4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * This file deals with various client/target related logic including recovery.
36 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
40 #define DEBUG_SUBSYSTEM S_LDLM
42 #include <linux/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_LIST_HEAD(&cli->cl_flight_waiters);
395 cli->cl_rpcs_in_flight = 0;
397 init_waitqueue_head(&cli->cl_destroy_waitq);
398 atomic_set(&cli->cl_destroy_in_flight, 0);
399 #ifdef ENABLE_CHECKSUM
400 /* Turn on checksumming by default. */
401 cli->cl_checksum = 1;
403 * The supported checksum types will be worked out at connect time
404 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
407 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
409 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
411 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
412 * from OFD after connecting. */
413 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
415 cli->cl_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
417 /* set cl_chunkbits default value to PAGE_SHIFT,
418 * it will be updated at OSC connection time. */
419 cli->cl_chunkbits = PAGE_SHIFT;
421 if (!strcmp(name, LUSTRE_MDC_NAME)) {
422 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
423 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
424 cli->cl_max_rpcs_in_flight = 2;
425 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
426 cli->cl_max_rpcs_in_flight = 3;
427 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
428 cli->cl_max_rpcs_in_flight = 4;
430 if (osc_on_mdt(obddev->obd_name))
431 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
433 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
436 spin_lock_init(&cli->cl_mod_rpcs_lock);
437 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
438 cli->cl_max_mod_rpcs_in_flight = 0;
439 cli->cl_mod_rpcs_in_flight = 0;
440 cli->cl_close_rpcs_in_flight = 0;
441 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
442 cli->cl_mod_tag_bitmap = NULL;
444 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
446 if (connect_op == MDS_CONNECT) {
447 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
448 OBD_ALLOC(cli->cl_mod_tag_bitmap,
449 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
450 if (cli->cl_mod_tag_bitmap == NULL)
451 GOTO(err, rc = -ENOMEM);
456 CERROR("ldlm_get_ref failed: %d\n", rc);
460 ptlrpc_init_client(rq_portal, rp_portal, name,
461 &obddev->obd_ldlm_client);
463 imp = class_new_import(obddev);
465 GOTO(err_ldlm, rc = -ENOENT);
466 imp->imp_client = &obddev->obd_ldlm_client;
467 imp->imp_connect_op = connect_op;
468 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
469 LUSTRE_CFG_BUFLEN(lcfg, 1));
470 class_import_put(imp);
472 if (lustre_cfg_buf(lcfg, 4)) {
473 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
475 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
477 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
478 obddev->obd_name, lustre_cfg_string(lcfg, 4),
480 GOTO(err_import, rc);
482 fake_conn.c_self = LNET_MKNID(refnet, 0);
483 imp->imp_connection = &fake_conn;
486 rc = client_import_add_conn(imp, &server_uuid, 1);
488 CERROR("can't add initial connection\n");
489 GOTO(err_import, rc);
491 imp->imp_connection = NULL;
493 cli->cl_import = imp;
494 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
495 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
497 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
498 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
499 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
500 name, obddev->obd_name,
501 cli->cl_target_uuid.uuid);
502 spin_lock(&imp->imp_lock);
503 imp->imp_deactive = 1;
504 spin_unlock(&imp->imp_lock);
508 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
509 LDLM_NAMESPACE_CLIENT,
510 LDLM_NAMESPACE_GREEDY,
512 if (obddev->obd_namespace == NULL) {
513 CERROR("Unable to create client namespace - %s\n",
515 GOTO(err_import, rc = -ENOMEM);
521 class_destroy_import(imp);
525 if (cli->cl_mod_tag_bitmap != NULL)
526 OBD_FREE(cli->cl_mod_tag_bitmap,
527 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
528 cli->cl_mod_tag_bitmap = NULL;
532 EXPORT_SYMBOL(client_obd_setup);
534 int client_obd_cleanup(struct obd_device *obddev)
536 struct client_obd *cli = &obddev->u.cli;
539 ldlm_namespace_free_post(obddev->obd_namespace);
540 obddev->obd_namespace = NULL;
542 obd_cleanup_client_import(obddev);
543 LASSERT(obddev->u.cli.cl_import == NULL);
547 if (cli->cl_mod_tag_bitmap != NULL)
548 OBD_FREE(cli->cl_mod_tag_bitmap,
549 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
550 cli->cl_mod_tag_bitmap = NULL;
554 EXPORT_SYMBOL(client_obd_cleanup);
556 /* ->o_connect() method for client side (OSC and MDC and MGC) */
557 int client_connect_import(const struct lu_env *env,
558 struct obd_export **exp,
559 struct obd_device *obd, struct obd_uuid *cluuid,
560 struct obd_connect_data *data, void *localdata)
562 struct client_obd *cli = &obd->u.cli;
563 struct obd_import *imp = cli->cl_import;
564 struct obd_connect_data *ocd;
565 struct lustre_handle conn = { 0 };
570 down_write(&cli->cl_sem);
571 if (cli->cl_conn_count > 0)
572 GOTO(out_sem, rc = -EALREADY);
574 rc = class_connect(&conn, obd, cluuid);
578 cli->cl_conn_count++;
579 *exp = class_conn2export(&conn);
581 LASSERT(obd->obd_namespace);
583 imp->imp_dlm_handle = conn;
584 rc = ptlrpc_init_import(imp);
588 ocd = &imp->imp_connect_data;
591 imp->imp_connect_flags_orig = data->ocd_connect_flags;
592 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
595 rc = ptlrpc_connect_import(imp);
597 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
600 LASSERT(*exp != NULL && (*exp)->exp_connection);
603 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
604 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
605 data->ocd_connect_flags, ocd->ocd_connect_flags);
606 data->ocd_connect_flags = ocd->ocd_connect_flags;
607 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
610 ptlrpc_pinger_add_import(imp);
616 cli->cl_conn_count--;
617 class_disconnect(*exp);
621 up_write(&cli->cl_sem);
625 EXPORT_SYMBOL(client_connect_import);
627 int client_disconnect_export(struct obd_export *exp)
629 struct obd_device *obd = class_exp2obd(exp);
630 struct client_obd *cli;
631 struct obd_import *imp;
636 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
637 exp, exp ? exp->exp_handle.h_cookie : -1);
642 imp = cli->cl_import;
644 down_write(&cli->cl_sem);
645 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
648 if (cli->cl_conn_count == 0) {
649 CERROR("disconnecting disconnected device (%s)\n",
651 GOTO(out_disconnect, rc = -EINVAL);
654 cli->cl_conn_count--;
655 if (cli->cl_conn_count != 0)
656 GOTO(out_disconnect, rc = 0);
658 /* Mark import deactivated now, so we don't try to reconnect if any
659 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
660 * fully deactivate the import, or that would drop all requests. */
661 spin_lock(&imp->imp_lock);
662 imp->imp_deactive = 1;
663 spin_unlock(&imp->imp_lock);
665 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
666 * delete it regardless. (It's safe to delete an import that was
668 (void)ptlrpc_pinger_del_import(imp);
670 if (obd->obd_namespace != NULL) {
671 /* obd_force == local only */
672 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
673 obd->obd_force ? LCF_LOCAL : 0, NULL);
674 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
677 /* There's no need to hold sem while disconnecting an import,
678 * and it may actually cause deadlock in GSS. */
679 up_write(&cli->cl_sem);
680 rc = ptlrpc_disconnect_import(imp, 0);
681 down_write(&cli->cl_sem);
683 ptlrpc_invalidate_import(imp);
688 /* Use server style - class_disconnect should be always called for
690 err = class_disconnect(exp);
694 up_write(&cli->cl_sem);
698 EXPORT_SYMBOL(client_disconnect_export);
700 #ifdef HAVE_SERVER_SUPPORT
701 int server_disconnect_export(struct obd_export *exp)
706 /* Disconnect early so that clients can't keep using export. */
707 rc = class_disconnect(exp);
708 /* Close import to avoid sending any requests. */
709 if (exp->exp_imp_reverse)
710 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
712 ldlm_bl_thread_wakeup();
714 /* complete all outstanding replies */
715 spin_lock(&exp->exp_lock);
716 while (!list_empty(&exp->exp_outstanding_replies)) {
717 struct ptlrpc_reply_state *rs =
718 list_entry(exp->exp_outstanding_replies.next,
719 struct ptlrpc_reply_state, rs_exp_list);
720 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
722 spin_lock(&svcpt->scp_rep_lock);
724 list_del_init(&rs->rs_exp_list);
726 spin_lock(&rs->rs_lock);
727 /* clear rs_convert_lock to make sure rs is handled and put */
728 rs->rs_convert_lock = 0;
729 ptlrpc_schedule_difficult_reply(rs);
730 spin_unlock(&rs->rs_lock);
732 spin_unlock(&svcpt->scp_rep_lock);
734 spin_unlock(&exp->exp_lock);
738 EXPORT_SYMBOL(server_disconnect_export);
740 /* --------------------------------------------------------------------------
741 * from old lib/target.c
742 * -------------------------------------------------------------------------- */
744 static int target_handle_reconnect(struct lustre_handle *conn,
745 struct obd_export *exp,
746 struct obd_uuid *cluuid)
748 struct obd_device *target;
749 struct lustre_handle *hdl;
756 hdl = &exp->exp_imp_reverse->imp_remote_handle;
757 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
758 conn->cookie = exp->exp_handle.h_cookie;
759 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
760 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
764 target = exp->exp_obd;
766 /* Might be a re-connect after a partition. */
767 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
768 LCONSOLE_WARN("%s: already connected client %s (at %s) "
769 "with handle %#llx. Rejecting client "
770 "with the same UUID trying to reconnect "
771 "with handle %#llx\n", target->obd_name,
772 obd_uuid2str(&exp->exp_client_uuid),
773 obd_export_nid2str(exp),
774 hdl->cookie, conn->cookie);
775 memset(conn, 0, sizeof *conn);
776 /* target_handle_connect() treats EALREADY and
777 * -EALREADY differently. -EALREADY is an error
778 * (same UUID, different handle). */
782 if (!target->obd_recovering) {
783 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
784 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
785 obd_export_nid2str(exp));
786 GOTO(out_already, rc);
789 now = ktime_get_seconds();
790 deadline = cfs_duration_sec(target->obd_recovery_timer.expires);
791 if (now < deadline) {
792 struct target_distribute_txn_data *tdtd;
797 timeout = deadline - now;
798 tdtd = class_exp2tgt(exp)->lut_tdtd;
799 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
800 buf = tdtd->tdtd_show_update_logs_retrievers(
801 tdtd->tdtd_show_retrievers_cbdata,
805 LCONSOLE_WARN("%s: Recovery already passed deadline "
806 "%lld:%.02lld. It is due to DNE recovery "
807 "failed/stuck on the %d MDT(s):%s. "
808 "Please wait until all MDTs recovered "
809 "or abort the recovery by force.\n",
810 target->obd_name, timeout / 60,
812 buf ? buf : "unknown (not enough RAM)");
814 LCONSOLE_WARN("%s: Recovery already passed deadline "
815 "%lld:%.02lld. If you do not want to wait "
816 "more, please abort the recovery by "
817 "force.\n", target->obd_name,
818 timeout / 60, timeout % 60);
823 timeout = now - deadline;
824 LCONSOLE_WARN("%s: Recovery already passed deadline"
825 " %lld:%.02lld, It is most likely due to DNE"
826 " recovery is failed or stuck, please wait a"
827 " few more minutes or abort the recovery.\n",
828 target->obd_name, timeout / 60, timeout % 60);
832 conn->cookie = exp->exp_handle.h_cookie;
833 /* target_handle_connect() treats EALREADY and
834 * -EALREADY differently. EALREADY means we are
835 * doing a valid reconnect from the same client. */
840 check_and_start_recovery_timer(struct obd_device *obd,
841 struct ptlrpc_request *req, int new_client);
844 * update flags for import during reconnect process
846 static int rev_import_flags_update(struct obd_import *revimp,
847 struct ptlrpc_request *req)
850 struct obd_connect_data *data;
852 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
854 if (data->ocd_connect_flags & OBD_CONNECT_AT)
855 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
857 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
859 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
861 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
863 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
864 revimp->imp_client->cli_name,
865 libcfs_id2str(req->rq_peer), rc);
873 * Allocate a new reverse import for an export.
