1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2003 Cluster File Systems, Inc.
6 * This file is part of the Lustre file system, http://www.lustre.org
7 * Lustre is a trademark of Cluster File Systems, Inc.
9 * You may have signed or agreed to another license before downloading
10 * this software. If so, you are bound by the terms and conditions
11 * of that agreement, and the following does not apply to you. See the
12 * LICENSE file included with this distribution for more information.
14 * If you did not agree to a different license, then this copy of Lustre
15 * is open source software; you can redistribute it and/or modify it
16 * under the terms of version 2 of the GNU General Public License as
17 * published by the Free Software Foundation.
19 * In either case, Lustre is distributed in the hope that it will be
20 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
21 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * license text for more details.
26 # define EXPORT_SYMTAB
28 #define DEBUG_SUBSYSTEM S_LDLM
31 # include <libcfs/libcfs.h>
33 # include <liblustre.h>
36 #include <lustre_mds.h>
37 #include <lustre_dlm.h>
38 #include <lustre_net.h>
39 #include <lustre_sec.h>
41 /* @priority: if non-zero, move the selected to the list head
42 * @create: if zero, only search in existed connections
44 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
45 int priority, int create)
47 struct ptlrpc_connection *ptlrpc_conn;
48 struct obd_import_conn *imp_conn = NULL, *item;
52 if (!create && !priority) {
53 CDEBUG(D_HA, "Nothing to do\n");
57 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
59 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
64 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
66 GOTO(out_put, rc = -ENOMEM);
70 spin_lock(&imp->imp_lock);
71 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
72 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
74 list_del(&item->oic_item);
75 list_add(&item->oic_item, &imp->imp_conn_list);
76 item->oic_last_attempt = 0;
78 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
79 imp, imp->imp_obd->obd_name, uuid->uuid,
80 (priority ? ", moved to head" : ""));
81 spin_unlock(&imp->imp_lock);
82 GOTO(out_free, rc = 0);
87 imp_conn->oic_conn = ptlrpc_conn;
88 imp_conn->oic_uuid = *uuid;
89 item->oic_last_attempt = 0;
91 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
93 list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
94 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
95 imp, imp->imp_obd->obd_name, uuid->uuid,
96 (priority ? "head" : "tail"));
98 spin_unlock(&imp->imp_lock);
99 GOTO(out_free, rc = -ENOENT);
102 spin_unlock(&imp->imp_lock);
106 OBD_FREE(imp_conn, sizeof(*imp_conn));
108 ptlrpc_put_connection(ptlrpc_conn);
112 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
114 return import_set_conn(imp, uuid, 1, 0);
117 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
120 return import_set_conn(imp, uuid, priority, 1);
123 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
125 struct obd_import_conn *imp_conn;
126 struct obd_export *dlmexp;
130 spin_lock(&imp->imp_lock);
131 if (list_empty(&imp->imp_conn_list)) {
132 LASSERT(!imp->imp_connection);
136 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
137 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
139 LASSERT(imp_conn->oic_conn);
141 /* is current conn? */
142 if (imp_conn == imp->imp_conn_current) {
143 LASSERT(imp_conn->oic_conn == imp->imp_connection);
145 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
146 imp->imp_state != LUSTRE_IMP_DISCON) {
147 CERROR("can't remove current connection\n");
148 GOTO(out, rc = -EBUSY);
151 ptlrpc_put_connection(imp->imp_connection);
152 imp->imp_connection = NULL;
154 dlmexp = class_conn2export(&imp->imp_dlm_handle);
155 if (dlmexp && dlmexp->exp_connection) {
156 LASSERT(dlmexp->exp_connection ==
158 ptlrpc_put_connection(dlmexp->exp_connection);
159 dlmexp->exp_connection = NULL;
163 list_del(&imp_conn->oic_item);
164 ptlrpc_put_connection(imp_conn->oic_conn);
165 OBD_FREE(imp_conn, sizeof(*imp_conn));
166 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
167 imp, imp->imp_obd->obd_name, uuid->uuid);
172 spin_unlock(&imp->imp_lock);
174 CERROR("connection %s not found\n", uuid->uuid);
178 static void destroy_import(struct obd_import *imp)
180 /* drop security policy instance after all rpc finished/aborted
181 * to let all busy credentials be released.
183 class_import_get(imp);
184 class_destroy_import(imp);
185 sptlrpc_import_put_sec(imp);
186 class_import_put(imp);
189 /* configure an RPC client OBD device
194 * 3 - inactive-on-startup
196 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
198 struct client_obd *cli = &obddev->u.cli;
199 struct obd_import *imp;
200 struct obd_uuid server_uuid;
201 int rq_portal, rp_portal, connect_op;
202 char *name = obddev->obd_type->typ_name;
206 /* In a more perfect world, we would hang a ptlrpc_client off of
207 * obd_type and just use the values from there. */
208 if (!strcmp(name, LUSTRE_OSC_NAME)) {
209 rq_portal = OST_REQUEST_PORTAL;
210 rp_portal = OSC_REPLY_PORTAL;
211 connect_op = OST_CONNECT;
212 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
213 rq_portal = MDS_REQUEST_PORTAL;
214 rp_portal = MDC_REPLY_PORTAL;
215 connect_op = MDS_CONNECT;
216 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
217 rq_portal = MGS_REQUEST_PORTAL;
218 rp_portal = MGC_REPLY_PORTAL;
219 connect_op = MGS_CONNECT;
221 CERROR("unknown client OBD type \"%s\", can't setup\n",
226 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
227 CERROR("requires a TARGET UUID\n");
231 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
232 CERROR("client UUID must be less than 38 characters\n");
236 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
237 CERROR("setup requires a SERVER UUID\n");
241 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
242 CERROR("target UUID must be less than 38 characters\n");
246 sema_init(&cli->cl_sem, 1);
247 sema_init(&cli->cl_mgc_sem, 1);
248 cli->cl_sec_conf.sfc_rpc_flavor = SPTLRPC_FLVR_NULL;
249 cli->cl_sec_conf.sfc_bulk_csum = BULK_CSUM_ALG_NULL;
250 cli->cl_sec_conf.sfc_bulk_priv = BULK_PRIV_ALG_NULL;
251 cli->cl_sec_conf.sfc_flags = 0;
252 cli->cl_conn_count = 0;
253 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
254 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
255 sizeof(server_uuid)));
258 cli->cl_avail_grant = 0;
259 /* FIXME: should limit this for the sum of all cl_dirty_max */
260 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
261 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
262 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
263 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
264 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
265 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
266 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
267 client_obd_list_lock_init(&cli->cl_loi_list_lock);
268 cli->cl_r_in_flight = 0;
269 cli->cl_w_in_flight = 0;
271 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
272 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
273 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
274 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
275 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
276 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
278 /* This value may be changed at connect time in
279 ptlrpc_connect_interpret. */
280 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
281 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
283 if (!strcmp(name, LUSTRE_MDC_NAME)) {
284 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
285 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
