4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * This file deals with various client/target related logic including recovery.
36 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
40 #define DEBUG_SUBSYSTEM S_LDLM
42 #include <linux/jiffies.h>
43 #include <linux/kernel.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"
53 /* @priority: If non-zero, move the selected connection to the list head.
54 * @create: If zero, only search in existing connections.
56 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
57 int priority, int create)
59 struct ptlrpc_connection *ptlrpc_conn;
60 struct obd_import_conn *imp_conn = NULL, *item;
61 lnet_nid_t nid4refnet = LNET_NID_ANY;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 if (imp->imp_connection &&
71 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
72 /* nid4refnet is used to restrict network connections */
73 nid4refnet = imp->imp_connection->c_self;
74 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
76 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
81 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
83 GOTO(out_put, rc = -ENOMEM);
86 spin_lock(&imp->imp_lock);
87 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
88 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
90 list_del(&item->oic_item);
91 list_add(&item->oic_item,
93 item->oic_last_attempt = 0;
95 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
96 imp, imp->imp_obd->obd_name, uuid->uuid,
97 (priority ? ", moved to head" : ""));
98 spin_unlock(&imp->imp_lock);
99 GOTO(out_free, rc = 0);
102 /* No existing import connection found for \a uuid. */
104 imp_conn->oic_conn = ptlrpc_conn;
105 imp_conn->oic_uuid = *uuid;
106 imp_conn->oic_last_attempt = 0;
108 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
110 list_add_tail(&imp_conn->oic_item,
111 &imp->imp_conn_list);
112 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
113 imp, imp->imp_obd->obd_name, uuid->uuid,
114 (priority ? "head" : "tail"));
116 spin_unlock(&imp->imp_lock);
117 GOTO(out_free, rc = -ENOENT);
120 spin_unlock(&imp->imp_lock);
124 OBD_FREE(imp_conn, sizeof(*imp_conn));
126 ptlrpc_connection_put(ptlrpc_conn);
130 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
132 return import_set_conn(imp, uuid, 1, 0);
135 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
138 return import_set_conn(imp, uuid, priority, 1);
140 EXPORT_SYMBOL(client_import_add_conn);
142 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
144 struct obd_import_conn *imp_conn;
145 struct obd_export *dlmexp;
149 spin_lock(&imp->imp_lock);
150 if (list_empty(&imp->imp_conn_list)) {
151 LASSERT(!imp->imp_connection);
155 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
156 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
158 LASSERT(imp_conn->oic_conn);
160 if (imp_conn == imp->imp_conn_current) {
161 LASSERT(imp_conn->oic_conn == imp->imp_connection);
163 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
164 imp->imp_state != LUSTRE_IMP_DISCON) {
165 CERROR("can't remove current connection\n");
166 GOTO(out, rc = -EBUSY);
169 ptlrpc_connection_put(imp->imp_connection);
170 imp->imp_connection = NULL;
172 dlmexp = class_conn2export(&imp->imp_dlm_handle);
173 if (dlmexp && dlmexp->exp_connection) {
174 LASSERT(dlmexp->exp_connection ==
176 ptlrpc_connection_put(dlmexp->exp_connection);
177 dlmexp->exp_connection = NULL;
181 class_export_put(dlmexp);
184 list_del(&imp_conn->oic_item);
185 ptlrpc_connection_put(imp_conn->oic_conn);
186 OBD_FREE(imp_conn, sizeof(*imp_conn));
187 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
188 imp, imp->imp_obd->obd_name, uuid->uuid);
193 spin_unlock(&imp->imp_lock);
195 CERROR("connection %s not found\n", uuid->uuid);
198 EXPORT_SYMBOL(client_import_del_conn);
201 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
202 * to find a conn uuid of \a imp which can reach \a peer.
204 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
205 struct obd_uuid *uuid)
207 struct obd_import_conn *conn;
211 spin_lock(&imp->imp_lock);
212 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
213 /* Check if conn UUID does have this peer NID. */
214 if (class_check_uuid(&conn->oic_uuid, peer)) {
215 *uuid = conn->oic_uuid;
220 spin_unlock(&imp->imp_lock);
223 EXPORT_SYMBOL(client_import_find_conn);
225 void client_destroy_import(struct obd_import *imp)
227 /* Drop security policy instance after all RPCs have finished/aborted
228 * to let all busy contexts be released. */
229 class_import_get(imp);
230 class_destroy_import(imp);
231 sptlrpc_import_sec_put(imp);
232 class_import_put(imp);
234 EXPORT_SYMBOL(client_destroy_import);
237 * Check whether or not the OSC is on MDT.
240 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
242 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
245 static int osc_on_mdt(char *obdname)
249 ptr = strrchr(obdname, '-');
253 if (strncmp(ptr + 1, "MDT", 3) == 0)
259 /* Configure an RPC client OBD device.
264 * 3 - inactive-on-startup
265 * 4 - restrictive net
267 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
269 struct client_obd *cli = &obddev->u.cli;
270 struct obd_import *imp;
271 struct obd_uuid server_uuid;
272 int rq_portal, rp_portal, connect_op;
273 char *name = obddev->obd_type->typ_name;
274 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
275 char *cli_name = lustre_cfg_buf(lcfg, 0);
276 struct ptlrpc_connection fake_conn = { .c_self = 0,
277 .c_remote_uuid.uuid[0] = 0 };
281 /* In a more perfect world, we would hang a ptlrpc_client off of
282 * obd_type and just use the values from there. */
283 if (!strcmp(name, LUSTRE_OSC_NAME)) {
284 rq_portal = OST_REQUEST_PORTAL;
285 rp_portal = OSC_REPLY_PORTAL;
286 connect_op = OST_CONNECT;
287 cli->cl_sp_me = LUSTRE_SP_CLI;
288 cli->cl_sp_to = LUSTRE_SP_OST;
289 ns_type = LDLM_NS_TYPE_OSC;
290 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
291 !strcmp(name, LUSTRE_LWP_NAME)) {
292 rq_portal = MDS_REQUEST_PORTAL;
293 rp_portal = MDC_REPLY_PORTAL;
294 connect_op = MDS_CONNECT;
295 if (is_lwp_on_ost(cli_name))
296 cli->cl_sp_me = LUSTRE_SP_OST;
297 else if (is_lwp_on_mdt(cli_name))
298 cli->cl_sp_me = LUSTRE_SP_MDT;
300 cli->cl_sp_me = LUSTRE_SP_CLI;
301 cli->cl_sp_to = LUSTRE_SP_MDT;
302 ns_type = LDLM_NS_TYPE_MDC;
303 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
304 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
305 /* OSP_on_MDT for other MDTs */
306 connect_op = MDS_CONNECT;
307 cli->cl_sp_to = LUSTRE_SP_MDT;
308 ns_type = LDLM_NS_TYPE_MDC;
309 rq_portal = OUT_PORTAL;
311 /* OSP on MDT for OST */
312 connect_op = OST_CONNECT;
313 cli->cl_sp_to = LUSTRE_SP_OST;
314 ns_type = LDLM_NS_TYPE_OSC;
315 rq_portal = OST_REQUEST_PORTAL;
317 rp_portal = OSC_REPLY_PORTAL;
318 cli->cl_sp_me = LUSTRE_SP_MDT;
319 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
320 rq_portal = MGS_REQUEST_PORTAL;
321 rp_portal = MGC_REPLY_PORTAL;
322 connect_op = MGS_CONNECT;
323 cli->cl_sp_me = LUSTRE_SP_MGC;
324 cli->cl_sp_to = LUSTRE_SP_MGS;
325 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
326 ns_type = LDLM_NS_TYPE_MGC;
328 CERROR("unknown client OBD type \"%s\", can't setup\n",
333 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
334 CERROR("requires a TARGET UUID\n");
338 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
339 CERROR("client UUID must be less than 38 characters\n");
343 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
344 CERROR("setup requires a SERVER UUID\n");
348 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
349 CERROR("target UUID must be less than 38 characters\n");
353 init_rwsem(&cli->cl_sem);
354 mutex_init(&cli->cl_mgc_mutex);
356 init_rwsem(&cli->cl_seq_rwsem);
357 cli->cl_conn_count = 0;
358 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
359 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
360 sizeof(server_uuid)));
362 cli->cl_dirty_pages = 0;
363 cli->cl_dirty_max_pages = 0;
364 cli->cl_avail_grant = 0;
365 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
366 /* cl_dirty_max_pages may be changed at connect time in
367 * ptlrpc_connect_interpret(). */
368 client_adjust_max_dirty(cli);
369 init_waitqueue_head(&cli->cl_cache_waiters);
370 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
371 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
372 INIT_LIST_HEAD(&cli->cl_loi_write_list);
373 INIT_LIST_HEAD(&cli->cl_loi_read_list);
374 spin_lock_init(&cli->cl_loi_list_lock);
375 atomic_set(&cli->cl_pending_w_pages, 0);
376 atomic_set(&cli->cl_pending_r_pages, 0);
377 cli->cl_r_in_flight = 0;
378 cli->cl_w_in_flight = 0;
380 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
381 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
382 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
383 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
384 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
385 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
388 INIT_LIST_HEAD(&cli->cl_lru_osc);
389 atomic_set(&cli->cl_lru_shrinkers, 0);
390 atomic_long_set(&cli->cl_lru_busy, 0);
391 atomic_long_set(&cli->cl_lru_in_list, 0);
392 INIT_LIST_HEAD(&cli->cl_lru_list);
393 spin_lock_init(&cli->cl_lru_list_lock);
394 atomic_long_set(&cli->cl_unstable_count, 0);
395 INIT_LIST_HEAD(&cli->cl_shrink_list);
396 INIT_LIST_HEAD(&cli->cl_grant_chain);
398 INIT_LIST_HEAD(&cli->cl_flight_waiters);
399 cli->cl_rpcs_in_flight = 0;
401 init_waitqueue_head(&cli->cl_destroy_waitq);
402 atomic_set(&cli->cl_destroy_in_flight, 0);
404 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
405 #ifdef ENABLE_CHECKSUM
406 /* Turn on checksumming by default. */
407 cli->cl_checksum = 1;
409 * The supported checksum types will be worked out at connect time
410 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
413 cli->cl_cksum_type = cli->cl_supp_cksum_types;
415 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
417 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
418 * from OFD after connecting. */
419 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
421 cli->cl_max_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
423 /* set cl_chunkbits default value to PAGE_SHIFT,
424 * it will be updated at OSC connection time. */
425 cli->cl_chunkbits = PAGE_SHIFT;
427 if (!strcmp(name, LUSTRE_MDC_NAME)) {
428 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
429 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
430 cli->cl_max_rpcs_in_flight = 2;
431 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
432 cli->cl_max_rpcs_in_flight = 3;
433 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
434 cli->cl_max_rpcs_in_flight = 4;
436 if (osc_on_mdt(obddev->obd_name))
437 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
439 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
442 spin_lock_init(&cli->cl_mod_rpcs_lock);
443 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
444 cli->cl_max_mod_rpcs_in_flight = 0;
445 cli->cl_mod_rpcs_in_flight = 0;
446 cli->cl_close_rpcs_in_flight = 0;
447 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
448 cli->cl_mod_tag_bitmap = NULL;
450 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
452 if (connect_op == MDS_CONNECT) {
453 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
454 OBD_ALLOC(cli->cl_mod_tag_bitmap,
455 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
456 if (cli->cl_mod_tag_bitmap == NULL)
457 GOTO(err, rc = -ENOMEM);
462 CERROR("ldlm_get_ref failed: %d\n", rc);
