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 item->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 static void 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_write_list);
281 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
282 client_obd_list_lock_init(&cli->cl_loi_list_lock);
283 cli->cl_r_in_flight = 0;
284 cli->cl_w_in_flight = 0;
286 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
287 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
288 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
289 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
290 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
291 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
292 cfs_waitq_init(&cli->cl_destroy_waitq);
293 atomic_set(&cli->cl_destroy_in_flight, 0);
294 #ifdef ENABLE_CHECKSUM
295 /* Turn on checksumming by default. */
296 cli->cl_checksum = 1;
298 * The supported checksum types will be worked out at connect time
299 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
302 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
304 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
306 /* This value may be changed at connect time in
307 ptlrpc_connect_interpret. */
308 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
309 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
311 if (!strcmp(name, LUSTRE_MDC_NAME)) {
312 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
313 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
314 cli->cl_max_rpcs_in_flight = 2;
315 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
316 cli->cl_max_rpcs_in_flight = 3;
317 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
318 cli->cl_max_rpcs_in_flight = 4;
320 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
325 CERROR("ldlm_get_ref failed: %d\n", rc);
329 ptlrpc_init_client(rq_portal, rp_portal, name,
330 &obddev->obd_ldlm_client);
332 imp = class_new_import(obddev);
334 GOTO(err_ldlm, rc = -ENOENT);
335 imp->imp_client = &obddev->obd_ldlm_client;
336 imp->imp_connect_op = connect_op;
337 imp->imp_initial_recov = 1;
338 imp->imp_initial_recov_bk = 0;
339 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
340 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
341 LUSTRE_CFG_BUFLEN(lcfg, 1));
342 class_import_put(imp);
344 rc = client_import_add_conn(imp, &server_uuid, 1);
346 CERROR("can't add initial connection\n");
347 GOTO(err_import, rc);
350 cli->cl_import = imp;
351 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
352 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
353 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
355 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
356 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
357 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
358 name, obddev->obd_name,
359 cli->cl_target_uuid.uuid);
360 spin_lock(&imp->imp_lock);
361 imp->imp_invalid = 1;
362 spin_unlock(&imp->imp_lock);
366 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
371 class_destroy_import(imp);
379 int client_obd_cleanup(struct obd_device *obddev)
386 /* ->o_connect() method for client side (OSC and MDC and MGC) */
387 int client_connect_import(const struct lu_env *env,
388 struct lustre_handle *dlm_handle,
389 struct obd_device *obd, struct obd_uuid *cluuid,
390 struct obd_connect_data *data, void *localdata)
392 struct client_obd *cli = &obd->u.cli;
393 struct obd_import *imp = cli->cl_import;
394 struct obd_export *exp;
395 struct obd_connect_data *ocd;
396 struct ldlm_namespace *to_be_freed = NULL;
400 down_write(&cli->cl_sem);
401 rc = class_connect(dlm_handle, obd, cluuid);
405 cli->cl_conn_count++;
406 if (cli->cl_conn_count > 1)
408 exp = class_conn2export(dlm_handle);
410 if (obd->obd_namespace != NULL)
411 CERROR("already have namespace!\n");
412 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
413 LDLM_NAMESPACE_CLIENT,
414 LDLM_NAMESPACE_GREEDY);
415 if (obd->obd_namespace == NULL)
416 GOTO(out_disco, rc = -ENOMEM);
418 imp->imp_dlm_handle = *dlm_handle;
419 rc = ptlrpc_init_import(imp);
423 ocd = &imp->imp_connect_data;
426 imp->imp_connect_flags_orig = data->ocd_connect_flags;
429 rc = ptlrpc_connect_import(imp, NULL);
431 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
434 LASSERT(exp->exp_connection);
437 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
438 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
439 data->ocd_connect_flags, ocd->ocd_connect_flags);
440 data->ocd_connect_flags = ocd->ocd_connect_flags;
443 ptlrpc_pinger_add_import(imp);
449 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
450 to_be_freed = obd->obd_namespace;
451 obd->obd_namespace = NULL;
453 cli->cl_conn_count--;
454 class_disconnect(exp);
456 class_export_put(exp);
459 up_write(&cli->cl_sem);
461 ldlm_namespace_free_post(to_be_freed);
466 int client_disconnect_export(struct obd_export *exp)
468 struct obd_device *obd = class_exp2obd(exp);
469 struct client_obd *cli;
470 struct obd_import *imp;
472 struct ldlm_namespace *to_be_freed = NULL;
476 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
477 exp, exp ? exp->exp_handle.h_cookie : -1);
482 imp = cli->cl_import;
484 down_write(&cli->cl_sem);
485 if (!cli->cl_conn_count) {
486 CERROR("disconnecting disconnected device (%s)\n",
488 GOTO(out_sem, rc = -EINVAL);
491 cli->cl_conn_count--;
492 if (cli->cl_conn_count)
493 GOTO(out_no_disconnect, rc = 0);
495 /* Mark import deactivated now, so we don't try to reconnect if any
496 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
497 * fully deactivate the import, or that would drop all requests. */
498 spin_lock(&imp->imp_lock);
499 imp->imp_deactive = 1;
500 spin_unlock(&imp->imp_lock);
502 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
503 * delete it regardless. (It's safe to delete an import that was
505 (void)ptlrpc_pinger_del_import(imp);
507 if (obd->obd_namespace != NULL) {
508 /* obd_force == local only */
509 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
510 obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
512 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
513 to_be_freed = obd->obd_namespace;
516 rc = ptlrpc_disconnect_import(imp, 0);
518 ptlrpc_invalidate_import(imp);
519 /* set obd_namespace to NULL only after invalidate, because we can have
520 * some connect requests in flight, and his need store a connect flags
521 * in obd_namespace. bug 14260 */
522 obd->obd_namespace = NULL;
524 if (imp->imp_rq_pool) {
525 ptlrpc_free_rq_pool(imp->imp_rq_pool);
526 imp->imp_rq_pool = NULL;
529 cli->cl_import = NULL;
533 err = class_disconnect(exp);
537 up_write(&cli->cl_sem);
539 ldlm_namespace_free_post(to_be_freed);
544 /* --------------------------------------------------------------------------
545 * from old lib/target.c
546 * -------------------------------------------------------------------------- */
548 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
549 struct obd_uuid *cluuid, int initial_conn)
552 if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
553 struct lustre_handle *hdl;
554 hdl = &exp->exp_imp_reverse->imp_remote_handle;
555 /* Might be a re-connect after a partition. */
556 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
557 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
559 conn->cookie = exp->exp_handle.h_cookie;
560 /* target_handle_connect() treats EALREADY and
561 * -EALREADY differently. EALREADY means we are
562 * doing a valid reconnect from the same client. */
565 CERROR("%s reconnecting from %s, "
566 "handle mismatch (ours "LPX64", theirs "
567 LPX64")\n", cluuid->uuid,
568 exp->exp_connection->c_remote_uuid.uuid,
569 hdl->cookie, conn->cookie);
570 memset(conn, 0, sizeof *conn);
571 /* target_handle_connect() treats EALREADY and
572 * -EALREADY differently. -EALREADY is an error
573 * (same UUID, different handle). */
578 conn->cookie = exp->exp_handle.h_cookie;
579 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
580 cluuid->uuid, exp, conn->cookie);
584 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
587 struct obd_export *exp = cb_data;
589 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
590 obd->obd_name, exp->exp_client_uuid.uuid);
592 spin_lock(&exp->exp_lock);
593 exp->exp_need_sync = 0;
594 spin_unlock(&exp->exp_lock);
596 EXPORT_SYMBOL(target_client_add_cb);
598 target_start_and_reset_recovery_timer(struct obd_device *obd,
599 struct ptlrpc_request *req,
602 int target_handle_connect(struct ptlrpc_request *req)
604 struct obd_device *target, *targref = NULL;
605 struct obd_export *export = NULL;
606 struct obd_import *revimp;
607 struct lustre_handle conn;
608 struct lustre_handle *tmp;
609 struct obd_uuid tgtuuid;
610 struct obd_uuid cluuid;
611 struct obd_uuid remote_uuid;
614 int initial_conn = 0;
615 struct obd_connect_data *data, *tmpdata;
616 lnet_nid_t *client_nid = NULL;
619 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
621 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
623 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
624 GOTO(out, rc = -EINVAL);
627 obd_str2uuid(&tgtuuid, str);
628 target = class_uuid2obd(&tgtuuid);
630 target = class_name2obd(str);
632 if (!target || target->obd_stopping || !target->obd_set_up) {
633 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
634 " for connect (%s)\n", str,
635 !target ? "no target" :
636 (target->obd_stopping ? "stopping" :
638 GOTO(out, rc = -ENODEV);
641 if (target->obd_no_conn) {
642 LCONSOLE_WARN("%s: temporarily refusing client connection "
643 "from %s\n", target->obd_name,
644 libcfs_nid2str(req->rq_peer.nid));
645 GOTO(out, rc = -EAGAIN);
648 /* Make sure the target isn't cleaned up while we're here. Yes,
649 there's still a race between the above check and our incref here.
