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_move(&item->oic_item,
95 item->oic_last_attempt = 0;
97 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
98 imp, imp->imp_obd->obd_name, uuid->uuid,
99 (priority ? ", moved to head" : ""));
100 spin_unlock(&imp->imp_lock);
101 GOTO(out_free, rc = 0);
104 /* No existing import connection found for \a uuid. */
106 imp_conn->oic_conn = ptlrpc_conn;
107 imp_conn->oic_uuid = *uuid;
108 imp_conn->oic_last_attempt = 0;
110 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
112 list_add_tail(&imp_conn->oic_item,
113 &imp->imp_conn_list);
114 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
115 imp, imp->imp_obd->obd_name, uuid->uuid,
116 (priority ? "head" : "tail"));
118 spin_unlock(&imp->imp_lock);
119 GOTO(out_free, rc = -ENOENT);
122 spin_unlock(&imp->imp_lock);
126 OBD_FREE(imp_conn, sizeof(*imp_conn));
128 ptlrpc_connection_put(ptlrpc_conn);
132 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
134 return import_set_conn(imp, uuid, 1, 0);
137 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
140 return import_set_conn(imp, uuid, priority, 1);
142 EXPORT_SYMBOL(client_import_add_conn);
144 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
146 struct obd_import_conn *imp_conn;
147 struct obd_export *dlmexp;
152 spin_lock(&imp->imp_lock);
153 if (list_empty(&imp->imp_conn_list)) {
154 LASSERT(!imp->imp_connection);
158 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
159 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
161 LASSERT(imp_conn->oic_conn);
163 if (imp_conn == imp->imp_conn_current) {
164 LASSERT(imp_conn->oic_conn == imp->imp_connection);
166 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
167 imp->imp_state != LUSTRE_IMP_DISCON) {
168 CERROR("can't remove current connection\n");
169 GOTO(out, rc = -EBUSY);
172 ptlrpc_connection_put(imp->imp_connection);
173 imp->imp_connection = NULL;
175 dlmexp = class_conn2export(&imp->imp_dlm_handle);
176 if (dlmexp && dlmexp->exp_connection) {
177 LASSERT(dlmexp->exp_connection ==
179 ptlrpc_connection_put(dlmexp->exp_connection);
180 dlmexp->exp_connection = NULL;
184 class_export_put(dlmexp);
187 list_del(&imp_conn->oic_item);
188 ptlrpc_connection_put(imp_conn->oic_conn);
189 OBD_FREE(imp_conn, sizeof(*imp_conn));
190 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
191 imp, imp->imp_obd->obd_name, uuid->uuid);
196 spin_unlock(&imp->imp_lock);
198 CERROR("connection %s not found\n", uuid->uuid);
201 EXPORT_SYMBOL(client_import_del_conn);
204 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
205 * to find a conn uuid of \a imp which can reach \a peer.
207 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
208 struct obd_uuid *uuid)
210 struct obd_import_conn *conn;
215 spin_lock(&imp->imp_lock);
216 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
217 /* Check if conn UUID does have this peer NID. */
218 if (class_check_uuid(&conn->oic_uuid, peer)) {
219 *uuid = conn->oic_uuid;
224 spin_unlock(&imp->imp_lock);
227 EXPORT_SYMBOL(client_import_find_conn);
229 void client_destroy_import(struct obd_import *imp)
232 * Drop security policy instance after all RPCs have finished/aborted
233 * to let all busy contexts be released.
235 class_import_get(imp);
236 class_destroy_import(imp);
237 sptlrpc_import_sec_put(imp);
238 class_import_put(imp);
240 EXPORT_SYMBOL(client_destroy_import);
243 * Check whether or not the OSC is on MDT.
246 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
248 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
251 static int osc_on_mdt(char *obdname)
255 ptr = strrchr(obdname, '-');
259 if (strncmp(ptr + 1, "MDT", 3) == 0)
266 * Configure an RPC client OBD device.
271 * 3 - inactive-on-startup
272 * 4 - restrictive net
274 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
276 struct client_obd *cli = &obddev->u.cli;
277 struct obd_import *imp;
278 struct obd_uuid server_uuid;
279 int rq_portal, rp_portal, connect_op;
280 const char *name = obddev->obd_type->typ_name;
281 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
282 char *cli_name = lustre_cfg_buf(lcfg, 0);
283 struct ptlrpc_connection fake_conn = { .c_self = 0,
284 .c_remote_uuid.uuid[0] = 0 };
290 * In a more perfect world, we would hang a ptlrpc_client off of
291 * obd_type and just use the values from there.
293 if (!strcmp(name, LUSTRE_OSC_NAME)) {
294 rq_portal = OST_REQUEST_PORTAL;
295 rp_portal = OSC_REPLY_PORTAL;
296 connect_op = OST_CONNECT;
297 cli->cl_sp_me = LUSTRE_SP_CLI;
298 cli->cl_sp_to = LUSTRE_SP_OST;
299 ns_type = LDLM_NS_TYPE_OSC;
300 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
301 !strcmp(name, LUSTRE_LWP_NAME)) {
302 rq_portal = MDS_REQUEST_PORTAL;
303 rp_portal = MDC_REPLY_PORTAL;
304 connect_op = MDS_CONNECT;
305 if (is_lwp_on_ost(cli_name))
306 cli->cl_sp_me = LUSTRE_SP_OST;
307 else if (is_lwp_on_mdt(cli_name))
308 cli->cl_sp_me = LUSTRE_SP_MDT;
310 cli->cl_sp_me = LUSTRE_SP_CLI;
311 cli->cl_sp_to = LUSTRE_SP_MDT;
312 ns_type = LDLM_NS_TYPE_MDC;
313 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
314 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
315 /* OSP_on_MDT for other MDTs */
316 connect_op = MDS_CONNECT;
317 cli->cl_sp_to = LUSTRE_SP_MDT;
318 ns_type = LDLM_NS_TYPE_MDC;
319 rq_portal = OUT_PORTAL;
321 /* OSP on MDT for OST */
322 connect_op = OST_CONNECT;
323 cli->cl_sp_to = LUSTRE_SP_OST;
324 ns_type = LDLM_NS_TYPE_OSC;
325 rq_portal = OST_REQUEST_PORTAL;
327 rp_portal = OSC_REPLY_PORTAL;
328 cli->cl_sp_me = LUSTRE_SP_MDT;
329 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
330 rq_portal = MGS_REQUEST_PORTAL;
331 rp_portal = MGC_REPLY_PORTAL;
332 connect_op = MGS_CONNECT;
333 cli->cl_sp_me = LUSTRE_SP_MGC;
334 cli->cl_sp_to = LUSTRE_SP_MGS;
335 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
336 ns_type = LDLM_NS_TYPE_MGC;
338 CERROR("unknown client OBD type \"%s\", can't setup\n",
343 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
344 CERROR("requires a TARGET UUID\n");
348 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
349 CERROR("client UUID must be less than 38 characters\n");
353 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
354 CERROR("setup requires a SERVER UUID\n");
358 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
359 CERROR("target UUID must be less than 38 characters\n");
363 init_rwsem(&cli->cl_sem);
364 mutex_init(&cli->cl_mgc_mutex);
366 init_rwsem(&cli->cl_seq_rwsem);
367 cli->cl_conn_count = 0;
368 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
369 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
370 sizeof(server_uuid)));
372 cli->cl_dirty_pages = 0;
373 cli->cl_dirty_max_pages = 0;
374 cli->cl_avail_grant = 0;
375 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
377 * cl_dirty_max_pages may be changed at connect time in
378 * ptlrpc_connect_interpret().
380 client_adjust_max_dirty(cli);
381 INIT_LIST_HEAD(&cli->cl_cache_waiters);
382 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
383 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
384 INIT_LIST_HEAD(&cli->cl_loi_write_list);
385 INIT_LIST_HEAD(&cli->cl_loi_read_list);
386 spin_lock_init(&cli->cl_loi_list_lock);
387 atomic_set(&cli->cl_pending_w_pages, 0);
388 atomic_set(&cli->cl_pending_r_pages, 0);
389 cli->cl_r_in_flight = 0;
390 cli->cl_w_in_flight = 0;
392 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
393 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
394 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
395 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
396 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
397 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
400 INIT_LIST_HEAD(&cli->cl_lru_osc);
401 atomic_set(&cli->cl_lru_shrinkers, 0);
402 atomic_long_set(&cli->cl_lru_busy, 0);
403 atomic_long_set(&cli->cl_lru_in_list, 0);
404 INIT_LIST_HEAD(&cli->cl_lru_list);
405 spin_lock_init(&cli->cl_lru_list_lock);
406 atomic_long_set(&cli->cl_unstable_count, 0);
407 INIT_LIST_HEAD(&cli->cl_shrink_list);
408 INIT_LIST_HEAD(&cli->cl_grant_chain);
410 INIT_LIST_HEAD(&cli->cl_flight_waiters);
411 cli->cl_rpcs_in_flight = 0;
413 init_waitqueue_head(&cli->cl_destroy_waitq);
414 atomic_set(&cli->cl_destroy_in_flight, 0);
417 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
418 cli->cl_preferred_cksum_type = 0;
419 #ifdef ENABLE_CHECKSUM
420 /* Turn on checksumming by default. */
421 cli->cl_checksum = 1;
423 * The supported checksum types will be worked out at connect time
424 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
427 cli->cl_cksum_type = cli->cl_supp_cksum_types;
429 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
432 * Set it to possible maximum size. It may be reduced by ocd_brw_size
433 * from OFD after connecting.
435 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
437 cli->cl_max_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
440 * set cl_chunkbits default value to PAGE_SHIFT,
441 * it will be updated at OSC connection time.