286 cli->cl_max_rpcs_in_flight = 2;
287 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
288 cli->cl_max_rpcs_in_flight = 3;
289 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
290 cli->cl_max_rpcs_in_flight = 4;
292 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
297 CERROR("ldlm_get_ref failed: %d\n", rc);
301 ptlrpc_init_client(rq_portal, rp_portal, name,
302 &obddev->obd_ldlm_client);
304 imp = class_new_import(obddev);
306 GOTO(err_ldlm, rc = -ENOENT);
307 imp->imp_client = &obddev->obd_ldlm_client;
308 imp->imp_connect_op = connect_op;
309 imp->imp_initial_recov = 1;
310 imp->imp_initial_recov_bk = 0;
311 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
312 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
313 LUSTRE_CFG_BUFLEN(lcfg, 1));
314 class_import_put(imp);
316 rc = client_import_add_conn(imp, &server_uuid, 1);
318 CERROR("can't add initial connection\n");
319 GOTO(err_import, rc);
322 cli->cl_import = imp;
323 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
324 cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
325 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
327 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
328 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
329 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
330 name, obddev->obd_name,
331 cli->cl_target_uuid.uuid);
332 spin_lock(&imp->imp_lock);
333 imp->imp_invalid = 1;
334 spin_unlock(&imp->imp_lock);
338 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
343 class_destroy_import(imp);
351 int client_obd_cleanup(struct obd_device *obddev)
354 ldlm_put_ref(obddev->obd_force);
359 /* ->o_connect() method for client side (OSC and MDC and MGC) */
360 int client_connect_import(const struct lu_env *env,
361 struct lustre_handle *dlm_handle,
362 struct obd_device *obd, struct obd_uuid *cluuid,
363 struct obd_connect_data *data)
365 struct client_obd *cli = &obd->u.cli;
366 struct obd_import *imp = cli->cl_import;
367 struct obd_export *exp;
368 struct obd_connect_data *ocd;
372 mutex_down(&cli->cl_sem);
373 rc = class_connect(dlm_handle, obd, cluuid);
377 cli->cl_conn_count++;
378 if (cli->cl_conn_count > 1)
380 exp = class_conn2export(dlm_handle);
382 if (obd->obd_namespace != NULL)
383 CERROR("already have namespace!\n");
384 obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
385 LDLM_NAMESPACE_CLIENT);
386 if (obd->obd_namespace == NULL)
387 GOTO(out_disco, rc = -ENOMEM);
389 imp->imp_dlm_handle = *dlm_handle;
390 rc = ptlrpc_init_import(imp);
394 rc = sptlrpc_import_get_sec(imp, NULL, cli->cl_sec_conf.sfc_rpc_flavor,
395 cli->cl_sec_conf.sfc_flags);
399 ocd = &imp->imp_connect_data;
402 imp->imp_connect_flags_orig = data->ocd_connect_flags;
405 rc = ptlrpc_connect_import(imp, NULL);
407 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
410 LASSERT(exp->exp_connection);
413 LASSERT((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
414 ocd->ocd_connect_flags);
415 data->ocd_connect_flags = ocd->ocd_connect_flags;
418 ptlrpc_pinger_add_import(imp);
424 ldlm_namespace_free(obd->obd_namespace, 0);
425 obd->obd_namespace = NULL;
427 cli->cl_conn_count--;
428 class_disconnect(exp);
430 class_export_put(exp);
433 mutex_up(&cli->cl_sem);
437 int client_disconnect_export(struct obd_export *exp)
439 struct obd_device *obd = class_exp2obd(exp);
440 struct client_obd *cli;
441 struct obd_import *imp;
446 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
447 exp, exp ? exp->exp_handle.h_cookie : -1);
452 imp = cli->cl_import;
454 mutex_down(&cli->cl_sem);
455 if (!cli->cl_conn_count) {
456 CERROR("disconnecting disconnected device (%s)\n",
458 GOTO(out_sem, rc = -EINVAL);
461 cli->cl_conn_count--;
462 if (cli->cl_conn_count)
463 GOTO(out_no_disconnect, rc = 0);
465 /* Mark import deactivated now, so we don't try to reconnect if any
466 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
467 * fully deactivate the import, or that would drop all requests. */
468 spin_lock(&imp->imp_lock);
469 imp->imp_deactive = 1;
470 spin_unlock(&imp->imp_lock);
472 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
473 * delete it regardless. (It's safe to delete an import that was
475 (void)ptlrpc_pinger_del_import(imp);
477 if (obd->obd_namespace != NULL) {
478 /* obd_force == local only */
479 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
480 obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
482 ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
483 obd->obd_namespace = NULL;
487 rc = ptlrpc_disconnect_import(imp, 0);
489 ptlrpc_invalidate_import(imp);
490 ptlrpc_free_rq_pool(imp->imp_rq_pool);
492 cli->cl_import = NULL;
496 err = class_disconnect(exp);
500 mutex_up(&cli->cl_sem);
504 /* --------------------------------------------------------------------------
505 * from old lib/target.c
506 * -------------------------------------------------------------------------- */
508 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
509 struct obd_uuid *cluuid, int initial_conn)
512 if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
513 struct lustre_handle *hdl;
514 hdl = &exp->exp_imp_reverse->imp_remote_handle;
515 /* Might be a re-connect after a partition. */
516 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
517 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
519 conn->cookie = exp->exp_handle.h_cookie;
520 /* target_handle_connect() treats EALREADY and
521 * -EALREADY differently. EALREADY means we are
522 * doing a valid reconnect from the same client. */
525 CERROR("%s reconnecting from %s, "
526 "handle mismatch (ours "LPX64", theirs "
527 LPX64")\n", cluuid->uuid,
528 exp->exp_connection->c_remote_uuid.uuid,
529 hdl->cookie, conn->cookie);
530 memset(conn, 0, sizeof *conn);
531 /* target_handle_connect() treats EALREADY and
532 * -EALREADY differently. -EALREADY is an error
533 * (same UUID, different handle). */
538 conn->cookie = exp->exp_handle.h_cookie;
539 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
540 cluuid->uuid, exp, conn->cookie);
544 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
547 struct obd_export *exp = cb_data;
549 CDEBUG(D_HA, "%s: committing for initial connect of %s\n",
550 obd->obd_name, exp->exp_client_uuid.uuid);
552 spin_lock(&exp->exp_lock);
553 exp->exp_need_sync = 0;
554 spin_unlock(&exp->exp_lock);
556 EXPORT_SYMBOL(target_client_add_cb);
558 int target_handle_connect(struct ptlrpc_request *req)
560 struct obd_device *target, *targref = NULL;
561 struct obd_export *export = NULL;
562 struct obd_import *revimp;
563 struct lustre_handle conn;
564 struct obd_uuid tgtuuid;
565 struct obd_uuid cluuid;
566 struct obd_uuid remote_uuid;
570 int initial_conn = 0;
571 struct obd_connect_data *data;
572 int size[2] = { sizeof(struct ptlrpc_body), sizeof(*data) };
575 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
577 LASSERT_REQSWAB(req, REQ_REC_OFF);
578 str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF, sizeof(tgtuuid)-1);
580 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
581 GOTO(out, rc = -EINVAL);
584 obd_str2uuid (&tgtuuid, str);
585 target = class_uuid2obd(&tgtuuid);
587 target = class_name2obd(str);
589 if (!target || target->obd_stopping || !target->obd_set_up) {
590 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
591 " for connect (%s)\n", str,
592 !target ? "no target" :
593 (target->obd_stopping ? "stopping" :
595 GOTO(out, rc = -ENODEV);
598 if (target->obd_no_conn) {
599 LCONSOLE_WARN("%s: temporarily refusing client connection "
600 "from %s\n", target->obd_name,
601 libcfs_nid2str(req->rq_peer.nid));
602 GOTO(out, rc = -EAGAIN);
605 /* Make sure the target isn't cleaned up while we're here. Yes,
606 there's still a race between the above check and our incref here.
