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 <cl_object.h>
43 #include <linux/jiffies.h>
44 #include <linux/kthread.h>
45 #include <libcfs/libcfs.h>
47 #include <obd_class.h>
48 #include <lustre_dlm.h>
49 #include <lustre_net.h>
50 #include <lustre_sec.h>
51 #include "ldlm_internal.h"
54 * @priority: If non-zero, move the selected connection to the list head.
55 * @create: If zero, only search in existing connections.
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
62 lnet_nid_t nid4refnet = LNET_NID_ANY;
67 if (!create && !priority) {
68 CDEBUG(D_HA, "Nothing to do\n");
72 if (imp->imp_connection &&
73 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
74 /* nid4refnet is used to restrict network connections */
75 nid4refnet = imp->imp_connection->c_self;
76 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
78 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
83 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
85 GOTO(out_put, rc = -ENOMEM);
88 spin_lock(&imp->imp_lock);
89 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
90 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
92 list_del(&item->oic_item);
93 list_add(&item->oic_item,
95 item->oic_last_attempt = 0;
97 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
98 imp, imp->imp_obd->obd_name, uuid->uuid,
99 (priority ? ", moved to head" : ""));
100 spin_unlock(&imp->imp_lock);
101 GOTO(out_free, rc = 0);
104 /* No existing import connection found for \a uuid. */
106 imp_conn->oic_conn = ptlrpc_conn;
107 imp_conn->oic_uuid = *uuid;
108 imp_conn->oic_last_attempt = 0;
110 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
112 list_add_tail(&imp_conn->oic_item,
113 &imp->imp_conn_list);
114 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
115 imp, imp->imp_obd->obd_name, uuid->uuid,
116 (priority ? "head" : "tail"));
118 spin_unlock(&imp->imp_lock);
119 GOTO(out_free, rc = -ENOENT);
122 spin_unlock(&imp->imp_lock);
126 OBD_FREE(imp_conn, sizeof(*imp_conn));
128 ptlrpc_connection_put(ptlrpc_conn);
132 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
134 return import_set_conn(imp, uuid, 1, 0);
137 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
140 return import_set_conn(imp, uuid, priority, 1);
142 EXPORT_SYMBOL(client_import_add_conn);
144 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
146 struct obd_import_conn *imp_conn;
147 struct obd_export *dlmexp;
152 spin_lock(&imp->imp_lock);
153 if (list_empty(&imp->imp_conn_list)) {
154 LASSERT(!imp->imp_connection);
158 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
159 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
161 LASSERT(imp_conn->oic_conn);
163 if (imp_conn == imp->imp_conn_current) {
164 LASSERT(imp_conn->oic_conn == imp->imp_connection);
166 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
167 imp->imp_state != LUSTRE_IMP_DISCON) {
168 CERROR("can't remove current connection\n");
169 GOTO(out, rc = -EBUSY);
172 ptlrpc_connection_put(imp->imp_connection);
173 imp->imp_connection = NULL;
175 dlmexp = class_conn2export(&imp->imp_dlm_handle);
176 if (dlmexp && dlmexp->exp_connection) {
177 LASSERT(dlmexp->exp_connection ==
179 ptlrpc_connection_put(dlmexp->exp_connection);
180 dlmexp->exp_connection = NULL;
184 class_export_put(dlmexp);
187 list_del(&imp_conn->oic_item);
188 ptlrpc_connection_put(imp_conn->oic_conn);
189 OBD_FREE(imp_conn, sizeof(*imp_conn));
190 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
191 imp, imp->imp_obd->obd_name, uuid->uuid);
196 spin_unlock(&imp->imp_lock);
198 CERROR("connection %s not found\n", uuid->uuid);
201 EXPORT_SYMBOL(client_import_del_conn);
204 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
205 * to find a conn uuid of \a imp which can reach \a peer.
207 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
208 struct obd_uuid *uuid)
210 struct obd_import_conn *conn;
215 spin_lock(&imp->imp_lock);
216 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
217 /* Check if conn UUID does have this peer NID. */
218 if (class_check_uuid(&conn->oic_uuid, peer)) {
219 *uuid = conn->oic_uuid;
224 spin_unlock(&imp->imp_lock);
227 EXPORT_SYMBOL(client_import_find_conn);
229 void client_destroy_import(struct obd_import *imp)
232 * Drop security policy instance after all RPCs have finished/aborted
233 * to let all busy contexts be released.
235 class_import_get(imp);
236 class_destroy_import(imp);
237 sptlrpc_import_sec_put(imp);
238 class_import_put(imp);
240 EXPORT_SYMBOL(client_destroy_import);
243 * Check whether or not the OSC is on MDT.
246 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
248 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
251 static int osc_on_mdt(char *obdname)
255 ptr = strrchr(obdname, '-');
259 if (strncmp(ptr + 1, "MDT", 3) == 0)
266 * Configure an RPC client OBD device.
271 * 3 - inactive-on-startup
272 * 4 - restrictive net
274 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
276 struct client_obd *cli = &obddev->u.cli;
277 struct obd_import *imp;
278 struct obd_uuid server_uuid;
279 int rq_portal, rp_portal, connect_op;
280 char *name = obddev->obd_type->typ_name;
281 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
282 char *cli_name = lustre_cfg_buf(lcfg, 0);
283 struct ptlrpc_connection fake_conn = { .c_self = 0,
284 .c_remote_uuid.uuid[0] = 0 };
290 * In a more perfect world, we would hang a ptlrpc_client off of
291 * obd_type and just use the values from there.
293 if (!strcmp(name, LUSTRE_OSC_NAME)) {
294 rq_portal = OST_REQUEST_PORTAL;
295 rp_portal = OSC_REPLY_PORTAL;
296 connect_op = OST_CONNECT;
297 cli->cl_sp_me = LUSTRE_SP_CLI;
298 cli->cl_sp_to = LUSTRE_SP_OST;
299 ns_type = LDLM_NS_TYPE_OSC;
300 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
301 !strcmp(name, LUSTRE_LWP_NAME)) {
302 rq_portal = MDS_REQUEST_PORTAL;
303 rp_portal = MDC_REPLY_PORTAL;
304 connect_op = MDS_CONNECT;
305 if (is_lwp_on_ost(cli_name))
306 cli->cl_sp_me = LUSTRE_SP_OST;
307 else if (is_lwp_on_mdt(cli_name))
308 cli->cl_sp_me = LUSTRE_SP_MDT;
310 cli->cl_sp_me = LUSTRE_SP_CLI;
311 cli->cl_sp_to = LUSTRE_SP_MDT;
312 ns_type = LDLM_NS_TYPE_MDC;
313 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
314 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
315 /* OSP_on_MDT for other MDTs */
316 connect_op = MDS_CONNECT;
317 cli->cl_sp_to = LUSTRE_SP_MDT;
318 ns_type = LDLM_NS_TYPE_MDC;
319 rq_portal = OUT_PORTAL;
321 /* OSP on MDT for OST */
322 connect_op = OST_CONNECT;
323 cli->cl_sp_to = LUSTRE_SP_OST;
324 ns_type = LDLM_NS_TYPE_OSC;
325 rq_portal = OST_REQUEST_PORTAL;
327 rp_portal = OSC_REPLY_PORTAL;
328 cli->cl_sp_me = LUSTRE_SP_MDT;
329 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
330 rq_portal = MGS_REQUEST_PORTAL;
331 rp_portal = MGC_REPLY_PORTAL;
332 connect_op = MGS_CONNECT;
333 cli->cl_sp_me = LUSTRE_SP_MGC;
334 cli->cl_sp_to = LUSTRE_SP_MGS;
335 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
336 ns_type = LDLM_NS_TYPE_MGC;
338 CERROR("unknown client OBD type \"%s\", can't setup\n",
343 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
344 CERROR("requires a TARGET UUID\n");
348 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
349 CERROR("client UUID must be less than 38 characters\n");
353 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
354 CERROR("setup requires a SERVER UUID\n");
358 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
359 CERROR("target UUID must be less than 38 characters\n");
363 init_rwsem(&cli->cl_sem);
364 mutex_init(&cli->cl_mgc_mutex);
366 init_rwsem(&cli->cl_seq_rwsem);
367 cli->cl_conn_count = 0;
368 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
369 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
370 sizeof(server_uuid)));
372 cli->cl_dirty_pages = 0;
373 cli->cl_dirty_max_pages = 0;
374 cli->cl_avail_grant = 0;
375 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
377 * cl_dirty_max_pages may be changed at connect time in
378 * ptlrpc_connect_interpret().
380 client_adjust_max_dirty(cli);
381 INIT_LIST_HEAD(&cli->cl_cache_waiters);
382 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
383 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
384 INIT_LIST_HEAD(&cli->cl_loi_write_list);
385 INIT_LIST_HEAD(&cli->cl_loi_read_list);
386 spin_lock_init(&cli->cl_loi_list_lock);
387 atomic_set(&cli->cl_pending_w_pages, 0);
388 atomic_set(&cli->cl_pending_r_pages, 0);
389 cli->cl_r_in_flight = 0;
390 cli->cl_w_in_flight = 0;
392 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
393 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
394 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
395 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
396 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
397 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
400 INIT_LIST_HEAD(&cli->cl_lru_osc);
401 atomic_set(&cli->cl_lru_shrinkers, 0);
402 atomic_long_set(&cli->cl_lru_busy, 0);
403 atomic_long_set(&cli->cl_lru_in_list, 0);
404 INIT_LIST_HEAD(&cli->cl_lru_list);
405 spin_lock_init(&cli->cl_lru_list_lock);
406 atomic_long_set(&cli->cl_unstable_count, 0);
407 INIT_LIST_HEAD(&cli->cl_shrink_list);
408 INIT_LIST_HEAD(&cli->cl_grant_chain);
410 INIT_LIST_HEAD(&cli->cl_flight_waiters);
411 cli->cl_rpcs_in_flight = 0;
413 init_waitqueue_head(&cli->cl_destroy_waitq);
414 atomic_set(&cli->cl_destroy_in_flight, 0);
416 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
417 #ifdef ENABLE_CHECKSUM
418 /* Turn on checksumming by default. */
419 cli->cl_checksum = 1;
421 * The supported checksum types will be worked out at connect time
422 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
425 cli->cl_cksum_type = cli->cl_supp_cksum_types;
427 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
430 * Set it to possible maximum size. It may be reduced by ocd_brw_size
431 * from OFD after connecting.
433 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
435 cli->cl_max_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
438 * set cl_chunkbits default value to PAGE_SHIFT,
439 * it will be updated at OSC connection time.