443 cli->cl_chunkbits = PAGE_SHIFT;
445 if (!strcmp(name, LUSTRE_MDC_NAME)) {
446 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
447 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
448 cli->cl_max_rpcs_in_flight = 2;
449 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
450 cli->cl_max_rpcs_in_flight = 3;
451 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
452 cli->cl_max_rpcs_in_flight = 4;
454 if (osc_on_mdt(obddev->obd_name))
455 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
457 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
460 spin_lock_init(&cli->cl_mod_rpcs_lock);
461 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
462 cli->cl_max_mod_rpcs_in_flight = 0;
463 cli->cl_mod_rpcs_in_flight = 0;
464 cli->cl_close_rpcs_in_flight = 0;
465 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
466 cli->cl_mod_tag_bitmap = NULL;
468 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
470 if (connect_op == MDS_CONNECT) {
471 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
472 OBD_ALLOC(cli->cl_mod_tag_bitmap,
473 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
474 if (cli->cl_mod_tag_bitmap == NULL)
475 GOTO(err, rc = -ENOMEM);
480 CERROR("ldlm_get_ref failed: %d\n", rc);
484 ptlrpc_init_client(rq_portal, rp_portal, name,
485 &obddev->obd_ldlm_client);
487 imp = class_new_import(obddev);
489 GOTO(err_ldlm, rc = -ENOENT);
490 imp->imp_client = &obddev->obd_ldlm_client;
491 imp->imp_connect_op = connect_op;
492 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
493 LUSTRE_CFG_BUFLEN(lcfg, 1));
494 class_import_put(imp);
496 if (lustre_cfg_buf(lcfg, 4)) {
497 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
499 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
501 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
502 obddev->obd_name, lustre_cfg_string(lcfg, 4),
504 GOTO(err_import, rc);
506 fake_conn.c_self = LNET_MKNID(refnet, 0);
507 imp->imp_connection = &fake_conn;
510 rc = client_import_add_conn(imp, &server_uuid, 1);
512 CERROR("can't add initial connection\n");
513 GOTO(err_import, rc);
515 imp->imp_connection = NULL;
517 cli->cl_import = imp;
518 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
519 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
521 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
522 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
523 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
524 name, obddev->obd_name,
525 cli->cl_target_uuid.uuid);
526 spin_lock(&imp->imp_lock);
527 imp->imp_deactive = 1;
528 spin_unlock(&imp->imp_lock);
532 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
533 LDLM_NAMESPACE_CLIENT,
534 LDLM_NAMESPACE_GREEDY,
536 if (obddev->obd_namespace == NULL) {
537 CERROR("Unable to create client namespace - %s\n",
539 GOTO(err_import, rc = -ENOMEM);
545 class_destroy_import(imp);
549 if (cli->cl_mod_tag_bitmap != NULL)
550 OBD_FREE(cli->cl_mod_tag_bitmap,
551 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
552 cli->cl_mod_tag_bitmap = NULL;
556 EXPORT_SYMBOL(client_obd_setup);
558 int client_obd_cleanup(struct obd_device *obddev)
560 struct client_obd *cli = &obddev->u.cli;
564 ldlm_namespace_free_post(obddev->obd_namespace);
565 obddev->obd_namespace = NULL;
567 obd_cleanup_client_import(obddev);
568 LASSERT(obddev->u.cli.cl_import == NULL);
572 if (cli->cl_mod_tag_bitmap != NULL)
573 OBD_FREE(cli->cl_mod_tag_bitmap,
574 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
575 cli->cl_mod_tag_bitmap = NULL;
579 EXPORT_SYMBOL(client_obd_cleanup);
581 /* ->o_connect() method for client side (OSC and MDC and MGC) */
582 int client_connect_import(const struct lu_env *env,
583 struct obd_export **exp,
584 struct obd_device *obd, struct obd_uuid *cluuid,
585 struct obd_connect_data *data, void *localdata)
587 struct client_obd *cli = &obd->u.cli;
588 struct obd_import *imp = cli->cl_import;
589 struct obd_connect_data *ocd;
590 struct lustre_handle conn = { 0 };
596 down_write(&cli->cl_sem);
597 if (cli->cl_conn_count > 0)
598 GOTO(out_sem, rc = -EALREADY);
600 rc = class_connect(&conn, obd, cluuid);
604 cli->cl_conn_count++;
605 *exp = class_conn2export(&conn);
607 LASSERT(obd->obd_namespace);
609 imp->imp_dlm_handle = conn;
610 rc = ptlrpc_init_import(imp);
614 ocd = &imp->imp_connect_data;
617 imp->imp_connect_flags_orig = data->ocd_connect_flags;
618 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
621 rc = ptlrpc_connect_import(imp);
623 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
626 LASSERT(*exp != NULL && (*exp)->exp_connection);
629 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
630 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
631 data->ocd_connect_flags, ocd->ocd_connect_flags);
632 data->ocd_connect_flags = ocd->ocd_connect_flags;
633 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
636 ptlrpc_pinger_add_import(imp);
642 cli->cl_conn_count--;
643 class_disconnect(*exp);
647 up_write(&cli->cl_sem);
649 if (!rc && localdata) {
650 LASSERT(cli->cl_cache == NULL); /* only once */
651 cli->cl_cache = (struct cl_client_cache *)localdata;
652 cl_cache_incref(cli->cl_cache);
653 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
655 /* add this osc into entity list */
656 LASSERT(list_empty(&cli->cl_lru_osc));
657 spin_lock(&cli->cl_cache->ccc_lru_lock);
658 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
659 spin_unlock(&cli->cl_cache->ccc_lru_lock);
664 EXPORT_SYMBOL(client_connect_import);
666 int client_disconnect_export(struct obd_export *exp)
668 struct obd_device *obd = class_exp2obd(exp);
669 struct client_obd *cli;
670 struct obd_import *imp;
676 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
677 exp, exp ? exp->exp_handle.h_cookie : -1);
682 imp = cli->cl_import;
684 down_write(&cli->cl_sem);
685 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
688 if (cli->cl_conn_count == 0) {
689 CERROR("disconnecting disconnected device (%s)\n",
691 GOTO(out_disconnect, rc = -EINVAL);
694 cli->cl_conn_count--;
695 if (cli->cl_conn_count != 0)
696 GOTO(out_disconnect, rc = 0);
699 * Mark import deactivated now, so we don't try to reconnect if any
700 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
701 * fully deactivate the import, or that would drop all requests.
703 spin_lock(&imp->imp_lock);
704 imp->imp_deactive = 1;
705 spin_unlock(&imp->imp_lock);
708 * Some non-replayable imports (MDS's OSCs) are pinged, so just
709 * delete it regardless. (It's safe to delete an import that was
712 (void)ptlrpc_pinger_del_import(imp);
714 if (obd->obd_namespace != NULL) {
715 /* obd_force == local only */
716 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
717 obd->obd_force ? LCF_LOCAL : 0, NULL);
718 ldlm_namespace_free_prior(obd->obd_namespace, imp,
723 * There's no need to hold sem while disconnecting an import,
724 * and it may actually cause deadlock in GSS.
726 up_write(&cli->cl_sem);
727 rc = ptlrpc_disconnect_import(imp, 0);
728 down_write(&cli->cl_sem);
730 ptlrpc_invalidate_import(imp);
736 * Use server style - class_disconnect should be always called for
739 err = class_disconnect(exp);
743 up_write(&cli->cl_sem);
747 EXPORT_SYMBOL(client_disconnect_export);
749 #ifdef HAVE_SERVER_SUPPORT
750 int server_disconnect_export(struct obd_export *exp)
756 /* Disconnect early so that clients can't keep using export. */
757 rc = class_disconnect(exp);
758 /* Close import to avoid sending any requests. */
759 if (exp->exp_imp_reverse)
760 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
762 ldlm_bl_thread_wakeup();
764 /* complete all outstanding replies */
765 spin_lock(&exp->exp_lock);
766 while (!list_empty(&exp->exp_outstanding_replies)) {
767 struct ptlrpc_reply_state *rs =
768 list_entry(exp->exp_outstanding_replies.next,
769 struct ptlrpc_reply_state, rs_exp_list);
770 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
772 spin_lock(&svcpt->scp_rep_lock);
774 list_del_init(&rs->rs_exp_list);
776 spin_lock(&rs->rs_lock);
777 /* clear rs_convert_lock to make sure rs is handled and put */
778 rs->rs_convert_lock = 0;
779 ptlrpc_schedule_difficult_reply(rs);
780 spin_unlock(&rs->rs_lock);
782 spin_unlock(&svcpt->scp_rep_lock);
784 spin_unlock(&exp->exp_lock);
788 EXPORT_SYMBOL(server_disconnect_export);
791 * --------------------------------------------------------------------------
792 * from old lib/target.c
793 * --------------------------------------------------------------------------
796 static int target_handle_reconnect(struct lustre_handle *conn,
797 struct obd_export *exp,
798 struct obd_uuid *cluuid)
800 struct obd_device *target;
801 struct lustre_handle *hdl;
807 hdl = &exp->exp_imp_reverse->imp_remote_handle;
808 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
809 conn->cookie = exp->exp_handle.h_cookie;
811 "connect export for UUID '%s' at %p, cookie %#llx\n",
812 cluuid->uuid, exp, conn->cookie);
816 target = exp->exp_obd;
818 /* Might be a re-connect after a partition. */
819 if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
820 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",
822 obd_uuid2str(&exp->exp_client_uuid),
823 obd_export_nid2str(exp),
824 hdl->cookie, conn->cookie);
825 memset(conn, 0, sizeof(*conn));
827 * target_handle_connect() treats EALREADY and
828 * -EALREADY differently. -EALREADY is an error
829 * (same UUID, different handle).
834 if (!target->obd_recovering) {
835 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
836 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
837 obd_export_nid2str(exp));
838 GOTO(out_already, rc);
841 remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
842 timeout = ktime_divns(remaining, NSEC_PER_SEC);
844 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
846 obd_uuid2str(&exp->exp_client_uuid),
847 obd_export_nid2str(exp),
848 target->obd_max_recoverable_clients,
849 timeout / 60, timeout % 60);
851 struct target_distribute_txn_data *tdtd;
856 tdtd = class_exp2tgt(exp)->lut_tdtd;
857 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
858 buf = tdtd->tdtd_show_update_logs_retrievers(
859 tdtd->tdtd_show_retrievers_cbdata,
863 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",
865 obd_uuid2str(&exp->exp_client_uuid),
866 obd_export_nid2str(exp), count,
867 buf ? buf : "unknown (not enough RAM)",
868 (abs(timeout) + target->obd_recovery_timeout) / 60,
869 (abs(timeout) + target->obd_recovery_timeout) % 60,
872 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",
873 target->obd_name, abs(timeout) / 60,
874 abs(timeout) % 60, target->obd_name);
881 conn->cookie = exp->exp_handle.h_cookie;
883 * target_handle_connect() treats EALREADY and
884 * -EALREADY differently. EALREADY means we are
885 * doing a valid reconnect from the same client.
891 check_and_start_recovery_timer(struct obd_device *obd,
892 struct ptlrpc_request *req, int new_client);
895 * update flags for import during reconnect process
897 static int rev_import_flags_update(struct obd_import *revimp,
898 struct ptlrpc_request *req)
901 struct obd_connect_data *data;
903 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
905 if (data->ocd_connect_flags & OBD_CONNECT_AT)
906 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
908 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
910 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
912 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
914 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
915 revimp->imp_client->cli_name,
916 libcfs_id2str(req->rq_peer), rc);
924 * Allocate a new reverse import for an export.
926 * \retval -errno in case error hit
927 * \retval 0 if reverse import correctly init
929 int rev_import_init(struct obd_export *export)
931 struct obd_device *obd = export->exp_obd;
932 struct obd_import *revimp;
934 LASSERT(export->exp_imp_reverse == NULL);
936 revimp = class_new_import(obd);
940 revimp->imp_remote_handle.cookie = 0ULL;
941 revimp->imp_client = &obd->obd_ldlm_client;
942 revimp->imp_dlm_fake = 1;
944 /* it is safe to connect import in new state as no sends possible */
945 spin_lock(&export->exp_lock);
946 export->exp_imp_reverse = revimp;
947 spin_unlock(&export->exp_lock);
948 class_import_put(revimp);
952 EXPORT_SYMBOL(rev_import_init);
955 * Handle reconnect for an export.
