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 *obd, struct lustre_cfg *lcfg)
276 struct client_obd *cli = &obd->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 = obd->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_waitqueue_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_DEF_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(obd->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 &obd->obd_ldlm_client);
487 imp = class_new_import(obd);
489 GOTO(err_ldlm, rc = -ENOENT);
490 imp->imp_client = &obd->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 obd->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",
525 cli->cl_target_uuid.uuid);
526 spin_lock(&imp->imp_lock);
527 imp->imp_deactive = 1;
528 spin_unlock(&imp->imp_lock);
532 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
533 LDLM_NAMESPACE_CLIENT,
534 LDLM_NAMESPACE_GREEDY,
536 if (obd->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 *obd)
560 struct client_obd *cli = &obd->u.cli;
564 ldlm_namespace_free_post(obd->obd_namespace);
565 obd->obd_namespace = NULL;
567 obd_cleanup_client_import(obd);
568 LASSERT(obd->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);
790 static inline int target_check_recovery_timer(struct obd_device *target)
795 if (!target->obd_recovering || target->obd_recovery_start == 0)
798 remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
799 timeout = ktime_divns(remaining, NSEC_PER_SEC);
803 /* the recovery timer should expire, but it isn't triggered,
804 * it's better to abort the recovery of this target to speed up
805 * the recovery of the whole cluster. */
806 spin_lock(&target->obd_dev_lock);
807 if (target->obd_recovering) {
808 CERROR("%s: Aborting recovery\n", target->obd_name);
809 target->obd_abort_recovery = 1;
810 wake_up(&target->obd_next_transno_waitq);
812 spin_unlock(&target->obd_dev_lock);
817 * --------------------------------------------------------------------------
818 * from old lib/target.c
819 * --------------------------------------------------------------------------
821 static int target_handle_reconnect(struct lustre_handle *conn,
822 struct obd_export *exp,
823 struct obd_uuid *cluuid)
825 struct obd_device *target;
826 struct lustre_handle *hdl;
832 hdl = &exp->exp_imp_reverse->imp_remote_handle;
833 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
834 conn->cookie = exp->exp_handle.h_cookie;
836 "connect export for UUID '%s' at %p, cookie %#llx\n",
837 cluuid->uuid, exp, conn->cookie);
841 target = exp->exp_obd;
843 /* Might be a re-connect after a partition. */
844 if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
845 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",
847 obd_uuid2str(&exp->exp_client_uuid),
848 obd_export_nid2str(exp),
849 hdl->cookie, conn->cookie);
850 memset(conn, 0, sizeof(*conn));
852 * target_handle_connect() treats EALREADY and
853 * -EALREADY differently. -EALREADY is an error
854 * (same UUID, different handle).
859 if (!target->obd_recovering) {
860 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
861 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
862 obd_export_nid2str(exp));
863 GOTO(out_already, rc);
866 remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
867 timeout = ktime_divns(remaining, NSEC_PER_SEC);
869 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
871 obd_uuid2str(&exp->exp_client_uuid),
872 obd_export_nid2str(exp),
873 atomic_read(&target->obd_max_recoverable_clients),
874 timeout / 60, timeout % 60);
876 struct target_distribute_txn_data *tdtd;
881 target_check_recovery_timer(target);
883 tdtd = class_exp2tgt(exp)->lut_tdtd;
884 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
885 buf = tdtd->tdtd_show_update_logs_retrievers(
886 tdtd->tdtd_show_retrievers_cbdata,
890 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",
892 obd_uuid2str(&exp->exp_client_uuid),
893 obd_export_nid2str(exp), count,
894 buf ? buf : "unknown (not enough RAM)",
895 (abs(timeout) + target->obd_recovery_timeout) / 60,
896 (abs(timeout) + target->obd_recovery_timeout) % 60,
899 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",
900 target->obd_name, abs(timeout) / 60,
901 abs(timeout) % 60, target->obd_name);
908 conn->cookie = exp->exp_handle.h_cookie;
910 * target_handle_connect() treats EALREADY and
911 * -EALREADY differently. EALREADY means we are
912 * doing a valid reconnect from the same client.
918 check_and_start_recovery_timer(struct obd_device *obd,
919 struct ptlrpc_request *req, int new_client);
922 * update flags for import during reconnect process
924 static int rev_import_flags_update(struct obd_import *revimp,
925 struct ptlrpc_request *req)
928 struct obd_connect_data *data;
930 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
932 if (data->ocd_connect_flags & OBD_CONNECT_AT)
933 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
935 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
937 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
939 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
941 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
942 revimp->imp_client->cli_name,
943 libcfs_id2str(req->rq_peer), rc);
951 * Allocate a new reverse import for an export.
953 * \retval -errno in case error hit
954 * \retval 0 if reverse import correctly init
956 int rev_import_init(struct obd_export *export)
958 struct obd_device *obd = export->exp_obd;
959 struct obd_import *revimp;
961 LASSERT(export->exp_imp_reverse == NULL);
963 revimp = class_new_import(obd);
967 revimp->imp_remote_handle.cookie = 0ULL;
968 revimp->imp_client = &obd->obd_ldlm_client;
969 revimp->imp_dlm_fake = 1;
971 /* it is safe to connect import in new state as no sends possible */
972 spin_lock(&export->exp_lock);
973 export->exp_imp_reverse = revimp;
974 spin_unlock(&export->exp_lock);
975 class_import_put(revimp);
979 EXPORT_SYMBOL(rev_import_init);
982 * Handle reconnect for an export.
984 * \param exp export to handle reconnect process
985 * \param req client reconnect request
987 * \retval -rc in case securitfy flavor can't be changed
988 * \retval 0 in case none problems
990 static int rev_import_reconnect(struct obd_export *exp,
991 struct ptlrpc_request *req)
993 struct obd_import *revimp = exp->exp_imp_reverse;
994 struct lustre_handle *lh;
997 /* avoid sending a request until import flags are changed */
998 ptlrpc_import_enter_resend(revimp);
1000 if (revimp->imp_connection != NULL)
1001 ptlrpc_connection_put(revimp->imp_connection);
1004 * client from recovery don't have a handle so we need to take from
1005 * request. it may produce situation when wrong client connected
1006 * to recovery as we trust a client uuid
1008 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1009 revimp->imp_remote_handle = *lh;
1012 * unknown versions will be caught in
1013 * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1015 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1017 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1019 rc = rev_import_flags_update(revimp, req);
1022 * it is safe to still be in RECOVERY phase as we are not able
1023 * to setup correct security flavor so requests are not able to
1024 * be delivered correctly
1029 /* resend all rpc's via new connection */
1030 return ptlrpc_import_recovery_state_machine(revimp);
1033 int target_handle_connect(struct ptlrpc_request *req)
1035 struct obd_device *target = NULL;
1036 struct obd_export *export = NULL;
1038 * connect handle - filled from target_handle_reconnect in
1041 struct lustre_handle conn;
1042 struct lustre_handle *tmp;
1043 struct obd_uuid cluuid;
1048 bool mds_conn = false, lw_client = false, initial_conn = false;
1049 bool mds_mds_conn = false;
1050 bool new_mds_mds_conn = false;
1051 struct obd_connect_data *data, *tmpdata;
1053 lnet_nid_t *client_nid = NULL;
1057 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1059 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1061 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1062 GOTO(out, rc = -EINVAL);
1065 target = class_dev_by_str(str);
1067 deuuidify(str, NULL, &target_start, &target_len);
1068 LCONSOLE_ERROR_MSG(0x137,
1069 "%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",
1070 str, libcfs_nid2str(req->rq_peer.nid));
1071 GOTO(out, rc = -ENODEV);
1074 spin_lock(&target->obd_dev_lock);
1076 target->obd_conn_inprogress++;
1078 if (target->obd_stopping || !target->obd_set_up) {
1079 spin_unlock(&target->obd_dev_lock);
1081 deuuidify(str, NULL, &target_start, &target_len);
1082 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1083 target_len, target_start,
1084 libcfs_nid2str(req->rq_peer.nid),
1085 (target->obd_stopping ?
1086 "stopping" : "not set up"));
1087 GOTO(out, rc = -ENODEV);
1090 if (target->obd_no_conn) {
1091 spin_unlock(&target->obd_dev_lock);
1094 "%s: Temporarily refusing client connection from %s\n",
1095 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1096 GOTO(out, rc = -EAGAIN);
1099 spin_unlock(&target->obd_dev_lock);
1101 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1103 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1104 GOTO(out, rc = -EINVAL);
1107 obd_str2uuid(&cluuid, str);
1109 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1111 GOTO(out, rc = -EPROTO);
1115 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1117 if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1118 GOTO(out, rc = -EPROTO);
1119 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1121 GOTO(out, rc = -EPROTO);
1123 rc = req_capsule_server_pack(&req->rq_pill);
1127 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1129 * Don't allow clients to connect that are using old 1.8 format
1130 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1131 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1132 * FULL20 flag should be set on all connections since 2.0, but no
1133 * longer affects behaviour.