441 cli->cl_chunkbits = PAGE_SHIFT;
443 if (!strcmp(name, LUSTRE_MDC_NAME)) {
444 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
445 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
446 cli->cl_max_rpcs_in_flight = 2;
447 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
448 cli->cl_max_rpcs_in_flight = 3;
449 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
450 cli->cl_max_rpcs_in_flight = 4;
452 if (osc_on_mdt(obddev->obd_name))
453 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
455 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
458 spin_lock_init(&cli->cl_mod_rpcs_lock);
459 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
460 cli->cl_max_mod_rpcs_in_flight = 0;
461 cli->cl_mod_rpcs_in_flight = 0;
462 cli->cl_close_rpcs_in_flight = 0;
463 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
464 cli->cl_mod_tag_bitmap = NULL;
466 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
468 if (connect_op == MDS_CONNECT) {
469 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
470 OBD_ALLOC(cli->cl_mod_tag_bitmap,
471 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
472 if (cli->cl_mod_tag_bitmap == NULL)
473 GOTO(err, rc = -ENOMEM);
478 CERROR("ldlm_get_ref failed: %d\n", rc);
482 ptlrpc_init_client(rq_portal, rp_portal, name,
483 &obddev->obd_ldlm_client);
485 imp = class_new_import(obddev);
487 GOTO(err_ldlm, rc = -ENOENT);
488 imp->imp_client = &obddev->obd_ldlm_client;
489 imp->imp_connect_op = connect_op;
490 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
491 LUSTRE_CFG_BUFLEN(lcfg, 1));
492 class_import_put(imp);
494 if (lustre_cfg_buf(lcfg, 4)) {
495 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
497 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
499 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
500 obddev->obd_name, lustre_cfg_string(lcfg, 4),
502 GOTO(err_import, rc);
504 fake_conn.c_self = LNET_MKNID(refnet, 0);
505 imp->imp_connection = &fake_conn;
508 rc = client_import_add_conn(imp, &server_uuid, 1);
510 CERROR("can't add initial connection\n");
511 GOTO(err_import, rc);
513 imp->imp_connection = NULL;
515 cli->cl_import = imp;
516 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
517 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
519 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
520 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
521 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
522 name, obddev->obd_name,
523 cli->cl_target_uuid.uuid);
524 spin_lock(&imp->imp_lock);
525 imp->imp_deactive = 1;
526 spin_unlock(&imp->imp_lock);
530 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
531 LDLM_NAMESPACE_CLIENT,
532 LDLM_NAMESPACE_GREEDY,
534 if (obddev->obd_namespace == NULL) {
535 CERROR("Unable to create client namespace - %s\n",
537 GOTO(err_import, rc = -ENOMEM);
543 class_destroy_import(imp);
547 if (cli->cl_mod_tag_bitmap != NULL)
548 OBD_FREE(cli->cl_mod_tag_bitmap,
549 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
550 cli->cl_mod_tag_bitmap = NULL;
554 EXPORT_SYMBOL(client_obd_setup);
556 int client_obd_cleanup(struct obd_device *obddev)
558 struct client_obd *cli = &obddev->u.cli;
562 ldlm_namespace_free_post(obddev->obd_namespace);
563 obddev->obd_namespace = NULL;
565 obd_cleanup_client_import(obddev);
566 LASSERT(obddev->u.cli.cl_import == NULL);
570 if (cli->cl_mod_tag_bitmap != NULL)
571 OBD_FREE(cli->cl_mod_tag_bitmap,
572 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
573 cli->cl_mod_tag_bitmap = NULL;
577 EXPORT_SYMBOL(client_obd_cleanup);
579 /* ->o_connect() method for client side (OSC and MDC and MGC) */
580 int client_connect_import(const struct lu_env *env,
581 struct obd_export **exp,
582 struct obd_device *obd, struct obd_uuid *cluuid,
583 struct obd_connect_data *data, void *localdata)
585 struct client_obd *cli = &obd->u.cli;
586 struct obd_import *imp = cli->cl_import;
587 struct obd_connect_data *ocd;
588 struct lustre_handle conn = { 0 };
594 down_write(&cli->cl_sem);
595 if (cli->cl_conn_count > 0)
596 GOTO(out_sem, rc = -EALREADY);
598 rc = class_connect(&conn, obd, cluuid);
602 cli->cl_conn_count++;
603 *exp = class_conn2export(&conn);
605 LASSERT(obd->obd_namespace);
607 imp->imp_dlm_handle = conn;
608 rc = ptlrpc_init_import(imp);
612 ocd = &imp->imp_connect_data;
615 imp->imp_connect_flags_orig = data->ocd_connect_flags;
616 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
619 rc = ptlrpc_connect_import(imp);
621 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
624 LASSERT(*exp != NULL && (*exp)->exp_connection);
627 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
628 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
629 data->ocd_connect_flags, ocd->ocd_connect_flags);
630 data->ocd_connect_flags = ocd->ocd_connect_flags;
631 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
634 ptlrpc_pinger_add_import(imp);
640 cli->cl_conn_count--;
641 class_disconnect(*exp);
645 up_write(&cli->cl_sem);
647 if (!rc && localdata) {
648 LASSERT(cli->cl_cache == NULL); /* only once */
649 cli->cl_cache = (struct cl_client_cache *)localdata;
650 cl_cache_incref(cli->cl_cache);
651 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
653 /* add this osc into entity list */
654 LASSERT(list_empty(&cli->cl_lru_osc));
655 spin_lock(&cli->cl_cache->ccc_lru_lock);
656 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
657 spin_unlock(&cli->cl_cache->ccc_lru_lock);
662 EXPORT_SYMBOL(client_connect_import);
664 int client_disconnect_export(struct obd_export *exp)
666 struct obd_device *obd = class_exp2obd(exp);
667 struct client_obd *cli;
668 struct obd_import *imp;
674 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
675 exp, exp ? exp->exp_handle.h_cookie : -1);
680 imp = cli->cl_import;
682 down_write(&cli->cl_sem);
683 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
686 if (cli->cl_conn_count == 0) {
687 CERROR("disconnecting disconnected device (%s)\n",
689 GOTO(out_disconnect, rc = -EINVAL);
692 cli->cl_conn_count--;
693 if (cli->cl_conn_count != 0)
694 GOTO(out_disconnect, rc = 0);
697 * Mark import deactivated now, so we don't try to reconnect if any
698 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
699 * fully deactivate the import, or that would drop all requests.
701 spin_lock(&imp->imp_lock);
702 imp->imp_deactive = 1;
703 spin_unlock(&imp->imp_lock);
706 * Some non-replayable imports (MDS's OSCs) are pinged, so just
707 * delete it regardless. (It's safe to delete an import that was
710 (void)ptlrpc_pinger_del_import(imp);
712 if (obd->obd_namespace != NULL) {
713 /* obd_force == local only */
714 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
715 obd->obd_force ? LCF_LOCAL : 0, NULL);
716 ldlm_namespace_free_prior(obd->obd_namespace, imp,
721 * There's no need to hold sem while disconnecting an import,
722 * and it may actually cause deadlock in GSS.
724 up_write(&cli->cl_sem);
725 rc = ptlrpc_disconnect_import(imp, 0);
726 down_write(&cli->cl_sem);
728 ptlrpc_invalidate_import(imp);
734 * Use server style - class_disconnect should be always called for
737 err = class_disconnect(exp);
741 up_write(&cli->cl_sem);
745 EXPORT_SYMBOL(client_disconnect_export);
747 #ifdef HAVE_SERVER_SUPPORT
748 int server_disconnect_export(struct obd_export *exp)
754 /* Disconnect early so that clients can't keep using export. */
755 rc = class_disconnect(exp);
756 /* Close import to avoid sending any requests. */
757 if (exp->exp_imp_reverse)
758 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
760 ldlm_bl_thread_wakeup();
762 /* complete all outstanding replies */
763 spin_lock(&exp->exp_lock);
764 while (!list_empty(&exp->exp_outstanding_replies)) {
765 struct ptlrpc_reply_state *rs =
766 list_entry(exp->exp_outstanding_replies.next,
767 struct ptlrpc_reply_state, rs_exp_list);
768 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
770 spin_lock(&svcpt->scp_rep_lock);
772 list_del_init(&rs->rs_exp_list);
774 spin_lock(&rs->rs_lock);
775 /* clear rs_convert_lock to make sure rs is handled and put */
776 rs->rs_convert_lock = 0;
777 ptlrpc_schedule_difficult_reply(rs);
778 spin_unlock(&rs->rs_lock);
780 spin_unlock(&svcpt->scp_rep_lock);
782 spin_unlock(&exp->exp_lock);
786 EXPORT_SYMBOL(server_disconnect_export);
789 * --------------------------------------------------------------------------
790 * from old lib/target.c
791 * --------------------------------------------------------------------------
794 static int target_handle_reconnect(struct lustre_handle *conn,
795 struct obd_export *exp,
796 struct obd_uuid *cluuid)
798 struct obd_device *target;
799 struct lustre_handle *hdl;
805 hdl = &exp->exp_imp_reverse->imp_remote_handle;
806 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
807 conn->cookie = exp->exp_handle.h_cookie;
809 "connect export for UUID '%s' at %p, cookie %#llx\n",
810 cluuid->uuid, exp, conn->cookie);
814 target = exp->exp_obd;
816 /* Might be a re-connect after a partition. */
817 if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
818 LCONSOLE_WARN("%s: already connected client %s (at %s) with handle %#llx. Rejecting client with the same UUID trying to reconnect with handle %#llx\n",
820 obd_uuid2str(&exp->exp_client_uuid),
821 obd_export_nid2str(exp),
822 hdl->cookie, conn->cookie);
823 memset(conn, 0, sizeof(*conn));
825 * target_handle_connect() treats EALREADY and
826 * -EALREADY differently. -EALREADY is an error
827 * (same UUID, different handle).
832 if (!target->obd_recovering) {
833 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
834 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
835 obd_export_nid2str(exp));
836 GOTO(out_already, rc);
839 remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
840 timeout = ktime_divns(remaining, NSEC_PER_SEC);
842 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
844 obd_uuid2str(&exp->exp_client_uuid),
845 obd_export_nid2str(exp),
846 target->obd_max_recoverable_clients,
847 timeout / 60, timeout % 60);
849 struct target_distribute_txn_data *tdtd;
854 tdtd = class_exp2tgt(exp)->lut_tdtd;
855 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
856 buf = tdtd->tdtd_show_update_logs_retrievers(
857 tdtd->tdtd_show_retrievers_cbdata,
861 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting, waiting for %d MDTs (%s) in recovery for %lld:%.02lld. Please wait until all MDTs recovered or you may force MDT evicition via 'lctl --device %s abort_recovery.\n",
863 obd_uuid2str(&exp->exp_client_uuid),
864 obd_export_nid2str(exp), count,
865 buf ? buf : "unknown (not enough RAM)",
866 (abs(timeout) + target->obd_recovery_timeout) / 60,
867 (abs(timeout) + target->obd_recovery_timeout) % 60,
870 LCONSOLE_WARN("%s: Recovery already passed deadline %lld:%.02lld. If you do not want to wait more, you may force taget eviction via 'lctl --device %s abort_recovery.\n",
871 target->obd_name, abs(timeout) / 60,
872 abs(timeout) % 60, target->obd_name);
879 conn->cookie = exp->exp_handle.h_cookie;
881 * target_handle_connect() treats EALREADY and
882 * -EALREADY differently. EALREADY means we are
883 * doing a valid reconnect from the same client.
889 check_and_start_recovery_timer(struct obd_device *obd,
890 struct ptlrpc_request *req, int new_client);
893 * update flags for import during reconnect process
895 static int rev_import_flags_update(struct obd_import *revimp,
896 struct ptlrpc_request *req)
899 struct obd_connect_data *data;
901 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
903 if (data->ocd_connect_flags & OBD_CONNECT_AT)
904 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
906 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
908 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
910 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
912 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
913 revimp->imp_client->cli_name,
914 libcfs_id2str(req->rq_peer), rc);
922 * Allocate a new reverse import for an export.
