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/kthread.h>
45 #include <libcfs/libcfs.h>
47 #include <obd_class.h>
48 #include <lustre_dlm.h>
49 #include <lustre_net.h>
50 #include <lustre_sec.h>
51 #include "ldlm_internal.h"
53 /* @priority: If non-zero, move the selected connection to the list head.
54 * @create: If zero, only search in existing connections.
56 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
57 int priority, int create)
59 struct ptlrpc_connection *ptlrpc_conn;
60 struct obd_import_conn *imp_conn = NULL, *item;
61 lnet_nid_t nid4refnet = LNET_NID_ANY;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 if (imp->imp_connection &&
71 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
72 /* nid4refnet is used to restrict network connections */
73 nid4refnet = imp->imp_connection->c_self;
74 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
76 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
81 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
83 GOTO(out_put, rc = -ENOMEM);
86 spin_lock(&imp->imp_lock);
87 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
88 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
90 list_del(&item->oic_item);
91 list_add(&item->oic_item,
93 item->oic_last_attempt = 0;
95 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
96 imp, imp->imp_obd->obd_name, uuid->uuid,
97 (priority ? ", moved to head" : ""));
98 spin_unlock(&imp->imp_lock);
99 GOTO(out_free, rc = 0);
102 /* No existing import connection found for \a uuid. */
104 imp_conn->oic_conn = ptlrpc_conn;
105 imp_conn->oic_uuid = *uuid;
106 imp_conn->oic_last_attempt = 0;
108 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
110 list_add_tail(&imp_conn->oic_item,
111 &imp->imp_conn_list);
112 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
113 imp, imp->imp_obd->obd_name, uuid->uuid,
114 (priority ? "head" : "tail"));
116 spin_unlock(&imp->imp_lock);
117 GOTO(out_free, rc = -ENOENT);
120 spin_unlock(&imp->imp_lock);
124 OBD_FREE(imp_conn, sizeof(*imp_conn));
126 ptlrpc_connection_put(ptlrpc_conn);
130 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
132 return import_set_conn(imp, uuid, 1, 0);
135 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
138 return import_set_conn(imp, uuid, priority, 1);
140 EXPORT_SYMBOL(client_import_add_conn);
142 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
144 struct obd_import_conn *imp_conn;
145 struct obd_export *dlmexp;
149 spin_lock(&imp->imp_lock);
150 if (list_empty(&imp->imp_conn_list)) {
151 LASSERT(!imp->imp_connection);
155 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
156 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
158 LASSERT(imp_conn->oic_conn);
160 if (imp_conn == imp->imp_conn_current) {
161 LASSERT(imp_conn->oic_conn == imp->imp_connection);
163 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
164 imp->imp_state != LUSTRE_IMP_DISCON) {
165 CERROR("can't remove current connection\n");
166 GOTO(out, rc = -EBUSY);
169 ptlrpc_connection_put(imp->imp_connection);
170 imp->imp_connection = NULL;
172 dlmexp = class_conn2export(&imp->imp_dlm_handle);
173 if (dlmexp && dlmexp->exp_connection) {
174 LASSERT(dlmexp->exp_connection ==
176 ptlrpc_connection_put(dlmexp->exp_connection);
177 dlmexp->exp_connection = NULL;
181 class_export_put(dlmexp);
184 list_del(&imp_conn->oic_item);
185 ptlrpc_connection_put(imp_conn->oic_conn);
186 OBD_FREE(imp_conn, sizeof(*imp_conn));
187 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
188 imp, imp->imp_obd->obd_name, uuid->uuid);
193 spin_unlock(&imp->imp_lock);
195 CERROR("connection %s not found\n", uuid->uuid);
198 EXPORT_SYMBOL(client_import_del_conn);
201 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
202 * to find a conn uuid of \a imp which can reach \a peer.
204 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
205 struct obd_uuid *uuid)
207 struct obd_import_conn *conn;
211 spin_lock(&imp->imp_lock);
212 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
213 /* Check if conn UUID does have this peer NID. */
214 if (class_check_uuid(&conn->oic_uuid, peer)) {
215 *uuid = conn->oic_uuid;
220 spin_unlock(&imp->imp_lock);
223 EXPORT_SYMBOL(client_import_find_conn);
225 void client_destroy_import(struct obd_import *imp)
227 /* Drop security policy instance after all RPCs have finished/aborted
228 * to let all busy contexts be released. */
229 class_import_get(imp);
230 class_destroy_import(imp);
231 sptlrpc_import_sec_put(imp);
232 class_import_put(imp);
234 EXPORT_SYMBOL(client_destroy_import);
237 * Check whether or not the OSC is on MDT.
240 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
242 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
245 static int osc_on_mdt(char *obdname)
249 ptr = strrchr(obdname, '-');
253 if (strncmp(ptr + 1, "MDT", 3) == 0)
259 /* Configure an RPC client OBD device.
264 * 3 - inactive-on-startup
265 * 4 - restrictive net
267 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
269 struct client_obd *cli = &obddev->u.cli;
270 struct obd_import *imp;
271 struct obd_uuid server_uuid;
272 int rq_portal, rp_portal, connect_op;
273 char *name = obddev->obd_type->typ_name;
274 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
275 char *cli_name = lustre_cfg_buf(lcfg, 0);
276 struct ptlrpc_connection fake_conn = { .c_self = 0,
277 .c_remote_uuid.uuid[0] = 0 };
281 /* In a more perfect world, we would hang a ptlrpc_client off of
282 * obd_type and just use the values from there. */
283 if (!strcmp(name, LUSTRE_OSC_NAME)) {
284 rq_portal = OST_REQUEST_PORTAL;
285 rp_portal = OSC_REPLY_PORTAL;
286 connect_op = OST_CONNECT;
287 cli->cl_sp_me = LUSTRE_SP_CLI;
288 cli->cl_sp_to = LUSTRE_SP_OST;
289 ns_type = LDLM_NS_TYPE_OSC;
290 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
291 !strcmp(name, LUSTRE_LWP_NAME)) {
292 rq_portal = MDS_REQUEST_PORTAL;
293 rp_portal = MDC_REPLY_PORTAL;
294 connect_op = MDS_CONNECT;
295 if (is_lwp_on_ost(cli_name))
296 cli->cl_sp_me = LUSTRE_SP_OST;
297 else if (is_lwp_on_mdt(cli_name))
298 cli->cl_sp_me = LUSTRE_SP_MDT;
300 cli->cl_sp_me = LUSTRE_SP_CLI;
301 cli->cl_sp_to = LUSTRE_SP_MDT;
302 ns_type = LDLM_NS_TYPE_MDC;
303 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
304 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
305 /* OSP_on_MDT for other MDTs */
306 connect_op = MDS_CONNECT;
307 cli->cl_sp_to = LUSTRE_SP_MDT;
308 ns_type = LDLM_NS_TYPE_MDC;
309 rq_portal = OUT_PORTAL;
311 /* OSP on MDT for OST */
312 connect_op = OST_CONNECT;
313 cli->cl_sp_to = LUSTRE_SP_OST;
314 ns_type = LDLM_NS_TYPE_OSC;
315 rq_portal = OST_REQUEST_PORTAL;
317 rp_portal = OSC_REPLY_PORTAL;
318 cli->cl_sp_me = LUSTRE_SP_MDT;
319 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
320 rq_portal = MGS_REQUEST_PORTAL;
321 rp_portal = MGC_REPLY_PORTAL;
322 connect_op = MGS_CONNECT;
323 cli->cl_sp_me = LUSTRE_SP_MGC;
324 cli->cl_sp_to = LUSTRE_SP_MGS;
325 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
326 ns_type = LDLM_NS_TYPE_MGC;
328 CERROR("unknown client OBD type \"%s\", can't setup\n",
333 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
334 CERROR("requires a TARGET UUID\n");
338 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
339 CERROR("client UUID must be less than 38 characters\n");
343 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
344 CERROR("setup requires a SERVER UUID\n");
348 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
349 CERROR("target UUID must be less than 38 characters\n");
353 init_rwsem(&cli->cl_sem);
354 mutex_init(&cli->cl_mgc_mutex);
356 init_rwsem(&cli->cl_seq_rwsem);
357 cli->cl_conn_count = 0;
358 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
359 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
360 sizeof(server_uuid)));
362 cli->cl_dirty_pages = 0;
363 cli->cl_dirty_max_pages = 0;
364 cli->cl_avail_grant = 0;
365 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
366 /* cl_dirty_max_pages may be changed at connect time in
367 * ptlrpc_connect_interpret(). */
368 client_adjust_max_dirty(cli);
369 INIT_LIST_HEAD(&cli->cl_cache_waiters);
370 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
371 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
372 INIT_LIST_HEAD(&cli->cl_loi_write_list);
373 INIT_LIST_HEAD(&cli->cl_loi_read_list);
374 spin_lock_init(&cli->cl_loi_list_lock);
375 atomic_set(&cli->cl_pending_w_pages, 0);
376 atomic_set(&cli->cl_pending_r_pages, 0);
377 cli->cl_r_in_flight = 0;
378 cli->cl_w_in_flight = 0;
380 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
381 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
382 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
383 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
384 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
385 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
388 INIT_LIST_HEAD(&cli->cl_lru_osc);
389 atomic_set(&cli->cl_lru_shrinkers, 0);
390 atomic_long_set(&cli->cl_lru_busy, 0);
391 atomic_long_set(&cli->cl_lru_in_list, 0);
392 INIT_LIST_HEAD(&cli->cl_lru_list);
393 spin_lock_init(&cli->cl_lru_list_lock);
394 atomic_long_set(&cli->cl_unstable_count, 0);
395 INIT_LIST_HEAD(&cli->cl_shrink_list);
396 INIT_LIST_HEAD(&cli->cl_grant_chain);
398 INIT_LIST_HEAD(&cli->cl_flight_waiters);
399 cli->cl_rpcs_in_flight = 0;
401 init_waitqueue_head(&cli->cl_destroy_waitq);
402 atomic_set(&cli->cl_destroy_in_flight, 0);
404 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
405 #ifdef ENABLE_CHECKSUM
406 /* Turn on checksumming by default. */
407 cli->cl_checksum = 1;
409 * The supported checksum types will be worked out at connect time
410 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
413 cli->cl_cksum_type = cli->cl_supp_cksum_types;
415 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
417 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
418 * from OFD after connecting. */
419 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
421 cli->cl_max_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
423 /* set cl_chunkbits default value to PAGE_SHIFT,
424 * it will be updated at OSC connection time. */
425 cli->cl_chunkbits = PAGE_SHIFT;
427 if (!strcmp(name, LUSTRE_MDC_NAME)) {
428 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
429 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
430 cli->cl_max_rpcs_in_flight = 2;
431 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
432 cli->cl_max_rpcs_in_flight = 3;
433 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
434 cli->cl_max_rpcs_in_flight = 4;
436 if (osc_on_mdt(obddev->obd_name))
437 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
439 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
442 spin_lock_init(&cli->cl_mod_rpcs_lock);
443 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
444 cli->cl_max_mod_rpcs_in_flight = 0;
445 cli->cl_mod_rpcs_in_flight = 0;
446 cli->cl_close_rpcs_in_flight = 0;
447 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
448 cli->cl_mod_tag_bitmap = NULL;
450 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
