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 <cl_object.h>
43 #include <linux/jiffies.h>
44 #include <linux/kernel.h>
45 #include <linux/kthread.h>
46 #include <libcfs/libcfs.h>
48 #include <obd_class.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
55 * @priority: If non-zero, move the selected connection to the list head.
56 * @create: If zero, only search in existing connections.
58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
59 int priority, int create)
61 struct ptlrpc_connection *ptlrpc_conn;
62 struct obd_import_conn *imp_conn = NULL, *item;
63 lnet_nid_t nid4refnet = LNET_NID_ANY;
68 if (!create && !priority) {
69 CDEBUG(D_HA, "Nothing to do\n");
73 if (imp->imp_connection &&
74 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
75 /* nid4refnet is used to restrict network connections */
76 nid4refnet = imp->imp_connection->c_self;
77 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
79 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
84 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
86 GOTO(out_put, rc = -ENOMEM);
89 spin_lock(&imp->imp_lock);
90 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
91 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
93 list_del(&item->oic_item);
94 list_add(&item->oic_item,
96 item->oic_last_attempt = 0;
98 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
99 imp, imp->imp_obd->obd_name, uuid->uuid,
100 (priority ? ", moved to head" : ""));
101 spin_unlock(&imp->imp_lock);
102 GOTO(out_free, rc = 0);
105 /* No existing import connection found for \a uuid. */
107 imp_conn->oic_conn = ptlrpc_conn;
108 imp_conn->oic_uuid = *uuid;
109 imp_conn->oic_last_attempt = 0;
111 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
113 list_add_tail(&imp_conn->oic_item,
114 &imp->imp_conn_list);
115 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
116 imp, imp->imp_obd->obd_name, uuid->uuid,
117 (priority ? "head" : "tail"));
119 spin_unlock(&imp->imp_lock);
120 GOTO(out_free, rc = -ENOENT);
123 spin_unlock(&imp->imp_lock);
127 OBD_FREE(imp_conn, sizeof(*imp_conn));
129 ptlrpc_connection_put(ptlrpc_conn);
133 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
135 return import_set_conn(imp, uuid, 1, 0);
138 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
141 return import_set_conn(imp, uuid, priority, 1);
143 EXPORT_SYMBOL(client_import_add_conn);
145 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
147 struct obd_import_conn *imp_conn;
148 struct obd_export *dlmexp;
153 spin_lock(&imp->imp_lock);
154 if (list_empty(&imp->imp_conn_list)) {
155 LASSERT(!imp->imp_connection);
159 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
160 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
162 LASSERT(imp_conn->oic_conn);
164 if (imp_conn == imp->imp_conn_current) {
165 LASSERT(imp_conn->oic_conn == imp->imp_connection);
167 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
168 imp->imp_state != LUSTRE_IMP_DISCON) {
169 CERROR("can't remove current connection\n");
170 GOTO(out, rc = -EBUSY);
173 ptlrpc_connection_put(imp->imp_connection);
174 imp->imp_connection = NULL;
176 dlmexp = class_conn2export(&imp->imp_dlm_handle);
177 if (dlmexp && dlmexp->exp_connection) {
178 LASSERT(dlmexp->exp_connection ==
180 ptlrpc_connection_put(dlmexp->exp_connection);
181 dlmexp->exp_connection = NULL;
185 class_export_put(dlmexp);
188 list_del(&imp_conn->oic_item);
189 ptlrpc_connection_put(imp_conn->oic_conn);
190 OBD_FREE(imp_conn, sizeof(*imp_conn));
191 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
192 imp, imp->imp_obd->obd_name, uuid->uuid);
197 spin_unlock(&imp->imp_lock);
199 CERROR("connection %s not found\n", uuid->uuid);
202 EXPORT_SYMBOL(client_import_del_conn);
205 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
206 * to find a conn uuid of \a imp which can reach \a peer.
208 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
209 struct obd_uuid *uuid)
211 struct obd_import_conn *conn;
216 spin_lock(&imp->imp_lock);
217 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
218 /* Check if conn UUID does have this peer NID. */
219 if (class_check_uuid(&conn->oic_uuid, peer)) {
220 *uuid = conn->oic_uuid;
225 spin_unlock(&imp->imp_lock);
228 EXPORT_SYMBOL(client_import_find_conn);
230 void client_destroy_import(struct obd_import *imp)
233 * Drop security policy instance after all RPCs have finished/aborted
234 * to let all busy contexts be released.
236 class_import_get(imp);
237 class_destroy_import(imp);
238 sptlrpc_import_sec_put(imp);
239 class_import_put(imp);
241 EXPORT_SYMBOL(client_destroy_import);
244 * Check whether or not the OSC is on MDT.
247 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
249 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
252 static int osc_on_mdt(char *obdname)
256 ptr = strrchr(obdname, '-');
260 if (strncmp(ptr + 1, "MDT", 3) == 0)
267 * Configure an RPC client OBD device.
272 * 3 - inactive-on-startup
273 * 4 - restrictive net
275 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
277 struct client_obd *cli = &obddev->u.cli;
278 struct obd_import *imp;
279 struct obd_uuid server_uuid;
280 int rq_portal, rp_portal, connect_op;
281 const char *name = obddev->obd_type->typ_name;
282 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
283 char *cli_name = lustre_cfg_buf(lcfg, 0);
284 struct ptlrpc_connection fake_conn = { .c_self = 0,
285 .c_remote_uuid.uuid[0] = 0 };
291 * In a more perfect world, we would hang a ptlrpc_client off of
292 * obd_type and just use the values from there.
294 if (!strcmp(name, LUSTRE_OSC_NAME)) {
295 rq_portal = OST_REQUEST_PORTAL;
296 rp_portal = OSC_REPLY_PORTAL;
297 connect_op = OST_CONNECT;
298 cli->cl_sp_me = LUSTRE_SP_CLI;
299 cli->cl_sp_to = LUSTRE_SP_OST;
300 ns_type = LDLM_NS_TYPE_OSC;
301 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
302 !strcmp(name, LUSTRE_LWP_NAME)) {
303 rq_portal = MDS_REQUEST_PORTAL;
304 rp_portal = MDC_REPLY_PORTAL;
305 connect_op = MDS_CONNECT;
306 if (is_lwp_on_ost(cli_name))
307 cli->cl_sp_me = LUSTRE_SP_OST;
308 else if (is_lwp_on_mdt(cli_name))
309 cli->cl_sp_me = LUSTRE_SP_MDT;
311 cli->cl_sp_me = LUSTRE_SP_CLI;
312 cli->cl_sp_to = LUSTRE_SP_MDT;
313 ns_type = LDLM_NS_TYPE_MDC;
314 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
315 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
316 /* OSP_on_MDT for other MDTs */
317 connect_op = MDS_CONNECT;
318 cli->cl_sp_to = LUSTRE_SP_MDT;
319 ns_type = LDLM_NS_TYPE_MDC;
320 rq_portal = OUT_PORTAL;
322 /* OSP on MDT for OST */
323 connect_op = OST_CONNECT;
324 cli->cl_sp_to = LUSTRE_SP_OST;
325 ns_type = LDLM_NS_TYPE_OSC;
326 rq_portal = OST_REQUEST_PORTAL;
328 rp_portal = OSC_REPLY_PORTAL;
329 cli->cl_sp_me = LUSTRE_SP_MDT;
330 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
331 rq_portal = MGS_REQUEST_PORTAL;
332 rp_portal = MGC_REPLY_PORTAL;
333 connect_op = MGS_CONNECT;
334 cli->cl_sp_me = LUSTRE_SP_MGC;
335 cli->cl_sp_to = LUSTRE_SP_MGS;
336 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
337 ns_type = LDLM_NS_TYPE_MGC;
339 CERROR("unknown client OBD type \"%s\", can't setup\n",
344 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
345 CERROR("requires a TARGET UUID\n");
349 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
350 CERROR("client UUID must be less than 38 characters\n");
354 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
355 CERROR("setup requires a SERVER UUID\n");
359 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
360 CERROR("target UUID must be less than 38 characters\n");
364 init_rwsem(&cli->cl_sem);
365 mutex_init(&cli->cl_mgc_mutex);
367 init_rwsem(&cli->cl_seq_rwsem);
368 cli->cl_conn_count = 0;
369 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
370 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
371 sizeof(server_uuid)));
373 cli->cl_dirty_pages = 0;
374 cli->cl_dirty_max_pages = 0;
375 cli->cl_avail_grant = 0;
376 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
378 * cl_dirty_max_pages may be changed at connect time in
379 * ptlrpc_connect_interpret().
381 client_adjust_max_dirty(cli);
382 INIT_LIST_HEAD(&cli->cl_cache_waiters);
383 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
384 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
385 INIT_LIST_HEAD(&cli->cl_loi_write_list);
386 INIT_LIST_HEAD(&cli->cl_loi_read_list);
387 spin_lock_init(&cli->cl_loi_list_lock);
388 atomic_set(&cli->cl_pending_w_pages, 0);
389 atomic_set(&cli->cl_pending_r_pages, 0);
390 cli->cl_r_in_flight = 0;
391 cli->cl_w_in_flight = 0;
393 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
394 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
395 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
396 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
397 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
398 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
401 INIT_LIST_HEAD(&cli->cl_lru_osc);
402 atomic_set(&cli->cl_lru_shrinkers, 0);
403 atomic_long_set(&cli->cl_lru_busy, 0);
404 atomic_long_set(&cli->cl_lru_in_list, 0);
405 INIT_LIST_HEAD(&cli->cl_lru_list);
406 spin_lock_init(&cli->cl_lru_list_lock);
407 atomic_long_set(&cli->cl_unstable_count, 0);
408 INIT_LIST_HEAD(&cli->cl_shrink_list);
409 INIT_LIST_HEAD(&cli->cl_grant_chain);
411 INIT_LIST_HEAD(&cli->cl_flight_waiters);
412 cli->cl_rpcs_in_flight = 0;
414 init_waitqueue_head(&cli->cl_destroy_waitq);
415 atomic_set(&cli->cl_destroy_in_flight, 0);
417 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
418 #ifdef ENABLE_CHECKSUM
419 /* Turn on checksumming by default. */
420 cli->cl_checksum = 1;
422 * The supported checksum types will be worked out at connect time
423 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
426 cli->cl_cksum_type = cli->cl_supp_cksum_types;
428 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
431 * Set it to possible maximum size. It may be reduced by ocd_brw_size
432 * from OFD after connecting.
434 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
436 cli->cl_max_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
439 * set cl_chunkbits default value to PAGE_SHIFT,
440 * it will be updated at OSC connection time.
