1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_LDLM
43 # include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_mds.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
54 /* @priority: if non-zero, move the selected to the list head
55 * @create: if zero, only search in existed connections
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
77 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79 GOTO(out_put, rc = -ENOMEM);
83 spin_lock(&imp->imp_lock);
84 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87 list_del(&item->oic_item);
88 list_add(&item->oic_item, &imp->imp_conn_list);
89 item->oic_last_attempt = 0;
91 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
92 imp, imp->imp_obd->obd_name, uuid->uuid,
93 (priority ? ", moved to head" : ""));
94 spin_unlock(&imp->imp_lock);
95 GOTO(out_free, rc = 0);
100 imp_conn->oic_conn = ptlrpc_conn;
101 imp_conn->oic_uuid = *uuid;
102 imp_conn->oic_last_attempt = 0;
104 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
106 list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
107 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
108 imp, imp->imp_obd->obd_name, uuid->uuid,
109 (priority ? "head" : "tail"));
111 spin_unlock(&imp->imp_lock);
112 GOTO(out_free, rc = -ENOENT);
115 spin_unlock(&imp->imp_lock);
119 OBD_FREE(imp_conn, sizeof(*imp_conn));
121 ptlrpc_connection_put(ptlrpc_conn);
125 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
127 return import_set_conn(imp, uuid, 1, 0);
130 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
133 return import_set_conn(imp, uuid, priority, 1);
136 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
138 struct obd_import_conn *imp_conn;
139 struct obd_export *dlmexp;
143 spin_lock(&imp->imp_lock);
144 if (list_empty(&imp->imp_conn_list)) {
145 LASSERT(!imp->imp_connection);
149 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
150 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
152 LASSERT(imp_conn->oic_conn);
154 /* is current conn? */
155 if (imp_conn == imp->imp_conn_current) {
156 LASSERT(imp_conn->oic_conn == imp->imp_connection);
158 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
159 imp->imp_state != LUSTRE_IMP_DISCON) {
160 CERROR("can't remove current connection\n");
161 GOTO(out, rc = -EBUSY);
164 ptlrpc_connection_put(imp->imp_connection);
165 imp->imp_connection = NULL;
167 dlmexp = class_conn2export(&imp->imp_dlm_handle);
168 if (dlmexp && dlmexp->exp_connection) {
169 LASSERT(dlmexp->exp_connection ==
171 ptlrpc_connection_put(dlmexp->exp_connection);
172 dlmexp->exp_connection = NULL;
176 list_del(&imp_conn->oic_item);
177 ptlrpc_connection_put(imp_conn->oic_conn);
178 OBD_FREE(imp_conn, sizeof(*imp_conn));
179 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
180 imp, imp->imp_obd->obd_name, uuid->uuid);
185 spin_unlock(&imp->imp_lock);
187 CERROR("connection %s not found\n", uuid->uuid);
191 void client_destroy_import(struct obd_import *imp)
193 /* drop security policy instance after all rpc finished/aborted
194 * to let all busy contexts be released. */
195 class_import_get(imp);
196 class_destroy_import(imp);
197 sptlrpc_import_sec_put(imp);
198 class_import_put(imp);
201 /* configure an RPC client OBD device
206 * 3 - inactive-on-startup
208 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
210 struct client_obd *cli = &obddev->u.cli;
211 struct obd_import *imp;
212 struct obd_uuid server_uuid;
213 int rq_portal, rp_portal, connect_op;
214 char *name = obddev->obd_type->typ_name;
218 /* In a more perfect world, we would hang a ptlrpc_client off of
219 * obd_type and just use the values from there. */
220 if (!strcmp(name, LUSTRE_OSC_NAME)) {
221 rq_portal = OST_REQUEST_PORTAL;
222 rp_portal = OSC_REPLY_PORTAL;
223 connect_op = OST_CONNECT;
224 cli->cl_sp_me = LUSTRE_SP_CLI;
225 cli->cl_sp_to = LUSTRE_SP_OST;
226 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
227 rq_portal = MDS_REQUEST_PORTAL;
228 rp_portal = MDC_REPLY_PORTAL;
229 connect_op = MDS_CONNECT;
230 cli->cl_sp_me = LUSTRE_SP_CLI;
231 cli->cl_sp_to = LUSTRE_SP_MDT;
232 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
233 rq_portal = MGS_REQUEST_PORTAL;
234 rp_portal = MGC_REPLY_PORTAL;
235 connect_op = MGS_CONNECT;
236 cli->cl_sp_me = LUSTRE_SP_MGC;
237 cli->cl_sp_to = LUSTRE_SP_MGS;
238 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
240 CERROR("unknown client OBD type \"%s\", can't setup\n",
245 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
246 CERROR("requires a TARGET UUID\n");
250 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
251 CERROR("client UUID must be less than 38 characters\n");
255 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
256 CERROR("setup requires a SERVER UUID\n");
260 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
261 CERROR("target UUID must be less than 38 characters\n");
265 init_rwsem(&cli->cl_sem);
266 sema_init(&cli->cl_mgc_sem, 1);
267 cli->cl_conn_count = 0;
268 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
269 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
270 sizeof(server_uuid)));
273 cli->cl_avail_grant = 0;
274 /* FIXME: should limit this for the sum of all cl_dirty_max */
275 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
276 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
277 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
278 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
279 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
280 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
281 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
282 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
283 client_obd_list_lock_init(&cli->cl_loi_list_lock);
284 cli->cl_r_in_flight = 0;
285 cli->cl_w_in_flight = 0;
287 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
288 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
289 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
290 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
291 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
292 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
293 cfs_waitq_init(&cli->cl_destroy_waitq);
294 atomic_set(&cli->cl_destroy_in_flight, 0);
295 #ifdef ENABLE_CHECKSUM
296 /* Turn on checksumming by default. */
297 cli->cl_checksum = 1;
299 * The supported checksum types will be worked out at connect time
300 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
303 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
305 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
306 atomic_set(&cli->cl_quota_resends, CLIENT_QUOTA_DEFAULT_RESENDS);
308 /* This value may be changed at connect time in
309 ptlrpc_connect_interpret. */
310 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
311 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
313 if (!strcmp(name, LUSTRE_MDC_NAME)) {
314 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
315 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
316 cli->cl_max_rpcs_in_flight = 2;
317 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
318 cli->cl_max_rpcs_in_flight = 3;
319 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
320 cli->cl_max_rpcs_in_flight = 4;
322 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
327 CERROR("ldlm_get_ref failed: %d\n", rc);
331 ptlrpc_init_client(rq_portal, rp_portal, name,
332 &obddev->obd_ldlm_client);
334 imp = class_new_import(obddev);
336 GOTO(err_ldlm, rc = -ENOENT);
337 imp->imp_client = &obddev->obd_ldlm_client;
338 imp->imp_connect_op = connect_op;
339 imp->imp_initial_recov = 1;
340 imp->imp_initial_recov_bk = 0;
341 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
342 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
343 LUSTRE_CFG_BUFLEN(lcfg, 1));
344 class_import_put(imp);
346 rc = client_import_add_conn(imp, &server_uuid, 1);
348 CERROR("can't add initial connection\n");
349 GOTO(err_import, rc);
352 cli->cl_import = imp;
353 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
354 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
355 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
357 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
358 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
359 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
360 name, obddev->obd_name,
361 cli->cl_target_uuid.uuid);
362 spin_lock(&imp->imp_lock);
363 imp->imp_deactive = 1;
364 spin_unlock(&imp->imp_lock);
368 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
369 LDLM_NAMESPACE_CLIENT,
370 LDLM_NAMESPACE_GREEDY);
371 if (obddev->obd_namespace == NULL) {
372 CERROR("Unable to create client namespace - %s\n",
374 GOTO(err_import, rc = -ENOMEM);
377 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
382 class_destroy_import(imp);
390 int client_obd_cleanup(struct obd_device *obddev)
394 ldlm_namespace_free_post(obddev->obd_namespace);
395 obddev->obd_namespace = NULL;
401 /* ->o_connect() method for client side (OSC and MDC and MGC) */
402 int client_connect_import(const struct lu_env *env,
403 struct obd_export **exp,
404 struct obd_device *obd, struct obd_uuid *cluuid,
405 struct obd_connect_data *data, void *localdata)
407 struct client_obd *cli = &obd->u.cli;
408 struct obd_import *imp = cli->cl_import;
409 struct obd_connect_data *ocd;
410 struct lustre_handle conn = { 0 };
415 down_write(&cli->cl_sem);
416 if (cli->cl_conn_count > 0 )
417 GOTO(out_sem, rc = -EALREADY);
419 rc = class_connect(&conn, obd, cluuid);
423 cli->cl_conn_count++;
424 *exp = class_conn2export(&conn);
426 LASSERT(obd->obd_namespace);
428 imp->imp_dlm_handle = conn;
429 rc = ptlrpc_init_import(imp);
433 ocd = &imp->imp_connect_data;
436 imp->imp_connect_flags_orig = data->ocd_connect_flags;
439 rc = ptlrpc_connect_import(imp, NULL);
441 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
444 LASSERT((*exp)->exp_connection);
447 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
448 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
449 data->ocd_connect_flags, ocd->ocd_connect_flags);
450 data->ocd_connect_flags = ocd->ocd_connect_flags;
453 ptlrpc_pinger_add_import(imp);
459 cli->cl_conn_count--;
460 class_disconnect(*exp);
464 up_write(&cli->cl_sem);
469 int client_disconnect_export(struct obd_export *exp)
471 struct obd_device *obd = class_exp2obd(exp);
472 struct client_obd *cli;
473 struct obd_import *imp;
478 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
479 exp, exp ? exp->exp_handle.h_cookie : -1);
484 imp = cli->cl_import;
486 down_write(&cli->cl_sem);
487 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
490 if (!cli->cl_conn_count) {
491 CERROR("disconnecting disconnected device (%s)\n",
493 GOTO(out_disconnect, rc = -EINVAL);
496 cli->cl_conn_count--;
497 if (cli->cl_conn_count)
498 GOTO(out_disconnect, rc = 0);
500 /* Mark import deactivated now, so we don't try to reconnect if any
501 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
502 * fully deactivate the import, or that would drop all requests. */
503 spin_lock(&imp->imp_lock);
504 imp->imp_deactive = 1;
505 spin_unlock(&imp->imp_lock);
507 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
508 * delete it regardless. (It's safe to delete an import that was
510 (void)ptlrpc_pinger_del_import(imp);
512 if (obd->obd_namespace != NULL) {
513 /* obd_force == local only */
514 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
515 obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
517 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
521 * there's no need to hold sem during disconnecting an import,
522 * and actually it may cause deadlock in gss.
