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;
239 CERROR("unknown client OBD type \"%s\", can't setup\n",
244 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
245 CERROR("requires a TARGET UUID\n");
249 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
250 CERROR("client UUID must be less than 38 characters\n");
254 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
255 CERROR("setup requires a SERVER UUID\n");
259 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
260 CERROR("target UUID must be less than 38 characters\n");
264 init_rwsem(&cli->cl_sem);
265 sema_init(&cli->cl_mgc_sem, 1);
266 cli->cl_conn_count = 0;
267 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
268 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
269 sizeof(server_uuid)));
272 cli->cl_avail_grant = 0;
273 /* FIXME: should limit this for the sum of all cl_dirty_max */
274 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
275 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
276 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
277 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
278 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
279 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
280 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
281 client_obd_list_lock_init(&cli->cl_loi_list_lock);
282 cli->cl_r_in_flight = 0;
283 cli->cl_w_in_flight = 0;
285 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
286 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
287 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
288 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
289 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
290 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
291 cfs_waitq_init(&cli->cl_destroy_waitq);
292 atomic_set(&cli->cl_destroy_in_flight, 0);
293 #ifdef ENABLE_CHECKSUM
294 /* Turn on checksumming by default. */
295 cli->cl_checksum = 1;
297 * The supported checksum types will be worked out at connect time
298 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
301 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
303 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
305 /* This value may be changed at connect time in
306 ptlrpc_connect_interpret. */
307 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
308 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
310 if (!strcmp(name, LUSTRE_MDC_NAME)) {
311 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
312 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
313 cli->cl_max_rpcs_in_flight = 2;
314 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
315 cli->cl_max_rpcs_in_flight = 3;
316 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
317 cli->cl_max_rpcs_in_flight = 4;
319 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
324 CERROR("ldlm_get_ref failed: %d\n", rc);
328 ptlrpc_init_client(rq_portal, rp_portal, name,
329 &obddev->obd_ldlm_client);
331 imp = class_new_import(obddev);
333 GOTO(err_ldlm, rc = -ENOENT);
334 imp->imp_client = &obddev->obd_ldlm_client;
335 imp->imp_connect_op = connect_op;
336 imp->imp_initial_recov = 1;
337 imp->imp_initial_recov_bk = 0;
338 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
339 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
340 LUSTRE_CFG_BUFLEN(lcfg, 1));
341 class_import_put(imp);
343 rc = client_import_add_conn(imp, &server_uuid, 1);
345 CERROR("can't add initial connection\n");
346 GOTO(err_import, rc);
349 cli->cl_import = imp;
350 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
351 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
352 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
354 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
355 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
356 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
357 name, obddev->obd_name,
358 cli->cl_target_uuid.uuid);
359 spin_lock(&imp->imp_lock);
360 imp->imp_invalid = 1;
361 spin_unlock(&imp->imp_lock);
365 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
370 class_destroy_import(imp);
378 int client_obd_cleanup(struct obd_device *obddev)
385 /* ->o_connect() method for client side (OSC and MDC and MGC) */
386 int client_connect_import(const struct lu_env *env,
387 struct lustre_handle *dlm_handle,
388 struct obd_device *obd, struct obd_uuid *cluuid,
389 struct obd_connect_data *data, void *localdata)
391 struct client_obd *cli = &obd->u.cli;
392 struct obd_import *imp = cli->cl_import;
393 struct obd_export *exp;
394 struct obd_connect_data *ocd;
395 struct ldlm_namespace *to_be_freed = NULL;
399 down_write(&cli->cl_sem);
400 rc = class_connect(dlm_handle, obd, cluuid);
404 cli->cl_conn_count++;
405 if (cli->cl_conn_count > 1)
407 exp = class_conn2export(dlm_handle);
409 if (obd->obd_namespace != NULL)
410 CERROR("already have namespace!\n");
411 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
412 LDLM_NAMESPACE_CLIENT,
413 LDLM_NAMESPACE_GREEDY);
414 if (obd->obd_namespace == NULL)
415 GOTO(out_disco, rc = -ENOMEM);
417 imp->imp_dlm_handle = *dlm_handle;
418 rc = ptlrpc_init_import(imp);
422 ocd = &imp->imp_connect_data;
425 imp->imp_connect_flags_orig = data->ocd_connect_flags;
428 rc = ptlrpc_connect_import(imp, NULL);
430 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
433 LASSERT(exp->exp_connection);
436 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
437 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
438 data->ocd_connect_flags, ocd->ocd_connect_flags);
439 data->ocd_connect_flags = ocd->ocd_connect_flags;
442 ptlrpc_pinger_add_import(imp);
448 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
449 to_be_freed = obd->obd_namespace;
450 obd->obd_namespace = NULL;
452 cli->cl_conn_count--;
453 class_disconnect(exp);
455 class_export_put(exp);
458 up_write(&cli->cl_sem);
460 ldlm_namespace_free_post(to_be_freed);
465 int client_disconnect_export(struct obd_export *exp)
467 struct obd_device *obd = class_exp2obd(exp);
468 struct client_obd *cli;
469 struct obd_import *imp;
471 struct ldlm_namespace *to_be_freed = NULL;
475 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
476 exp, exp ? exp->exp_handle.h_cookie : -1);
481 imp = cli->cl_import;
483 down_write(&cli->cl_sem);
484 if (!cli->cl_conn_count) {
485 CERROR("disconnecting disconnected device (%s)\n",
487 GOTO(out_sem, rc = -EINVAL);
490 cli->cl_conn_count--;
491 if (cli->cl_conn_count)
492 GOTO(out_no_disconnect, rc = 0);
494 /* Mark import deactivated now, so we don't try to reconnect if any
495 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
496 * fully deactivate the import, or that would drop all requests. */
497 spin_lock(&imp->imp_lock);
498 imp->imp_deactive = 1;
499 spin_unlock(&imp->imp_lock);
501 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
502 * delete it regardless. (It's safe to delete an import that was
504 (void)ptlrpc_pinger_del_import(imp);
506 if (obd->obd_namespace != NULL) {
507 /* obd_force == local only */
508 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
509 obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
511 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
512 to_be_freed = obd->obd_namespace;
515 rc = ptlrpc_disconnect_import(imp, 0);
517 ptlrpc_invalidate_import(imp);
518 /* set obd_namespace to NULL only after invalidate, because we can have
519 * some connect requests in flight, and his need store a connect flags
520 * in obd_namespace. bug 14260 */
521 obd->obd_namespace = NULL;
523 if (imp->imp_rq_pool) {
524 ptlrpc_free_rq_pool(imp->imp_rq_pool);
525 imp->imp_rq_pool = NULL;
528 cli->cl_import = NULL;
532 err = class_disconnect(exp);
536 up_write(&cli->cl_sem);
538 ldlm_namespace_free_post(to_be_freed);
543 /* --------------------------------------------------------------------------
544 * from old lib/target.c
545 * -------------------------------------------------------------------------- */
547 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
548 struct obd_uuid *cluuid, int initial_conn)
551 if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
552 struct lustre_handle *hdl;
553 hdl = &exp->exp_imp_reverse->imp_remote_handle;
554 /* Might be a re-connect after a partition. */
555 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
556 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
558 conn->cookie = exp->exp_handle.h_cookie;
559 /* target_handle_connect() treats EALREADY and
560 * -EALREADY differently. EALREADY means we are
561 * doing a valid reconnect from the same client. */
564 CERROR("%s reconnecting from %s, "
565 "handle mismatch (ours "LPX64", theirs "
566 LPX64")\n", cluuid->uuid,
567 exp->exp_connection->c_remote_uuid.uuid,
568 hdl->cookie, conn->cookie);