650 Really, class_uuid2obd should take the ref. */
651 targref = class_incref(target, __FUNCTION__, cfs_current());
654 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
656 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
657 GOTO(out, rc = -EINVAL);
660 obd_str2uuid(&cluuid, str);
662 /* XXX extract a nettype and format accordingly */
663 switch (sizeof(lnet_nid_t)) {
664 /* NB the casts only avoid compiler warnings */
666 snprintf(remote_uuid.uuid, sizeof remote_uuid,
667 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
670 snprintf(remote_uuid.uuid, sizeof remote_uuid,
671 "NET_%x_UUID", (__u32)req->rq_peer.nid);
677 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
679 GOTO(out, rc = -EPROTO);
683 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
685 GOTO(out, rc = -EPROTO);
687 rc = req_capsule_server_pack(&req->rq_pill);
691 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
693 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
694 "libclient connection attempt");
695 GOTO(out, rc = -EPROTO);
696 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
697 LUSTRE_VERSION_ALLOWED_OFFSET ||
698 data->ocd_version > LUSTRE_VERSION_CODE +
699 LUSTRE_VERSION_ALLOWED_OFFSET) {
700 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
701 "libclient connection attempt",
702 data->ocd_version < LUSTRE_VERSION_CODE ?
704 OBD_OCD_VERSION_MAJOR(data->ocd_version),
705 OBD_OCD_VERSION_MINOR(data->ocd_version),
706 OBD_OCD_VERSION_PATCH(data->ocd_version),
707 OBD_OCD_VERSION_FIX(data->ocd_version));
708 data = req_capsule_server_sized_get(&req->rq_pill,
710 offsetof(typeof(*data), ocd_version) +
711 sizeof(data->ocd_version));
713 data->ocd_connect_flags = OBD_CONNECT_VERSION;
714 data->ocd_version = LUSTRE_VERSION_CODE;
716 GOTO(out, rc = -EPROTO);
720 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
723 /* lctl gets a backstage, all-access pass. */
724 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
725 goto dont_check_exports;
727 spin_lock(&target->obd_dev_lock);
728 export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
730 if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
731 CWARN("%s: exp %p already connecting\n",
732 export->exp_obd->obd_name, export);
733 class_export_put(export);
736 } else if (export != NULL && export->exp_connection != NULL &&
737 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
738 /* make darn sure this is coming from the same peer
739 * if the UUIDs matched */
740 if (data && data->ocd_connect_flags & OBD_CONNECT_MDS) {
741 /* the MDS UUID can be reused, don't need to wait
742 * for the export to be evicted */
743 CWARN("%s: received MDS connection from a new NID %s,"
744 " removing former export from NID %s\n",
746 libcfs_nid2str(req->rq_peer.nid),
747 libcfs_nid2str(export->exp_connection->c_peer.nid));
748 class_fail_export(export);
750 CWARN("%s: cookie %s seen on new NID %s when "
751 "existing NID %s is already connected\n",
752 target->obd_name, cluuid.uuid,
753 libcfs_nid2str(req->rq_peer.nid),
754 libcfs_nid2str(export->exp_connection->c_peer.nid));
757 class_export_put(export);
759 } else if (export != NULL) {
760 spin_lock(&export->exp_lock);
761 export->exp_connecting = 1;
762 spin_unlock(&export->exp_lock);
763 class_export_put(export);
764 spin_unlock(&target->obd_dev_lock);
765 LASSERT(export->exp_obd == target);
767 rc = target_handle_reconnect(&conn, export, &cluuid, initial_conn);
770 /* If we found an export, we already unlocked. */
772 spin_unlock(&target->obd_dev_lock);
773 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
774 } else if (req->rq_export == NULL &&
775 atomic_read(&export->exp_rpc_count) > 0) {
776 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
777 target->obd_name, cluuid.uuid,
778 libcfs_nid2str(req->rq_peer.nid),
779 export, atomic_read(&export->exp_refcount));
780 GOTO(out, rc = -EBUSY);
781 } else if (req->rq_export != NULL &&
782 (atomic_read(&export->exp_rpc_count) > 1)) {
783 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
784 target->obd_name, cluuid.uuid,
785 libcfs_nid2str(req->rq_peer.nid),
786 export, atomic_read(&export->exp_rpc_count));
787 GOTO(out, rc = -EBUSY);
788 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
790 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
791 "cookies not random?\n", target->obd_name,
792 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
793 GOTO(out, rc = -EALREADY);
795 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
796 if (req->rq_export == NULL && initial_conn)
797 export->exp_last_request_time =
798 max(export->exp_last_request_time,
799 (time_t)cfs_time_current_sec());
806 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
807 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
808 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
809 export, (long)cfs_time_current_sec(),
810 export ? (long)export->exp_last_request_time : 0);
812 /* Tell the client if we're in recovery. */
813 if (target->obd_recovering) {
814 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
815 /* If this is the first time a client connects,
816 reset the recovery timer */
818 target_start_and_reset_recovery_timer(target, req,
822 /* We want to handle EALREADY but *not* -EALREADY from
823 * target_handle_reconnect(), return reconnection state in a flag */
824 if (rc == EALREADY) {
825 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
831 /* Tell the client if we support replayable requests */
832 if (target->obd_replayable)
833 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
834 client_nid = &req->rq_peer.nid;
836 if (export == NULL) {
837 if (target->obd_recovering) {
840 t = cfs_timer_deadline(&target->obd_recovery_timer);
841 t = cfs_time_sub(t, cfs_time_current());
842 CERROR("%s: denying connection for new client %s (%s): "
843 "%d clients in recovery for "CFS_TIME_T"s\n",
845 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
846 target->obd_recoverable_clients,
847 cfs_duration_sec(t));
851 rc = obd_connect(req->rq_svc_thread->t_env,
852 &conn, target, &cluuid, data,
856 rc = obd_reconnect(req->rq_svc_thread->t_env,
857 export, target, &cluuid, data, client_nid);
861 /* Return only the parts of obd_connect_data that we understand, so the
862 * client knows that we don't understand the rest. */
864 tmpdata = req_capsule_server_get(&req->rq_pill,
866 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
870 /* If all else goes well, this is our RPC return code. */
873 lustre_msg_set_handle(req->rq_repmsg, &conn);
875 /* ownership of this export ref transfers to the request AFTER we
876 * drop any previous reference the request had, but we don't want
877 * that to go to zero before we get our new export reference. */
878 export = class_conn2export(&conn);
880 DEBUG_REQ(D_ERROR, req, "Missing export!");
881 GOTO(out, rc = -ENODEV);
884 /* If the client and the server are the same node, we will already
885 * have an export that really points to the client's DLM export,
886 * because we have a shared handles table.
