4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * This file deals with various client/target related logic including recovery.
36 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
40 #define DEBUG_SUBSYSTEM S_LDLM
42 #include <linux/jiffies.h>
43 #include <linux/kthread.h>
44 #include <libcfs/libcfs.h>
46 #include <obd_class.h>
47 #include <lustre_dlm.h>
48 #include <lustre_net.h>
49 #include <lustre_sec.h>
50 #include "ldlm_internal.h"
52 /* @priority: If non-zero, move the selected connection to the list head.
53 * @create: If zero, only search in existing connections.
55 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
56 int priority, int create)
58 struct ptlrpc_connection *ptlrpc_conn;
59 struct obd_import_conn *imp_conn = NULL, *item;
60 lnet_nid_t nid4refnet = LNET_NID_ANY;
64 if (!create && !priority) {
65 CDEBUG(D_HA, "Nothing to do\n");
69 if (imp->imp_connection &&
70 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
71 /* nid4refnet is used to restrict network connections */
72 nid4refnet = imp->imp_connection->c_self;
73 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
75 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
80 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
82 GOTO(out_put, rc = -ENOMEM);
85 spin_lock(&imp->imp_lock);
86 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
87 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
89 list_del(&item->oic_item);
90 list_add(&item->oic_item,
92 item->oic_last_attempt = 0;
94 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
95 imp, imp->imp_obd->obd_name, uuid->uuid,
96 (priority ? ", moved to head" : ""));
97 spin_unlock(&imp->imp_lock);
98 GOTO(out_free, rc = 0);
101 /* No existing import connection found for \a uuid. */
103 imp_conn->oic_conn = ptlrpc_conn;
104 imp_conn->oic_uuid = *uuid;
105 imp_conn->oic_last_attempt = 0;
107 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
109 list_add_tail(&imp_conn->oic_item,
110 &imp->imp_conn_list);
111 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
112 imp, imp->imp_obd->obd_name, uuid->uuid,
113 (priority ? "head" : "tail"));
115 spin_unlock(&imp->imp_lock);
116 GOTO(out_free, rc = -ENOENT);
119 spin_unlock(&imp->imp_lock);
123 OBD_FREE(imp_conn, sizeof(*imp_conn));
125 ptlrpc_connection_put(ptlrpc_conn);
129 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
131 return import_set_conn(imp, uuid, 1, 0);
134 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
137 return import_set_conn(imp, uuid, priority, 1);
139 EXPORT_SYMBOL(client_import_add_conn);
141 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
143 struct obd_import_conn *imp_conn;
144 struct obd_export *dlmexp;
148 spin_lock(&imp->imp_lock);
149 if (list_empty(&imp->imp_conn_list)) {
150 LASSERT(!imp->imp_connection);
154 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
155 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
157 LASSERT(imp_conn->oic_conn);
159 if (imp_conn == imp->imp_conn_current) {
160 LASSERT(imp_conn->oic_conn == imp->imp_connection);
162 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
163 imp->imp_state != LUSTRE_IMP_DISCON) {
164 CERROR("can't remove current connection\n");
165 GOTO(out, rc = -EBUSY);
168 ptlrpc_connection_put(imp->imp_connection);
169 imp->imp_connection = NULL;
171 dlmexp = class_conn2export(&imp->imp_dlm_handle);
172 if (dlmexp && dlmexp->exp_connection) {
173 LASSERT(dlmexp->exp_connection ==
175 ptlrpc_connection_put(dlmexp->exp_connection);
176 dlmexp->exp_connection = NULL;
180 class_export_put(dlmexp);
183 list_del(&imp_conn->oic_item);
184 ptlrpc_connection_put(imp_conn->oic_conn);
185 OBD_FREE(imp_conn, sizeof(*imp_conn));
186 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
187 imp, imp->imp_obd->obd_name, uuid->uuid);
192 spin_unlock(&imp->imp_lock);
194 CERROR("connection %s not found\n", uuid->uuid);
197 EXPORT_SYMBOL(client_import_del_conn);
200 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
201 * to find a conn uuid of \a imp which can reach \a peer.
203 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
204 struct obd_uuid *uuid)
206 struct obd_import_conn *conn;
210 spin_lock(&imp->imp_lock);
211 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
212 /* Check if conn UUID does have this peer NID. */
213 if (class_check_uuid(&conn->oic_uuid, peer)) {
214 *uuid = conn->oic_uuid;
219 spin_unlock(&imp->imp_lock);
222 EXPORT_SYMBOL(client_import_find_conn);
224 void client_destroy_import(struct obd_import *imp)
226 /* Drop security policy instance after all RPCs have finished/aborted
227 * to let all busy contexts be released. */
228 class_import_get(imp);
229 class_destroy_import(imp);
230 sptlrpc_import_sec_put(imp);
231 class_import_put(imp);
233 EXPORT_SYMBOL(client_destroy_import);
236 * Check whether or not the OSC is on MDT.
239 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
241 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
244 static int osc_on_mdt(char *obdname)
248 ptr = strrchr(obdname, '-');
252 if (strncmp(ptr + 1, "MDT", 3) == 0)
258 /* Configure an RPC client OBD device.
263 * 3 - inactive-on-startup
264 * 4 - restrictive net
266 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
268 struct client_obd *cli = &obddev->u.cli;
269 struct obd_import *imp;
270 struct obd_uuid server_uuid;
271 int rq_portal, rp_portal, connect_op;
272 char *name = obddev->obd_type->typ_name;
273 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
274 char *cli_name = lustre_cfg_buf(lcfg, 0);
275 struct ptlrpc_connection fake_conn = { .c_self = 0,
276 .c_remote_uuid.uuid[0] = 0 };
280 /* In a more perfect world, we would hang a ptlrpc_client off of
281 * obd_type and just use the values from there. */
282 if (!strcmp(name, LUSTRE_OSC_NAME)) {
283 rq_portal = OST_REQUEST_PORTAL;
284 rp_portal = OSC_REPLY_PORTAL;
285 connect_op = OST_CONNECT;
286 cli->cl_sp_me = LUSTRE_SP_CLI;
287 cli->cl_sp_to = LUSTRE_SP_OST;
288 ns_type = LDLM_NS_TYPE_OSC;
289 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
290 !strcmp(name, LUSTRE_LWP_NAME)) {
291 rq_portal = MDS_REQUEST_PORTAL;
292 rp_portal = MDC_REPLY_PORTAL;
293 connect_op = MDS_CONNECT;
294 if (is_lwp_on_ost(cli_name))
295 cli->cl_sp_me = LUSTRE_SP_OST;
296 else if (is_lwp_on_mdt(cli_name))
297 cli->cl_sp_me = LUSTRE_SP_MDT;
299 cli->cl_sp_me = LUSTRE_SP_CLI;
300 cli->cl_sp_to = LUSTRE_SP_MDT;
301 ns_type = LDLM_NS_TYPE_MDC;
302 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
303 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
304 /* OSP_on_MDT for other MDTs */
305 connect_op = MDS_CONNECT;
306 cli->cl_sp_to = LUSTRE_SP_MDT;
307 ns_type = LDLM_NS_TYPE_MDC;
308 rq_portal = OUT_PORTAL;
310 /* OSP on MDT for OST */
311 connect_op = OST_CONNECT;
312 cli->cl_sp_to = LUSTRE_SP_OST;
313 ns_type = LDLM_NS_TYPE_OSC;
314 rq_portal = OST_REQUEST_PORTAL;
316 rp_portal = OSC_REPLY_PORTAL;
317 cli->cl_sp_me = LUSTRE_SP_MDT;
318 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
319 rq_portal = MGS_REQUEST_PORTAL;
320 rp_portal = MGC_REPLY_PORTAL;
321 connect_op = MGS_CONNECT;
322 cli->cl_sp_me = LUSTRE_SP_MGC;
323 cli->cl_sp_to = LUSTRE_SP_MGS;
324 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
325 ns_type = LDLM_NS_TYPE_MGC;
327 CERROR("unknown client OBD type \"%s\", can't setup\n",
332 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
333 CERROR("requires a TARGET UUID\n");
337 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
338 CERROR("client UUID must be less than 38 characters\n");
342 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
343 CERROR("setup requires a SERVER UUID\n");
347 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
348 CERROR("target UUID must be less than 38 characters\n");
352 init_rwsem(&cli->cl_sem);
353 mutex_init(&cli->cl_mgc_mutex);
355 init_rwsem(&cli->cl_seq_rwsem);
356 cli->cl_conn_count = 0;
357 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
358 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
359 sizeof(server_uuid)));
361 cli->cl_dirty_pages = 0;
362 cli->cl_avail_grant = 0;
363 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
364 /* cl_dirty_max_pages may be changed at connect time in
365 * ptlrpc_connect_interpret(). */
366 client_adjust_max_dirty(cli);
367 INIT_LIST_HEAD(&cli->cl_cache_waiters);
368 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
369 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
370 INIT_LIST_HEAD(&cli->cl_loi_write_list);
371 INIT_LIST_HEAD(&cli->cl_loi_read_list);
372 spin_lock_init(&cli->cl_loi_list_lock);
373 atomic_set(&cli->cl_pending_w_pages, 0);
374 atomic_set(&cli->cl_pending_r_pages, 0);
375 cli->cl_r_in_flight = 0;
376 cli->cl_w_in_flight = 0;
378 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
379 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
380 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
381 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
382 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
383 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
386 INIT_LIST_HEAD(&cli->cl_lru_osc);
387 atomic_set(&cli->cl_lru_shrinkers, 0);
388 atomic_long_set(&cli->cl_lru_busy, 0);
389 atomic_long_set(&cli->cl_lru_in_list, 0);
390 INIT_LIST_HEAD(&cli->cl_lru_list);
391 spin_lock_init(&cli->cl_lru_list_lock);
392 atomic_long_set(&cli->cl_unstable_count, 0);
393 INIT_LIST_HEAD(&cli->cl_shrink_list);
394 INIT_LIST_HEAD(&cli->cl_grant_chain);
396 INIT_LIST_HEAD(&cli->cl_flight_waiters);
397 cli->cl_rpcs_in_flight = 0;
399 init_waitqueue_head(&cli->cl_destroy_waitq);
400 atomic_set(&cli->cl_destroy_in_flight, 0);
402 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
403 #ifdef ENABLE_CHECKSUM
404 /* Turn on checksumming by default. */
405 cli->cl_checksum = 1;
407 * The supported checksum types will be worked out at connect time
408 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
411 cli->cl_cksum_type = cli->cl_supp_cksum_types;
413 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
415 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
416 * from OFD after connecting. */
417 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
419 cli->cl_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
421 /* set cl_chunkbits default value to PAGE_SHIFT,
422 * it will be updated at OSC connection time. */
423 cli->cl_chunkbits = PAGE_SHIFT;
425 if (!strcmp(name, LUSTRE_MDC_NAME)) {
426 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
427 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
428 cli->cl_max_rpcs_in_flight = 2;
429 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
430 cli->cl_max_rpcs_in_flight = 3;
431 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
432 cli->cl_max_rpcs_in_flight = 4;
434 if (osc_on_mdt(obddev->obd_name))
435 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
437 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