442 cli->cl_chunkbits = PAGE_SHIFT;
444 if (!strcmp(name, LUSTRE_MDC_NAME)) {
445 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
446 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
447 cli->cl_max_rpcs_in_flight = 2;
448 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
449 cli->cl_max_rpcs_in_flight = 3;
450 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
451 cli->cl_max_rpcs_in_flight = 4;
453 if (osc_on_mdt(obddev->obd_name))
454 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
456 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
459 spin_lock_init(&cli->cl_mod_rpcs_lock);
460 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
461 cli->cl_max_mod_rpcs_in_flight = 0;
462 cli->cl_mod_rpcs_in_flight = 0;
463 cli->cl_close_rpcs_in_flight = 0;
464 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
465 cli->cl_mod_tag_bitmap = NULL;
467 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
469 if (connect_op == MDS_CONNECT) {
470 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
471 OBD_ALLOC(cli->cl_mod_tag_bitmap,
472 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
473 if (cli->cl_mod_tag_bitmap == NULL)
474 GOTO(err, rc = -ENOMEM);
479 CERROR("ldlm_get_ref failed: %d\n", rc);
483 ptlrpc_init_client(rq_portal, rp_portal, name,
484 &obddev->obd_ldlm_client);
486 imp = class_new_import(obddev);
488 GOTO(err_ldlm, rc = -ENOENT);
489 imp->imp_client = &obddev->obd_ldlm_client;
490 imp->imp_connect_op = connect_op;
491 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
492 LUSTRE_CFG_BUFLEN(lcfg, 1));
493 class_import_put(imp);
495 if (lustre_cfg_buf(lcfg, 4)) {
496 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
498 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
500 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
501 obddev->obd_name, lustre_cfg_string(lcfg, 4),
503 GOTO(err_import, rc);
505 fake_conn.c_self = LNET_MKNID(refnet, 0);
506 imp->imp_connection = &fake_conn;
509 rc = client_import_add_conn(imp, &server_uuid, 1);
511 CERROR("can't add initial connection\n");
512 GOTO(err_import, rc);
514 imp->imp_connection = NULL;
516 cli->cl_import = imp;
517 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
518 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
520 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
521 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
522 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
523 name, obddev->obd_name,
524 cli->cl_target_uuid.uuid);
525 spin_lock(&imp->imp_lock);
526 imp->imp_deactive = 1;
527 spin_unlock(&imp->imp_lock);
531 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
532 LDLM_NAMESPACE_CLIENT,
533 LDLM_NAMESPACE_GREEDY,
535 if (obddev->obd_namespace == NULL) {
536 CERROR("Unable to create client namespace - %s\n",
538 GOTO(err_import, rc = -ENOMEM);
544 class_destroy_import(imp);
548 if (cli->cl_mod_tag_bitmap != NULL)
549 OBD_FREE(cli->cl_mod_tag_bitmap,
550 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
551 cli->cl_mod_tag_bitmap = NULL;
555 EXPORT_SYMBOL(client_obd_setup);
557 int client_obd_cleanup(struct obd_device *obddev)
559 struct client_obd *cli = &obddev->u.cli;
563 ldlm_namespace_free_post(obddev->obd_namespace);
564 obddev->obd_namespace = NULL;
566 obd_cleanup_client_import(obddev);
567 LASSERT(obddev->u.cli.cl_import == NULL);
571 if (cli->cl_mod_tag_bitmap != NULL)
572 OBD_FREE(cli->cl_mod_tag_bitmap,
573 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
574 cli->cl_mod_tag_bitmap = NULL;
578 EXPORT_SYMBOL(client_obd_cleanup);
580 /* ->o_connect() method for client side (OSC and MDC and MGC) */
581 int client_connect_import(const struct lu_env *env,
582 struct obd_export **exp,
583 struct obd_device *obd, struct obd_uuid *cluuid,
584 struct obd_connect_data *data, void *localdata)
586 struct client_obd *cli = &obd->u.cli;
587 struct obd_import *imp = cli->cl_import;
588 struct obd_connect_data *ocd;
589 struct lustre_handle conn = { 0 };
595 down_write(&cli->cl_sem);
596 if (cli->cl_conn_count > 0)
597 GOTO(out_sem, rc = -EALREADY);
599 rc = class_connect(&conn, obd, cluuid);
603 cli->cl_conn_count++;
604 *exp = class_conn2export(&conn);
606 LASSERT(obd->obd_namespace);
608 imp->imp_dlm_handle = conn;
609 rc = ptlrpc_init_import(imp);
613 ocd = &imp->imp_connect_data;
616 imp->imp_connect_flags_orig = data->ocd_connect_flags;
617 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
620 rc = ptlrpc_connect_import(imp);
622 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
625 LASSERT(*exp != NULL && (*exp)->exp_connection);
628 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
629 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
630 data->ocd_connect_flags, ocd->ocd_connect_flags);
631 data->ocd_connect_flags = ocd->ocd_connect_flags;
632 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
635 ptlrpc_pinger_add_import(imp);
641 cli->cl_conn_count--;
642 class_disconnect(*exp);
646 up_write(&cli->cl_sem);
648 if (!rc && localdata) {
649 LASSERT(cli->cl_cache == NULL); /* only once */
650 cli->cl_cache = (struct cl_client_cache *)localdata;
651 cl_cache_incref(cli->cl_cache);
652 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
654 /* add this osc into entity list */
655 LASSERT(list_empty(&cli->cl_lru_osc));
656 spin_lock(&cli->cl_cache->ccc_lru_lock);
657 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
658 spin_unlock(&cli->cl_cache->ccc_lru_lock);
663 EXPORT_SYMBOL(client_connect_import);
665 int client_disconnect_export(struct obd_export *exp)
667 struct obd_device *obd = class_exp2obd(exp);
668 struct client_obd *cli;
669 struct obd_import *imp;
675 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
676 exp, exp ? exp->exp_handle.h_cookie : -1);
681 imp = cli->cl_import;
683 down_write(&cli->cl_sem);
684 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
687 if (cli->cl_conn_count == 0) {
688 CERROR("disconnecting disconnected device (%s)\n",
690 GOTO(out_disconnect, rc = -EINVAL);
693 cli->cl_conn_count--;
694 if (cli->cl_conn_count != 0)
695 GOTO(out_disconnect, rc = 0);
698 * Mark import deactivated now, so we don't try to reconnect if any
699 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
700 * fully deactivate the import, or that would drop all requests.
702 spin_lock(&imp->imp_lock);
703 imp->imp_deactive = 1;
704 spin_unlock(&imp->imp_lock);
707 * Some non-replayable imports (MDS's OSCs) are pinged, so just
708 * delete it regardless. (It's safe to delete an import that was
711 (void)ptlrpc_pinger_del_import(imp);
713 if (obd->obd_namespace != NULL) {
714 /* obd_force == local only */
715 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
716 obd->obd_force ? LCF_LOCAL : 0, NULL);
717 ldlm_namespace_free_prior(obd->obd_namespace, imp,
722 * There's no need to hold sem while disconnecting an import,
723 * and it may actually cause deadlock in GSS.
725 up_write(&cli->cl_sem);
726 rc = ptlrpc_disconnect_import(imp, 0);
727 down_write(&cli->cl_sem);
729 ptlrpc_invalidate_import(imp);
735 * Use server style - class_disconnect should be always called for
738 err = class_disconnect(exp);
742 up_write(&cli->cl_sem);
746 EXPORT_SYMBOL(client_disconnect_export);
748 #ifdef HAVE_SERVER_SUPPORT
749 int server_disconnect_export(struct obd_export *exp)
755 /* Disconnect early so that clients can't keep using export. */
756 rc = class_disconnect(exp);
757 /* Close import to avoid sending any requests. */
758 if (exp->exp_imp_reverse)
759 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
761 ldlm_bl_thread_wakeup();
763 /* complete all outstanding replies */
764 spin_lock(&exp->exp_lock);
765 while (!list_empty(&exp->exp_outstanding_replies)) {
766 struct ptlrpc_reply_state *rs =
767 list_entry(exp->exp_outstanding_replies.next,
768 struct ptlrpc_reply_state, rs_exp_list);
769 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
771 spin_lock(&svcpt->scp_rep_lock);
773 list_del_init(&rs->rs_exp_list);
775 spin_lock(&rs->rs_lock);
776 /* clear rs_convert_lock to make sure rs is handled and put */
777 rs->rs_convert_lock = 0;
778 ptlrpc_schedule_difficult_reply(rs);
779 spin_unlock(&rs->rs_lock);
781 spin_unlock(&svcpt->scp_rep_lock);
783 spin_unlock(&exp->exp_lock);
787 EXPORT_SYMBOL(server_disconnect_export);
790 * --------------------------------------------------------------------------
791 * from old lib/target.c
792 * --------------------------------------------------------------------------
795 static int target_handle_reconnect(struct lustre_handle *conn,
796 struct obd_export *exp,
797 struct obd_uuid *cluuid)
799 struct obd_device *target;
800 struct lustre_handle *hdl;
806 hdl = &exp->exp_imp_reverse->imp_remote_handle;
807 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
808 conn->cookie = exp->exp_handle.h_cookie;
810 "connect export for UUID '%s' at %p, cookie %#llx\n",
811 cluuid->uuid, exp, conn->cookie);
815 target = exp->exp_obd;
817 /* Might be a re-connect after a partition. */
818 if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
819 LCONSOLE_WARN("%s: already connected client %s (at %s) with handle %#llx. Rejecting client with the same UUID trying to reconnect with handle %#llx\n",
821 obd_uuid2str(&exp->exp_client_uuid),
822 obd_export_nid2str(exp),
823 hdl->cookie, conn->cookie);
824 memset(conn, 0, sizeof(*conn));
826 * target_handle_connect() treats EALREADY and
827 * -EALREADY differently. -EALREADY is an error
828 * (same UUID, different handle).
833 if (!target->obd_recovering) {
834 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
835 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
836 obd_export_nid2str(exp));
837 GOTO(out_already, rc);
840 remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
841 timeout = ktime_divns(remaining, NSEC_PER_SEC);
843 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
845 obd_uuid2str(&exp->exp_client_uuid),
846 obd_export_nid2str(exp),
847 target->obd_max_recoverable_clients,
848 timeout / 60, timeout % 60);
850 struct target_distribute_txn_data *tdtd;
855 tdtd = class_exp2tgt(exp)->lut_tdtd;
856 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
857 buf = tdtd->tdtd_show_update_logs_retrievers(
858 tdtd->tdtd_show_retrievers_cbdata,
862 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting, waiting for %d MDTs (%s) in recovery for %lld:%.02lld. Please wait until all MDTs recovered or you may force MDT evicition via 'lctl --device %s abort_recovery.\n",
864 obd_uuid2str(&exp->exp_client_uuid),
865 obd_export_nid2str(exp), count,
866 buf ? buf : "unknown (not enough RAM)",
867 (abs(timeout) + target->obd_recovery_timeout) / 60,
868 (abs(timeout) + target->obd_recovery_timeout) % 60,
871 LCONSOLE_WARN("%s: Recovery already passed deadline %lld:%.02lld. If you do not want to wait more, you may force taget eviction via 'lctl --device %s abort_recovery.\n",
872 target->obd_name, abs(timeout) / 60,
873 abs(timeout) % 60, target->obd_name);
880 conn->cookie = exp->exp_handle.h_cookie;
882 * target_handle_connect() treats EALREADY and
883 * -EALREADY differently. EALREADY means we are
884 * doing a valid reconnect from the same client.
890 check_and_start_recovery_timer(struct obd_device *obd,
891 struct ptlrpc_request *req, int new_client);
894 * update flags for import during reconnect process
896 static int rev_import_flags_update(struct obd_import *revimp,
897 struct ptlrpc_request *req)
900 struct obd_connect_data *data;
902 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
904 if (data->ocd_connect_flags & OBD_CONNECT_AT)
905 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
907 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
909 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
911 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
913 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
914 revimp->imp_client->cli_name,
915 libcfs_id2str(req->rq_peer), rc);
923 * Allocate a new reverse import for an export.