466 ptlrpc_init_client(rq_portal, rp_portal, name,
467 &obddev->obd_ldlm_client);
469 imp = class_new_import(obddev);
471 GOTO(err_ldlm, rc = -ENOENT);
472 imp->imp_client = &obddev->obd_ldlm_client;
473 imp->imp_connect_op = connect_op;
474 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
475 LUSTRE_CFG_BUFLEN(lcfg, 1));
476 class_import_put(imp);
478 if (lustre_cfg_buf(lcfg, 4)) {
479 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
481 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
483 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
484 obddev->obd_name, lustre_cfg_string(lcfg, 4),
486 GOTO(err_import, rc);
488 fake_conn.c_self = LNET_MKNID(refnet, 0);
489 imp->imp_connection = &fake_conn;
492 rc = client_import_add_conn(imp, &server_uuid, 1);
494 CERROR("can't add initial connection\n");
495 GOTO(err_import, rc);
497 imp->imp_connection = NULL;
499 cli->cl_import = imp;
500 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
501 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
503 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
504 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
505 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
506 name, obddev->obd_name,
507 cli->cl_target_uuid.uuid);
508 spin_lock(&imp->imp_lock);
509 imp->imp_deactive = 1;
510 spin_unlock(&imp->imp_lock);
514 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
515 LDLM_NAMESPACE_CLIENT,
516 LDLM_NAMESPACE_GREEDY,
518 if (obddev->obd_namespace == NULL) {
519 CERROR("Unable to create client namespace - %s\n",
521 GOTO(err_import, rc = -ENOMEM);
527 class_destroy_import(imp);
531 if (cli->cl_mod_tag_bitmap != NULL)
532 OBD_FREE(cli->cl_mod_tag_bitmap,
533 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
534 cli->cl_mod_tag_bitmap = NULL;
538 EXPORT_SYMBOL(client_obd_setup);
540 int client_obd_cleanup(struct obd_device *obddev)
542 struct client_obd *cli = &obddev->u.cli;
545 ldlm_namespace_free_post(obddev->obd_namespace);
546 obddev->obd_namespace = NULL;
548 obd_cleanup_client_import(obddev);
549 LASSERT(obddev->u.cli.cl_import == NULL);
553 if (cli->cl_mod_tag_bitmap != NULL)
554 OBD_FREE(cli->cl_mod_tag_bitmap,
555 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
556 cli->cl_mod_tag_bitmap = NULL;
560 EXPORT_SYMBOL(client_obd_cleanup);
562 /* ->o_connect() method for client side (OSC and MDC and MGC) */
563 int client_connect_import(const struct lu_env *env,
564 struct obd_export **exp,
565 struct obd_device *obd, struct obd_uuid *cluuid,
566 struct obd_connect_data *data, void *localdata)
568 struct client_obd *cli = &obd->u.cli;
569 struct obd_import *imp = cli->cl_import;
570 struct obd_connect_data *ocd;
571 struct lustre_handle conn = { 0 };
576 down_write(&cli->cl_sem);
577 if (cli->cl_conn_count > 0)
578 GOTO(out_sem, rc = -EALREADY);
580 rc = class_connect(&conn, obd, cluuid);
584 cli->cl_conn_count++;
585 *exp = class_conn2export(&conn);
587 LASSERT(obd->obd_namespace);
589 imp->imp_dlm_handle = conn;
590 rc = ptlrpc_init_import(imp);
594 ocd = &imp->imp_connect_data;
597 imp->imp_connect_flags_orig = data->ocd_connect_flags;
598 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
601 rc = ptlrpc_connect_import(imp);
603 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
606 LASSERT(*exp != NULL && (*exp)->exp_connection);
609 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
610 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
611 data->ocd_connect_flags, ocd->ocd_connect_flags);
612 data->ocd_connect_flags = ocd->ocd_connect_flags;
613 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
616 ptlrpc_pinger_add_import(imp);
622 cli->cl_conn_count--;
623 class_disconnect(*exp);
627 up_write(&cli->cl_sem);
631 EXPORT_SYMBOL(client_connect_import);
633 int client_disconnect_export(struct obd_export *exp)
635 struct obd_device *obd = class_exp2obd(exp);
636 struct client_obd *cli;
637 struct obd_import *imp;
642 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
643 exp, exp ? exp->exp_handle.h_cookie : -1);
648 imp = cli->cl_import;
650 down_write(&cli->cl_sem);
651 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
654 if (cli->cl_conn_count == 0) {
655 CERROR("disconnecting disconnected device (%s)\n",
657 GOTO(out_disconnect, rc = -EINVAL);
660 cli->cl_conn_count--;
661 if (cli->cl_conn_count != 0)
662 GOTO(out_disconnect, rc = 0);
664 /* Mark import deactivated now, so we don't try to reconnect if any
665 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
666 * fully deactivate the import, or that would drop all requests. */
667 spin_lock(&imp->imp_lock);
668 imp->imp_deactive = 1;
669 spin_unlock(&imp->imp_lock);
671 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
672 * delete it regardless. (It's safe to delete an import that was
674 (void)ptlrpc_pinger_del_import(imp);
676 if (obd->obd_namespace != NULL) {
677 /* obd_force == local only */
678 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
679 obd->obd_force ? LCF_LOCAL : 0, NULL);
680 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
683 /* There's no need to hold sem while disconnecting an import,
684 * and it may actually cause deadlock in GSS. */
685 up_write(&cli->cl_sem);
686 rc = ptlrpc_disconnect_import(imp, 0);
687 down_write(&cli->cl_sem);
689 ptlrpc_invalidate_import(imp);
694 /* Use server style - class_disconnect should be always called for
696 err = class_disconnect(exp);
700 up_write(&cli->cl_sem);
704 EXPORT_SYMBOL(client_disconnect_export);
706 #ifdef HAVE_SERVER_SUPPORT
707 int server_disconnect_export(struct obd_export *exp)
712 /* Disconnect early so that clients can't keep using export. */
713 rc = class_disconnect(exp);
714 /* Close import to avoid sending any requests. */
715 if (exp->exp_imp_reverse)
716 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
718 ldlm_bl_thread_wakeup();
720 /* complete all outstanding replies */
721 spin_lock(&exp->exp_lock);
722 while (!list_empty(&exp->exp_outstanding_replies)) {
723 struct ptlrpc_reply_state *rs =
724 list_entry(exp->exp_outstanding_replies.next,
725 struct ptlrpc_reply_state, rs_exp_list);
726 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
728 spin_lock(&svcpt->scp_rep_lock);
730 list_del_init(&rs->rs_exp_list);
732 spin_lock(&rs->rs_lock);
733 /* clear rs_convert_lock to make sure rs is handled and put */
734 rs->rs_convert_lock = 0;
735 ptlrpc_schedule_difficult_reply(rs);
736 spin_unlock(&rs->rs_lock);
738 spin_unlock(&svcpt->scp_rep_lock);
740 spin_unlock(&exp->exp_lock);
744 EXPORT_SYMBOL(server_disconnect_export);
746 static inline int target_check_recovery_timer(struct obd_device *target)
751 if (!target->obd_recovering || target->obd_recovery_start == 0)
754 remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
755 timeout = ktime_divns(remaining, NSEC_PER_SEC);
759 /* the recovery timer should expire, but it isn't triggered,
760 * it's better to abort the recovery of this target to speed up
761 * the recovery of the whole cluster. */
762 spin_lock(&target->obd_dev_lock);
763 if (target->obd_recovering) {
764 CERROR("%s: Aborting recovery\n", target->obd_name);
765 target->obd_abort_recovery = 1;
766 wake_up(&target->obd_next_transno_waitq);
768 spin_unlock(&target->obd_dev_lock);
772 /* --------------------------------------------------------------------------
773 * from old lib/target.c
774 * -------------------------------------------------------------------------- */
776 static int target_handle_reconnect(struct lustre_handle *conn,
777 struct obd_export *exp,
778 struct obd_uuid *cluuid)
780 struct obd_device *target;
781 struct lustre_handle *hdl;
787 hdl = &exp->exp_imp_reverse->imp_remote_handle;
788 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
789 conn->cookie = exp->exp_handle.h_cookie;
790 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
791 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
795 target = exp->exp_obd;
797 /* Might be a re-connect after a partition. */
798 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
799 LCONSOLE_WARN("%s: already connected client %s (at %s) "
800 "with handle %#llx. Rejecting client "
801 "with the same UUID trying to reconnect "
802 "with handle %#llx\n", target->obd_name,
803 obd_uuid2str(&exp->exp_client_uuid),
804 obd_export_nid2str(exp),
805 hdl->cookie, conn->cookie);
806 memset(conn, 0, sizeof *conn);
807 /* target_handle_connect() treats EALREADY and
808 * -EALREADY differently. -EALREADY is an error
809 * (same UUID, different handle). */
813 if (!target->obd_recovering) {
814 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
815 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
816 obd_export_nid2str(exp));
817 GOTO(out_already, rc);
820 remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
821 timeout = ktime_divns(remaining, NSEC_PER_SEC);
823 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
825 obd_uuid2str(&exp->exp_client_uuid),
826 obd_export_nid2str(exp),
827 atomic_read(&target->obd_max_recoverable_clients),
828 timeout / 60, timeout % 60);
830 struct target_distribute_txn_data *tdtd;
835 target_check_recovery_timer(target);
837 tdtd = class_exp2tgt(exp)->lut_tdtd;
838 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
839 buf = tdtd->tdtd_show_update_logs_retrievers(
840 tdtd->tdtd_show_retrievers_cbdata,
844 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",
846 obd_uuid2str(&exp->exp_client_uuid),
847 obd_export_nid2str(exp), count,
848 buf ? buf : "unknown (not enough RAM)",
849 (abs(timeout) + target->obd_recovery_timeout) / 60,
850 (abs(timeout) + target->obd_recovery_timeout) % 60,
853 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",
854 target->obd_name, abs(timeout) / 60,
855 abs(timeout) % 60, target->obd_name);
862 conn->cookie = exp->exp_handle.h_cookie;
863 /* target_handle_connect() treats EALREADY and
864 * -EALREADY differently. EALREADY means we are
865 * doing a valid reconnect from the same client. */
870 check_and_start_recovery_timer(struct obd_device *obd,
871 struct ptlrpc_request *req, int new_client);
874 * update flags for import during reconnect process
876 static int rev_import_flags_update(struct obd_import *revimp,
877 struct ptlrpc_request *req)
880 struct obd_connect_data *data;
882 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
884 if (data->ocd_connect_flags & OBD_CONNECT_AT)
885 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
887 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
889 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
891 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
893 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
894 revimp->imp_client->cli_name,
895 libcfs_id2str(req->rq_peer), rc);
903 * Allocate a new reverse import for an export.
905 * \retval -errno in case error hit
906 * \retval 0 if reverse import correctly init
908 int rev_import_init(struct obd_export *export)
910 struct obd_device *obd = export->exp_obd;
911 struct obd_import *revimp;
913 LASSERT(export->exp_imp_reverse == NULL);
915 revimp = class_new_import(obd);
919 revimp->imp_remote_handle.cookie = 0ULL;
920 revimp->imp_client = &obd->obd_ldlm_client;
921 revimp->imp_dlm_fake = 1;
923 /* it is safe to connect import in new state as no sends possible */
924 spin_lock(&export->exp_lock);
925 export->exp_imp_reverse = revimp;
926 spin_unlock(&export->exp_lock);
927 class_import_put(revimp);
931 EXPORT_SYMBOL(rev_import_init);
934 * Handle reconnect for an export.
