4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * This file deals with various client/target related logic including recovery.
36 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
40 #define DEBUG_SUBSYSTEM S_LDLM
42 #include <linux/kthread.h>
43 #include <libcfs/libcfs.h>
45 #include <obd_class.h>
46 #include <lustre_dlm.h>
47 #include <lustre_net.h>
48 #include <lustre_sec.h>
49 #include "ldlm_internal.h"
51 /* @priority: If non-zero, move the selected connection to the list head.
52 * @create: If zero, only search in existing connections.
54 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
55 int priority, int create)
57 struct ptlrpc_connection *ptlrpc_conn;
58 struct obd_import_conn *imp_conn = NULL, *item;
59 lnet_nid_t nid4refnet = LNET_NID_ANY;
63 if (!create && !priority) {
64 CDEBUG(D_HA, "Nothing to do\n");
68 if (imp->imp_connection &&
69 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
70 /* nid4refnet is used to restrict network connections */
71 nid4refnet = imp->imp_connection->c_self;
72 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
74 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
79 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
81 GOTO(out_put, rc = -ENOMEM);
84 spin_lock(&imp->imp_lock);
85 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
86 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
88 list_del(&item->oic_item);
89 list_add(&item->oic_item,
91 item->oic_last_attempt = 0;
93 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
94 imp, imp->imp_obd->obd_name, uuid->uuid,
95 (priority ? ", moved to head" : ""));
96 spin_unlock(&imp->imp_lock);
97 GOTO(out_free, rc = 0);
100 /* No existing import connection found for \a uuid. */
102 imp_conn->oic_conn = ptlrpc_conn;
103 imp_conn->oic_uuid = *uuid;
104 imp_conn->oic_last_attempt = 0;
106 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
108 list_add_tail(&imp_conn->oic_item,
109 &imp->imp_conn_list);
110 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
111 imp, imp->imp_obd->obd_name, uuid->uuid,
112 (priority ? "head" : "tail"));
114 spin_unlock(&imp->imp_lock);
115 GOTO(out_free, rc = -ENOENT);
118 spin_unlock(&imp->imp_lock);
122 OBD_FREE(imp_conn, sizeof(*imp_conn));
124 ptlrpc_connection_put(ptlrpc_conn);
128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
130 return import_set_conn(imp, uuid, 1, 0);
133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
136 return import_set_conn(imp, uuid, priority, 1);
138 EXPORT_SYMBOL(client_import_add_conn);
140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
142 struct obd_import_conn *imp_conn;
143 struct obd_export *dlmexp;
147 spin_lock(&imp->imp_lock);
148 if (list_empty(&imp->imp_conn_list)) {
149 LASSERT(!imp->imp_connection);
153 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
154 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
156 LASSERT(imp_conn->oic_conn);
158 if (imp_conn == imp->imp_conn_current) {
159 LASSERT(imp_conn->oic_conn == imp->imp_connection);
161 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
162 imp->imp_state != LUSTRE_IMP_DISCON) {
163 CERROR("can't remove current connection\n");
164 GOTO(out, rc = -EBUSY);
167 ptlrpc_connection_put(imp->imp_connection);
168 imp->imp_connection = NULL;
170 dlmexp = class_conn2export(&imp->imp_dlm_handle);
171 if (dlmexp && dlmexp->exp_connection) {
172 LASSERT(dlmexp->exp_connection ==
174 ptlrpc_connection_put(dlmexp->exp_connection);
175 dlmexp->exp_connection = NULL;
179 class_export_put(dlmexp);
182 list_del(&imp_conn->oic_item);
183 ptlrpc_connection_put(imp_conn->oic_conn);
184 OBD_FREE(imp_conn, sizeof(*imp_conn));
185 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
186 imp, imp->imp_obd->obd_name, uuid->uuid);
191 spin_unlock(&imp->imp_lock);
193 CERROR("connection %s not found\n", uuid->uuid);
196 EXPORT_SYMBOL(client_import_del_conn);
199 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
200 * to find a conn uuid of \a imp which can reach \a peer.
202 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
203 struct obd_uuid *uuid)
205 struct obd_import_conn *conn;
209 spin_lock(&imp->imp_lock);
210 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
211 /* Check if conn UUID does have this peer NID. */
212 if (class_check_uuid(&conn->oic_uuid, peer)) {
213 *uuid = conn->oic_uuid;
218 spin_unlock(&imp->imp_lock);
221 EXPORT_SYMBOL(client_import_find_conn);
223 void client_destroy_import(struct obd_import *imp)
225 /* Drop security policy instance after all RPCs have finished/aborted
226 * to let all busy contexts be released. */
227 class_import_get(imp);
228 class_destroy_import(imp);
229 sptlrpc_import_sec_put(imp);
230 class_import_put(imp);
232 EXPORT_SYMBOL(client_destroy_import);
235 * Check whether or not the OSC is on MDT.
238 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
240 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
243 static int osc_on_mdt(char *obdname)
247 ptr = strrchr(obdname, '-');
251 if (strncmp(ptr + 1, "MDT", 3) == 0)
257 /* Configure an RPC client OBD device.
262 * 3 - inactive-on-startup
263 * 4 - restrictive net
265 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
267 struct client_obd *cli = &obddev->u.cli;
268 struct obd_import *imp;
269 struct obd_uuid server_uuid;
270 int rq_portal, rp_portal, connect_op;
271 char *name = obddev->obd_type->typ_name;
272 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
273 char *cli_name = lustre_cfg_buf(lcfg, 0);
274 struct ptlrpc_connection fake_conn = { .c_self = 0,
275 .c_remote_uuid.uuid[0] = 0 };
279 /* In a more perfect world, we would hang a ptlrpc_client off of
280 * obd_type and just use the values from there. */
281 if (!strcmp(name, LUSTRE_OSC_NAME)) {
282 rq_portal = OST_REQUEST_PORTAL;
283 rp_portal = OSC_REPLY_PORTAL;
284 connect_op = OST_CONNECT;
285 cli->cl_sp_me = LUSTRE_SP_CLI;
286 cli->cl_sp_to = LUSTRE_SP_OST;
287 ns_type = LDLM_NS_TYPE_OSC;
288 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
289 !strcmp(name, LUSTRE_LWP_NAME)) {
290 rq_portal = MDS_REQUEST_PORTAL;
291 rp_portal = MDC_REPLY_PORTAL;
292 connect_op = MDS_CONNECT;
293 if (is_lwp_on_ost(cli_name))
294 cli->cl_sp_me = LUSTRE_SP_OST;
295 else if (is_lwp_on_mdt(cli_name))
296 cli->cl_sp_me = LUSTRE_SP_MDT;
298 cli->cl_sp_me = LUSTRE_SP_CLI;
299 cli->cl_sp_to = LUSTRE_SP_MDT;
300 ns_type = LDLM_NS_TYPE_MDC;
301 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
302 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
303 /* OSP_on_MDT for other MDTs */
304 connect_op = MDS_CONNECT;
305 cli->cl_sp_to = LUSTRE_SP_MDT;
306 ns_type = LDLM_NS_TYPE_MDC;
307 rq_portal = OUT_PORTAL;
309 /* OSP on MDT for OST */
310 connect_op = OST_CONNECT;
311 cli->cl_sp_to = LUSTRE_SP_OST;
312 ns_type = LDLM_NS_TYPE_OSC;
313 rq_portal = OST_REQUEST_PORTAL;
315 rp_portal = OSC_REPLY_PORTAL;
316 cli->cl_sp_me = LUSTRE_SP_MDT;
317 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
318 rq_portal = MGS_REQUEST_PORTAL;
319 rp_portal = MGC_REPLY_PORTAL;
320 connect_op = MGS_CONNECT;
321 cli->cl_sp_me = LUSTRE_SP_MGC;
322 cli->cl_sp_to = LUSTRE_SP_MGS;
323 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
324 ns_type = LDLM_NS_TYPE_MGC;
326 CERROR("unknown client OBD type \"%s\", can't setup\n",
331 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
332 CERROR("requires a TARGET UUID\n");
336 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
337 CERROR("client UUID must be less than 38 characters\n");
341 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
342 CERROR("setup requires a SERVER UUID\n");
346 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
347 CERROR("target UUID must be less than 38 characters\n");
351 init_rwsem(&cli->cl_sem);
352 mutex_init(&cli->cl_mgc_mutex);
354 init_rwsem(&cli->cl_seq_rwsem);
355 cli->cl_conn_count = 0;
356 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
357 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
358 sizeof(server_uuid)));
360 cli->cl_dirty_pages = 0;
361 cli->cl_avail_grant = 0;
362 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
363 /* cl_dirty_max_pages may be changed at connect time in
364 * ptlrpc_connect_interpret(). */
365 client_adjust_max_dirty(cli);
366 INIT_LIST_HEAD(&cli->cl_cache_waiters);
367 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
368 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
369 INIT_LIST_HEAD(&cli->cl_loi_write_list);
370 INIT_LIST_HEAD(&cli->cl_loi_read_list);
371 spin_lock_init(&cli->cl_loi_list_lock);
372 atomic_set(&cli->cl_pending_w_pages, 0);
373 atomic_set(&cli->cl_pending_r_pages, 0);
374 cli->cl_r_in_flight = 0;
375 cli->cl_w_in_flight = 0;
377 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
378 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
379 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
380 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
381 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
382 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
385 INIT_LIST_HEAD(&cli->cl_lru_osc);
386 atomic_set(&cli->cl_lru_shrinkers, 0);
387 atomic_long_set(&cli->cl_lru_busy, 0);
388 atomic_long_set(&cli->cl_lru_in_list, 0);
389 INIT_LIST_HEAD(&cli->cl_lru_list);
390 spin_lock_init(&cli->cl_lru_list_lock);
391 atomic_long_set(&cli->cl_unstable_count, 0);
392 INIT_LIST_HEAD(&cli->cl_shrink_list);
394 init_waitqueue_head(&cli->cl_destroy_waitq);
395 atomic_set(&cli->cl_destroy_in_flight, 0);
396 #ifdef ENABLE_CHECKSUM
397 /* Turn on checksumming by default. */
398 cli->cl_checksum = 1;
400 * The supported checksum types will be worked out at connect time
401 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
404 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
406 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
408 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
409 * from OFD after connecting. */
410 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
412 /* set cl_chunkbits default value to PAGE_SHIFT,
413 * it will be updated at OSC connection time. */
414 cli->cl_chunkbits = PAGE_SHIFT;
416 if (!strcmp(name, LUSTRE_MDC_NAME)) {
417 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
418 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
419 cli->cl_max_rpcs_in_flight = 2;
420 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
421 cli->cl_max_rpcs_in_flight = 3;
422 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
423 cli->cl_max_rpcs_in_flight = 4;
425 if (osc_on_mdt(obddev->obd_name))
426 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
428 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
431 spin_lock_init(&cli->cl_mod_rpcs_lock);
432 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
433 cli->cl_max_mod_rpcs_in_flight = 0;
434 cli->cl_mod_rpcs_in_flight = 0;
435 cli->cl_close_rpcs_in_flight = 0;
436 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
437 cli->cl_mod_tag_bitmap = NULL;
439 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
441 if (connect_op == MDS_CONNECT) {
442 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
443 OBD_ALLOC(cli->cl_mod_tag_bitmap,