524 up_write(&cli->cl_sem);
525 rc = ptlrpc_disconnect_import(imp, 0);
526 down_write(&cli->cl_sem);
528 ptlrpc_invalidate_import(imp);
530 if (imp->imp_rq_pool) {
531 ptlrpc_free_rq_pool(imp->imp_rq_pool);
532 imp->imp_rq_pool = NULL;
534 client_destroy_import(imp);
535 cli->cl_import = NULL;
540 /* use server style - class_disconnect should be always called for
542 err = class_disconnect(exp);
546 up_write(&cli->cl_sem);
551 int server_disconnect_export(struct obd_export *exp)
556 /* Disconnect early so that clients can't keep using export */
557 rc = class_disconnect(exp);
558 /* close import for avoid sending any requests */
559 if (exp->exp_imp_reverse)
560 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
562 if (exp->exp_obd->obd_namespace != NULL)
563 ldlm_cancel_locks_for_export(exp);
565 /* complete all outstanding replies */
566 spin_lock(&exp->exp_lock);
567 while (!list_empty(&exp->exp_outstanding_replies)) {
568 struct ptlrpc_reply_state *rs =
569 list_entry(exp->exp_outstanding_replies.next,
570 struct ptlrpc_reply_state, rs_exp_list);
571 struct ptlrpc_service *svc = rs->rs_service;
573 spin_lock(&svc->srv_lock);
574 list_del_init(&rs->rs_exp_list);
575 spin_lock(&rs->rs_lock);
576 ptlrpc_schedule_difficult_reply(rs);
577 spin_unlock(&rs->rs_lock);
578 spin_unlock(&svc->srv_lock);
580 spin_unlock(&exp->exp_lock);
582 /* release nid stat refererence */
583 lprocfs_exp_cleanup(exp);
588 /* --------------------------------------------------------------------------
589 * from old lib/target.c
590 * -------------------------------------------------------------------------- */
592 static int target_handle_reconnect(struct lustre_handle *conn,
593 struct obd_export *exp,
594 struct obd_uuid *cluuid)
597 if (exp->exp_connection && exp->exp_imp_reverse) {
598 struct lustre_handle *hdl;
599 hdl = &exp->exp_imp_reverse->imp_remote_handle;
600 /* Might be a re-connect after a partition. */
601 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
602 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
604 conn->cookie = exp->exp_handle.h_cookie;
605 /* target_handle_connect() treats EALREADY and
606 * -EALREADY differently. EALREADY means we are
607 * doing a valid reconnect from the same client. */
610 CERROR("%s reconnecting from %s, "
611 "handle mismatch (ours "LPX64", theirs "
612 LPX64")\n", cluuid->uuid,
613 exp->exp_connection->c_remote_uuid.uuid,
614 hdl->cookie, conn->cookie);
615 memset(conn, 0, sizeof *conn);
616 /* target_handle_connect() treats EALREADY and
617 * -EALREADY differently. -EALREADY is an error
618 * (same UUID, different handle). */
623 conn->cookie = exp->exp_handle.h_cookie;
624 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
625 cluuid->uuid, exp, conn->cookie);
629 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
632 struct obd_export *exp = cb_data;
634 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
635 obd->obd_name, exp->exp_client_uuid.uuid);
637 spin_lock(&exp->exp_lock);
638 exp->exp_need_sync = 0;
639 spin_unlock(&exp->exp_lock);
640 class_export_cb_put(exp);
642 EXPORT_SYMBOL(target_client_add_cb);
645 target_start_and_reset_recovery_timer(struct obd_device *obd,
646 struct ptlrpc_request *req,
649 int target_handle_connect(struct ptlrpc_request *req)
651 struct obd_device *target, *targref = NULL;
652 struct obd_export *export = NULL;
653 struct obd_import *revimp;
654 struct lustre_handle conn;
655 struct lustre_handle *tmp;
656 struct obd_uuid tgtuuid;
657 struct obd_uuid cluuid;
658 struct obd_uuid remote_uuid;
662 struct obd_connect_data *data, *tmpdata;
663 lnet_nid_t *client_nid = NULL;
666 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
668 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
670 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
671 GOTO(out, rc = -EINVAL);
674 obd_str2uuid(&tgtuuid, str);
675 target = class_uuid2obd(&tgtuuid);
677 target = class_name2obd(str);
679 if (!target || target->obd_stopping || !target->obd_set_up) {
680 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
681 " for connect (%s)\n", str,
682 !target ? "no target" :
683 (target->obd_stopping ? "stopping" :
685 GOTO(out, rc = -ENODEV);
688 if (target->obd_no_conn) {
689 LCONSOLE_WARN("%s: temporarily refusing client connection "
690 "from %s\n", target->obd_name,
691 libcfs_nid2str(req->rq_peer.nid));
692 GOTO(out, rc = -EAGAIN);
695 /* Make sure the target isn't cleaned up while we're here. Yes,
696 there's still a race between the above check and our incref here.
697 Really, class_uuid2obd should take the ref. */
698 targref = class_incref(target, __FUNCTION__, cfs_current());
701 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
703 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
704 GOTO(out, rc = -EINVAL);
707 obd_str2uuid(&cluuid, str);
709 /* XXX extract a nettype and format accordingly */
710 switch (sizeof(lnet_nid_t)) {
711 /* NB the casts only avoid compiler warnings */
713 snprintf(remote_uuid.uuid, sizeof remote_uuid,
714 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
717 snprintf(remote_uuid.uuid, sizeof remote_uuid,
718 "NET_%x_UUID", (__u32)req->rq_peer.nid);
724 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
726 GOTO(out, rc = -EPROTO);
730 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
732 GOTO(out, rc = -EPROTO);
734 rc = req_capsule_server_pack(&req->rq_pill);
738 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
740 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
741 "libclient connection attempt");
742 GOTO(out, rc = -EPROTO);
743 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
744 LUSTRE_VERSION_ALLOWED_OFFSET ||
745 data->ocd_version > LUSTRE_VERSION_CODE +
746 LUSTRE_VERSION_ALLOWED_OFFSET) {
747 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
748 "libclient connection attempt",
749 data->ocd_version < LUSTRE_VERSION_CODE ?
