4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * This file deals with various client/target related logic including recovery.
36 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
40 #define DEBUG_SUBSYSTEM S_LDLM
42 #include <linux/kthread.h>
43 #include <libcfs/libcfs.h>
45 #include <obd_class.h>
46 #include <lustre_dlm.h>
47 #include <lustre_net.h>
48 #include <lustre_sec.h>
49 #include "ldlm_internal.h"
51 /* @priority: If non-zero, move the selected connection to the list head.
52 * @create: If zero, only search in existing connections.
54 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
55 int priority, int create)
57 struct ptlrpc_connection *ptlrpc_conn;
58 struct obd_import_conn *imp_conn = NULL, *item;
59 lnet_nid_t nid4refnet = LNET_NID_ANY;
63 if (!create && !priority) {
64 CDEBUG(D_HA, "Nothing to do\n");
68 if (imp->imp_connection &&
69 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
70 /* nid4refnet is used to restrict network connections */
71 nid4refnet = imp->imp_connection->c_self;
72 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
74 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
79 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
81 GOTO(out_put, rc = -ENOMEM);
84 spin_lock(&imp->imp_lock);
85 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
86 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
88 list_del(&item->oic_item);
89 list_add(&item->oic_item,
91 item->oic_last_attempt = 0;
93 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
94 imp, imp->imp_obd->obd_name, uuid->uuid,
95 (priority ? ", moved to head" : ""));
96 spin_unlock(&imp->imp_lock);
97 GOTO(out_free, rc = 0);
100 /* No existing import connection found for \a uuid. */
102 imp_conn->oic_conn = ptlrpc_conn;
103 imp_conn->oic_uuid = *uuid;
104 imp_conn->oic_last_attempt = 0;
106 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
108 list_add_tail(&imp_conn->oic_item,
109 &imp->imp_conn_list);
110 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
111 imp, imp->imp_obd->obd_name, uuid->uuid,
112 (priority ? "head" : "tail"));
114 spin_unlock(&imp->imp_lock);
115 GOTO(out_free, rc = -ENOENT);
118 spin_unlock(&imp->imp_lock);
122 OBD_FREE(imp_conn, sizeof(*imp_conn));
124 ptlrpc_connection_put(ptlrpc_conn);
128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
130 return import_set_conn(imp, uuid, 1, 0);
133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
136 return import_set_conn(imp, uuid, priority, 1);
138 EXPORT_SYMBOL(client_import_add_conn);
140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
142 struct obd_import_conn *imp_conn;
143 struct obd_export *dlmexp;
147 spin_lock(&imp->imp_lock);
148 if (list_empty(&imp->imp_conn_list)) {
149 LASSERT(!imp->imp_connection);
153 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
154 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
156 LASSERT(imp_conn->oic_conn);
158 if (imp_conn == imp->imp_conn_current) {
159 LASSERT(imp_conn->oic_conn == imp->imp_connection);
161 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
162 imp->imp_state != LUSTRE_IMP_DISCON) {
163 CERROR("can't remove current connection\n");
164 GOTO(out, rc = -EBUSY);
167 ptlrpc_connection_put(imp->imp_connection);
168 imp->imp_connection = NULL;
170 dlmexp = class_conn2export(&imp->imp_dlm_handle);
171 if (dlmexp && dlmexp->exp_connection) {
172 LASSERT(dlmexp->exp_connection ==
174 ptlrpc_connection_put(dlmexp->exp_connection);
175 dlmexp->exp_connection = NULL;
179 class_export_put(dlmexp);
182 list_del(&imp_conn->oic_item);
183 ptlrpc_connection_put(imp_conn->oic_conn);
184 OBD_FREE(imp_conn, sizeof(*imp_conn));
185 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
186 imp, imp->imp_obd->obd_name, uuid->uuid);
191 spin_unlock(&imp->imp_lock);
193 CERROR("connection %s not found\n", uuid->uuid);
196 EXPORT_SYMBOL(client_import_del_conn);
199 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
200 * to find a conn uuid of \a imp which can reach \a peer.
202 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
203 struct obd_uuid *uuid)
205 struct obd_import_conn *conn;
209 spin_lock(&imp->imp_lock);
210 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
211 /* Check if conn UUID does have this peer NID. */
212 if (class_check_uuid(&conn->oic_uuid, peer)) {
213 *uuid = conn->oic_uuid;
218 spin_unlock(&imp->imp_lock);
221 EXPORT_SYMBOL(client_import_find_conn);
223 void client_destroy_import(struct obd_import *imp)
225 /* Drop security policy instance after all RPCs have finished/aborted
226 * to let all busy contexts be released. */
227 class_import_get(imp);
228 class_destroy_import(imp);
229 sptlrpc_import_sec_put(imp);
230 class_import_put(imp);
232 EXPORT_SYMBOL(client_destroy_import);
235 * Check whether or not the OSC is on MDT.
238 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
240 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
243 static int osc_on_mdt(char *obdname)
247 ptr = strrchr(obdname, '-');
251 if (strncmp(ptr + 1, "MDT", 3) == 0)
257 /* Configure an RPC client OBD device.
262 * 3 - inactive-on-startup
263 * 4 - restrictive net
265 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
267 struct client_obd *cli = &obddev->u.cli;
268 struct obd_import *imp;
269 struct obd_uuid server_uuid;
270 int rq_portal, rp_portal, connect_op;
271 char *name = obddev->obd_type->typ_name;
272 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
273 char *cli_name = lustre_cfg_buf(lcfg, 0);
274 struct ptlrpc_connection fake_conn = { .c_self = 0,
275 .c_remote_uuid.uuid[0] = 0 };
279 /* In a more perfect world, we would hang a ptlrpc_client off of
280 * obd_type and just use the values from there. */
281 if (!strcmp(name, LUSTRE_OSC_NAME)) {
282 rq_portal = OST_REQUEST_PORTAL;
283 rp_portal = OSC_REPLY_PORTAL;
284 connect_op = OST_CONNECT;
285 cli->cl_sp_me = LUSTRE_SP_CLI;
286 cli->cl_sp_to = LUSTRE_SP_OST;
287 ns_type = LDLM_NS_TYPE_OSC;
288 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
289 !strcmp(name, LUSTRE_LWP_NAME)) {
290 rq_portal = MDS_REQUEST_PORTAL;
291 rp_portal = MDC_REPLY_PORTAL;
292 connect_op = MDS_CONNECT;
293 if (is_lwp_on_ost(cli_name))
294 cli->cl_sp_me = LUSTRE_SP_OST;
295 else if (is_lwp_on_mdt(cli_name))
296 cli->cl_sp_me = LUSTRE_SP_MDT;
298 cli->cl_sp_me = LUSTRE_SP_CLI;
299 cli->cl_sp_to = LUSTRE_SP_MDT;
300 ns_type = LDLM_NS_TYPE_MDC;
301 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
302 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
303 /* OSP_on_MDT for other MDTs */
304 connect_op = MDS_CONNECT;
305 cli->cl_sp_to = LUSTRE_SP_MDT;
306 ns_type = LDLM_NS_TYPE_MDC;
307 rq_portal = OUT_PORTAL;
309 /* OSP on MDT for OST */
310 connect_op = OST_CONNECT;
311 cli->cl_sp_to = LUSTRE_SP_OST;
312 ns_type = LDLM_NS_TYPE_OSC;
313 rq_portal = OST_REQUEST_PORTAL;
315 rp_portal = OSC_REPLY_PORTAL;
316 cli->cl_sp_me = LUSTRE_SP_MDT;
317 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
318 rq_portal = MGS_REQUEST_PORTAL;
319 rp_portal = MGC_REPLY_PORTAL;
320 connect_op = MGS_CONNECT;
321 cli->cl_sp_me = LUSTRE_SP_MGC;
322 cli->cl_sp_to = LUSTRE_SP_MGS;
323 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
324 ns_type = LDLM_NS_TYPE_MGC;
326 CERROR("unknown client OBD type \"%s\", can't setup\n",
331 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
332 CERROR("requires a TARGET UUID\n");
336 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
337 CERROR("client UUID must be less than 38 characters\n");
341 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
342 CERROR("setup requires a SERVER UUID\n");
346 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
347 CERROR("target UUID must be less than 38 characters\n");
351 init_rwsem(&cli->cl_sem);
352 mutex_init(&cli->cl_mgc_mutex);
354 init_rwsem(&cli->cl_seq_rwsem);
355 cli->cl_conn_count = 0;
356 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
357 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
358 sizeof(server_uuid)));
360 cli->cl_dirty_pages = 0;
361 cli->cl_avail_grant = 0;
362 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
363 /* cl_dirty_max_pages may be changed at connect time in
364 * ptlrpc_connect_interpret(). */
365 client_adjust_max_dirty(cli);
366 INIT_LIST_HEAD(&cli->cl_cache_waiters);
367 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
368 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
369 INIT_LIST_HEAD(&cli->cl_loi_write_list);
370 INIT_LIST_HEAD(&cli->cl_loi_read_list);
371 spin_lock_init(&cli->cl_loi_list_lock);
372 atomic_set(&cli->cl_pending_w_pages, 0);
373 atomic_set(&cli->cl_pending_r_pages, 0);
374 cli->cl_r_in_flight = 0;
375 cli->cl_w_in_flight = 0;
377 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
378 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
379 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
380 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
381 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
382 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
385 INIT_LIST_HEAD(&cli->cl_lru_osc);
386 atomic_set(&cli->cl_lru_shrinkers, 0);
387 atomic_long_set(&cli->cl_lru_busy, 0);
388 atomic_long_set(&cli->cl_lru_in_list, 0);
389 INIT_LIST_HEAD(&cli->cl_lru_list);
390 spin_lock_init(&cli->cl_lru_list_lock);
391 atomic_long_set(&cli->cl_unstable_count, 0);
392 INIT_LIST_HEAD(&cli->cl_shrink_list);
394 init_waitqueue_head(&cli->cl_destroy_waitq);
395 atomic_set(&cli->cl_destroy_in_flight, 0);
396 #ifdef ENABLE_CHECKSUM
397 /* Turn on checksumming by default. */
398 cli->cl_checksum = 1;
400 * The supported checksum types will be worked out at connect time
401 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
404 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
406 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
408 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
409 * from OFD after connecting. */
410 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
412 /* set cl_chunkbits default value to PAGE_SHIFT,
413 * it will be updated at OSC connection time. */
414 cli->cl_chunkbits = PAGE_SHIFT;
416 if (!strcmp(name, LUSTRE_MDC_NAME)) {
417 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
418 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
419 cli->cl_max_rpcs_in_flight = 2;
420 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
421 cli->cl_max_rpcs_in_flight = 3;
422 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
423 cli->cl_max_rpcs_in_flight = 4;
425 if (osc_on_mdt(obddev->obd_name))
426 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
428 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
431 spin_lock_init(&cli->cl_mod_rpcs_lock);
432 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
433 cli->cl_max_mod_rpcs_in_flight = 0;
434 cli->cl_mod_rpcs_in_flight = 0;
435 cli->cl_close_rpcs_in_flight = 0;
436 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
437 cli->cl_mod_tag_bitmap = NULL;
439 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
441 if (connect_op == MDS_CONNECT) {
442 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