1135 * Later this check will be disabled and the flag can be retired
1136 * completely once interop with 3.0 is no longer needed.
1138 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1139 GOTO(out, rc = -EPROTO);
1142 * Don't allow liblustre clients to connect.
1143 * - testing was disabled in v2_2_50_0-61-g6a75d65
1144 * - building was disabled in v2_5_58_0-28-g7277179
1145 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1146 * - clients were refused connect for version difference > 0.0.1.32
1148 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1149 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1150 GOTO(out, rc = -EPROTO);
1155 * Note: lw_client is needed in MDS-MDS failover during update log
1156 * processing, so we needs to allow lw_client to be connected at
1157 * anytime, instead of only the initial connection
1159 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1161 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1162 initial_conn = true;
1163 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1164 mds_mds_conn = (data->ocd_connect_flags &
1165 OBD_CONNECT_MDS_MDS) != 0;
1168 * OBD_CONNECT_MNE_SWAB is removed at 2.14
1169 * Checking OBD_CONNECT_FID can be removed in the future.
1171 * Via check OBD_CONNECT_FID, we can distinguish whether
1172 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1173 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1176 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1177 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1178 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1179 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1180 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1181 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1184 * We do not support the MDT-MDT interoperations with
1185 * different version MDT because of protocol changes.
1187 if (unlikely(major != LUSTRE_MAJOR ||
1188 minor != LUSTRE_MINOR ||
1189 abs(patch - LUSTRE_PATCH) > 3)) {
1190 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1191 target->obd_name, LUSTRE_MAJOR,
1192 LUSTRE_MINOR, LUSTRE_PATCH,
1193 LUSTRE_FIX, major, minor, patch,
1194 OBD_OCD_VERSION_FIX(data->ocd_version),
1195 libcfs_nid2str(req->rq_peer.nid),
1197 GOTO(out, rc = -EPROTO);
1202 /* lctl gets a backstage, all-access pass. */
1203 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1204 goto dont_check_exports;
1206 export = obd_uuid_lookup(target, &cluuid);
1210 /* We've found an export in the hash. */
1212 spin_lock(&export->exp_lock);
1214 if (export->exp_connecting) { /* b=9635, et. al. */
1215 spin_unlock(&export->exp_lock);
1216 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1217 export->exp_obd->obd_name, export,
1218 libcfs_nid2str(req->rq_peer.nid));
1219 class_export_put(export);
1222 } else if ((mds_conn || (lw_client && initial_conn) ||
1223 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1224 export->exp_connection != NULL) {
1225 spin_unlock(&export->exp_lock);
1226 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1227 /* MDS or LWP reconnected after failover. */
1228 LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1230 mds_conn ? "MDS" : "LWP",
1231 libcfs_nid2str(req->rq_peer.nid),
1232 libcfs_nid2str(export->exp_connection->c_peer.nid));
1234 /* New MDS connection from the same NID. */
1235 LCONSOLE_WARN("%s: Received new %s connection from %s, removing former export from same NID\n",
1237 mds_conn ? "MDS" : "LWP",
1238 libcfs_nid2str(req->rq_peer.nid));
1241 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1242 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1244 * Because exports between MDTs will always be
1245 * kept, let's do not fail such export if they
1246 * come from the same NID, otherwise it might
1247 * cause eviction between MDTs, which might
1248 * cause namespace inconsistency
1250 spin_lock(&export->exp_lock);
1251 export->exp_connecting = 1;
1252 export->exp_conn_cnt = 0;
1253 spin_unlock(&export->exp_lock);
1254 conn.cookie = export->exp_handle.h_cookie;
1257 class_fail_export(export);
1258 class_export_put(export);
1262 } else if (export->exp_connection != NULL && initial_conn &&
1263 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1264 spin_unlock(&export->exp_lock);
1265 /* In MDS failover we have static UUID but NID can change. */
1266 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1267 target->obd_name, cluuid.uuid,
1268 libcfs_nid2str(req->rq_peer.nid),
1270 export->exp_connection->c_peer.nid));
1272 class_export_put(export);
1274 } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1276 spin_unlock(&export->exp_lock);
1279 export->exp_connecting = 1;
1280 spin_unlock(&export->exp_lock);
1281 LASSERT(export->exp_obd == target);
1283 rc = target_handle_reconnect(&conn, export, &cluuid);
1286 /* If we found an export, we already unlocked. */
1289 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1290 } else if (req->rq_export == NULL &&
1291 atomic_read(&export->exp_rpc_count) > 0) {
1292 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1293 target->obd_name, cluuid.uuid,
1294 libcfs_nid2str(req->rq_peer.nid),
1295 refcount_read(&export->exp_handle.h_ref));
1296 GOTO(out, rc = -EBUSY);
1297 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1299 if (!strstr(cluuid.uuid, "mdt"))
1300 LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client",
1301 target->obd_name, cluuid.uuid,
1302 libcfs_nid2str(req->rq_peer.nid));
1303 GOTO(out, rc = -EALREADY);
1305 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1311 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1312 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1313 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1314 export, ktime_get_seconds(),
1315 export ? export->exp_last_request_time : 0);
1318 * If this is the first time a client connects, reset the recovery
1319 * timer. Discard lightweight connections which might be local.
1321 if (!lw_client && rc == 0 && target->obd_recovering)
1322 check_and_start_recovery_timer(target, req, export == NULL);
1325 * We want to handle EALREADY but *not* -EALREADY from
1326 * target_handle_reconnect(), return reconnection state in a flag.
1328 if (rc == EALREADY) {
1329 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1335 /* Tell the client if we support replayable requests. */
1336 if (target->obd_replayable)
1337 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1338 client_nid = &req->rq_peer.nid;
1340 if (export == NULL) {
1341 /* allow lightweight connections during recovery */
1343 * allow "new" MDT to be connected during recovery, since we
1344 * need retrieve recovery update records from it
1346 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1347 struct hrtimer *timer = &target->obd_recovery_timer;
1356 connected = atomic_read(&target->obd_connected_clients);
1357 in_progress = atomic_read(&target->obd_lock_replay_clients);
1359 atomic_read(&target->obd_max_recoverable_clients);
1360 stale = target->obd_stale_clients;
1361 remaining = hrtimer_expires_remaining(timer);
1362 left = ktime_divns(remaining, NSEC_PER_SEC);
1364 if (ktime_to_ns(remaining) > 0) {
1365 msg = "to recover in";
1368 msg = "already passed deadline";
1371 target_check_recovery_timer(target);
1374 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",
1375 target->obd_name, cluuid.uuid,
1376 libcfs_nid2str(req->rq_peer.nid), known,
1377 connected - in_progress, in_progress,
1378 stale, msg, timeout / 60, timeout % 60);
1382 rc = obd_connect(req->rq_svc_thread->t_env,
1383 &export, target, &cluuid, data,
1385 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1386 lustre_msg_add_op_flags(req->rq_repmsg,
1387 MSG_CONNECT_RECOVERING);
1389 conn.cookie = export->exp_handle.h_cookie;
1390 rc = rev_import_init(export);
1394 new_mds_mds_conn = true;
1397 rc = obd_reconnect(req->rq_svc_thread->t_env,
1398 export, target, &cluuid, data, client_nid);
1403 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1404 data->ocd_instance = target->u.obt.obt_instance;
1407 * Return only the parts of obd_connect_data that we understand, so the
1408 * client knows that we don't understand the rest.
1411 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1413 tmpdata = req_capsule_server_get(&req->rq_pill,
1416 * Don't use struct assignment here, because the client reply
1417 * buffer may be smaller/larger than the local struct
1420 memcpy(tmpdata, data, min(tmpsize, size));
1424 * If the client and the server are the same node, we will already
1425 * have an export that really points to the client's DLM export,
1426 * because we have a shared handles table.
1428 * XXX this will go away when shaver stops sending the "connect" handle
1429 * in the real "remote handle" field of the request --phik 24 Apr 2003
1431 ptlrpc_request_change_export(req, export);
1433 spin_lock(&export->exp_lock);
1434 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1435 spin_unlock(&export->exp_lock);
1437 "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1438 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1439 export->exp_conn_cnt,
1440 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1442 GOTO(out, rc = -EALREADY);
1444 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1445 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1446 spin_unlock(&export->exp_lock);
1448 if (export->exp_connection != NULL) {
1449 /* Check to see if connection came from another NID. */
1450 if (export->exp_connection->c_peer.nid != req->rq_peer.nid)
1451 obd_nid_del(export->exp_obd, export);
1453 ptlrpc_connection_put(export->exp_connection);
1456 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1459 obd_nid_add(export->exp_obd, export);
1461 lustre_msg_set_handle(req->rq_repmsg, &conn);
1463 rc = rev_import_reconnect(export, req);
1467 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1469 __u64 transno = data->ocd_transno;
1471 spin_lock(&export->exp_lock);
1473 * possible race with class_disconnect_stale_exports,
1474 * export may be already in the eviction process
1476 if (export->exp_failed) {
1477 spin_unlock(&export->exp_lock);
1478 GOTO(out, rc = -ENODEV);
1480 export->exp_in_recovery = 1;
1481 export->exp_req_replay_needed = 1;
1482 export->exp_lock_replay_needed = 1;
1483 spin_unlock(&export->exp_lock);
1485 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1486 MSG_CONNECT_TRANSNO);
1487 if (has_transno && transno == 0)
1488 CWARN("Connect with zero transno!\n");
1490 if (has_transno && transno > 0 &&
1491 transno < target->obd_next_recovery_transno &&
1492 transno > target->obd_last_committed) {
1493 /* Another way is to use cmpxchg() to be lock-free. */
1494 spin_lock(&target->obd_recovery_task_lock);
1495 if (transno < target->obd_next_recovery_transno)
1496 target->obd_next_recovery_transno = transno;
1497 spin_unlock(&target->obd_recovery_task_lock);
1500 atomic_inc(&target->obd_req_replay_clients);
1501 atomic_inc(&target->obd_lock_replay_clients);
1503 * Note: MDS-MDS connection is allowed to be connected during
1504 * recovery, no matter if the exports needs to be recoveried.