452 if (connect_op == MDS_CONNECT) {
453 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
454 OBD_ALLOC(cli->cl_mod_tag_bitmap,
455 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
456 if (cli->cl_mod_tag_bitmap == NULL)
457 GOTO(err, rc = -ENOMEM);
462 CERROR("ldlm_get_ref failed: %d\n", rc);
466 ptlrpc_init_client(rq_portal, rp_portal, name,
467 &obddev->obd_ldlm_client);
469 imp = class_new_import(obddev);
471 GOTO(err_ldlm, rc = -ENOENT);
472 imp->imp_client = &obddev->obd_ldlm_client;
473 imp->imp_connect_op = connect_op;
474 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
475 LUSTRE_CFG_BUFLEN(lcfg, 1));
476 class_import_put(imp);
478 if (lustre_cfg_buf(lcfg, 4)) {
479 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
481 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
483 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
484 obddev->obd_name, lustre_cfg_string(lcfg, 4),
486 GOTO(err_import, rc);
488 fake_conn.c_self = LNET_MKNID(refnet, 0);
489 imp->imp_connection = &fake_conn;
492 rc = client_import_add_conn(imp, &server_uuid, 1);
494 CERROR("can't add initial connection\n");
495 GOTO(err_import, rc);
497 imp->imp_connection = NULL;
499 cli->cl_import = imp;
500 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
501 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
503 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
504 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
505 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
506 name, obddev->obd_name,
507 cli->cl_target_uuid.uuid);
508 spin_lock(&imp->imp_lock);
509 imp->imp_deactive = 1;
510 spin_unlock(&imp->imp_lock);
514 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
515 LDLM_NAMESPACE_CLIENT,
516 LDLM_NAMESPACE_GREEDY,
518 if (obddev->obd_namespace == NULL) {
519 CERROR("Unable to create client namespace - %s\n",
521 GOTO(err_import, rc = -ENOMEM);
527 class_destroy_import(imp);
531 if (cli->cl_mod_tag_bitmap != NULL)
532 OBD_FREE(cli->cl_mod_tag_bitmap,
533 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
534 cli->cl_mod_tag_bitmap = NULL;
538 EXPORT_SYMBOL(client_obd_setup);
540 int client_obd_cleanup(struct obd_device *obddev)
542 struct client_obd *cli = &obddev->u.cli;
545 ldlm_namespace_free_post(obddev->obd_namespace);
546 obddev->obd_namespace = NULL;
548 obd_cleanup_client_import(obddev);
549 LASSERT(obddev->u.cli.cl_import == NULL);
553 if (cli->cl_mod_tag_bitmap != NULL)
554 OBD_FREE(cli->cl_mod_tag_bitmap,
555 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
556 cli->cl_mod_tag_bitmap = NULL;
560 EXPORT_SYMBOL(client_obd_cleanup);
562 /* ->o_connect() method for client side (OSC and MDC and MGC) */
563 int client_connect_import(const struct lu_env *env,
564 struct obd_export **exp,
565 struct obd_device *obd, struct obd_uuid *cluuid,
566 struct obd_connect_data *data, void *localdata)
568 struct client_obd *cli = &obd->u.cli;
569 struct obd_import *imp = cli->cl_import;
570 struct obd_connect_data *ocd;
571 struct lustre_handle conn = { 0 };
576 down_write(&cli->cl_sem);
577 if (cli->cl_conn_count > 0)
578 GOTO(out_sem, rc = -EALREADY);
580 rc = class_connect(&conn, obd, cluuid);
584 cli->cl_conn_count++;
585 *exp = class_conn2export(&conn);
587 LASSERT(obd->obd_namespace);
589 imp->imp_dlm_handle = conn;
590 rc = ptlrpc_init_import(imp);
594 ocd = &imp->imp_connect_data;
597 imp->imp_connect_flags_orig = data->ocd_connect_flags;
598 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
601 rc = ptlrpc_connect_import(imp);
603 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
606 LASSERT(*exp != NULL && (*exp)->exp_connection);
609 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
610 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
611 data->ocd_connect_flags, ocd->ocd_connect_flags);
612 data->ocd_connect_flags = ocd->ocd_connect_flags;
613 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
616 ptlrpc_pinger_add_import(imp);
622 cli->cl_conn_count--;
623 class_disconnect(*exp);
627 up_write(&cli->cl_sem);
629 if (!rc && localdata) {
630 LASSERT(cli->cl_cache == NULL); /* only once */
631 cli->cl_cache = (struct cl_client_cache *)localdata;
632 cl_cache_incref(cli->cl_cache);
633 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
635 /* add this osc into entity list */
636 LASSERT(list_empty(&cli->cl_lru_osc));
637 spin_lock(&cli->cl_cache->ccc_lru_lock);
638 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
639 spin_unlock(&cli->cl_cache->ccc_lru_lock);
644 EXPORT_SYMBOL(client_connect_import);
646 int client_disconnect_export(struct obd_export *exp)
648 struct obd_device *obd = class_exp2obd(exp);
649 struct client_obd *cli;
650 struct obd_import *imp;
655 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
656 exp, exp ? exp->exp_handle.h_cookie : -1);
661 imp = cli->cl_import;
663 down_write(&cli->cl_sem);
664 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
667 if (cli->cl_conn_count == 0) {
668 CERROR("disconnecting disconnected device (%s)\n",
670 GOTO(out_disconnect, rc = -EINVAL);
673 cli->cl_conn_count--;
674 if (cli->cl_conn_count != 0)
675 GOTO(out_disconnect, rc = 0);
677 /* Mark import deactivated now, so we don't try to reconnect if any
678 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
679 * fully deactivate the import, or that would drop all requests. */
680 spin_lock(&imp->imp_lock);
681 imp->imp_deactive = 1;
682 spin_unlock(&imp->imp_lock);
684 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
685 * delete it regardless. (It's safe to delete an import that was
687 (void)ptlrpc_pinger_del_import(imp);
689 if (obd->obd_namespace != NULL) {
690 /* obd_force == local only */
691 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
692 obd->obd_force ? LCF_LOCAL : 0, NULL);
693 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
696 /* There's no need to hold sem while disconnecting an import,
697 * and it may actually cause deadlock in GSS. */
698 up_write(&cli->cl_sem);
699 rc = ptlrpc_disconnect_import(imp, 0);
700 down_write(&cli->cl_sem);
702 ptlrpc_invalidate_import(imp);
707 /* Use server style - class_disconnect should be always called for
709 err = class_disconnect(exp);
713 up_write(&cli->cl_sem);
717 EXPORT_SYMBOL(client_disconnect_export);
719 #ifdef HAVE_SERVER_SUPPORT
720 int server_disconnect_export(struct obd_export *exp)
725 /* Disconnect early so that clients can't keep using export. */
726 rc = class_disconnect(exp);
727 /* Close import to avoid sending any requests. */
728 if (exp->exp_imp_reverse)
729 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
731 ldlm_bl_thread_wakeup();
733 /* complete all outstanding replies */
734 spin_lock(&exp->exp_lock);
735 while (!list_empty(&exp->exp_outstanding_replies)) {
736 struct ptlrpc_reply_state *rs =
737 list_entry(exp->exp_outstanding_replies.next,
738 struct ptlrpc_reply_state, rs_exp_list);
739 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
741 spin_lock(&svcpt->scp_rep_lock);
743 list_del_init(&rs->rs_exp_list);
745 spin_lock(&rs->rs_lock);
746 /* clear rs_convert_lock to make sure rs is handled and put */
747 rs->rs_convert_lock = 0;
748 ptlrpc_schedule_difficult_reply(rs);
749 spin_unlock(&rs->rs_lock);
751 spin_unlock(&svcpt->scp_rep_lock);
753 spin_unlock(&exp->exp_lock);
757 EXPORT_SYMBOL(server_disconnect_export);
759 /* --------------------------------------------------------------------------
760 * from old lib/target.c
761 * -------------------------------------------------------------------------- */
763 static int target_handle_reconnect(struct lustre_handle *conn,
764 struct obd_export *exp,
765 struct obd_uuid *cluuid)
767 struct obd_device *target;
768 struct lustre_handle *hdl;
775 hdl = &exp->exp_imp_reverse->imp_remote_handle;
776 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
777 conn->cookie = exp->exp_handle.h_cookie;
778 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
779 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
783 target = exp->exp_obd;
785 /* Might be a re-connect after a partition. */
786 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
787 LCONSOLE_WARN("%s: already connected client %s (at %s) "
788 "with handle %#llx. Rejecting client "
789 "with the same UUID trying to reconnect "
790 "with handle %#llx\n", target->obd_name,
791 obd_uuid2str(&exp->exp_client_uuid),
792 obd_export_nid2str(exp),
793 hdl->cookie, conn->cookie);
794 memset(conn, 0, sizeof *conn);
795 /* target_handle_connect() treats EALREADY and
796 * -EALREADY differently. -EALREADY is an error
797 * (same UUID, different handle). */
801 if (!target->obd_recovering) {
802 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
803 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
804 obd_export_nid2str(exp));
805 GOTO(out_already, rc);
808 now = ktime_get_seconds();
809 deadline = jiffies_to_msecs(target->obd_recovery_timer.expires) /
811 if (now < deadline) {
812 struct target_distribute_txn_data *tdtd;
817 timeout = deadline - now;
818 tdtd = class_exp2tgt(exp)->lut_tdtd;
819 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
820 buf = tdtd->tdtd_show_update_logs_retrievers(
821 tdtd->tdtd_show_retrievers_cbdata,
825 LCONSOLE_WARN("%s: Recovery already passed deadline "
826 "%lld:%.02lld. It is due to DNE recovery "
827 "failed/stuck on the %d MDT(s):%s. "
828 "Please wait until all MDTs recovered "
829 "or abort the recovery by force.\n",
830 target->obd_name, timeout / 60,
832 buf ? buf : "unknown (not enough RAM)");
834 LCONSOLE_WARN("%s: Recovery already passed deadline "
835 "%lld:%.02lld. If you do not want to wait "
836 "more, please abort the recovery by "
837 "force.\n", target->obd_name,
838 timeout / 60, timeout % 60);
843 timeout = now - deadline;
844 LCONSOLE_WARN("%s: Recovery already passed deadline"
845 " %lld:%.02lld, It is most likely due to DNE"
846 " recovery is failed or stuck, please wait a"
847 " few more minutes or abort the recovery.\n",
848 target->obd_name, timeout / 60, timeout % 60);
852 conn->cookie = exp->exp_handle.h_cookie;
853 /* target_handle_connect() treats EALREADY and
854 * -EALREADY differently. EALREADY means we are
855 * doing a valid reconnect from the same client. */
860 check_and_start_recovery_timer(struct obd_device *obd,
861 struct ptlrpc_request *req, int new_client);
864 * update flags for import during reconnect process
866 static int rev_import_flags_update(struct obd_import *revimp,
867 struct ptlrpc_request *req)
870 struct obd_connect_data *data;
872 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
874 if (data->ocd_connect_flags & OBD_CONNECT_AT)
875 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
877 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
879 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
881 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
883 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
884 revimp->imp_client->cli_name,
885 libcfs_id2str(req->rq_peer), rc);
893 * Allocate a new reverse import for an export.