444 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
445 if (cli->cl_mod_tag_bitmap == NULL)
446 GOTO(err, rc = -ENOMEM);
451 CERROR("ldlm_get_ref failed: %d\n", rc);
455 ptlrpc_init_client(rq_portal, rp_portal, name,
456 &obddev->obd_ldlm_client);
458 imp = class_new_import(obddev);
460 GOTO(err_ldlm, rc = -ENOENT);
461 imp->imp_client = &obddev->obd_ldlm_client;
462 imp->imp_connect_op = connect_op;
463 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
464 LUSTRE_CFG_BUFLEN(lcfg, 1));
465 class_import_put(imp);
467 if (lustre_cfg_buf(lcfg, 4)) {
468 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
470 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
472 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
473 obddev->obd_name, lustre_cfg_string(lcfg, 4),
475 GOTO(err_import, rc);
477 fake_conn.c_self = LNET_MKNID(refnet, 0);
478 imp->imp_connection = &fake_conn;
481 rc = client_import_add_conn(imp, &server_uuid, 1);
483 CERROR("can't add initial connection\n");
484 GOTO(err_import, rc);
486 imp->imp_connection = NULL;
488 cli->cl_import = imp;
489 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
490 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
492 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
493 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
494 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
495 name, obddev->obd_name,
496 cli->cl_target_uuid.uuid);
497 spin_lock(&imp->imp_lock);
498 imp->imp_deactive = 1;
499 spin_unlock(&imp->imp_lock);
503 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
504 LDLM_NAMESPACE_CLIENT,
505 LDLM_NAMESPACE_GREEDY,
507 if (obddev->obd_namespace == NULL) {
508 CERROR("Unable to create client namespace - %s\n",
510 GOTO(err_import, rc = -ENOMEM);
516 class_destroy_import(imp);
520 if (cli->cl_mod_tag_bitmap != NULL)
521 OBD_FREE(cli->cl_mod_tag_bitmap,
522 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
523 cli->cl_mod_tag_bitmap = NULL;
527 EXPORT_SYMBOL(client_obd_setup);
529 int client_obd_cleanup(struct obd_device *obddev)
531 struct client_obd *cli = &obddev->u.cli;
534 ldlm_namespace_free_post(obddev->obd_namespace);
535 obddev->obd_namespace = NULL;
537 obd_cleanup_client_import(obddev);
538 LASSERT(obddev->u.cli.cl_import == NULL);
542 if (cli->cl_mod_tag_bitmap != NULL)
543 OBD_FREE(cli->cl_mod_tag_bitmap,
544 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
545 cli->cl_mod_tag_bitmap = NULL;
549 EXPORT_SYMBOL(client_obd_cleanup);
551 /* ->o_connect() method for client side (OSC and MDC and MGC) */
552 int client_connect_import(const struct lu_env *env,
553 struct obd_export **exp,
554 struct obd_device *obd, struct obd_uuid *cluuid,
555 struct obd_connect_data *data, void *localdata)
557 struct client_obd *cli = &obd->u.cli;
558 struct obd_import *imp = cli->cl_import;
559 struct obd_connect_data *ocd;
560 struct lustre_handle conn = { 0 };
565 down_write(&cli->cl_sem);
566 if (cli->cl_conn_count > 0)
567 GOTO(out_sem, rc = -EALREADY);
569 rc = class_connect(&conn, obd, cluuid);
573 cli->cl_conn_count++;
574 *exp = class_conn2export(&conn);
576 LASSERT(obd->obd_namespace);
578 imp->imp_dlm_handle = conn;
579 rc = ptlrpc_init_import(imp);
583 ocd = &imp->imp_connect_data;
586 imp->imp_connect_flags_orig = data->ocd_connect_flags;
587 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
590 rc = ptlrpc_connect_import(imp);
592 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
595 LASSERT(*exp != NULL && (*exp)->exp_connection);
598 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
599 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
600 data->ocd_connect_flags, ocd->ocd_connect_flags);
601 data->ocd_connect_flags = ocd->ocd_connect_flags;
604 ptlrpc_pinger_add_import(imp);
610 cli->cl_conn_count--;
611 class_disconnect(*exp);
615 up_write(&cli->cl_sem);
619 EXPORT_SYMBOL(client_connect_import);
621 int client_disconnect_export(struct obd_export *exp)
623 struct obd_device *obd = class_exp2obd(exp);
624 struct client_obd *cli;
625 struct obd_import *imp;
630 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
631 exp, exp ? exp->exp_handle.h_cookie : -1);
636 imp = cli->cl_import;
638 down_write(&cli->cl_sem);
639 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
642 if (cli->cl_conn_count == 0) {
643 CERROR("disconnecting disconnected device (%s)\n",
645 GOTO(out_disconnect, rc = -EINVAL);
648 cli->cl_conn_count--;
649 if (cli->cl_conn_count != 0)
650 GOTO(out_disconnect, rc = 0);
652 /* Mark import deactivated now, so we don't try to reconnect if any
653 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
654 * fully deactivate the import, or that would drop all requests. */
655 spin_lock(&imp->imp_lock);
656 imp->imp_deactive = 1;
657 spin_unlock(&imp->imp_lock);
659 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
660 * delete it regardless. (It's safe to delete an import that was
662 (void)ptlrpc_pinger_del_import(imp);
664 if (obd->obd_namespace != NULL) {
665 /* obd_force == local only */
666 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
667 obd->obd_force ? LCF_LOCAL : 0, NULL);
668 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
671 /* There's no need to hold sem while disconnecting an import,
672 * and it may actually cause deadlock in GSS. */
673 up_write(&cli->cl_sem);
674 rc = ptlrpc_disconnect_import(imp, 0);
675 down_write(&cli->cl_sem);
677 ptlrpc_invalidate_import(imp);
682 /* Use server style - class_disconnect should be always called for
684 err = class_disconnect(exp);
688 up_write(&cli->cl_sem);
692 EXPORT_SYMBOL(client_disconnect_export);
694 #ifdef HAVE_SERVER_SUPPORT
695 int server_disconnect_export(struct obd_export *exp)
700 /* Disconnect early so that clients can't keep using export. */
701 rc = class_disconnect(exp);
702 /* Close import to avoid sending any requests. */
703 if (exp->exp_imp_reverse)
704 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
706 ldlm_bl_thread_wakeup();
708 /* complete all outstanding replies */
709 spin_lock(&exp->exp_lock);
710 while (!list_empty(&exp->exp_outstanding_replies)) {
711 struct ptlrpc_reply_state *rs =
712 list_entry(exp->exp_outstanding_replies.next,
713 struct ptlrpc_reply_state, rs_exp_list);
714 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
716 spin_lock(&svcpt->scp_rep_lock);
718 list_del_init(&rs->rs_exp_list);
720 spin_lock(&rs->rs_lock);
721 /* clear rs_convert_lock to make sure rs is handled and put */
722 rs->rs_convert_lock = 0;
723 ptlrpc_schedule_difficult_reply(rs);
724 spin_unlock(&rs->rs_lock);
726 spin_unlock(&svcpt->scp_rep_lock);
728 spin_unlock(&exp->exp_lock);
732 EXPORT_SYMBOL(server_disconnect_export);
734 /* --------------------------------------------------------------------------
735 * from old lib/target.c
736 * -------------------------------------------------------------------------- */
738 static int target_handle_reconnect(struct lustre_handle *conn,
739 struct obd_export *exp,
740 struct obd_uuid *cluuid)
742 struct obd_device *target;
743 struct lustre_handle *hdl;
750 hdl = &exp->exp_imp_reverse->imp_remote_handle;
751 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
752 conn->cookie = exp->exp_handle.h_cookie;
753 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
754 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
758 target = exp->exp_obd;
760 /* Might be a re-connect after a partition. */
761 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
762 LCONSOLE_WARN("%s: already connected client %s (at %s) "
763 "with handle %#llx. Rejecting client "
764 "with the same UUID trying to reconnect "
765 "with handle %#llx\n", target->obd_name,
766 obd_uuid2str(&exp->exp_client_uuid),
767 obd_export_nid2str(exp),
768 hdl->cookie, conn->cookie);
769 memset(conn, 0, sizeof *conn);
770 /* target_handle_connect() treats EALREADY and
771 * -EALREADY differently. -EALREADY is an error
772 * (same UUID, different handle). */
776 if (!target->obd_recovering) {
777 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
778 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
779 obd_export_nid2str(exp));
780 GOTO(out_already, rc);
783 now = cfs_time_current();
784 deadline = target->obd_recovery_timer.expires;
785 if (cfs_time_before(now, deadline)) {
786 struct target_distribute_txn_data *tdtd =
787 class_exp2tgt(exp)->lut_tdtd;
792 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
793 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
794 buf = tdtd->tdtd_show_update_logs_retrievers(
795 tdtd->tdtd_show_retrievers_cbdata,
799 LCONSOLE_WARN("%s: Recovery already passed deadline "
800 "%d:%.02d. It is due to DNE recovery "
801 "failed/stuck on the %d MDT(s):%s. "
802 "Please wait until all MDTs recovered "
803 "or abort the recovery by force.\n",
804 target->obd_name, timeout / 60,
806 buf ? buf : "unknown (not enough RAM)");
808 LCONSOLE_WARN("%s: Recovery already passed deadline "
809 "%d:%.02d. If you do not want to wait "
810 "more, please abort the recovery by "
811 "force.\n", target->obd_name,
812 timeout / 60, timeout % 60);
817 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
818 LCONSOLE_WARN("%s: Recovery already passed deadline"
819 " %d:%.02d, It is most likely due to DNE"
820 " recovery is failed or stuck, please wait a"
821 " few more minutes or abort the recovery.\n",
822 target->obd_name, timeout / 60, timeout % 60);
826 conn->cookie = exp->exp_handle.h_cookie;
827 /* target_handle_connect() treats EALREADY and
828 * -EALREADY differently. EALREADY means we are
829 * doing a valid reconnect from the same client. */
834 check_and_start_recovery_timer(struct obd_device *obd,
835 struct ptlrpc_request *req, int new_client);
838 * update flags for import during reconnect process
840 static int rev_import_flags_update(struct obd_import *revimp,
841 struct ptlrpc_request *req)
844 struct obd_connect_data *data;
846 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
848 if (data->ocd_connect_flags & OBD_CONNECT_AT)
849 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
851 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
853 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
855 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
857 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
858 revimp->imp_client->cli_name,
859 libcfs_id2str(req->rq_peer), rc);
867 * Allocate a new reverse import for an export.
869 * \retval -errno in case error hit
870 * \retval 0 if reverse import correctly init
872 int rev_import_init(struct obd_export *export)
874 struct obd_device *obd = export->exp_obd;
875 struct obd_import *revimp;
877 LASSERT(export->exp_imp_reverse == NULL);
879 revimp = class_new_import(obd);
883 revimp->imp_remote_handle.cookie = 0ULL;
884 revimp->imp_client = &obd->obd_ldlm_client;
885 revimp->imp_dlm_fake = 1;
887 /* it is safe to connect import in new state as no sends possible */
888 spin_lock(&export->exp_lock);
889 export->exp_imp_reverse = revimp;
890 spin_unlock(&export->exp_lock);
891 class_import_put(revimp);
895 EXPORT_SYMBOL(rev_import_init);
898 * Handle reconnect for an export.