607 Really, class_uuid2obd should take the ref. */
608 targref = class_incref(target);
610 LASSERT_REQSWAB(req, REQ_REC_OFF + 1);
611 str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF + 1,
614 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
615 GOTO(out, rc = -EINVAL);
618 obd_str2uuid (&cluuid, str);
620 /* XXX extract a nettype and format accordingly */
621 switch (sizeof(lnet_nid_t)) {
622 /* NB the casts only avoid compiler warnings */
624 snprintf(remote_uuid.uuid, sizeof remote_uuid,
625 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
628 snprintf(remote_uuid.uuid, sizeof remote_uuid,
629 "NET_%x_UUID", (__u32)req->rq_peer.nid);
635 tmp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn);
637 GOTO(out, rc = -EPROTO);
639 memcpy(&conn, tmp, sizeof conn);
641 data = lustre_swab_reqbuf(req, REQ_REC_OFF + 3, sizeof(*data),
642 lustre_swab_connect);
645 GOTO(out, rc = -EPROTO);
647 rc = lustre_pack_reply(req, 2, size, NULL);
651 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
653 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
654 "libclient connection attempt");
655 GOTO(out, rc = -EPROTO);
656 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
657 LUSTRE_VERSION_ALLOWED_OFFSET ||
658 data->ocd_version > LUSTRE_VERSION_CODE +
659 LUSTRE_VERSION_ALLOWED_OFFSET) {
660 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
661 "libclient connection attempt",
662 data->ocd_version < LUSTRE_VERSION_CODE ?
664 OBD_OCD_VERSION_MAJOR(data->ocd_version),
665 OBD_OCD_VERSION_MINOR(data->ocd_version),
666 OBD_OCD_VERSION_PATCH(data->ocd_version),
667 OBD_OCD_VERSION_FIX(data->ocd_version));
668 data = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
669 offsetof(typeof(*data),
671 sizeof(data->ocd_version));
673 data->ocd_connect_flags = OBD_CONNECT_VERSION;
674 data->ocd_version = LUSTRE_VERSION_CODE;
676 GOTO(out, rc = -EPROTO);
680 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
683 /* lctl gets a backstage, all-access pass. */
684 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
685 goto dont_check_exports;
687 spin_lock(&target->obd_dev_lock);
688 list_for_each(p, &target->obd_exports) {
689 export = list_entry(p, struct obd_export, exp_obd_chain);
690 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
691 if (export->exp_connecting) { /* bug 9635, et. al. */
692 CWARN("%s: exp %p already connecting\n",
693 export->exp_obd->obd_name, export);
699 /* make darn sure this is coming from the same peer
700 * if the UUIDs matched */
701 if ((export->exp_connection != NULL) &&
702 (strcmp(libcfs_nid2str(req->rq_peer.nid),
703 libcfs_nid2str(export->exp_connection->c_peer.nid)))) {
704 CWARN("%s: cookie %s seen on new NID %s when "
705 "existing NID %s is already connected\n",
706 target->obd_name, cluuid.uuid,
707 libcfs_nid2str(req->rq_peer.nid),
708 libcfs_nid2str(export->exp_connection->c_peer.nid));
714 spin_lock(&export->exp_lock);
715 export->exp_connecting = 1;
716 spin_unlock(&export->exp_lock);
717 spin_unlock(&target->obd_dev_lock);
718 LASSERT(export->exp_obd == target);
720 rc = target_handle_reconnect(&conn, export, &cluuid,
727 /* If we found an export, we already unlocked. */
729 spin_unlock(&target->obd_dev_lock);
730 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
731 } else if (req->rq_export == NULL &&
732 atomic_read(&export->exp_rpc_count) > 0) {
733 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
734 target->obd_name, cluuid.uuid,
735 libcfs_nid2str(req->rq_peer.nid),
736 export, atomic_read(&export->exp_refcount));
737 GOTO(out, rc = -EBUSY);
738 } else if (req->rq_export != NULL &&
739 (atomic_read(&export->exp_rpc_count) > 1)) {
740 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
741 target->obd_name, cluuid.uuid,
742 libcfs_nid2str(req->rq_peer.nid),
743 export, atomic_read(&export->exp_rpc_count));
744 GOTO(out, rc = -EBUSY);
745 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
747 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
748 "cookies not random?\n", target->obd_name,
749 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
750 GOTO(out, rc = -EALREADY);
752 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
753 if (req->rq_export == NULL && initial_conn)
754 export->exp_last_request_time =
755 max(export->exp_last_request_time,
756 (time_t)CURRENT_SECONDS);
759 /* We want to handle EALREADY but *not* -EALREADY from
760 * target_handle_reconnect(), return reconnection state in a flag */
761 if (rc == EALREADY) {
762 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
767 /* Tell the client if we're in recovery. */
768 /* If this is the first client, start the recovery timer */
769 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
770 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
771 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
772 export, (long)CURRENT_SECONDS,
773 export ? (long)export->exp_last_request_time : 0);
776 if (target->obd_recovering) {
777 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
778 target_start_recovery_timer(target);
781 /* Tell the client if we support replayable requests */
782 if (target->obd_replayable)
783 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
785 if (export == NULL) {
786 if (target->obd_recovering) {
787 CERROR("%s: denying connection for new client %s (%s): "
788 "%d clients in recovery for %lds\n",
790 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
791 target->obd_recoverable_clients,
792 cfs_duration_sec(cfs_time_sub(cfs_timer_deadline(&target->obd_recovery_timer),
793 cfs_time_current())));
797 rc = obd_connect(req->rq_svc_thread->t_env,
798 &conn, target, &cluuid, data);
801 rc = obd_reconnect(export, target, &cluuid, data);
805 /* Return only the parts of obd_connect_data that we understand, so the
806 * client knows that we don't understand the rest. */
808 memcpy(lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
810 data, sizeof(*data));
812 /* If all else goes well, this is our RPC return code. */
815 lustre_msg_set_handle(req->rq_repmsg, &conn);
817 /* ownership of this export ref transfers to the request AFTER we
818 * drop any previous reference the request had, but we don't want
819 * that to go to zero before we get our new export reference. */
820 export = class_conn2export(&conn);
822 DEBUG_REQ(D_ERROR, req, "Missing export!");
823 GOTO(out, rc = -ENODEV);
826 /* If the client and the server are the same node, we will already
827 * have an export that really points to the client's DLM export,
828 * because we have a shared handles table.