443 OBD_ALLOC(cli->cl_mod_tag_bitmap,
444 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
445 if (cli->cl_mod_tag_bitmap == NULL)
446 GOTO(err, rc = -ENOMEM);
451 CERROR("ldlm_get_ref failed: %d\n", rc);
455 ptlrpc_init_client(rq_portal, rp_portal, name,
456 &obddev->obd_ldlm_client);
458 imp = class_new_import(obddev);
460 GOTO(err_ldlm, rc = -ENOENT);
461 imp->imp_client = &obddev->obd_ldlm_client;
462 imp->imp_connect_op = connect_op;
463 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
464 LUSTRE_CFG_BUFLEN(lcfg, 1));
465 class_import_put(imp);
467 if (lustre_cfg_buf(lcfg, 4)) {
468 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
470 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
472 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
473 obddev->obd_name, lustre_cfg_string(lcfg, 4),
475 GOTO(err_import, rc);
477 fake_conn.c_self = LNET_MKNID(refnet, 0);
478 imp->imp_connection = &fake_conn;
481 rc = client_import_add_conn(imp, &server_uuid, 1);
483 CERROR("can't add initial connection\n");
484 GOTO(err_import, rc);
486 imp->imp_connection = NULL;
488 cli->cl_import = imp;
489 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
490 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
492 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
493 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
494 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
495 name, obddev->obd_name,
496 cli->cl_target_uuid.uuid);
497 spin_lock(&imp->imp_lock);
498 imp->imp_deactive = 1;
499 spin_unlock(&imp->imp_lock);
503 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
504 LDLM_NAMESPACE_CLIENT,
505 LDLM_NAMESPACE_GREEDY,
507 if (obddev->obd_namespace == NULL) {
508 CERROR("Unable to create client namespace - %s\n",
510 GOTO(err_import, rc = -ENOMEM);
516 class_destroy_import(imp);
520 if (cli->cl_mod_tag_bitmap != NULL)
521 OBD_FREE(cli->cl_mod_tag_bitmap,
522 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
523 cli->cl_mod_tag_bitmap = NULL;
527 EXPORT_SYMBOL(client_obd_setup);
529 int client_obd_cleanup(struct obd_device *obddev)
531 struct client_obd *cli = &obddev->u.cli;
534 ldlm_namespace_free_post(obddev->obd_namespace);
535 obddev->obd_namespace = NULL;
537 obd_cleanup_client_import(obddev);
538 LASSERT(obddev->u.cli.cl_import == NULL);
542 if (cli->cl_mod_tag_bitmap != NULL)
543 OBD_FREE(cli->cl_mod_tag_bitmap,
544 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
545 cli->cl_mod_tag_bitmap = NULL;
549 EXPORT_SYMBOL(client_obd_cleanup);
551 /* ->o_connect() method for client side (OSC and MDC and MGC) */
552 int client_connect_import(const struct lu_env *env,
553 struct obd_export **exp,
554 struct obd_device *obd, struct obd_uuid *cluuid,
555 struct obd_connect_data *data, void *localdata)
557 struct client_obd *cli = &obd->u.cli;
558 struct obd_import *imp = cli->cl_import;
559 struct obd_connect_data *ocd;
560 struct lustre_handle conn = { 0 };
565 down_write(&cli->cl_sem);
566 if (cli->cl_conn_count > 0)
567 GOTO(out_sem, rc = -EALREADY);
569 rc = class_connect(&conn, obd, cluuid);
573 cli->cl_conn_count++;
574 *exp = class_conn2export(&conn);
576 LASSERT(obd->obd_namespace);
578 imp->imp_dlm_handle = conn;
579 rc = ptlrpc_init_import(imp);
583 ocd = &imp->imp_connect_data;
586 imp->imp_connect_flags_orig = data->ocd_connect_flags;
587 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
590 rc = ptlrpc_connect_import(imp);
592 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
595 LASSERT(*exp != NULL && (*exp)->exp_connection);
598 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
599 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
600 data->ocd_connect_flags, ocd->ocd_connect_flags);
601 data->ocd_connect_flags = ocd->ocd_connect_flags;
604 ptlrpc_pinger_add_import(imp);
610 cli->cl_conn_count--;
611 class_disconnect(*exp);
615 up_write(&cli->cl_sem);
619 EXPORT_SYMBOL(client_connect_import);
621 int client_disconnect_export(struct obd_export *exp)
623 struct obd_device *obd = class_exp2obd(exp);
624 struct client_obd *cli;
625 struct obd_import *imp;
630 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
631 exp, exp ? exp->exp_handle.h_cookie : -1);
636 imp = cli->cl_import;
638 down_write(&cli->cl_sem);
639 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
642 if (cli->cl_conn_count == 0) {
643 CERROR("disconnecting disconnected device (%s)\n",
645 GOTO(out_disconnect, rc = -EINVAL);
648 cli->cl_conn_count--;
649 if (cli->cl_conn_count != 0)
650 GOTO(out_disconnect, rc = 0);
652 /* Mark import deactivated now, so we don't try to reconnect if any
653 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
654 * fully deactivate the import, or that would drop all requests. */
655 spin_lock(&imp->imp_lock);
656 imp->imp_deactive = 1;
657 spin_unlock(&imp->imp_lock);
659 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
660 * delete it regardless. (It's safe to delete an import that was
662 (void)ptlrpc_pinger_del_import(imp);
664 if (obd->obd_namespace != NULL) {
665 /* obd_force == local only */
666 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
667 obd->obd_force ? LCF_LOCAL : 0, NULL);
668 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
671 /* There's no need to hold sem while disconnecting an import,
672 * and it may actually cause deadlock in GSS. */
673 up_write(&cli->cl_sem);
674 rc = ptlrpc_disconnect_import(imp, 0);
675 down_write(&cli->cl_sem);
677 ptlrpc_invalidate_import(imp);
682 /* Use server style - class_disconnect should be always called for
684 err = class_disconnect(exp);
688 up_write(&cli->cl_sem);
692 EXPORT_SYMBOL(client_disconnect_export);
694 #ifdef HAVE_SERVER_SUPPORT
695 int server_disconnect_export(struct obd_export *exp)
700 /* Disconnect early so that clients can't keep using export. */
701 rc = class_disconnect(exp);
702 /* Close import to avoid sending any requests. */
703 if (exp->exp_imp_reverse)
704 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
706 ldlm_bl_thread_wakeup();
708 /* complete all outstanding replies */
709 spin_lock(&exp->exp_lock);
710 while (!list_empty(&exp->exp_outstanding_replies)) {
711 struct ptlrpc_reply_state *rs =
712 list_entry(exp->exp_outstanding_replies.next,
713 struct ptlrpc_reply_state, rs_exp_list);
714 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
716 spin_lock(&svcpt->scp_rep_lock);
718 list_del_init(&rs->rs_exp_list);
719 spin_lock(&rs->rs_lock);
720 ptlrpc_schedule_difficult_reply(rs);
721 spin_unlock(&rs->rs_lock);
723 spin_unlock(&svcpt->scp_rep_lock);
725 spin_unlock(&exp->exp_lock);
729 EXPORT_SYMBOL(server_disconnect_export);
731 /* --------------------------------------------------------------------------
732 * from old lib/target.c
733 * -------------------------------------------------------------------------- */
735 static int target_handle_reconnect(struct lustre_handle *conn,
736 struct obd_export *exp,
737 struct obd_uuid *cluuid)
739 struct obd_device *target;
740 struct lustre_handle *hdl;
747 hdl = &exp->exp_imp_reverse->imp_remote_handle;
748 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
749 conn->cookie = exp->exp_handle.h_cookie;
750 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
751 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
755 target = exp->exp_obd;
757 /* Might be a re-connect after a partition. */
758 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
759 LCONSOLE_WARN("%s: already connected client %s (at %s) "
760 "with handle %#llx. Rejecting client "
761 "with the same UUID trying to reconnect "
762 "with handle %#llx\n", target->obd_name,
763 obd_uuid2str(&exp->exp_client_uuid),
764 obd_export_nid2str(exp),
765 hdl->cookie, conn->cookie);
766 memset(conn, 0, sizeof *conn);
767 /* target_handle_connect() treats EALREADY and
768 * -EALREADY differently. -EALREADY is an error
769 * (same UUID, different handle). */
773 if (!target->obd_recovering) {
774 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
775 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
776 obd_export_nid2str(exp));
777 GOTO(out_already, rc);
780 now = cfs_time_current();
781 deadline = target->obd_recovery_timer.expires;
782 if (cfs_time_before(now, deadline)) {
783 struct target_distribute_txn_data *tdtd =
784 class_exp2tgt(exp)->lut_tdtd;
789 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
790 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
791 buf = tdtd->tdtd_show_update_logs_retrievers(
792 tdtd->tdtd_show_retrievers_cbdata,
796 LCONSOLE_WARN("%s: Recovery already passed deadline "
797 "%d:%.02d. It is due to DNE recovery "
798 "failed/stuck on the %d MDT(s):%s. "
799 "Please wait until all MDTs recovered "
800 "or abort the recovery by force.\n",
801 target->obd_name, timeout / 60,
803 buf ? buf : "unknown (not enough RAM)");
805 LCONSOLE_WARN("%s: Recovery already passed deadline "
806 "%d:%.02d. If you do not want to wait "
807 "more, please abort the recovery by "
808 "force.\n", target->obd_name,
809 timeout / 60, timeout % 60);
814 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
815 LCONSOLE_WARN("%s: Recovery already passed deadline"
816 " %d:%.02d, It is most likely due to DNE"
817 " recovery is failed or stuck, please wait a"
818 " few more minutes or abort the recovery.\n",
819 target->obd_name, timeout / 60, timeout % 60);
823 conn->cookie = exp->exp_handle.h_cookie;
824 /* target_handle_connect() treats EALREADY and
825 * -EALREADY differently. EALREADY means we are
826 * doing a valid reconnect from the same client. */
831 check_and_start_recovery_timer(struct obd_device *obd,
832 struct ptlrpc_request *req, int new_client);
835 * update flags for import during reconnect process
837 static int rev_import_flags_update(struct obd_import *revimp,
838 struct ptlrpc_request *req)
841 struct obd_connect_data *data;
843 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
845 if (data->ocd_connect_flags & OBD_CONNECT_AT)
846 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
848 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
850 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
852 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
854 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
855 revimp->imp_client->cli_name,
856 libcfs_id2str(req->rq_peer), rc);
864 * Allocate a new reverse import for an export.
866 * \retval -errno in case error hit
867 * \retval 0 if reverse import correctly init
869 int rev_import_init(struct obd_export *export)
871 struct obd_device *obd = export->exp_obd;
872 struct obd_import *revimp;
874 LASSERT(export->exp_imp_reverse == NULL);
876 revimp = class_new_import(obd);
880 revimp->imp_remote_handle.cookie = 0ULL;
881 revimp->imp_client = &obd->obd_ldlm_client;
882 revimp->imp_dlm_fake = 1;
884 /* it is safe to connect import in new state as no sends possible */
885 spin_lock(&export->exp_lock);
886 export->exp_imp_reverse = revimp;
887 spin_unlock(&export->exp_lock);
888 class_import_put(revimp);
892 EXPORT_SYMBOL(rev_import_init);
895 * Handle reconnect for an export.
