196 /* If we found an export, we already unlocked. */
1199 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1200 } else if (req->rq_export == NULL &&
1201 atomic_read(&export->exp_rpc_count) > 0) {
1202 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1203 "still busy with %d references\n",
1204 target->obd_name, cluuid.uuid,
1205 libcfs_nid2str(req->rq_peer.nid),
1206 atomic_read(&export->exp_refcount));
1207 GOTO(out, rc = -EBUSY);
1208 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1210 if (!strstr(cluuid.uuid, "mdt"))
1211 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1212 "known client %s (at %s) because it "
1213 "is indicating it is a new client",
1214 target->obd_name, cluuid.uuid,
1215 libcfs_nid2str(req->rq_peer.nid));
1216 GOTO(out, rc = -EALREADY);
1218 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1225 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1226 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1227 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1228 export, ktime_get_real_seconds(),
1229 export ? export->exp_last_request_time : 0);
1231 /* If this is the first time a client connects, reset the recovery
1232 * timer. Discard lightweight connections which might be local. */
1233 if (!lw_client && rc == 0 && target->obd_recovering)
1234 check_and_start_recovery_timer(target, req, export == NULL);
1236 /* We want to handle EALREADY but *not* -EALREADY from
1237 * target_handle_reconnect(), return reconnection state in a flag. */
1238 if (rc == EALREADY) {
1239 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1245 /* Tell the client if we support replayable requests. */
1246 if (target->obd_replayable)
1247 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1248 client_nid = &req->rq_peer.nid;
1250 if (export == NULL) {
1251 /* allow lightweight connections during recovery */
1252 /* allow "new" MDT to be connected during recovery, since we
1253 * need retrieve recovery update records from it */
1254 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1258 int c; /* connected */
1259 int i; /* in progress */
1261 int s; /* stale/evicted */
1263 c = atomic_read(&target->obd_connected_clients);
1264 i = atomic_read(&target->obd_lock_replay_clients);
1265 k = target->obd_max_recoverable_clients;
1266 s = target->obd_stale_clients;
1267 t = jiffies_to_msecs(target->obd_recovery_timer.expires);
1269 now = ktime_get_seconds();
1272 msg = "already passed deadline";
1275 msg = "to recover in";
1278 LCONSOLE_WARN("%s: Denying connection for new client %s"
1279 "(at %s), waiting for %d known clients "
1280 "(%d recovered, %d in progress, and %d "
1281 "evicted) %s %lld:%.02lld\n",
1282 target->obd_name, cluuid.uuid,
1283 libcfs_nid2str(req->rq_peer.nid), k,
1284 c - i, i, s, msg, t / 60, t % 60);
1288 rc = obd_connect(req->rq_svc_thread->t_env,
1289 &export, target, &cluuid, data,
1291 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1292 lustre_msg_add_op_flags(req->rq_repmsg,
1293 MSG_CONNECT_RECOVERING);
1295 conn.cookie = export->exp_handle.h_cookie;
1296 rc = rev_import_init(export);
1300 new_mds_mds_conn = true;
1303 rc = obd_reconnect(req->rq_svc_thread->t_env,
1304 export, target, &cluuid, data, client_nid);
1309 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1310 data->ocd_instance = target->u.obt.obt_instance;
1312 /* Return only the parts of obd_connect_data that we understand, so the
1313 * client knows that we don't understand the rest. */
1315 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1317 tmpdata = req_capsule_server_get(&req->rq_pill,
1319 /* Don't use struct assignment here, because the client reply
1320 * buffer may be smaller/larger than the local struct
1321 * obd_connect_data. */
1322 memcpy(tmpdata, data, min(tmpsize, size));
1325 /* If the client and the server are the same node, we will already
1326 * have an export that really points to the client's DLM export,
1327 * because we have a shared handles table.
1329 * XXX this will go away when shaver stops sending the "connect" handle
1330 * in the real "remote handle" field of the request --phik 24 Apr 2003
1332 ptlrpc_request_change_export(req, export);
1334 spin_lock(&export->exp_lock);
1335 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1336 spin_unlock(&export->exp_lock);
1337 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1338 "or equal conn_cnt: %d >= %d\n",
1339 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1340 export->exp_conn_cnt,
1341 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1343 GOTO(out, rc = -EALREADY);
1345 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1346 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1347 spin_unlock(&export->exp_lock);
1349 if (export->exp_connection != NULL) {
1350 /* Check to see if connection came from another NID. */
1351 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1352 !hlist_unhashed(&export->exp_nid_hash))
1353 cfs_hash_del(export->exp_obd->obd_nid_hash,
1354 &export->exp_connection->c_peer.nid,
1355 &export->exp_nid_hash);
1357 ptlrpc_connection_put(export->exp_connection);
1360 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1363 if (hlist_unhashed(&export->exp_nid_hash))
1364 cfs_hash_add(export->exp_obd->obd_nid_hash,
1365 &export->exp_connection->c_peer.nid,
1366 &export->exp_nid_hash);
1368 lustre_msg_set_handle(req->rq_repmsg, &conn);
1370 rc = rev_import_reconnect(export, req);
1374 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1376 __u64 transno = data->ocd_transno;
1378 spin_lock(&export->exp_lock);
1379 /* possible race with class_disconnect_stale_exports,
1380 * export may be already in the eviction process */
1381 if (export->exp_failed) {
1382 spin_unlock(&export->exp_lock);
1383 GOTO(out, rc = -ENODEV);
1385 export->exp_in_recovery = 1;
1386 export->exp_req_replay_needed = 1;
1387 export->exp_lock_replay_needed = 1;
1388 spin_unlock(&export->exp_lock);
1390 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1391 MSG_CONNECT_TRANSNO);
1392 if (has_transno && transno == 0)
1393 CWARN("Connect with zero transno!\n");
1395 if (has_transno && transno > 0 &&
1396 transno < target->obd_next_recovery_transno &&
1397 transno > target->obd_last_committed) {
1398 /* Another way is to use cmpxchg() to be lock-free. */
1399 spin_lock(&target->obd_recovery_task_lock);
1400 if (transno < target->obd_next_recovery_transno)
1401 target->obd_next_recovery_transno = transno;
1402 spin_unlock(&target->obd_recovery_task_lock);
1405 atomic_inc(&target->obd_req_replay_clients);
1406 atomic_inc(&target->obd_lock_replay_clients);
1407 /* Note: MDS-MDS connection is allowed to be connected during
1408 * recovery, no matter if the exports needs to be recoveried.
1409 * Because we need retrieve updates logs from all other MDTs.