830 * XXX this will go away when shaver stops sending the "connect" handle
831 * in the real "remote handle" field of the request --phik 24 Apr 2003
833 if (req->rq_export != NULL)
834 class_export_put(req->rq_export);
836 req->rq_export = export;
838 spin_lock(&export->exp_lock);
840 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
841 } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
842 spin_unlock(&export->exp_lock);
843 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
844 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
845 export->exp_conn_cnt,
846 lustre_msg_get_conn_cnt(req->rq_reqmsg));
848 GOTO(out, rc = -EALREADY);
850 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
852 /* request from liblustre? Don't evict it for not pinging. */
853 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
854 export->exp_libclient = 1;
855 spin_unlock(&export->exp_lock);
857 spin_lock(&target->obd_dev_lock);
858 list_del_init(&export->exp_obd_chain_timed);
859 spin_unlock(&target->obd_dev_lock);
861 spin_unlock(&export->exp_lock);
864 if (export->exp_connection != NULL)
865 ptlrpc_put_connection(export->exp_connection);
866 export->exp_connection = ptlrpc_get_connection(req->rq_peer,
870 if (lustre_msg_get_op_flags(req->rq_repmsg) & MSG_CONNECT_RECONNECT) {
871 LASSERT(export->exp_imp_reverse);
872 sptlrpc_svc_install_rvs_ctx(export->exp_imp_reverse,
877 spin_lock_bh(&target->obd_processing_task_lock);
878 if (target->obd_recovering && !export->exp_in_recovery) {
879 spin_lock(&export->exp_lock);
880 export->exp_in_recovery = 1;
881 export->exp_req_replay_needed = 1;
882 export->exp_lock_replay_needed = 1;
883 spin_unlock(&export->exp_lock);
884 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
885 && data->ocd_transno < target->obd_next_recovery_transno)
886 target->obd_next_recovery_transno = data->ocd_transno;
887 target->obd_connected_clients++;
888 /* each connected client is counted as recoverable */
889 target->obd_recoverable_clients++;
890 atomic_inc(&target->obd_req_replay_clients);
891 atomic_inc(&target->obd_lock_replay_clients);
892 if (target->obd_connected_clients ==
893 target->obd_max_recoverable_clients)
894 wake_up(&target->obd_next_transno_waitq);
896 spin_unlock_bh(&target->obd_processing_task_lock);
898 lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn),
901 if (export->exp_imp_reverse != NULL) {
902 /* destroyed import can be still referenced in ctxt */
903 obd_set_info_async(export, strlen(KEY_REVIMP_UPD),
904 KEY_REVIMP_UPD, 0, NULL, NULL);
905 destroy_import(export->exp_imp_reverse);
908 /* for the rest part, we return -ENOTCONN in case of errors
909 * in order to let client initialize connection again.
911 revimp = export->exp_imp_reverse = class_new_import(target);
913 CERROR("fail to alloc new reverse import.\n");
914 GOTO(out, rc = -ENOTCONN);
917 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
918 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
919 revimp->imp_remote_handle = conn;
920 revimp->imp_dlm_fake = 1;
921 revimp->imp_state = LUSTRE_IMP_FULL;
923 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_NEXT_VER) {
924 revimp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
925 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_NEXT_VER);
928 rc = sptlrpc_import_get_sec(revimp, req->rq_svc_ctx,
929 req->rq_sec_flavor, 0);
931 CERROR("Failed to get sec for reverse import: %d\n", rc);
932 export->exp_imp_reverse = NULL;
933 class_destroy_import(revimp);
936 class_import_put(revimp);
939 spin_lock(&export->exp_lock);
940 export->exp_connecting = 0;
941 spin_unlock(&export->exp_lock);
944 class_decref(targref);
950 int target_handle_disconnect(struct ptlrpc_request *req)
955 rc = lustre_pack_reply(req, 1, NULL, NULL);
959 /* keep the rq_export around so we can send the reply */
960 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
965 void target_destroy_export(struct obd_export *exp)
967 /* exports created from last_rcvd data, and "fake"
968 exports created by lctl don't have an import */
969 if (exp->exp_imp_reverse != NULL)
970 destroy_import(exp->exp_imp_reverse);
972 /* We cancel locks at disconnect time, but this will catch any locks
973 * granted in a race with recovery-induced disconnect. */
974 if (exp->exp_obd->obd_namespace != NULL)
975 ldlm_cancel_locks_for_export(exp);
982 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
984 struct ptlrpc_request *copy_req;
985 struct lustre_msg *copy_reqmsg;
986 struct ptlrpc_user_desc *udesc = NULL;
988 OBD_ALLOC_PTR(copy_req);
991 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
993 OBD_FREE_PTR(copy_req);
997 if (orig_req->rq_user_desc) {
998 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1000 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1002 OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1003 OBD_FREE_PTR(copy_req);
1006 memcpy(udesc, orig_req->rq_user_desc,
1007 sptlrpc_user_desc_size(ngroups));
1010 *copy_req = *orig_req;
1011 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1012 copy_req->rq_reqmsg = copy_reqmsg;
1013 copy_req->rq_user_desc = udesc;
1015 class_export_get(copy_req->rq_export);
1016 CFS_INIT_LIST_HEAD(©_req->rq_list);
1017 sptlrpc_svc_ctx_addref(copy_req);
1019 if (copy_req->rq_reply_state) {
1020 /* the copied req takes over the reply state */
1021 orig_req->rq_reply_state = NULL;
1022 /* to catch further access */
1023 orig_req->rq_repmsg = NULL;
1024 orig_req->rq_replen = 0;
1030 void ptlrpc_free_clone( struct ptlrpc_request *req)
1032 if (req->rq_reply_state) {
1033 ptlrpc_rs_decref(req->rq_reply_state);
1034 req->rq_reply_state = NULL;
1037 sptlrpc_svc_ctx_decref(req);
1038 class_export_put(req->rq_export);
1039 list_del(&req->rq_list);
1041 if (req->rq_user_desc) {
1042 int ngroups = req->rq_user_desc->pud_ngroups;
1043 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1045 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1050 static void target_finish_recovery(struct obd_device *obd)
1053 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1056 ldlm_reprocess_all_ns(obd->obd_namespace);
1058 /* when recovery finished, cleanup orphans on mds and ost */
1059 if (OBT(obd) && OBP(obd, postrecov)) {
1060 int rc = OBP(obd, postrecov)(obd);
1061 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1062 rc < 0 ? "failed" : "complete", rc);
1065 obd->obd_recovery_end = CURRENT_SECONDS;
1069 static void abort_req_replay_queue(struct obd_device *obd)
1071 struct ptlrpc_request *req, *n;
1073 list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1074 DEBUG_REQ(D_WARNING, req, "aborted:");
1075 req->rq_status = -ENOTCONN;
1076 if (ptlrpc_error(req)) {
1077 DEBUG_REQ(D_ERROR, req,
1078 "failed abort_req_reply; skipping");
1080 ptlrpc_free_clone(req);
1084 static void abort_lock_replay_queue(struct obd_device *obd)
1086 struct ptlrpc_request *req, *n;
1088 list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1089 DEBUG_REQ(D_ERROR, req, "aborted:");
1090 req->rq_status = -ENOTCONN;
1091 if (ptlrpc_error(req)) {
1092 DEBUG_REQ(D_ERROR, req,
1093 "failed abort_lock_reply; skipping");
1095 ptlrpc_free_clone(req);