895 * \retval -errno in case error hit
896 * \retval 0 if reverse import correctly init
898 int rev_import_init(struct obd_export *export)
900 struct obd_device *obd = export->exp_obd;
901 struct obd_import *revimp;
903 LASSERT(export->exp_imp_reverse == NULL);
905 revimp = class_new_import(obd);
909 revimp->imp_remote_handle.cookie = 0ULL;
910 revimp->imp_client = &obd->obd_ldlm_client;
911 revimp->imp_dlm_fake = 1;
913 /* it is safe to connect import in new state as no sends possible */
914 spin_lock(&export->exp_lock);
915 export->exp_imp_reverse = revimp;
916 spin_unlock(&export->exp_lock);
917 class_import_put(revimp);
921 EXPORT_SYMBOL(rev_import_init);
924 * Handle reconnect for an export.
926 * \param exp export to handle reconnect process
927 * \param req client reconnect request
929 * \retval -rc in case securitfy flavor can't be changed
930 * \retval 0 in case none problems
932 static int rev_import_reconnect(struct obd_export *exp,
933 struct ptlrpc_request *req)
935 struct obd_import *revimp = exp->exp_imp_reverse;
936 struct lustre_handle *lh;
939 /* avoid sending a request until import flags are changed */
940 ptlrpc_import_enter_resend(revimp);
942 if (revimp->imp_connection != NULL)
943 ptlrpc_connection_put(revimp->imp_connection);
946 * client from recovery don't have a handle so we need to take from
947 * request. it may produce situation when wrong client connected
948 * to recovery as we trust a client uuid
950 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
951 revimp->imp_remote_handle = *lh;
953 /* unknown versions will be caught in
954 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
955 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
957 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
959 rc = rev_import_flags_update(revimp, req);
961 /* it is safe to still be in RECOVERY phase as we are not able
962 * to setup correct security flavor so requests are not able to
963 * be delivered correctly */
967 /* resend all rpc's via new connection */
968 return ptlrpc_import_recovery_state_machine(revimp);
971 int target_handle_connect(struct ptlrpc_request *req)
973 struct obd_device *target = NULL;
974 struct obd_export *export = NULL;
975 /* connect handle - filled from target_handle_reconnect in
977 struct lustre_handle conn;
978 struct lustre_handle *tmp;
979 struct obd_uuid cluuid;
984 bool mds_conn = false, lw_client = false, initial_conn = false;
985 bool mds_mds_conn = false;
986 bool new_mds_mds_conn = false;
987 struct obd_connect_data *data, *tmpdata;
989 lnet_nid_t *client_nid = NULL;
992 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
994 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
996 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
997 GOTO(out, rc = -EINVAL);
1000 target = class_dev_by_str(str);
1002 deuuidify(str, NULL, &target_start, &target_len);
1003 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
1004 "from %s (no target). If you are running "
1005 "an HA pair check that the target is "
1006 "mounted on the other server.\n", str,
1007 libcfs_nid2str(req->rq_peer.nid));
1008 GOTO(out, rc = -ENODEV);
1011 spin_lock(&target->obd_dev_lock);
1013 target->obd_conn_inprogress++;
1015 if (target->obd_stopping || !target->obd_set_up) {
1016 spin_unlock(&target->obd_dev_lock);
1018 deuuidify(str, NULL, &target_start, &target_len);
1019 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1020 target_len, target_start,
1021 libcfs_nid2str(req->rq_peer.nid),
1022 (target->obd_stopping ?
1023 "stopping" : "not set up"));
1024 GOTO(out, rc = -ENODEV);
1027 if (target->obd_no_conn) {
1028 spin_unlock(&target->obd_dev_lock);
1030 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1031 "from %s\n", target->obd_name,
1032 libcfs_nid2str(req->rq_peer.nid));
1033 GOTO(out, rc = -EAGAIN);
1036 spin_unlock(&target->obd_dev_lock);
1038 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1040 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1041 GOTO(out, rc = -EINVAL);
1044 obd_str2uuid(&cluuid, str);
1046 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1048 GOTO(out, rc = -EPROTO);
1052 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1054 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1056 GOTO(out, rc = -EPROTO);
1058 rc = req_capsule_server_pack(&req->rq_pill);
1062 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1063 /* Don't allow clients to connect that are using old 1.8 format
1064 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1065 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1066 * FULL20 flag should be set on all connections since 2.0, but no
1067 * longer affects behaviour.
1069 * Later this check will be disabled and the flag can be retired
1070 * completely once interop with 3.0 is no longer needed.
1072 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1073 GOTO(out, rc = -EPROTO);
1075 /* Don't allow liblustre clients to connect.
1076 * - testing was disabled in v2_2_50_0-61-g6a75d65
1077 * - building was disabled in v2_5_58_0-28-g7277179
1078 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1079 * - clients were refused connect for version difference > 0.0.1.32 */
1080 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1081 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1082 GOTO(out, rc = -EPROTO);
1086 /* Note: lw_client is needed in MDS-MDS failover during update log
1087 * processing, so we needs to allow lw_client to be connected at
1088 * anytime, instead of only the initial connection */
1089 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1091 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1092 initial_conn = true;
1093 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1094 mds_mds_conn = (data->ocd_connect_flags &
1095 OBD_CONNECT_MDS_MDS) != 0;
1097 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1098 * for Imperative Recovery connection from MGC to MGS.
1100 * Via check OBD_CONNECT_FID, we can distinguish whether
1101 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1102 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1105 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1106 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1107 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1108 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1109 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1110 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1112 /* We do not support the MDT-MDT interoperations with
1113 * different version MDT because of protocol changes. */
1114 if (unlikely(major != LUSTRE_MAJOR ||
1115 minor != LUSTRE_MINOR ||
1116 abs(patch - LUSTRE_PATCH) > 3)) {
1117 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1118 "connection from different version MDT "
1119 "(%d.%d.%d.%d) %s %s\n",
1120 target->obd_name, LUSTRE_MAJOR,
1121 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1122 major, minor, patch,
1123 OBD_OCD_VERSION_FIX(data->ocd_version),
1124 libcfs_nid2str(req->rq_peer.nid), str);
1126 GOTO(out, rc = -EPROTO);
1131 /* lctl gets a backstage, all-access pass. */
1132 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1133 goto dont_check_exports;
1135 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1139 /* We've found an export in the hash. */
1141 spin_lock(&export->exp_lock);
1143 if (export->exp_connecting) { /* bug 9635, et. al. */
1144 spin_unlock(&export->exp_lock);
1145 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1146 export->exp_obd->obd_name, export,
1147 libcfs_nid2str(req->rq_peer.nid));
1148 class_export_put(export);
1151 } else if ((mds_conn || (lw_client && initial_conn) ||
1152 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1153 export->exp_connection != NULL) {
1154 spin_unlock(&export->exp_lock);
1155 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1156 /* MDS or LWP reconnected after failover. */
1157 LCONSOLE_WARN("%s: Received %s connection from "
1158 "%s, removing former export from %s\n",
1159 target->obd_name, mds_conn ? "MDS" : "LWP",
1160 libcfs_nid2str(req->rq_peer.nid),
1161 libcfs_nid2str(export->exp_connection->c_peer.nid));
1163 /* New MDS connection from the same NID. */
1164 LCONSOLE_WARN("%s: Received new %s connection from "
1165 "%s, removing former export from same NID\n",
1166 target->obd_name, mds_conn ? "MDS" : "LWP",
1167 libcfs_nid2str(req->rq_peer.nid));
1170 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1171 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1172 /* Because exports between MDTs will always be
1173 * kept, let's do not fail such export if they
1174 * come from the same NID, otherwise it might
1175 * cause eviction between MDTs, which might
1176 * cause namespace inconsistency */
1177 spin_lock(&export->exp_lock);
1178 export->exp_connecting = 1;
1179 export->exp_conn_cnt = 0;
1180 spin_unlock(&export->exp_lock);
1181 conn.cookie = export->exp_handle.h_cookie;
1184 class_fail_export(export);
1185 class_export_put(export);
1189 } else if (export->exp_connection != NULL && initial_conn &&
1190 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1191 spin_unlock(&export->exp_lock);
1192 /* In MDS failover we have static UUID but NID can change. */
1193 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1194 "existing nid %s is already connected\n",
1195 target->obd_name, cluuid.uuid,
1196 libcfs_nid2str(req->rq_peer.nid),
1198 export->exp_connection->c_peer.nid));
1200 class_export_put(export);
1203 export->exp_connecting = 1;
1204 spin_unlock(&export->exp_lock);
1205 LASSERT(export->exp_obd == target);
1207 rc = target_handle_reconnect(&conn, export, &cluuid);
1210 /* If we found an export, we already unlocked. */
1213 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1214 } else if (req->rq_export == NULL &&
1215 atomic_read(&export->exp_rpc_count) > 0) {
1216 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1217 "still busy with %d references\n",
1218 target->obd_name, cluuid.uuid,
1219 libcfs_nid2str(req->rq_peer.nid),
1220 atomic_read(&export->exp_refcount));
1221 GOTO(out, rc = -EBUSY);
1222 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1224 if (!strstr(cluuid.uuid, "mdt"))
1225 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1226 "known client %s (at %s) because it "
1227 "is indicating it is a new client",
1228 target->obd_name, cluuid.uuid,
1229 libcfs_nid2str(req->rq_peer.nid));
1230 GOTO(out, rc = -EALREADY);
1232 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1239 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1240 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1241 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1242 export, ktime_get_real_seconds(),
1243 export ? export->exp_last_request_time : 0);
1245 /* If this is the first time a client connects, reset the recovery
1246 * timer. Discard lightweight connections which might be local. */
1247 if (!lw_client && rc == 0 && target->obd_recovering)
1248 check_and_start_recovery_timer(target, req, export == NULL);
1250 /* We want to handle EALREADY but *not* -EALREADY from
1251 * target_handle_reconnect(), return reconnection state in a flag. */
1252 if (rc == EALREADY) {
1253 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1259 /* Tell the client if we support replayable requests. */
1260 if (target->obd_replayable)
1261 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1262 client_nid = &req->rq_peer.nid;
1264 if (export == NULL) {
1265 /* allow lightweight connections during recovery */
1266 /* allow "new" MDT to be connected during recovery, since we
1267 * need retrieve recovery update records from it */
1268 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1272 int c; /* connected */
1273 int i; /* in progress */
1275 int s; /* stale/evicted */
1277 c = atomic_read(&target->obd_connected_clients);
1278 i = atomic_read(&target->obd_lock_replay_clients);
1279 k = target->obd_max_recoverable_clients;
1280 s = target->obd_stale_clients;
1281 t = jiffies_to_msecs(target->obd_recovery_timer.expires);
1283 now = ktime_get_seconds();
1286 msg = "already passed deadline";
1289 msg = "to recover in";
1292 LCONSOLE_WARN("%s: Denying connection for new client %s"
1293 "(at %s), waiting for %d known clients "
1294 "(%d recovered, %d in progress, and %d "
1295 "evicted) %s %lld:%.02lld\n",
1296 target->obd_name, cluuid.uuid,
1297 libcfs_nid2str(req->rq_peer.nid), k,
1298 c - i, i, s, msg, t / 60, t % 60);
1302 rc = obd_connect(req->rq_svc_thread->t_env,
1303 &export, target, &cluuid, data,
1305 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1306 lustre_msg_add_op_flags(req->rq_repmsg,
1307 MSG_CONNECT_RECOVERING);
1309 conn.cookie = export->exp_handle.h_cookie;
1310 rc = rev_import_init(export);
1314 new_mds_mds_conn = true;
1317 rc = obd_reconnect(req->rq_svc_thread->t_env,
1318 export, target, &cluuid, data, client_nid);
1323 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1324 data->ocd_instance = target->u.obt.obt_instance;
1326 /* Return only the parts of obd_connect_data that we understand, so the
1327 * client knows that we don't understand the rest. */
1329 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1331 tmpdata = req_capsule_server_get(&req->rq_pill,
1333 /* Don't use struct assignment here, because the client reply
1334 * buffer may be smaller/larger than the local struct
1335 * obd_connect_data. */
1336 memcpy(tmpdata, data, min(tmpsize, size));
1339 /* If the client and the server are the same node, we will already
1340 * have an export that really points to the client's DLM export,
1341 * because we have a shared handles table.