957 * \param exp export to handle reconnect process
958 * \param req client reconnect request
960 * \retval -rc in case securitfy flavor can't be changed
961 * \retval 0 in case none problems
963 static int rev_import_reconnect(struct obd_export *exp,
964 struct ptlrpc_request *req)
966 struct obd_import *revimp = exp->exp_imp_reverse;
967 struct lustre_handle *lh;
970 /* avoid sending a request until import flags are changed */
971 ptlrpc_import_enter_resend(revimp);
973 if (revimp->imp_connection != NULL)
974 ptlrpc_connection_put(revimp->imp_connection);
977 * client from recovery don't have a handle so we need to take from
978 * request. it may produce situation when wrong client connected
979 * to recovery as we trust a client uuid
981 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
982 revimp->imp_remote_handle = *lh;
985 * unknown versions will be caught in
986 * ptlrpc_handle_server_req_in->lustre_unpack_msg()
988 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
990 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
992 rc = rev_import_flags_update(revimp, req);
995 * it is safe to still be in RECOVERY phase as we are not able
996 * to setup correct security flavor so requests are not able to
997 * be delivered correctly
1002 /* resend all rpc's via new connection */
1003 return ptlrpc_import_recovery_state_machine(revimp);
1006 int target_handle_connect(struct ptlrpc_request *req)
1008 struct obd_device *target = NULL;
1009 struct obd_export *export = NULL;
1011 * connect handle - filled from target_handle_reconnect in
1014 struct lustre_handle conn;
1015 struct lustre_handle *tmp;
1016 struct obd_uuid cluuid;
1021 bool mds_conn = false, lw_client = false, initial_conn = false;
1022 bool mds_mds_conn = false;
1023 bool new_mds_mds_conn = false;
1024 struct obd_connect_data *data, *tmpdata;
1026 lnet_nid_t *client_nid = NULL;
1030 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1032 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1034 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1035 GOTO(out, rc = -EINVAL);
1038 target = class_dev_by_str(str);
1040 deuuidify(str, NULL, &target_start, &target_len);
1041 LCONSOLE_ERROR_MSG(0x137,
1042 "%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",
1043 str, libcfs_nid2str(req->rq_peer.nid));
1044 GOTO(out, rc = -ENODEV);
1047 spin_lock(&target->obd_dev_lock);
1049 target->obd_conn_inprogress++;
1051 if (target->obd_stopping || !target->obd_set_up) {
1052 spin_unlock(&target->obd_dev_lock);
1054 deuuidify(str, NULL, &target_start, &target_len);
1055 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1056 target_len, target_start,
1057 libcfs_nid2str(req->rq_peer.nid),
1058 (target->obd_stopping ?
1059 "stopping" : "not set up"));
1060 GOTO(out, rc = -ENODEV);
1063 if (target->obd_no_conn) {
1064 spin_unlock(&target->obd_dev_lock);
1067 "%s: Temporarily refusing client connection from %s\n",
1068 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1069 GOTO(out, rc = -EAGAIN);
1072 spin_unlock(&target->obd_dev_lock);
1074 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1076 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1077 GOTO(out, rc = -EINVAL);
1080 obd_str2uuid(&cluuid, str);
1082 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1084 GOTO(out, rc = -EPROTO);
1088 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1090 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1092 GOTO(out, rc = -EPROTO);
1094 rc = req_capsule_server_pack(&req->rq_pill);
1098 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1100 * Don't allow clients to connect that are using old 1.8 format
1101 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1102 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1103 * FULL20 flag should be set on all connections since 2.0, but no
1104 * longer affects behaviour.
1106 * Later this check will be disabled and the flag can be retired
1107 * completely once interop with 3.0 is no longer needed.
1109 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1110 GOTO(out, rc = -EPROTO);
1113 * Don't allow liblustre clients to connect.
1114 * - testing was disabled in v2_2_50_0-61-g6a75d65
1115 * - building was disabled in v2_5_58_0-28-g7277179
1116 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1117 * - clients were refused connect for version difference > 0.0.1.32
1119 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1120 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1121 GOTO(out, rc = -EPROTO);
1126 * Note: lw_client is needed in MDS-MDS failover during update log
1127 * processing, so we needs to allow lw_client to be connected at
1128 * anytime, instead of only the initial connection
1130 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1132 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1133 initial_conn = true;
1134 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1135 mds_mds_conn = (data->ocd_connect_flags &
1136 OBD_CONNECT_MDS_MDS) != 0;
1139 * OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1140 * for Imperative Recovery connection from MGC to MGS.
1142 * Via check OBD_CONNECT_FID, we can distinguish whether
1143 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1144 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1147 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1148 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1149 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1150 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1151 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1152 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1155 * We do not support the MDT-MDT interoperations with
1156 * different version MDT because of protocol changes.
1158 if (unlikely(major != LUSTRE_MAJOR ||
1159 minor != LUSTRE_MINOR ||
1160 abs(patch - LUSTRE_PATCH) > 3)) {
1161 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1162 target->obd_name, LUSTRE_MAJOR,
1163 LUSTRE_MINOR, LUSTRE_PATCH,
1164 LUSTRE_FIX, major, minor, patch,
1165 OBD_OCD_VERSION_FIX(data->ocd_version),
1166 libcfs_nid2str(req->rq_peer.nid),
1168 GOTO(out, rc = -EPROTO);
1173 /* lctl gets a backstage, all-access pass. */
1174 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1175 goto dont_check_exports;
1177 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1181 /* We've found an export in the hash. */
1183 spin_lock(&export->exp_lock);
1185 if (export->exp_connecting) { /* b=9635, et. al. */
1186 spin_unlock(&export->exp_lock);
1187 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1188 export->exp_obd->obd_name, export,
1189 libcfs_nid2str(req->rq_peer.nid));
1190 class_export_put(export);
1193 } else if ((mds_conn || (lw_client && initial_conn) ||
1194 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1195 export->exp_connection != NULL) {
1196 spin_unlock(&export->exp_lock);
1197 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1198 /* MDS or LWP reconnected after failover. */
1199 LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1201 mds_conn ? "MDS" : "LWP",
1202 libcfs_nid2str(req->rq_peer.nid),
1203 libcfs_nid2str(export->exp_connection->c_peer.nid));
1205 /* New MDS connection from the same NID. */
1206 LCONSOLE_WARN("%s: Received new %s connection from %s, removing former export from same NID\n",
1208 mds_conn ? "MDS" : "LWP",
1209 libcfs_nid2str(req->rq_peer.nid));
1212 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1213 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1215 * Because exports between MDTs will always be
1216 * kept, let's do not fail such export if they
1217 * come from the same NID, otherwise it might
1218 * cause eviction between MDTs, which might
1219 * cause namespace inconsistency
1221 spin_lock(&export->exp_lock);
1222 export->exp_connecting = 1;
1223 export->exp_conn_cnt = 0;
1224 spin_unlock(&export->exp_lock);
1225 conn.cookie = export->exp_handle.h_cookie;
1228 class_fail_export(export);
1229 class_export_put(export);
1233 } else if (export->exp_connection != NULL && initial_conn &&
1234 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1235 spin_unlock(&export->exp_lock);
1236 /* In MDS failover we have static UUID but NID can change. */
1237 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1238 target->obd_name, cluuid.uuid,
1239 libcfs_nid2str(req->rq_peer.nid),
1241 export->exp_connection->c_peer.nid));
1243 class_export_put(export);
1246 export->exp_connecting = 1;
1247 spin_unlock(&export->exp_lock);
1248 LASSERT(export->exp_obd == target);
1250 rc = target_handle_reconnect(&conn, export, &cluuid);
1253 /* If we found an export, we already unlocked. */
1256 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1257 } else if (req->rq_export == NULL &&
1258 atomic_read(&export->exp_rpc_count) > 0) {
1259 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1260 target->obd_name, cluuid.uuid,
1261 libcfs_nid2str(req->rq_peer.nid),
1262 atomic_read(&export->exp_refcount));
1263 GOTO(out, rc = -EBUSY);
1264 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1266 if (!strstr(cluuid.uuid, "mdt"))
1267 LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client",
1268 target->obd_name, cluuid.uuid,
1269 libcfs_nid2str(req->rq_peer.nid));
1270 GOTO(out, rc = -EALREADY);
1272 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1278 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1279 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1280 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1281 export, ktime_get_real_seconds(),
1282 export ? export->exp_last_request_time : 0);
1285 * If this is the first time a client connects, reset the recovery
1286 * timer. Discard lightweight connections which might be local.
1288 if (!lw_client && rc == 0 && target->obd_recovering)
1289 check_and_start_recovery_timer(target, req, export == NULL);
1292 * We want to handle EALREADY but *not* -EALREADY from
1293 * target_handle_reconnect(), return reconnection state in a flag.
1295 if (rc == EALREADY) {
1296 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1302 /* Tell the client if we support replayable requests. */
1303 if (target->obd_replayable)
1304 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1305 client_nid = &req->rq_peer.nid;
1307 if (export == NULL) {
1308 /* allow lightweight connections during recovery */
1310 * allow "new" MDT to be connected during recovery, since we
1311 * need retrieve recovery update records from it
1313 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1314 struct hrtimer *timer = &target->obd_recovery_timer;
1323 connected = atomic_read(&target->obd_connected_clients);
1324 in_progress = atomic_read(&target->obd_lock_replay_clients);
1325 known = target->obd_max_recoverable_clients;
1326 stale = target->obd_stale_clients;
1327 remaining = hrtimer_expires_remaining(timer);
1328 left = ktime_divns(remaining, NSEC_PER_SEC);
1330 if (ktime_to_ns(remaining) > 0) {
1331 msg = "to recover in";
1334 msg = "already passed deadline";
1338 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",
1339 target->obd_name, cluuid.uuid,
1340 libcfs_nid2str(req->rq_peer.nid), known,
1341 connected - in_progress, in_progress,
1342 stale, msg, timeout / 60, timeout % 60);
1346 rc = obd_connect(req->rq_svc_thread->t_env,
1347 &export, target, &cluuid, data,
1349 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1350 lustre_msg_add_op_flags(req->rq_repmsg,
1351 MSG_CONNECT_RECOVERING);
1353 conn.cookie = export->exp_handle.h_cookie;
1354 rc = rev_import_init(export);
1358 new_mds_mds_conn = true;
1361 rc = obd_reconnect(req->rq_svc_thread->t_env,
1362 export, target, &cluuid, data, client_nid);
1367 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1368 data->ocd_instance = target->u.obt.obt_instance;
1371 * Return only the parts of obd_connect_data that we understand, so the
1372 * client knows that we don't understand the rest.
1375 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1377 tmpdata = req_capsule_server_get(&req->rq_pill,
1380 * Don't use struct assignment here, because the client reply
1381 * buffer may be smaller/larger than the local struct
1384 memcpy(tmpdata, data, min(tmpsize, size));
1388 * If the client and the server are the same node, we will already
1389 * have an export that really points to the client's DLM export,
1390 * because we have a shared handles table.