569 memset(conn, 0, sizeof *conn);
570 /* target_handle_connect() treats EALREADY and
571 * -EALREADY differently. -EALREADY is an error
572 * (same UUID, different handle). */
577 conn->cookie = exp->exp_handle.h_cookie;
578 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
579 cluuid->uuid, exp, conn->cookie);
583 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
586 struct obd_export *exp = cb_data;
588 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
589 obd->obd_name, exp->exp_client_uuid.uuid);
591 spin_lock(&exp->exp_lock);
592 exp->exp_need_sync = 0;
593 spin_unlock(&exp->exp_lock);
595 EXPORT_SYMBOL(target_client_add_cb);
597 target_start_and_reset_recovery_timer(struct obd_device *obd,
598 struct ptlrpc_request *req,
601 int target_handle_connect(struct ptlrpc_request *req)
603 struct obd_device *target, *targref = NULL;
604 struct obd_export *export = NULL;
605 struct obd_import *revimp;
606 struct lustre_handle conn;
607 struct lustre_handle *tmp;
608 struct obd_uuid tgtuuid;
609 struct obd_uuid cluuid;
610 struct obd_uuid remote_uuid;
613 int initial_conn = 0;
614 struct obd_connect_data *data, *tmpdata;
615 lnet_nid_t *client_nid = NULL;
618 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
620 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
622 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
623 GOTO(out, rc = -EINVAL);
626 obd_str2uuid(&tgtuuid, str);
627 target = class_uuid2obd(&tgtuuid);
629 target = class_name2obd(str);
631 if (!target || target->obd_stopping || !target->obd_set_up) {
632 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
633 " for connect (%s)\n", str,
634 !target ? "no target" :
635 (target->obd_stopping ? "stopping" :
637 GOTO(out, rc = -ENODEV);
640 if (target->obd_no_conn) {
641 LCONSOLE_WARN("%s: temporarily refusing client connection "
642 "from %s\n", target->obd_name,
643 libcfs_nid2str(req->rq_peer.nid));
644 GOTO(out, rc = -EAGAIN);
647 /* Make sure the target isn't cleaned up while we're here. Yes,
648 there's still a race between the above check and our incref here.
649 Really, class_uuid2obd should take the ref. */
650 targref = class_incref(target, __FUNCTION__, cfs_current());
653 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
655 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
656 GOTO(out, rc = -EINVAL);
659 obd_str2uuid(&cluuid, str);
661 /* XXX extract a nettype and format accordingly */
662 switch (sizeof(lnet_nid_t)) {
663 /* NB the casts only avoid compiler warnings */
665 snprintf(remote_uuid.uuid, sizeof remote_uuid,
666 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
669 snprintf(remote_uuid.uuid, sizeof remote_uuid,
670 "NET_%x_UUID", (__u32)req->rq_peer.nid);
676 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
678 GOTO(out, rc = -EPROTO);
682 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
684 GOTO(out, rc = -EPROTO);
686 rc = req_capsule_server_pack(&req->rq_pill);
690 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
692 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
693 "libclient connection attempt");
694 GOTO(out, rc = -EPROTO);
695 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
696 LUSTRE_VERSION_ALLOWED_OFFSET ||
697 data->ocd_version > LUSTRE_VERSION_CODE +
698 LUSTRE_VERSION_ALLOWED_OFFSET) {
699 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
700 "libclient connection attempt",
701 data->ocd_version < LUSTRE_VERSION_CODE ?
703 OBD_OCD_VERSION_MAJOR(data->ocd_version),
704 OBD_OCD_VERSION_MINOR(data->ocd_version),
705 OBD_OCD_VERSION_PATCH(data->ocd_version),
706 OBD_OCD_VERSION_FIX(data->ocd_version));
707 data = req_capsule_server_sized_get(&req->rq_pill,
709 offsetof(typeof(*data), ocd_version) +
710 sizeof(data->ocd_version));
712 data->ocd_connect_flags = OBD_CONNECT_VERSION;
713 data->ocd_version = LUSTRE_VERSION_CODE;
715 GOTO(out, rc = -EPROTO);
719 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
722 /* lctl gets a backstage, all-access pass. */
723 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
724 goto dont_check_exports;
726 spin_lock(&target->obd_dev_lock);
727 export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
729 if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
730 CWARN("%s: exp %p already connecting\n",
731 export->exp_obd->obd_name, export);
732 class_export_put(export);
735 } else if (export != NULL && export->exp_connection != NULL &&
736 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
737 /* make darn sure this is coming from the same peer
738 * if the UUIDs matched */
739 CWARN("%s: cookie %s seen on new NID %s when "
740 "existing NID %s is already connected\n",
741 target->obd_name, cluuid.uuid,
742 libcfs_nid2str(req->rq_peer.nid),
743 libcfs_nid2str(export->exp_connection->c_peer.nid));
744 class_export_put(export);
747 } else if (export != NULL) {
748 spin_lock(&export->exp_lock);
749 export->exp_connecting = 1;
750 spin_unlock(&export->exp_lock);
751 class_export_put(export);
752 spin_unlock(&target->obd_dev_lock);
753 LASSERT(export->exp_obd == target);
755 rc = target_handle_reconnect(&conn, export, &cluuid, initial_conn);
758 /* If we found an export, we already unlocked. */
760 spin_unlock(&target->obd_dev_lock);
761 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
762 } else if (req->rq_export == NULL &&
763 atomic_read(&export->exp_rpc_count) > 0) {
764 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
765 target->obd_name, cluuid.uuid,
766 libcfs_nid2str(req->rq_peer.nid),
767 export, atomic_read(&export->exp_refcount));
768 GOTO(out, rc = -EBUSY);
769 } else if (req->rq_export != NULL &&
770 (atomic_read(&export->exp_rpc_count) > 1)) {
771 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
772 target->obd_name, cluuid.uuid,
773 libcfs_nid2str(req->rq_peer.nid),
774 export, atomic_read(&export->exp_rpc_count));
775 GOTO(out, rc = -EBUSY);
776 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
778 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
779 "cookies not random?\n", target->obd_name,
780 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
781 GOTO(out, rc = -EALREADY);
783 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
784 if (req->rq_export == NULL && initial_conn)
785 export->exp_last_request_time =
786 max(export->exp_last_request_time,
787 (time_t)cfs_time_current_sec());
794 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
795 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
796 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
797 export, (long)cfs_time_current_sec(),
798 export ? (long)export->exp_last_request_time : 0);
800 /* Tell the client if we're in recovery. */
801 if (target->obd_recovering) {
802 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
803 /* If this is the first time a client connects,
804 reset the recovery timer */
806 target_start_and_reset_recovery_timer(target, req,
810 /* We want to handle EALREADY but *not* -EALREADY from
811 * target_handle_reconnect(), return reconnection state in a flag */
812 if (rc == EALREADY) {
813 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
819 /* Tell the client if we support replayable requests */
820 if (target->obd_replayable)
821 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
822 client_nid = &req->rq_peer.nid;
824 if (export == NULL) {
825 if (target->obd_recovering) {
828 t = cfs_timer_deadline(&target->obd_recovery_timer);
829 t = cfs_time_sub(t, cfs_time_current());
830 CERROR("%s: denying connection for new client %s (%s): "
831 "%d clients in recovery for "CFS_TIME_T"s\n",
833 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
834 target->obd_recoverable_clients,
835 cfs_duration_sec(t));
839 rc = obd_connect(req->rq_svc_thread->t_env,
840 &conn, target, &cluuid, data,
844 rc = obd_reconnect(req->rq_svc_thread->t_env,
845 export, target, &cluuid, data, client_nid);
849 /* Return only the parts of obd_connect_data that we understand, so the
850 * client knows that we don't understand the rest. */
852 tmpdata = req_capsule_server_get(&req->rq_pill,
854 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
858 /* If all else goes well, this is our RPC return code. */
861 lustre_msg_set_handle(req->rq_repmsg, &conn);
863 /* ownership of this export ref transfers to the request AFTER we
864 * drop any previous reference the request had, but we don't want
865 * that to go to zero before we get our new export reference. */
866 export = class_conn2export(&conn);
868 DEBUG_REQ(D_ERROR, req, "Missing export!");
869 GOTO(out, rc = -ENODEV);
872 /* If the client and the server are the same node, we will already
873 * have an export that really points to the client's DLM export,
874 * because we have a shared handles table.