751 OBD_OCD_VERSION_MAJOR(data->ocd_version),
752 OBD_OCD_VERSION_MINOR(data->ocd_version),
753 OBD_OCD_VERSION_PATCH(data->ocd_version),
754 OBD_OCD_VERSION_FIX(data->ocd_version));
755 data = req_capsule_server_sized_get(&req->rq_pill,
757 offsetof(typeof(*data), ocd_version) +
758 sizeof(data->ocd_version));
760 data->ocd_connect_flags = OBD_CONNECT_VERSION;
761 data->ocd_version = LUSTRE_VERSION_CODE;
763 GOTO(out, rc = -EPROTO);
767 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
768 (data->ocd_connect_flags & OBD_CONNECT_MDS))
771 /* lctl gets a backstage, all-access pass. */
772 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
773 goto dont_check_exports;
775 export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
779 /* we've found an export in the hash */
780 if (export->exp_connecting) { /* bug 9635, et. al. */
781 CWARN("%s: exp %p already connecting\n",
782 export->exp_obd->obd_name, export);
783 class_export_put(export);
786 } else if (mds_conn && export->exp_connection) {
787 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
788 /* mds reconnected after failover */
789 CWARN("%s: received MDS connection from NID %s,"
790 " removing former export from NID %s\n",
791 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
792 libcfs_nid2str(export->exp_connection->c_peer.nid));
794 /* new mds connection from the same nid */
795 CWARN("%s: received new MDS connection from NID %s,"
796 " removing former export from same NID\n",
797 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
798 class_fail_export(export);
799 class_export_put(export);
802 } else if (export->exp_connection != NULL &&
803 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
804 (lustre_msg_get_op_flags(req->rq_reqmsg) &
805 MSG_CONNECT_INITIAL)) {
806 /* in mds failover we have static uuid but nid can be
808 CWARN("%s: cookie %s seen on new NID %s when "
809 "existing NID %s is already connected\n",
810 target->obd_name, cluuid.uuid,
811 libcfs_nid2str(req->rq_peer.nid),
812 libcfs_nid2str(export->exp_connection->c_peer.nid));
814 class_export_put(export);
817 spin_lock(&export->exp_lock);
818 export->exp_connecting = 1;
819 spin_unlock(&export->exp_lock);
820 class_export_put(export);
821 LASSERT(export->exp_obd == target);
823 rc = target_handle_reconnect(&conn, export, &cluuid);
826 /* If we found an export, we already unlocked. */
829 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
830 } else if (req->rq_export == NULL &&
831 atomic_read(&export->exp_rpc_count) > 0) {
832 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
833 target->obd_name, cluuid.uuid,
834 libcfs_nid2str(req->rq_peer.nid),
835 export, atomic_read(&export->exp_refcount));
836 GOTO(out, rc = -EBUSY);
837 } else if (req->rq_export != NULL &&
838 (atomic_read(&export->exp_rpc_count) > 1)) {
839 /* the current connect rpc has increased exp_rpc_count */
840 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
841 target->obd_name, cluuid.uuid,
842 libcfs_nid2str(req->rq_peer.nid),
843 export, atomic_read(&export->exp_rpc_count) - 1);
844 spin_lock(&export->exp_lock);
845 if (req->rq_export->exp_conn_cnt <
846 lustre_msg_get_conn_cnt(req->rq_reqmsg))
847 /* try to abort active requests */
848 req->rq_export->exp_abort_active_req = 1;
849 spin_unlock(&export->exp_lock);
850 GOTO(out, rc = -EBUSY);
851 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
852 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
853 "cookies not random?\n", target->obd_name,
854 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
855 GOTO(out, rc = -EALREADY);
857 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
864 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
865 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
866 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
867 export, (long)cfs_time_current_sec(),
868 export ? (long)export->exp_last_request_time : 0);
870 /* Tell the client if we're in recovery. */
871 if (target->obd_recovering) {
872 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
873 /* If this is the first time a client connects,
874 reset the recovery timer */
876 target_start_and_reset_recovery_timer(target, req,
880 /* We want to handle EALREADY but *not* -EALREADY from
881 * target_handle_reconnect(), return reconnection state in a flag */
882 if (rc == EALREADY) {
883 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
889 /* Tell the client if we support replayable requests */
890 if (target->obd_replayable)
891 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
892 client_nid = &req->rq_peer.nid;
894 if (export == NULL) {
895 if (target->obd_recovering) {
898 t = cfs_timer_deadline(&target->obd_recovery_timer);
899 t = cfs_time_sub(t, cfs_time_current());
900 CERROR("%s: denying connection for new client %s (%s): "
901 "%d clients in recovery for "CFS_TIME_T"s\n",
903 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
904 atomic_read(&target->obd_lock_replay_clients),
905 cfs_duration_sec(t));
909 rc = obd_connect(req->rq_svc_thread->t_env,
910 &export, target, &cluuid, data,
913 conn.cookie = export->exp_handle.h_cookie;
916 rc = obd_reconnect(req->rq_svc_thread->t_env,
917 export, target, &cluuid, data, client_nid);
919 /* prevous done via class_conn2export */
920 class_export_get(export);
924 /* Return only the parts of obd_connect_data that we understand, so the
925 * client knows that we don't understand the rest. */
927 tmpdata = req_capsule_server_get(&req->rq_pill,
929 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
933 /* If all else goes well, this is our RPC return code. */
936 lustre_msg_set_handle(req->rq_repmsg, &conn);
938 /* If the client and the server are the same node, we will already
939 * have an export that really points to the client's DLM export,
940 * because we have a shared handles table.
942 * XXX this will go away when shaver stops sending the "connect" handle
943 * in the real "remote handle" field of the request --phik 24 Apr 2003
945 if (req->rq_export != NULL)
946 class_export_put(req->rq_export);
948 req->rq_export = export;
950 spin_lock(&export->exp_lock);
951 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
952 spin_unlock(&export->exp_lock);
953 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
954 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
955 export->exp_conn_cnt,
956 lustre_msg_get_conn_cnt(req->rq_reqmsg));
958 GOTO(out, rc = -EALREADY);
960 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
961 export->exp_abort_active_req = 0;
963 /* request from liblustre? Don't evict it for not pinging. */
964 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
965 export->exp_libclient = 1;
966 spin_unlock(&export->exp_lock);
968 spin_lock(&target->obd_dev_lock);
969 list_del_init(&export->exp_obd_chain_timed);
970 spin_unlock(&target->obd_dev_lock);
972 spin_unlock(&export->exp_lock);
975 if (export->exp_connection != NULL) {
976 /* Check to see if connection came from another NID */
977 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
978 !hlist_unhashed(&export->exp_nid_hash))
979 lustre_hash_del(export->exp_obd->obd_nid_hash,
980 &export->exp_connection->c_peer.nid,
981 &export->exp_nid_hash);
983 ptlrpc_connection_put(export->exp_connection);
986 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
989 if (hlist_unhashed(&export->exp_nid_hash)) {
990 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
991 &export->exp_connection->c_peer.nid,
992 &export->exp_nid_hash);
995 spin_lock_bh(&target->obd_processing_task_lock);
996 if (target->obd_recovering && !export->exp_in_recovery) {
997 spin_lock(&export->exp_lock);
998 export->exp_in_recovery = 1;
999 export->exp_req_replay_needed = 1;
1000 export->exp_lock_replay_needed = 1;
1001 spin_unlock(&export->exp_lock);
1002 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1003 && (data->ocd_transno == 0))
1004 CWARN("Connect with zero transno!\n");
1006 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1007 && data->ocd_transno < target->obd_next_recovery_transno)
1008 target->obd_next_recovery_transno = data->ocd_transno;
1009 target->obd_connected_clients++;
1010 atomic_inc(&target->obd_req_replay_clients);
1011 atomic_inc(&target->obd_lock_replay_clients);
1012 if (target->obd_connected_clients ==
1013 target->obd_max_recoverable_clients)
1014 cfs_waitq_signal(&target->obd_next_transno_waitq);
1016 spin_unlock_bh(&target->obd_processing_task_lock);
1017 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1020 if (export->exp_imp_reverse != NULL) {
1021 /* destroyed import can be still referenced in ctxt */
1022 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1023 KEY_REVIMP_UPD, 0, NULL, NULL);
1025 client_destroy_import(export->exp_imp_reverse);
1028 /* for the rest part, we return -ENOTCONN in case of errors
1029 * in order to let client initialize connection again.