1505 * Because we need retrieve updates logs from all other MDTs.
1506 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1507 * also needs to be increased to match other recovery checking
1510 if (new_mds_mds_conn)
1511 atomic_inc(&target->obd_max_recoverable_clients);
1513 if (atomic_inc_return(&target->obd_connected_clients) ==
1514 atomic_read(&target->obd_max_recoverable_clients))
1515 wake_up(&target->obd_next_transno_waitq);
1518 /* Tell the client we're in recovery, when client is involved in it. */
1519 if (target->obd_recovering && !lw_client)
1520 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1524 spin_lock(&export->exp_lock);
1525 export->exp_connecting = 0;
1526 spin_unlock(&export->exp_lock);
1528 class_export_put(export);
1530 if (target != NULL) {
1531 spin_lock(&target->obd_dev_lock);
1532 target->obd_conn_inprogress--;
1533 spin_unlock(&target->obd_dev_lock);
1534 class_decref(target, "find", current);
1536 req->rq_status = rc;
1540 int target_handle_disconnect(struct ptlrpc_request *req)
1546 rc = req_capsule_server_pack(&req->rq_pill);
1550 /* Keep the rq_export around so we can send the reply. */
1551 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1556 void target_destroy_export(struct obd_export *exp)
1558 struct obd_import *imp = NULL;
1560 * exports created from last_rcvd data, and "fake"
1561 * exports created by lctl don't have an import
1563 spin_lock(&exp->exp_lock);
1564 if (exp->exp_imp_reverse != NULL) {
1565 imp = exp->exp_imp_reverse;
1566 exp->exp_imp_reverse = NULL;
1568 spin_unlock(&exp->exp_lock);
1570 client_destroy_import(imp);
1572 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1573 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1574 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1575 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1577 EXPORT_SYMBOL(target_destroy_export);
1580 * Recovery functions
1582 static void target_request_copy_get(struct ptlrpc_request *req)
1584 class_export_rpc_inc(req->rq_export);
1585 LASSERT(list_empty(&req->rq_list));
1586 INIT_LIST_HEAD(&req->rq_replay_list);
1588 /* Increase refcount to keep request in queue. */
1589 atomic_inc(&req->rq_refcount);
1590 /* Let export know it has replays to be handled. */
1591 atomic_inc(&req->rq_export->exp_replay_count);
1594 static void target_request_copy_put(struct ptlrpc_request *req)
1596 LASSERT(list_empty(&req->rq_replay_list));
1597 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1599 atomic_dec(&req->rq_export->exp_replay_count);
1600 class_export_rpc_dec(req->rq_export);
1601 ptlrpc_server_drop_request(req);
1604 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1606 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1607 struct obd_export *exp = req->rq_export;
1608 struct ptlrpc_request *reqiter;
1609 struct ptlrpc_request *dup_req = NULL;
1614 spin_lock(&exp->exp_lock);
1615 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1617 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1625 /* We expect it with RESENT and REPLAY flags. */
1626 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1627 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1628 CERROR("invalid flags %x of resent replay\n",
1629 lustre_msg_get_flags(req->rq_reqmsg));
1631 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1634 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1636 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1637 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1641 list_add_tail(&req->rq_replay_list,
1642 &exp->exp_req_replay_queue);
1645 spin_unlock(&exp->exp_lock);
1649 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1651 LASSERT(!list_empty(&req->rq_replay_list));
1652 LASSERT(req->rq_export);
1654 spin_lock(&req->rq_export->exp_lock);
1655 list_del_init(&req->rq_replay_list);
1656 spin_unlock(&req->rq_export->exp_lock);
1659 static void target_finish_recovery(struct lu_target *lut)
1661 struct obd_device *obd = lut->lut_obd;
1665 /* Only log a recovery message when recovery has occurred. */
1666 if (obd->obd_recovery_start) {
1667 time64_t now = ktime_get_seconds();
1668 time64_t elapsed_time;
1670 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1672 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1673 obd->obd_name, elapsed_time / 60,
1675 atomic_read(&obd->obd_max_recoverable_clients),
1676 atomic_read(&obd->obd_connected_clients),
1677 obd->obd_stale_clients,
1678 obd->obd_stale_clients == 1 ? "was" : "were");
1681 ldlm_reprocess_recovery_done(obd->obd_namespace);
1682 spin_lock(&obd->obd_recovery_task_lock);
1683 if (!list_empty(&obd->obd_req_replay_queue) ||
1684 !list_empty(&obd->obd_lock_replay_queue) ||
1685 !list_empty(&obd->obd_final_req_queue)) {
1686 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1688 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1689 list_empty(&obd->obd_lock_replay_queue) ? \
1691 list_empty(&obd->obd_final_req_queue) ? \
1693 spin_unlock(&obd->obd_recovery_task_lock);
1696 spin_unlock(&obd->obd_recovery_task_lock);
1698 obd->obd_recovery_end = ktime_get_seconds();
1700 /* When recovery finished, cleanup orphans on MDS and OST. */
1701 if (obd->obd_type && OBP(obd, postrecov)) {
1702 int rc = OBP(obd, postrecov)(obd);
1705 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1711 static void abort_req_replay_queue(struct obd_device *obd)
1713 struct ptlrpc_request *req, *n;
1714 LIST_HEAD(abort_list);
1716 spin_lock(&obd->obd_recovery_task_lock);
1717 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1718 spin_unlock(&obd->obd_recovery_task_lock);
1719 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1720 DEBUG_REQ(D_WARNING, req, "aborted:");
1721 req->rq_status = -ENOTCONN;
1722 if (ptlrpc_error(req)) {
1723 DEBUG_REQ(D_ERROR, req,
1724 "failed abort_req_reply; skipping");
1726 target_exp_dequeue_req_replay(req);
1727 target_request_copy_put(req);
1731 static void abort_lock_replay_queue(struct obd_device *obd)
1733 struct ptlrpc_request *req, *n;
1734 LIST_HEAD(abort_list);
1736 spin_lock(&obd->obd_recovery_task_lock);
1737 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1738 spin_unlock(&obd->obd_recovery_task_lock);
1739 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1740 DEBUG_REQ(D_ERROR, req, "aborted:");
1741 req->rq_status = -ENOTCONN;
1742 if (ptlrpc_error(req)) {
1743 DEBUG_REQ(D_ERROR, req,
1744 "failed abort_lock_reply; skipping");
1746 target_request_copy_put(req);
1751 * Called from a cleanup function if the device is being cleaned up
1752 * forcefully. The exports should all have been disconnected already,
1753 * the only thing left to do is
1754 * - clear the recovery flags
1755 * - cancel the timer
1756 * - free queued requests and replies, but don't send replies
1757 * Because the obd_stopping flag is set, no new requests should be received.
1759 void target_cleanup_recovery(struct obd_device *obd)
1761 struct ptlrpc_request *req, *n;
1762 LIST_HEAD(clean_list);
1764 spin_lock(&obd->obd_dev_lock);
1765 if (!obd->obd_recovering) {
1766 spin_unlock(&obd->obd_dev_lock);
1770 obd->obd_recovering = obd->obd_abort_recovery = 0;
1771 spin_unlock(&obd->obd_dev_lock);
1773 spin_lock(&obd->obd_recovery_task_lock);
1774 target_cancel_recovery_timer(obd);
1775 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1776 spin_unlock(&obd->obd_recovery_task_lock);
1778 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1779 LASSERT(req->rq_reply_state == NULL);
1780 target_exp_dequeue_req_replay(req);
1781 target_request_copy_put(req);
1784 spin_lock(&obd->obd_recovery_task_lock);
1785 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1786 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1787 spin_unlock(&obd->obd_recovery_task_lock);
1789 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1790 LASSERT(req->rq_reply_state == NULL);
1791 target_request_copy_put(req);
1796 EXPORT_SYMBOL(target_cleanup_recovery);
1798 /* obd_recovery_task_lock should be held */
1799 void target_cancel_recovery_timer(struct obd_device *obd)
1801 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1802 hrtimer_cancel(&obd->obd_recovery_timer);
1805 static void target_start_recovery_timer(struct obd_device *obd)
1809 if (obd->obd_recovery_start != 0)
1812 spin_lock(&obd->obd_dev_lock);
1813 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1814 spin_unlock(&obd->obd_dev_lock);
1818 LASSERT(obd->obd_recovery_timeout != 0);
1820 if (obd->obd_recovery_start != 0) {
1821 spin_unlock(&obd->obd_dev_lock);
1825 obd->obd_recovery_start = ktime_get_seconds();
1826 delay = ktime_set(obd->obd_recovery_start +
1827 obd->obd_recovery_timeout, 0);
1828 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1829 spin_unlock(&obd->obd_dev_lock);
1831 LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1833 obd->obd_recovery_timeout / 60,
1834 obd->obd_recovery_timeout % 60,
1835 atomic_read(&obd->obd_max_recoverable_clients),
1836 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1838 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1843 * extend recovery window.