925 * \retval -errno in case error hit
926 * \retval 0 if reverse import correctly init
928 int rev_import_init(struct obd_export *export)
930 struct obd_device *obd = export->exp_obd;
931 struct obd_import *revimp;
933 LASSERT(export->exp_imp_reverse == NULL);
935 revimp = class_new_import(obd);
939 revimp->imp_remote_handle.cookie = 0ULL;
940 revimp->imp_client = &obd->obd_ldlm_client;
941 revimp->imp_dlm_fake = 1;
943 /* it is safe to connect import in new state as no sends possible */
944 spin_lock(&export->exp_lock);
945 export->exp_imp_reverse = revimp;
946 spin_unlock(&export->exp_lock);
947 class_import_put(revimp);
951 EXPORT_SYMBOL(rev_import_init);
954 * Handle reconnect for an export.
956 * \param exp export to handle reconnect process
957 * \param req client reconnect request
959 * \retval -rc in case securitfy flavor can't be changed
960 * \retval 0 in case none problems
962 static int rev_import_reconnect(struct obd_export *exp,
963 struct ptlrpc_request *req)
965 struct obd_import *revimp = exp->exp_imp_reverse;
966 struct lustre_handle *lh;
969 /* avoid sending a request until import flags are changed */
970 ptlrpc_import_enter_resend(revimp);
972 if (revimp->imp_connection != NULL)
973 ptlrpc_connection_put(revimp->imp_connection);
976 * client from recovery don't have a handle so we need to take from
977 * request. it may produce situation when wrong client connected
978 * to recovery as we trust a client uuid
980 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
981 revimp->imp_remote_handle = *lh;
984 * unknown versions will be caught in
985 * ptlrpc_handle_server_req_in->lustre_unpack_msg()
987 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
989 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
991 rc = rev_import_flags_update(revimp, req);
994 * it is safe to still be in RECOVERY phase as we are not able
995 * to setup correct security flavor so requests are not able to
996 * be delivered correctly
1001 /* resend all rpc's via new connection */
1002 return ptlrpc_import_recovery_state_machine(revimp);
1005 int target_handle_connect(struct ptlrpc_request *req)
1007 struct obd_device *target = NULL;
1008 struct obd_export *export = NULL;
1010 * connect handle - filled from target_handle_reconnect in
1013 struct lustre_handle conn;
1014 struct lustre_handle *tmp;
1015 struct obd_uuid cluuid;
1020 bool mds_conn = false, lw_client = false, initial_conn = false;
1021 bool mds_mds_conn = false;
1022 bool new_mds_mds_conn = false;
1023 struct obd_connect_data *data, *tmpdata;
1025 lnet_nid_t *client_nid = NULL;
1029 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1031 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1033 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1034 GOTO(out, rc = -EINVAL);
1037 target = class_dev_by_str(str);
1039 deuuidify(str, NULL, &target_start, &target_len);
1040 LCONSOLE_ERROR_MSG(0x137,
1041 "%s: not available for connect from %s (no target). If you are running an HA pair check that the target is mounted on the other server.\n",
1042 str, libcfs_nid2str(req->rq_peer.nid));
1043 GOTO(out, rc = -ENODEV);
1046 spin_lock(&target->obd_dev_lock);
1048 target->obd_conn_inprogress++;
1050 if (target->obd_stopping || !target->obd_set_up) {
1051 spin_unlock(&target->obd_dev_lock);
1053 deuuidify(str, NULL, &target_start, &target_len);
1054 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1055 target_len, target_start,
1056 libcfs_nid2str(req->rq_peer.nid),
1057 (target->obd_stopping ?
1058 "stopping" : "not set up"));
1059 GOTO(out, rc = -ENODEV);
1062 if (target->obd_no_conn) {
1063 spin_unlock(&target->obd_dev_lock);
1066 "%s: Temporarily refusing client connection from %s\n",
1067 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1068 GOTO(out, rc = -EAGAIN);
1071 spin_unlock(&target->obd_dev_lock);
1073 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1075 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1076 GOTO(out, rc = -EINVAL);
1079 obd_str2uuid(&cluuid, str);
1081 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1083 GOTO(out, rc = -EPROTO);
1087 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1089 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1091 GOTO(out, rc = -EPROTO);
1093 rc = req_capsule_server_pack(&req->rq_pill);
1097 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1099 * Don't allow clients to connect that are using old 1.8 format
1100 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1101 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1102 * FULL20 flag should be set on all connections since 2.0, but no
1103 * longer affects behaviour.
1105 * Later this check will be disabled and the flag can be retired
1106 * completely once interop with 3.0 is no longer needed.
1108 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1109 GOTO(out, rc = -EPROTO);
1112 * Don't allow liblustre clients to connect.
1113 * - testing was disabled in v2_2_50_0-61-g6a75d65
1114 * - building was disabled in v2_5_58_0-28-g7277179
1115 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1116 * - clients were refused connect for version difference > 0.0.1.32
1118 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1119 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1120 GOTO(out, rc = -EPROTO);
1125 * Note: lw_client is needed in MDS-MDS failover during update log
1126 * processing, so we needs to allow lw_client to be connected at
1127 * anytime, instead of only the initial connection
1129 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1131 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1132 initial_conn = true;
1133 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1134 mds_mds_conn = (data->ocd_connect_flags &
1135 OBD_CONNECT_MDS_MDS) != 0;
1138 * OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1139 * for Imperative Recovery connection from MGC to MGS.
1141 * Via check OBD_CONNECT_FID, we can distinguish whether
1142 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1143 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1146 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1147 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1148 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1149 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1150 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1151 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1154 * We do not support the MDT-MDT interoperations with
1155 * different version MDT because of protocol changes.
1157 if (unlikely(major != LUSTRE_MAJOR ||
1158 minor != LUSTRE_MINOR ||
1159 abs(patch - LUSTRE_PATCH) > 3)) {
1160 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1161 target->obd_name, LUSTRE_MAJOR,
1162 LUSTRE_MINOR, LUSTRE_PATCH,
1163 LUSTRE_FIX, major, minor, patch,
1164 OBD_OCD_VERSION_FIX(data->ocd_version),
1165 libcfs_nid2str(req->rq_peer.nid),
1167 GOTO(out, rc = -EPROTO);
1172 /* lctl gets a backstage, all-access pass. */
1173 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1174 goto dont_check_exports;
1176 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1180 /* We've found an export in the hash. */
1182 spin_lock(&export->exp_lock);
1184 if (export->exp_connecting) { /* b=9635, et. al. */
1185 spin_unlock(&export->exp_lock);
1186 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1187 export->exp_obd->obd_name, export,
1188 libcfs_nid2str(req->rq_peer.nid));
1189 class_export_put(export);
1192 } else if ((mds_conn || (lw_client && initial_conn) ||
1193 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1194 export->exp_connection != NULL) {
1195 spin_unlock(&export->exp_lock);
1196 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1197 /* MDS or LWP reconnected after failover. */
1198 LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1200 mds_conn ? "MDS" : "LWP",
1201 libcfs_nid2str(req->rq_peer.nid),
1202 libcfs_nid2str(export->exp_connection->c_peer.nid));
1204 /* New MDS connection from the same NID. */
1205 LCONSOLE_WARN("%s: Received new %s connection from %s, removing former export from same NID\n",
1207 mds_conn ? "MDS" : "LWP",
1208 libcfs_nid2str(req->rq_peer.nid));
1211 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1212 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1214 * Because exports between MDTs will always be
1215 * kept, let's do not fail such export if they
1216 * come from the same NID, otherwise it might
1217 * cause eviction between MDTs, which might
1218 * cause namespace inconsistency
1220 spin_lock(&export->exp_lock);
1221 export->exp_connecting = 1;
1222 export->exp_conn_cnt = 0;
1223 spin_unlock(&export->exp_lock);
1224 conn.cookie = export->exp_handle.h_cookie;
1227 class_fail_export(export);
1228 class_export_put(export);
1232 } else if (export->exp_connection != NULL && initial_conn &&
1233 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1234 spin_unlock(&export->exp_lock);
1235 /* In MDS failover we have static UUID but NID can change. */
1236 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1237 target->obd_name, cluuid.uuid,
1238 libcfs_nid2str(req->rq_peer.nid),
1240 export->exp_connection->c_peer.nid));
1242 class_export_put(export);
1245 export->exp_connecting = 1;
1246 spin_unlock(&export->exp_lock);
1247 LASSERT(export->exp_obd == target);
1249 rc = target_handle_reconnect(&conn, export, &cluuid);
1252 /* If we found an export, we already unlocked. */
1255 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1256 } else if (req->rq_export == NULL &&
1257 atomic_read(&export->exp_rpc_count) > 0) {
1258 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1259 target->obd_name, cluuid.uuid,
1260 libcfs_nid2str(req->rq_peer.nid),
1261 atomic_read(&export->exp_refcount));
1262 GOTO(out, rc = -EBUSY);
1263 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1265 if (!strstr(cluuid.uuid, "mdt"))
1266 LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client",
1267 target->obd_name, cluuid.uuid,
1268 libcfs_nid2str(req->rq_peer.nid));
1269 GOTO(out, rc = -EALREADY);
1271 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1277 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1278 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1279 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1280 export, ktime_get_real_seconds(),
1281 export ? export->exp_last_request_time : 0);
1284 * If this is the first time a client connects, reset the recovery
1285 * timer. Discard lightweight connections which might be local.
1287 if (!lw_client && rc == 0 && target->obd_recovering)
1288 check_and_start_recovery_timer(target, req, export == NULL);
1291 * We want to handle EALREADY but *not* -EALREADY from
1292 * target_handle_reconnect(), return reconnection state in a flag.
1294 if (rc == EALREADY) {
1295 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1301 /* Tell the client if we support replayable requests. */
1302 if (target->obd_replayable)
1303 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1304 client_nid = &req->rq_peer.nid;
1306 if (export == NULL) {
1307 /* allow lightweight connections during recovery */
1309 * allow "new" MDT to be connected during recovery, since we
1310 * need retrieve recovery update records from it
1312 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1313 struct hrtimer *timer = &target->obd_recovery_timer;
1322 connected = atomic_read(&target->obd_connected_clients);
1323 in_progress = atomic_read(&target->obd_lock_replay_clients);
1324 known = target->obd_max_recoverable_clients;
1325 stale = target->obd_stale_clients;
1326 remaining = hrtimer_expires_remaining(timer);
1327 left = ktime_divns(remaining, NSEC_PER_SEC);
1329 if (ktime_to_ns(remaining) > 0) {
1330 msg = "to recover in";
1333 msg = "already passed deadline";
1337 LCONSOLE_WARN("%s: Denying connection for new client %s (at %s), waiting for %d known clients (%d recovered, %d in progress, and %d evicted) %s %lld:%.02lld\n",
1338 target->obd_name, cluuid.uuid,
1339 libcfs_nid2str(req->rq_peer.nid), known,
1340 connected - in_progress, in_progress,
1341 stale, msg, timeout / 60, timeout % 60);
1345 rc = obd_connect(req->rq_svc_thread->t_env,
1346 &export, target, &cluuid, data,
1348 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1349 lustre_msg_add_op_flags(req->rq_repmsg,
1350 MSG_CONNECT_RECOVERING);
1352 conn.cookie = export->exp_handle.h_cookie;
1353 rc = rev_import_init(export);
1357 new_mds_mds_conn = true;
1360 rc = obd_reconnect(req->rq_svc_thread->t_env,
1361 export, target, &cluuid, data, client_nid);
1366 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1367 data->ocd_instance = target->u.obt.obt_instance;
1370 * Return only the parts of obd_connect_data that we understand, so the
1371 * client knows that we don't understand the rest.
