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"
55 /* @priority: if non-zero, move the selected to the list head
56 * @create: if zero, only search in existed connections
58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
59 int priority, int create)
61 struct ptlrpc_connection *ptlrpc_conn;
62 struct obd_import_conn *imp_conn = NULL, *item;
66 if (!create && !priority) {
67 CDEBUG(D_HA, "Nothing to do\n");
71 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
73 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
78 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
80 GOTO(out_put, rc = -ENOMEM);
84 spin_lock(&imp->imp_lock);
85 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
86 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
88 list_del(&item->oic_item);
89 list_add(&item->oic_item, &imp->imp_conn_list);
90 item->oic_last_attempt = 0;
92 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
93 imp, imp->imp_obd->obd_name, uuid->uuid,
94 (priority ? ", moved to head" : ""));
95 spin_unlock(&imp->imp_lock);
96 GOTO(out_free, rc = 0);
101 imp_conn->oic_conn = ptlrpc_conn;
102 imp_conn->oic_uuid = *uuid;
103 item->oic_last_attempt = 0;
105 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107 list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
108 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
109 imp, imp->imp_obd->obd_name, uuid->uuid,
110 (priority ? "head" : "tail"));
112 spin_unlock(&imp->imp_lock);
113 GOTO(out_free, rc = -ENOENT);
116 spin_unlock(&imp->imp_lock);
120 OBD_FREE(imp_conn, sizeof(*imp_conn));
122 ptlrpc_connection_put(ptlrpc_conn);
126 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
128 return import_set_conn(imp, uuid, 1, 0);
131 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
134 return import_set_conn(imp, uuid, priority, 1);
137 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
139 struct obd_import_conn *imp_conn;
140 struct obd_export *dlmexp;
144 spin_lock(&imp->imp_lock);
145 if (list_empty(&imp->imp_conn_list)) {
146 LASSERT(!imp->imp_connection);
150 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
151 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
153 LASSERT(imp_conn->oic_conn);
155 /* is current conn? */
156 if (imp_conn == imp->imp_conn_current) {
157 LASSERT(imp_conn->oic_conn == imp->imp_connection);
159 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
160 imp->imp_state != LUSTRE_IMP_DISCON) {
161 CERROR("can't remove current connection\n");
162 GOTO(out, rc = -EBUSY);
165 ptlrpc_connection_put(imp->imp_connection);
166 imp->imp_connection = NULL;
168 dlmexp = class_conn2export(&imp->imp_dlm_handle);
169 if (dlmexp && dlmexp->exp_connection) {
170 LASSERT(dlmexp->exp_connection ==
172 ptlrpc_connection_put(dlmexp->exp_connection);
173 dlmexp->exp_connection = NULL;
177 list_del(&imp_conn->oic_item);
178 ptlrpc_connection_put(imp_conn->oic_conn);
179 OBD_FREE(imp_conn, sizeof(*imp_conn));
180 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
181 imp, imp->imp_obd->obd_name, uuid->uuid);
186 spin_unlock(&imp->imp_lock);
188 CERROR("connection %s not found\n", uuid->uuid);
192 static void destroy_import(struct obd_import *imp)
194 /* drop security policy instance after all rpc finished/aborted
195 * to let all busy contexts be released. */
196 class_import_get(imp);
197 class_destroy_import(imp);
198 sptlrpc_import_sec_put(imp);
199 class_import_put(imp);
202 /* configure an RPC client OBD device
207 * 3 - inactive-on-startup
209 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
211 struct client_obd *cli = &obddev->u.cli;
212 struct obd_import *imp;
213 struct obd_uuid server_uuid;
214 int rq_portal, rp_portal, connect_op;
215 char *name = obddev->obd_type->typ_name;
219 /* In a more perfect world, we would hang a ptlrpc_client off of
220 * obd_type and just use the values from there. */
221 if (!strcmp(name, LUSTRE_OSC_NAME)) {
222 rq_portal = OST_REQUEST_PORTAL;
223 rp_portal = OSC_REPLY_PORTAL;
224 connect_op = OST_CONNECT;
225 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
226 rq_portal = MDS_REQUEST_PORTAL;
227 rp_portal = MDC_REPLY_PORTAL;
228 connect_op = MDS_CONNECT;
229 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
230 rq_portal = MGS_REQUEST_PORTAL;
231 rp_portal = MGC_REPLY_PORTAL;
232 connect_op = MGS_CONNECT;
234 CERROR("unknown client OBD type \"%s\", can't setup\n",
239 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
240 CERROR("requires a TARGET UUID\n");
244 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
245 CERROR("client UUID must be less than 38 characters\n");
249 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
250 CERROR("setup requires a SERVER UUID\n");
254 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
255 CERROR("target UUID must be less than 38 characters\n");
259 init_rwsem(&cli->cl_sem);
260 sema_init(&cli->cl_mgc_sem, 1);
261 sptlrpc_rule_set_init(&cli->cl_sptlrpc_rset);
262 cli->cl_sec_part = LUSTRE_SP_ANY;
263 cli->cl_conn_count = 0;
264 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
265 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
266 sizeof(server_uuid)));
269 cli->cl_avail_grant = 0;
270 /* FIXME: should limit this for the sum of all cl_dirty_max */
271 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
272 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
273 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
274 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
275 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
276 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
277 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
278 client_obd_list_lock_init(&cli->cl_loi_list_lock);
279 cli->cl_r_in_flight = 0;
280 cli->cl_w_in_flight = 0;
282 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
283 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
284 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
285 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
286 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
287 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
288 cfs_waitq_init(&cli->cl_destroy_waitq);
289 atomic_set(&cli->cl_destroy_in_flight, 0);
290 #ifdef ENABLE_CHECKSUM
291 /* Turn on checksumming by default. */
292 cli->cl_checksum = 1;
294 * The supported checksum types will be worked out at connect time
295 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
298 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
300 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
302 /* This value may be changed at connect time in
303 ptlrpc_connect_interpret. */
304 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
305 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
307 if (!strcmp(name, LUSTRE_MDC_NAME)) {
308 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
309 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
310 cli->cl_max_rpcs_in_flight = 2;
311 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
312 cli->cl_max_rpcs_in_flight = 3;
313 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
314 cli->cl_max_rpcs_in_flight = 4;
316 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
321 CERROR("ldlm_get_ref failed: %d\n", rc);
325 ptlrpc_init_client(rq_portal, rp_portal, name,
326 &obddev->obd_ldlm_client);
328 imp = class_new_import(obddev);
330 GOTO(err_ldlm, rc = -ENOENT);
331 imp->imp_client = &obddev->obd_ldlm_client;
332 imp->imp_connect_op = connect_op;
333 imp->imp_initial_recov = 1;
334 imp->imp_initial_recov_bk = 0;
335 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
