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;
517 * there's no necessary to hold sem during diconnecting an import,
518 * and actually it may cause deadlock in gss.
520 up_write(&cli->cl_sem);
521 rc = ptlrpc_disconnect_import(imp, 0);
522 down_write(&cli->cl_sem);
524 ptlrpc_invalidate_import(imp);
525 /* set obd_namespace to NULL only after invalidate, because we can have
526 * some connect requests in flight, and his need store a connect flags
527 * in obd_namespace. bug 14260 */
528 obd->obd_namespace = NULL;
530 if (imp->imp_rq_pool) {
531 ptlrpc_free_rq_pool(imp->imp_rq_pool);
532 imp->imp_rq_pool = NULL;
535 cli->cl_import = NULL;
539 err = class_disconnect(exp);
543 up_write(&cli->cl_sem);
545 ldlm_namespace_free_post(to_be_freed);
550 /* --------------------------------------------------------------------------
551 * from old lib/target.c
552 * -------------------------------------------------------------------------- */
554 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
555 struct obd_uuid *cluuid, int mds_conn)
558 if (exp->exp_connection && exp->exp_imp_reverse && !mds_conn) {
559 struct lustre_handle *hdl;
560 hdl = &exp->exp_imp_reverse->imp_remote_handle;
561 /* Might be a re-connect after a partition. */
562 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
563 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
565 conn->cookie = exp->exp_handle.h_cookie;
566 /* target_handle_connect() treats EALREADY and
567 * -EALREADY differently. EALREADY means we are
568 * doing a valid reconnect from the same client. */
571 CERROR("%s reconnecting from %s, "
572 "handle mismatch (ours "LPX64", theirs "
573 LPX64")\n", cluuid->uuid,
574 exp->exp_connection->c_remote_uuid.uuid,
575 hdl->cookie, conn->cookie);
576 memset(conn, 0, sizeof *conn);
577 /* target_handle_connect() treats EALREADY and
578 * -EALREADY differently. -EALREADY is an error
579 * (same UUID, different handle). */
584 conn->cookie = exp->exp_handle.h_cookie;
585 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
586 cluuid->uuid, exp, conn->cookie);
590 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
593 struct obd_export *exp = cb_data;
595 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
596 obd->obd_name, exp->exp_client_uuid.uuid);
598 spin_lock(&exp->exp_lock);
599 exp->exp_need_sync = 0;
600 spin_unlock(&exp->exp_lock);
602 EXPORT_SYMBOL(target_client_add_cb);
604 target_start_and_reset_recovery_timer(struct obd_device *obd,
605 struct ptlrpc_request *req,
608 int target_handle_connect(struct ptlrpc_request *req)
610 struct obd_device *target, *targref = NULL;
611 struct obd_export *export = NULL;
612 struct obd_import *revimp;
613 struct lustre_handle conn;
614 struct lustre_handle *tmp;
615 struct obd_uuid tgtuuid;
616 struct obd_uuid cluuid;
617 struct obd_uuid remote_uuid;
621 struct obd_connect_data *data, *tmpdata;
622 lnet_nid_t *client_nid = NULL;
625 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
627 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
629 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
630 GOTO(out, rc = -EINVAL);
633 obd_str2uuid(&tgtuuid, str);
634 target = class_uuid2obd(&tgtuuid);
636 target = class_name2obd(str);
638 if (!target || target->obd_stopping || !target->obd_set_up) {
639 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
640 " for connect (%s)\n", str,
641 !target ? "no target" :
642 (target->obd_stopping ? "stopping" :
644 GOTO(out, rc = -ENODEV);
647 if (target->obd_no_conn) {
648 LCONSOLE_WARN("%s: temporarily refusing client connection "
649 "from %s\n", target->obd_name,
650 libcfs_nid2str(req->rq_peer.nid));
651 GOTO(out, rc = -EAGAIN);
654 /* Make sure the target isn't cleaned up while we're here. Yes,
655 there's still a race between the above check and our incref here.
656 Really, class_uuid2obd should take the ref. */
657 targref = class_incref(target, __FUNCTION__, cfs_current());
660 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
662 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
663 GOTO(out, rc = -EINVAL);
666 obd_str2uuid(&cluuid, str);
668 /* XXX extract a nettype and format accordingly */
669 switch (sizeof(lnet_nid_t)) {
670 /* NB the casts only avoid compiler warnings */
672 snprintf(remote_uuid.uuid, sizeof remote_uuid,
673 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
676 snprintf(remote_uuid.uuid, sizeof remote_uuid,
677 "NET_%x_UUID", (__u32)req->rq_peer.nid);
683 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
685 GOTO(out, rc = -EPROTO);
689 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
691 GOTO(out, rc = -EPROTO);
693 rc = req_capsule_server_pack(&req->rq_pill);
697 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
699 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
700 "libclient connection attempt");
701 GOTO(out, rc = -EPROTO);
702 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
703 LUSTRE_VERSION_ALLOWED_OFFSET ||
704 data->ocd_version > LUSTRE_VERSION_CODE +
705 LUSTRE_VERSION_ALLOWED_OFFSET) {
706 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
707 "libclient connection attempt",
708 data->ocd_version < LUSTRE_VERSION_CODE ?