1343 * XXX this will go away when shaver stops sending the "connect" handle
1344 * in the real "remote handle" field of the request --phik 24 Apr 2003
1346 ptlrpc_request_change_export(req, export);
1348 spin_lock(&export->exp_lock);
1349 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1350 spin_unlock(&export->exp_lock);
1351 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1352 "or equal conn_cnt: %d >= %d\n",
1353 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1354 export->exp_conn_cnt,
1355 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1357 GOTO(out, rc = -EALREADY);
1359 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1360 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1361 spin_unlock(&export->exp_lock);
1363 if (export->exp_connection != NULL) {
1364 /* Check to see if connection came from another NID. */
1365 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1366 !hlist_unhashed(&export->exp_nid_hash))
1367 cfs_hash_del(export->exp_obd->obd_nid_hash,
1368 &export->exp_connection->c_peer.nid,
1369 &export->exp_nid_hash);
1371 ptlrpc_connection_put(export->exp_connection);
1374 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1377 if (hlist_unhashed(&export->exp_nid_hash))
1378 cfs_hash_add(export->exp_obd->obd_nid_hash,
1379 &export->exp_connection->c_peer.nid,
1380 &export->exp_nid_hash);
1382 lustre_msg_set_handle(req->rq_repmsg, &conn);
1384 rc = rev_import_reconnect(export, req);
1388 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1390 __u64 transno = data->ocd_transno;
1392 spin_lock(&export->exp_lock);
1393 /* possible race with class_disconnect_stale_exports,
1394 * export may be already in the eviction process */
1395 if (export->exp_failed) {
1396 spin_unlock(&export->exp_lock);
1397 GOTO(out, rc = -ENODEV);
1399 export->exp_in_recovery = 1;
1400 export->exp_req_replay_needed = 1;
1401 export->exp_lock_replay_needed = 1;
1402 spin_unlock(&export->exp_lock);
1404 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1405 MSG_CONNECT_TRANSNO);
1406 if (has_transno && transno == 0)
1407 CWARN("Connect with zero transno!\n");
1409 if (has_transno && transno > 0 &&
1410 transno < target->obd_next_recovery_transno &&
1411 transno > target->obd_last_committed) {
1412 /* Another way is to use cmpxchg() to be lock-free. */
1413 spin_lock(&target->obd_recovery_task_lock);
1414 if (transno < target->obd_next_recovery_transno)
1415 target->obd_next_recovery_transno = transno;
1416 spin_unlock(&target->obd_recovery_task_lock);
1419 atomic_inc(&target->obd_req_replay_clients);
1420 atomic_inc(&target->obd_lock_replay_clients);
1421 /* Note: MDS-MDS connection is allowed to be connected during
1422 * recovery, no matter if the exports needs to be recoveried.
1423 * Because we need retrieve updates logs from all other MDTs.
1424 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1425 * also needs to be increased to match other recovery checking
1427 if (new_mds_mds_conn)
1428 target->obd_max_recoverable_clients++;
1429 if (atomic_inc_return(&target->obd_connected_clients) ==
1430 target->obd_max_recoverable_clients)
1431 wake_up(&target->obd_next_transno_waitq);
1434 /* Tell the client we're in recovery, when client is involved in it. */
1435 if (target->obd_recovering && !lw_client)
1436 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1440 spin_lock(&export->exp_lock);
1441 export->exp_connecting = 0;
1442 spin_unlock(&export->exp_lock);
1444 class_export_put(export);
1446 if (target != NULL) {
1447 spin_lock(&target->obd_dev_lock);
1448 target->obd_conn_inprogress--;
1449 spin_unlock(&target->obd_dev_lock);
1450 class_decref(target, "find", current);
1452 req->rq_status = rc;
1456 int target_handle_disconnect(struct ptlrpc_request *req)
1461 rc = req_capsule_server_pack(&req->rq_pill);
1465 /* Keep the rq_export around so we can send the reply. */
1466 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1471 void target_destroy_export(struct obd_export *exp)
1473 struct obd_import *imp = NULL;
1474 /* exports created from last_rcvd data, and "fake"
1475 exports created by lctl don't have an import */
1476 spin_lock(&exp->exp_lock);
1477 if (exp->exp_imp_reverse != NULL) {
1478 imp = exp->exp_imp_reverse;
1479 exp->exp_imp_reverse = NULL;
1481 spin_unlock(&exp->exp_lock);
1483 client_destroy_import(imp);
1485 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1486 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1487 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1488 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1490 EXPORT_SYMBOL(target_destroy_export);
1493 * Recovery functions
1495 static void target_request_copy_get(struct ptlrpc_request *req)
1497 class_export_rpc_inc(req->rq_export);
1498 LASSERT(list_empty(&req->rq_list));
1499 INIT_LIST_HEAD(&req->rq_replay_list);
1501 /* Increase refcount to keep request in queue. */
1502 atomic_inc(&req->rq_refcount);
1503 /* Let export know it has replays to be handled. */
1504 atomic_inc(&req->rq_export->exp_replay_count);
1507 static void target_request_copy_put(struct ptlrpc_request *req)
1509 LASSERT(list_empty(&req->rq_replay_list));
1510 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1512 atomic_dec(&req->rq_export->exp_replay_count);
1513 class_export_rpc_dec(req->rq_export);
1514 ptlrpc_server_drop_request(req);
1517 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1519 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1520 struct obd_export *exp = req->rq_export;
1521 struct ptlrpc_request *reqiter;
1522 struct ptlrpc_request *dup_req = NULL;
1527 spin_lock(&exp->exp_lock);
1528 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1530 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1538 /* We expect it with RESENT and REPLAY flags. */
1539 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1540 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1541 CERROR("invalid flags %x of resent replay\n",
1542 lustre_msg_get_flags(req->rq_reqmsg));
1544 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1547 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1549 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1550 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1554 list_add_tail(&req->rq_replay_list,
1555 &exp->exp_req_replay_queue);
1558 spin_unlock(&exp->exp_lock);
1562 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1564 LASSERT(!list_empty(&req->rq_replay_list));
1565 LASSERT(req->rq_export);
1567 spin_lock(&req->rq_export->exp_lock);
1568 list_del_init(&req->rq_replay_list);
1569 spin_unlock(&req->rq_export->exp_lock);
1572 static void target_finish_recovery(struct lu_target *lut)
1574 struct obd_device *obd = lut->lut_obd;
1577 /* Only log a recovery message when recovery has occurred. */
1578 if (obd->obd_recovery_start) {
1579 time64_t now = ktime_get_real_seconds();
1580 time64_t elapsed_time;
1582 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1583 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1584 "%d recovered and %d %s evicted.\n", obd->obd_name,
1585 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1586 obd->obd_max_recoverable_clients,
1587 atomic_read(&obd->obd_connected_clients),
1588 obd->obd_stale_clients,
1589 obd->obd_stale_clients == 1 ? "was" : "were");
1592 ldlm_reprocess_recovery_done(obd->obd_namespace);
1593 spin_lock(&obd->obd_recovery_task_lock);
1594 if (!list_empty(&obd->obd_req_replay_queue) ||
1595 !list_empty(&obd->obd_lock_replay_queue) ||
1596 !list_empty(&obd->obd_final_req_queue)) {
1597 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1599 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1600 list_empty(&obd->obd_lock_replay_queue) ? \
1602 list_empty(&obd->obd_final_req_queue) ? \
1604 spin_unlock(&obd->obd_recovery_task_lock);
1607 spin_unlock(&obd->obd_recovery_task_lock);
1609 obd->obd_recovery_end = ktime_get_real_seconds();
1611 /* When recovery finished, cleanup orphans on MDS and OST. */
1612 if (obd->obd_type && OBP(obd, postrecov)) {
1613 int rc = OBP(obd, postrecov)(obd);
1616 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1622 static void abort_req_replay_queue(struct obd_device *obd)
1624 struct ptlrpc_request *req, *n;
1625 struct list_head abort_list;
1627 INIT_LIST_HEAD(&abort_list);
1628 spin_lock(&obd->obd_recovery_task_lock);
1629 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1630 spin_unlock(&obd->obd_recovery_task_lock);
1631 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1632 DEBUG_REQ(D_WARNING, req, "aborted:");
1633 req->rq_status = -ENOTCONN;
1634 if (ptlrpc_error(req)) {
1635 DEBUG_REQ(D_ERROR, req,
1636 "failed abort_req_reply; skipping");
1638 target_exp_dequeue_req_replay(req);
1639 target_request_copy_put(req);
1643 static void abort_lock_replay_queue(struct obd_device *obd)
1645 struct ptlrpc_request *req, *n;
1646 struct list_head abort_list;
1648 INIT_LIST_HEAD(&abort_list);
1649 spin_lock(&obd->obd_recovery_task_lock);
1650 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1651 spin_unlock(&obd->obd_recovery_task_lock);
1652 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1653 DEBUG_REQ(D_ERROR, req, "aborted:");
1654 req->rq_status = -ENOTCONN;
1655 if (ptlrpc_error(req)) {
1656 DEBUG_REQ(D_ERROR, req,
1657 "failed abort_lock_reply; skipping");
1659 target_request_copy_put(req);
1663 /* Called from a cleanup function if the device is being cleaned up
1664 forcefully. The exports should all have been disconnected already,
1665 the only thing left to do is
1666 - clear the recovery flags
1668 - free queued requests and replies, but don't send replies
1669 Because the obd_stopping flag is set, no new requests should be received.
1672 void target_cleanup_recovery(struct obd_device *obd)
1674 struct ptlrpc_request *req, *n;
1675 struct list_head clean_list;
1677 INIT_LIST_HEAD(&clean_list);
1678 spin_lock(&obd->obd_dev_lock);
1679 if (!obd->obd_recovering) {
1680 spin_unlock(&obd->obd_dev_lock);
1684 obd->obd_recovering = obd->obd_abort_recovery = 0;
1685 spin_unlock(&obd->obd_dev_lock);
1687 spin_lock(&obd->obd_recovery_task_lock);
1688 target_cancel_recovery_timer(obd);
1689 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1690 spin_unlock(&obd->obd_recovery_task_lock);
1692 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1693 LASSERT(req->rq_reply_state == NULL);
1694 target_exp_dequeue_req_replay(req);
1695 target_request_copy_put(req);
1698 spin_lock(&obd->obd_recovery_task_lock);
1699 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1700 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1701 spin_unlock(&obd->obd_recovery_task_lock);
1703 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1704 LASSERT(req->rq_reply_state == NULL);
1705 target_request_copy_put(req);
1710 EXPORT_SYMBOL(target_cleanup_recovery);
1712 /* obd_recovery_task_lock should be held */
1713 void target_cancel_recovery_timer(struct obd_device *obd)
1715 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1716 del_timer(&obd->obd_recovery_timer);
1719 static void target_start_recovery_timer(struct obd_device *obd)
1721 if (obd->obd_recovery_start != 0)
1724 spin_lock(&obd->obd_dev_lock);
1725 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1726 spin_unlock(&obd->obd_dev_lock);
1730 LASSERT(obd->obd_recovery_timeout != 0);
1732 if (obd->obd_recovery_start != 0) {
1733 spin_unlock(&obd->obd_dev_lock);
1737 mod_timer(&obd->obd_recovery_timer,
1738 jiffies + cfs_time_seconds(obd->obd_recovery_timeout));
1739 obd->obd_recovery_start = ktime_get_real_seconds();
1740 spin_unlock(&obd->obd_dev_lock);
1742 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1744 obd->obd_recovery_timeout / 60,
1745 obd->obd_recovery_timeout % 60,
1746 obd->obd_max_recoverable_clients,
1747 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1748 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1752 * extend recovery window.