875 * \retval -errno in case error hit
876 * \retval 0 if reverse import correctly init
878 int rev_import_init(struct obd_export *export)
880 struct obd_device *obd = export->exp_obd;
881 struct obd_import *revimp;
883 LASSERT(export->exp_imp_reverse == NULL);
885 revimp = class_new_import(obd);
889 revimp->imp_remote_handle.cookie = 0ULL;
890 revimp->imp_client = &obd->obd_ldlm_client;
891 revimp->imp_dlm_fake = 1;
893 /* it is safe to connect import in new state as no sends possible */
894 spin_lock(&export->exp_lock);
895 export->exp_imp_reverse = revimp;
896 spin_unlock(&export->exp_lock);
897 class_import_put(revimp);
901 EXPORT_SYMBOL(rev_import_init);
904 * Handle reconnect for an export.
906 * \param exp export to handle reconnect process
907 * \param req client reconnect request
909 * \retval -rc in case securitfy flavor can't be changed
910 * \retval 0 in case none problems
912 static int rev_import_reconnect(struct obd_export *exp,
913 struct ptlrpc_request *req)
915 struct obd_import *revimp = exp->exp_imp_reverse;
916 struct lustre_handle *lh;
919 /* avoid sending a request until import flags are changed */
920 ptlrpc_import_enter_resend(revimp);
922 if (revimp->imp_connection != NULL)
923 ptlrpc_connection_put(revimp->imp_connection);
926 * client from recovery don't have a handle so we need to take from
927 * request. it may produce situation when wrong client connected
928 * to recovery as we trust a client uuid
930 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
931 revimp->imp_remote_handle = *lh;
933 /* unknown versions will be caught in
934 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
935 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
937 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
939 rc = rev_import_flags_update(revimp, req);
941 /* it is safe to still be in RECOVERY phase as we are not able
942 * to setup correct security flavor so requests are not able to
943 * be delivered correctly */
947 /* resend all rpc's via new connection */
948 return ptlrpc_import_recovery_state_machine(revimp);
951 int target_handle_connect(struct ptlrpc_request *req)
953 struct obd_device *target = NULL;
954 struct obd_export *export = NULL;
955 /* connect handle - filled from target_handle_reconnect in
957 struct lustre_handle conn;
958 struct lustre_handle *tmp;
959 struct obd_uuid cluuid;
964 bool mds_conn = false, lw_client = false, initial_conn = false;
965 bool mds_mds_conn = false;
966 bool new_mds_mds_conn = false;
967 struct obd_connect_data *data, *tmpdata;
969 lnet_nid_t *client_nid = NULL;
972 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
974 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
976 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
977 GOTO(out, rc = -EINVAL);
980 target = class_dev_by_str(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);
993 target->obd_conn_inprogress++;
995 if (target->obd_stopping || !target->obd_set_up) {
996 spin_unlock(&target->obd_dev_lock);
998 deuuidify(str, NULL, &target_start, &target_len);
999 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1000 target_len, target_start,
1001 libcfs_nid2str(req->rq_peer.nid),
1002 (target->obd_stopping ?
1003 "stopping" : "not set up"));
1004 GOTO(out, rc = -ENODEV);
1007 if (target->obd_no_conn) {
1008 spin_unlock(&target->obd_dev_lock);
1010 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1011 "from %s\n", target->obd_name,
1012 libcfs_nid2str(req->rq_peer.nid));
1013 GOTO(out, rc = -EAGAIN);
1016 spin_unlock(&target->obd_dev_lock);
1018 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1020 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1021 GOTO(out, rc = -EINVAL);
1024 obd_str2uuid(&cluuid, str);
1026 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1028 GOTO(out, rc = -EPROTO);
1032 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1034 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1036 GOTO(out, rc = -EPROTO);
1038 rc = req_capsule_server_pack(&req->rq_pill);
1042 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1043 /* Don't allow clients to connect that are using old 1.8 format
1044 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1045 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1046 * FULL20 flag should be set on all connections since 2.0, but no
1047 * longer affects behaviour.
1049 * Later this check will be disabled and the flag can be retired
1050 * completely once interop with 3.0 is no longer needed.
1052 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1053 GOTO(out, rc = -EPROTO);
1055 /* Don't allow liblustre clients to connect.
1056 * - testing was disabled in v2_2_50_0-61-g6a75d65
1057 * - building was disabled in v2_5_58_0-28-g7277179
1058 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1059 * - clients were refused connect for version difference > 0.0.1.32 */
1060 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1061 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1062 GOTO(out, rc = -EPROTO);
1066 /* Note: lw_client is needed in MDS-MDS failover during update log
1067 * processing, so we needs to allow lw_client to be connected at
1068 * anytime, instead of only the initial connection */
1069 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1071 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1072 initial_conn = true;
1073 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1074 mds_mds_conn = (data->ocd_connect_flags &
1075 OBD_CONNECT_MDS_MDS) != 0;
1077 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1078 * for Imperative Recovery connection from MGC to MGS.