336 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
337 LUSTRE_CFG_BUFLEN(lcfg, 1));
338 class_import_put(imp);
340 rc = client_import_add_conn(imp, &server_uuid, 1);
342 CERROR("can't add initial connection\n");
343 GOTO(err_import, rc);
346 cli->cl_import = imp;
347 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
348 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
349 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
351 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
352 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
353 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
354 name, obddev->obd_name,
355 cli->cl_target_uuid.uuid);
356 spin_lock(&imp->imp_lock);
357 imp->imp_invalid = 1;
358 spin_unlock(&imp->imp_lock);
362 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
367 class_destroy_import(imp);
375 int client_obd_cleanup(struct obd_device *obddev)
378 sptlrpc_rule_set_free(&obddev->u.cli.cl_sptlrpc_rset);
383 /* ->o_connect() method for client side (OSC and MDC and MGC) */
384 int client_connect_import(const struct lu_env *env,
385 struct lustre_handle *dlm_handle,
386 struct obd_device *obd, struct obd_uuid *cluuid,
387 struct obd_connect_data *data, void *localdata)
389 struct client_obd *cli = &obd->u.cli;
390 struct obd_import *imp = cli->cl_import;
391 struct obd_export *exp;
392 struct obd_connect_data *ocd;
393 struct ldlm_namespace *to_be_freed = NULL;
397 down_write(&cli->cl_sem);
398 rc = class_connect(dlm_handle, obd, cluuid);
402 cli->cl_conn_count++;
403 if (cli->cl_conn_count > 1)
405 exp = class_conn2export(dlm_handle);
407 if (obd->obd_namespace != NULL)
408 CERROR("already have namespace!\n");
409 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
410 LDLM_NAMESPACE_CLIENT,
411 LDLM_NAMESPACE_GREEDY);
412 if (obd->obd_namespace == NULL)
413 GOTO(out_disco, rc = -ENOMEM);
415 imp->imp_dlm_handle = *dlm_handle;
416 rc = ptlrpc_init_import(imp);
420 ocd = &imp->imp_connect_data;
423 imp->imp_connect_flags_orig = data->ocd_connect_flags;
426 rc = ptlrpc_connect_import(imp, NULL);
428 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
431 LASSERT(exp->exp_connection);
434 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
435 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
436 data->ocd_connect_flags, ocd->ocd_connect_flags);
437 data->ocd_connect_flags = ocd->ocd_connect_flags;
440 ptlrpc_pinger_add_import(imp);
446 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
447 to_be_freed = obd->obd_namespace;
448 obd->obd_namespace = NULL;
450 cli->cl_conn_count--;
451 class_disconnect(exp);
453 class_export_put(exp);
456 up_write(&cli->cl_sem);
458 ldlm_namespace_free_post(to_be_freed);
463 int client_disconnect_export(struct obd_export *exp)
465 struct obd_device *obd = class_exp2obd(exp);
466 struct client_obd *cli;
467 struct obd_import *imp;
469 struct ldlm_namespace *to_be_freed = NULL;
473 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
474 exp, exp ? exp->exp_handle.h_cookie : -1);
479 imp = cli->cl_import;
481 down_write(&cli->cl_sem);
482 if (!cli->cl_conn_count) {
483 CERROR("disconnecting disconnected device (%s)\n",
485 GOTO(out_sem, rc = -EINVAL);
488 cli->cl_conn_count--;
489 if (cli->cl_conn_count)
490 GOTO(out_no_disconnect, rc = 0);
492 /* Mark import deactivated now, so we don't try to reconnect if any
493 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
494 * fully deactivate the import, or that would drop all requests. */
495 spin_lock(&imp->imp_lock);
496 imp->imp_deactive = 1;
497 spin_unlock(&imp->imp_lock);
499 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
500 * delete it regardless. (It's safe to delete an import that was
502 (void)ptlrpc_pinger_del_import(imp);
504 if (obd->obd_namespace != NULL) {
505 /* obd_force == local only */
506 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
507 obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
509 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
510 to_be_freed = obd->obd_namespace;
513 rc = ptlrpc_disconnect_import(imp, 0);
515 ptlrpc_invalidate_import(imp);
516 /* set obd_namespace to NULL only after invalidate, because we can have
517 * some connect requests in flight, and his need store a connect flags
518 * in obd_namespace. bug 14260 */
519 obd->obd_namespace = NULL;
521 ptlrpc_free_rq_pool(imp->imp_rq_pool);
523 cli->cl_import = NULL;
527 err = class_disconnect(exp);
531 up_write(&cli->cl_sem);
533 ldlm_namespace_free_post(to_be_freed);
538 /* --------------------------------------------------------------------------
539 * from old lib/target.c
540 * -------------------------------------------------------------------------- */
542 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
543 struct obd_uuid *cluuid, int initial_conn)
546 if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
547 struct lustre_handle *hdl;
548 hdl = &exp->exp_imp_reverse->imp_remote_handle;
549 /* Might be a re-connect after a partition. */
550 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
551 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
553 conn->cookie = exp->exp_handle.h_cookie;
554 /* target_handle_connect() treats EALREADY and
555 * -EALREADY differently. EALREADY means we are
556 * doing a valid reconnect from the same client. */
559 CERROR("%s reconnecting from %s, "
560 "handle mismatch (ours "LPX64", theirs "
561 LPX64")\n", cluuid->uuid,
562 exp->exp_connection->c_remote_uuid.uuid,
563 hdl->cookie, conn->cookie);
564 memset(conn, 0, sizeof *conn);
565 /* target_handle_connect() treats EALREADY and
566 * -EALREADY differently. -EALREADY is an error
567 * (same UUID, different handle). */
572 conn->cookie = exp->exp_handle.h_cookie;
573 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
574 cluuid->uuid, exp, conn->cookie);
578 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
581 struct obd_export *exp = cb_data;
583 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
584 obd->obd_name, exp->exp_client_uuid.uuid);
586 spin_lock(&exp->exp_lock);
587 exp->exp_need_sync = 0;
588 spin_unlock(&exp->exp_lock);
590 EXPORT_SYMBOL(target_client_add_cb);
592 target_start_and_reset_recovery_timer(struct obd_device *obd,
593 struct ptlrpc_request *req,
596 int target_handle_connect(struct ptlrpc_request *req)
598 struct obd_device *target, *targref = NULL;
599 struct obd_export *export = NULL;
600 struct obd_import *revimp;
601 struct lustre_handle conn;
602 struct lustre_handle *tmp;
603 struct obd_uuid tgtuuid;
604 struct obd_uuid cluuid;
605 struct obd_uuid remote_uuid;
608 int initial_conn = 0;
609 struct obd_connect_data *data, *tmpdata;
610 lnet_nid_t *client_nid = NULL;
613 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
615 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
617 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
618 GOTO(out, rc = -EINVAL);
621 obd_str2uuid(&tgtuuid, str);
622 target = class_uuid2obd(&tgtuuid);
624 target = class_name2obd(str);
626 if (!target || target->obd_stopping || !target->obd_set_up) {
627 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
628 " for connect (%s)\n", str,
629 !target ? "no target" :
630 (target->obd_stopping ? "stopping" :
632 GOTO(out, rc = -ENODEV);
635 if (target->obd_no_conn) {
636 LCONSOLE_WARN("%s: temporarily refusing client connection "
637 "from %s\n", target->obd_name,
638 libcfs_nid2str(req->rq_peer.nid));
639 GOTO(out, rc = -EAGAIN);
642 /* Make sure the target isn't cleaned up while we're here. Yes,
643 there's still a race between the above check and our incref here.