1031 revimp = export->exp_imp_reverse = class_new_import(target);
1033 CERROR("fail to alloc new reverse import.\n");
1034 GOTO(out, rc = -ENOTCONN);
1037 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1038 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1039 revimp->imp_remote_handle = conn;
1040 revimp->imp_dlm_fake = 1;
1041 revimp->imp_state = LUSTRE_IMP_FULL;
1043 /* unknown versions will be caught in
1044 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1045 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1047 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1048 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1049 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1051 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1053 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1055 CERROR("Failed to get sec for reverse import: %d\n", rc);
1056 export->exp_imp_reverse = NULL;
1057 class_destroy_import(revimp);
1060 class_import_put(revimp);
1063 spin_lock(&export->exp_lock);
1064 export->exp_connecting = 0;
1065 spin_unlock(&export->exp_lock);
1068 class_decref(targref, __FUNCTION__, cfs_current());
1070 req->rq_status = rc;
1074 int target_handle_disconnect(struct ptlrpc_request *req)
1079 rc = req_capsule_server_pack(&req->rq_pill);
1083 /* keep the rq_export around so we can send the reply */
1084 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1089 void target_destroy_export(struct obd_export *exp)
1091 /* exports created from last_rcvd data, and "fake"
1092 exports created by lctl don't have an import */
1093 if (exp->exp_imp_reverse != NULL)
1094 client_destroy_import(exp->exp_imp_reverse);
1096 LASSERT(atomic_read(&exp->exp_locks_count) == 0);
1097 LASSERT(atomic_read(&exp->exp_rpc_count) == 0);
1098 LASSERT(atomic_read(&exp->exp_cb_count) == 0);
1099 LASSERT(atomic_read(&exp->exp_replay_count) == 0);
1103 * Recovery functions
1105 static void target_request_copy_get(struct ptlrpc_request *req)
1107 class_export_rpc_get(req->rq_export);
1108 LASSERT(list_empty(&req->rq_list));
1109 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1110 /* increase refcount to keep request in queue */
1111 LASSERT(atomic_read(&req->rq_refcount));
1112 atomic_inc(&req->rq_refcount);
1113 /** let export know it has replays to be handled */
1114 atomic_inc(&req->rq_export->exp_replay_count);
1117 static void target_request_copy_put(struct ptlrpc_request *req)
1119 LASSERT(list_empty(&req->rq_replay_list));
1120 LASSERT(atomic_read(&req->rq_export->exp_replay_count) > 0);
1121 atomic_dec(&req->rq_export->exp_replay_count);
1122 class_export_rpc_put(req->rq_export);
1123 ptlrpc_server_drop_request(req);
1126 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1128 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1129 struct obd_export *exp = req->rq_export;
1130 struct ptlrpc_request *reqiter;
1135 spin_lock(&exp->exp_lock);
1136 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1138 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1145 /* we expect it with RESENT and REPLAY flags */
1146 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1147 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1148 CERROR("invalid flags %x of resent replay\n",
1149 lustre_msg_get_flags(req->rq_reqmsg));
1151 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1154 spin_unlock(&exp->exp_lock);
1158 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1160 LASSERT(!list_empty(&req->rq_replay_list));
1161 LASSERT(req->rq_export);
1163 spin_lock(&req->rq_export->exp_lock);
1164 list_del_init(&req->rq_replay_list);
1165 spin_unlock(&req->rq_export->exp_lock);
1169 static void target_finish_recovery(struct obd_device *obd)
1172 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1175 ldlm_reprocess_all_ns(obd->obd_namespace);
1176 spin_lock_bh(&obd->obd_processing_task_lock);
1177 if (!list_empty(&obd->obd_req_replay_queue) ||
1178 !list_empty(&obd->obd_lock_replay_queue) ||
1179 !list_empty(&obd->obd_final_req_queue)) {
1180 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1182 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1183 list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1184 list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1185 spin_unlock_bh(&obd->obd_processing_task_lock);
1188 spin_unlock_bh(&obd->obd_processing_task_lock);
1190 obd->obd_recovery_end = cfs_time_current_sec();
1192 /* when recovery finished, cleanup orphans on mds and ost */
1193 if (OBT(obd) && OBP(obd, postrecov)) {
1194 int rc = OBP(obd, postrecov)(obd);
1196 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1202 static void abort_req_replay_queue(struct obd_device *obd)
1204 struct ptlrpc_request *req, *n;
1205 struct list_head abort_list;
1207 CFS_INIT_LIST_HEAD(&abort_list);
1208 spin_lock_bh(&obd->obd_processing_task_lock);
1209 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1210 spin_unlock_bh(&obd->obd_processing_task_lock);
1211 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1212 DEBUG_REQ(D_WARNING, req, "aborted:");
1213 req->rq_status = -ENOTCONN;
1214 if (ptlrpc_error(req)) {
1215 DEBUG_REQ(D_ERROR, req,
1216 "failed abort_req_reply; skipping");
1218 target_exp_dequeue_req_replay(req);
1219 target_request_copy_put(req);
1223 static void abort_lock_replay_queue(struct obd_device *obd)
1225 struct ptlrpc_request *req, *n;
1226 struct list_head abort_list;
1228 CFS_INIT_LIST_HEAD(&abort_list);
1229 spin_lock_bh(&obd->obd_processing_task_lock);
1230 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1231 spin_unlock_bh(&obd->obd_processing_task_lock);
1232 list_for_each_entry_safe(req, n, &abort_list, rq_list){
1233 DEBUG_REQ(D_ERROR, req, "aborted:");
1234 req->rq_status = -ENOTCONN;
1235 if (ptlrpc_error(req)) {
1236 DEBUG_REQ(D_ERROR, req,
1237 "failed abort_lock_reply; skipping");
1239 target_request_copy_put(req);
1244 /* Called from a cleanup function if the device is being cleaned up
1245 forcefully. The exports should all have been disconnected already,
1246 the only thing left to do is
1247 - clear the recovery flags
1249 - free queued requests and replies, but don't send replies
1250 Because the obd_stopping flag is set, no new requests should be received.
1253 void target_cleanup_recovery(struct obd_device *obd)
1255 struct ptlrpc_request *req, *n;
1256 struct list_head clean_list;
1259 CFS_INIT_LIST_HEAD(&clean_list);
1260 spin_lock_bh(&obd->obd_processing_task_lock);
1261 if (!obd->obd_recovering) {
1262 spin_unlock_bh(&obd->obd_processing_task_lock);
1266 obd->obd_recovering = obd->obd_abort_recovery = 0;
1267 target_cancel_recovery_timer(obd);
1269 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1270 spin_unlock_bh(&obd->obd_processing_task_lock);
1272 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1273 LASSERT(req->rq_reply_state == 0);
1274 target_exp_dequeue_req_replay(req);
1275 target_request_copy_put(req);
1278 spin_lock_bh(&obd->obd_processing_task_lock);
1279 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1280 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1281 spin_unlock_bh(&obd->obd_processing_task_lock);
1283 list_for_each_entry_safe(req, n, &clean_list, rq_list){
1284 LASSERT(req->rq_reply_state == 0);
1285 target_request_copy_put(req);
1291 /* obd_processing_task_lock should be held */
1292 void target_cancel_recovery_timer(struct obd_device *obd)
1294 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1295 cfs_timer_disarm(&obd->obd_recovery_timer);
1298 /* extend = 1 means require at least "duration" seconds left in the timer,
1299 extend = 0 means set the total duration (start_recovery_timer) */
1300 static void reset_recovery_timer(struct obd_device *obd, int duration,
1303 cfs_time_t now = cfs_time_current_sec();
1304 cfs_duration_t left;
1306 spin_lock_bh(&obd->obd_processing_task_lock);
1307 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1308 spin_unlock_bh(&obd->obd_processing_task_lock);
1312 left = cfs_time_sub(obd->obd_recovery_end, now);
1314 if (extend && (duration > left))
1315 obd->obd_recovery_timeout += duration - left;
1316 else if (!extend && (duration > obd->obd_recovery_timeout))
1317 /* Track the client's largest expected replay time */
1318 obd->obd_recovery_timeout = duration;
1321 * If total recovery time already exceed the
1322 * obd_recovery_max_time, then CRAY XT3 will
1323 * abort the recovery
1325 if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1326 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1328 obd->obd_recovery_end = obd->obd_recovery_start +
1329 obd->obd_recovery_timeout;
1330 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1331 cfs_time_before(now, obd->obd_recovery_end)) {
1332 left = cfs_time_sub(obd->obd_recovery_end, now);
1333 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1335 spin_unlock_bh(&obd->obd_processing_task_lock);
1336 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1337 obd->obd_name, (unsigned)left);
1340 static void check_and_start_recovery_timer(struct obd_device *obd)
1342 spin_lock_bh(&obd->obd_processing_task_lock);
1343 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1344 spin_unlock_bh(&obd->obd_processing_task_lock);
1347 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1348 obd->obd_recovery_start = cfs_time_current_sec();
1350 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1351 spin_unlock_bh(&obd->obd_processing_task_lock);
1353 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1356 /* Reset the timer with each new client connection */
1358 * This timer is actually reconnect_timer, which is for making sure
1359 * the total recovery window is at least as big as my reconnect
1360 * attempt timing. So the initial recovery time_out will be set to
1361 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1362 * from client is bigger than this, then the recovery time_out will
1363 * be extended to make sure the client could be reconnected, in the
1364 * process, the timeout from the new client should be ignored.