924 * \retval -errno in case error hit
925 * \retval 0 if reverse import correctly init
927 int rev_import_init(struct obd_export *export)
929 struct obd_device *obd = export->exp_obd;
930 struct obd_import *revimp;
932 LASSERT(export->exp_imp_reverse == NULL);
934 revimp = class_new_import(obd);
938 revimp->imp_remote_handle.cookie = 0ULL;
939 revimp->imp_client = &obd->obd_ldlm_client;
940 revimp->imp_dlm_fake = 1;
942 /* it is safe to connect import in new state as no sends possible */
943 spin_lock(&export->exp_lock);
944 export->exp_imp_reverse = revimp;
945 spin_unlock(&export->exp_lock);
946 class_import_put(revimp);
950 EXPORT_SYMBOL(rev_import_init);
953 * Handle reconnect for an export.
955 * \param exp export to handle reconnect process
956 * \param req client reconnect request
958 * \retval -rc in case securitfy flavor can't be changed
959 * \retval 0 in case none problems
961 static int rev_import_reconnect(struct obd_export *exp,
962 struct ptlrpc_request *req)
964 struct obd_import *revimp = exp->exp_imp_reverse;
965 struct lustre_handle *lh;
968 /* avoid sending a request until import flags are changed */
969 ptlrpc_import_enter_resend(revimp);
971 if (revimp->imp_connection != NULL)
972 ptlrpc_connection_put(revimp->imp_connection);
975 * client from recovery don't have a handle so we need to take from
976 * request. it may produce situation when wrong client connected
977 * to recovery as we trust a client uuid
979 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
980 revimp->imp_remote_handle = *lh;
983 * unknown versions will be caught in
984 * ptlrpc_handle_server_req_in->lustre_unpack_msg()
986 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
988 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
990 rc = rev_import_flags_update(revimp, req);
993 * it is safe to still be in RECOVERY phase as we are not able
994 * to setup correct security flavor so requests are not able to
995 * be delivered correctly
1000 /* resend all rpc's via new connection */
1001 return ptlrpc_import_recovery_state_machine(revimp);
1004 int target_handle_connect(struct ptlrpc_request *req)
1006 struct obd_device *target = NULL;
1007 struct obd_export *export = NULL;
1009 * connect handle - filled from target_handle_reconnect in
1012 struct lustre_handle conn;
1013 struct lustre_handle *tmp;
1014 struct obd_uuid cluuid;
1019 bool mds_conn = false, lw_client = false, initial_conn = false;
1020 bool mds_mds_conn = false;
1021 bool new_mds_mds_conn = false;
1022 struct obd_connect_data *data, *tmpdata;
1024 lnet_nid_t *client_nid = NULL;
1028 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1030 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1032 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1033 GOTO(out, rc = -EINVAL);
1036 target = class_dev_by_str(str);
1038 deuuidify(str, NULL, &target_start, &target_len);
1039 LCONSOLE_ERROR_MSG(0x137,
1040 "%s: not available for connect from %s (no target). If you are running an HA pair check that the target is mounted on the other server.\n",
1041 str, libcfs_nid2str(req->rq_peer.nid));
1042 GOTO(out, rc = -ENODEV);
1045 spin_lock(&target->obd_dev_lock);
1047 target->obd_conn_inprogress++;
1049 if (target->obd_stopping || !target->obd_set_up) {
1050 spin_unlock(&target->obd_dev_lock);
1052 deuuidify(str, NULL, &target_start, &target_len);
1053 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1054 target_len, target_start,
1055 libcfs_nid2str(req->rq_peer.nid),
1056 (target->obd_stopping ?
1057 "stopping" : "not set up"));
1058 GOTO(out, rc = -ENODEV);
1061 if (target->obd_no_conn) {
1062 spin_unlock(&target->obd_dev_lock);
1065 "%s: Temporarily refusing client connection from %s\n",
1066 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1067 GOTO(out, rc = -EAGAIN);
1070 spin_unlock(&target->obd_dev_lock);
1072 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1074 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1075 GOTO(out, rc = -EINVAL);
1078 obd_str2uuid(&cluuid, str);
1080 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1082 GOTO(out, rc = -EPROTO);
1086 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1088 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1090 GOTO(out, rc = -EPROTO);
1092 rc = req_capsule_server_pack(&req->rq_pill);
1096 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1098 * Don't allow clients to connect that are using old 1.8 format
1099 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1100 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1101 * FULL20 flag should be set on all connections since 2.0, but no
1102 * longer affects behaviour.
1104 * Later this check will be disabled and the flag can be retired
1105 * completely once interop with 3.0 is no longer needed.
1107 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1108 GOTO(out, rc = -EPROTO);
1111 * Don't allow liblustre clients to connect.
1112 * - testing was disabled in v2_2_50_0-61-g6a75d65
1113 * - building was disabled in v2_5_58_0-28-g7277179
1114 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1115 * - clients were refused connect for version difference > 0.0.1.32
1117 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1118 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1119 GOTO(out, rc = -EPROTO);
1124 * Note: lw_client is needed in MDS-MDS failover during update log
1125 * processing, so we needs to allow lw_client to be connected at
1126 * anytime, instead of only the initial connection
1128 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1130 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1131 initial_conn = true;
1132 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1133 mds_mds_conn = (data->ocd_connect_flags &
1134 OBD_CONNECT_MDS_MDS) != 0;
1137 * OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1138 * for Imperative Recovery connection from MGC to MGS.
1140 * Via check OBD_CONNECT_FID, we can distinguish whether
1141 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1142 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1145 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1146 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1147 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1148 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1149 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1150 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1153 * We do not support the MDT-MDT interoperations with
1154 * different version MDT because of protocol changes.
1156 if (unlikely(major != LUSTRE_MAJOR ||
1157 minor != LUSTRE_MINOR ||
1158 abs(patch - LUSTRE_PATCH) > 3)) {
1159 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1160 target->obd_name, LUSTRE_MAJOR,
1161 LUSTRE_MINOR, LUSTRE_PATCH,
1162 LUSTRE_FIX, major, minor, patch,
1163 OBD_OCD_VERSION_FIX(data->ocd_version),
1164 libcfs_nid2str(req->rq_peer.nid),
1166 GOTO(out, rc = -EPROTO);
1171 /* lctl gets a backstage, all-access pass. */
1172 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1173 goto dont_check_exports;
1175 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1179 /* We've found an export in the hash. */
1181 spin_lock(&export->exp_lock);
1183 if (export->exp_connecting) { /* b=9635, et. al. */
1184 spin_unlock(&export->exp_lock);
1185 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1186 export->exp_obd->obd_name, export,
1187 libcfs_nid2str(req->rq_peer.nid));
1188 class_export_put(export);
1191 } else if ((mds_conn || (lw_client && initial_conn) ||
1192 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1193 export->exp_connection != NULL) {
1194 spin_unlock(&export->exp_lock);
1195 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1196 /* MDS or LWP reconnected after failover. */
1197 LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1199 mds_conn ? "MDS" : "LWP",
1200 libcfs_nid2str(req->rq_peer.nid),
1201 libcfs_nid2str(export->exp_connection->c_peer.nid));
1203 /* New MDS connection from the same NID. */
1204 LCONSOLE_WARN("%s: Received new %s connection from %s, removing former export from same NID\n",
1206 mds_conn ? "MDS" : "LWP",
1207 libcfs_nid2str(req->rq_peer.nid));
1210 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1211 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1213 * Because exports between MDTs will always be
1214 * kept, let's do not fail such export if they
1215 * come from the same NID, otherwise it might
1216 * cause eviction between MDTs, which might
1217 * cause namespace inconsistency
1219 spin_lock(&export->exp_lock);
1220 export->exp_connecting = 1;
1221 export->exp_conn_cnt = 0;
1222 spin_unlock(&export->exp_lock);
1223 conn.cookie = export->exp_handle.h_cookie;
1226 class_fail_export(export);
1227 class_export_put(export);
1231 } else if (export->exp_connection != NULL && initial_conn &&
1232 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1233 spin_unlock(&export->exp_lock);
1234 /* In MDS failover we have static UUID but NID can change. */
1235 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1236 target->obd_name, cluuid.uuid,
1237 libcfs_nid2str(req->rq_peer.nid),
1239 export->exp_connection->c_peer.nid));
1241 class_export_put(export);
1244 export->exp_connecting = 1;
1245 spin_unlock(&export->exp_lock);
1246 LASSERT(export->exp_obd == target);
1248 rc = target_handle_reconnect(&conn, export, &cluuid);
1251 /* If we found an export, we already unlocked. */
1254 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1255 } else if (req->rq_export == NULL &&
1256 atomic_read(&export->exp_rpc_count) > 0) {
1257 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1258 target->obd_name, cluuid.uuid,
1259 libcfs_nid2str(req->rq_peer.nid),
1260 atomic_read(&export->exp_refcount));
1261 GOTO(out, rc = -EBUSY);
1262 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1264 if (!strstr(cluuid.uuid, "mdt"))
1265 LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client",
1266 target->obd_name, cluuid.uuid,
1267 libcfs_nid2str(req->rq_peer.nid));
1268 GOTO(out, rc = -EALREADY);
1270 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1276 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1277 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1278 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1279 export, ktime_get_real_seconds(),
1280 export ? export->exp_last_request_time : 0);
1283 * If this is the first time a client connects, reset the recovery
1284 * timer. Discard lightweight connections which might be local.
1286 if (!lw_client && rc == 0 && target->obd_recovering)
1287 check_and_start_recovery_timer(target, req, export == NULL);
1290 * We want to handle EALREADY but *not* -EALREADY from
1291 * target_handle_reconnect(), return reconnection state in a flag.
1293 if (rc == EALREADY) {
1294 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1300 /* Tell the client if we support replayable requests. */
1301 if (target->obd_replayable)
1302 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1303 client_nid = &req->rq_peer.nid;
1305 if (export == NULL) {
1306 /* allow lightweight connections during recovery */
1308 * allow "new" MDT to be connected during recovery, since we
1309 * need retrieve recovery update records from it
1311 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1312 struct hrtimer *timer = &target->obd_recovery_timer;
1321 connected = atomic_read(&target->obd_connected_clients);
1322 in_progress = atomic_read(&target->obd_lock_replay_clients);
1323 known = target->obd_max_recoverable_clients;
1324 stale = target->obd_stale_clients;
1325 remaining = hrtimer_expires_remaining(timer);
1326 left = ktime_divns(remaining, NSEC_PER_SEC);
1328 if (ktime_to_ns(remaining) > 0) {
1329 msg = "to recover in";
1332 msg = "already passed deadline";
1336 LCONSOLE_WARN("%s: Denying connection for new client %s (at %s), waiting for %d known clients (%d recovered, %d in progress, and %d evicted) %s %lld:%.02lld\n",
1337 target->obd_name, cluuid.uuid,
1338 libcfs_nid2str(req->rq_peer.nid), known,
1339 connected - in_progress, in_progress,
1340 stale, msg, timeout / 60, timeout % 60);
1344 rc = obd_connect(req->rq_svc_thread->t_env,
1345 &export, target, &cluuid, data,
1347 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1348 lustre_msg_add_op_flags(req->rq_repmsg,
1349 MSG_CONNECT_RECOVERING);
1351 conn.cookie = export->exp_handle.h_cookie;
1352 rc = rev_import_init(export);
1356 new_mds_mds_conn = true;
1359 rc = obd_reconnect(req->rq_svc_thread->t_env,
1360 export, target, &cluuid, data, client_nid);
1365 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1366 data->ocd_instance = target->u.obt.obt_instance;
1369 * Return only the parts of obd_connect_data that we understand, so the
1370 * client knows that we don't understand the rest.