936 * \param exp export to handle reconnect process
937 * \param req client reconnect request
939 * \retval -rc in case securitfy flavor can't be changed
940 * \retval 0 in case none problems
942 static int rev_import_reconnect(struct obd_export *exp,
943 struct ptlrpc_request *req)
945 struct obd_import *revimp = exp->exp_imp_reverse;
946 struct lustre_handle *lh;
949 /* avoid sending a request until import flags are changed */
950 ptlrpc_import_enter_resend(revimp);
952 if (revimp->imp_connection != NULL)
953 ptlrpc_connection_put(revimp->imp_connection);
956 * client from recovery don't have a handle so we need to take from
957 * request. it may produce situation when wrong client connected
958 * to recovery as we trust a client uuid
960 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
961 revimp->imp_remote_handle = *lh;
963 /* unknown versions will be caught in
964 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
965 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
967 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
969 rc = rev_import_flags_update(revimp, req);
971 /* it is safe to still be in RECOVERY phase as we are not able
972 * to setup correct security flavor so requests are not able to
973 * be delivered correctly */
977 /* resend all rpc's via new connection */
978 return ptlrpc_import_recovery_state_machine(revimp);
981 int target_handle_connect(struct ptlrpc_request *req)
983 struct obd_device *target = NULL;
984 struct obd_export *export = NULL;
985 /* connect handle - filled from target_handle_reconnect in
987 struct lustre_handle conn;
988 struct lustre_handle *tmp;
989 struct obd_uuid cluuid;
994 bool mds_conn = false, lw_client = false, initial_conn = false;
995 bool mds_mds_conn = false;
996 bool new_mds_mds_conn = false;
997 struct obd_connect_data *data, *tmpdata;
999 lnet_nid_t *client_nid = NULL;
1002 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1004 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1006 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1007 GOTO(out, rc = -EINVAL);
1010 target = class_dev_by_str(str);
1012 deuuidify(str, NULL, &target_start, &target_len);
1013 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
1014 "from %s (no target). If you are running "
1015 "an HA pair check that the target is "
1016 "mounted on the other server.\n", str,
1017 libcfs_nid2str(req->rq_peer.nid));
1018 GOTO(out, rc = -ENODEV);
1021 spin_lock(&target->obd_dev_lock);
1023 target->obd_conn_inprogress++;
1025 if (target->obd_stopping || !target->obd_set_up) {
1026 spin_unlock(&target->obd_dev_lock);
1028 deuuidify(str, NULL, &target_start, &target_len);
1029 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1030 target_len, target_start,
1031 libcfs_nid2str(req->rq_peer.nid),
1032 (target->obd_stopping ?
1033 "stopping" : "not set up"));
1034 GOTO(out, rc = -ENODEV);
1037 if (target->obd_no_conn) {
1038 spin_unlock(&target->obd_dev_lock);
1040 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1041 "from %s\n", target->obd_name,
1042 libcfs_nid2str(req->rq_peer.nid));
1043 GOTO(out, rc = -EAGAIN);
1046 spin_unlock(&target->obd_dev_lock);
1048 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1050 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1051 GOTO(out, rc = -EINVAL);
1054 obd_str2uuid(&cluuid, str);
1056 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1058 GOTO(out, rc = -EPROTO);
1062 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1064 if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1065 GOTO(out, rc = -EPROTO);
1066 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1068 GOTO(out, rc = -EPROTO);
1070 rc = req_capsule_server_pack(&req->rq_pill);
1074 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1075 /* Don't allow clients to connect that are using old 1.8 format
1076 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1077 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1078 * FULL20 flag should be set on all connections since 2.0, but no
1079 * longer affects behaviour.
1081 * Later this check will be disabled and the flag can be retired
1082 * completely once interop with 3.0 is no longer needed.
1084 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1085 GOTO(out, rc = -EPROTO);
1087 /* Don't allow liblustre clients to connect.
1088 * - testing was disabled in v2_2_50_0-61-g6a75d65
1089 * - building was disabled in v2_5_58_0-28-g7277179
1090 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1091 * - clients were refused connect for version difference > 0.0.1.32 */
1092 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1093 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1094 GOTO(out, rc = -EPROTO);
1098 /* Note: lw_client is needed in MDS-MDS failover during update log
1099 * processing, so we needs to allow lw_client to be connected at
1100 * anytime, instead of only the initial connection
1102 lw_client = OCD_HAS_FLAG(data, LIGHTWEIGHT);
1104 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1105 initial_conn = true;
1106 mds_conn = OCD_HAS_FLAG(data, MDS);
1107 mds_mds_conn = OCD_HAS_FLAG(data, MDS_MDS);
1109 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1110 * for Imperative Recovery connection from MGC to MGS.
1112 * Via check OBD_CONNECT_FID, we can distinguish whether
1113 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1114 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1117 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1118 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1119 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1120 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1121 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1122 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1124 /* We do not support the MDT-MDT interoperations with
1125 * different version MDT because of protocol changes. */
1126 if (unlikely(major != LUSTRE_MAJOR ||
1127 minor != LUSTRE_MINOR ||
1128 abs(patch - LUSTRE_PATCH) > 3)) {
1129 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1130 "connection from different version MDT "
1131 "(%d.%d.%d.%d) %s %s\n",
1132 target->obd_name, LUSTRE_MAJOR,
1133 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1134 major, minor, patch,
1135 OBD_OCD_VERSION_FIX(data->ocd_version),
1136 libcfs_nid2str(req->rq_peer.nid), str);
1138 GOTO(out, rc = -EPROTO);
1143 /* lctl gets a backstage, all-access pass. */
1144 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1145 goto dont_check_exports;
1147 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1151 /* We've found an export in the hash. */
1153 spin_lock(&export->exp_lock);
1155 if (export->exp_connecting) { /* bug 9635, et. al. */
1156 spin_unlock(&export->exp_lock);
1157 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1158 export->exp_obd->obd_name, export,
1159 libcfs_nid2str(req->rq_peer.nid));
1160 class_export_put(export);
1163 } else if ((mds_conn || (lw_client && initial_conn) ||
1164 OCD_HAS_FLAG(data, MDS_MDS)) && export->exp_connection) {
1165 spin_unlock(&export->exp_lock);
1166 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1167 /* MDS or LWP reconnected after failover. */
1168 LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1170 lw_client ? "LWP" : "MDS",
1171 libcfs_nid2str(req->rq_peer.nid),
1172 libcfs_nid2str(export->exp_connection->c_peer.nid));
1174 /* New connection from the same NID. */
1175 LCONSOLE_WARN("%s: Received new %s connection from %s, %s former export from same NID\n",
1177 lw_client ? "LWP" : "MDS",
1178 libcfs_nid2str(req->rq_peer.nid),
1179 OCD_HAS_FLAG(data, MDS_MDS) ?
1183 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1184 OCD_HAS_FLAG(data, MDS_MDS)) {
1186 * Because exports between MDTs will always be
1187 * kept, let's do not fail such export if they
1188 * come from the same NID, otherwise it might
1189 * cause eviction between MDTs, which might
1190 * cause namespace inconsistency */
1191 spin_lock(&export->exp_lock);
1192 export->exp_connecting = 1;
1193 export->exp_conn_cnt = 0;
1194 spin_unlock(&export->exp_lock);
1195 conn.cookie = export->exp_handle.h_cookie;
1198 class_fail_export(export);
1199 class_export_put(export);
1203 } else if (export->exp_connection != NULL && initial_conn &&
1204 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1205 spin_unlock(&export->exp_lock);
1206 /* In MDS failover we have static UUID but NID can change. */
1207 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1208 "existing nid %s is already connected\n",
1209 target->obd_name, cluuid.uuid,
1210 libcfs_nid2str(req->rq_peer.nid),
1212 export->exp_connection->c_peer.nid));
1214 class_export_put(export);
1217 export->exp_connecting = 1;
1218 spin_unlock(&export->exp_lock);
1219 LASSERT(export->exp_obd == target);
1221 rc = target_handle_reconnect(&conn, export, &cluuid);
1224 /* If we found an export, we already unlocked. */
1227 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1228 } else if (req->rq_export == NULL &&
1229 atomic_read(&export->exp_rpc_count) > 0) {
1230 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1231 "still busy with %d references\n",
1232 target->obd_name, cluuid.uuid,
1233 libcfs_nid2str(req->rq_peer.nid),
1234 atomic_read(&export->exp_refcount));
1235 GOTO(out, rc = -EBUSY);
1236 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1238 if (!strstr(cluuid.uuid, "mdt"))
1239 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1240 "known client %s (at %s) because it "
1241 "is indicating it is a new client",
1242 target->obd_name, cluuid.uuid,
1243 libcfs_nid2str(req->rq_peer.nid));
1244 GOTO(out, rc = -EALREADY);
1246 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1253 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1254 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1255 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1256 export, ktime_get_seconds(),
1257 export ? export->exp_last_request_time : 0);
1259 /* If this is the first time a client connects, reset the recovery
1260 * timer. Discard lightweight connections which might be local. */
1261 if (!lw_client && rc == 0 && target->obd_recovering)
1262 check_and_start_recovery_timer(target, req, export == NULL);
1264 /* We want to handle EALREADY but *not* -EALREADY from
1265 * target_handle_reconnect(), return reconnection state in a flag. */
1266 if (rc == EALREADY) {
1267 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1273 /* Tell the client if we support replayable requests. */
1274 if (target->obd_replayable)
1275 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1276 client_nid = &req->rq_peer.nid;
1278 if (export == NULL) {
1279 /* allow lightweight connections during recovery */
1280 /* allow "new" MDT to be connected during recovery, since we
1281 * need retrieve recovery update records from it */
1282 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1283 struct hrtimer *timer = &target->obd_recovery_timer;
1292 connected = atomic_read(&target->obd_connected_clients);
1293 in_progress = atomic_read(&target->obd_lock_replay_clients);
1295 atomic_read(&target->obd_max_recoverable_clients);
1296 stale = target->obd_stale_clients;
1297 remaining = hrtimer_expires_remaining(timer);
1298 left = ktime_divns(remaining, NSEC_PER_SEC);
1299 if (ktime_to_ns(remaining) > 0) {
1300 msg = "to recover in";
1303 msg = "already passed deadline";
1306 target_check_recovery_timer(target);
1309 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",
1310 target->obd_name, cluuid.uuid,
1311 libcfs_nid2str(req->rq_peer.nid), known,
1312 connected - in_progress, in_progress,
1313 stale, msg, timeout / 60, timeout % 60);
1317 rc = obd_connect(req->rq_svc_thread->t_env,
1318 &export, target, &cluuid, data,
1320 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1321 lustre_msg_add_op_flags(req->rq_repmsg,
1322 MSG_CONNECT_RECOVERING);
1324 conn.cookie = export->exp_handle.h_cookie;
1325 rc = rev_import_init(export);
1329 new_mds_mds_conn = true;
1332 rc = obd_reconnect(req->rq_svc_thread->t_env,
1333 export, target, &cluuid, data, client_nid);
1338 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1339 data->ocd_instance = target->u.obt.obt_instance;
1341 /* Return only the parts of obd_connect_data that we understand, so the
1342 * client knows that we don't understand the rest. */
1344 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1346 tmpdata = req_capsule_server_get(&req->rq_pill,
1348 /* Don't use struct assignment here, because the client reply
1349 * buffer may be smaller/larger than the local struct
1350 * obd_connect_data. */
1351 memcpy(tmpdata, data, min(tmpsize, size));
1354 /* If the client and the server are the same node, we will already
1355 * have an export that really points to the client's DLM export,
1356 * because we have a shared handles table.