440 spin_lock_init(&cli->cl_mod_rpcs_lock);
441 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
442 cli->cl_max_mod_rpcs_in_flight = 0;
443 cli->cl_mod_rpcs_in_flight = 0;
444 cli->cl_close_rpcs_in_flight = 0;
445 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
446 cli->cl_mod_tag_bitmap = NULL;
448 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
450 if (connect_op == MDS_CONNECT) {
451 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
452 OBD_ALLOC(cli->cl_mod_tag_bitmap,
453 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
454 if (cli->cl_mod_tag_bitmap == NULL)
455 GOTO(err, rc = -ENOMEM);
460 CERROR("ldlm_get_ref failed: %d\n", rc);
464 ptlrpc_init_client(rq_portal, rp_portal, name,
465 &obddev->obd_ldlm_client);
467 imp = class_new_import(obddev);
469 GOTO(err_ldlm, rc = -ENOENT);
470 imp->imp_client = &obddev->obd_ldlm_client;
471 imp->imp_connect_op = connect_op;
472 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
473 LUSTRE_CFG_BUFLEN(lcfg, 1));
474 class_import_put(imp);
476 if (lustre_cfg_buf(lcfg, 4)) {
477 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
479 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
481 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
482 obddev->obd_name, lustre_cfg_string(lcfg, 4),
484 GOTO(err_import, rc);
486 fake_conn.c_self = LNET_MKNID(refnet, 0);
487 imp->imp_connection = &fake_conn;
490 rc = client_import_add_conn(imp, &server_uuid, 1);
492 CERROR("can't add initial connection\n");
493 GOTO(err_import, rc);
495 imp->imp_connection = NULL;
497 cli->cl_import = imp;
498 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
499 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
501 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
502 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
503 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
504 name, obddev->obd_name,
505 cli->cl_target_uuid.uuid);
506 spin_lock(&imp->imp_lock);
507 imp->imp_deactive = 1;
508 spin_unlock(&imp->imp_lock);
512 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
513 LDLM_NAMESPACE_CLIENT,
514 LDLM_NAMESPACE_GREEDY,
516 if (obddev->obd_namespace == NULL) {
517 CERROR("Unable to create client namespace - %s\n",
519 GOTO(err_import, rc = -ENOMEM);
525 class_destroy_import(imp);
529 if (cli->cl_mod_tag_bitmap != NULL)
530 OBD_FREE(cli->cl_mod_tag_bitmap,
531 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
532 cli->cl_mod_tag_bitmap = NULL;
536 EXPORT_SYMBOL(client_obd_setup);
538 int client_obd_cleanup(struct obd_device *obddev)
540 struct client_obd *cli = &obddev->u.cli;
543 ldlm_namespace_free_post(obddev->obd_namespace);
544 obddev->obd_namespace = NULL;
546 obd_cleanup_client_import(obddev);
547 LASSERT(obddev->u.cli.cl_import == NULL);
551 if (cli->cl_mod_tag_bitmap != NULL)
552 OBD_FREE(cli->cl_mod_tag_bitmap,
553 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
554 cli->cl_mod_tag_bitmap = NULL;
558 EXPORT_SYMBOL(client_obd_cleanup);
560 /* ->o_connect() method for client side (OSC and MDC and MGC) */
561 int client_connect_import(const struct lu_env *env,
562 struct obd_export **exp,
563 struct obd_device *obd, struct obd_uuid *cluuid,
564 struct obd_connect_data *data, void *localdata)
566 struct client_obd *cli = &obd->u.cli;
567 struct obd_import *imp = cli->cl_import;
568 struct obd_connect_data *ocd;
569 struct lustre_handle conn = { 0 };
574 down_write(&cli->cl_sem);
575 if (cli->cl_conn_count > 0)
576 GOTO(out_sem, rc = -EALREADY);
578 rc = class_connect(&conn, obd, cluuid);
582 cli->cl_conn_count++;
583 *exp = class_conn2export(&conn);
585 LASSERT(obd->obd_namespace);
587 imp->imp_dlm_handle = conn;
588 rc = ptlrpc_init_import(imp);
592 ocd = &imp->imp_connect_data;
595 imp->imp_connect_flags_orig = data->ocd_connect_flags;
596 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
599 rc = ptlrpc_connect_import(imp);
601 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
604 LASSERT(*exp != NULL && (*exp)->exp_connection);
607 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
608 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
609 data->ocd_connect_flags, ocd->ocd_connect_flags);
610 data->ocd_connect_flags = ocd->ocd_connect_flags;
611 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
614 ptlrpc_pinger_add_import(imp);
620 cli->cl_conn_count--;
621 class_disconnect(*exp);
625 up_write(&cli->cl_sem);
629 EXPORT_SYMBOL(client_connect_import);
631 int client_disconnect_export(struct obd_export *exp)
633 struct obd_device *obd = class_exp2obd(exp);
634 struct client_obd *cli;
635 struct obd_import *imp;
640 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
641 exp, exp ? exp->exp_handle.h_cookie : -1);
646 imp = cli->cl_import;
648 down_write(&cli->cl_sem);
649 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
652 if (cli->cl_conn_count == 0) {
653 CERROR("disconnecting disconnected device (%s)\n",
655 GOTO(out_disconnect, rc = -EINVAL);
658 cli->cl_conn_count--;
659 if (cli->cl_conn_count != 0)
660 GOTO(out_disconnect, rc = 0);
662 /* Mark import deactivated now, so we don't try to reconnect if any
663 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
664 * fully deactivate the import, or that would drop all requests. */
665 spin_lock(&imp->imp_lock);
666 imp->imp_deactive = 1;
667 spin_unlock(&imp->imp_lock);
669 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
670 * delete it regardless. (It's safe to delete an import that was
672 (void)ptlrpc_pinger_del_import(imp);
674 if (obd->obd_namespace != NULL) {
675 /* obd_force == local only */
676 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
677 obd->obd_force ? LCF_LOCAL : 0, NULL);
678 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
681 /* There's no need to hold sem while disconnecting an import,
682 * and it may actually cause deadlock in GSS. */
683 up_write(&cli->cl_sem);
684 rc = ptlrpc_disconnect_import(imp, 0);
685 down_write(&cli->cl_sem);
687 ptlrpc_invalidate_import(imp);
692 /* Use server style - class_disconnect should be always called for
694 err = class_disconnect(exp);
698 up_write(&cli->cl_sem);
702 EXPORT_SYMBOL(client_disconnect_export);
704 #ifdef HAVE_SERVER_SUPPORT
705 int server_disconnect_export(struct obd_export *exp)
710 /* Disconnect early so that clients can't keep using export. */
711 rc = class_disconnect(exp);
712 /* Close import to avoid sending any requests. */
713 if (exp->exp_imp_reverse)
714 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
716 ldlm_bl_thread_wakeup();
718 /* complete all outstanding replies */
719 spin_lock(&exp->exp_lock);
720 while (!list_empty(&exp->exp_outstanding_replies)) {
721 struct ptlrpc_reply_state *rs =
722 list_entry(exp->exp_outstanding_replies.next,
723 struct ptlrpc_reply_state, rs_exp_list);
724 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
726 spin_lock(&svcpt->scp_rep_lock);
728 list_del_init(&rs->rs_exp_list);
730 spin_lock(&rs->rs_lock);
731 /* clear rs_convert_lock to make sure rs is handled and put */
732 rs->rs_convert_lock = 0;
733 ptlrpc_schedule_difficult_reply(rs);
734 spin_unlock(&rs->rs_lock);
736 spin_unlock(&svcpt->scp_rep_lock);
738 spin_unlock(&exp->exp_lock);
742 EXPORT_SYMBOL(server_disconnect_export);
744 /* --------------------------------------------------------------------------
745 * from old lib/target.c
746 * -------------------------------------------------------------------------- */
748 static int target_handle_reconnect(struct lustre_handle *conn,
749 struct obd_export *exp,
750 struct obd_uuid *cluuid)
752 struct obd_device *target;
753 struct lustre_handle *hdl;
760 hdl = &exp->exp_imp_reverse->imp_remote_handle;
761 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
762 conn->cookie = exp->exp_handle.h_cookie;
763 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
764 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
768 target = exp->exp_obd;
770 /* Might be a re-connect after a partition. */
771 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
772 LCONSOLE_WARN("%s: already connected client %s (at %s) "
773 "with handle %#llx. Rejecting client "
774 "with the same UUID trying to reconnect "
775 "with handle %#llx\n", target->obd_name,
776 obd_uuid2str(&exp->exp_client_uuid),
777 obd_export_nid2str(exp),
778 hdl->cookie, conn->cookie);
779 memset(conn, 0, sizeof *conn);
780 /* target_handle_connect() treats EALREADY and
781 * -EALREADY differently. -EALREADY is an error
782 * (same UUID, different handle). */
786 if (!target->obd_recovering) {
787 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
788 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
789 obd_export_nid2str(exp));
790 GOTO(out_already, rc);
793 now = ktime_get_seconds();
794 deadline = jiffies_to_msecs(target->obd_recovery_timer.expires) /
796 if (now < deadline) {
797 struct target_distribute_txn_data *tdtd;
802 timeout = deadline - now;
803 tdtd = class_exp2tgt(exp)->lut_tdtd;
804 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
805 buf = tdtd->tdtd_show_update_logs_retrievers(
806 tdtd->tdtd_show_retrievers_cbdata,
810 LCONSOLE_WARN("%s: Recovery already passed deadline "
811 "%lld:%.02lld. It is due to DNE recovery "
812 "failed/stuck on the %d MDT(s):%s. "
813 "Please wait until all MDTs recovered "
814 "or abort the recovery by force.\n",
815 target->obd_name, timeout / 60,
817 buf ? buf : "unknown (not enough RAM)");
819 LCONSOLE_WARN("%s: Recovery already passed deadline "
820 "%lld:%.02lld. If you do not want to wait "
821 "more, please abort the recovery by "
822 "force.\n", target->obd_name,
823 timeout / 60, timeout % 60);
828 timeout = now - deadline;
829 LCONSOLE_WARN("%s: Recovery already passed deadline"
830 " %lld:%.02lld, It is most likely due to DNE"
831 " recovery is failed or stuck, please wait a"
832 " few more minutes or abort the recovery.\n",
833 target->obd_name, timeout / 60, timeout % 60);
837 conn->cookie = exp->exp_handle.h_cookie;
838 /* target_handle_connect() treats EALREADY and
839 * -EALREADY differently. EALREADY means we are
840 * doing a valid reconnect from the same client. */
845 check_and_start_recovery_timer(struct obd_device *obd,
846 struct ptlrpc_request *req, int new_client);
849 * update flags for import during reconnect process
851 static int rev_import_flags_update(struct obd_import *revimp,
852 struct ptlrpc_request *req)
855 struct obd_connect_data *data;
857 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
859 if (data->ocd_connect_flags & OBD_CONNECT_AT)
860 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
862 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
864 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
866 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
868 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
869 revimp->imp_client->cli_name,
870 libcfs_id2str(req->rq_peer), rc);
878 * Allocate a new reverse import for an export.