897 * \param exp export to handle reconnect process
898 * \param req client reconnect request
900 * \retval -rc in case securitfy flavor can't be changed
901 * \retval 0 in case none problems
903 static int rev_import_reconnect(struct obd_export *exp,
904 struct ptlrpc_request *req)
906 struct obd_import *revimp = exp->exp_imp_reverse;
907 struct lustre_handle *lh;
910 /* avoid sending a request until import flags are changed */
911 ptlrpc_import_enter_resend(revimp);
913 if (revimp->imp_connection != NULL)
914 ptlrpc_connection_put(revimp->imp_connection);
917 * client from recovery don't have a handle so we need to take from
918 * request. it may produce situation when wrong client connected
919 * to recovery as we trust a client uuid
921 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
922 revimp->imp_remote_handle = *lh;
924 /* unknown versions will be caught in
925 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
926 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
928 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
930 rc = rev_import_flags_update(revimp, req);
932 /* it is safe to still be in RECOVERY phase as we are not able
933 * to setup correct security flavor so requests are not able to
934 * be delivered correctly */
938 /* resend all rpc's via new connection */
939 return ptlrpc_import_recovery_state_machine(revimp);
942 int target_handle_connect(struct ptlrpc_request *req)
944 struct obd_device *target = NULL;
945 struct obd_export *export = NULL;
946 /* connect handle - filled from target_handle_reconnect in
948 struct lustre_handle conn;
949 struct lustre_handle *tmp;
950 struct obd_uuid tgtuuid;
951 struct obd_uuid cluuid;
956 bool mds_conn = false, lw_client = false, initial_conn = false;
957 bool mds_mds_conn = false;
958 bool new_mds_mds_conn = false;
959 bool target_referenced = false;
960 struct obd_connect_data *data, *tmpdata;
962 lnet_nid_t *client_nid = NULL;
965 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
967 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
969 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
970 GOTO(out, rc = -EINVAL);
973 obd_str2uuid(&tgtuuid, str);
974 target = class_uuid2obd(&tgtuuid);
976 target = class_name2obd(str);
979 deuuidify(str, NULL, &target_start, &target_len);
980 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
981 "from %s (no target). If you are running "
982 "an HA pair check that the target is "
983 "mounted on the other server.\n", str,
984 libcfs_nid2str(req->rq_peer.nid));
985 GOTO(out, rc = -ENODEV);
988 spin_lock(&target->obd_dev_lock);
989 if (target->obd_stopping || !target->obd_set_up) {
990 spin_unlock(&target->obd_dev_lock);
992 deuuidify(str, NULL, &target_start, &target_len);
993 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
994 target_len, target_start,
995 libcfs_nid2str(req->rq_peer.nid),
996 (target->obd_stopping ?
997 "stopping" : "not set up"));
998 GOTO(out, rc = -ENODEV);
1001 if (target->obd_no_conn) {
1002 spin_unlock(&target->obd_dev_lock);
1004 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1005 "from %s\n", target->obd_name,
1006 libcfs_nid2str(req->rq_peer.nid));
1007 GOTO(out, rc = -EAGAIN);
1010 /* Make sure the target isn't cleaned up while we're here. Yes,
1011 * there's still a race between the above check and our incref here.
1012 * Really, class_uuid2obd should take the ref. */
1013 class_incref(target, __func__, current);
1014 target_referenced = true;
1016 target->obd_conn_inprogress++;
1017 spin_unlock(&target->obd_dev_lock);
1019 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1021 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1022 GOTO(out, rc = -EINVAL);
1025 obd_str2uuid(&cluuid, str);
1027 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1029 GOTO(out, rc = -EPROTO);
1033 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1035 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1037 GOTO(out, rc = -EPROTO);
1039 rc = req_capsule_server_pack(&req->rq_pill);
1043 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1044 /* Don't allow clients to connect that are using old 1.8 format
1045 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1046 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1047 * FULL20 flag should be set on all connections since 2.0, but no
1048 * longer affects behaviour.
1050 * Later this check will be disabled and the flag can be retired
1051 * completely once interop with 3.0 is no longer needed.
1053 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1054 GOTO(out, rc = -EPROTO);
1057 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1058 if (data->ocd_version < LUSTRE_VERSION_CODE -
1059 LUSTRE_VERSION_ALLOWED_OFFSET ||
1060 data->ocd_version > LUSTRE_VERSION_CODE +
1061 LUSTRE_VERSION_ALLOWED_OFFSET) {
1062 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1063 "libclient connection attempt",
1064 data->ocd_version < LUSTRE_VERSION_CODE ?
1066 OBD_OCD_VERSION_MAJOR(data->ocd_version),
1067 OBD_OCD_VERSION_MINOR(data->ocd_version),
1068 OBD_OCD_VERSION_PATCH(data->ocd_version),
1069 OBD_OCD_VERSION_FIX(data->ocd_version));
1070 data = req_capsule_server_sized_get(&req->rq_pill,
1072 offsetof(typeof(*data), ocd_version) +
1073 sizeof(data->ocd_version));
1075 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1076 data->ocd_version = LUSTRE_VERSION_CODE;
1078 GOTO(out, rc = -EPROTO);
1082 /* Note: lw_client is needed in MDS-MDS failover during update log
1083 * processing, so we needs to allow lw_client to be connected at
1084 * anytime, instead of only the initial connection */
1085 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1087 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1088 initial_conn = true;
1089 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1090 mds_mds_conn = (data->ocd_connect_flags &
1091 OBD_CONNECT_MDS_MDS) != 0;
1093 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1094 * for Imperative Recovery connection from MGC to MGS.
1096 * Via check OBD_CONNECT_FID, we can distinguish whether
1097 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1100 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1101 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1102 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1103 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1104 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1105 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1107 /* We do not support the MDT-MDT interoperations with
1108 * different version MDT because of protocol changes. */
1109 if (unlikely(major != LUSTRE_MAJOR ||
1110 minor != LUSTRE_MINOR ||
1111 abs(patch - LUSTRE_PATCH) > 3)) {
1112 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1113 "connection from different version MDT "
1114 "(%d.%d.%d.%d) %s %s\n",
1115 target->obd_name, LUSTRE_MAJOR,
1116 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1117 major, minor, patch,
1118 OBD_OCD_VERSION_FIX(data->ocd_version),
1119 libcfs_nid2str(req->rq_peer.nid), str);
1121 GOTO(out, rc = -EPROTO);
1126 /* lctl gets a backstage, all-access pass. */
1127 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1128 goto dont_check_exports;
1130 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1134 /* We've found an export in the hash. */
1136 spin_lock(&export->exp_lock);
1138 if (export->exp_connecting) { /* bug 9635, et. al. */
1139 spin_unlock(&export->exp_lock);
1140 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1141 export->exp_obd->obd_name, export,
1142 libcfs_nid2str(req->rq_peer.nid));
1143 class_export_put(export);
1146 } else if ((mds_conn || (lw_client && initial_conn) ||
1147 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1148 export->exp_connection != NULL) {
1149 spin_unlock(&export->exp_lock);
1150 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1151 /* MDS or LWP reconnected after failover. */
1152 LCONSOLE_WARN("%s: Received %s connection from "
1153 "%s, removing former export from %s\n",
1154 target->obd_name, mds_conn ? "MDS" : "LWP",
1155 libcfs_nid2str(req->rq_peer.nid),
1156 libcfs_nid2str(export->exp_connection->c_peer.nid));
1158 /* New MDS connection from the same NID. */
1159 LCONSOLE_WARN("%s: Received new %s connection from "
1160 "%s, removing former export from same NID\n",
1161 target->obd_name, mds_conn ? "MDS" : "LWP",
1162 libcfs_nid2str(req->rq_peer.nid));
1165 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1166 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1167 /* Because exports between MDTs will always be
1168 * kept, let's do not fail such export if they
1169 * come from the same NID, otherwise it might
1170 * cause eviction between MDTs, which might
1171 * cause namespace inconsistency */
1172 spin_lock(&export->exp_lock);
1173 export->exp_connecting = 1;
1174 spin_unlock(&export->exp_lock);
1175 conn.cookie = export->exp_handle.h_cookie;
1178 class_fail_export(export);
1179 class_export_put(export);
1183 } else if (export->exp_connection != NULL && initial_conn &&
1184 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1185 spin_unlock(&export->exp_lock);
1186 /* In MDS failover we have static UUID but NID can change. */
1187 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1188 "existing nid %s is already connected\n",
1189 target->obd_name, cluuid.uuid,
1190 libcfs_nid2str(req->rq_peer.nid),
1192 export->exp_connection->c_peer.nid));
1194 class_export_put(export);
1197 export->exp_connecting = 1;
1198 spin_unlock(&export->exp_lock);
1199 LASSERT(export->exp_obd == target);
1201 rc = target_handle_reconnect(&conn, export, &cluuid);
1204 /* If we found an export, we already unlocked. */
1207 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1208 } else if (req->rq_export == NULL &&
1209 atomic_read(&export->exp_rpc_count) > 0) {
1210 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1211 "still busy with %d references\n",
1212 target->obd_name, cluuid.uuid,
1213 libcfs_nid2str(req->rq_peer.nid),
1214 atomic_read(&export->exp_refcount));
1215 GOTO(out, rc = -EBUSY);
1216 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1217 if (!strstr(cluuid.uuid, "mdt"))
1218 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1219 "known client %s (at %s) because it "
1220 "is indicating it is a new client",
1221 target->obd_name, cluuid.uuid,
1222 libcfs_nid2str(req->rq_peer.nid));
1223 GOTO(out, rc = -EALREADY);
1225 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1232 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1233 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1234 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1235 export, (long)cfs_time_current_sec(),
1236 export ? (long)export->exp_last_request_time : 0);
1238 /* If this is the first time a client connects, reset the recovery
1239 * timer. Discard lightweight connections which might be local. */
1240 if (!lw_client && rc == 0 && target->obd_recovering)
1241 check_and_start_recovery_timer(target, req, export == NULL);
1243 /* We want to handle EALREADY but *not* -EALREADY from
1244 * target_handle_reconnect(), return reconnection state in a flag. */
1245 if (rc == EALREADY) {
1246 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1252 /* Tell the client if we support replayable requests. */
1253 if (target->obd_replayable)
1254 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1255 client_nid = &req->rq_peer.nid;
1257 if (export == NULL) {
1258 /* allow lightweight connections during recovery */
1259 /* allow "new" MDT to be connected during recovery, since we
1260 * need retrieve recovery update records from it */
1261 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1263 int c; /* connected */
1264 int i; /* in progress */
1266 int s; /* stale/evicted */
1268 c = atomic_read(&target->obd_connected_clients);
1269 i = atomic_read(&target->obd_lock_replay_clients);
1270 k = target->obd_max_recoverable_clients;
1271 s = target->obd_stale_clients;
1272 t = target->obd_recovery_timer.expires;
1273 t = cfs_time_sub(t, cfs_time_current());
1274 t = cfs_duration_sec(t);
1275 LCONSOLE_WARN("%s: Denying connection for new client %s"
1276 "(at %s), waiting for %d known clients "
1277 "(%d recovered, %d in progress, and %d "
1278 "evicted) to recover in %d:%.02d\n",
1279 target->obd_name, cluuid.uuid,
1280 libcfs_nid2str(req->rq_peer.nid), k,
1281 c - i, i, s, (int)t / 60,
1286 rc = obd_connect(req->rq_svc_thread->t_env,
1287 &export, target, &cluuid, data,
1289 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1290 lustre_msg_add_op_flags(req->rq_repmsg,
1291 MSG_CONNECT_RECOVERING);
1293 conn.cookie = export->exp_handle.h_cookie;
1294 rc = rev_import_init(export);
1298 new_mds_mds_conn = true;
1301 rc = obd_reconnect(req->rq_svc_thread->t_env,
1302 export, target, &cluuid, data, client_nid);
1307 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1308 data->ocd_instance = target->u.obt.obt_instance;
1310 /* Return only the parts of obd_connect_data that we understand, so the
1311 * client knows that we don't understand the rest. */
1313 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1315 tmpdata = req_capsule_server_get(&req->rq_pill,
1317 /* Don't use struct assignment here, because the client reply
1318 * buffer may be smaller/larger than the local struct
1319 * obd_connect_data. */
1320 memcpy(tmpdata, data, min(tmpsize, size));
1323 /* If the client and the server are the same node, we will already
1324 * have an export that really points to the client's DLM export,
1325 * because we have a shared handles table.