1368 target_start_and_reset_recovery_timer(struct obd_device *obd,
1369 struct ptlrpc_request *req,
1372 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1374 if (!new_client && service_time)
1375 /* Teach server about old server's estimates, as first guess
1376 * at how long new requests will take. */
1377 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1380 check_and_start_recovery_timer(obd);
1382 /* convert the service time to rpc timeout,
1383 * reuse service_time to limit stack usage */
1384 service_time = at_est2timeout(service_time);
1386 /* We expect other clients to timeout within service_time, then try
1387 * to reconnect, then try the failover server. The max delay between
1388 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1389 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1390 INITIAL_CONNECT_TIMEOUT);
1391 if (service_time > obd->obd_recovery_timeout && !new_client)
1392 reset_recovery_timer(obd, service_time, 0);
1397 /** Health checking routines */
1398 static inline int exp_connect_healthy(struct obd_export *exp)
1400 return (exp->exp_in_recovery);
1403 /** if export done req_replay or has replay in queue */
1404 static inline int exp_req_replay_healthy(struct obd_export *exp)
1406 return (!exp->exp_req_replay_needed ||
1407 atomic_read(&exp->exp_replay_count) > 0);
1409 /** if export done lock_replay or has replay in queue */
1410 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1412 return (!exp->exp_lock_replay_needed ||
1413 atomic_read(&exp->exp_replay_count) > 0);
1416 static inline int exp_vbr_healthy(struct obd_export *exp)
1418 return (!exp->exp_vbr_failed);
1421 static inline int exp_finished(struct obd_export *exp)
1423 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1426 /** Checking routines for recovery */
1427 static int check_for_clients(struct obd_device *obd)
1429 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1431 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1432 if (obd->obd_no_conn == 0 &&
1433 obd->obd_connected_clients + obd->obd_stale_clients ==
1434 obd->obd_max_recoverable_clients)
1439 static int check_for_next_transno(struct obd_device *obd)
1441 struct ptlrpc_request *req = NULL;
1442 int wake_up = 0, connected, completed, queue_len;
1443 __u64 next_transno, req_transno;
1445 spin_lock_bh(&obd->obd_processing_task_lock);
1447 if (!list_empty(&obd->obd_req_replay_queue)) {
1448 req = list_entry(obd->obd_req_replay_queue.next,
1449 struct ptlrpc_request, rq_list);
1450 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1455 connected = obd->obd_connected_clients;
1456 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1457 queue_len = obd->obd_requests_queued_for_recovery;
1458 next_transno = obd->obd_next_recovery_transno;
1460 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1461 "req_transno: "LPU64", next_transno: "LPU64"\n",
1462 obd->obd_max_recoverable_clients, connected, completed,
1463 queue_len, req_transno, next_transno);
1465 if (obd->obd_abort_recovery) {
1466 CDEBUG(D_HA, "waking for aborted recovery\n");
1468 } else if (obd->obd_recovery_expired) {
1469 CDEBUG(D_HA, "waking for expired recovery\n");
1471 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1472 CDEBUG(D_HA, "waking for completed recovery\n");
1474 } else if (req_transno == next_transno) {
1475 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1477 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1479 /** handle gaps occured due to lost reply or VBR */
1480 LASSERTF(req_transno >= next_transno,
1481 "req_transno: "LPU64", next_transno: "LPU64"\n",
1482 req_transno, next_transno);
1483 if (req_transno > obd->obd_last_committed &&
1484 !obd->obd_version_recov)
1487 "%s: waking for gap in transno, VBR is %s (skip: "
1488 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1489 ", last_committed: "LPD64")\n",
1490 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1491 next_transno, queue_len, completed, connected,
1492 req_transno, obd->obd_last_committed);
1493 obd->obd_next_recovery_transno = req_transno;
1495 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1496 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1497 " by fail_lock, waking up ("LPD64")\n", next_transno);
1498 obd->obd_next_recovery_transno = req_transno;
1501 spin_unlock_bh(&obd->obd_processing_task_lock);
1505 static int check_for_next_lock(struct obd_device *obd)
1509 spin_lock_bh(&obd->obd_processing_task_lock);
1510 if (!list_empty(&obd->obd_lock_replay_queue)) {
1511 CDEBUG(D_HA, "waking for next lock\n");
1513 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1514 CDEBUG(D_HA, "waking for completed lock replay\n");
1516 } else if (obd->obd_abort_recovery) {
1517 CDEBUG(D_HA, "waking for aborted recovery\n");
1519 } else if (obd->obd_recovery_expired) {
1520 CDEBUG(D_HA, "waking for expired recovery\n");
1523 spin_unlock_bh(&obd->obd_processing_task_lock);
1529 * wait for recovery events,
1530 * check its status with help of check_routine
1531 * evict dead clients via health_check
1533 static int target_recovery_overseer(struct obd_device *obd,
1534 int (*check_routine)(struct obd_device *),
1535 int (*health_check)(struct obd_export *))
1537 int abort = 0, expired = 1;
1540 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1541 spin_lock_bh(&obd->obd_processing_task_lock);
1542 abort = obd->obd_abort_recovery;
1543 expired = obd->obd_recovery_expired;
1544 obd->obd_recovery_expired = 0;
1545 spin_unlock_bh(&obd->obd_processing_task_lock);
1547 CWARN("recovery is aborted, evict exports in recovery\n");
1548 /** evict exports which didn't finish recovery yet */
1549 class_disconnect_stale_exports(obd, exp_finished);
1550 } else if (expired) {
1551 /** If some clients died being recovered, evict them */
1552 CDEBUG(D_WARNING, "recovery is timed out, evict stale exports\n");
1553 /** evict cexports with no replay in queue, they are stalled */
1554 class_disconnect_stale_exports(obd, health_check);
1555 /** continue with VBR */
1556 spin_lock_bh(&obd->obd_processing_task_lock);
1557 obd->obd_version_recov = 1;
1558 spin_unlock_bh(&obd->obd_processing_task_lock);
1560 * reset timer, recovery will proceed with versions now,
1561 * timeout is set just to handle reconnection delays
1563 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1564 /** Wait for recovery events again, after evicting bad clients */
1566 } while (!abort && expired);
1571 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1573 struct ptlrpc_request *req = NULL;
1576 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1577 obd->obd_next_recovery_transno);
1579 if (target_recovery_overseer(obd, check_for_next_transno,
1580 exp_req_replay_healthy)) {
1581 abort_req_replay_queue(obd);
1582 abort_lock_replay_queue(obd);
1585 spin_lock_bh(&obd->obd_processing_task_lock);
1586 if (!list_empty(&obd->obd_req_replay_queue)) {
1587 req = list_entry(obd->obd_req_replay_queue.next,
1588 struct ptlrpc_request, rq_list);
1589 list_del_init(&req->rq_list);
1590 obd->obd_requests_queued_for_recovery--;
1591 spin_unlock_bh(&obd->obd_processing_task_lock);
1593 spin_unlock_bh(&obd->obd_processing_task_lock);
1594 LASSERT(list_empty(&obd->obd_req_replay_queue));
1595 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1596 /** evict exports failed VBR */
1597 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1602 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1604 struct ptlrpc_request *req = NULL;