1080 * Via check OBD_CONNECT_FID, we can distinguish whether
1081 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1082 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1085 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1086 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1087 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1088 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1089 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1090 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1092 /* We do not support the MDT-MDT interoperations with
1093 * different version MDT because of protocol changes. */
1094 if (unlikely(major != LUSTRE_MAJOR ||
1095 minor != LUSTRE_MINOR ||
1096 abs(patch - LUSTRE_PATCH) > 3)) {
1097 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1098 "connection from different version MDT "
1099 "(%d.%d.%d.%d) %s %s\n",
1100 target->obd_name, LUSTRE_MAJOR,
1101 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1102 major, minor, patch,
1103 OBD_OCD_VERSION_FIX(data->ocd_version),
1104 libcfs_nid2str(req->rq_peer.nid), str);
1106 GOTO(out, rc = -EPROTO);
1111 /* lctl gets a backstage, all-access pass. */
1112 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1113 goto dont_check_exports;
1115 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1119 /* We've found an export in the hash. */
1121 spin_lock(&export->exp_lock);
1123 if (export->exp_connecting) { /* bug 9635, et. al. */
1124 spin_unlock(&export->exp_lock);
1125 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1126 export->exp_obd->obd_name, export,
1127 libcfs_nid2str(req->rq_peer.nid));
1128 class_export_put(export);
1131 } else if ((mds_conn || (lw_client && initial_conn) ||
1132 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1133 export->exp_connection != NULL) {
1134 spin_unlock(&export->exp_lock);
1135 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1136 /* MDS or LWP reconnected after failover. */
1137 LCONSOLE_WARN("%s: Received %s connection from "
1138 "%s, removing former export from %s\n",
1139 target->obd_name, mds_conn ? "MDS" : "LWP",
1140 libcfs_nid2str(req->rq_peer.nid),
1141 libcfs_nid2str(export->exp_connection->c_peer.nid));
1143 /* New MDS connection from the same NID. */
1144 LCONSOLE_WARN("%s: Received new %s connection from "
1145 "%s, removing former export from same NID\n",
1146 target->obd_name, mds_conn ? "MDS" : "LWP",
1147 libcfs_nid2str(req->rq_peer.nid));
1150 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1151 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1152 /* Because exports between MDTs will always be
1153 * kept, let's do not fail such export if they
1154 * come from the same NID, otherwise it might
1155 * cause eviction between MDTs, which might
1156 * cause namespace inconsistency */
1157 spin_lock(&export->exp_lock);
1158 export->exp_connecting = 1;
1159 spin_unlock(&export->exp_lock);
1160 conn.cookie = export->exp_handle.h_cookie;
1163 class_fail_export(export);
1164 class_export_put(export);
1168 } else if (export->exp_connection != NULL && initial_conn &&
1169 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1170 spin_unlock(&export->exp_lock);
1171 /* In MDS failover we have static UUID but NID can change. */
1172 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1173 "existing nid %s is already connected\n",
1174 target->obd_name, cluuid.uuid,
1175 libcfs_nid2str(req->rq_peer.nid),
1177 export->exp_connection->c_peer.nid));
1179 class_export_put(export);
1182 export->exp_connecting = 1;
1183 spin_unlock(&export->exp_lock);
1184 LASSERT(export->exp_obd == target);
1186 rc = target_handle_reconnect(&conn, export, &cluuid);
1189 /* If we found an export, we already unlocked. */
1192 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1193 } else if (req->rq_export == NULL &&
1194 atomic_read(&export->exp_rpc_count) > 0) {
1195 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1196 "still busy with %d references\n",
1197 target->obd_name, cluuid.uuid,
1198 libcfs_nid2str(req->rq_peer.nid),
1199 atomic_read(&export->exp_refcount));
1200 GOTO(out, rc = -EBUSY);
1201 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1202 if (!strstr(cluuid.uuid, "mdt"))
1203 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1204 "known client %s (at %s) because it "
1205 "is indicating it is a new client",
1206 target->obd_name, cluuid.uuid,
1207 libcfs_nid2str(req->rq_peer.nid));
1208 GOTO(out, rc = -EALREADY);
1210 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1217 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1218 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1219 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1220 export, ktime_get_real_seconds(),
1221 export ? export->exp_last_request_time : 0);
1223 /* If this is the first time a client connects, reset the recovery
1224 * timer. Discard lightweight connections which might be local. */
1225 if (!lw_client && rc == 0 && target->obd_recovering)
1226 check_and_start_recovery_timer(target, req, export == NULL);
1228 /* We want to handle EALREADY but *not* -EALREADY from
1229 * target_handle_reconnect(), return reconnection state in a flag. */
1230 if (rc == EALREADY) {
1231 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1237 /* Tell the client if we support replayable requests. */
1238 if (target->obd_replayable)
1239 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1240 client_nid = &req->rq_peer.nid;
1242 if (export == NULL) {
1243 /* allow lightweight connections during recovery */
1244 /* allow "new" MDT to be connected during recovery, since we
1245 * need retrieve recovery update records from it */
1246 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1248 int c; /* connected */
1249 int i; /* in progress */
1251 int s; /* stale/evicted */
1253 c = atomic_read(&target->obd_connected_clients);
1254 i = atomic_read(&target->obd_lock_replay_clients);
1255 k = target->obd_max_recoverable_clients;
1256 s = target->obd_stale_clients;
1257 t = target->obd_recovery_timer.expires;
1258 t = cfs_duration_sec(target->obd_recovery_timer.expires);
1259 t -= ktime_get_seconds();
1260 LCONSOLE_WARN("%s: Denying connection for new client %s"
1261 "(at %s), waiting for %d known clients "
1262 "(%d recovered, %d in progress, and %d "
1263 "evicted) to recover in %lld:%.02lld\n",
1264 target->obd_name, cluuid.uuid,
1265 libcfs_nid2str(req->rq_peer.nid), k,
1266 c - i, i, s, t / 60, t % 60);
1270 rc = obd_connect(req->rq_svc_thread->t_env,
1271 &export, target, &cluuid, data,
1273 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1274 lustre_msg_add_op_flags(req->rq_repmsg,
1275 MSG_CONNECT_RECOVERING);
1277 conn.cookie = export->exp_handle.h_cookie;
1278 rc = rev_import_init(export);
1282 new_mds_mds_conn = true;
1285 rc = obd_reconnect(req->rq_svc_thread->t_env,
1286 export, target, &cluuid, data, client_nid);
1291 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1292 data->ocd_instance = target->u.obt.obt_instance;
1294 /* Return only the parts of obd_connect_data that we understand, so the
1295 * client knows that we don't understand the rest. */
1297 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1299 tmpdata = req_capsule_server_get(&req->rq_pill,
1301 /* Don't use struct assignment here, because the client reply
1302 * buffer may be smaller/larger than the local struct
1303 * obd_connect_data. */
1304 memcpy(tmpdata, data, min(tmpsize, size));
1307 /* If the client and the server are the same node, we will already
1308 * have an export that really points to the client's DLM export,
1309 * because we have a shared handles table.
1311 * XXX this will go away when shaver stops sending the "connect" handle
1312 * in the real "remote handle" field of the request --phik 24 Apr 2003
1314 ptlrpc_request_change_export(req, export);
1316 spin_lock(&export->exp_lock);
1317 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1318 spin_unlock(&export->exp_lock);
1319 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1320 "or equal conn_cnt: %d >= %d\n",
1321 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1322 export->exp_conn_cnt,
1323 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1325 GOTO(out, rc = -EALREADY);
1327 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1328 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1329 spin_unlock(&export->exp_lock);
1331 if (export->exp_connection != NULL) {
1332 /* Check to see if connection came from another NID. */
1333 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1334 !hlist_unhashed(&export->exp_nid_hash))
1335 cfs_hash_del(export->exp_obd->obd_nid_hash,
1336 &export->exp_connection->c_peer.nid,
1337 &export->exp_nid_hash);
1339 ptlrpc_connection_put(export->exp_connection);
1342 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1345 if (hlist_unhashed(&export->exp_nid_hash))
1346 cfs_hash_add(export->exp_obd->obd_nid_hash,
1347 &export->exp_connection->c_peer.nid,
1348 &export->exp_nid_hash);
1350 lustre_msg_set_handle(req->rq_repmsg, &conn);
1352 rc = rev_import_reconnect(export, req);
1356 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1358 __u64 transno = data->ocd_transno;
1360 spin_lock(&export->exp_lock);
1361 /* possible race with class_disconnect_stale_exports,
1362 * export may be already in the eviction process */
1363 if (export->exp_failed) {
1364 spin_unlock(&export->exp_lock);
1365 GOTO(out, rc = -ENODEV);
1367 export->exp_in_recovery = 1;
1368 export->exp_req_replay_needed = 1;
1369 export->exp_lock_replay_needed = 1;
1370 spin_unlock(&export->exp_lock);
1372 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1373 MSG_CONNECT_TRANSNO);
1374 if (has_transno && transno == 0)
1375 CWARN("Connect with zero transno!\n");
1377 if (has_transno && transno > 0 &&
1378 transno < target->obd_next_recovery_transno &&
1379 transno > target->obd_last_committed) {
1380 /* Another way is to use cmpxchg() to be lock-free. */
1381 spin_lock(&target->obd_recovery_task_lock);
1382 if (transno < target->obd_next_recovery_transno)
1383 target->obd_next_recovery_transno = transno;
1384 spin_unlock(&target->obd_recovery_task_lock);
1387 atomic_inc(&target->obd_req_replay_clients);
1388 atomic_inc(&target->obd_lock_replay_clients);
1389 /* Note: MDS-MDS connection is allowed to be connected during
1390 * recovery, no matter if the exports needs to be recoveried.
1391 * Because we need retrieve updates logs from all other MDTs.