888 * XXX this will go away when shaver stops sending the "connect" handle
889 * in the real "remote handle" field of the request --phik 24 Apr 2003
891 if (req->rq_export != NULL)
892 class_export_put(req->rq_export);
894 req->rq_export = export;
896 spin_lock(&export->exp_lock);
898 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
899 } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
900 spin_unlock(&export->exp_lock);
901 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
902 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
903 export->exp_conn_cnt,
904 lustre_msg_get_conn_cnt(req->rq_reqmsg));
906 GOTO(out, rc = -EALREADY);
908 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
910 /* request from liblustre? Don't evict it for not pinging. */
911 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
912 export->exp_libclient = 1;
913 spin_unlock(&export->exp_lock);
915 spin_lock(&target->obd_dev_lock);
916 list_del_init(&export->exp_obd_chain_timed);
917 spin_unlock(&target->obd_dev_lock);
919 spin_unlock(&export->exp_lock);
922 if (export->exp_connection != NULL)
923 ptlrpc_connection_put(export->exp_connection);
924 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
927 if (hlist_unhashed(&export->exp_nid_hash)) {
928 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
929 &export->exp_connection->c_peer.nid,
930 &export->exp_nid_hash);
933 spin_lock_bh(&target->obd_processing_task_lock);
934 if (target->obd_recovering && !export->exp_in_recovery) {
935 spin_lock(&export->exp_lock);
936 export->exp_in_recovery = 1;
937 export->exp_req_replay_needed = 1;
938 export->exp_lock_replay_needed = 1;
939 spin_unlock(&export->exp_lock);
940 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
941 && (data->ocd_transno == 0))
942 CWARN("Connect with zero transno!\n");
944 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
945 && data->ocd_transno < target->obd_next_recovery_transno)
946 target->obd_next_recovery_transno = data->ocd_transno;
947 target->obd_connected_clients++;
948 /* each connected client is counted as recoverable */
949 target->obd_recoverable_clients++;
950 atomic_inc(&target->obd_req_replay_clients);
951 atomic_inc(&target->obd_lock_replay_clients);
952 if (target->obd_connected_clients ==
953 target->obd_max_recoverable_clients)
954 wake_up(&target->obd_next_transno_waitq);
956 spin_unlock_bh(&target->obd_processing_task_lock);
957 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
960 if (export->exp_imp_reverse != NULL) {
961 /* destroyed import can be still referenced in ctxt */
962 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
963 KEY_REVIMP_UPD, 0, NULL, NULL);
965 /* in some recovery senarios, previous ctx init rpc handled
966 * in sptlrpc_target_export_check() might be used to install
967 * a reverse ctx in this reverse import, and later OBD_CONNECT
968 * using the same gss ctx could reach here and following new
969 * reverse import. note all reverse ctx in new/old import are
970 * actually based on the same gss ctx. so we invalidate ctx
971 * here before destroy import, otherwise flush old import will
972 * lead to remote reverse ctx be destroied, thus the reverse
973 * ctx of new import will lost its peer.
974 * there might be a better way to deal with this???
976 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
978 destroy_import(export->exp_imp_reverse);
981 /* for the rest part, we return -ENOTCONN in case of errors
982 * in order to let client initialize connection again.
984 revimp = export->exp_imp_reverse = class_new_import(target);
986 CERROR("fail to alloc new reverse import.\n");
987 GOTO(out, rc = -ENOTCONN);
990 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
991 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
992 revimp->imp_remote_handle = conn;
993 revimp->imp_dlm_fake = 1;
994 revimp->imp_state = LUSTRE_IMP_FULL;
996 /* unknown versions will be caught in
997 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
998 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1000 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1001 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1002 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1004 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1006 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx,
1007 req->rq_flvr.sf_rpc);
1009 CERROR("Failed to get sec for reverse import: %d\n", rc);
1010 export->exp_imp_reverse = NULL;
1011 class_destroy_import(revimp);
1014 class_import_put(revimp);
1017 spin_lock(&export->exp_lock);
1018 export->exp_connecting = 0;
1019 spin_unlock(&export->exp_lock);
1022 class_decref(targref, __FUNCTION__, cfs_current());
1024 req->rq_status = rc;
1028 int target_handle_disconnect(struct ptlrpc_request *req)
1033 rc = req_capsule_server_pack(&req->rq_pill);
1037 /* keep the rq_export around so we can send the reply */
1038 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1043 void target_destroy_export(struct obd_export *exp)
1045 /* exports created from last_rcvd data, and "fake"
1046 exports created by lctl don't have an import */
1047 if (exp->exp_imp_reverse != NULL)
1048 destroy_import(exp->exp_imp_reverse);
1050 /* We cancel locks at disconnect time, but this will catch any locks
1051 * granted in a race with recovery-induced disconnect. */
1052 if (exp->exp_obd->obd_namespace != NULL)
1053 ldlm_cancel_locks_for_export(exp);
1057 * Recovery functions
1060 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1062 struct ptlrpc_request *copy_req;
1063 struct lustre_msg *copy_reqmsg;
1064 struct ptlrpc_user_desc *udesc = NULL;
1066 OBD_ALLOC_PTR(copy_req);
1069 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1071 OBD_FREE_PTR(copy_req);
1075 if (orig_req->rq_user_desc) {
1076 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1078 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1080 OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1081 OBD_FREE_PTR(copy_req);
1084 memcpy(udesc, orig_req->rq_user_desc,
1085 sptlrpc_user_desc_size(ngroups));
1088 *copy_req = *orig_req;
1089 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1090 copy_req->rq_reqmsg = copy_reqmsg;
1091 copy_req->rq_user_desc = udesc;
1093 class_export_rpc_get(copy_req->rq_export);
1094 CFS_INIT_LIST_HEAD(©_req->rq_list);
1095 CFS_INIT_LIST_HEAD(©_req->rq_replay_list);
1096 sptlrpc_svc_ctx_addref(copy_req);
1098 if (copy_req->rq_reply_state) {
1099 /* the copied req takes over the reply state */
1100 orig_req->rq_reply_state = NULL;
1101 /* to catch further access */
1102 orig_req->rq_repmsg = NULL;
