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);
353 cli->cl_conn_count = 0;
354 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
355 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
356 sizeof(server_uuid)));
358 cli->cl_dirty_pages = 0;
359 cli->cl_avail_grant = 0;
360 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
361 /* cl_dirty_max_pages may be changed at connect time in
362 * ptlrpc_connect_interpret(). */
363 client_adjust_max_dirty(cli);
364 INIT_LIST_HEAD(&cli->cl_cache_waiters);
365 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
366 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
367 INIT_LIST_HEAD(&cli->cl_loi_write_list);
368 INIT_LIST_HEAD(&cli->cl_loi_read_list);
369 spin_lock_init(&cli->cl_loi_list_lock);
370 atomic_set(&cli->cl_pending_w_pages, 0);
371 atomic_set(&cli->cl_pending_r_pages, 0);
372 cli->cl_r_in_flight = 0;
373 cli->cl_w_in_flight = 0;
375 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
376 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
377 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
378 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
379 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
380 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
383 INIT_LIST_HEAD(&cli->cl_lru_osc);
384 atomic_set(&cli->cl_lru_shrinkers, 0);
385 atomic_long_set(&cli->cl_lru_busy, 0);
386 atomic_long_set(&cli->cl_lru_in_list, 0);
387 INIT_LIST_HEAD(&cli->cl_lru_list);
388 spin_lock_init(&cli->cl_lru_list_lock);
389 atomic_long_set(&cli->cl_unstable_count, 0);
390 INIT_LIST_HEAD(&cli->cl_shrink_list);
392 init_waitqueue_head(&cli->cl_destroy_waitq);
393 atomic_set(&cli->cl_destroy_in_flight, 0);
394 #ifdef ENABLE_CHECKSUM
395 /* Turn on checksumming by default. */
396 cli->cl_checksum = 1;
398 * The supported checksum types will be worked out at connect time
399 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
402 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
404 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
406 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
407 * from OFD after connecting. */
408 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
410 /* set cl_chunkbits default value to PAGE_SHIFT,
411 * it will be updated at OSC connection time. */
412 cli->cl_chunkbits = PAGE_SHIFT;
414 if (!strcmp(name, LUSTRE_MDC_NAME)) {
415 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
416 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
417 cli->cl_max_rpcs_in_flight = 2;
418 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
419 cli->cl_max_rpcs_in_flight = 3;
420 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
421 cli->cl_max_rpcs_in_flight = 4;
423 if (osc_on_mdt(obddev->obd_name))
424 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
426 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
429 spin_lock_init(&cli->cl_mod_rpcs_lock);
430 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
431 cli->cl_max_mod_rpcs_in_flight = 0;
432 cli->cl_mod_rpcs_in_flight = 0;
433 cli->cl_close_rpcs_in_flight = 0;
434 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
435 cli->cl_mod_tag_bitmap = NULL;
437 if (connect_op == MDS_CONNECT) {
438 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
439 OBD_ALLOC(cli->cl_mod_tag_bitmap,
440 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
441 if (cli->cl_mod_tag_bitmap == NULL)
442 GOTO(err, rc = -ENOMEM);
447 CERROR("ldlm_get_ref failed: %d\n", rc);
451 ptlrpc_init_client(rq_portal, rp_portal, name,
452 &obddev->obd_ldlm_client);
454 imp = class_new_import(obddev);
456 GOTO(err_ldlm, rc = -ENOENT);
457 imp->imp_client = &obddev->obd_ldlm_client;
458 imp->imp_connect_op = connect_op;
459 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
460 LUSTRE_CFG_BUFLEN(lcfg, 1));
461 class_import_put(imp);
463 if (lustre_cfg_buf(lcfg, 4)) {
464 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
466 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
468 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
469 obddev->obd_name, lustre_cfg_string(lcfg, 4),
471 GOTO(err_import, rc);
473 fake_conn.c_self = LNET_MKNID(refnet, 0);
474 imp->imp_connection = &fake_conn;
477 rc = client_import_add_conn(imp, &server_uuid, 1);
479 CERROR("can't add initial connection\n");
480 GOTO(err_import, rc);
482 imp->imp_connection = NULL;
484 cli->cl_import = imp;
485 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
486 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
488 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
489 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
490 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
491 name, obddev->obd_name,
492 cli->cl_target_uuid.uuid);
493 spin_lock(&imp->imp_lock);
494 imp->imp_deactive = 1;
495 spin_unlock(&imp->imp_lock);
499 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
500 LDLM_NAMESPACE_CLIENT,
501 LDLM_NAMESPACE_GREEDY,
503 if (obddev->obd_namespace == NULL) {
504 CERROR("Unable to create client namespace - %s\n",
506 GOTO(err_import, rc = -ENOMEM);
512 class_destroy_import(imp);
516 if (cli->cl_mod_tag_bitmap != NULL)
517 OBD_FREE(cli->cl_mod_tag_bitmap,
518 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
519 cli->cl_mod_tag_bitmap = NULL;
523 EXPORT_SYMBOL(client_obd_setup);
525 int client_obd_cleanup(struct obd_device *obddev)
527 struct client_obd *cli = &obddev->u.cli;
530 ldlm_namespace_free_post(obddev->obd_namespace);
531 obddev->obd_namespace = NULL;
533 obd_cleanup_client_import(obddev);
534 LASSERT(obddev->u.cli.cl_import == NULL);
538 if (cli->cl_mod_tag_bitmap != NULL)
539 OBD_FREE(cli->cl_mod_tag_bitmap,
540 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
541 cli->cl_mod_tag_bitmap = NULL;
545 EXPORT_SYMBOL(client_obd_cleanup);
547 /* ->o_connect() method for client side (OSC and MDC and MGC) */
548 int client_connect_import(const struct lu_env *env,
549 struct obd_export **exp,
550 struct obd_device *obd, struct obd_uuid *cluuid,
551 struct obd_connect_data *data, void *localdata)
553 struct client_obd *cli = &obd->u.cli;
554 struct obd_import *imp = cli->cl_import;
555 struct obd_connect_data *ocd;
556 struct lustre_handle conn = { 0 };
561 down_write(&cli->cl_sem);
562 if (cli->cl_conn_count > 0)
563 GOTO(out_sem, rc = -EALREADY);
565 rc = class_connect(&conn, obd, cluuid);
569 cli->cl_conn_count++;
570 *exp = class_conn2export(&conn);
572 LASSERT(obd->obd_namespace);
574 imp->imp_dlm_handle = conn;
575 rc = ptlrpc_init_import(imp);
579 ocd = &imp->imp_connect_data;
582 imp->imp_connect_flags_orig = data->ocd_connect_flags;
583 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
586 rc = ptlrpc_connect_import(imp);
588 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
591 LASSERT(*exp != NULL && (*exp)->exp_connection);
594 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
595 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
596 data->ocd_connect_flags, ocd->ocd_connect_flags);
597 data->ocd_connect_flags = ocd->ocd_connect_flags;
600 ptlrpc_pinger_add_import(imp);
606 cli->cl_conn_count--;
607 class_disconnect(*exp);
611 up_write(&cli->cl_sem);
615 EXPORT_SYMBOL(client_connect_import);
617 int client_disconnect_export(struct obd_export *exp)
619 struct obd_device *obd = class_exp2obd(exp);
620 struct client_obd *cli;
621 struct obd_import *imp;
626 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
627 exp, exp ? exp->exp_handle.h_cookie : -1);
632 imp = cli->cl_import;
634 down_write(&cli->cl_sem);
635 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
638 if (cli->cl_conn_count == 0) {
639 CERROR("disconnecting disconnected device (%s)\n",
641 GOTO(out_disconnect, rc = -EINVAL);
644 cli->cl_conn_count--;
645 if (cli->cl_conn_count != 0)
646 GOTO(out_disconnect, rc = 0);
648 /* Mark import deactivated now, so we don't try to reconnect if any
649 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
650 * fully deactivate the import, or that would drop all requests. */
651 spin_lock(&imp->imp_lock);
652 imp->imp_deactive = 1;
653 spin_unlock(&imp->imp_lock);
655 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
656 * delete it regardless. (It's safe to delete an import that was
658 (void)ptlrpc_pinger_del_import(imp);
660 if (obd->obd_namespace != NULL) {
661 /* obd_force == local only */
662 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
663 obd->obd_force ? LCF_LOCAL : 0, NULL);
664 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
667 /* There's no need to hold sem while disconnecting an import,
668 * and it may actually cause deadlock in GSS. */
669 up_write(&cli->cl_sem);
670 rc = ptlrpc_disconnect_import(imp, 0);
671 down_write(&cli->cl_sem);
673 ptlrpc_invalidate_import(imp);
678 /* Use server style - class_disconnect should be always called for
680 err = class_disconnect(exp);
684 up_write(&cli->cl_sem);
688 EXPORT_SYMBOL(client_disconnect_export);
690 #ifdef HAVE_SERVER_SUPPORT
691 int server_disconnect_export(struct obd_export *exp)
696 /* Disconnect early so that clients can't keep using export. */
697 rc = class_disconnect(exp);
698 /* Close import to avoid sending any requests. */
699 if (exp->exp_imp_reverse)
700 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
702 ldlm_bl_thread_wakeup();
704 /* complete all outstanding replies */
705 spin_lock(&exp->exp_lock);
706 while (!list_empty(&exp->exp_outstanding_replies)) {
707 struct ptlrpc_reply_state *rs =
708 list_entry(exp->exp_outstanding_replies.next,
709 struct ptlrpc_reply_state, rs_exp_list);
710 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
712 spin_lock(&svcpt->scp_rep_lock);
714 list_del_init(&rs->rs_exp_list);
715 spin_lock(&rs->rs_lock);
716 ptlrpc_schedule_difficult_reply(rs);
717 spin_unlock(&rs->rs_lock);
719 spin_unlock(&svcpt->scp_rep_lock);
721 spin_unlock(&exp->exp_lock);
725 EXPORT_SYMBOL(server_disconnect_export);
727 /* --------------------------------------------------------------------------
728 * from old lib/target.c
729 * -------------------------------------------------------------------------- */
731 static int target_handle_reconnect(struct lustre_handle *conn,