710 OBD_OCD_VERSION_MAJOR(data->ocd_version),
711 OBD_OCD_VERSION_MINOR(data->ocd_version),
712 OBD_OCD_VERSION_PATCH(data->ocd_version),
713 OBD_OCD_VERSION_FIX(data->ocd_version));
714 data = req_capsule_server_sized_get(&req->rq_pill,
716 offsetof(typeof(*data), ocd_version) +
717 sizeof(data->ocd_version));
719 data->ocd_connect_flags = OBD_CONNECT_VERSION;
720 data->ocd_version = LUSTRE_VERSION_CODE;
722 GOTO(out, rc = -EPROTO);
726 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
727 (data->ocd_connect_flags & OBD_CONNECT_MDS))
730 /* lctl gets a backstage, all-access pass. */
731 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
732 goto dont_check_exports;
734 spin_lock(&target->obd_dev_lock);
735 export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
737 if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
738 CWARN("%s: exp %p already connecting\n",
739 export->exp_obd->obd_name, export);
740 class_export_put(export);
743 } else if (export != NULL && export->exp_connection != NULL &&
744 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
746 /* in mds failover we have static uuid but nid can be
748 CWARN("%s: cookie %s seen on new NID %s when "
749 "existing NID %s is already connected\n",
750 target->obd_name, cluuid.uuid,
751 libcfs_nid2str(req->rq_peer.nid),
752 libcfs_nid2str(export->exp_connection->c_peer.nid));
754 class_export_put(export);
756 } else if (export != NULL) {
757 spin_lock(&export->exp_lock);
758 export->exp_connecting = 1;
759 spin_unlock(&export->exp_lock);
760 class_export_put(export);
761 spin_unlock(&target->obd_dev_lock);
762 LASSERT(export->exp_obd == target);
764 rc = target_handle_reconnect(&conn, export, &cluuid, mds_conn);
767 /* If we found an export, we already unlocked. */
769 spin_unlock(&target->obd_dev_lock);
770 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
771 } else if (req->rq_export == NULL &&
772 atomic_read(&export->exp_rpc_count) > 0) {
773 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
774 target->obd_name, cluuid.uuid,
775 libcfs_nid2str(req->rq_peer.nid),
776 export, atomic_read(&export->exp_refcount));
777 GOTO(out, rc = -EBUSY);
778 } else if (req->rq_export != NULL &&
779 (atomic_read(&export->exp_rpc_count) > 1)) {
780 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
781 target->obd_name, cluuid.uuid,
782 libcfs_nid2str(req->rq_peer.nid),
783 export, atomic_read(&export->exp_rpc_count));
784 GOTO(out, rc = -EBUSY);
785 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
787 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
788 "cookies not random?\n", target->obd_name,
789 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
790 GOTO(out, rc = -EALREADY);
792 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
793 if (req->rq_export == NULL && mds_conn)
794 export->exp_last_request_time =
795 max(export->exp_last_request_time,
796 (time_t)cfs_time_current_sec());
803 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
804 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
805 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
806 export, (long)cfs_time_current_sec(),
807 export ? (long)export->exp_last_request_time : 0);
809 /* Tell the client if we're in recovery. */
810 if (target->obd_recovering) {
811 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
812 /* If this is the first time a client connects,
813 reset the recovery timer */
815 target_start_and_reset_recovery_timer(target, req,
819 /* We want to handle EALREADY but *not* -EALREADY from
820 * target_handle_reconnect(), return reconnection state in a flag */
821 if (rc == EALREADY) {
822 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
828 /* Tell the client if we support replayable requests */
829 if (target->obd_replayable)
830 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
831 client_nid = &req->rq_peer.nid;
833 if (export == NULL) {
834 if (target->obd_recovering) {
837 t = cfs_timer_deadline(&target->obd_recovery_timer);
838 t = cfs_time_sub(t, cfs_time_current());
839 CERROR("%s: denying connection for new client %s (%s): "
840 "%d clients in recovery for "CFS_TIME_T"s\n",
842 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
843 target->obd_recoverable_clients,
844 cfs_duration_sec(t));
848 rc = obd_connect(req->rq_svc_thread->t_env,
849 &conn, target, &cluuid, data,
853 rc = obd_reconnect(req->rq_svc_thread->t_env,
854 export, target, &cluuid, data, client_nid);
858 /* Return only the parts of obd_connect_data that we understand, so the
859 * client knows that we don't understand the rest. */
861 tmpdata = req_capsule_server_get(&req->rq_pill,
863 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
867 /* If all else goes well, this is our RPC return code. */
870 lustre_msg_set_handle(req->rq_repmsg, &conn);
872 /* ownership of this export ref transfers to the request AFTER we
873 * drop any previous reference the request had, but we don't want
874 * that to go to zero before we get our new export reference. */
875 export = class_conn2export(&conn);
877 DEBUG_REQ(D_ERROR, req, "Missing export!");
878 GOTO(out, rc = -ENODEV);
881 /* If the client and the server are the same node, we will already
882 * have an export that really points to the client's DLM export,
883 * because we have a shared handles table.
885 * XXX this will go away when shaver stops sending the "connect" handle
886 * in the real "remote handle" field of the request --phik 24 Apr 2003
888 if (req->rq_export != NULL)
889 class_export_put(req->rq_export);
891 req->rq_export = export;
893 spin_lock(&export->exp_lock);
895 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
896 } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
897 spin_unlock(&export->exp_lock);
898 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
899 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
900 export->exp_conn_cnt,
901 lustre_msg_get_conn_cnt(req->rq_reqmsg));
903 GOTO(out, rc = -EALREADY);
905 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
907 /* request from liblustre? Don't evict it for not pinging. */
908 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
909 export->exp_libclient = 1;
910 spin_unlock(&export->exp_lock);
912 spin_lock(&target->obd_dev_lock);
913 list_del_init(&export->exp_obd_chain_timed);
914 spin_unlock(&target->obd_dev_lock);
916 spin_unlock(&export->exp_lock);
919 if (export->exp_connection != NULL)
920 ptlrpc_connection_put(export->exp_connection);
921 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
924 if (hlist_unhashed(&export->exp_nid_hash)) {
925 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
926 &export->exp_connection->c_peer.nid,
927 &export->exp_nid_hash);
930 spin_lock_bh(&target->obd_processing_task_lock);
931 if (target->obd_recovering && !export->exp_in_recovery) {
932 spin_lock(&export->exp_lock);
933 export->exp_in_recovery = 1;
934 export->exp_req_replay_needed = 1;
935 export->exp_lock_replay_needed = 1;
936 spin_unlock(&export->exp_lock);
937 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
938 && (data->ocd_transno == 0))
939 CWARN("Connect with zero transno!\n");
941 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
942 && data->ocd_transno < target->obd_next_recovery_transno)
943 target->obd_next_recovery_transno = data->ocd_transno;
944 target->obd_connected_clients++;
945 /* each connected client is counted as recoverable */
946 target->obd_recoverable_clients++;
947 atomic_inc(&target->obd_req_replay_clients);
948 atomic_inc(&target->obd_lock_replay_clients);
949 if (target->obd_connected_clients ==
950 target->obd_max_recoverable_clients)
951 wake_up(&target->obd_next_transno_waitq);
953 spin_unlock_bh(&target->obd_processing_task_lock);
954 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
957 if (export->exp_imp_reverse != NULL) {
958 /* destroyed import can be still referenced in ctxt */
959 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
960 KEY_REVIMP_UPD, 0, NULL, NULL);
962 /* in some recovery senarios, previous ctx init rpc handled
963 * in sptlrpc_target_export_check() might be used to install
964 * a reverse ctx in this reverse import, and later OBD_CONNECT
965 * using the same gss ctx could reach here and following new
966 * reverse import. note all reverse ctx in new/old import are
967 * actually based on the same gss ctx. so we invalidate ctx
968 * here before destroy import, otherwise flush old import will
969 * lead to remote reverse ctx be destroied, thus the reverse
970 * ctx of new import will lost its peer.
971 * there might be a better way to deal with this???
973 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
975 destroy_import(export->exp_imp_reverse);
978 /* for the rest part, we return -ENOTCONN in case of errors
979 * in order to let client initialize connection again.