1410 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1411 * also needs to be increased to match other recovery checking
1413 if (new_mds_mds_conn)
1414 target->obd_max_recoverable_clients++;
1415 if (atomic_inc_return(&target->obd_connected_clients) ==
1416 target->obd_max_recoverable_clients)
1417 wake_up(&target->obd_next_transno_waitq);
1420 /* Tell the client we're in recovery, when client is involved in it. */
1421 if (target->obd_recovering && !lw_client)
1422 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1426 spin_lock(&export->exp_lock);
1427 export->exp_connecting = 0;
1428 spin_unlock(&export->exp_lock);
1430 class_export_put(export);
1432 if (target != NULL) {
1433 spin_lock(&target->obd_dev_lock);
1434 target->obd_conn_inprogress--;
1435 spin_unlock(&target->obd_dev_lock);
1436 class_decref(target, "find", current);
1438 req->rq_status = rc;
1442 int target_handle_disconnect(struct ptlrpc_request *req)
1447 rc = req_capsule_server_pack(&req->rq_pill);
1451 /* Keep the rq_export around so we can send the reply. */
1452 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1457 void target_destroy_export(struct obd_export *exp)
1459 struct obd_import *imp = NULL;
1460 /* exports created from last_rcvd data, and "fake"
1461 exports created by lctl don't have an import */
1462 spin_lock(&exp->exp_lock);
1463 if (exp->exp_imp_reverse != NULL) {
1464 imp = exp->exp_imp_reverse;
1465 exp->exp_imp_reverse = NULL;
1467 spin_unlock(&exp->exp_lock);
1469 client_destroy_import(imp);
1471 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1472 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1473 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1474 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1476 EXPORT_SYMBOL(target_destroy_export);
1479 * Recovery functions
1481 static void target_request_copy_get(struct ptlrpc_request *req)
1483 class_export_rpc_inc(req->rq_export);
1484 LASSERT(list_empty(&req->rq_list));
1485 INIT_LIST_HEAD(&req->rq_replay_list);
1487 /* Increase refcount to keep request in queue. */
1488 atomic_inc(&req->rq_refcount);
1489 /* Let export know it has replays to be handled. */
1490 atomic_inc(&req->rq_export->exp_replay_count);
1493 static void target_request_copy_put(struct ptlrpc_request *req)
1495 LASSERT(list_empty(&req->rq_replay_list));
1496 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1498 atomic_dec(&req->rq_export->exp_replay_count);
1499 class_export_rpc_dec(req->rq_export);
1500 ptlrpc_server_drop_request(req);
1503 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1505 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1506 struct obd_export *exp = req->rq_export;
1507 struct ptlrpc_request *reqiter;
1508 struct ptlrpc_request *dup_req = NULL;
1513 spin_lock(&exp->exp_lock);
1514 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1516 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1524 /* We expect it with RESENT and REPLAY flags. */
1525 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1526 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1527 CERROR("invalid flags %x of resent replay\n",
1528 lustre_msg_get_flags(req->rq_reqmsg));
1530 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1533 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1535 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1536 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1540 list_add_tail(&req->rq_replay_list,
1541 &exp->exp_req_replay_queue);
1544 spin_unlock(&exp->exp_lock);
1548 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1550 LASSERT(!list_empty(&req->rq_replay_list));
1551 LASSERT(req->rq_export);
1553 spin_lock(&req->rq_export->exp_lock);
1554 list_del_init(&req->rq_replay_list);
1555 spin_unlock(&req->rq_export->exp_lock);
1558 static void target_finish_recovery(struct lu_target *lut)
1560 struct obd_device *obd = lut->lut_obd;
1563 /* Only log a recovery message when recovery has occurred. */
1564 if (obd->obd_recovery_start) {
1565 time64_t now = ktime_get_real_seconds();
1566 time64_t elapsed_time;
1568 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1569 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1570 "%d recovered and %d %s evicted.\n", obd->obd_name,
1571 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1572 obd->obd_max_recoverable_clients,
1573 atomic_read(&obd->obd_connected_clients),
1574 obd->obd_stale_clients,
1575 obd->obd_stale_clients == 1 ? "was" : "were");
1578 ldlm_reprocess_recovery_done(obd->obd_namespace);
1579 spin_lock(&obd->obd_recovery_task_lock);
1580 if (!list_empty(&obd->obd_req_replay_queue) ||
1581 !list_empty(&obd->obd_lock_replay_queue) ||
1582 !list_empty(&obd->obd_final_req_queue)) {
1583 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1585 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1586 list_empty(&obd->obd_lock_replay_queue) ? \
1588 list_empty(&obd->obd_final_req_queue) ? \
1590 spin_unlock(&obd->obd_recovery_task_lock);
1593 spin_unlock(&obd->obd_recovery_task_lock);
1595 obd->obd_recovery_end = ktime_get_real_seconds();
1597 /* When recovery finished, cleanup orphans on MDS and OST. */
1598 if (obd->obd_type && OBP(obd, postrecov)) {
1599 int rc = OBP(obd, postrecov)(obd);
1602 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1608 static void abort_req_replay_queue(struct obd_device *obd)
1610 struct ptlrpc_request *req, *n;
1611 struct list_head abort_list;
1613 INIT_LIST_HEAD(&abort_list);
1614 spin_lock(&obd->obd_recovery_task_lock);
1615 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1616 spin_unlock(&obd->obd_recovery_task_lock);
1617 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1618 DEBUG_REQ(D_WARNING, req, "aborted:");
1619 req->rq_status = -ENOTCONN;
1620 if (ptlrpc_error(req)) {
1621 DEBUG_REQ(D_ERROR, req,
1622 "failed abort_req_reply; skipping");
1624 target_exp_dequeue_req_replay(req);
1625 target_request_copy_put(req);
1629 static void abort_lock_replay_queue(struct obd_device *obd)
1631 struct ptlrpc_request *req, *n;
1632 struct list_head abort_list;
1634 INIT_LIST_HEAD(&abort_list);
1635 spin_lock(&obd->obd_recovery_task_lock);
1636 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1637 spin_unlock(&obd->obd_recovery_task_lock);
1638 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1639 DEBUG_REQ(D_ERROR, req, "aborted:");
1640 req->rq_status = -ENOTCONN;
1641 if (ptlrpc_error(req)) {
1642 DEBUG_REQ(D_ERROR, req,
1643 "failed abort_lock_reply; skipping");
1645 target_request_copy_put(req);
1649 /* Called from a cleanup function if the device is being cleaned up
1650 forcefully. The exports should all have been disconnected already,
1651 the only thing left to do is
1652 - clear the recovery flags
1654 - free queued requests and replies, but don't send replies
1655 Because the obd_stopping flag is set, no new requests should be received.
1658 void target_cleanup_recovery(struct obd_device *obd)
1660 struct ptlrpc_request *req, *n;
1661 struct list_head clean_list;
1663 INIT_LIST_HEAD(&clean_list);
1664 spin_lock(&obd->obd_dev_lock);
1665 if (!obd->obd_recovering) {
1666 spin_unlock(&obd->obd_dev_lock);
1670 obd->obd_recovering = obd->obd_abort_recovery = 0;
1671 spin_unlock(&obd->obd_dev_lock);
1673 spin_lock(&obd->obd_recovery_task_lock);
1674 target_cancel_recovery_timer(obd);
1675 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1676 spin_unlock(&obd->obd_recovery_task_lock);
1678 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1679 LASSERT(req->rq_reply_state == NULL);
1680 target_exp_dequeue_req_replay(req);
1681 target_request_copy_put(req);
1684 spin_lock(&obd->obd_recovery_task_lock);
1685 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1686 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1687 spin_unlock(&obd->obd_recovery_task_lock);
1689 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1690 LASSERT(req->rq_reply_state == NULL);
1691 target_request_copy_put(req);
1696 EXPORT_SYMBOL(target_cleanup_recovery);
1698 /* obd_recovery_task_lock should be held */
1699 void target_cancel_recovery_timer(struct obd_device *obd)
1701 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1702 del_timer(&obd->obd_recovery_timer);
1705 static void target_start_recovery_timer(struct obd_device *obd)
1707 if (obd->obd_recovery_start != 0)
1710 spin_lock(&obd->obd_dev_lock);
1711 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1712 spin_unlock(&obd->obd_dev_lock);
1716 LASSERT(obd->obd_recovery_timeout != 0);
1718 if (obd->obd_recovery_start != 0) {
1719 spin_unlock(&obd->obd_dev_lock);
1723 mod_timer(&obd->obd_recovery_timer,
1724 jiffies + cfs_time_seconds(obd->obd_recovery_timeout));
1725 obd->obd_recovery_start = ktime_get_real_seconds();
1726 spin_unlock(&obd->obd_dev_lock);
1728 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1730 obd->obd_recovery_timeout / 60,
1731 obd->obd_recovery_timeout % 60,
1732 obd->obd_max_recoverable_clients,
1733 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1734 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1738 * extend recovery window.