1374 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1376 tmpdata = req_capsule_server_get(&req->rq_pill,
1379 * Don't use struct assignment here, because the client reply
1380 * buffer may be smaller/larger than the local struct
1383 memcpy(tmpdata, data, min(tmpsize, size));
1387 * If the client and the server are the same node, we will already
1388 * have an export that really points to the client's DLM export,
1389 * because we have a shared handles table.
1391 * XXX this will go away when shaver stops sending the "connect" handle
1392 * in the real "remote handle" field of the request --phik 24 Apr 2003
1394 ptlrpc_request_change_export(req, export);
1396 spin_lock(&export->exp_lock);
1397 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1398 spin_unlock(&export->exp_lock);
1400 "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1401 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1402 export->exp_conn_cnt,
1403 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1405 GOTO(out, rc = -EALREADY);
1407 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1408 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1409 spin_unlock(&export->exp_lock);
1411 if (export->exp_connection != NULL) {
1412 /* Check to see if connection came from another NID. */
1413 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1414 !hlist_unhashed(&export->exp_nid_hash))
1415 cfs_hash_del(export->exp_obd->obd_nid_hash,
1416 &export->exp_connection->c_peer.nid,
1417 &export->exp_nid_hash);
1419 ptlrpc_connection_put(export->exp_connection);
1422 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1425 if (hlist_unhashed(&export->exp_nid_hash))
1426 cfs_hash_add(export->exp_obd->obd_nid_hash,
1427 &export->exp_connection->c_peer.nid,
1428 &export->exp_nid_hash);
1430 lustre_msg_set_handle(req->rq_repmsg, &conn);
1432 rc = rev_import_reconnect(export, req);
1436 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1438 __u64 transno = data->ocd_transno;
1440 spin_lock(&export->exp_lock);
1442 * possible race with class_disconnect_stale_exports,
1443 * export may be already in the eviction process
1445 if (export->exp_failed) {
1446 spin_unlock(&export->exp_lock);
1447 GOTO(out, rc = -ENODEV);
1449 export->exp_in_recovery = 1;
1450 export->exp_req_replay_needed = 1;
1451 export->exp_lock_replay_needed = 1;
1452 spin_unlock(&export->exp_lock);
1454 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1455 MSG_CONNECT_TRANSNO);
1456 if (has_transno && transno == 0)
1457 CWARN("Connect with zero transno!\n");
1459 if (has_transno && transno > 0 &&
1460 transno < target->obd_next_recovery_transno &&
1461 transno > target->obd_last_committed) {
1462 /* Another way is to use cmpxchg() to be lock-free. */
1463 spin_lock(&target->obd_recovery_task_lock);
1464 if (transno < target->obd_next_recovery_transno)
1465 target->obd_next_recovery_transno = transno;
1466 spin_unlock(&target->obd_recovery_task_lock);
1469 atomic_inc(&target->obd_req_replay_clients);
1470 atomic_inc(&target->obd_lock_replay_clients);
1472 * Note: MDS-MDS connection is allowed to be connected during
1473 * recovery, no matter if the exports needs to be recoveried.
1474 * Because we need retrieve updates logs from all other MDTs.
1475 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1476 * also needs to be increased to match other recovery checking
1479 if (new_mds_mds_conn)
1480 target->obd_max_recoverable_clients++;
1481 if (atomic_inc_return(&target->obd_connected_clients) ==
1482 target->obd_max_recoverable_clients)
1483 wake_up(&target->obd_next_transno_waitq);
1486 /* Tell the client we're in recovery, when client is involved in it. */
1487 if (target->obd_recovering && !lw_client)
1488 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1492 spin_lock(&export->exp_lock);
1493 export->exp_connecting = 0;
1494 spin_unlock(&export->exp_lock);
1496 class_export_put(export);
1498 if (target != NULL) {
1499 spin_lock(&target->obd_dev_lock);
1500 target->obd_conn_inprogress--;
1501 spin_unlock(&target->obd_dev_lock);
1502 class_decref(target, "find", current);
1504 req->rq_status = rc;
1508 int target_handle_disconnect(struct ptlrpc_request *req)
1514 rc = req_capsule_server_pack(&req->rq_pill);
1518 /* Keep the rq_export around so we can send the reply. */
1519 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1524 void target_destroy_export(struct obd_export *exp)
1526 struct obd_import *imp = NULL;
1528 * exports created from last_rcvd data, and "fake"
1529 * exports created by lctl don't have an import
1531 spin_lock(&exp->exp_lock);
1532 if (exp->exp_imp_reverse != NULL) {
1533 imp = exp->exp_imp_reverse;
1534 exp->exp_imp_reverse = NULL;
1536 spin_unlock(&exp->exp_lock);
1538 client_destroy_import(imp);
1540 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1541 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1542 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1543 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1545 EXPORT_SYMBOL(target_destroy_export);
1548 * Recovery functions
1550 static void target_request_copy_get(struct ptlrpc_request *req)
1552 class_export_rpc_inc(req->rq_export);
1553 LASSERT(list_empty(&req->rq_list));
1554 INIT_LIST_HEAD(&req->rq_replay_list);
1556 /* Increase refcount to keep request in queue. */
1557 atomic_inc(&req->rq_refcount);
1558 /* Let export know it has replays to be handled. */
1559 atomic_inc(&req->rq_export->exp_replay_count);
1562 static void target_request_copy_put(struct ptlrpc_request *req)
1564 LASSERT(list_empty(&req->rq_replay_list));
1565 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1567 atomic_dec(&req->rq_export->exp_replay_count);
1568 class_export_rpc_dec(req->rq_export);
1569 ptlrpc_server_drop_request(req);
1572 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1574 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1575 struct obd_export *exp = req->rq_export;
1576 struct ptlrpc_request *reqiter;
1577 struct ptlrpc_request *dup_req = NULL;
1582 spin_lock(&exp->exp_lock);
1583 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1585 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1593 /* We expect it with RESENT and REPLAY flags. */
1594 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1595 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1596 CERROR("invalid flags %x of resent replay\n",
1597 lustre_msg_get_flags(req->rq_reqmsg));
1599 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1602 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1604 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1605 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1609 list_add_tail(&req->rq_replay_list,
1610 &exp->exp_req_replay_queue);
1613 spin_unlock(&exp->exp_lock);
1617 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1619 LASSERT(!list_empty(&req->rq_replay_list));
1620 LASSERT(req->rq_export);
1622 spin_lock(&req->rq_export->exp_lock);
1623 list_del_init(&req->rq_replay_list);
1624 spin_unlock(&req->rq_export->exp_lock);
1627 static void target_finish_recovery(struct lu_target *lut)
1629 struct obd_device *obd = lut->lut_obd;
1633 /* Only log a recovery message when recovery has occurred. */
1634 if (obd->obd_recovery_start) {
1635 time64_t now = ktime_get_real_seconds();
1636 time64_t elapsed_time;
1638 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1640 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1641 obd->obd_name, (s64)elapsed_time / 60,
1642 (s64)elapsed_time % 60,
1643 obd->obd_max_recoverable_clients,
1644 atomic_read(&obd->obd_connected_clients),
1645 obd->obd_stale_clients,
1646 obd->obd_stale_clients == 1 ? "was" : "were");
1649 ldlm_reprocess_recovery_done(obd->obd_namespace);
1650 spin_lock(&obd->obd_recovery_task_lock);
1651 if (!list_empty(&obd->obd_req_replay_queue) ||
1652 !list_empty(&obd->obd_lock_replay_queue) ||
1653 !list_empty(&obd->obd_final_req_queue)) {
1654 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1656 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1657 list_empty(&obd->obd_lock_replay_queue) ? \
1659 list_empty(&obd->obd_final_req_queue) ? \
1661 spin_unlock(&obd->obd_recovery_task_lock);
1664 spin_unlock(&obd->obd_recovery_task_lock);
1666 obd->obd_recovery_end = ktime_get_real_seconds();
1668 /* When recovery finished, cleanup orphans on MDS and OST. */
1669 if (obd->obd_type && OBP(obd, postrecov)) {
1670 int rc = OBP(obd, postrecov)(obd);
1673 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1679 static void abort_req_replay_queue(struct obd_device *obd)
1681 struct ptlrpc_request *req, *n;
1682 struct list_head abort_list;
1684 INIT_LIST_HEAD(&abort_list);
1685 spin_lock(&obd->obd_recovery_task_lock);
1686 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1687 spin_unlock(&obd->obd_recovery_task_lock);
1688 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1689 DEBUG_REQ(D_WARNING, req, "aborted:");
1690 req->rq_status = -ENOTCONN;
1691 if (ptlrpc_error(req)) {
1692 DEBUG_REQ(D_ERROR, req,
1693 "failed abort_req_reply; skipping");
1695 target_exp_dequeue_req_replay(req);
1696 target_request_copy_put(req);
1700 static void abort_lock_replay_queue(struct obd_device *obd)
1702 struct ptlrpc_request *req, *n;
1703 struct list_head abort_list;
1705 INIT_LIST_HEAD(&abort_list);
1706 spin_lock(&obd->obd_recovery_task_lock);
1707 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1708 spin_unlock(&obd->obd_recovery_task_lock);
1709 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1710 DEBUG_REQ(D_ERROR, req, "aborted:");
1711 req->rq_status = -ENOTCONN;
1712 if (ptlrpc_error(req)) {
1713 DEBUG_REQ(D_ERROR, req,
1714 "failed abort_lock_reply; skipping");
1716 target_request_copy_put(req);
1721 * Called from a cleanup function if the device is being cleaned up
1722 * forcefully. The exports should all have been disconnected already,
1723 * the only thing left to do is
1724 * - clear the recovery flags
1725 * - cancel the timer
1726 * - free queued requests and replies, but don't send replies
1727 * Because the obd_stopping flag is set, no new requests should be received.