644 Really, class_uuid2obd should take the ref. */
645 targref = class_incref(target, __FUNCTION__, cfs_current());
648 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
650 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
651 GOTO(out, rc = -EINVAL);
654 obd_str2uuid(&cluuid, str);
656 /* XXX extract a nettype and format accordingly */
657 switch (sizeof(lnet_nid_t)) {
658 /* NB the casts only avoid compiler warnings */
660 snprintf(remote_uuid.uuid, sizeof remote_uuid,
661 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
664 snprintf(remote_uuid.uuid, sizeof remote_uuid,
665 "NET_%x_UUID", (__u32)req->rq_peer.nid);
671 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
673 GOTO(out, rc = -EPROTO);
677 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
679 GOTO(out, rc = -EPROTO);
681 rc = req_capsule_server_pack(&req->rq_pill);
685 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
687 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
688 "libclient connection attempt");
689 GOTO(out, rc = -EPROTO);
690 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
691 LUSTRE_VERSION_ALLOWED_OFFSET ||
692 data->ocd_version > LUSTRE_VERSION_CODE +
693 LUSTRE_VERSION_ALLOWED_OFFSET) {
694 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
695 "libclient connection attempt",
696 data->ocd_version < LUSTRE_VERSION_CODE ?
698 OBD_OCD_VERSION_MAJOR(data->ocd_version),
699 OBD_OCD_VERSION_MINOR(data->ocd_version),
700 OBD_OCD_VERSION_PATCH(data->ocd_version),
701 OBD_OCD_VERSION_FIX(data->ocd_version));
702 data = req_capsule_server_sized_get(&req->rq_pill,
704 offsetof(typeof(*data), ocd_version) +
705 sizeof(data->ocd_version));
707 data->ocd_connect_flags = OBD_CONNECT_VERSION;
708 data->ocd_version = LUSTRE_VERSION_CODE;
710 GOTO(out, rc = -EPROTO);
714 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
717 /* lctl gets a backstage, all-access pass. */
718 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
719 goto dont_check_exports;
721 spin_lock(&target->obd_dev_lock);
722 export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
724 if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
725 CWARN("%s: exp %p already connecting\n",
726 export->exp_obd->obd_name, export);
727 class_export_put(export);
730 } else if (export != NULL && export->exp_connection != NULL &&
731 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
732 /* make darn sure this is coming from the same peer
733 * if the UUIDs matched */
734 CWARN("%s: cookie %s seen on new NID %s when "
735 "existing NID %s is already connected\n",
736 target->obd_name, cluuid.uuid,
737 libcfs_nid2str(req->rq_peer.nid),
738 libcfs_nid2str(export->exp_connection->c_peer.nid));
739 class_export_put(export);
742 } else if (export != NULL) {
743 spin_lock(&export->exp_lock);
744 export->exp_connecting = 1;
745 spin_unlock(&export->exp_lock);
746 class_export_put(export);
747 spin_unlock(&target->obd_dev_lock);
748 LASSERT(export->exp_obd == target);
750 rc = target_handle_reconnect(&conn, export, &cluuid, initial_conn);
753 /* If we found an export, we already unlocked. */
755 spin_unlock(&target->obd_dev_lock);
756 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
757 } else if (req->rq_export == NULL &&
758 atomic_read(&export->exp_rpc_count) > 0) {
759 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
760 target->obd_name, cluuid.uuid,
761 libcfs_nid2str(req->rq_peer.nid),
762 export, atomic_read(&export->exp_refcount));
763 GOTO(out, rc = -EBUSY);
764 } else if (req->rq_export != NULL &&
765 (atomic_read(&export->exp_rpc_count) > 1)) {
766 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
767 target->obd_name, cluuid.uuid,
768 libcfs_nid2str(req->rq_peer.nid),
769 export, atomic_read(&export->exp_rpc_count));
770 GOTO(out, rc = -EBUSY);
771 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
773 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
774 "cookies not random?\n", target->obd_name,
775 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
776 GOTO(out, rc = -EALREADY);
778 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
779 if (req->rq_export == NULL && initial_conn)
780 export->exp_last_request_time =
781 max(export->exp_last_request_time,
782 (time_t)cfs_time_current_sec());
789 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
790 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
791 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
792 export, (long)cfs_time_current_sec(),
793 export ? (long)export->exp_last_request_time : 0);
795 /* Tell the client if we're in recovery. */
796 if (target->obd_recovering) {
797 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
798 /* If this is the first time a client connects,
799 reset the recovery timer */
801 target_start_and_reset_recovery_timer(target, req,
805 /* We want to handle EALREADY but *not* -EALREADY from
806 * target_handle_reconnect(), return reconnection state in a flag */
807 if (rc == EALREADY) {
808 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
814 /* Tell the client if we support replayable requests */
815 if (target->obd_replayable)
816 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
817 client_nid = &req->rq_peer.nid;
819 if (export == NULL) {
820 if (target->obd_recovering) {
823 t = cfs_timer_deadline(&target->obd_recovery_timer);
824 t = cfs_time_sub(t, cfs_time_current());
825 CERROR("%s: denying connection for new client %s (%s): "
826 "%d clients in recovery for "CFS_TIME_T"s\n",
828 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
829 target->obd_recoverable_clients,
830 cfs_duration_sec(t));
834 rc = obd_connect(req->rq_svc_thread->t_env,
835 &conn, target, &cluuid, data,
839 rc = obd_reconnect(req->rq_svc_thread->t_env,
840 export, target, &cluuid, data, client_nid);
844 /* Return only the parts of obd_connect_data that we understand, so the
845 * client knows that we don't understand the rest. */
847 tmpdata = req_capsule_server_get(&req->rq_pill,
849 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
853 /* If all else goes well, this is our RPC return code. */
856 lustre_msg_set_handle(req->rq_repmsg, &conn);
858 /* ownership of this export ref transfers to the request AFTER we
859 * drop any previous reference the request had, but we don't want
860 * that to go to zero before we get our new export reference. */
861 export = class_conn2export(&conn);
863 DEBUG_REQ(D_ERROR, req, "Missing export!");
864 GOTO(out, rc = -ENODEV);
867 /* If the client and the server are the same node, we will already
868 * have an export that really points to the client's DLM export,
869 * because we have a shared handles table.