1358 * XXX this will go away when shaver stops sending the "connect" handle
1359 * in the real "remote handle" field of the request --phik 24 Apr 2003
1361 ptlrpc_request_change_export(req, export);
1363 spin_lock(&export->exp_lock);
1364 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1365 spin_unlock(&export->exp_lock);
1366 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1367 "or equal conn_cnt: %d >= %d\n",
1368 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1369 export->exp_conn_cnt,
1370 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1372 GOTO(out, rc = -EALREADY);
1374 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1375 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1376 spin_unlock(&export->exp_lock);
1378 if (export->exp_connection != NULL) {
1379 /* Check to see if connection came from another NID. */
1380 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1381 !hlist_unhashed(&export->exp_nid_hash))
1382 cfs_hash_del(export->exp_obd->obd_nid_hash,
1383 &export->exp_connection->c_peer.nid,
1384 &export->exp_nid_hash);
1386 ptlrpc_connection_put(export->exp_connection);
1389 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1392 if (hlist_unhashed(&export->exp_nid_hash))
1393 cfs_hash_add(export->exp_obd->obd_nid_hash,
1394 &export->exp_connection->c_peer.nid,
1395 &export->exp_nid_hash);
1397 lustre_msg_set_handle(req->rq_repmsg, &conn);
1399 rc = rev_import_reconnect(export, req);
1403 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1405 __u64 transno = data->ocd_transno;
1407 spin_lock(&export->exp_lock);
1408 /* possible race with class_disconnect_stale_exports,
1409 * export may be already in the eviction process */
1410 if (export->exp_failed) {
1411 spin_unlock(&export->exp_lock);
1412 GOTO(out, rc = -ENODEV);
1414 export->exp_in_recovery = 1;
1415 export->exp_req_replay_needed = 1;
1416 export->exp_lock_replay_needed = 1;
1417 spin_unlock(&export->exp_lock);
1419 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1420 MSG_CONNECT_TRANSNO);
1421 if (has_transno && transno == 0)
1422 CWARN("Connect with zero transno!\n");
1424 if (has_transno && transno > 0 &&
1425 transno < target->obd_next_recovery_transno &&
1426 transno > target->obd_last_committed) {
1427 /* Another way is to use cmpxchg() to be lock-free. */
1428 spin_lock(&target->obd_recovery_task_lock);
1429 if (transno < target->obd_next_recovery_transno)
1430 target->obd_next_recovery_transno = transno;
1431 spin_unlock(&target->obd_recovery_task_lock);
1434 atomic_inc(&target->obd_req_replay_clients);
1435 atomic_inc(&target->obd_lock_replay_clients);
1436 /* Note: MDS-MDS connection is allowed to be connected during
1437 * recovery, no matter if the exports needs to be recoveried.
1438 * Because we need retrieve updates logs from all other MDTs.
1439 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1440 * also needs to be increased to match other recovery checking
1442 if (new_mds_mds_conn)
1443 atomic_inc(&target->obd_max_recoverable_clients);
1445 if (atomic_inc_return(&target->obd_connected_clients) ==
1446 atomic_read(&target->obd_max_recoverable_clients))
1447 wake_up(&target->obd_next_transno_waitq);
1450 /* Tell the client we're in recovery, when client is involved in it. */
1451 if (target->obd_recovering && !lw_client)
1452 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1456 spin_lock(&export->exp_lock);
1457 export->exp_connecting = 0;
1458 spin_unlock(&export->exp_lock);
1460 class_export_put(export);
1462 if (target != NULL) {
1463 spin_lock(&target->obd_dev_lock);
1464 target->obd_conn_inprogress--;
1465 spin_unlock(&target->obd_dev_lock);
1466 class_decref(target, "find", current);
1468 req->rq_status = rc;
1472 int target_handle_disconnect(struct ptlrpc_request *req)
1477 rc = req_capsule_server_pack(&req->rq_pill);
1481 /* In case of target disconnect, updating sec ctx immediately is
1482 * required in order to record latest sequence number used.
1483 * Sequence is normally updated on export destroy, but this event
1484 * can occur too late, ie after a new target connect request has
1486 * Maintaining correct sequence when client connection becomes idle
1487 * ensures that GSS does not erroneously consider requests as replays.
1489 rc = sptlrpc_export_update_ctx(req->rq_export);
1493 /* Keep the rq_export around so we can send the reply. */
1494 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1499 void target_destroy_export(struct obd_export *exp)
1501 struct obd_import *imp = NULL;
1502 /* exports created from last_rcvd data, and "fake"
1503 exports created by lctl don't have an import */
1504 spin_lock(&exp->exp_lock);
1505 if (exp->exp_imp_reverse != NULL) {
1506 imp = exp->exp_imp_reverse;
1507 exp->exp_imp_reverse = NULL;
1509 spin_unlock(&exp->exp_lock);
1511 client_destroy_import(imp);
1513 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1514 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1515 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1516 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1518 EXPORT_SYMBOL(target_destroy_export);
1521 * Recovery functions
1523 static void target_request_copy_get(struct ptlrpc_request *req)
1525 class_export_rpc_inc(req->rq_export);
1526 LASSERT(list_empty(&req->rq_list));
1527 INIT_LIST_HEAD(&req->rq_replay_list);
1529 /* Increase refcount to keep request in queue. */
1530 atomic_inc(&req->rq_refcount);
1531 /* Let export know it has replays to be handled. */
1532 atomic_inc(&req->rq_export->exp_replay_count);
1535 static void target_request_copy_put(struct ptlrpc_request *req)
1537 LASSERT(list_empty(&req->rq_replay_list));
1538 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1540 atomic_dec(&req->rq_export->exp_replay_count);
1541 class_export_rpc_dec(req->rq_export);
1542 ptlrpc_server_drop_request(req);
1545 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1547 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1548 struct obd_export *exp = req->rq_export;
1549 struct ptlrpc_request *reqiter;
1550 struct ptlrpc_request *dup_req = NULL;
1555 spin_lock(&exp->exp_lock);
1556 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1558 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1566 /* We expect it with RESENT and REPLAY flags. */
1567 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1568 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1569 CERROR("invalid flags %x of resent replay\n",
1570 lustre_msg_get_flags(req->rq_reqmsg));
1572 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1575 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1577 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1578 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1582 list_add_tail(&req->rq_replay_list,
1583 &exp->exp_req_replay_queue);
1586 spin_unlock(&exp->exp_lock);
1590 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1592 LASSERT(!list_empty(&req->rq_replay_list));
1593 LASSERT(req->rq_export);
1595 spin_lock(&req->rq_export->exp_lock);
1596 list_del_init(&req->rq_replay_list);
1597 spin_unlock(&req->rq_export->exp_lock);
1600 static void target_finish_recovery(struct lu_target *lut)
1602 struct obd_device *obd = lut->lut_obd;
1605 /* Only log a recovery message when recovery has occurred. */
1606 if (obd->obd_recovery_start) {
1607 time64_t now = ktime_get_seconds();
1608 time64_t elapsed_time;
1610 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1611 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1612 "%d recovered and %d %s evicted.\n", obd->obd_name,
1613 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1614 atomic_read(&obd->obd_max_recoverable_clients),
1615 atomic_read(&obd->obd_connected_clients),
1616 obd->obd_stale_clients,
1617 obd->obd_stale_clients == 1 ? "was" : "were");
1620 ldlm_reprocess_recovery_done(obd->obd_namespace);
1621 spin_lock(&obd->obd_recovery_task_lock);
1622 if (!list_empty(&obd->obd_req_replay_queue) ||
1623 !list_empty(&obd->obd_lock_replay_queue) ||
1624 !list_empty(&obd->obd_final_req_queue)) {
1625 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1627 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1628 list_empty(&obd->obd_lock_replay_queue) ? \
1630 list_empty(&obd->obd_final_req_queue) ? \
1632 spin_unlock(&obd->obd_recovery_task_lock);
1635 spin_unlock(&obd->obd_recovery_task_lock);
1637 obd->obd_recovery_end = ktime_get_seconds();
1639 /* When recovery finished, cleanup orphans on MDS and OST. */
1640 if (obd->obd_type && OBP(obd, postrecov)) {
1641 int rc = OBP(obd, postrecov)(obd);
1644 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1650 static void abort_req_replay_queue(struct obd_device *obd)
1652 struct ptlrpc_request *req, *n;
1653 struct list_head abort_list;
1655 INIT_LIST_HEAD(&abort_list);
1656 spin_lock(&obd->obd_recovery_task_lock);
1657 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1658 spin_unlock(&obd->obd_recovery_task_lock);
1659 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1660 DEBUG_REQ(D_WARNING, req, "aborted:");
1661 req->rq_status = -ENOTCONN;
1662 if (ptlrpc_error(req)) {
1663 DEBUG_REQ(D_ERROR, req,
1664 "failed abort_req_reply; skipping");
1666 target_exp_dequeue_req_replay(req);
1667 target_request_copy_put(req);
1671 static void abort_lock_replay_queue(struct obd_device *obd)
1673 struct ptlrpc_request *req, *n;
1674 struct list_head abort_list;
1676 INIT_LIST_HEAD(&abort_list);
1677 spin_lock(&obd->obd_recovery_task_lock);
1678 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1679 spin_unlock(&obd->obd_recovery_task_lock);
1680 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1681 DEBUG_REQ(D_ERROR, req, "aborted:");
1682 req->rq_status = -ENOTCONN;
1683 if (ptlrpc_error(req)) {
1684 DEBUG_REQ(D_ERROR, req,
1685 "failed abort_lock_reply; skipping");
1687 target_request_copy_put(req);
1691 /* Called from a cleanup function if the device is being cleaned up
1692 forcefully. The exports should all have been disconnected already,
1693 the only thing left to do is
1694 - clear the recovery flags
1696 - free queued requests and replies, but don't send replies
1697 Because the obd_stopping flag is set, no new requests should be received.
1700 void target_cleanup_recovery(struct obd_device *obd)
1702 struct ptlrpc_request *req, *n;
1703 struct list_head clean_list;
1705 INIT_LIST_HEAD(&clean_list);
1706 spin_lock(&obd->obd_dev_lock);
1707 if (!obd->obd_recovering) {
1708 spin_unlock(&obd->obd_dev_lock);
1712 obd->obd_recovering = obd->obd_abort_recovery = 0;
1713 spin_unlock(&obd->obd_dev_lock);
1715 spin_lock(&obd->obd_recovery_task_lock);
1716 target_cancel_recovery_timer(obd);
1717 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1718 spin_unlock(&obd->obd_recovery_task_lock);
1720 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1721 LASSERT(req->rq_reply_state == NULL);
1722 target_exp_dequeue_req_replay(req);
1723 target_request_copy_put(req);
1726 spin_lock(&obd->obd_recovery_task_lock);
1727 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1728 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1729 spin_unlock(&obd->obd_recovery_task_lock);
1731 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1732 LASSERT(req->rq_reply_state == NULL);
1733 target_request_copy_put(req);
1738 EXPORT_SYMBOL(target_cleanup_recovery);
1740 /* obd_recovery_task_lock should be held */
1741 void target_cancel_recovery_timer(struct obd_device *obd)
1743 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1744 hrtimer_cancel(&obd->obd_recovery_timer);
1747 static void target_start_recovery_timer(struct obd_device *obd)
1751 if (obd->obd_recovery_start != 0)
1754 spin_lock(&obd->obd_dev_lock);
1755 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1756 spin_unlock(&obd->obd_dev_lock);
1760 LASSERT(obd->obd_recovery_timeout != 0);
1762 if (obd->obd_recovery_start != 0) {
1763 spin_unlock(&obd->obd_dev_lock);
1767 obd->obd_recovery_start = ktime_get_seconds();
1768 delay = ktime_set(obd->obd_recovery_start +
1769 obd->obd_recovery_timeout, 0);
1770 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1771 spin_unlock(&obd->obd_dev_lock);
1773 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1775 obd->obd_recovery_timeout / 60,
1776 obd->obd_recovery_timeout % 60,
1777 atomic_read(&obd->obd_max_recoverable_clients),
1778 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1780 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1785 * extend recovery window.