732 struct obd_export *exp,
733 struct obd_uuid *cluuid)
735 struct obd_device *target;
736 struct lustre_handle *hdl;
743 hdl = &exp->exp_imp_reverse->imp_remote_handle;
744 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
745 conn->cookie = exp->exp_handle.h_cookie;
746 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
747 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
751 target = exp->exp_obd;
753 /* Might be a re-connect after a partition. */
754 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
755 LCONSOLE_WARN("%s: already connected client %s (at %s) "
756 "with handle %#llx. Rejecting client "
757 "with the same UUID trying to reconnect "
758 "with handle %#llx\n", target->obd_name,
759 obd_uuid2str(&exp->exp_client_uuid),
760 obd_export_nid2str(exp),
761 hdl->cookie, conn->cookie);
762 memset(conn, 0, sizeof *conn);
763 /* target_handle_connect() treats EALREADY and
764 * -EALREADY differently. -EALREADY is an error
765 * (same UUID, different handle). */
769 if (!target->obd_recovering) {
770 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
771 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
772 obd_export_nid2str(exp));
773 GOTO(out_already, rc);
776 now = cfs_time_current();
777 deadline = target->obd_recovery_timer.expires;
778 if (cfs_time_before(now, deadline)) {
779 struct target_distribute_txn_data *tdtd =
780 class_exp2tgt(exp)->lut_tdtd;
785 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
786 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
787 buf = tdtd->tdtd_show_update_logs_retrievers(
788 tdtd->tdtd_show_retrievers_cbdata,
792 LCONSOLE_WARN("%s: Recovery already passed deadline "
793 "%d:%.02d. It is due to DNE recovery "
794 "failed/stuck on the %d MDT(s):%s. "
795 "Please wait until all MDTs recovered "
796 "or abort the recovery by force.\n",
797 target->obd_name, timeout / 60,
799 buf ? buf : "unknown (not enough RAM)");
801 LCONSOLE_WARN("%s: Recovery already passed deadline "
802 "%d:%.02d. If you do not want to wait "
803 "more, please abort the recovery by "
804 "force.\n", target->obd_name,
805 timeout / 60, timeout % 60);
810 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
811 LCONSOLE_WARN("%s: Recovery already passed deadline"
812 " %d:%.02d, It is most likely due to DNE"
813 " recovery is failed or stuck, please wait a"
814 " few more minutes or abort the recovery.\n",
815 target->obd_name, timeout / 60, timeout % 60);
819 conn->cookie = exp->exp_handle.h_cookie;
820 /* target_handle_connect() treats EALREADY and
821 * -EALREADY differently. EALREADY means we are
822 * doing a valid reconnect from the same client. */
827 check_and_start_recovery_timer(struct obd_device *obd,
828 struct ptlrpc_request *req, int new_client);
831 * update flags for import during reconnect process
833 static int rev_import_flags_update(struct obd_import *revimp,
834 struct ptlrpc_request *req)
837 struct obd_connect_data *data;
839 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
841 if (data->ocd_connect_flags & OBD_CONNECT_AT)
842 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
844 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
846 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
848 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
850 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
851 revimp->imp_client->cli_name,
852 libcfs_id2str(req->rq_peer), rc);
860 * Allocate a new reverse import for an export.
862 * \retval -errno in case error hit
863 * \retval 0 if reverse import correctly init
865 int rev_import_init(struct obd_export *export)
867 struct obd_device *obd = export->exp_obd;
868 struct obd_import *revimp;
870 LASSERT(export->exp_imp_reverse == NULL);
872 revimp = class_new_import(obd);
876 revimp->imp_remote_handle.cookie = 0ULL;
877 revimp->imp_client = &obd->obd_ldlm_client;
878 revimp->imp_dlm_fake = 1;
880 /* it is safe to connect import in new state as no sends possible */
881 spin_lock(&export->exp_lock);
882 export->exp_imp_reverse = revimp;
883 spin_unlock(&export->exp_lock);
884 class_import_put(revimp);
888 EXPORT_SYMBOL(rev_import_init);
891 * Handle reconnect for an export.
893 * \param exp export to handle reconnect process
894 * \param req client reconnect request
896 * \retval -rc in case securitfy flavor can't be changed
897 * \retval 0 in case none problems
899 static int rev_import_reconnect(struct obd_export *exp,
900 struct ptlrpc_request *req)
902 struct obd_import *revimp = exp->exp_imp_reverse;
903 struct lustre_handle *lh;
906 /* avoid sending a request until import flags are changed */
907 ptlrpc_import_enter_resend(revimp);
909 if (revimp->imp_connection != NULL)
910 ptlrpc_connection_put(revimp->imp_connection);
913 * client from recovery don't have a handle so we need to take from
914 * request. it may produce situation when wrong client connected
915 * to recovery as we trust a client uuid
917 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
918 revimp->imp_remote_handle = *lh;
920 /* unknown versions will be caught in
921 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
922 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
924 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
926 rc = rev_import_flags_update(revimp, req);
928 /* it is safe to still be in RECOVERY phase as we are not able
929 * to setup correct security flavor so requests are not able to
930 * be delivered correctly */
934 /* resend all rpc's via new connection */
935 return ptlrpc_import_recovery_state_machine(revimp);
938 int target_handle_connect(struct ptlrpc_request *req)
940 struct obd_device *target = NULL;
941 struct obd_export *export = NULL;
942 /* connect handle - filled from target_handle_reconnect in
944 struct lustre_handle conn;
945 struct lustre_handle *tmp;
946 struct obd_uuid tgtuuid;
947 struct obd_uuid cluuid;
952 bool mds_conn = false, lw_client = false, initial_conn = false;
953 bool mds_mds_conn = false;
954 bool new_mds_mds_conn = false;
955 bool target_referenced = false;
956 struct obd_connect_data *data, *tmpdata;
958 lnet_nid_t *client_nid = NULL;
961 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
963 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
965 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
966 GOTO(out, rc = -EINVAL);
969 obd_str2uuid(&tgtuuid, str);
970 target = class_uuid2obd(&tgtuuid);
972 target = class_name2obd(str);
975 deuuidify(str, NULL, &target_start, &target_len);
976 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
977 "from %s (no target). If you are running "
978 "an HA pair check that the target is "
979 "mounted on the other server.\n", str,
980 libcfs_nid2str(req->rq_peer.nid));
981 GOTO(out, rc = -ENODEV);
984 spin_lock(&target->obd_dev_lock);
985 if (target->obd_stopping || !target->obd_set_up) {
986 spin_unlock(&target->obd_dev_lock);
988 deuuidify(str, NULL, &target_start, &target_len);
989 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
990 target_len, target_start,
991 libcfs_nid2str(req->rq_peer.nid),
992 (target->obd_stopping ?
993 "stopping" : "not set up"));
994 GOTO(out, rc = -ENODEV);
997 if (target->obd_no_conn) {
998 spin_unlock(&target->obd_dev_lock);
1000 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1001 "from %s\n", target->obd_name,
1002 libcfs_nid2str(req->rq_peer.nid));
1003 GOTO(out, rc = -EAGAIN);
1006 /* Make sure the target isn't cleaned up while we're here. Yes,
1007 * there's still a race between the above check and our incref here.
1008 * Really, class_uuid2obd should take the ref. */
1009 class_incref(target, __func__, current);
1010 target_referenced = true;
1012 target->obd_conn_inprogress++;
1013 spin_unlock(&target->obd_dev_lock);
1015 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1017 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1018 GOTO(out, rc = -EINVAL);
1021 obd_str2uuid(&cluuid, str);
1023 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1025 GOTO(out, rc = -EPROTO);
1029 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1031 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1033 GOTO(out, rc = -EPROTO);
1035 rc = req_capsule_server_pack(&req->rq_pill);
1039 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1040 /* Don't allow clients to connect that are using old 1.8 format
1041 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1042 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1043 * FULL20 flag should be set on all connections since 2.0, but no
1044 * longer affects behaviour.
1046 * Later this check will be disabled and the flag can be retired
1047 * completely once interop with 3.0 is no longer needed.
1049 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1050 GOTO(out, rc = -EPROTO);
1053 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1054 if (data->ocd_version < LUSTRE_VERSION_CODE -
1055 LUSTRE_VERSION_ALLOWED_OFFSET ||
1056 data->ocd_version > LUSTRE_VERSION_CODE +
1057 LUSTRE_VERSION_ALLOWED_OFFSET) {
1058 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1059 "libclient connection attempt",
1060 data->ocd_version < LUSTRE_VERSION_CODE ?
1062 OBD_OCD_VERSION_MAJOR(data->ocd_version),
1063 OBD_OCD_VERSION_MINOR(data->ocd_version),
1064 OBD_OCD_VERSION_PATCH(data->ocd_version),
1065 OBD_OCD_VERSION_FIX(data->ocd_version));
1066 data = req_capsule_server_sized_get(&req->rq_pill,
1068 offsetof(typeof(*data), ocd_version) +
1069 sizeof(data->ocd_version));
1071 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1072 data->ocd_version = LUSTRE_VERSION_CODE;
1074 GOTO(out, rc = -EPROTO);
1078 /* Note: lw_client is needed in MDS-MDS failover during update log
1079 * processing, so we needs to allow lw_client to be connected at
1080 * anytime, instead of only the initial connection */
1081 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1083 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1084 initial_conn = true;
1085 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1086 mds_mds_conn = (data->ocd_connect_flags &
1087 OBD_CONNECT_MDS_MDS) != 0;
1089 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1090 * for Imperative Recovery connection from MGC to MGS.