1845 * if @extend is true, extend recovery window to have @dr_timeout remaining
1846 * at least; otherwise, make sure the recovery timeout value is not less
1849 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1856 spin_lock(&obd->obd_dev_lock);
1857 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1858 obd->obd_stopping) {
1859 spin_unlock(&obd->obd_dev_lock);
1862 LASSERT(obd->obd_recovery_start != 0);
1864 left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1865 left = ktime_divns(left_ns, NSEC_PER_SEC);
1868 timeout = obd->obd_recovery_timeout;
1869 /* dr_timeout will happen after the hrtimer has expired.
1870 * Add the excess time to the soft recovery timeout without
1871 * exceeding the hard recovery timeout.
1873 if (dr_timeout > left) {
1874 timeout += dr_timeout - left;
1875 timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1879 timeout = clamp_t(timeout_t, dr_timeout,
1880 obd->obd_recovery_timeout,
1881 obd->obd_recovery_time_hard);
1884 if (timeout == obd->obd_recovery_time_hard)
1885 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1886 obd->obd_name, timeout, extend);
1888 if (obd->obd_recovery_timeout < timeout) {
1891 obd->obd_recovery_timeout = timeout;
1892 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1893 now = ktime_set(ktime_get_seconds(), 0);
1894 left_ns = ktime_sub(end, now);
1895 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1896 left = ktime_divns(left_ns, NSEC_PER_SEC);
1898 spin_unlock(&obd->obd_dev_lock);
1900 CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
1901 obd->obd_name, left);
1904 /* Reset the timer with each new client connection */
1906 * This timer is actually reconnect_timer, which is for making sure
1907 * the total recovery window is at least as big as my reconnect
1908 * attempt timing. So the initial recovery time_out will be set to
1909 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1910 * from client is bigger than this, then the recovery time_out will
1911 * be extended to make sure the client could be reconnected, in the
1912 * process, the timeout from the new client should be ignored.
1915 check_and_start_recovery_timer(struct obd_device *obd,
1916 struct ptlrpc_request *req,
1919 timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
1920 struct obd_device_target *obt = &obd->u.obt;
1922 if (!new_client && service_timeout)
1924 * Teach server about old server's estimates, as first guess
1925 * at how long new requests will take.
1927 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1930 target_start_recovery_timer(obd);
1933 * Convert the service time to RPC timeout,
1934 * and reuse service_timeout to limit stack usage.
1936 service_timeout = at_est2timeout(service_timeout);
1938 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1939 service_timeout < at_extra)
1940 service_timeout = at_extra;
1943 * We expect other clients to timeout within service_timeout, then try
1944 * to reconnect, then try the failover server. The max delay between
1945 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
1947 service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
1949 LASSERT(obt->obt_magic == OBT_MAGIC);
1950 service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1951 if (service_timeout > obd->obd_recovery_timeout && !new_client)
1952 extend_recovery_timer(obd, service_timeout, false);
1955 /** Health checking routines */
1956 static inline int exp_connect_healthy(struct obd_export *exp)
1958 return exp->exp_in_recovery;
1961 /** if export done req_replay or has replay in queue */
1962 static inline int exp_req_replay_healthy(struct obd_export *exp)
1964 return (!exp->exp_req_replay_needed ||
1965 atomic_read(&exp->exp_replay_count) > 0);
1969 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1971 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1972 exp_req_replay_healthy(exp);
1975 /** if export done lock_replay or has replay in queue */
1976 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1978 return (!exp->exp_lock_replay_needed ||
1979 atomic_read(&exp->exp_replay_count) > 0);
1982 static inline int exp_vbr_healthy(struct obd_export *exp)
1984 return !exp->exp_vbr_failed;
1987 static inline int exp_finished(struct obd_export *exp)
1989 return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
1992 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1994 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1998 static int check_for_next_transno(struct lu_target *lut)
2000 struct ptlrpc_request *req = NULL;
2001 struct obd_device *obd = lut->lut_obd;
2002 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2003 int wake_up = 0, connected, completed, queue_len;
2004 __u64 req_transno = 0;
2005 __u64 update_transno = 0;
2006 __u64 next_transno = 0;
2010 spin_lock(&obd->obd_recovery_task_lock);
2011 if (!list_empty(&obd->obd_req_replay_queue)) {
2012 req = list_entry(obd->obd_req_replay_queue.next,
2013 struct ptlrpc_request, rq_list);
2014 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2018 update_transno = distribute_txn_get_next_transno(tdtd);
2020 connected = atomic_read(&obd->obd_connected_clients);
2021 completed = connected - atomic_read(&obd->obd_req_replay_clients);
2022 queue_len = obd->obd_requests_queued_for_recovery;
2023 next_transno = obd->obd_next_recovery_transno;
2026 "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2027 atomic_read(&obd->obd_max_recoverable_clients),
2028 connected, completed,
2029 queue_len, req_transno, next_transno);
2031 if (obd->obd_abort_recovery) {
2032 CDEBUG(D_HA, "waking for aborted recovery\n");
2034 } else if (obd->obd_recovery_expired) {
2035 CDEBUG(D_HA, "waking for expired recovery\n");
2037 } else if (tdtd != NULL && req != NULL &&
2038 is_req_replayed_by_update(req)) {
2039 LASSERTF(req_transno < next_transno,
2040 "req_transno %llu next_transno%llu\n", req_transno,
2042 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2045 } else if (req_transno == next_transno ||
2046 (update_transno != 0 && update_transno <= next_transno)) {
2047 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2049 } else if (queue_len > 0 &&
2050 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2051 /** handle gaps occured due to lost reply or VBR */
2052 LASSERTF(req_transno >= next_transno,
2053 "req_transno: %llu, next_transno: %llu\n",
2054 req_transno, next_transno);
2056 "%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",
2057 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2058 next_transno, queue_len, completed, connected,
2059 req_transno, update_transno, obd->obd_last_committed);
2060 obd->obd_next_recovery_transno = req_transno;
2062 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2063 CDEBUG(D_HA, "waking for completed recovery\n");
2065 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2067 "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2069 obd->obd_next_recovery_transno = req_transno;
2072 spin_unlock(&obd->obd_recovery_task_lock);
2076 static int check_for_next_lock(struct lu_target *lut)
2078 struct obd_device *obd = lut->lut_obd;
2081 spin_lock(&obd->obd_recovery_task_lock);
2082 if (!list_empty(&obd->obd_lock_replay_queue)) {
2083 CDEBUG(D_HA, "waking for next lock\n");
2085 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2086 CDEBUG(D_HA, "waking for completed lock replay\n");
2088 } else if (obd->obd_abort_recovery) {
2089 CDEBUG(D_HA, "waking for aborted recovery\n");
2091 } else if (obd->obd_recovery_expired) {
2092 CDEBUG(D_HA, "waking for expired recovery\n");
2095 spin_unlock(&obd->obd_recovery_task_lock);
2101 * wait for recovery events,
2102 * check its status with help of check_routine
2103 * evict dead clients via health_check
2105 static int target_recovery_overseer(struct lu_target *lut,
2106 int (*check_routine)(struct lu_target *),
2107 int (*health_check)(struct obd_export *))
2109 struct obd_device *obd = lut->lut_obd;
2110 struct target_distribute_txn_data *tdtd;
2114 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2115 now = ktime_get_seconds();
2116 if (now - last > 600) {
2117 LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2122 if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2123 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2124 __u64 next_update_transno = 0;
2127 * Only abort the recovery if there are no update recovery
2130 spin_lock(&obd->obd_recovery_task_lock);
2131 if (lut->lut_tdtd != NULL) {
2132 next_update_transno =
2133 distribute_txn_get_next_transno(lut->lut_tdtd);
2135 tdtd = lut->lut_tdtd;
2137 * If next_update_transno == 0, it probably because
2138 * updatelog retrieve threads did not get any records
2139 * yet, let's wait those threads stopped
2141 if (next_update_transno == 0) {
2142 spin_unlock(&obd->obd_recovery_task_lock);
2144 tdtd->tdtd_recovery_threads_waitq,
2145 atomic_read(&tdtd->tdtd_recovery_threads_count)
2148 spin_lock(&obd->obd_recovery_task_lock);
2149 next_update_transno =
2150 distribute_txn_get_next_transno(
2155 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2156 obd->obd_next_recovery_transno = next_update_transno;