871 * XXX this will go away when shaver stops sending the "connect" handle
872 * in the real "remote handle" field of the request --phik 24 Apr 2003
874 if (req->rq_export != NULL)
875 class_export_put(req->rq_export);
877 req->rq_export = export;
879 spin_lock(&export->exp_lock);
881 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
882 } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
883 spin_unlock(&export->exp_lock);
884 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
885 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
886 export->exp_conn_cnt,
887 lustre_msg_get_conn_cnt(req->rq_reqmsg));
889 GOTO(out, rc = -EALREADY);
891 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
893 /* request from liblustre? Don't evict it for not pinging. */
894 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
895 export->exp_libclient = 1;
896 spin_unlock(&export->exp_lock);
898 spin_lock(&target->obd_dev_lock);
899 list_del_init(&export->exp_obd_chain_timed);
900 spin_unlock(&target->obd_dev_lock);
902 spin_unlock(&export->exp_lock);
905 if (export->exp_connection != NULL)
906 ptlrpc_connection_put(export->exp_connection);
907 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
911 spin_lock(&target->obd_dev_lock);
913 /* Export might be hashed already, e.g. if this is reconnect */
914 if (hlist_unhashed(&export->exp_nid_hash))
915 lustre_hash_add(export->exp_obd->obd_nid_hash,
916 &export->exp_connection->c_peer.nid,
917 &export->exp_nid_hash);
919 spin_unlock(&target->obd_dev_lock);
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(&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);
1165 /* when recovery finished, cleanup orphans on mds and ost */
1166 if (OBT(obd) && OBP(obd, postrecov)) {
1167 int rc = OBP(obd, postrecov)(obd);
1168 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1169 rc < 0 ? "failed" : "complete", rc);
1172 obd->obd_recovery_end = cfs_time_current_sec();
1176 static void abort_req_replay_queue(struct obd_device *obd)
1178 struct ptlrpc_request *req, *n;
1180 list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1181 DEBUG_REQ(D_WARNING, req, "aborted:");
1182 req->rq_status = -ENOTCONN;
1183 if (ptlrpc_error(req)) {
1184 DEBUG_REQ(D_ERROR, req,
1185 "failed abort_req_reply; skipping");
1187 target_exp_dequeue_req_replay(req);
1188 ptlrpc_free_clone(req);
1192 static void abort_lock_replay_queue(struct obd_device *obd)
1194 struct ptlrpc_request *req, *n;
1196 list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1197 DEBUG_REQ(D_ERROR, req, "aborted:");
1198 req->rq_status = -ENOTCONN;
1199 if (ptlrpc_error(req)) {
1200 DEBUG_REQ(D_ERROR, req,
1201 "failed abort_lock_reply; skipping");
1203 ptlrpc_free_clone(req);
1208 /* Called from a cleanup function if the device is being cleaned up
1209 forcefully. The exports should all have been disconnected already,
1210 the only thing left to do is
1211 - clear the recovery flags
1213 - free queued requests and replies, but don't send replies
1214 Because the obd_stopping flag is set, no new requests should be received.
1217 void target_cleanup_recovery(struct obd_device *obd)
1219 struct ptlrpc_request *req, *n;
1222 LASSERT(obd->obd_stopping);
1224 spin_lock_bh(&obd->obd_processing_task_lock);
1225 if (!obd->obd_recovering) {
1226 spin_unlock_bh(&obd->obd_processing_task_lock);
1230 obd->obd_recovering = obd->obd_abort_recovery = 0;
1231 target_cancel_recovery_timer(obd);
1232 spin_unlock_bh(&obd->obd_processing_task_lock);
1234 list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1235 LASSERT (req->rq_reply_state == 0);
1236 target_exp_dequeue_req_replay(req);
1237 ptlrpc_free_clone(req);
1239 list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1240 LASSERT (req->rq_reply_state == 0);
1241 ptlrpc_free_clone(req);
1243 list_for_each_entry_safe(req, n, &obd->obd_final_req_queue, rq_list) {
1244 LASSERT (req->rq_reply_state == 0);
1245 ptlrpc_free_clone(req);
1251 /* obd_processing_task_lock should be held */
1252 void target_cancel_recovery_timer(struct obd_device *obd)
1254 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1255 cfs_timer_disarm(&obd->obd_recovery_timer);
1258 /* extend = 1 means require at least "duration" seconds left in the timer,
1259 extend = 0 means set the total duration (start_recovery_timer) */
1260 static void reset_recovery_timer(struct obd_device *obd, int duration,
1263 cfs_time_t now = cfs_time_current_sec();
1264 cfs_duration_t left;
1266 spin_lock_bh(&obd->obd_processing_task_lock);
1267 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1268 spin_unlock_bh(&obd->obd_processing_task_lock);
1272 left = cfs_time_sub(obd->obd_recovery_end, now);
1274 if (extend && (duration > left))
1275 obd->obd_recovery_timeout += duration - left;
1276 else if (!extend && (duration > obd->obd_recovery_timeout))
1277 /* Track the client's largest expected replay time */
1278 obd->obd_recovery_timeout = duration;
1281 * If total recovery time already exceed the
1282 * obd_recovery_max_time, then CRAY XT3 will
1283 * abort the recovery
1285 if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1286 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1288 obd->obd_recovery_end = obd->obd_recovery_start +
1289 obd->obd_recovery_timeout;
1290 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1291 cfs_time_before(now, obd->obd_recovery_end)) {
1292 left = cfs_time_sub(obd->obd_recovery_end, now);
1293 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1295 spin_unlock_bh(&obd->obd_processing_task_lock);
1296 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1297 obd->obd_name, (unsigned)left);
1300 static void check_and_start_recovery_timer(struct obd_device *obd)
1302 spin_lock_bh(&obd->obd_processing_task_lock);
1303 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1304 spin_unlock_bh(&obd->obd_processing_task_lock);
1307 CWARN("%s: starting recovery timer\n", obd->obd_name);
1308 obd->obd_recovery_start = cfs_time_current_sec();
1310 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1311 spin_unlock_bh(&obd->obd_processing_task_lock);
1313 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1316 /* Reset the timer with each new client connection */
1318 * This timer is actually reconnect_timer, which is for making sure
1319 * the total recovery window is at least as big as my reconnect
1320 * attempt timing. So the initial recovery time_out will be set to
1321 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1322 * from client is bigger than this, then the recovery time_out will
1323 * be extend to make sure the client could be reconnected, in the
1324 * process, the timeout from the new client should be ignored.