1092 * Via check OBD_CONNECT_FID, we can distinguish whether
1093 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1096 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1097 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1098 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1099 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1100 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1101 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1103 /* We do not support the MDT-MDT interoperations with
1104 * different version MDT because of protocol changes. */
1105 if (unlikely(major != LUSTRE_MAJOR ||
1106 minor != LUSTRE_MINOR ||
1107 abs(patch - LUSTRE_PATCH) > 3)) {
1108 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1109 "connection from different version MDT "
1110 "(%d.%d.%d.%d) %s %s\n",
1111 target->obd_name, LUSTRE_MAJOR,
1112 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1113 major, minor, patch,
1114 OBD_OCD_VERSION_FIX(data->ocd_version),
1115 libcfs_nid2str(req->rq_peer.nid), str);
1117 GOTO(out, rc = -EPROTO);
1122 /* lctl gets a backstage, all-access pass. */
1123 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1124 goto dont_check_exports;
1126 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1130 /* We've found an export in the hash. */
1132 spin_lock(&export->exp_lock);
1134 if (export->exp_connecting) { /* bug 9635, et. al. */
1135 spin_unlock(&export->exp_lock);
1136 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1137 export->exp_obd->obd_name, export,
1138 libcfs_nid2str(req->rq_peer.nid));
1139 class_export_put(export);
1142 } else if ((mds_conn || (lw_client && initial_conn) ||
1143 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1144 export->exp_connection != NULL) {
1145 spin_unlock(&export->exp_lock);
1146 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1147 /* MDS or LWP reconnected after failover. */
1148 LCONSOLE_WARN("%s: Received %s connection from "
1149 "%s, removing former export from %s\n",
1150 target->obd_name, mds_conn ? "MDS" : "LWP",
1151 libcfs_nid2str(req->rq_peer.nid),
1152 libcfs_nid2str(export->exp_connection->c_peer.nid));
1154 /* New MDS connection from the same NID. */
1155 LCONSOLE_WARN("%s: Received new %s connection from "
1156 "%s, removing former export from same NID\n",
1157 target->obd_name, mds_conn ? "MDS" : "LWP",
1158 libcfs_nid2str(req->rq_peer.nid));
1161 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1162 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1163 /* Because exports between MDTs will always be
1164 * kept, let's do not fail such export if they
1165 * come from the same NID, otherwise it might
1166 * cause eviction between MDTs, which might
1167 * cause namespace inconsistency */
1168 spin_lock(&export->exp_lock);
1169 export->exp_connecting = 1;
1170 spin_unlock(&export->exp_lock);
1171 conn.cookie = export->exp_handle.h_cookie;
1174 class_fail_export(export);
1175 class_export_put(export);
1179 } else if (export->exp_connection != NULL && initial_conn &&
1180 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1181 spin_unlock(&export->exp_lock);
1182 /* In MDS failover we have static UUID but NID can change. */
1183 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1184 "existing nid %s is already connected\n",
1185 target->obd_name, cluuid.uuid,
1186 libcfs_nid2str(req->rq_peer.nid),
1188 export->exp_connection->c_peer.nid));
1190 class_export_put(export);
1193 export->exp_connecting = 1;
1194 spin_unlock(&export->exp_lock);
1195 LASSERT(export->exp_obd == target);
1197 rc = target_handle_reconnect(&conn, export, &cluuid);
1200 /* If we found an export, we already unlocked. */
1203 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1204 } else if (req->rq_export == NULL &&
1205 atomic_read(&export->exp_rpc_count) > 0) {
1206 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1207 "still busy with %d references\n",
1208 target->obd_name, cluuid.uuid,
1209 libcfs_nid2str(req->rq_peer.nid),
1210 atomic_read(&export->exp_refcount));
1211 GOTO(out, rc = -EBUSY);
1212 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1213 if (!strstr(cluuid.uuid, "mdt"))
1214 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1215 "known client %s (at %s) because it "
1216 "is indicating it is a new client",
1217 target->obd_name, cluuid.uuid,
1218 libcfs_nid2str(req->rq_peer.nid));
1219 GOTO(out, rc = -EALREADY);
1221 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1228 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1229 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1230 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1231 export, (long)cfs_time_current_sec(),
1232 export ? (long)export->exp_last_request_time : 0);
1234 /* If this is the first time a client connects, reset the recovery
1235 * timer. Discard lightweight connections which might be local. */
1236 if (!lw_client && rc == 0 && target->obd_recovering)
1237 check_and_start_recovery_timer(target, req, export == NULL);
1239 /* We want to handle EALREADY but *not* -EALREADY from
1240 * target_handle_reconnect(), return reconnection state in a flag. */
1241 if (rc == EALREADY) {
1242 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1248 /* Tell the client if we support replayable requests. */
1249 if (target->obd_replayable)
1250 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1251 client_nid = &req->rq_peer.nid;
1253 if (export == NULL) {
1254 /* allow lightweight connections during recovery */
1255 /* allow "new" MDT to be connected during recovery, since we
1256 * need retrieve recovery update records from it */
1257 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1259 int c; /* connected */
1260 int i; /* in progress */
1262 int s; /* stale/evicted */
1264 c = atomic_read(&target->obd_connected_clients);
1265 i = atomic_read(&target->obd_lock_replay_clients);
1266 k = target->obd_max_recoverable_clients;
1267 s = target->obd_stale_clients;
1268 t = target->obd_recovery_timer.expires;
1269 t = cfs_time_sub(t, cfs_time_current());
1270 t = cfs_duration_sec(t);
1271 LCONSOLE_WARN("%s: Denying connection for new client %s"
1272 "(at %s), waiting for %d known clients "
1273 "(%d recovered, %d in progress, and %d "
1274 "evicted) to recover in %d:%.02d\n",
1275 target->obd_name, cluuid.uuid,
1276 libcfs_nid2str(req->rq_peer.nid), k,
1277 c - i, i, s, (int)t / 60,
1282 rc = obd_connect(req->rq_svc_thread->t_env,
1283 &export, target, &cluuid, data,
1285 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1286 lustre_msg_add_op_flags(req->rq_repmsg,
1287 MSG_CONNECT_RECOVERING);
1289 conn.cookie = export->exp_handle.h_cookie;
1290 rc = rev_import_init(export);
1294 new_mds_mds_conn = true;
1297 rc = obd_reconnect(req->rq_svc_thread->t_env,
1298 export, target, &cluuid, data, client_nid);
1303 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1304 data->ocd_instance = target->u.obt.obt_instance;
1306 /* Return only the parts of obd_connect_data that we understand, so the
1307 * client knows that we don't understand the rest. */
1309 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1311 tmpdata = req_capsule_server_get(&req->rq_pill,
1313 /* Don't use struct assignment here, because the client reply
1314 * buffer may be smaller/larger than the local struct
1315 * obd_connect_data. */
1316 memcpy(tmpdata, data, min(tmpsize, size));
1319 /* If the client and the server are the same node, we will already
1320 * have an export that really points to the client's DLM export,
1321 * because we have a shared handles table.
1323 * XXX this will go away when shaver stops sending the "connect" handle
1324 * in the real "remote handle" field of the request --phik 24 Apr 2003
1326 ptlrpc_request_change_export(req, export);
1328 spin_lock(&export->exp_lock);
1329 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1330 spin_unlock(&export->exp_lock);
1331 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1332 "or equal conn_cnt: %d >= %d\n",
1333 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1334 export->exp_conn_cnt,
1335 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1337 GOTO(out, rc = -EALREADY);
1339 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1340 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1342 /* Don't evict liblustre clients for not pinging. */
1343 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1344 export->exp_libclient = 1;
1345 spin_unlock(&export->exp_lock);
1347 spin_lock(&target->obd_dev_lock);
1348 list_del_init(&export->exp_obd_chain_timed);
1349 spin_unlock(&target->obd_dev_lock);
1351 spin_unlock(&export->exp_lock);
1354 if (export->exp_connection != NULL) {
1355 /* Check to see if connection came from another NID. */
1356 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1357 !hlist_unhashed(&export->exp_nid_hash))
1358 cfs_hash_del(export->exp_obd->obd_nid_hash,
1359 &export->exp_connection->c_peer.nid,
1360 &export->exp_nid_hash);
1362 ptlrpc_connection_put(export->exp_connection);
1365 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1368 if (hlist_unhashed(&export->exp_nid_hash))
1369 cfs_hash_add(export->exp_obd->obd_nid_hash,
1370 &export->exp_connection->c_peer.nid,
1371 &export->exp_nid_hash);
1373 lustre_msg_set_handle(req->rq_repmsg, &conn);
1375 rc = rev_import_reconnect(export, req);
1379 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1381 __u64 transno = data->ocd_transno;
1383 spin_lock(&export->exp_lock);
1384 /* possible race with class_disconnect_stale_exports,
1385 * export may be already in the eviction process */
1386 if (export->exp_failed) {
1387 spin_unlock(&export->exp_lock);
1388 GOTO(out, rc = -ENODEV);
1390 export->exp_in_recovery = 1;
1391 export->exp_req_replay_needed = 1;
1392 export->exp_lock_replay_needed = 1;
1393 spin_unlock(&export->exp_lock);
1395 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1396 MSG_CONNECT_TRANSNO);
1397 if (has_transno && transno == 0)
1398 CWARN("Connect with zero transno!\n");
1400 if (has_transno && transno > 0 &&
1401 transno < target->obd_next_recovery_transno &&
1402 transno > target->obd_last_committed) {
1403 /* Another way is to use cmpxchg() to be lock-free. */
1404 spin_lock(&target->obd_recovery_task_lock);
1405 if (transno < target->obd_next_recovery_transno)
1406 target->obd_next_recovery_transno = transno;
1407 spin_unlock(&target->obd_recovery_task_lock);
1410 atomic_inc(&target->obd_req_replay_clients);
1411 atomic_inc(&target->obd_lock_replay_clients);
1412 /* Note: MDS-MDS connection is allowed to be connected during
1413 * recovery, no matter if the exports needs to be recoveried.
1414 * Because we need retrieve updates logs from all other MDTs.