981 revimp = export->exp_imp_reverse = class_new_import(target);
983 CERROR("fail to alloc new reverse import.\n");
984 GOTO(out, rc = -ENOTCONN);
987 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
988 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
989 revimp->imp_remote_handle = conn;
990 revimp->imp_dlm_fake = 1;
991 revimp->imp_state = LUSTRE_IMP_FULL;
993 /* unknown versions will be caught in
994 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
995 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
997 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
998 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
999 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1001 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1003 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1005 CERROR("Failed to get sec for reverse import: %d\n", rc);
1006 export->exp_imp_reverse = NULL;
1007 class_destroy_import(revimp);
1010 class_import_put(revimp);
1013 spin_lock(&export->exp_lock);
1014 export->exp_connecting = 0;
1015 spin_unlock(&export->exp_lock);
1018 class_decref(targref, __FUNCTION__, cfs_current());
1020 req->rq_status = rc;
1024 int target_handle_disconnect(struct ptlrpc_request *req)
1029 rc = req_capsule_server_pack(&req->rq_pill);
1033 /* keep the rq_export around so we can send the reply */
1034 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1039 void target_destroy_export(struct obd_export *exp)
1041 /* exports created from last_rcvd data, and "fake"
1042 exports created by lctl don't have an import */
1043 if (exp->exp_imp_reverse != NULL)
1044 destroy_import(exp->exp_imp_reverse);
1046 /* We cancel locks at disconnect time, but this will catch any locks
1047 * granted in a race with recovery-induced disconnect. */
1048 if (exp->exp_obd->obd_namespace != NULL)
1049 ldlm_cancel_locks_for_export(exp);
1053 * Recovery functions
1056 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1058 struct ptlrpc_request *copy_req;
1059 struct lustre_msg *copy_reqmsg;
1060 struct ptlrpc_user_desc *udesc = NULL;
1062 OBD_ALLOC_PTR(copy_req);
1065 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1067 OBD_FREE_PTR(copy_req);
1071 if (orig_req->rq_user_desc) {
1072 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1074 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1076 OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1077 OBD_FREE_PTR(copy_req);
1080 memcpy(udesc, orig_req->rq_user_desc,
1081 sptlrpc_user_desc_size(ngroups));
1084 *copy_req = *orig_req;
1085 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1086 copy_req->rq_reqmsg = copy_reqmsg;
1087 copy_req->rq_user_desc = udesc;
1089 class_export_rpc_get(copy_req->rq_export);
1090 CFS_INIT_LIST_HEAD(©_req->rq_list);
1091 CFS_INIT_LIST_HEAD(©_req->rq_replay_list);
1092 sptlrpc_svc_ctx_addref(copy_req);
1094 if (copy_req->rq_reply_state) {
1095 /* the copied req takes over the reply state */
1096 orig_req->rq_reply_state = NULL;
1097 /* to catch further access */
1098 orig_req->rq_repmsg = NULL;
1099 orig_req->rq_replen = 0;
1105 void ptlrpc_free_clone(struct ptlrpc_request *req)
1107 LASSERT(list_empty(&req->rq_replay_list));
1109 ptlrpc_req_drop_rs(req);
1110 sptlrpc_svc_ctx_decref(req);
1111 class_export_rpc_put(req->rq_export);
1112 list_del_init(&req->rq_list);
1114 if (req->rq_user_desc) {
1115 int ngroups = req->rq_user_desc->pud_ngroups;
1116 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1118 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1122 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1124 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1125 struct obd_export *exp = req->rq_export;
1126 struct ptlrpc_request *reqiter;
1131 spin_lock(&exp->exp_lock);
1132 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1134 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1141 /* we expect it with RESENT and REPLAY flags */
1142 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1143 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1144 CERROR("invalid flags %x of resent replay\n",
1145 lustre_msg_get_flags(req->rq_reqmsg));
1147 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1150 spin_unlock(&exp->exp_lock);
1154 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1156 LASSERT(!list_empty(&req->rq_replay_list));
1157 LASSERT(req->rq_export);
1159 spin_lock(&req->rq_export->exp_lock);
1160 list_del_init(&req->rq_replay_list);
1161 spin_unlock(&req->rq_export->exp_lock);
1165 static void target_finish_recovery(struct obd_device *obd)
1168 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1171 ldlm_reprocess_all_ns(obd->obd_namespace);
1172 spin_lock_bh(&obd->obd_processing_task_lock);
1173 if (list_empty(&obd->obd_req_replay_queue) &&
1174 list_empty(&obd->obd_lock_replay_queue) &&
1175 list_empty(&obd->obd_final_req_queue)) {
1176 obd->obd_processing_task = 0;
1178 CERROR("%s: Recovery queues ( %s%s%s) are empty\n",
1180 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1181 list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1182 list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1183 spin_unlock_bh(&obd->obd_processing_task_lock);
1186 spin_unlock_bh(&obd->obd_processing_task_lock);
1188 /* when recovery finished, cleanup orphans on mds and ost */
1189 if (OBT(obd) && OBP(obd, postrecov)) {
1190 int rc = OBP(obd, postrecov)(obd);
1191 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1192 rc < 0 ? "failed" : "complete", rc);
1195 obd->obd_recovery_end = cfs_time_current_sec();
1199 static void abort_req_replay_queue(struct obd_device *obd)
1201 struct ptlrpc_request *req, *n;
1202 struct list_head abort_list;
1204 CFS_INIT_LIST_HEAD(&abort_list);
1205 spin_lock_bh(&obd->obd_processing_task_lock);
1206 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1207 spin_unlock_bh(&obd->obd_processing_task_lock);
1208 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1209 DEBUG_REQ(D_WARNING, req, "aborted:");
1210 req->rq_status = -ENOTCONN;
1211 if (ptlrpc_error(req)) {
1212 DEBUG_REQ(D_ERROR, req,
1213 "failed abort_req_reply; skipping");
1215 target_exp_dequeue_req_replay(req);
1216 ptlrpc_free_clone(req);
1220 static void abort_lock_replay_queue(struct obd_device *obd)
1222 struct ptlrpc_request *req, *n;
1223 struct list_head abort_list;
1225 CFS_INIT_LIST_HEAD(&abort_list);
1226 spin_lock_bh(&obd->obd_processing_task_lock);
1227 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1228 spin_unlock_bh(&obd->obd_processing_task_lock);
1229 list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1230 DEBUG_REQ(D_ERROR, req, "aborted:");
1231 req->rq_status = -ENOTCONN;
1232 if (ptlrpc_error(req)) {
1233 DEBUG_REQ(D_ERROR, req,
1234 "failed abort_lock_reply; skipping");
1236 ptlrpc_free_clone(req);