1606 CDEBUG(D_HA, "Waiting for lock\n");
1607 if (target_recovery_overseer(obd, check_for_next_lock,
1608 exp_lock_replay_healthy))
1609 abort_lock_replay_queue(obd);
1611 spin_lock_bh(&obd->obd_processing_task_lock);
1612 if (!list_empty(&obd->obd_lock_replay_queue)) {
1613 req = list_entry(obd->obd_lock_replay_queue.next,
1614 struct ptlrpc_request, rq_list);
1615 list_del_init(&req->rq_list);
1616 spin_unlock_bh(&obd->obd_processing_task_lock);
1618 spin_unlock_bh(&obd->obd_processing_task_lock);
1619 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1620 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1621 /** evict exports failed VBR */
1622 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1627 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1629 struct ptlrpc_request *req = NULL;
1631 spin_lock_bh(&obd->obd_processing_task_lock);
1632 if (!list_empty(&obd->obd_final_req_queue)) {
1633 req = list_entry(obd->obd_final_req_queue.next,
1634 struct ptlrpc_request, rq_list);
1635 list_del_init(&req->rq_list);
1636 if (req->rq_export->exp_in_recovery) {
1637 spin_lock(&req->rq_export->exp_lock);
1638 req->rq_export->exp_in_recovery = 0;
1639 spin_unlock(&req->rq_export->exp_lock);
1642 spin_unlock_bh(&obd->obd_processing_task_lock);
1646 static int handle_recovery_req(struct ptlrpc_thread *thread,
1647 struct ptlrpc_request *req,
1648 svc_handler_t handler)
1653 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1655 CERROR("Failure to initialize session: %d\n", rc);
1656 GOTO(reqcopy_put, rc);
1659 * export can be evicted during recovery, no need to handle replays for
1660 * it after that, discard such request silently
1662 if (req->rq_export->exp_disconnected)
1663 GOTO(reqcopy_put, rc);
1665 req->rq_recov_session.lc_thread = thread;
1666 lu_context_enter(&req->rq_recov_session);
1667 req->rq_svc_thread = thread;
1668 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1670 /* thread context */
1671 lu_context_enter(&thread->t_env->le_ctx);
1673 lu_context_exit(&thread->t_env->le_ctx);
1675 lu_context_exit(&req->rq_recov_session);
1676 lu_context_fini(&req->rq_recov_session);
1677 /* don't reset timer for final stage */
1678 if (!exp_finished(req->rq_export)) {
1680 * Add request timeout to the recovery time so next request from
1681 * this client may come in recovery time
1683 reset_recovery_timer(class_exp2obd(req->rq_export),
1684 AT_OFF ? obd_timeout :
1685 lustre_msg_get_timeout(req->rq_reqmsg), 1);
1691 static int target_recovery_thread(void *arg)
1693 struct lu_target *lut = arg;
1694 struct obd_device *obd = lut->lut_obd;
1695 struct ptlrpc_request *req;
1696 struct target_recovery_data *trd = &obd->obd_recovery_data;
1697 unsigned long delta;
1698 unsigned long flags;
1700 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1704 cfs_daemonize("tgt_recov");
1706 SIGNAL_MASK_LOCK(current, flags);
1707 sigfillset(¤t->blocked);
1709 SIGNAL_MASK_UNLOCK(current, flags);
1711 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1715 thread->t_env = &env;
1716 env.le_ctx.lc_thread = thread;
1718 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1720 trd->trd_processing_task = cfs_curproc_pid();
1722 obd->obd_recovering = 1;
1723 complete(&trd->trd_starting);
1725 /* first of all, we have to know the first transno to replay */
1726 if (target_recovery_overseer(obd, check_for_clients,
1727 exp_connect_healthy)) {
1728 abort_req_replay_queue(obd);
1729 abort_lock_replay_queue(obd);
1732 /* next stage: replay requests */
1734 obd->obd_req_replaying = 1;
1735 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1736 atomic_read(&obd->obd_req_replay_clients),
1737 obd->obd_next_recovery_transno);
1738 while ((req = target_next_replay_req(obd))) {
1739 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1740 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1741 lustre_msg_get_transno(req->rq_reqmsg),
1742 libcfs_nid2str(req->rq_peer.nid));
1743 handle_recovery_req(thread, req,
1744 trd->trd_recovery_handler);
1746 * bz18031: increase next_recovery_transno before
1747 * target_request_copy_put() will drop exp_rpc reference
1749 spin_lock_bh(&obd->obd_processing_task_lock);
1750 obd->obd_next_recovery_transno++;
1751 spin_unlock_bh(&obd->obd_processing_task_lock);
1752 target_exp_dequeue_req_replay(req);
1753 target_request_copy_put(req);
1754 obd->obd_replayed_requests++;
1758 * The second stage: replay locks
1760 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1761 atomic_read(&obd->obd_lock_replay_clients));
1762 while ((req = target_next_replay_lock(obd))) {
1763 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1764 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1765 libcfs_nid2str(req->rq_peer.nid));
1766 handle_recovery_req(thread, req,
1767 trd->trd_recovery_handler);
1768 target_request_copy_put(req);
1769 obd->obd_replayed_locks++;
1773 * The third stage: reply on final pings, at this moment all clients
1774 * must have request in final queue
1776 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1777 /** Update server last boot epoch */
1778 lut_boot_epoch_update(lut);
1779 /* We drop recoverying flag to forward all new requests
1780 * to regular mds_handle() since now */
1781 spin_lock_bh(&obd->obd_processing_task_lock);
1782 obd->obd_recovering = obd->obd_abort_recovery = 0;
1783 target_cancel_recovery_timer(obd);
1784 spin_unlock_bh(&obd->obd_processing_task_lock);
1785 while ((req = target_next_final_ping(obd))) {
1786 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1787 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1788 libcfs_nid2str(req->rq_peer.nid));
1789 handle_recovery_req(thread, req,
1790 trd->trd_recovery_handler);
1791 target_request_copy_put(req);
1794 delta = (jiffies - delta) / HZ;
1795 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1796 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1797 if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1798 CWARN("too long recovery - read logs\n");
1799 libcfs_debug_dumplog();
1802 target_finish_recovery(obd);
1804 lu_context_fini(&env.le_ctx);
1805 trd->trd_processing_task = 0;
1806 complete(&trd->trd_finishing);
1810 static int target_start_recovery_thread(struct lu_target *lut,
1811 svc_handler_t handler)
1813 struct obd_device *obd = lut->lut_obd;
1815 struct target_recovery_data *trd = &obd->obd_recovery_data;
1817 memset(trd, 0, sizeof(*trd));
1818 init_completion(&trd->trd_starting);
1819 init_completion(&trd->trd_finishing);
1820 trd->trd_recovery_handler = handler;
1822 if (kernel_thread(target_recovery_thread, lut, 0) > 0) {
1823 wait_for_completion(&trd->trd_starting);
1824 LASSERT(obd->obd_recovering != 0);
1831 void target_stop_recovery_thread(struct obd_device *obd)
1833 spin_lock_bh(&obd->obd_processing_task_lock);
1834 if (obd->obd_recovery_data.trd_processing_task > 0) {
1835 struct target_recovery_data *trd = &obd->obd_recovery_data;
1836 /** recovery can be done but postrecovery is not yet */
1837 if (obd->obd_recovering) {
1838 CERROR("%s: Aborting recovery\n", obd->obd_name);
1839 obd->obd_abort_recovery = 1;
1840 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1842 spin_unlock_bh(&obd->obd_processing_task_lock);
1843 wait_for_completion(&trd->trd_finishing);
1845 spin_unlock_bh(&obd->obd_processing_task_lock);
1849 void target_recovery_fini(struct obd_device *obd)