1328 target_start_and_reset_recovery_timer(struct obd_device *obd,
1329 struct ptlrpc_request *req,
1332 int req_timeout = lustre_msg_get_timeout(req->rq_reqmsg);
1334 /* teach server about old server's estimates */
1336 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1337 at_timeout2est(req_timeout));
1339 check_and_start_recovery_timer(obd);
1341 req_timeout *= OBD_RECOVERY_FACTOR;
1342 if (req_timeout > obd->obd_recovery_timeout && !new_client)
1343 reset_recovery_timer(obd, req_timeout, 0);
1347 static int check_for_next_transno(struct obd_device *obd)
1349 struct ptlrpc_request *req = NULL;
1350 int wake_up = 0, connected, completed, queue_len, max;
1351 __u64 next_transno, req_transno;
1353 spin_lock_bh(&obd->obd_processing_task_lock);
1355 if (!list_empty(&obd->obd_req_replay_queue)) {
1356 req = list_entry(obd->obd_req_replay_queue.next,
1357 struct ptlrpc_request, rq_list);
1358 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1363 max = obd->obd_max_recoverable_clients;
1364 connected = obd->obd_connected_clients;
1365 completed = connected - obd->obd_recoverable_clients;
1366 queue_len = obd->obd_requests_queued_for_recovery;
1367 next_transno = obd->obd_next_recovery_transno;
1369 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1370 "req_transno: "LPU64", next_transno: "LPU64"\n",
1371 max, connected, completed, queue_len, req_transno, next_transno);
1373 if (obd->obd_abort_recovery) {
1374 CDEBUG(D_HA, "waking for aborted recovery\n");
1376 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1377 CDEBUG(D_HA, "waking for completed recovery\n");
1379 } else if (req_transno == next_transno) {
1380 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1382 } else if (queue_len + completed == max) {
1383 /* handle gaps occured due to lost reply. It is allowed gaps
1384 * because all clients are connected and there will be resend
1385 * for missed transaction */
1386 LASSERTF(req_transno >= next_transno,
1387 "req_transno: "LPU64", next_transno: "LPU64"\n",
1388 req_transno, next_transno);
1390 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1391 "waking for skipped transno (skip: "LPD64
1392 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1393 next_transno, queue_len, completed, connected, req_transno);
1394 obd->obd_next_recovery_transno = req_transno;
1396 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1397 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1398 " by fail_lock, waking up ("LPD64")\n", next_transno);
1399 obd->obd_next_recovery_transno = req_transno;
1401 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1402 /* some clients haven't connected in time, but we can try
1403 * to replay requests that demand on already committed ones
1404 * also, we can replay first non-committed transation */
1405 LASSERT(req_transno != 0);
1406 if (req_transno == obd->obd_last_committed + 1) {
1407 obd->obd_next_recovery_transno = req_transno;
1408 } else if (req_transno > obd->obd_last_committed) {
1409 /* can't continue recovery: have no needed transno */
1410 obd->obd_abort_recovery = 1;
1411 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1412 "connected: %d, completed: %d, queue_len: %d, "
1413 "req_transno: "LPU64", next_transno: "LPU64"\n",
1414 max, connected, completed, queue_len,
1415 req_transno, next_transno);
1420 spin_unlock_bh(&obd->obd_processing_task_lock);
1424 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1426 struct l_wait_info lwi = { 0 };
1427 struct ptlrpc_request *req;
1429 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1430 obd->obd_next_recovery_transno);
1431 l_wait_event(obd->obd_next_transno_waitq,
1432 check_for_next_transno(obd), &lwi);
1434 spin_lock_bh(&obd->obd_processing_task_lock);
1435 if (obd->obd_abort_recovery) {
1437 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1438 req = list_entry(obd->obd_req_replay_queue.next,
1439 struct ptlrpc_request, rq_list);
1440 target_exp_dequeue_req_replay(req);
1441 list_del_init(&req->rq_list);
1442 obd->obd_requests_queued_for_recovery--;
1446 spin_unlock_bh(&obd->obd_processing_task_lock);
1450 static int check_for_next_lock(struct obd_device *obd)
1452 struct ptlrpc_request *req = NULL;
1455 spin_lock_bh(&obd->obd_processing_task_lock);
1456 if (!list_empty(&obd->obd_lock_replay_queue)) {
1457 req = list_entry(obd->obd_lock_replay_queue.next,
1458 struct ptlrpc_request, rq_list);
1459 CDEBUG(D_HA, "waking for next lock\n");
1461 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1462 CDEBUG(D_HA, "waking for completed lock replay\n");
1464 } else if (obd->obd_abort_recovery) {
1465 CDEBUG(D_HA, "waking for aborted recovery\n");
1468 spin_unlock_bh(&obd->obd_processing_task_lock);
1473 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1475 struct l_wait_info lwi = { 0 };
1476 struct ptlrpc_request *req;
1478 CDEBUG(D_HA, "Waiting for lock\n");
1479 l_wait_event(obd->obd_next_transno_waitq,
1480 check_for_next_lock(obd), &lwi);
1482 spin_lock_bh(&obd->obd_processing_task_lock);
1483 if (obd->obd_abort_recovery) {
1485 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1486 req = list_entry(obd->obd_lock_replay_queue.next,
1487 struct ptlrpc_request, rq_list);
1488 list_del_init(&req->rq_list);
1492 spin_unlock_bh(&obd->obd_processing_task_lock);
1496 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1498 struct ptlrpc_request *req;
1500 spin_lock_bh(&obd->obd_processing_task_lock);
1501 if (!list_empty(&obd->obd_final_req_queue)) {
1502 req = list_entry(obd->obd_final_req_queue.next,
1503 struct ptlrpc_request, rq_list);
1504 list_del_init(&req->rq_list);
1508 spin_unlock_bh(&obd->obd_processing_task_lock);
1512 static inline int req_replay_done(struct obd_export *exp)
1514 return (exp->exp_req_replay_needed == 0);
1517 static inline int lock_replay_done(struct obd_export *exp)
1519 return (exp->exp_lock_replay_needed == 0);
1522 static inline int connect_done(struct obd_export *exp)
1524 return (exp->exp_in_recovery != 0);
1527 static int check_for_clients(struct obd_device *obd)
1529 if (obd->obd_abort_recovery)
1531 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1532 if (obd->obd_no_conn == 0 &&
1533 obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1538 static int handle_recovery_req(struct ptlrpc_thread *thread,
1539 struct ptlrpc_request *req,
1540 svc_handler_t handler)
1545 rc = lu_context_init(&req->rq_session, LCT_SESSION);
1547 CERROR("Failure to initialize session: %d\n", rc);
1550 req->rq_session.lc_thread = thread;
1551 lu_context_enter(&req->rq_session);
1552 req->rq_svc_thread = thread;
1553 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1555 /* thread context */
1556 lu_context_enter(&thread->t_env->le_ctx);
1558 lu_context_exit(&thread->t_env->le_ctx);
1560 lu_context_exit(&req->rq_session);
1561 lu_context_fini(&req->rq_session);
1562 /* don't reset timer for final stage */
1563 if (!req_replay_done(req->rq_export) ||