1373 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1375 tmpdata = req_capsule_server_get(&req->rq_pill,
1378 * Don't use struct assignment here, because the client reply
1379 * buffer may be smaller/larger than the local struct
1382 memcpy(tmpdata, data, min(tmpsize, size));
1386 * If the client and the server are the same node, we will already
1387 * have an export that really points to the client's DLM export,
1388 * because we have a shared handles table.
1390 * XXX this will go away when shaver stops sending the "connect" handle
1391 * in the real "remote handle" field of the request --phik 24 Apr 2003
1393 ptlrpc_request_change_export(req, export);
1395 spin_lock(&export->exp_lock);
1396 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1397 spin_unlock(&export->exp_lock);
1399 "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1400 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1401 export->exp_conn_cnt,
1402 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1404 GOTO(out, rc = -EALREADY);
1406 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1407 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1408 spin_unlock(&export->exp_lock);
1410 if (export->exp_connection != NULL) {
1411 /* Check to see if connection came from another NID. */
1412 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1413 !hlist_unhashed(&export->exp_nid_hash))
1414 cfs_hash_del(export->exp_obd->obd_nid_hash,
1415 &export->exp_connection->c_peer.nid,
1416 &export->exp_nid_hash);
1418 ptlrpc_connection_put(export->exp_connection);
1421 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1424 if (hlist_unhashed(&export->exp_nid_hash))
1425 cfs_hash_add(export->exp_obd->obd_nid_hash,
1426 &export->exp_connection->c_peer.nid,
1427 &export->exp_nid_hash);
1429 lustre_msg_set_handle(req->rq_repmsg, &conn);
1431 rc = rev_import_reconnect(export, req);
1435 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1437 __u64 transno = data->ocd_transno;
1439 spin_lock(&export->exp_lock);
1441 * possible race with class_disconnect_stale_exports,
1442 * export may be already in the eviction process
1444 if (export->exp_failed) {
1445 spin_unlock(&export->exp_lock);
1446 GOTO(out, rc = -ENODEV);
1448 export->exp_in_recovery = 1;
1449 export->exp_req_replay_needed = 1;
1450 export->exp_lock_replay_needed = 1;
1451 spin_unlock(&export->exp_lock);
1453 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1454 MSG_CONNECT_TRANSNO);
1455 if (has_transno && transno == 0)
1456 CWARN("Connect with zero transno!\n");
1458 if (has_transno && transno > 0 &&
1459 transno < target->obd_next_recovery_transno &&
1460 transno > target->obd_last_committed) {
1461 /* Another way is to use cmpxchg() to be lock-free. */
1462 spin_lock(&target->obd_recovery_task_lock);
1463 if (transno < target->obd_next_recovery_transno)
1464 target->obd_next_recovery_transno = transno;
1465 spin_unlock(&target->obd_recovery_task_lock);
1468 atomic_inc(&target->obd_req_replay_clients);
1469 atomic_inc(&target->obd_lock_replay_clients);
1471 * Note: MDS-MDS connection is allowed to be connected during
1472 * recovery, no matter if the exports needs to be recoveried.
1473 * Because we need retrieve updates logs from all other MDTs.
1474 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1475 * also needs to be increased to match other recovery checking
1478 if (new_mds_mds_conn)
1479 target->obd_max_recoverable_clients++;
1480 if (atomic_inc_return(&target->obd_connected_clients) ==
1481 target->obd_max_recoverable_clients)
1482 wake_up(&target->obd_next_transno_waitq);
1485 /* Tell the client we're in recovery, when client is involved in it. */
1486 if (target->obd_recovering && !lw_client)
1487 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1491 spin_lock(&export->exp_lock);
1492 export->exp_connecting = 0;
1493 spin_unlock(&export->exp_lock);
1495 class_export_put(export);
1497 if (target != NULL) {
1498 spin_lock(&target->obd_dev_lock);
1499 target->obd_conn_inprogress--;
1500 spin_unlock(&target->obd_dev_lock);
1501 class_decref(target, "find", current);
1503 req->rq_status = rc;
1507 int target_handle_disconnect(struct ptlrpc_request *req)
1513 rc = req_capsule_server_pack(&req->rq_pill);
1517 /* Keep the rq_export around so we can send the reply. */
1518 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1523 void target_destroy_export(struct obd_export *exp)
1525 struct obd_import *imp = NULL;
1527 * exports created from last_rcvd data, and "fake"
1528 * exports created by lctl don't have an import
1530 spin_lock(&exp->exp_lock);
1531 if (exp->exp_imp_reverse != NULL) {
1532 imp = exp->exp_imp_reverse;
1533 exp->exp_imp_reverse = NULL;
1535 spin_unlock(&exp->exp_lock);
1537 client_destroy_import(imp);
1539 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1540 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1541 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1542 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1544 EXPORT_SYMBOL(target_destroy_export);
1547 * Recovery functions
1549 static void target_request_copy_get(struct ptlrpc_request *req)
1551 class_export_rpc_inc(req->rq_export);
1552 LASSERT(list_empty(&req->rq_list));
1553 INIT_LIST_HEAD(&req->rq_replay_list);
1555 /* Increase refcount to keep request in queue. */
1556 atomic_inc(&req->rq_refcount);
1557 /* Let export know it has replays to be handled. */
1558 atomic_inc(&req->rq_export->exp_replay_count);
1561 static void target_request_copy_put(struct ptlrpc_request *req)
1563 LASSERT(list_empty(&req->rq_replay_list));
1564 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1566 atomic_dec(&req->rq_export->exp_replay_count);
1567 class_export_rpc_dec(req->rq_export);
1568 ptlrpc_server_drop_request(req);
1571 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1573 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1574 struct obd_export *exp = req->rq_export;
1575 struct ptlrpc_request *reqiter;
1576 struct ptlrpc_request *dup_req = NULL;
1581 spin_lock(&exp->exp_lock);
1582 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1584 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1592 /* We expect it with RESENT and REPLAY flags. */
1593 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1594 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1595 CERROR("invalid flags %x of resent replay\n",
1596 lustre_msg_get_flags(req->rq_reqmsg));
1598 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1601 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1603 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1604 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1608 list_add_tail(&req->rq_replay_list,
1609 &exp->exp_req_replay_queue);
1612 spin_unlock(&exp->exp_lock);
1616 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1618 LASSERT(!list_empty(&req->rq_replay_list));
1619 LASSERT(req->rq_export);
1621 spin_lock(&req->rq_export->exp_lock);
1622 list_del_init(&req->rq_replay_list);
1623 spin_unlock(&req->rq_export->exp_lock);
1626 static void target_finish_recovery(struct lu_target *lut)
1628 struct obd_device *obd = lut->lut_obd;
1632 /* Only log a recovery message when recovery has occurred. */
1633 if (obd->obd_recovery_start) {
1634 time64_t now = ktime_get_real_seconds();
1635 time64_t elapsed_time;
1637 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1639 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1640 obd->obd_name, (s64)elapsed_time / 60,
1641 (s64)elapsed_time % 60,
1642 obd->obd_max_recoverable_clients,
1643 atomic_read(&obd->obd_connected_clients),
1644 obd->obd_stale_clients,
1645 obd->obd_stale_clients == 1 ? "was" : "were");
1648 ldlm_reprocess_recovery_done(obd->obd_namespace);
1649 spin_lock(&obd->obd_recovery_task_lock);
1650 if (!list_empty(&obd->obd_req_replay_queue) ||
1651 !list_empty(&obd->obd_lock_replay_queue) ||
1652 !list_empty(&obd->obd_final_req_queue)) {
1653 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1655 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1656 list_empty(&obd->obd_lock_replay_queue) ? \
1658 list_empty(&obd->obd_final_req_queue) ? \
1660 spin_unlock(&obd->obd_recovery_task_lock);
1663 spin_unlock(&obd->obd_recovery_task_lock);
1665 obd->obd_recovery_end = ktime_get_real_seconds();
1667 /* When recovery finished, cleanup orphans on MDS and OST. */
1668 if (obd->obd_type && OBP(obd, postrecov)) {
1669 int rc = OBP(obd, postrecov)(obd);
1672 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1678 static void abort_req_replay_queue(struct obd_device *obd)
1680 struct ptlrpc_request *req, *n;
1681 struct list_head abort_list;
1683 INIT_LIST_HEAD(&abort_list);
1684 spin_lock(&obd->obd_recovery_task_lock);
1685 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1686 spin_unlock(&obd->obd_recovery_task_lock);
1687 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1688 DEBUG_REQ(D_WARNING, req, "aborted:");
1689 req->rq_status = -ENOTCONN;
1690 if (ptlrpc_error(req)) {
1691 DEBUG_REQ(D_ERROR, req,
1692 "failed abort_req_reply; skipping");
1694 target_exp_dequeue_req_replay(req);
1695 target_request_copy_put(req);
1699 static void abort_lock_replay_queue(struct obd_device *obd)
1701 struct ptlrpc_request *req, *n;
1702 struct list_head abort_list;
1704 INIT_LIST_HEAD(&abort_list);
1705 spin_lock(&obd->obd_recovery_task_lock);
1706 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1707 spin_unlock(&obd->obd_recovery_task_lock);
1708 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1709 DEBUG_REQ(D_ERROR, req, "aborted:");
1710 req->rq_status = -ENOTCONN;
1711 if (ptlrpc_error(req)) {
1712 DEBUG_REQ(D_ERROR, req,
1713 "failed abort_lock_reply; skipping");
1715 target_request_copy_put(req);
1720 * Called from a cleanup function if the device is being cleaned up
1721 * forcefully. The exports should all have been disconnected already,
1722 * the only thing left to do is
1723 * - clear the recovery flags
1724 * - cancel the timer
1725 * - free queued requests and replies, but don't send replies
1726 * Because the obd_stopping flag is set, no new requests should be received.