2157 spin_unlock(&obd->obd_recovery_task_lock);
2159 * Disconnect unfinished exports from clients, and
2160 * keep connection from MDT to make sure the update
2161 * recovery will still keep trying until some one
2162 * manually abort the recovery
2164 class_disconnect_stale_exports(obd,
2165 exp_finished_or_from_mdt);
2166 /* Abort all of replay & replay lock req from clients */
2167 abort_req_replay_queue(obd);
2168 abort_lock_replay_queue(obd);
2170 "%s: there are still update replay (%#llx)in the queue.\n",
2171 obd->obd_name, next_update_transno);
2173 obd->obd_abort_recovery = 1;
2174 spin_unlock(&obd->obd_recovery_task_lock);
2175 CWARN("%s recovery is aborted by hard timeout\n",
2180 while (wait_event_timeout(obd->obd_next_transno_waitq,
2182 cfs_time_seconds(60)) == 0)
2183 ; /* wait indefinitely for event, but don't trigger watchdog */
2185 if (obd->obd_abort_recovery) {
2186 CWARN("recovery is aborted, evict exports in recovery\n");
2187 if (lut->lut_tdtd != NULL) {
2188 tdtd = lut->lut_tdtd;
2190 * Let's wait all of the update log recovery thread
2194 tdtd->tdtd_recovery_threads_waitq,
2195 atomic_read(&tdtd->tdtd_recovery_threads_count)
2197 /* Then abort the update recovery list */
2198 dtrq_list_destroy(lut->lut_tdtd);
2201 /** evict exports which didn't finish recovery yet */
2202 class_disconnect_stale_exports(obd, exp_finished);
2204 } else if (obd->obd_recovery_expired &&
2205 obd->obd_recovery_timeout < obd->obd_recovery_time_hard) {
2206 obd->obd_recovery_expired = 0;
2208 /** If some clients died being recovered, evict them */
2209 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2211 /** evict cexports with no replay in queue, they are stalled */
2212 class_disconnect_stale_exports(obd, health_check);
2214 /** continue with VBR */
2215 spin_lock(&obd->obd_dev_lock);
2216 obd->obd_version_recov = 1;
2217 spin_unlock(&obd->obd_dev_lock);
2219 * reset timer, recovery will proceed with versions now,
2220 * timeout is set just to handle reconnection delays
2222 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2224 * Wait for recovery events again, after evicting bad clients
2231 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2233 struct obd_device *obd = lut->lut_obd;
2234 struct ptlrpc_request *req = NULL;
2236 CDEBUG(D_HA, "Waiting for lock\n");
2237 if (target_recovery_overseer(lut, check_for_next_lock,
2238 exp_lock_replay_healthy))
2239 abort_lock_replay_queue(obd);
2241 spin_lock(&obd->obd_recovery_task_lock);
2242 if (!list_empty(&obd->obd_lock_replay_queue)) {
2243 req = list_entry(obd->obd_lock_replay_queue.next,
2244 struct ptlrpc_request, rq_list);
2245 list_del_init(&req->rq_list);
2246 spin_unlock(&obd->obd_recovery_task_lock);
2248 spin_unlock(&obd->obd_recovery_task_lock);
2249 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2250 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2251 /** evict exports failed VBR */
2252 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2257 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2259 struct ptlrpc_request *req = NULL;
2261 spin_lock(&obd->obd_recovery_task_lock);
2262 if (!list_empty(&obd->obd_final_req_queue)) {
2263 req = list_entry(obd->obd_final_req_queue.next,
2264 struct ptlrpc_request, rq_list);
2265 list_del_init(&req->rq_list);
2266 spin_unlock(&obd->obd_recovery_task_lock);
2267 if (req->rq_export->exp_in_recovery) {
2268 spin_lock(&req->rq_export->exp_lock);
2269 req->rq_export->exp_in_recovery = 0;
2270 spin_unlock(&req->rq_export->exp_lock);
2273 spin_unlock(&obd->obd_recovery_task_lock);
2278 static void handle_recovery_req(struct ptlrpc_thread *thread,
2279 struct ptlrpc_request *req,
2280 svc_handler_t handler)
2285 * export can be evicted during recovery, no need to handle replays for
2286 * it after that, discard such request silently
2288 if (req->rq_export->exp_disconnected)
2291 req->rq_session.lc_thread = thread;
2292 req->rq_svc_thread = thread;
2293 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2295 /* thread context */
2296 lu_context_enter(&thread->t_env->le_ctx);
2298 lu_context_exit(&thread->t_env->le_ctx);
2300 req->rq_svc_thread->t_env->le_ses = NULL;
2302 /* don't reset timer for final stage */
2303 if (!exp_finished(req->rq_export)) {
2304 timeout_t timeout = obd_timeout;
2307 * Add request @timeout to the recovery time so next request from
2308 * this client may come in recovery time
2311 struct ptlrpc_service_part *svcpt;
2312 timeout_t est_timeout;
2314 svcpt = req->rq_rqbd->rqbd_svcpt;
2316 * If the server sent early reply for this request,
2317 * the client will recalculate the timeout according to
2318 * current server estimate service time, so we will
2319 * use the maxium timeout here for waiting the client
2320 * sending the next req
2322 est_timeout = at_get(&svcpt->scp_at_estimate);
2323 timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2324 lustre_msg_get_timeout(req->rq_reqmsg));
2326 * Add 2 net_latency, one for balance rq_deadline
2327 * (see ptl_send_rpc), one for resend the req to server,
2328 * Note: client will pack net_latency in replay req
2329 * (see ptlrpc_replay_req)
2331 timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2333 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2339 /** Checking routines for recovery */
2340 static int check_for_recovery_ready(struct lu_target *lut)
2342 struct obd_device *obd = lut->lut_obd;
2343 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2346 "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2347 clnts, obd->obd_stale_clients,
2348 atomic_read(&obd->obd_max_recoverable_clients),
2349 obd->obd_abort_recovery, obd->obd_recovery_expired);
2351 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2353 atomic_read(&obd->obd_max_recoverable_clients));
2354 if (clnts + obd->obd_stale_clients <
2355 atomic_read(&obd->obd_max_recoverable_clients))
2359 if (lut->lut_tdtd != NULL) {
2360 if (!lut->lut_tdtd->tdtd_replay_ready &&
2361 !obd->obd_abort_recovery && !obd->obd_stopping) {
2363 * Let's extend recovery timer, in case the recovery
2364 * timer expired, and some clients got evicted
2366 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2369 "%s update recovery is not ready, extend recovery %d\n",
2370 obd->obd_name, obd->obd_recovery_timeout);
2379 REQUEST_RECOVERY = 1,
2380 UPDATE_RECOVERY = 2,
2383 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2387 if (!list_empty(&obd->obd_req_replay_queue)) {
2388 struct ptlrpc_request *req;
2390 req = list_entry(obd->obd_req_replay_queue.next,
2391 struct ptlrpc_request, rq_list);
2392 transno = lustre_msg_get_transno(req->rq_reqmsg);
2398 static __u64 get_next_transno(struct lu_target *lut, int *type)
2400 struct obd_device *obd = lut->lut_obd;
2401 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2403 __u64 update_transno;
2407 transno = get_next_replay_req_transno(obd);
2409 *type = REQUEST_RECOVERY;
2414 update_transno = distribute_txn_get_next_transno(tdtd);
2415 if (transno == 0 || (transno >= update_transno &&
2416 update_transno != 0)) {
2417 transno = update_transno;
2419 *type = UPDATE_RECOVERY;
2426 * drop duplicate replay request
2428 * Because the operation has been replayed by update recovery, the request
2429 * with the same transno will be dropped and also notify the client to send
2430 * next replay request.
2432 * \param[in] env execution environment
2433 * \param[in] obd failover obd device
2434 * \param[in] req request to be dropped
2436 static void drop_duplicate_replay_req(struct lu_env *env,
2437 struct obd_device *obd,
2438 struct ptlrpc_request *req)
2440 DEBUG_REQ(D_HA, req,
2441 "remove t%lld from %s because duplicate update records found",
2442 lustre_msg_get_transno(req->rq_reqmsg),
2443 libcfs_nid2str(req->rq_peer.nid));
2446 * Right now, only for MDS reint operation update replay and
2447 * normal request replay can have the same transno
2449 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2450 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2451 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2452 if (likely(req->rq_export))
2453 target_committed_to_req(req);
2454 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2455 target_send_reply(req, req->rq_status, 0);
2457 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2458 libcfs_nid2str(req->rq_peer.nid));
2460 target_exp_dequeue_req_replay(req);
2461 target_request_copy_put(req);
2462 obd->obd_replayed_requests++;
2465 static void replay_request_or_update(struct lu_env *env,
2466 struct lu_target *lut,
2467 struct target_recovery_data *trd,
2468 struct ptlrpc_thread *thread)
2470 struct obd_device *obd = lut->lut_obd;
2471 struct ptlrpc_request *req = NULL;
2477 CDEBUG(D_HA, "Waiting for transno %lld\n",
2478 obd->obd_next_recovery_transno);
2480 /* Replay all of request and update by transno */
2482 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2484 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2487 * It is needed to extend recovery window above
2488 * recovery_time_soft. Extending is possible only in the
2489 * end of recovery window (see more details in
2490 * handle_recovery_req()).