876 * XXX this will go away when shaver stops sending the "connect" handle
877 * in the real "remote handle" field of the request --phik 24 Apr 2003
879 if (req->rq_export != NULL)
880 class_export_put(req->rq_export);
882 req->rq_export = export;
884 spin_lock(&export->exp_lock);
886 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
887 } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
888 spin_unlock(&export->exp_lock);
889 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
890 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
891 export->exp_conn_cnt,
892 lustre_msg_get_conn_cnt(req->rq_reqmsg));
894 GOTO(out, rc = -EALREADY);
896 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
898 /* request from liblustre? Don't evict it for not pinging. */
899 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
900 export->exp_libclient = 1;
901 spin_unlock(&export->exp_lock);
903 spin_lock(&target->obd_dev_lock);
904 list_del_init(&export->exp_obd_chain_timed);
905 spin_unlock(&target->obd_dev_lock);
907 spin_unlock(&export->exp_lock);
910 if (export->exp_connection != NULL)
911 ptlrpc_connection_put(export->exp_connection);
912 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
915 if (hlist_unhashed(&export->exp_nid_hash)) {
916 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
917 &export->exp_connection->c_peer.nid,
918 &export->exp_nid_hash);
921 spin_lock_bh(&target->obd_processing_task_lock);
922 if (target->obd_recovering && !export->exp_in_recovery) {
923 spin_lock(&export->exp_lock);
924 export->exp_in_recovery = 1;
925 export->exp_req_replay_needed = 1;
926 export->exp_lock_replay_needed = 1;
927 spin_unlock(&export->exp_lock);
928 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
929 && (data->ocd_transno == 0))
930 CWARN("Connect with zero transno!\n");
932 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
933 && data->ocd_transno < target->obd_next_recovery_transno)
934 target->obd_next_recovery_transno = data->ocd_transno;
935 target->obd_connected_clients++;
936 /* each connected client is counted as recoverable */
937 target->obd_recoverable_clients++;
938 atomic_inc(&target->obd_req_replay_clients);
939 atomic_inc(&target->obd_lock_replay_clients);
940 if (target->obd_connected_clients ==
941 target->obd_max_recoverable_clients)
942 wake_up(&target->obd_next_transno_waitq);
944 spin_unlock_bh(&target->obd_processing_task_lock);
945 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
948 if (export->exp_imp_reverse != NULL) {
949 /* destroyed import can be still referenced in ctxt */
950 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
951 KEY_REVIMP_UPD, 0, NULL, NULL);
953 /* in some recovery senarios, previous ctx init rpc handled
954 * in sptlrpc_target_export_check() might be used to install
955 * a reverse ctx in this reverse import, and later OBD_CONNECT
956 * using the same gss ctx could reach here and following new
957 * reverse import. note all reverse ctx in new/old import are
958 * actually based on the same gss ctx. so we invalidate ctx
959 * here before destroy import, otherwise flush old import will
960 * lead to remote reverse ctx be destroied, thus the reverse
961 * ctx of new import will lost its peer.
962 * there might be a better way to deal with this???
964 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
966 destroy_import(export->exp_imp_reverse);
969 /* for the rest part, we return -ENOTCONN in case of errors
970 * in order to let client initialize connection again.
972 revimp = export->exp_imp_reverse = class_new_import(target);
974 CERROR("fail to alloc new reverse import.\n");
975 GOTO(out, rc = -ENOTCONN);
978 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
979 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
980 revimp->imp_remote_handle = conn;
981 revimp->imp_dlm_fake = 1;
982 revimp->imp_state = LUSTRE_IMP_FULL;
984 /* unknown versions will be caught in
985 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
986 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
988 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
989 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
990 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
992 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
994 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx,
995 req->rq_flvr.sf_rpc);
997 CERROR("Failed to get sec for reverse import: %d\n", rc);
998 export->exp_imp_reverse = NULL;
999 class_destroy_import(revimp);
1002 class_import_put(revimp);
1005 spin_lock(&export->exp_lock);
1006 export->exp_connecting = 0;
1007 spin_unlock(&export->exp_lock);
1010 class_decref(targref, __FUNCTION__, cfs_current());
1012 req->rq_status = rc;
1016 int target_handle_disconnect(struct ptlrpc_request *req)
1021 rc = req_capsule_server_pack(&req->rq_pill);
1025 /* keep the rq_export around so we can send the reply */
1026 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1031 void target_destroy_export(struct obd_export *exp)
1033 /* exports created from last_rcvd data, and "fake"
1034 exports created by lctl don't have an import */
1035 if (exp->exp_imp_reverse != NULL)
1036 destroy_import(exp->exp_imp_reverse);
1038 /* We cancel locks at disconnect time, but this will catch any locks
1039 * granted in a race with recovery-induced disconnect. */
1040 if (exp->exp_obd->obd_namespace != NULL)
1041 ldlm_cancel_locks_for_export(exp);
1045 * Recovery functions
1048 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1050 struct ptlrpc_request *copy_req;
1051 struct lustre_msg *copy_reqmsg;
1052 struct ptlrpc_user_desc *udesc = NULL;
1054 OBD_ALLOC_PTR(copy_req);
1057 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1059 OBD_FREE_PTR(copy_req);
1063 if (orig_req->rq_user_desc) {
1064 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1066 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1068 OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1069 OBD_FREE_PTR(copy_req);
1072 memcpy(udesc, orig_req->rq_user_desc,
1073 sptlrpc_user_desc_size(ngroups));
1076 *copy_req = *orig_req;
1077 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1078 copy_req->rq_reqmsg = copy_reqmsg;
1079 copy_req->rq_user_desc = udesc;
1081 class_export_rpc_get(copy_req->rq_export);
1082 CFS_INIT_LIST_HEAD(©_req->rq_list);
1083 CFS_INIT_LIST_HEAD(©_req->rq_replay_list);
1084 sptlrpc_svc_ctx_addref(copy_req);
1086 if (copy_req->rq_reply_state) {
1087 /* the copied req takes over the reply state */
1088 orig_req->rq_reply_state = NULL;
1089 /* to catch further access */
1090 orig_req->rq_repmsg = NULL;
1091 orig_req->rq_replen = 0;
1097 void ptlrpc_free_clone(struct ptlrpc_request *req)
1099 LASSERT(list_empty(&req->rq_replay_list));
1101 ptlrpc_req_drop_rs(req);
1102 sptlrpc_svc_ctx_decref(req);
1103 class_export_rpc_put(req->rq_export);
1104 list_del_init(&req->rq_list);
1106 if (req->rq_user_desc) {
1107 int ngroups = req->rq_user_desc->pud_ngroups;
1108 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1110 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1114 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1116 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1117 struct obd_export *exp = req->rq_export;
1118 struct ptlrpc_request *reqiter;