900 * \param exp export to handle reconnect process
901 * \param req client reconnect request
903 * \retval -rc in case securitfy flavor can't be changed
904 * \retval 0 in case none problems
906 static int rev_import_reconnect(struct obd_export *exp,
907 struct ptlrpc_request *req)
909 struct obd_import *revimp = exp->exp_imp_reverse;
910 struct lustre_handle *lh;
913 /* avoid sending a request until import flags are changed */
914 ptlrpc_import_enter_resend(revimp);
916 if (revimp->imp_connection != NULL)
917 ptlrpc_connection_put(revimp->imp_connection);
920 * client from recovery don't have a handle so we need to take from
921 * request. it may produce situation when wrong client connected
922 * to recovery as we trust a client uuid
924 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
925 revimp->imp_remote_handle = *lh;
927 /* unknown versions will be caught in
928 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
929 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
931 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
933 rc = rev_import_flags_update(revimp, req);
935 /* it is safe to still be in RECOVERY phase as we are not able
936 * to setup correct security flavor so requests are not able to
937 * be delivered correctly */
941 /* resend all rpc's via new connection */
942 return ptlrpc_import_recovery_state_machine(revimp);
945 int target_handle_connect(struct ptlrpc_request *req)
947 struct obd_device *target = NULL;
948 struct obd_export *export = NULL;
949 /* connect handle - filled from target_handle_reconnect in
951 struct lustre_handle conn;
952 struct lustre_handle *tmp;
953 struct obd_uuid tgtuuid;
954 struct obd_uuid cluuid;
959 bool mds_conn = false, lw_client = false, initial_conn = false;
960 bool mds_mds_conn = false;
961 bool new_mds_mds_conn = false;
962 bool target_referenced = false;
963 struct obd_connect_data *data, *tmpdata;
965 lnet_nid_t *client_nid = NULL;
968 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
970 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
972 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
973 GOTO(out, rc = -EINVAL);
976 obd_str2uuid(&tgtuuid, str);
977 target = class_uuid2obd(&tgtuuid);
979 target = class_name2obd(str);
982 deuuidify(str, NULL, &target_start, &target_len);
983 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
984 "from %s (no target). If you are running "
985 "an HA pair check that the target is "
986 "mounted on the other server.\n", str,
987 libcfs_nid2str(req->rq_peer.nid));
988 GOTO(out, rc = -ENODEV);
991 spin_lock(&target->obd_dev_lock);
992 if (target->obd_stopping || !target->obd_set_up) {
993 spin_unlock(&target->obd_dev_lock);
995 deuuidify(str, NULL, &target_start, &target_len);
996 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
997 target_len, target_start,
998 libcfs_nid2str(req->rq_peer.nid),
999 (target->obd_stopping ?
1000 "stopping" : "not set up"));
1001 GOTO(out, rc = -ENODEV);
1004 if (target->obd_no_conn) {
1005 spin_unlock(&target->obd_dev_lock);
1007 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1008 "from %s\n", target->obd_name,
1009 libcfs_nid2str(req->rq_peer.nid));
1010 GOTO(out, rc = -EAGAIN);
1013 /* Make sure the target isn't cleaned up while we're here. Yes,
1014 * there's still a race between the above check and our incref here.
1015 * Really, class_uuid2obd should take the ref. */
1016 class_incref(target, __func__, current);
1017 target_referenced = true;
1019 target->obd_conn_inprogress++;
1020 spin_unlock(&target->obd_dev_lock);
1022 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1024 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1025 GOTO(out, rc = -EINVAL);
1028 obd_str2uuid(&cluuid, str);
1030 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1032 GOTO(out, rc = -EPROTO);
1036 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1038 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1040 GOTO(out, rc = -EPROTO);
1042 rc = req_capsule_server_pack(&req->rq_pill);
1046 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1047 /* Don't allow clients to connect that are using old 1.8 format
1048 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1049 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1050 * FULL20 flag should be set on all connections since 2.0, but no
1051 * longer affects behaviour.
1053 * Later this check will be disabled and the flag can be retired
1054 * completely once interop with 3.0 is no longer needed.
1056 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1057 GOTO(out, rc = -EPROTO);
1060 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1061 if (data->ocd_version < LUSTRE_VERSION_CODE -
1062 LUSTRE_VERSION_ALLOWED_OFFSET ||
1063 data->ocd_version > LUSTRE_VERSION_CODE +
1064 LUSTRE_VERSION_ALLOWED_OFFSET) {
1065 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1066 "libclient connection attempt",
1067 data->ocd_version < LUSTRE_VERSION_CODE ?
1069 OBD_OCD_VERSION_MAJOR(data->ocd_version),
1070 OBD_OCD_VERSION_MINOR(data->ocd_version),
1071 OBD_OCD_VERSION_PATCH(data->ocd_version),
1072 OBD_OCD_VERSION_FIX(data->ocd_version));
1073 data = req_capsule_server_sized_get(&req->rq_pill,
1075 offsetof(typeof(*data), ocd_version) +
1076 sizeof(data->ocd_version));
1078 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1079 data->ocd_version = LUSTRE_VERSION_CODE;
1081 GOTO(out, rc = -EPROTO);
1085 /* Note: lw_client is needed in MDS-MDS failover during update log
1086 * processing, so we needs to allow lw_client to be connected at
1087 * anytime, instead of only the initial connection */
1088 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1090 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1091 initial_conn = true;
1092 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1093 mds_mds_conn = (data->ocd_connect_flags &
1094 OBD_CONNECT_MDS_MDS) != 0;
1096 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1097 * for Imperative Recovery connection from MGC to MGS.
1099 * Via check OBD_CONNECT_FID, we can distinguish whether
1100 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1103 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1104 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1105 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1106 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1107 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1108 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1110 /* We do not support the MDT-MDT interoperations with
1111 * different version MDT because of protocol changes. */
1112 if (unlikely(major != LUSTRE_MAJOR ||
1113 minor != LUSTRE_MINOR ||
1114 abs(patch - LUSTRE_PATCH) > 3)) {
1115 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1116 "connection from different version MDT "
1117 "(%d.%d.%d.%d) %s %s\n",
1118 target->obd_name, LUSTRE_MAJOR,
1119 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1120 major, minor, patch,
1121 OBD_OCD_VERSION_FIX(data->ocd_version),
1122 libcfs_nid2str(req->rq_peer.nid), str);
1124 GOTO(out, rc = -EPROTO);
1129 /* lctl gets a backstage, all-access pass. */
1130 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1131 goto dont_check_exports;
1133 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1137 /* We've found an export in the hash. */
1139 spin_lock(&export->exp_lock);
1141 if (export->exp_connecting) { /* bug 9635, et. al. */
1142 spin_unlock(&export->exp_lock);
1143 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1144 export->exp_obd->obd_name, export,
1145 libcfs_nid2str(req->rq_peer.nid));
1146 class_export_put(export);
1149 } else if ((mds_conn || (lw_client && initial_conn) ||
1150 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1151 export->exp_connection != NULL) {
1152 spin_unlock(&export->exp_lock);
1153 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1154 /* MDS or LWP reconnected after failover. */
1155 LCONSOLE_WARN("%s: Received %s connection from "
1156 "%s, removing former export from %s\n",
1157 target->obd_name, mds_conn ? "MDS" : "LWP",
1158 libcfs_nid2str(req->rq_peer.nid),
1159 libcfs_nid2str(export->exp_connection->c_peer.nid));
1161 /* New MDS connection from the same NID. */
1162 LCONSOLE_WARN("%s: Received new %s connection from "
1163 "%s, removing former export from same NID\n",
1164 target->obd_name, mds_conn ? "MDS" : "LWP",
1165 libcfs_nid2str(req->rq_peer.nid));
1168 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1169 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1170 /* Because exports between MDTs will always be
1171 * kept, let's do not fail such export if they
1172 * come from the same NID, otherwise it might
1173 * cause eviction between MDTs, which might
1174 * cause namespace inconsistency */
1175 spin_lock(&export->exp_lock);
1176 export->exp_connecting = 1;
1177 spin_unlock(&export->exp_lock);
1178 conn.cookie = export->exp_handle.h_cookie;
1181 class_fail_export(export);
1182 class_export_put(export);
1186 } else if (export->exp_connection != NULL && initial_conn &&
1187 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1188 spin_unlock(&export->exp_lock);
1189 /* In MDS failover we have static UUID but NID can change. */
1190 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1191 "existing nid %s is already connected\n",
1192 target->obd_name, cluuid.uuid,
1193 libcfs_nid2str(req->rq_peer.nid),
1195 export->exp_connection->c_peer.nid));
1197 class_export_put(export);
1200 export->exp_connecting = 1;
1201 spin_unlock(&export->exp_lock);
1202 LASSERT(export->exp_obd == target);
1204 rc = target_handle_reconnect(&conn, export, &cluuid);
1207 /* If we found an export, we already unlocked. */
1210 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1211 } else if (req->rq_export == NULL &&
1212 atomic_read(&export->exp_rpc_count) > 0) {
1213 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1214 "still busy with %d references\n",
1215 target->obd_name, cluuid.uuid,
1216 libcfs_nid2str(req->rq_peer.nid),
1217 atomic_read(&export->exp_refcount));
1218 GOTO(out, rc = -EBUSY);
1219 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1220 if (!strstr(cluuid.uuid, "mdt"))
1221 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1222 "known client %s (at %s) because it "
1223 "is indicating it is a new client",
1224 target->obd_name, cluuid.uuid,
1225 libcfs_nid2str(req->rq_peer.nid));
1226 GOTO(out, rc = -EALREADY);
1228 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1235 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1236 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1237 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1238 export, (long)cfs_time_current_sec(),
1239 export ? (long)export->exp_last_request_time : 0);
1241 /* If this is the first time a client connects, reset the recovery
1242 * timer. Discard lightweight connections which might be local. */
1243 if (!lw_client && rc == 0 && target->obd_recovering)
1244 check_and_start_recovery_timer(target, req, export == NULL);
1246 /* We want to handle EALREADY but *not* -EALREADY from
1247 * target_handle_reconnect(), return reconnection state in a flag. */
1248 if (rc == EALREADY) {
1249 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1255 /* Tell the client if we support replayable requests. */
1256 if (target->obd_replayable)
1257 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1258 client_nid = &req->rq_peer.nid;
1260 if (export == NULL) {
1261 /* allow lightweight connections during recovery */
1262 /* allow "new" MDT to be connected during recovery, since we
1263 * need retrieve recovery update records from it */
1264 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1266 int c; /* connected */
1267 int i; /* in progress */
1269 int s; /* stale/evicted */
1271 c = atomic_read(&target->obd_connected_clients);
1272 i = atomic_read(&target->obd_lock_replay_clients);
1273 k = target->obd_max_recoverable_clients;
1274 s = target->obd_stale_clients;
1275 t = target->obd_recovery_timer.expires;
1276 t = cfs_time_sub(t, cfs_time_current());
1277 t = cfs_duration_sec(t);
1278 LCONSOLE_WARN("%s: Denying connection for new client %s"
1279 "(at %s), waiting for %d known clients "
1280 "(%d recovered, %d in progress, and %d "
1281 "evicted) to recover in %d:%.02d\n",
1282 target->obd_name, cluuid.uuid,
1283 libcfs_nid2str(req->rq_peer.nid), k,
1284 c - i, i, s, (int)t / 60,
1289 rc = obd_connect(req->rq_svc_thread->t_env,
1290 &export, target, &cluuid, data,
1292 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1293 lustre_msg_add_op_flags(req->rq_repmsg,
1294 MSG_CONNECT_RECOVERING);
1296 conn.cookie = export->exp_handle.h_cookie;
1297 rc = rev_import_init(export);
1301 new_mds_mds_conn = true;
1304 rc = obd_reconnect(req->rq_svc_thread->t_env,
1305 export, target, &cluuid, data, client_nid);
1310 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1311 data->ocd_instance = target->u.obt.obt_instance;
1313 /* Return only the parts of obd_connect_data that we understand, so the
1314 * client knows that we don't understand the rest. */
1316 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1318 tmpdata = req_capsule_server_get(&req->rq_pill,
1320 /* Don't use struct assignment here, because the client reply
1321 * buffer may be smaller/larger than the local struct
1322 * obd_connect_data. */
1323 memcpy(tmpdata, data, min(tmpsize, size));
1326 /* If the client and the server are the same node, we will already
1327 * have an export that really points to the client's DLM export,
1328 * because we have a shared handles table.