1415 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1416 * also needs to be increased to match other recovery checking
1418 if (new_mds_mds_conn)
1419 target->obd_max_recoverable_clients++;
1420 if (atomic_inc_return(&target->obd_connected_clients) ==
1421 target->obd_max_recoverable_clients)
1422 wake_up(&target->obd_next_transno_waitq);
1425 /* Tell the client we're in recovery, when client is involved in it. */
1426 if (target->obd_recovering && !lw_client)
1427 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1431 spin_lock(&export->exp_lock);
1432 export->exp_connecting = 0;
1433 spin_unlock(&export->exp_lock);
1435 class_export_put(export);
1437 if (target_referenced == true && target != NULL) {
1438 spin_lock(&target->obd_dev_lock);
1439 target->obd_conn_inprogress--;
1440 spin_unlock(&target->obd_dev_lock);
1442 class_decref(target, __func__, current);
1444 req->rq_status = rc;
1448 int target_handle_disconnect(struct ptlrpc_request *req)
1453 rc = req_capsule_server_pack(&req->rq_pill);
1457 /* Keep the rq_export around so we can send the reply. */
1458 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1463 void target_destroy_export(struct obd_export *exp)
1465 struct obd_import *imp = NULL;
1466 /* exports created from last_rcvd data, and "fake"
1467 exports created by lctl don't have an import */
1468 spin_lock(&exp->exp_lock);
1469 if (exp->exp_imp_reverse != NULL) {
1470 imp = exp->exp_imp_reverse;
1471 exp->exp_imp_reverse = NULL;
1473 spin_unlock(&exp->exp_lock);
1475 client_destroy_import(imp);
1477 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1478 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1479 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1480 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1482 EXPORT_SYMBOL(target_destroy_export);
1485 * Recovery functions
1487 static void target_request_copy_get(struct ptlrpc_request *req)
1489 class_export_rpc_inc(req->rq_export);
1490 LASSERT(list_empty(&req->rq_list));
1491 INIT_LIST_HEAD(&req->rq_replay_list);
1493 /* Increase refcount to keep request in queue. */
1494 atomic_inc(&req->rq_refcount);
1495 /* Let export know it has replays to be handled. */
1496 atomic_inc(&req->rq_export->exp_replay_count);
1499 static void target_request_copy_put(struct ptlrpc_request *req)
1501 LASSERT(list_empty(&req->rq_replay_list));
1502 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1504 atomic_dec(&req->rq_export->exp_replay_count);
1505 class_export_rpc_dec(req->rq_export);
1506 ptlrpc_server_drop_request(req);
1509 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1511 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1512 struct obd_export *exp = req->rq_export;
1513 struct ptlrpc_request *reqiter;
1514 struct ptlrpc_request *dup_req = NULL;
1519 spin_lock(&exp->exp_lock);
1520 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1522 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1530 /* We expect it with RESENT and REPLAY flags. */
1531 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1532 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1533 CERROR("invalid flags %x of resent replay\n",
1534 lustre_msg_get_flags(req->rq_reqmsg));
1536 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1539 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1541 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1542 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1546 list_add_tail(&req->rq_replay_list,
1547 &exp->exp_req_replay_queue);
1550 spin_unlock(&exp->exp_lock);
1554 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1556 LASSERT(!list_empty(&req->rq_replay_list));
1557 LASSERT(req->rq_export);
1559 spin_lock(&req->rq_export->exp_lock);
1560 list_del_init(&req->rq_replay_list);
1561 spin_unlock(&req->rq_export->exp_lock);
1564 static void target_finish_recovery(struct lu_target *lut)
1566 struct obd_device *obd = lut->lut_obd;
1569 /* Only log a recovery message when recovery has occurred. */
1570 if (obd->obd_recovery_start) {
1571 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1572 obd->obd_recovery_start);
1573 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1574 "%d recovered and %d %s evicted.\n", obd->obd_name,
1575 (int)elapsed_time / 60, (int)elapsed_time % 60,
1576 obd->obd_max_recoverable_clients,
1577 atomic_read(&obd->obd_connected_clients),
1578 obd->obd_stale_clients,
1579 obd->obd_stale_clients == 1 ? "was" : "were");
1582 ldlm_reprocess_recovery_done(obd->obd_namespace);
1583 spin_lock(&obd->obd_recovery_task_lock);
1584 if (!list_empty(&obd->obd_req_replay_queue) ||
1585 !list_empty(&obd->obd_lock_replay_queue) ||
1586 !list_empty(&obd->obd_final_req_queue)) {
1587 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1589 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1590 list_empty(&obd->obd_lock_replay_queue) ? \
1592 list_empty(&obd->obd_final_req_queue) ? \
1594 spin_unlock(&obd->obd_recovery_task_lock);
1597 spin_unlock(&obd->obd_recovery_task_lock);
1599 obd->obd_recovery_end = cfs_time_current_sec();
1601 /* When recovery finished, cleanup orphans on MDS and OST. */
1602 if (OBT(obd) && OBP(obd, postrecov)) {
1603 int rc = OBP(obd, postrecov)(obd);
1605 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1611 static void abort_req_replay_queue(struct obd_device *obd)
1613 struct ptlrpc_request *req, *n;
1614 struct list_head abort_list;
1616 INIT_LIST_HEAD(&abort_list);
1617 spin_lock(&obd->obd_recovery_task_lock);
1618 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1619 spin_unlock(&obd->obd_recovery_task_lock);
1620 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1621 DEBUG_REQ(D_WARNING, req, "aborted:");
1622 req->rq_status = -ENOTCONN;
1623 if (ptlrpc_error(req)) {
1624 DEBUG_REQ(D_ERROR, req,
1625 "failed abort_req_reply; skipping");
1627 target_exp_dequeue_req_replay(req);
1628 target_request_copy_put(req);
1632 static void abort_lock_replay_queue(struct obd_device *obd)
1634 struct ptlrpc_request *req, *n;
1635 struct list_head abort_list;
1637 INIT_LIST_HEAD(&abort_list);
1638 spin_lock(&obd->obd_recovery_task_lock);
1639 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1640 spin_unlock(&obd->obd_recovery_task_lock);
1641 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1642 DEBUG_REQ(D_ERROR, req, "aborted:");
1643 req->rq_status = -ENOTCONN;
1644 if (ptlrpc_error(req)) {
1645 DEBUG_REQ(D_ERROR, req,
1646 "failed abort_lock_reply; skipping");
1648 target_request_copy_put(req);
1652 /* Called from a cleanup function if the device is being cleaned up
1653 forcefully. The exports should all have been disconnected already,
1654 the only thing left to do is
1655 - clear the recovery flags
1657 - free queued requests and replies, but don't send replies
1658 Because the obd_stopping flag is set, no new requests should be received.
1661 void target_cleanup_recovery(struct obd_device *obd)
1663 struct ptlrpc_request *req, *n;
1664 struct list_head clean_list;
1666 INIT_LIST_HEAD(&clean_list);
1667 spin_lock(&obd->obd_dev_lock);
1668 if (!obd->obd_recovering) {
1669 spin_unlock(&obd->obd_dev_lock);
1673 obd->obd_recovering = obd->obd_abort_recovery = 0;
1674 spin_unlock(&obd->obd_dev_lock);
1676 spin_lock(&obd->obd_recovery_task_lock);
1677 target_cancel_recovery_timer(obd);
1678 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1679 spin_unlock(&obd->obd_recovery_task_lock);
1681 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1682 LASSERT(req->rq_reply_state == NULL);
1683 target_exp_dequeue_req_replay(req);
1684 target_request_copy_put(req);
1687 spin_lock(&obd->obd_recovery_task_lock);
1688 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1689 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1690 spin_unlock(&obd->obd_recovery_task_lock);
1692 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1693 LASSERT(req->rq_reply_state == NULL);
1694 target_request_copy_put(req);
1699 EXPORT_SYMBOL(target_cleanup_recovery);
1701 /* obd_recovery_task_lock should be held */
1702 void target_cancel_recovery_timer(struct obd_device *obd)
1704 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1705 del_timer(&obd->obd_recovery_timer);
1708 static void target_start_recovery_timer(struct obd_device *obd)
1710 if (obd->obd_recovery_start != 0)
1713 spin_lock(&obd->obd_dev_lock);
1714 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1715 spin_unlock(&obd->obd_dev_lock);
1719 LASSERT(obd->obd_recovery_timeout != 0);
1721 if (obd->obd_recovery_start != 0) {
1722 spin_unlock(&obd->obd_dev_lock);
1726 mod_timer(&obd->obd_recovery_timer,
1727 cfs_time_shift(obd->obd_recovery_timeout));
1728 obd->obd_recovery_start = cfs_time_current_sec();
1729 spin_unlock(&obd->obd_dev_lock);
1731 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1732 "or until %d client%s reconnect%s\n",
1734 obd->obd_recovery_timeout / 60,
1735 obd->obd_recovery_timeout % 60,
1736 obd->obd_max_recoverable_clients,
1737 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1738 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1742 * extend recovery window.
1744 * if @extend is true, extend recovery window to have @drt remaining at least;
1745 * otherwise, make sure the recovery timeout value is not less than @drt.
1747 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1751 cfs_duration_t left;
1754 spin_lock(&obd->obd_dev_lock);
1755 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1756 spin_unlock(&obd->obd_dev_lock);
1759 LASSERT(obd->obd_recovery_start != 0);
1761 now = cfs_time_current_sec();
1762 to = obd->obd_recovery_timeout;
1763 end = obd->obd_recovery_start + to;
1764 left = cfs_time_sub(end, now);
1766 if (extend && (drt > left)) {
1768 } else if (!extend && (drt > to)) {
1772 if (to > obd->obd_recovery_time_hard) {
1773 to = obd->obd_recovery_time_hard;
1774 CWARN("%s: extended recovery timer reaching hard "
1775 "limit: %d, extend: %d\n",
1776 obd->obd_name, to, extend);
1779 if (obd->obd_recovery_timeout < to) {
1780 obd->obd_recovery_timeout = to;
1781 end = obd->obd_recovery_start + to;
1782 mod_timer(&obd->obd_recovery_timer,
1783 cfs_time_shift(end - now));
1785 spin_unlock(&obd->obd_dev_lock);
1787 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1788 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1791 /* Reset the timer with each new client connection */
1793 * This timer is actually reconnect_timer, which is for making sure
1794 * the total recovery window is at least as big as my reconnect
1795 * attempt timing. So the initial recovery time_out will be set to
1796 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1797 * from client is bigger than this, then the recovery time_out will
1798 * be extended to make sure the client could be reconnected, in the
1799 * process, the timeout from the new client should be ignored.