1392 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1393 * also needs to be increased to match other recovery checking
1395 if (new_mds_mds_conn)
1396 target->obd_max_recoverable_clients++;
1397 if (atomic_inc_return(&target->obd_connected_clients) ==
1398 target->obd_max_recoverable_clients)
1399 wake_up(&target->obd_next_transno_waitq);
1402 /* Tell the client we're in recovery, when client is involved in it. */
1403 if (target->obd_recovering && !lw_client)
1404 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1408 spin_lock(&export->exp_lock);
1409 export->exp_connecting = 0;
1410 spin_unlock(&export->exp_lock);
1412 class_export_put(export);
1414 if (target != NULL) {
1415 spin_lock(&target->obd_dev_lock);
1416 target->obd_conn_inprogress--;
1417 spin_unlock(&target->obd_dev_lock);
1418 class_decref(target, "find", current);
1420 req->rq_status = rc;
1424 int target_handle_disconnect(struct ptlrpc_request *req)
1429 rc = req_capsule_server_pack(&req->rq_pill);
1433 /* Keep the rq_export around so we can send the reply. */
1434 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1439 void target_destroy_export(struct obd_export *exp)
1441 struct obd_import *imp = NULL;
1442 /* exports created from last_rcvd data, and "fake"
1443 exports created by lctl don't have an import */
1444 spin_lock(&exp->exp_lock);
1445 if (exp->exp_imp_reverse != NULL) {
1446 imp = exp->exp_imp_reverse;
1447 exp->exp_imp_reverse = NULL;
1449 spin_unlock(&exp->exp_lock);
1451 client_destroy_import(imp);
1453 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1454 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1455 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1456 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1458 EXPORT_SYMBOL(target_destroy_export);
1461 * Recovery functions
1463 static void target_request_copy_get(struct ptlrpc_request *req)
1465 class_export_rpc_inc(req->rq_export);
1466 LASSERT(list_empty(&req->rq_list));
1467 INIT_LIST_HEAD(&req->rq_replay_list);
1469 /* Increase refcount to keep request in queue. */
1470 atomic_inc(&req->rq_refcount);
1471 /* Let export know it has replays to be handled. */
1472 atomic_inc(&req->rq_export->exp_replay_count);
1475 static void target_request_copy_put(struct ptlrpc_request *req)
1477 LASSERT(list_empty(&req->rq_replay_list));
1478 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1480 atomic_dec(&req->rq_export->exp_replay_count);
1481 class_export_rpc_dec(req->rq_export);
1482 ptlrpc_server_drop_request(req);
1485 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1487 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1488 struct obd_export *exp = req->rq_export;
1489 struct ptlrpc_request *reqiter;
1490 struct ptlrpc_request *dup_req = NULL;
1495 spin_lock(&exp->exp_lock);
1496 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1498 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1506 /* We expect it with RESENT and REPLAY flags. */
1507 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1508 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1509 CERROR("invalid flags %x of resent replay\n",
1510 lustre_msg_get_flags(req->rq_reqmsg));
1512 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1515 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1517 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1518 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1522 list_add_tail(&req->rq_replay_list,
1523 &exp->exp_req_replay_queue);
1526 spin_unlock(&exp->exp_lock);
1530 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1532 LASSERT(!list_empty(&req->rq_replay_list));
1533 LASSERT(req->rq_export);
1535 spin_lock(&req->rq_export->exp_lock);
1536 list_del_init(&req->rq_replay_list);
1537 spin_unlock(&req->rq_export->exp_lock);
1540 static void target_finish_recovery(struct lu_target *lut)
1542 struct obd_device *obd = lut->lut_obd;
1545 /* Only log a recovery message when recovery has occurred. */
1546 if (obd->obd_recovery_start) {
1547 time64_t now = ktime_get_real_seconds();
1548 time64_t elapsed_time;
1550 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1551 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1552 "%d recovered and %d %s evicted.\n", obd->obd_name,
1553 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1554 obd->obd_max_recoverable_clients,
1555 atomic_read(&obd->obd_connected_clients),
1556 obd->obd_stale_clients,
1557 obd->obd_stale_clients == 1 ? "was" : "were");
1560 ldlm_reprocess_recovery_done(obd->obd_namespace);
1561 spin_lock(&obd->obd_recovery_task_lock);
1562 if (!list_empty(&obd->obd_req_replay_queue) ||
1563 !list_empty(&obd->obd_lock_replay_queue) ||
1564 !list_empty(&obd->obd_final_req_queue)) {
1565 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1567 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1568 list_empty(&obd->obd_lock_replay_queue) ? \
1570 list_empty(&obd->obd_final_req_queue) ? \
1572 spin_unlock(&obd->obd_recovery_task_lock);
1575 spin_unlock(&obd->obd_recovery_task_lock);
1577 obd->obd_recovery_end = ktime_get_real_seconds();
1579 /* When recovery finished, cleanup orphans on MDS and OST. */
1580 if (obd->obd_type && OBP(obd, postrecov)) {
1581 int rc = OBP(obd, postrecov)(obd);
1584 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1590 static void abort_req_replay_queue(struct obd_device *obd)
1592 struct ptlrpc_request *req, *n;
1593 struct list_head abort_list;
1595 INIT_LIST_HEAD(&abort_list);
1596 spin_lock(&obd->obd_recovery_task_lock);
1597 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1598 spin_unlock(&obd->obd_recovery_task_lock);
1599 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1600 DEBUG_REQ(D_WARNING, req, "aborted:");
1601 req->rq_status = -ENOTCONN;
1602 if (ptlrpc_error(req)) {
1603 DEBUG_REQ(D_ERROR, req,
1604 "failed abort_req_reply; skipping");
1606 target_exp_dequeue_req_replay(req);
1607 target_request_copy_put(req);
1611 static void abort_lock_replay_queue(struct obd_device *obd)
1613 struct ptlrpc_request *req, *n;
1614 struct list_head abort_list;
1616 INIT_LIST_HEAD(&abort_list);
1617 spin_lock(&obd->obd_recovery_task_lock);
1618 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1619 spin_unlock(&obd->obd_recovery_task_lock);
1620 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1621 DEBUG_REQ(D_ERROR, req, "aborted:");
1622 req->rq_status = -ENOTCONN;
1623 if (ptlrpc_error(req)) {
1624 DEBUG_REQ(D_ERROR, req,
1625 "failed abort_lock_reply; skipping");
1627 target_request_copy_put(req);
1631 /* Called from a cleanup function if the device is being cleaned up
1632 forcefully. The exports should all have been disconnected already,
1633 the only thing left to do is
1634 - clear the recovery flags
1636 - free queued requests and replies, but don't send replies
1637 Because the obd_stopping flag is set, no new requests should be received.
1640 void target_cleanup_recovery(struct obd_device *obd)
1642 struct ptlrpc_request *req, *n;
1643 struct list_head clean_list;
1645 INIT_LIST_HEAD(&clean_list);
1646 spin_lock(&obd->obd_dev_lock);
1647 if (!obd->obd_recovering) {
1648 spin_unlock(&obd->obd_dev_lock);
1652 obd->obd_recovering = obd->obd_abort_recovery = 0;
1653 spin_unlock(&obd->obd_dev_lock);
1655 spin_lock(&obd->obd_recovery_task_lock);
1656 target_cancel_recovery_timer(obd);
1657 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1658 spin_unlock(&obd->obd_recovery_task_lock);
1660 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1661 LASSERT(req->rq_reply_state == NULL);
1662 target_exp_dequeue_req_replay(req);
1663 target_request_copy_put(req);
1666 spin_lock(&obd->obd_recovery_task_lock);
1667 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1668 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1669 spin_unlock(&obd->obd_recovery_task_lock);
1671 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1672 LASSERT(req->rq_reply_state == NULL);
1673 target_request_copy_put(req);
1678 EXPORT_SYMBOL(target_cleanup_recovery);
1680 /* obd_recovery_task_lock should be held */
1681 void target_cancel_recovery_timer(struct obd_device *obd)
1683 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1684 del_timer(&obd->obd_recovery_timer);
1687 static void target_start_recovery_timer(struct obd_device *obd)
1689 if (obd->obd_recovery_start != 0)
1692 spin_lock(&obd->obd_dev_lock);
1693 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1694 spin_unlock(&obd->obd_dev_lock);
1698 LASSERT(obd->obd_recovery_timeout != 0);
1700 if (obd->obd_recovery_start != 0) {
1701 spin_unlock(&obd->obd_dev_lock);
1705 mod_timer(&obd->obd_recovery_timer,
1706 jiffies + cfs_time_seconds(obd->obd_recovery_timeout));
1707 obd->obd_recovery_start = ktime_get_real_seconds();
1708 spin_unlock(&obd->obd_dev_lock);
1710 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1712 obd->obd_recovery_timeout / 60,
1713 obd->obd_recovery_timeout % 60,
1714 obd->obd_max_recoverable_clients,
1715 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1716 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1720 * extend recovery window.
1722 * if @extend is true, extend recovery window to have @drt remaining at least;
1723 * otherwise, make sure the recovery timeout value is not less than @drt.
1725 static void extend_recovery_timer(struct obd_device *obd, time64_t drt,
1733 spin_lock(&obd->obd_dev_lock);
1734 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1735 spin_unlock(&obd->obd_dev_lock);
1738 LASSERT(obd->obd_recovery_start != 0);
1740 now = ktime_get_real_seconds();
1741 to = obd->obd_recovery_timeout;
1742 end = obd->obd_recovery_start + to;
1745 if (extend && (drt > left)) {
1747 } else if (!extend && (drt > to)) {
1751 if (to > obd->obd_recovery_time_hard) {
1752 to = obd->obd_recovery_time_hard;
1753 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1754 obd->obd_name, to, extend);
1757 if (obd->obd_recovery_timeout < to) {
1758 obd->obd_recovery_timeout = to;
1759 end = obd->obd_recovery_start + to;
1760 mod_timer(&obd->obd_recovery_timer,
1761 jiffies + cfs_time_seconds(end - now));
1763 spin_unlock(&obd->obd_dev_lock);
1765 CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1766 obd->obd_name, (s64)(end - now));
1769 /* Reset the timer with each new client connection */
1771 * This timer is actually reconnect_timer, which is for making sure
1772 * the total recovery window is at least as big as my reconnect
1773 * attempt timing. So the initial recovery time_out will be set to
1774 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1775 * from client is bigger than this, then the recovery time_out will
1776 * be extended to make sure the client could be reconnected, in the
1777 * process, the timeout from the new client should be ignored.
1781 check_and_start_recovery_timer(struct obd_device *obd,
1782 struct ptlrpc_request *req,
1785 time64_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1786 struct obd_device_target *obt = &obd->u.obt;
1788 if (!new_client && service_time)
1789 /* Teach server about old server's estimates, as first guess
1790 * at how long new requests will take. */
1791 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1794 target_start_recovery_timer(obd);
1796 /* Convert the service time to RPC timeout,
1797 * and reuse service_time to limit stack usage.