1330 * XXX this will go away when shaver stops sending the "connect" handle
1331 * in the real "remote handle" field of the request --phik 24 Apr 2003
1333 ptlrpc_request_change_export(req, export);
1335 spin_lock(&export->exp_lock);
1336 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1337 spin_unlock(&export->exp_lock);
1338 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1339 "or equal conn_cnt: %d >= %d\n",
1340 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1341 export->exp_conn_cnt,
1342 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1344 GOTO(out, rc = -EALREADY);
1346 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1347 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1349 /* Don't evict liblustre clients for not pinging. */
1350 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1351 export->exp_libclient = 1;
1352 spin_unlock(&export->exp_lock);
1354 spin_lock(&target->obd_dev_lock);
1355 list_del_init(&export->exp_obd_chain_timed);
1356 spin_unlock(&target->obd_dev_lock);
1358 spin_unlock(&export->exp_lock);
1361 if (export->exp_connection != NULL) {
1362 /* Check to see if connection came from another NID. */
1363 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1364 !hlist_unhashed(&export->exp_nid_hash))
1365 cfs_hash_del(export->exp_obd->obd_nid_hash,
1366 &export->exp_connection->c_peer.nid,
1367 &export->exp_nid_hash);
1369 ptlrpc_connection_put(export->exp_connection);
1372 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1375 if (hlist_unhashed(&export->exp_nid_hash))
1376 cfs_hash_add(export->exp_obd->obd_nid_hash,
1377 &export->exp_connection->c_peer.nid,
1378 &export->exp_nid_hash);
1380 lustre_msg_set_handle(req->rq_repmsg, &conn);
1382 rc = rev_import_reconnect(export, req);
1386 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1388 __u64 transno = data->ocd_transno;
1390 spin_lock(&export->exp_lock);
1391 /* possible race with class_disconnect_stale_exports,
1392 * export may be already in the eviction process */
1393 if (export->exp_failed) {
1394 spin_unlock(&export->exp_lock);
1395 GOTO(out, rc = -ENODEV);
1397 export->exp_in_recovery = 1;
1398 export->exp_req_replay_needed = 1;
1399 export->exp_lock_replay_needed = 1;
1400 spin_unlock(&export->exp_lock);
1402 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1403 MSG_CONNECT_TRANSNO);
1404 if (has_transno && transno == 0)
1405 CWARN("Connect with zero transno!\n");
1407 if (has_transno && transno > 0 &&
1408 transno < target->obd_next_recovery_transno &&
1409 transno > target->obd_last_committed) {
1410 /* Another way is to use cmpxchg() to be lock-free. */
1411 spin_lock(&target->obd_recovery_task_lock);
1412 if (transno < target->obd_next_recovery_transno)
1413 target->obd_next_recovery_transno = transno;
1414 spin_unlock(&target->obd_recovery_task_lock);
1417 atomic_inc(&target->obd_req_replay_clients);
1418 atomic_inc(&target->obd_lock_replay_clients);
1419 /* Note: MDS-MDS connection is allowed to be connected during
1420 * recovery, no matter if the exports needs to be recoveried.
1421 * Because we need retrieve updates logs from all other MDTs.
1422 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1423 * also needs to be increased to match other recovery checking
1425 if (new_mds_mds_conn)
1426 target->obd_max_recoverable_clients++;
1427 if (atomic_inc_return(&target->obd_connected_clients) ==
1428 target->obd_max_recoverable_clients)
1429 wake_up(&target->obd_next_transno_waitq);
1432 /* Tell the client we're in recovery, when client is involved in it. */
1433 if (target->obd_recovering && !lw_client)
1434 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1438 spin_lock(&export->exp_lock);
1439 export->exp_connecting = 0;
1440 spin_unlock(&export->exp_lock);
1442 class_export_put(export);
1444 if (target_referenced == true && target != NULL) {
1445 spin_lock(&target->obd_dev_lock);
1446 target->obd_conn_inprogress--;
1447 spin_unlock(&target->obd_dev_lock);
1449 class_decref(target, __func__, current);
1451 req->rq_status = rc;
1455 int target_handle_disconnect(struct ptlrpc_request *req)
1460 rc = req_capsule_server_pack(&req->rq_pill);
1464 /* Keep the rq_export around so we can send the reply. */
1465 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1470 void target_destroy_export(struct obd_export *exp)
1472 struct obd_import *imp = NULL;
1473 /* exports created from last_rcvd data, and "fake"
1474 exports created by lctl don't have an import */
1475 spin_lock(&exp->exp_lock);
1476 if (exp->exp_imp_reverse != NULL) {
1477 imp = exp->exp_imp_reverse;
1478 exp->exp_imp_reverse = NULL;
1480 spin_unlock(&exp->exp_lock);
1482 client_destroy_import(imp);
1484 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1485 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1486 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1487 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1489 EXPORT_SYMBOL(target_destroy_export);
1492 * Recovery functions
1494 static void target_request_copy_get(struct ptlrpc_request *req)
1496 class_export_rpc_inc(req->rq_export);
1497 LASSERT(list_empty(&req->rq_list));
1498 INIT_LIST_HEAD(&req->rq_replay_list);
1500 /* Increase refcount to keep request in queue. */
1501 atomic_inc(&req->rq_refcount);
1502 /* Let export know it has replays to be handled. */
1503 atomic_inc(&req->rq_export->exp_replay_count);
1506 static void target_request_copy_put(struct ptlrpc_request *req)
1508 LASSERT(list_empty(&req->rq_replay_list));
1509 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1511 atomic_dec(&req->rq_export->exp_replay_count);
1512 class_export_rpc_dec(req->rq_export);
1513 ptlrpc_server_drop_request(req);
1516 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1518 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1519 struct obd_export *exp = req->rq_export;
1520 struct ptlrpc_request *reqiter;
1521 struct ptlrpc_request *dup_req = NULL;
1526 spin_lock(&exp->exp_lock);
1527 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1529 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1537 /* We expect it with RESENT and REPLAY flags. */
1538 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1539 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1540 CERROR("invalid flags %x of resent replay\n",
1541 lustre_msg_get_flags(req->rq_reqmsg));
1543 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1546 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1548 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1549 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1553 list_add_tail(&req->rq_replay_list,
1554 &exp->exp_req_replay_queue);
1557 spin_unlock(&exp->exp_lock);
1561 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1563 LASSERT(!list_empty(&req->rq_replay_list));
1564 LASSERT(req->rq_export);
1566 spin_lock(&req->rq_export->exp_lock);
1567 list_del_init(&req->rq_replay_list);
1568 spin_unlock(&req->rq_export->exp_lock);
1571 static void target_finish_recovery(struct lu_target *lut)
1573 struct obd_device *obd = lut->lut_obd;
1576 /* Only log a recovery message when recovery has occurred. */
1577 if (obd->obd_recovery_start) {
1578 time64_t now = ktime_get_real_seconds();
1579 time64_t elapsed_time;
1581 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1582 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1583 "%d recovered and %d %s evicted.\n", obd->obd_name,
1584 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1585 obd->obd_max_recoverable_clients,
1586 atomic_read(&obd->obd_connected_clients),
1587 obd->obd_stale_clients,
1588 obd->obd_stale_clients == 1 ? "was" : "were");
1591 ldlm_reprocess_recovery_done(obd->obd_namespace);
1592 spin_lock(&obd->obd_recovery_task_lock);
1593 if (!list_empty(&obd->obd_req_replay_queue) ||
1594 !list_empty(&obd->obd_lock_replay_queue) ||
1595 !list_empty(&obd->obd_final_req_queue)) {
1596 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1598 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1599 list_empty(&obd->obd_lock_replay_queue) ? \
1601 list_empty(&obd->obd_final_req_queue) ? \
1603 spin_unlock(&obd->obd_recovery_task_lock);
1606 spin_unlock(&obd->obd_recovery_task_lock);
1608 obd->obd_recovery_end = ktime_get_real_seconds();
1610 /* When recovery finished, cleanup orphans on MDS and OST. */
1611 if (OBT(obd) && OBP(obd, postrecov)) {
1612 int rc = OBP(obd, postrecov)(obd);
1614 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1620 static void abort_req_replay_queue(struct obd_device *obd)
1622 struct ptlrpc_request *req, *n;
1623 struct list_head abort_list;
1625 INIT_LIST_HEAD(&abort_list);
1626 spin_lock(&obd->obd_recovery_task_lock);
1627 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1628 spin_unlock(&obd->obd_recovery_task_lock);
1629 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1630 DEBUG_REQ(D_WARNING, req, "aborted:");
1631 req->rq_status = -ENOTCONN;
1632 if (ptlrpc_error(req)) {
1633 DEBUG_REQ(D_ERROR, req,
1634 "failed abort_req_reply; skipping");
1636 target_exp_dequeue_req_replay(req);
1637 target_request_copy_put(req);
1641 static void abort_lock_replay_queue(struct obd_device *obd)
1643 struct ptlrpc_request *req, *n;
1644 struct list_head abort_list;
1646 INIT_LIST_HEAD(&abort_list);
1647 spin_lock(&obd->obd_recovery_task_lock);
1648 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1649 spin_unlock(&obd->obd_recovery_task_lock);
1650 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1651 DEBUG_REQ(D_ERROR, req, "aborted:");
1652 req->rq_status = -ENOTCONN;
1653 if (ptlrpc_error(req)) {
1654 DEBUG_REQ(D_ERROR, req,
1655 "failed abort_lock_reply; skipping");
1657 target_request_copy_put(req);