880 * \retval -errno in case error hit
881 * \retval 0 if reverse import correctly init
883 int rev_import_init(struct obd_export *export)
885 struct obd_device *obd = export->exp_obd;
886 struct obd_import *revimp;
888 LASSERT(export->exp_imp_reverse == NULL);
890 revimp = class_new_import(obd);
894 revimp->imp_remote_handle.cookie = 0ULL;
895 revimp->imp_client = &obd->obd_ldlm_client;
896 revimp->imp_dlm_fake = 1;
898 /* it is safe to connect import in new state as no sends possible */
899 spin_lock(&export->exp_lock);
900 export->exp_imp_reverse = revimp;
901 spin_unlock(&export->exp_lock);
902 class_import_put(revimp);
906 EXPORT_SYMBOL(rev_import_init);
909 * Handle reconnect for an export.
911 * \param exp export to handle reconnect process
912 * \param req client reconnect request
914 * \retval -rc in case securitfy flavor can't be changed
915 * \retval 0 in case none problems
917 static int rev_import_reconnect(struct obd_export *exp,
918 struct ptlrpc_request *req)
920 struct obd_import *revimp = exp->exp_imp_reverse;
921 struct lustre_handle *lh;
924 /* avoid sending a request until import flags are changed */
925 ptlrpc_import_enter_resend(revimp);
927 if (revimp->imp_connection != NULL)
928 ptlrpc_connection_put(revimp->imp_connection);
931 * client from recovery don't have a handle so we need to take from
932 * request. it may produce situation when wrong client connected
933 * to recovery as we trust a client uuid
935 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
936 revimp->imp_remote_handle = *lh;
938 /* unknown versions will be caught in
939 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
940 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
942 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
944 rc = rev_import_flags_update(revimp, req);
946 /* it is safe to still be in RECOVERY phase as we are not able
947 * to setup correct security flavor so requests are not able to
948 * be delivered correctly */
952 /* resend all rpc's via new connection */
953 return ptlrpc_import_recovery_state_machine(revimp);
956 int target_handle_connect(struct ptlrpc_request *req)
958 struct obd_device *target = NULL;
959 struct obd_export *export = NULL;
960 /* connect handle - filled from target_handle_reconnect in
962 struct lustre_handle conn;
963 struct lustre_handle *tmp;
964 struct obd_uuid cluuid;
969 bool mds_conn = false, lw_client = false, initial_conn = false;
970 bool mds_mds_conn = false;
971 bool new_mds_mds_conn = false;
972 struct obd_connect_data *data, *tmpdata;
974 lnet_nid_t *client_nid = NULL;
977 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
979 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
981 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
982 GOTO(out, rc = -EINVAL);
985 target = class_dev_by_str(str);
987 deuuidify(str, NULL, &target_start, &target_len);
988 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
989 "from %s (no target). If you are running "
990 "an HA pair check that the target is "
991 "mounted on the other server.\n", str,
992 libcfs_nid2str(req->rq_peer.nid));
993 GOTO(out, rc = -ENODEV);
996 spin_lock(&target->obd_dev_lock);
998 target->obd_conn_inprogress++;
1000 if (target->obd_stopping || !target->obd_set_up) {
1001 spin_unlock(&target->obd_dev_lock);
1003 deuuidify(str, NULL, &target_start, &target_len);
1004 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1005 target_len, target_start,
1006 libcfs_nid2str(req->rq_peer.nid),
1007 (target->obd_stopping ?
1008 "stopping" : "not set up"));
1009 GOTO(out, rc = -ENODEV);
1012 if (target->obd_no_conn) {
1013 spin_unlock(&target->obd_dev_lock);
1015 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1016 "from %s\n", target->obd_name,
1017 libcfs_nid2str(req->rq_peer.nid));
1018 GOTO(out, rc = -EAGAIN);
1021 spin_unlock(&target->obd_dev_lock);
1023 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1025 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1026 GOTO(out, rc = -EINVAL);
1029 obd_str2uuid(&cluuid, str);
1031 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1033 GOTO(out, rc = -EPROTO);
1037 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1039 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1041 GOTO(out, rc = -EPROTO);
1043 rc = req_capsule_server_pack(&req->rq_pill);
1047 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1048 /* Don't allow clients to connect that are using old 1.8 format
1049 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1050 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1051 * FULL20 flag should be set on all connections since 2.0, but no
1052 * longer affects behaviour.
1054 * Later this check will be disabled and the flag can be retired
1055 * completely once interop with 3.0 is no longer needed.
1057 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1058 GOTO(out, rc = -EPROTO);
1060 /* Don't allow liblustre clients to connect.
1061 * - testing was disabled in v2_2_50_0-61-g6a75d65
1062 * - building was disabled in v2_5_58_0-28-g7277179
1063 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1064 * - clients were refused connect for version difference > 0.0.1.32 */
1065 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1066 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1067 GOTO(out, rc = -EPROTO);
1071 /* Note: lw_client is needed in MDS-MDS failover during update log
1072 * processing, so we needs to allow lw_client to be connected at
1073 * anytime, instead of only the initial connection */
1074 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1076 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1077 initial_conn = true;
1078 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1079 mds_mds_conn = (data->ocd_connect_flags &
1080 OBD_CONNECT_MDS_MDS) != 0;
1082 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1083 * for Imperative Recovery connection from MGC to MGS.
1085 * Via check OBD_CONNECT_FID, we can distinguish whether
1086 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1087 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1090 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1091 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1092 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1093 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1094 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1095 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1097 /* We do not support the MDT-MDT interoperations with
1098 * different version MDT because of protocol changes. */
1099 if (unlikely(major != LUSTRE_MAJOR ||
1100 minor != LUSTRE_MINOR ||
1101 abs(patch - LUSTRE_PATCH) > 3)) {
1102 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1103 "connection from different version MDT "
1104 "(%d.%d.%d.%d) %s %s\n",
1105 target->obd_name, LUSTRE_MAJOR,
1106 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1107 major, minor, patch,
1108 OBD_OCD_VERSION_FIX(data->ocd_version),
1109 libcfs_nid2str(req->rq_peer.nid), str);
1111 GOTO(out, rc = -EPROTO);
1116 /* lctl gets a backstage, all-access pass. */
1117 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1118 goto dont_check_exports;
1120 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1124 /* We've found an export in the hash. */
1126 spin_lock(&export->exp_lock);
1128 if (export->exp_connecting) { /* bug 9635, et. al. */
1129 spin_unlock(&export->exp_lock);
1130 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1131 export->exp_obd->obd_name, export,
1132 libcfs_nid2str(req->rq_peer.nid));
1133 class_export_put(export);
1136 } else if ((mds_conn || (lw_client && initial_conn) ||
1137 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1138 export->exp_connection != NULL) {
1139 spin_unlock(&export->exp_lock);
1140 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1141 /* MDS or LWP reconnected after failover. */
1142 LCONSOLE_WARN("%s: Received %s connection from "
1143 "%s, removing former export from %s\n",
1144 target->obd_name, mds_conn ? "MDS" : "LWP",
1145 libcfs_nid2str(req->rq_peer.nid),
1146 libcfs_nid2str(export->exp_connection->c_peer.nid));
1148 /* New MDS connection from the same NID. */
1149 LCONSOLE_WARN("%s: Received new %s connection from "
1150 "%s, removing former export from same NID\n",
1151 target->obd_name, mds_conn ? "MDS" : "LWP",
1152 libcfs_nid2str(req->rq_peer.nid));
1155 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1156 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1157 /* Because exports between MDTs will always be
1158 * kept, let's do not fail such export if they
1159 * come from the same NID, otherwise it might
1160 * cause eviction between MDTs, which might
1161 * cause namespace inconsistency */
1162 spin_lock(&export->exp_lock);
1163 export->exp_connecting = 1;
1164 spin_unlock(&export->exp_lock);
1165 conn.cookie = export->exp_handle.h_cookie;
1168 class_fail_export(export);
1169 class_export_put(export);
1173 } else if (export->exp_connection != NULL && initial_conn &&
1174 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1175 spin_unlock(&export->exp_lock);
1176 /* In MDS failover we have static UUID but NID can change. */
1177 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1178 "existing nid %s is already connected\n",
1179 target->obd_name, cluuid.uuid,
1180 libcfs_nid2str(req->rq_peer.nid),
1182 export->exp_connection->c_peer.nid));
1184 class_export_put(export);
1187 export->exp_connecting = 1;
1188 spin_unlock(&export->exp_lock);
1189 LASSERT(export->exp_obd == target);
1191 rc = target_handle_reconnect(&conn, export, &cluuid);
1194 /* If we found an export, we already unlocked. */
1197 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1198 } else if (req->rq_export == NULL &&
1199 atomic_read(&export->exp_rpc_count) > 0) {
1200 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1201 "still busy with %d references\n",
1202 target->obd_name, cluuid.uuid,
1203 libcfs_nid2str(req->rq_peer.nid),
1204 atomic_read(&export->exp_refcount));
1205 GOTO(out, rc = -EBUSY);
1206 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1207 if (!strstr(cluuid.uuid, "mdt"))
1208 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1209 "known client %s (at %s) because it "
1210 "is indicating it is a new client",
1211 target->obd_name, cluuid.uuid,
1212 libcfs_nid2str(req->rq_peer.nid));
1213 GOTO(out, rc = -EALREADY);
1215 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1222 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1223 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1224 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1225 export, ktime_get_real_seconds(),
1226 export ? export->exp_last_request_time : 0);
1228 /* If this is the first time a client connects, reset the recovery
1229 * timer. Discard lightweight connections which might be local. */
1230 if (!lw_client && rc == 0 && target->obd_recovering)
1231 check_and_start_recovery_timer(target, req, export == NULL);
1233 /* We want to handle EALREADY but *not* -EALREADY from
1234 * target_handle_reconnect(), return reconnection state in a flag. */
1235 if (rc == EALREADY) {
1236 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1242 /* Tell the client if we support replayable requests. */
1243 if (target->obd_replayable)
1244 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1245 client_nid = &req->rq_peer.nid;
1247 if (export == NULL) {
1248 /* allow lightweight connections during recovery */
1249 /* allow "new" MDT to be connected during recovery, since we
1250 * need retrieve recovery update records from it */
1251 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1253 int c; /* connected */
1254 int i; /* in progress */
1256 int s; /* stale/evicted */
1258 c = atomic_read(&target->obd_connected_clients);
1259 i = atomic_read(&target->obd_lock_replay_clients);
1260 k = target->obd_max_recoverable_clients;
1261 s = target->obd_stale_clients;
1262 t = jiffies_to_msecs(target->obd_recovery_timer.expires);
1264 t -= ktime_get_seconds();
1265 LCONSOLE_WARN("%s: Denying connection for new client %s"
1266 "(at %s), waiting for %d known clients "
1267 "(%d recovered, %d in progress, and %d "
1268 "evicted) to recover in %lld:%.02lld\n",
1269 target->obd_name, cluuid.uuid,
1270 libcfs_nid2str(req->rq_peer.nid), k,
1271 c - i, i, s, t / 60, t % 60);
1275 rc = obd_connect(req->rq_svc_thread->t_env,
1276 &export, target, &cluuid, data,
1278 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1279 lustre_msg_add_op_flags(req->rq_repmsg,
1280 MSG_CONNECT_RECOVERING);
1282 conn.cookie = export->exp_handle.h_cookie;
1283 rc = rev_import_init(export);
1287 new_mds_mds_conn = true;
1290 rc = obd_reconnect(req->rq_svc_thread->t_env,
1291 export, target, &cluuid, data, client_nid);
1296 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1297 data->ocd_instance = target->u.obt.obt_instance;
1299 /* Return only the parts of obd_connect_data that we understand, so the
1300 * client knows that we don't understand the rest. */
1302 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1304 tmpdata = req_capsule_server_get(&req->rq_pill,
1306 /* Don't use struct assignment here, because the client reply
1307 * buffer may be smaller/larger than the local struct
1308 * obd_connect_data. */
1309 memcpy(tmpdata, data, min(tmpsize, size));
1312 /* If the client and the server are the same node, we will already
1313 * have an export that really points to the client's DLM export,
1314 * because we have a shared handles table.