1740 * if @extend is true, extend recovery window to have @drt remaining at least;
1741 * otherwise, make sure the recovery timeout value is not less than @drt.
1743 static void extend_recovery_timer(struct obd_device *obd, time64_t drt,
1751 spin_lock(&obd->obd_dev_lock);
1752 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1753 spin_unlock(&obd->obd_dev_lock);
1756 LASSERT(obd->obd_recovery_start != 0);
1758 now = ktime_get_real_seconds();
1759 to = obd->obd_recovery_timeout;
1760 end = obd->obd_recovery_start + to;
1763 if (extend && (drt > left)) {
1765 } else if (!extend && (drt > to)) {
1769 if (to > obd->obd_recovery_time_hard) {
1770 to = obd->obd_recovery_time_hard;
1771 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1772 obd->obd_name, to, extend);
1775 if (obd->obd_recovery_timeout < to) {
1776 obd->obd_recovery_timeout = to;
1777 end = obd->obd_recovery_start + to;
1778 mod_timer(&obd->obd_recovery_timer,
1779 jiffies + cfs_time_seconds(end - now));
1781 spin_unlock(&obd->obd_dev_lock);
1783 CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1784 obd->obd_name, (s64)(end - now));
1787 /* Reset the timer with each new client connection */
1789 * This timer is actually reconnect_timer, which is for making sure
1790 * the total recovery window is at least as big as my reconnect
1791 * attempt timing. So the initial recovery time_out will be set to
1792 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1793 * from client is bigger than this, then the recovery time_out will
1794 * be extended to make sure the client could be reconnected, in the
1795 * process, the timeout from the new client should be ignored.
1799 check_and_start_recovery_timer(struct obd_device *obd,
1800 struct ptlrpc_request *req,
1803 time64_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1804 struct obd_device_target *obt = &obd->u.obt;
1806 if (!new_client && service_time)
1807 /* Teach server about old server's estimates, as first guess
1808 * at how long new requests will take. */
1809 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1812 target_start_recovery_timer(obd);
1814 /* Convert the service time to RPC timeout,
1815 * and reuse service_time to limit stack usage.
1817 service_time = at_est2timeout(service_time);
1819 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1820 service_time < at_extra)
1821 service_time = at_extra;
1823 /* We expect other clients to timeout within service_time, then try
1824 * to reconnect, then try the failover server. The max delay between
1825 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1826 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1828 LASSERT(obt->obt_magic == OBT_MAGIC);
1829 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1830 if (service_time > obd->obd_recovery_timeout && !new_client)
1831 extend_recovery_timer(obd, service_time, false);
1834 /** Health checking routines */
1835 static inline int exp_connect_healthy(struct obd_export *exp)
1837 return (exp->exp_in_recovery);
1840 /** if export done req_replay or has replay in queue */
1841 static inline int exp_req_replay_healthy(struct obd_export *exp)
1843 return (!exp->exp_req_replay_needed ||
1844 atomic_read(&exp->exp_replay_count) > 0);
1848 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1850 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1851 exp_req_replay_healthy(exp);
1854 /** if export done lock_replay or has replay in queue */
1855 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1857 return (!exp->exp_lock_replay_needed ||
1858 atomic_read(&exp->exp_replay_count) > 0);
1861 static inline int exp_vbr_healthy(struct obd_export *exp)
1863 return (!exp->exp_vbr_failed);
1866 static inline int exp_finished(struct obd_export *exp)
1868 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1871 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1873 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1877 static int check_for_next_transno(struct lu_target *lut)
1879 struct ptlrpc_request *req = NULL;
1880 struct obd_device *obd = lut->lut_obd;
1881 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1882 int wake_up = 0, connected, completed, queue_len;
1883 __u64 req_transno = 0;
1884 __u64 update_transno = 0;
1885 __u64 next_transno = 0;
1888 spin_lock(&obd->obd_recovery_task_lock);
1889 if (!list_empty(&obd->obd_req_replay_queue)) {
1890 req = list_entry(obd->obd_req_replay_queue.next,
1891 struct ptlrpc_request, rq_list);
1892 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1896 update_transno = distribute_txn_get_next_transno(tdtd);
1898 connected = atomic_read(&obd->obd_connected_clients);
1899 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1900 queue_len = obd->obd_requests_queued_for_recovery;
1901 next_transno = obd->obd_next_recovery_transno;
1903 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1904 "req_transno: %llu, next_transno: %llu\n",
1905 obd->obd_max_recoverable_clients, connected, completed,
1906 queue_len, req_transno, next_transno);
1908 if (obd->obd_abort_recovery) {
1909 CDEBUG(D_HA, "waking for aborted recovery\n");
1911 } else if (obd->obd_recovery_expired) {
1912 CDEBUG(D_HA, "waking for expired recovery\n");
1914 } else if (tdtd != NULL && req != NULL &&
1915 is_req_replayed_by_update(req)) {
1916 LASSERTF(req_transno < next_transno, "req_transno %llu"
1917 "next_transno%llu\n", req_transno, next_transno);
1918 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1921 } else if (req_transno == next_transno ||
1922 (update_transno != 0 && update_transno <= next_transno)) {
1923 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1925 } else if (queue_len > 0 &&
1926 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1927 /** handle gaps occured due to lost reply or VBR */
1928 LASSERTF(req_transno >= next_transno,
1929 "req_transno: %llu, next_transno: %llu\n",
1930 req_transno, next_transno);
1932 "%s: waking for gap in transno, VBR is %s (skip: "
1933 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1934 ", next_update %lld last_committed: %lld)\n",
1935 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1936 next_transno, queue_len, completed, connected,
1937 req_transno, update_transno, obd->obd_last_committed);
1938 obd->obd_next_recovery_transno = req_transno;
1940 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1941 CDEBUG(D_HA, "waking for completed recovery\n");
1943 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1944 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1945 " by fail_lock, waking up (%lld)\n", next_transno);
1946 obd->obd_next_recovery_transno = req_transno;
1949 spin_unlock(&obd->obd_recovery_task_lock);
1953 static int check_for_next_lock(struct lu_target *lut)
1955 struct obd_device *obd = lut->lut_obd;
1958 spin_lock(&obd->obd_recovery_task_lock);
1959 if (!list_empty(&obd->obd_lock_replay_queue)) {
1960 CDEBUG(D_HA, "waking for next lock\n");
1962 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1963 CDEBUG(D_HA, "waking for completed lock replay\n");
1965 } else if (obd->obd_abort_recovery) {
1966 CDEBUG(D_HA, "waking for aborted recovery\n");
1968 } else if (obd->obd_recovery_expired) {
1969 CDEBUG(D_HA, "waking for expired recovery\n");
1972 spin_unlock(&obd->obd_recovery_task_lock);
1978 * wait for recovery events,
1979 * check its status with help of check_routine
1980 * evict dead clients via health_check
1982 static int target_recovery_overseer(struct lu_target *lut,
1983 int (*check_routine)(struct lu_target *),
1984 int (*health_check)(struct obd_export *))
1986 struct obd_device *obd = lut->lut_obd;
1987 struct target_distribute_txn_data *tdtd;
1991 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
1992 now = ktime_get_seconds();
1993 if (now - last > 600) {
1994 LCONSOLE_INFO("%s: in recovery but waiting for "
1995 "the first client to connect\n",
2000 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
2001 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2002 __u64 next_update_transno = 0;
2004 /* Only abort the recovery if there are no update recovery