1661 /* Called from a cleanup function if the device is being cleaned up
1662 forcefully. The exports should all have been disconnected already,
1663 the only thing left to do is
1664 - clear the recovery flags
1666 - free queued requests and replies, but don't send replies
1667 Because the obd_stopping flag is set, no new requests should be received.
1670 void target_cleanup_recovery(struct obd_device *obd)
1672 struct ptlrpc_request *req, *n;
1673 struct list_head clean_list;
1675 INIT_LIST_HEAD(&clean_list);
1676 spin_lock(&obd->obd_dev_lock);
1677 if (!obd->obd_recovering) {
1678 spin_unlock(&obd->obd_dev_lock);
1682 obd->obd_recovering = obd->obd_abort_recovery = 0;
1683 spin_unlock(&obd->obd_dev_lock);
1685 spin_lock(&obd->obd_recovery_task_lock);
1686 target_cancel_recovery_timer(obd);
1687 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1688 spin_unlock(&obd->obd_recovery_task_lock);
1690 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1691 LASSERT(req->rq_reply_state == NULL);
1692 target_exp_dequeue_req_replay(req);
1693 target_request_copy_put(req);
1696 spin_lock(&obd->obd_recovery_task_lock);
1697 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1698 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1699 spin_unlock(&obd->obd_recovery_task_lock);
1701 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1702 LASSERT(req->rq_reply_state == NULL);
1703 target_request_copy_put(req);
1708 EXPORT_SYMBOL(target_cleanup_recovery);
1710 /* obd_recovery_task_lock should be held */
1711 void target_cancel_recovery_timer(struct obd_device *obd)
1713 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1714 del_timer(&obd->obd_recovery_timer);
1717 static void target_start_recovery_timer(struct obd_device *obd)
1719 if (obd->obd_recovery_start != 0)
1722 spin_lock(&obd->obd_dev_lock);
1723 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1724 spin_unlock(&obd->obd_dev_lock);
1728 LASSERT(obd->obd_recovery_timeout != 0);
1730 if (obd->obd_recovery_start != 0) {
1731 spin_unlock(&obd->obd_dev_lock);
1735 mod_timer(&obd->obd_recovery_timer,
1736 cfs_time_shift(obd->obd_recovery_timeout));
1737 obd->obd_recovery_start = ktime_get_real_seconds();
1738 spin_unlock(&obd->obd_dev_lock);
1740 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1742 obd->obd_recovery_timeout / 60,
1743 obd->obd_recovery_timeout % 60,
1744 obd->obd_max_recoverable_clients,
1745 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1746 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1750 * extend recovery window.
1752 * if @extend is true, extend recovery window to have @drt remaining at least;
1753 * otherwise, make sure the recovery timeout value is not less than @drt.
1755 static void extend_recovery_timer(struct obd_device *obd, int drt,
1763 spin_lock(&obd->obd_dev_lock);
1764 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1765 spin_unlock(&obd->obd_dev_lock);
1768 LASSERT(obd->obd_recovery_start != 0);
1770 now = ktime_get_real_seconds();
1771 to = obd->obd_recovery_timeout;
1772 end = obd->obd_recovery_start + to;
1775 if (extend && (drt > left)) {
1777 } else if (!extend && (drt > to)) {
1781 if (to > obd->obd_recovery_time_hard) {
1782 to = obd->obd_recovery_time_hard;
1783 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1784 obd->obd_name, to, extend);
1787 if (obd->obd_recovery_timeout < to) {
1788 obd->obd_recovery_timeout = to;
1789 end = obd->obd_recovery_start + to;
1790 mod_timer(&obd->obd_recovery_timer,
1791 cfs_time_shift(end - now));
1793 spin_unlock(&obd->obd_dev_lock);
1795 CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1796 obd->obd_name, (s64)(end - now));
1799 /* Reset the timer with each new client connection */
1801 * This timer is actually reconnect_timer, which is for making sure
1802 * the total recovery window is at least as big as my reconnect
1803 * attempt timing. So the initial recovery time_out will be set to
1804 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1805 * from client is bigger than this, then the recovery time_out will
1806 * be extended to make sure the client could be reconnected, in the
1807 * process, the timeout from the new client should be ignored.
1811 check_and_start_recovery_timer(struct obd_device *obd,
1812 struct ptlrpc_request *req,
1815 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1816 struct obd_device_target *obt = &obd->u.obt;
1818 if (!new_client && service_time)
1819 /* Teach server about old server's estimates, as first guess
1820 * at how long new requests will take. */
1821 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1824 target_start_recovery_timer(obd);
1826 /* Convert the service time to RPC timeout,
1827 * and reuse service_time to limit stack usage. */
1828 service_time = at_est2timeout(service_time);
1830 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1831 service_time < at_extra)
1832 service_time = at_extra;
1834 /* We expect other clients to timeout within service_time, then try
1835 * to reconnect, then try the failover server. The max delay between
1836 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1837 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1839 LASSERT(obt->obt_magic == OBT_MAGIC);
1840 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1841 if (service_time > obd->obd_recovery_timeout && !new_client)
1842 extend_recovery_timer(obd, service_time, false);
1845 /** Health checking routines */
1846 static inline int exp_connect_healthy(struct obd_export *exp)
1848 return (exp->exp_in_recovery);
1851 /** if export done req_replay or has replay in queue */
1852 static inline int exp_req_replay_healthy(struct obd_export *exp)
1854 return (!exp->exp_req_replay_needed ||
1855 atomic_read(&exp->exp_replay_count) > 0);
1859 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1861 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1862 exp_req_replay_healthy(exp);
1865 /** if export done lock_replay or has replay in queue */
1866 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1868 return (!exp->exp_lock_replay_needed ||
1869 atomic_read(&exp->exp_replay_count) > 0);
1872 static inline int exp_vbr_healthy(struct obd_export *exp)
1874 return (!exp->exp_vbr_failed);
1877 static inline int exp_finished(struct obd_export *exp)
1879 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1882 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1884 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1888 static int check_for_next_transno(struct lu_target *lut)
1890 struct ptlrpc_request *req = NULL;
1891 struct obd_device *obd = lut->lut_obd;
1892 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1893 int wake_up = 0, connected, completed, queue_len;
1894 __u64 req_transno = 0;
1895 __u64 update_transno = 0;
1896 __u64 next_transno = 0;
1899 spin_lock(&obd->obd_recovery_task_lock);
1900 if (!list_empty(&obd->obd_req_replay_queue)) {
1901 req = list_entry(obd->obd_req_replay_queue.next,
1902 struct ptlrpc_request, rq_list);
1903 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1907 update_transno = distribute_txn_get_next_transno(tdtd);
1909 connected = atomic_read(&obd->obd_connected_clients);
1910 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1911 queue_len = obd->obd_requests_queued_for_recovery;
1912 next_transno = obd->obd_next_recovery_transno;
1914 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1915 "req_transno: %llu, next_transno: %llu\n",
1916 obd->obd_max_recoverable_clients, connected, completed,
1917 queue_len, req_transno, next_transno);
1919 if (obd->obd_abort_recovery) {
1920 CDEBUG(D_HA, "waking for aborted recovery\n");
1922 } else if (obd->obd_recovery_expired) {
1923 CDEBUG(D_HA, "waking for expired recovery\n");
1925 } else if (tdtd != NULL && req != NULL &&
1926 is_req_replayed_by_update(req)) {
1927 LASSERTF(req_transno < next_transno, "req_transno %llu"
1928 "next_transno%llu\n", req_transno, next_transno);
1929 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1932 } else if (req_transno == next_transno ||
1933 (update_transno != 0 && update_transno <= next_transno)) {
1934 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1936 } else if (queue_len > 0 &&
1937 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1938 /** handle gaps occured due to lost reply or VBR */
1939 LASSERTF(req_transno >= next_transno,
1940 "req_transno: %llu, next_transno: %llu\n",
1941 req_transno, next_transno);
1943 "%s: waking for gap in transno, VBR is %s (skip: "
1944 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1945 ", next_update %lld last_committed: %lld)\n",
1946 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1947 next_transno, queue_len, completed, connected,
1948 req_transno, update_transno, obd->obd_last_committed);
1949 obd->obd_next_recovery_transno = req_transno;
1951 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1952 CDEBUG(D_HA, "waking for completed recovery\n");
1954 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1955 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1956 " by fail_lock, waking up (%lld)\n", next_transno);
1957 obd->obd_next_recovery_transno = req_transno;
1960 spin_unlock(&obd->obd_recovery_task_lock);
1964 static int check_for_next_lock(struct lu_target *lut)
1966 struct obd_device *obd = lut->lut_obd;
1969 spin_lock(&obd->obd_recovery_task_lock);
1970 if (!list_empty(&obd->obd_lock_replay_queue)) {
1971 CDEBUG(D_HA, "waking for next lock\n");
1973 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1974 CDEBUG(D_HA, "waking for completed lock replay\n");
1976 } else if (obd->obd_abort_recovery) {
1977 CDEBUG(D_HA, "waking for aborted recovery\n");
1979 } else if (obd->obd_recovery_expired) {
1980 CDEBUG(D_HA, "waking for expired recovery\n");
1983 spin_unlock(&obd->obd_recovery_task_lock);
1989 * wait for recovery events,
1990 * check its status with help of check_routine
1991 * evict dead clients via health_check
1993 static int target_recovery_overseer(struct lu_target *lut,
1994 int (*check_routine)(struct lu_target *),
1995 int (*health_check)(struct obd_export *))
1997 struct obd_device *obd = lut->lut_obd;
1998 struct target_distribute_txn_data *tdtd;
2000 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
2001 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2002 __u64 next_update_transno = 0;
2004 /* Only abort the recovery if there are no update recovery
2005 * left in the queue */
2006 spin_lock(&obd->obd_recovery_task_lock);
2007 if (lut->lut_tdtd != NULL) {
2008 next_update_transno =
2009 distribute_txn_get_next_transno(lut->lut_tdtd);
2011 tdtd = lut->lut_tdtd;
2012 /* If next_update_transno == 0, it probably because
2013 * updatelog retrieve threads did not get any records
2014 * yet, let's wait those threads stopped */
2015 if (next_update_transno == 0) {
2016 struct l_wait_info lwi = { 0 };
2018 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2020 &tdtd->tdtd_recovery_threads_count) == 0,
2023 next_update_transno =
2024 distribute_txn_get_next_transno(
2029 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2030 obd->obd_next_recovery_transno = next_update_transno;
2031 spin_unlock(&obd->obd_recovery_task_lock);
2032 /* Disconnect unfinished exports from clients, and