1103 orig_req->rq_replen = 0;
1109 void ptlrpc_free_clone(struct ptlrpc_request *req)
1111 LASSERT(list_empty(&req->rq_replay_list));
1113 ptlrpc_req_drop_rs(req);
1114 sptlrpc_svc_ctx_decref(req);
1115 class_export_rpc_put(req->rq_export);
1116 list_del_init(&req->rq_list);
1118 if (req->rq_user_desc) {
1119 int ngroups = req->rq_user_desc->pud_ngroups;
1120 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1122 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
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 obd->obd_processing_task = 0;
1182 CERROR("%s: Recovery queues ( %s%s%s) are empty\n",
1184 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1185 list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1186 list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1187 spin_unlock_bh(&obd->obd_processing_task_lock);
1190 spin_unlock_bh(&obd->obd_processing_task_lock);
1192 /* when recovery finished, cleanup orphans on mds and ost */
1193 if (OBT(obd) && OBP(obd, postrecov)) {
1194 int rc = OBP(obd, postrecov)(obd);
1195 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1196 rc < 0 ? "failed" : "complete", rc);
1199 obd->obd_recovery_end = cfs_time_current_sec();
1203 static void abort_req_replay_queue(struct obd_device *obd)
1205 struct ptlrpc_request *req, *n;
1206 struct list_head abort_list;
1208 CFS_INIT_LIST_HEAD(&abort_list);
1209 spin_lock_bh(&obd->obd_processing_task_lock);
1210 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1211 spin_unlock_bh(&obd->obd_processing_task_lock);
1212 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1213 DEBUG_REQ(D_WARNING, req, "aborted:");
1214 req->rq_status = -ENOTCONN;
1215 if (ptlrpc_error(req)) {
1216 DEBUG_REQ(D_ERROR, req,
1217 "failed abort_req_reply; skipping");
1219 target_exp_dequeue_req_replay(req);
1220 ptlrpc_free_clone(req);
1224 static void abort_lock_replay_queue(struct obd_device *obd)
1226 struct ptlrpc_request *req, *n;
1227 struct list_head abort_list;
1229 CFS_INIT_LIST_HEAD(&abort_list);
1230 spin_lock_bh(&obd->obd_processing_task_lock);
1231 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1232 spin_unlock_bh(&obd->obd_processing_task_lock);
1233 list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1234 DEBUG_REQ(D_ERROR, req, "aborted:");
1235 req->rq_status = -ENOTCONN;
1236 if (ptlrpc_error(req)) {
1237 DEBUG_REQ(D_ERROR, req,
1238 "failed abort_lock_reply; skipping");
1240 ptlrpc_free_clone(req);
1245 /* Called from a cleanup function if the device is being cleaned up
1246 forcefully. The exports should all have been disconnected already,
1247 the only thing left to do is
1248 - clear the recovery flags
1250 - free queued requests and replies, but don't send replies
1251 Because the obd_stopping flag is set, no new requests should be received.
1254 void target_cleanup_recovery(struct obd_device *obd)
1256 struct ptlrpc_request *req, *n;
1257 struct list_head clean_list;
1260 LASSERT(obd->obd_stopping);
1262 CFS_INIT_LIST_HEAD(&clean_list);
1263 spin_lock_bh(&obd->obd_processing_task_lock);
1264 if (!obd->obd_recovering) {
1265 spin_unlock_bh(&obd->obd_processing_task_lock);
1269 obd->obd_recovering = obd->obd_abort_recovery = 0;
1270 target_cancel_recovery_timer(obd);
1272 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1273 spin_unlock_bh(&obd->obd_processing_task_lock);
1275 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1276 LASSERT(req->rq_reply_state == 0);
1277 target_exp_dequeue_req_replay(req);
1278 ptlrpc_free_clone(req);
1281 spin_lock_bh(&obd->obd_processing_task_lock);
1282 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1283 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1284 spin_unlock_bh(&obd->obd_processing_task_lock);
1286 list_for_each_entry_safe(req, n, &clean_list, rq_list){
1287 LASSERT(req->rq_reply_state == 0);
1288 ptlrpc_free_clone(req);
1294 /* obd_processing_task_lock should be held */
1295 void target_cancel_recovery_timer(struct obd_device *obd)
1297 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1298 cfs_timer_disarm(&obd->obd_recovery_timer);
1301 /* extend = 1 means require at least "duration" seconds left in the timer,
1302 extend = 0 means set the total duration (start_recovery_timer) */
1303 static void reset_recovery_timer(struct obd_device *obd, int duration,
1306 cfs_time_t now = cfs_time_current_sec();
1307 cfs_duration_t left;
1309 spin_lock_bh(&obd->obd_processing_task_lock);
1310 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1311 spin_unlock_bh(&obd->obd_processing_task_lock);
1315 left = cfs_time_sub(obd->obd_recovery_end, now);
1317 if (extend && (duration > left))
1318 obd->obd_recovery_timeout += duration - left;
1319 else if (!extend && (duration > obd->obd_recovery_timeout))
1320 /* Track the client's largest expected replay time */
1321 obd->obd_recovery_timeout = duration;
1324 * If total recovery time already exceed the
1325 * obd_recovery_max_time, then CRAY XT3 will
1326 * abort the recovery
1328 if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1329 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1331 obd->obd_recovery_end = obd->obd_recovery_start +
1332 obd->obd_recovery_timeout;
1333 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1334 cfs_time_before(now, obd->obd_recovery_end)) {
1335 left = cfs_time_sub(obd->obd_recovery_end, now);
1336 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1338 spin_unlock_bh(&obd->obd_processing_task_lock);
1339 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1340 obd->obd_name, (unsigned)left);
1343 static void check_and_start_recovery_timer(struct obd_device *obd)
1345 spin_lock_bh(&obd->obd_processing_task_lock);
1346 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1347 spin_unlock_bh(&obd->obd_processing_task_lock);
1350 CWARN("%s: starting recovery timer\n", obd->obd_name);
1351 obd->obd_recovery_start = cfs_time_current_sec();
1353 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1354 spin_unlock_bh(&obd->obd_processing_task_lock);
1356 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1359 /* Reset the timer with each new client connection */
1361 * This timer is actually reconnect_timer, which is for making sure
1362 * the total recovery window is at least as big as my reconnect
1363 * attempt timing. So the initial recovery time_out will be set to
1364 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1365 * from client is bigger than this, then the recovery time_out will
1366 * be extend to make sure the client could be reconnected, in the
1367 * process, the timeout from the new client should be ignored.