1123 spin_lock(&exp->exp_lock);
1124 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1126 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1133 /* we expect it with RESENT and REPLAY flags */
1134 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1135 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1136 CERROR("invalid flags %x of resent replay\n",
1137 lustre_msg_get_flags(req->rq_reqmsg));
1139 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1142 spin_unlock(&exp->exp_lock);
1146 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1148 LASSERT(!list_empty(&req->rq_replay_list));
1149 LASSERT(req->rq_export);
1151 spin_lock(&req->rq_export->exp_lock);
1152 list_del_init(&req->rq_replay_list);
1153 spin_unlock(&req->rq_export->exp_lock);
1157 static void target_finish_recovery(struct obd_device *obd)
1160 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1163 ldlm_reprocess_all_ns(obd->obd_namespace);
1164 spin_lock_bh(&obd->obd_processing_task_lock);
1165 if (list_empty(&obd->obd_req_replay_queue) &&
1166 list_empty(&obd->obd_lock_replay_queue) &&
1167 list_empty(&obd->obd_final_req_queue)) {
1168 obd->obd_processing_task = 0;
1170 CERROR("%s: Recovery queues ( %s%s%s) are empty\n",
1172 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1173 list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1174 list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1175 spin_unlock_bh(&obd->obd_processing_task_lock);
1178 spin_unlock_bh(&obd->obd_processing_task_lock);
1180 /* when recovery finished, cleanup orphans on mds and ost */
1181 if (OBT(obd) && OBP(obd, postrecov)) {
1182 int rc = OBP(obd, postrecov)(obd);
1183 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1184 rc < 0 ? "failed" : "complete", rc);
1187 obd->obd_recovery_end = cfs_time_current_sec();
1191 static void abort_req_replay_queue(struct obd_device *obd)
1193 struct ptlrpc_request *req, *n;
1194 struct list_head abort_list;
1196 CFS_INIT_LIST_HEAD(&abort_list);
1197 spin_lock_bh(&obd->obd_processing_task_lock);
1198 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1199 spin_unlock_bh(&obd->obd_processing_task_lock);
1200 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1201 DEBUG_REQ(D_WARNING, req, "aborted:");
1202 req->rq_status = -ENOTCONN;
1203 if (ptlrpc_error(req)) {
1204 DEBUG_REQ(D_ERROR, req,
1205 "failed abort_req_reply; skipping");
1207 target_exp_dequeue_req_replay(req);
1208 ptlrpc_free_clone(req);
1212 static void abort_lock_replay_queue(struct obd_device *obd)
1214 struct ptlrpc_request *req, *n;
1215 struct list_head abort_list;
1217 CFS_INIT_LIST_HEAD(&abort_list);
1218 spin_lock_bh(&obd->obd_processing_task_lock);
1219 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1220 spin_unlock_bh(&obd->obd_processing_task_lock);
1221 list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1222 DEBUG_REQ(D_ERROR, req, "aborted:");
1223 req->rq_status = -ENOTCONN;
1224 if (ptlrpc_error(req)) {
1225 DEBUG_REQ(D_ERROR, req,
1226 "failed abort_lock_reply; skipping");
1228 ptlrpc_free_clone(req);
1233 /* Called from a cleanup function if the device is being cleaned up
1234 forcefully. The exports should all have been disconnected already,
1235 the only thing left to do is
1236 - clear the recovery flags
1238 - free queued requests and replies, but don't send replies
1239 Because the obd_stopping flag is set, no new requests should be received.
1242 void target_cleanup_recovery(struct obd_device *obd)
1244 struct ptlrpc_request *req, *n;
1245 struct list_head clean_list;
1248 LASSERT(obd->obd_stopping);
1250 CFS_INIT_LIST_HEAD(&clean_list);
1251 spin_lock_bh(&obd->obd_processing_task_lock);
1252 if (!obd->obd_recovering) {
1253 spin_unlock_bh(&obd->obd_processing_task_lock);
1257 obd->obd_recovering = obd->obd_abort_recovery = 0;
1258 target_cancel_recovery_timer(obd);
1260 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1261 spin_unlock_bh(&obd->obd_processing_task_lock);
1263 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1264 LASSERT(req->rq_reply_state == 0);
1265 target_exp_dequeue_req_replay(req);
1266 ptlrpc_free_clone(req);
1269 spin_lock_bh(&obd->obd_processing_task_lock);
1270 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1271 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1272 spin_unlock_bh(&obd->obd_processing_task_lock);
1274 list_for_each_entry_safe(req, n, &clean_list, rq_list){
1275 LASSERT(req->rq_reply_state == 0);
1276 ptlrpc_free_clone(req);
1282 /* obd_processing_task_lock should be held */
1283 void target_cancel_recovery_timer(struct obd_device *obd)
1285 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1286 cfs_timer_disarm(&obd->obd_recovery_timer);
1289 /* extend = 1 means require at least "duration" seconds left in the timer,
1290 extend = 0 means set the total duration (start_recovery_timer) */
1291 static void reset_recovery_timer(struct obd_device *obd, int duration,
1294 cfs_time_t now = cfs_time_current_sec();
1295 cfs_duration_t left;
1297 spin_lock_bh(&obd->obd_processing_task_lock);
1298 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1299 spin_unlock_bh(&obd->obd_processing_task_lock);
1303 left = cfs_time_sub(obd->obd_recovery_end, now);
1305 if (extend && (duration > left))
1306 obd->obd_recovery_timeout += duration - left;
1307 else if (!extend && (duration > obd->obd_recovery_timeout))
1308 /* Track the client's largest expected replay time */
1309 obd->obd_recovery_timeout = duration;
1312 * If total recovery time already exceed the
1313 * obd_recovery_max_time, then CRAY XT3 will
1314 * abort the recovery
1316 if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1317 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1319 obd->obd_recovery_end = obd->obd_recovery_start +
1320 obd->obd_recovery_timeout;
1321 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1322 cfs_time_before(now, obd->obd_recovery_end)) {
1323 left = cfs_time_sub(obd->obd_recovery_end, now);
1324 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1326 spin_unlock_bh(&obd->obd_processing_task_lock);
1327 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1328 obd->obd_name, (unsigned)left);
1331 static void check_and_start_recovery_timer(struct obd_device *obd)
1333 spin_lock_bh(&obd->obd_processing_task_lock);
1334 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1335 spin_unlock_bh(&obd->obd_processing_task_lock);
1338 CWARN("%s: starting recovery timer\n", obd->obd_name);
1339 obd->obd_recovery_start = cfs_time_current_sec();
1341 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1342 spin_unlock_bh(&obd->obd_processing_task_lock);
1344 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1347 /* Reset the timer with each new client connection */
1349 * This timer is actually reconnect_timer, which is for making sure
1350 * the total recovery window is at least as big as my reconnect
1351 * attempt timing. So the initial recovery time_out will be set to
1352 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1353 * from client is bigger than this, then the recovery time_out will
1354 * be extend to make sure the client could be reconnected, in the
1355 * process, the timeout from the new client should be ignored.