1241 /* Called from a cleanup function if the device is being cleaned up
1242 forcefully. The exports should all have been disconnected already,
1243 the only thing left to do is
1244 - clear the recovery flags
1246 - free queued requests and replies, but don't send replies
1247 Because the obd_stopping flag is set, no new requests should be received.
1250 void target_cleanup_recovery(struct obd_device *obd)
1252 struct ptlrpc_request *req, *n;
1253 struct list_head clean_list;
1256 LASSERT(obd->obd_stopping);
1258 CFS_INIT_LIST_HEAD(&clean_list);
1259 spin_lock_bh(&obd->obd_processing_task_lock);
1260 if (!obd->obd_recovering) {
1261 spin_unlock_bh(&obd->obd_processing_task_lock);
1265 obd->obd_recovering = obd->obd_abort_recovery = 0;
1266 target_cancel_recovery_timer(obd);
1268 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1269 spin_unlock_bh(&obd->obd_processing_task_lock);
1271 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1272 LASSERT(req->rq_reply_state == 0);
1273 target_exp_dequeue_req_replay(req);
1274 ptlrpc_free_clone(req);
1277 spin_lock_bh(&obd->obd_processing_task_lock);
1278 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1279 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1280 spin_unlock_bh(&obd->obd_processing_task_lock);
1282 list_for_each_entry_safe(req, n, &clean_list, rq_list){
1283 LASSERT(req->rq_reply_state == 0);
1284 ptlrpc_free_clone(req);
1290 /* obd_processing_task_lock should be held */
1291 void target_cancel_recovery_timer(struct obd_device *obd)
1293 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1294 cfs_timer_disarm(&obd->obd_recovery_timer);
1297 /* extend = 1 means require at least "duration" seconds left in the timer,
1298 extend = 0 means set the total duration (start_recovery_timer) */
1299 static void reset_recovery_timer(struct obd_device *obd, int duration,
1302 cfs_time_t now = cfs_time_current_sec();
1303 cfs_duration_t left;
1305 spin_lock_bh(&obd->obd_processing_task_lock);
1306 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1307 spin_unlock_bh(&obd->obd_processing_task_lock);
1311 left = cfs_time_sub(obd->obd_recovery_end, now);
1313 if (extend && (duration > left))
1314 obd->obd_recovery_timeout += duration - left;
1315 else if (!extend && (duration > obd->obd_recovery_timeout))
1316 /* Track the client's largest expected replay time */
1317 obd->obd_recovery_timeout = duration;
1320 * If total recovery time already exceed the
1321 * obd_recovery_max_time, then CRAY XT3 will
1322 * abort the recovery
1324 if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1325 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1327 obd->obd_recovery_end = obd->obd_recovery_start +
1328 obd->obd_recovery_timeout;
1329 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1330 cfs_time_before(now, obd->obd_recovery_end)) {
1331 left = cfs_time_sub(obd->obd_recovery_end, now);
1332 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1334 spin_unlock_bh(&obd->obd_processing_task_lock);
1335 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1336 obd->obd_name, (unsigned)left);
1339 static void check_and_start_recovery_timer(struct obd_device *obd)
1341 spin_lock_bh(&obd->obd_processing_task_lock);
1342 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1343 spin_unlock_bh(&obd->obd_processing_task_lock);
1346 CWARN("%s: starting recovery timer\n", obd->obd_name);
1347 obd->obd_recovery_start = cfs_time_current_sec();
1349 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1350 spin_unlock_bh(&obd->obd_processing_task_lock);
1352 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1355 /* Reset the timer with each new client connection */
1357 * This timer is actually reconnect_timer, which is for making sure
1358 * the total recovery window is at least as big as my reconnect
1359 * attempt timing. So the initial recovery time_out will be set to
1360 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1361 * from client is bigger than this, then the recovery time_out will
1362 * be extend to make sure the client could be reconnected, in the
1363 * process, the timeout from the new client should be ignored.
1367 target_start_and_reset_recovery_timer(struct obd_device *obd,
1368 struct ptlrpc_request *req,
1371 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1373 if (!new_client && service_time)
1374 /* Teach server about old server's estimates, as first guess
1375 * at how long new requests will take. */
1376 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1379 check_and_start_recovery_timer(obd);
1381 /* convert the service time to rpc timeout,
1382 * reuse service_time to limit stack usage */
1383 service_time = at_est2timeout(service_time);
1385 /* We expect other clients to timeout within service_time, then try
1386 * to reconnect, then try the failover server. The max delay between
1387 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1388 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1389 INITIAL_CONNECT_TIMEOUT);
1390 if (service_time > obd->obd_recovery_timeout && !new_client)
1391 reset_recovery_timer(obd, service_time, 0);
1395 static int check_for_next_transno(struct obd_device *obd)
1397 struct ptlrpc_request *req = NULL;
1398 int wake_up = 0, connected, completed, queue_len, max;
1399 __u64 next_transno, req_transno;
1401 spin_lock_bh(&obd->obd_processing_task_lock);
1403 if (!list_empty(&obd->obd_req_replay_queue)) {
1404 req = list_entry(obd->obd_req_replay_queue.next,
1405 struct ptlrpc_request, rq_list);
1406 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1411 max = obd->obd_max_recoverable_clients;
1412 connected = obd->obd_connected_clients;
1413 completed = connected - obd->obd_recoverable_clients;
1414 queue_len = obd->obd_requests_queued_for_recovery;
1415 next_transno = obd->obd_next_recovery_transno;
1417 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1418 "req_transno: "LPU64", next_transno: "LPU64"\n",
1419 max, connected, completed, queue_len, req_transno, next_transno);
1421 if (obd->obd_abort_recovery) {
1422 CDEBUG(D_HA, "waking for aborted recovery\n");
1424 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1425 CDEBUG(D_HA, "waking for completed recovery\n");
1427 } else if (req_transno == next_transno) {
1428 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1430 } else if (queue_len + completed == max) {
1431 /* handle gaps occured due to lost reply. It is allowed gaps
1432 * because all clients are connected and there will be resend