2033 * keep connection from MDT to make sure the update
2034 * recovery will still keep trying until some one
2035 * manually abort the recovery */
2036 class_disconnect_stale_exports(obd,
2037 exp_finished_or_from_mdt);
2038 /* Abort all of replay and replay lock req from
2040 abort_req_replay_queue(obd);
2041 abort_lock_replay_queue(obd);
2042 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2043 ")in the queue.\n", obd->obd_name,
2044 next_update_transno);
2046 obd->obd_abort_recovery = 1;
2047 spin_unlock(&obd->obd_recovery_task_lock);
2048 CWARN("%s recovery is aborted by hard timeout\n",
2053 while (wait_event_timeout(obd->obd_next_transno_waitq,
2055 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2056 /* wait indefinitely for event, but don't trigger watchdog */;
2058 if (obd->obd_abort_recovery) {
2059 CWARN("recovery is aborted, evict exports in recovery\n");
2060 if (lut->lut_tdtd != NULL) {
2061 struct l_wait_info lwi = { 0 };
2063 tdtd = lut->lut_tdtd;
2064 /* Let's wait all of the update log recovery thread
2066 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2067 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2069 /* Then abort the update recovery list */
2070 dtrq_list_destroy(lut->lut_tdtd);
2073 /** evict exports which didn't finish recovery yet */
2074 class_disconnect_stale_exports(obd, exp_finished);
2076 } else if (obd->obd_recovery_expired) {
2077 obd->obd_recovery_expired = 0;
2078 /** If some clients died being recovered, evict them */
2079 LCONSOLE_WARN("%s: recovery is timed out, "
2080 "evict stale exports\n", obd->obd_name);
2081 /** evict cexports with no replay in queue, they are stalled */
2082 class_disconnect_stale_exports(obd, health_check);
2084 /** continue with VBR */
2085 spin_lock(&obd->obd_dev_lock);
2086 obd->obd_version_recov = 1;
2087 spin_unlock(&obd->obd_dev_lock);
2089 * reset timer, recovery will proceed with versions now,
2090 * timeout is set just to handle reconnection delays
2092 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2093 /** Wait for recovery events again, after evicting bad clients */
2099 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2101 struct obd_device *obd = lut->lut_obd;
2102 struct ptlrpc_request *req = NULL;
2104 CDEBUG(D_HA, "Waiting for lock\n");
2105 if (target_recovery_overseer(lut, check_for_next_lock,
2106 exp_lock_replay_healthy))
2107 abort_lock_replay_queue(obd);
2109 spin_lock(&obd->obd_recovery_task_lock);
2110 if (!list_empty(&obd->obd_lock_replay_queue)) {
2111 req = list_entry(obd->obd_lock_replay_queue.next,
2112 struct ptlrpc_request, rq_list);
2113 list_del_init(&req->rq_list);
2114 spin_unlock(&obd->obd_recovery_task_lock);
2116 spin_unlock(&obd->obd_recovery_task_lock);
2117 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2118 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2119 /** evict exports failed VBR */
2120 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2125 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2127 struct ptlrpc_request *req = NULL;
2129 spin_lock(&obd->obd_recovery_task_lock);
2130 if (!list_empty(&obd->obd_final_req_queue)) {
2131 req = list_entry(obd->obd_final_req_queue.next,
2132 struct ptlrpc_request, rq_list);
2133 list_del_init(&req->rq_list);
2134 spin_unlock(&obd->obd_recovery_task_lock);
2135 if (req->rq_export->exp_in_recovery) {
2136 spin_lock(&req->rq_export->exp_lock);
2137 req->rq_export->exp_in_recovery = 0;
2138 spin_unlock(&req->rq_export->exp_lock);
2141 spin_unlock(&obd->obd_recovery_task_lock);
2146 static void handle_recovery_req(struct ptlrpc_thread *thread,
2147 struct ptlrpc_request *req,
2148 svc_handler_t handler)
2153 * export can be evicted during recovery, no need to handle replays for
2154 * it after that, discard such request silently
2156 if (req->rq_export->exp_disconnected)
2159 req->rq_session.lc_thread = thread;
2160 req->rq_svc_thread = thread;
2161 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2163 /* thread context */
2164 lu_context_enter(&thread->t_env->le_ctx);
2166 lu_context_exit(&thread->t_env->le_ctx);
2168 /* don't reset timer for final stage */
2169 if (!exp_finished(req->rq_export)) {
2170 int to = obd_timeout;
2173 * Add request timeout to the recovery time so next request from
2174 * this client may come in recovery time
2177 struct ptlrpc_service_part *svcpt;
2179 svcpt = req->rq_rqbd->rqbd_svcpt;
2180 /* If the server sent early reply for this request,
2181 * the client will recalculate the timeout according to
2182 * current server estimate service time, so we will
2183 * use the maxium timeout here for waiting the client
2184 * sending the next req */
2185 to = max((int)at_est2timeout(
2186 at_get(&svcpt->scp_at_estimate)),
2187 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2188 /* Add 2 net_latency, one for balance rq_deadline
2189 * (see ptl_send_rpc), one for resend the req to server,
2190 * Note: client will pack net_latency in replay req
2191 * (see ptlrpc_replay_req) */
2192 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2194 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2199 /** Checking routines for recovery */
2200 static int check_for_recovery_ready(struct lu_target *lut)
2202 struct obd_device *obd = lut->lut_obd;
2203 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2205 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2206 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2207 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2208 obd->obd_recovery_expired);
2210 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2211 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2212 if (clnts + obd->obd_stale_clients <
2213 obd->obd_max_recoverable_clients)
2217 if (lut->lut_tdtd != NULL) {
2218 if (!lut->lut_tdtd->tdtd_replay_ready &&
2219 !obd->obd_abort_recovery) {
2220 /* Let's extend recovery timer, in case the recovery
2221 * timer expired, and some clients got evicted */
2222 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2224 CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2225 obd->obd_name, obd->obd_recovery_timeout);
2234 REQUEST_RECOVERY = 1,
2235 UPDATE_RECOVERY = 2,
2238 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2242 if (!list_empty(&obd->obd_req_replay_queue)) {
2243 struct ptlrpc_request *req;
2245 req = list_entry(obd->obd_req_replay_queue.next,
2246 struct ptlrpc_request, rq_list);
2247 transno = lustre_msg_get_transno(req->rq_reqmsg);
2253 static __u64 get_next_transno(struct lu_target *lut, int *type)
2255 struct obd_device *obd = lut->lut_obd;
2256 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2258 __u64 update_transno;
2261 transno = get_next_replay_req_transno(obd);
2263 *type = REQUEST_RECOVERY;
2268 update_transno = distribute_txn_get_next_transno(tdtd);
2269 if (transno == 0 || (transno >= update_transno &&
2270 update_transno != 0)) {
2271 transno = update_transno;
2273 *type = UPDATE_RECOVERY;
2280 * drop duplicate replay request
2282 * Because the operation has been replayed by update recovery, the request
2283 * with the same transno will be dropped and also notify the client to send
2284 * next replay request.
2286 * \param[in] env execution environment
2287 * \param[in] obd failover obd device
2288 * \param[in] req request to be dropped
2290 static void drop_duplicate_replay_req(struct lu_env *env,
2291 struct obd_device *obd,
2292 struct ptlrpc_request *req)
2294 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2295 " update records are found.\n",
2296 lustre_msg_get_transno(req->rq_reqmsg),
2297 libcfs_nid2str(req->rq_peer.nid));
2299 /* Right now, only for MDS reint operation update replay and
2300 * normal request replay can have the same transno */
2301 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2302 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2303 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2304 if (likely(req->rq_export))
2305 target_committed_to_req(req);
2306 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2307 target_send_reply(req, req->rq_status, 0);
2309 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2310 libcfs_nid2str(req->rq_peer.nid));
2312 target_exp_dequeue_req_replay(req);
2313 target_request_copy_put(req);
2314 obd->obd_replayed_requests++;
2317 static void replay_request_or_update(struct lu_env *env,
2318 struct lu_target *lut,
2319 struct target_recovery_data *trd,
2320 struct ptlrpc_thread *thread)
2322 struct obd_device *obd = lut->lut_obd;
2323 struct ptlrpc_request *req = NULL;
2328 CDEBUG(D_HA, "Waiting for transno %lld\n",
2329 obd->obd_next_recovery_transno);
2331 /* Replay all of request and update by transno */
2333 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2335 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2337 /** It is needed to extend recovery window above
2338 * recovery_time_soft. Extending is possible only in the
2339 * end of recovery window (see more details in
2340 * handle_recovery_req()).
2342 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2344 if (target_recovery_overseer(lut, check_for_next_transno,
2345 exp_req_replay_healthy_or_from_mdt)) {
2346 abort_req_replay_queue(obd);
2347 abort_lock_replay_queue(obd);
2351 spin_lock(&obd->obd_recovery_task_lock);
2352 transno = get_next_transno(lut, &type);
2353 if (type == REQUEST_RECOVERY && transno != 0) {
2354 /* Drop replay request from client side, if the
2355 * replay has been executed by update with the
2357 req = list_entry(obd->obd_req_replay_queue.next,
2358 struct ptlrpc_request, rq_list);
2360 list_del_init(&req->rq_list);
2361 obd->obd_requests_queued_for_recovery--;
2362 spin_unlock(&obd->obd_recovery_task_lock);
2364 /* Let's check if the request has been redone by
2366 if (is_req_replayed_by_update(req)) {
2367 struct distribute_txn_replay_req *dtrq;
2369 dtrq = distribute_txn_lookup_finish_list(tdtd,
2371 LASSERT(dtrq != NULL);
2372 spin_lock(&tdtd->tdtd_replay_list_lock);
2373 list_del_init(&dtrq->dtrq_list);
2374 spin_unlock(&tdtd->tdtd_replay_list_lock);
2377 drop_duplicate_replay_req(env, obd, req);
2382 LASSERT(trd->trd_processing_task == current_pid());
2383 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2384 lustre_msg_get_transno(req->rq_reqmsg),
2385 libcfs_nid2str(req->rq_peer.nid));
2387 handle_recovery_req(thread, req,
2388 trd->trd_recovery_handler);
2390 * bz18031: increase next_recovery_transno before
2391 * target_request_copy_put() will drop exp_rpc reference
2393 spin_lock(&obd->obd_recovery_task_lock);
2394 obd->obd_next_recovery_transno++;
2395 spin_unlock(&obd->obd_recovery_task_lock);
2396 target_exp_dequeue_req_replay(req);
2397 target_request_copy_put(req);
2398 obd->obd_replayed_requests++;
2399 } else if (type == UPDATE_RECOVERY && transno != 0) {
2400 struct distribute_txn_replay_req *dtrq;
2403 spin_unlock(&obd->obd_recovery_task_lock);
2405 LASSERT(tdtd != NULL);
2406 dtrq = distribute_txn_get_next_req(tdtd);
2407 lu_context_enter(&thread->t_env->le_ctx);
2408 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2409 lu_context_exit(&thread->t_env->le_ctx);
2410 extend_recovery_timer(obd, obd_timeout, true);
2412 if (rc == 0 && dtrq->dtrq_xid != 0) {