1799 service_time = at_est2timeout(service_time);
1801 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1802 service_time < at_extra)
1803 service_time = at_extra;
1805 /* We expect other clients to timeout within service_time, then try
1806 * to reconnect, then try the failover server. The max delay between
1807 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1808 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1810 LASSERT(obt->obt_magic == OBT_MAGIC);
1811 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1812 if (service_time > obd->obd_recovery_timeout && !new_client)
1813 extend_recovery_timer(obd, service_time, false);
1816 /** Health checking routines */
1817 static inline int exp_connect_healthy(struct obd_export *exp)
1819 return (exp->exp_in_recovery);
1822 /** if export done req_replay or has replay in queue */
1823 static inline int exp_req_replay_healthy(struct obd_export *exp)
1825 return (!exp->exp_req_replay_needed ||
1826 atomic_read(&exp->exp_replay_count) > 0);
1830 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1832 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1833 exp_req_replay_healthy(exp);
1836 /** if export done lock_replay or has replay in queue */
1837 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1839 return (!exp->exp_lock_replay_needed ||
1840 atomic_read(&exp->exp_replay_count) > 0);
1843 static inline int exp_vbr_healthy(struct obd_export *exp)
1845 return (!exp->exp_vbr_failed);
1848 static inline int exp_finished(struct obd_export *exp)
1850 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1853 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1855 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1859 static int check_for_next_transno(struct lu_target *lut)
1861 struct ptlrpc_request *req = NULL;
1862 struct obd_device *obd = lut->lut_obd;
1863 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1864 int wake_up = 0, connected, completed, queue_len;
1865 __u64 req_transno = 0;
1866 __u64 update_transno = 0;
1867 __u64 next_transno = 0;
1870 spin_lock(&obd->obd_recovery_task_lock);
1871 if (!list_empty(&obd->obd_req_replay_queue)) {
1872 req = list_entry(obd->obd_req_replay_queue.next,
1873 struct ptlrpc_request, rq_list);
1874 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1878 update_transno = distribute_txn_get_next_transno(tdtd);
1880 connected = atomic_read(&obd->obd_connected_clients);
1881 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1882 queue_len = obd->obd_requests_queued_for_recovery;
1883 next_transno = obd->obd_next_recovery_transno;
1885 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1886 "req_transno: %llu, next_transno: %llu\n",
1887 obd->obd_max_recoverable_clients, connected, completed,
1888 queue_len, req_transno, next_transno);
1890 if (obd->obd_abort_recovery) {
1891 CDEBUG(D_HA, "waking for aborted recovery\n");
1893 } else if (obd->obd_recovery_expired) {
1894 CDEBUG(D_HA, "waking for expired recovery\n");
1896 } else if (tdtd != NULL && req != NULL &&
1897 is_req_replayed_by_update(req)) {
1898 LASSERTF(req_transno < next_transno, "req_transno %llu"
1899 "next_transno%llu\n", req_transno, next_transno);
1900 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1903 } else if (req_transno == next_transno ||
1904 (update_transno != 0 && update_transno <= next_transno)) {
1905 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1907 } else if (queue_len > 0 &&
1908 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1909 /** handle gaps occured due to lost reply or VBR */
1910 LASSERTF(req_transno >= next_transno,
1911 "req_transno: %llu, next_transno: %llu\n",
1912 req_transno, next_transno);
1914 "%s: waking for gap in transno, VBR is %s (skip: "
1915 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1916 ", next_update %lld last_committed: %lld)\n",
1917 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1918 next_transno, queue_len, completed, connected,
1919 req_transno, update_transno, obd->obd_last_committed);
1920 obd->obd_next_recovery_transno = req_transno;
1922 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1923 CDEBUG(D_HA, "waking for completed recovery\n");
1925 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1926 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1927 " by fail_lock, waking up (%lld)\n", next_transno);
1928 obd->obd_next_recovery_transno = req_transno;
1931 spin_unlock(&obd->obd_recovery_task_lock);
1935 static int check_for_next_lock(struct lu_target *lut)
1937 struct obd_device *obd = lut->lut_obd;
1940 spin_lock(&obd->obd_recovery_task_lock);
1941 if (!list_empty(&obd->obd_lock_replay_queue)) {
1942 CDEBUG(D_HA, "waking for next lock\n");
1944 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1945 CDEBUG(D_HA, "waking for completed lock replay\n");
1947 } else if (obd->obd_abort_recovery) {
1948 CDEBUG(D_HA, "waking for aborted recovery\n");
1950 } else if (obd->obd_recovery_expired) {
1951 CDEBUG(D_HA, "waking for expired recovery\n");
1954 spin_unlock(&obd->obd_recovery_task_lock);
1960 * wait for recovery events,
1961 * check its status with help of check_routine
1962 * evict dead clients via health_check
1964 static int target_recovery_overseer(struct lu_target *lut,
1965 int (*check_routine)(struct lu_target *),
1966 int (*health_check)(struct obd_export *))
1968 struct obd_device *obd = lut->lut_obd;
1969 struct target_distribute_txn_data *tdtd;
1973 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
1974 now = ktime_get_seconds();
1975 if (now - last > 600) {
1976 LCONSOLE_INFO("%s: in recovery but waiting for "
1977 "the first client to connect\n",
1982 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
1983 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
1984 __u64 next_update_transno = 0;
1986 /* Only abort the recovery if there are no update recovery
1987 * left in the queue */
1988 spin_lock(&obd->obd_recovery_task_lock);
1989 if (lut->lut_tdtd != NULL) {
1990 next_update_transno =
1991 distribute_txn_get_next_transno(lut->lut_tdtd);
1993 tdtd = lut->lut_tdtd;
1994 /* If next_update_transno == 0, it probably because
1995 * updatelog retrieve threads did not get any records
1996 * yet, let's wait those threads stopped */
1997 if (next_update_transno == 0) {
1998 struct l_wait_info lwi = { 0 };
2000 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2002 &tdtd->tdtd_recovery_threads_count) == 0,
2005 next_update_transno =
2006 distribute_txn_get_next_transno(
2011 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2012 obd->obd_next_recovery_transno = next_update_transno;
2013 spin_unlock(&obd->obd_recovery_task_lock);
2014 /* Disconnect unfinished exports from clients, and
2015 * keep connection from MDT to make sure the update
2016 * recovery will still keep trying until some one
2017 * manually abort the recovery */
2018 class_disconnect_stale_exports(obd,
2019 exp_finished_or_from_mdt);
2020 /* Abort all of replay and replay lock req from
2022 abort_req_replay_queue(obd);
2023 abort_lock_replay_queue(obd);
2024 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2025 ")in the queue.\n", obd->obd_name,
2026 next_update_transno);
2028 obd->obd_abort_recovery = 1;
2029 spin_unlock(&obd->obd_recovery_task_lock);
2030 CWARN("%s recovery is aborted by hard timeout\n",
2035 while (wait_event_timeout(obd->obd_next_transno_waitq,
2037 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2038 /* wait indefinitely for event, but don't trigger watchdog */;
2040 if (obd->obd_abort_recovery) {
2041 CWARN("recovery is aborted, evict exports in recovery\n");
2042 if (lut->lut_tdtd != NULL) {
2043 struct l_wait_info lwi = { 0 };
2045 tdtd = lut->lut_tdtd;
2046 /* Let's wait all of the update log recovery thread
2048 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2049 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2051 /* Then abort the update recovery list */
2052 dtrq_list_destroy(lut->lut_tdtd);
2055 /** evict exports which didn't finish recovery yet */
2056 class_disconnect_stale_exports(obd, exp_finished);
2058 } else if (obd->obd_recovery_expired) {
2059 obd->obd_recovery_expired = 0;
2060 /** If some clients died being recovered, evict them */
2061 LCONSOLE_WARN("%s: recovery is timed out, "
2062 "evict stale exports\n", obd->obd_name);
2063 /** evict cexports with no replay in queue, they are stalled */
2064 class_disconnect_stale_exports(obd, health_check);
2066 /** continue with VBR */
2067 spin_lock(&obd->obd_dev_lock);
2068 obd->obd_version_recov = 1;
2069 spin_unlock(&obd->obd_dev_lock);
2071 * reset timer, recovery will proceed with versions now,
2072 * timeout is set just to handle reconnection delays
2074 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2075 /** Wait for recovery events again, after evicting bad clients */
2081 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2083 struct obd_device *obd = lut->lut_obd;
2084 struct ptlrpc_request *req = NULL;
2086 CDEBUG(D_HA, "Waiting for lock\n");
2087 if (target_recovery_overseer(lut, check_for_next_lock,
2088 exp_lock_replay_healthy))
2089 abort_lock_replay_queue(obd);
2091 spin_lock(&obd->obd_recovery_task_lock);
2092 if (!list_empty(&obd->obd_lock_replay_queue)) {
2093 req = list_entry(obd->obd_lock_replay_queue.next,
2094 struct ptlrpc_request, rq_list);
2095 list_del_init(&req->rq_list);
2096 spin_unlock(&obd->obd_recovery_task_lock);
2098 spin_unlock(&obd->obd_recovery_task_lock);
2099 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2100 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2101 /** evict exports failed VBR */
2102 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2107 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2109 struct ptlrpc_request *req = NULL;
2111 spin_lock(&obd->obd_recovery_task_lock);
2112 if (!list_empty(&obd->obd_final_req_queue)) {
2113 req = list_entry(obd->obd_final_req_queue.next,
2114 struct ptlrpc_request, rq_list);
2115 list_del_init(&req->rq_list);
2116 spin_unlock(&obd->obd_recovery_task_lock);
2117 if (req->rq_export->exp_in_recovery) {
2118 spin_lock(&req->rq_export->exp_lock);
2119 req->rq_export->exp_in_recovery = 0;
2120 spin_unlock(&req->rq_export->exp_lock);
2123 spin_unlock(&obd->obd_recovery_task_lock);
2128 static void handle_recovery_req(struct ptlrpc_thread *thread,
2129 struct ptlrpc_request *req,
2130 svc_handler_t handler)
2135 * export can be evicted during recovery, no need to handle replays for
2136 * it after that, discard such request silently
2138 if (req->rq_export->exp_disconnected)
2141 req->rq_session.lc_thread = thread;
2142 req->rq_svc_thread = thread;
2143 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2145 /* thread context */
2146 lu_context_enter(&thread->t_env->le_ctx);
2148 lu_context_exit(&thread->t_env->le_ctx);
2150 /* don't reset timer for final stage */
2151 if (!exp_finished(req->rq_export)) {
2152 time64_t to = obd_timeout;
2155 * Add request timeout to the recovery time so next request from
2156 * this client may come in recovery time
2159 struct ptlrpc_service_part *svcpt;
2161 svcpt = req->rq_rqbd->rqbd_svcpt;
2162 /* If the server sent early reply for this request,
2163 * the client will recalculate the timeout according to
2164 * current server estimate service time, so we will
2165 * use the maxium timeout here for waiting the client
2166 * sending the next req */
2167 to = max((int)at_est2timeout(
2168 at_get(&svcpt->scp_at_estimate)),
2169 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2170 /* Add 2 net_latency, one for balance rq_deadline
2171 * (see ptl_send_rpc), one for resend the req to server,
2172 * Note: client will pack net_latency in replay req
2173 * (see ptlrpc_replay_req) */
2174 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2176 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2181 /** Checking routines for recovery */
2182 static int check_for_recovery_ready(struct lu_target *lut)
2184 struct obd_device *obd = lut->lut_obd;
2185 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2187 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2188 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2189 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2190 obd->obd_recovery_expired);
2192 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2193 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2194 if (clnts + obd->obd_stale_clients <
2195 obd->obd_max_recoverable_clients)
2199 if (lut->lut_tdtd != NULL) {
2200 if (!lut->lut_tdtd->tdtd_replay_ready &&
2201 !obd->obd_abort_recovery) {
2202 /* Let's extend recovery timer, in case the recovery
2203 * timer expired, and some clients got evicted */
2204 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2206 CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2207 obd->obd_name, obd->obd_recovery_timeout);
2216 REQUEST_RECOVERY = 1,
2217 UPDATE_RECOVERY = 2,
2220 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2224 if (!list_empty(&obd->obd_req_replay_queue)) {
2225 struct ptlrpc_request *req;
2227 req = list_entry(obd->obd_req_replay_queue.next,
2228 struct ptlrpc_request, rq_list);
2229 transno = lustre_msg_get_transno(req->rq_reqmsg);
2235 static __u64 get_next_transno(struct lu_target *lut, int *type)
2237 struct obd_device *obd = lut->lut_obd;
2238 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2240 __u64 update_transno;
2243 transno = get_next_replay_req_transno(obd);
2245 *type = REQUEST_RECOVERY;
2250 update_transno = distribute_txn_get_next_transno(tdtd);
2251 if (transno == 0 || (transno >= update_transno &&
2252 update_transno != 0)) {
2253 transno = update_transno;
2255 *type = UPDATE_RECOVERY;
2262 * drop duplicate replay request
2264 * Because the operation has been replayed by update recovery, the request
2265 * with the same transno will be dropped and also notify the client to send
2266 * next replay request.