1564 !lock_replay_done(req->rq_export))
1565 reset_recovery_timer(class_exp2obd(req->rq_export),
1566 OBD_RECOVERY_FACTOR * AT_OFF ? obd_timeout :
1567 at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1568 ptlrpc_free_clone(req);
1572 static void resume_recovery_timer(struct obd_device *obd)
1574 /* to be safe, make it at least OBD_RECOVERY_FACTOR * obd_timeout */
1575 reset_recovery_timer(obd, OBD_RECOVERY_FACTOR * obd_timeout, 1);
1578 static int target_recovery_thread(void *arg)
1580 struct obd_device *obd = arg;
1581 struct ptlrpc_request *req;
1582 struct target_recovery_data *trd = &obd->obd_recovery_data;
1583 struct l_wait_info lwi = { 0 };
1584 unsigned long delta;
1585 unsigned long flags;
1587 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1588 __u32 recov_ctx_tags = LCT_MD_THREAD;
1592 cfs_daemonize("tgt_recov");
1594 SIGNAL_MASK_LOCK(current, flags);
1595 sigfillset(¤t->blocked);
1597 SIGNAL_MASK_UNLOCK(current, flags);
1599 rc = lu_context_init(&env.le_ctx, recov_ctx_tags);
1603 thread->t_env = &env;
1604 env.le_ctx.lc_thread = thread;
1606 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1608 trd->trd_processing_task = current->pid;
1610 obd->obd_recovering = 1;
1611 complete(&trd->trd_starting);
1613 /* first of all, we have to know the first transno to replay */
1614 obd->obd_abort_recovery = 0;
1615 l_wait_event(obd->obd_next_transno_waitq,
1616 check_for_clients(obd), &lwi);
1618 spin_lock_bh(&obd->obd_processing_task_lock);
1619 target_cancel_recovery_timer(obd);
1620 spin_unlock_bh(&obd->obd_processing_task_lock);
1622 /* If some clients haven't connected in time, evict them */
1623 if (obd->obd_abort_recovery) {
1624 CWARN("Some clients haven't connect in time (%d/%d),"
1625 "evict them\n", obd->obd_connected_clients,
1626 obd->obd_max_recoverable_clients);
1627 obd->obd_abort_recovery = obd->obd_stopping;
1628 class_disconnect_stale_exports(obd, connect_done);
1630 /* next stage: replay requests */
1632 obd->obd_req_replaying = 1;
1633 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1634 atomic_read(&obd->obd_req_replay_clients),
1635 obd->obd_next_recovery_transno);
1636 resume_recovery_timer(obd);
1637 while ((req = target_next_replay_req(obd))) {
1638 LASSERT(trd->trd_processing_task == current->pid);
1639 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1640 lustre_msg_get_transno(req->rq_reqmsg),
1641 libcfs_nid2str(req->rq_peer.nid));
1642 handle_recovery_req(thread, req,
1643 trd->trd_recovery_handler);
1644 obd->obd_replayed_requests++;
1645 spin_lock_bh(&obd->obd_processing_task_lock);
1646 obd->obd_next_recovery_transno++;
1647 spin_unlock_bh(&obd->obd_processing_task_lock);
1650 spin_lock_bh(&obd->obd_processing_task_lock);
1651 target_cancel_recovery_timer(obd);
1652 spin_unlock_bh(&obd->obd_processing_task_lock);
1654 /* If some clients haven't replayed requests in time, evict them */
1655 if (obd->obd_abort_recovery) {
1656 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1657 obd->obd_abort_recovery = obd->obd_stopping;
1658 class_disconnect_stale_exports(obd, req_replay_done);
1659 abort_req_replay_queue(obd);
1662 /* The second stage: replay locks */
1663 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1664 atomic_read(&obd->obd_lock_replay_clients));
1665 resume_recovery_timer(obd);
1666 while ((req = target_next_replay_lock(obd))) {
1667 LASSERT(trd->trd_processing_task == current->pid);
1668 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1669 libcfs_nid2str(req->rq_peer.nid));
1670 handle_recovery_req(thread, req,
1671 trd->trd_recovery_handler);
1672 obd->obd_replayed_locks++;
1675 spin_lock_bh(&obd->obd_processing_task_lock);
1676 target_cancel_recovery_timer(obd);
1677 spin_unlock_bh(&obd->obd_processing_task_lock);
1678 /* If some clients haven't replayed requests in time, evict them */
1679 if (obd->obd_abort_recovery) {
1681 CERROR("lock replay timed out, aborting ...\n");
1682 obd->obd_abort_recovery = obd->obd_stopping;
1683 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1684 abort_lock_replay_queue(obd);
1687 /* We drop recoverying flag to forward all new requests
1688 * to regular mds_handle() since now */
1689 spin_lock_bh(&obd->obd_processing_task_lock);
1690 obd->obd_recovering = obd->obd_abort_recovery = 0;
1691 spin_unlock_bh(&obd->obd_processing_task_lock);
1692 /* The third stage: reply on final pings */
1693 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1694 while ((req = target_next_final_ping(obd))) {
1695 LASSERT(trd->trd_processing_task == current->pid);
1696 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1697 libcfs_nid2str(req->rq_peer.nid));
1698 handle_recovery_req(thread, req,
1699 trd->trd_recovery_handler);
1702 delta = (jiffies - delta) / HZ;
1703 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1704 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1705 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1706 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1707 if (delta > obd_timeout * 2) {
1708 CWARN("too long recovery - read logs\n");
1709 libcfs_debug_dumplog();
1712 target_finish_recovery(obd);
1714 lu_context_fini(&env.le_ctx);
1715 trd->trd_processing_task = 0;
1716 complete(&trd->trd_finishing);
1720 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1723 struct target_recovery_data *trd = &obd->obd_recovery_data;
1725 memset(trd, 0, sizeof(*trd));
1726 init_completion(&trd->trd_starting);
1727 init_completion(&trd->trd_finishing);
1728 trd->trd_recovery_handler = handler;
1730 if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1731 wait_for_completion(&trd->trd_starting);
1732 LASSERT(obd->obd_recovering != 0);
1739 void target_stop_recovery_thread(struct obd_device *obd)
1741 spin_lock_bh(&obd->obd_processing_task_lock);
1742 if (obd->obd_recovery_data.trd_processing_task > 0) {
1743 struct target_recovery_data *trd = &obd->obd_recovery_data;
1744 CERROR("%s: Aborting recovery\n", obd->obd_name);
1745 obd->obd_abort_recovery = 1;
1746 wake_up(&obd->obd_next_transno_waitq);
1747 spin_unlock_bh(&obd->obd_processing_task_lock);
1748 wait_for_completion(&trd->trd_finishing);
1750 spin_unlock_bh(&obd->obd_processing_task_lock);
1754 void target_recovery_fini(struct obd_device *obd)
1756 class_disconnect_exports(obd);
1757 target_stop_recovery_thread(obd);
1758 target_cleanup_recovery(obd);
1760 EXPORT_SYMBOL(target_recovery_fini);
1762 static void target_recovery_expired(unsigned long castmeharder)
1764 struct obd_device *obd = (struct obd_device *)castmeharder;
1765 LCONSOLE_WARN("%s: recovery timed out; %d clients never reconnected "
1766 "after %lds (%d clients did)\n",
1767 obd->obd_name, obd->obd_recoverable_clients,
1768 cfs_time_current_sec()- obd->obd_recovery_start,
1769 obd->obd_connected_clients);
1770 spin_lock_bh(&obd->obd_processing_task_lock);
1771 if (obd->obd_recovering)
1772 obd->obd_abort_recovery = 1;