1316 * XXX this will go away when shaver stops sending the "connect" handle
1317 * in the real "remote handle" field of the request --phik 24 Apr 2003
1319 ptlrpc_request_change_export(req, export);
1321 spin_lock(&export->exp_lock);
1322 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1323 spin_unlock(&export->exp_lock);
1324 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1325 "or equal conn_cnt: %d >= %d\n",
1326 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1327 export->exp_conn_cnt,
1328 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1330 GOTO(out, rc = -EALREADY);
1332 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1333 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1334 spin_unlock(&export->exp_lock);
1336 if (export->exp_connection != NULL) {
1337 /* Check to see if connection came from another NID. */
1338 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1339 !hlist_unhashed(&export->exp_nid_hash))
1340 cfs_hash_del(export->exp_obd->obd_nid_hash,
1341 &export->exp_connection->c_peer.nid,
1342 &export->exp_nid_hash);
1344 ptlrpc_connection_put(export->exp_connection);
1347 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1350 if (hlist_unhashed(&export->exp_nid_hash))
1351 cfs_hash_add(export->exp_obd->obd_nid_hash,
1352 &export->exp_connection->c_peer.nid,
1353 &export->exp_nid_hash);
1355 lustre_msg_set_handle(req->rq_repmsg, &conn);
1357 rc = rev_import_reconnect(export, req);
1361 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1363 __u64 transno = data->ocd_transno;
1365 spin_lock(&export->exp_lock);
1366 /* possible race with class_disconnect_stale_exports,
1367 * export may be already in the eviction process */
1368 if (export->exp_failed) {
1369 spin_unlock(&export->exp_lock);
1370 GOTO(out, rc = -ENODEV);
1372 export->exp_in_recovery = 1;
1373 export->exp_req_replay_needed = 1;
1374 export->exp_lock_replay_needed = 1;
1375 spin_unlock(&export->exp_lock);
1377 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1378 MSG_CONNECT_TRANSNO);
1379 if (has_transno && transno == 0)
1380 CWARN("Connect with zero transno!\n");
1382 if (has_transno && transno > 0 &&
1383 transno < target->obd_next_recovery_transno &&
1384 transno > target->obd_last_committed) {
1385 /* Another way is to use cmpxchg() to be lock-free. */
1386 spin_lock(&target->obd_recovery_task_lock);
1387 if (transno < target->obd_next_recovery_transno)
1388 target->obd_next_recovery_transno = transno;
1389 spin_unlock(&target->obd_recovery_task_lock);
1392 atomic_inc(&target->obd_req_replay_clients);
1393 atomic_inc(&target->obd_lock_replay_clients);
1394 /* Note: MDS-MDS connection is allowed to be connected during
1395 * recovery, no matter if the exports needs to be recoveried.
1396 * Because we need retrieve updates logs from all other MDTs.
1397 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1398 * also needs to be increased to match other recovery checking
1400 if (new_mds_mds_conn)
1401 target->obd_max_recoverable_clients++;
1402 if (atomic_inc_return(&target->obd_connected_clients) ==
1403 target->obd_max_recoverable_clients)
1404 wake_up(&target->obd_next_transno_waitq);
1407 /* Tell the client we're in recovery, when client is involved in it. */
1408 if (target->obd_recovering && !lw_client)
1409 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1413 spin_lock(&export->exp_lock);
1414 export->exp_connecting = 0;
1415 spin_unlock(&export->exp_lock);
1417 class_export_put(export);
1419 if (target != NULL) {
1420 spin_lock(&target->obd_dev_lock);
1421 target->obd_conn_inprogress--;
1422 spin_unlock(&target->obd_dev_lock);
1423 class_decref(target, "find", current);
1425 req->rq_status = rc;
1429 int target_handle_disconnect(struct ptlrpc_request *req)
1434 rc = req_capsule_server_pack(&req->rq_pill);
1438 /* Keep the rq_export around so we can send the reply. */
1439 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1444 void target_destroy_export(struct obd_export *exp)
1446 struct obd_import *imp = NULL;
1447 /* exports created from last_rcvd data, and "fake"
1448 exports created by lctl don't have an import */
1449 spin_lock(&exp->exp_lock);
1450 if (exp->exp_imp_reverse != NULL) {
1451 imp = exp->exp_imp_reverse;
1452 exp->exp_imp_reverse = NULL;
1454 spin_unlock(&exp->exp_lock);
1456 client_destroy_import(imp);
1458 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1459 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1460 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1461 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1463 EXPORT_SYMBOL(target_destroy_export);
1466 * Recovery functions
1468 static void target_request_copy_get(struct ptlrpc_request *req)
1470 class_export_rpc_inc(req->rq_export);
1471 LASSERT(list_empty(&req->rq_list));
1472 INIT_LIST_HEAD(&req->rq_replay_list);
1474 /* Increase refcount to keep request in queue. */
1475 atomic_inc(&req->rq_refcount);
1476 /* Let export know it has replays to be handled. */
1477 atomic_inc(&req->rq_export->exp_replay_count);
1480 static void target_request_copy_put(struct ptlrpc_request *req)
1482 LASSERT(list_empty(&req->rq_replay_list));
1483 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1485 atomic_dec(&req->rq_export->exp_replay_count);
1486 class_export_rpc_dec(req->rq_export);
1487 ptlrpc_server_drop_request(req);
1490 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1492 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1493 struct obd_export *exp = req->rq_export;
1494 struct ptlrpc_request *reqiter;
1495 struct ptlrpc_request *dup_req = NULL;
1500 spin_lock(&exp->exp_lock);
1501 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1503 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1511 /* We expect it with RESENT and REPLAY flags. */
1512 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1513 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1514 CERROR("invalid flags %x of resent replay\n",
1515 lustre_msg_get_flags(req->rq_reqmsg));
1517 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1520 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1522 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1523 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1527 list_add_tail(&req->rq_replay_list,
1528 &exp->exp_req_replay_queue);
1531 spin_unlock(&exp->exp_lock);
1535 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1537 LASSERT(!list_empty(&req->rq_replay_list));
1538 LASSERT(req->rq_export);
1540 spin_lock(&req->rq_export->exp_lock);
1541 list_del_init(&req->rq_replay_list);
1542 spin_unlock(&req->rq_export->exp_lock);
1545 static void target_finish_recovery(struct lu_target *lut)
1547 struct obd_device *obd = lut->lut_obd;
1550 /* Only log a recovery message when recovery has occurred. */
1551 if (obd->obd_recovery_start) {
1552 time64_t now = ktime_get_real_seconds();
1553 time64_t elapsed_time;
1555 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1556 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1557 "%d recovered and %d %s evicted.\n", obd->obd_name,
1558 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1559 obd->obd_max_recoverable_clients,
1560 atomic_read(&obd->obd_connected_clients),
1561 obd->obd_stale_clients,
1562 obd->obd_stale_clients == 1 ? "was" : "were");
1565 ldlm_reprocess_recovery_done(obd->obd_namespace);
1566 spin_lock(&obd->obd_recovery_task_lock);
1567 if (!list_empty(&obd->obd_req_replay_queue) ||
1568 !list_empty(&obd->obd_lock_replay_queue) ||
1569 !list_empty(&obd->obd_final_req_queue)) {
1570 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1572 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1573 list_empty(&obd->obd_lock_replay_queue) ? \
1575 list_empty(&obd->obd_final_req_queue) ? \
1577 spin_unlock(&obd->obd_recovery_task_lock);
1580 spin_unlock(&obd->obd_recovery_task_lock);
1582 obd->obd_recovery_end = ktime_get_real_seconds();
1584 /* When recovery finished, cleanup orphans on MDS and OST. */
1585 if (obd->obd_type && OBP(obd, postrecov)) {
1586 int rc = OBP(obd, postrecov)(obd);
1589 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1595 static void abort_req_replay_queue(struct obd_device *obd)
1597 struct ptlrpc_request *req, *n;
1598 struct list_head abort_list;
1600 INIT_LIST_HEAD(&abort_list);
1601 spin_lock(&obd->obd_recovery_task_lock);
1602 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1603 spin_unlock(&obd->obd_recovery_task_lock);
1604 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1605 DEBUG_REQ(D_WARNING, req, "aborted:");
1606 req->rq_status = -ENOTCONN;
1607 if (ptlrpc_error(req)) {
1608 DEBUG_REQ(D_ERROR, req,
1609 "failed abort_req_reply; skipping");
1611 target_exp_dequeue_req_replay(req);
1612 target_request_copy_put(req);
1616 static void abort_lock_replay_queue(struct obd_device *obd)
1618 struct ptlrpc_request *req, *n;
1619 struct list_head abort_list;
1621 INIT_LIST_HEAD(&abort_list);
1622 spin_lock(&obd->obd_recovery_task_lock);
1623 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1624 spin_unlock(&obd->obd_recovery_task_lock);
1625 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1626 DEBUG_REQ(D_ERROR, req, "aborted:");
1627 req->rq_status = -ENOTCONN;
1628 if (ptlrpc_error(req)) {
1629 DEBUG_REQ(D_ERROR, req,
1630 "failed abort_lock_reply; skipping");
1632 target_request_copy_put(req);
1636 /* Called from a cleanup function if the device is being cleaned up
1637 forcefully. The exports should all have been disconnected already,
1638 the only thing left to do is
1639 - clear the recovery flags
1641 - free queued requests and replies, but don't send replies
1642 Because the obd_stopping flag is set, no new requests should be received.
1645 void target_cleanup_recovery(struct obd_device *obd)
1647 struct ptlrpc_request *req, *n;
1648 struct list_head clean_list;
1650 INIT_LIST_HEAD(&clean_list);
1651 spin_lock(&obd->obd_dev_lock);
1652 if (!obd->obd_recovering) {
1653 spin_unlock(&obd->obd_dev_lock);
1657 obd->obd_recovering = obd->obd_abort_recovery = 0;
1658 spin_unlock(&obd->obd_dev_lock);
1660 spin_lock(&obd->obd_recovery_task_lock);
1661 target_cancel_recovery_timer(obd);
1662 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1663 spin_unlock(&obd->obd_recovery_task_lock);
1665 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1666 LASSERT(req->rq_reply_state == NULL);
1667 target_exp_dequeue_req_replay(req);
1668 target_request_copy_put(req);
1671 spin_lock(&obd->obd_recovery_task_lock);
1672 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1673 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1674 spin_unlock(&obd->obd_recovery_task_lock);
1676 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1677 LASSERT(req->rq_reply_state == NULL);
1678 target_request_copy_put(req);
1683 EXPORT_SYMBOL(target_cleanup_recovery);
1685 /* obd_recovery_task_lock should be held */
1686 void target_cancel_recovery_timer(struct obd_device *obd)
1688 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1689 del_timer(&obd->obd_recovery_timer);
1692 static void target_start_recovery_timer(struct obd_device *obd)
1694 if (obd->obd_recovery_start != 0)
1697 spin_lock(&obd->obd_dev_lock);
1698 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1699 spin_unlock(&obd->obd_dev_lock);
1703 LASSERT(obd->obd_recovery_timeout != 0);
1705 if (obd->obd_recovery_start != 0) {
1706 spin_unlock(&obd->obd_dev_lock);
1710 mod_timer(&obd->obd_recovery_timer,
1711 jiffies + cfs_time_seconds(obd->obd_recovery_timeout));
1712 obd->obd_recovery_start = ktime_get_real_seconds();
1713 spin_unlock(&obd->obd_dev_lock);
1715 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1717 obd->obd_recovery_timeout / 60,
1718 obd->obd_recovery_timeout % 60,
1719 obd->obd_max_recoverable_clients,
1720 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1721 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1725 * extend recovery window.