2268 * \param[in] env execution environment
2269 * \param[in] obd failover obd device
2270 * \param[in] req request to be dropped
2272 static void drop_duplicate_replay_req(struct lu_env *env,
2273 struct obd_device *obd,
2274 struct ptlrpc_request *req)
2276 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2277 " update records are found.\n",
2278 lustre_msg_get_transno(req->rq_reqmsg),
2279 libcfs_nid2str(req->rq_peer.nid));
2281 /* Right now, only for MDS reint operation update replay and
2282 * normal request replay can have the same transno */
2283 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2284 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2285 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2286 if (likely(req->rq_export))
2287 target_committed_to_req(req);
2288 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2289 target_send_reply(req, req->rq_status, 0);
2291 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2292 libcfs_nid2str(req->rq_peer.nid));
2294 target_exp_dequeue_req_replay(req);
2295 target_request_copy_put(req);
2296 obd->obd_replayed_requests++;
2299 static void replay_request_or_update(struct lu_env *env,
2300 struct lu_target *lut,
2301 struct target_recovery_data *trd,
2302 struct ptlrpc_thread *thread)
2304 struct obd_device *obd = lut->lut_obd;
2305 struct ptlrpc_request *req = NULL;
2310 CDEBUG(D_HA, "Waiting for transno %lld\n",
2311 obd->obd_next_recovery_transno);
2313 /* Replay all of request and update by transno */
2315 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2317 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2319 /** It is needed to extend recovery window above
2320 * recovery_time_soft. Extending is possible only in the
2321 * end of recovery window (see more details in
2322 * handle_recovery_req()).
2324 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2326 if (target_recovery_overseer(lut, check_for_next_transno,
2327 exp_req_replay_healthy_or_from_mdt)) {
2328 abort_req_replay_queue(obd);
2329 abort_lock_replay_queue(obd);
2333 spin_lock(&obd->obd_recovery_task_lock);
2334 transno = get_next_transno(lut, &type);
2335 if (type == REQUEST_RECOVERY && transno != 0) {
2336 /* Drop replay request from client side, if the
2337 * replay has been executed by update with the
2339 req = list_entry(obd->obd_req_replay_queue.next,
2340 struct ptlrpc_request, rq_list);
2342 list_del_init(&req->rq_list);
2343 obd->obd_requests_queued_for_recovery--;
2344 spin_unlock(&obd->obd_recovery_task_lock);
2346 /* Let's check if the request has been redone by
2348 if (is_req_replayed_by_update(req)) {
2349 struct distribute_txn_replay_req *dtrq;
2351 dtrq = distribute_txn_lookup_finish_list(tdtd,
2353 LASSERT(dtrq != NULL);
2354 spin_lock(&tdtd->tdtd_replay_list_lock);
2355 list_del_init(&dtrq->dtrq_list);
2356 spin_unlock(&tdtd->tdtd_replay_list_lock);
2359 drop_duplicate_replay_req(env, obd, req);
2364 LASSERT(trd->trd_processing_task == current_pid());
2365 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2366 lustre_msg_get_transno(req->rq_reqmsg),
2367 libcfs_nid2str(req->rq_peer.nid));
2369 handle_recovery_req(thread, req,
2370 trd->trd_recovery_handler);
2372 * bz18031: increase next_recovery_transno before
2373 * target_request_copy_put() will drop exp_rpc reference
2375 spin_lock(&obd->obd_recovery_task_lock);
2376 obd->obd_next_recovery_transno++;
2377 spin_unlock(&obd->obd_recovery_task_lock);
2378 target_exp_dequeue_req_replay(req);
2379 target_request_copy_put(req);
2380 obd->obd_replayed_requests++;
2381 } else if (type == UPDATE_RECOVERY && transno != 0) {
2382 struct distribute_txn_replay_req *dtrq;
2385 spin_unlock(&obd->obd_recovery_task_lock);
2387 LASSERT(tdtd != NULL);
2388 dtrq = distribute_txn_get_next_req(tdtd);
2389 lu_context_enter(&thread->t_env->le_ctx);
2390 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2391 lu_context_exit(&thread->t_env->le_ctx);
2392 extend_recovery_timer(obd, obd_timeout, true);
2394 if (rc == 0 && dtrq->dtrq_xid != 0) {
2395 CDEBUG(D_HA, "Move x%llu t%llu"
2396 " to finish list\n", dtrq->dtrq_xid,
2397 dtrq->dtrq_master_transno);
2399 /* Add it to the replay finish list */
2400 spin_lock(&tdtd->tdtd_replay_list_lock);
2401 list_add(&dtrq->dtrq_list,
2402 &tdtd->tdtd_replay_finish_list);
2403 spin_unlock(&tdtd->tdtd_replay_list_lock);
2405 spin_lock(&obd->obd_recovery_task_lock);
2406 if (transno == obd->obd_next_recovery_transno)
2407 obd->obd_next_recovery_transno++;
2409 obd->obd_next_recovery_transno)
2410 obd->obd_next_recovery_transno =
2412 spin_unlock(&obd->obd_recovery_task_lock);
2417 spin_unlock(&obd->obd_recovery_task_lock);
2419 LASSERT(list_empty(&obd->obd_req_replay_queue));
2420 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2421 /** evict exports failed VBR */
2422 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2428 static int target_recovery_thread(void *arg)
2430 struct lu_target *lut = arg;
2431 struct obd_device *obd = lut->lut_obd;
2432 struct ptlrpc_request *req;
2433 struct target_recovery_data *trd = &obd->obd_recovery_data;
2434 unsigned long delta;
2436 struct ptlrpc_thread *thread = NULL;
2440 unshare_fs_struct();
2441 OBD_ALLOC_PTR(thread);
2447 OBD_FREE_PTR(thread);
2451 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2453 OBD_FREE_PTR(thread);
2458 thread->t_env = env;
2459 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2460 env->le_ctx.lc_thread = thread;
2461 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2462 thread->t_watchdog = NULL;
2464 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2466 trd->trd_processing_task = current_pid();
2468 spin_lock(&obd->obd_dev_lock);
2469 obd->obd_recovering = 1;
2470 spin_unlock(&obd->obd_dev_lock);
2471 complete(&trd->trd_starting);
2473 /* first of all, we have to know the first transno to replay */
2474 if (target_recovery_overseer(lut, check_for_recovery_ready,
2475 exp_connect_healthy)) {
2476 abort_req_replay_queue(obd);
2477 abort_lock_replay_queue(obd);
2478 if (lut->lut_tdtd != NULL)
2479 dtrq_list_destroy(lut->lut_tdtd);
2482 /* next stage: replay requests or update */
2484 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2485 atomic_read(&obd->obd_req_replay_clients),
2486 obd->obd_next_recovery_transno);
2487 replay_request_or_update(env, lut, trd, thread);
2490 * The second stage: replay locks
2492 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2493 atomic_read(&obd->obd_lock_replay_clients));
2494 while ((req = target_next_replay_lock(lut))) {
2495 LASSERT(trd->trd_processing_task == current_pid());
2496 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2497 libcfs_nid2str(req->rq_peer.nid));
2498 handle_recovery_req(thread, req,
2499 trd->trd_recovery_handler);
2500 target_request_copy_put(req);
2501 obd->obd_replayed_locks++;
2505 * The third stage: reply on final pings, at this moment all clients
2506 * must have request in final queue
2508 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2509 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2510 /** Update server last boot epoch */
2511 tgt_boot_epoch_update(lut);
2512 /* We drop recoverying flag to forward all new requests
2513 * to regular mds_handle() since now */
2514 spin_lock(&obd->obd_dev_lock);
2515 obd->obd_recovering = obd->obd_abort_recovery = 0;
2516 spin_unlock(&obd->obd_dev_lock);
2517 spin_lock(&obd->obd_recovery_task_lock);
2518 target_cancel_recovery_timer(obd);
2519 spin_unlock(&obd->obd_recovery_task_lock);
2520 while ((req = target_next_final_ping(obd))) {
2521 LASSERT(trd->trd_processing_task == current_pid());
2522 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2523 libcfs_nid2str(req->rq_peer.nid));
2524 handle_recovery_req(thread, req,
2525 trd->trd_recovery_handler);
2526 /* Because the waiting client can not send ping to server,
2527 * so we need refresh the last_request_time, to avoid the
2528 * export is being evicted */
2529 ptlrpc_update_export_timer(req->rq_export, 0);
2530 target_request_copy_put(req);
2533 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2534 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2535 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2536 if (delta > OBD_RECOVERY_TIME_SOFT) {
2537 CWARN("too long recovery - read logs\n");
2538 libcfs_debug_dumplog();
2541 target_finish_recovery(lut);
2543 lu_context_fini(&env->le_ctx);
2544 trd->trd_processing_task = 0;
2545 complete(&trd->trd_finishing);
2547 tgt_io_thread_done(thread);
2548 OBD_FREE_PTR(thread);
2553 static int target_start_recovery_thread(struct lu_target *lut,
2554 svc_handler_t handler)
2556 struct obd_device *obd = lut->lut_obd;
2558 struct target_recovery_data *trd = &obd->obd_recovery_data;
2561 memset(trd, 0, sizeof(*trd));
2562 init_completion(&trd->trd_starting);
2563 init_completion(&trd->trd_finishing);
2564 trd->trd_recovery_handler = handler;
2566 rc = server_name2index(obd->obd_name, &index, NULL);
2570 if (!IS_ERR(kthread_run(target_recovery_thread,