1773 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1774 spin_unlock_bh(&obd->obd_processing_task_lock);
1777 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1779 if (obd->obd_max_recoverable_clients == 0)
1782 CWARN("RECOVERY: service %s, %d recoverable clients, "
1783 "last_transno "LPU64"\n", obd->obd_name,
1784 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1785 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1786 obd->obd_recovery_start = 0;
1787 obd->obd_recovery_end = 0;
1788 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1789 /* bz13079: this should be set to desired value for ost but not for mds */
1790 obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1791 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1792 target_start_recovery_thread(obd, handler);
1794 EXPORT_SYMBOL(target_recovery_init);
1798 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1800 struct obd_export *exp = req->rq_export;
1801 LASSERT(exp != NULL);
1802 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1803 /* client declares he's ready to replay locks */
1804 spin_lock_bh(&obd->obd_processing_task_lock);
1805 if (exp->exp_req_replay_needed) {
1806 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1807 spin_lock(&exp->exp_lock);
1808 exp->exp_req_replay_needed = 0;
1809 spin_unlock(&exp->exp_lock);
1810 atomic_dec(&obd->obd_req_replay_clients);
1811 LASSERT(obd->obd_recoverable_clients > 0);
1812 obd->obd_recoverable_clients--;
1813 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1814 CDEBUG(D_HA, "all clients have replayed reqs\n");
1815 wake_up(&obd->obd_next_transno_waitq);
1817 spin_unlock_bh(&obd->obd_processing_task_lock);
1819 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1820 /* client declares he's ready to complete recovery
1821 * so, we put the request on th final queue */
1822 spin_lock_bh(&obd->obd_processing_task_lock);
1823 if (exp->exp_lock_replay_needed) {
1824 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1825 spin_lock(&exp->exp_lock);
1826 exp->exp_lock_replay_needed = 0;
1827 spin_unlock(&exp->exp_lock);
1828 atomic_dec(&obd->obd_lock_replay_clients);
1829 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1830 CDEBUG(D_HA, "all clients have replayed locks\n");
1831 wake_up(&obd->obd_next_transno_waitq);
1833 spin_unlock_bh(&obd->obd_processing_task_lock);
1839 int target_queue_recovery_request(struct ptlrpc_request *req,
1840 struct obd_device *obd)
1842 struct list_head *tmp;
1844 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1848 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1849 /* Processing the queue right now, don't re-add. */
1853 target_process_req_flags(obd, req);
1855 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1856 /* client declares he's ready to complete recovery
1857 * so, we put the request on th final queue */
1858 req = ptlrpc_clone_req(req);
1861 DEBUG_REQ(D_HA, req, "queue final req");
1862 spin_lock_bh(&obd->obd_processing_task_lock);
1863 if (obd->obd_recovering)
1864 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1866 spin_unlock_bh(&obd->obd_processing_task_lock);
1867 ptlrpc_free_clone(req);
1868 if (obd->obd_stopping) {
1874 spin_unlock_bh(&obd->obd_processing_task_lock);
1877 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1878 /* client declares he's ready to replay locks */
1879 req = ptlrpc_clone_req(req);
1882 DEBUG_REQ(D_HA, req, "queue lock replay req");
1883 spin_lock_bh(&obd->obd_processing_task_lock);
1884 LASSERT(obd->obd_recovering);
1885 /* usually due to recovery abort */
1886 if (!req->rq_export->exp_in_recovery) {
1887 spin_unlock_bh(&obd->obd_processing_task_lock);
1888 ptlrpc_free_clone(req);
1891 LASSERT(req->rq_export->exp_lock_replay_needed);
1892 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1893 spin_unlock_bh(&obd->obd_processing_task_lock);
1894 wake_up(&obd->obd_next_transno_waitq);
1898 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1899 * (i.e. buflens etc are in my own byte order), but type-dependent
1900 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1903 CFS_INIT_LIST_HEAD(&req->rq_list);
1904 DEBUG_REQ(D_HA, req, "not queueing");
1908 spin_lock_bh(&obd->obd_processing_task_lock);
1910 /* If we're processing the queue, we want don't want to queue this
1913 * Also, if this request has a transno less than the one we're waiting
1914 * for, we should process it now. It could (and currently always will)
1915 * be an open request for a descriptor that was opened some time ago.
1917 * Also, a resent, replayed request that has already been
1918 * handled will pass through here and be processed immediately.
1920 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1921 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1922 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1923 /* Processing the queue right now, don't re-add. */
1924 LASSERT(list_empty(&req->rq_list));
1925 spin_unlock_bh(&obd->obd_processing_task_lock);
1928 spin_unlock_bh(&obd->obd_processing_task_lock);
1930 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
1933 req = ptlrpc_clone_req(req);
1937 spin_lock_bh(&obd->obd_processing_task_lock);
1938 LASSERT(obd->obd_recovering);
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_req_replay_needed);
1946 if (target_exp_enqueue_req_replay(req)) {
1947 spin_unlock_bh(&obd->obd_processing_task_lock);
1948 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1949 ptlrpc_free_clone(req);
1954 list_for_each(tmp, &obd->obd_req_replay_queue) {
1955 struct ptlrpc_request *reqiter =
1956 list_entry(tmp, struct ptlrpc_request, rq_list);
1958 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1959 list_add_tail(&req->rq_list, &reqiter->rq_list);
1964 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
1966 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
1967 "has been claimed by another client");
1968 spin_unlock_bh(&obd->obd_processing_task_lock);
1969 target_exp_dequeue_req_replay(req);
1970 ptlrpc_free_clone(req);
1976 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1978 obd->obd_requests_queued_for_recovery++;
1979 wake_up(&obd->obd_next_transno_waitq);
1980 spin_unlock_bh(&obd->obd_processing_task_lock);
1984 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1986 return req->rq_export->exp_obd;
1989 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
1991 LASSERT(exp != NULL);
1992 return &exp->exp_obd->obd_namespace->ns_pool;
1996 * Packs current SLV and Limit into \a req.
1998 int target_pack_pool_reply(struct ptlrpc_request *req)
2000 struct obd_device *obd;
2004 * Check that we still have all structures alive as this may
2005 * be some late rpc in shutdown time.
2007 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2008 !exp_connect_lru_resize(req->rq_export))) {
2009 lustre_msg_set_slv(req->rq_repmsg, 0);
2010 lustre_msg_set_limit(req->rq_repmsg, 0);
2015 * OBD is alive here as export is alive, which we checked above.