1729 void target_cleanup_recovery(struct obd_device *obd)
1731 struct ptlrpc_request *req, *n;
1732 struct list_head clean_list;
1734 INIT_LIST_HEAD(&clean_list);
1735 spin_lock(&obd->obd_dev_lock);
1736 if (!obd->obd_recovering) {
1737 spin_unlock(&obd->obd_dev_lock);
1741 obd->obd_recovering = obd->obd_abort_recovery = 0;
1742 spin_unlock(&obd->obd_dev_lock);
1744 spin_lock(&obd->obd_recovery_task_lock);
1745 target_cancel_recovery_timer(obd);
1746 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1747 spin_unlock(&obd->obd_recovery_task_lock);
1749 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1750 LASSERT(req->rq_reply_state == NULL);
1751 target_exp_dequeue_req_replay(req);
1752 target_request_copy_put(req);
1755 spin_lock(&obd->obd_recovery_task_lock);
1756 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1757 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1758 spin_unlock(&obd->obd_recovery_task_lock);
1760 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1761 LASSERT(req->rq_reply_state == NULL);
1762 target_request_copy_put(req);
1767 EXPORT_SYMBOL(target_cleanup_recovery);
1769 /* obd_recovery_task_lock should be held */
1770 void target_cancel_recovery_timer(struct obd_device *obd)
1772 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1773 hrtimer_cancel(&obd->obd_recovery_timer);
1776 static void target_start_recovery_timer(struct obd_device *obd)
1780 if (obd->obd_recovery_start != 0)
1783 spin_lock(&obd->obd_dev_lock);
1784 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1785 spin_unlock(&obd->obd_dev_lock);
1789 LASSERT(obd->obd_recovery_timeout != 0);
1791 if (obd->obd_recovery_start != 0) {
1792 spin_unlock(&obd->obd_dev_lock);
1796 delay = ktime_set(obd->obd_recovery_timeout, 0);
1797 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_REL);
1798 obd->obd_recovery_start = ktime_get_real_seconds();
1799 spin_unlock(&obd->obd_dev_lock);
1801 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1803 obd->obd_recovery_timeout / 60,
1804 obd->obd_recovery_timeout % 60,
1805 obd->obd_max_recoverable_clients,
1806 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1807 (obd->obd_max_recoverable_clients == 1) ? "s" : "");
1811 * extend recovery window.
1813 * if @extend is true, extend recovery window to have @dr_timeout remaining
1814 * at least; otherwise, make sure the recovery timeout value is not less
1817 static void extend_recovery_timer(struct obd_device *obd, time_t dr_timeout,
1824 spin_lock(&obd->obd_dev_lock);
1825 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1826 spin_unlock(&obd->obd_dev_lock);
1829 LASSERT(obd->obd_recovery_start != 0);
1831 left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1832 left = ktime_divns(left_ns, NSEC_PER_SEC);
1835 timeout = obd->obd_recovery_timeout;
1836 /* dr_timeout will happen after the hrtimer has expired.
1837 * Add the excess time to the soft recovery timeout without
1838 * exceeding the hard recovery timeout.
1840 if (dr_timeout > left) {
1841 timeout += dr_timeout - left;
1842 timeout = min_t(time_t, obd->obd_recovery_time_hard,
1846 timeout = clamp_t(time_t, dr_timeout, obd->obd_recovery_timeout,
1847 obd->obd_recovery_time_hard);
1850 if (timeout == obd->obd_recovery_time_hard)
1851 CWARN("%s: extended recovery timer reached hard limit: %ld, extend: %d\n",
1852 obd->obd_name, timeout, extend);
1854 if (obd->obd_recovery_timeout < timeout) {
1855 ktime_t now = ktime_get_real();
1858 obd->obd_recovery_timeout = timeout;
1859 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1860 left_ns = ktime_sub(end, now);
1861 hrtimer_forward_now(&obd->obd_recovery_timer, left_ns);
1862 left = ktime_divns(left_ns, NSEC_PER_MSEC);
1864 spin_unlock(&obd->obd_dev_lock);
1866 CDEBUG(D_HA, "%s: recovery timer will expire in %ld seconds\n",
1867 obd->obd_name, left);
1870 /* Reset the timer with each new client connection */
1872 * This timer is actually reconnect_timer, which is for making sure
1873 * the total recovery window is at least as big as my reconnect
1874 * attempt timing. So the initial recovery time_out will be set to
1875 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1876 * from client is bigger than this, then the recovery time_out will
1877 * be extended to make sure the client could be reconnected, in the
1878 * process, the timeout from the new client should be ignored.
1881 check_and_start_recovery_timer(struct obd_device *obd,
1882 struct ptlrpc_request *req,
1885 time_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1886 struct obd_device_target *obt = &obd->u.obt;
1888 if (!new_client && service_time)
1890 * Teach server about old server's estimates, as first guess
1891 * at how long new requests will take.
1893 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1896 target_start_recovery_timer(obd);
1899 * Convert the service time to RPC timeout,
1900 * and reuse service_time to limit stack usage.
1902 service_time = at_est2timeout(service_time);
1904 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1905 service_time < at_extra)
1906 service_time = at_extra;
1909 * We expect other clients to timeout within service_time, then try
1910 * to reconnect, then try the failover server. The max delay between
1911 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
1913 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1915 LASSERT(obt->obt_magic == OBT_MAGIC);
1916 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1917 if (service_time > obd->obd_recovery_timeout && !new_client)
1918 extend_recovery_timer(obd, service_time, false);
1921 /** Health checking routines */
1922 static inline int exp_connect_healthy(struct obd_export *exp)
1924 return exp->exp_in_recovery;
1927 /** if export done req_replay or has replay in queue */
1928 static inline int exp_req_replay_healthy(struct obd_export *exp)
1930 return (!exp->exp_req_replay_needed ||
1931 atomic_read(&exp->exp_replay_count) > 0);
1935 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1937 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1938 exp_req_replay_healthy(exp);
1941 /** if export done lock_replay or has replay in queue */
1942 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1944 return (!exp->exp_lock_replay_needed ||
1945 atomic_read(&exp->exp_replay_count) > 0);
1948 static inline int exp_vbr_healthy(struct obd_export *exp)
1950 return !exp->exp_vbr_failed;
1953 static inline int exp_finished(struct obd_export *exp)
1955 return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
1958 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1960 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1964 static int check_for_next_transno(struct lu_target *lut)
1966 struct ptlrpc_request *req = NULL;
1967 struct obd_device *obd = lut->lut_obd;
1968 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1969 int wake_up = 0, connected, completed, queue_len;
1970 __u64 req_transno = 0;
1971 __u64 update_transno = 0;
1972 __u64 next_transno = 0;
1976 spin_lock(&obd->obd_recovery_task_lock);
1977 if (!list_empty(&obd->obd_req_replay_queue)) {
1978 req = list_entry(obd->obd_req_replay_queue.next,
1979 struct ptlrpc_request, rq_list);
1980 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1984 update_transno = distribute_txn_get_next_transno(tdtd);
1986 connected = atomic_read(&obd->obd_connected_clients);
1987 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1988 queue_len = obd->obd_requests_queued_for_recovery;
1989 next_transno = obd->obd_next_recovery_transno;
1992 "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
1993 obd->obd_max_recoverable_clients, connected, completed,
1994 queue_len, req_transno, next_transno);
1996 if (obd->obd_abort_recovery) {
1997 CDEBUG(D_HA, "waking for aborted recovery\n");
1999 } else if (obd->obd_recovery_expired) {
2000 CDEBUG(D_HA, "waking for expired recovery\n");
2002 } else if (tdtd != NULL && req != NULL &&
2003 is_req_replayed_by_update(req)) {
2004 LASSERTF(req_transno < next_transno,
2005 "req_transno %llu next_transno%llu\n", req_transno,
2007 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2010 } else if (req_transno == next_transno ||
2011 (update_transno != 0 && update_transno <= next_transno)) {
2012 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2014 } else if (queue_len > 0 &&
2015 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2016 /** handle gaps occured due to lost reply or VBR */
2017 LASSERTF(req_transno >= next_transno,
2018 "req_transno: %llu, next_transno: %llu\n",
2019 req_transno, next_transno);
2021 "%s: waking for gap in transno, VBR is %s (skip: %lld, ql: %d, comp: %d, conn: %d, next: %lld, next_update %lld last_committed: %lld)\n",
2022 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2023 next_transno, queue_len, completed, connected,
2024 req_transno, update_transno, obd->obd_last_committed);
2025 obd->obd_next_recovery_transno = req_transno;
2027 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2028 CDEBUG(D_HA, "waking for completed recovery\n");
2030 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2032 "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2034 obd->obd_next_recovery_transno = req_transno;
2037 spin_unlock(&obd->obd_recovery_task_lock);
2041 static int check_for_next_lock(struct lu_target *lut)
2043 struct obd_device *obd = lut->lut_obd;
2046 spin_lock(&obd->obd_recovery_task_lock);
2047 if (!list_empty(&obd->obd_lock_replay_queue)) {
2048 CDEBUG(D_HA, "waking for next lock\n");
2050 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2051 CDEBUG(D_HA, "waking for completed lock replay\n");
2053 } else if (obd->obd_abort_recovery) {
2054 CDEBUG(D_HA, "waking for aborted recovery\n");
2056 } else if (obd->obd_recovery_expired) {
2057 CDEBUG(D_HA, "waking for expired recovery\n");
2060 spin_unlock(&obd->obd_recovery_task_lock);
2066 * wait for recovery events,
2067 * check its status with help of check_routine
2068 * evict dead clients via health_check
2070 static int target_recovery_overseer(struct lu_target *lut,
2071 int (*check_routine)(struct lu_target *),
2072 int (*health_check)(struct obd_export *))
2074 struct obd_device *obd = lut->lut_obd;
2075 struct target_distribute_txn_data *tdtd;
2079 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2080 now = ktime_get_seconds();
2081 if (now - last > 600) {
2082 LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2087 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
2088 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2089 __u64 next_update_transno = 0;
2092 * Only abort the recovery if there are no update recovery
2095 spin_lock(&obd->obd_recovery_task_lock);
2096 if (lut->lut_tdtd != NULL) {
2097 next_update_transno =
2098 distribute_txn_get_next_transno(lut->lut_tdtd);
2100 tdtd = lut->lut_tdtd;
2102 * If next_update_transno == 0, it probably because
2103 * updatelog retrieve threads did not get any records
2104 * yet, let's wait those threads stopped
2106 if (next_update_transno == 0) {
2107 struct l_wait_info lwi = { 0 };
2109 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2111 &tdtd->tdtd_recovery_threads_count) == 0,
2114 next_update_transno =
2115 distribute_txn_get_next_transno(
2120 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2121 obd->obd_next_recovery_transno = next_update_transno;