1433 * for missed transaction */
1434 LASSERTF(req_transno >= next_transno,
1435 "req_transno: "LPU64", next_transno: "LPU64"\n",
1436 req_transno, next_transno);
1438 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1439 "waking for skipped transno (skip: "LPD64
1440 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1441 next_transno, queue_len, completed, connected, req_transno);
1442 obd->obd_next_recovery_transno = req_transno;
1444 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1445 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1446 " by fail_lock, waking up ("LPD64")\n", next_transno);
1447 obd->obd_next_recovery_transno = req_transno;
1449 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1450 /* some clients haven't connected in time, but we can try
1451 * to replay requests that demand on already committed ones
1452 * also, we can replay first non-committed transation */
1453 LASSERT(req_transno != 0);
1454 if (req_transno == obd->obd_last_committed + 1) {
1455 obd->obd_next_recovery_transno = req_transno;
1456 } else if (req_transno > obd->obd_last_committed) {
1457 /* can't continue recovery: have no needed transno */
1458 obd->obd_abort_recovery = 1;
1459 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1460 "connected: %d, completed: %d, queue_len: %d, "
1461 "req_transno: "LPU64", next_transno: "LPU64"\n",
1462 max, connected, completed, queue_len,
1463 req_transno, next_transno);
1468 spin_unlock_bh(&obd->obd_processing_task_lock);
1472 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1474 struct l_wait_info lwi = { 0 };
1475 struct ptlrpc_request *req;
1477 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1478 obd->obd_next_recovery_transno);
1479 l_wait_event(obd->obd_next_transno_waitq,
1480 check_for_next_transno(obd), &lwi);
1482 spin_lock_bh(&obd->obd_processing_task_lock);
1483 if (obd->obd_abort_recovery) {
1485 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1486 req = list_entry(obd->obd_req_replay_queue.next,
1487 struct ptlrpc_request, rq_list);
1488 target_exp_dequeue_req_replay(req);
1489 list_del_init(&req->rq_list);
1490 obd->obd_requests_queued_for_recovery--;
1494 spin_unlock_bh(&obd->obd_processing_task_lock);
1498 static int check_for_next_lock(struct obd_device *obd)
1500 struct ptlrpc_request *req = NULL;
1503 spin_lock_bh(&obd->obd_processing_task_lock);
1504 if (!list_empty(&obd->obd_lock_replay_queue)) {
1505 req = list_entry(obd->obd_lock_replay_queue.next,
1506 struct ptlrpc_request, rq_list);
1507 CDEBUG(D_HA, "waking for next lock\n");
1509 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1510 CDEBUG(D_HA, "waking for completed lock replay\n");
1512 } else if (obd->obd_abort_recovery) {
1513 CDEBUG(D_HA, "waking for aborted recovery\n");
1516 spin_unlock_bh(&obd->obd_processing_task_lock);
1521 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1523 struct l_wait_info lwi = { 0 };
1524 struct ptlrpc_request *req;
1526 CDEBUG(D_HA, "Waiting for lock\n");
1527 l_wait_event(obd->obd_next_transno_waitq,
1528 check_for_next_lock(obd), &lwi);
1530 spin_lock_bh(&obd->obd_processing_task_lock);
1531 if (obd->obd_abort_recovery) {
1533 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1534 req = list_entry(obd->obd_lock_replay_queue.next,
1535 struct ptlrpc_request, rq_list);
1536 list_del_init(&req->rq_list);
1540 spin_unlock_bh(&obd->obd_processing_task_lock);
1544 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1546 struct ptlrpc_request *req;
1548 spin_lock_bh(&obd->obd_processing_task_lock);
1549 if (!list_empty(&obd->obd_final_req_queue)) {
1550 req = list_entry(obd->obd_final_req_queue.next,
1551 struct ptlrpc_request, rq_list);
1552 list_del_init(&req->rq_list);
1556 spin_unlock_bh(&obd->obd_processing_task_lock);
1560 static inline int req_replay_done(struct obd_export *exp)
1562 return (exp->exp_req_replay_needed == 0);
1565 static inline int lock_replay_done(struct obd_export *exp)
1567 return (exp->exp_lock_replay_needed == 0);
1570 static inline int connect_done(struct obd_export *exp)
1572 return (exp->exp_in_recovery != 0);
1575 static int check_for_clients(struct obd_device *obd)
1577 if (obd->obd_abort_recovery)
1579 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1580 if (obd->obd_no_conn == 0 &&
1581 obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1586 static int handle_recovery_req(struct ptlrpc_thread *thread,
1587 struct ptlrpc_request *req,
1588 svc_handler_t handler)
1593 rc = lu_context_init(&req->rq_session, LCT_SESSION);
1595 CERROR("Failure to initialize session: %d\n", rc);
1598 req->rq_session.lc_thread = thread;
1599 lu_context_enter(&req->rq_session);
1600 req->rq_svc_thread = thread;
1601 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1603 /* thread context */
1604 lu_context_enter(&thread->t_env->le_ctx);
1606 lu_context_exit(&thread->t_env->le_ctx);
1608 lu_context_exit(&req->rq_session);
1609 lu_context_fini(&req->rq_session);
1610 /* don't reset timer for final stage */
1611 if (!req_replay_done(req->rq_export) ||
1612 !lock_replay_done(req->rq_export))
1613 reset_recovery_timer(class_exp2obd(req->rq_export),
1614 AT_OFF ? obd_timeout :
1615 at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1616 ptlrpc_free_clone(req);
1620 static void resume_recovery_timer(struct obd_device *obd)
1622 /* to be safe, make it at least OBD_RECOVERY_FACTOR * obd_timeout */
1623 reset_recovery_timer(obd, OBD_RECOVERY_FACTOR * obd_timeout, 1);
1626 static int target_recovery_thread(void *arg)
1628 struct obd_device *obd = arg;
1629 struct ptlrpc_request *req;
1630 struct target_recovery_data *trd = &obd->obd_recovery_data;
1631 struct l_wait_info lwi = { 0 };
1632 unsigned long delta;
1633 unsigned long flags;
1635 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1639 cfs_daemonize("tgt_recov");
1641 SIGNAL_MASK_LOCK(current, flags);
1642 sigfillset(¤t->blocked);
1644 SIGNAL_MASK_UNLOCK(current, flags);
1646 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1650 thread->t_env = &env;
1651 env.le_ctx.lc_thread = thread;
1653 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1655 trd->trd_processing_task = current->pid;
1657 obd->obd_recovering = 1;
1658 complete(&trd->trd_starting);
1660 /* first of all, we have to know the first transno to replay */
1661 obd->obd_abort_recovery = 0;
1662 l_wait_event(obd->obd_next_transno_waitq,
1663 check_for_clients(obd), &lwi);
1665 spin_lock_bh(&obd->obd_processing_task_lock);
1666 target_cancel_recovery_timer(obd);