1392 * XXX this will go away when shaver stops sending the "connect" handle
1393 * in the real "remote handle" field of the request --phik 24 Apr 2003
1395 ptlrpc_request_change_export(req, export);
1397 spin_lock(&export->exp_lock);
1398 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1399 spin_unlock(&export->exp_lock);
1401 "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1402 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1403 export->exp_conn_cnt,
1404 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1406 GOTO(out, rc = -EALREADY);
1408 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1409 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1410 spin_unlock(&export->exp_lock);
1412 if (export->exp_connection != NULL) {
1413 /* Check to see if connection came from another NID. */
1414 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1415 !hlist_unhashed(&export->exp_nid_hash))
1416 cfs_hash_del(export->exp_obd->obd_nid_hash,
1417 &export->exp_connection->c_peer.nid,
1418 &export->exp_nid_hash);
1420 ptlrpc_connection_put(export->exp_connection);
1423 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1426 if (hlist_unhashed(&export->exp_nid_hash))
1427 cfs_hash_add(export->exp_obd->obd_nid_hash,
1428 &export->exp_connection->c_peer.nid,
1429 &export->exp_nid_hash);
1431 lustre_msg_set_handle(req->rq_repmsg, &conn);
1433 rc = rev_import_reconnect(export, req);
1437 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1439 __u64 transno = data->ocd_transno;
1441 spin_lock(&export->exp_lock);
1443 * possible race with class_disconnect_stale_exports,
1444 * export may be already in the eviction process
1446 if (export->exp_failed) {
1447 spin_unlock(&export->exp_lock);
1448 GOTO(out, rc = -ENODEV);
1450 export->exp_in_recovery = 1;
1451 export->exp_req_replay_needed = 1;
1452 export->exp_lock_replay_needed = 1;
1453 spin_unlock(&export->exp_lock);
1455 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1456 MSG_CONNECT_TRANSNO);
1457 if (has_transno && transno == 0)
1458 CWARN("Connect with zero transno!\n");
1460 if (has_transno && transno > 0 &&
1461 transno < target->obd_next_recovery_transno &&
1462 transno > target->obd_last_committed) {
1463 /* Another way is to use cmpxchg() to be lock-free. */
1464 spin_lock(&target->obd_recovery_task_lock);
1465 if (transno < target->obd_next_recovery_transno)
1466 target->obd_next_recovery_transno = transno;
1467 spin_unlock(&target->obd_recovery_task_lock);
1470 atomic_inc(&target->obd_req_replay_clients);
1471 atomic_inc(&target->obd_lock_replay_clients);
1473 * Note: MDS-MDS connection is allowed to be connected during
1474 * recovery, no matter if the exports needs to be recoveried.
1475 * Because we need retrieve updates logs from all other MDTs.
1476 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1477 * also needs to be increased to match other recovery checking
1480 if (new_mds_mds_conn)
1481 target->obd_max_recoverable_clients++;
1482 if (atomic_inc_return(&target->obd_connected_clients) ==
1483 target->obd_max_recoverable_clients)
1484 wake_up(&target->obd_next_transno_waitq);
1487 /* Tell the client we're in recovery, when client is involved in it. */
1488 if (target->obd_recovering && !lw_client)
1489 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1493 spin_lock(&export->exp_lock);
1494 export->exp_connecting = 0;
1495 spin_unlock(&export->exp_lock);
1497 class_export_put(export);
1499 if (target != NULL) {
1500 spin_lock(&target->obd_dev_lock);
1501 target->obd_conn_inprogress--;
1502 spin_unlock(&target->obd_dev_lock);
1503 class_decref(target, "find", current);
1505 req->rq_status = rc;
1509 int target_handle_disconnect(struct ptlrpc_request *req)
1515 rc = req_capsule_server_pack(&req->rq_pill);
1519 /* Keep the rq_export around so we can send the reply. */
1520 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1525 void target_destroy_export(struct obd_export *exp)
1527 struct obd_import *imp = NULL;
1529 * exports created from last_rcvd data, and "fake"
1530 * exports created by lctl don't have an import
1532 spin_lock(&exp->exp_lock);
1533 if (exp->exp_imp_reverse != NULL) {
1534 imp = exp->exp_imp_reverse;
1535 exp->exp_imp_reverse = NULL;
1537 spin_unlock(&exp->exp_lock);
1539 client_destroy_import(imp);
1541 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1542 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1543 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1544 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1546 EXPORT_SYMBOL(target_destroy_export);
1549 * Recovery functions
1551 static void target_request_copy_get(struct ptlrpc_request *req)
1553 class_export_rpc_inc(req->rq_export);
1554 LASSERT(list_empty(&req->rq_list));
1555 INIT_LIST_HEAD(&req->rq_replay_list);
1557 /* Increase refcount to keep request in queue. */
1558 atomic_inc(&req->rq_refcount);
1559 /* Let export know it has replays to be handled. */
1560 atomic_inc(&req->rq_export->exp_replay_count);
1563 static void target_request_copy_put(struct ptlrpc_request *req)
1565 LASSERT(list_empty(&req->rq_replay_list));
1566 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1568 atomic_dec(&req->rq_export->exp_replay_count);
1569 class_export_rpc_dec(req->rq_export);
1570 ptlrpc_server_drop_request(req);
1573 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1575 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1576 struct obd_export *exp = req->rq_export;
1577 struct ptlrpc_request *reqiter;
1578 struct ptlrpc_request *dup_req = NULL;
1583 spin_lock(&exp->exp_lock);
1584 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1586 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1594 /* We expect it with RESENT and REPLAY flags. */
1595 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1596 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1597 CERROR("invalid flags %x of resent replay\n",
1598 lustre_msg_get_flags(req->rq_reqmsg));
1600 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1603 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1605 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1606 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1610 list_add_tail(&req->rq_replay_list,
1611 &exp->exp_req_replay_queue);
1614 spin_unlock(&exp->exp_lock);
1618 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1620 LASSERT(!list_empty(&req->rq_replay_list));
1621 LASSERT(req->rq_export);
1623 spin_lock(&req->rq_export->exp_lock);
1624 list_del_init(&req->rq_replay_list);
1625 spin_unlock(&req->rq_export->exp_lock);
1628 static void target_finish_recovery(struct lu_target *lut)
1630 struct obd_device *obd = lut->lut_obd;
1634 /* Only log a recovery message when recovery has occurred. */
1635 if (obd->obd_recovery_start) {
1636 time64_t now = ktime_get_real_seconds();
1637 time64_t elapsed_time;
1639 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1641 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1642 obd->obd_name, (s64)elapsed_time / 60,
1643 (s64)elapsed_time % 60,
1644 obd->obd_max_recoverable_clients,
1645 atomic_read(&obd->obd_connected_clients),
1646 obd->obd_stale_clients,
1647 obd->obd_stale_clients == 1 ? "was" : "were");
1650 ldlm_reprocess_recovery_done(obd->obd_namespace);
1651 spin_lock(&obd->obd_recovery_task_lock);
1652 if (!list_empty(&obd->obd_req_replay_queue) ||
1653 !list_empty(&obd->obd_lock_replay_queue) ||
1654 !list_empty(&obd->obd_final_req_queue)) {
1655 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1657 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1658 list_empty(&obd->obd_lock_replay_queue) ? \
1660 list_empty(&obd->obd_final_req_queue) ? \
1662 spin_unlock(&obd->obd_recovery_task_lock);
1665 spin_unlock(&obd->obd_recovery_task_lock);
1667 obd->obd_recovery_end = ktime_get_real_seconds();
1669 /* When recovery finished, cleanup orphans on MDS and OST. */
1670 if (obd->obd_type && OBP(obd, postrecov)) {
1671 int rc = OBP(obd, postrecov)(obd);
1674 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1680 static void abort_req_replay_queue(struct obd_device *obd)
1682 struct ptlrpc_request *req, *n;
1683 struct list_head abort_list;
1685 INIT_LIST_HEAD(&abort_list);
1686 spin_lock(&obd->obd_recovery_task_lock);
1687 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1688 spin_unlock(&obd->obd_recovery_task_lock);
1689 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1690 DEBUG_REQ(D_WARNING, req, "aborted:");
1691 req->rq_status = -ENOTCONN;
1692 if (ptlrpc_error(req)) {
1693 DEBUG_REQ(D_ERROR, req,
1694 "failed abort_req_reply; skipping");
1696 target_exp_dequeue_req_replay(req);
1697 target_request_copy_put(req);
1701 static void abort_lock_replay_queue(struct obd_device *obd)
1703 struct ptlrpc_request *req, *n;
1704 struct list_head abort_list;
1706 INIT_LIST_HEAD(&abort_list);
1707 spin_lock(&obd->obd_recovery_task_lock);
1708 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1709 spin_unlock(&obd->obd_recovery_task_lock);
1710 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1711 DEBUG_REQ(D_ERROR, req, "aborted:");
1712 req->rq_status = -ENOTCONN;
1713 if (ptlrpc_error(req)) {
1714 DEBUG_REQ(D_ERROR, req,
1715 "failed abort_lock_reply; skipping");
1717 target_request_copy_put(req);
1722 * Called from a cleanup function if the device is being cleaned up
1723 * forcefully. The exports should all have been disconnected already,
1724 * the only thing left to do is
1725 * - clear the recovery flags
1726 * - cancel the timer
1727 * - free queued requests and replies, but don't send replies
1728 * Because the obd_stopping flag is set, no new requests should be received.
1730 void target_cleanup_recovery(struct obd_device *obd)
1732 struct ptlrpc_request *req, *n;
1733 struct list_head clean_list;
1735 INIT_LIST_HEAD(&clean_list);
1736 spin_lock(&obd->obd_dev_lock);
1737 if (!obd->obd_recovering) {
1738 spin_unlock(&obd->obd_dev_lock);
1742 obd->obd_recovering = obd->obd_abort_recovery = 0;
1743 spin_unlock(&obd->obd_dev_lock);
1745 spin_lock(&obd->obd_recovery_task_lock);
1746 target_cancel_recovery_timer(obd);
1747 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1748 spin_unlock(&obd->obd_recovery_task_lock);
1750 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1751 LASSERT(req->rq_reply_state == NULL);
1752 target_exp_dequeue_req_replay(req);
1753 target_request_copy_put(req);
1756 spin_lock(&obd->obd_recovery_task_lock);
1757 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1758 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1759 spin_unlock(&obd->obd_recovery_task_lock);
1761 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1762 LASSERT(req->rq_reply_state == NULL);
1763 target_request_copy_put(req);
1768 EXPORT_SYMBOL(target_cleanup_recovery);
1770 /* obd_recovery_task_lock should be held */
1771 void target_cancel_recovery_timer(struct obd_device *obd)
1773 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1774 hrtimer_cancel(&obd->obd_recovery_timer);
1777 static void target_start_recovery_timer(struct obd_device *obd)
1781 if (obd->obd_recovery_start != 0)
1784 spin_lock(&obd->obd_dev_lock);
1785 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1786 spin_unlock(&obd->obd_dev_lock);
1790 LASSERT(obd->obd_recovery_timeout != 0);
1792 if (obd->obd_recovery_start != 0) {
1793 spin_unlock(&obd->obd_dev_lock);
1797 delay = ktime_set(obd->obd_recovery_timeout, 0);
1798 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_REL);
1799 obd->obd_recovery_start = ktime_get_real_seconds();
1800 spin_unlock(&obd->obd_dev_lock);
1802 LCONSOLE_WARN("%s: Will be in recovery for at least %lu:%02lu, or until %d client%s reconnect%s\n",
1804 obd->obd_recovery_timeout / 60,
1805 obd->obd_recovery_timeout % 60,
1806 obd->obd_max_recoverable_clients,
1807 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1808 (obd->obd_max_recoverable_clients == 1) ? "s" : "");
1812 * extend recovery window.