1359 target_start_and_reset_recovery_timer(struct obd_device *obd,
1360 struct ptlrpc_request *req,
1363 int req_timeout = lustre_msg_get_timeout(req->rq_reqmsg);
1365 /* teach server about old server's estimates */
1367 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1368 at_timeout2est(req_timeout));
1370 check_and_start_recovery_timer(obd);
1372 req_timeout *= OBD_RECOVERY_FACTOR;
1373 if (req_timeout > obd->obd_recovery_timeout && !new_client)
1374 reset_recovery_timer(obd, req_timeout, 0);
1378 static int check_for_next_transno(struct obd_device *obd)
1380 struct ptlrpc_request *req = NULL;
1381 int wake_up = 0, connected, completed, queue_len, max;
1382 __u64 next_transno, req_transno;
1384 spin_lock_bh(&obd->obd_processing_task_lock);
1386 if (!list_empty(&obd->obd_req_replay_queue)) {
1387 req = list_entry(obd->obd_req_replay_queue.next,
1388 struct ptlrpc_request, rq_list);
1389 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1394 max = obd->obd_max_recoverable_clients;
1395 connected = obd->obd_connected_clients;
1396 completed = connected - obd->obd_recoverable_clients;
1397 queue_len = obd->obd_requests_queued_for_recovery;
1398 next_transno = obd->obd_next_recovery_transno;
1400 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1401 "req_transno: "LPU64", next_transno: "LPU64"\n",
1402 max, connected, completed, queue_len, req_transno, next_transno);
1404 if (obd->obd_abort_recovery) {
1405 CDEBUG(D_HA, "waking for aborted recovery\n");
1407 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1408 CDEBUG(D_HA, "waking for completed recovery\n");
1410 } else if (req_transno == next_transno) {
1411 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1413 } else if (queue_len + completed == max) {
1414 /* handle gaps occured due to lost reply. It is allowed gaps
1415 * because all clients are connected and there will be resend
1416 * for missed transaction */
1417 LASSERTF(req_transno >= next_transno,
1418 "req_transno: "LPU64", next_transno: "LPU64"\n",
1419 req_transno, next_transno);
1421 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1422 "waking for skipped transno (skip: "LPD64
1423 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1424 next_transno, queue_len, completed, connected, req_transno);
1425 obd->obd_next_recovery_transno = req_transno;
1427 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1428 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1429 " by fail_lock, waking up ("LPD64")\n", next_transno);
1430 obd->obd_next_recovery_transno = req_transno;
1432 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1433 /* some clients haven't connected in time, but we can try
1434 * to replay requests that demand on already committed ones
1435 * also, we can replay first non-committed transation */
1436 LASSERT(req_transno != 0);
1437 if (req_transno == obd->obd_last_committed + 1) {
1438 obd->obd_next_recovery_transno = req_transno;
1439 } else if (req_transno > obd->obd_last_committed) {
1440 /* can't continue recovery: have no needed transno */
1441 obd->obd_abort_recovery = 1;
1442 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1443 "connected: %d, completed: %d, queue_len: %d, "
1444 "req_transno: "LPU64", next_transno: "LPU64"\n",
1445 max, connected, completed, queue_len,
1446 req_transno, next_transno);
1451 spin_unlock_bh(&obd->obd_processing_task_lock);
1455 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1457 struct l_wait_info lwi = { 0 };
1458 struct ptlrpc_request *req;
1460 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1461 obd->obd_next_recovery_transno);
1462 l_wait_event(obd->obd_next_transno_waitq,
1463 check_for_next_transno(obd), &lwi);
1465 spin_lock_bh(&obd->obd_processing_task_lock);
1466 if (obd->obd_abort_recovery) {
1468 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1469 req = list_entry(obd->obd_req_replay_queue.next,
1470 struct ptlrpc_request, rq_list);
1471 target_exp_dequeue_req_replay(req);
1472 list_del_init(&req->rq_list);
1473 obd->obd_requests_queued_for_recovery--;
1477 spin_unlock_bh(&obd->obd_processing_task_lock);
1481 static int check_for_next_lock(struct obd_device *obd)
1483 struct ptlrpc_request *req = NULL;
1486 spin_lock_bh(&obd->obd_processing_task_lock);
1487 if (!list_empty(&obd->obd_lock_replay_queue)) {
1488 req = list_entry(obd->obd_lock_replay_queue.next,
1489 struct ptlrpc_request, rq_list);
1490 CDEBUG(D_HA, "waking for next lock\n");
1492 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1493 CDEBUG(D_HA, "waking for completed lock replay\n");
1495 } else if (obd->obd_abort_recovery) {
1496 CDEBUG(D_HA, "waking for aborted recovery\n");
1499 spin_unlock_bh(&obd->obd_processing_task_lock);
1504 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1506 struct l_wait_info lwi = { 0 };
1507 struct ptlrpc_request *req;
1509 CDEBUG(D_HA, "Waiting for lock\n");
1510 l_wait_event(obd->obd_next_transno_waitq,
1511 check_for_next_lock(obd), &lwi);
1513 spin_lock_bh(&obd->obd_processing_task_lock);
1514 if (obd->obd_abort_recovery) {
1516 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1517 req = list_entry(obd->obd_lock_replay_queue.next,
1518 struct ptlrpc_request, rq_list);
1519 list_del_init(&req->rq_list);
1523 spin_unlock_bh(&obd->obd_processing_task_lock);
1527 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1529 struct ptlrpc_request *req;
1531 spin_lock_bh(&obd->obd_processing_task_lock);
1532 if (!list_empty(&obd->obd_final_req_queue)) {
1533 req = list_entry(obd->obd_final_req_queue.next,
1534 struct ptlrpc_request, rq_list);
1535 list_del_init(&req->rq_list);
1539 spin_unlock_bh(&obd->obd_processing_task_lock);
1543 static inline int req_replay_done(struct obd_export *exp)
1545 return (exp->exp_req_replay_needed == 0);
1548 static inline int lock_replay_done(struct obd_export *exp)
1550 return (exp->exp_lock_replay_needed == 0);
1553 static inline int connect_done(struct obd_export *exp)
1555 return (exp->exp_in_recovery != 0);
1558 static int check_for_clients(struct obd_device *obd)
1560 if (obd->obd_abort_recovery)
1562 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1563 if (obd->obd_no_conn == 0 &&
1564 obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1569 static int handle_recovery_req(struct ptlrpc_thread *thread,
1570 struct ptlrpc_request *req,
1571 svc_handler_t handler)
1576 rc = lu_context_init(&req->rq_session, LCT_SESSION);
1578 CERROR("Failure to initialize session: %d\n", rc);
1581 req->rq_session.lc_thread = thread;
1582 lu_context_enter(&req->rq_session);
1583 req->rq_svc_thread = thread;
1584 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1586 /* thread context */
1587 lu_context_enter(&thread->t_env->le_ctx);
1589 lu_context_exit(&thread->t_env->le_ctx);
1591 lu_context_exit(&req->rq_session);
1592 lu_context_fini(&req->rq_session);
1593 /* don't reset timer for final stage */
1594 if (!req_replay_done(req->rq_export) ||
1595 !lock_replay_done(req->rq_export))