1727 * if @extend is true, extend recovery window to have @drt remaining at least;
1728 * otherwise, make sure the recovery timeout value is not less than @drt.
1730 static void extend_recovery_timer(struct obd_device *obd, time64_t drt,
1738 spin_lock(&obd->obd_dev_lock);
1739 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1740 spin_unlock(&obd->obd_dev_lock);
1743 LASSERT(obd->obd_recovery_start != 0);
1745 now = ktime_get_real_seconds();
1746 to = obd->obd_recovery_timeout;
1747 end = obd->obd_recovery_start + to;
1750 if (extend && (drt > left)) {
1752 } else if (!extend && (drt > to)) {
1756 if (to > obd->obd_recovery_time_hard) {
1757 to = obd->obd_recovery_time_hard;
1758 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1759 obd->obd_name, to, extend);
1762 if (obd->obd_recovery_timeout < to) {
1763 obd->obd_recovery_timeout = to;
1764 end = obd->obd_recovery_start + to;
1765 mod_timer(&obd->obd_recovery_timer,
1766 jiffies + cfs_time_seconds(end - now));
1768 spin_unlock(&obd->obd_dev_lock);
1770 CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1771 obd->obd_name, (s64)(end - now));
1774 /* Reset the timer with each new client connection */
1776 * This timer is actually reconnect_timer, which is for making sure
1777 * the total recovery window is at least as big as my reconnect
1778 * attempt timing. So the initial recovery time_out will be set to
1779 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1780 * from client is bigger than this, then the recovery time_out will
1781 * be extended to make sure the client could be reconnected, in the
1782 * process, the timeout from the new client should be ignored.
1786 check_and_start_recovery_timer(struct obd_device *obd,
1787 struct ptlrpc_request *req,
1790 time64_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1791 struct obd_device_target *obt = &obd->u.obt;
1793 if (!new_client && service_time)
1794 /* Teach server about old server's estimates, as first guess
1795 * at how long new requests will take. */
1796 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1799 target_start_recovery_timer(obd);
1801 /* Convert the service time to RPC timeout,
1802 * and reuse service_time to limit stack usage.
1804 service_time = at_est2timeout(service_time);
1806 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1807 service_time < at_extra)
1808 service_time = at_extra;
1810 /* We expect other clients to timeout within service_time, then try
1811 * to reconnect, then try the failover server. The max delay between
1812 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1813 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1815 LASSERT(obt->obt_magic == OBT_MAGIC);
1816 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1817 if (service_time > obd->obd_recovery_timeout && !new_client)
1818 extend_recovery_timer(obd, service_time, false);
1821 /** Health checking routines */
1822 static inline int exp_connect_healthy(struct obd_export *exp)
1824 return (exp->exp_in_recovery);
1827 /** if export done req_replay or has replay in queue */
1828 static inline int exp_req_replay_healthy(struct obd_export *exp)
1830 return (!exp->exp_req_replay_needed ||
1831 atomic_read(&exp->exp_replay_count) > 0);
1835 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1837 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1838 exp_req_replay_healthy(exp);
1841 /** if export done lock_replay or has replay in queue */
1842 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1844 return (!exp->exp_lock_replay_needed ||
1845 atomic_read(&exp->exp_replay_count) > 0);
1848 static inline int exp_vbr_healthy(struct obd_export *exp)
1850 return (!exp->exp_vbr_failed);
1853 static inline int exp_finished(struct obd_export *exp)
1855 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1858 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1860 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1864 static int check_for_next_transno(struct lu_target *lut)
1866 struct ptlrpc_request *req = NULL;
1867 struct obd_device *obd = lut->lut_obd;
1868 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1869 int wake_up = 0, connected, completed, queue_len;
1870 __u64 req_transno = 0;
1871 __u64 update_transno = 0;
1872 __u64 next_transno = 0;
1875 spin_lock(&obd->obd_recovery_task_lock);
1876 if (!list_empty(&obd->obd_req_replay_queue)) {
1877 req = list_entry(obd->obd_req_replay_queue.next,
1878 struct ptlrpc_request, rq_list);
1879 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1883 update_transno = distribute_txn_get_next_transno(tdtd);
1885 connected = atomic_read(&obd->obd_connected_clients);
1886 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1887 queue_len = obd->obd_requests_queued_for_recovery;
1888 next_transno = obd->obd_next_recovery_transno;
1890 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1891 "req_transno: %llu, next_transno: %llu\n",
1892 obd->obd_max_recoverable_clients, connected, completed,
1893 queue_len, req_transno, next_transno);
1895 if (obd->obd_abort_recovery) {
1896 CDEBUG(D_HA, "waking for aborted recovery\n");
1898 } else if (obd->obd_recovery_expired) {
1899 CDEBUG(D_HA, "waking for expired recovery\n");
1901 } else if (tdtd != NULL && req != NULL &&
1902 is_req_replayed_by_update(req)) {
1903 LASSERTF(req_transno < next_transno, "req_transno %llu"
1904 "next_transno%llu\n", req_transno, next_transno);
1905 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1908 } else if (req_transno == next_transno ||
1909 (update_transno != 0 && update_transno <= next_transno)) {
1910 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1912 } else if (queue_len > 0 &&
1913 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1914 /** handle gaps occured due to lost reply or VBR */
1915 LASSERTF(req_transno >= next_transno,
1916 "req_transno: %llu, next_transno: %llu\n",
1917 req_transno, next_transno);
1919 "%s: waking for gap in transno, VBR is %s (skip: "
1920 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1921 ", next_update %lld last_committed: %lld)\n",
1922 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1923 next_transno, queue_len, completed, connected,
1924 req_transno, update_transno, obd->obd_last_committed);
1925 obd->obd_next_recovery_transno = req_transno;
1927 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1928 CDEBUG(D_HA, "waking for completed recovery\n");
1930 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1931 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1932 " by fail_lock, waking up (%lld)\n", next_transno);
1933 obd->obd_next_recovery_transno = req_transno;
1936 spin_unlock(&obd->obd_recovery_task_lock);
1940 static int check_for_next_lock(struct lu_target *lut)
1942 struct obd_device *obd = lut->lut_obd;
1945 spin_lock(&obd->obd_recovery_task_lock);
1946 if (!list_empty(&obd->obd_lock_replay_queue)) {
1947 CDEBUG(D_HA, "waking for next lock\n");
1949 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1950 CDEBUG(D_HA, "waking for completed lock replay\n");
1952 } else if (obd->obd_abort_recovery) {
1953 CDEBUG(D_HA, "waking for aborted recovery\n");
1955 } else if (obd->obd_recovery_expired) {
1956 CDEBUG(D_HA, "waking for expired recovery\n");
1959 spin_unlock(&obd->obd_recovery_task_lock);
1965 * wait for recovery events,
1966 * check its status with help of check_routine
1967 * evict dead clients via health_check
1969 static int target_recovery_overseer(struct lu_target *lut,
1970 int (*check_routine)(struct lu_target *),
1971 int (*health_check)(struct obd_export *))
1973 struct obd_device *obd = lut->lut_obd;
1974 struct target_distribute_txn_data *tdtd;
1978 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
1979 now = ktime_get_seconds();
1980 if (now - last > 600) {
1981 LCONSOLE_INFO("%s: in recovery but waiting for "
1982 "the first client to connect\n",
1987 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
1988 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
1989 __u64 next_update_transno = 0;
1991 /* Only abort the recovery if there are no update recovery
1992 * left in the queue */
1993 spin_lock(&obd->obd_recovery_task_lock);
1994 if (lut->lut_tdtd != NULL) {
1995 next_update_transno =
1996 distribute_txn_get_next_transno(lut->lut_tdtd);
1998 tdtd = lut->lut_tdtd;
1999 /* If next_update_transno == 0, it probably because
2000 * updatelog retrieve threads did not get any records
2001 * yet, let's wait those threads stopped */
2002 if (next_update_transno == 0) {
2003 struct l_wait_info lwi = { 0 };
2005 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2007 &tdtd->tdtd_recovery_threads_count) == 0,
2010 next_update_transno =
2011 distribute_txn_get_next_transno(
2016 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2017 obd->obd_next_recovery_transno = next_update_transno;
2018 spin_unlock(&obd->obd_recovery_task_lock);
2019 /* Disconnect unfinished exports from clients, and
2020 * keep connection from MDT to make sure the update
2021 * recovery will still keep trying until some one
2022 * manually abort the recovery */
2023 class_disconnect_stale_exports(obd,
2024 exp_finished_or_from_mdt);
2025 /* Abort all of replay and replay lock req from
2027 abort_req_replay_queue(obd);
2028 abort_lock_replay_queue(obd);
2029 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2030 ")in the queue.\n", obd->obd_name,
2031 next_update_transno);
2033 obd->obd_abort_recovery = 1;
2034 spin_unlock(&obd->obd_recovery_task_lock);
2035 CWARN("%s recovery is aborted by hard timeout\n",
2040 while (wait_event_timeout(obd->obd_next_transno_waitq,
2042 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2043 /* wait indefinitely for event, but don't trigger watchdog */;
2045 if (obd->obd_abort_recovery) {
2046 CWARN("recovery is aborted, evict exports in recovery\n");
2047 if (lut->lut_tdtd != NULL) {
2048 struct l_wait_info lwi = { 0 };
2050 tdtd = lut->lut_tdtd;
2051 /* Let's wait all of the update log recovery thread
2053 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2054 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2056 /* Then abort the update recovery list */
2057 dtrq_list_destroy(lut->lut_tdtd);
2060 /** evict exports which didn't finish recovery yet */
2061 class_disconnect_stale_exports(obd, exp_finished);
2063 } else if (obd->obd_recovery_expired) {
2064 obd->obd_recovery_expired = 0;
2065 /** If some clients died being recovered, evict them */
2066 LCONSOLE_WARN("%s: recovery is timed out, "
2067 "evict stale exports\n", obd->obd_name);