1327 * XXX this will go away when shaver stops sending the "connect" handle
1328 * in the real "remote handle" field of the request --phik 24 Apr 2003
1330 ptlrpc_request_change_export(req, export);
1332 spin_lock(&export->exp_lock);
1333 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1334 spin_unlock(&export->exp_lock);
1335 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1336 "or equal conn_cnt: %d >= %d\n",
1337 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1338 export->exp_conn_cnt,
1339 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1341 GOTO(out, rc = -EALREADY);
1343 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1344 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1346 /* Don't evict liblustre clients for not pinging. */
1347 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1348 export->exp_libclient = 1;
1349 spin_unlock(&export->exp_lock);
1351 spin_lock(&target->obd_dev_lock);
1352 list_del_init(&export->exp_obd_chain_timed);
1353 spin_unlock(&target->obd_dev_lock);
1355 spin_unlock(&export->exp_lock);
1358 if (export->exp_connection != NULL) {
1359 /* Check to see if connection came from another NID. */
1360 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1361 !hlist_unhashed(&export->exp_nid_hash))
1362 cfs_hash_del(export->exp_obd->obd_nid_hash,
1363 &export->exp_connection->c_peer.nid,
1364 &export->exp_nid_hash);
1366 ptlrpc_connection_put(export->exp_connection);
1369 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1372 if (hlist_unhashed(&export->exp_nid_hash))
1373 cfs_hash_add(export->exp_obd->obd_nid_hash,
1374 &export->exp_connection->c_peer.nid,
1375 &export->exp_nid_hash);
1377 lustre_msg_set_handle(req->rq_repmsg, &conn);
1379 rc = rev_import_reconnect(export, req);
1383 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1385 __u64 transno = data->ocd_transno;
1387 spin_lock(&export->exp_lock);
1388 /* possible race with class_disconnect_stale_exports,
1389 * export may be already in the eviction process */
1390 if (export->exp_failed) {
1391 spin_unlock(&export->exp_lock);
1392 GOTO(out, rc = -ENODEV);
1394 export->exp_in_recovery = 1;
1395 export->exp_req_replay_needed = 1;
1396 export->exp_lock_replay_needed = 1;
1397 spin_unlock(&export->exp_lock);
1399 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1400 MSG_CONNECT_TRANSNO);
1401 if (has_transno && transno == 0)
1402 CWARN("Connect with zero transno!\n");
1404 if (has_transno && transno > 0 &&
1405 transno < target->obd_next_recovery_transno &&
1406 transno > target->obd_last_committed) {
1407 /* Another way is to use cmpxchg() to be lock-free. */
1408 spin_lock(&target->obd_recovery_task_lock);
1409 if (transno < target->obd_next_recovery_transno)
1410 target->obd_next_recovery_transno = transno;
1411 spin_unlock(&target->obd_recovery_task_lock);
1414 atomic_inc(&target->obd_req_replay_clients);
1415 atomic_inc(&target->obd_lock_replay_clients);
1416 /* Note: MDS-MDS connection is allowed to be connected during
1417 * recovery, no matter if the exports needs to be recoveried.
1418 * Because we need retrieve updates logs from all other MDTs.
1419 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1420 * also needs to be increased to match other recovery checking
1422 if (new_mds_mds_conn)
1423 target->obd_max_recoverable_clients++;
1424 if (atomic_inc_return(&target->obd_connected_clients) ==
1425 target->obd_max_recoverable_clients)
1426 wake_up(&target->obd_next_transno_waitq);
1429 /* Tell the client we're in recovery, when client is involved in it. */
1430 if (target->obd_recovering && !lw_client)
1431 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1435 spin_lock(&export->exp_lock);
1436 export->exp_connecting = 0;
1437 spin_unlock(&export->exp_lock);
1439 class_export_put(export);
1441 if (target_referenced == true && target != NULL) {
1442 spin_lock(&target->obd_dev_lock);
1443 target->obd_conn_inprogress--;
1444 spin_unlock(&target->obd_dev_lock);
1446 class_decref(target, __func__, current);
1448 req->rq_status = rc;
1452 int target_handle_disconnect(struct ptlrpc_request *req)
1457 rc = req_capsule_server_pack(&req->rq_pill);
1461 /* Keep the rq_export around so we can send the reply. */
1462 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1467 void target_destroy_export(struct obd_export *exp)
1469 struct obd_import *imp = NULL;
1470 /* exports created from last_rcvd data, and "fake"
1471 exports created by lctl don't have an import */
1472 spin_lock(&exp->exp_lock);
1473 if (exp->exp_imp_reverse != NULL) {
1474 imp = exp->exp_imp_reverse;
1475 exp->exp_imp_reverse = NULL;
1477 spin_unlock(&exp->exp_lock);
1479 client_destroy_import(imp);
1481 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1482 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1483 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1484 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1486 EXPORT_SYMBOL(target_destroy_export);
1489 * Recovery functions
1491 static void target_request_copy_get(struct ptlrpc_request *req)
1493 class_export_rpc_inc(req->rq_export);
1494 LASSERT(list_empty(&req->rq_list));
1495 INIT_LIST_HEAD(&req->rq_replay_list);
1497 /* Increase refcount to keep request in queue. */
1498 atomic_inc(&req->rq_refcount);
1499 /* Let export know it has replays to be handled. */
1500 atomic_inc(&req->rq_export->exp_replay_count);
1503 static void target_request_copy_put(struct ptlrpc_request *req)
1505 LASSERT(list_empty(&req->rq_replay_list));
1506 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1508 atomic_dec(&req->rq_export->exp_replay_count);
1509 class_export_rpc_dec(req->rq_export);
1510 ptlrpc_server_drop_request(req);
1513 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1515 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1516 struct obd_export *exp = req->rq_export;
1517 struct ptlrpc_request *reqiter;
1518 struct ptlrpc_request *dup_req = NULL;
1523 spin_lock(&exp->exp_lock);
1524 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1526 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1534 /* We expect it with RESENT and REPLAY flags. */
1535 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1536 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1537 CERROR("invalid flags %x of resent replay\n",
1538 lustre_msg_get_flags(req->rq_reqmsg));
1540 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1543 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1545 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1546 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1550 list_add_tail(&req->rq_replay_list,
1551 &exp->exp_req_replay_queue);
1554 spin_unlock(&exp->exp_lock);
1558 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1560 LASSERT(!list_empty(&req->rq_replay_list));
1561 LASSERT(req->rq_export);
1563 spin_lock(&req->rq_export->exp_lock);
1564 list_del_init(&req->rq_replay_list);
1565 spin_unlock(&req->rq_export->exp_lock);
1568 static void target_finish_recovery(struct lu_target *lut)
1570 struct obd_device *obd = lut->lut_obd;
1573 /* Only log a recovery message when recovery has occurred. */
1574 if (obd->obd_recovery_start) {
1575 time64_t now = ktime_get_real_seconds();
1576 time64_t elapsed_time;
1578 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1579 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1580 "%d recovered and %d %s evicted.\n", obd->obd_name,
1581 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1582 obd->obd_max_recoverable_clients,
1583 atomic_read(&obd->obd_connected_clients),
1584 obd->obd_stale_clients,
1585 obd->obd_stale_clients == 1 ? "was" : "were");
1588 ldlm_reprocess_recovery_done(obd->obd_namespace);
1589 spin_lock(&obd->obd_recovery_task_lock);
1590 if (!list_empty(&obd->obd_req_replay_queue) ||
1591 !list_empty(&obd->obd_lock_replay_queue) ||
1592 !list_empty(&obd->obd_final_req_queue)) {
1593 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1595 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1596 list_empty(&obd->obd_lock_replay_queue) ? \
1598 list_empty(&obd->obd_final_req_queue) ? \
1600 spin_unlock(&obd->obd_recovery_task_lock);
1603 spin_unlock(&obd->obd_recovery_task_lock);
1605 obd->obd_recovery_end = ktime_get_real_seconds();
1607 /* When recovery finished, cleanup orphans on MDS and OST. */
1608 if (OBT(obd) && OBP(obd, postrecov)) {
1609 int rc = OBP(obd, postrecov)(obd);
1611 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1617 static void abort_req_replay_queue(struct obd_device *obd)
1619 struct ptlrpc_request *req, *n;
1620 struct list_head abort_list;
1622 INIT_LIST_HEAD(&abort_list);
1623 spin_lock(&obd->obd_recovery_task_lock);
1624 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1625 spin_unlock(&obd->obd_recovery_task_lock);
1626 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1627 DEBUG_REQ(D_WARNING, req, "aborted:");
1628 req->rq_status = -ENOTCONN;
1629 if (ptlrpc_error(req)) {
1630 DEBUG_REQ(D_ERROR, req,
1631 "failed abort_req_reply; skipping");
1633 target_exp_dequeue_req_replay(req);
1634 target_request_copy_put(req);
1638 static void abort_lock_replay_queue(struct obd_device *obd)
1640 struct ptlrpc_request *req, *n;
1641 struct list_head abort_list;
1643 INIT_LIST_HEAD(&abort_list);
1644 spin_lock(&obd->obd_recovery_task_lock);
1645 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1646 spin_unlock(&obd->obd_recovery_task_lock);
1647 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1648 DEBUG_REQ(D_ERROR, req, "aborted:");
1649 req->rq_status = -ENOTCONN;
1650 if (ptlrpc_error(req)) {
1651 DEBUG_REQ(D_ERROR, req,
1652 "failed abort_lock_reply; skipping");
1654 target_request_copy_put(req);