1814 * if @extend is true, extend recovery window to have @dr_timeout remaining
1815 * at least; otherwise, make sure the recovery timeout value is not less
1818 static void extend_recovery_timer(struct obd_device *obd, time_t dr_timeout,
1825 spin_lock(&obd->obd_dev_lock);
1826 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1827 spin_unlock(&obd->obd_dev_lock);
1830 LASSERT(obd->obd_recovery_start != 0);
1832 left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1833 left = ktime_divns(left_ns, NSEC_PER_SEC);
1836 timeout = obd->obd_recovery_timeout;
1837 /* dr_timeout will happen after the hrtimer has expired.
1838 * Add the excess time to the soft recovery timeout without
1839 * exceeding the hard recovery timeout.
1841 if (dr_timeout > left) {
1842 timeout += dr_timeout - left;
1843 timeout = min_t(time_t, obd->obd_recovery_time_hard,
1847 timeout = clamp_t(time_t, dr_timeout, obd->obd_recovery_timeout,
1848 obd->obd_recovery_time_hard);
1851 if (timeout == obd->obd_recovery_time_hard)
1852 CWARN("%s: extended recovery timer reached hard limit: %ld, extend: %d\n",
1853 obd->obd_name, timeout, extend);
1855 if (obd->obd_recovery_timeout < timeout) {
1856 ktime_t now = ktime_get_real();
1859 obd->obd_recovery_timeout = timeout;
1860 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1861 left_ns = ktime_sub(end, now);
1862 hrtimer_forward_now(&obd->obd_recovery_timer, left_ns);
1863 left = ktime_divns(left_ns, NSEC_PER_MSEC);
1865 spin_unlock(&obd->obd_dev_lock);
1867 CDEBUG(D_HA, "%s: recovery timer will expire in %ld seconds\n",
1868 obd->obd_name, left);
1871 /* Reset the timer with each new client connection */
1873 * This timer is actually reconnect_timer, which is for making sure
1874 * the total recovery window is at least as big as my reconnect
1875 * attempt timing. So the initial recovery time_out will be set to
1876 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1877 * from client is bigger than this, then the recovery time_out will
1878 * be extended to make sure the client could be reconnected, in the
1879 * process, the timeout from the new client should be ignored.
1882 check_and_start_recovery_timer(struct obd_device *obd,
1883 struct ptlrpc_request *req,
1886 time_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1887 struct obd_device_target *obt = &obd->u.obt;
1889 if (!new_client && service_time)
1891 * Teach server about old server's estimates, as first guess
1892 * at how long new requests will take.
1894 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1897 target_start_recovery_timer(obd);
1900 * Convert the service time to RPC timeout,
1901 * and reuse service_time to limit stack usage.
1903 service_time = at_est2timeout(service_time);
1905 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1906 service_time < at_extra)
1907 service_time = at_extra;
1910 * We expect other clients to timeout within service_time, then try
1911 * to reconnect, then try the failover server. The max delay between
1912 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
1914 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1916 LASSERT(obt->obt_magic == OBT_MAGIC);
1917 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1918 if (service_time > obd->obd_recovery_timeout && !new_client)
1919 extend_recovery_timer(obd, service_time, false);
1922 /** Health checking routines */
1923 static inline int exp_connect_healthy(struct obd_export *exp)
1925 return exp->exp_in_recovery;
1928 /** if export done req_replay or has replay in queue */
1929 static inline int exp_req_replay_healthy(struct obd_export *exp)
1931 return (!exp->exp_req_replay_needed ||
1932 atomic_read(&exp->exp_replay_count) > 0);
1936 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1938 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1939 exp_req_replay_healthy(exp);
1942 /** if export done lock_replay or has replay in queue */
1943 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1945 return (!exp->exp_lock_replay_needed ||
1946 atomic_read(&exp->exp_replay_count) > 0);
1949 static inline int exp_vbr_healthy(struct obd_export *exp)
1951 return !exp->exp_vbr_failed;
1954 static inline int exp_finished(struct obd_export *exp)
1956 return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
1959 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1961 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1965 static int check_for_next_transno(struct lu_target *lut)
1967 struct ptlrpc_request *req = NULL;
1968 struct obd_device *obd = lut->lut_obd;
1969 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1970 int wake_up = 0, connected, completed, queue_len;
1971 __u64 req_transno = 0;
1972 __u64 update_transno = 0;
1973 __u64 next_transno = 0;
1977 spin_lock(&obd->obd_recovery_task_lock);
1978 if (!list_empty(&obd->obd_req_replay_queue)) {
1979 req = list_entry(obd->obd_req_replay_queue.next,
1980 struct ptlrpc_request, rq_list);
1981 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1985 update_transno = distribute_txn_get_next_transno(tdtd);
1987 connected = atomic_read(&obd->obd_connected_clients);
1988 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1989 queue_len = obd->obd_requests_queued_for_recovery;
1990 next_transno = obd->obd_next_recovery_transno;
1993 "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
1994 obd->obd_max_recoverable_clients, connected, completed,
1995 queue_len, req_transno, next_transno);
1997 if (obd->obd_abort_recovery) {
1998 CDEBUG(D_HA, "waking for aborted recovery\n");
2000 } else if (obd->obd_recovery_expired) {
2001 CDEBUG(D_HA, "waking for expired recovery\n");
2003 } else if (tdtd != NULL && req != NULL &&
2004 is_req_replayed_by_update(req)) {
2005 LASSERTF(req_transno < next_transno,
2006 "req_transno %llu next_transno%llu\n", req_transno,
2008 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2011 } else if (req_transno == next_transno ||
2012 (update_transno != 0 && update_transno <= next_transno)) {
2013 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2015 } else if (queue_len > 0 &&
2016 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2017 /** handle gaps occured due to lost reply or VBR */
2018 LASSERTF(req_transno >= next_transno,
2019 "req_transno: %llu, next_transno: %llu\n",
2020 req_transno, next_transno);
2022 "%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",
2023 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2024 next_transno, queue_len, completed, connected,
2025 req_transno, update_transno, obd->obd_last_committed);
2026 obd->obd_next_recovery_transno = req_transno;
2028 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2029 CDEBUG(D_HA, "waking for completed recovery\n");
2031 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2033 "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2035 obd->obd_next_recovery_transno = req_transno;
2038 spin_unlock(&obd->obd_recovery_task_lock);
2042 static int check_for_next_lock(struct lu_target *lut)
2044 struct obd_device *obd = lut->lut_obd;
2047 spin_lock(&obd->obd_recovery_task_lock);
2048 if (!list_empty(&obd->obd_lock_replay_queue)) {
2049 CDEBUG(D_HA, "waking for next lock\n");
2051 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2052 CDEBUG(D_HA, "waking for completed lock replay\n");
2054 } else if (obd->obd_abort_recovery) {
2055 CDEBUG(D_HA, "waking for aborted recovery\n");
2057 } else if (obd->obd_recovery_expired) {
2058 CDEBUG(D_HA, "waking for expired recovery\n");
2061 spin_unlock(&obd->obd_recovery_task_lock);
2067 * wait for recovery events,
2068 * check its status with help of check_routine
2069 * evict dead clients via health_check
2071 static int target_recovery_overseer(struct lu_target *lut,
2072 int (*check_routine)(struct lu_target *),
2073 int (*health_check)(struct obd_export *))
2075 struct obd_device *obd = lut->lut_obd;
2076 struct target_distribute_txn_data *tdtd;
2080 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2081 now = ktime_get_seconds();
2082 if (now - last > 600) {
2083 LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2088 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
2089 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2090 __u64 next_update_transno = 0;
2093 * Only abort the recovery if there are no update recovery
2096 spin_lock(&obd->obd_recovery_task_lock);
2097 if (lut->lut_tdtd != NULL) {
2098 next_update_transno =
2099 distribute_txn_get_next_transno(lut->lut_tdtd);
2101 tdtd = lut->lut_tdtd;
2103 * If next_update_transno == 0, it probably because
2104 * updatelog retrieve threads did not get any records
2105 * yet, let's wait those threads stopped
2107 if (next_update_transno == 0) {
2108 struct l_wait_info lwi = { 0 };
2110 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2112 &tdtd->tdtd_recovery_threads_count) == 0,
2115 next_update_transno =
2116 distribute_txn_get_next_transno(
2121 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2122 obd->obd_next_recovery_transno = next_update_transno;
2123 spin_unlock(&obd->obd_recovery_task_lock);
2125 * Disconnect unfinished exports from clients, and
2126 * keep connection from MDT to make sure the update
2127 * recovery will still keep trying until some one
2128 * manually abort the recovery
2130 class_disconnect_stale_exports(obd,
2131 exp_finished_or_from_mdt);
2132 /* Abort all of replay & replay lock req from clients */
2133 abort_req_replay_queue(obd);
2134 abort_lock_replay_queue(obd);
2136 "%s: there are still update replay (%#llx)in the queue.\n",
2137 obd->obd_name, next_update_transno);
2139 obd->obd_abort_recovery = 1;
2140 spin_unlock(&obd->obd_recovery_task_lock);
2141 CWARN("%s recovery is aborted by hard timeout\n",
2146 while (wait_event_timeout(obd->obd_next_transno_waitq,
2148 cfs_time_seconds(60)) == 0)
2149 ; /* wait indefinitely for event, but don't trigger watchdog */
2151 if (obd->obd_abort_recovery) {
2152 CWARN("recovery is aborted, evict exports in recovery\n");
2153 if (lut->lut_tdtd != NULL) {
2154 struct l_wait_info lwi = { 0 };
2156 tdtd = lut->lut_tdtd;
2158 * Let's wait all of the update log recovery thread
2161 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2162 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2164 /* Then abort the update recovery list */
2165 dtrq_list_destroy(lut->lut_tdtd);
2168 /** evict exports which didn't finish recovery yet */
2169 class_disconnect_stale_exports(obd, exp_finished);
2171 } else if (obd->obd_recovery_expired) {
2172 obd->obd_recovery_expired = 0;