1596 reset_recovery_timer(class_exp2obd(req->rq_export),
1597 OBD_RECOVERY_FACTOR * AT_OFF ? obd_timeout :
1598 at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1599 ptlrpc_free_clone(req);
1603 static void resume_recovery_timer(struct obd_device *obd)
1605 /* to be safe, make it at least OBD_RECOVERY_FACTOR * obd_timeout */
1606 reset_recovery_timer(obd, OBD_RECOVERY_FACTOR * obd_timeout, 1);
1609 static int target_recovery_thread(void *arg)
1611 struct obd_device *obd = arg;
1612 struct ptlrpc_request *req;
1613 struct target_recovery_data *trd = &obd->obd_recovery_data;
1614 struct l_wait_info lwi = { 0 };
1615 unsigned long delta;
1616 unsigned long flags;
1618 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1622 cfs_daemonize("tgt_recov");
1624 SIGNAL_MASK_LOCK(current, flags);
1625 sigfillset(¤t->blocked);
1627 SIGNAL_MASK_UNLOCK(current, flags);
1629 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1633 thread->t_env = &env;
1634 env.le_ctx.lc_thread = thread;
1636 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1638 trd->trd_processing_task = current->pid;
1640 obd->obd_recovering = 1;
1641 complete(&trd->trd_starting);
1643 /* first of all, we have to know the first transno to replay */
1644 obd->obd_abort_recovery = 0;
1645 l_wait_event(obd->obd_next_transno_waitq,
1646 check_for_clients(obd), &lwi);
1648 spin_lock_bh(&obd->obd_processing_task_lock);
1649 target_cancel_recovery_timer(obd);
1650 spin_unlock_bh(&obd->obd_processing_task_lock);
1652 /* If some clients haven't connected in time, evict them */
1653 if (obd->obd_abort_recovery) {
1654 CWARN("Some clients haven't connect in time (%d/%d),"
1655 "evict them\n", obd->obd_connected_clients,
1656 obd->obd_max_recoverable_clients);
1657 obd->obd_abort_recovery = obd->obd_stopping;
1658 class_disconnect_stale_exports(obd, connect_done);
1660 /* next stage: replay requests */
1662 obd->obd_req_replaying = 1;
1663 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1664 atomic_read(&obd->obd_req_replay_clients),
1665 obd->obd_next_recovery_transno);
1666 resume_recovery_timer(obd);
1667 while ((req = target_next_replay_req(obd))) {
1668 LASSERT(trd->trd_processing_task == current->pid);
1669 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1670 lustre_msg_get_transno(req->rq_reqmsg),
1671 libcfs_nid2str(req->rq_peer.nid));
1672 handle_recovery_req(thread, req,
1673 trd->trd_recovery_handler);
1674 obd->obd_replayed_requests++;
1675 spin_lock_bh(&obd->obd_processing_task_lock);
1676 obd->obd_next_recovery_transno++;
1677 spin_unlock_bh(&obd->obd_processing_task_lock);
1680 spin_lock_bh(&obd->obd_processing_task_lock);
1681 target_cancel_recovery_timer(obd);
1682 spin_unlock_bh(&obd->obd_processing_task_lock);
1684 /* If some clients haven't replayed requests in time, evict them */
1685 if (obd->obd_abort_recovery) {
1686 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1687 obd->obd_abort_recovery = obd->obd_stopping;
1688 class_disconnect_stale_exports(obd, req_replay_done);
1689 abort_req_replay_queue(obd);
1692 /* The second stage: replay locks */
1693 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1694 atomic_read(&obd->obd_lock_replay_clients));
1695 resume_recovery_timer(obd);
1696 while ((req = target_next_replay_lock(obd))) {
1697 LASSERT(trd->trd_processing_task == current->pid);
1698 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1699 libcfs_nid2str(req->rq_peer.nid));
1700 handle_recovery_req(thread, req,
1701 trd->trd_recovery_handler);
1702 obd->obd_replayed_locks++;
1705 spin_lock_bh(&obd->obd_processing_task_lock);
1706 target_cancel_recovery_timer(obd);
1707 spin_unlock_bh(&obd->obd_processing_task_lock);
1708 /* If some clients haven't replayed requests in time, evict them */
1709 if (obd->obd_abort_recovery) {
1711 CERROR("lock replay timed out, aborting ...\n");
1712 obd->obd_abort_recovery = obd->obd_stopping;
1713 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1714 abort_lock_replay_queue(obd);
1717 /* We drop recoverying flag to forward all new requests
1718 * to regular mds_handle() since now */
1719 spin_lock_bh(&obd->obd_processing_task_lock);
1720 obd->obd_recovering = obd->obd_abort_recovery = 0;
1721 spin_unlock_bh(&obd->obd_processing_task_lock);
1722 /* The third stage: reply on final pings */
1723 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1724 while ((req = target_next_final_ping(obd))) {
1725 LASSERT(trd->trd_processing_task == current->pid);
1726 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1727 libcfs_nid2str(req->rq_peer.nid));
1728 handle_recovery_req(thread, req,
1729 trd->trd_recovery_handler);
1732 delta = (jiffies - delta) / HZ;
1733 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1734 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1735 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1736 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1737 if (delta > obd_timeout * 2) {
1738 CWARN("too long recovery - read logs\n");
1739 libcfs_debug_dumplog();
1742 target_finish_recovery(obd);
1744 lu_context_fini(&env.le_ctx);
1745 trd->trd_processing_task = 0;
1746 complete(&trd->trd_finishing);
1750 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1753 struct target_recovery_data *trd = &obd->obd_recovery_data;
1755 memset(trd, 0, sizeof(*trd));
1756 init_completion(&trd->trd_starting);
1757 init_completion(&trd->trd_finishing);
1758 trd->trd_recovery_handler = handler;
1760 if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1761 wait_for_completion(&trd->trd_starting);
1762 LASSERT(obd->obd_recovering != 0);
1769 void target_stop_recovery_thread(struct obd_device *obd)
1771 spin_lock_bh(&obd->obd_processing_task_lock);
1772 if (obd->obd_recovery_data.trd_processing_task > 0) {
1773 struct target_recovery_data *trd = &obd->obd_recovery_data;
1774 CERROR("%s: Aborting recovery\n", obd->obd_name);
1775 obd->obd_abort_recovery = 1;
1776 wake_up(&obd->obd_next_transno_waitq);
1777 spin_unlock_bh(&obd->obd_processing_task_lock);
1778 wait_for_completion(&trd->trd_finishing);
1780 spin_unlock_bh(&obd->obd_processing_task_lock);
1784 void target_recovery_fini(struct obd_device *obd)
1786 class_disconnect_exports(obd);
1787 target_stop_recovery_thread(obd);
1788 target_cleanup_recovery(obd);
1790 EXPORT_SYMBOL(target_recovery_fini);
1792 static void target_recovery_expired(unsigned long castmeharder)
1794 struct obd_device *obd = (struct obd_device *)castmeharder;
1795 LCONSOLE_WARN("%s: recovery timed out; %d clients never reconnected "
1796 "after %lds (%d clients did)\n",
1797 obd->obd_name, obd->obd_recoverable_clients,
1798 cfs_time_current_sec()- obd->obd_recovery_start,
1799 obd->obd_connected_clients);
1800 spin_lock_bh(&obd->obd_processing_task_lock);
1801 if (obd->obd_recovering)
1802 obd->obd_abort_recovery = 1;
1803 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1804 spin_unlock_bh(&obd->obd_processing_task_lock);