2492 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2494 if (target_recovery_overseer(lut, check_for_next_transno,
2495 exp_req_replay_healthy_or_from_mdt)) {
2496 abort_req_replay_queue(obd);
2497 abort_lock_replay_queue(obd);
2501 spin_lock(&obd->obd_recovery_task_lock);
2502 transno = get_next_transno(lut, &type);
2503 if (type == REQUEST_RECOVERY && transno != 0) {
2505 * Drop replay request from client side, if the
2506 * replay has been executed by update with the
2509 req = list_entry(obd->obd_req_replay_queue.next,
2510 struct ptlrpc_request, rq_list);
2512 list_del_init(&req->rq_list);
2513 obd->obd_requests_queued_for_recovery--;
2514 spin_unlock(&obd->obd_recovery_task_lock);
2517 * Let's check if the request has been redone by
2520 if (is_req_replayed_by_update(req)) {
2521 struct distribute_txn_replay_req *dtrq;
2523 dtrq = distribute_txn_lookup_finish_list(tdtd,
2525 LASSERT(dtrq != NULL);
2526 spin_lock(&tdtd->tdtd_replay_list_lock);
2527 list_del_init(&dtrq->dtrq_list);
2528 spin_unlock(&tdtd->tdtd_replay_list_lock);
2531 drop_duplicate_replay_req(env, obd, req);
2536 LASSERT(trd->trd_processing_task == current->pid);
2537 DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2539 lustre_msg_get_transno(req->rq_reqmsg),
2540 libcfs_nid2str(req->rq_peer.nid));
2542 handle_recovery_req(thread, req,
2543 trd->trd_recovery_handler);
2545 * bz18031: increase next_recovery_transno before
2546 * target_request_copy_put() will drop exp_rpc reference
2548 spin_lock(&obd->obd_recovery_task_lock);
2549 obd->obd_next_recovery_transno++;
2550 spin_unlock(&obd->obd_recovery_task_lock);
2551 target_exp_dequeue_req_replay(req);
2552 target_request_copy_put(req);
2553 obd->obd_replayed_requests++;
2554 } else if (type == UPDATE_RECOVERY && transno != 0) {
2555 struct distribute_txn_replay_req *dtrq;
2558 spin_unlock(&obd->obd_recovery_task_lock);
2560 LASSERT(tdtd != NULL);
2561 dtrq = distribute_txn_get_next_req(tdtd);
2562 lu_context_enter(&thread->t_env->le_ctx);
2563 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2564 lu_context_exit(&thread->t_env->le_ctx);
2565 extend_recovery_timer(obd, obd_timeout, true);
2567 if (rc == 0 && dtrq->dtrq_xid != 0) {
2569 "Move x%llu t%llu to finish list\n",
2571 dtrq->dtrq_master_transno);
2573 /* Add it to the replay finish list */
2574 spin_lock(&tdtd->tdtd_replay_list_lock);
2575 list_add(&dtrq->dtrq_list,
2576 &tdtd->tdtd_replay_finish_list);
2577 spin_unlock(&tdtd->tdtd_replay_list_lock);
2579 spin_lock(&obd->obd_recovery_task_lock);
2580 if (transno == obd->obd_next_recovery_transno)
2581 obd->obd_next_recovery_transno++;
2583 obd->obd_next_recovery_transno)
2584 obd->obd_next_recovery_transno =
2586 spin_unlock(&obd->obd_recovery_task_lock);
2591 spin_unlock(&obd->obd_recovery_task_lock);
2593 LASSERT(list_empty(&obd->obd_req_replay_queue));
2594 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2595 /** evict exports failed VBR */
2596 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2602 static int target_recovery_thread(void *arg)
2604 struct lu_target *lut = arg;
2605 struct obd_device *obd = lut->lut_obd;
2606 struct ptlrpc_request *req;
2607 struct target_recovery_data *trd = &obd->obd_recovery_data;
2608 unsigned long delta;
2610 struct ptlrpc_thread *thread = NULL;
2615 unshare_fs_struct();
2616 OBD_ALLOC_PTR(thread);
2622 GOTO(out_thread, rc = -ENOMEM);
2623 rc = lu_env_add(env);
2627 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2629 GOTO(out_env_remove, rc);
2631 thread->t_env = env;
2632 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2633 env->le_ctx.lc_thread = thread;
2634 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2636 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2638 trd->trd_processing_task = current->pid;
2640 spin_lock(&obd->obd_dev_lock);
2641 obd->obd_recovering = 1;
2642 spin_unlock(&obd->obd_dev_lock);
2643 complete(&trd->trd_starting);
2645 /* first of all, we have to know the first transno to replay */
2646 if (target_recovery_overseer(lut, check_for_recovery_ready,
2647 exp_connect_healthy)) {
2648 abort_req_replay_queue(obd);
2649 abort_lock_replay_queue(obd);
2650 if (lut->lut_tdtd != NULL)
2651 dtrq_list_destroy(lut->lut_tdtd);
2654 /* next stage: replay requests or update */
2656 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2657 atomic_read(&obd->obd_req_replay_clients),
2658 obd->obd_next_recovery_transno);
2659 replay_request_or_update(env, lut, trd, thread);
2662 * The second stage: replay locks
2664 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2665 atomic_read(&obd->obd_lock_replay_clients));
2666 while ((req = target_next_replay_lock(lut))) {
2667 LASSERT(trd->trd_processing_task == current->pid);
2668 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2669 libcfs_nid2str(req->rq_peer.nid));
2670 handle_recovery_req(thread, req,
2671 trd->trd_recovery_handler);
2672 target_request_copy_put(req);
2673 obd->obd_replayed_locks++;
2677 * The third stage: reply on final pings, at this moment all clients
2678 * must have request in final queue
2680 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2681 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2682 /** Update server last boot epoch */
2683 tgt_boot_epoch_update(lut);
2685 * We drop recoverying flag to forward all new requests
2686 * to regular mds_handle() since now
2688 spin_lock(&obd->obd_dev_lock);
2689 obd->obd_recovering = obd->obd_abort_recovery = 0;
2690 spin_unlock(&obd->obd_dev_lock);
2691 spin_lock(&obd->obd_recovery_task_lock);
2692 target_cancel_recovery_timer(obd);
2693 spin_unlock(&obd->obd_recovery_task_lock);
2694 while ((req = target_next_final_ping(obd))) {
2695 LASSERT(trd->trd_processing_task == current->pid);
2696 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2697 libcfs_nid2str(req->rq_peer.nid));
2698 handle_recovery_req(thread, req,
2699 trd->trd_recovery_handler);
2701 * Because the waiting client can not send ping to server,
2702 * so we need refresh the last_request_time, to avoid the
2703 * export is being evicted
2705 ptlrpc_update_export_timer(req->rq_export, 0);
2706 target_request_copy_put(req);
2709 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2710 CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2711 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2712 if (delta > OBD_RECOVERY_TIME_SOFT) {
2713 CWARN("too long recovery - read logs\n");
2714 libcfs_debug_dumplog();
2717 target_finish_recovery(lut);
2718 lu_context_fini(&env->le_ctx);
2719 trd->trd_processing_task = 0;
2720 complete_all(&trd->trd_finishing);
2721 tgt_io_thread_done(thread);
2727 OBD_FREE_PTR(thread);
2731 static int target_start_recovery_thread(struct lu_target *lut,
2732 svc_handler_t handler)
2734 struct obd_device *obd = lut->lut_obd;
2736 struct target_recovery_data *trd = &obd->obd_recovery_data;
2739 memset(trd, 0, sizeof(*trd));
2740 init_completion(&trd->trd_starting);
2741 init_completion(&trd->trd_finishing);
2742 trd->trd_recovery_handler = handler;
2744 rc = server_name2index(obd->obd_name, &index, NULL);
2748 if (!IS_ERR(kthread_run(target_recovery_thread,
2749 lut, "tgt_recover_%d", index))) {
2750 wait_for_completion(&trd->trd_starting);
2751 LASSERT(obd->obd_recovering != 0);
2759 void target_stop_recovery_thread(struct obd_device *obd)
2761 if (obd->obd_recovery_data.trd_processing_task > 0) {
2762 struct target_recovery_data *trd = &obd->obd_recovery_data;
2763 /** recovery can be done but postrecovery is not yet */
2764 spin_lock(&obd->obd_dev_lock);
2765 if (obd->obd_recovering) {
2766 CERROR("%s: Aborting recovery\n", obd->obd_name);
2767 obd->obd_abort_recovery = 1;
2768 wake_up(&obd->obd_next_transno_waitq);
2770 spin_unlock(&obd->obd_dev_lock);
2771 wait_for_completion(&trd->trd_finishing);
2774 EXPORT_SYMBOL(target_stop_recovery_thread);
2776 void target_recovery_fini(struct obd_device *obd)
2778 class_disconnect_exports(obd);
2779 target_stop_recovery_thread(obd);
2780 target_cleanup_recovery(obd);
2782 EXPORT_SYMBOL(target_recovery_fini);