1100 /* Called from a cleanup function if the device is being cleaned up
1101 forcefully. The exports should all have been disconnected already,
1102 the only thing left to do is
1103 - clear the recovery flags
1105 - free queued requests and replies, but don't send replies
1106 Because the obd_stopping flag is set, no new requests should be received.
1109 void target_cleanup_recovery(struct obd_device *obd)
1111 struct ptlrpc_request *req, *n;
1114 LASSERT(obd->obd_stopping);
1116 spin_lock_bh(&obd->obd_processing_task_lock);
1117 if (!obd->obd_recovering) {
1118 spin_unlock_bh(&obd->obd_processing_task_lock);
1122 obd->obd_recovering = obd->obd_abort_recovery = 0;
1123 target_cancel_recovery_timer(obd);
1124 spin_unlock_bh(&obd->obd_processing_task_lock);
1126 list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1127 LASSERT (req->rq_reply_state == 0);
1128 ptlrpc_free_clone(req);
1130 list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1131 LASSERT (req->rq_reply_state == 0);
1132 ptlrpc_free_clone(req);
1134 list_for_each_entry_safe(req, n, &obd->obd_final_req_queue, rq_list) {
1135 LASSERT (req->rq_reply_state == 0);
1136 ptlrpc_free_clone(req);
1142 static void target_recovery_expired(unsigned long castmeharder)
1144 struct obd_device *obd = (struct obd_device *)castmeharder;
1145 CERROR("%s: recovery timed out, aborting\n", obd->obd_name);
1146 spin_lock_bh(&obd->obd_processing_task_lock);
1147 if (obd->obd_recovering)
1148 obd->obd_abort_recovery = 1;
1149 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1150 spin_unlock_bh(&obd->obd_processing_task_lock);
1154 /* obd_processing_task_lock should be held */
1155 void target_cancel_recovery_timer(struct obd_device *obd)
1157 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1158 cfs_timer_disarm(&obd->obd_recovery_timer);
1161 static void reset_recovery_timer(struct obd_device *obd)
1163 spin_lock_bh(&obd->obd_processing_task_lock);
1164 if (!obd->obd_recovering) {
1165 spin_unlock_bh(&obd->obd_processing_task_lock);
1168 cfs_timer_arm(&obd->obd_recovery_timer,
1169 cfs_time_shift(OBD_RECOVERY_TIMEOUT));
1170 spin_unlock_bh(&obd->obd_processing_task_lock);
1171 CDEBUG(D_HA, "%s: timer will expire in %u seconds\n", obd->obd_name,
1172 OBD_RECOVERY_TIMEOUT);
1173 /* Only used for lprocfs_status */
1174 obd->obd_recovery_end = CURRENT_SECONDS + OBD_RECOVERY_TIMEOUT;
1178 /* Only start it the first time called */
1179 void target_start_recovery_timer(struct obd_device *obd)
1181 spin_lock_bh(&obd->obd_processing_task_lock);
1182 if (obd->obd_recovery_handler
1183 || timer_pending((struct timer_list *)&obd->obd_recovery_timer)) {
1184 spin_unlock_bh(&obd->obd_processing_task_lock);
1187 CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1188 OBD_RECOVERY_TIMEOUT);
1189 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1190 spin_unlock_bh(&obd->obd_processing_task_lock);
1192 reset_recovery_timer(obd);
1196 static int check_for_next_transno(struct obd_device *obd)
1198 struct ptlrpc_request *req = NULL;
1199 int wake_up = 0, connected, completed, queue_len, max;
1200 __u64 next_transno, req_transno;
1202 spin_lock_bh(&obd->obd_processing_task_lock);
1204 if (!list_empty(&obd->obd_req_replay_queue)) {
1205 req = list_entry(obd->obd_req_replay_queue.next,
1206 struct ptlrpc_request, rq_list);
1207 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1212 max = obd->obd_max_recoverable_clients;
1213 connected = obd->obd_connected_clients;
1214 completed = connected - obd->obd_recoverable_clients;
1215 queue_len = obd->obd_requests_queued_for_recovery;
1216 next_transno = obd->obd_next_recovery_transno;
1218 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1219 "req_transno: "LPU64", next_transno: "LPU64"\n",
1220 max, connected, completed, queue_len, req_transno, next_transno);
1222 if (obd->obd_abort_recovery) {
1223 CDEBUG(D_HA, "waking for aborted recovery\n");
1225 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1226 CDEBUG(D_HA, "waking for completed recovery\n");
1228 } else if (req_transno == next_transno) {
1229 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1231 } else if (queue_len + completed == max) {
1232 /* handle gaps occured due to lost reply. It is allowed gaps
1233 * because all clients are connected and there will be resend
1234 * for missed transaction */
1235 LASSERTF(req_transno >= next_transno,
1236 "req_transno: "LPU64", next_transno: "LPU64"\n",
1237 req_transno, next_transno);
1239 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1240 "waking for skipped transno (skip: "LPD64
1241 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1242 next_transno, queue_len, completed, connected, req_transno);
1243 obd->obd_next_recovery_transno = req_transno;
1245 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1246 /* some clients haven't connected in time, but we can try
1247 * to replay requests that demand on already committed ones
1248 * also, we can replay first non-committed transation */
1249 LASSERT(req_transno != 0);
1250 if (req_transno == obd->obd_last_committed + 1) {
1251 obd->obd_next_recovery_transno = req_transno;
1252 } else if (req_transno > obd->obd_last_committed) {
1253 /* can't continue recovery: have no needed transno */
1254 obd->obd_abort_recovery = 1;
1255 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1256 "connected: %d, completed: %d, queue_len: %d, "
1257 "req_transno: "LPU64", next_transno: "LPU64"\n",
1258 max, connected, completed, queue_len,
1259 req_transno, next_transno);
1264 spin_unlock_bh(&obd->obd_processing_task_lock);
1268 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1270 struct l_wait_info lwi = { 0 };
1271 struct ptlrpc_request *req;
1273 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1274 obd->obd_next_recovery_transno);
1275 l_wait_event(obd->obd_next_transno_waitq,
1276 check_for_next_transno(obd), &lwi);
1278 spin_lock_bh(&obd->obd_processing_task_lock);
1279 if (obd->obd_abort_recovery) {
1281 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1282 req = list_entry(obd->obd_req_replay_queue.next,
1283 struct ptlrpc_request, rq_list);
1284 list_del_init(&req->rq_list);
1285 obd->obd_requests_queued_for_recovery--;
1289 spin_unlock_bh(&obd->obd_processing_task_lock);
1293 static int check_for_next_lock(struct obd_device *obd)
1295 struct ptlrpc_request *req = NULL;
1298 spin_lock_bh(&obd->obd_processing_task_lock);
1299 if (!list_empty(&obd->obd_lock_replay_queue)) {
1300 req = list_entry(obd->obd_lock_replay_queue.next,
1301 struct ptlrpc_request, rq_list);
1302 CDEBUG(D_HA, "waking for next lock\n");