2122 spin_unlock(&obd->obd_recovery_task_lock);
2124 * Disconnect unfinished exports from clients, and
2125 * keep connection from MDT to make sure the update
2126 * recovery will still keep trying until some one
2127 * manually abort the recovery
2129 class_disconnect_stale_exports(obd,
2130 exp_finished_or_from_mdt);
2131 /* Abort all of replay & replay lock req from clients */
2132 abort_req_replay_queue(obd);
2133 abort_lock_replay_queue(obd);
2135 "%s: there are still update replay (%#llx)in the queue.\n",
2136 obd->obd_name, next_update_transno);
2138 obd->obd_abort_recovery = 1;
2139 spin_unlock(&obd->obd_recovery_task_lock);
2140 CWARN("%s recovery is aborted by hard timeout\n",
2145 while (wait_event_timeout(obd->obd_next_transno_waitq,
2147 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2148 ; /* wait indefinitely for event, but don't trigger watchdog */
2150 if (obd->obd_abort_recovery) {
2151 CWARN("recovery is aborted, evict exports in recovery\n");
2152 if (lut->lut_tdtd != NULL) {
2153 struct l_wait_info lwi = { 0 };
2155 tdtd = lut->lut_tdtd;
2157 * Let's wait all of the update log recovery thread
2160 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2161 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2163 /* Then abort the update recovery list */
2164 dtrq_list_destroy(lut->lut_tdtd);
2167 /** evict exports which didn't finish recovery yet */
2168 class_disconnect_stale_exports(obd, exp_finished);
2170 } else if (obd->obd_recovery_expired) {
2171 obd->obd_recovery_expired = 0;
2172 /** If some clients died being recovered, evict them */
2173 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2175 /** evict cexports with no replay in queue, they are stalled */
2176 class_disconnect_stale_exports(obd, health_check);
2178 /** continue with VBR */
2179 spin_lock(&obd->obd_dev_lock);
2180 obd->obd_version_recov = 1;
2181 spin_unlock(&obd->obd_dev_lock);
2183 * reset timer, recovery will proceed with versions now,
2184 * timeout is set just to handle reconnection delays
2186 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2188 * Wait for recovery events again, after evicting bad clients
2195 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2197 struct obd_device *obd = lut->lut_obd;
2198 struct ptlrpc_request *req = NULL;
2200 CDEBUG(D_HA, "Waiting for lock\n");
2201 if (target_recovery_overseer(lut, check_for_next_lock,
2202 exp_lock_replay_healthy))
2203 abort_lock_replay_queue(obd);
2205 spin_lock(&obd->obd_recovery_task_lock);
2206 if (!list_empty(&obd->obd_lock_replay_queue)) {
2207 req = list_entry(obd->obd_lock_replay_queue.next,
2208 struct ptlrpc_request, rq_list);
2209 list_del_init(&req->rq_list);
2210 spin_unlock(&obd->obd_recovery_task_lock);
2212 spin_unlock(&obd->obd_recovery_task_lock);
2213 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2214 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2215 /** evict exports failed VBR */
2216 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2221 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2223 struct ptlrpc_request *req = NULL;
2225 spin_lock(&obd->obd_recovery_task_lock);
2226 if (!list_empty(&obd->obd_final_req_queue)) {
2227 req = list_entry(obd->obd_final_req_queue.next,
2228 struct ptlrpc_request, rq_list);
2229 list_del_init(&req->rq_list);
2230 spin_unlock(&obd->obd_recovery_task_lock);
2231 if (req->rq_export->exp_in_recovery) {
2232 spin_lock(&req->rq_export->exp_lock);
2233 req->rq_export->exp_in_recovery = 0;
2234 spin_unlock(&req->rq_export->exp_lock);
2237 spin_unlock(&obd->obd_recovery_task_lock);
2242 static void handle_recovery_req(struct ptlrpc_thread *thread,
2243 struct ptlrpc_request *req,
2244 svc_handler_t handler)
2249 * export can be evicted during recovery, no need to handle replays for
2250 * it after that, discard such request silently
2252 if (req->rq_export->exp_disconnected)
2255 req->rq_session.lc_thread = thread;
2256 req->rq_svc_thread = thread;
2257 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2259 /* thread context */
2260 lu_context_enter(&thread->t_env->le_ctx);
2262 lu_context_exit(&thread->t_env->le_ctx);
2264 /* don't reset timer for final stage */
2265 if (!exp_finished(req->rq_export)) {
2266 time_t to = obd_timeout;
2269 * Add request timeout to the recovery time so next request from
2270 * this client may come in recovery time
2273 struct ptlrpc_service_part *svcpt;
2275 svcpt = req->rq_rqbd->rqbd_svcpt;
2277 * If the server sent early reply for this request,
2278 * the client will recalculate the timeout according to
2279 * current server estimate service time, so we will
2280 * use the maxium timeout here for waiting the client
2281 * sending the next req
2284 at_est2timeout(at_get(&svcpt->scp_at_estimate)),
2285 lustre_msg_get_timeout(req->rq_reqmsg));
2287 * Add 2 net_latency, one for balance rq_deadline
2288 * (see ptl_send_rpc), one for resend the req to server,
2289 * Note: client will pack net_latency in replay req
2290 * (see ptlrpc_replay_req)
2292 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2294 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2299 /** Checking routines for recovery */
2300 static int check_for_recovery_ready(struct lu_target *lut)
2302 struct obd_device *obd = lut->lut_obd;
2303 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2306 "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2307 clnts, obd->obd_stale_clients, obd->obd_max_recoverable_clients,
2308 obd->obd_abort_recovery, obd->obd_recovery_expired);
2310 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2311 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2312 if (clnts + obd->obd_stale_clients <
2313 obd->obd_max_recoverable_clients)
2317 if (lut->lut_tdtd != NULL) {
2318 if (!lut->lut_tdtd->tdtd_replay_ready &&
2319 !obd->obd_abort_recovery) {
2321 * Let's extend recovery timer, in case the recovery
2322 * timer expired, and some clients got evicted
2324 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2327 "%s update recovery is not ready, extend recovery %llu\n",
2328 obd->obd_name, obd->obd_recovery_timeout);
2337 REQUEST_RECOVERY = 1,
2338 UPDATE_RECOVERY = 2,
2341 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2345 if (!list_empty(&obd->obd_req_replay_queue)) {
2346 struct ptlrpc_request *req;
2348 req = list_entry(obd->obd_req_replay_queue.next,
2349 struct ptlrpc_request, rq_list);
2350 transno = lustre_msg_get_transno(req->rq_reqmsg);
2356 static __u64 get_next_transno(struct lu_target *lut, int *type)
2358 struct obd_device *obd = lut->lut_obd;
2359 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2361 __u64 update_transno;
2365 transno = get_next_replay_req_transno(obd);
2367 *type = REQUEST_RECOVERY;
2372 update_transno = distribute_txn_get_next_transno(tdtd);
2373 if (transno == 0 || (transno >= update_transno &&
2374 update_transno != 0)) {
2375 transno = update_transno;
2377 *type = UPDATE_RECOVERY;
2384 * drop duplicate replay request
2386 * Because the operation has been replayed by update recovery, the request
2387 * with the same transno will be dropped and also notify the client to send
2388 * next replay request.
2390 * \param[in] env execution environment
2391 * \param[in] obd failover obd device
2392 * \param[in] req request to be dropped
2394 static void drop_duplicate_replay_req(struct lu_env *env,
2395 struct obd_device *obd,
2396 struct ptlrpc_request *req)
2398 DEBUG_REQ(D_HA, req,
2399 "remove t%lld from %s because of duplicate update records are found.\n",
2400 lustre_msg_get_transno(req->rq_reqmsg),
2401 libcfs_nid2str(req->rq_peer.nid));
2404 * Right now, only for MDS reint operation update replay and
2405 * normal request replay can have the same transno
2407 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2408 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2409 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2410 if (likely(req->rq_export))
2411 target_committed_to_req(req);
2412 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2413 target_send_reply(req, req->rq_status, 0);
2415 DEBUG_REQ(D_ERROR, req, "wrong opc from %s\n",
2416 libcfs_nid2str(req->rq_peer.nid));
2418 target_exp_dequeue_req_replay(req);
2419 target_request_copy_put(req);
2420 obd->obd_replayed_requests++;
2423 static void replay_request_or_update(struct lu_env *env,
2424 struct lu_target *lut,
2425 struct target_recovery_data *trd,
2426 struct ptlrpc_thread *thread)
2428 struct obd_device *obd = lut->lut_obd;
2429 struct ptlrpc_request *req = NULL;
2435 CDEBUG(D_HA, "Waiting for transno %lld\n",
2436 obd->obd_next_recovery_transno);
2438 /* Replay all of request and update by transno */
2440 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2442 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2445 * It is needed to extend recovery window above
2446 * recovery_time_soft. Extending is possible only in the
2447 * end of recovery window (see more details in
2448 * handle_recovery_req()).
2450 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2452 if (target_recovery_overseer(lut, check_for_next_transno,
2453 exp_req_replay_healthy_or_from_mdt)) {
2454 abort_req_replay_queue(obd);
2455 abort_lock_replay_queue(obd);
2459 spin_lock(&obd->obd_recovery_task_lock);
2460 transno = get_next_transno(lut, &type);
2461 if (type == REQUEST_RECOVERY && transno != 0) {
2463 * Drop replay request from client side, if the
2464 * replay has been executed by update with the
2467 req = list_entry(obd->obd_req_replay_queue.next,
2468 struct ptlrpc_request, rq_list);
2470 list_del_init(&req->rq_list);
2471 obd->obd_requests_queued_for_recovery--;
2472 spin_unlock(&obd->obd_recovery_task_lock);
2475 * Let's check if the request has been redone by
2478 if (is_req_replayed_by_update(req)) {
2479 struct distribute_txn_replay_req *dtrq;
2481 dtrq = distribute_txn_lookup_finish_list(tdtd,
2483 LASSERT(dtrq != NULL);
2484 spin_lock(&tdtd->tdtd_replay_list_lock);
2485 list_del_init(&dtrq->dtrq_list);
2486 spin_unlock(&tdtd->tdtd_replay_list_lock);
2489 drop_duplicate_replay_req(env, obd, req);
2494 LASSERT(trd->trd_processing_task == current_pid());
2495 DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2497 lustre_msg_get_transno(req->rq_reqmsg),
2498 libcfs_nid2str(req->rq_peer.nid));
2500 handle_recovery_req(thread, req,
2501 trd->trd_recovery_handler);
2503 * bz18031: increase next_recovery_transno before
2504 * target_request_copy_put() will drop exp_rpc reference
2506 spin_lock(&obd->obd_recovery_task_lock);
2507 obd->obd_next_recovery_transno++;
2508 spin_unlock(&obd->obd_recovery_task_lock);
2509 target_exp_dequeue_req_replay(req);
2510 target_request_copy_put(req);
2511 obd->obd_replayed_requests++;
2512 } else if (type == UPDATE_RECOVERY && transno != 0) {
2513 struct distribute_txn_replay_req *dtrq;
2516 spin_unlock(&obd->obd_recovery_task_lock);
2518 LASSERT(tdtd != NULL);
2519 dtrq = distribute_txn_get_next_req(tdtd);
2520 lu_context_enter(&thread->t_env->le_ctx);
2521 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2522 lu_context_exit(&thread->t_env->le_ctx);
2523 extend_recovery_timer(obd, obd_timeout, true);
2525 if (rc == 0 && dtrq->dtrq_xid != 0) {
2527 "Move x%llu t%llu to finish list\n",
2529 dtrq->dtrq_master_transno);
2531 /* Add it to the replay finish list */