2017 obd = req->rq_export->exp_obd;
2019 read_lock(&obd->obd_pool_lock);
2020 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2021 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2022 read_unlock(&obd->obd_pool_lock);
2027 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2029 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2030 DEBUG_REQ(D_ERROR, req, "dropping reply");
2035 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2036 req->rq_status = rc;
2037 return (ptlrpc_send_error(req, 1));
2039 DEBUG_REQ(D_NET, req, "sending reply");
2042 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2045 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2048 struct ptlrpc_reply_state *rs;
2049 struct obd_device *obd;
2050 struct obd_export *exp;
2051 struct ptlrpc_service *svc;
2054 if (req->rq_no_reply) {
2059 svc = req->rq_rqbd->rqbd_service;
2060 rs = req->rq_reply_state;
2061 if (rs == NULL || !rs->rs_difficult) {
2063 target_send_reply_msg (req, rc, fail_id);
2068 /* must be an export if locks saved */
2069 LASSERT (req->rq_export != NULL);
2070 /* req/reply consistent */
2071 LASSERT (rs->rs_service == svc);
2074 LASSERT (!rs->rs_scheduled);
2075 LASSERT (!rs->rs_scheduled_ever);
2076 LASSERT (!rs->rs_handled);
2077 LASSERT (!rs->rs_on_net);
2078 LASSERT (rs->rs_export == NULL);
2079 LASSERT (list_empty(&rs->rs_obd_list));
2080 LASSERT (list_empty(&rs->rs_exp_list));
2082 exp = class_export_get (req->rq_export);
2085 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2086 rs->rs_scheduled = 1;
2088 rs->rs_xid = req->rq_xid;
2089 rs->rs_transno = req->rq_transno;
2090 rs->rs_export = exp;
2092 spin_lock(&obd->obd_uncommitted_replies_lock);
2094 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2095 rs->rs_transno, obd->obd_last_committed);
2096 if (rs->rs_transno > obd->obd_last_committed) {
2097 /* not committed already */
2098 list_add_tail (&rs->rs_obd_list,
2099 &obd->obd_uncommitted_replies);
2102 spin_unlock (&obd->obd_uncommitted_replies_lock);
2103 spin_lock (&exp->exp_lock);
2105 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
2107 spin_unlock(&exp->exp_lock);
2109 netrc = target_send_reply_msg (req, rc, fail_id);
2111 spin_lock(&svc->srv_lock);
2113 svc->srv_n_difficult_replies++;
2116 /* error sending: reply is off the net. Also we need +1
2117 * reply ref until ptlrpc_server_handle_reply() is done
2118 * with the reply state (if the send was successful, there
2119 * would have been +1 ref for the net, which
2120 * reply_out_callback leaves alone) */
2122 ptlrpc_rs_addref(rs);
2123 atomic_inc (&svc->srv_outstanding_replies);
2126 if (rs->rs_transno <= obd->obd_last_committed ||
2127 (!rs->rs_on_net && !rs->rs_no_ack) ||
2128 list_empty(&rs->rs_exp_list) || /* completed already */
2129 list_empty(&rs->rs_obd_list)) {
2130 CDEBUG(D_HA, "Schedule reply immediately\n");
2131 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
2132 cfs_waitq_signal (&svc->srv_waitq);
2134 list_add (&rs->rs_list, &svc->srv_active_replies);
2135 rs->rs_scheduled = 0; /* allow notifier to schedule */
2138 spin_unlock(&svc->srv_lock);
2142 int target_handle_ping(struct ptlrpc_request *req)
2144 obd_ping(req->rq_export);
2145 return req_capsule_server_pack(&req->rq_pill);
2148 void target_committed_to_req(struct ptlrpc_request *req)
2150 struct obd_device *obd;
2152 if (req == NULL || req->rq_export == NULL)
2155 obd = req->rq_export->exp_obd;
2159 if (!obd->obd_no_transno && req->rq_repmsg != NULL)
2160 lustre_msg_set_last_committed(req->rq_repmsg,
2161 obd->obd_last_committed);
2163 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2164 "%d)", obd->obd_no_transno, req->rq_repmsg == NULL);
2166 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2167 obd->obd_last_committed, req->rq_transno, req->rq_xid);
2170 EXPORT_SYMBOL(target_committed_to_req);
2172 #ifdef HAVE_QUOTA_SUPPORT
2173 int target_handle_qc_callback(struct ptlrpc_request *req)
2175 struct obd_quotactl *oqctl;
2176 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2178 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2182 cli->cl_qchk_stat = oqctl->qc_stat;
2187 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2190 struct obd_device *obd = req->rq_export->exp_obd;
2191 struct obd_device *master_obd;
2192 struct lustre_quota_ctxt *qctxt;
2193 struct qunit_data *qdata;
2195 struct qunit_data_old *qdata_old;
2199 rc = req_capsule_server_pack(&req->rq_pill);
2201 CERROR("packing reply failed!: rc = %d\n", rc);
2205 LASSERT(req->rq_export);
2207 /* fixed for bug10707 */
2208 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2209 !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2210 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
2211 rep = req_capsule_server_get(&req->rq_pill,
2214 qdata = req_capsule_client_swab_get(&req->rq_pill,
2216 (void*)lustre_swab_qdata);
2218 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
2219 rep = req_capsule_server_get(&req->rq_pill, &RMF_QUNIT_DATA);
2221 qdata_old = req_capsule_client_swab_get(&req->rq_pill,
2223 (void*)lustre_swab_qdata_old);
2224 qdata = lustre_quota_old_to_new(qdata_old);
2230 /* we use the observer */
2231 LASSERT(obd->obd_observer && obd->obd_observer->obd_observer);
2232 master_obd = obd->obd_observer->obd_observer;
2233 qctxt = &master_obd->u.obt.obt_qctxt;
2235 LASSERT(qctxt->lqc_handler);
2236 rc = qctxt->lqc_handler(master_obd, qdata,
2237 lustre_msg_get_opc(req->rq_reqmsg));
2238 if (rc && rc != -EDQUOT)
2239 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2240 "dqacq failed! (rc:%d)\n", rc);
2242 /* the qd_count might be changed in lqc_handler */
2243 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2244 !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2245 memcpy(rep, qdata, sizeof(*qdata));
2247 qdata_old = lustre_quota_new_to_old(qdata);
2248 memcpy(rep, qdata_old, sizeof(*qdata_old));
2250 req->rq_status = rc;
2251 rc = ptlrpc_reply(req);
2256 #endif /* !__KERNEL__ */
2258 #endif /* HAVE_QUOTA_SUPPORT */
2260 ldlm_mode_t lck_compat_array[] = {
2261 [LCK_EX] LCK_COMPAT_EX,
2262 [LCK_PW] LCK_COMPAT_PW,
2263 [LCK_PR] LCK_COMPAT_PR,
2264 [LCK_CW] LCK_COMPAT_CW,
2265 [LCK_CR] LCK_COMPAT_CR,
2266 [LCK_NL] LCK_COMPAT_NL,
2267 [LCK_GROUP] LCK_COMPAT_GROUP,
2268 [LCK_COS] LCK_COMPAT_COS,
2272 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2273 * not escape to the user level.
2275 int ldlm_error2errno(ldlm_error_t error)
2283 case ELDLM_LOCK_CHANGED:
2286 case ELDLM_LOCK_ABORTED:
2289 case ELDLM_LOCK_REPLACED:
2292 case ELDLM_NO_LOCK_DATA:
2295 case ELDLM_NAMESPACE_EXISTS:
2298 case ELDLM_BAD_NAMESPACE:
2302 if (((int)error) < 0) /* cast to signed type */
2303 result = error; /* as ldlm_error_t can be unsigned */
2305 CERROR("Invalid DLM result code: %i\n", error);
2311 EXPORT_SYMBOL(ldlm_error2errno);
2314 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2316 ldlm_error_t ldlm_errno2error(int err_no)
2325 error = ELDLM_LOCK_CHANGED;
2328 error = ELDLM_LOCK_ABORTED;
2331 error = ELDLM_LOCK_REPLACED;
2334 error = ELDLM_NO_LOCK_DATA;
2337 error = ELDLM_NAMESPACE_EXISTS;
2340 error = ELDLM_BAD_NAMESPACE;
2347 EXPORT_SYMBOL(ldlm_errno2error);