1803 check_and_start_recovery_timer(struct obd_device *obd,
1804 struct ptlrpc_request *req,
1807 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1808 struct obd_device_target *obt = &obd->u.obt;
1810 if (!new_client && service_time)
1811 /* Teach server about old server's estimates, as first guess
1812 * at how long new requests will take. */
1813 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1816 target_start_recovery_timer(obd);
1818 /* Convert the service time to RPC timeout,
1819 * and reuse service_time to limit stack usage. */
1820 service_time = at_est2timeout(service_time);
1822 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1823 service_time < at_extra)
1824 service_time = at_extra;
1826 /* We expect other clients to timeout within service_time, then try
1827 * to reconnect, then try the failover server. The max delay between
1828 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1829 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1831 LASSERT(obt->obt_magic == OBT_MAGIC);
1832 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1833 if (service_time > obd->obd_recovery_timeout && !new_client)
1834 extend_recovery_timer(obd, service_time, false);
1837 /** Health checking routines */
1838 static inline int exp_connect_healthy(struct obd_export *exp)
1840 return (exp->exp_in_recovery);
1843 /** if export done req_replay or has replay in queue */
1844 static inline int exp_req_replay_healthy(struct obd_export *exp)
1846 return (!exp->exp_req_replay_needed ||
1847 atomic_read(&exp->exp_replay_count) > 0);
1851 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1853 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1854 exp_req_replay_healthy(exp);
1857 /** if export done lock_replay or has replay in queue */
1858 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1860 return (!exp->exp_lock_replay_needed ||
1861 atomic_read(&exp->exp_replay_count) > 0);
1864 static inline int exp_vbr_healthy(struct obd_export *exp)
1866 return (!exp->exp_vbr_failed);
1869 static inline int exp_finished(struct obd_export *exp)
1871 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1874 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1876 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1880 static int check_for_next_transno(struct lu_target *lut)
1882 struct ptlrpc_request *req = NULL;
1883 struct obd_device *obd = lut->lut_obd;
1884 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1885 int wake_up = 0, connected, completed, queue_len;
1886 __u64 req_transno = 0;
1887 __u64 update_transno = 0;
1888 __u64 next_transno = 0;
1891 spin_lock(&obd->obd_recovery_task_lock);
1892 if (!list_empty(&obd->obd_req_replay_queue)) {
1893 req = list_entry(obd->obd_req_replay_queue.next,
1894 struct ptlrpc_request, rq_list);
1895 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1899 update_transno = distribute_txn_get_next_transno(tdtd);
1901 connected = atomic_read(&obd->obd_connected_clients);
1902 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1903 queue_len = obd->obd_requests_queued_for_recovery;
1904 next_transno = obd->obd_next_recovery_transno;
1906 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1907 "req_transno: %llu, next_transno: %llu\n",
1908 obd->obd_max_recoverable_clients, connected, completed,
1909 queue_len, req_transno, next_transno);
1911 if (obd->obd_abort_recovery) {
1912 CDEBUG(D_HA, "waking for aborted recovery\n");
1914 } else if (obd->obd_recovery_expired) {
1915 CDEBUG(D_HA, "waking for expired recovery\n");
1917 } else if (tdtd != NULL && req != NULL &&
1918 is_req_replayed_by_update(req)) {
1919 LASSERTF(req_transno < next_transno, "req_transno %llu"
1920 "next_transno%llu\n", req_transno, next_transno);
1921 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1924 } else if (req_transno == next_transno ||
1925 (update_transno != 0 && update_transno <= next_transno)) {
1926 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1928 } else if (queue_len > 0 &&
1929 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1930 /** handle gaps occured due to lost reply or VBR */
1931 LASSERTF(req_transno >= next_transno,
1932 "req_transno: %llu, next_transno: %llu\n",
1933 req_transno, next_transno);
1935 "%s: waking for gap in transno, VBR is %s (skip: "
1936 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1937 ", next_update %lld last_committed: %lld)\n",
1938 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1939 next_transno, queue_len, completed, connected,
1940 req_transno, update_transno, obd->obd_last_committed);
1941 obd->obd_next_recovery_transno = req_transno;
1943 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1944 CDEBUG(D_HA, "waking for completed recovery\n");
1946 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1947 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1948 " by fail_lock, waking up (%lld)\n", next_transno);
1949 obd->obd_next_recovery_transno = req_transno;
1952 spin_unlock(&obd->obd_recovery_task_lock);
1956 static int check_for_next_lock(struct lu_target *lut)
1958 struct obd_device *obd = lut->lut_obd;
1961 spin_lock(&obd->obd_recovery_task_lock);
1962 if (!list_empty(&obd->obd_lock_replay_queue)) {
1963 CDEBUG(D_HA, "waking for next lock\n");
1965 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1966 CDEBUG(D_HA, "waking for completed lock replay\n");
1968 } else if (obd->obd_abort_recovery) {
1969 CDEBUG(D_HA, "waking for aborted recovery\n");
1971 } else if (obd->obd_recovery_expired) {
1972 CDEBUG(D_HA, "waking for expired recovery\n");
1975 spin_unlock(&obd->obd_recovery_task_lock);
1981 * wait for recovery events,
1982 * check its status with help of check_routine
1983 * evict dead clients via health_check
1985 static int target_recovery_overseer(struct lu_target *lut,
1986 int (*check_routine)(struct lu_target *),
1987 int (*health_check)(struct obd_export *))
1989 struct obd_device *obd = lut->lut_obd;
1990 struct target_distribute_txn_data *tdtd;
1992 if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1993 (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1994 __u64 next_update_transno = 0;
1996 /* Only abort the recovery if there are no update recovery
1997 * left in the queue */
1998 spin_lock(&obd->obd_recovery_task_lock);
1999 if (lut->lut_tdtd != NULL) {
2000 next_update_transno =
2001 distribute_txn_get_next_transno(lut->lut_tdtd);
2003 tdtd = lut->lut_tdtd;
2004 /* If next_update_transno == 0, it probably because
2005 * updatelog retrieve threads did not get any records
2006 * yet, let's wait those threads stopped */
2007 if (next_update_transno == 0) {
2008 struct l_wait_info lwi = { 0 };
2010 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2012 &tdtd->tdtd_recovery_threads_count) == 0,
2015 next_update_transno =
2016 distribute_txn_get_next_transno(
2021 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2022 obd->obd_next_recovery_transno = next_update_transno;
2023 spin_unlock(&obd->obd_recovery_task_lock);
2024 /* Disconnect unfinished exports from clients, and
2025 * keep connection from MDT to make sure the update
2026 * recovery will still keep trying until some one
2027 * manually abort the recovery */
2028 class_disconnect_stale_exports(obd,
2029 exp_finished_or_from_mdt);
2030 /* Abort all of replay and replay lock req from
2032 abort_req_replay_queue(obd);
2033 abort_lock_replay_queue(obd);
2034 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2035 ")in the queue.\n", obd->obd_name,
2036 next_update_transno);
2038 obd->obd_abort_recovery = 1;
2039 spin_unlock(&obd->obd_recovery_task_lock);
2040 CWARN("%s recovery is aborted by hard timeout\n",
2045 while (wait_event_timeout(obd->obd_next_transno_waitq,
2047 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2048 /* wait indefinitely for event, but don't trigger watchdog */;
2050 if (obd->obd_abort_recovery) {
2051 CWARN("recovery is aborted, evict exports in recovery\n");
2052 if (lut->lut_tdtd != NULL) {
2053 struct l_wait_info lwi = { 0 };
2055 tdtd = lut->lut_tdtd;
2056 /* Let's wait all of the update log recovery thread
2058 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2059 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2061 /* Then abort the update recovery list */
2062 dtrq_list_destroy(lut->lut_tdtd);
2065 /** evict exports which didn't finish recovery yet */
2066 class_disconnect_stale_exports(obd, exp_finished);
2068 } else if (obd->obd_recovery_expired) {
2069 obd->obd_recovery_expired = 0;
2070 /** If some clients died being recovered, evict them */
2071 LCONSOLE_WARN("%s: recovery is timed out, "
2072 "evict stale exports\n", obd->obd_name);
2073 /** evict cexports with no replay in queue, they are stalled */
2074 class_disconnect_stale_exports(obd, health_check);
2076 /** continue with VBR */
2077 spin_lock(&obd->obd_dev_lock);
2078 obd->obd_version_recov = 1;
2079 spin_unlock(&obd->obd_dev_lock);
2081 * reset timer, recovery will proceed with versions now,
2082 * timeout is set just to handle reconnection delays
2084 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2085 /** Wait for recovery events again, after evicting bad clients */
2091 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2093 struct obd_device *obd = lut->lut_obd;
2094 struct ptlrpc_request *req = NULL;
2096 CDEBUG(D_HA, "Waiting for lock\n");
2097 if (target_recovery_overseer(lut, check_for_next_lock,
2098 exp_lock_replay_healthy))
2099 abort_lock_replay_queue(obd);
2101 spin_lock(&obd->obd_recovery_task_lock);
2102 if (!list_empty(&obd->obd_lock_replay_queue)) {
2103 req = list_entry(obd->obd_lock_replay_queue.next,
2104 struct ptlrpc_request, rq_list);
2105 list_del_init(&req->rq_list);
2106 spin_unlock(&obd->obd_recovery_task_lock);
2108 spin_unlock(&obd->obd_recovery_task_lock);
2109 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2110 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2111 /** evict exports failed VBR */
2112 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2117 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2119 struct ptlrpc_request *req = NULL;
2121 spin_lock(&obd->obd_recovery_task_lock);
2122 if (!list_empty(&obd->obd_final_req_queue)) {
2123 req = list_entry(obd->obd_final_req_queue.next,
2124 struct ptlrpc_request, rq_list);
2125 list_del_init(&req->rq_list);
2126 spin_unlock(&obd->obd_recovery_task_lock);
2127 if (req->rq_export->exp_in_recovery) {
2128 spin_lock(&req->rq_export->exp_lock);
2129 req->rq_export->exp_in_recovery = 0;
2130 spin_unlock(&req->rq_export->exp_lock);
2133 spin_unlock(&obd->obd_recovery_task_lock);
2138 static void handle_recovery_req(struct ptlrpc_thread *thread,
2139 struct ptlrpc_request *req,
2140 svc_handler_t handler)
2145 * export can be evicted during recovery, no need to handle replays for
2146 * it after that, discard such request silently
2148 if (req->rq_export->exp_disconnected)
2151 req->rq_session.lc_thread = thread;
2152 req->rq_svc_thread = thread;
2153 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2155 /* thread context */
2156 lu_context_enter(&thread->t_env->le_ctx);
2158 lu_context_exit(&thread->t_env->le_ctx);
2160 /* don't reset timer for final stage */
2161 if (!exp_finished(req->rq_export)) {
2162 int to = obd_timeout;
2165 * Add request timeout to the recovery time so next request from
2166 * this client may come in recovery time
2169 struct ptlrpc_service_part *svcpt;
2171 svcpt = req->rq_rqbd->rqbd_svcpt;
2172 /* If the server sent early reply for this request,
2173 * the client will recalculate the timeout according to
2174 * current server estimate service time, so we will
2175 * use the maxium timeout here for waiting the client
2176 * sending the next req */
2177 to = max((int)at_est2timeout(
2178 at_get(&svcpt->scp_at_estimate)),
2179 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2180 /* Add 2 net_latency, one for balance rq_deadline
2181 * (see ptl_send_rpc), one for resend the req to server,
2182 * Note: client will pack net_latency in replay req
2183 * (see ptlrpc_replay_req) */
2184 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2186 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2191 /** Checking routines for recovery */
2192 static int check_for_recovery_ready(struct lu_target *lut)
2194 struct obd_device *obd = lut->lut_obd;
2195 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2197 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2198 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2199 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2200 obd->obd_recovery_expired);
2202 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2203 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2204 if (clnts + obd->obd_stale_clients <
2205 obd->obd_max_recoverable_clients)
2209 if (lut->lut_tdtd != NULL) {
2210 if (!lut->lut_tdtd->tdtd_replay_ready &&
2211 !obd->obd_abort_recovery) {
2212 /* Let's extend recovery timer, in case the recovery
2213 * timer expired, and some clients got evicted */
2214 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2216 CDEBUG(D_HA, "%s update recovery is not ready,"
2217 " extend recovery %d\n", obd->obd_name,
2218 obd->obd_recovery_timeout);
2227 REQUEST_RECOVERY = 1,
2228 UPDATE_RECOVERY = 2,
2231 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2235 if (!list_empty(&obd->obd_req_replay_queue)) {
2236 struct ptlrpc_request *req;
2238 req = list_entry(obd->obd_req_replay_queue.next,
2239 struct ptlrpc_request, rq_list);
2240 transno = lustre_msg_get_transno(req->rq_reqmsg);
2246 static __u64 get_next_transno(struct lu_target *lut, int *type)
2248 struct obd_device *obd = lut->lut_obd;
2249 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2251 __u64 update_transno;
2254 transno = get_next_replay_req_transno(obd);
2256 *type = REQUEST_RECOVERY;
2261 update_transno = distribute_txn_get_next_transno(tdtd);
2262 if (transno == 0 || (transno >= update_transno &&
2263 update_transno != 0)) {
2264 transno = update_transno;
2266 *type = UPDATE_RECOVERY;
2273 * drop duplicate replay request
2275 * Because the operation has been replayed by update recovery, the request
2276 * with the same transno will be dropped and also notify the client to send
2277 * next replay request.