2005 * left in the queue */
2006 spin_lock(&obd->obd_recovery_task_lock);
2007 if (lut->lut_tdtd != NULL) {
2008 next_update_transno =
2009 distribute_txn_get_next_transno(lut->lut_tdtd);
2011 tdtd = lut->lut_tdtd;
2012 /* If next_update_transno == 0, it probably because
2013 * updatelog retrieve threads did not get any records
2014 * yet, let's wait those threads stopped */
2015 if (next_update_transno == 0) {
2016 struct l_wait_info lwi = { 0 };
2018 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2020 &tdtd->tdtd_recovery_threads_count) == 0,
2023 next_update_transno =
2024 distribute_txn_get_next_transno(
2029 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2030 obd->obd_next_recovery_transno = next_update_transno;
2031 spin_unlock(&obd->obd_recovery_task_lock);
2032 /* Disconnect unfinished exports from clients, and
2033 * keep connection from MDT to make sure the update
2034 * recovery will still keep trying until some one
2035 * manually abort the recovery */
2036 class_disconnect_stale_exports(obd,
2037 exp_finished_or_from_mdt);
2038 /* Abort all of replay and replay lock req from
2040 abort_req_replay_queue(obd);
2041 abort_lock_replay_queue(obd);
2042 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2043 ")in the queue.\n", obd->obd_name,
2044 next_update_transno);
2046 obd->obd_abort_recovery = 1;
2047 spin_unlock(&obd->obd_recovery_task_lock);
2048 CWARN("%s recovery is aborted by hard timeout\n",
2053 while (wait_event_timeout(obd->obd_next_transno_waitq,
2055 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2056 /* wait indefinitely for event, but don't trigger watchdog */;
2058 if (obd->obd_abort_recovery) {
2059 CWARN("recovery is aborted, evict exports in recovery\n");
2060 if (lut->lut_tdtd != NULL) {
2061 struct l_wait_info lwi = { 0 };
2063 tdtd = lut->lut_tdtd;
2064 /* Let's wait all of the update log recovery thread
2066 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2067 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2069 /* Then abort the update recovery list */
2070 dtrq_list_destroy(lut->lut_tdtd);
2073 /** evict exports which didn't finish recovery yet */
2074 class_disconnect_stale_exports(obd, exp_finished);
2076 } else if (obd->obd_recovery_expired) {
2077 obd->obd_recovery_expired = 0;
2078 /** If some clients died being recovered, evict them */
2079 LCONSOLE_WARN("%s: recovery is timed out, "
2080 "evict stale exports\n", obd->obd_name);
2081 /** evict cexports with no replay in queue, they are stalled */
2082 class_disconnect_stale_exports(obd, health_check);
2084 /** continue with VBR */
2085 spin_lock(&obd->obd_dev_lock);
2086 obd->obd_version_recov = 1;
2087 spin_unlock(&obd->obd_dev_lock);
2089 * reset timer, recovery will proceed with versions now,
2090 * timeout is set just to handle reconnection delays
2092 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2093 /** Wait for recovery events again, after evicting bad clients */
2099 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2101 struct obd_device *obd = lut->lut_obd;
2102 struct ptlrpc_request *req = NULL;
2104 CDEBUG(D_HA, "Waiting for lock\n");
2105 if (target_recovery_overseer(lut, check_for_next_lock,
2106 exp_lock_replay_healthy))
2107 abort_lock_replay_queue(obd);
2109 spin_lock(&obd->obd_recovery_task_lock);
2110 if (!list_empty(&obd->obd_lock_replay_queue)) {
2111 req = list_entry(obd->obd_lock_replay_queue.next,
2112 struct ptlrpc_request, rq_list);
2113 list_del_init(&req->rq_list);
2114 spin_unlock(&obd->obd_recovery_task_lock);
2116 spin_unlock(&obd->obd_recovery_task_lock);
2117 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2118 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2119 /** evict exports failed VBR */
2120 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2125 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2127 struct ptlrpc_request *req = NULL;
2129 spin_lock(&obd->obd_recovery_task_lock);
2130 if (!list_empty(&obd->obd_final_req_queue)) {
2131 req = list_entry(obd->obd_final_req_queue.next,
2132 struct ptlrpc_request, rq_list);
2133 list_del_init(&req->rq_list);
2134 spin_unlock(&obd->obd_recovery_task_lock);
2135 if (req->rq_export->exp_in_recovery) {
2136 spin_lock(&req->rq_export->exp_lock);
2137 req->rq_export->exp_in_recovery = 0;
2138 spin_unlock(&req->rq_export->exp_lock);
2141 spin_unlock(&obd->obd_recovery_task_lock);
2146 static void handle_recovery_req(struct ptlrpc_thread *thread,
2147 struct ptlrpc_request *req,
2148 svc_handler_t handler)
2153 * export can be evicted during recovery, no need to handle replays for
2154 * it after that, discard such request silently
2156 if (req->rq_export->exp_disconnected)
2159 req->rq_session.lc_thread = thread;
2160 req->rq_svc_thread = thread;
2161 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2163 /* thread context */
2164 lu_context_enter(&thread->t_env->le_ctx);
2166 lu_context_exit(&thread->t_env->le_ctx);
2168 /* don't reset timer for final stage */
2169 if (!exp_finished(req->rq_export)) {
2170 time64_t to = obd_timeout;
2173 * Add request timeout to the recovery time so next request from
2174 * this client may come in recovery time
2177 struct ptlrpc_service_part *svcpt;
2179 svcpt = req->rq_rqbd->rqbd_svcpt;
2180 /* If the server sent early reply for this request,
2181 * the client will recalculate the timeout according to
2182 * current server estimate service time, so we will
2183 * use the maxium timeout here for waiting the client
2184 * sending the next req */
2185 to = max((int)at_est2timeout(
2186 at_get(&svcpt->scp_at_estimate)),
2187 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2188 /* Add 2 net_latency, one for balance rq_deadline
2189 * (see ptl_send_rpc), one for resend the req to server,
2190 * Note: client will pack net_latency in replay req
2191 * (see ptlrpc_replay_req) */
2192 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2194 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2199 /** Checking routines for recovery */
2200 static int check_for_recovery_ready(struct lu_target *lut)
2202 struct obd_device *obd = lut->lut_obd;
2203 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2205 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2206 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2207 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2208 obd->obd_recovery_expired);
2210 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2211 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2212 if (clnts + obd->obd_stale_clients <
2213 obd->obd_max_recoverable_clients)
2217 if (lut->lut_tdtd != NULL) {
2218 if (!lut->lut_tdtd->tdtd_replay_ready &&
2219 !obd->obd_abort_recovery) {
2220 /* Let's extend recovery timer, in case the recovery
2221 * timer expired, and some clients got evicted */
2222 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2224 CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2225 obd->obd_name, obd->obd_recovery_timeout);
2234 REQUEST_RECOVERY = 1,
2235 UPDATE_RECOVERY = 2,
2238 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2242 if (!list_empty(&obd->obd_req_replay_queue)) {
2243 struct ptlrpc_request *req;
2245 req = list_entry(obd->obd_req_replay_queue.next,
2246 struct ptlrpc_request, rq_list);
2247 transno = lustre_msg_get_transno(req->rq_reqmsg);
2253 static __u64 get_next_transno(struct lu_target *lut, int *type)
2255 struct obd_device *obd = lut->lut_obd;
2256 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2258 __u64 update_transno;
2261 transno = get_next_replay_req_transno(obd);
2263 *type = REQUEST_RECOVERY;
2268 update_transno = distribute_txn_get_next_transno(tdtd);
2269 if (transno == 0 || (transno >= update_transno &&
2270 update_transno != 0)) {
2271 transno = update_transno;
2273 *type = UPDATE_RECOVERY;
2280 * drop duplicate replay request
2282 * Because the operation has been replayed by update recovery, the request
2283 * with the same transno will be dropped and also notify the client to send
2284 * next replay request.