1304 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1305 CDEBUG(D_HA, "waking for completed lock replay\n");
1307 } else if (obd->obd_abort_recovery) {
1308 CDEBUG(D_HA, "waking for aborted recovery\n");
1311 spin_unlock_bh(&obd->obd_processing_task_lock);
1316 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1318 struct l_wait_info lwi = { 0 };
1319 struct ptlrpc_request *req;
1321 CDEBUG(D_HA, "Waiting for lock\n");
1322 l_wait_event(obd->obd_next_transno_waitq,
1323 check_for_next_lock(obd), &lwi);
1325 spin_lock_bh(&obd->obd_processing_task_lock);
1326 if (obd->obd_abort_recovery) {
1328 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1329 req = list_entry(obd->obd_lock_replay_queue.next,
1330 struct ptlrpc_request, rq_list);
1331 list_del_init(&req->rq_list);
1335 spin_unlock_bh(&obd->obd_processing_task_lock);
1339 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1341 struct ptlrpc_request *req;
1343 spin_lock_bh(&obd->obd_processing_task_lock);
1344 if (!list_empty(&obd->obd_final_req_queue)) {
1345 req = list_entry(obd->obd_final_req_queue.next,
1346 struct ptlrpc_request, rq_list);
1347 list_del_init(&req->rq_list);
1351 spin_unlock_bh(&obd->obd_processing_task_lock);
1355 static inline int req_replay_done(struct obd_export *exp)
1357 return (exp->exp_req_replay_needed == 0);
1360 static inline int lock_replay_done(struct obd_export *exp)
1362 return (exp->exp_lock_replay_needed == 0);
1365 static inline int connect_done(struct obd_export *exp)
1367 return (exp->exp_in_recovery != 0);
1370 static int check_for_clients(struct obd_device *obd)
1372 if (obd->obd_abort_recovery)
1374 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1375 if (obd->obd_no_conn == 0 &&
1376 obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1381 static int handle_recovery_req(struct ptlrpc_thread *thread,
1382 struct ptlrpc_request *req,
1383 svc_handler_t handler)
1388 rc = lu_context_init(&req->rq_session, LCT_SESSION);
1390 CERROR("Failure to initialize session: %d\n", rc);
1393 req->rq_session.lc_thread = thread;
1394 lu_context_enter(&req->rq_session);
1395 req->rq_svc_thread = thread;
1396 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1400 lu_context_exit(&req->rq_session);
1401 lu_context_fini(&req->rq_session);
1402 /* don't reset timer for final stage */
1403 if (!req_replay_done(req->rq_export) ||
1404 !lock_replay_done(req->rq_export))
1405 reset_recovery_timer(class_exp2obd(req->rq_export));
1406 ptlrpc_free_clone(req);
1410 static int target_recovery_thread(void *arg)
1412 struct obd_device *obd = arg;
1413 struct ptlrpc_request *req;
1414 struct target_recovery_data *trd = &obd->obd_recovery_data;
1415 struct l_wait_info lwi = { 0 };
1416 unsigned long delta;
1417 unsigned long flags;
1419 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1420 __u32 recov_ctx_tags = LCT_MD_THREAD;
1424 cfs_daemonize("tgt_recov");
1426 SIGNAL_MASK_LOCK(current, flags);
1427 sigfillset(¤t->blocked);
1429 SIGNAL_MASK_UNLOCK(current, flags);
1431 rc = lu_context_init(&env.le_ctx, recov_ctx_tags);
1435 thread->t_env = &env;
1436 env.le_ctx.lc_thread = thread;
1438 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1440 trd->trd_processing_task = current->pid;
1442 obd->obd_recovering = 1;
1443 complete(&trd->trd_starting);
1445 /* first of all, we have to know the first transno to replay */
1446 obd->obd_abort_recovery = 0;
1447 l_wait_event(obd->obd_next_transno_waitq,
1448 check_for_clients(obd), &lwi);
1450 spin_lock_bh(&obd->obd_processing_task_lock);
1451 target_cancel_recovery_timer(obd);
1452 spin_unlock_bh(&obd->obd_processing_task_lock);
1454 /* If some clients haven't connected in time, evict them */
1455 if (obd->obd_abort_recovery) {
1456 CWARN("Some clients haven't connect in time (%d/%d),"
1457 "evict them\n", obd->obd_connected_clients,
1458 obd->obd_max_recoverable_clients);
1459 obd->obd_abort_recovery = obd->obd_stopping;
1460 class_disconnect_stale_exports(obd, connect_done);
1462 /* next stage: replay requests */
1464 obd->obd_req_replaying = 1;
1465 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1466 atomic_read(&obd->obd_req_replay_clients),
1467 obd->obd_next_recovery_transno);
1468 while ((req = target_next_replay_req(obd))) {
1469 LASSERT(trd->trd_processing_task == current->pid);
1470 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1471 lustre_msg_get_transno(req->rq_reqmsg),
1472 libcfs_nid2str(req->rq_peer.nid));
1474 handle_recovery_req(thread, req,
1475 trd->trd_recovery_handler);
1476 obd->obd_replayed_requests++;
1477 spin_lock_bh(&obd->obd_processing_task_lock);
1478 obd->obd_next_recovery_transno++;
1479 spin_unlock_bh(&obd->obd_processing_task_lock);
1482 spin_lock_bh(&obd->obd_processing_task_lock);
1483 target_cancel_recovery_timer(obd);
1484 spin_unlock_bh(&obd->obd_processing_task_lock);
1486 /* If some clients haven't replayed requests in time, evict them */
1487 if (obd->obd_abort_recovery) {
1488 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1489 obd->obd_abort_recovery = obd->obd_stopping;
1490 class_disconnect_stale_exports(obd, req_replay_done);
1491 abort_req_replay_queue(obd);
1493 /* The second stage: replay locks */
1494 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1495 atomic_read(&obd->obd_lock_replay_clients));
1496 while ((req = target_next_replay_lock(obd))) {
1497 LASSERT(trd->trd_processing_task == current->pid);
1498 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1499 libcfs_nid2str(req->rq_peer.nid));
1500 handle_recovery_req(thread, req,
1501 trd->trd_recovery_handler);
1502 obd->obd_replayed_locks++;
1505 spin_lock_bh(&obd->obd_processing_task_lock);
1506 target_cancel_recovery_timer(obd);
1507 spin_unlock_bh(&obd->obd_processing_task_lock);
1508 /* If some clients haven't replayed requests in time, evict them */
1509 if (obd->obd_abort_recovery) {
1511 CERROR("lock replay timed out, aborting ...\n");
1512 obd->obd_abort_recovery = obd->obd_stopping;
1513 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1514 abort_lock_replay_queue(obd);
1517 /* We drop recoverying flag to forward all new requests
1518 * to regular mds_handle() since now */
1519 spin_lock_bh(&obd->obd_processing_task_lock);
1520 obd->obd_recovering = obd->obd_abort_recovery = 0;
1521 spin_unlock_bh(&obd->obd_processing_task_lock);
1522 /* The third stage: reply on final pings */
1523 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1524 while ((req = target_next_final_ping(obd))) {
1525 LASSERT(trd->trd_processing_task == current->pid);
1526 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1527 libcfs_nid2str(req->rq_peer.nid));