1658 /* Called from a cleanup function if the device is being cleaned up
1659 forcefully. The exports should all have been disconnected already,
1660 the only thing left to do is
1661 - clear the recovery flags
1663 - free queued requests and replies, but don't send replies
1664 Because the obd_stopping flag is set, no new requests should be received.
1667 void target_cleanup_recovery(struct obd_device *obd)
1669 struct ptlrpc_request *req, *n;
1670 struct list_head clean_list;
1672 INIT_LIST_HEAD(&clean_list);
1673 spin_lock(&obd->obd_dev_lock);
1674 if (!obd->obd_recovering) {
1675 spin_unlock(&obd->obd_dev_lock);
1679 obd->obd_recovering = obd->obd_abort_recovery = 0;
1680 spin_unlock(&obd->obd_dev_lock);
1682 spin_lock(&obd->obd_recovery_task_lock);
1683 target_cancel_recovery_timer(obd);
1684 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1685 spin_unlock(&obd->obd_recovery_task_lock);
1687 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1688 LASSERT(req->rq_reply_state == NULL);
1689 target_exp_dequeue_req_replay(req);
1690 target_request_copy_put(req);
1693 spin_lock(&obd->obd_recovery_task_lock);
1694 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1695 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1696 spin_unlock(&obd->obd_recovery_task_lock);
1698 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1699 LASSERT(req->rq_reply_state == NULL);
1700 target_request_copy_put(req);
1705 EXPORT_SYMBOL(target_cleanup_recovery);
1707 /* obd_recovery_task_lock should be held */
1708 void target_cancel_recovery_timer(struct obd_device *obd)
1710 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1711 del_timer(&obd->obd_recovery_timer);
1714 static void target_start_recovery_timer(struct obd_device *obd)
1716 if (obd->obd_recovery_start != 0)
1719 spin_lock(&obd->obd_dev_lock);
1720 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1721 spin_unlock(&obd->obd_dev_lock);
1725 LASSERT(obd->obd_recovery_timeout != 0);
1727 if (obd->obd_recovery_start != 0) {
1728 spin_unlock(&obd->obd_dev_lock);
1732 mod_timer(&obd->obd_recovery_timer,
1733 cfs_time_shift(obd->obd_recovery_timeout));
1734 obd->obd_recovery_start = ktime_get_real_seconds();
1735 spin_unlock(&obd->obd_dev_lock);
1737 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1739 obd->obd_recovery_timeout / 60,
1740 obd->obd_recovery_timeout % 60,
1741 obd->obd_max_recoverable_clients,
1742 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1743 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1747 * extend recovery window.
1749 * if @extend is true, extend recovery window to have @drt remaining at least;
1750 * otherwise, make sure the recovery timeout value is not less than @drt.
1752 static void extend_recovery_timer(struct obd_device *obd, int drt,
1760 spin_lock(&obd->obd_dev_lock);
1761 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1762 spin_unlock(&obd->obd_dev_lock);
1765 LASSERT(obd->obd_recovery_start != 0);
1767 now = ktime_get_real_seconds();
1768 to = obd->obd_recovery_timeout;
1769 end = obd->obd_recovery_start + to;
1772 if (extend && (drt > left)) {
1774 } else if (!extend && (drt > to)) {
1778 if (to > obd->obd_recovery_time_hard) {
1779 to = obd->obd_recovery_time_hard;
1780 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1781 obd->obd_name, to, extend);
1784 if (obd->obd_recovery_timeout < to) {
1785 obd->obd_recovery_timeout = to;
1786 end = obd->obd_recovery_start + to;
1787 mod_timer(&obd->obd_recovery_timer,
1788 cfs_time_shift(end - now));
1790 spin_unlock(&obd->obd_dev_lock);
1792 CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1793 obd->obd_name, (s64)(end - now));
1796 /* Reset the timer with each new client connection */
1798 * This timer is actually reconnect_timer, which is for making sure
1799 * the total recovery window is at least as big as my reconnect
1800 * attempt timing. So the initial recovery time_out will be set to
1801 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1802 * from client is bigger than this, then the recovery time_out will
1803 * be extended to make sure the client could be reconnected, in the
1804 * process, the timeout from the new client should be ignored.
1808 check_and_start_recovery_timer(struct obd_device *obd,
1809 struct ptlrpc_request *req,
1812 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1813 struct obd_device_target *obt = &obd->u.obt;
1815 if (!new_client && service_time)
1816 /* Teach server about old server's estimates, as first guess
1817 * at how long new requests will take. */
1818 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1821 target_start_recovery_timer(obd);
1823 /* Convert the service time to RPC timeout,
1824 * and reuse service_time to limit stack usage. */
1825 service_time = at_est2timeout(service_time);
1827 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1828 service_time < at_extra)
1829 service_time = at_extra;
1831 /* We expect other clients to timeout within service_time, then try
1832 * to reconnect, then try the failover server. The max delay between
1833 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1834 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1836 LASSERT(obt->obt_magic == OBT_MAGIC);
1837 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1838 if (service_time > obd->obd_recovery_timeout && !new_client)
1839 extend_recovery_timer(obd, service_time, false);
1842 /** Health checking routines */
1843 static inline int exp_connect_healthy(struct obd_export *exp)
1845 return (exp->exp_in_recovery);
1848 /** if export done req_replay or has replay in queue */
1849 static inline int exp_req_replay_healthy(struct obd_export *exp)
1851 return (!exp->exp_req_replay_needed ||
1852 atomic_read(&exp->exp_replay_count) > 0);
1856 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1858 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1859 exp_req_replay_healthy(exp);
1862 /** if export done lock_replay or has replay in queue */
1863 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1865 return (!exp->exp_lock_replay_needed ||
1866 atomic_read(&exp->exp_replay_count) > 0);
1869 static inline int exp_vbr_healthy(struct obd_export *exp)
1871 return (!exp->exp_vbr_failed);
1874 static inline int exp_finished(struct obd_export *exp)
1876 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1879 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1881 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1885 static int check_for_next_transno(struct lu_target *lut)
1887 struct ptlrpc_request *req = NULL;
1888 struct obd_device *obd = lut->lut_obd;
1889 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1890 int wake_up = 0, connected, completed, queue_len;
1891 __u64 req_transno = 0;
1892 __u64 update_transno = 0;
1893 __u64 next_transno = 0;
1896 spin_lock(&obd->obd_recovery_task_lock);
1897 if (!list_empty(&obd->obd_req_replay_queue)) {
1898 req = list_entry(obd->obd_req_replay_queue.next,
1899 struct ptlrpc_request, rq_list);
1900 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1904 update_transno = distribute_txn_get_next_transno(tdtd);
1906 connected = atomic_read(&obd->obd_connected_clients);
1907 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1908 queue_len = obd->obd_requests_queued_for_recovery;
1909 next_transno = obd->obd_next_recovery_transno;
1911 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1912 "req_transno: %llu, next_transno: %llu\n",
1913 obd->obd_max_recoverable_clients, connected, completed,
1914 queue_len, req_transno, next_transno);
1916 if (obd->obd_abort_recovery) {
1917 CDEBUG(D_HA, "waking for aborted recovery\n");
1919 } else if (obd->obd_recovery_expired) {
1920 CDEBUG(D_HA, "waking for expired recovery\n");
1922 } else if (tdtd != NULL && req != NULL &&
1923 is_req_replayed_by_update(req)) {
1924 LASSERTF(req_transno < next_transno, "req_transno %llu"
1925 "next_transno%llu\n", req_transno, next_transno);
1926 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1929 } else if (req_transno == next_transno ||
1930 (update_transno != 0 && update_transno <= next_transno)) {
1931 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1933 } else if (queue_len > 0 &&
1934 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1935 /** handle gaps occured due to lost reply or VBR */
1936 LASSERTF(req_transno >= next_transno,
1937 "req_transno: %llu, next_transno: %llu\n",
1938 req_transno, next_transno);
1940 "%s: waking for gap in transno, VBR is %s (skip: "
1941 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1942 ", next_update %lld last_committed: %lld)\n",
1943 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1944 next_transno, queue_len, completed, connected,
1945 req_transno, update_transno, obd->obd_last_committed);
1946 obd->obd_next_recovery_transno = req_transno;
1948 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1949 CDEBUG(D_HA, "waking for completed recovery\n");
1951 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1952 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1953 " by fail_lock, waking up (%lld)\n", next_transno);
1954 obd->obd_next_recovery_transno = req_transno;
1957 spin_unlock(&obd->obd_recovery_task_lock);
1961 static int check_for_next_lock(struct lu_target *lut)
1963 struct obd_device *obd = lut->lut_obd;
1966 spin_lock(&obd->obd_recovery_task_lock);
1967 if (!list_empty(&obd->obd_lock_replay_queue)) {
1968 CDEBUG(D_HA, "waking for next lock\n");
1970 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1971 CDEBUG(D_HA, "waking for completed lock replay\n");
1973 } else if (obd->obd_abort_recovery) {
1974 CDEBUG(D_HA, "waking for aborted recovery\n");
1976 } else if (obd->obd_recovery_expired) {
1977 CDEBUG(D_HA, "waking for expired recovery\n");
1980 spin_unlock(&obd->obd_recovery_task_lock);
1986 * wait for recovery events,
1987 * check its status with help of check_routine
1988 * evict dead clients via health_check
1990 static int target_recovery_overseer(struct lu_target *lut,
1991 int (*check_routine)(struct lu_target *),
1992 int (*health_check)(struct obd_export *))
1994 struct obd_device *obd = lut->lut_obd;
1995 struct target_distribute_txn_data *tdtd;
1997 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
1998 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
1999 __u64 next_update_transno = 0;
2001 /* Only abort the recovery if there are no update recovery
2002 * left in the queue */
2003 spin_lock(&obd->obd_recovery_task_lock);
2004 if (lut->lut_tdtd != NULL) {
2005 next_update_transno =
2006 distribute_txn_get_next_transno(lut->lut_tdtd);
2008 tdtd = lut->lut_tdtd;
2009 /* If next_update_transno == 0, it probably because
2010 * updatelog retrieve threads did not get any records
2011 * yet, let's wait those threads stopped */
2012 if (next_update_transno == 0) {
2013 struct l_wait_info lwi = { 0 };
2015 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2017 &tdtd->tdtd_recovery_threads_count) == 0,
2020 next_update_transno =
2021 distribute_txn_get_next_transno(
2026 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2027 obd->obd_next_recovery_transno = next_update_transno;
2028 spin_unlock(&obd->obd_recovery_task_lock);
2029 /* Disconnect unfinished exports from clients, and