2286 * \param[in] env execution environment
2287 * \param[in] obd failover obd device
2288 * \param[in] req request to be dropped
2290 static void drop_duplicate_replay_req(struct lu_env *env,
2291 struct obd_device *obd,
2292 struct ptlrpc_request *req)
2294 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2295 " update records are found.\n",
2296 lustre_msg_get_transno(req->rq_reqmsg),
2297 libcfs_nid2str(req->rq_peer.nid));
2299 /* Right now, only for MDS reint operation update replay and
2300 * normal request replay can have the same transno */
2301 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2302 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2303 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2304 if (likely(req->rq_export))
2305 target_committed_to_req(req);
2306 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2307 target_send_reply(req, req->rq_status, 0);
2309 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2310 libcfs_nid2str(req->rq_peer.nid));
2312 target_exp_dequeue_req_replay(req);
2313 target_request_copy_put(req);
2314 obd->obd_replayed_requests++;
2317 static void replay_request_or_update(struct lu_env *env,
2318 struct lu_target *lut,
2319 struct target_recovery_data *trd,
2320 struct ptlrpc_thread *thread)
2322 struct obd_device *obd = lut->lut_obd;
2323 struct ptlrpc_request *req = NULL;
2328 CDEBUG(D_HA, "Waiting for transno %lld\n",
2329 obd->obd_next_recovery_transno);
2331 /* Replay all of request and update by transno */
2333 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2335 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2337 /** It is needed to extend recovery window above
2338 * recovery_time_soft. Extending is possible only in the
2339 * end of recovery window (see more details in
2340 * handle_recovery_req()).
2342 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2344 if (target_recovery_overseer(lut, check_for_next_transno,
2345 exp_req_replay_healthy_or_from_mdt)) {
2346 abort_req_replay_queue(obd);
2347 abort_lock_replay_queue(obd);
2351 spin_lock(&obd->obd_recovery_task_lock);
2352 transno = get_next_transno(lut, &type);
2353 if (type == REQUEST_RECOVERY && transno != 0) {
2354 /* Drop replay request from client side, if the
2355 * replay has been executed by update with the
2357 req = list_entry(obd->obd_req_replay_queue.next,
2358 struct ptlrpc_request, rq_list);
2360 list_del_init(&req->rq_list);
2361 obd->obd_requests_queued_for_recovery--;
2362 spin_unlock(&obd->obd_recovery_task_lock);
2364 /* Let's check if the request has been redone by
2366 if (is_req_replayed_by_update(req)) {
2367 struct distribute_txn_replay_req *dtrq;
2369 dtrq = distribute_txn_lookup_finish_list(tdtd,
2371 LASSERT(dtrq != NULL);
2372 spin_lock(&tdtd->tdtd_replay_list_lock);
2373 list_del_init(&dtrq->dtrq_list);
2374 spin_unlock(&tdtd->tdtd_replay_list_lock);
2377 drop_duplicate_replay_req(env, obd, req);
2382 LASSERT(trd->trd_processing_task == current_pid());
2383 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2384 lustre_msg_get_transno(req->rq_reqmsg),
2385 libcfs_nid2str(req->rq_peer.nid));
2387 handle_recovery_req(thread, req,
2388 trd->trd_recovery_handler);
2390 * bz18031: increase next_recovery_transno before
2391 * target_request_copy_put() will drop exp_rpc reference
2393 spin_lock(&obd->obd_recovery_task_lock);
2394 obd->obd_next_recovery_transno++;
2395 spin_unlock(&obd->obd_recovery_task_lock);
2396 target_exp_dequeue_req_replay(req);
2397 target_request_copy_put(req);
2398 obd->obd_replayed_requests++;
2399 } else if (type == UPDATE_RECOVERY && transno != 0) {
2400 struct distribute_txn_replay_req *dtrq;
2403 spin_unlock(&obd->obd_recovery_task_lock);
2405 LASSERT(tdtd != NULL);
2406 dtrq = distribute_txn_get_next_req(tdtd);
2407 lu_context_enter(&thread->t_env->le_ctx);
2408 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2409 lu_context_exit(&thread->t_env->le_ctx);
2410 extend_recovery_timer(obd, obd_timeout, true);
2412 if (rc == 0 && dtrq->dtrq_xid != 0) {
2413 CDEBUG(D_HA, "Move x%llu t%llu"
2414 " to finish list\n", dtrq->dtrq_xid,
2415 dtrq->dtrq_master_transno);
2417 /* Add it to the replay finish list */
2418 spin_lock(&tdtd->tdtd_replay_list_lock);
2419 list_add(&dtrq->dtrq_list,
2420 &tdtd->tdtd_replay_finish_list);
2421 spin_unlock(&tdtd->tdtd_replay_list_lock);
2423 spin_lock(&obd->obd_recovery_task_lock);
2424 if (transno == obd->obd_next_recovery_transno)
2425 obd->obd_next_recovery_transno++;
2427 obd->obd_next_recovery_transno)
2428 obd->obd_next_recovery_transno =
2430 spin_unlock(&obd->obd_recovery_task_lock);
2435 spin_unlock(&obd->obd_recovery_task_lock);
2437 LASSERT(list_empty(&obd->obd_req_replay_queue));
2438 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2439 /** evict exports failed VBR */
2440 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2446 static int target_recovery_thread(void *arg)
2448 struct lu_target *lut = arg;
2449 struct obd_device *obd = lut->lut_obd;
2450 struct ptlrpc_request *req;
2451 struct target_recovery_data *trd = &obd->obd_recovery_data;
2452 unsigned long delta;
2454 struct ptlrpc_thread *thread = NULL;
2458 unshare_fs_struct();
2459 OBD_ALLOC_PTR(thread);
2465 OBD_FREE_PTR(thread);
2469 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2471 OBD_FREE_PTR(thread);
2476 thread->t_env = env;
2477 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2478 env->le_ctx.lc_thread = thread;
2479 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2480 thread->t_watchdog = NULL;
2482 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2484 trd->trd_processing_task = current_pid();
2486 spin_lock(&obd->obd_dev_lock);
2487 obd->obd_recovering = 1;
2488 spin_unlock(&obd->obd_dev_lock);
2489 complete(&trd->trd_starting);
2491 /* first of all, we have to know the first transno to replay */
2492 if (target_recovery_overseer(lut, check_for_recovery_ready,
2493 exp_connect_healthy)) {
2494 abort_req_replay_queue(obd);
2495 abort_lock_replay_queue(obd);
2496 if (lut->lut_tdtd != NULL)
2497 dtrq_list_destroy(lut->lut_tdtd);
2500 /* next stage: replay requests or update */
2502 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2503 atomic_read(&obd->obd_req_replay_clients),
2504 obd->obd_next_recovery_transno);
2505 replay_request_or_update(env, lut, trd, thread);
2508 * The second stage: replay locks
2510 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2511 atomic_read(&obd->obd_lock_replay_clients));
2512 while ((req = target_next_replay_lock(lut))) {
2513 LASSERT(trd->trd_processing_task == current_pid());
2514 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2515 libcfs_nid2str(req->rq_peer.nid));