1371 target_start_and_reset_recovery_timer(struct obd_device *obd,
1372 struct ptlrpc_request *req,
1375 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1377 if (!new_client && service_time)
1378 /* Teach server about old server's estimates, as first guess
1379 * at how long new requests will take. */
1380 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1383 check_and_start_recovery_timer(obd);
1385 /* convert the service time to rpc timeout,
1386 * reuse service_time to limit stack usage */
1387 service_time = at_est2timeout(service_time);
1389 /* We expect other clients to timeout within service_time, then try
1390 * to reconnect, then try the failover server. The max delay between
1391 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1392 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1393 INITIAL_CONNECT_TIMEOUT);
1394 if (service_time > obd->obd_recovery_timeout && !new_client)
1395 reset_recovery_timer(obd, service_time, 0);
1399 static int check_for_next_transno(struct obd_device *obd)
1401 struct ptlrpc_request *req = NULL;
1402 int wake_up = 0, connected, completed, queue_len, max;
1403 __u64 next_transno, req_transno;
1405 spin_lock_bh(&obd->obd_processing_task_lock);
1407 if (!list_empty(&obd->obd_req_replay_queue)) {
1408 req = list_entry(obd->obd_req_replay_queue.next,
1409 struct ptlrpc_request, rq_list);
1410 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1415 max = obd->obd_max_recoverable_clients;
1416 connected = obd->obd_connected_clients;
1417 completed = connected - obd->obd_recoverable_clients;
1418 queue_len = obd->obd_requests_queued_for_recovery;
1419 next_transno = obd->obd_next_recovery_transno;
1421 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1422 "req_transno: "LPU64", next_transno: "LPU64"\n",
1423 max, connected, completed, queue_len, req_transno, next_transno);
1425 if (obd->obd_abort_recovery) {
1426 CDEBUG(D_HA, "waking for aborted recovery\n");
1428 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1429 CDEBUG(D_HA, "waking for completed recovery\n");
1431 } else if (req_transno == next_transno) {
1432 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1434 } else if (queue_len + completed == max) {
1435 /* handle gaps occured due to lost reply. It is allowed gaps
1436 * because all clients are connected and there will be resend
1437 * for missed transaction */
1438 LASSERTF(req_transno >= next_transno,
1439 "req_transno: "LPU64", next_transno: "LPU64"\n",
1440 req_transno, next_transno);
1442 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1443 "waking for skipped transno (skip: "LPD64
1444 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1445 next_transno, queue_len, completed, connected, req_transno);
1446 obd->obd_next_recovery_transno = req_transno;
1448 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1449 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1450 " by fail_lock, waking up ("LPD64")\n", next_transno);
1451 obd->obd_next_recovery_transno = req_transno;
1453 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1454 /* some clients haven't connected in time, but we can try
1455 * to replay requests that demand on already committed ones
1456 * also, we can replay first non-committed transation */
1457 LASSERT(req_transno != 0);
1458 if (req_transno == obd->obd_last_committed + 1) {
1459 obd->obd_next_recovery_transno = req_transno;
1460 } else if (req_transno > obd->obd_last_committed) {
1461 /* can't continue recovery: have no needed transno */
1462 obd->obd_abort_recovery = 1;
1463 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1464 "connected: %d, completed: %d, queue_len: %d, "
1465 "req_transno: "LPU64", next_transno: "LPU64"\n",
1466 max, connected, completed, queue_len,
1467 req_transno, next_transno);
1472 spin_unlock_bh(&obd->obd_processing_task_lock);
1476 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1478 struct l_wait_info lwi = { 0 };
1479 struct ptlrpc_request *req;
1481 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1482 obd->obd_next_recovery_transno);
1483 l_wait_event(obd->obd_next_transno_waitq,
1484 check_for_next_transno(obd), &lwi);
1486 spin_lock_bh(&obd->obd_processing_task_lock);
1487 if (obd->obd_abort_recovery) {
1489 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1490 req = list_entry(obd->obd_req_replay_queue.next,
1491 struct ptlrpc_request, rq_list);
1492 target_exp_dequeue_req_replay(req);
1493 list_del_init(&req->rq_list);
1494 obd->obd_requests_queued_for_recovery--;
1498 spin_unlock_bh(&obd->obd_processing_task_lock);
1502 static int check_for_next_lock(struct obd_device *obd)
1504 struct ptlrpc_request *req = NULL;
1507 spin_lock_bh(&obd->obd_processing_task_lock);
1508 if (!list_empty(&obd->obd_lock_replay_queue)) {
1509 req = list_entry(obd->obd_lock_replay_queue.next,
1510 struct ptlrpc_request, rq_list);
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");
1520 spin_unlock_bh(&obd->obd_processing_task_lock);
1525 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1527 struct l_wait_info lwi = { 0 };
1528 struct ptlrpc_request *req;
1530 CDEBUG(D_HA, "Waiting for lock\n");
1531 l_wait_event(obd->obd_next_transno_waitq,
1532 check_for_next_lock(obd), &lwi);
1534 spin_lock_bh(&obd->obd_processing_task_lock);
1535 if (obd->obd_abort_recovery) {
1537 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1538 req = list_entry(obd->obd_lock_replay_queue.next,
1539 struct ptlrpc_request, rq_list);
1540 list_del_init(&req->rq_list);
1544 spin_unlock_bh(&obd->obd_processing_task_lock);
1548 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1550 struct ptlrpc_request *req;
1552 spin_lock_bh(&obd->obd_processing_task_lock);
1553 if (!list_empty(&obd->obd_final_req_queue)) {
1554 req = list_entry(obd->obd_final_req_queue.next,
1555 struct ptlrpc_request, rq_list);
1556 list_del_init(&req->rq_list);
1560 spin_unlock_bh(&obd->obd_processing_task_lock);
1564 static inline int req_replay_done(struct obd_export *exp)
1566 return (exp->exp_req_replay_needed == 0);
1569 static inline int lock_replay_done(struct obd_export *exp)
1571 return (exp->exp_lock_replay_needed == 0);
1574 static inline int connect_done(struct obd_export *exp)
1576 return (exp->exp_in_recovery != 0);
1579 static int check_for_clients(struct obd_device *obd)
1581 if (obd->obd_abort_recovery)
1583 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1584 if (obd->obd_no_conn == 0 &&
1585 obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1590 static int handle_recovery_req(struct ptlrpc_thread *thread,
1591 struct ptlrpc_request *req,
1592 svc_handler_t handler)
1597 rc = lu_context_init(&req->rq_session, LCT_SESSION);
1599 CERROR("Failure to initialize session: %d\n", rc);
1602 req->rq_session.lc_thread = thread;
1603 lu_context_enter(&req->rq_session);
1604 req->rq_svc_thread = thread;
1605 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1607 /* thread context */
1608 lu_context_enter(&thread->t_env->le_ctx);
1610 lu_context_exit(&thread->t_env->le_ctx);