2279 * \param[in] env execution environment
2280 * \param[in] obd failover obd device
2281 * \param[in] req request to be dropped
2283 static void drop_duplicate_replay_req(struct lu_env *env,
2284 struct obd_device *obd,
2285 struct ptlrpc_request *req)
2287 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2288 " update records are found.\n",
2289 lustre_msg_get_transno(req->rq_reqmsg),
2290 libcfs_nid2str(req->rq_peer.nid));
2292 /* Right now, only for MDS reint operation update replay and
2293 * normal request replay can have the same transno */
2294 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2295 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2296 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2297 if (likely(req->rq_export))
2298 target_committed_to_req(req);
2299 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2300 target_send_reply(req, req->rq_status, 0);
2302 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2303 libcfs_nid2str(req->rq_peer.nid));
2305 target_exp_dequeue_req_replay(req);
2306 target_request_copy_put(req);
2307 obd->obd_replayed_requests++;
2310 static void replay_request_or_update(struct lu_env *env,
2311 struct lu_target *lut,
2312 struct target_recovery_data *trd,
2313 struct ptlrpc_thread *thread)
2315 struct obd_device *obd = lut->lut_obd;
2316 struct ptlrpc_request *req = NULL;
2321 CDEBUG(D_HA, "Waiting for transno %lld\n",
2322 obd->obd_next_recovery_transno);
2324 /* Replay all of request and update by transno */
2326 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2328 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2330 /** It is needed to extend recovery window above
2331 * recovery_time_soft. Extending is possible only in the
2332 * end of recovery window (see more details in
2333 * handle_recovery_req()).
2335 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2337 if (target_recovery_overseer(lut, check_for_next_transno,
2338 exp_req_replay_healthy_or_from_mdt)) {
2339 abort_req_replay_queue(obd);
2340 abort_lock_replay_queue(obd);
2344 spin_lock(&obd->obd_recovery_task_lock);
2345 transno = get_next_transno(lut, &type);
2346 if (type == REQUEST_RECOVERY && transno != 0) {
2347 /* Drop replay request from client side, if the
2348 * replay has been executed by update with the
2350 req = list_entry(obd->obd_req_replay_queue.next,
2351 struct ptlrpc_request, rq_list);
2353 list_del_init(&req->rq_list);
2354 obd->obd_requests_queued_for_recovery--;
2355 spin_unlock(&obd->obd_recovery_task_lock);
2357 /* Let's check if the request has been redone by
2359 if (is_req_replayed_by_update(req)) {
2360 struct distribute_txn_replay_req *dtrq;
2362 dtrq = distribute_txn_lookup_finish_list(tdtd,
2364 LASSERT(dtrq != NULL);
2365 spin_lock(&tdtd->tdtd_replay_list_lock);
2366 list_del_init(&dtrq->dtrq_list);
2367 spin_unlock(&tdtd->tdtd_replay_list_lock);
2370 drop_duplicate_replay_req(env, obd, req);
2375 LASSERT(trd->trd_processing_task == current_pid());
2376 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2377 lustre_msg_get_transno(req->rq_reqmsg),
2378 libcfs_nid2str(req->rq_peer.nid));
2380 handle_recovery_req(thread, req,
2381 trd->trd_recovery_handler);
2383 * bz18031: increase next_recovery_transno before
2384 * target_request_copy_put() will drop exp_rpc reference
2386 spin_lock(&obd->obd_recovery_task_lock);
2387 obd->obd_next_recovery_transno++;
2388 spin_unlock(&obd->obd_recovery_task_lock);
2389 target_exp_dequeue_req_replay(req);
2390 target_request_copy_put(req);
2391 obd->obd_replayed_requests++;
2392 } else if (type == UPDATE_RECOVERY && transno != 0) {
2393 struct distribute_txn_replay_req *dtrq;
2396 spin_unlock(&obd->obd_recovery_task_lock);
2398 LASSERT(tdtd != NULL);
2399 dtrq = distribute_txn_get_next_req(tdtd);
2400 lu_context_enter(&thread->t_env->le_ctx);
2401 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2402 lu_context_exit(&thread->t_env->le_ctx);
2403 extend_recovery_timer(obd, obd_timeout, true);
2405 if (rc == 0 && dtrq->dtrq_xid != 0) {
2406 CDEBUG(D_HA, "Move x%llu t%llu"
2407 " to finish list\n", dtrq->dtrq_xid,
2408 dtrq->dtrq_master_transno);
2410 /* Add it to the replay finish list */
2411 spin_lock(&tdtd->tdtd_replay_list_lock);
2412 list_add(&dtrq->dtrq_list,
2413 &tdtd->tdtd_replay_finish_list);
2414 spin_unlock(&tdtd->tdtd_replay_list_lock);
2416 spin_lock(&obd->obd_recovery_task_lock);
2417 if (transno == obd->obd_next_recovery_transno)
2418 obd->obd_next_recovery_transno++;
2420 obd->obd_next_recovery_transno)
2421 obd->obd_next_recovery_transno =
2423 spin_unlock(&obd->obd_recovery_task_lock);
2428 spin_unlock(&obd->obd_recovery_task_lock);
2430 LASSERT(list_empty(&obd->obd_req_replay_queue));
2431 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2432 /** evict exports failed VBR */
2433 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2439 static int target_recovery_thread(void *arg)
2441 struct lu_target *lut = arg;
2442 struct obd_device *obd = lut->lut_obd;
2443 struct ptlrpc_request *req;
2444 struct target_recovery_data *trd = &obd->obd_recovery_data;
2445 unsigned long delta;
2447 struct ptlrpc_thread *thread = NULL;
2451 unshare_fs_struct();
2452 OBD_ALLOC_PTR(thread);
2458 OBD_FREE_PTR(thread);
2462 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2464 OBD_FREE_PTR(thread);
2469 thread->t_env = env;
2470 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2471 env->le_ctx.lc_thread = thread;
2472 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2473 thread->t_watchdog = NULL;
2475 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2477 trd->trd_processing_task = current_pid();
2479 spin_lock(&obd->obd_dev_lock);
2480 obd->obd_recovering = 1;
2481 spin_unlock(&obd->obd_dev_lock);
2482 complete(&trd->trd_starting);
2484 /* first of all, we have to know the first transno to replay */
2485 if (target_recovery_overseer(lut, check_for_recovery_ready,
2486 exp_connect_healthy)) {
2487 abort_req_replay_queue(obd);
2488 abort_lock_replay_queue(obd);
2489 if (lut->lut_tdtd != NULL)
2490 dtrq_list_destroy(lut->lut_tdtd);
2493 /* next stage: replay requests or update */
2495 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2496 atomic_read(&obd->obd_req_replay_clients),
2497 obd->obd_next_recovery_transno);
2498 replay_request_or_update(env, lut, trd, thread);
2501 * The second stage: replay locks
2503 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2504 atomic_read(&obd->obd_lock_replay_clients));
2505 while ((req = target_next_replay_lock(lut))) {
2506 LASSERT(trd->trd_processing_task == current_pid());
2507 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2508 libcfs_nid2str(req->rq_peer.nid));
2509 handle_recovery_req(thread, req,
2510 trd->trd_recovery_handler);
2511 target_request_copy_put(req);
2512 obd->obd_replayed_locks++;
2516 * The third stage: reply on final pings, at this moment all clients
2517 * must have request in final queue
2519 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2520 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2521 /** Update server last boot epoch */
2522 tgt_boot_epoch_update(lut);
2523 /* We drop recoverying flag to forward all new requests
2524 * to regular mds_handle() since now */
2525 spin_lock(&obd->obd_dev_lock);
2526 obd->obd_recovering = obd->obd_abort_recovery = 0;
2527 spin_unlock(&obd->obd_dev_lock);
2528 spin_lock(&obd->obd_recovery_task_lock);
2529 target_cancel_recovery_timer(obd);
2530 spin_unlock(&obd->obd_recovery_task_lock);
2531 while ((req = target_next_final_ping(obd))) {
2532 LASSERT(trd->trd_processing_task == current_pid());
2533 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2534 libcfs_nid2str(req->rq_peer.nid));
2535 handle_recovery_req(thread, req,
2536 trd->trd_recovery_handler);
2537 /* Because the waiting client can not send ping to server,
2538 * so we need refresh the last_request_time, to avoid the
2539 * export is being evicted */
2540 ptlrpc_update_export_timer(req->rq_export, 0);
2541 target_request_copy_put(req);
2544 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2545 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2546 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2547 if (delta > OBD_RECOVERY_TIME_SOFT) {
2548 CWARN("too long recovery - read logs\n");
2549 libcfs_debug_dumplog();
2552 target_finish_recovery(lut);
2554 lu_context_fini(&env->le_ctx);
2555 trd->trd_processing_task = 0;
2556 complete(&trd->trd_finishing);
2558 tgt_io_thread_done(thread);
2559 OBD_FREE_PTR(thread);
2564 static int target_start_recovery_thread(struct lu_target *lut,
2565 svc_handler_t handler)
2567 struct obd_device *obd = lut->lut_obd;
2569 struct target_recovery_data *trd = &obd->obd_recovery_data;
2572 memset(trd, 0, sizeof(*trd));
2573 init_completion(&trd->trd_starting);
2574 init_completion(&trd->trd_finishing);
2575 trd->trd_recovery_handler = handler;