1728 void target_cleanup_recovery(struct obd_device *obd)
1730 struct ptlrpc_request *req, *n;
1731 struct list_head clean_list;
1733 INIT_LIST_HEAD(&clean_list);
1734 spin_lock(&obd->obd_dev_lock);
1735 if (!obd->obd_recovering) {
1736 spin_unlock(&obd->obd_dev_lock);
1740 obd->obd_recovering = obd->obd_abort_recovery = 0;
1741 spin_unlock(&obd->obd_dev_lock);
1743 spin_lock(&obd->obd_recovery_task_lock);
1744 target_cancel_recovery_timer(obd);
1745 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1746 spin_unlock(&obd->obd_recovery_task_lock);
1748 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1749 LASSERT(req->rq_reply_state == NULL);
1750 target_exp_dequeue_req_replay(req);
1751 target_request_copy_put(req);
1754 spin_lock(&obd->obd_recovery_task_lock);
1755 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1756 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1757 spin_unlock(&obd->obd_recovery_task_lock);
1759 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1760 LASSERT(req->rq_reply_state == NULL);
1761 target_request_copy_put(req);
1766 EXPORT_SYMBOL(target_cleanup_recovery);
1768 /* obd_recovery_task_lock should be held */
1769 void target_cancel_recovery_timer(struct obd_device *obd)
1771 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1772 hrtimer_cancel(&obd->obd_recovery_timer);
1775 static void target_start_recovery_timer(struct obd_device *obd)
1779 if (obd->obd_recovery_start != 0)
1782 spin_lock(&obd->obd_dev_lock);
1783 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1784 spin_unlock(&obd->obd_dev_lock);
1788 LASSERT(obd->obd_recovery_timeout != 0);
1790 if (obd->obd_recovery_start != 0) {
1791 spin_unlock(&obd->obd_dev_lock);
1795 delay = ktime_set(obd->obd_recovery_timeout, 0);
1796 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_REL);
1797 obd->obd_recovery_start = ktime_get_real_seconds();
1798 spin_unlock(&obd->obd_dev_lock);
1800 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1802 obd->obd_recovery_timeout / 60,
1803 obd->obd_recovery_timeout % 60,
1804 obd->obd_max_recoverable_clients,
1805 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1806 (obd->obd_max_recoverable_clients == 1) ? "s" : "");
1810 * extend recovery window.
1812 * if @extend is true, extend recovery window to have @drt remaining at least;
1813 * otherwise, make sure the recovery timeout value is not less than @drt.
1815 static void extend_recovery_timer(struct obd_device *obd, time_t drt,
1822 spin_lock(&obd->obd_dev_lock);
1823 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1824 spin_unlock(&obd->obd_dev_lock);
1827 LASSERT(obd->obd_recovery_start != 0);
1829 to = obd->obd_recovery_timeout;
1830 left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1831 left = ktime_divns(left_ns, NSEC_PER_SEC);
1833 if (extend && (drt > left))
1835 else if (!extend && (drt > to))
1838 if (to > obd->obd_recovery_time_hard) {
1839 to = obd->obd_recovery_time_hard;
1840 CWARN("%s: extended recovery timer reaching hard limit: %ld, extend: %d\n",
1841 obd->obd_name, to, extend);
1844 if (obd->obd_recovery_timeout < to) {
1845 ktime_t now = ktime_get_real();
1848 obd->obd_recovery_timeout = to;
1849 end = ktime_set(obd->obd_recovery_start + to, 0);
1850 left_ns = ktime_sub(end, now);
1851 hrtimer_forward_now(&obd->obd_recovery_timer, left_ns);
1852 left = ktime_divns(left_ns, NSEC_PER_MSEC);
1854 spin_unlock(&obd->obd_dev_lock);
1856 CDEBUG(D_HA, "%s: recovery timer will expire in %ld seconds\n",
1857 obd->obd_name, left);
1860 /* Reset the timer with each new client connection */
1862 * This timer is actually reconnect_timer, which is for making sure
1863 * the total recovery window is at least as big as my reconnect
1864 * attempt timing. So the initial recovery time_out will be set to
1865 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1866 * from client is bigger than this, then the recovery time_out will
1867 * be extended to make sure the client could be reconnected, in the
1868 * process, the timeout from the new client should be ignored.
1871 check_and_start_recovery_timer(struct obd_device *obd,
1872 struct ptlrpc_request *req,
1875 time_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1876 struct obd_device_target *obt = &obd->u.obt;
1878 if (!new_client && service_time)
1880 * Teach server about old server's estimates, as first guess
1881 * at how long new requests will take.
1883 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1886 target_start_recovery_timer(obd);
1889 * Convert the service time to RPC timeout,
1890 * and reuse service_time to limit stack usage.
1892 service_time = at_est2timeout(service_time);
1894 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1895 service_time < at_extra)
1896 service_time = at_extra;
1899 * We expect other clients to timeout within service_time, then try
1900 * to reconnect, then try the failover server. The max delay between
1901 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
1903 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1905 LASSERT(obt->obt_magic == OBT_MAGIC);
1906 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1907 if (service_time > obd->obd_recovery_timeout && !new_client)
1908 extend_recovery_timer(obd, service_time, false);
1911 /** Health checking routines */
1912 static inline int exp_connect_healthy(struct obd_export *exp)
1914 return exp->exp_in_recovery;
1917 /** if export done req_replay or has replay in queue */
1918 static inline int exp_req_replay_healthy(struct obd_export *exp)
1920 return (!exp->exp_req_replay_needed ||
1921 atomic_read(&exp->exp_replay_count) > 0);
1925 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1927 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1928 exp_req_replay_healthy(exp);
1931 /** if export done lock_replay or has replay in queue */
1932 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1934 return (!exp->exp_lock_replay_needed ||
1935 atomic_read(&exp->exp_replay_count) > 0);
1938 static inline int exp_vbr_healthy(struct obd_export *exp)
1940 return !exp->exp_vbr_failed;
1943 static inline int exp_finished(struct obd_export *exp)
1945 return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
1948 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1950 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1954 static int check_for_next_transno(struct lu_target *lut)
1956 struct ptlrpc_request *req = NULL;
1957 struct obd_device *obd = lut->lut_obd;
1958 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1959 int wake_up = 0, connected, completed, queue_len;
1960 __u64 req_transno = 0;
1961 __u64 update_transno = 0;
1962 __u64 next_transno = 0;
1966 spin_lock(&obd->obd_recovery_task_lock);
1967 if (!list_empty(&obd->obd_req_replay_queue)) {
1968 req = list_entry(obd->obd_req_replay_queue.next,
1969 struct ptlrpc_request, rq_list);
1970 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1974 update_transno = distribute_txn_get_next_transno(tdtd);
1976 connected = atomic_read(&obd->obd_connected_clients);
1977 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1978 queue_len = obd->obd_requests_queued_for_recovery;
1979 next_transno = obd->obd_next_recovery_transno;
1982 "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
1983 obd->obd_max_recoverable_clients, connected, completed,
1984 queue_len, req_transno, next_transno);
1986 if (obd->obd_abort_recovery) {
1987 CDEBUG(D_HA, "waking for aborted recovery\n");
1989 } else if (obd->obd_recovery_expired) {
1990 CDEBUG(D_HA, "waking for expired recovery\n");
1992 } else if (tdtd != NULL && req != NULL &&
1993 is_req_replayed_by_update(req)) {
1994 LASSERTF(req_transno < next_transno,
1995 "req_transno %llu next_transno%llu\n", req_transno,
1997 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2000 } else if (req_transno == next_transno ||
2001 (update_transno != 0 && update_transno <= next_transno)) {
2002 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2004 } else if (queue_len > 0 &&
2005 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2006 /** handle gaps occured due to lost reply or VBR */
2007 LASSERTF(req_transno >= next_transno,
2008 "req_transno: %llu, next_transno: %llu\n",
2009 req_transno, next_transno);
2011 "%s: waking for gap in transno, VBR is %s (skip: %lld, ql: %d, comp: %d, conn: %d, next: %lld, next_update %lld last_committed: %lld)\n",
2012 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2013 next_transno, queue_len, completed, connected,
2014 req_transno, update_transno, obd->obd_last_committed);
2015 obd->obd_next_recovery_transno = req_transno;
2017 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2018 CDEBUG(D_HA, "waking for completed recovery\n");
2020 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2022 "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2024 obd->obd_next_recovery_transno = req_transno;
2027 spin_unlock(&obd->obd_recovery_task_lock);
2031 static int check_for_next_lock(struct lu_target *lut)
2033 struct obd_device *obd = lut->lut_obd;
2036 spin_lock(&obd->obd_recovery_task_lock);
2037 if (!list_empty(&obd->obd_lock_replay_queue)) {
2038 CDEBUG(D_HA, "waking for next lock\n");
2040 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2041 CDEBUG(D_HA, "waking for completed lock replay\n");
2043 } else if (obd->obd_abort_recovery) {
2044 CDEBUG(D_HA, "waking for aborted recovery\n");
2046 } else if (obd->obd_recovery_expired) {
2047 CDEBUG(D_HA, "waking for expired recovery\n");
2050 spin_unlock(&obd->obd_recovery_task_lock);
2056 * wait for recovery events,
2057 * check its status with help of check_routine
2058 * evict dead clients via health_check
2060 static int target_recovery_overseer(struct lu_target *lut,
2061 int (*check_routine)(struct lu_target *),
2062 int (*health_check)(struct obd_export *))
2064 struct obd_device *obd = lut->lut_obd;
2065 struct target_distribute_txn_data *tdtd;
2069 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2070 now = ktime_get_seconds();
2071 if (now - last > 600) {
2072 LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2077 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
2078 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2079 __u64 next_update_transno = 0;
2082 * Only abort the recovery if there are no update recovery
2085 spin_lock(&obd->obd_recovery_task_lock);
2086 if (lut->lut_tdtd != NULL) {
2087 next_update_transno =
2088 distribute_txn_get_next_transno(lut->lut_tdtd);
2090 tdtd = lut->lut_tdtd;
2092 * If next_update_transno == 0, it probably because
2093 * updatelog retrieve threads did not get any records
2094 * yet, let's wait those threads stopped
2096 if (next_update_transno == 0) {
2097 struct l_wait_info lwi = { 0 };
2099 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2101 &tdtd->tdtd_recovery_threads_count) == 0,
2104 next_update_transno =
2105 distribute_txn_get_next_transno(
2110 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2111 obd->obd_next_recovery_transno = next_update_transno;
2112 spin_unlock(&obd->obd_recovery_task_lock);
2114 * Disconnect unfinished exports from clients, and
2115 * keep connection from MDT to make sure the update
2116 * recovery will still keep trying until some one