1612 lu_context_exit(&req->rq_session);
1613 lu_context_fini(&req->rq_session);
1614 /* don't reset timer for final stage */
1615 if (!req_replay_done(req->rq_export) ||
1616 !lock_replay_done(req->rq_export))
1617 reset_recovery_timer(class_exp2obd(req->rq_export),
1618 AT_OFF ? obd_timeout :
1619 at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1620 ptlrpc_free_clone(req);
1624 static void resume_recovery_timer(struct obd_device *obd)
1626 /* to be safe, make it at least OBD_RECOVERY_FACTOR * obd_timeout */
1627 reset_recovery_timer(obd, OBD_RECOVERY_FACTOR * obd_timeout, 1);
1630 static int target_recovery_thread(void *arg)
1632 struct obd_device *obd = arg;
1633 struct ptlrpc_request *req;
1634 struct target_recovery_data *trd = &obd->obd_recovery_data;
1635 struct l_wait_info lwi = { 0 };
1636 unsigned long delta;
1637 unsigned long flags;
1639 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1643 cfs_daemonize("tgt_recov");
1645 SIGNAL_MASK_LOCK(current, flags);
1646 sigfillset(¤t->blocked);
1648 SIGNAL_MASK_UNLOCK(current, flags);
1650 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1654 thread->t_env = &env;
1655 env.le_ctx.lc_thread = thread;
1657 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1659 trd->trd_processing_task = current->pid;
1661 obd->obd_recovering = 1;
1662 complete(&trd->trd_starting);
1664 /* first of all, we have to know the first transno to replay */
1665 obd->obd_abort_recovery = 0;
1666 l_wait_event(obd->obd_next_transno_waitq,
1667 check_for_clients(obd), &lwi);
1669 spin_lock_bh(&obd->obd_processing_task_lock);
1670 target_cancel_recovery_timer(obd);
1671 spin_unlock_bh(&obd->obd_processing_task_lock);
1673 /* If some clients haven't connected in time, evict them */
1674 if (obd->obd_abort_recovery) {
1675 CWARN("Some clients haven't connect in time (%d/%d),"
1676 "evict them\n", obd->obd_connected_clients,
1677 obd->obd_max_recoverable_clients);
1678 obd->obd_abort_recovery = obd->obd_stopping;
1679 class_disconnect_stale_exports(obd, connect_done);
1681 /* next stage: replay requests */
1683 obd->obd_req_replaying = 1;
1684 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1685 atomic_read(&obd->obd_req_replay_clients),
1686 obd->obd_next_recovery_transno);
1687 resume_recovery_timer(obd);
1688 while ((req = target_next_replay_req(obd))) {
1689 LASSERT(trd->trd_processing_task == current->pid);
1690 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1691 lustre_msg_get_transno(req->rq_reqmsg),
1692 libcfs_nid2str(req->rq_peer.nid));
1693 handle_recovery_req(thread, req,
1694 trd->trd_recovery_handler);
1695 obd->obd_replayed_requests++;
1696 spin_lock_bh(&obd->obd_processing_task_lock);
1697 obd->obd_next_recovery_transno++;
1698 spin_unlock_bh(&obd->obd_processing_task_lock);
1701 spin_lock_bh(&obd->obd_processing_task_lock);
1702 target_cancel_recovery_timer(obd);
1703 spin_unlock_bh(&obd->obd_processing_task_lock);
1705 /* If some clients haven't replayed requests in time, evict them */
1706 if (obd->obd_abort_recovery) {
1707 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1708 obd->obd_abort_recovery = obd->obd_stopping;
1709 class_disconnect_stale_exports(obd, req_replay_done);
1710 abort_req_replay_queue(obd);
1713 /* The second stage: replay locks */
1714 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1715 atomic_read(&obd->obd_lock_replay_clients));
1716 resume_recovery_timer(obd);
1717 while ((req = target_next_replay_lock(obd))) {
1718 LASSERT(trd->trd_processing_task == current->pid);
1719 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1720 libcfs_nid2str(req->rq_peer.nid));
1721 handle_recovery_req(thread, req,
1722 trd->trd_recovery_handler);
1723 obd->obd_replayed_locks++;
1726 spin_lock_bh(&obd->obd_processing_task_lock);
1727 target_cancel_recovery_timer(obd);
1728 spin_unlock_bh(&obd->obd_processing_task_lock);
1729 /* If some clients haven't replayed requests in time, evict them */
1730 if (obd->obd_abort_recovery) {
1732 CERROR("lock replay timed out, aborting ...\n");
1733 obd->obd_abort_recovery = obd->obd_stopping;
1734 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1735 abort_lock_replay_queue(obd);
1738 /* We drop recoverying flag to forward all new requests
1739 * to regular mds_handle() since now */
1740 spin_lock_bh(&obd->obd_processing_task_lock);
1741 obd->obd_recovering = obd->obd_abort_recovery = 0;
1742 spin_unlock_bh(&obd->obd_processing_task_lock);
1743 /* The third stage: reply on final pings */
1744 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1745 while ((req = target_next_final_ping(obd))) {
1746 LASSERT(trd->trd_processing_task == current->pid);
1747 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1748 libcfs_nid2str(req->rq_peer.nid));
1749 handle_recovery_req(thread, req,
1750 trd->trd_recovery_handler);
1753 delta = (jiffies - delta) / HZ;
1754 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1755 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1756 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1757 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1758 if (delta > obd_timeout * 2) {
1759 CWARN("too long recovery - read logs\n");
1760 libcfs_debug_dumplog();
1763 target_finish_recovery(obd);
1765 lu_context_fini(&env.le_ctx);
1766 trd->trd_processing_task = 0;
1767 complete(&trd->trd_finishing);
1771 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1774 struct target_recovery_data *trd = &obd->obd_recovery_data;
1776 memset(trd, 0, sizeof(*trd));
1777 init_completion(&trd->trd_starting);
1778 init_completion(&trd->trd_finishing);
1779 trd->trd_recovery_handler = handler;
1781 if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1782 wait_for_completion(&trd->trd_starting);
1783 LASSERT(obd->obd_recovering != 0);
1790 void target_stop_recovery_thread(struct obd_device *obd)
1792 spin_lock_bh(&obd->obd_processing_task_lock);
1793 if (obd->obd_recovery_data.trd_processing_task > 0) {
1794 struct target_recovery_data *trd = &obd->obd_recovery_data;
1795 CERROR("%s: Aborting recovery\n", obd->obd_name);
1796 obd->obd_abort_recovery = 1;
1797 wake_up(&obd->obd_next_transno_waitq);
1798 spin_unlock_bh(&obd->obd_processing_task_lock);
1799 wait_for_completion(&trd->trd_finishing);
1801 spin_unlock_bh(&obd->obd_processing_task_lock);
1805 void target_recovery_fini(struct obd_device *obd)
1807 class_disconnect_exports(obd);
1808 target_stop_recovery_thread(obd);
1809 target_cleanup_recovery(obd);
1811 EXPORT_SYMBOL(target_recovery_fini);
1813 static void target_recovery_expired(unsigned long castmeharder)
1815 struct obd_device *obd = (struct obd_device *)castmeharder;
1816 LCONSOLE_WARN("%s: recovery timed out; %d clients never reconnected "
1817 "after %lds (%d clients did)\n",
1818 obd->obd_name, obd->obd_recoverable_clients,
1819 cfs_time_current_sec()- obd->obd_recovery_start,
1820 obd->obd_connected_clients);
1821 spin_lock_bh(&obd->obd_processing_task_lock);
1822 if (obd->obd_recovering)
1823 obd->obd_abort_recovery = 1;