2577 rc = server_name2index(obd->obd_name, &index, NULL);
2581 if (!IS_ERR(kthread_run(target_recovery_thread,
2582 lut, "tgt_recover_%d", index))) {
2583 wait_for_completion(&trd->trd_starting);
2584 LASSERT(obd->obd_recovering != 0);
2592 void target_stop_recovery_thread(struct obd_device *obd)
2594 if (obd->obd_recovery_data.trd_processing_task > 0) {
2595 struct target_recovery_data *trd = &obd->obd_recovery_data;
2596 /** recovery can be done but postrecovery is not yet */
2597 spin_lock(&obd->obd_dev_lock);
2598 if (obd->obd_recovering) {
2599 CERROR("%s: Aborting recovery\n", obd->obd_name);
2600 obd->obd_abort_recovery = 1;
2601 wake_up(&obd->obd_next_transno_waitq);
2603 spin_unlock(&obd->obd_dev_lock);
2604 wait_for_completion(&trd->trd_finishing);
2607 EXPORT_SYMBOL(target_stop_recovery_thread);
2609 void target_recovery_fini(struct obd_device *obd)
2611 class_disconnect_exports(obd);
2612 target_stop_recovery_thread(obd);
2613 target_cleanup_recovery(obd);
2615 EXPORT_SYMBOL(target_recovery_fini);
2617 static void target_recovery_expired(unsigned long castmeharder)
2619 struct obd_device *obd = (struct obd_device *)castmeharder;
2620 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2621 " after %lds (%d clients connected)\n",
2622 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2623 cfs_time_current_sec()- obd->obd_recovery_start,
2624 atomic_read(&obd->obd_connected_clients));
2626 obd->obd_recovery_expired = 1;
2627 wake_up(&obd->obd_next_transno_waitq);
2630 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2632 struct obd_device *obd = lut->lut_obd;
2634 if (lut->lut_bottom->dd_rdonly)
2637 if (obd->obd_max_recoverable_clients == 0) {
2638 /** Update server last boot epoch */
2639 tgt_boot_epoch_update(lut);
2643 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2644 "last_transno %llu\n", obd->obd_name,
2645 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2646 LASSERT(obd->obd_stopping == 0);
2647 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2648 obd->obd_recovery_start = 0;
2649 obd->obd_recovery_end = 0;
2651 setup_timer(&obd->obd_recovery_timer, target_recovery_expired,
2652 (unsigned long)obd);
2653 target_start_recovery_thread(lut, handler);
2655 EXPORT_SYMBOL(target_recovery_init);
2657 static int target_process_req_flags(struct obd_device *obd,
2658 struct ptlrpc_request *req)
2660 struct obd_export *exp = req->rq_export;
2661 LASSERT(exp != NULL);
2662 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2663 /* client declares he's ready to replay locks */
2664 spin_lock(&exp->exp_lock);
2665 if (exp->exp_req_replay_needed) {
2666 exp->exp_req_replay_needed = 0;
2667 spin_unlock(&exp->exp_lock);
2669 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2670 atomic_dec(&obd->obd_req_replay_clients);
2672 spin_unlock(&exp->exp_lock);
2675 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2676 /* client declares he's ready to complete recovery
2677 * so, we put the request on th final queue */
2678 spin_lock(&exp->exp_lock);
2679 if (exp->exp_lock_replay_needed) {
2680 exp->exp_lock_replay_needed = 0;
2681 spin_unlock(&exp->exp_lock);
2683 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2684 atomic_dec(&obd->obd_lock_replay_clients);
2686 spin_unlock(&exp->exp_lock);
2692 int target_queue_recovery_request(struct ptlrpc_request *req,
2693 struct obd_device *obd)
2695 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2696 struct ptlrpc_request *reqiter;
2700 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2701 /* Processing the queue right now, don't re-add. */
2705 target_process_req_flags(obd, req);
2707 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2708 /* client declares he's ready to complete recovery
2709 * so, we put the request on th final queue */
2710 target_request_copy_get(req);
2711 DEBUG_REQ(D_HA, req, "queue final req");
2712 wake_up(&obd->obd_next_transno_waitq);
2713 spin_lock(&obd->obd_recovery_task_lock);
2714 if (obd->obd_recovering) {
2715 struct ptlrpc_request *tmp;
2716 struct ptlrpc_request *duplicate = NULL;
2718 if (likely(!req->rq_export->exp_replay_done)) {
2719 req->rq_export->exp_replay_done = 1;
2720 list_add_tail(&req->rq_list,
2721 &obd->obd_final_req_queue);
2722 spin_unlock(&obd->obd_recovery_task_lock);
2726 /* XXX O(n), but only happens if final ping is
2727 * timed out, probably reorganize the list as
2728 * a hash list later */
2729 list_for_each_entry_safe(reqiter, tmp,
2730 &obd->obd_final_req_queue,
2732 if (reqiter->rq_export == req->rq_export) {
2733 list_del_init(&reqiter->rq_list);
2734 duplicate = reqiter;
2739 list_add_tail(&req->rq_list,
2740 &obd->obd_final_req_queue);
2741 req->rq_export->exp_replay_done = 1;
2742 spin_unlock(&obd->obd_recovery_task_lock);
2744 if (duplicate != NULL) {
2745 DEBUG_REQ(D_HA, duplicate,
2746 "put prev final req\n");
2747 target_request_copy_put(duplicate);
2751 spin_unlock(&obd->obd_recovery_task_lock);
2752 target_request_copy_put(req);
2753 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2756 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2757 /* client declares he's ready to replay locks */
2758 target_request_copy_get(req);
2759 DEBUG_REQ(D_HA, req, "queue lock replay req");
2760 wake_up(&obd->obd_next_transno_waitq);
2761 spin_lock(&obd->obd_recovery_task_lock);
2762 LASSERT(obd->obd_recovering);
2763 /* usually due to recovery abort */
2764 if (!req->rq_export->exp_in_recovery) {
2765 spin_unlock(&obd->obd_recovery_task_lock);
2766 target_request_copy_put(req);
2769 LASSERT(req->rq_export->exp_lock_replay_needed);
2770 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2771 spin_unlock(&obd->obd_recovery_task_lock);
2775 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2776 * (i.e. buflens etc are in my own byte order), but type-dependent
2777 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2780 INIT_LIST_HEAD(&req->rq_list);
2781 DEBUG_REQ(D_HA, req, "not queueing");
2785 /* If we're processing the queue, we want don't want to queue this
2788 * Also, if this request has a transno less than the one we're waiting
2789 * for, we should process it now. It could (and currently always will)
2790 * be an open request for a descriptor that was opened some time ago.
2792 * Also, a resent, replayed request that has already been
2793 * handled will pass through here and be processed immediately.
2795 CDEBUG(D_HA, "Next recovery transno: %llu"
2796 ", current: %llu, replaying\n",
2797 obd->obd_next_recovery_transno, transno);
2799 /* If the request has been replayed by update replay, then sends this
2800 * request to the recovery thread (replay_request_or_update()), where
2801 * it will be handled */
2802 spin_lock(&obd->obd_recovery_task_lock);
2803 if (transno < obd->obd_next_recovery_transno &&
2804 !is_req_replayed_by_update(req)) {
2805 /* Processing the queue right now, don't re-add. */
2806 LASSERT(list_empty(&req->rq_list));
2807 spin_unlock(&obd->obd_recovery_task_lock);
2810 spin_unlock(&obd->obd_recovery_task_lock);
2812 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2815 target_request_copy_get(req);
2816 if (!req->rq_export->exp_in_recovery) {
2817 target_request_copy_put(req);
2820 LASSERT(req->rq_export->exp_req_replay_needed);
2822 if (target_exp_enqueue_req_replay(req)) {
2823 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2824 target_request_copy_put(req);
2829 spin_lock(&obd->obd_recovery_task_lock);
2830 LASSERT(obd->obd_recovering);
2831 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2832 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2833 list_add_tail(&req->rq_list, &reqiter->rq_list);
2838 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2840 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2841 "has been claimed by another client");
2842 spin_unlock(&obd->obd_recovery_task_lock);
2843 target_exp_dequeue_req_replay(req);
2844 target_request_copy_put(req);
2850 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2852 obd->obd_requests_queued_for_recovery++;
2853 spin_unlock(&obd->obd_recovery_task_lock);
2854 wake_up(&obd->obd_next_transno_waitq);
2858 int target_handle_ping(struct ptlrpc_request *req)
2860 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2861 return req_capsule_server_pack(&req->rq_pill);
2864 void target_committed_to_req(struct ptlrpc_request *req)
2866 struct obd_export *exp = req->rq_export;
2868 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2869 lustre_msg_set_last_committed(req->rq_repmsg,
2870 exp->exp_last_committed);
2872 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2873 "%d)", exp->exp_obd->obd_no_transno,
2874 req->rq_repmsg == NULL);
2876 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2877 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2880 #endif /* HAVE_SERVER_SUPPORT */
2883 * Packs current SLV and Limit into \a req.