1787 * if @extend is true, extend recovery window to have @dr_timeout remaining
1788 * at least; otherwise, make sure the recovery timeout value is not less
1791 static void extend_recovery_timer(struct obd_device *obd, time_t dr_timeout,
1798 spin_lock(&obd->obd_dev_lock);
1799 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1800 spin_unlock(&obd->obd_dev_lock);
1803 LASSERT(obd->obd_recovery_start != 0);
1805 left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1806 left = ktime_divns(left_ns, NSEC_PER_SEC);
1809 timeout = obd->obd_recovery_timeout;
1810 /* dr_timeout will happen after the hrtimer has expired.
1811 * Add the excess time to the soft recovery timeout without
1812 * exceeding the hard recovery timeout.
1814 if (dr_timeout > left) {
1815 timeout += dr_timeout - left;
1816 timeout = min_t(time_t, obd->obd_recovery_time_hard,
1820 timeout = clamp_t(time_t, dr_timeout, obd->obd_recovery_timeout,
1821 obd->obd_recovery_time_hard);
1824 if (timeout == obd->obd_recovery_time_hard)
1825 CWARN("%s: extended recovery timer reached hard limit: %ld, extend: %d\n",
1826 obd->obd_name, timeout, extend);
1828 if (obd->obd_recovery_timeout < timeout) {
1831 obd->obd_recovery_timeout = timeout;
1832 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1833 now = ktime_set(ktime_get_seconds(), 0);
1834 left_ns = ktime_sub(end, now);
1835 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1836 left = ktime_divns(left_ns, NSEC_PER_SEC);
1838 spin_unlock(&obd->obd_dev_lock);
1840 CDEBUG(D_HA, "%s: recovery timer will expire in %ld seconds\n",
1841 obd->obd_name, left);
1844 /* Reset the timer with each new client connection */
1846 * This timer is actually reconnect_timer, which is for making sure
1847 * the total recovery window is at least as big as my reconnect
1848 * attempt timing. So the initial recovery time_out will be set to
1849 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1850 * from client is bigger than this, then the recovery time_out will
1851 * be extended to make sure the client could be reconnected, in the
1852 * process, the timeout from the new client should be ignored.
1855 check_and_start_recovery_timer(struct obd_device *obd,
1856 struct ptlrpc_request *req,
1859 time_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1860 struct obd_device_target *obt = &obd->u.obt;
1862 if (!new_client && service_time)
1863 /* Teach server about old server's estimates, as first guess
1864 * at how long new requests will take.
1866 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1869 target_start_recovery_timer(obd);
1871 /* Convert the service time to RPC timeout,
1872 * and reuse service_time to limit stack usage.
1874 service_time = at_est2timeout(service_time);
1876 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1877 service_time < at_extra)
1878 service_time = at_extra;
1880 /* We expect other clients to timeout within service_time, then try
1881 * to reconnect, then try the failover server. The max delay between
1882 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1883 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1885 LASSERT(obt->obt_magic == OBT_MAGIC);
1886 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1887 if (service_time > obd->obd_recovery_timeout && !new_client)
1888 extend_recovery_timer(obd, service_time, false);
1891 /** Health checking routines */
1892 static inline int exp_connect_healthy(struct obd_export *exp)
1894 return (exp->exp_in_recovery);
1897 /** if export done req_replay or has replay in queue */
1898 static inline int exp_req_replay_healthy(struct obd_export *exp)
1900 return (!exp->exp_req_replay_needed ||
1901 atomic_read(&exp->exp_replay_count) > 0);
1905 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1907 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1908 exp_req_replay_healthy(exp);
1911 /** if export done lock_replay or has replay in queue */
1912 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1914 return (!exp->exp_lock_replay_needed ||
1915 atomic_read(&exp->exp_replay_count) > 0);
1918 static inline int exp_vbr_healthy(struct obd_export *exp)
1920 return (!exp->exp_vbr_failed);
1923 static inline int exp_finished(struct obd_export *exp)
1925 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1928 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1930 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1934 static int check_for_next_transno(struct lu_target *lut)
1936 struct ptlrpc_request *req = NULL;
1937 struct obd_device *obd = lut->lut_obd;
1938 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1939 int wake_up = 0, connected, completed, queue_len;
1940 __u64 req_transno = 0;
1941 __u64 update_transno = 0;
1942 __u64 next_transno = 0;
1945 spin_lock(&obd->obd_recovery_task_lock);
1946 if (!list_empty(&obd->obd_req_replay_queue)) {
1947 req = list_entry(obd->obd_req_replay_queue.next,
1948 struct ptlrpc_request, rq_list);
1949 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1953 update_transno = distribute_txn_get_next_transno(tdtd);
1955 connected = atomic_read(&obd->obd_connected_clients);
1956 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1957 queue_len = obd->obd_requests_queued_for_recovery;
1958 next_transno = obd->obd_next_recovery_transno;
1961 "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
1962 atomic_read(&obd->obd_max_recoverable_clients),
1963 connected, completed,
1964 queue_len, req_transno, next_transno);
1966 if (obd->obd_abort_recovery) {
1967 CDEBUG(D_HA, "waking for aborted recovery\n");
1969 } else if (obd->obd_recovery_expired) {
1970 CDEBUG(D_HA, "waking for expired recovery\n");
1972 } else if (tdtd != NULL && req != NULL &&
1973 is_req_replayed_by_update(req)) {
1974 LASSERTF(req_transno < next_transno, "req_transno %llu"
1975 "next_transno%llu\n", req_transno, next_transno);
1976 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1979 } else if (req_transno == next_transno ||
1980 (update_transno != 0 && update_transno <= next_transno)) {
1981 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1983 } else if (queue_len > 0 &&
1984 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1985 /** handle gaps occured due to lost reply or VBR */
1986 LASSERTF(req_transno >= next_transno,
1987 "req_transno: %llu, next_transno: %llu\n",
1988 req_transno, next_transno);
1990 "%s: waking for gap in transno, VBR is %s (skip: "
1991 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1992 ", next_update %lld last_committed: %lld)\n",
1993 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1994 next_transno, queue_len, completed, connected,
1995 req_transno, update_transno, obd->obd_last_committed);
1996 obd->obd_next_recovery_transno = req_transno;
1998 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1999 CDEBUG(D_HA, "waking for completed recovery\n");
2001 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2002 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
2003 " by fail_lock, waking up (%lld)\n", next_transno);
2004 obd->obd_next_recovery_transno = req_transno;
2007 spin_unlock(&obd->obd_recovery_task_lock);
2011 static int check_for_next_lock(struct lu_target *lut)
2013 struct obd_device *obd = lut->lut_obd;
2016 spin_lock(&obd->obd_recovery_task_lock);
2017 if (!list_empty(&obd->obd_lock_replay_queue)) {
2018 CDEBUG(D_HA, "waking for next lock\n");
2020 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2021 CDEBUG(D_HA, "waking for completed lock replay\n");
2023 } else if (obd->obd_abort_recovery) {
2024 CDEBUG(D_HA, "waking for aborted recovery\n");
2026 } else if (obd->obd_recovery_expired) {
2027 CDEBUG(D_HA, "waking for expired recovery\n");
2030 spin_unlock(&obd->obd_recovery_task_lock);
2035 static int check_update_llog(struct lu_target *lut)
2037 struct obd_device *obd = lut->lut_obd;
2038 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2040 if (obd->obd_abort_recovery) {
2041 CDEBUG(D_HA, "waking for aborted recovery\n");
2045 if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2046 CDEBUG(D_HA, "waking for completion of reading update log\n");
2054 * wait for recovery events,
2055 * check its status with help of check_routine
2056 * evict dead clients via health_check
2058 static int target_recovery_overseer(struct lu_target *lut,
2059 int (*check_routine)(struct lu_target *),
2060 int (*health_check)(struct obd_export *))
2062 struct obd_device *obd = lut->lut_obd;
2063 struct target_distribute_txn_data *tdtd;
2067 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2068 now = ktime_get_seconds();
2069 if (now - last > 600) {
2070 LCONSOLE_INFO("%s: in recovery but waiting for "
2071 "the first client to connect\n",
2076 if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2077 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2078 __u64 next_update_transno = 0;
2080 /* Only abort the recovery if there are no update recovery
2081 * left in the queue */
2082 spin_lock(&obd->obd_recovery_task_lock);
2083 if (lut->lut_tdtd != NULL) {
2084 next_update_transno =
2085 distribute_txn_get_next_transno(lut->lut_tdtd);
2087 tdtd = lut->lut_tdtd;
2088 /* If next_update_transno == 0, it probably because
2089 * updatelog retrieve threads did not get any records
2090 * yet, let's wait those threads stopped */
2091 if (next_update_transno == 0) {
2092 spin_unlock(&obd->obd_recovery_task_lock);
2094 while (wait_event_timeout(
2095 tdtd->tdtd_recovery_threads_waitq,
2096 check_update_llog(lut),
2097 cfs_time_seconds(60)) == 0);
2099 spin_lock(&obd->obd_recovery_task_lock);
2100 next_update_transno =
2101 distribute_txn_get_next_transno(tdtd);
2105 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2106 obd->obd_next_recovery_transno = next_update_transno;
2107 spin_unlock(&obd->obd_recovery_task_lock);
2108 /* Disconnect unfinished exports from clients, and
2109 * keep connection from MDT to make sure the update
2110 * recovery will still keep trying until some one
2111 * manually abort the recovery */
2112 class_disconnect_stale_exports(obd,
2113 exp_finished_or_from_mdt);
2114 /* Abort all of replay and replay lock req from
2116 abort_req_replay_queue(obd);
2117 abort_lock_replay_queue(obd);
2118 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2119 ")in the queue.\n", obd->obd_name,
2120 next_update_transno);
2122 obd->obd_abort_recovery = 1;
2123 spin_unlock(&obd->obd_recovery_task_lock);
2124 CWARN("%s recovery is aborted by hard timeout\n",
2129 while (wait_event_timeout(obd->obd_next_transno_waitq,
2131 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2132 /* wait indefinitely for event, but don't trigger watchdog */;
2134 if (obd->obd_abort_recovery) {
2135 CWARN("recovery is aborted, evict exports in recovery\n");
2136 if (lut->lut_tdtd != NULL) {
2137 struct l_wait_info lwi = { 0 };
2139 tdtd = lut->lut_tdtd;
2140 /* Let's wait all of the update log recovery thread
2142 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2143 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2145 /* Then abort the update recovery list */