1851 class_disconnect_exports(obd);
1852 target_stop_recovery_thread(obd);
1853 target_cleanup_recovery(obd);
1855 EXPORT_SYMBOL(target_recovery_fini);
1857 static void target_recovery_expired(unsigned long castmeharder)
1859 struct obd_device *obd = (struct obd_device *)castmeharder;
1860 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1861 " after %lds (%d clients connected)\n",
1862 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
1863 cfs_time_current_sec()- obd->obd_recovery_start,
1864 obd->obd_connected_clients);
1866 spin_lock_bh(&obd->obd_processing_task_lock);
1867 obd->obd_recovery_expired = 1;
1868 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1869 spin_unlock_bh(&obd->obd_processing_task_lock);
1872 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1874 struct obd_device *obd = lut->lut_obd;
1875 if (obd->obd_max_recoverable_clients == 0) {
1876 /** Update server last boot epoch */
1877 lut_boot_epoch_update(lut);
1881 CWARN("RECOVERY: service %s, %d recoverable clients, "
1882 "last_transno "LPU64"\n", obd->obd_name,
1883 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1884 LASSERT(obd->obd_stopping == 0);
1885 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1886 obd->obd_recovery_start = 0;
1887 obd->obd_recovery_end = 0;
1888 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1889 /* bz13079: this should be set to desired value for ost but not for mds */
1890 obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1891 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1892 target_start_recovery_thread(lut, handler);
1894 EXPORT_SYMBOL(target_recovery_init);
1898 static int target_process_req_flags(struct obd_device *obd,
1899 struct ptlrpc_request *req)
1901 struct obd_export *exp = req->rq_export;
1902 LASSERT(exp != NULL);
1903 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1904 /* client declares he's ready to replay locks */
1905 spin_lock_bh(&obd->obd_processing_task_lock);
1906 if (exp->exp_req_replay_needed) {
1907 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1908 spin_lock(&exp->exp_lock);
1909 exp->exp_req_replay_needed = 0;
1910 spin_unlock(&exp->exp_lock);
1911 atomic_dec(&obd->obd_req_replay_clients);
1913 spin_unlock_bh(&obd->obd_processing_task_lock);
1915 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1916 /* client declares he's ready to complete recovery
1917 * so, we put the request on th final queue */
1918 spin_lock_bh(&obd->obd_processing_task_lock);
1919 if (exp->exp_lock_replay_needed) {
1920 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1921 spin_lock(&exp->exp_lock);
1922 exp->exp_lock_replay_needed = 0;
1923 spin_unlock(&exp->exp_lock);
1924 atomic_dec(&obd->obd_lock_replay_clients);
1926 spin_unlock_bh(&obd->obd_processing_task_lock);
1932 int target_queue_recovery_request(struct ptlrpc_request *req,
1933 struct obd_device *obd)
1935 struct list_head *tmp;
1937 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1940 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1941 /* Processing the queue right now, don't re-add. */
1945 target_process_req_flags(obd, req);
1947 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1948 /* client declares he's ready to complete recovery
1949 * so, we put the request on th final queue */
1950 target_request_copy_get(req);
1951 DEBUG_REQ(D_HA, req, "queue final req");
1952 spin_lock_bh(&obd->obd_processing_task_lock);
1953 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1954 if (obd->obd_recovering) {
1955 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1957 spin_unlock_bh(&obd->obd_processing_task_lock);
1958 target_request_copy_put(req);
1959 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1961 spin_unlock_bh(&obd->obd_processing_task_lock);
1964 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1965 /* client declares he's ready to replay locks */
1966 target_request_copy_get(req);
1967 DEBUG_REQ(D_HA, req, "queue lock replay req");
1968 spin_lock_bh(&obd->obd_processing_task_lock);
1969 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1970 LASSERT(obd->obd_recovering);
1971 /* usually due to recovery abort */
1972 if (!req->rq_export->exp_in_recovery) {
1973 spin_unlock_bh(&obd->obd_processing_task_lock);
1974 target_request_copy_put(req);
1977 LASSERT(req->rq_export->exp_lock_replay_needed);
1978 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1979 spin_unlock_bh(&obd->obd_processing_task_lock);
1983 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1984 * (i.e. buflens etc are in my own byte order), but type-dependent
1985 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1988 CFS_INIT_LIST_HEAD(&req->rq_list);
1989 DEBUG_REQ(D_HA, req, "not queueing");
1993 spin_lock_bh(&obd->obd_processing_task_lock);
1995 /* If we're processing the queue, we want don't want to queue this
1998 * Also, if this request has a transno less than the one we're waiting
1999 * for, we should process it now. It could (and currently always will)
2000 * be an open request for a descriptor that was opened some time ago.
2002 * Also, a resent, replayed request that has already been
2003 * handled will pass through here and be processed immediately.
2005 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2006 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2007 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2008 /* Processing the queue right now, don't re-add. */
2009 LASSERT(list_empty(&req->rq_list));
2010 spin_unlock_bh(&obd->obd_processing_task_lock);
2013 spin_unlock_bh(&obd->obd_processing_task_lock);
2015 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2018 target_request_copy_get(req);
2019 spin_lock_bh(&obd->obd_processing_task_lock);
2020 LASSERT(obd->obd_recovering);
2021 if (!req->rq_export->exp_in_recovery) {
2022 spin_unlock_bh(&obd->obd_processing_task_lock);
2023 target_request_copy_put(req);
2026 LASSERT(req->rq_export->exp_req_replay_needed);
2028 if (target_exp_enqueue_req_replay(req)) {
2029 spin_unlock_bh(&obd->obd_processing_task_lock);
2030 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2031 target_request_copy_put(req);
2036 list_for_each(tmp, &obd->obd_req_replay_queue) {
2037 struct ptlrpc_request *reqiter =
2038 list_entry(tmp, struct ptlrpc_request, rq_list);
2040 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2041 list_add_tail(&req->rq_list, &reqiter->rq_list);
2046 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2048 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2049 "has been claimed by another client");
2050 spin_unlock_bh(&obd->obd_processing_task_lock);
2051 target_exp_dequeue_req_replay(req);
2052 target_request_copy_put(req);
2058 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2060 obd->obd_requests_queued_for_recovery++;
2061 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2062 spin_unlock_bh(&obd->obd_processing_task_lock);
2067 * Packs current SLV and Limit into \a req.
2069 int target_pack_pool_reply(struct ptlrpc_request *req)
2071 struct obd_device *obd;
2075 * Check that we still have all structures alive as this may
2076 * be some late rpc in shutdown time.
2078 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2079 !exp_connect_lru_resize(req->rq_export))) {
2080 lustre_msg_set_slv(req->rq_repmsg, 0);
2081 lustre_msg_set_limit(req->rq_repmsg, 0);
2086 * OBD is alive here as export is alive, which we checked above.