2030 * keep connection from MDT to make sure the update
2031 * recovery will still keep trying until some one
2032 * manually abort the recovery */
2033 class_disconnect_stale_exports(obd,
2034 exp_finished_or_from_mdt);
2035 /* Abort all of replay and replay lock req from
2037 abort_req_replay_queue(obd);
2038 abort_lock_replay_queue(obd);
2039 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2040 ")in the queue.\n", obd->obd_name,
2041 next_update_transno);
2043 obd->obd_abort_recovery = 1;
2044 spin_unlock(&obd->obd_recovery_task_lock);
2045 CWARN("%s recovery is aborted by hard timeout\n",
2050 while (wait_event_timeout(obd->obd_next_transno_waitq,
2052 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2053 /* wait indefinitely for event, but don't trigger watchdog */;
2055 if (obd->obd_abort_recovery) {
2056 CWARN("recovery is aborted, evict exports in recovery\n");
2057 if (lut->lut_tdtd != NULL) {
2058 struct l_wait_info lwi = { 0 };
2060 tdtd = lut->lut_tdtd;
2061 /* Let's wait all of the update log recovery thread
2063 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2064 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2066 /* Then abort the update recovery list */
2067 dtrq_list_destroy(lut->lut_tdtd);
2070 /** evict exports which didn't finish recovery yet */
2071 class_disconnect_stale_exports(obd, exp_finished);
2073 } else if (obd->obd_recovery_expired) {
2074 obd->obd_recovery_expired = 0;
2075 /** If some clients died being recovered, evict them */
2076 LCONSOLE_WARN("%s: recovery is timed out, "
2077 "evict stale exports\n", obd->obd_name);
2078 /** evict cexports with no replay in queue, they are stalled */
2079 class_disconnect_stale_exports(obd, health_check);
2081 /** continue with VBR */
2082 spin_lock(&obd->obd_dev_lock);
2083 obd->obd_version_recov = 1;
2084 spin_unlock(&obd->obd_dev_lock);
2086 * reset timer, recovery will proceed with versions now,
2087 * timeout is set just to handle reconnection delays
2089 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2090 /** Wait for recovery events again, after evicting bad clients */
2096 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2098 struct obd_device *obd = lut->lut_obd;
2099 struct ptlrpc_request *req = NULL;
2101 CDEBUG(D_HA, "Waiting for lock\n");
2102 if (target_recovery_overseer(lut, check_for_next_lock,
2103 exp_lock_replay_healthy))
2104 abort_lock_replay_queue(obd);
2106 spin_lock(&obd->obd_recovery_task_lock);
2107 if (!list_empty(&obd->obd_lock_replay_queue)) {
2108 req = list_entry(obd->obd_lock_replay_queue.next,
2109 struct ptlrpc_request, rq_list);
2110 list_del_init(&req->rq_list);
2111 spin_unlock(&obd->obd_recovery_task_lock);
2113 spin_unlock(&obd->obd_recovery_task_lock);
2114 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2115 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2116 /** evict exports failed VBR */
2117 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2122 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2124 struct ptlrpc_request *req = NULL;
2126 spin_lock(&obd->obd_recovery_task_lock);
2127 if (!list_empty(&obd->obd_final_req_queue)) {
2128 req = list_entry(obd->obd_final_req_queue.next,
2129 struct ptlrpc_request, rq_list);
2130 list_del_init(&req->rq_list);
2131 spin_unlock(&obd->obd_recovery_task_lock);
2132 if (req->rq_export->exp_in_recovery) {
2133 spin_lock(&req->rq_export->exp_lock);
2134 req->rq_export->exp_in_recovery = 0;
2135 spin_unlock(&req->rq_export->exp_lock);
2138 spin_unlock(&obd->obd_recovery_task_lock);
2143 static void handle_recovery_req(struct ptlrpc_thread *thread,
2144 struct ptlrpc_request *req,
2145 svc_handler_t handler)
2150 * export can be evicted during recovery, no need to handle replays for
2151 * it after that, discard such request silently
2153 if (req->rq_export->exp_disconnected)
2156 req->rq_session.lc_thread = thread;
2157 req->rq_svc_thread = thread;
2158 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2160 /* thread context */
2161 lu_context_enter(&thread->t_env->le_ctx);
2163 lu_context_exit(&thread->t_env->le_ctx);
2165 /* don't reset timer for final stage */
2166 if (!exp_finished(req->rq_export)) {
2167 int to = obd_timeout;
2170 * Add request timeout to the recovery time so next request from
2171 * this client may come in recovery time
2174 struct ptlrpc_service_part *svcpt;
2176 svcpt = req->rq_rqbd->rqbd_svcpt;
2177 /* If the server sent early reply for this request,
2178 * the client will recalculate the timeout according to
2179 * current server estimate service time, so we will
2180 * use the maxium timeout here for waiting the client
2181 * sending the next req */
2182 to = max((int)at_est2timeout(
2183 at_get(&svcpt->scp_at_estimate)),
2184 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2185 /* Add 2 net_latency, one for balance rq_deadline
2186 * (see ptl_send_rpc), one for resend the req to server,
2187 * Note: client will pack net_latency in replay req
2188 * (see ptlrpc_replay_req) */
2189 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2191 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2196 /** Checking routines for recovery */
2197 static int check_for_recovery_ready(struct lu_target *lut)
2199 struct obd_device *obd = lut->lut_obd;
2200 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2202 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2203 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2204 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2205 obd->obd_recovery_expired);
2207 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2208 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2209 if (clnts + obd->obd_stale_clients <
2210 obd->obd_max_recoverable_clients)
2214 if (lut->lut_tdtd != NULL) {
2215 if (!lut->lut_tdtd->tdtd_replay_ready &&
2216 !obd->obd_abort_recovery) {
2217 /* Let's extend recovery timer, in case the recovery
2218 * timer expired, and some clients got evicted */
2219 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2221 CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2222 obd->obd_name, obd->obd_recovery_timeout);
2231 REQUEST_RECOVERY = 1,
2232 UPDATE_RECOVERY = 2,
2235 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2239 if (!list_empty(&obd->obd_req_replay_queue)) {
2240 struct ptlrpc_request *req;
2242 req = list_entry(obd->obd_req_replay_queue.next,
2243 struct ptlrpc_request, rq_list);
2244 transno = lustre_msg_get_transno(req->rq_reqmsg);
2250 static __u64 get_next_transno(struct lu_target *lut, int *type)
2252 struct obd_device *obd = lut->lut_obd;
2253 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2255 __u64 update_transno;
2258 transno = get_next_replay_req_transno(obd);
2260 *type = REQUEST_RECOVERY;
2265 update_transno = distribute_txn_get_next_transno(tdtd);
2266 if (transno == 0 || (transno >= update_transno &&
2267 update_transno != 0)) {
2268 transno = update_transno;
2270 *type = UPDATE_RECOVERY;
2277 * drop duplicate replay request
2279 * Because the operation has been replayed by update recovery, the request
2280 * with the same transno will be dropped and also notify the client to send
2281 * next replay request.
2283 * \param[in] env execution environment
2284 * \param[in] obd failover obd device
2285 * \param[in] req request to be dropped
2287 static void drop_duplicate_replay_req(struct lu_env *env,
2288 struct obd_device *obd,
2289 struct ptlrpc_request *req)
2291 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2292 " update records are found.\n",
2293 lustre_msg_get_transno(req->rq_reqmsg),
2294 libcfs_nid2str(req->rq_peer.nid));
2296 /* Right now, only for MDS reint operation update replay and
2297 * normal request replay can have the same transno */
2298 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2299 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2300 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2301 if (likely(req->rq_export))
2302 target_committed_to_req(req);
2303 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2304 target_send_reply(req, req->rq_status, 0);
2306 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2307 libcfs_nid2str(req->rq_peer.nid));
2309 target_exp_dequeue_req_replay(req);
2310 target_request_copy_put(req);
2311 obd->obd_replayed_requests++;
2314 static void replay_request_or_update(struct lu_env *env,
2315 struct lu_target *lut,
2316 struct target_recovery_data *trd,
2317 struct ptlrpc_thread *thread)
2319 struct obd_device *obd = lut->lut_obd;
2320 struct ptlrpc_request *req = NULL;
2325 CDEBUG(D_HA, "Waiting for transno %lld\n",
2326 obd->obd_next_recovery_transno);
2328 /* Replay all of request and update by transno */
2330 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2332 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2334 /** It is needed to extend recovery window above
2335 * recovery_time_soft. Extending is possible only in the
2336 * end of recovery window (see more details in
2337 * handle_recovery_req()).
2339 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2341 if (target_recovery_overseer(lut, check_for_next_transno,
2342 exp_req_replay_healthy_or_from_mdt)) {
2343 abort_req_replay_queue(obd);
2344 abort_lock_replay_queue(obd);
2348 spin_lock(&obd->obd_recovery_task_lock);
2349 transno = get_next_transno(lut, &type);
2350 if (type == REQUEST_RECOVERY && transno != 0) {
2351 /* Drop replay request from client side, if the
2352 * replay has been executed by update with the
2354 req = list_entry(obd->obd_req_replay_queue.next,
2355 struct ptlrpc_request, rq_list);
2357 list_del_init(&req->rq_list);
2358 obd->obd_requests_queued_for_recovery--;
2359 spin_unlock(&obd->obd_recovery_task_lock);
2361 /* Let's check if the request has been redone by
2363 if (is_req_replayed_by_update(req)) {
2364 struct distribute_txn_replay_req *dtrq;
2366 dtrq = distribute_txn_lookup_finish_list(tdtd,
2368 LASSERT(dtrq != NULL);
2369 spin_lock(&tdtd->tdtd_replay_list_lock);
2370 list_del_init(&dtrq->dtrq_list);
2371 spin_unlock(&tdtd->tdtd_replay_list_lock);
2374 drop_duplicate_replay_req(env, obd, req);
2379 LASSERT(trd->trd_processing_task == current_pid());
2380 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2381 lustre_msg_get_transno(req->rq_reqmsg),
2382 libcfs_nid2str(req->rq_peer.nid));
2384 handle_recovery_req(thread, req,
2385 trd->trd_recovery_handler);
2387 * bz18031: increase next_recovery_transno before
2388 * target_request_copy_put() will drop exp_rpc reference
2390 spin_lock(&obd->obd_recovery_task_lock);
2391 obd->obd_next_recovery_transno++;
2392 spin_unlock(&obd->obd_recovery_task_lock);
2393 target_exp_dequeue_req_replay(req);
2394 target_request_copy_put(req);
2395 obd->obd_replayed_requests++;
2396 } else if (type == UPDATE_RECOVERY && transno != 0) {
2397 struct distribute_txn_replay_req *dtrq;
2400 spin_unlock(&obd->obd_recovery_task_lock);
2402 LASSERT(tdtd != NULL);
2403 dtrq = distribute_txn_get_next_req(tdtd);
2404 lu_context_enter(&thread->t_env->le_ctx);
2405 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2406 lu_context_exit(&thread->t_env->le_ctx);
2407 extend_recovery_timer(obd, obd_timeout, true);
2409 if (rc == 0 && dtrq->dtrq_xid != 0) {