1667 spin_unlock_bh(&obd->obd_processing_task_lock);
1669 /* If some clients haven't connected in time, evict them */
1670 if (obd->obd_abort_recovery) {
1671 CWARN("Some clients haven't connect in time (%d/%d),"
1672 "evict them\n", obd->obd_connected_clients,
1673 obd->obd_max_recoverable_clients);
1674 obd->obd_abort_recovery = obd->obd_stopping;
1675 class_disconnect_stale_exports(obd, connect_done,
1676 exp_flags_from_obd(obd) |
1677 OBD_OPT_ABORT_RECOV);
1679 /* next stage: replay requests */
1681 obd->obd_req_replaying = 1;
1682 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1683 atomic_read(&obd->obd_req_replay_clients),
1684 obd->obd_next_recovery_transno);
1685 resume_recovery_timer(obd);
1686 while ((req = target_next_replay_req(obd))) {
1687 LASSERT(trd->trd_processing_task == current->pid);
1688 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1689 lustre_msg_get_transno(req->rq_reqmsg),
1690 libcfs_nid2str(req->rq_peer.nid));
1691 handle_recovery_req(thread, req,
1692 trd->trd_recovery_handler);
1693 obd->obd_replayed_requests++;
1694 spin_lock_bh(&obd->obd_processing_task_lock);
1695 obd->obd_next_recovery_transno++;
1696 spin_unlock_bh(&obd->obd_processing_task_lock);
1699 spin_lock_bh(&obd->obd_processing_task_lock);
1700 target_cancel_recovery_timer(obd);
1701 spin_unlock_bh(&obd->obd_processing_task_lock);
1703 /* If some clients haven't replayed requests in time, evict them */
1704 if (obd->obd_abort_recovery) {
1705 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1706 obd->obd_abort_recovery = obd->obd_stopping;
1707 class_disconnect_stale_exports(obd, req_replay_done,
1708 exp_flags_from_obd(obd) |
1709 OBD_OPT_ABORT_RECOV);
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 exp_flags_from_obd(obd) |
1736 OBD_OPT_ABORT_RECOV);
1737 abort_lock_replay_queue(obd);
1740 /* We drop recoverying flag to forward all new requests
1741 * to regular mds_handle() since now */
1742 spin_lock_bh(&obd->obd_processing_task_lock);
1743 obd->obd_recovering = obd->obd_abort_recovery = 0;
1744 spin_unlock_bh(&obd->obd_processing_task_lock);
1745 /* The third stage: reply on final pings */
1746 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1747 while ((req = target_next_final_ping(obd))) {
1748 LASSERT(trd->trd_processing_task == current->pid);
1749 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1750 libcfs_nid2str(req->rq_peer.nid));
1751 handle_recovery_req(thread, req,
1752 trd->trd_recovery_handler);
1755 delta = (jiffies - delta) / HZ;
1756 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1757 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1758 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1759 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1760 if (delta > obd_timeout * 2) {
1761 CWARN("too long recovery - read logs\n");
1762 libcfs_debug_dumplog();
1765 target_finish_recovery(obd);
1767 lu_context_fini(&env.le_ctx);
1768 trd->trd_processing_task = 0;
1769 complete(&trd->trd_finishing);
1773 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1776 struct target_recovery_data *trd = &obd->obd_recovery_data;
1778 memset(trd, 0, sizeof(*trd));
1779 init_completion(&trd->trd_starting);
1780 init_completion(&trd->trd_finishing);
1781 trd->trd_recovery_handler = handler;
1783 if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1784 wait_for_completion(&trd->trd_starting);
1785 LASSERT(obd->obd_recovering != 0);
1792 void target_stop_recovery_thread(struct obd_device *obd)
1794 spin_lock_bh(&obd->obd_processing_task_lock);
1795 if (obd->obd_recovery_data.trd_processing_task > 0) {
1796 struct target_recovery_data *trd = &obd->obd_recovery_data;
1797 CERROR("%s: Aborting recovery\n", obd->obd_name);
1798 obd->obd_abort_recovery = 1;
1799 wake_up(&obd->obd_next_transno_waitq);
1800 spin_unlock_bh(&obd->obd_processing_task_lock);
1801 wait_for_completion(&trd->trd_finishing);
1803 spin_unlock_bh(&obd->obd_processing_task_lock);
1807 void target_recovery_fini(struct obd_device *obd)
1809 class_disconnect_exports(obd);
1810 target_stop_recovery_thread(obd);
1811 target_cleanup_recovery(obd);
1813 EXPORT_SYMBOL(target_recovery_fini);
1815 static void target_recovery_expired(unsigned long castmeharder)
1817 struct obd_device *obd = (struct obd_device *)castmeharder;
1818 LCONSOLE_WARN("%s: recovery timed out; %d clients never reconnected "
1819 "after %lds (%d clients did)\n",
1820 obd->obd_name, obd->obd_recoverable_clients,
1821 cfs_time_current_sec()- obd->obd_recovery_start,
1822 obd->obd_connected_clients);
1823 spin_lock_bh(&obd->obd_processing_task_lock);
1824 if (obd->obd_recovering)
1825 obd->obd_abort_recovery = 1;
1826 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1827 spin_unlock_bh(&obd->obd_processing_task_lock);
1830 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1832 if (obd->obd_max_recoverable_clients == 0)
1835 CWARN("RECOVERY: service %s, %d recoverable clients, "
1836 "last_transno "LPU64"\n", obd->obd_name,
1837 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1838 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1839 obd->obd_recovery_start = 0;
1840 obd->obd_recovery_end = 0;
1841 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1842 /* bz13079: this should be set to desired value for ost but not for mds */
1843 obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1844 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1845 target_start_recovery_thread(obd, handler);
1847 EXPORT_SYMBOL(target_recovery_init);
1851 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1853 struct obd_export *exp = req->rq_export;
1854 LASSERT(exp != NULL);
1855 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1856 /* client declares he's ready to replay locks */
1857 spin_lock_bh(&obd->obd_processing_task_lock);
1858 if (exp->exp_req_replay_needed) {
1859 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1860 spin_lock(&exp->exp_lock);
1861 exp->exp_req_replay_needed = 0;
1862 spin_unlock(&exp->exp_lock);
1863 atomic_dec(&obd->obd_req_replay_clients);
1864 LASSERT(obd->obd_recoverable_clients > 0);
1865 obd->obd_recoverable_clients--;
1866 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1867 CDEBUG(D_HA, "all clients have replayed reqs\n");
1868 wake_up(&obd->obd_next_transno_waitq);
1870 spin_unlock_bh(&obd->obd_processing_task_lock);
1872 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1873 /* client declares he's ready to complete recovery