2146 dtrq_list_destroy(lut->lut_tdtd);
2149 /** evict exports which didn't finish recovery yet */
2150 class_disconnect_stale_exports(obd, exp_finished);
2152 } else if (obd->obd_recovery_expired) {
2153 obd->obd_recovery_expired = 0;
2155 /** If some clients died being recovered, evict them */
2156 LCONSOLE_WARN("%s: recovery is timed out, "
2157 "evict stale exports\n", obd->obd_name);
2158 /** evict cexports with no replay in queue, they are stalled */
2159 class_disconnect_stale_exports(obd, health_check);
2161 /** continue with VBR */
2162 spin_lock(&obd->obd_dev_lock);
2163 obd->obd_version_recov = 1;
2164 spin_unlock(&obd->obd_dev_lock);
2166 * reset timer, recovery will proceed with versions now,
2167 * timeout is set just to handle reconnection delays
2169 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2170 /** Wait for recovery events again, after evicting bad clients */
2176 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2178 struct obd_device *obd = lut->lut_obd;
2179 struct ptlrpc_request *req = NULL;
2181 CDEBUG(D_HA, "Waiting for lock\n");
2182 if (target_recovery_overseer(lut, check_for_next_lock,
2183 exp_lock_replay_healthy))
2184 abort_lock_replay_queue(obd);
2186 spin_lock(&obd->obd_recovery_task_lock);
2187 if (!list_empty(&obd->obd_lock_replay_queue)) {
2188 req = list_entry(obd->obd_lock_replay_queue.next,
2189 struct ptlrpc_request, rq_list);
2190 list_del_init(&req->rq_list);
2191 spin_unlock(&obd->obd_recovery_task_lock);
2193 spin_unlock(&obd->obd_recovery_task_lock);
2194 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2195 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2196 /** evict exports failed VBR */
2197 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2202 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2204 struct ptlrpc_request *req = NULL;
2206 spin_lock(&obd->obd_recovery_task_lock);
2207 if (!list_empty(&obd->obd_final_req_queue)) {
2208 req = list_entry(obd->obd_final_req_queue.next,
2209 struct ptlrpc_request, rq_list);
2210 list_del_init(&req->rq_list);
2211 spin_unlock(&obd->obd_recovery_task_lock);
2212 if (req->rq_export->exp_in_recovery) {
2213 spin_lock(&req->rq_export->exp_lock);
2214 req->rq_export->exp_in_recovery = 0;
2215 spin_unlock(&req->rq_export->exp_lock);
2218 spin_unlock(&obd->obd_recovery_task_lock);
2223 static void handle_recovery_req(struct ptlrpc_thread *thread,
2224 struct ptlrpc_request *req,
2225 svc_handler_t handler)
2230 * export can be evicted during recovery, no need to handle replays for
2231 * it after that, discard such request silently
2233 if (req->rq_export->exp_disconnected)
2236 req->rq_session.lc_thread = thread;
2237 req->rq_svc_thread = thread;
2238 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2240 /* thread context */
2241 lu_context_enter(&thread->t_env->le_ctx);
2243 lu_context_exit(&thread->t_env->le_ctx);
2245 req->rq_svc_thread->t_env->le_ses = NULL;
2247 /* don't reset timer for final stage */
2248 if (!exp_finished(req->rq_export)) {
2249 time_t to = obd_timeout;
2252 * Add request timeout to the recovery time so next request from
2253 * this client may come in recovery time
2256 struct ptlrpc_service_part *svcpt;
2258 svcpt = req->rq_rqbd->rqbd_svcpt;
2259 /* If the server sent early reply for this request,
2260 * the client will recalculate the timeout according to
2261 * current server estimate service time, so we will
2262 * use the maxium timeout here for waiting the client
2263 * sending the next req
2265 to = max_t(time_t, at_est2timeout(at_get(&svcpt->scp_at_estimate)),
2266 lustre_msg_get_timeout(req->rq_reqmsg));
2267 /* Add 2 net_latency, one for balance rq_deadline
2268 * (see ptl_send_rpc), one for resend the req to server,
2269 * Note: client will pack net_latency in replay req
2270 * (see ptlrpc_replay_req) */
2271 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2273 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2278 /** Checking routines for recovery */
2279 static int check_for_recovery_ready(struct lu_target *lut)
2281 struct obd_device *obd = lut->lut_obd;
2282 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2285 "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2286 clnts, obd->obd_stale_clients,
2287 atomic_read(&obd->obd_max_recoverable_clients),
2288 obd->obd_abort_recovery, obd->obd_recovery_expired);
2290 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2292 atomic_read(&obd->obd_max_recoverable_clients));
2293 if (clnts + obd->obd_stale_clients <
2294 atomic_read(&obd->obd_max_recoverable_clients))
2298 if (lut->lut_tdtd != NULL) {
2299 if (!lut->lut_tdtd->tdtd_replay_ready &&
2300 !obd->obd_abort_recovery) {
2301 /* Let's extend recovery timer, in case the recovery
2302 * timer expired, and some clients got evicted */
2303 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2305 CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2306 obd->obd_name, obd->obd_recovery_timeout);
2315 REQUEST_RECOVERY = 1,
2316 UPDATE_RECOVERY = 2,
2319 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2323 if (!list_empty(&obd->obd_req_replay_queue)) {
2324 struct ptlrpc_request *req;
2326 req = list_entry(obd->obd_req_replay_queue.next,
2327 struct ptlrpc_request, rq_list);
2328 transno = lustre_msg_get_transno(req->rq_reqmsg);
2334 static __u64 get_next_transno(struct lu_target *lut, int *type)
2336 struct obd_device *obd = lut->lut_obd;
2337 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2339 __u64 update_transno;
2342 transno = get_next_replay_req_transno(obd);
2344 *type = REQUEST_RECOVERY;
2349 update_transno = distribute_txn_get_next_transno(tdtd);
2350 if (transno == 0 || (transno >= update_transno &&
2351 update_transno != 0)) {
2352 transno = update_transno;
2354 *type = UPDATE_RECOVERY;
2361 * drop duplicate replay request
2363 * Because the operation has been replayed by update recovery, the request
2364 * with the same transno will be dropped and also notify the client to send
2365 * next replay request.
2367 * \param[in] env execution environment
2368 * \param[in] obd failover obd device
2369 * \param[in] req request to be dropped
2371 static void drop_duplicate_replay_req(struct lu_env *env,
2372 struct obd_device *obd,
2373 struct ptlrpc_request *req)
2375 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2376 " update records are found.\n",
2377 lustre_msg_get_transno(req->rq_reqmsg),
2378 libcfs_nid2str(req->rq_peer.nid));
2380 /* Right now, only for MDS reint operation update replay and
2381 * normal request replay can have the same transno */
2382 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2383 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2384 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2385 if (likely(req->rq_export))
2386 target_committed_to_req(req);
2387 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2388 target_send_reply(req, req->rq_status, 0);
2390 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2391 libcfs_nid2str(req->rq_peer.nid));
2393 target_exp_dequeue_req_replay(req);
2394 target_request_copy_put(req);
2395 obd->obd_replayed_requests++;
2398 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2400 static void replay_request_or_update(struct lu_env *env,
2401 struct lu_target *lut,
2402 struct target_recovery_data *trd,
2403 struct ptlrpc_thread *thread)
2405 struct obd_device *obd = lut->lut_obd;
2406 struct ptlrpc_request *req = NULL;
2411 CDEBUG(D_HA, "Waiting for transno %lld\n",
2412 obd->obd_next_recovery_transno);
2414 /* Replay all of request and update by transno */
2416 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2418 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2420 /** It is needed to extend recovery window above
2421 * recovery_time_soft. Extending is possible only in the
2422 * end of recovery window (see more details in
2423 * handle_recovery_req()).
2425 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2427 if (target_recovery_overseer(lut, check_for_next_transno,
2428 exp_req_replay_healthy_or_from_mdt)) {
2429 abort_req_replay_queue(obd);
2430 abort_lock_replay_queue(obd);
2434 spin_lock(&obd->obd_recovery_task_lock);
2435 transno = get_next_transno(lut, &type);
2436 if (type == REQUEST_RECOVERY && transno != 0) {
2437 /* Drop replay request from client side, if the
2438 * replay has been executed by update with the
2440 req = list_entry(obd->obd_req_replay_queue.next,
2441 struct ptlrpc_request, rq_list);
2443 list_del_init(&req->rq_list);
2444 obd->obd_requests_queued_for_recovery--;
2445 spin_unlock(&obd->obd_recovery_task_lock);
2447 /* Let's check if the request has been redone by
2449 if (is_req_replayed_by_update(req)) {
2450 struct distribute_txn_replay_req *dtrq;
2452 dtrq = distribute_txn_lookup_finish_list(tdtd,
2454 LASSERT(dtrq != NULL);
2455 spin_lock(&tdtd->tdtd_replay_list_lock);
2456 list_del_init(&dtrq->dtrq_list);
2457 spin_unlock(&tdtd->tdtd_replay_list_lock);
2460 drop_duplicate_replay_req(env, obd, req);
2465 LASSERT(trd->trd_processing_task == current_pid());
2466 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2467 lustre_msg_get_transno(req->rq_reqmsg),
2468 libcfs_nid2str(req->rq_peer.nid));
2470 thread->t_watchdog = lc_watchdog_add(WATCHDOG_TIMEOUT,
2472 handle_recovery_req(thread, req,
2473 trd->trd_recovery_handler);
2474 lc_watchdog_delete(thread->t_watchdog);
2475 thread->t_watchdog = NULL;
2478 * bz18031: increase next_recovery_transno before
2479 * target_request_copy_put() will drop exp_rpc reference
2481 spin_lock(&obd->obd_recovery_task_lock);
2482 obd->obd_next_recovery_transno++;
2483 spin_unlock(&obd->obd_recovery_task_lock);
2484 target_exp_dequeue_req_replay(req);
2485 target_request_copy_put(req);
2486 obd->obd_replayed_requests++;
2487 } else if (type == UPDATE_RECOVERY && transno != 0) {
2488 struct distribute_txn_replay_req *dtrq;
2491 spin_unlock(&obd->obd_recovery_task_lock);
2493 LASSERT(tdtd != NULL);
2494 dtrq = distribute_txn_get_next_req(tdtd);
2495 lu_context_enter(&thread->t_env->le_ctx);
2496 thread->t_watchdog = lc_watchdog_add(WATCHDOG_TIMEOUT,
2498 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2499 lc_watchdog_delete(thread->t_watchdog);
2500 thread->t_watchdog = NULL;
2501 lu_context_exit(&thread->t_env->le_ctx);
2502 extend_recovery_timer(obd, obd_timeout, true);
2504 if (rc == 0 && dtrq->dtrq_xid != 0) {
2505 CDEBUG(D_HA, "Move x%llu t%llu"
2506 " to finish list\n", dtrq->dtrq_xid,
2507 dtrq->dtrq_master_transno);
2509 /* Add it to the replay finish list */
2510 spin_lock(&tdtd->tdtd_replay_list_lock);
2511 list_add(&dtrq->dtrq_list,
2512 &tdtd->tdtd_replay_finish_list);
2513 spin_unlock(&tdtd->tdtd_replay_list_lock);
2515 spin_lock(&obd->obd_recovery_task_lock);
2516 if (transno == obd->obd_next_recovery_transno)
2517 obd->obd_next_recovery_transno++;
2519 obd->obd_next_recovery_transno)
2520 obd->obd_next_recovery_transno =
2522 spin_unlock(&obd->obd_recovery_task_lock);
2527 spin_unlock(&obd->obd_recovery_task_lock);