2532 spin_lock(&tdtd->tdtd_replay_list_lock);
2533 list_add(&dtrq->dtrq_list,
2534 &tdtd->tdtd_replay_finish_list);
2535 spin_unlock(&tdtd->tdtd_replay_list_lock);
2537 spin_lock(&obd->obd_recovery_task_lock);
2538 if (transno == obd->obd_next_recovery_transno)
2539 obd->obd_next_recovery_transno++;
2541 obd->obd_next_recovery_transno)
2542 obd->obd_next_recovery_transno =
2544 spin_unlock(&obd->obd_recovery_task_lock);
2549 spin_unlock(&obd->obd_recovery_task_lock);
2551 LASSERT(list_empty(&obd->obd_req_replay_queue));
2552 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2553 /** evict exports failed VBR */
2554 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2560 static int target_recovery_thread(void *arg)
2562 struct lu_target *lut = arg;
2563 struct obd_device *obd = lut->lut_obd;
2564 struct ptlrpc_request *req;
2565 struct target_recovery_data *trd = &obd->obd_recovery_data;
2566 unsigned long delta;
2568 struct ptlrpc_thread *thread = NULL;
2573 unshare_fs_struct();
2574 OBD_ALLOC_PTR(thread);
2580 GOTO(out_thread, rc = -ENOMEM);
2581 rc = lu_env_add(env);
2585 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2587 GOTO(out_env_remove, rc);
2589 thread->t_env = env;
2590 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2591 env->le_ctx.lc_thread = thread;
2592 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2594 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2596 trd->trd_processing_task = current_pid();
2598 spin_lock(&obd->obd_dev_lock);
2599 obd->obd_recovering = 1;
2600 spin_unlock(&obd->obd_dev_lock);
2601 complete(&trd->trd_starting);
2603 /* first of all, we have to know the first transno to replay */
2604 if (target_recovery_overseer(lut, check_for_recovery_ready,
2605 exp_connect_healthy)) {
2606 abort_req_replay_queue(obd);
2607 abort_lock_replay_queue(obd);
2608 if (lut->lut_tdtd != NULL)
2609 dtrq_list_destroy(lut->lut_tdtd);
2612 /* next stage: replay requests or update */
2614 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2615 atomic_read(&obd->obd_req_replay_clients),
2616 obd->obd_next_recovery_transno);
2617 replay_request_or_update(env, lut, trd, thread);
2620 * The second stage: replay locks
2622 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2623 atomic_read(&obd->obd_lock_replay_clients));
2624 while ((req = target_next_replay_lock(lut))) {
2625 LASSERT(trd->trd_processing_task == current_pid());
2626 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2627 libcfs_nid2str(req->rq_peer.nid));
2628 handle_recovery_req(thread, req,
2629 trd->trd_recovery_handler);
2630 target_request_copy_put(req);
2631 obd->obd_replayed_locks++;
2635 * The third stage: reply on final pings, at this moment all clients
2636 * must have request in final queue
2638 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2639 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2640 /** Update server last boot epoch */
2641 tgt_boot_epoch_update(lut);
2643 * We drop recoverying flag to forward all new requests
2644 * to regular mds_handle() since now
2646 spin_lock(&obd->obd_dev_lock);
2647 obd->obd_recovering = obd->obd_abort_recovery = 0;
2648 spin_unlock(&obd->obd_dev_lock);
2649 spin_lock(&obd->obd_recovery_task_lock);
2650 target_cancel_recovery_timer(obd);
2651 spin_unlock(&obd->obd_recovery_task_lock);
2652 while ((req = target_next_final_ping(obd))) {
2653 LASSERT(trd->trd_processing_task == current_pid());
2654 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2655 libcfs_nid2str(req->rq_peer.nid));
2656 handle_recovery_req(thread, req,
2657 trd->trd_recovery_handler);
2659 * Because the waiting client can not send ping to server,
2660 * so we need refresh the last_request_time, to avoid the
2661 * export is being evicted
2663 ptlrpc_update_export_timer(req->rq_export, 0);
2664 target_request_copy_put(req);
2667 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2668 CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2669 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2670 if (delta > OBD_RECOVERY_TIME_SOFT) {
2671 CWARN("too long recovery - read logs\n");
2672 libcfs_debug_dumplog();
2675 target_finish_recovery(lut);
2676 lu_context_fini(&env->le_ctx);
2677 trd->trd_processing_task = 0;
2678 complete_all(&trd->trd_finishing);
2679 tgt_io_thread_done(thread);
2685 OBD_FREE_PTR(thread);
2689 static int target_start_recovery_thread(struct lu_target *lut,
2690 svc_handler_t handler)
2692 struct obd_device *obd = lut->lut_obd;
2694 struct target_recovery_data *trd = &obd->obd_recovery_data;
2697 memset(trd, 0, sizeof(*trd));
2698 init_completion(&trd->trd_starting);
2699 init_completion(&trd->trd_finishing);
2700 trd->trd_recovery_handler = handler;
2702 rc = server_name2index(obd->obd_name, &index, NULL);
2706 if (!IS_ERR(kthread_run(target_recovery_thread,
2707 lut, "tgt_recover_%d", index))) {
2708 wait_for_completion(&trd->trd_starting);
2709 LASSERT(obd->obd_recovering != 0);
2717 void target_stop_recovery_thread(struct obd_device *obd)
2719 if (obd->obd_recovery_data.trd_processing_task > 0) {
2720 struct target_recovery_data *trd = &obd->obd_recovery_data;
2721 /** recovery can be done but postrecovery is not yet */
2722 spin_lock(&obd->obd_dev_lock);
2723 if (obd->obd_recovering) {
2724 CERROR("%s: Aborting recovery\n", obd->obd_name);
2725 obd->obd_abort_recovery = 1;
2726 wake_up(&obd->obd_next_transno_waitq);
2728 spin_unlock(&obd->obd_dev_lock);
2729 wait_for_completion(&trd->trd_finishing);
2732 EXPORT_SYMBOL(target_stop_recovery_thread);
2734 void target_recovery_fini(struct obd_device *obd)
2736 class_disconnect_exports(obd);
2737 target_stop_recovery_thread(obd);
2738 target_cleanup_recovery(obd);
2740 EXPORT_SYMBOL(target_recovery_fini);
2742 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2744 struct obd_device *obd = container_of(timer, struct obd_device,
2745 obd_recovery_timer);
2748 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2749 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2750 ktime_get_real_seconds() - obd->obd_recovery_start,
2751 atomic_read(&obd->obd_connected_clients));
2753 obd->obd_recovery_expired = 1;
2754 wake_up(&obd->obd_next_transno_waitq);
2755 return HRTIMER_NORESTART;
2758 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2760 struct obd_device *obd = lut->lut_obd;
2762 if (lut->lut_bottom->dd_rdonly)
2765 if (obd->obd_max_recoverable_clients == 0) {
2766 /** Update server last boot epoch */
2767 tgt_boot_epoch_update(lut);
2771 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2772 "last_transno %llu\n", obd->obd_name,
2773 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2774 LASSERT(obd->obd_stopping == 0);
2775 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2776 obd->obd_recovery_start = 0;
2777 obd->obd_recovery_end = 0;
2779 hrtimer_init(&obd->obd_recovery_timer, CLOCK_REALTIME,
2781 obd->obd_recovery_timer.function = &target_recovery_expired;
2782 target_start_recovery_thread(lut, handler);
2784 EXPORT_SYMBOL(target_recovery_init);
2786 static int target_process_req_flags(struct obd_device *obd,
2787 struct ptlrpc_request *req)
2789 struct obd_export *exp = req->rq_export;
2791 LASSERT(exp != NULL);
2792 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2793 /* client declares he's ready to replay locks */
2794 spin_lock(&exp->exp_lock);
2795 if (exp->exp_req_replay_needed) {
2796 exp->exp_req_replay_needed = 0;
2797 spin_unlock(&exp->exp_lock);
2799 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2800 atomic_dec(&obd->obd_req_replay_clients);
2802 spin_unlock(&exp->exp_lock);
2805 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2807 * client declares he's ready to complete recovery
2808 * so, we put the request on th final queue
2810 spin_lock(&exp->exp_lock);
2811 if (exp->exp_lock_replay_needed) {
2812 exp->exp_lock_replay_needed = 0;
2813 spin_unlock(&exp->exp_lock);
2815 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2816 atomic_dec(&obd->obd_lock_replay_clients);
2818 spin_unlock(&exp->exp_lock);
2824 int target_queue_recovery_request(struct ptlrpc_request *req,
2825 struct obd_device *obd)
2827 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2828 struct ptlrpc_request *reqiter;
2833 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2834 /* Processing the queue right now, don't re-add. */
2838 target_process_req_flags(obd, req);
2840 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2841 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2842 if (cfs_fail_val == 1) {
2845 wake_up(&cfs_race_waitq);
2847 set_current_state(TASK_INTERRUPTIBLE);
2848 schedule_timeout(cfs_time_seconds(1));
2853 * client declares he's ready to complete recovery
2854 * so, we put the request on th final queue
2856 target_request_copy_get(req);
2857 DEBUG_REQ(D_HA, req, "queue final req");
2858 wake_up(&obd->obd_next_transno_waitq);
2859 spin_lock(&obd->obd_recovery_task_lock);
2860 if (obd->obd_recovering) {
2861 struct ptlrpc_request *tmp;
2862 struct ptlrpc_request *duplicate = NULL;
2864 if (likely(!req->rq_export->exp_replay_done)) {
2865 req->rq_export->exp_replay_done = 1;
2866 list_add_tail(&req->rq_list,
2867 &obd->obd_final_req_queue);
2868 spin_unlock(&obd->obd_recovery_task_lock);
2873 * XXX O(n), but only happens if final ping is
2874 * timed out, probably reorganize the list as
2877 list_for_each_entry_safe(reqiter, tmp,
2878 &obd->obd_final_req_queue,
2880 if (reqiter->rq_export == req->rq_export) {
2881 list_del_init(&reqiter->rq_list);
2882 duplicate = reqiter;
2887 list_add_tail(&req->rq_list,
2888 &obd->obd_final_req_queue);
2889 req->rq_export->exp_replay_done = 1;
2890 spin_unlock(&obd->obd_recovery_task_lock);
2892 if (duplicate != NULL) {
2893 DEBUG_REQ(D_HA, duplicate,
2894 "put prev final req\n");
2895 target_request_copy_put(duplicate);
2899 spin_unlock(&obd->obd_recovery_task_lock);
2900 target_request_copy_put(req);
2901 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2904 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2905 /* client declares he's ready to replay locks */
2906 target_request_copy_get(req);
2907 DEBUG_REQ(D_HA, req, "queue lock replay req");
2908 wake_up(&obd->obd_next_transno_waitq);
2909 spin_lock(&obd->obd_recovery_task_lock);
2910 LASSERT(obd->obd_recovering);
2911 /* usually due to recovery abort */
2912 if (!req->rq_export->exp_in_recovery) {
2913 spin_unlock(&obd->obd_recovery_task_lock);
2914 target_request_copy_put(req);
2917 LASSERT(req->rq_export->exp_lock_replay_needed);
2918 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2919 spin_unlock(&obd->obd_recovery_task_lock);
2924 * CAVEAT EMPTOR: The incoming request message has been swabbed
2925 * (i.e. buflens etc are in my own byte order), but type-dependent
2926 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
2930 INIT_LIST_HEAD(&req->rq_list);
2931 DEBUG_REQ(D_HA, req, "not queueing");
2936 * If we're processing the queue, we want don't want to queue this
2939 * Also, if this request has a transno less than the one we're waiting
2940 * for, we should process it now. It could (and currently always will)
2941 * be an open request for a descriptor that was opened some time ago.
2943 * Also, a resent, replayed request that has already been
2944 * handled will pass through here and be processed immediately.