2173 /** If some clients died being recovered, evict them */
2174 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2176 /** evict cexports with no replay in queue, they are stalled */
2177 class_disconnect_stale_exports(obd, health_check);
2179 /** continue with VBR */
2180 spin_lock(&obd->obd_dev_lock);
2181 obd->obd_version_recov = 1;
2182 spin_unlock(&obd->obd_dev_lock);
2184 * reset timer, recovery will proceed with versions now,
2185 * timeout is set just to handle reconnection delays
2187 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2189 * Wait for recovery events again, after evicting bad clients
2196 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2198 struct obd_device *obd = lut->lut_obd;
2199 struct ptlrpc_request *req = NULL;
2201 CDEBUG(D_HA, "Waiting for lock\n");
2202 if (target_recovery_overseer(lut, check_for_next_lock,
2203 exp_lock_replay_healthy))
2204 abort_lock_replay_queue(obd);
2206 spin_lock(&obd->obd_recovery_task_lock);
2207 if (!list_empty(&obd->obd_lock_replay_queue)) {
2208 req = list_entry(obd->obd_lock_replay_queue.next,
2209 struct ptlrpc_request, rq_list);
2210 list_del_init(&req->rq_list);
2211 spin_unlock(&obd->obd_recovery_task_lock);
2213 spin_unlock(&obd->obd_recovery_task_lock);
2214 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2215 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2216 /** evict exports failed VBR */
2217 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2222 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2224 struct ptlrpc_request *req = NULL;
2226 spin_lock(&obd->obd_recovery_task_lock);
2227 if (!list_empty(&obd->obd_final_req_queue)) {
2228 req = list_entry(obd->obd_final_req_queue.next,
2229 struct ptlrpc_request, rq_list);
2230 list_del_init(&req->rq_list);
2231 spin_unlock(&obd->obd_recovery_task_lock);
2232 if (req->rq_export->exp_in_recovery) {
2233 spin_lock(&req->rq_export->exp_lock);
2234 req->rq_export->exp_in_recovery = 0;
2235 spin_unlock(&req->rq_export->exp_lock);
2238 spin_unlock(&obd->obd_recovery_task_lock);
2243 static void handle_recovery_req(struct ptlrpc_thread *thread,
2244 struct ptlrpc_request *req,
2245 svc_handler_t handler)
2250 * export can be evicted during recovery, no need to handle replays for
2251 * it after that, discard such request silently
2253 if (req->rq_export->exp_disconnected)
2256 req->rq_session.lc_thread = thread;
2257 req->rq_svc_thread = thread;
2258 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2260 /* thread context */
2261 lu_context_enter(&thread->t_env->le_ctx);
2263 lu_context_exit(&thread->t_env->le_ctx);
2265 /* don't reset timer for final stage */
2266 if (!exp_finished(req->rq_export)) {
2267 time_t to = obd_timeout;
2270 * Add request timeout to the recovery time so next request from
2271 * this client may come in recovery time
2274 struct ptlrpc_service_part *svcpt;
2276 svcpt = req->rq_rqbd->rqbd_svcpt;
2278 * If the server sent early reply for this request,
2279 * the client will recalculate the timeout according to
2280 * current server estimate service time, so we will
2281 * use the maxium timeout here for waiting the client
2282 * sending the next req
2285 at_est2timeout(at_get(&svcpt->scp_at_estimate)),
2286 lustre_msg_get_timeout(req->rq_reqmsg));
2288 * Add 2 net_latency, one for balance rq_deadline
2289 * (see ptl_send_rpc), one for resend the req to server,
2290 * Note: client will pack net_latency in replay req
2291 * (see ptlrpc_replay_req)
2293 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2295 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2300 /** Checking routines for recovery */
2301 static int check_for_recovery_ready(struct lu_target *lut)
2303 struct obd_device *obd = lut->lut_obd;
2304 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2307 "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2308 clnts, obd->obd_stale_clients, obd->obd_max_recoverable_clients,
2309 obd->obd_abort_recovery, obd->obd_recovery_expired);
2311 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2312 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2313 if (clnts + obd->obd_stale_clients <
2314 obd->obd_max_recoverable_clients)
2318 if (lut->lut_tdtd != NULL) {
2319 if (!lut->lut_tdtd->tdtd_replay_ready &&
2320 !obd->obd_abort_recovery) {
2322 * Let's extend recovery timer, in case the recovery
2323 * timer expired, and some clients got evicted
2325 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2328 "%s update recovery is not ready, extend recovery %lu\n",
2329 obd->obd_name, obd->obd_recovery_timeout);
2338 REQUEST_RECOVERY = 1,
2339 UPDATE_RECOVERY = 2,
2342 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2346 if (!list_empty(&obd->obd_req_replay_queue)) {
2347 struct ptlrpc_request *req;
2349 req = list_entry(obd->obd_req_replay_queue.next,
2350 struct ptlrpc_request, rq_list);
2351 transno = lustre_msg_get_transno(req->rq_reqmsg);
2357 static __u64 get_next_transno(struct lu_target *lut, int *type)
2359 struct obd_device *obd = lut->lut_obd;
2360 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2362 __u64 update_transno;
2366 transno = get_next_replay_req_transno(obd);
2368 *type = REQUEST_RECOVERY;
2373 update_transno = distribute_txn_get_next_transno(tdtd);
2374 if (transno == 0 || (transno >= update_transno &&
2375 update_transno != 0)) {
2376 transno = update_transno;
2378 *type = UPDATE_RECOVERY;
2385 * drop duplicate replay request
2387 * Because the operation has been replayed by update recovery, the request
2388 * with the same transno will be dropped and also notify the client to send
2389 * next replay request.
2391 * \param[in] env execution environment
2392 * \param[in] obd failover obd device
2393 * \param[in] req request to be dropped
2395 static void drop_duplicate_replay_req(struct lu_env *env,
2396 struct obd_device *obd,
2397 struct ptlrpc_request *req)
2399 DEBUG_REQ(D_HA, req,
2400 "remove t%lld from %s because of duplicate update records are found.\n",
2401 lustre_msg_get_transno(req->rq_reqmsg),
2402 libcfs_nid2str(req->rq_peer.nid));
2405 * Right now, only for MDS reint operation update replay and
2406 * normal request replay can have the same transno
2408 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2409 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2410 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2411 if (likely(req->rq_export))
2412 target_committed_to_req(req);
2413 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2414 target_send_reply(req, req->rq_status, 0);
2416 DEBUG_REQ(D_ERROR, req, "wrong opc from %s\n",
2417 libcfs_nid2str(req->rq_peer.nid));
2419 target_exp_dequeue_req_replay(req);
2420 target_request_copy_put(req);
2421 obd->obd_replayed_requests++;
2424 static void replay_request_or_update(struct lu_env *env,
2425 struct lu_target *lut,
2426 struct target_recovery_data *trd,
2427 struct ptlrpc_thread *thread)
2429 struct obd_device *obd = lut->lut_obd;
2430 struct ptlrpc_request *req = NULL;
2436 CDEBUG(D_HA, "Waiting for transno %lld\n",
2437 obd->obd_next_recovery_transno);
2439 /* Replay all of request and update by transno */
2441 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2443 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2446 * It is needed to extend recovery window above
2447 * recovery_time_soft. Extending is possible only in the
2448 * end of recovery window (see more details in
2449 * handle_recovery_req()).
2451 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2453 if (target_recovery_overseer(lut, check_for_next_transno,
2454 exp_req_replay_healthy_or_from_mdt)) {
2455 abort_req_replay_queue(obd);
2456 abort_lock_replay_queue(obd);
2460 spin_lock(&obd->obd_recovery_task_lock);
2461 transno = get_next_transno(lut, &type);
2462 if (type == REQUEST_RECOVERY && transno != 0) {
2464 * Drop replay request from client side, if the
2465 * replay has been executed by update with the
2468 req = list_entry(obd->obd_req_replay_queue.next,
2469 struct ptlrpc_request, rq_list);
2471 list_del_init(&req->rq_list);
2472 obd->obd_requests_queued_for_recovery--;
2473 spin_unlock(&obd->obd_recovery_task_lock);
2476 * Let's check if the request has been redone by
2479 if (is_req_replayed_by_update(req)) {
2480 struct distribute_txn_replay_req *dtrq;
2482 dtrq = distribute_txn_lookup_finish_list(tdtd,
2484 LASSERT(dtrq != NULL);
2485 spin_lock(&tdtd->tdtd_replay_list_lock);
2486 list_del_init(&dtrq->dtrq_list);
2487 spin_unlock(&tdtd->tdtd_replay_list_lock);
2490 drop_duplicate_replay_req(env, obd, req);
2495 LASSERT(trd->trd_processing_task == current_pid());
2496 DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2498 lustre_msg_get_transno(req->rq_reqmsg),
2499 libcfs_nid2str(req->rq_peer.nid));
2501 handle_recovery_req(thread, req,
2502 trd->trd_recovery_handler);
2504 * bz18031: increase next_recovery_transno before
2505 * target_request_copy_put() will drop exp_rpc reference
2507 spin_lock(&obd->obd_recovery_task_lock);
2508 obd->obd_next_recovery_transno++;
2509 spin_unlock(&obd->obd_recovery_task_lock);
2510 target_exp_dequeue_req_replay(req);
2511 target_request_copy_put(req);
2512 obd->obd_replayed_requests++;
2513 } else if (type == UPDATE_RECOVERY && transno != 0) {
2514 struct distribute_txn_replay_req *dtrq;
2517 spin_unlock(&obd->obd_recovery_task_lock);
2519 LASSERT(tdtd != NULL);
2520 dtrq = distribute_txn_get_next_req(tdtd);
2521 lu_context_enter(&thread->t_env->le_ctx);
2522 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2523 lu_context_exit(&thread->t_env->le_ctx);
2524 extend_recovery_timer(obd, obd_timeout, true);
2526 if (rc == 0 && dtrq->dtrq_xid != 0) {
2528 "Move x%llu t%llu to finish list\n",
2530 dtrq->dtrq_master_transno);
2532 /* Add it to the replay finish list */
2533 spin_lock(&tdtd->tdtd_replay_list_lock);
2534 list_add(&dtrq->dtrq_list,
2535 &tdtd->tdtd_replay_finish_list);
2536 spin_unlock(&tdtd->tdtd_replay_list_lock);
2538 spin_lock(&obd->obd_recovery_task_lock);
2539 if (transno == obd->obd_next_recovery_transno)
2540 obd->obd_next_recovery_transno++;
2542 obd->obd_next_recovery_transno)
2543 obd->obd_next_recovery_transno =
2545 spin_unlock(&obd->obd_recovery_task_lock);
2550 spin_unlock(&obd->obd_recovery_task_lock);
2552 LASSERT(list_empty(&obd->obd_req_replay_queue));
2553 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2554 /** evict exports failed VBR */
2555 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2561 static int target_recovery_thread(void *arg)
2563 struct lu_target *lut = arg;