1807 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1809 if (obd->obd_max_recoverable_clients == 0)
1812 CWARN("RECOVERY: service %s, %d recoverable clients, "
1813 "last_transno "LPU64"\n", obd->obd_name,
1814 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1815 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1816 obd->obd_recovery_start = 0;
1817 obd->obd_recovery_end = 0;
1818 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1819 /* bz13079: this should be set to desired value for ost but not for mds */
1820 obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1821 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1822 target_start_recovery_thread(obd, handler);
1824 EXPORT_SYMBOL(target_recovery_init);
1828 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1830 struct obd_export *exp = req->rq_export;
1831 LASSERT(exp != NULL);
1832 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1833 /* client declares he's ready to replay locks */
1834 spin_lock_bh(&obd->obd_processing_task_lock);
1835 if (exp->exp_req_replay_needed) {
1836 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1837 spin_lock(&exp->exp_lock);
1838 exp->exp_req_replay_needed = 0;
1839 spin_unlock(&exp->exp_lock);
1840 atomic_dec(&obd->obd_req_replay_clients);
1841 LASSERT(obd->obd_recoverable_clients > 0);
1842 obd->obd_recoverable_clients--;
1843 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1844 CDEBUG(D_HA, "all clients have replayed reqs\n");
1845 wake_up(&obd->obd_next_transno_waitq);
1847 spin_unlock_bh(&obd->obd_processing_task_lock);
1849 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1850 /* client declares he's ready to complete recovery
1851 * so, we put the request on th final queue */
1852 spin_lock_bh(&obd->obd_processing_task_lock);
1853 if (exp->exp_lock_replay_needed) {
1854 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1855 spin_lock(&exp->exp_lock);
1856 exp->exp_lock_replay_needed = 0;
1857 spin_unlock(&exp->exp_lock);
1858 atomic_dec(&obd->obd_lock_replay_clients);
1859 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1860 CDEBUG(D_HA, "all clients have replayed locks\n");
1861 wake_up(&obd->obd_next_transno_waitq);
1863 spin_unlock_bh(&obd->obd_processing_task_lock);
1869 int target_queue_recovery_request(struct ptlrpc_request *req,
1870 struct obd_device *obd)
1872 struct list_head *tmp;
1874 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1878 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1879 /* Processing the queue right now, don't re-add. */
1883 target_process_req_flags(obd, req);
1885 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1886 /* client declares he's ready to complete recovery
1887 * so, we put the request on th final queue */
1888 req = ptlrpc_clone_req(req);
1891 DEBUG_REQ(D_HA, req, "queue final req");
1892 spin_lock_bh(&obd->obd_processing_task_lock);
1893 if (obd->obd_recovering)
1894 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1896 spin_unlock_bh(&obd->obd_processing_task_lock);
1897 ptlrpc_free_clone(req);
1898 if (obd->obd_stopping) {
1904 spin_unlock_bh(&obd->obd_processing_task_lock);
1907 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1908 /* client declares he's ready to replay locks */
1909 req = ptlrpc_clone_req(req);
1912 DEBUG_REQ(D_HA, req, "queue lock replay req");
1913 spin_lock_bh(&obd->obd_processing_task_lock);
1914 LASSERT(obd->obd_recovering);
1915 /* usually due to recovery abort */
1916 if (!req->rq_export->exp_in_recovery) {
1917 spin_unlock_bh(&obd->obd_processing_task_lock);
1918 ptlrpc_free_clone(req);
1921 LASSERT(req->rq_export->exp_lock_replay_needed);
1922 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1923 spin_unlock_bh(&obd->obd_processing_task_lock);
1924 wake_up(&obd->obd_next_transno_waitq);
1928 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1929 * (i.e. buflens etc are in my own byte order), but type-dependent
1930 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1933 CFS_INIT_LIST_HEAD(&req->rq_list);
1934 DEBUG_REQ(D_HA, req, "not queueing");
1938 spin_lock_bh(&obd->obd_processing_task_lock);
1940 /* If we're processing the queue, we want don't want to queue this
1943 * Also, if this request has a transno less than the one we're waiting
1944 * for, we should process it now. It could (and currently always will)
1945 * be an open request for a descriptor that was opened some time ago.
1947 * Also, a resent, replayed request that has already been
1948 * handled will pass through here and be processed immediately.
1950 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1951 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1952 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1953 /* Processing the queue right now, don't re-add. */
1954 LASSERT(list_empty(&req->rq_list));
1955 spin_unlock_bh(&obd->obd_processing_task_lock);
1958 spin_unlock_bh(&obd->obd_processing_task_lock);
1960 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
1963 req = ptlrpc_clone_req(req);
1967 spin_lock_bh(&obd->obd_processing_task_lock);
1968 LASSERT(obd->obd_recovering);
1969 if (!req->rq_export->exp_in_recovery) {
1970 spin_unlock_bh(&obd->obd_processing_task_lock);
1971 ptlrpc_free_clone(req);
1974 LASSERT(req->rq_export->exp_req_replay_needed);
1976 if (target_exp_enqueue_req_replay(req)) {
1977 spin_unlock_bh(&obd->obd_processing_task_lock);
1978 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1979 ptlrpc_free_clone(req);
1984 list_for_each(tmp, &obd->obd_req_replay_queue) {
1985 struct ptlrpc_request *reqiter =
1986 list_entry(tmp, struct ptlrpc_request, rq_list);
1988 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1989 list_add_tail(&req->rq_list, &reqiter->rq_list);
1994 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
1996 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
1997 "has been claimed by another client");
1998 spin_unlock_bh(&obd->obd_processing_task_lock);
1999 target_exp_dequeue_req_replay(req);
2000 ptlrpc_free_clone(req);
2006 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2008 obd->obd_requests_queued_for_recovery++;
2009 wake_up(&obd->obd_next_transno_waitq);
2010 spin_unlock_bh(&obd->obd_processing_task_lock);
2014 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2016 return req->rq_export->exp_obd;
2019 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2021 LASSERT(exp != NULL);
2022 return &exp->exp_obd->obd_namespace->ns_pool;
2026 * Packs current SLV and Limit into \a req.
2028 int target_pack_pool_reply(struct ptlrpc_request *req)
2030 struct obd_device *obd;
2034 * Check that we still have all structures alive as this may
2035 * be some late rpc in shutdown time.
2037 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2038 !exp_connect_lru_resize(req->rq_export))) {
2039 lustre_msg_set_slv(req->rq_repmsg, 0);
2040 lustre_msg_set_limit(req->rq_repmsg, 0);
2045 * OBD is alive here as export is alive, which we checked above.