1824 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1825 spin_unlock_bh(&obd->obd_processing_task_lock);
1828 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1830 if (obd->obd_max_recoverable_clients == 0)
1833 CWARN("RECOVERY: service %s, %d recoverable clients, "
1834 "last_transno "LPU64"\n", obd->obd_name,
1835 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1836 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1837 obd->obd_recovery_start = 0;
1838 obd->obd_recovery_end = 0;
1839 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1840 /* bz13079: this should be set to desired value for ost but not for mds */
1841 obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1842 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1843 target_start_recovery_thread(obd, handler);
1845 EXPORT_SYMBOL(target_recovery_init);
1849 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1851 struct obd_export *exp = req->rq_export;
1852 LASSERT(exp != NULL);
1853 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1854 /* client declares he's ready to replay locks */
1855 spin_lock_bh(&obd->obd_processing_task_lock);
1856 if (exp->exp_req_replay_needed) {
1857 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1858 spin_lock(&exp->exp_lock);
1859 exp->exp_req_replay_needed = 0;
1860 spin_unlock(&exp->exp_lock);
1861 atomic_dec(&obd->obd_req_replay_clients);
1862 LASSERT(obd->obd_recoverable_clients > 0);
1863 obd->obd_recoverable_clients--;
1864 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1865 CDEBUG(D_HA, "all clients have replayed reqs\n");
1866 wake_up(&obd->obd_next_transno_waitq);
1868 spin_unlock_bh(&obd->obd_processing_task_lock);
1870 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1871 /* client declares he's ready to complete recovery
1872 * so, we put the request on th final queue */
1873 spin_lock_bh(&obd->obd_processing_task_lock);
1874 if (exp->exp_lock_replay_needed) {
1875 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1876 spin_lock(&exp->exp_lock);
1877 exp->exp_lock_replay_needed = 0;
1878 spin_unlock(&exp->exp_lock);
1879 atomic_dec(&obd->obd_lock_replay_clients);
1880 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1881 CDEBUG(D_HA, "all clients have replayed locks\n");
1882 wake_up(&obd->obd_next_transno_waitq);
1884 spin_unlock_bh(&obd->obd_processing_task_lock);
1890 int target_queue_recovery_request(struct ptlrpc_request *req,
1891 struct obd_device *obd)
1893 struct list_head *tmp;
1895 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1899 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1900 /* Processing the queue right now, don't re-add. */
1904 target_process_req_flags(obd, req);
1906 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1907 /* client declares he's ready to complete recovery
1908 * so, we put the request on th final queue */
1909 req = ptlrpc_clone_req(req);
1912 DEBUG_REQ(D_HA, req, "queue final req");
1913 spin_lock_bh(&obd->obd_processing_task_lock);
1914 if (obd->obd_recovering)
1915 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1917 spin_unlock_bh(&obd->obd_processing_task_lock);
1918 ptlrpc_free_clone(req);
1919 if (obd->obd_stopping) {
1925 spin_unlock_bh(&obd->obd_processing_task_lock);
1928 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1929 /* client declares he's ready to replay locks */
1930 req = ptlrpc_clone_req(req);
1933 DEBUG_REQ(D_HA, req, "queue lock replay req");
1934 spin_lock_bh(&obd->obd_processing_task_lock);
1935 LASSERT(obd->obd_recovering);
1936 /* usually due to recovery abort */
1937 if (!req->rq_export->exp_in_recovery) {
1938 spin_unlock_bh(&obd->obd_processing_task_lock);
1939 ptlrpc_free_clone(req);
1942 LASSERT(req->rq_export->exp_lock_replay_needed);
1943 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1944 spin_unlock_bh(&obd->obd_processing_task_lock);
1945 wake_up(&obd->obd_next_transno_waitq);
1949 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1950 * (i.e. buflens etc are in my own byte order), but type-dependent
1951 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1954 CFS_INIT_LIST_HEAD(&req->rq_list);
1955 DEBUG_REQ(D_HA, req, "not queueing");
1959 spin_lock_bh(&obd->obd_processing_task_lock);
1961 /* If we're processing the queue, we want don't want to queue this
1964 * Also, if this request has a transno less than the one we're waiting
1965 * for, we should process it now. It could (and currently always will)
1966 * be an open request for a descriptor that was opened some time ago.
1968 * Also, a resent, replayed request that has already been
1969 * handled will pass through here and be processed immediately.
1971 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1972 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1973 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1974 /* Processing the queue right now, don't re-add. */
1975 LASSERT(list_empty(&req->rq_list));
1976 spin_unlock_bh(&obd->obd_processing_task_lock);
1979 spin_unlock_bh(&obd->obd_processing_task_lock);
1981 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
1984 req = ptlrpc_clone_req(req);
1988 spin_lock_bh(&obd->obd_processing_task_lock);
1989 LASSERT(obd->obd_recovering);
1990 if (!req->rq_export->exp_in_recovery) {
1991 spin_unlock_bh(&obd->obd_processing_task_lock);
1992 ptlrpc_free_clone(req);
1995 LASSERT(req->rq_export->exp_req_replay_needed);
1997 if (target_exp_enqueue_req_replay(req)) {
1998 spin_unlock_bh(&obd->obd_processing_task_lock);
1999 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2000 ptlrpc_free_clone(req);
2005 list_for_each(tmp, &obd->obd_req_replay_queue) {
2006 struct ptlrpc_request *reqiter =
2007 list_entry(tmp, struct ptlrpc_request, rq_list);
2009 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2010 list_add_tail(&req->rq_list, &reqiter->rq_list);
2015 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2017 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2018 "has been claimed by another client");
2019 spin_unlock_bh(&obd->obd_processing_task_lock);
2020 target_exp_dequeue_req_replay(req);
2021 ptlrpc_free_clone(req);
2027 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2029 obd->obd_requests_queued_for_recovery++;
2030 wake_up(&obd->obd_next_transno_waitq);
2031 spin_unlock_bh(&obd->obd_processing_task_lock);
2035 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2037 return req->rq_export->exp_obd;
2040 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2042 LASSERT(exp != NULL);
2043 return &exp->exp_obd->obd_namespace->ns_pool;
2047 * Packs current SLV and Limit into \a req.
2049 int target_pack_pool_reply(struct ptlrpc_request *req)
2051 struct obd_device *obd;
2055 * Check that we still have all structures alive as this may
2056 * be some late rpc in shutdown time.
2058 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2059 !exp_connect_lru_resize(req->rq_export))) {
2060 lustre_msg_set_slv(req->rq_repmsg, 0);
2061 lustre_msg_set_limit(req->rq_repmsg, 0);
2066 * OBD is alive here as export is alive, which we checked above.