2529 LASSERT(list_empty(&obd->obd_req_replay_queue));
2530 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2531 /** evict exports failed VBR */
2532 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2538 static int target_recovery_thread(void *arg)
2540 struct lu_target *lut = arg;
2541 struct obd_device *obd = lut->lut_obd;
2542 struct ptlrpc_request *req;
2543 struct target_recovery_data *trd = &obd->obd_recovery_data;
2544 unsigned long delta;
2546 struct ptlrpc_thread *thread = NULL;
2550 unshare_fs_struct();
2551 OBD_ALLOC_PTR(thread);
2557 GOTO(out_thread, rc = -ENOMEM);
2558 rc = lu_env_add(env);
2562 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2564 GOTO(out_env_remove, rc);
2566 thread->t_env = env;
2567 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2568 env->le_ctx.lc_thread = thread;
2569 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2570 thread->t_watchdog = NULL;
2572 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2574 trd->trd_processing_task = current_pid();
2576 spin_lock(&obd->obd_dev_lock);
2577 obd->obd_recovering = 1;
2578 spin_unlock(&obd->obd_dev_lock);
2579 complete(&trd->trd_starting);
2581 /* first of all, we have to know the first transno to replay */
2582 if (target_recovery_overseer(lut, check_for_recovery_ready,
2583 exp_connect_healthy)) {
2584 abort_req_replay_queue(obd);
2585 abort_lock_replay_queue(obd);
2586 if (lut->lut_tdtd != NULL)
2587 dtrq_list_destroy(lut->lut_tdtd);
2590 /* next stage: replay requests or update */
2592 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2593 atomic_read(&obd->obd_req_replay_clients),
2594 obd->obd_next_recovery_transno);
2595 replay_request_or_update(env, lut, trd, thread);
2598 * The second stage: replay locks
2600 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2601 atomic_read(&obd->obd_lock_replay_clients));
2602 while ((req = target_next_replay_lock(lut))) {
2603 LASSERT(trd->trd_processing_task == current_pid());
2604 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2605 libcfs_nid2str(req->rq_peer.nid));
2606 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2607 req->rq_status = -ENODEV;
2608 target_request_copy_put(req);
2611 handle_recovery_req(thread, req,
2612 trd->trd_recovery_handler);
2613 target_request_copy_put(req);
2614 obd->obd_replayed_locks++;
2618 * The third stage: reply on final pings, at this moment all clients
2619 * must have request in final queue
2621 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2622 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2623 /** Update server last boot epoch */
2624 tgt_boot_epoch_update(lut);
2625 /* We drop recoverying flag to forward all new requests
2626 * to regular mds_handle() since now */
2627 spin_lock(&obd->obd_dev_lock);
2628 obd->obd_recovering = obd->obd_abort_recovery = 0;
2629 spin_unlock(&obd->obd_dev_lock);
2630 spin_lock(&obd->obd_recovery_task_lock);
2631 target_cancel_recovery_timer(obd);
2632 spin_unlock(&obd->obd_recovery_task_lock);
2633 while ((req = target_next_final_ping(obd))) {
2634 LASSERT(trd->trd_processing_task == current_pid());
2635 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2636 libcfs_nid2str(req->rq_peer.nid));
2637 handle_recovery_req(thread, req,
2638 trd->trd_recovery_handler);
2639 /* Because the waiting client can not send ping to server,
2640 * so we need refresh the last_request_time, to avoid the
2641 * export is being evicted */
2642 ptlrpc_update_export_timer(req->rq_export, 0);
2643 target_request_copy_put(req);
2646 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2647 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2648 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2649 if (delta > OBD_RECOVERY_TIME_SOFT) {
2650 CWARN("too long recovery - read logs\n");
2651 libcfs_debug_dumplog();
2654 target_finish_recovery(lut);
2656 lu_context_fini(&env->le_ctx);
2657 trd->trd_processing_task = 0;
2658 complete(&trd->trd_finishing);
2660 tgt_io_thread_done(thread);
2666 OBD_FREE_PTR(thread);
2670 static int target_start_recovery_thread(struct lu_target *lut,
2671 svc_handler_t handler)
2673 struct obd_device *obd = lut->lut_obd;
2675 struct target_recovery_data *trd = &obd->obd_recovery_data;
2678 memset(trd, 0, sizeof(*trd));
2679 init_completion(&trd->trd_starting);
2680 init_completion(&trd->trd_finishing);
2681 trd->trd_recovery_handler = handler;
2683 rc = server_name2index(obd->obd_name, &index, NULL);
2687 if (!IS_ERR(kthread_run(target_recovery_thread,
2688 lut, "tgt_recover_%d", index))) {
2689 wait_for_completion(&trd->trd_starting);
2690 LASSERT(obd->obd_recovering != 0);
2698 void target_stop_recovery_thread(struct obd_device *obd)
2700 if (obd->obd_recovery_data.trd_processing_task > 0) {
2701 struct target_recovery_data *trd = &obd->obd_recovery_data;
2702 /** recovery can be done but postrecovery is not yet */
2703 spin_lock(&obd->obd_dev_lock);
2704 if (obd->obd_recovering) {
2705 CERROR("%s: Aborting recovery\n", obd->obd_name);
2706 obd->obd_abort_recovery = 1;
2707 wake_up(&obd->obd_next_transno_waitq);
2709 spin_unlock(&obd->obd_dev_lock);
2710 wait_for_completion(&trd->trd_finishing);
2713 EXPORT_SYMBOL(target_stop_recovery_thread);
2715 void target_recovery_fini(struct obd_device *obd)
2717 class_disconnect_exports(obd);
2718 target_stop_recovery_thread(obd);
2719 target_cleanup_recovery(obd);
2721 EXPORT_SYMBOL(target_recovery_fini);
2723 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2725 struct obd_device *obd = container_of(timer, struct obd_device,
2726 obd_recovery_timer);
2729 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2730 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2731 ktime_get_real_seconds() - obd->obd_recovery_start,
2732 atomic_read(&obd->obd_connected_clients));
2734 obd->obd_recovery_expired = 1;
2735 wake_up(&obd->obd_next_transno_waitq);
2736 return HRTIMER_NORESTART;
2739 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2741 struct obd_device *obd = lut->lut_obd;
2743 if (lut->lut_bottom->dd_rdonly)
2746 if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2747 /** Update server last boot epoch */
2748 tgt_boot_epoch_update(lut);
2752 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2753 "last_transno %llu\n", obd->obd_name,
2754 atomic_read(&obd->obd_max_recoverable_clients),
2755 obd->obd_last_committed);
2756 LASSERT(obd->obd_stopping == 0);
2757 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2758 obd->obd_recovery_start = 0;
2759 obd->obd_recovery_end = 0;
2761 hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2763 obd->obd_recovery_timer.function = &target_recovery_expired;
2764 target_start_recovery_thread(lut, handler);
2766 EXPORT_SYMBOL(target_recovery_init);
2768 static int target_process_req_flags(struct obd_device *obd,
2769 struct ptlrpc_request *req)
2771 struct obd_export *exp = req->rq_export;
2772 LASSERT(exp != NULL);
2773 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2774 /* client declares he's ready to replay locks */
2775 spin_lock(&exp->exp_lock);
2776 if (exp->exp_req_replay_needed) {
2777 exp->exp_req_replay_needed = 0;
2778 spin_unlock(&exp->exp_lock);
2780 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2781 atomic_dec(&obd->obd_req_replay_clients);
2783 spin_unlock(&exp->exp_lock);
2786 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2787 /* client declares he's ready to complete recovery
2788 * so, we put the request on th final queue */
2789 spin_lock(&exp->exp_lock);
2790 if (exp->exp_lock_replay_needed) {
2791 exp->exp_lock_replay_needed = 0;
2792 spin_unlock(&exp->exp_lock);
2794 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2795 atomic_dec(&obd->obd_lock_replay_clients);
2797 spin_unlock(&exp->exp_lock);
2803 int target_queue_recovery_request(struct ptlrpc_request *req,
2804 struct obd_device *obd)
2806 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2807 struct ptlrpc_request *reqiter;
2811 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2812 /* Processing the queue right now, don't re-add. */
2816 target_process_req_flags(obd, req);
2818 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2819 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2820 if (cfs_fail_val == 1) {
2823 wake_up(&cfs_race_waitq);
2825 set_current_state(TASK_INTERRUPTIBLE);
2826 schedule_timeout(cfs_time_seconds(1));
2830 /* client declares he's ready to complete recovery
2831 * so, we put the request on th final queue */
2832 target_request_copy_get(req);
2833 DEBUG_REQ(D_HA, req, "queue final req");
2834 wake_up(&obd->obd_next_transno_waitq);
2835 spin_lock(&obd->obd_recovery_task_lock);
2836 if (obd->obd_recovering) {
2837 struct ptlrpc_request *tmp;
2838 struct ptlrpc_request *duplicate = NULL;
2840 if (likely(!req->rq_export->exp_replay_done)) {
2841 req->rq_export->exp_replay_done = 1;
2842 list_add_tail(&req->rq_list,
2843 &obd->obd_final_req_queue);
2844 spin_unlock(&obd->obd_recovery_task_lock);
2848 /* XXX O(n), but only happens if final ping is
2849 * timed out, probably reorganize the list as
2850 * a hash list later */
2851 list_for_each_entry_safe(reqiter, tmp,
2852 &obd->obd_final_req_queue,
2854 if (reqiter->rq_export == req->rq_export) {
2855 list_del_init(&reqiter->rq_list);
2856 duplicate = reqiter;
2861 list_add_tail(&req->rq_list,
2862 &obd->obd_final_req_queue);
2863 req->rq_export->exp_replay_done = 1;
2864 spin_unlock(&obd->obd_recovery_task_lock);
2866 if (duplicate != NULL) {
2867 DEBUG_REQ(D_HA, duplicate,
2868 "put prev final req\n");
2869 target_request_copy_put(duplicate);
2873 spin_unlock(&obd->obd_recovery_task_lock);
2874 target_request_copy_put(req);
2875 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2878 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2879 /* client declares he's ready to replay locks */
2880 target_request_copy_get(req);
2881 DEBUG_REQ(D_HA, req, "queue lock replay req");
2882 wake_up(&obd->obd_next_transno_waitq);
2883 spin_lock(&obd->obd_recovery_task_lock);
2884 LASSERT(obd->obd_recovering);
2885 /* usually due to recovery abort */
2886 if (!req->rq_export->exp_in_recovery) {
2887 spin_unlock(&obd->obd_recovery_task_lock);
2888 target_request_copy_put(req);
2891 LASSERT(req->rq_export->exp_lock_replay_needed);
2892 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2893 spin_unlock(&obd->obd_recovery_task_lock);
2897 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2898 * (i.e. buflens etc are in my own byte order), but type-dependent
2899 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2902 INIT_LIST_HEAD(&req->rq_list);
2903 DEBUG_REQ(D_HA, req, "not queueing");
2907 /* If we're processing the queue, we want don't want to queue this
2910 * Also, if this request has a transno less than the one we're waiting
2911 * for, we should process it now. It could (and currently always will)
2912 * be an open request for a descriptor that was opened some time ago.
2914 * Also, a resent, replayed request that has already been
2915 * handled will pass through here and be processed immediately.