2068 /** evict cexports with no replay in queue, they are stalled */
2069 class_disconnect_stale_exports(obd, health_check);
2071 /** continue with VBR */
2072 spin_lock(&obd->obd_dev_lock);
2073 obd->obd_version_recov = 1;
2074 spin_unlock(&obd->obd_dev_lock);
2076 * reset timer, recovery will proceed with versions now,
2077 * timeout is set just to handle reconnection delays
2079 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2080 /** Wait for recovery events again, after evicting bad clients */
2086 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2088 struct obd_device *obd = lut->lut_obd;
2089 struct ptlrpc_request *req = NULL;
2091 CDEBUG(D_HA, "Waiting for lock\n");
2092 if (target_recovery_overseer(lut, check_for_next_lock,
2093 exp_lock_replay_healthy))
2094 abort_lock_replay_queue(obd);
2096 spin_lock(&obd->obd_recovery_task_lock);
2097 if (!list_empty(&obd->obd_lock_replay_queue)) {
2098 req = list_entry(obd->obd_lock_replay_queue.next,
2099 struct ptlrpc_request, rq_list);
2100 list_del_init(&req->rq_list);
2101 spin_unlock(&obd->obd_recovery_task_lock);
2103 spin_unlock(&obd->obd_recovery_task_lock);
2104 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2105 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2106 /** evict exports failed VBR */
2107 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2112 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2114 struct ptlrpc_request *req = NULL;
2116 spin_lock(&obd->obd_recovery_task_lock);
2117 if (!list_empty(&obd->obd_final_req_queue)) {
2118 req = list_entry(obd->obd_final_req_queue.next,
2119 struct ptlrpc_request, rq_list);
2120 list_del_init(&req->rq_list);
2121 spin_unlock(&obd->obd_recovery_task_lock);
2122 if (req->rq_export->exp_in_recovery) {
2123 spin_lock(&req->rq_export->exp_lock);
2124 req->rq_export->exp_in_recovery = 0;
2125 spin_unlock(&req->rq_export->exp_lock);
2128 spin_unlock(&obd->obd_recovery_task_lock);
2133 static void handle_recovery_req(struct ptlrpc_thread *thread,
2134 struct ptlrpc_request *req,
2135 svc_handler_t handler)
2140 * export can be evicted during recovery, no need to handle replays for
2141 * it after that, discard such request silently
2143 if (req->rq_export->exp_disconnected)
2146 req->rq_session.lc_thread = thread;
2147 req->rq_svc_thread = thread;
2148 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2150 /* thread context */
2151 lu_context_enter(&thread->t_env->le_ctx);
2153 lu_context_exit(&thread->t_env->le_ctx);
2155 /* don't reset timer for final stage */
2156 if (!exp_finished(req->rq_export)) {
2157 time64_t to = obd_timeout;
2160 * Add request timeout to the recovery time so next request from
2161 * this client may come in recovery time
2164 struct ptlrpc_service_part *svcpt;
2166 svcpt = req->rq_rqbd->rqbd_svcpt;
2167 /* If the server sent early reply for this request,
2168 * the client will recalculate the timeout according to
2169 * current server estimate service time, so we will
2170 * use the maxium timeout here for waiting the client
2171 * sending the next req */
2172 to = max((int)at_est2timeout(
2173 at_get(&svcpt->scp_at_estimate)),
2174 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2175 /* Add 2 net_latency, one for balance rq_deadline
2176 * (see ptl_send_rpc), one for resend the req to server,
2177 * Note: client will pack net_latency in replay req
2178 * (see ptlrpc_replay_req) */
2179 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2181 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2186 /** Checking routines for recovery */
2187 static int check_for_recovery_ready(struct lu_target *lut)
2189 struct obd_device *obd = lut->lut_obd;
2190 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2192 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2193 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2194 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2195 obd->obd_recovery_expired);
2197 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2198 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2199 if (clnts + obd->obd_stale_clients <
2200 obd->obd_max_recoverable_clients)
2204 if (lut->lut_tdtd != NULL) {
2205 if (!lut->lut_tdtd->tdtd_replay_ready &&
2206 !obd->obd_abort_recovery) {
2207 /* Let's extend recovery timer, in case the recovery
2208 * timer expired, and some clients got evicted */
2209 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2211 CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2212 obd->obd_name, obd->obd_recovery_timeout);
2221 REQUEST_RECOVERY = 1,
2222 UPDATE_RECOVERY = 2,
2225 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2229 if (!list_empty(&obd->obd_req_replay_queue)) {
2230 struct ptlrpc_request *req;
2232 req = list_entry(obd->obd_req_replay_queue.next,
2233 struct ptlrpc_request, rq_list);
2234 transno = lustre_msg_get_transno(req->rq_reqmsg);
2240 static __u64 get_next_transno(struct lu_target *lut, int *type)
2242 struct obd_device *obd = lut->lut_obd;
2243 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2245 __u64 update_transno;
2248 transno = get_next_replay_req_transno(obd);
2250 *type = REQUEST_RECOVERY;
2255 update_transno = distribute_txn_get_next_transno(tdtd);
2256 if (transno == 0 || (transno >= update_transno &&
2257 update_transno != 0)) {
2258 transno = update_transno;
2260 *type = UPDATE_RECOVERY;
2267 * drop duplicate replay request
2269 * Because the operation has been replayed by update recovery, the request
2270 * with the same transno will be dropped and also notify the client to send
2271 * next replay request.
2273 * \param[in] env execution environment
2274 * \param[in] obd failover obd device
2275 * \param[in] req request to be dropped
2277 static void drop_duplicate_replay_req(struct lu_env *env,
2278 struct obd_device *obd,
2279 struct ptlrpc_request *req)
2281 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2282 " update records are found.\n",
2283 lustre_msg_get_transno(req->rq_reqmsg),
2284 libcfs_nid2str(req->rq_peer.nid));
2286 /* Right now, only for MDS reint operation update replay and
2287 * normal request replay can have the same transno */
2288 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2289 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2290 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2291 if (likely(req->rq_export))
2292 target_committed_to_req(req);
2293 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2294 target_send_reply(req, req->rq_status, 0);
2296 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2297 libcfs_nid2str(req->rq_peer.nid));
2299 target_exp_dequeue_req_replay(req);
2300 target_request_copy_put(req);
2301 obd->obd_replayed_requests++;
2304 static void replay_request_or_update(struct lu_env *env,
2305 struct lu_target *lut,
2306 struct target_recovery_data *trd,
2307 struct ptlrpc_thread *thread)
2309 struct obd_device *obd = lut->lut_obd;
2310 struct ptlrpc_request *req = NULL;
2315 CDEBUG(D_HA, "Waiting for transno %lld\n",
2316 obd->obd_next_recovery_transno);
2318 /* Replay all of request and update by transno */
2320 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2322 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2324 /** It is needed to extend recovery window above
2325 * recovery_time_soft. Extending is possible only in the
2326 * end of recovery window (see more details in
2327 * handle_recovery_req()).
2329 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2331 if (target_recovery_overseer(lut, check_for_next_transno,
2332 exp_req_replay_healthy_or_from_mdt)) {
2333 abort_req_replay_queue(obd);
2334 abort_lock_replay_queue(obd);
2338 spin_lock(&obd->obd_recovery_task_lock);
2339 transno = get_next_transno(lut, &type);
2340 if (type == REQUEST_RECOVERY && transno != 0) {
2341 /* Drop replay request from client side, if the
2342 * replay has been executed by update with the
2344 req = list_entry(obd->obd_req_replay_queue.next,
2345 struct ptlrpc_request, rq_list);
2347 list_del_init(&req->rq_list);
2348 obd->obd_requests_queued_for_recovery--;
2349 spin_unlock(&obd->obd_recovery_task_lock);
2351 /* Let's check if the request has been redone by
2353 if (is_req_replayed_by_update(req)) {
2354 struct distribute_txn_replay_req *dtrq;
2356 dtrq = distribute_txn_lookup_finish_list(tdtd,
2358 LASSERT(dtrq != NULL);
2359 spin_lock(&tdtd->tdtd_replay_list_lock);
2360 list_del_init(&dtrq->dtrq_list);
2361 spin_unlock(&tdtd->tdtd_replay_list_lock);
2364 drop_duplicate_replay_req(env, obd, req);
2369 LASSERT(trd->trd_processing_task == current_pid());
2370 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2371 lustre_msg_get_transno(req->rq_reqmsg),
2372 libcfs_nid2str(req->rq_peer.nid));
2374 handle_recovery_req(thread, req,
2375 trd->trd_recovery_handler);
2377 * bz18031: increase next_recovery_transno before
2378 * target_request_copy_put() will drop exp_rpc reference
2380 spin_lock(&obd->obd_recovery_task_lock);
2381 obd->obd_next_recovery_transno++;
2382 spin_unlock(&obd->obd_recovery_task_lock);
2383 target_exp_dequeue_req_replay(req);
2384 target_request_copy_put(req);
2385 obd->obd_replayed_requests++;
2386 } else if (type == UPDATE_RECOVERY && transno != 0) {
2387 struct distribute_txn_replay_req *dtrq;
2390 spin_unlock(&obd->obd_recovery_task_lock);
2392 LASSERT(tdtd != NULL);
2393 dtrq = distribute_txn_get_next_req(tdtd);
2394 lu_context_enter(&thread->t_env->le_ctx);
2395 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2396 lu_context_exit(&thread->t_env->le_ctx);
2397 extend_recovery_timer(obd, obd_timeout, true);
2399 if (rc == 0 && dtrq->dtrq_xid != 0) {
2400 CDEBUG(D_HA, "Move x%llu t%llu"
2401 " to finish list\n", dtrq->dtrq_xid,
2402 dtrq->dtrq_master_transno);
2404 /* Add it to the replay finish list */
2405 spin_lock(&tdtd->tdtd_replay_list_lock);
2406 list_add(&dtrq->dtrq_list,
2407 &tdtd->tdtd_replay_finish_list);
2408 spin_unlock(&tdtd->tdtd_replay_list_lock);
2410 spin_lock(&obd->obd_recovery_task_lock);
2411 if (transno == obd->obd_next_recovery_transno)
2412 obd->obd_next_recovery_transno++;
2414 obd->obd_next_recovery_transno)
2415 obd->obd_next_recovery_transno =
2417 spin_unlock(&obd->obd_recovery_task_lock);
2422 spin_unlock(&obd->obd_recovery_task_lock);
2424 LASSERT(list_empty(&obd->obd_req_replay_queue));
2425 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2426 /** evict exports failed VBR */
2427 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2433 static int target_recovery_thread(void *arg)
2435 struct lu_target *lut = arg;