2516 handle_recovery_req(thread, req,
2517 trd->trd_recovery_handler);
2518 target_request_copy_put(req);
2519 obd->obd_replayed_locks++;
2523 * The third stage: reply on final pings, at this moment all clients
2524 * must have request in final queue
2526 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2527 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2528 /** Update server last boot epoch */
2529 tgt_boot_epoch_update(lut);
2530 /* We drop recoverying flag to forward all new requests
2531 * to regular mds_handle() since now */
2532 spin_lock(&obd->obd_dev_lock);
2533 obd->obd_recovering = obd->obd_abort_recovery = 0;
2534 spin_unlock(&obd->obd_dev_lock);
2535 spin_lock(&obd->obd_recovery_task_lock);
2536 target_cancel_recovery_timer(obd);
2537 spin_unlock(&obd->obd_recovery_task_lock);
2538 while ((req = target_next_final_ping(obd))) {
2539 LASSERT(trd->trd_processing_task == current_pid());
2540 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2541 libcfs_nid2str(req->rq_peer.nid));
2542 handle_recovery_req(thread, req,
2543 trd->trd_recovery_handler);
2544 /* Because the waiting client can not send ping to server,
2545 * so we need refresh the last_request_time, to avoid the
2546 * export is being evicted */
2547 ptlrpc_update_export_timer(req->rq_export, 0);
2548 target_request_copy_put(req);
2551 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2552 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2553 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2554 if (delta > OBD_RECOVERY_TIME_SOFT) {
2555 CWARN("too long recovery - read logs\n");
2556 libcfs_debug_dumplog();
2559 target_finish_recovery(lut);
2561 lu_context_fini(&env->le_ctx);
2562 trd->trd_processing_task = 0;
2563 complete(&trd->trd_finishing);
2565 tgt_io_thread_done(thread);
2566 OBD_FREE_PTR(thread);
2571 static int target_start_recovery_thread(struct lu_target *lut,
2572 svc_handler_t handler)
2574 struct obd_device *obd = lut->lut_obd;
2576 struct target_recovery_data *trd = &obd->obd_recovery_data;
2579 memset(trd, 0, sizeof(*trd));
2580 init_completion(&trd->trd_starting);
2581 init_completion(&trd->trd_finishing);
2582 trd->trd_recovery_handler = handler;
2584 rc = server_name2index(obd->obd_name, &index, NULL);
2588 if (!IS_ERR(kthread_run(target_recovery_thread,
2589 lut, "tgt_recover_%d", index))) {
2590 wait_for_completion(&trd->trd_starting);
2591 LASSERT(obd->obd_recovering != 0);
2599 void target_stop_recovery_thread(struct obd_device *obd)
2601 if (obd->obd_recovery_data.trd_processing_task > 0) {
2602 struct target_recovery_data *trd = &obd->obd_recovery_data;
2603 /** recovery can be done but postrecovery is not yet */
2604 spin_lock(&obd->obd_dev_lock);
2605 if (obd->obd_recovering) {
2606 CERROR("%s: Aborting recovery\n", obd->obd_name);
2607 obd->obd_abort_recovery = 1;
2608 wake_up(&obd->obd_next_transno_waitq);
2610 spin_unlock(&obd->obd_dev_lock);
2611 wait_for_completion(&trd->trd_finishing);
2614 EXPORT_SYMBOL(target_stop_recovery_thread);
2616 void target_recovery_fini(struct obd_device *obd)
2618 class_disconnect_exports(obd);
2619 target_stop_recovery_thread(obd);
2620 target_cleanup_recovery(obd);
2622 EXPORT_SYMBOL(target_recovery_fini);
2624 static void target_recovery_expired(cfs_timer_cb_arg_t data)
2626 struct obd_device *obd = cfs_from_timer(obd, data, obd_recovery_timer);
2627 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2628 " after %llus (%d clients connected)\n",
2629 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2630 (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2631 atomic_read(&obd->obd_connected_clients));
2633 obd->obd_recovery_expired = 1;
2634 wake_up(&obd->obd_next_transno_waitq);
2637 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2639 struct obd_device *obd = lut->lut_obd;
2641 if (lut->lut_bottom->dd_rdonly)
2644 if (obd->obd_max_recoverable_clients == 0) {
2645 /** Update server last boot epoch */
2646 tgt_boot_epoch_update(lut);
2650 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2651 "last_transno %llu\n", obd->obd_name,
2652 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2653 LASSERT(obd->obd_stopping == 0);
2654 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2655 obd->obd_recovery_start = 0;
2656 obd->obd_recovery_end = 0;
2658 cfs_timer_setup(&obd->obd_recovery_timer, target_recovery_expired,
2659 (unsigned long)obd, 0);
2660 target_start_recovery_thread(lut, handler);
2662 EXPORT_SYMBOL(target_recovery_init);
2664 static int target_process_req_flags(struct obd_device *obd,
2665 struct ptlrpc_request *req)
2667 struct obd_export *exp = req->rq_export;
2668 LASSERT(exp != NULL);
2669 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2670 /* client declares he's ready to replay locks */
2671 spin_lock(&exp->exp_lock);
2672 if (exp->exp_req_replay_needed) {
2673 exp->exp_req_replay_needed = 0;
2674 spin_unlock(&exp->exp_lock);
2676 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2677 atomic_dec(&obd->obd_req_replay_clients);
2679 spin_unlock(&exp->exp_lock);
2682 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2683 /* client declares he's ready to complete recovery
2684 * so, we put the request on th final queue */
2685 spin_lock(&exp->exp_lock);
2686 if (exp->exp_lock_replay_needed) {
2687 exp->exp_lock_replay_needed = 0;
2688 spin_unlock(&exp->exp_lock);
2690 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2691 atomic_dec(&obd->obd_lock_replay_clients);
2693 spin_unlock(&exp->exp_lock);
2699 int target_queue_recovery_request(struct ptlrpc_request *req,
2700 struct obd_device *obd)
2702 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2703 struct ptlrpc_request *reqiter;
2707 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2708 /* Processing the queue right now, don't re-add. */
2712 target_process_req_flags(obd, req);
2714 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2715 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2716 if (cfs_fail_val == 1) {
2719 wake_up(&cfs_race_waitq);
2721 set_current_state(TASK_INTERRUPTIBLE);
2722 schedule_timeout(cfs_time_seconds(1));
2726 /* client declares he's ready to complete recovery
2727 * so, we put the request on th final queue */
2728 target_request_copy_get(req);
2729 DEBUG_REQ(D_HA, req, "queue final req");
2730 wake_up(&obd->obd_next_transno_waitq);
2731 spin_lock(&obd->obd_recovery_task_lock);
2732 if (obd->obd_recovering) {
2733 struct ptlrpc_request *tmp;
2734 struct ptlrpc_request *duplicate = NULL;
2736 if (likely(!req->rq_export->exp_replay_done)) {
2737 req->rq_export->exp_replay_done = 1;
2738 list_add_tail(&req->rq_list,
2739 &obd->obd_final_req_queue);