2784 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2786 struct obd_device *obd = container_of(timer, struct obd_device,
2787 obd_recovery_timer);
2790 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2791 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2792 ktime_get_seconds() - obd->obd_recovery_start,
2793 atomic_read(&obd->obd_connected_clients));
2795 obd->obd_recovery_expired = 1;
2796 wake_up(&obd->obd_next_transno_waitq);
2797 return HRTIMER_NORESTART;
2800 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2802 struct obd_device *obd = lut->lut_obd;
2804 if (lut->lut_bottom->dd_rdonly)
2807 if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2808 /** Update server last boot epoch */
2809 tgt_boot_epoch_update(lut);
2813 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2814 "last_transno %llu\n", obd->obd_name,
2815 atomic_read(&obd->obd_max_recoverable_clients),
2816 obd->obd_last_committed);
2817 LASSERT(obd->obd_stopping == 0);
2818 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2819 obd->obd_recovery_start = 0;
2820 obd->obd_recovery_end = 0;
2822 hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2824 obd->obd_recovery_timer.function = &target_recovery_expired;
2825 target_start_recovery_thread(lut, handler);
2827 EXPORT_SYMBOL(target_recovery_init);
2829 static int target_process_req_flags(struct obd_device *obd,
2830 struct ptlrpc_request *req)
2832 struct obd_export *exp = req->rq_export;
2834 LASSERT(exp != NULL);
2835 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2836 /* client declares he's ready to replay locks */
2837 spin_lock(&exp->exp_lock);
2838 if (exp->exp_req_replay_needed) {
2839 exp->exp_req_replay_needed = 0;
2840 spin_unlock(&exp->exp_lock);
2842 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2843 atomic_dec(&obd->obd_req_replay_clients);
2845 spin_unlock(&exp->exp_lock);
2848 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2850 * client declares he's ready to complete recovery
2851 * so, we put the request on th final queue
2853 spin_lock(&exp->exp_lock);
2854 if (exp->exp_lock_replay_needed) {
2855 exp->exp_lock_replay_needed = 0;
2856 spin_unlock(&exp->exp_lock);
2858 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2859 atomic_dec(&obd->obd_lock_replay_clients);
2861 spin_unlock(&exp->exp_lock);
2867 int target_queue_recovery_request(struct ptlrpc_request *req,
2868 struct obd_device *obd)
2870 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2871 struct ptlrpc_request *reqiter;
2876 if (obd->obd_recovery_data.trd_processing_task == current->pid) {
2877 /* Processing the queue right now, don't re-add. */
2881 target_process_req_flags(obd, req);
2883 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2884 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2885 if (cfs_fail_val == 1) {
2888 wake_up(&cfs_race_waitq);
2890 schedule_timeout_interruptible(
2891 cfs_time_seconds(1));
2896 * client declares he's ready to complete recovery
2897 * so, we put the request on th final queue
2899 target_request_copy_get(req);
2900 DEBUG_REQ(D_HA, req, "queue final req");
2901 wake_up(&obd->obd_next_transno_waitq);
2902 spin_lock(&obd->obd_recovery_task_lock);
2903 if (obd->obd_recovering) {
2904 struct ptlrpc_request *tmp;
2905 struct ptlrpc_request *duplicate = NULL;
2907 if (likely(!req->rq_export->exp_replay_done)) {
2908 req->rq_export->exp_replay_done = 1;
2909 list_add_tail(&req->rq_list,
2910 &obd->obd_final_req_queue);
2911 spin_unlock(&obd->obd_recovery_task_lock);
2916 * XXX O(n), but only happens if final ping is
2917 * timed out, probably reorganize the list as
2920 list_for_each_entry_safe(reqiter, tmp,
2921 &obd->obd_final_req_queue,
2923 if (reqiter->rq_export == req->rq_export) {
2924 list_del_init(&reqiter->rq_list);
2925 duplicate = reqiter;
2930 list_add_tail(&req->rq_list,
2931 &obd->obd_final_req_queue);
2932 req->rq_export->exp_replay_done = 1;
2933 spin_unlock(&obd->obd_recovery_task_lock);
2935 if (duplicate != NULL) {
2936 DEBUG_REQ(D_HA, duplicate,
2937 "put prev final req");
2938 target_request_copy_put(duplicate);
2942 spin_unlock(&obd->obd_recovery_task_lock);
2943 target_request_copy_put(req);
2944 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2947 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2948 /* client declares he's ready to replay locks */
2949 target_request_copy_get(req);
2950 DEBUG_REQ(D_HA, req, "queue lock replay req");
2951 wake_up(&obd->obd_next_transno_waitq);
2952 spin_lock(&obd->obd_recovery_task_lock);
2953 LASSERT(obd->obd_recovering);
2954 /* usually due to recovery abort */
2955 if (!req->rq_export->exp_in_recovery) {
2956 spin_unlock(&obd->obd_recovery_task_lock);
2957 target_request_copy_put(req);
2960 LASSERT(req->rq_export->exp_lock_replay_needed);
2961 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2962 spin_unlock(&obd->obd_recovery_task_lock);
2967 * CAVEAT EMPTOR: The incoming request message has been swabbed
2968 * (i.e. buflens etc are in my own byte order), but type-dependent
2969 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
2973 INIT_LIST_HEAD(&req->rq_list);
2974 DEBUG_REQ(D_HA, req, "not queueing");
2979 * If we're processing the queue, we want don't want to queue this
2982 * Also, if this request has a transno less than the one we're waiting
2983 * for, we should process it now. It could (and currently always will)
2984 * be an open request for a descriptor that was opened some time ago.
2986 * Also, a resent, replayed request that has already been
2987 * handled will pass through here and be processed immediately.
2990 "Next recovery transno: %llu, current: %llu, replaying\n",
2991 obd->obd_next_recovery_transno, transno);
2994 * If the request has been replayed by update replay, then sends this
2995 * request to the recovery thread (replay_request_or_update()), where
2996 * it will be handled
2998 spin_lock(&obd->obd_recovery_task_lock);
2999 if (transno < obd->obd_next_recovery_transno &&
3000 !is_req_replayed_by_update(req)) {
3001 /* Processing the queue right now, don't re-add. */
3002 LASSERT(list_empty(&req->rq_list));
3003 spin_unlock(&obd->obd_recovery_task_lock);
3006 spin_unlock(&obd->obd_recovery_task_lock);
3008 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3011 target_request_copy_get(req);
3012 if (!req->rq_export->exp_in_recovery) {
3013 target_request_copy_put(req);
3016 LASSERT(req->rq_export->exp_req_replay_needed);
3018 if (target_exp_enqueue_req_replay(req)) {
3019 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3020 target_request_copy_put(req);
3025 spin_lock(&obd->obd_recovery_task_lock);
3026 LASSERT(obd->obd_recovering);
3027 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3028 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3029 list_add_tail(&req->rq_list, &reqiter->rq_list);
3034 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3036 DEBUG_REQ(D_ERROR, req,
3037 "dropping replay: transno has been claimed by another client");
3038 spin_unlock(&obd->obd_recovery_task_lock);
3039 target_exp_dequeue_req_replay(req);
3040 target_request_copy_put(req);
3046 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3048 obd->obd_requests_queued_for_recovery++;
3049 spin_unlock(&obd->obd_recovery_task_lock);
3050 wake_up(&obd->obd_next_transno_waitq);
3054 void target_committed_to_req(struct ptlrpc_request *req)
3056 struct obd_export *exp = req->rq_export;
3058 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3059 lustre_msg_set_last_committed(req->rq_repmsg,
3060 exp->exp_last_committed);
3062 DEBUG_REQ(D_IOCTL, req,
3063 "not sending last_committed update (%d/%d)",
3064 exp->exp_obd->obd_no_transno,
3065 req->rq_repmsg == NULL);
3067 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3068 exp->exp_last_committed, req->rq_transno, req->rq_xid);
3071 #endif /* HAVE_SERVER_SUPPORT */
3074 * Packs current SLV and Limit into \a req.
3076 int target_pack_pool_reply(struct ptlrpc_request *req)
3078 struct obd_device *obd;
3083 * Check that we still have all structures alive as this may
3084 * be some late RPC at shutdown time.