2088 obd = req->rq_export->exp_obd;
2090 read_lock(&obd->obd_pool_lock);
2091 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2092 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2093 read_unlock(&obd->obd_pool_lock);
2098 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2100 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2101 DEBUG_REQ(D_ERROR, req, "dropping reply");
2106 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2107 req->rq_status = rc;
2108 return (ptlrpc_send_error(req, 1));
2110 DEBUG_REQ(D_NET, req, "sending reply");
2113 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2116 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2119 struct ptlrpc_reply_state *rs;
2120 struct obd_device *obd;
2121 struct obd_export *exp;
2122 struct ptlrpc_service *svc;
2125 if (req->rq_no_reply) {
2130 svc = req->rq_rqbd->rqbd_service;
2131 rs = req->rq_reply_state;
2132 if (rs == NULL || !rs->rs_difficult) {
2134 target_send_reply_msg (req, rc, fail_id);
2139 /* must be an export if locks saved */
2140 LASSERT (req->rq_export != NULL);
2141 /* req/reply consistent */
2142 LASSERT (rs->rs_service == svc);
2145 LASSERT (!rs->rs_scheduled);
2146 LASSERT (!rs->rs_scheduled_ever);
2147 LASSERT (!rs->rs_handled);
2148 LASSERT (!rs->rs_on_net);
2149 LASSERT (rs->rs_export == NULL);
2150 LASSERT (list_empty(&rs->rs_obd_list));
2151 LASSERT (list_empty(&rs->rs_exp_list));
2153 exp = class_export_get (req->rq_export);
2156 /* disable reply scheduling while I'm setting up */
2157 rs->rs_scheduled = 1;
2159 rs->rs_xid = req->rq_xid;
2160 rs->rs_transno = req->rq_transno;
2161 rs->rs_export = exp;
2162 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2164 spin_lock(&exp->exp_uncommitted_replies_lock);
2165 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2166 rs->rs_transno, exp->exp_last_committed);
2167 if (rs->rs_transno > exp->exp_last_committed) {
2168 /* not committed already */
2169 list_add_tail(&rs->rs_obd_list,
2170 &exp->exp_uncommitted_replies);
2172 spin_unlock (&exp->exp_uncommitted_replies_lock);
2174 spin_lock(&exp->exp_lock);
2175 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2176 spin_unlock(&exp->exp_lock);
2178 netrc = target_send_reply_msg (req, rc, fail_id);
2180 spin_lock(&svc->srv_lock);
2182 atomic_inc(&svc->srv_n_difficult_replies);
2185 /* error sending: reply is off the net. Also we need +1
2186 * reply ref until ptlrpc_handle_rs() is done
2187 * with the reply state (if the send was successful, there
2188 * would have been +1 ref for the net, which
2189 * reply_out_callback leaves alone) */
2191 ptlrpc_rs_addref(rs);
2192 atomic_inc (&svc->srv_outstanding_replies);
2195 spin_lock(&rs->rs_lock);
2196 if (rs->rs_transno <= exp->exp_last_committed ||
2197 (!rs->rs_on_net && !rs->rs_no_ack) ||
2198 list_empty(&rs->rs_exp_list) || /* completed already */
2199 list_empty(&rs->rs_obd_list)) {
2200 CDEBUG(D_HA, "Schedule reply immediately\n");
2201 ptlrpc_dispatch_difficult_reply(rs);
2203 list_add (&rs->rs_list, &svc->srv_active_replies);
2204 rs->rs_scheduled = 0; /* allow notifier to schedule */
2206 spin_unlock(&rs->rs_lock);
2207 spin_unlock(&svc->srv_lock);
2211 int target_handle_ping(struct ptlrpc_request *req)
2213 obd_ping(req->rq_export);
2214 return req_capsule_server_pack(&req->rq_pill);
2217 void target_committed_to_req(struct ptlrpc_request *req)
2219 struct obd_export *exp = req->rq_export;
2221 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2222 lustre_msg_set_last_committed(req->rq_repmsg,
2223 exp->exp_last_committed);
2225 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2226 "%d)", exp->exp_obd->obd_no_transno,
2227 req->rq_repmsg == NULL);
2229 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2230 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2232 EXPORT_SYMBOL(target_committed_to_req);
2234 int target_handle_qc_callback(struct ptlrpc_request *req)
2236 struct obd_quotactl *oqctl;
2237 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2239 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2240 if (oqctl == NULL) {
2241 CERROR("Can't unpack obd_quotactl\n");
2245 cli->cl_qchk_stat = oqctl->qc_stat;
2250 #ifdef HAVE_QUOTA_SUPPORT
2251 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2254 struct obd_device *obd = req->rq_export->exp_obd;
2255 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2256 struct obd_device_target *obt;
2257 struct lustre_quota_ctxt *qctxt;
2258 struct qunit_data *qdata = NULL;
2262 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2265 rc = req_capsule_server_pack(&req->rq_pill);
2267 CERROR("packing reply failed!: rc = %d\n", rc);
2271 LASSERT(req->rq_export);
2273 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2274 if (IS_ERR(qdata)) {
2275 rc = PTR_ERR(qdata);
2276 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2277 req->rq_status = rc;
2281 /* we use the observer */
2282 if (obd_pin_observer(obd, &lov_obd) ||
2283 obd_pin_observer(lov_obd, &master_obd)) {
2284 CERROR("Can't find the observer, it is recovering\n");
2285 req->rq_status = -EAGAIN;
2289 obt = &master_obd->u.obt;
2290 qctxt = &obt->obt_qctxt;
2292 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2293 /* quota_type has not been processed yet, return EAGAIN
2294 * until we know whether or not quotas are supposed to
2296 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2298 req->rq_status = -EAGAIN;
2302 down_read(&obt->obt_rwsem);
2303 if (qctxt->lqc_lqs_hash == NULL) {
2304 up_read(&obt->obt_rwsem);
2305 /* quota_type has not been processed yet, return EAGAIN
2306 * until we know whether or not quotas are supposed to
2308 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2310 req->rq_status = -EAGAIN;
2314 LASSERT(qctxt->lqc_handler);
2315 rc = qctxt->lqc_handler(master_obd, qdata,
2316 lustre_msg_get_opc(req->rq_reqmsg));
2317 up_read(&obt->obt_rwsem);
2318 if (rc && rc != -EDQUOT)
2319 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2320 "dqacq/dqrel failed! (rc:%d)\n", rc);
2321 req->rq_status = rc;
2323 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2325 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2329 /* Block the quota req. b=14840 */
2330 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2335 obd_unpin_observer(lov_obd);
2337 obd_unpin_observer(obd);
2339 rc = ptlrpc_reply(req);
2343 #endif /* !__KERNEL__ */
2345 #endif /* HAVE_QUOTA_SUPPORT */
2347 /* Send a remote set_info_async.
2348 * This may go from client to server or server to client
2350 int target_set_info_rpc(struct obd_import *imp, int opcode,
2351 obd_count keylen, void *key,
2352 obd_count vallen, void *val,
2353 struct ptlrpc_request_set *set)
2355 struct ptlrpc_request *req;
2360 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
2364 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2365 RCL_CLIENT, keylen);
2366 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2367 RCL_CLIENT, vallen);
2368 rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, opcode);
2370 ptlrpc_request_free(req);
2374 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2375 memcpy(tmp, key, keylen);
2376 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2377 memcpy(tmp, val, vallen);
2379 ptlrpc_request_set_replen(req);
2382 ptlrpc_set_add_req(set, req);
2383 ptlrpc_check_set(NULL, set);
2385 rc = ptlrpc_queue_wait(req);
2386 ptlrpc_req_finished(req);
2391 EXPORT_SYMBOL(target_set_info_rpc);
2394 ldlm_mode_t lck_compat_array[] = {
2395 [LCK_EX] LCK_COMPAT_EX,
2396 [LCK_PW] LCK_COMPAT_PW,
2397 [LCK_PR] LCK_COMPAT_PR,
2398 [LCK_CW] LCK_COMPAT_CW,
2399 [LCK_CR] LCK_COMPAT_CR,
2400 [LCK_NL] LCK_COMPAT_NL,
2401 [LCK_GROUP] LCK_COMPAT_GROUP,
2402 [LCK_COS] LCK_COMPAT_COS,
2406 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2407 * not escape to the user level.
2409 int ldlm_error2errno(ldlm_error_t error)
2417 case ELDLM_LOCK_CHANGED:
2420 case ELDLM_LOCK_ABORTED:
2423 case ELDLM_LOCK_REPLACED:
2426 case ELDLM_NO_LOCK_DATA:
2429 case ELDLM_NAMESPACE_EXISTS:
2432 case ELDLM_BAD_NAMESPACE:
2436 if (((int)error) < 0) /* cast to signed type */
2437 result = error; /* as ldlm_error_t can be unsigned */
2439 CERROR("Invalid DLM result code: %i\n", error);
2445 EXPORT_SYMBOL(ldlm_error2errno);
2448 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2450 ldlm_error_t ldlm_errno2error(int err_no)
2459 error = ELDLM_LOCK_CHANGED;
2462 error = ELDLM_LOCK_ABORTED;
2465 error = ELDLM_LOCK_REPLACED;
2468 error = ELDLM_NO_LOCK_DATA;
2471 error = ELDLM_NAMESPACE_EXISTS;
2474 error = ELDLM_BAD_NAMESPACE;
2481 EXPORT_SYMBOL(ldlm_errno2error);
2483 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2484 void ldlm_dump_export_locks(struct obd_export *exp)
2486 spin_lock(&exp->exp_locks_list_guard);
2487 if (!list_empty(&exp->exp_locks_list)) {
2488 struct ldlm_lock *lock;
2490 CERROR("dumping locks for export %p,"
2491 "ignore if the unmount doesn't hang\n", exp);
2492 list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2493 ldlm_lock_dump(D_ERROR, lock, 0);
2495 spin_unlock(&exp->exp_locks_list_guard);