2740 spin_unlock(&obd->obd_recovery_task_lock);
2744 /* XXX O(n), but only happens if final ping is
2745 * timed out, probably reorganize the list as
2746 * a hash list later */
2747 list_for_each_entry_safe(reqiter, tmp,
2748 &obd->obd_final_req_queue,
2750 if (reqiter->rq_export == req->rq_export) {
2751 list_del_init(&reqiter->rq_list);
2752 duplicate = reqiter;
2757 list_add_tail(&req->rq_list,
2758 &obd->obd_final_req_queue);
2759 req->rq_export->exp_replay_done = 1;
2760 spin_unlock(&obd->obd_recovery_task_lock);
2762 if (duplicate != NULL) {
2763 DEBUG_REQ(D_HA, duplicate,
2764 "put prev final req\n");
2765 target_request_copy_put(duplicate);
2769 spin_unlock(&obd->obd_recovery_task_lock);
2770 target_request_copy_put(req);
2771 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2774 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2775 /* client declares he's ready to replay locks */
2776 target_request_copy_get(req);
2777 DEBUG_REQ(D_HA, req, "queue lock replay req");
2778 wake_up(&obd->obd_next_transno_waitq);
2779 spin_lock(&obd->obd_recovery_task_lock);
2780 LASSERT(obd->obd_recovering);
2781 /* usually due to recovery abort */
2782 if (!req->rq_export->exp_in_recovery) {
2783 spin_unlock(&obd->obd_recovery_task_lock);
2784 target_request_copy_put(req);
2787 LASSERT(req->rq_export->exp_lock_replay_needed);
2788 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2789 spin_unlock(&obd->obd_recovery_task_lock);
2793 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2794 * (i.e. buflens etc are in my own byte order), but type-dependent
2795 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2798 INIT_LIST_HEAD(&req->rq_list);
2799 DEBUG_REQ(D_HA, req, "not queueing");
2803 /* If we're processing the queue, we want don't want to queue this
2806 * Also, if this request has a transno less than the one we're waiting
2807 * for, we should process it now. It could (and currently always will)
2808 * be an open request for a descriptor that was opened some time ago.
2810 * Also, a resent, replayed request that has already been
2811 * handled will pass through here and be processed immediately.
2813 CDEBUG(D_HA, "Next recovery transno: %llu"
2814 ", current: %llu, replaying\n",
2815 obd->obd_next_recovery_transno, transno);
2817 /* If the request has been replayed by update replay, then sends this
2818 * request to the recovery thread (replay_request_or_update()), where
2819 * it will be handled */
2820 spin_lock(&obd->obd_recovery_task_lock);
2821 if (transno < obd->obd_next_recovery_transno &&
2822 !is_req_replayed_by_update(req)) {
2823 /* Processing the queue right now, don't re-add. */
2824 LASSERT(list_empty(&req->rq_list));
2825 spin_unlock(&obd->obd_recovery_task_lock);
2828 spin_unlock(&obd->obd_recovery_task_lock);
2830 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2833 target_request_copy_get(req);
2834 if (!req->rq_export->exp_in_recovery) {
2835 target_request_copy_put(req);
2838 LASSERT(req->rq_export->exp_req_replay_needed);
2840 if (target_exp_enqueue_req_replay(req)) {
2841 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2842 target_request_copy_put(req);
2847 spin_lock(&obd->obd_recovery_task_lock);
2848 LASSERT(obd->obd_recovering);
2849 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2850 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2851 list_add_tail(&req->rq_list, &reqiter->rq_list);
2856 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2858 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2859 "has been claimed by another client");
2860 spin_unlock(&obd->obd_recovery_task_lock);
2861 target_exp_dequeue_req_replay(req);
2862 target_request_copy_put(req);
2868 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2870 obd->obd_requests_queued_for_recovery++;
2871 spin_unlock(&obd->obd_recovery_task_lock);
2872 wake_up(&obd->obd_next_transno_waitq);
2876 void target_committed_to_req(struct ptlrpc_request *req)
2878 struct obd_export *exp = req->rq_export;
2880 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2881 lustre_msg_set_last_committed(req->rq_repmsg,
2882 exp->exp_last_committed);
2884 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2885 "%d)", exp->exp_obd->obd_no_transno,
2886 req->rq_repmsg == NULL);
2888 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2889 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2892 #endif /* HAVE_SERVER_SUPPORT */
2895 * Packs current SLV and Limit into \a req.
2897 int target_pack_pool_reply(struct ptlrpc_request *req)
2899 struct obd_device *obd;
2902 /* Check that we still have all structures alive as this may
2903 * be some late RPC at shutdown time. */
2904 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2905 !exp_connect_lru_resize(req->rq_export))) {
2906 lustre_msg_set_slv(req->rq_repmsg, 0);
2907 lustre_msg_set_limit(req->rq_repmsg, 0);
2911 /* OBD is alive here as export is alive, which we checked above. */
2912 obd = req->rq_export->exp_obd;
2914 read_lock(&obd->obd_pool_lock);
2915 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2916 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2917 read_unlock(&obd->obd_pool_lock);
2922 static int target_send_reply_msg(struct ptlrpc_request *req,
2923 int rc, int fail_id)
2925 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2926 DEBUG_REQ(D_ERROR, req, "dropping reply");
2929 /* We can have a null rq_reqmsg in the event of bad signature or
2930 * no context when unwrapping */
2931 if (req->rq_reqmsg &&
2932 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2933 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2937 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2938 req->rq_status = rc;
2939 return ptlrpc_send_error(req, 1);
2941 DEBUG_REQ(D_NET, req, "sending reply");
2944 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2947 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2949 struct ptlrpc_service_part *svcpt;
2951 struct ptlrpc_reply_state *rs;
2952 struct obd_export *exp;
2955 if (req->rq_no_reply) {
2960 svcpt = req->rq_rqbd->rqbd_svcpt;
2961 rs = req->rq_reply_state;
2962 if (rs == NULL || !rs->rs_difficult) {
2964 target_send_reply_msg (req, rc, fail_id);
2969 /* must be an export if locks saved */
2970 LASSERT(req->rq_export != NULL);
2971 /* req/reply consistent */
2972 LASSERT(rs->rs_svcpt == svcpt);
2975 LASSERT(!rs->rs_scheduled);
2976 LASSERT(!rs->rs_scheduled_ever);
2977 LASSERT(!rs->rs_handled);
2978 LASSERT(!rs->rs_on_net);
2979 LASSERT(rs->rs_export == NULL);
2980 LASSERT(list_empty(&rs->rs_obd_list));
2981 LASSERT(list_empty(&rs->rs_exp_list));
2983 exp = class_export_get(req->rq_export);
2985 /* disable reply scheduling while I'm setting up */