2885 int target_pack_pool_reply(struct ptlrpc_request *req)
2887 struct obd_device *obd;
2890 /* Check that we still have all structures alive as this may
2891 * be some late RPC at shutdown time. */
2892 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2893 !exp_connect_lru_resize(req->rq_export))) {
2894 lustre_msg_set_slv(req->rq_repmsg, 0);
2895 lustre_msg_set_limit(req->rq_repmsg, 0);
2899 /* OBD is alive here as export is alive, which we checked above. */
2900 obd = req->rq_export->exp_obd;
2902 read_lock(&obd->obd_pool_lock);
2903 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2904 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2905 read_unlock(&obd->obd_pool_lock);
2910 static int target_send_reply_msg(struct ptlrpc_request *req,
2911 int rc, int fail_id)
2913 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2914 DEBUG_REQ(D_ERROR, req, "dropping reply");
2917 /* We can have a null rq_reqmsg in the event of bad signature or
2918 * no context when unwrapping */
2919 if (req->rq_reqmsg &&
2920 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2921 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2925 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2926 req->rq_status = rc;
2927 return ptlrpc_send_error(req, 1);
2929 DEBUG_REQ(D_NET, req, "sending reply");
2932 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2935 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2937 struct ptlrpc_service_part *svcpt;
2939 struct ptlrpc_reply_state *rs;
2940 struct obd_export *exp;
2943 if (req->rq_no_reply) {
2948 svcpt = req->rq_rqbd->rqbd_svcpt;
2949 rs = req->rq_reply_state;
2950 if (rs == NULL || !rs->rs_difficult) {
2952 target_send_reply_msg (req, rc, fail_id);
2957 /* must be an export if locks saved */
2958 LASSERT(req->rq_export != NULL);
2959 /* req/reply consistent */
2960 LASSERT(rs->rs_svcpt == svcpt);
2963 LASSERT(!rs->rs_scheduled);
2964 LASSERT(!rs->rs_scheduled_ever);
2965 LASSERT(!rs->rs_handled);
2966 LASSERT(!rs->rs_on_net);
2967 LASSERT(rs->rs_export == NULL);
2968 LASSERT(list_empty(&rs->rs_obd_list));
2969 LASSERT(list_empty(&rs->rs_exp_list));
2971 exp = class_export_get(req->rq_export);
2973 /* disable reply scheduling while I'm setting up */
2974 rs->rs_scheduled = 1;
2976 rs->rs_xid = req->rq_xid;
2977 rs->rs_transno = req->rq_transno;
2978 rs->rs_export = exp;
2979 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2981 spin_lock(&exp->exp_uncommitted_replies_lock);
2982 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2983 rs->rs_transno, exp->exp_last_committed);
2984 if (rs->rs_transno > exp->exp_last_committed) {
2985 /* not committed already */
2986 list_add_tail(&rs->rs_obd_list,
2987 &exp->exp_uncommitted_replies);
2989 spin_unlock(&exp->exp_uncommitted_replies_lock);
2991 spin_lock(&exp->exp_lock);
2992 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2993 spin_unlock(&exp->exp_lock);
2995 netrc = target_send_reply_msg(req, rc, fail_id);
2997 spin_lock(&svcpt->scp_rep_lock);
2999 atomic_inc(&svcpt->scp_nreps_difficult);
3002 /* error sending: reply is off the net. Also we need +1
3003 * reply ref until ptlrpc_handle_rs() is done
3004 * with the reply state (if the send was successful, there
3005 * would have been +1 ref for the net, which
3006 * reply_out_callback leaves alone) */
3008 ptlrpc_rs_addref(rs);
3011 spin_lock(&rs->rs_lock);
3012 if (rs->rs_transno <= exp->exp_last_committed ||
3013 (!rs->rs_on_net && !rs->rs_no_ack) ||
3014 list_empty(&rs->rs_exp_list) || /* completed already */
3015 list_empty(&rs->rs_obd_list)) {
3016 CDEBUG(D_HA, "Schedule reply immediately\n");
3017 ptlrpc_dispatch_difficult_reply(rs);
3019 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3020 rs->rs_scheduled = 0; /* allow notifier to schedule */
3022 spin_unlock(&rs->rs_lock);
3023 spin_unlock(&svcpt->scp_rep_lock);
3027 enum ldlm_mode lck_compat_array[] = {
3028 [LCK_EX] = LCK_COMPAT_EX,
3029 [LCK_PW] = LCK_COMPAT_PW,
3030 [LCK_PR] = LCK_COMPAT_PR,
3031 [LCK_CW] = LCK_COMPAT_CW,
3032 [LCK_CR] = LCK_COMPAT_CR,
3033 [LCK_NL] = LCK_COMPAT_NL,
3034 [LCK_GROUP] = LCK_COMPAT_GROUP,
3035 [LCK_COS] = LCK_COMPAT_COS,
3039 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3040 * not escape to the user level.
3042 int ldlm_error2errno(enum ldlm_error error)
3048 case ELDLM_LOCK_MATCHED:
3051 case ELDLM_LOCK_CHANGED:
3054 case ELDLM_LOCK_ABORTED:
3057 case ELDLM_LOCK_REPLACED:
3060 case ELDLM_NO_LOCK_DATA:
3063 case ELDLM_NAMESPACE_EXISTS:
3066 case ELDLM_BAD_NAMESPACE:
3070 if (((int)error) < 0) { /* cast to signed type */
3071 result = error; /* as ldlm_error can be unsigned */
3073 CERROR("Invalid DLM result code: %d\n", error);
3079 EXPORT_SYMBOL(ldlm_error2errno);
3082 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3084 enum ldlm_error ldlm_errno2error(int err_no)
3093 error = ELDLM_LOCK_CHANGED;
3096 error = ELDLM_LOCK_ABORTED;
3099 error = ELDLM_LOCK_REPLACED;
3102 error = ELDLM_NO_LOCK_DATA;
3105 error = ELDLM_NAMESPACE_EXISTS;
3108 error = ELDLM_BAD_NAMESPACE;
3116 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3117 void ldlm_dump_export_locks(struct obd_export *exp)
3119 spin_lock(&exp->exp_locks_list_guard);
3120 if (!list_empty(&exp->exp_locks_list)) {
3121 struct ldlm_lock *lock;
3123 CERROR("dumping locks for export %p,"
3124 "ignore if the unmount doesn't hang\n", exp);
3125 list_for_each_entry(lock, &exp->exp_locks_list,
3127 LDLM_ERROR(lock, "lock:");
3129 spin_unlock(&exp->exp_locks_list_guard);
3133 #ifdef HAVE_SERVER_SUPPORT
3134 static int target_bulk_timeout(void *data)
3137 /* We don't fail the connection here, because having the export
3138 * killed makes the (vital) call to commitrw very sad.
3143 static inline const char *bulk2type(struct ptlrpc_request *req)
3145 if (req->rq_bulk_read)
3147 if (req->rq_bulk_write)
3152 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3153 struct l_wait_info *lwi)
3155 struct ptlrpc_request *req = desc->bd_req;
3156 time_t start = cfs_time_current_sec();
3162 /* If there is eviction in progress, wait for it to finish. */
3163 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3164 *lwi = LWI_INTR(NULL, NULL);
3165 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3166 !atomic_read(&exp->exp_obd->
3167 obd_evict_inprogress),
3171 /* Check if client was evicted or reconnected already. */
3172 if (exp->exp_failed ||
3173 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3176 if (req->rq_bulk_read)
3177 rc = sptlrpc_svc_wrap_bulk(req, desc);
3179 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3180 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3181 else /* old version, bulk matchbits is rq_xid */
3182 req->rq_mbits = req->rq_xid;
3185 rc = ptlrpc_start_bulk_transfer(desc);
3189 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3190 bulk2type(req), rc);
3194 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3195 ptlrpc_abort_bulk(desc);
3199 /* limit actual bulk transfer to bulk_timeout seconds */
3200 deadline = start + bulk_timeout;
3201 if (deadline > req->rq_deadline)
3202 deadline = req->rq_deadline;
3205 long timeoutl = deadline - cfs_time_current_sec();
3206 cfs_duration_t timeout = timeoutl <= 0 ?
3207 CFS_TICK : cfs_time_seconds(timeoutl);
3210 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3211 target_bulk_timeout, desc);
3212 rc = l_wait_event(desc->bd_waitq,
3213 !ptlrpc_server_bulk_active(desc) ||
3216 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3218 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3219 /* Wait again if we changed rq_deadline. */
3220 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3221 deadline = start + bulk_timeout;
3222 if (deadline > rq_deadline)
3223 deadline = rq_deadline;
3224 } while ((rc == -ETIMEDOUT) &&
3225 (deadline > cfs_time_current_sec()));
3227 if (rc == -ETIMEDOUT) {
3228 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3229 bulk2type(req), deadline - start,
3230 cfs_time_current_sec() - deadline);
3231 ptlrpc_abort_bulk(desc);
3232 } else if (exp->exp_failed) {
3233 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3236 ptlrpc_abort_bulk(desc);
3237 } else if (exp->exp_conn_cnt >
3238 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3239 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3241 /* We don't reply anyway. */
3243 ptlrpc_abort_bulk(desc);
3244 } else if (desc->bd_failure) {
3245 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3247 /* XXX should this be a different errno? */
3250 if (req->rq_bulk_write)
3251 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3252 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3253 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3254 bulk2type(req), desc->bd_nob_transferred,
3256 /* XXX should this be a different errno? */
3263 EXPORT_SYMBOL(target_bulk_io);
3265 #endif /* HAVE_SERVER_SUPPORT */