1528 handle_recovery_req(thread, req,
1529 trd->trd_recovery_handler);
1532 delta = (jiffies - delta) / HZ;
1533 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1534 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1535 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1536 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1537 if (delta > obd_timeout * 2) {
1538 CWARN("too long recovery - read logs\n");
1539 libcfs_debug_dumplog();
1542 target_finish_recovery(obd);
1545 trd->trd_processing_task = 0;
1546 complete(&trd->trd_finishing);
1550 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1553 struct target_recovery_data *trd = &obd->obd_recovery_data;
1555 memset(trd, 0, sizeof(*trd));
1556 init_completion(&trd->trd_starting);
1557 init_completion(&trd->trd_finishing);
1558 trd->trd_recovery_handler = handler;
1560 if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1561 wait_for_completion(&trd->trd_starting);
1562 LASSERT(obd->obd_recovering != 0);
1569 void target_stop_recovery_thread(struct obd_device *obd)
1571 spin_lock_bh(&obd->obd_processing_task_lock);
1572 if (obd->obd_recovery_data.trd_processing_task > 0) {
1573 struct target_recovery_data *trd = &obd->obd_recovery_data;
1574 CERROR("%s: Aborting recovery\n", obd->obd_name);
1575 obd->obd_abort_recovery = 1;
1576 wake_up(&obd->obd_next_transno_waitq);
1577 spin_unlock_bh(&obd->obd_processing_task_lock);
1578 wait_for_completion(&trd->trd_finishing);
1580 spin_unlock_bh(&obd->obd_processing_task_lock);
1584 void target_recovery_fini(struct obd_device *obd)
1586 class_disconnect_exports(obd);
1587 target_stop_recovery_thread(obd);
1588 target_cleanup_recovery(obd);
1590 EXPORT_SYMBOL(target_recovery_fini);
1592 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1594 if (obd->obd_max_recoverable_clients == 0)
1597 CWARN("RECOVERY: service %s, %d recoverable clients, "
1598 "last_transno "LPU64"\n", obd->obd_name,
1599 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1600 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1601 target_start_recovery_thread(obd, handler);
1602 obd->obd_recovery_start = CURRENT_SECONDS;
1603 /* Only used for lprocfs_status */
1604 obd->obd_recovery_end = obd->obd_recovery_start + OBD_RECOVERY_TIMEOUT;
1606 EXPORT_SYMBOL(target_recovery_init);
1610 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1612 struct obd_export *exp = req->rq_export;
1613 LASSERT(exp != NULL);
1614 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1615 /* client declares he's ready to replay locks */
1616 spin_lock_bh(&obd->obd_processing_task_lock);
1617 if (exp->exp_req_replay_needed) {
1618 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1619 spin_lock(&exp->exp_lock);
1620 exp->exp_req_replay_needed = 0;
1621 spin_unlock(&exp->exp_lock);
1622 atomic_dec(&obd->obd_req_replay_clients);
1623 LASSERT(obd->obd_recoverable_clients > 0);
1624 obd->obd_recoverable_clients--;
1625 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1626 CDEBUG(D_HA, "all clients have replayed reqs\n");
1627 wake_up(&obd->obd_next_transno_waitq);
1629 spin_unlock_bh(&obd->obd_processing_task_lock);
1631 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1632 /* client declares he's ready to complete recovery
1633 * so, we put the request on th final queue */
1634 spin_lock_bh(&obd->obd_processing_task_lock);
1635 if (exp->exp_lock_replay_needed) {
1636 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1637 spin_lock(&exp->exp_lock);
1638 exp->exp_lock_replay_needed = 0;
1639 spin_unlock(&exp->exp_lock);
1640 atomic_dec(&obd->obd_lock_replay_clients);
1641 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1642 CDEBUG(D_HA, "all clients have replayed locks\n");
1643 wake_up(&obd->obd_next_transno_waitq);
1645 spin_unlock_bh(&obd->obd_processing_task_lock);
1651 int target_queue_recovery_request(struct ptlrpc_request *req,
1652 struct obd_device *obd)
1654 struct list_head *tmp;
1656 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1660 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1661 /* Processing the queue right now, don't re-add. */
1665 target_process_req_flags(obd, req);
1667 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1668 /* client declares he's ready to complete recovery
1669 * so, we put the request on th final queue */
1670 req = ptlrpc_clone_req(req);
1673 DEBUG_REQ(D_HA, req, "queue final req");
1674 spin_lock_bh(&obd->obd_processing_task_lock);
1675 if (obd->obd_recovering)
1676 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1678 spin_unlock_bh(&obd->obd_processing_task_lock);
1679 ptlrpc_free_clone(req);
1680 if (obd->obd_stopping) {
1686 spin_unlock_bh(&obd->obd_processing_task_lock);
1689 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1690 /* client declares he's ready to replay locks */
1691 req = ptlrpc_clone_req(req);
1694 DEBUG_REQ(D_HA, req, "queue lock replay req");
1695 spin_lock_bh(&obd->obd_processing_task_lock);
1696 LASSERT(obd->obd_recovering);
1697 /* usually due to recovery abort */
1698 if (!req->rq_export->exp_in_recovery) {
1699 spin_unlock_bh(&obd->obd_processing_task_lock);
1700 ptlrpc_free_clone(req);
1703 LASSERT(req->rq_export->exp_lock_replay_needed);
1704 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1705 spin_unlock_bh(&obd->obd_processing_task_lock);
1706 wake_up(&obd->obd_next_transno_waitq);
1710 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1711 * (i.e. buflens etc are in my own byte order), but type-dependent
1712 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1715 CFS_INIT_LIST_HEAD(&req->rq_list);
1716 DEBUG_REQ(D_HA, req, "not queueing");
1720 spin_lock_bh(&obd->obd_processing_task_lock);
1722 /* If we're processing the queue, we want don't want to queue this
1725 * Also, if this request has a transno less than the one we're waiting
1726 * for, we should process it now. It could (and currently always will)
1727 * be an open request for a descriptor that was opened some time ago.
1729 * Also, a resent, replayed request that has already been
1730 * handled will pass through here and be processed immediately.
1732 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1733 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1734 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1735 /* Processing the queue right now, don't re-add. */
1736 LASSERT(list_empty(&req->rq_list));
1737 spin_unlock_bh(&obd->obd_processing_task_lock);
1740 spin_unlock_bh(&obd->obd_processing_task_lock);
1742 /* A resent, replayed request that is still on the queue; just drop it.