2986 rs->rs_scheduled = 1;
2988 rs->rs_xid = req->rq_xid;
2989 rs->rs_transno = req->rq_transno;
2990 rs->rs_export = exp;
2991 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2993 spin_lock(&exp->exp_uncommitted_replies_lock);
2994 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2995 rs->rs_transno, exp->exp_last_committed);
2996 if (rs->rs_transno > exp->exp_last_committed) {
2997 /* not committed already */
2998 list_add_tail(&rs->rs_obd_list,
2999 &exp->exp_uncommitted_replies);
3001 spin_unlock(&exp->exp_uncommitted_replies_lock);
3003 spin_lock(&exp->exp_lock);
3004 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3005 spin_unlock(&exp->exp_lock);
3007 netrc = target_send_reply_msg(req, rc, fail_id);
3009 spin_lock(&svcpt->scp_rep_lock);
3011 atomic_inc(&svcpt->scp_nreps_difficult);
3014 /* error sending: reply is off the net. Also we need +1
3015 * reply ref until ptlrpc_handle_rs() is done
3016 * with the reply state (if the send was successful, there
3017 * would have been +1 ref for the net, which
3018 * reply_out_callback leaves alone) */
3020 ptlrpc_rs_addref(rs);
3023 spin_lock(&rs->rs_lock);
3024 if (rs->rs_transno <= exp->exp_last_committed ||
3025 (!rs->rs_on_net && !rs->rs_no_ack) ||
3026 list_empty(&rs->rs_exp_list) || /* completed already */
3027 list_empty(&rs->rs_obd_list)) {
3028 CDEBUG(D_HA, "Schedule reply immediately\n");
3029 ptlrpc_dispatch_difficult_reply(rs);
3031 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3032 rs->rs_scheduled = 0; /* allow notifier to schedule */
3034 spin_unlock(&rs->rs_lock);
3035 spin_unlock(&svcpt->scp_rep_lock);
3039 enum ldlm_mode lck_compat_array[] = {
3040 [LCK_EX] = LCK_COMPAT_EX,
3041 [LCK_PW] = LCK_COMPAT_PW,
3042 [LCK_PR] = LCK_COMPAT_PR,
3043 [LCK_CW] = LCK_COMPAT_CW,
3044 [LCK_CR] = LCK_COMPAT_CR,
3045 [LCK_NL] = LCK_COMPAT_NL,
3046 [LCK_GROUP] = LCK_COMPAT_GROUP,
3047 [LCK_COS] = LCK_COMPAT_COS,
3051 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3052 * not escape to the user level.
3054 int ldlm_error2errno(enum ldlm_error error)
3060 case ELDLM_LOCK_MATCHED:
3063 case ELDLM_LOCK_CHANGED:
3066 case ELDLM_LOCK_ABORTED:
3069 case ELDLM_LOCK_REPLACED:
3072 case ELDLM_NO_LOCK_DATA:
3075 case ELDLM_NAMESPACE_EXISTS:
3078 case ELDLM_BAD_NAMESPACE:
3082 if (((int)error) < 0) { /* cast to signed type */
3083 result = error; /* as ldlm_error can be unsigned */
3085 CERROR("Invalid DLM result code: %d\n", error);
3091 EXPORT_SYMBOL(ldlm_error2errno);
3094 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3096 enum ldlm_error ldlm_errno2error(int err_no)
3105 error = ELDLM_LOCK_CHANGED;
3108 error = ELDLM_LOCK_ABORTED;
3111 error = ELDLM_LOCK_REPLACED;
3114 error = ELDLM_NO_LOCK_DATA;
3117 error = ELDLM_NAMESPACE_EXISTS;
3120 error = ELDLM_BAD_NAMESPACE;
3128 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3129 void ldlm_dump_export_locks(struct obd_export *exp)
3131 spin_lock(&exp->exp_locks_list_guard);
3132 if (!list_empty(&exp->exp_locks_list)) {
3133 struct ldlm_lock *lock;
3135 CERROR("dumping locks for export %p,"
3136 "ignore if the unmount doesn't hang\n", exp);
3137 list_for_each_entry(lock, &exp->exp_locks_list,
3139 LDLM_ERROR(lock, "lock:");
3141 spin_unlock(&exp->exp_locks_list_guard);
3145 #ifdef HAVE_SERVER_SUPPORT
3146 static int target_bulk_timeout(void *data)
3149 /* We don't fail the connection here, because having the export
3150 * killed makes the (vital) call to commitrw very sad.
3155 static inline const char *bulk2type(struct ptlrpc_request *req)
3157 if (req->rq_bulk_read)
3159 if (req->rq_bulk_write)
3164 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3165 struct l_wait_info *lwi)
3167 struct ptlrpc_request *req = desc->bd_req;
3168 time64_t start = ktime_get_real_seconds();
3174 /* If there is eviction in progress, wait for it to finish. */
3175 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3176 *lwi = LWI_INTR(NULL, NULL);
3177 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3178 !atomic_read(&exp->exp_obd->
3179 obd_evict_inprogress),
3183 /* Check if client was evicted or reconnected already. */
3184 if (exp->exp_failed ||
3185 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3188 if (req->rq_bulk_read)
3189 rc = sptlrpc_svc_wrap_bulk(req, desc);
3191 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3192 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3193 else /* old version, bulk matchbits is rq_xid */
3194 req->rq_mbits = req->rq_xid;
3197 rc = ptlrpc_start_bulk_transfer(desc);
3201 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3202 bulk2type(req), rc);
3206 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3207 ptlrpc_abort_bulk(desc);
3211 /* limit actual bulk transfer to bulk_timeout seconds */
3212 deadline = start + bulk_timeout;
3213 if (deadline > req->rq_deadline)
3214 deadline = req->rq_deadline;
3217 time64_t timeoutl = deadline - ktime_get_real_seconds();
3218 long timeout_jiffies = timeoutl <= 0 ?
3219 1 : cfs_time_seconds(timeoutl);
3220 time64_t rq_deadline;
3222 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3223 cfs_time_seconds(1),
3224 target_bulk_timeout, desc);
3225 rc = l_wait_event(desc->bd_waitq,
3226 !ptlrpc_server_bulk_active(desc) ||
3229 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3231 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3232 /* Wait again if we changed rq_deadline. */
3233 rq_deadline = READ_ONCE(req->rq_deadline);
3234 deadline = start + bulk_timeout;
3235 if (deadline > rq_deadline)
3236 deadline = rq_deadline;
3237 } while (rc == -ETIMEDOUT &&
3238 deadline > ktime_get_real_seconds());
3240 if (rc == -ETIMEDOUT) {
3241 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3242 bulk2type(req), deadline - start,
3243 ktime_get_real_seconds() - deadline);
3244 ptlrpc_abort_bulk(desc);
3245 } else if (exp->exp_failed) {
3246 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3249 ptlrpc_abort_bulk(desc);
3250 } else if (exp->exp_conn_cnt >
3251 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3252 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3254 /* We don't reply anyway. */
3256 ptlrpc_abort_bulk(desc);
3257 } else if (desc->bd_failure) {
3258 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3260 /* XXX should this be a different errno? */
3263 if (req->rq_bulk_write)
3264 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3265 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3266 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3267 bulk2type(req), desc->bd_nob_transferred,
3269 /* XXX should this be a different errno? */
3276 EXPORT_SYMBOL(target_bulk_io);
3278 #endif /* HAVE_SERVER_SUPPORT */