2413 CDEBUG(D_HA, "Move x%llu t%llu"
2414 " to finish list\n", dtrq->dtrq_xid,
2415 dtrq->dtrq_master_transno);
2417 /* Add it to the replay finish list */
2418 spin_lock(&tdtd->tdtd_replay_list_lock);
2419 list_add(&dtrq->dtrq_list,
2420 &tdtd->tdtd_replay_finish_list);
2421 spin_unlock(&tdtd->tdtd_replay_list_lock);
2423 spin_lock(&obd->obd_recovery_task_lock);
2424 if (transno == obd->obd_next_recovery_transno)
2425 obd->obd_next_recovery_transno++;
2427 obd->obd_next_recovery_transno)
2428 obd->obd_next_recovery_transno =
2430 spin_unlock(&obd->obd_recovery_task_lock);
2435 spin_unlock(&obd->obd_recovery_task_lock);
2437 LASSERT(list_empty(&obd->obd_req_replay_queue));
2438 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2439 /** evict exports failed VBR */
2440 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2446 static int target_recovery_thread(void *arg)
2448 struct lu_target *lut = arg;
2449 struct obd_device *obd = lut->lut_obd;
2450 struct ptlrpc_request *req;
2451 struct target_recovery_data *trd = &obd->obd_recovery_data;
2452 unsigned long delta;
2454 struct ptlrpc_thread *thread = NULL;
2458 unshare_fs_struct();
2459 OBD_ALLOC_PTR(thread);
2465 OBD_FREE_PTR(thread);
2469 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2471 OBD_FREE_PTR(thread);
2476 thread->t_env = env;
2477 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2478 env->le_ctx.lc_thread = thread;
2479 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2480 thread->t_watchdog = NULL;
2482 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2484 trd->trd_processing_task = current_pid();
2486 spin_lock(&obd->obd_dev_lock);
2487 obd->obd_recovering = 1;
2488 spin_unlock(&obd->obd_dev_lock);
2489 complete(&trd->trd_starting);
2491 /* first of all, we have to know the first transno to replay */
2492 if (target_recovery_overseer(lut, check_for_recovery_ready,
2493 exp_connect_healthy)) {
2494 abort_req_replay_queue(obd);
2495 abort_lock_replay_queue(obd);
2496 if (lut->lut_tdtd != NULL)
2497 dtrq_list_destroy(lut->lut_tdtd);
2500 /* next stage: replay requests or update */
2502 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2503 atomic_read(&obd->obd_req_replay_clients),
2504 obd->obd_next_recovery_transno);
2505 replay_request_or_update(env, lut, trd, thread);
2508 * The second stage: replay locks
2510 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2511 atomic_read(&obd->obd_lock_replay_clients));
2512 while ((req = target_next_replay_lock(lut))) {
2513 LASSERT(trd->trd_processing_task == current_pid());
2514 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2515 libcfs_nid2str(req->rq_peer.nid));
2516 handle_recovery_req(thread, req,
2517 trd->trd_recovery_handler);
2518 target_request_copy_put(req);
2519 obd->obd_replayed_locks++;
2523 * The third stage: reply on final pings, at this moment all clients
2524 * must have request in final queue
2526 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2527 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2528 /** Update server last boot epoch */
2529 tgt_boot_epoch_update(lut);
2530 /* We drop recoverying flag to forward all new requests
2531 * to regular mds_handle() since now */
2532 spin_lock(&obd->obd_dev_lock);
2533 obd->obd_recovering = obd->obd_abort_recovery = 0;
2534 spin_unlock(&obd->obd_dev_lock);
2535 spin_lock(&obd->obd_recovery_task_lock);
2536 target_cancel_recovery_timer(obd);
2537 spin_unlock(&obd->obd_recovery_task_lock);
2538 while ((req = target_next_final_ping(obd))) {
2539 LASSERT(trd->trd_processing_task == current_pid());
2540 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2541 libcfs_nid2str(req->rq_peer.nid));
2542 handle_recovery_req(thread, req,
2543 trd->trd_recovery_handler);
2544 /* Because the waiting client can not send ping to server,
2545 * so we need refresh the last_request_time, to avoid the
2546 * export is being evicted */
2547 ptlrpc_update_export_timer(req->rq_export, 0);
2548 target_request_copy_put(req);
2551 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2552 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2553 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2554 if (delta > OBD_RECOVERY_TIME_SOFT) {
2555 CWARN("too long recovery - read logs\n");
2556 libcfs_debug_dumplog();
2559 target_finish_recovery(lut);
2561 lu_context_fini(&env->le_ctx);
2562 trd->trd_processing_task = 0;
2563 complete(&trd->trd_finishing);
2565 tgt_io_thread_done(thread);
2566 OBD_FREE_PTR(thread);
2571 static int target_start_recovery_thread(struct lu_target *lut,
2572 svc_handler_t handler)
2574 struct obd_device *obd = lut->lut_obd;
2576 struct target_recovery_data *trd = &obd->obd_recovery_data;
2579 memset(trd, 0, sizeof(*trd));
2580 init_completion(&trd->trd_starting);
2581 init_completion(&trd->trd_finishing);
2582 trd->trd_recovery_handler = handler;
2584 rc = server_name2index(obd->obd_name, &index, NULL);
2588 if (!IS_ERR(kthread_run(target_recovery_thread,
2589 lut, "tgt_recover_%d", index))) {
2590 wait_for_completion(&trd->trd_starting);
2591 LASSERT(obd->obd_recovering != 0);
2599 void target_stop_recovery_thread(struct obd_device *obd)
2601 if (obd->obd_recovery_data.trd_processing_task > 0) {
2602 struct target_recovery_data *trd = &obd->obd_recovery_data;
2603 /** recovery can be done but postrecovery is not yet */
2604 spin_lock(&obd->obd_dev_lock);
2605 if (obd->obd_recovering) {
2606 CERROR("%s: Aborting recovery\n", obd->obd_name);
2607 obd->obd_abort_recovery = 1;
2608 wake_up(&obd->obd_next_transno_waitq);
2610 spin_unlock(&obd->obd_dev_lock);
2611 wait_for_completion(&trd->trd_finishing);
2614 EXPORT_SYMBOL(target_stop_recovery_thread);
2616 void target_recovery_fini(struct obd_device *obd)
2618 class_disconnect_exports(obd);
2619 target_stop_recovery_thread(obd);
2620 target_cleanup_recovery(obd);
2622 EXPORT_SYMBOL(target_recovery_fini);
2624 static void target_recovery_expired(unsigned long castmeharder)
2626 struct obd_device *obd = (struct obd_device *)castmeharder;
2627 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2628 " after %llus (%d clients connected)\n",
2629 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2630 (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2631 atomic_read(&obd->obd_connected_clients));
2633 obd->obd_recovery_expired = 1;
2634 wake_up(&obd->obd_next_transno_waitq);
2637 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2639 struct obd_device *obd = lut->lut_obd;
2641 if (lut->lut_bottom->dd_rdonly)
2644 if (obd->obd_max_recoverable_clients == 0) {
2645 /** Update server last boot epoch */
2646 tgt_boot_epoch_update(lut);
2650 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2651 "last_transno %llu\n", obd->obd_name,
2652 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2653 LASSERT(obd->obd_stopping == 0);
2654 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2655 obd->obd_recovery_start = 0;
2656 obd->obd_recovery_end = 0;
2658 setup_timer(&obd->obd_recovery_timer, target_recovery_expired,
2659 (unsigned long)obd);
2660 target_start_recovery_thread(lut, handler);
2662 EXPORT_SYMBOL(target_recovery_init);
2664 static int target_process_req_flags(struct obd_device *obd,
2665 struct ptlrpc_request *req)
2667 struct obd_export *exp = req->rq_export;
2668 LASSERT(exp != NULL);
2669 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2670 /* client declares he's ready to replay locks */
2671 spin_lock(&exp->exp_lock);
2672 if (exp->exp_req_replay_needed) {
2673 exp->exp_req_replay_needed = 0;
2674 spin_unlock(&exp->exp_lock);
2676 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2677 atomic_dec(&obd->obd_req_replay_clients);
2679 spin_unlock(&exp->exp_lock);
2682 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2683 /* client declares he's ready to complete recovery
2684 * so, we put the request on th final queue */
2685 spin_lock(&exp->exp_lock);
2686 if (exp->exp_lock_replay_needed) {
2687 exp->exp_lock_replay_needed = 0;
2688 spin_unlock(&exp->exp_lock);
2690 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2691 atomic_dec(&obd->obd_lock_replay_clients);
2693 spin_unlock(&exp->exp_lock);
2699 int target_queue_recovery_request(struct ptlrpc_request *req,
2700 struct obd_device *obd)
2702 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2703 struct ptlrpc_request *reqiter;
2707 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2708 /* Processing the queue right now, don't re-add. */
2712 target_process_req_flags(obd, req);
2714 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2715 /* client declares he's ready to complete recovery
2716 * so, we put the request on th final queue */
2717 target_request_copy_get(req);
2718 DEBUG_REQ(D_HA, req, "queue final req");
2719 wake_up(&obd->obd_next_transno_waitq);
2720 spin_lock(&obd->obd_recovery_task_lock);
2721 if (obd->obd_recovering) {
2722 struct ptlrpc_request *tmp;
2723 struct ptlrpc_request *duplicate = NULL;
2725 if (likely(!req->rq_export->exp_replay_done)) {
2726 req->rq_export->exp_replay_done = 1;
2727 list_add_tail(&req->rq_list,
2728 &obd->obd_final_req_queue);
2729 spin_unlock(&obd->obd_recovery_task_lock);
2733 /* XXX O(n), but only happens if final ping is
2734 * timed out, probably reorganize the list as
2735 * a hash list later */
2736 list_for_each_entry_safe(reqiter, tmp,
2737 &obd->obd_final_req_queue,
2739 if (reqiter->rq_export == req->rq_export) {
2740 list_del_init(&reqiter->rq_list);
2741 duplicate = reqiter;
2746 list_add_tail(&req->rq_list,
2747 &obd->obd_final_req_queue);
2748 req->rq_export->exp_replay_done = 1;
2749 spin_unlock(&obd->obd_recovery_task_lock);
2751 if (duplicate != NULL) {
2752 DEBUG_REQ(D_HA, duplicate,
2753 "put prev final req\n");
2754 target_request_copy_put(duplicate);
2758 spin_unlock(&obd->obd_recovery_task_lock);
2759 target_request_copy_put(req);
2760 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2763 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2764 /* client declares he's ready to replay locks */
2765 target_request_copy_get(req);
2766 DEBUG_REQ(D_HA, req, "queue lock replay req");
2767 wake_up(&obd->obd_next_transno_waitq);
2768 spin_lock(&obd->obd_recovery_task_lock);
2769 LASSERT(obd->obd_recovering);
2770 /* usually due to recovery abort */
2771 if (!req->rq_export->exp_in_recovery) {
2772 spin_unlock(&obd->obd_recovery_task_lock);
2773 target_request_copy_put(req);
2776 LASSERT(req->rq_export->exp_lock_replay_needed);
2777 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2778 spin_unlock(&obd->obd_recovery_task_lock);
2782 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2783 * (i.e. buflens etc are in my own byte order), but type-dependent
2784 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2787 INIT_LIST_HEAD(&req->rq_list);
2788 DEBUG_REQ(D_HA, req, "not queueing");
2792 /* If we're processing the queue, we want don't want to queue this
2795 * Also, if this request has a transno less than the one we're waiting
2796 * for, we should process it now. It could (and currently always will)
2797 * be an open request for a descriptor that was opened some time ago.
2799 * Also, a resent, replayed request that has already been
2800 * handled will pass through here and be processed immediately.