1754 * if @extend is true, extend recovery window to have @drt remaining at least;
1755 * otherwise, make sure the recovery timeout value is not less than @drt.
1757 static void extend_recovery_timer(struct obd_device *obd, time64_t drt,
1765 spin_lock(&obd->obd_dev_lock);
1766 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1767 spin_unlock(&obd->obd_dev_lock);
1770 LASSERT(obd->obd_recovery_start != 0);
1772 now = ktime_get_real_seconds();
1773 to = obd->obd_recovery_timeout;
1774 end = obd->obd_recovery_start + to;
1777 if (extend && (drt > left)) {
1779 } else if (!extend && (drt > to)) {
1783 if (to > obd->obd_recovery_time_hard) {
1784 to = obd->obd_recovery_time_hard;
1785 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1786 obd->obd_name, to, extend);
1789 if (obd->obd_recovery_timeout < to) {
1790 obd->obd_recovery_timeout = to;
1791 end = obd->obd_recovery_start + to;
1792 mod_timer(&obd->obd_recovery_timer,
1793 jiffies + cfs_time_seconds(end - now));
1795 spin_unlock(&obd->obd_dev_lock);
1797 CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1798 obd->obd_name, (s64)(end - now));
1801 /* Reset the timer with each new client connection */
1803 * This timer is actually reconnect_timer, which is for making sure
1804 * the total recovery window is at least as big as my reconnect
1805 * attempt timing. So the initial recovery time_out will be set to
1806 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1807 * from client is bigger than this, then the recovery time_out will
1808 * be extended to make sure the client could be reconnected, in the
1809 * process, the timeout from the new client should be ignored.
1813 check_and_start_recovery_timer(struct obd_device *obd,
1814 struct ptlrpc_request *req,
1817 time64_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1818 struct obd_device_target *obt = &obd->u.obt;
1820 if (!new_client && service_time)
1821 /* Teach server about old server's estimates, as first guess
1822 * at how long new requests will take. */
1823 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1826 target_start_recovery_timer(obd);
1828 /* Convert the service time to RPC timeout,
1829 * and reuse service_time to limit stack usage.
1831 service_time = at_est2timeout(service_time);
1833 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1834 service_time < at_extra)
1835 service_time = at_extra;
1837 /* We expect other clients to timeout within service_time, then try
1838 * to reconnect, then try the failover server. The max delay between
1839 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1840 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1842 LASSERT(obt->obt_magic == OBT_MAGIC);
1843 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1844 if (service_time > obd->obd_recovery_timeout && !new_client)
1845 extend_recovery_timer(obd, service_time, false);
1848 /** Health checking routines */
1849 static inline int exp_connect_healthy(struct obd_export *exp)
1851 return (exp->exp_in_recovery);
1854 /** if export done req_replay or has replay in queue */
1855 static inline int exp_req_replay_healthy(struct obd_export *exp)
1857 return (!exp->exp_req_replay_needed ||
1858 atomic_read(&exp->exp_replay_count) > 0);
1862 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1864 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1865 exp_req_replay_healthy(exp);
1868 /** if export done lock_replay or has replay in queue */
1869 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1871 return (!exp->exp_lock_replay_needed ||
1872 atomic_read(&exp->exp_replay_count) > 0);
1875 static inline int exp_vbr_healthy(struct obd_export *exp)
1877 return (!exp->exp_vbr_failed);
1880 static inline int exp_finished(struct obd_export *exp)
1882 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1885 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1887 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1891 static int check_for_next_transno(struct lu_target *lut)
1893 struct ptlrpc_request *req = NULL;
1894 struct obd_device *obd = lut->lut_obd;
1895 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1896 int wake_up = 0, connected, completed, queue_len;
1897 __u64 req_transno = 0;
1898 __u64 update_transno = 0;
1899 __u64 next_transno = 0;
1902 spin_lock(&obd->obd_recovery_task_lock);
1903 if (!list_empty(&obd->obd_req_replay_queue)) {
1904 req = list_entry(obd->obd_req_replay_queue.next,
1905 struct ptlrpc_request, rq_list);
1906 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1910 update_transno = distribute_txn_get_next_transno(tdtd);
1912 connected = atomic_read(&obd->obd_connected_clients);
1913 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1914 queue_len = obd->obd_requests_queued_for_recovery;
1915 next_transno = obd->obd_next_recovery_transno;
1917 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1918 "req_transno: %llu, next_transno: %llu\n",
1919 obd->obd_max_recoverable_clients, connected, completed,
1920 queue_len, req_transno, next_transno);
1922 if (obd->obd_abort_recovery) {
1923 CDEBUG(D_HA, "waking for aborted recovery\n");
1925 } else if (obd->obd_recovery_expired) {
1926 CDEBUG(D_HA, "waking for expired recovery\n");
1928 } else if (tdtd != NULL && req != NULL &&
1929 is_req_replayed_by_update(req)) {
1930 LASSERTF(req_transno < next_transno, "req_transno %llu"
1931 "next_transno%llu\n", req_transno, next_transno);
1932 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1935 } else if (req_transno == next_transno ||
1936 (update_transno != 0 && update_transno <= next_transno)) {
1937 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1939 } else if (queue_len > 0 &&
1940 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1941 /** handle gaps occured due to lost reply or VBR */
1942 LASSERTF(req_transno >= next_transno,
1943 "req_transno: %llu, next_transno: %llu\n",
1944 req_transno, next_transno);
1946 "%s: waking for gap in transno, VBR is %s (skip: "
1947 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1948 ", next_update %lld last_committed: %lld)\n",
1949 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1950 next_transno, queue_len, completed, connected,
1951 req_transno, update_transno, obd->obd_last_committed);
1952 obd->obd_next_recovery_transno = req_transno;
1954 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1955 CDEBUG(D_HA, "waking for completed recovery\n");
1957 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1958 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1959 " by fail_lock, waking up (%lld)\n", next_transno);
1960 obd->obd_next_recovery_transno = req_transno;
1963 spin_unlock(&obd->obd_recovery_task_lock);
1967 static int check_for_next_lock(struct lu_target *lut)
1969 struct obd_device *obd = lut->lut_obd;
1972 spin_lock(&obd->obd_recovery_task_lock);
1973 if (!list_empty(&obd->obd_lock_replay_queue)) {
1974 CDEBUG(D_HA, "waking for next lock\n");
1976 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1977 CDEBUG(D_HA, "waking for completed lock replay\n");
1979 } else if (obd->obd_abort_recovery) {
1980 CDEBUG(D_HA, "waking for aborted recovery\n");
1982 } else if (obd->obd_recovery_expired) {
1983 CDEBUG(D_HA, "waking for expired recovery\n");
1986 spin_unlock(&obd->obd_recovery_task_lock);
1992 * wait for recovery events,
1993 * check its status with help of check_routine
1994 * evict dead clients via health_check
1996 static int target_recovery_overseer(struct lu_target *lut,
1997 int (*check_routine)(struct lu_target *),
1998 int (*health_check)(struct obd_export *))
2000 struct obd_device *obd = lut->lut_obd;
2001 struct target_distribute_txn_data *tdtd;
2005 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2006 now = ktime_get_seconds();
2007 if (now - last > 600) {
2008 LCONSOLE_INFO("%s: in recovery but waiting for "
2009 "the first client to connect\n",
2014 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
2015 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2016 __u64 next_update_transno = 0;
2018 /* Only abort the recovery if there are no update recovery
2019 * left in the queue */
2020 spin_lock(&obd->obd_recovery_task_lock);
2021 if (lut->lut_tdtd != NULL) {
2022 next_update_transno =
2023 distribute_txn_get_next_transno(lut->lut_tdtd);
2025 tdtd = lut->lut_tdtd;
2026 /* If next_update_transno == 0, it probably because
2027 * updatelog retrieve threads did not get any records
2028 * yet, let's wait those threads stopped */
2029 if (next_update_transno == 0) {
2030 struct l_wait_info lwi = { 0 };
2032 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2034 &tdtd->tdtd_recovery_threads_count) == 0,
2037 next_update_transno =
2038 distribute_txn_get_next_transno(
2043 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2044 obd->obd_next_recovery_transno = next_update_transno;
2045 spin_unlock(&obd->obd_recovery_task_lock);
2046 /* Disconnect unfinished exports from clients, and
2047 * keep connection from MDT to make sure the update
2048 * recovery will still keep trying until some one
2049 * manually abort the recovery */
2050 class_disconnect_stale_exports(obd,
2051 exp_finished_or_from_mdt);
2052 /* Abort all of replay and replay lock req from
2054 abort_req_replay_queue(obd);
2055 abort_lock_replay_queue(obd);
2056 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2057 ")in the queue.\n", obd->obd_name,
2058 next_update_transno);
2060 obd->obd_abort_recovery = 1;
2061 spin_unlock(&obd->obd_recovery_task_lock);
2062 CWARN("%s recovery is aborted by hard timeout\n",
2067 while (wait_event_timeout(obd->obd_next_transno_waitq,
2069 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2070 /* wait indefinitely for event, but don't trigger watchdog */;
2072 if (obd->obd_abort_recovery) {
2073 CWARN("recovery is aborted, evict exports in recovery\n");
2074 if (lut->lut_tdtd != NULL) {
2075 struct l_wait_info lwi = { 0 };
2077 tdtd = lut->lut_tdtd;
2078 /* Let's wait all of the update log recovery thread
2080 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2081 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2083 /* Then abort the update recovery list */
2084 dtrq_list_destroy(lut->lut_tdtd);
2087 /** evict exports which didn't finish recovery yet */
2088 class_disconnect_stale_exports(obd, exp_finished);
2090 } else if (obd->obd_recovery_expired) {
2091 obd->obd_recovery_expired = 0;
2092 /** If some clients died being recovered, evict them */
2093 LCONSOLE_WARN("%s: recovery is timed out, "
2094 "evict stale exports\n", obd->obd_name);