2117 * manually abort the recovery
2119 class_disconnect_stale_exports(obd,
2120 exp_finished_or_from_mdt);
2121 /* Abort all of replay & replay lock req from clients */
2122 abort_req_replay_queue(obd);
2123 abort_lock_replay_queue(obd);
2125 "%s: there are still update replay (%#llx)in the queue.\n",
2126 obd->obd_name, next_update_transno);
2128 obd->obd_abort_recovery = 1;
2129 spin_unlock(&obd->obd_recovery_task_lock);
2130 CWARN("%s recovery is aborted by hard timeout\n",
2135 while (wait_event_timeout(obd->obd_next_transno_waitq,
2137 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2138 ; /* wait indefinitely for event, but don't trigger watchdog */
2140 if (obd->obd_abort_recovery) {
2141 CWARN("recovery is aborted, evict exports in recovery\n");
2142 if (lut->lut_tdtd != NULL) {
2143 struct l_wait_info lwi = { 0 };
2145 tdtd = lut->lut_tdtd;
2147 * Let's wait all of the update log recovery thread
2150 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2151 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2153 /* Then abort the update recovery list */
2154 dtrq_list_destroy(lut->lut_tdtd);
2157 /** evict exports which didn't finish recovery yet */
2158 class_disconnect_stale_exports(obd, exp_finished);
2160 } else if (obd->obd_recovery_expired) {
2161 obd->obd_recovery_expired = 0;
2162 /** If some clients died being recovered, evict them */
2163 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2165 /** evict cexports with no replay in queue, they are stalled */
2166 class_disconnect_stale_exports(obd, health_check);
2168 /** continue with VBR */
2169 spin_lock(&obd->obd_dev_lock);
2170 obd->obd_version_recov = 1;
2171 spin_unlock(&obd->obd_dev_lock);
2173 * reset timer, recovery will proceed with versions now,
2174 * timeout is set just to handle reconnection delays
2176 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2178 * Wait for recovery events again, after evicting bad clients
2185 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2187 struct obd_device *obd = lut->lut_obd;
2188 struct ptlrpc_request *req = NULL;
2190 CDEBUG(D_HA, "Waiting for lock\n");
2191 if (target_recovery_overseer(lut, check_for_next_lock,
2192 exp_lock_replay_healthy))
2193 abort_lock_replay_queue(obd);
2195 spin_lock(&obd->obd_recovery_task_lock);
2196 if (!list_empty(&obd->obd_lock_replay_queue)) {
2197 req = list_entry(obd->obd_lock_replay_queue.next,
2198 struct ptlrpc_request, rq_list);
2199 list_del_init(&req->rq_list);
2200 spin_unlock(&obd->obd_recovery_task_lock);
2202 spin_unlock(&obd->obd_recovery_task_lock);
2203 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2204 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2205 /** evict exports failed VBR */
2206 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2211 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2213 struct ptlrpc_request *req = NULL;
2215 spin_lock(&obd->obd_recovery_task_lock);
2216 if (!list_empty(&obd->obd_final_req_queue)) {
2217 req = list_entry(obd->obd_final_req_queue.next,
2218 struct ptlrpc_request, rq_list);
2219 list_del_init(&req->rq_list);
2220 spin_unlock(&obd->obd_recovery_task_lock);
2221 if (req->rq_export->exp_in_recovery) {
2222 spin_lock(&req->rq_export->exp_lock);
2223 req->rq_export->exp_in_recovery = 0;
2224 spin_unlock(&req->rq_export->exp_lock);
2227 spin_unlock(&obd->obd_recovery_task_lock);
2232 static void handle_recovery_req(struct ptlrpc_thread *thread,
2233 struct ptlrpc_request *req,
2234 svc_handler_t handler)
2239 * export can be evicted during recovery, no need to handle replays for
2240 * it after that, discard such request silently
2242 if (req->rq_export->exp_disconnected)
2245 req->rq_session.lc_thread = thread;
2246 req->rq_svc_thread = thread;
2247 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2249 /* thread context */
2250 lu_context_enter(&thread->t_env->le_ctx);
2252 lu_context_exit(&thread->t_env->le_ctx);
2254 /* don't reset timer for final stage */
2255 if (!exp_finished(req->rq_export)) {
2256 time_t to = obd_timeout;
2259 * Add request timeout to the recovery time so next request from
2260 * this client may come in recovery time
2263 struct ptlrpc_service_part *svcpt;
2265 svcpt = req->rq_rqbd->rqbd_svcpt;
2267 * If the server sent early reply for this request,
2268 * the client will recalculate the timeout according to
2269 * current server estimate service time, so we will
2270 * use the maxium timeout here for waiting the client
2271 * sending the next req
2274 at_est2timeout(at_get(&svcpt->scp_at_estimate)),
2275 lustre_msg_get_timeout(req->rq_reqmsg));
2277 * Add 2 net_latency, one for balance rq_deadline
2278 * (see ptl_send_rpc), one for resend the req to server,
2279 * Note: client will pack net_latency in replay req
2280 * (see ptlrpc_replay_req)
2282 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2284 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2289 /** Checking routines for recovery */
2290 static int check_for_recovery_ready(struct lu_target *lut)
2292 struct obd_device *obd = lut->lut_obd;
2293 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2296 "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2297 clnts, obd->obd_stale_clients, obd->obd_max_recoverable_clients,
2298 obd->obd_abort_recovery, obd->obd_recovery_expired);
2300 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2301 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2302 if (clnts + obd->obd_stale_clients <
2303 obd->obd_max_recoverable_clients)
2307 if (lut->lut_tdtd != NULL) {
2308 if (!lut->lut_tdtd->tdtd_replay_ready &&
2309 !obd->obd_abort_recovery) {
2311 * Let's extend recovery timer, in case the recovery
2312 * timer expired, and some clients got evicted
2314 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2317 "%s update recovery is not ready, extend recovery %llu\n",
2318 obd->obd_name, obd->obd_recovery_timeout);
2327 REQUEST_RECOVERY = 1,
2328 UPDATE_RECOVERY = 2,
2331 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2335 if (!list_empty(&obd->obd_req_replay_queue)) {
2336 struct ptlrpc_request *req;
2338 req = list_entry(obd->obd_req_replay_queue.next,
2339 struct ptlrpc_request, rq_list);
2340 transno = lustre_msg_get_transno(req->rq_reqmsg);
2346 static __u64 get_next_transno(struct lu_target *lut, int *type)
2348 struct obd_device *obd = lut->lut_obd;
2349 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2351 __u64 update_transno;
2355 transno = get_next_replay_req_transno(obd);
2357 *type = REQUEST_RECOVERY;
2362 update_transno = distribute_txn_get_next_transno(tdtd);
2363 if (transno == 0 || (transno >= update_transno &&
2364 update_transno != 0)) {
2365 transno = update_transno;
2367 *type = UPDATE_RECOVERY;
2374 * drop duplicate replay request
2376 * Because the operation has been replayed by update recovery, the request
2377 * with the same transno will be dropped and also notify the client to send
2378 * next replay request.
2380 * \param[in] env execution environment
2381 * \param[in] obd failover obd device
2382 * \param[in] req request to be dropped
2384 static void drop_duplicate_replay_req(struct lu_env *env,
2385 struct obd_device *obd,
2386 struct ptlrpc_request *req)
2388 DEBUG_REQ(D_HA, req,
2389 "remove t%lld from %s because of duplicate update records are found.\n",
2390 lustre_msg_get_transno(req->rq_reqmsg),
2391 libcfs_nid2str(req->rq_peer.nid));
2394 * Right now, only for MDS reint operation update replay and
2395 * normal request replay can have the same transno
2397 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2398 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2399 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2400 if (likely(req->rq_export))
2401 target_committed_to_req(req);
2402 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2403 target_send_reply(req, req->rq_status, 0);
2405 DEBUG_REQ(D_ERROR, req, "wrong opc from %s\n",
2406 libcfs_nid2str(req->rq_peer.nid));
2408 target_exp_dequeue_req_replay(req);
2409 target_request_copy_put(req);
2410 obd->obd_replayed_requests++;
2413 static void replay_request_or_update(struct lu_env *env,
2414 struct lu_target *lut,
2415 struct target_recovery_data *trd,
2416 struct ptlrpc_thread *thread)
2418 struct obd_device *obd = lut->lut_obd;
2419 struct ptlrpc_request *req = NULL;
2425 CDEBUG(D_HA, "Waiting for transno %lld\n",
2426 obd->obd_next_recovery_transno);
2428 /* Replay all of request and update by transno */
2430 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2432 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2435 * It is needed to extend recovery window above
2436 * recovery_time_soft. Extending is possible only in the
2437 * end of recovery window (see more details in
2438 * handle_recovery_req()).
2440 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2442 if (target_recovery_overseer(lut, check_for_next_transno,
2443 exp_req_replay_healthy_or_from_mdt)) {
2444 abort_req_replay_queue(obd);
2445 abort_lock_replay_queue(obd);
2449 spin_lock(&obd->obd_recovery_task_lock);
2450 transno = get_next_transno(lut, &type);
2451 if (type == REQUEST_RECOVERY && transno != 0) {
2453 * Drop replay request from client side, if the
2454 * replay has been executed by update with the
2457 req = list_entry(obd->obd_req_replay_queue.next,
2458 struct ptlrpc_request, rq_list);
2460 list_del_init(&req->rq_list);
2461 obd->obd_requests_queued_for_recovery--;
2462 spin_unlock(&obd->obd_recovery_task_lock);
2465 * Let's check if the request has been redone by
2468 if (is_req_replayed_by_update(req)) {
2469 struct distribute_txn_replay_req *dtrq;
2471 dtrq = distribute_txn_lookup_finish_list(tdtd,
2473 LASSERT(dtrq != NULL);
2474 spin_lock(&tdtd->tdtd_replay_list_lock);
2475 list_del_init(&dtrq->dtrq_list);
2476 spin_unlock(&tdtd->tdtd_replay_list_lock);
2479 drop_duplicate_replay_req(env, obd, req);
2484 LASSERT(trd->trd_processing_task == current_pid());
2485 DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2487 lustre_msg_get_transno(req->rq_reqmsg),
2488 libcfs_nid2str(req->rq_peer.nid));
2490 handle_recovery_req(thread, req,
2491 trd->trd_recovery_handler);
2493 * bz18031: increase next_recovery_transno before
2494 * target_request_copy_put() will drop exp_rpc reference
2496 spin_lock(&obd->obd_recovery_task_lock);
2497 obd->obd_next_recovery_transno++;
2498 spin_unlock(&obd->obd_recovery_task_lock);
2499 target_exp_dequeue_req_replay(req);
2500 target_request_copy_put(req);
2501 obd->obd_replayed_requests++;
2502 } else if (type == UPDATE_RECOVERY && transno != 0) {
2503 struct distribute_txn_replay_req *dtrq;
2506 spin_unlock(&obd->obd_recovery_task_lock);
2508 LASSERT(tdtd != NULL);
2509 dtrq = distribute_txn_get_next_req(tdtd);
2510 lu_context_enter(&thread->t_env->le_ctx);
2511 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2512 lu_context_exit(&thread->t_env->le_ctx);
2513 extend_recovery_timer(obd, obd_timeout, true);
2515 if (rc == 0 && dtrq->dtrq_xid != 0) {
2517 "Move x%llu t%llu to finish list\n",
2519 dtrq->dtrq_master_transno);
2521 /* Add it to the replay finish list */
2522 spin_lock(&tdtd->tdtd_replay_list_lock);
2523 list_add(&dtrq->dtrq_list,
2524 &tdtd->tdtd_replay_finish_list);