2068 obd = req->rq_export->exp_obd;
2070 read_lock(&obd->obd_pool_lock);
2071 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2072 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2073 read_unlock(&obd->obd_pool_lock);
2078 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2080 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2081 DEBUG_REQ(D_ERROR, req, "dropping reply");
2086 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2087 req->rq_status = rc;
2088 return (ptlrpc_send_error(req, 1));
2090 DEBUG_REQ(D_NET, req, "sending reply");
2093 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2096 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2099 struct ptlrpc_reply_state *rs;
2100 struct obd_device *obd;
2101 struct obd_export *exp;
2102 struct ptlrpc_service *svc;
2105 if (req->rq_no_reply) {
2110 svc = req->rq_rqbd->rqbd_service;
2111 rs = req->rq_reply_state;
2112 if (rs == NULL || !rs->rs_difficult) {
2114 target_send_reply_msg (req, rc, fail_id);
2119 /* must be an export if locks saved */
2120 LASSERT (req->rq_export != NULL);
2121 /* req/reply consistent */
2122 LASSERT (rs->rs_service == svc);
2125 LASSERT (!rs->rs_scheduled);
2126 LASSERT (!rs->rs_scheduled_ever);
2127 LASSERT (!rs->rs_handled);
2128 LASSERT (!rs->rs_on_net);
2129 LASSERT (rs->rs_export == NULL);
2130 LASSERT (list_empty(&rs->rs_obd_list));
2131 LASSERT (list_empty(&rs->rs_exp_list));
2133 exp = class_export_get (req->rq_export);
2136 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2137 rs->rs_scheduled = 1;
2139 rs->rs_xid = req->rq_xid;
2140 rs->rs_transno = req->rq_transno;
2141 rs->rs_export = exp;
2143 spin_lock(&obd->obd_uncommitted_replies_lock);
2145 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2146 rs->rs_transno, obd->obd_last_committed);
2147 if (rs->rs_transno > obd->obd_last_committed) {
2148 /* not committed already */
2149 list_add_tail (&rs->rs_obd_list,
2150 &obd->obd_uncommitted_replies);
2153 spin_unlock (&obd->obd_uncommitted_replies_lock);
2154 spin_lock (&exp->exp_lock);
2156 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
2158 spin_unlock(&exp->exp_lock);
2160 netrc = target_send_reply_msg (req, rc, fail_id);
2162 spin_lock(&svc->srv_lock);
2164 atomic_inc(&svc->srv_n_difficult_replies);
2167 /* error sending: reply is off the net. Also we need +1
2168 * reply ref until ptlrpc_server_handle_reply() is done
2169 * with the reply state (if the send was successful, there
2170 * would have been +1 ref for the net, which
2171 * reply_out_callback leaves alone) */
2173 ptlrpc_rs_addref(rs);
2174 atomic_inc (&svc->srv_outstanding_replies);
2177 spin_lock(&rs->rs_lock);
2178 if (rs->rs_transno <= obd->obd_last_committed ||
2179 (!rs->rs_on_net && !rs->rs_no_ack) ||
2180 list_empty(&rs->rs_exp_list) || /* completed already */
2181 list_empty(&rs->rs_obd_list)) {
2182 CDEBUG(D_HA, "Schedule reply immediately\n");
2183 ptlrpc_dispatch_difficult_reply(rs);
2185 list_add (&rs->rs_list, &svc->srv_active_replies);
2186 rs->rs_scheduled = 0; /* allow notifier to schedule */
2188 spin_unlock(&rs->rs_lock);
2189 spin_unlock(&svc->srv_lock);
2193 int target_handle_ping(struct ptlrpc_request *req)
2195 obd_ping(req->rq_export);
2196 return req_capsule_server_pack(&req->rq_pill);
2199 void target_committed_to_req(struct ptlrpc_request *req)
2201 struct obd_device *obd;
2203 if (req == NULL || req->rq_export == NULL)
2206 obd = req->rq_export->exp_obd;
2210 if (!obd->obd_no_transno && req->rq_repmsg != NULL)
2211 lustre_msg_set_last_committed(req->rq_repmsg,
2212 obd->obd_last_committed);
2214 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2215 "%d)", obd->obd_no_transno, req->rq_repmsg == NULL);
2217 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2218 obd->obd_last_committed, req->rq_transno, req->rq_xid);
2221 EXPORT_SYMBOL(target_committed_to_req);
2223 int target_handle_qc_callback(struct ptlrpc_request *req)
2225 struct obd_quotactl *oqctl;
2226 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2228 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2229 if (oqctl == NULL) {
2230 CERROR("Can't unpack obd_quotactl\n");
2234 cli->cl_qchk_stat = oqctl->qc_stat;
2239 #ifdef HAVE_QUOTA_SUPPORT
2240 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2243 struct obd_device *obd = req->rq_export->exp_obd;
2244 struct obd_device *master_obd;
2245 struct obd_device_target *obt;
2246 struct lustre_quota_ctxt *qctxt;
2247 struct qunit_data *qdata = NULL;
2251 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2254 rc = req_capsule_server_pack(&req->rq_pill);
2256 CERROR("packing reply failed!: rc = %d\n", rc);
2260 LASSERT(req->rq_export);
2262 OBD_ALLOC(qdata, sizeof(struct qunit_data));
2265 rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
2267 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2271 /* we use the observer */
2272 if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2273 CERROR("Can't find the observer, it is recovering\n");
2274 req->rq_status = -EAGAIN;
2275 GOTO(send_reply, rc = -EAGAIN);
2278 master_obd = obd->obd_observer->obd_observer;
2279 obt = &master_obd->u.obt;
2280 qctxt = &obt->obt_qctxt;
2282 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2283 /* quota_type has not been processed yet, return EAGAIN
2284 * until we know whether or not quotas are supposed to
2286 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2288 req->rq_status = -EAGAIN;
2289 rc = ptlrpc_reply(req);
2293 down_read(&obt->obt_rwsem);
2294 if (qctxt->lqc_lqs_hash == NULL) {
2295 up_read(&obt->obt_rwsem);
2296 /* quota_type has not been processed yet, return EAGAIN
2297 * until we know whether or not quotas are supposed to
2299 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2301 req->rq_status = -EAGAIN;
2302 rc = ptlrpc_reply(req);
2306 LASSERT(qctxt->lqc_handler);
2307 rc = qctxt->lqc_handler(master_obd, qdata,
2308 lustre_msg_get_opc(req->rq_reqmsg));
2309 up_read(&obt->obt_rwsem);
2310 if (rc && rc != -EDQUOT)
2311 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2312 "dqacq failed! (rc:%d)\n", rc);
2313 req->rq_status = rc;
2315 /* there are three forms of qunit(historic causes), so we need to
2316 * adjust the same form to different forms slaves needed */
2317 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2319 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2323 /* Block the quota req. b=14840 */
2324 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2326 rc = ptlrpc_reply(req);
2328 OBD_FREE(qdata, sizeof(struct qunit_data));
2332 #endif /* !__KERNEL__ */
2334 #endif /* HAVE_QUOTA_SUPPORT */
2336 ldlm_mode_t lck_compat_array[] = {
2337 [LCK_EX] LCK_COMPAT_EX,
2338 [LCK_PW] LCK_COMPAT_PW,
2339 [LCK_PR] LCK_COMPAT_PR,
2340 [LCK_CW] LCK_COMPAT_CW,
2341 [LCK_CR] LCK_COMPAT_CR,
2342 [LCK_NL] LCK_COMPAT_NL,
2343 [LCK_GROUP] LCK_COMPAT_GROUP,
2344 [LCK_COS] LCK_COMPAT_COS,
2348 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2349 * not escape to the user level.
2351 int ldlm_error2errno(ldlm_error_t error)
2359 case ELDLM_LOCK_CHANGED:
2362 case ELDLM_LOCK_ABORTED:
2365 case ELDLM_LOCK_REPLACED:
2368 case ELDLM_NO_LOCK_DATA:
2371 case ELDLM_NAMESPACE_EXISTS:
2374 case ELDLM_BAD_NAMESPACE:
2378 if (((int)error) < 0) /* cast to signed type */
2379 result = error; /* as ldlm_error_t can be unsigned */
2381 CERROR("Invalid DLM result code: %i\n", error);
2387 EXPORT_SYMBOL(ldlm_error2errno);
2390 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2392 ldlm_error_t ldlm_errno2error(int err_no)
2401 error = ELDLM_LOCK_CHANGED;
2404 error = ELDLM_LOCK_ABORTED;
2407 error = ELDLM_LOCK_REPLACED;
2410 error = ELDLM_NO_LOCK_DATA;
2413 error = ELDLM_NAMESPACE_EXISTS;
2416 error = ELDLM_BAD_NAMESPACE;
2423 EXPORT_SYMBOL(ldlm_errno2error);