2095 /** evict cexports with no replay in queue, they are stalled */
2096 class_disconnect_stale_exports(obd, health_check);
2098 /** continue with VBR */
2099 spin_lock(&obd->obd_dev_lock);
2100 obd->obd_version_recov = 1;
2101 spin_unlock(&obd->obd_dev_lock);
2103 * reset timer, recovery will proceed with versions now,
2104 * timeout is set just to handle reconnection delays
2106 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2107 /** Wait for recovery events again, after evicting bad clients */
2113 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2115 struct obd_device *obd = lut->lut_obd;
2116 struct ptlrpc_request *req = NULL;
2118 CDEBUG(D_HA, "Waiting for lock\n");
2119 if (target_recovery_overseer(lut, check_for_next_lock,
2120 exp_lock_replay_healthy))
2121 abort_lock_replay_queue(obd);
2123 spin_lock(&obd->obd_recovery_task_lock);
2124 if (!list_empty(&obd->obd_lock_replay_queue)) {
2125 req = list_entry(obd->obd_lock_replay_queue.next,
2126 struct ptlrpc_request, rq_list);
2127 list_del_init(&req->rq_list);
2128 spin_unlock(&obd->obd_recovery_task_lock);
2130 spin_unlock(&obd->obd_recovery_task_lock);
2131 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2132 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2133 /** evict exports failed VBR */
2134 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2139 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2141 struct ptlrpc_request *req = NULL;
2143 spin_lock(&obd->obd_recovery_task_lock);
2144 if (!list_empty(&obd->obd_final_req_queue)) {
2145 req = list_entry(obd->obd_final_req_queue.next,
2146 struct ptlrpc_request, rq_list);
2147 list_del_init(&req->rq_list);
2148 spin_unlock(&obd->obd_recovery_task_lock);
2149 if (req->rq_export->exp_in_recovery) {
2150 spin_lock(&req->rq_export->exp_lock);
2151 req->rq_export->exp_in_recovery = 0;
2152 spin_unlock(&req->rq_export->exp_lock);
2155 spin_unlock(&obd->obd_recovery_task_lock);
2160 static void handle_recovery_req(struct ptlrpc_thread *thread,
2161 struct ptlrpc_request *req,
2162 svc_handler_t handler)
2167 * export can be evicted during recovery, no need to handle replays for
2168 * it after that, discard such request silently
2170 if (req->rq_export->exp_disconnected)
2173 req->rq_session.lc_thread = thread;
2174 req->rq_svc_thread = thread;
2175 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2177 /* thread context */
2178 lu_context_enter(&thread->t_env->le_ctx);
2180 lu_context_exit(&thread->t_env->le_ctx);
2182 /* don't reset timer for final stage */
2183 if (!exp_finished(req->rq_export)) {
2184 time64_t to = obd_timeout;
2187 * Add request timeout to the recovery time so next request from
2188 * this client may come in recovery time
2191 struct ptlrpc_service_part *svcpt;
2193 svcpt = req->rq_rqbd->rqbd_svcpt;
2194 /* If the server sent early reply for this request,
2195 * the client will recalculate the timeout according to
2196 * current server estimate service time, so we will
2197 * use the maxium timeout here for waiting the client
2198 * sending the next req */
2199 to = max((int)at_est2timeout(
2200 at_get(&svcpt->scp_at_estimate)),
2201 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2202 /* Add 2 net_latency, one for balance rq_deadline
2203 * (see ptl_send_rpc), one for resend the req to server,
2204 * Note: client will pack net_latency in replay req
2205 * (see ptlrpc_replay_req) */
2206 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2208 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2213 /** Checking routines for recovery */
2214 static int check_for_recovery_ready(struct lu_target *lut)
2216 struct obd_device *obd = lut->lut_obd;
2217 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2219 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2220 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2221 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2222 obd->obd_recovery_expired);
2224 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2225 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2226 if (clnts + obd->obd_stale_clients <
2227 obd->obd_max_recoverable_clients)
2231 if (lut->lut_tdtd != NULL) {
2232 if (!lut->lut_tdtd->tdtd_replay_ready &&
2233 !obd->obd_abort_recovery) {
2234 /* Let's extend recovery timer, in case the recovery
2235 * timer expired, and some clients got evicted */
2236 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2238 CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2239 obd->obd_name, obd->obd_recovery_timeout);
2248 REQUEST_RECOVERY = 1,
2249 UPDATE_RECOVERY = 2,
2252 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2256 if (!list_empty(&obd->obd_req_replay_queue)) {
2257 struct ptlrpc_request *req;
2259 req = list_entry(obd->obd_req_replay_queue.next,
2260 struct ptlrpc_request, rq_list);
2261 transno = lustre_msg_get_transno(req->rq_reqmsg);
2267 static __u64 get_next_transno(struct lu_target *lut, int *type)
2269 struct obd_device *obd = lut->lut_obd;
2270 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2272 __u64 update_transno;
2275 transno = get_next_replay_req_transno(obd);
2277 *type = REQUEST_RECOVERY;
2282 update_transno = distribute_txn_get_next_transno(tdtd);
2283 if (transno == 0 || (transno >= update_transno &&
2284 update_transno != 0)) {
2285 transno = update_transno;
2287 *type = UPDATE_RECOVERY;
2294 * drop duplicate replay request
2296 * Because the operation has been replayed by update recovery, the request
2297 * with the same transno will be dropped and also notify the client to send
2298 * next replay request.
2300 * \param[in] env execution environment
2301 * \param[in] obd failover obd device
2302 * \param[in] req request to be dropped
2304 static void drop_duplicate_replay_req(struct lu_env *env,
2305 struct obd_device *obd,
2306 struct ptlrpc_request *req)
2308 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2309 " update records are found.\n",
2310 lustre_msg_get_transno(req->rq_reqmsg),
2311 libcfs_nid2str(req->rq_peer.nid));
2313 /* Right now, only for MDS reint operation update replay and
2314 * normal request replay can have the same transno */
2315 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2316 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2317 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2318 if (likely(req->rq_export))
2319 target_committed_to_req(req);
2320 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2321 target_send_reply(req, req->rq_status, 0);
2323 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2324 libcfs_nid2str(req->rq_peer.nid));
2326 target_exp_dequeue_req_replay(req);
2327 target_request_copy_put(req);
2328 obd->obd_replayed_requests++;
2331 static void replay_request_or_update(struct lu_env *env,
2332 struct lu_target *lut,
2333 struct target_recovery_data *trd,
2334 struct ptlrpc_thread *thread)
2336 struct obd_device *obd = lut->lut_obd;
2337 struct ptlrpc_request *req = NULL;
2342 CDEBUG(D_HA, "Waiting for transno %lld\n",
2343 obd->obd_next_recovery_transno);
2345 /* Replay all of request and update by transno */
2347 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2349 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2351 /** It is needed to extend recovery window above
2352 * recovery_time_soft. Extending is possible only in the
2353 * end of recovery window (see more details in
2354 * handle_recovery_req()).
2356 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2358 if (target_recovery_overseer(lut, check_for_next_transno,
2359 exp_req_replay_healthy_or_from_mdt)) {
2360 abort_req_replay_queue(obd);
2361 abort_lock_replay_queue(obd);
2365 spin_lock(&obd->obd_recovery_task_lock);
2366 transno = get_next_transno(lut, &type);
2367 if (type == REQUEST_RECOVERY && transno != 0) {
2368 /* Drop replay request from client side, if the
2369 * replay has been executed by update with the
2371 req = list_entry(obd->obd_req_replay_queue.next,
2372 struct ptlrpc_request, rq_list);
2374 list_del_init(&req->rq_list);
2375 obd->obd_requests_queued_for_recovery--;
2376 spin_unlock(&obd->obd_recovery_task_lock);
2378 /* Let's check if the request has been redone by
2380 if (is_req_replayed_by_update(req)) {
2381 struct distribute_txn_replay_req *dtrq;
2383 dtrq = distribute_txn_lookup_finish_list(tdtd,
2385 LASSERT(dtrq != NULL);
2386 spin_lock(&tdtd->tdtd_replay_list_lock);
2387 list_del_init(&dtrq->dtrq_list);
2388 spin_unlock(&tdtd->tdtd_replay_list_lock);
2391 drop_duplicate_replay_req(env, obd, req);
2396 LASSERT(trd->trd_processing_task == current_pid());
2397 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2398 lustre_msg_get_transno(req->rq_reqmsg),
2399 libcfs_nid2str(req->rq_peer.nid));
2401 handle_recovery_req(thread, req,
2402 trd->trd_recovery_handler);
2404 * bz18031: increase next_recovery_transno before
2405 * target_request_copy_put() will drop exp_rpc reference
2407 spin_lock(&obd->obd_recovery_task_lock);
2408 obd->obd_next_recovery_transno++;
2409 spin_unlock(&obd->obd_recovery_task_lock);
2410 target_exp_dequeue_req_replay(req);
2411 target_request_copy_put(req);
2412 obd->obd_replayed_requests++;
2413 } else if (type == UPDATE_RECOVERY && transno != 0) {
2414 struct distribute_txn_replay_req *dtrq;
2417 spin_unlock(&obd->obd_recovery_task_lock);
2419 LASSERT(tdtd != NULL);
2420 dtrq = distribute_txn_get_next_req(tdtd);
2421 lu_context_enter(&thread->t_env->le_ctx);
2422 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2423 lu_context_exit(&thread->t_env->le_ctx);
2424 extend_recovery_timer(obd, obd_timeout, true);
2426 if (rc == 0 && dtrq->dtrq_xid != 0) {
2427 CDEBUG(D_HA, "Move x%llu t%llu"
2428 " to finish list\n", dtrq->dtrq_xid,
2429 dtrq->dtrq_master_transno);
2431 /* Add it to the replay finish list */
2432 spin_lock(&tdtd->tdtd_replay_list_lock);
2433 list_add(&dtrq->dtrq_list,
2434 &tdtd->tdtd_replay_finish_list);
2435 spin_unlock(&tdtd->tdtd_replay_list_lock);
2437 spin_lock(&obd->obd_recovery_task_lock);
2438 if (transno == obd->obd_next_recovery_transno)
2439 obd->obd_next_recovery_transno++;
2441 obd->obd_next_recovery_transno)
2442 obd->obd_next_recovery_transno =
2444 spin_unlock(&obd->obd_recovery_task_lock);
2449 spin_unlock(&obd->obd_recovery_task_lock);
2451 LASSERT(list_empty(&obd->obd_req_replay_queue));
2452 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2453 /** evict exports failed VBR */
2454 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2460 static int target_recovery_thread(void *arg)
2462 struct lu_target *lut = arg;