2564 struct obd_device *obd = lut->lut_obd;
2565 struct ptlrpc_request *req;
2566 struct target_recovery_data *trd = &obd->obd_recovery_data;
2567 unsigned long delta;
2569 struct ptlrpc_thread *thread = NULL;
2574 unshare_fs_struct();
2575 OBD_ALLOC_PTR(thread);
2581 GOTO(out_thread, rc = -ENOMEM);
2582 rc = lu_env_add(env);
2586 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2588 GOTO(out_env_remove, rc);
2590 thread->t_env = env;
2591 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2592 env->le_ctx.lc_thread = thread;
2593 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2595 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2597 trd->trd_processing_task = current_pid();
2599 spin_lock(&obd->obd_dev_lock);
2600 obd->obd_recovering = 1;
2601 spin_unlock(&obd->obd_dev_lock);
2602 complete(&trd->trd_starting);
2604 /* first of all, we have to know the first transno to replay */
2605 if (target_recovery_overseer(lut, check_for_recovery_ready,
2606 exp_connect_healthy)) {
2607 abort_req_replay_queue(obd);
2608 abort_lock_replay_queue(obd);
2609 if (lut->lut_tdtd != NULL)
2610 dtrq_list_destroy(lut->lut_tdtd);
2613 /* next stage: replay requests or update */
2615 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2616 atomic_read(&obd->obd_req_replay_clients),
2617 obd->obd_next_recovery_transno);
2618 replay_request_or_update(env, lut, trd, thread);
2621 * The second stage: replay locks
2623 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2624 atomic_read(&obd->obd_lock_replay_clients));
2625 while ((req = target_next_replay_lock(lut))) {
2626 LASSERT(trd->trd_processing_task == current_pid());
2627 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2628 libcfs_nid2str(req->rq_peer.nid));
2629 handle_recovery_req(thread, req,
2630 trd->trd_recovery_handler);
2631 target_request_copy_put(req);
2632 obd->obd_replayed_locks++;
2636 * The third stage: reply on final pings, at this moment all clients
2637 * must have request in final queue
2639 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2640 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2641 /** Update server last boot epoch */
2642 tgt_boot_epoch_update(lut);
2644 * We drop recoverying flag to forward all new requests
2645 * to regular mds_handle() since now
2647 spin_lock(&obd->obd_dev_lock);
2648 obd->obd_recovering = obd->obd_abort_recovery = 0;
2649 spin_unlock(&obd->obd_dev_lock);
2650 spin_lock(&obd->obd_recovery_task_lock);
2651 target_cancel_recovery_timer(obd);
2652 spin_unlock(&obd->obd_recovery_task_lock);
2653 while ((req = target_next_final_ping(obd))) {
2654 LASSERT(trd->trd_processing_task == current_pid());
2655 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2656 libcfs_nid2str(req->rq_peer.nid));
2657 handle_recovery_req(thread, req,
2658 trd->trd_recovery_handler);
2660 * Because the waiting client can not send ping to server,
2661 * so we need refresh the last_request_time, to avoid the
2662 * export is being evicted
2664 ptlrpc_update_export_timer(req->rq_export, 0);
2665 target_request_copy_put(req);
2668 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2669 CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2670 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2671 if (delta > OBD_RECOVERY_TIME_SOFT) {
2672 CWARN("too long recovery - read logs\n");
2673 libcfs_debug_dumplog();
2676 target_finish_recovery(lut);
2677 lu_context_fini(&env->le_ctx);
2678 trd->trd_processing_task = 0;
2679 complete_all(&trd->trd_finishing);
2680 tgt_io_thread_done(thread);
2686 OBD_FREE_PTR(thread);
2690 static int target_start_recovery_thread(struct lu_target *lut,
2691 svc_handler_t handler)
2693 struct obd_device *obd = lut->lut_obd;
2695 struct target_recovery_data *trd = &obd->obd_recovery_data;
2698 memset(trd, 0, sizeof(*trd));
2699 init_completion(&trd->trd_starting);
2700 init_completion(&trd->trd_finishing);
2701 trd->trd_recovery_handler = handler;
2703 rc = server_name2index(obd->obd_name, &index, NULL);
2707 if (!IS_ERR(kthread_run(target_recovery_thread,
2708 lut, "tgt_recover_%d", index))) {
2709 wait_for_completion(&trd->trd_starting);
2710 LASSERT(obd->obd_recovering != 0);
2718 void target_stop_recovery_thread(struct obd_device *obd)
2720 if (obd->obd_recovery_data.trd_processing_task > 0) {
2721 struct target_recovery_data *trd = &obd->obd_recovery_data;
2722 /** recovery can be done but postrecovery is not yet */
2723 spin_lock(&obd->obd_dev_lock);
2724 if (obd->obd_recovering) {
2725 CERROR("%s: Aborting recovery\n", obd->obd_name);
2726 obd->obd_abort_recovery = 1;
2727 wake_up(&obd->obd_next_transno_waitq);
2729 spin_unlock(&obd->obd_dev_lock);
2730 wait_for_completion(&trd->trd_finishing);
2733 EXPORT_SYMBOL(target_stop_recovery_thread);
2735 void target_recovery_fini(struct obd_device *obd)
2737 class_disconnect_exports(obd);
2738 target_stop_recovery_thread(obd);
2739 target_cleanup_recovery(obd);
2741 EXPORT_SYMBOL(target_recovery_fini);
2743 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2745 struct obd_device *obd = container_of(timer, struct obd_device,
2746 obd_recovery_timer);
2749 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2750 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2751 ktime_get_real_seconds() - obd->obd_recovery_start,
2752 atomic_read(&obd->obd_connected_clients));
2754 obd->obd_recovery_expired = 1;
2755 wake_up(&obd->obd_next_transno_waitq);
2756 return HRTIMER_NORESTART;
2759 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2761 struct obd_device *obd = lut->lut_obd;
2763 if (lut->lut_bottom->dd_rdonly)
2766 if (obd->obd_max_recoverable_clients == 0) {
2767 /** Update server last boot epoch */
2768 tgt_boot_epoch_update(lut);
2772 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2773 "last_transno %llu\n", obd->obd_name,
2774 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2775 LASSERT(obd->obd_stopping == 0);
2776 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2777 obd->obd_recovery_start = 0;
2778 obd->obd_recovery_end = 0;
2780 hrtimer_init(&obd->obd_recovery_timer, CLOCK_REALTIME,
2782 obd->obd_recovery_timer.function = &target_recovery_expired;
2783 target_start_recovery_thread(lut, handler);
2785 EXPORT_SYMBOL(target_recovery_init);
2787 static int target_process_req_flags(struct obd_device *obd,
2788 struct ptlrpc_request *req)
2790 struct obd_export *exp = req->rq_export;
2792 LASSERT(exp != NULL);
2793 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2794 /* client declares he's ready to replay locks */
2795 spin_lock(&exp->exp_lock);
2796 if (exp->exp_req_replay_needed) {
2797 exp->exp_req_replay_needed = 0;
2798 spin_unlock(&exp->exp_lock);
2800 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2801 atomic_dec(&obd->obd_req_replay_clients);
2803 spin_unlock(&exp->exp_lock);
2806 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2808 * client declares he's ready to complete recovery
2809 * so, we put the request on th final queue
2811 spin_lock(&exp->exp_lock);
2812 if (exp->exp_lock_replay_needed) {
2813 exp->exp_lock_replay_needed = 0;
2814 spin_unlock(&exp->exp_lock);
2816 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2817 atomic_dec(&obd->obd_lock_replay_clients);
2819 spin_unlock(&exp->exp_lock);
2825 int target_queue_recovery_request(struct ptlrpc_request *req,
2826 struct obd_device *obd)
2828 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2829 struct ptlrpc_request *reqiter;
2834 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2835 /* Processing the queue right now, don't re-add. */
2839 target_process_req_flags(obd, req);
2841 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2842 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2843 if (cfs_fail_val == 1) {
2846 wake_up(&cfs_race_waitq);
2848 set_current_state(TASK_INTERRUPTIBLE);
2849 schedule_timeout(cfs_time_seconds(1));
2854 * client declares he's ready to complete recovery
2855 * so, we put the request on th final queue
2857 target_request_copy_get(req);
2858 DEBUG_REQ(D_HA, req, "queue final req");
2859 wake_up(&obd->obd_next_transno_waitq);
2860 spin_lock(&obd->obd_recovery_task_lock);
2861 if (obd->obd_recovering) {
2862 struct ptlrpc_request *tmp;
2863 struct ptlrpc_request *duplicate = NULL;
2865 if (likely(!req->rq_export->exp_replay_done)) {
2866 req->rq_export->exp_replay_done = 1;
2867 list_add_tail(&req->rq_list,
2868 &obd->obd_final_req_queue);
2869 spin_unlock(&obd->obd_recovery_task_lock);
2874 * XXX O(n), but only happens if final ping is
2875 * timed out, probably reorganize the list as
2878 list_for_each_entry_safe(reqiter, tmp,
2879 &obd->obd_final_req_queue,
2881 if (reqiter->rq_export == req->rq_export) {
2882 list_del_init(&reqiter->rq_list);
2883 duplicate = reqiter;
2888 list_add_tail(&req->rq_list,
2889 &obd->obd_final_req_queue);
2890 req->rq_export->exp_replay_done = 1;
2891 spin_unlock(&obd->obd_recovery_task_lock);
2893 if (duplicate != NULL) {
2894 DEBUG_REQ(D_HA, duplicate,
2895 "put prev final req\n");
2896 target_request_copy_put(duplicate);
2900 spin_unlock(&obd->obd_recovery_task_lock);
2901 target_request_copy_put(req);
2902 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2905 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2906 /* client declares he's ready to replay locks */
2907 target_request_copy_get(req);
2908 DEBUG_REQ(D_HA, req, "queue lock replay req");
2909 wake_up(&obd->obd_next_transno_waitq);
2910 spin_lock(&obd->obd_recovery_task_lock);
2911 LASSERT(obd->obd_recovering);
2912 /* usually due to recovery abort */
2913 if (!req->rq_export->exp_in_recovery) {
2914 spin_unlock(&obd->obd_recovery_task_lock);
2915 target_request_copy_put(req);
2918 LASSERT(req->rq_export->exp_lock_replay_needed);
2919 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2920 spin_unlock(&obd->obd_recovery_task_lock);
2925 * CAVEAT EMPTOR: The incoming request message has been swabbed
2926 * (i.e. buflens etc are in my own byte order), but type-dependent
2927 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
2931 INIT_LIST_HEAD(&req->rq_list);
2932 DEBUG_REQ(D_HA, req, "not queueing");
2937 * If we're processing the queue, we want don't want to queue this
2940 * Also, if this request has a transno less than the one we're waiting
2941 * for, we should process it now. It could (and currently always will)
2942 * be an open request for a descriptor that was opened some time ago.
2944 * Also, a resent, replayed request that has already been
2945 * handled will pass through here and be processed immediately.