2947 "Next recovery transno: %llu, current: %llu, replaying\n",
2948 obd->obd_next_recovery_transno, transno);
2951 * If the request has been replayed by update replay, then sends this
2952 * request to the recovery thread (replay_request_or_update()), where
2953 * it will be handled
2955 spin_lock(&obd->obd_recovery_task_lock);
2956 if (transno < obd->obd_next_recovery_transno &&
2957 !is_req_replayed_by_update(req)) {
2958 /* Processing the queue right now, don't re-add. */
2959 LASSERT(list_empty(&req->rq_list));
2960 spin_unlock(&obd->obd_recovery_task_lock);
2963 spin_unlock(&obd->obd_recovery_task_lock);
2965 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2968 target_request_copy_get(req);
2969 if (!req->rq_export->exp_in_recovery) {
2970 target_request_copy_put(req);
2973 LASSERT(req->rq_export->exp_req_replay_needed);
2975 if (target_exp_enqueue_req_replay(req)) {
2976 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2977 target_request_copy_put(req);
2982 spin_lock(&obd->obd_recovery_task_lock);
2983 LASSERT(obd->obd_recovering);
2984 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2985 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2986 list_add_tail(&req->rq_list, &reqiter->rq_list);
2991 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2993 DEBUG_REQ(D_ERROR, req,
2994 "dropping replay: transno has been claimed by another client");
2995 spin_unlock(&obd->obd_recovery_task_lock);
2996 target_exp_dequeue_req_replay(req);
2997 target_request_copy_put(req);
3003 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3005 obd->obd_requests_queued_for_recovery++;
3006 spin_unlock(&obd->obd_recovery_task_lock);
3007 wake_up(&obd->obd_next_transno_waitq);
3011 void target_committed_to_req(struct ptlrpc_request *req)
3013 struct obd_export *exp = req->rq_export;
3015 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3016 lustre_msg_set_last_committed(req->rq_repmsg,
3017 exp->exp_last_committed);
3019 DEBUG_REQ(D_IOCTL, req,
3020 "not sending last_committed update (%d/%d)",
3021 exp->exp_obd->obd_no_transno,
3022 req->rq_repmsg == NULL);
3024 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3025 exp->exp_last_committed, req->rq_transno, req->rq_xid);
3028 #endif /* HAVE_SERVER_SUPPORT */
3031 * Packs current SLV and Limit into \a req.
3033 int target_pack_pool_reply(struct ptlrpc_request *req)
3035 struct obd_device *obd;
3040 * Check that we still have all structures alive as this may
3041 * be some late RPC at shutdown time.
3043 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3044 !exp_connect_lru_resize(req->rq_export))) {
3045 lustre_msg_set_slv(req->rq_repmsg, 0);
3046 lustre_msg_set_limit(req->rq_repmsg, 0);
3050 /* OBD is alive here as export is alive, which we checked above. */
3051 obd = req->rq_export->exp_obd;
3053 read_lock(&obd->obd_pool_lock);
3054 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3055 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3056 read_unlock(&obd->obd_pool_lock);
3061 static int target_send_reply_msg(struct ptlrpc_request *req,
3062 int rc, int fail_id)
3064 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3065 DEBUG_REQ(D_ERROR, req, "dropping reply");
3069 * We can have a null rq_reqmsg in the event of bad signature or
3070 * no context when unwrapping
3072 if (req->rq_reqmsg &&
3073 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3074 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3078 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3079 req->rq_status = rc;
3080 return ptlrpc_send_error(req, 1);
3082 DEBUG_REQ(D_NET, req, "sending reply");
3084 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3087 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3089 struct ptlrpc_service_part *svcpt;
3091 struct ptlrpc_reply_state *rs;
3092 struct obd_export *exp;
3096 if (req->rq_no_reply) {
3101 svcpt = req->rq_rqbd->rqbd_svcpt;
3102 rs = req->rq_reply_state;
3103 if (rs == NULL || !rs->rs_difficult) {
3105 target_send_reply_msg(req, rc, fail_id);
3110 /* must be an export if locks saved */
3111 LASSERT(req->rq_export != NULL);
3112 /* req/reply consistent */
3113 LASSERT(rs->rs_svcpt == svcpt);
3116 LASSERT(!rs->rs_scheduled);
3117 LASSERT(!rs->rs_scheduled_ever);
3118 LASSERT(!rs->rs_handled);
3119 LASSERT(!rs->rs_on_net);
3120 LASSERT(rs->rs_export == NULL);
3121 LASSERT(list_empty(&rs->rs_obd_list));
3122 LASSERT(list_empty(&rs->rs_exp_list));
3124 exp = class_export_get(req->rq_export);
3126 /* disable reply scheduling while I'm setting up */
3127 rs->rs_scheduled = 1;
3129 rs->rs_xid = req->rq_xid;
3130 rs->rs_transno = req->rq_transno;
3131 rs->rs_export = exp;
3132 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3134 spin_lock(&exp->exp_uncommitted_replies_lock);
3135 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3136 rs->rs_transno, exp->exp_last_committed);
3137 if (rs->rs_transno > exp->exp_last_committed) {
3138 /* not committed already */
3139 list_add_tail(&rs->rs_obd_list,
3140 &exp->exp_uncommitted_replies);
3142 spin_unlock(&exp->exp_uncommitted_replies_lock);
3144 spin_lock(&exp->exp_lock);
3145 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3146 spin_unlock(&exp->exp_lock);
3148 netrc = target_send_reply_msg(req, rc, fail_id);
3150 spin_lock(&svcpt->scp_rep_lock);
3152 atomic_inc(&svcpt->scp_nreps_difficult);
3156 * error sending: reply is off the net. Also we need +1
3157 * reply ref until ptlrpc_handle_rs() is done
3158 * with the reply state (if the send was successful, there
3159 * would have been +1 ref for the net, which
3160 * reply_out_callback leaves alone)
3163 ptlrpc_rs_addref(rs);
3166 spin_lock(&rs->rs_lock);
3167 if (rs->rs_transno <= exp->exp_last_committed ||
3168 (!rs->rs_on_net && !rs->rs_no_ack) ||
3169 list_empty(&rs->rs_exp_list) || /* completed already */
3170 list_empty(&rs->rs_obd_list)) {
3171 CDEBUG(D_HA, "Schedule reply immediately\n");
3172 ptlrpc_dispatch_difficult_reply(rs);
3174 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3175 rs->rs_scheduled = 0; /* allow notifier to schedule */
3177 spin_unlock(&rs->rs_lock);
3178 spin_unlock(&svcpt->scp_rep_lock);
3182 enum ldlm_mode lck_compat_array[] = {
3183 [LCK_EX] = LCK_COMPAT_EX,
3184 [LCK_PW] = LCK_COMPAT_PW,
3185 [LCK_PR] = LCK_COMPAT_PR,
3186 [LCK_CW] = LCK_COMPAT_CW,
3187 [LCK_CR] = LCK_COMPAT_CR,
3188 [LCK_NL] = LCK_COMPAT_NL,
3189 [LCK_GROUP] = LCK_COMPAT_GROUP,
3190 [LCK_COS] = LCK_COMPAT_COS,
3194 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3195 * not escape to the user level.
3197 int ldlm_error2errno(enum ldlm_error error)
3203 case ELDLM_LOCK_MATCHED:
3206 case ELDLM_LOCK_CHANGED:
3209 case ELDLM_LOCK_ABORTED:
3212 case ELDLM_LOCK_REPLACED:
3215 case ELDLM_NO_LOCK_DATA:
3218 case ELDLM_NAMESPACE_EXISTS:
3221 case ELDLM_BAD_NAMESPACE:
3225 if (((int)error) < 0) { /* cast to signed type */
3226 result = error; /* as ldlm_error can be unsigned */
3228 CERROR("Invalid DLM result code: %d\n", error);
3234 EXPORT_SYMBOL(ldlm_error2errno);
3237 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3239 enum ldlm_error ldlm_errno2error(int err_no)
3248 error = ELDLM_LOCK_CHANGED;
3251 error = ELDLM_LOCK_ABORTED;
3254 error = ELDLM_LOCK_REPLACED;
3257 error = ELDLM_NO_LOCK_DATA;
3260 error = ELDLM_NAMESPACE_EXISTS;
3263 error = ELDLM_BAD_NAMESPACE;
3271 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3272 void ldlm_dump_export_locks(struct obd_export *exp)
3274 spin_lock(&exp->exp_locks_list_guard);
3275 if (!list_empty(&exp->exp_locks_list)) {
3276 struct ldlm_lock *lock;
3278 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3280 list_for_each_entry(lock, &exp->exp_locks_list,
3282 LDLM_ERROR(lock, "lock:");
3284 spin_unlock(&exp->exp_locks_list_guard);
3288 #ifdef HAVE_SERVER_SUPPORT
3289 static int target_bulk_timeout(void *data)
3293 * We don't fail the connection here, because having the export
3294 * killed makes the (vital) call to commitrw very sad.
3299 static inline const char *bulk2type(struct ptlrpc_request *req)
3301 if (req->rq_bulk_read)
3303 if (req->rq_bulk_write)
3308 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3309 struct l_wait_info *lwi)
3311 struct ptlrpc_request *req = desc->bd_req;
3312 time64_t start = ktime_get_real_seconds();
3318 /* If there is eviction in progress, wait for it to finish. */
3319 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3320 *lwi = LWI_INTR(NULL, NULL);
3321 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3322 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
3326 /* Check if client was evicted or reconnected already. */
3327 if (exp->exp_failed ||
3328 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3331 if (req->rq_bulk_read)
3332 rc = sptlrpc_svc_wrap_bulk(req, desc);
3334 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3335 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3336 else /* old version, bulk matchbits is rq_xid */
3337 req->rq_mbits = req->rq_xid;
3340 rc = ptlrpc_start_bulk_transfer(desc);
3344 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3345 bulk2type(req), rc);
3349 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3350 ptlrpc_abort_bulk(desc);
3354 /* limit actual bulk transfer to bulk_timeout seconds */
3355 deadline = start + bulk_timeout;
3356 if (deadline > req->rq_deadline)
3357 deadline = req->rq_deadline;
3360 time64_t timeoutl = deadline - ktime_get_real_seconds();
3361 long timeout_jiffies = timeoutl <= 0 ?
3362 1 : cfs_time_seconds(timeoutl);
3363 time64_t rq_deadline;
3365 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3366 cfs_time_seconds(1),
3367 target_bulk_timeout, desc);
3368 rc = l_wait_event(desc->bd_waitq,
3369 !ptlrpc_server_bulk_active(desc) ||
3372 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3374 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3375 /* Wait again if we changed rq_deadline. */
3376 rq_deadline = READ_ONCE(req->rq_deadline);
3377 deadline = start + bulk_timeout;
3378 if (deadline > rq_deadline)
3379 deadline = rq_deadline;
3380 } while (rc == -ETIMEDOUT &&
3381 deadline > ktime_get_real_seconds());
3383 if (rc == -ETIMEDOUT) {
3384 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3385 bulk2type(req), deadline - start,
3386 ktime_get_real_seconds() - deadline);
3387 ptlrpc_abort_bulk(desc);
3388 } else if (exp->exp_failed) {
3389 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3392 ptlrpc_abort_bulk(desc);
3393 } else if (exp->exp_conn_cnt >
3394 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3395 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3397 /* We don't reply anyway. */
3399 ptlrpc_abort_bulk(desc);
3400 } else if (desc->bd_failure) {
3401 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3403 /* XXX should this be a different errno? */
3406 if (req->rq_bulk_write)
3407 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3408 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3409 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3410 bulk2type(req), desc->bd_nob_transferred,
3412 /* XXX should this be a different errno? */
3419 EXPORT_SYMBOL(target_bulk_io);
3421 #endif /* HAVE_SERVER_SUPPORT */