2463 struct obd_device *obd = lut->lut_obd;
2464 struct ptlrpc_request *req;
2465 struct target_recovery_data *trd = &obd->obd_recovery_data;
2466 unsigned long delta;
2468 struct ptlrpc_thread *thread = NULL;
2472 unshare_fs_struct();
2473 OBD_ALLOC_PTR(thread);
2479 OBD_FREE_PTR(thread);
2483 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2485 OBD_FREE_PTR(thread);
2490 thread->t_env = env;
2491 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2492 env->le_ctx.lc_thread = thread;
2493 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2495 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2497 trd->trd_processing_task = current_pid();
2499 spin_lock(&obd->obd_dev_lock);
2500 obd->obd_recovering = 1;
2501 spin_unlock(&obd->obd_dev_lock);
2502 complete(&trd->trd_starting);
2504 /* first of all, we have to know the first transno to replay */
2505 if (target_recovery_overseer(lut, check_for_recovery_ready,
2506 exp_connect_healthy)) {
2507 abort_req_replay_queue(obd);
2508 abort_lock_replay_queue(obd);
2509 if (lut->lut_tdtd != NULL)
2510 dtrq_list_destroy(lut->lut_tdtd);
2513 /* next stage: replay requests or update */
2515 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2516 atomic_read(&obd->obd_req_replay_clients),
2517 obd->obd_next_recovery_transno);
2518 replay_request_or_update(env, lut, trd, thread);
2521 * The second stage: replay locks
2523 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2524 atomic_read(&obd->obd_lock_replay_clients));
2525 while ((req = target_next_replay_lock(lut))) {
2526 LASSERT(trd->trd_processing_task == current_pid());
2527 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2528 libcfs_nid2str(req->rq_peer.nid));
2529 handle_recovery_req(thread, req,
2530 trd->trd_recovery_handler);
2531 target_request_copy_put(req);
2532 obd->obd_replayed_locks++;
2536 * The third stage: reply on final pings, at this moment all clients
2537 * must have request in final queue
2539 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2540 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2541 /** Update server last boot epoch */
2542 tgt_boot_epoch_update(lut);
2543 /* We drop recoverying flag to forward all new requests
2544 * to regular mds_handle() since now */
2545 spin_lock(&obd->obd_dev_lock);
2546 obd->obd_recovering = obd->obd_abort_recovery = 0;
2547 spin_unlock(&obd->obd_dev_lock);
2548 spin_lock(&obd->obd_recovery_task_lock);
2549 target_cancel_recovery_timer(obd);
2550 spin_unlock(&obd->obd_recovery_task_lock);
2551 while ((req = target_next_final_ping(obd))) {
2552 LASSERT(trd->trd_processing_task == current_pid());
2553 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2554 libcfs_nid2str(req->rq_peer.nid));
2555 handle_recovery_req(thread, req,
2556 trd->trd_recovery_handler);
2557 /* Because the waiting client can not send ping to server,
2558 * so we need refresh the last_request_time, to avoid the
2559 * export is being evicted */
2560 ptlrpc_update_export_timer(req->rq_export, 0);
2561 target_request_copy_put(req);
2564 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2565 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2566 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2567 if (delta > OBD_RECOVERY_TIME_SOFT) {
2568 CWARN("too long recovery - read logs\n");
2569 libcfs_debug_dumplog();
2572 target_finish_recovery(lut);
2574 lu_context_fini(&env->le_ctx);
2575 trd->trd_processing_task = 0;
2576 complete(&trd->trd_finishing);
2578 tgt_io_thread_done(thread);
2579 OBD_FREE_PTR(thread);
2584 static int target_start_recovery_thread(struct lu_target *lut,
2585 svc_handler_t handler)
2587 struct obd_device *obd = lut->lut_obd;
2589 struct target_recovery_data *trd = &obd->obd_recovery_data;
2592 memset(trd, 0, sizeof(*trd));
2593 init_completion(&trd->trd_starting);
2594 init_completion(&trd->trd_finishing);
2595 trd->trd_recovery_handler = handler;
2597 rc = server_name2index(obd->obd_name, &index, NULL);
2601 if (!IS_ERR(kthread_run(target_recovery_thread,
2602 lut, "tgt_recover_%d", index))) {
2603 wait_for_completion(&trd->trd_starting);
2604 LASSERT(obd->obd_recovering != 0);
2612 void target_stop_recovery_thread(struct obd_device *obd)
2614 if (obd->obd_recovery_data.trd_processing_task > 0) {
2615 struct target_recovery_data *trd = &obd->obd_recovery_data;
2616 /** recovery can be done but postrecovery is not yet */
2617 spin_lock(&obd->obd_dev_lock);
2618 if (obd->obd_recovering) {
2619 CERROR("%s: Aborting recovery\n", obd->obd_name);
2620 obd->obd_abort_recovery = 1;
2621 wake_up(&obd->obd_next_transno_waitq);
2623 spin_unlock(&obd->obd_dev_lock);
2624 wait_for_completion(&trd->trd_finishing);
2627 EXPORT_SYMBOL(target_stop_recovery_thread);
2629 void target_recovery_fini(struct obd_device *obd)
2631 class_disconnect_exports(obd);
2632 target_stop_recovery_thread(obd);
2633 target_cleanup_recovery(obd);
2635 EXPORT_SYMBOL(target_recovery_fini);
2637 static void target_recovery_expired(cfs_timer_cb_arg_t data)
2639 struct obd_device *obd = cfs_from_timer(obd, data, obd_recovery_timer);
2640 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2641 " after %llus (%d clients connected)\n",
2642 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2643 (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2644 atomic_read(&obd->obd_connected_clients));
2646 obd->obd_recovery_expired = 1;
2647 wake_up(&obd->obd_next_transno_waitq);
2650 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2652 struct obd_device *obd = lut->lut_obd;
2654 if (lut->lut_bottom->dd_rdonly)
2657 if (obd->obd_max_recoverable_clients == 0) {
2658 /** Update server last boot epoch */
2659 tgt_boot_epoch_update(lut);
2663 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2664 "last_transno %llu\n", obd->obd_name,
2665 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2666 LASSERT(obd->obd_stopping == 0);
2667 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2668 obd->obd_recovery_start = 0;
2669 obd->obd_recovery_end = 0;
2671 cfs_timer_setup(&obd->obd_recovery_timer, target_recovery_expired,
2672 (unsigned long)obd, 0);
2673 target_start_recovery_thread(lut, handler);
2675 EXPORT_SYMBOL(target_recovery_init);
2677 static int target_process_req_flags(struct obd_device *obd,
2678 struct ptlrpc_request *req)
2680 struct obd_export *exp = req->rq_export;
2681 LASSERT(exp != NULL);
2682 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2683 /* client declares he's ready to replay locks */
2684 spin_lock(&exp->exp_lock);
2685 if (exp->exp_req_replay_needed) {
2686 exp->exp_req_replay_needed = 0;
2687 spin_unlock(&exp->exp_lock);
2689 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2690 atomic_dec(&obd->obd_req_replay_clients);
2692 spin_unlock(&exp->exp_lock);
2695 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2696 /* client declares he's ready to complete recovery
2697 * so, we put the request on th final queue */
2698 spin_lock(&exp->exp_lock);
2699 if (exp->exp_lock_replay_needed) {
2700 exp->exp_lock_replay_needed = 0;
2701 spin_unlock(&exp->exp_lock);
2703 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2704 atomic_dec(&obd->obd_lock_replay_clients);
2706 spin_unlock(&exp->exp_lock);
2712 int target_queue_recovery_request(struct ptlrpc_request *req,
2713 struct obd_device *obd)
2715 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2716 struct ptlrpc_request *reqiter;
2720 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2721 /* Processing the queue right now, don't re-add. */
2725 target_process_req_flags(obd, req);
2727 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2728 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2729 if (cfs_fail_val == 1) {
2732 wake_up(&cfs_race_waitq);
2734 set_current_state(TASK_INTERRUPTIBLE);
2735 schedule_timeout(cfs_time_seconds(1));
2739 /* client declares he's ready to complete recovery
2740 * so, we put the request on th final queue */
2741 target_request_copy_get(req);
2742 DEBUG_REQ(D_HA, req, "queue final req");
2743 wake_up(&obd->obd_next_transno_waitq);
2744 spin_lock(&obd->obd_recovery_task_lock);
2745 if (obd->obd_recovering) {
2746 struct ptlrpc_request *tmp;
2747 struct ptlrpc_request *duplicate = NULL;
2749 if (likely(!req->rq_export->exp_replay_done)) {
2750 req->rq_export->exp_replay_done = 1;
2751 list_add_tail(&req->rq_list,
2752 &obd->obd_final_req_queue);
2753 spin_unlock(&obd->obd_recovery_task_lock);
2757 /* XXX O(n), but only happens if final ping is
2758 * timed out, probably reorganize the list as
2759 * a hash list later */
2760 list_for_each_entry_safe(reqiter, tmp,
2761 &obd->obd_final_req_queue,
2763 if (reqiter->rq_export == req->rq_export) {
2764 list_del_init(&reqiter->rq_list);
2765 duplicate = reqiter;
2770 list_add_tail(&req->rq_list,
2771 &obd->obd_final_req_queue);
2772 req->rq_export->exp_replay_done = 1;
2773 spin_unlock(&obd->obd_recovery_task_lock);
2775 if (duplicate != NULL) {
2776 DEBUG_REQ(D_HA, duplicate,
2777 "put prev final req\n");
2778 target_request_copy_put(duplicate);
2782 spin_unlock(&obd->obd_recovery_task_lock);
2783 target_request_copy_put(req);
2784 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2787 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2788 /* client declares he's ready to replay locks */
2789 target_request_copy_get(req);
2790 DEBUG_REQ(D_HA, req, "queue lock replay req");
2791 wake_up(&obd->obd_next_transno_waitq);
2792 spin_lock(&obd->obd_recovery_task_lock);
2793 LASSERT(obd->obd_recovering);
2794 /* usually due to recovery abort */
2795 if (!req->rq_export->exp_in_recovery) {
2796 spin_unlock(&obd->obd_recovery_task_lock);
2797 target_request_copy_put(req);
2800 LASSERT(req->rq_export->exp_lock_replay_needed);
2801 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2802 spin_unlock(&obd->obd_recovery_task_lock);
2806 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2807 * (i.e. buflens etc are in my own byte order), but type-dependent
2808 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2811 INIT_LIST_HEAD(&req->rq_list);
2812 DEBUG_REQ(D_HA, req, "not queueing");
2816 /* If we're processing the queue, we want don't want to queue this
2819 * Also, if this request has a transno less than the one we're waiting
2820 * for, we should process it now. It could (and currently always will)
2821 * be an open request for a descriptor that was opened some time ago.
2823 * Also, a resent, replayed request that has already been
2824 * handled will pass through here and be processed immediately.