2948 "Next recovery transno: %llu, current: %llu, replaying\n",
2949 obd->obd_next_recovery_transno, transno);
2952 * If the request has been replayed by update replay, then sends this
2953 * request to the recovery thread (replay_request_or_update()), where
2954 * it will be handled
2956 spin_lock(&obd->obd_recovery_task_lock);
2957 if (transno < obd->obd_next_recovery_transno &&
2958 !is_req_replayed_by_update(req)) {
2959 /* Processing the queue right now, don't re-add. */
2960 LASSERT(list_empty(&req->rq_list));
2961 spin_unlock(&obd->obd_recovery_task_lock);
2964 spin_unlock(&obd->obd_recovery_task_lock);
2966 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2969 target_request_copy_get(req);
2970 if (!req->rq_export->exp_in_recovery) {
2971 target_request_copy_put(req);
2974 LASSERT(req->rq_export->exp_req_replay_needed);
2976 if (target_exp_enqueue_req_replay(req)) {
2977 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2978 target_request_copy_put(req);
2983 spin_lock(&obd->obd_recovery_task_lock);
2984 LASSERT(obd->obd_recovering);
2985 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2986 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2987 list_add_tail(&req->rq_list, &reqiter->rq_list);
2992 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2994 DEBUG_REQ(D_ERROR, req,
2995 "dropping replay: transno has been claimed by another client");
2996 spin_unlock(&obd->obd_recovery_task_lock);
2997 target_exp_dequeue_req_replay(req);
2998 target_request_copy_put(req);
3004 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3006 obd->obd_requests_queued_for_recovery++;
3007 spin_unlock(&obd->obd_recovery_task_lock);
3008 wake_up(&obd->obd_next_transno_waitq);
3012 void target_committed_to_req(struct ptlrpc_request *req)
3014 struct obd_export *exp = req->rq_export;
3016 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3017 lustre_msg_set_last_committed(req->rq_repmsg,
3018 exp->exp_last_committed);
3020 DEBUG_REQ(D_IOCTL, req,
3021 "not sending last_committed update (%d/%d)",
3022 exp->exp_obd->obd_no_transno,
3023 req->rq_repmsg == NULL);
3025 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3026 exp->exp_last_committed, req->rq_transno, req->rq_xid);
3029 #endif /* HAVE_SERVER_SUPPORT */
3032 * Packs current SLV and Limit into \a req.
3034 int target_pack_pool_reply(struct ptlrpc_request *req)
3036 struct obd_device *obd;
3041 * Check that we still have all structures alive as this may
3042 * be some late RPC at shutdown time.
3044 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3045 !exp_connect_lru_resize(req->rq_export))) {
3046 lustre_msg_set_slv(req->rq_repmsg, 0);
3047 lustre_msg_set_limit(req->rq_repmsg, 0);
3051 /* OBD is alive here as export is alive, which we checked above. */
3052 obd = req->rq_export->exp_obd;
3054 read_lock(&obd->obd_pool_lock);
3055 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3056 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3057 read_unlock(&obd->obd_pool_lock);
3062 static int target_send_reply_msg(struct ptlrpc_request *req,
3063 int rc, int fail_id)
3065 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3066 DEBUG_REQ(D_ERROR, req, "dropping reply");
3070 * We can have a null rq_reqmsg in the event of bad signature or
3071 * no context when unwrapping
3073 if (req->rq_reqmsg &&
3074 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3075 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3079 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3080 req->rq_status = rc;
3081 return ptlrpc_send_error(req, 1);
3083 DEBUG_REQ(D_NET, req, "sending reply");
3085 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3088 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3090 struct ptlrpc_service_part *svcpt;
3092 struct ptlrpc_reply_state *rs;
3093 struct obd_export *exp;
3097 if (req->rq_no_reply) {
3102 svcpt = req->rq_rqbd->rqbd_svcpt;
3103 rs = req->rq_reply_state;
3104 if (rs == NULL || !rs->rs_difficult) {
3106 target_send_reply_msg(req, rc, fail_id);
3111 /* must be an export if locks saved */
3112 LASSERT(req->rq_export != NULL);
3113 /* req/reply consistent */
3114 LASSERT(rs->rs_svcpt == svcpt);
3117 LASSERT(!rs->rs_scheduled);
3118 LASSERT(!rs->rs_scheduled_ever);
3119 LASSERT(!rs->rs_handled);
3120 LASSERT(!rs->rs_on_net);
3121 LASSERT(rs->rs_export == NULL);
3122 LASSERT(list_empty(&rs->rs_obd_list));
3123 LASSERT(list_empty(&rs->rs_exp_list));
3125 exp = class_export_get(req->rq_export);
3127 /* disable reply scheduling while I'm setting up */
3128 rs->rs_scheduled = 1;
3130 rs->rs_xid = req->rq_xid;
3131 rs->rs_transno = req->rq_transno;
3132 rs->rs_export = exp;
3133 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3135 spin_lock(&exp->exp_uncommitted_replies_lock);
3136 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3137 rs->rs_transno, exp->exp_last_committed);
3138 if (rs->rs_transno > exp->exp_last_committed) {
3139 /* not committed already */
3140 list_add_tail(&rs->rs_obd_list,
3141 &exp->exp_uncommitted_replies);
3143 spin_unlock(&exp->exp_uncommitted_replies_lock);
3145 spin_lock(&exp->exp_lock);
3146 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3147 spin_unlock(&exp->exp_lock);
3149 netrc = target_send_reply_msg(req, rc, fail_id);
3151 spin_lock(&svcpt->scp_rep_lock);
3153 atomic_inc(&svcpt->scp_nreps_difficult);
3157 * error sending: reply is off the net. Also we need +1
3158 * reply ref until ptlrpc_handle_rs() is done
3159 * with the reply state (if the send was successful, there
3160 * would have been +1 ref for the net, which
3161 * reply_out_callback leaves alone)
3164 ptlrpc_rs_addref(rs);
3167 spin_lock(&rs->rs_lock);
3168 if (rs->rs_transno <= exp->exp_last_committed ||
3169 (!rs->rs_on_net && !rs->rs_no_ack) ||
3170 list_empty(&rs->rs_exp_list) || /* completed already */
3171 list_empty(&rs->rs_obd_list)) {
3172 CDEBUG(D_HA, "Schedule reply immediately\n");
3173 ptlrpc_dispatch_difficult_reply(rs);
3175 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3176 rs->rs_scheduled = 0; /* allow notifier to schedule */
3178 spin_unlock(&rs->rs_lock);
3179 spin_unlock(&svcpt->scp_rep_lock);
3183 enum ldlm_mode lck_compat_array[] = {
3184 [LCK_EX] = LCK_COMPAT_EX,
3185 [LCK_PW] = LCK_COMPAT_PW,
3186 [LCK_PR] = LCK_COMPAT_PR,
3187 [LCK_CW] = LCK_COMPAT_CW,
3188 [LCK_CR] = LCK_COMPAT_CR,
3189 [LCK_NL] = LCK_COMPAT_NL,
3190 [LCK_GROUP] = LCK_COMPAT_GROUP,
3191 [LCK_COS] = LCK_COMPAT_COS,
3195 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3196 * not escape to the user level.
3198 int ldlm_error2errno(enum ldlm_error error)
3204 case ELDLM_LOCK_MATCHED:
3207 case ELDLM_LOCK_CHANGED:
3210 case ELDLM_LOCK_ABORTED:
3213 case ELDLM_LOCK_REPLACED:
3216 case ELDLM_NO_LOCK_DATA:
3219 case ELDLM_NAMESPACE_EXISTS:
3222 case ELDLM_BAD_NAMESPACE:
3226 if (((int)error) < 0) { /* cast to signed type */
3227 result = error; /* as ldlm_error can be unsigned */
3229 CERROR("Invalid DLM result code: %d\n", error);
3235 EXPORT_SYMBOL(ldlm_error2errno);
3238 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3240 enum ldlm_error ldlm_errno2error(int err_no)
3249 error = ELDLM_LOCK_CHANGED;
3252 error = ELDLM_LOCK_ABORTED;
3255 error = ELDLM_LOCK_REPLACED;
3258 error = ELDLM_NO_LOCK_DATA;
3261 error = ELDLM_NAMESPACE_EXISTS;
3264 error = ELDLM_BAD_NAMESPACE;
3272 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3273 void ldlm_dump_export_locks(struct obd_export *exp)
3275 spin_lock(&exp->exp_locks_list_guard);
3276 if (!list_empty(&exp->exp_locks_list)) {
3277 struct ldlm_lock *lock;
3279 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3281 list_for_each_entry(lock, &exp->exp_locks_list,
3283 LDLM_ERROR(lock, "lock:");
3285 spin_unlock(&exp->exp_locks_list_guard);
3289 #ifdef HAVE_SERVER_SUPPORT
3290 static int target_bulk_timeout(void *data)
3294 * We don't fail the connection here, because having the export
3295 * killed makes the (vital) call to commitrw very sad.
3300 static inline const char *bulk2type(struct ptlrpc_request *req)
3302 if (req->rq_bulk_read)
3304 if (req->rq_bulk_write)
3309 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3310 struct l_wait_info *lwi)
3312 struct ptlrpc_request *req = desc->bd_req;
3313 time64_t start = ktime_get_real_seconds();
3319 /* If there is eviction in progress, wait for it to finish. */
3320 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3321 *lwi = LWI_INTR(NULL, NULL);
3322 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3323 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
3327 /* Check if client was evicted or reconnected already. */
3328 if (exp->exp_failed ||
3329 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3332 if (req->rq_bulk_read)
3333 rc = sptlrpc_svc_wrap_bulk(req, desc);
3335 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3336 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3337 else /* old version, bulk matchbits is rq_xid */
3338 req->rq_mbits = req->rq_xid;
3341 rc = ptlrpc_start_bulk_transfer(desc);
3345 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3346 bulk2type(req), rc);
3350 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3351 ptlrpc_abort_bulk(desc);
3355 /* limit actual bulk transfer to bulk_timeout seconds */
3356 deadline = start + bulk_timeout;
3357 if (deadline > req->rq_deadline)
3358 deadline = req->rq_deadline;
3361 time64_t timeoutl = deadline - ktime_get_real_seconds();
3362 long timeout_jiffies = timeoutl <= 0 ?
3363 1 : cfs_time_seconds(timeoutl);
3364 time64_t rq_deadline;
3366 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3367 cfs_time_seconds(1),
3368 target_bulk_timeout, desc);
3369 rc = l_wait_event(desc->bd_waitq,
3370 !ptlrpc_server_bulk_active(desc) ||
3373 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3375 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3376 /* Wait again if we changed rq_deadline. */
3377 rq_deadline = READ_ONCE(req->rq_deadline);
3378 deadline = start + bulk_timeout;
3379 if (deadline > rq_deadline)
3380 deadline = rq_deadline;
3381 } while (rc == -ETIMEDOUT &&
3382 deadline > ktime_get_real_seconds());
3384 if (rc == -ETIMEDOUT) {
3385 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3386 bulk2type(req), deadline - start,
3387 ktime_get_real_seconds() - deadline);
3388 ptlrpc_abort_bulk(desc);
3389 } else if (exp->exp_failed) {
3390 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3393 ptlrpc_abort_bulk(desc);
3394 } else if (exp->exp_conn_cnt >
3395 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3396 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3398 /* We don't reply anyway. */
3400 ptlrpc_abort_bulk(desc);
3401 } else if (desc->bd_failure) {
3402 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3404 /* XXX should this be a different errno? */
3407 if (req->rq_bulk_write)
3408 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3409 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3410 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3411 bulk2type(req), desc->bd_nob_transferred,
3413 /* XXX should this be a different errno? */
3420 EXPORT_SYMBOL(target_bulk_io);
3422 #endif /* HAVE_SERVER_SUPPORT */