2917 CDEBUG(D_HA, "Next recovery transno: %llu"
2918 ", current: %llu, replaying\n",
2919 obd->obd_next_recovery_transno, transno);
2921 /* If the request has been replayed by update replay, then sends this
2922 * request to the recovery thread (replay_request_or_update()), where
2923 * it will be handled */
2924 spin_lock(&obd->obd_recovery_task_lock);
2925 if (transno < obd->obd_next_recovery_transno &&
2926 !is_req_replayed_by_update(req)) {
2927 /* Processing the queue right now, don't re-add. */
2928 LASSERT(list_empty(&req->rq_list));
2929 spin_unlock(&obd->obd_recovery_task_lock);
2932 spin_unlock(&obd->obd_recovery_task_lock);
2934 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2937 target_request_copy_get(req);
2938 if (!req->rq_export->exp_in_recovery) {
2939 target_request_copy_put(req);
2942 LASSERT(req->rq_export->exp_req_replay_needed);
2944 if (target_exp_enqueue_req_replay(req)) {
2945 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2946 target_request_copy_put(req);
2951 spin_lock(&obd->obd_recovery_task_lock);
2952 LASSERT(obd->obd_recovering);
2953 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2954 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2955 list_add_tail(&req->rq_list, &reqiter->rq_list);
2960 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2962 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2963 "has been claimed by another client");
2964 spin_unlock(&obd->obd_recovery_task_lock);
2965 target_exp_dequeue_req_replay(req);
2966 target_request_copy_put(req);
2972 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2974 obd->obd_requests_queued_for_recovery++;
2975 spin_unlock(&obd->obd_recovery_task_lock);
2976 wake_up(&obd->obd_next_transno_waitq);
2980 void target_committed_to_req(struct ptlrpc_request *req)
2982 struct obd_export *exp = req->rq_export;
2984 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2985 lustre_msg_set_last_committed(req->rq_repmsg,
2986 exp->exp_last_committed);
2988 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2989 "%d)", exp->exp_obd->obd_no_transno,
2990 req->rq_repmsg == NULL);
2992 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2993 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2996 #endif /* HAVE_SERVER_SUPPORT */
2999 * Packs current SLV and Limit into \a req.
3001 int target_pack_pool_reply(struct ptlrpc_request *req)
3003 struct obd_device *obd;
3006 /* Check that we still have all structures alive as this may
3007 * be some late RPC at shutdown time. */
3008 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3009 !exp_connect_lru_resize(req->rq_export))) {
3010 lustre_msg_set_slv(req->rq_repmsg, 0);
3011 lustre_msg_set_limit(req->rq_repmsg, 0);
3015 /* OBD is alive here as export is alive, which we checked above. */
3016 obd = req->rq_export->exp_obd;
3018 read_lock(&obd->obd_pool_lock);
3019 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3020 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3021 read_unlock(&obd->obd_pool_lock);
3026 static int target_send_reply_msg(struct ptlrpc_request *req,
3027 int rc, int fail_id)
3029 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3030 DEBUG_REQ(D_ERROR, req, "dropping reply");
3033 /* We can have a null rq_reqmsg in the event of bad signature or
3034 * no context when unwrapping */
3035 if (req->rq_reqmsg &&
3036 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3037 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3041 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3042 req->rq_status = rc;
3043 return ptlrpc_send_error(req, 1);
3045 DEBUG_REQ(D_NET, req, "sending reply");
3048 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3051 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3053 struct ptlrpc_service_part *svcpt;
3055 struct ptlrpc_reply_state *rs;
3056 struct obd_export *exp;
3059 if (req->rq_no_reply) {
3064 svcpt = req->rq_rqbd->rqbd_svcpt;
3065 rs = req->rq_reply_state;
3066 if (rs == NULL || !rs->rs_difficult) {
3068 target_send_reply_msg (req, rc, fail_id);
3073 /* must be an export if locks saved */
3074 LASSERT(req->rq_export != NULL);
3075 /* req/reply consistent */
3076 LASSERT(rs->rs_svcpt == svcpt);
3079 LASSERT(!rs->rs_scheduled);
3080 LASSERT(!rs->rs_scheduled_ever);
3081 LASSERT(!rs->rs_handled);
3082 LASSERT(!rs->rs_on_net);
3083 LASSERT(rs->rs_export == NULL);
3084 LASSERT(list_empty(&rs->rs_obd_list));
3085 LASSERT(list_empty(&rs->rs_exp_list));
3087 exp = class_export_get(req->rq_export);
3089 /* disable reply scheduling while I'm setting up */
3090 rs->rs_scheduled = 1;
3092 rs->rs_xid = req->rq_xid;
3093 rs->rs_transno = req->rq_transno;
3094 rs->rs_export = exp;
3095 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3097 spin_lock(&exp->exp_uncommitted_replies_lock);
3098 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3099 rs->rs_transno, exp->exp_last_committed);
3100 if (rs->rs_transno > exp->exp_last_committed) {
3101 /* not committed already */
3102 list_add_tail(&rs->rs_obd_list,
3103 &exp->exp_uncommitted_replies);
3105 spin_unlock(&exp->exp_uncommitted_replies_lock);
3107 spin_lock(&exp->exp_lock);
3108 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3109 spin_unlock(&exp->exp_lock);
3111 netrc = target_send_reply_msg(req, rc, fail_id);
3113 spin_lock(&svcpt->scp_rep_lock);
3115 atomic_inc(&svcpt->scp_nreps_difficult);
3118 /* error sending: reply is off the net. Also we need +1
3119 * reply ref until ptlrpc_handle_rs() is done
3120 * with the reply state (if the send was successful, there
3121 * would have been +1 ref for the net, which
3122 * reply_out_callback leaves alone) */
3124 ptlrpc_rs_addref(rs);
3127 spin_lock(&rs->rs_lock);
3128 if (rs->rs_transno <= exp->exp_last_committed ||
3129 (!rs->rs_on_net && !rs->rs_no_ack) ||
3130 list_empty(&rs->rs_exp_list) || /* completed already */
3131 list_empty(&rs->rs_obd_list)) {
3132 CDEBUG(D_HA, "Schedule reply immediately\n");
3133 ptlrpc_dispatch_difficult_reply(rs);
3135 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3136 rs->rs_scheduled = 0; /* allow notifier to schedule */
3138 spin_unlock(&rs->rs_lock);
3139 spin_unlock(&svcpt->scp_rep_lock);
3143 enum ldlm_mode lck_compat_array[] = {
3144 [LCK_EX] = LCK_COMPAT_EX,
3145 [LCK_PW] = LCK_COMPAT_PW,
3146 [LCK_PR] = LCK_COMPAT_PR,
3147 [LCK_CW] = LCK_COMPAT_CW,
3148 [LCK_CR] = LCK_COMPAT_CR,
3149 [LCK_NL] = LCK_COMPAT_NL,
3150 [LCK_GROUP] = LCK_COMPAT_GROUP,
3151 [LCK_COS] = LCK_COMPAT_COS,
3155 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3156 * not escape to the user level.
3158 int ldlm_error2errno(enum ldlm_error error)
3164 case ELDLM_LOCK_MATCHED:
3167 case ELDLM_LOCK_CHANGED:
3170 case ELDLM_LOCK_ABORTED:
3173 case ELDLM_LOCK_REPLACED:
3176 case ELDLM_NO_LOCK_DATA:
3179 case ELDLM_NAMESPACE_EXISTS:
3182 case ELDLM_BAD_NAMESPACE:
3186 if (((int)error) < 0) { /* cast to signed type */
3187 result = error; /* as ldlm_error can be unsigned */
3189 CERROR("Invalid DLM result code: %d\n", error);
3195 EXPORT_SYMBOL(ldlm_error2errno);
3198 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3200 enum ldlm_error ldlm_errno2error(int err_no)
3209 error = ELDLM_LOCK_CHANGED;
3212 error = ELDLM_LOCK_ABORTED;
3215 error = ELDLM_LOCK_REPLACED;
3218 error = ELDLM_NO_LOCK_DATA;
3221 error = ELDLM_NAMESPACE_EXISTS;
3224 error = ELDLM_BAD_NAMESPACE;
3232 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3233 void ldlm_dump_export_locks(struct obd_export *exp)
3235 spin_lock(&exp->exp_locks_list_guard);
3236 if (!list_empty(&exp->exp_locks_list)) {
3237 struct ldlm_lock *lock;
3239 CERROR("dumping locks for export %p,"
3240 "ignore if the unmount doesn't hang\n", exp);
3241 list_for_each_entry(lock, &exp->exp_locks_list,
3243 LDLM_ERROR(lock, "lock:");
3245 spin_unlock(&exp->exp_locks_list_guard);
3249 #ifdef HAVE_SERVER_SUPPORT
3250 static int target_bulk_timeout(void *data)
3253 /* We don't fail the connection here, because having the export
3254 * killed makes the (vital) call to commitrw very sad.
3259 static inline const char *bulk2type(struct ptlrpc_request *req)
3261 if (req->rq_bulk_read)
3263 if (req->rq_bulk_write)
3268 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3269 struct l_wait_info *lwi)
3271 struct ptlrpc_request *req = desc->bd_req;
3272 time64_t start = ktime_get_seconds();
3278 /* If there is eviction in progress, wait for it to finish. */
3279 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3280 *lwi = LWI_INTR(NULL, NULL);
3281 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3282 !atomic_read(&exp->exp_obd->
3283 obd_evict_inprogress),
3287 /* Check if client was evicted or reconnected already. */
3288 if (exp->exp_failed ||
3289 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3292 if (req->rq_bulk_read)
3293 rc = sptlrpc_svc_wrap_bulk(req, desc);
3295 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3296 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3297 else /* old version, bulk matchbits is rq_xid */
3298 req->rq_mbits = req->rq_xid;
3301 rc = ptlrpc_start_bulk_transfer(desc);
3305 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3306 bulk2type(req), rc);
3310 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3311 ptlrpc_abort_bulk(desc);
3315 /* limit actual bulk transfer to bulk_timeout seconds */
3316 deadline = start + bulk_timeout;
3317 if (deadline > req->rq_deadline)
3318 deadline = req->rq_deadline;
3321 time64_t timeoutl = deadline - ktime_get_seconds();
3322 long timeout_jiffies = timeoutl <= 0 ?
3323 1 : cfs_time_seconds(timeoutl);
3324 time64_t rq_deadline;
3326 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3327 cfs_time_seconds(1),
3328 target_bulk_timeout, desc);
3329 rc = l_wait_event(desc->bd_waitq,
3330 !ptlrpc_server_bulk_active(desc) ||
3333 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3335 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3336 /* Wait again if we changed rq_deadline. */
3337 rq_deadline = READ_ONCE(req->rq_deadline);
3338 deadline = start + bulk_timeout;
3339 if (deadline > rq_deadline)
3340 deadline = rq_deadline;
3341 } while (rc == -ETIMEDOUT &&
3342 deadline > ktime_get_seconds());
3344 if (rc == -ETIMEDOUT) {
3345 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3346 bulk2type(req), deadline - start,
3347 ktime_get_real_seconds() - deadline);
3348 ptlrpc_abort_bulk(desc);
3349 } else if (exp->exp_failed) {
3350 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3353 ptlrpc_abort_bulk(desc);
3354 } else if (exp->exp_conn_cnt >
3355 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3356 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3358 /* We don't reply anyway. */
3360 ptlrpc_abort_bulk(desc);
3361 } else if (desc->bd_failure) {
3362 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3364 /* XXX should this be a different errno? */
3367 if (req->rq_bulk_write)
3368 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3369 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3370 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3371 bulk2type(req), desc->bd_nob_transferred,
3373 /* XXX should this be a different errno? */
3380 EXPORT_SYMBOL(target_bulk_io);
3382 #endif /* HAVE_SERVER_SUPPORT */