2571 lut, "tgt_recover_%d", index))) {
2572 wait_for_completion(&trd->trd_starting);
2573 LASSERT(obd->obd_recovering != 0);
2581 void target_stop_recovery_thread(struct obd_device *obd)
2583 if (obd->obd_recovery_data.trd_processing_task > 0) {
2584 struct target_recovery_data *trd = &obd->obd_recovery_data;
2585 /** recovery can be done but postrecovery is not yet */
2586 spin_lock(&obd->obd_dev_lock);
2587 if (obd->obd_recovering) {
2588 CERROR("%s: Aborting recovery\n", obd->obd_name);
2589 obd->obd_abort_recovery = 1;
2590 wake_up(&obd->obd_next_transno_waitq);
2592 spin_unlock(&obd->obd_dev_lock);
2593 wait_for_completion(&trd->trd_finishing);
2596 EXPORT_SYMBOL(target_stop_recovery_thread);
2598 void target_recovery_fini(struct obd_device *obd)
2600 class_disconnect_exports(obd);
2601 target_stop_recovery_thread(obd);
2602 target_cleanup_recovery(obd);
2604 EXPORT_SYMBOL(target_recovery_fini);
2606 static void target_recovery_expired(unsigned long castmeharder)
2608 struct obd_device *obd = (struct obd_device *)castmeharder;
2609 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2610 " after %llus (%d clients connected)\n",
2611 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2612 (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2613 atomic_read(&obd->obd_connected_clients));
2615 obd->obd_recovery_expired = 1;
2616 wake_up(&obd->obd_next_transno_waitq);
2619 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2621 struct obd_device *obd = lut->lut_obd;
2623 if (lut->lut_bottom->dd_rdonly)
2626 if (obd->obd_max_recoverable_clients == 0) {
2627 /** Update server last boot epoch */
2628 tgt_boot_epoch_update(lut);
2632 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2633 "last_transno %llu\n", obd->obd_name,
2634 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2635 LASSERT(obd->obd_stopping == 0);
2636 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2637 obd->obd_recovery_start = 0;
2638 obd->obd_recovery_end = 0;
2640 setup_timer(&obd->obd_recovery_timer, target_recovery_expired,
2641 (unsigned long)obd);
2642 target_start_recovery_thread(lut, handler);
2644 EXPORT_SYMBOL(target_recovery_init);
2646 static int target_process_req_flags(struct obd_device *obd,
2647 struct ptlrpc_request *req)
2649 struct obd_export *exp = req->rq_export;
2650 LASSERT(exp != NULL);
2651 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2652 /* client declares he's ready to replay locks */
2653 spin_lock(&exp->exp_lock);
2654 if (exp->exp_req_replay_needed) {
2655 exp->exp_req_replay_needed = 0;
2656 spin_unlock(&exp->exp_lock);
2658 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2659 atomic_dec(&obd->obd_req_replay_clients);
2661 spin_unlock(&exp->exp_lock);
2664 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2665 /* client declares he's ready to complete recovery
2666 * so, we put the request on th final queue */
2667 spin_lock(&exp->exp_lock);
2668 if (exp->exp_lock_replay_needed) {
2669 exp->exp_lock_replay_needed = 0;
2670 spin_unlock(&exp->exp_lock);
2672 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2673 atomic_dec(&obd->obd_lock_replay_clients);
2675 spin_unlock(&exp->exp_lock);
2681 int target_queue_recovery_request(struct ptlrpc_request *req,
2682 struct obd_device *obd)
2684 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2685 struct ptlrpc_request *reqiter;
2689 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2690 /* Processing the queue right now, don't re-add. */
2694 target_process_req_flags(obd, req);
2696 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2697 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2698 if (cfs_fail_val == 1) {
2701 wake_up(&cfs_race_waitq);
2703 set_current_state(TASK_INTERRUPTIBLE);
2704 schedule_timeout(cfs_time_seconds(1));
2708 /* client declares he's ready to complete recovery
2709 * so, we put the request on th final queue */
2710 target_request_copy_get(req);
2711 DEBUG_REQ(D_HA, req, "queue final req");
2712 wake_up(&obd->obd_next_transno_waitq);
2713 spin_lock(&obd->obd_recovery_task_lock);
2714 if (obd->obd_recovering) {
2715 struct ptlrpc_request *tmp;
2716 struct ptlrpc_request *duplicate = NULL;
2718 if (likely(!req->rq_export->exp_replay_done)) {
2719 req->rq_export->exp_replay_done = 1;
2720 list_add_tail(&req->rq_list,
2721 &obd->obd_final_req_queue);
2722 spin_unlock(&obd->obd_recovery_task_lock);
2726 /* XXX O(n), but only happens if final ping is
2727 * timed out, probably reorganize the list as
2728 * a hash list later */
2729 list_for_each_entry_safe(reqiter, tmp,
2730 &obd->obd_final_req_queue,
2732 if (reqiter->rq_export == req->rq_export) {
2733 list_del_init(&reqiter->rq_list);
2734 duplicate = reqiter;
2739 list_add_tail(&req->rq_list,
2740 &obd->obd_final_req_queue);
2741 req->rq_export->exp_replay_done = 1;
2742 spin_unlock(&obd->obd_recovery_task_lock);
2744 if (duplicate != NULL) {
2745 DEBUG_REQ(D_HA, duplicate,
2746 "put prev final req\n");
2747 target_request_copy_put(duplicate);
2751 spin_unlock(&obd->obd_recovery_task_lock);
2752 target_request_copy_put(req);
2753 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2756 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2757 /* client declares he's ready to replay locks */
2758 target_request_copy_get(req);
2759 DEBUG_REQ(D_HA, req, "queue lock replay req");
2760 wake_up(&obd->obd_next_transno_waitq);
2761 spin_lock(&obd->obd_recovery_task_lock);
2762 LASSERT(obd->obd_recovering);
2763 /* usually due to recovery abort */
2764 if (!req->rq_export->exp_in_recovery) {
2765 spin_unlock(&obd->obd_recovery_task_lock);
2766 target_request_copy_put(req);
2769 LASSERT(req->rq_export->exp_lock_replay_needed);
2770 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2771 spin_unlock(&obd->obd_recovery_task_lock);
2775 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2776 * (i.e. buflens etc are in my own byte order), but type-dependent
2777 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2780 INIT_LIST_HEAD(&req->rq_list);
2781 DEBUG_REQ(D_HA, req, "not queueing");
2785 /* If we're processing the queue, we want don't want to queue this
2788 * Also, if this request has a transno less than the one we're waiting
2789 * for, we should process it now. It could (and currently always will)
2790 * be an open request for a descriptor that was opened some time ago.
2792 * Also, a resent, replayed request that has already been
2793 * handled will pass through here and be processed immediately.
2795 CDEBUG(D_HA, "Next recovery transno: %llu"
2796 ", current: %llu, replaying\n",
2797 obd->obd_next_recovery_transno, transno);
2799 /* If the request has been replayed by update replay, then sends this
2800 * request to the recovery thread (replay_request_or_update()), where
2801 * it will be handled */
2802 spin_lock(&obd->obd_recovery_task_lock);
2803 if (transno < obd->obd_next_recovery_transno &&
2804 !is_req_replayed_by_update(req)) {
2805 /* Processing the queue right now, don't re-add. */
2806 LASSERT(list_empty(&req->rq_list));
2807 spin_unlock(&obd->obd_recovery_task_lock);
2810 spin_unlock(&obd->obd_recovery_task_lock);
2812 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2815 target_request_copy_get(req);
2816 if (!req->rq_export->exp_in_recovery) {
2817 target_request_copy_put(req);
2820 LASSERT(req->rq_export->exp_req_replay_needed);
2822 if (target_exp_enqueue_req_replay(req)) {
2823 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2824 target_request_copy_put(req);
2829 spin_lock(&obd->obd_recovery_task_lock);
2830 LASSERT(obd->obd_recovering);
2831 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2832 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2833 list_add_tail(&req->rq_list, &reqiter->rq_list);
2838 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2840 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2841 "has been claimed by another client");
2842 spin_unlock(&obd->obd_recovery_task_lock);
2843 target_exp_dequeue_req_replay(req);
2844 target_request_copy_put(req);
2850 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2852 obd->obd_requests_queued_for_recovery++;
2853 spin_unlock(&obd->obd_recovery_task_lock);
2854 wake_up(&obd->obd_next_transno_waitq);
2858 int target_handle_ping(struct ptlrpc_request *req)
2860 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2861 return req_capsule_server_pack(&req->rq_pill);
2864 void target_committed_to_req(struct ptlrpc_request *req)
2866 struct obd_export *exp = req->rq_export;
2868 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2869 lustre_msg_set_last_committed(req->rq_repmsg,
2870 exp->exp_last_committed);
2872 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2873 "%d)", exp->exp_obd->obd_no_transno,
2874 req->rq_repmsg == NULL);
2876 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2877 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2880 #endif /* HAVE_SERVER_SUPPORT */
2883 * Packs current SLV and Limit into \a req.