2410 CDEBUG(D_HA, "Move x%llu t%llu"
2411 " to finish list\n", dtrq->dtrq_xid,
2412 dtrq->dtrq_master_transno);
2414 /* Add it to the replay finish list */
2415 spin_lock(&tdtd->tdtd_replay_list_lock);
2416 list_add(&dtrq->dtrq_list,
2417 &tdtd->tdtd_replay_finish_list);
2418 spin_unlock(&tdtd->tdtd_replay_list_lock);
2420 spin_lock(&obd->obd_recovery_task_lock);
2421 if (transno == obd->obd_next_recovery_transno)
2422 obd->obd_next_recovery_transno++;
2424 obd->obd_next_recovery_transno)
2425 obd->obd_next_recovery_transno =
2427 spin_unlock(&obd->obd_recovery_task_lock);
2432 spin_unlock(&obd->obd_recovery_task_lock);
2434 LASSERT(list_empty(&obd->obd_req_replay_queue));
2435 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2436 /** evict exports failed VBR */
2437 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2443 static int target_recovery_thread(void *arg)
2445 struct lu_target *lut = arg;
2446 struct obd_device *obd = lut->lut_obd;
2447 struct ptlrpc_request *req;
2448 struct target_recovery_data *trd = &obd->obd_recovery_data;
2449 unsigned long delta;
2451 struct ptlrpc_thread *thread = NULL;
2455 unshare_fs_struct();
2456 OBD_ALLOC_PTR(thread);
2462 OBD_FREE_PTR(thread);
2466 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2468 OBD_FREE_PTR(thread);
2473 thread->t_env = env;
2474 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2475 env->le_ctx.lc_thread = thread;
2476 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2477 thread->t_watchdog = NULL;
2479 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2481 trd->trd_processing_task = current_pid();
2483 spin_lock(&obd->obd_dev_lock);
2484 obd->obd_recovering = 1;
2485 spin_unlock(&obd->obd_dev_lock);
2486 complete(&trd->trd_starting);
2488 /* first of all, we have to know the first transno to replay */
2489 if (target_recovery_overseer(lut, check_for_recovery_ready,
2490 exp_connect_healthy)) {
2491 abort_req_replay_queue(obd);
2492 abort_lock_replay_queue(obd);
2493 if (lut->lut_tdtd != NULL)
2494 dtrq_list_destroy(lut->lut_tdtd);
2497 /* next stage: replay requests or update */
2499 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2500 atomic_read(&obd->obd_req_replay_clients),
2501 obd->obd_next_recovery_transno);
2502 replay_request_or_update(env, lut, trd, thread);
2505 * The second stage: replay locks
2507 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2508 atomic_read(&obd->obd_lock_replay_clients));
2509 while ((req = target_next_replay_lock(lut))) {
2510 LASSERT(trd->trd_processing_task == current_pid());
2511 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2512 libcfs_nid2str(req->rq_peer.nid));
2513 handle_recovery_req(thread, req,
2514 trd->trd_recovery_handler);
2515 target_request_copy_put(req);
2516 obd->obd_replayed_locks++;
2520 * The third stage: reply on final pings, at this moment all clients
2521 * must have request in final queue
2523 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2524 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2525 /** Update server last boot epoch */
2526 tgt_boot_epoch_update(lut);
2527 /* We drop recoverying flag to forward all new requests
2528 * to regular mds_handle() since now */
2529 spin_lock(&obd->obd_dev_lock);
2530 obd->obd_recovering = obd->obd_abort_recovery = 0;
2531 spin_unlock(&obd->obd_dev_lock);
2532 spin_lock(&obd->obd_recovery_task_lock);
2533 target_cancel_recovery_timer(obd);
2534 spin_unlock(&obd->obd_recovery_task_lock);
2535 while ((req = target_next_final_ping(obd))) {
2536 LASSERT(trd->trd_processing_task == current_pid());
2537 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2538 libcfs_nid2str(req->rq_peer.nid));
2539 handle_recovery_req(thread, req,
2540 trd->trd_recovery_handler);
2541 /* Because the waiting client can not send ping to server,
2542 * so we need refresh the last_request_time, to avoid the
2543 * export is being evicted */
2544 ptlrpc_update_export_timer(req->rq_export, 0);
2545 target_request_copy_put(req);
2548 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2549 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2550 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2551 if (delta > OBD_RECOVERY_TIME_SOFT) {
2552 CWARN("too long recovery - read logs\n");
2553 libcfs_debug_dumplog();
2556 target_finish_recovery(lut);
2558 lu_context_fini(&env->le_ctx);
2559 trd->trd_processing_task = 0;
2560 complete(&trd->trd_finishing);
2562 tgt_io_thread_done(thread);
2563 OBD_FREE_PTR(thread);
2568 static int target_start_recovery_thread(struct lu_target *lut,
2569 svc_handler_t handler)
2571 struct obd_device *obd = lut->lut_obd;
2573 struct target_recovery_data *trd = &obd->obd_recovery_data;
2576 memset(trd, 0, sizeof(*trd));
2577 init_completion(&trd->trd_starting);
2578 init_completion(&trd->trd_finishing);
2579 trd->trd_recovery_handler = handler;
2581 rc = server_name2index(obd->obd_name, &index, NULL);
2585 if (!IS_ERR(kthread_run(target_recovery_thread,
2586 lut, "tgt_recover_%d", index))) {
2587 wait_for_completion(&trd->trd_starting);
2588 LASSERT(obd->obd_recovering != 0);
2596 void target_stop_recovery_thread(struct obd_device *obd)
2598 if (obd->obd_recovery_data.trd_processing_task > 0) {
2599 struct target_recovery_data *trd = &obd->obd_recovery_data;
2600 /** recovery can be done but postrecovery is not yet */
2601 spin_lock(&obd->obd_dev_lock);
2602 if (obd->obd_recovering) {
2603 CERROR("%s: Aborting recovery\n", obd->obd_name);
2604 obd->obd_abort_recovery = 1;
2605 wake_up(&obd->obd_next_transno_waitq);
2607 spin_unlock(&obd->obd_dev_lock);
2608 wait_for_completion(&trd->trd_finishing);
2611 EXPORT_SYMBOL(target_stop_recovery_thread);
2613 void target_recovery_fini(struct obd_device *obd)
2615 class_disconnect_exports(obd);
2616 target_stop_recovery_thread(obd);
2617 target_cleanup_recovery(obd);
2619 EXPORT_SYMBOL(target_recovery_fini);
2621 static void target_recovery_expired(unsigned long castmeharder)
2623 struct obd_device *obd = (struct obd_device *)castmeharder;
2624 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2625 " after %llus (%d clients connected)\n",
2626 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2627 (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2628 atomic_read(&obd->obd_connected_clients));
2630 obd->obd_recovery_expired = 1;
2631 wake_up(&obd->obd_next_transno_waitq);
2634 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2636 struct obd_device *obd = lut->lut_obd;
2638 if (lut->lut_bottom->dd_rdonly)
2641 if (obd->obd_max_recoverable_clients == 0) {
2642 /** Update server last boot epoch */
2643 tgt_boot_epoch_update(lut);
2647 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2648 "last_transno %llu\n", obd->obd_name,
2649 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2650 LASSERT(obd->obd_stopping == 0);
2651 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2652 obd->obd_recovery_start = 0;
2653 obd->obd_recovery_end = 0;
2655 setup_timer(&obd->obd_recovery_timer, target_recovery_expired,
2656 (unsigned long)obd);
2657 target_start_recovery_thread(lut, handler);
2659 EXPORT_SYMBOL(target_recovery_init);
2661 static int target_process_req_flags(struct obd_device *obd,
2662 struct ptlrpc_request *req)
2664 struct obd_export *exp = req->rq_export;
2665 LASSERT(exp != NULL);
2666 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2667 /* client declares he's ready to replay locks */
2668 spin_lock(&exp->exp_lock);
2669 if (exp->exp_req_replay_needed) {
2670 exp->exp_req_replay_needed = 0;
2671 spin_unlock(&exp->exp_lock);
2673 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2674 atomic_dec(&obd->obd_req_replay_clients);
2676 spin_unlock(&exp->exp_lock);
2679 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2680 /* client declares he's ready to complete recovery
2681 * so, we put the request on th final queue */
2682 spin_lock(&exp->exp_lock);
2683 if (exp->exp_lock_replay_needed) {
2684 exp->exp_lock_replay_needed = 0;
2685 spin_unlock(&exp->exp_lock);
2687 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2688 atomic_dec(&obd->obd_lock_replay_clients);
2690 spin_unlock(&exp->exp_lock);
2696 int target_queue_recovery_request(struct ptlrpc_request *req,
2697 struct obd_device *obd)
2699 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2700 struct ptlrpc_request *reqiter;
2704 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2705 /* Processing the queue right now, don't re-add. */
2709 target_process_req_flags(obd, req);
2711 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2712 /* client declares he's ready to complete recovery
2713 * so, we put the request on th final queue */
2714 target_request_copy_get(req);
2715 DEBUG_REQ(D_HA, req, "queue final req");
2716 wake_up(&obd->obd_next_transno_waitq);
2717 spin_lock(&obd->obd_recovery_task_lock);
2718 if (obd->obd_recovering) {
2719 struct ptlrpc_request *tmp;
2720 struct ptlrpc_request *duplicate = NULL;
2722 if (likely(!req->rq_export->exp_replay_done)) {
2723 req->rq_export->exp_replay_done = 1;
2724 list_add_tail(&req->rq_list,
2725 &obd->obd_final_req_queue);
2726 spin_unlock(&obd->obd_recovery_task_lock);
2730 /* XXX O(n), but only happens if final ping is
2731 * timed out, probably reorganize the list as
2732 * a hash list later */
2733 list_for_each_entry_safe(reqiter, tmp,
2734 &obd->obd_final_req_queue,
2736 if (reqiter->rq_export == req->rq_export) {
2737 list_del_init(&reqiter->rq_list);
2738 duplicate = reqiter;
2743 list_add_tail(&req->rq_list,
2744 &obd->obd_final_req_queue);
2745 req->rq_export->exp_replay_done = 1;
2746 spin_unlock(&obd->obd_recovery_task_lock);
2748 if (duplicate != NULL) {
2749 DEBUG_REQ(D_HA, duplicate,
2750 "put prev final req\n");
2751 target_request_copy_put(duplicate);
2755 spin_unlock(&obd->obd_recovery_task_lock);
2756 target_request_copy_put(req);
2757 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2760 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2761 /* client declares he's ready to replay locks */
2762 target_request_copy_get(req);
2763 DEBUG_REQ(D_HA, req, "queue lock replay req");
2764 wake_up(&obd->obd_next_transno_waitq);
2765 spin_lock(&obd->obd_recovery_task_lock);
2766 LASSERT(obd->obd_recovering);
2767 /* usually due to recovery abort */
2768 if (!req->rq_export->exp_in_recovery) {
2769 spin_unlock(&obd->obd_recovery_task_lock);
2770 target_request_copy_put(req);
2773 LASSERT(req->rq_export->exp_lock_replay_needed);
2774 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2775 spin_unlock(&obd->obd_recovery_task_lock);
2779 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2780 * (i.e. buflens etc are in my own byte order), but type-dependent
2781 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2784 INIT_LIST_HEAD(&req->rq_list);
2785 DEBUG_REQ(D_HA, req, "not queueing");
2789 /* If we're processing the queue, we want don't want to queue this
2792 * Also, if this request has a transno less than the one we're waiting
2793 * for, we should process it now. It could (and currently always will)
2794 * be an open request for a descriptor that was opened some time ago.
2796 * Also, a resent, replayed request that has already been
2797 * handled will pass through here and be processed immediately.