2826 CDEBUG(D_HA, "Next recovery transno: %llu"
2827 ", current: %llu, replaying\n",
2828 obd->obd_next_recovery_transno, transno);
2830 /* If the request has been replayed by update replay, then sends this
2831 * request to the recovery thread (replay_request_or_update()), where
2832 * it will be handled */
2833 spin_lock(&obd->obd_recovery_task_lock);
2834 if (transno < obd->obd_next_recovery_transno &&
2835 !is_req_replayed_by_update(req)) {
2836 /* Processing the queue right now, don't re-add. */
2837 LASSERT(list_empty(&req->rq_list));
2838 spin_unlock(&obd->obd_recovery_task_lock);
2841 spin_unlock(&obd->obd_recovery_task_lock);
2843 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2846 target_request_copy_get(req);
2847 if (!req->rq_export->exp_in_recovery) {
2848 target_request_copy_put(req);
2851 LASSERT(req->rq_export->exp_req_replay_needed);
2853 if (target_exp_enqueue_req_replay(req)) {
2854 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2855 target_request_copy_put(req);
2860 spin_lock(&obd->obd_recovery_task_lock);
2861 LASSERT(obd->obd_recovering);
2862 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2863 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2864 list_add_tail(&req->rq_list, &reqiter->rq_list);
2869 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2871 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2872 "has been claimed by another client");
2873 spin_unlock(&obd->obd_recovery_task_lock);
2874 target_exp_dequeue_req_replay(req);
2875 target_request_copy_put(req);
2881 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2883 obd->obd_requests_queued_for_recovery++;
2884 spin_unlock(&obd->obd_recovery_task_lock);
2885 wake_up(&obd->obd_next_transno_waitq);
2889 void target_committed_to_req(struct ptlrpc_request *req)
2891 struct obd_export *exp = req->rq_export;
2893 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2894 lustre_msg_set_last_committed(req->rq_repmsg,
2895 exp->exp_last_committed);
2897 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2898 "%d)", exp->exp_obd->obd_no_transno,
2899 req->rq_repmsg == NULL);
2901 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2902 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2905 #endif /* HAVE_SERVER_SUPPORT */
2908 * Packs current SLV and Limit into \a req.
2910 int target_pack_pool_reply(struct ptlrpc_request *req)
2912 struct obd_device *obd;
2915 /* Check that we still have all structures alive as this may
2916 * be some late RPC at shutdown time. */
2917 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2918 !exp_connect_lru_resize(req->rq_export))) {
2919 lustre_msg_set_slv(req->rq_repmsg, 0);
2920 lustre_msg_set_limit(req->rq_repmsg, 0);
2924 /* OBD is alive here as export is alive, which we checked above. */
2925 obd = req->rq_export->exp_obd;
2927 read_lock(&obd->obd_pool_lock);
2928 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2929 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2930 read_unlock(&obd->obd_pool_lock);
2935 static int target_send_reply_msg(struct ptlrpc_request *req,
2936 int rc, int fail_id)
2938 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2939 DEBUG_REQ(D_ERROR, req, "dropping reply");
2942 /* We can have a null rq_reqmsg in the event of bad signature or
2943 * no context when unwrapping */
2944 if (req->rq_reqmsg &&
2945 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2946 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2950 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2951 req->rq_status = rc;
2952 return ptlrpc_send_error(req, 1);
2954 DEBUG_REQ(D_NET, req, "sending reply");
2957 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2960 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2962 struct ptlrpc_service_part *svcpt;
2964 struct ptlrpc_reply_state *rs;
2965 struct obd_export *exp;
2968 if (req->rq_no_reply) {
2973 svcpt = req->rq_rqbd->rqbd_svcpt;
2974 rs = req->rq_reply_state;
2975 if (rs == NULL || !rs->rs_difficult) {
2977 target_send_reply_msg (req, rc, fail_id);
2982 /* must be an export if locks saved */
2983 LASSERT(req->rq_export != NULL);
2984 /* req/reply consistent */
2985 LASSERT(rs->rs_svcpt == svcpt);
2988 LASSERT(!rs->rs_scheduled);
2989 LASSERT(!rs->rs_scheduled_ever);
2990 LASSERT(!rs->rs_handled);
2991 LASSERT(!rs->rs_on_net);
2992 LASSERT(rs->rs_export == NULL);
2993 LASSERT(list_empty(&rs->rs_obd_list));
2994 LASSERT(list_empty(&rs->rs_exp_list));
2996 exp = class_export_get(req->rq_export);
2998 /* disable reply scheduling while I'm setting up */
2999 rs->rs_scheduled = 1;
3001 rs->rs_xid = req->rq_xid;
3002 rs->rs_transno = req->rq_transno;
3003 rs->rs_export = exp;
3004 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3006 spin_lock(&exp->exp_uncommitted_replies_lock);
3007 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3008 rs->rs_transno, exp->exp_last_committed);
3009 if (rs->rs_transno > exp->exp_last_committed) {
3010 /* not committed already */
3011 list_add_tail(&rs->rs_obd_list,
3012 &exp->exp_uncommitted_replies);
3014 spin_unlock(&exp->exp_uncommitted_replies_lock);
3016 spin_lock(&exp->exp_lock);
3017 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3018 spin_unlock(&exp->exp_lock);
3020 netrc = target_send_reply_msg(req, rc, fail_id);
3022 spin_lock(&svcpt->scp_rep_lock);
3024 atomic_inc(&svcpt->scp_nreps_difficult);
3027 /* error sending: reply is off the net. Also we need +1
3028 * reply ref until ptlrpc_handle_rs() is done
3029 * with the reply state (if the send was successful, there
3030 * would have been +1 ref for the net, which
3031 * reply_out_callback leaves alone) */
3033 ptlrpc_rs_addref(rs);
3036 spin_lock(&rs->rs_lock);
3037 if (rs->rs_transno <= exp->exp_last_committed ||
3038 (!rs->rs_on_net && !rs->rs_no_ack) ||
3039 list_empty(&rs->rs_exp_list) || /* completed already */
3040 list_empty(&rs->rs_obd_list)) {
3041 CDEBUG(D_HA, "Schedule reply immediately\n");
3042 ptlrpc_dispatch_difficult_reply(rs);
3044 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3045 rs->rs_scheduled = 0; /* allow notifier to schedule */
3047 spin_unlock(&rs->rs_lock);
3048 spin_unlock(&svcpt->scp_rep_lock);
3052 enum ldlm_mode lck_compat_array[] = {
3053 [LCK_EX] = LCK_COMPAT_EX,
3054 [LCK_PW] = LCK_COMPAT_PW,
3055 [LCK_PR] = LCK_COMPAT_PR,
3056 [LCK_CW] = LCK_COMPAT_CW,
3057 [LCK_CR] = LCK_COMPAT_CR,
3058 [LCK_NL] = LCK_COMPAT_NL,
3059 [LCK_GROUP] = LCK_COMPAT_GROUP,
3060 [LCK_COS] = LCK_COMPAT_COS,
3064 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3065 * not escape to the user level.
3067 int ldlm_error2errno(enum ldlm_error error)
3073 case ELDLM_LOCK_MATCHED:
3076 case ELDLM_LOCK_CHANGED:
3079 case ELDLM_LOCK_ABORTED:
3082 case ELDLM_LOCK_REPLACED:
3085 case ELDLM_NO_LOCK_DATA:
3088 case ELDLM_NAMESPACE_EXISTS:
3091 case ELDLM_BAD_NAMESPACE:
3095 if (((int)error) < 0) { /* cast to signed type */
3096 result = error; /* as ldlm_error can be unsigned */
3098 CERROR("Invalid DLM result code: %d\n", error);
3104 EXPORT_SYMBOL(ldlm_error2errno);
3107 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3109 enum ldlm_error ldlm_errno2error(int err_no)
3118 error = ELDLM_LOCK_CHANGED;
3121 error = ELDLM_LOCK_ABORTED;
3124 error = ELDLM_LOCK_REPLACED;
3127 error = ELDLM_NO_LOCK_DATA;
3130 error = ELDLM_NAMESPACE_EXISTS;
3133 error = ELDLM_BAD_NAMESPACE;
3141 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3142 void ldlm_dump_export_locks(struct obd_export *exp)
3144 spin_lock(&exp->exp_locks_list_guard);
3145 if (!list_empty(&exp->exp_locks_list)) {
3146 struct ldlm_lock *lock;
3148 CERROR("dumping locks for export %p,"
3149 "ignore if the unmount doesn't hang\n", exp);
3150 list_for_each_entry(lock, &exp->exp_locks_list,
3152 LDLM_ERROR(lock, "lock:");
3154 spin_unlock(&exp->exp_locks_list_guard);
3158 #ifdef HAVE_SERVER_SUPPORT
3159 static int target_bulk_timeout(void *data)
3162 /* We don't fail the connection here, because having the export
3163 * killed makes the (vital) call to commitrw very sad.
3168 static inline const char *bulk2type(struct ptlrpc_request *req)
3170 if (req->rq_bulk_read)
3172 if (req->rq_bulk_write)
3177 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3178 struct l_wait_info *lwi)
3180 struct ptlrpc_request *req = desc->bd_req;
3181 time64_t start = ktime_get_real_seconds();
3187 /* If there is eviction in progress, wait for it to finish. */
3188 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3189 *lwi = LWI_INTR(NULL, NULL);
3190 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3191 !atomic_read(&exp->exp_obd->
3192 obd_evict_inprogress),
3196 /* Check if client was evicted or reconnected already. */
3197 if (exp->exp_failed ||
3198 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3201 if (req->rq_bulk_read)
3202 rc = sptlrpc_svc_wrap_bulk(req, desc);
3204 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3205 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3206 else /* old version, bulk matchbits is rq_xid */
3207 req->rq_mbits = req->rq_xid;
3210 rc = ptlrpc_start_bulk_transfer(desc);
3214 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3215 bulk2type(req), rc);
3219 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3220 ptlrpc_abort_bulk(desc);
3224 /* limit actual bulk transfer to bulk_timeout seconds */
3225 deadline = start + bulk_timeout;
3226 if (deadline > req->rq_deadline)
3227 deadline = req->rq_deadline;
3230 time64_t timeoutl = deadline - ktime_get_real_seconds();
3231 long timeout_jiffies = timeoutl <= 0 ?
3232 1 : cfs_time_seconds(timeoutl);
3233 time64_t rq_deadline;
3235 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3236 cfs_time_seconds(1),
3237 target_bulk_timeout, desc);
3238 rc = l_wait_event(desc->bd_waitq,
3239 !ptlrpc_server_bulk_active(desc) ||
3242 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3244 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3245 /* Wait again if we changed rq_deadline. */
3246 rq_deadline = READ_ONCE(req->rq_deadline);
3247 deadline = start + bulk_timeout;
3248 if (deadline > rq_deadline)
3249 deadline = rq_deadline;
3250 } while (rc == -ETIMEDOUT &&
3251 deadline > ktime_get_real_seconds());
3253 if (rc == -ETIMEDOUT) {
3254 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3255 bulk2type(req), deadline - start,
3256 ktime_get_real_seconds() - deadline);
3257 ptlrpc_abort_bulk(desc);
3258 } else if (exp->exp_failed) {
3259 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3262 ptlrpc_abort_bulk(desc);
3263 } else if (exp->exp_conn_cnt >
3264 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3265 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3267 /* We don't reply anyway. */
3269 ptlrpc_abort_bulk(desc);
3270 } else if (desc->bd_failure) {
3271 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3273 /* XXX should this be a different errno? */
3276 if (req->rq_bulk_write)
3277 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3278 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3279 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3280 bulk2type(req), desc->bd_nob_transferred,
3282 /* XXX should this be a different errno? */
3289 EXPORT_SYMBOL(target_bulk_io);
3291 #endif /* HAVE_SERVER_SUPPORT */