2885 int target_pack_pool_reply(struct ptlrpc_request *req)
2887 struct obd_device *obd;
2890 /* Check that we still have all structures alive as this may
2891 * be some late RPC at shutdown time. */
2892 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2893 !exp_connect_lru_resize(req->rq_export))) {
2894 lustre_msg_set_slv(req->rq_repmsg, 0);
2895 lustre_msg_set_limit(req->rq_repmsg, 0);
2899 /* OBD is alive here as export is alive, which we checked above. */
2900 obd = req->rq_export->exp_obd;
2902 read_lock(&obd->obd_pool_lock);
2903 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2904 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2905 read_unlock(&obd->obd_pool_lock);
2910 static int target_send_reply_msg(struct ptlrpc_request *req,
2911 int rc, int fail_id)
2913 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2914 DEBUG_REQ(D_ERROR, req, "dropping reply");
2917 /* We can have a null rq_reqmsg in the event of bad signature or
2918 * no context when unwrapping */
2919 if (req->rq_reqmsg &&
2920 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2921 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2925 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2926 req->rq_status = rc;
2927 return ptlrpc_send_error(req, 1);
2929 DEBUG_REQ(D_NET, req, "sending reply");
2932 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2935 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2937 struct ptlrpc_service_part *svcpt;
2939 struct ptlrpc_reply_state *rs;
2940 struct obd_export *exp;
2943 if (req->rq_no_reply) {
2948 svcpt = req->rq_rqbd->rqbd_svcpt;
2949 rs = req->rq_reply_state;
2950 if (rs == NULL || !rs->rs_difficult) {
2952 target_send_reply_msg (req, rc, fail_id);
2957 /* must be an export if locks saved */
2958 LASSERT(req->rq_export != NULL);
2959 /* req/reply consistent */
2960 LASSERT(rs->rs_svcpt == svcpt);
2963 LASSERT(!rs->rs_scheduled);
2964 LASSERT(!rs->rs_scheduled_ever);
2965 LASSERT(!rs->rs_handled);
2966 LASSERT(!rs->rs_on_net);
2967 LASSERT(rs->rs_export == NULL);
2968 LASSERT(list_empty(&rs->rs_obd_list));
2969 LASSERT(list_empty(&rs->rs_exp_list));
2971 exp = class_export_get(req->rq_export);
2973 /* disable reply scheduling while I'm setting up */
2974 rs->rs_scheduled = 1;
2976 rs->rs_xid = req->rq_xid;
2977 rs->rs_transno = req->rq_transno;
2978 rs->rs_export = exp;
2979 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2981 spin_lock(&exp->exp_uncommitted_replies_lock);
2982 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2983 rs->rs_transno, exp->exp_last_committed);
2984 if (rs->rs_transno > exp->exp_last_committed) {
2985 /* not committed already */
2986 list_add_tail(&rs->rs_obd_list,
2987 &exp->exp_uncommitted_replies);
2989 spin_unlock(&exp->exp_uncommitted_replies_lock);
2991 spin_lock(&exp->exp_lock);
2992 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2993 spin_unlock(&exp->exp_lock);
2995 netrc = target_send_reply_msg(req, rc, fail_id);
2997 spin_lock(&svcpt->scp_rep_lock);
2999 atomic_inc(&svcpt->scp_nreps_difficult);
3002 /* error sending: reply is off the net. Also we need +1
3003 * reply ref until ptlrpc_handle_rs() is done
3004 * with the reply state (if the send was successful, there
3005 * would have been +1 ref for the net, which
3006 * reply_out_callback leaves alone) */
3008 ptlrpc_rs_addref(rs);
3011 spin_lock(&rs->rs_lock);
3012 if (rs->rs_transno <= exp->exp_last_committed ||
3013 (!rs->rs_on_net && !rs->rs_no_ack) ||
3014 list_empty(&rs->rs_exp_list) || /* completed already */
3015 list_empty(&rs->rs_obd_list)) {
3016 CDEBUG(D_HA, "Schedule reply immediately\n");
3017 ptlrpc_dispatch_difficult_reply(rs);
3019 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3020 rs->rs_scheduled = 0; /* allow notifier to schedule */
3022 spin_unlock(&rs->rs_lock);
3023 spin_unlock(&svcpt->scp_rep_lock);
3027 enum ldlm_mode lck_compat_array[] = {
3028 [LCK_EX] = LCK_COMPAT_EX,
3029 [LCK_PW] = LCK_COMPAT_PW,
3030 [LCK_PR] = LCK_COMPAT_PR,
3031 [LCK_CW] = LCK_COMPAT_CW,
3032 [LCK_CR] = LCK_COMPAT_CR,
3033 [LCK_NL] = LCK_COMPAT_NL,
3034 [LCK_GROUP] = LCK_COMPAT_GROUP,
3035 [LCK_COS] = LCK_COMPAT_COS,
3039 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3040 * not escape to the user level.
3042 int ldlm_error2errno(enum ldlm_error error)
3048 case ELDLM_LOCK_MATCHED:
3051 case ELDLM_LOCK_CHANGED:
3054 case ELDLM_LOCK_ABORTED:
3057 case ELDLM_LOCK_REPLACED:
3060 case ELDLM_NO_LOCK_DATA:
3063 case ELDLM_NAMESPACE_EXISTS:
3066 case ELDLM_BAD_NAMESPACE:
3070 if (((int)error) < 0) { /* cast to signed type */
3071 result = error; /* as ldlm_error can be unsigned */
3073 CERROR("Invalid DLM result code: %d\n", error);
3079 EXPORT_SYMBOL(ldlm_error2errno);
3082 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3084 enum ldlm_error ldlm_errno2error(int err_no)
3093 error = ELDLM_LOCK_CHANGED;
3096 error = ELDLM_LOCK_ABORTED;
3099 error = ELDLM_LOCK_REPLACED;
3102 error = ELDLM_NO_LOCK_DATA;
3105 error = ELDLM_NAMESPACE_EXISTS;
3108 error = ELDLM_BAD_NAMESPACE;
3116 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3117 void ldlm_dump_export_locks(struct obd_export *exp)
3119 spin_lock(&exp->exp_locks_list_guard);
3120 if (!list_empty(&exp->exp_locks_list)) {
3121 struct ldlm_lock *lock;
3123 CERROR("dumping locks for export %p,"
3124 "ignore if the unmount doesn't hang\n", exp);
3125 list_for_each_entry(lock, &exp->exp_locks_list,
3127 LDLM_ERROR(lock, "lock:");
3129 spin_unlock(&exp->exp_locks_list_guard);
3133 #ifdef HAVE_SERVER_SUPPORT
3134 static int target_bulk_timeout(void *data)
3137 /* We don't fail the connection here, because having the export
3138 * killed makes the (vital) call to commitrw very sad.
3143 static inline const char *bulk2type(struct ptlrpc_request *req)
3145 if (req->rq_bulk_read)
3147 if (req->rq_bulk_write)
3152 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3153 struct l_wait_info *lwi)
3155 struct ptlrpc_request *req = desc->bd_req;
3156 time64_t start = ktime_get_real_seconds();
3162 /* If there is eviction in progress, wait for it to finish. */
3163 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3164 *lwi = LWI_INTR(NULL, NULL);
3165 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3166 !atomic_read(&exp->exp_obd->
3167 obd_evict_inprogress),
3171 /* Check if client was evicted or reconnected already. */
3172 if (exp->exp_failed ||
3173 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3176 if (req->rq_bulk_read)
3177 rc = sptlrpc_svc_wrap_bulk(req, desc);
3179 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3180 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3181 else /* old version, bulk matchbits is rq_xid */
3182 req->rq_mbits = req->rq_xid;
3185 rc = ptlrpc_start_bulk_transfer(desc);
3189 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3190 bulk2type(req), rc);
3194 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3195 ptlrpc_abort_bulk(desc);
3199 /* limit actual bulk transfer to bulk_timeout seconds */
3200 deadline = start + bulk_timeout;
3201 if (deadline > req->rq_deadline)
3202 deadline = req->rq_deadline;
3205 time64_t timeoutl = deadline - ktime_get_real_seconds();
3206 long timeout_jiffies = timeoutl <= 0 ?
3207 1 : cfs_time_seconds(timeoutl);
3208 time64_t rq_deadline;
3210 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3211 cfs_time_seconds(1),
3212 target_bulk_timeout, desc);
3213 rc = l_wait_event(desc->bd_waitq,
3214 !ptlrpc_server_bulk_active(desc) ||
3217 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3219 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3220 /* Wait again if we changed rq_deadline. */
3221 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3222 deadline = start + bulk_timeout;
3223 if (deadline > rq_deadline)
3224 deadline = rq_deadline;
3225 } while (rc == -ETIMEDOUT &&
3226 deadline > ktime_get_real_seconds());
3228 if (rc == -ETIMEDOUT) {
3229 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3230 bulk2type(req), deadline - start,
3231 ktime_get_real_seconds() - deadline);
3232 ptlrpc_abort_bulk(desc);
3233 } else if (exp->exp_failed) {
3234 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3237 ptlrpc_abort_bulk(desc);
3238 } else if (exp->exp_conn_cnt >
3239 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3240 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3242 /* We don't reply anyway. */
3244 ptlrpc_abort_bulk(desc);
3245 } else if (desc->bd_failure) {
3246 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3248 /* XXX should this be a different errno? */
3251 if (req->rq_bulk_write)
3252 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3253 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3254 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3255 bulk2type(req), desc->bd_nob_transferred,
3257 /* XXX should this be a different errno? */
3264 EXPORT_SYMBOL(target_bulk_io);
3266 #endif /* HAVE_SERVER_SUPPORT */