1874 * so, we put the request on th final queue */
1875 spin_lock_bh(&obd->obd_processing_task_lock);
1876 if (exp->exp_lock_replay_needed) {
1877 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1878 spin_lock(&exp->exp_lock);
1879 exp->exp_lock_replay_needed = 0;
1880 spin_unlock(&exp->exp_lock);
1881 atomic_dec(&obd->obd_lock_replay_clients);
1882 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1883 CDEBUG(D_HA, "all clients have replayed locks\n");
1884 wake_up(&obd->obd_next_transno_waitq);
1886 spin_unlock_bh(&obd->obd_processing_task_lock);
1892 int target_queue_recovery_request(struct ptlrpc_request *req,
1893 struct obd_device *obd)
1895 struct list_head *tmp;
1897 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1901 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1902 /* Processing the queue right now, don't re-add. */
1906 target_process_req_flags(obd, req);
1908 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1909 /* client declares he's ready to complete recovery
1910 * so, we put the request on th final queue */
1911 req = ptlrpc_clone_req(req);
1914 DEBUG_REQ(D_HA, req, "queue final req");
1915 spin_lock_bh(&obd->obd_processing_task_lock);
1916 if (obd->obd_recovering)
1917 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1919 spin_unlock_bh(&obd->obd_processing_task_lock);
1920 ptlrpc_free_clone(req);
1921 if (obd->obd_stopping) {
1927 spin_unlock_bh(&obd->obd_processing_task_lock);
1930 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1931 /* client declares he's ready to replay locks */
1932 req = ptlrpc_clone_req(req);
1935 DEBUG_REQ(D_HA, req, "queue lock replay req");
1936 spin_lock_bh(&obd->obd_processing_task_lock);
1937 LASSERT(obd->obd_recovering);
1938 /* usually due to recovery abort */
1939 if (!req->rq_export->exp_in_recovery) {
1940 spin_unlock_bh(&obd->obd_processing_task_lock);
1941 ptlrpc_free_clone(req);
1944 LASSERT(req->rq_export->exp_lock_replay_needed);
1945 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1946 spin_unlock_bh(&obd->obd_processing_task_lock);
1947 wake_up(&obd->obd_next_transno_waitq);
1951 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1952 * (i.e. buflens etc are in my own byte order), but type-dependent
1953 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1956 CFS_INIT_LIST_HEAD(&req->rq_list);
1957 DEBUG_REQ(D_HA, req, "not queueing");
1961 spin_lock_bh(&obd->obd_processing_task_lock);
1963 /* If we're processing the queue, we want don't want to queue this
1966 * Also, if this request has a transno less than the one we're waiting
1967 * for, we should process it now. It could (and currently always will)
1968 * be an open request for a descriptor that was opened some time ago.
1970 * Also, a resent, replayed request that has already been
1971 * handled will pass through here and be processed immediately.
1973 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1974 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1975 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1976 /* Processing the queue right now, don't re-add. */
1977 LASSERT(list_empty(&req->rq_list));
1978 spin_unlock_bh(&obd->obd_processing_task_lock);
1981 spin_unlock_bh(&obd->obd_processing_task_lock);
1983 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
1986 req = ptlrpc_clone_req(req);
1990 spin_lock_bh(&obd->obd_processing_task_lock);
1991 LASSERT(obd->obd_recovering);
1992 if (!req->rq_export->exp_in_recovery) {
1993 spin_unlock_bh(&obd->obd_processing_task_lock);
1994 ptlrpc_free_clone(req);
1997 LASSERT(req->rq_export->exp_req_replay_needed);
1999 if (target_exp_enqueue_req_replay(req)) {
2000 spin_unlock_bh(&obd->obd_processing_task_lock);
2001 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2002 ptlrpc_free_clone(req);
2007 list_for_each(tmp, &obd->obd_req_replay_queue) {
2008 struct ptlrpc_request *reqiter =
2009 list_entry(tmp, struct ptlrpc_request, rq_list);
2011 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2012 list_add_tail(&req->rq_list, &reqiter->rq_list);
2017 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2019 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2020 "has been claimed by another client");
2021 spin_unlock_bh(&obd->obd_processing_task_lock);
2022 target_exp_dequeue_req_replay(req);
2023 ptlrpc_free_clone(req);
2029 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2031 obd->obd_requests_queued_for_recovery++;
2032 wake_up(&obd->obd_next_transno_waitq);
2033 spin_unlock_bh(&obd->obd_processing_task_lock);
2037 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2039 return req->rq_export->exp_obd;
2042 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2044 LASSERT(exp != NULL);
2045 return &exp->exp_obd->obd_namespace->ns_pool;
2049 * Packs current SLV and Limit into \a req.
2051 int target_pack_pool_reply(struct ptlrpc_request *req)
2053 struct obd_device *obd;
2057 * Check that we still have all structures alive as this may
2058 * be some late rpc in shutdown time.
2060 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2061 !exp_connect_lru_resize(req->rq_export))) {
2062 lustre_msg_set_slv(req->rq_repmsg, 0);
2063 lustre_msg_set_limit(req->rq_repmsg, 0);
2068 * OBD is alive here as export is alive, which we checked above.
2070 obd = req->rq_export->exp_obd;
2072 read_lock(&obd->obd_pool_lock);
2073 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2074 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2075 read_unlock(&obd->obd_pool_lock);
2080 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2082 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2083 DEBUG_REQ(D_ERROR, req, "dropping reply");
2088 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2089 req->rq_status = rc;
2090 return (ptlrpc_send_error(req, 1));
2092 DEBUG_REQ(D_NET, req, "sending reply");
2095 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2098 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2101 struct ptlrpc_reply_state *rs;
2102 struct obd_device *obd;
2103 struct obd_export *exp;
2104 struct ptlrpc_service *svc;
2107 if (req->rq_no_reply) {
2112 svc = req->rq_rqbd->rqbd_service;
2113 rs = req->rq_reply_state;
2114 if (rs == NULL || !rs->rs_difficult) {
2116 target_send_reply_msg (req, rc, fail_id);
2121 /* must be an export if locks saved */
2122 LASSERT (req->rq_export != NULL);
2123 /* req/reply consistent */
2124 LASSERT (rs->rs_service == svc);
2127 LASSERT (!rs->rs_scheduled);
2128 LASSERT (!rs->rs_scheduled_ever);
2129 LASSERT (!rs->rs_handled);