2799 CDEBUG(D_HA, "Next recovery transno: %llu"
2800 ", current: %llu, replaying\n",
2801 obd->obd_next_recovery_transno, transno);
2803 /* If the request has been replayed by update replay, then sends this
2804 * request to the recovery thread (replay_request_or_update()), where
2805 * it will be handled */
2806 spin_lock(&obd->obd_recovery_task_lock);
2807 if (transno < obd->obd_next_recovery_transno &&
2808 !is_req_replayed_by_update(req)) {
2809 /* Processing the queue right now, don't re-add. */
2810 LASSERT(list_empty(&req->rq_list));
2811 spin_unlock(&obd->obd_recovery_task_lock);
2814 spin_unlock(&obd->obd_recovery_task_lock);
2816 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2819 target_request_copy_get(req);
2820 if (!req->rq_export->exp_in_recovery) {
2821 target_request_copy_put(req);
2824 LASSERT(req->rq_export->exp_req_replay_needed);
2826 if (target_exp_enqueue_req_replay(req)) {
2827 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2828 target_request_copy_put(req);
2833 spin_lock(&obd->obd_recovery_task_lock);
2834 LASSERT(obd->obd_recovering);
2835 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2836 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2837 list_add_tail(&req->rq_list, &reqiter->rq_list);
2842 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2844 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2845 "has been claimed by another client");
2846 spin_unlock(&obd->obd_recovery_task_lock);
2847 target_exp_dequeue_req_replay(req);
2848 target_request_copy_put(req);
2854 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2856 obd->obd_requests_queued_for_recovery++;
2857 spin_unlock(&obd->obd_recovery_task_lock);
2858 wake_up(&obd->obd_next_transno_waitq);
2862 int target_handle_ping(struct ptlrpc_request *req)
2864 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2865 return req_capsule_server_pack(&req->rq_pill);
2868 void target_committed_to_req(struct ptlrpc_request *req)
2870 struct obd_export *exp = req->rq_export;
2872 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2873 lustre_msg_set_last_committed(req->rq_repmsg,
2874 exp->exp_last_committed);
2876 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2877 "%d)", exp->exp_obd->obd_no_transno,
2878 req->rq_repmsg == NULL);
2880 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2881 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2884 #endif /* HAVE_SERVER_SUPPORT */
2887 * Packs current SLV and Limit into \a req.
2889 int target_pack_pool_reply(struct ptlrpc_request *req)
2891 struct obd_device *obd;
2894 /* Check that we still have all structures alive as this may
2895 * be some late RPC at shutdown time. */
2896 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2897 !exp_connect_lru_resize(req->rq_export))) {
2898 lustre_msg_set_slv(req->rq_repmsg, 0);
2899 lustre_msg_set_limit(req->rq_repmsg, 0);
2903 /* OBD is alive here as export is alive, which we checked above. */
2904 obd = req->rq_export->exp_obd;
2906 read_lock(&obd->obd_pool_lock);
2907 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2908 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2909 read_unlock(&obd->obd_pool_lock);
2914 static int target_send_reply_msg(struct ptlrpc_request *req,
2915 int rc, int fail_id)
2917 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2918 DEBUG_REQ(D_ERROR, req, "dropping reply");
2921 /* We can have a null rq_reqmsg in the event of bad signature or
2922 * no context when unwrapping */
2923 if (req->rq_reqmsg &&
2924 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2925 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2929 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2930 req->rq_status = rc;
2931 return ptlrpc_send_error(req, 1);
2933 DEBUG_REQ(D_NET, req, "sending reply");
2936 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2939 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2941 struct ptlrpc_service_part *svcpt;
2943 struct ptlrpc_reply_state *rs;
2944 struct obd_export *exp;
2947 if (req->rq_no_reply) {
2952 svcpt = req->rq_rqbd->rqbd_svcpt;
2953 rs = req->rq_reply_state;
2954 if (rs == NULL || !rs->rs_difficult) {
2956 target_send_reply_msg (req, rc, fail_id);
2961 /* must be an export if locks saved */
2962 LASSERT(req->rq_export != NULL);
2963 /* req/reply consistent */
2964 LASSERT(rs->rs_svcpt == svcpt);
2967 LASSERT(!rs->rs_scheduled);
2968 LASSERT(!rs->rs_scheduled_ever);
2969 LASSERT(!rs->rs_handled);
2970 LASSERT(!rs->rs_on_net);
2971 LASSERT(rs->rs_export == NULL);
2972 LASSERT(list_empty(&rs->rs_obd_list));
2973 LASSERT(list_empty(&rs->rs_exp_list));
2975 exp = class_export_get(req->rq_export);
2977 /* disable reply scheduling while I'm setting up */
2978 rs->rs_scheduled = 1;
2980 rs->rs_xid = req->rq_xid;
2981 rs->rs_transno = req->rq_transno;
2982 rs->rs_export = exp;
2983 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2985 spin_lock(&exp->exp_uncommitted_replies_lock);
2986 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2987 rs->rs_transno, exp->exp_last_committed);
2988 if (rs->rs_transno > exp->exp_last_committed) {
2989 /* not committed already */
2990 list_add_tail(&rs->rs_obd_list,
2991 &exp->exp_uncommitted_replies);
2993 spin_unlock(&exp->exp_uncommitted_replies_lock);
2995 spin_lock(&exp->exp_lock);
2996 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2997 spin_unlock(&exp->exp_lock);
2999 netrc = target_send_reply_msg(req, rc, fail_id);
3001 spin_lock(&svcpt->scp_rep_lock);
3003 atomic_inc(&svcpt->scp_nreps_difficult);
3006 /* error sending: reply is off the net. Also we need +1
3007 * reply ref until ptlrpc_handle_rs() is done
3008 * with the reply state (if the send was successful, there
3009 * would have been +1 ref for the net, which
3010 * reply_out_callback leaves alone) */
3012 ptlrpc_rs_addref(rs);
3015 spin_lock(&rs->rs_lock);
3016 if (rs->rs_transno <= exp->exp_last_committed ||
3017 (!rs->rs_on_net && !rs->rs_no_ack) ||
3018 list_empty(&rs->rs_exp_list) || /* completed already */
3019 list_empty(&rs->rs_obd_list)) {
3020 CDEBUG(D_HA, "Schedule reply immediately\n");
3021 ptlrpc_dispatch_difficult_reply(rs);
3023 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3024 rs->rs_scheduled = 0; /* allow notifier to schedule */
3026 spin_unlock(&rs->rs_lock);
3027 spin_unlock(&svcpt->scp_rep_lock);
3031 enum ldlm_mode lck_compat_array[] = {
3032 [LCK_EX] = LCK_COMPAT_EX,
3033 [LCK_PW] = LCK_COMPAT_PW,
3034 [LCK_PR] = LCK_COMPAT_PR,
3035 [LCK_CW] = LCK_COMPAT_CW,
3036 [LCK_CR] = LCK_COMPAT_CR,
3037 [LCK_NL] = LCK_COMPAT_NL,
3038 [LCK_GROUP] = LCK_COMPAT_GROUP,
3039 [LCK_COS] = LCK_COMPAT_COS,
3043 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3044 * not escape to the user level.
3046 int ldlm_error2errno(enum ldlm_error error)
3052 case ELDLM_LOCK_MATCHED:
3055 case ELDLM_LOCK_CHANGED:
3058 case ELDLM_LOCK_ABORTED:
3061 case ELDLM_LOCK_REPLACED:
3064 case ELDLM_NO_LOCK_DATA:
3067 case ELDLM_NAMESPACE_EXISTS:
3070 case ELDLM_BAD_NAMESPACE:
3074 if (((int)error) < 0) { /* cast to signed type */
3075 result = error; /* as ldlm_error can be unsigned */
3077 CERROR("Invalid DLM result code: %d\n", error);
3083 EXPORT_SYMBOL(ldlm_error2errno);
3086 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3088 enum ldlm_error ldlm_errno2error(int err_no)
3097 error = ELDLM_LOCK_CHANGED;
3100 error = ELDLM_LOCK_ABORTED;
3103 error = ELDLM_LOCK_REPLACED;
3106 error = ELDLM_NO_LOCK_DATA;
3109 error = ELDLM_NAMESPACE_EXISTS;
3112 error = ELDLM_BAD_NAMESPACE;
3120 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3121 void ldlm_dump_export_locks(struct obd_export *exp)
3123 spin_lock(&exp->exp_locks_list_guard);
3124 if (!list_empty(&exp->exp_locks_list)) {
3125 struct ldlm_lock *lock;
3127 CERROR("dumping locks for export %p,"
3128 "ignore if the unmount doesn't hang\n", exp);
3129 list_for_each_entry(lock, &exp->exp_locks_list,
3131 LDLM_ERROR(lock, "lock:");
3133 spin_unlock(&exp->exp_locks_list_guard);
3137 #ifdef HAVE_SERVER_SUPPORT
3138 static int target_bulk_timeout(void *data)
3141 /* We don't fail the connection here, because having the export
3142 * killed makes the (vital) call to commitrw very sad.
3147 static inline const char *bulk2type(struct ptlrpc_request *req)
3149 if (req->rq_bulk_read)
3151 if (req->rq_bulk_write)
3156 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3157 struct l_wait_info *lwi)
3159 struct ptlrpc_request *req = desc->bd_req;
3160 time_t start = cfs_time_current_sec();
3166 /* If there is eviction in progress, wait for it to finish. */
3167 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3168 *lwi = LWI_INTR(NULL, NULL);
3169 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3170 !atomic_read(&exp->exp_obd->
3171 obd_evict_inprogress),
3175 /* Check if client was evicted or reconnected already. */
3176 if (exp->exp_failed ||
3177 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3180 if (req->rq_bulk_read)
3181 rc = sptlrpc_svc_wrap_bulk(req, desc);
3183 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3184 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3185 else /* old version, bulk matchbits is rq_xid */
3186 req->rq_mbits = req->rq_xid;
3189 rc = ptlrpc_start_bulk_transfer(desc);
3193 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3194 bulk2type(req), rc);
3198 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3199 ptlrpc_abort_bulk(desc);
3203 /* limit actual bulk transfer to bulk_timeout seconds */
3204 deadline = start + bulk_timeout;
3205 if (deadline > req->rq_deadline)
3206 deadline = req->rq_deadline;
3209 long timeoutl = deadline - cfs_time_current_sec();
3210 cfs_duration_t timeout = timeoutl <= 0 ?
3211 CFS_TICK : cfs_time_seconds(timeoutl);
3214 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3215 target_bulk_timeout, desc);
3216 rc = l_wait_event(desc->bd_waitq,
3217 !ptlrpc_server_bulk_active(desc) ||
3220 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3222 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3223 /* Wait again if we changed rq_deadline. */
3224 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3225 deadline = start + bulk_timeout;
3226 if (deadline > rq_deadline)
3227 deadline = rq_deadline;
3228 } while ((rc == -ETIMEDOUT) &&
3229 (deadline > cfs_time_current_sec()));
3231 if (rc == -ETIMEDOUT) {
3232 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3233 bulk2type(req), deadline - start,
3234 cfs_time_current_sec() - deadline);
3235 ptlrpc_abort_bulk(desc);
3236 } else if (exp->exp_failed) {
3237 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3240 ptlrpc_abort_bulk(desc);
3241 } else if (exp->exp_conn_cnt >
3242 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3243 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3245 /* We don't reply anyway. */
3247 ptlrpc_abort_bulk(desc);
3248 } else if (desc->bd_failure) {
3249 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3251 /* XXX should this be a different errno? */
3254 if (req->rq_bulk_write)
3255 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3256 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3257 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3258 bulk2type(req), desc->bd_nob_transferred,
3260 /* XXX should this be a different errno? */
3267 EXPORT_SYMBOL(target_bulk_io);
3269 #endif /* HAVE_SERVER_SUPPORT */