2047 obd = req->rq_export->exp_obd;
2049 read_lock(&obd->obd_pool_lock);
2050 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2051 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2052 read_unlock(&obd->obd_pool_lock);
2057 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2059 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2060 DEBUG_REQ(D_ERROR, req, "dropping reply");
2065 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2066 req->rq_status = rc;
2067 return (ptlrpc_send_error(req, 1));
2069 DEBUG_REQ(D_NET, req, "sending reply");
2072 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2075 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2078 struct ptlrpc_reply_state *rs;
2079 struct obd_device *obd;
2080 struct obd_export *exp;
2081 struct ptlrpc_service *svc;
2084 if (req->rq_no_reply) {
2089 svc = req->rq_rqbd->rqbd_service;
2090 rs = req->rq_reply_state;
2091 if (rs == NULL || !rs->rs_difficult) {
2093 target_send_reply_msg (req, rc, fail_id);
2098 /* must be an export if locks saved */
2099 LASSERT (req->rq_export != NULL);
2100 /* req/reply consistent */
2101 LASSERT (rs->rs_service == svc);
2104 LASSERT (!rs->rs_scheduled);
2105 LASSERT (!rs->rs_scheduled_ever);
2106 LASSERT (!rs->rs_handled);
2107 LASSERT (!rs->rs_on_net);
2108 LASSERT (rs->rs_export == NULL);
2109 LASSERT (list_empty(&rs->rs_obd_list));
2110 LASSERT (list_empty(&rs->rs_exp_list));
2112 exp = class_export_get (req->rq_export);
2115 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2116 rs->rs_scheduled = 1;
2118 rs->rs_xid = req->rq_xid;
2119 rs->rs_transno = req->rq_transno;
2120 rs->rs_export = exp;
2122 spin_lock(&obd->obd_uncommitted_replies_lock);
2124 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2125 rs->rs_transno, obd->obd_last_committed);
2126 if (rs->rs_transno > obd->obd_last_committed) {
2127 /* not committed already */
2128 list_add_tail (&rs->rs_obd_list,
2129 &obd->obd_uncommitted_replies);
2132 spin_unlock (&obd->obd_uncommitted_replies_lock);
2133 spin_lock (&exp->exp_lock);
2135 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
2137 spin_unlock(&exp->exp_lock);
2139 netrc = target_send_reply_msg (req, rc, fail_id);
2141 spin_lock(&svc->srv_lock);
2143 svc->srv_n_difficult_replies++;
2146 /* error sending: reply is off the net. Also we need +1
2147 * reply ref until ptlrpc_server_handle_reply() is done
2148 * with the reply state (if the send was successful, there
2149 * would have been +1 ref for the net, which
2150 * reply_out_callback leaves alone) */
2152 ptlrpc_rs_addref(rs);
2153 atomic_inc (&svc->srv_outstanding_replies);
2156 if (rs->rs_transno <= obd->obd_last_committed ||
2157 (!rs->rs_on_net && !rs->rs_no_ack) ||
2158 list_empty(&rs->rs_exp_list) || /* completed already */
2159 list_empty(&rs->rs_obd_list)) {
2160 CDEBUG(D_HA, "Schedule reply immediately\n");
2161 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
2162 cfs_waitq_signal (&svc->srv_waitq);
2164 list_add (&rs->rs_list, &svc->srv_active_replies);
2165 rs->rs_scheduled = 0; /* allow notifier to schedule */
2168 spin_unlock(&svc->srv_lock);
2172 int target_handle_ping(struct ptlrpc_request *req)
2174 obd_ping(req->rq_export);
2175 return req_capsule_server_pack(&req->rq_pill);
2178 void target_committed_to_req(struct ptlrpc_request *req)
2180 struct obd_device *obd;
2182 if (req == NULL || req->rq_export == NULL)
2185 obd = req->rq_export->exp_obd;
2189 if (!obd->obd_no_transno && req->rq_repmsg != NULL)
2190 lustre_msg_set_last_committed(req->rq_repmsg,
2191 obd->obd_last_committed);
2193 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2194 "%d)", obd->obd_no_transno, req->rq_repmsg == NULL);
2196 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2197 obd->obd_last_committed, req->rq_transno, req->rq_xid);
2200 EXPORT_SYMBOL(target_committed_to_req);
2202 int target_handle_qc_callback(struct ptlrpc_request *req)
2204 struct obd_quotactl *oqctl;
2205 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2207 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2208 if (oqctl == NULL) {
2209 CERROR("Can't unpack obd_quotactl\n");
2213 cli->cl_qchk_stat = oqctl->qc_stat;
2218 #ifdef HAVE_QUOTA_SUPPORT
2219 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2222 struct obd_device *obd = req->rq_export->exp_obd;
2223 struct obd_device *master_obd;
2224 struct obd_device_target *obt;
2225 struct lustre_quota_ctxt *qctxt;
2226 struct qunit_data *qdata = NULL;
2230 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2233 rc = req_capsule_server_pack(&req->rq_pill);
2235 CERROR("packing reply failed!: rc = %d\n", rc);
2239 LASSERT(req->rq_export);
2241 OBD_ALLOC(qdata, sizeof(struct qunit_data));
2244 rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
2246 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2250 /* we use the observer */
2251 if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2252 CERROR("Can't find the observer, it is recovering\n");
2253 req->rq_status = -EIO;
2254 GOTO(send_reply, rc = -EIO);
2257 master_obd = obd->obd_observer->obd_observer;
2258 obt = &master_obd->u.obt;
2259 qctxt = &obt->obt_qctxt;
2261 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2262 /* quota_type has not been processed yet, return EAGAIN
2263 * until we know whether or not quotas are supposed to
2265 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2267 req->rq_status = -EAGAIN;
2268 rc = ptlrpc_reply(req);
2272 down_read(&obt->obt_rwsem);
2273 if (qctxt->lqc_lqs_hash == NULL) {
2274 up_read(&obt->obt_rwsem);
2275 /* quota_type has not been processed yet, return EAGAIN
2276 * until we know whether or not quotas are supposed to
2278 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2280 req->rq_status = -EAGAIN;
2281 rc = ptlrpc_reply(req);
2285 LASSERT(qctxt->lqc_handler);
2286 rc = qctxt->lqc_handler(master_obd, qdata,
2287 lustre_msg_get_opc(req->rq_reqmsg));
2288 up_read(&obt->obt_rwsem);
2289 if (rc && rc != -EDQUOT)
2290 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2291 "dqacq failed! (rc:%d)\n", rc);
2292 req->rq_status = rc;
2294 /* there are three forms of qunit(historic causes), so we need to
2295 * adjust the same form to different forms slaves needed */
2296 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2298 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2302 /* Block the quota req. b=14840 */
2303 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2305 rc = ptlrpc_reply(req);
2307 OBD_FREE(qdata, sizeof(struct qunit_data));
2311 #endif /* !__KERNEL__ */
2313 #endif /* HAVE_QUOTA_SUPPORT */
2315 ldlm_mode_t lck_compat_array[] = {
2316 [LCK_EX] LCK_COMPAT_EX,
2317 [LCK_PW] LCK_COMPAT_PW,
2318 [LCK_PR] LCK_COMPAT_PR,
2319 [LCK_CW] LCK_COMPAT_CW,
2320 [LCK_CR] LCK_COMPAT_CR,
2321 [LCK_NL] LCK_COMPAT_NL,
2322 [LCK_GROUP] LCK_COMPAT_GROUP,
2323 [LCK_COS] LCK_COMPAT_COS,
2327 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2328 * not escape to the user level.
2330 int ldlm_error2errno(ldlm_error_t error)
2338 case ELDLM_LOCK_CHANGED:
2341 case ELDLM_LOCK_ABORTED:
2344 case ELDLM_LOCK_REPLACED:
2347 case ELDLM_NO_LOCK_DATA:
2350 case ELDLM_NAMESPACE_EXISTS:
2353 case ELDLM_BAD_NAMESPACE:
2357 if (((int)error) < 0) /* cast to signed type */
2358 result = error; /* as ldlm_error_t can be unsigned */
2360 CERROR("Invalid DLM result code: %i\n", error);
2366 EXPORT_SYMBOL(ldlm_error2errno);
2369 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2371 ldlm_error_t ldlm_errno2error(int err_no)
2380 error = ELDLM_LOCK_CHANGED;
2383 error = ELDLM_LOCK_ABORTED;
2386 error = ELDLM_LOCK_REPLACED;
2389 error = ELDLM_NO_LOCK_DATA;
2392 error = ELDLM_NAMESPACE_EXISTS;
2395 error = ELDLM_BAD_NAMESPACE;
2402 EXPORT_SYMBOL(ldlm_errno2error);