2525 spin_unlock(&tdtd->tdtd_replay_list_lock);
2527 spin_lock(&obd->obd_recovery_task_lock);
2528 if (transno == obd->obd_next_recovery_transno)
2529 obd->obd_next_recovery_transno++;
2531 obd->obd_next_recovery_transno)
2532 obd->obd_next_recovery_transno =
2534 spin_unlock(&obd->obd_recovery_task_lock);
2539 spin_unlock(&obd->obd_recovery_task_lock);
2541 LASSERT(list_empty(&obd->obd_req_replay_queue));
2542 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2543 /** evict exports failed VBR */
2544 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2550 static int target_recovery_thread(void *arg)
2552 struct lu_target *lut = arg;
2553 struct obd_device *obd = lut->lut_obd;
2554 struct ptlrpc_request *req;
2555 struct target_recovery_data *trd = &obd->obd_recovery_data;
2556 unsigned long delta;
2558 struct ptlrpc_thread *thread = NULL;
2563 unshare_fs_struct();
2564 OBD_ALLOC_PTR(thread);
2570 GOTO(out_thread, rc = -ENOMEM);
2571 rc = lu_env_add(env);
2575 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2577 GOTO(out_env_remove, rc);
2579 thread->t_env = env;
2580 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2581 env->le_ctx.lc_thread = thread;
2582 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2584 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2586 trd->trd_processing_task = current_pid();
2588 spin_lock(&obd->obd_dev_lock);
2589 obd->obd_recovering = 1;
2590 spin_unlock(&obd->obd_dev_lock);
2591 complete(&trd->trd_starting);
2593 /* first of all, we have to know the first transno to replay */
2594 if (target_recovery_overseer(lut, check_for_recovery_ready,
2595 exp_connect_healthy)) {
2596 abort_req_replay_queue(obd);
2597 abort_lock_replay_queue(obd);
2598 if (lut->lut_tdtd != NULL)
2599 dtrq_list_destroy(lut->lut_tdtd);
2602 /* next stage: replay requests or update */
2604 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2605 atomic_read(&obd->obd_req_replay_clients),
2606 obd->obd_next_recovery_transno);
2607 replay_request_or_update(env, lut, trd, thread);
2610 * The second stage: replay locks
2612 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2613 atomic_read(&obd->obd_lock_replay_clients));
2614 while ((req = target_next_replay_lock(lut))) {
2615 LASSERT(trd->trd_processing_task == current_pid());
2616 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2617 libcfs_nid2str(req->rq_peer.nid));
2618 handle_recovery_req(thread, req,
2619 trd->trd_recovery_handler);
2620 target_request_copy_put(req);
2621 obd->obd_replayed_locks++;
2625 * The third stage: reply on final pings, at this moment all clients
2626 * must have request in final queue
2628 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2629 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2630 /** Update server last boot epoch */
2631 tgt_boot_epoch_update(lut);
2633 * We drop recoverying flag to forward all new requests
2634 * to regular mds_handle() since now
2636 spin_lock(&obd->obd_dev_lock);
2637 obd->obd_recovering = obd->obd_abort_recovery = 0;
2638 spin_unlock(&obd->obd_dev_lock);
2639 spin_lock(&obd->obd_recovery_task_lock);
2640 target_cancel_recovery_timer(obd);
2641 spin_unlock(&obd->obd_recovery_task_lock);
2642 while ((req = target_next_final_ping(obd))) {
2643 LASSERT(trd->trd_processing_task == current_pid());
2644 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2645 libcfs_nid2str(req->rq_peer.nid));
2646 handle_recovery_req(thread, req,
2647 trd->trd_recovery_handler);
2649 * Because the waiting client can not send ping to server,
2650 * so we need refresh the last_request_time, to avoid the
2651 * export is being evicted
2653 ptlrpc_update_export_timer(req->rq_export, 0);
2654 target_request_copy_put(req);
2657 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2658 CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2659 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2660 if (delta > OBD_RECOVERY_TIME_SOFT) {
2661 CWARN("too long recovery - read logs\n");
2662 libcfs_debug_dumplog();
2665 target_finish_recovery(lut);
2666 lu_context_fini(&env->le_ctx);
2667 trd->trd_processing_task = 0;
2668 complete_all(&trd->trd_finishing);
2669 tgt_io_thread_done(thread);
2675 OBD_FREE_PTR(thread);
2679 static int target_start_recovery_thread(struct lu_target *lut,
2680 svc_handler_t handler)
2682 struct obd_device *obd = lut->lut_obd;
2684 struct target_recovery_data *trd = &obd->obd_recovery_data;
2687 memset(trd, 0, sizeof(*trd));
2688 init_completion(&trd->trd_starting);
2689 init_completion(&trd->trd_finishing);
2690 trd->trd_recovery_handler = handler;
2692 rc = server_name2index(obd->obd_name, &index, NULL);
2696 if (!IS_ERR(kthread_run(target_recovery_thread,
2697 lut, "tgt_recover_%d", index))) {
2698 wait_for_completion(&trd->trd_starting);
2699 LASSERT(obd->obd_recovering != 0);
2707 void target_stop_recovery_thread(struct obd_device *obd)
2709 if (obd->obd_recovery_data.trd_processing_task > 0) {
2710 struct target_recovery_data *trd = &obd->obd_recovery_data;
2711 /** recovery can be done but postrecovery is not yet */
2712 spin_lock(&obd->obd_dev_lock);
2713 if (obd->obd_recovering) {
2714 CERROR("%s: Aborting recovery\n", obd->obd_name);
2715 obd->obd_abort_recovery = 1;
2716 wake_up(&obd->obd_next_transno_waitq);
2718 spin_unlock(&obd->obd_dev_lock);
2719 wait_for_completion(&trd->trd_finishing);
2722 EXPORT_SYMBOL(target_stop_recovery_thread);
2724 void target_recovery_fini(struct obd_device *obd)
2726 class_disconnect_exports(obd);
2727 target_stop_recovery_thread(obd);
2728 target_cleanup_recovery(obd);
2730 EXPORT_SYMBOL(target_recovery_fini);
2732 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2734 struct obd_device *obd = container_of(timer, struct obd_device,
2735 obd_recovery_timer);
2738 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2739 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2740 ktime_get_real_seconds() - obd->obd_recovery_start,
2741 atomic_read(&obd->obd_connected_clients));
2743 obd->obd_recovery_expired = 1;
2744 wake_up(&obd->obd_next_transno_waitq);
2745 return HRTIMER_NORESTART;
2748 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2750 struct obd_device *obd = lut->lut_obd;
2752 if (lut->lut_bottom->dd_rdonly)
2755 if (obd->obd_max_recoverable_clients == 0) {
2756 /** Update server last boot epoch */
2757 tgt_boot_epoch_update(lut);
2761 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2762 "last_transno %llu\n", obd->obd_name,
2763 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2764 LASSERT(obd->obd_stopping == 0);
2765 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2766 obd->obd_recovery_start = 0;
2767 obd->obd_recovery_end = 0;
2769 hrtimer_init(&obd->obd_recovery_timer, CLOCK_REALTIME,
2771 obd->obd_recovery_timer.function = &target_recovery_expired;
2772 target_start_recovery_thread(lut, handler);
2774 EXPORT_SYMBOL(target_recovery_init);
2776 static int target_process_req_flags(struct obd_device *obd,
2777 struct ptlrpc_request *req)
2779 struct obd_export *exp = req->rq_export;
2781 LASSERT(exp != NULL);
2782 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2783 /* client declares he's ready to replay locks */
2784 spin_lock(&exp->exp_lock);
2785 if (exp->exp_req_replay_needed) {
2786 exp->exp_req_replay_needed = 0;
2787 spin_unlock(&exp->exp_lock);
2789 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2790 atomic_dec(&obd->obd_req_replay_clients);
2792 spin_unlock(&exp->exp_lock);
2795 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2797 * client declares he's ready to complete recovery
2798 * so, we put the request on th final queue
2800 spin_lock(&exp->exp_lock);
2801 if (exp->exp_lock_replay_needed) {
2802 exp->exp_lock_replay_needed = 0;
2803 spin_unlock(&exp->exp_lock);
2805 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2806 atomic_dec(&obd->obd_lock_replay_clients);
2808 spin_unlock(&exp->exp_lock);
2814 int target_queue_recovery_request(struct ptlrpc_request *req,
2815 struct obd_device *obd)
2817 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2818 struct ptlrpc_request *reqiter;
2823 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2824 /* Processing the queue right now, don't re-add. */
2828 target_process_req_flags(obd, req);
2830 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2831 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2832 if (cfs_fail_val == 1) {
2835 wake_up(&cfs_race_waitq);
2837 set_current_state(TASK_INTERRUPTIBLE);
2838 schedule_timeout(cfs_time_seconds(1));
2843 * client declares he's ready to complete recovery
2844 * so, we put the request on th final queue
2846 target_request_copy_get(req);
2847 DEBUG_REQ(D_HA, req, "queue final req");
2848 wake_up(&obd->obd_next_transno_waitq);
2849 spin_lock(&obd->obd_recovery_task_lock);
2850 if (obd->obd_recovering) {
2851 struct ptlrpc_request *tmp;
2852 struct ptlrpc_request *duplicate = NULL;
2854 if (likely(!req->rq_export->exp_replay_done)) {
2855 req->rq_export->exp_replay_done = 1;
2856 list_add_tail(&req->rq_list,
2857 &obd->obd_final_req_queue);
2858 spin_unlock(&obd->obd_recovery_task_lock);
2863 * XXX O(n), but only happens if final ping is
2864 * timed out, probably reorganize the list as
2867 list_for_each_entry_safe(reqiter, tmp,
2868 &obd->obd_final_req_queue,
2870 if (reqiter->rq_export == req->rq_export) {
2871 list_del_init(&reqiter->rq_list);
2872 duplicate = reqiter;
2877 list_add_tail(&req->rq_list,
2878 &obd->obd_final_req_queue);
2879 req->rq_export->exp_replay_done = 1;
2880 spin_unlock(&obd->obd_recovery_task_lock);
2882 if (duplicate != NULL) {
2883 DEBUG_REQ(D_HA, duplicate,
2884 "put prev final req\n");
2885 target_request_copy_put(duplicate);
2889 spin_unlock(&obd->obd_recovery_task_lock);
2890 target_request_copy_put(req);
2891 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2894 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2895 /* client declares he's ready to replay locks */
2896 target_request_copy_get(req);
2897 DEBUG_REQ(D_HA, req, "queue lock replay req");
2898 wake_up(&obd->obd_next_transno_waitq);
2899 spin_lock(&obd->obd_recovery_task_lock);
2900 LASSERT(obd->obd_recovering);
2901 /* usually due to recovery abort */
2902 if (!req->rq_export->exp_in_recovery) {
2903 spin_unlock(&obd->obd_recovery_task_lock);
2904 target_request_copy_put(req);
2907 LASSERT(req->rq_export->exp_lock_replay_needed);
2908 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2909 spin_unlock(&obd->obd_recovery_task_lock);
2914 * CAVEAT EMPTOR: The incoming request message has been swabbed
2915 * (i.e. buflens etc are in my own byte order), but type-dependent
2916 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
2920 INIT_LIST_HEAD(&req->rq_list);
2921 DEBUG_REQ(D_HA, req, "not queueing");
2926 * If we're processing the queue, we want don't want to queue this
2929 * Also, if this request has a transno less than the one we're waiting
2930 * for, we should process it now. It could (and currently always will)
2931 * be an open request for a descriptor that was opened some time ago.
2933 * Also, a resent, replayed request that has already been
2934 * handled will pass through here and be processed immediately.