2436 struct obd_device *obd = lut->lut_obd;
2437 struct ptlrpc_request *req;
2438 struct target_recovery_data *trd = &obd->obd_recovery_data;
2439 unsigned long delta;
2441 struct ptlrpc_thread *thread = NULL;
2445 unshare_fs_struct();
2446 OBD_ALLOC_PTR(thread);
2452 OBD_FREE_PTR(thread);
2456 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2458 OBD_FREE_PTR(thread);
2463 thread->t_env = env;
2464 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2465 env->le_ctx.lc_thread = thread;
2466 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2467 thread->t_watchdog = NULL;
2469 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2471 trd->trd_processing_task = current_pid();
2473 spin_lock(&obd->obd_dev_lock);
2474 obd->obd_recovering = 1;
2475 spin_unlock(&obd->obd_dev_lock);
2476 complete(&trd->trd_starting);
2478 /* first of all, we have to know the first transno to replay */
2479 if (target_recovery_overseer(lut, check_for_recovery_ready,
2480 exp_connect_healthy)) {
2481 abort_req_replay_queue(obd);
2482 abort_lock_replay_queue(obd);
2483 if (lut->lut_tdtd != NULL)
2484 dtrq_list_destroy(lut->lut_tdtd);
2487 /* next stage: replay requests or update */
2489 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2490 atomic_read(&obd->obd_req_replay_clients),
2491 obd->obd_next_recovery_transno);
2492 replay_request_or_update(env, lut, trd, thread);
2495 * The second stage: replay locks
2497 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2498 atomic_read(&obd->obd_lock_replay_clients));
2499 while ((req = target_next_replay_lock(lut))) {
2500 LASSERT(trd->trd_processing_task == current_pid());
2501 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2502 libcfs_nid2str(req->rq_peer.nid));
2503 handle_recovery_req(thread, req,
2504 trd->trd_recovery_handler);
2505 target_request_copy_put(req);
2506 obd->obd_replayed_locks++;
2510 * The third stage: reply on final pings, at this moment all clients
2511 * must have request in final queue
2513 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2514 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2515 /** Update server last boot epoch */
2516 tgt_boot_epoch_update(lut);
2517 /* We drop recoverying flag to forward all new requests
2518 * to regular mds_handle() since now */
2519 spin_lock(&obd->obd_dev_lock);
2520 obd->obd_recovering = obd->obd_abort_recovery = 0;
2521 spin_unlock(&obd->obd_dev_lock);
2522 spin_lock(&obd->obd_recovery_task_lock);
2523 target_cancel_recovery_timer(obd);
2524 spin_unlock(&obd->obd_recovery_task_lock);
2525 while ((req = target_next_final_ping(obd))) {
2526 LASSERT(trd->trd_processing_task == current_pid());
2527 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2528 libcfs_nid2str(req->rq_peer.nid));
2529 handle_recovery_req(thread, req,
2530 trd->trd_recovery_handler);
2531 /* Because the waiting client can not send ping to server,
2532 * so we need refresh the last_request_time, to avoid the
2533 * export is being evicted */
2534 ptlrpc_update_export_timer(req->rq_export, 0);
2535 target_request_copy_put(req);
2538 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2539 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2540 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2541 if (delta > OBD_RECOVERY_TIME_SOFT) {
2542 CWARN("too long recovery - read logs\n");
2543 libcfs_debug_dumplog();
2546 target_finish_recovery(lut);
2548 lu_context_fini(&env->le_ctx);
2549 trd->trd_processing_task = 0;
2550 complete(&trd->trd_finishing);
2552 tgt_io_thread_done(thread);
2553 OBD_FREE_PTR(thread);
2558 static int target_start_recovery_thread(struct lu_target *lut,
2559 svc_handler_t handler)
2561 struct obd_device *obd = lut->lut_obd;
2563 struct target_recovery_data *trd = &obd->obd_recovery_data;
2566 memset(trd, 0, sizeof(*trd));
2567 init_completion(&trd->trd_starting);
2568 init_completion(&trd->trd_finishing);
2569 trd->trd_recovery_handler = handler;
2571 rc = server_name2index(obd->obd_name, &index, NULL);
2575 if (!IS_ERR(kthread_run(target_recovery_thread,
2576 lut, "tgt_recover_%d", index))) {
2577 wait_for_completion(&trd->trd_starting);
2578 LASSERT(obd->obd_recovering != 0);
2586 void target_stop_recovery_thread(struct obd_device *obd)
2588 if (obd->obd_recovery_data.trd_processing_task > 0) {
2589 struct target_recovery_data *trd = &obd->obd_recovery_data;
2590 /** recovery can be done but postrecovery is not yet */
2591 spin_lock(&obd->obd_dev_lock);
2592 if (obd->obd_recovering) {
2593 CERROR("%s: Aborting recovery\n", obd->obd_name);
2594 obd->obd_abort_recovery = 1;
2595 wake_up(&obd->obd_next_transno_waitq);
2597 spin_unlock(&obd->obd_dev_lock);
2598 wait_for_completion(&trd->trd_finishing);
2601 EXPORT_SYMBOL(target_stop_recovery_thread);
2603 void target_recovery_fini(struct obd_device *obd)
2605 class_disconnect_exports(obd);
2606 target_stop_recovery_thread(obd);
2607 target_cleanup_recovery(obd);
2609 EXPORT_SYMBOL(target_recovery_fini);
2611 static void target_recovery_expired(cfs_timer_cb_arg_t data)
2613 struct obd_device *obd = cfs_from_timer(obd, data, obd_recovery_timer);
2614 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2615 " after %llus (%d clients connected)\n",
2616 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2617 (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2618 atomic_read(&obd->obd_connected_clients));
2620 obd->obd_recovery_expired = 1;
2621 wake_up(&obd->obd_next_transno_waitq);
2624 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2626 struct obd_device *obd = lut->lut_obd;
2628 if (lut->lut_bottom->dd_rdonly)
2631 if (obd->obd_max_recoverable_clients == 0) {
2632 /** Update server last boot epoch */
2633 tgt_boot_epoch_update(lut);
2637 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2638 "last_transno %llu\n", obd->obd_name,
2639 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2640 LASSERT(obd->obd_stopping == 0);
2641 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2642 obd->obd_recovery_start = 0;
2643 obd->obd_recovery_end = 0;
2645 cfs_timer_setup(&obd->obd_recovery_timer, target_recovery_expired,
2646 (unsigned long)obd, 0);
2647 target_start_recovery_thread(lut, handler);
2649 EXPORT_SYMBOL(target_recovery_init);
2651 static int target_process_req_flags(struct obd_device *obd,
2652 struct ptlrpc_request *req)
2654 struct obd_export *exp = req->rq_export;
2655 LASSERT(exp != NULL);
2656 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2657 /* client declares he's ready to replay locks */
2658 spin_lock(&exp->exp_lock);
2659 if (exp->exp_req_replay_needed) {
2660 exp->exp_req_replay_needed = 0;
2661 spin_unlock(&exp->exp_lock);
2663 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2664 atomic_dec(&obd->obd_req_replay_clients);
2666 spin_unlock(&exp->exp_lock);
2669 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2670 /* client declares he's ready to complete recovery
2671 * so, we put the request on th final queue */
2672 spin_lock(&exp->exp_lock);
2673 if (exp->exp_lock_replay_needed) {
2674 exp->exp_lock_replay_needed = 0;
2675 spin_unlock(&exp->exp_lock);
2677 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2678 atomic_dec(&obd->obd_lock_replay_clients);
2680 spin_unlock(&exp->exp_lock);
2686 int target_queue_recovery_request(struct ptlrpc_request *req,
2687 struct obd_device *obd)
2689 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2690 struct ptlrpc_request *reqiter;
2694 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2695 /* Processing the queue right now, don't re-add. */
2699 target_process_req_flags(obd, req);
2701 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2702 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2703 if (cfs_fail_val == 1) {
2706 wake_up(&cfs_race_waitq);
2708 set_current_state(TASK_INTERRUPTIBLE);
2709 schedule_timeout(cfs_time_seconds(1));
2713 /* client declares he's ready to complete recovery
2714 * so, we put the request on th final queue */
2715 target_request_copy_get(req);
2716 DEBUG_REQ(D_HA, req, "queue final req");
2717 wake_up(&obd->obd_next_transno_waitq);
2718 spin_lock(&obd->obd_recovery_task_lock);
2719 if (obd->obd_recovering) {
2720 struct ptlrpc_request *tmp;
2721 struct ptlrpc_request *duplicate = NULL;
2723 if (likely(!req->rq_export->exp_replay_done)) {
2724 req->rq_export->exp_replay_done = 1;
2725 list_add_tail(&req->rq_list,
2726 &obd->obd_final_req_queue);
2727 spin_unlock(&obd->obd_recovery_task_lock);
2731 /* XXX O(n), but only happens if final ping is
2732 * timed out, probably reorganize the list as
2733 * a hash list later */
2734 list_for_each_entry_safe(reqiter, tmp,
2735 &obd->obd_final_req_queue,
2737 if (reqiter->rq_export == req->rq_export) {
2738 list_del_init(&reqiter->rq_list);
2739 duplicate = reqiter;
2744 list_add_tail(&req->rq_list,
2745 &obd->obd_final_req_queue);
2746 req->rq_export->exp_replay_done = 1;
2747 spin_unlock(&obd->obd_recovery_task_lock);
2749 if (duplicate != NULL) {
2750 DEBUG_REQ(D_HA, duplicate,
2751 "put prev final req\n");
2752 target_request_copy_put(duplicate);
2756 spin_unlock(&obd->obd_recovery_task_lock);
2757 target_request_copy_put(req);
2758 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2761 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2762 /* client declares he's ready to replay locks */
2763 target_request_copy_get(req);
2764 DEBUG_REQ(D_HA, req, "queue lock replay req");
2765 wake_up(&obd->obd_next_transno_waitq);
2766 spin_lock(&obd->obd_recovery_task_lock);
2767 LASSERT(obd->obd_recovering);
2768 /* usually due to recovery abort */
2769 if (!req->rq_export->exp_in_recovery) {
2770 spin_unlock(&obd->obd_recovery_task_lock);
2771 target_request_copy_put(req);
2774 LASSERT(req->rq_export->exp_lock_replay_needed);
2775 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2776 spin_unlock(&obd->obd_recovery_task_lock);
2780 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2781 * (i.e. buflens etc are in my own byte order), but type-dependent
2782 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2785 INIT_LIST_HEAD(&req->rq_list);
2786 DEBUG_REQ(D_HA, req, "not queueing");
2790 /* If we're processing the queue, we want don't want to queue this
2793 * Also, if this request has a transno less than the one we're waiting
2794 * for, we should process it now. It could (and currently always will)
2795 * be an open request for a descriptor that was opened some time ago.
2797 * Also, a resent, replayed request that has already been
2798 * handled will pass through here and be processed immediately.