2130 LASSERT (!rs->rs_on_net);
2131 LASSERT (rs->rs_export == NULL);
2132 LASSERT (list_empty(&rs->rs_obd_list));
2133 LASSERT (list_empty(&rs->rs_exp_list));
2135 exp = class_export_get (req->rq_export);
2138 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2139 rs->rs_scheduled = 1;
2141 rs->rs_xid = req->rq_xid;
2142 rs->rs_transno = req->rq_transno;
2143 rs->rs_export = exp;
2145 spin_lock(&obd->obd_uncommitted_replies_lock);
2147 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2148 rs->rs_transno, obd->obd_last_committed);
2149 if (rs->rs_transno > obd->obd_last_committed) {
2150 /* not committed already */
2151 list_add_tail (&rs->rs_obd_list,
2152 &obd->obd_uncommitted_replies);
2155 spin_unlock (&obd->obd_uncommitted_replies_lock);
2156 spin_lock (&exp->exp_lock);
2158 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
2160 spin_unlock(&exp->exp_lock);
2162 netrc = target_send_reply_msg (req, rc, fail_id);
2164 spin_lock(&svc->srv_lock);
2166 atomic_inc(&svc->srv_n_difficult_replies);
2169 /* error sending: reply is off the net. Also we need +1
2170 * reply ref until ptlrpc_server_handle_reply() is done
2171 * with the reply state (if the send was successful, there
2172 * would have been +1 ref for the net, which
2173 * reply_out_callback leaves alone) */
2175 ptlrpc_rs_addref(rs);
2176 atomic_inc (&svc->srv_outstanding_replies);
2179 spin_lock(&rs->rs_lock);
2180 if (rs->rs_transno <= obd->obd_last_committed ||
2181 (!rs->rs_on_net && !rs->rs_no_ack) ||
2182 list_empty(&rs->rs_exp_list) || /* completed already */
2183 list_empty(&rs->rs_obd_list)) {
2184 CDEBUG(D_HA, "Schedule reply immediately\n");
2185 ptlrpc_dispatch_difficult_reply(rs);
2187 list_add (&rs->rs_list, &svc->srv_active_replies);
2188 rs->rs_scheduled = 0; /* allow notifier to schedule */
2190 spin_unlock(&rs->rs_lock);
2191 spin_unlock(&svc->srv_lock);
2195 int target_handle_ping(struct ptlrpc_request *req)
2197 obd_ping(req->rq_export);
2198 return req_capsule_server_pack(&req->rq_pill);
2201 void target_committed_to_req(struct ptlrpc_request *req)
2203 struct obd_device *obd;
2205 if (req == NULL || req->rq_export == NULL)
2208 obd = req->rq_export->exp_obd;
2212 if (!obd->obd_no_transno && req->rq_repmsg != NULL)
2213 lustre_msg_set_last_committed(req->rq_repmsg,
2214 obd->obd_last_committed);
2216 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2217 "%d)", obd->obd_no_transno, req->rq_repmsg == NULL);
2219 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2220 obd->obd_last_committed, req->rq_transno, req->rq_xid);
2223 EXPORT_SYMBOL(target_committed_to_req);
2225 int target_handle_qc_callback(struct ptlrpc_request *req)
2227 struct obd_quotactl *oqctl;
2228 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2230 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2231 if (oqctl == NULL) {
2232 CERROR("Can't unpack obd_quotactl\n");
2236 cli->cl_qchk_stat = oqctl->qc_stat;
2241 #ifdef HAVE_QUOTA_SUPPORT
2242 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2245 struct obd_device *obd = req->rq_export->exp_obd;
2246 struct obd_device *master_obd;
2247 struct obd_device_target *obt;
2248 struct lustre_quota_ctxt *qctxt;
2249 struct qunit_data *qdata = NULL;
2253 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2256 rc = req_capsule_server_pack(&req->rq_pill);
2258 CERROR("packing reply failed!: rc = %d\n", rc);
2262 LASSERT(req->rq_export);
2264 OBD_ALLOC(qdata, sizeof(struct qunit_data));
2267 rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
2269 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2273 /* we use the observer */
2274 if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2275 CERROR("Can't find the observer, it is recovering\n");
2276 req->rq_status = -EAGAIN;
2277 GOTO(send_reply, rc = -EAGAIN);
2280 master_obd = obd->obd_observer->obd_observer;
2281 obt = &master_obd->u.obt;
2282 qctxt = &obt->obt_qctxt;
2284 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2285 /* quota_type has not been processed yet, return EAGAIN
2286 * until we know whether or not quotas are supposed to
2288 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2290 req->rq_status = -EAGAIN;
2291 rc = ptlrpc_reply(req);
2295 down_read(&obt->obt_rwsem);
2296 if (qctxt->lqc_lqs_hash == NULL) {
2297 up_read(&obt->obt_rwsem);
2298 /* quota_type has not been processed yet, return EAGAIN
2299 * until we know whether or not quotas are supposed to
2301 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2303 req->rq_status = -EAGAIN;
2304 rc = ptlrpc_reply(req);
2308 LASSERT(qctxt->lqc_handler);
2309 rc = qctxt->lqc_handler(master_obd, qdata,
2310 lustre_msg_get_opc(req->rq_reqmsg));
2311 up_read(&obt->obt_rwsem);
2312 if (rc && rc != -EDQUOT)
2313 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2314 "dqacq failed! (rc:%d)\n", rc);
2315 req->rq_status = rc;
2317 /* there are three forms of qunit(historic causes), so we need to
2318 * adjust the same form to different forms slaves needed */
2319 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2321 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2325 /* Block the quota req. b=14840 */
2326 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2328 rc = ptlrpc_reply(req);
2330 OBD_FREE(qdata, sizeof(struct qunit_data));
2334 #endif /* !__KERNEL__ */
2336 #endif /* HAVE_QUOTA_SUPPORT */
2338 ldlm_mode_t lck_compat_array[] = {
2339 [LCK_EX] LCK_COMPAT_EX,
2340 [LCK_PW] LCK_COMPAT_PW,
2341 [LCK_PR] LCK_COMPAT_PR,
2342 [LCK_CW] LCK_COMPAT_CW,
2343 [LCK_CR] LCK_COMPAT_CR,
2344 [LCK_NL] LCK_COMPAT_NL,
2345 [LCK_GROUP] LCK_COMPAT_GROUP,
2346 [LCK_COS] LCK_COMPAT_COS,
2350 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2351 * not escape to the user level.
2353 int ldlm_error2errno(ldlm_error_t error)
2361 case ELDLM_LOCK_CHANGED:
2364 case ELDLM_LOCK_ABORTED:
2367 case ELDLM_LOCK_REPLACED:
2370 case ELDLM_NO_LOCK_DATA:
2373 case ELDLM_NAMESPACE_EXISTS:
2376 case ELDLM_BAD_NAMESPACE:
2380 if (((int)error) < 0) /* cast to signed type */
2381 result = error; /* as ldlm_error_t can be unsigned */
2383 CERROR("Invalid DLM result code: %i\n", error);
2389 EXPORT_SYMBOL(ldlm_error2errno);
2392 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2394 ldlm_error_t ldlm_errno2error(int err_no)
2403 error = ELDLM_LOCK_CHANGED;
2406 error = ELDLM_LOCK_ABORTED;
2409 error = ELDLM_LOCK_REPLACED;
2412 error = ELDLM_NO_LOCK_DATA;
2415 error = ELDLM_NAMESPACE_EXISTS;
2418 error = ELDLM_BAD_NAMESPACE;
2425 EXPORT_SYMBOL(ldlm_errno2error);