2937 "Next recovery transno: %llu, current: %llu, replaying\n",
2938 obd->obd_next_recovery_transno, transno);
2941 * If the request has been replayed by update replay, then sends this
2942 * request to the recovery thread (replay_request_or_update()), where
2943 * it will be handled
2945 spin_lock(&obd->obd_recovery_task_lock);
2946 if (transno < obd->obd_next_recovery_transno &&
2947 !is_req_replayed_by_update(req)) {
2948 /* Processing the queue right now, don't re-add. */
2949 LASSERT(list_empty(&req->rq_list));
2950 spin_unlock(&obd->obd_recovery_task_lock);
2953 spin_unlock(&obd->obd_recovery_task_lock);
2955 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2958 target_request_copy_get(req);
2959 if (!req->rq_export->exp_in_recovery) {
2960 target_request_copy_put(req);
2963 LASSERT(req->rq_export->exp_req_replay_needed);
2965 if (target_exp_enqueue_req_replay(req)) {
2966 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2967 target_request_copy_put(req);
2972 spin_lock(&obd->obd_recovery_task_lock);
2973 LASSERT(obd->obd_recovering);
2974 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2975 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2976 list_add_tail(&req->rq_list, &reqiter->rq_list);
2981 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2983 DEBUG_REQ(D_ERROR, req,
2984 "dropping replay: transno has been claimed by another client");
2985 spin_unlock(&obd->obd_recovery_task_lock);
2986 target_exp_dequeue_req_replay(req);
2987 target_request_copy_put(req);
2993 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2995 obd->obd_requests_queued_for_recovery++;
2996 spin_unlock(&obd->obd_recovery_task_lock);
2997 wake_up(&obd->obd_next_transno_waitq);
3001 void target_committed_to_req(struct ptlrpc_request *req)
3003 struct obd_export *exp = req->rq_export;
3005 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3006 lustre_msg_set_last_committed(req->rq_repmsg,
3007 exp->exp_last_committed);
3009 DEBUG_REQ(D_IOCTL, req,
3010 "not sending last_committed update (%d/%d)",
3011 exp->exp_obd->obd_no_transno,
3012 req->rq_repmsg == NULL);
3014 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3015 exp->exp_last_committed, req->rq_transno, req->rq_xid);
3018 #endif /* HAVE_SERVER_SUPPORT */
3021 * Packs current SLV and Limit into \a req.
3023 int target_pack_pool_reply(struct ptlrpc_request *req)
3025 struct obd_device *obd;
3030 * Check that we still have all structures alive as this may
3031 * be some late RPC at shutdown time.
3033 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3034 !exp_connect_lru_resize(req->rq_export))) {
3035 lustre_msg_set_slv(req->rq_repmsg, 0);
3036 lustre_msg_set_limit(req->rq_repmsg, 0);
3040 /* OBD is alive here as export is alive, which we checked above. */
3041 obd = req->rq_export->exp_obd;
3043 read_lock(&obd->obd_pool_lock);
3044 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3045 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3046 read_unlock(&obd->obd_pool_lock);
3051 static int target_send_reply_msg(struct ptlrpc_request *req,
3052 int rc, int fail_id)
3054 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3055 DEBUG_REQ(D_ERROR, req, "dropping reply");
3059 * We can have a null rq_reqmsg in the event of bad signature or
3060 * no context when unwrapping
3062 if (req->rq_reqmsg &&
3063 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3064 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3068 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3069 req->rq_status = rc;
3070 return ptlrpc_send_error(req, 1);
3072 DEBUG_REQ(D_NET, req, "sending reply");
3074 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3077 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3079 struct ptlrpc_service_part *svcpt;
3081 struct ptlrpc_reply_state *rs;
3082 struct obd_export *exp;
3086 if (req->rq_no_reply) {
3091 svcpt = req->rq_rqbd->rqbd_svcpt;
3092 rs = req->rq_reply_state;
3093 if (rs == NULL || !rs->rs_difficult) {
3095 target_send_reply_msg(req, rc, fail_id);
3100 /* must be an export if locks saved */
3101 LASSERT(req->rq_export != NULL);
3102 /* req/reply consistent */
3103 LASSERT(rs->rs_svcpt == svcpt);
3106 LASSERT(!rs->rs_scheduled);
3107 LASSERT(!rs->rs_scheduled_ever);
3108 LASSERT(!rs->rs_handled);
3109 LASSERT(!rs->rs_on_net);
3110 LASSERT(rs->rs_export == NULL);
3111 LASSERT(list_empty(&rs->rs_obd_list));
3112 LASSERT(list_empty(&rs->rs_exp_list));
3114 exp = class_export_get(req->rq_export);
3116 /* disable reply scheduling while I'm setting up */
3117 rs->rs_scheduled = 1;
3119 rs->rs_xid = req->rq_xid;
3120 rs->rs_transno = req->rq_transno;
3121 rs->rs_export = exp;
3122 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3124 spin_lock(&exp->exp_uncommitted_replies_lock);
3125 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3126 rs->rs_transno, exp->exp_last_committed);
3127 if (rs->rs_transno > exp->exp_last_committed) {
3128 /* not committed already */
3129 list_add_tail(&rs->rs_obd_list,
3130 &exp->exp_uncommitted_replies);
3132 spin_unlock(&exp->exp_uncommitted_replies_lock);
3134 spin_lock(&exp->exp_lock);
3135 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3136 spin_unlock(&exp->exp_lock);
3138 netrc = target_send_reply_msg(req, rc, fail_id);
3140 spin_lock(&svcpt->scp_rep_lock);
3142 atomic_inc(&svcpt->scp_nreps_difficult);
3146 * error sending: reply is off the net. Also we need +1
3147 * reply ref until ptlrpc_handle_rs() is done
3148 * with the reply state (if the send was successful, there
3149 * would have been +1 ref for the net, which
3150 * reply_out_callback leaves alone)
3153 ptlrpc_rs_addref(rs);
3156 spin_lock(&rs->rs_lock);
3157 if (rs->rs_transno <= exp->exp_last_committed ||
3158 (!rs->rs_on_net && !rs->rs_no_ack) ||
3159 list_empty(&rs->rs_exp_list) || /* completed already */
3160 list_empty(&rs->rs_obd_list)) {
3161 CDEBUG(D_HA, "Schedule reply immediately\n");
3162 ptlrpc_dispatch_difficult_reply(rs);
3164 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3165 rs->rs_scheduled = 0; /* allow notifier to schedule */
3167 spin_unlock(&rs->rs_lock);
3168 spin_unlock(&svcpt->scp_rep_lock);
3172 enum ldlm_mode lck_compat_array[] = {
3173 [LCK_EX] = LCK_COMPAT_EX,
3174 [LCK_PW] = LCK_COMPAT_PW,
3175 [LCK_PR] = LCK_COMPAT_PR,
3176 [LCK_CW] = LCK_COMPAT_CW,
3177 [LCK_CR] = LCK_COMPAT_CR,
3178 [LCK_NL] = LCK_COMPAT_NL,
3179 [LCK_GROUP] = LCK_COMPAT_GROUP,
3180 [LCK_COS] = LCK_COMPAT_COS,
3184 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3185 * not escape to the user level.
3187 int ldlm_error2errno(enum ldlm_error error)
3193 case ELDLM_LOCK_MATCHED:
3196 case ELDLM_LOCK_CHANGED:
3199 case ELDLM_LOCK_ABORTED:
3202 case ELDLM_LOCK_REPLACED:
3205 case ELDLM_NO_LOCK_DATA:
3208 case ELDLM_NAMESPACE_EXISTS:
3211 case ELDLM_BAD_NAMESPACE:
3215 if (((int)error) < 0) { /* cast to signed type */
3216 result = error; /* as ldlm_error can be unsigned */
3218 CERROR("Invalid DLM result code: %d\n", error);
3224 EXPORT_SYMBOL(ldlm_error2errno);
3227 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3229 enum ldlm_error ldlm_errno2error(int err_no)
3238 error = ELDLM_LOCK_CHANGED;
3241 error = ELDLM_LOCK_ABORTED;
3244 error = ELDLM_LOCK_REPLACED;
3247 error = ELDLM_NO_LOCK_DATA;
3250 error = ELDLM_NAMESPACE_EXISTS;
3253 error = ELDLM_BAD_NAMESPACE;
3261 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3262 void ldlm_dump_export_locks(struct obd_export *exp)
3264 spin_lock(&exp->exp_locks_list_guard);
3265 if (!list_empty(&exp->exp_locks_list)) {
3266 struct ldlm_lock *lock;
3268 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3270 list_for_each_entry(lock, &exp->exp_locks_list,
3272 LDLM_ERROR(lock, "lock:");
3274 spin_unlock(&exp->exp_locks_list_guard);
3278 #ifdef HAVE_SERVER_SUPPORT
3279 static int target_bulk_timeout(void *data)
3283 * We don't fail the connection here, because having the export
3284 * killed makes the (vital) call to commitrw very sad.
3289 static inline const char *bulk2type(struct ptlrpc_request *req)
3291 if (req->rq_bulk_read)
3293 if (req->rq_bulk_write)
3298 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3299 struct l_wait_info *lwi)
3301 struct ptlrpc_request *req = desc->bd_req;
3302 time64_t start = ktime_get_real_seconds();
3308 /* If there is eviction in progress, wait for it to finish. */
3309 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3310 *lwi = LWI_INTR(NULL, NULL);
3311 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3312 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
3316 /* Check if client was evicted or reconnected already. */
3317 if (exp->exp_failed ||
3318 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3321 if (req->rq_bulk_read)
3322 rc = sptlrpc_svc_wrap_bulk(req, desc);
3324 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3325 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3326 else /* old version, bulk matchbits is rq_xid */
3327 req->rq_mbits = req->rq_xid;
3330 rc = ptlrpc_start_bulk_transfer(desc);
3334 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3335 bulk2type(req), rc);
3339 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3340 ptlrpc_abort_bulk(desc);
3344 /* limit actual bulk transfer to bulk_timeout seconds */
3345 deadline = start + bulk_timeout;
3346 if (deadline > req->rq_deadline)
3347 deadline = req->rq_deadline;
3350 time64_t timeoutl = deadline - ktime_get_real_seconds();
3351 long timeout_jiffies = timeoutl <= 0 ?
3352 1 : cfs_time_seconds(timeoutl);
3353 time64_t rq_deadline;
3355 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3356 cfs_time_seconds(1),
3357 target_bulk_timeout, desc);
3358 rc = l_wait_event(desc->bd_waitq,
3359 !ptlrpc_server_bulk_active(desc) ||
3362 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3364 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3365 /* Wait again if we changed rq_deadline. */
3366 rq_deadline = READ_ONCE(req->rq_deadline);
3367 deadline = start + bulk_timeout;
3368 if (deadline > rq_deadline)
3369 deadline = rq_deadline;
3370 } while (rc == -ETIMEDOUT &&
3371 deadline > ktime_get_real_seconds());
3373 if (rc == -ETIMEDOUT) {
3374 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3375 bulk2type(req), deadline - start,
3376 ktime_get_real_seconds() - deadline);
3377 ptlrpc_abort_bulk(desc);
3378 } else if (exp->exp_failed) {
3379 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3382 ptlrpc_abort_bulk(desc);
3383 } else if (exp->exp_conn_cnt >
3384 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3385 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3387 /* We don't reply anyway. */
3389 ptlrpc_abort_bulk(desc);
3390 } else if (desc->bd_failure) {
3391 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3393 /* XXX should this be a different errno? */
3396 if (req->rq_bulk_write)
3397 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3398 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3399 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3400 bulk2type(req), desc->bd_nob_transferred,
3402 /* XXX should this be a different errno? */
3409 EXPORT_SYMBOL(target_bulk_io);
3411 #endif /* HAVE_SERVER_SUPPORT */