2802 CDEBUG(D_HA, "Next recovery transno: %llu"
2803 ", current: %llu, replaying\n",
2804 obd->obd_next_recovery_transno, transno);
2806 /* If the request has been replayed by update replay, then sends this
2807 * request to the recovery thread (replay_request_or_update()), where
2808 * it will be handled */
2809 spin_lock(&obd->obd_recovery_task_lock);
2810 if (transno < obd->obd_next_recovery_transno &&
2811 !is_req_replayed_by_update(req)) {
2812 /* Processing the queue right now, don't re-add. */
2813 LASSERT(list_empty(&req->rq_list));
2814 spin_unlock(&obd->obd_recovery_task_lock);
2817 spin_unlock(&obd->obd_recovery_task_lock);
2819 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2822 target_request_copy_get(req);
2823 if (!req->rq_export->exp_in_recovery) {
2824 target_request_copy_put(req);
2827 LASSERT(req->rq_export->exp_req_replay_needed);
2829 if (target_exp_enqueue_req_replay(req)) {
2830 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2831 target_request_copy_put(req);
2836 spin_lock(&obd->obd_recovery_task_lock);
2837 LASSERT(obd->obd_recovering);
2838 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2839 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2840 list_add_tail(&req->rq_list, &reqiter->rq_list);
2845 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2847 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2848 "has been claimed by another client");
2849 spin_unlock(&obd->obd_recovery_task_lock);
2850 target_exp_dequeue_req_replay(req);
2851 target_request_copy_put(req);
2857 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2859 obd->obd_requests_queued_for_recovery++;
2860 spin_unlock(&obd->obd_recovery_task_lock);
2861 wake_up(&obd->obd_next_transno_waitq);
2865 int target_handle_ping(struct ptlrpc_request *req)
2867 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2868 return req_capsule_server_pack(&req->rq_pill);
2871 void target_committed_to_req(struct ptlrpc_request *req)
2873 struct obd_export *exp = req->rq_export;
2875 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2876 lustre_msg_set_last_committed(req->rq_repmsg,
2877 exp->exp_last_committed);
2879 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2880 "%d)", exp->exp_obd->obd_no_transno,
2881 req->rq_repmsg == NULL);
2883 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2884 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2887 #endif /* HAVE_SERVER_SUPPORT */
2890 * Packs current SLV and Limit into \a req.
2892 int target_pack_pool_reply(struct ptlrpc_request *req)
2894 struct obd_device *obd;
2897 /* Check that we still have all structures alive as this may
2898 * be some late RPC at shutdown time. */
2899 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2900 !exp_connect_lru_resize(req->rq_export))) {
2901 lustre_msg_set_slv(req->rq_repmsg, 0);
2902 lustre_msg_set_limit(req->rq_repmsg, 0);
2906 /* OBD is alive here as export is alive, which we checked above. */
2907 obd = req->rq_export->exp_obd;
2909 read_lock(&obd->obd_pool_lock);
2910 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2911 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2912 read_unlock(&obd->obd_pool_lock);
2917 static int target_send_reply_msg(struct ptlrpc_request *req,
2918 int rc, int fail_id)
2920 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2921 DEBUG_REQ(D_ERROR, req, "dropping reply");
2924 /* We can have a null rq_reqmsg in the event of bad signature or
2925 * no context when unwrapping */
2926 if (req->rq_reqmsg &&
2927 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2928 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2932 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2933 req->rq_status = rc;
2934 return ptlrpc_send_error(req, 1);
2936 DEBUG_REQ(D_NET, req, "sending reply");
2939 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2942 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2944 struct ptlrpc_service_part *svcpt;
2946 struct ptlrpc_reply_state *rs;
2947 struct obd_export *exp;
2950 if (req->rq_no_reply) {
2955 svcpt = req->rq_rqbd->rqbd_svcpt;
2956 rs = req->rq_reply_state;
2957 if (rs == NULL || !rs->rs_difficult) {
2959 target_send_reply_msg (req, rc, fail_id);
2964 /* must be an export if locks saved */
2965 LASSERT(req->rq_export != NULL);
2966 /* req/reply consistent */
2967 LASSERT(rs->rs_svcpt == svcpt);
2970 LASSERT(!rs->rs_scheduled);
2971 LASSERT(!rs->rs_scheduled_ever);
2972 LASSERT(!rs->rs_handled);
2973 LASSERT(!rs->rs_on_net);
2974 LASSERT(rs->rs_export == NULL);
2975 LASSERT(list_empty(&rs->rs_obd_list));
2976 LASSERT(list_empty(&rs->rs_exp_list));
2978 exp = class_export_get(req->rq_export);
2980 /* disable reply scheduling while I'm setting up */
2981 rs->rs_scheduled = 1;
2983 rs->rs_xid = req->rq_xid;
2984 rs->rs_transno = req->rq_transno;
2985 rs->rs_export = exp;
2986 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2988 spin_lock(&exp->exp_uncommitted_replies_lock);
2989 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2990 rs->rs_transno, exp->exp_last_committed);
2991 if (rs->rs_transno > exp->exp_last_committed) {
2992 /* not committed already */
2993 list_add_tail(&rs->rs_obd_list,
2994 &exp->exp_uncommitted_replies);
2996 spin_unlock(&exp->exp_uncommitted_replies_lock);
2998 spin_lock(&exp->exp_lock);
2999 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3000 spin_unlock(&exp->exp_lock);
3002 netrc = target_send_reply_msg(req, rc, fail_id);
3004 spin_lock(&svcpt->scp_rep_lock);
3006 atomic_inc(&svcpt->scp_nreps_difficult);
3009 /* error sending: reply is off the net. Also we need +1
3010 * reply ref until ptlrpc_handle_rs() is done
3011 * with the reply state (if the send was successful, there
3012 * would have been +1 ref for the net, which
3013 * reply_out_callback leaves alone) */
3015 ptlrpc_rs_addref(rs);
3018 spin_lock(&rs->rs_lock);
3019 if (rs->rs_transno <= exp->exp_last_committed ||
3020 (!rs->rs_on_net && !rs->rs_no_ack) ||
3021 list_empty(&rs->rs_exp_list) || /* completed already */
3022 list_empty(&rs->rs_obd_list)) {
3023 CDEBUG(D_HA, "Schedule reply immediately\n");
3024 ptlrpc_dispatch_difficult_reply(rs);
3026 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3027 rs->rs_scheduled = 0; /* allow notifier to schedule */
3029 spin_unlock(&rs->rs_lock);
3030 spin_unlock(&svcpt->scp_rep_lock);
3034 enum ldlm_mode lck_compat_array[] = {
3035 [LCK_EX] = LCK_COMPAT_EX,
3036 [LCK_PW] = LCK_COMPAT_PW,
3037 [LCK_PR] = LCK_COMPAT_PR,
3038 [LCK_CW] = LCK_COMPAT_CW,
3039 [LCK_CR] = LCK_COMPAT_CR,
3040 [LCK_NL] = LCK_COMPAT_NL,
3041 [LCK_GROUP] = LCK_COMPAT_GROUP,
3042 [LCK_COS] = LCK_COMPAT_COS,
3046 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3047 * not escape to the user level.
3049 int ldlm_error2errno(enum ldlm_error error)
3055 case ELDLM_LOCK_MATCHED:
3058 case ELDLM_LOCK_CHANGED:
3061 case ELDLM_LOCK_ABORTED:
3064 case ELDLM_LOCK_REPLACED:
3067 case ELDLM_NO_LOCK_DATA:
3070 case ELDLM_NAMESPACE_EXISTS:
3073 case ELDLM_BAD_NAMESPACE:
3077 if (((int)error) < 0) { /* cast to signed type */
3078 result = error; /* as ldlm_error can be unsigned */
3080 CERROR("Invalid DLM result code: %d\n", error);
3086 EXPORT_SYMBOL(ldlm_error2errno);
3089 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3091 enum ldlm_error ldlm_errno2error(int err_no)
3100 error = ELDLM_LOCK_CHANGED;
3103 error = ELDLM_LOCK_ABORTED;
3106 error = ELDLM_LOCK_REPLACED;
3109 error = ELDLM_NO_LOCK_DATA;
3112 error = ELDLM_NAMESPACE_EXISTS;
3115 error = ELDLM_BAD_NAMESPACE;
3123 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3124 void ldlm_dump_export_locks(struct obd_export *exp)
3126 spin_lock(&exp->exp_locks_list_guard);
3127 if (!list_empty(&exp->exp_locks_list)) {
3128 struct ldlm_lock *lock;
3130 CERROR("dumping locks for export %p,"
3131 "ignore if the unmount doesn't hang\n", exp);
3132 list_for_each_entry(lock, &exp->exp_locks_list,
3134 LDLM_ERROR(lock, "lock:");
3136 spin_unlock(&exp->exp_locks_list_guard);
3140 #ifdef HAVE_SERVER_SUPPORT
3141 static int target_bulk_timeout(void *data)
3144 /* We don't fail the connection here, because having the export
3145 * killed makes the (vital) call to commitrw very sad.
3150 static inline const char *bulk2type(struct ptlrpc_request *req)
3152 if (req->rq_bulk_read)
3154 if (req->rq_bulk_write)
3159 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3160 struct l_wait_info *lwi)
3162 struct ptlrpc_request *req = desc->bd_req;
3163 time_t start = cfs_time_current_sec();
3169 /* If there is eviction in progress, wait for it to finish. */
3170 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3171 *lwi = LWI_INTR(NULL, NULL);
3172 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3173 !atomic_read(&exp->exp_obd->
3174 obd_evict_inprogress),
3178 /* Check if client was evicted or reconnected already. */
3179 if (exp->exp_failed ||
3180 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3183 if (req->rq_bulk_read)
3184 rc = sptlrpc_svc_wrap_bulk(req, desc);
3186 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3187 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3188 else /* old version, bulk matchbits is rq_xid */
3189 req->rq_mbits = req->rq_xid;
3192 rc = ptlrpc_start_bulk_transfer(desc);
3196 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3197 bulk2type(req), rc);
3201 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3202 ptlrpc_abort_bulk(desc);
3206 /* limit actual bulk transfer to bulk_timeout seconds */
3207 deadline = start + bulk_timeout;
3208 if (deadline > req->rq_deadline)
3209 deadline = req->rq_deadline;
3212 long timeoutl = deadline - cfs_time_current_sec();
3213 cfs_duration_t timeout = timeoutl <= 0 ?
3214 CFS_TICK : cfs_time_seconds(timeoutl);
3217 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3218 target_bulk_timeout, desc);
3219 rc = l_wait_event(desc->bd_waitq,
3220 !ptlrpc_server_bulk_active(desc) ||
3223 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3225 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3226 /* Wait again if we changed rq_deadline. */
3227 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3228 deadline = start + bulk_timeout;
3229 if (deadline > rq_deadline)
3230 deadline = rq_deadline;
3231 } while ((rc == -ETIMEDOUT) &&
3232 (deadline > cfs_time_current_sec()));
3234 if (rc == -ETIMEDOUT) {
3235 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3236 bulk2type(req), deadline - start,
3237 cfs_time_current_sec() - deadline);
3238 ptlrpc_abort_bulk(desc);
3239 } else if (exp->exp_failed) {
3240 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3243 ptlrpc_abort_bulk(desc);
3244 } else if (exp->exp_conn_cnt >
3245 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3246 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3248 /* We don't reply anyway. */
3250 ptlrpc_abort_bulk(desc);
3251 } else if (desc->bd_failure) {
3252 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3254 /* XXX should this be a different errno? */
3257 if (req->rq_bulk_write)
3258 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3259 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3260 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3261 bulk2type(req), desc->bd_nob_transferred,
3263 /* XXX should this be a different errno? */
3270 EXPORT_SYMBOL(target_bulk_io);
3272 #endif /* HAVE_SERVER_SUPPORT */