3086 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3087 !exp_connect_lru_resize(req->rq_export))) {
3088 lustre_msg_set_slv(req->rq_repmsg, 0);
3089 lustre_msg_set_limit(req->rq_repmsg, 0);
3093 /* OBD is alive here as export is alive, which we checked above. */
3094 obd = req->rq_export->exp_obd;
3096 read_lock(&obd->obd_pool_lock);
3097 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3098 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3099 read_unlock(&obd->obd_pool_lock);
3104 static int target_send_reply_msg(struct ptlrpc_request *req,
3105 int rc, int fail_id)
3107 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3108 DEBUG_REQ(D_ERROR, req, "dropping reply");
3112 * We can have a null rq_reqmsg in the event of bad signature or
3113 * no context when unwrapping
3115 if (req->rq_reqmsg &&
3116 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3117 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3121 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3122 req->rq_status = rc;
3123 return ptlrpc_send_error(req, 1);
3125 DEBUG_REQ(D_NET, req, "sending reply");
3127 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3130 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3132 struct ptlrpc_service_part *svcpt;
3134 struct ptlrpc_reply_state *rs;
3135 struct obd_export *exp;
3139 if (req->rq_no_reply) {
3144 svcpt = req->rq_rqbd->rqbd_svcpt;
3145 rs = req->rq_reply_state;
3146 if (rs == NULL || !rs->rs_difficult) {
3148 target_send_reply_msg(req, rc, fail_id);
3153 /* must be an export if locks saved */
3154 LASSERT(req->rq_export != NULL);
3155 /* req/reply consistent */
3156 LASSERT(rs->rs_svcpt == svcpt);
3159 LASSERT(!rs->rs_scheduled);
3160 LASSERT(!rs->rs_scheduled_ever);
3161 LASSERT(!rs->rs_handled);
3162 LASSERT(!rs->rs_on_net);
3163 LASSERT(rs->rs_export == NULL);
3164 LASSERT(list_empty(&rs->rs_obd_list));
3165 LASSERT(list_empty(&rs->rs_exp_list));
3167 exp = class_export_get(req->rq_export);
3169 /* disable reply scheduling while I'm setting up */
3170 rs->rs_scheduled = 1;
3172 rs->rs_xid = req->rq_xid;
3173 rs->rs_transno = req->rq_transno;
3174 rs->rs_export = exp;
3175 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3177 spin_lock(&exp->exp_uncommitted_replies_lock);
3178 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3179 rs->rs_transno, exp->exp_last_committed);
3180 if (rs->rs_transno > exp->exp_last_committed) {
3181 /* not committed already */
3182 list_add_tail(&rs->rs_obd_list,
3183 &exp->exp_uncommitted_replies);
3185 spin_unlock(&exp->exp_uncommitted_replies_lock);
3187 spin_lock(&exp->exp_lock);
3188 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3189 spin_unlock(&exp->exp_lock);
3191 netrc = target_send_reply_msg(req, rc, fail_id);
3193 spin_lock(&svcpt->scp_rep_lock);
3195 atomic_inc(&svcpt->scp_nreps_difficult);
3199 * error sending: reply is off the net. Also we need +1
3200 * reply ref until ptlrpc_handle_rs() is done
3201 * with the reply state (if the send was successful, there
3202 * would have been +1 ref for the net, which
3203 * reply_out_callback leaves alone)
3206 ptlrpc_rs_addref(rs);
3209 spin_lock(&rs->rs_lock);
3210 if (rs->rs_transno <= exp->exp_last_committed ||
3211 (!rs->rs_on_net && !rs->rs_no_ack) ||
3212 list_empty(&rs->rs_exp_list) || /* completed already */
3213 list_empty(&rs->rs_obd_list)) {
3214 CDEBUG(D_HA, "Schedule reply immediately\n");
3215 ptlrpc_dispatch_difficult_reply(rs);
3217 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3218 rs->rs_scheduled = 0; /* allow notifier to schedule */
3220 spin_unlock(&rs->rs_lock);
3221 spin_unlock(&svcpt->scp_rep_lock);
3225 enum ldlm_mode lck_compat_array[] = {
3226 [LCK_EX] = LCK_COMPAT_EX,
3227 [LCK_PW] = LCK_COMPAT_PW,
3228 [LCK_PR] = LCK_COMPAT_PR,
3229 [LCK_CW] = LCK_COMPAT_CW,
3230 [LCK_CR] = LCK_COMPAT_CR,
3231 [LCK_NL] = LCK_COMPAT_NL,
3232 [LCK_GROUP] = LCK_COMPAT_GROUP,
3233 [LCK_COS] = LCK_COMPAT_COS,
3237 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3238 * not escape to the user level.
3240 int ldlm_error2errno(enum ldlm_error error)
3246 case ELDLM_LOCK_MATCHED:
3249 case ELDLM_LOCK_CHANGED:
3252 case ELDLM_LOCK_ABORTED:
3255 case ELDLM_LOCK_REPLACED:
3258 case ELDLM_NO_LOCK_DATA:
3261 case ELDLM_NAMESPACE_EXISTS:
3264 case ELDLM_BAD_NAMESPACE:
3268 if (((int)error) < 0) { /* cast to signed type */
3269 result = error; /* as ldlm_error can be unsigned */
3271 CERROR("Invalid DLM result code: %d\n", error);
3277 EXPORT_SYMBOL(ldlm_error2errno);
3280 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3282 enum ldlm_error ldlm_errno2error(int err_no)
3291 error = ELDLM_LOCK_CHANGED;
3294 error = ELDLM_LOCK_ABORTED;
3297 error = ELDLM_LOCK_REPLACED;
3300 error = ELDLM_NO_LOCK_DATA;
3303 error = ELDLM_NAMESPACE_EXISTS;
3306 error = ELDLM_BAD_NAMESPACE;
3314 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3315 void ldlm_dump_export_locks(struct obd_export *exp)
3317 spin_lock(&exp->exp_locks_list_guard);
3318 if (!list_empty(&exp->exp_locks_list)) {
3319 struct ldlm_lock *lock;
3321 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3323 list_for_each_entry(lock, &exp->exp_locks_list,
3325 LDLM_ERROR(lock, "lock:");
3327 spin_unlock(&exp->exp_locks_list_guard);
3331 #ifdef HAVE_SERVER_SUPPORT
3332 static inline const char *bulk2type(struct ptlrpc_request *req)
3334 if (req->rq_bulk_read)
3336 if (req->rq_bulk_write)
3341 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3343 struct ptlrpc_request *req = desc->bd_req;
3344 time64_t start = ktime_get_seconds();
3350 /* If there is eviction in progress, wait for it to finish. */
3352 exp->exp_obd->obd_evict_inprogress_waitq,
3353 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3355 /* Check if client was evicted or reconnected already. */
3356 if (exp->exp_failed ||
3357 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3360 if (req->rq_bulk_read)
3361 rc = sptlrpc_svc_wrap_bulk(req, desc);
3363 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3364 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3365 else /* old version, bulk matchbits is rq_xid */
3366 req->rq_mbits = req->rq_xid;
3369 rc = ptlrpc_start_bulk_transfer(desc);
3373 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3374 bulk2type(req), rc);
3378 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3379 ptlrpc_abort_bulk(desc);
3383 /* limit actual bulk transfer to bulk_timeout seconds */
3384 deadline = start + bulk_timeout;
3385 if (deadline > req->rq_deadline)
3386 deadline = req->rq_deadline;
3389 time64_t timeoutl = deadline - ktime_get_seconds();
3390 time64_t rq_deadline;
3392 while (timeoutl >= 0 &&
3393 wait_event_idle_timeout(
3395 !ptlrpc_server_bulk_active(desc) ||
3398 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3399 timeoutl ? cfs_time_seconds(1) : 1) == 0)
3401 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3403 /* Wait again if we changed rq_deadline. */
3404 rq_deadline = READ_ONCE(req->rq_deadline);
3405 deadline = start + bulk_timeout;
3406 if (deadline > rq_deadline)
3407 deadline = rq_deadline;
3408 } while (rc == -ETIMEDOUT &&
3409 deadline > ktime_get_seconds());
3411 if (rc == -ETIMEDOUT) {
3412 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3413 bulk2type(req), deadline - start,
3414 ktime_get_real_seconds() - deadline);
3415 ptlrpc_abort_bulk(desc);
3416 } else if (exp->exp_failed) {
3417 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3420 ptlrpc_abort_bulk(desc);
3421 } else if (exp->exp_conn_cnt >
3422 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3423 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3425 /* We don't reply anyway. */
3427 ptlrpc_abort_bulk(desc);
3428 } else if (desc->bd_failure) {
3429 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3431 /* XXX should this be a different errno? */
3434 if (req->rq_bulk_write)
3435 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3436 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3437 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3438 bulk2type(req), desc->bd_nob_transferred,
3440 /* XXX should this be a different errno? */
3447 EXPORT_SYMBOL(target_bulk_io);
3449 #endif /* HAVE_SERVER_SUPPORT */