1743 The queued request will handle this. */
1744 if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT|MSG_REPLAY)) ==
1745 (MSG_RESENT | MSG_REPLAY)) {
1746 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1750 req = ptlrpc_clone_req(req);
1754 spin_lock_bh(&obd->obd_processing_task_lock);
1755 LASSERT(obd->obd_recovering);
1756 if (!req->rq_export->exp_in_recovery) {
1757 spin_unlock_bh(&obd->obd_processing_task_lock);
1758 ptlrpc_free_clone(req);
1761 LASSERT(req->rq_export->exp_req_replay_needed);
1764 list_for_each(tmp, &obd->obd_req_replay_queue) {
1765 struct ptlrpc_request *reqiter =
1766 list_entry(tmp, struct ptlrpc_request, rq_list);
1768 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1769 list_add_tail(&req->rq_list, &reqiter->rq_list);
1776 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1778 obd->obd_requests_queued_for_recovery++;
1779 wake_up(&obd->obd_next_transno_waitq);
1780 spin_unlock_bh(&obd->obd_processing_task_lock);
1785 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1787 return req->rq_export->exp_obd;
1790 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
1792 if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1793 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1794 DEBUG_REQ(D_ERROR, req, "dropping reply");
1799 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1800 req->rq_status = rc;
1801 return (ptlrpc_error(req));
1803 DEBUG_REQ(D_NET, req, "sending reply");
1806 return (ptlrpc_send_reply(req, 1));
1809 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1812 struct ptlrpc_reply_state *rs;
1813 struct obd_device *obd;
1814 struct obd_export *exp;
1815 struct ptlrpc_service *svc;
1817 svc = req->rq_rqbd->rqbd_service;
1818 rs = req->rq_reply_state;
1819 if (rs == NULL || !rs->rs_difficult) {
1821 target_send_reply_msg (req, rc, fail_id);
1825 /* must be an export if locks saved */
1826 LASSERT (req->rq_export != NULL);
1827 /* req/reply consistent */
1828 LASSERT (rs->rs_service == svc);
1831 LASSERT (!rs->rs_scheduled);
1832 LASSERT (!rs->rs_scheduled_ever);
1833 LASSERT (!rs->rs_handled);
1834 LASSERT (!rs->rs_on_net);
1835 LASSERT (rs->rs_export == NULL);
1836 LASSERT (list_empty(&rs->rs_obd_list));
1837 LASSERT (list_empty(&rs->rs_exp_list));
1839 exp = class_export_get (req->rq_export);
1842 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1843 rs->rs_scheduled = 1;
1845 rs->rs_xid = req->rq_xid;
1846 rs->rs_transno = req->rq_transno;
1847 rs->rs_export = exp;
1849 spin_lock(&obd->obd_uncommitted_replies_lock);
1851 if (rs->rs_transno > obd->obd_last_committed) {
1852 /* not committed already */
1853 list_add_tail (&rs->rs_obd_list,
1854 &obd->obd_uncommitted_replies);
1857 spin_unlock (&obd->obd_uncommitted_replies_lock);
1858 spin_lock (&exp->exp_lock);
1860 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1862 spin_unlock(&exp->exp_lock);
1864 netrc = target_send_reply_msg (req, rc, fail_id);
1866 spin_lock(&svc->srv_lock);
1868 svc->srv_n_difficult_replies++;
1871 /* error sending: reply is off the net. Also we need +1
1872 * reply ref until ptlrpc_server_handle_reply() is done
1873 * with the reply state (if the send was successful, there
1874 * would have been +1 ref for the net, which
1875 * reply_out_callback leaves alone) */
1877 ptlrpc_rs_addref(rs);
1878 atomic_inc (&svc->srv_outstanding_replies);
1881 if (!rs->rs_on_net || /* some notifier */
1882 list_empty(&rs->rs_exp_list) || /* completed already */
1883 list_empty(&rs->rs_obd_list)) {
1884 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1885 cfs_waitq_signal (&svc->srv_waitq);
1887 list_add (&rs->rs_list, &svc->srv_active_replies);
1888 rs->rs_scheduled = 0; /* allow notifier to schedule */
1891 spin_unlock(&svc->srv_lock);
1894 int target_handle_ping(struct ptlrpc_request *req)
1896 obd_ping(req->rq_export);
1897 return lustre_pack_reply(req, 1, NULL, NULL);
1900 void target_committed_to_req(struct ptlrpc_request *req)
1902 struct obd_device *obd;
1904 if (req == NULL || req->rq_export == NULL)
1907 obd = req->rq_export->exp_obd;
1911 if (!obd->obd_no_transno && req->rq_repmsg != NULL)
1912 lustre_msg_set_last_committed(req->rq_repmsg,
1913 obd->obd_last_committed);
1915 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update");
1917 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
1918 obd->obd_last_committed, req->rq_transno, req->rq_xid);
1921 EXPORT_SYMBOL(target_committed_to_req);
1923 #ifdef HAVE_QUOTA_SUPPORT
1924 int target_handle_qc_callback(struct ptlrpc_request *req)
1926 struct obd_quotactl *oqctl;
1927 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
1929 oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1930 lustre_swab_obd_quotactl);
1931 if (oqctl == NULL) {
1932 CERROR("Can't unpack obd_quotactl\n");
1936 cli->cl_qchk_stat = oqctl->qc_stat;
1941 int target_handle_dqacq_callback(struct ptlrpc_request *req)
1944 struct obd_device *obd = req->rq_export->exp_obd;
1945 struct obd_device *master_obd;
1946 struct lustre_quota_ctxt *qctxt;
1947 struct qunit_data *qdata;
1949 struct qunit_data_old *qdata_old;
1951 int repsize[2] = { sizeof(struct ptlrpc_body),
1952 sizeof(struct qunit_data) };
1955 rc = lustre_pack_reply(req, 2, repsize, NULL);
1957 CERROR("packing reply failed!: rc = %d\n", rc);
1960 LASSERT(req->rq_export);
1962 /* fixed for bug10707 */
1963 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
1964 !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
1965 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
1966 rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1967 sizeof(struct qunit_data));
1969 qdata = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata),
1972 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
1973 rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1974 sizeof(struct qunit_data_old));
1976 qdata_old = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata_old),
1977 lustre_swab_qdata_old);
1978 qdata = lustre_quota_old_to_new(qdata_old);
1981 if (qdata == NULL) {
1982 CERROR("Can't unpack qunit_data\n");
1986 /* we use the observer */
1987 LASSERT(obd->obd_observer && obd->obd_observer->obd_observer);
1988 master_obd = obd->obd_observer->obd_observer;
1989 qctxt = &master_obd->u.obt.obt_qctxt;
1991 LASSERT(qctxt->lqc_handler);
1992 rc = qctxt->lqc_handler(master_obd, qdata,
1993 lustre_msg_get_opc(req->rq_reqmsg));
1994 if (rc && rc != -EDQUOT)
1995 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
1996 "dqacq failed! (rc:%d)\n", rc);
1998 /* the qd_count might be changed in lqc_handler */
1999 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2000 !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2001 memcpy(rep,qdata,sizeof(*qdata));
2003 qdata_old = lustre_quota_new_to_old(qdata);
2004 memcpy(rep,qdata_old,sizeof(*qdata_old));
2006 req->rq_status = rc;
2007 rc = ptlrpc_reply(req);
2012 #endif /* !__KERNEL__ */
2014 #endif /* HAVE_QUOTA_SUPPORT */
2016 ldlm_mode_t lck_compat_array[] = {
2017 [LCK_EX] LCK_COMPAT_EX,
2018 [LCK_PW] LCK_COMPAT_PW,
2019 [LCK_PR] LCK_COMPAT_PR,
2020 [LCK_CW] LCK_COMPAT_CW,
2021 [LCK_CR] LCK_COMPAT_CR,
2022 [LCK_NL] LCK_COMPAT_NL,
2023 [LCK_GROUP] LCK_COMPAT_GROUP