2800 CDEBUG(D_HA, "Next recovery transno: %llu"
2801 ", current: %llu, replaying\n",
2802 obd->obd_next_recovery_transno, transno);
2804 /* If the request has been replayed by update replay, then sends this
2805 * request to the recovery thread (replay_request_or_update()), where
2806 * it will be handled */
2807 spin_lock(&obd->obd_recovery_task_lock);
2808 if (transno < obd->obd_next_recovery_transno &&
2809 !is_req_replayed_by_update(req)) {
2810 /* Processing the queue right now, don't re-add. */
2811 LASSERT(list_empty(&req->rq_list));
2812 spin_unlock(&obd->obd_recovery_task_lock);
2815 spin_unlock(&obd->obd_recovery_task_lock);
2817 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2820 target_request_copy_get(req);
2821 if (!req->rq_export->exp_in_recovery) {
2822 target_request_copy_put(req);
2825 LASSERT(req->rq_export->exp_req_replay_needed);
2827 if (target_exp_enqueue_req_replay(req)) {
2828 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2829 target_request_copy_put(req);
2834 spin_lock(&obd->obd_recovery_task_lock);
2835 LASSERT(obd->obd_recovering);
2836 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2837 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2838 list_add_tail(&req->rq_list, &reqiter->rq_list);
2843 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2845 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2846 "has been claimed by another client");
2847 spin_unlock(&obd->obd_recovery_task_lock);
2848 target_exp_dequeue_req_replay(req);
2849 target_request_copy_put(req);
2855 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2857 obd->obd_requests_queued_for_recovery++;
2858 spin_unlock(&obd->obd_recovery_task_lock);
2859 wake_up(&obd->obd_next_transno_waitq);
2863 int target_handle_ping(struct ptlrpc_request *req)
2865 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2866 return req_capsule_server_pack(&req->rq_pill);
2869 void target_committed_to_req(struct ptlrpc_request *req)
2871 struct obd_export *exp = req->rq_export;
2873 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2874 lustre_msg_set_last_committed(req->rq_repmsg,
2875 exp->exp_last_committed);
2877 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2878 "%d)", exp->exp_obd->obd_no_transno,
2879 req->rq_repmsg == NULL);
2881 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2882 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2885 #endif /* HAVE_SERVER_SUPPORT */
2888 * Packs current SLV and Limit into \a req.
2890 int target_pack_pool_reply(struct ptlrpc_request *req)
2892 struct obd_device *obd;
2895 /* Check that we still have all structures alive as this may
2896 * be some late RPC at shutdown time. */
2897 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2898 !exp_connect_lru_resize(req->rq_export))) {
2899 lustre_msg_set_slv(req->rq_repmsg, 0);
2900 lustre_msg_set_limit(req->rq_repmsg, 0);
2904 /* OBD is alive here as export is alive, which we checked above. */
2905 obd = req->rq_export->exp_obd;
2907 read_lock(&obd->obd_pool_lock);
2908 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2909 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2910 read_unlock(&obd->obd_pool_lock);
2915 static int target_send_reply_msg(struct ptlrpc_request *req,
2916 int rc, int fail_id)
2918 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2919 DEBUG_REQ(D_ERROR, req, "dropping reply");
2922 /* We can have a null rq_reqmsg in the event of bad signature or
2923 * no context when unwrapping */
2924 if (req->rq_reqmsg &&
2925 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2926 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2930 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2931 req->rq_status = rc;
2932 return ptlrpc_send_error(req, 1);
2934 DEBUG_REQ(D_NET, req, "sending reply");
2937 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2940 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2942 struct ptlrpc_service_part *svcpt;
2944 struct ptlrpc_reply_state *rs;
2945 struct obd_export *exp;
2948 if (req->rq_no_reply) {
2953 svcpt = req->rq_rqbd->rqbd_svcpt;
2954 rs = req->rq_reply_state;
2955 if (rs == NULL || !rs->rs_difficult) {
2957 target_send_reply_msg (req, rc, fail_id);
2962 /* must be an export if locks saved */
2963 LASSERT(req->rq_export != NULL);
2964 /* req/reply consistent */
2965 LASSERT(rs->rs_svcpt == svcpt);
2968 LASSERT(!rs->rs_scheduled);
2969 LASSERT(!rs->rs_scheduled_ever);
2970 LASSERT(!rs->rs_handled);
2971 LASSERT(!rs->rs_on_net);
2972 LASSERT(rs->rs_export == NULL);
2973 LASSERT(list_empty(&rs->rs_obd_list));
2974 LASSERT(list_empty(&rs->rs_exp_list));
2976 exp = class_export_get(req->rq_export);
2978 /* disable reply scheduling while I'm setting up */
2979 rs->rs_scheduled = 1;
2981 rs->rs_xid = req->rq_xid;
2982 rs->rs_transno = req->rq_transno;
2983 rs->rs_export = exp;
2984 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2986 spin_lock(&exp->exp_uncommitted_replies_lock);
2987 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2988 rs->rs_transno, exp->exp_last_committed);
2989 if (rs->rs_transno > exp->exp_last_committed) {
2990 /* not committed already */
2991 list_add_tail(&rs->rs_obd_list,
2992 &exp->exp_uncommitted_replies);
2994 spin_unlock(&exp->exp_uncommitted_replies_lock);
2996 spin_lock(&exp->exp_lock);
2997 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2998 spin_unlock(&exp->exp_lock);
3000 netrc = target_send_reply_msg(req, rc, fail_id);
3002 spin_lock(&svcpt->scp_rep_lock);
3004 atomic_inc(&svcpt->scp_nreps_difficult);
3007 /* error sending: reply is off the net. Also we need +1
3008 * reply ref until ptlrpc_handle_rs() is done
3009 * with the reply state (if the send was successful, there
3010 * would have been +1 ref for the net, which
3011 * reply_out_callback leaves alone) */
3013 ptlrpc_rs_addref(rs);
3016 spin_lock(&rs->rs_lock);
3017 if (rs->rs_transno <= exp->exp_last_committed ||
3018 (!rs->rs_on_net && !rs->rs_no_ack) ||
3019 list_empty(&rs->rs_exp_list) || /* completed already */
3020 list_empty(&rs->rs_obd_list)) {
3021 CDEBUG(D_HA, "Schedule reply immediately\n");
3022 ptlrpc_dispatch_difficult_reply(rs);
3024 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3025 rs->rs_scheduled = 0; /* allow notifier to schedule */
3027 spin_unlock(&rs->rs_lock);
3028 spin_unlock(&svcpt->scp_rep_lock);
3032 enum ldlm_mode lck_compat_array[] = {
3033 [LCK_EX] = LCK_COMPAT_EX,
3034 [LCK_PW] = LCK_COMPAT_PW,
3035 [LCK_PR] = LCK_COMPAT_PR,
3036 [LCK_CW] = LCK_COMPAT_CW,
3037 [LCK_CR] = LCK_COMPAT_CR,
3038 [LCK_NL] = LCK_COMPAT_NL,
3039 [LCK_GROUP] = LCK_COMPAT_GROUP,
3040 [LCK_COS] = LCK_COMPAT_COS,
3044 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3045 * not escape to the user level.
3047 int ldlm_error2errno(enum ldlm_error error)
3053 case ELDLM_LOCK_MATCHED:
3056 case ELDLM_LOCK_CHANGED:
3059 case ELDLM_LOCK_ABORTED:
3062 case ELDLM_LOCK_REPLACED:
3065 case ELDLM_NO_LOCK_DATA:
3068 case ELDLM_NAMESPACE_EXISTS:
3071 case ELDLM_BAD_NAMESPACE:
3075 if (((int)error) < 0) { /* cast to signed type */
3076 result = error; /* as ldlm_error can be unsigned */
3078 CERROR("Invalid DLM result code: %d\n", error);
3084 EXPORT_SYMBOL(ldlm_error2errno);
3087 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3089 enum ldlm_error ldlm_errno2error(int err_no)
3098 error = ELDLM_LOCK_CHANGED;
3101 error = ELDLM_LOCK_ABORTED;
3104 error = ELDLM_LOCK_REPLACED;
3107 error = ELDLM_NO_LOCK_DATA;
3110 error = ELDLM_NAMESPACE_EXISTS;
3113 error = ELDLM_BAD_NAMESPACE;
3121 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3122 void ldlm_dump_export_locks(struct obd_export *exp)
3124 spin_lock(&exp->exp_locks_list_guard);
3125 if (!list_empty(&exp->exp_locks_list)) {
3126 struct ldlm_lock *lock;
3128 CERROR("dumping locks for export %p,"
3129 "ignore if the unmount doesn't hang\n", exp);
3130 list_for_each_entry(lock, &exp->exp_locks_list,
3132 LDLM_ERROR(lock, "lock:");
3134 spin_unlock(&exp->exp_locks_list_guard);
3138 #ifdef HAVE_SERVER_SUPPORT
3139 static int target_bulk_timeout(void *data)
3142 /* We don't fail the connection here, because having the export
3143 * killed makes the (vital) call to commitrw very sad.
3148 static inline const char *bulk2type(struct ptlrpc_request *req)
3150 if (req->rq_bulk_read)
3152 if (req->rq_bulk_write)
3157 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3158 struct l_wait_info *lwi)
3160 struct ptlrpc_request *req = desc->bd_req;
3161 time64_t start = ktime_get_real_seconds();
3167 /* If there is eviction in progress, wait for it to finish. */
3168 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3169 *lwi = LWI_INTR(NULL, NULL);
3170 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3171 !atomic_read(&exp->exp_obd->
3172 obd_evict_inprogress),
3176 /* Check if client was evicted or reconnected already. */
3177 if (exp->exp_failed ||
3178 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3181 if (req->rq_bulk_read)
3182 rc = sptlrpc_svc_wrap_bulk(req, desc);
3184 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3185 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3186 else /* old version, bulk matchbits is rq_xid */
3187 req->rq_mbits = req->rq_xid;
3190 rc = ptlrpc_start_bulk_transfer(desc);
3194 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3195 bulk2type(req), rc);
3199 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3200 ptlrpc_abort_bulk(desc);
3204 /* limit actual bulk transfer to bulk_timeout seconds */
3205 deadline = start + bulk_timeout;
3206 if (deadline > req->rq_deadline)
3207 deadline = req->rq_deadline;
3210 time64_t timeoutl = deadline - ktime_get_real_seconds();
3211 long timeout_jiffies = timeoutl <= 0 ?
3212 1 : cfs_time_seconds(timeoutl);
3213 time64_t rq_deadline;
3215 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3216 cfs_time_seconds(1),
3217 target_bulk_timeout, desc);
3218 rc = l_wait_event(desc->bd_waitq,
3219 !ptlrpc_server_bulk_active(desc) ||
3222 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3224 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3225 /* Wait again if we changed rq_deadline. */
3226 rq_deadline = READ_ONCE(req->rq_deadline);
3227 deadline = start + bulk_timeout;
3228 if (deadline > rq_deadline)
3229 deadline = rq_deadline;
3230 } while (rc == -ETIMEDOUT &&
3231 deadline > ktime_get_real_seconds());
3233 if (rc == -ETIMEDOUT) {
3234 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3235 bulk2type(req), deadline - start,
3236 ktime_get_real_seconds() - deadline);
3237 ptlrpc_abort_bulk(desc);
3238 } else if (exp->exp_failed) {
3239 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3242 ptlrpc_abort_bulk(desc);
3243 } else if (exp->exp_conn_cnt >
3244 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3245 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3247 /* We don't reply anyway. */
3249 ptlrpc_abort_bulk(desc);
3250 } else if (desc->bd_failure) {
3251 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3253 /* XXX should this be a different errno? */
3256 if (req->rq_bulk_write)
3257 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3258 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3259 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3260 bulk2type(req), desc->bd_nob_transferred,
3262 /* XXX should this be a different errno? */
3269 EXPORT_SYMBOL(target_bulk_io);
3271 #endif /* HAVE_SERVER_SUPPORT */