4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdt/mdt_open.c
34 * Lustre Metadata Target (mdt) open/close file handling
36 * Author: Huang Hua <huanghua@clusterfs.com>
39 #define DEBUG_SUBSYSTEM S_MDS
41 #include <lustre_acl.h>
42 #include <lustre_mds.h>
43 #include <lustre_swab.h>
44 #include "mdt_internal.h"
45 #include <lustre_nodemap.h>
47 /* we do nothing because we do not have refcount now */
48 static void mdt_mfd_get(void *mfdp)
52 static struct portals_handle_ops mfd_open_handle_ops = {
53 .hop_addref = mdt_mfd_get,
57 /* Create a new mdt_file_data struct, initialize it,
58 * and insert it to global hash table */
59 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
61 struct mdt_file_data *mfd;
66 INIT_LIST_HEAD(&mfd->mfd_open_handle.h_link);
67 mfd->mfd_open_handle.h_owner = med;
68 INIT_LIST_HEAD(&mfd->mfd_list);
69 class_handle_hash(&mfd->mfd_open_handle, &mfd_open_handle_ops);
76 * Find the mfd pointed to by handle in global hash table.
77 * In case of replay the handle is obsoleted
78 * but mfd can be found in mfd list by that handle.
79 * Callers need to be holding med_open_lock.
81 struct mdt_file_data *mdt_open_handle2mfd(struct mdt_export_data *med,
82 const struct lustre_handle *open_handle,
83 bool is_replay_or_resent)
85 struct mdt_file_data *mfd;
88 LASSERT(open_handle != NULL);
89 mfd = class_handle2object(open_handle->cookie, med);
90 /* during dw/setattr replay the mfd can be found by old handle */
91 if (mfd == NULL && is_replay_or_resent) {
92 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
93 if (mfd->mfd_open_handle_old.cookie ==
104 void mdt_mfd_free(struct mdt_file_data *mfd)
106 LASSERT(list_empty(&mfd->mfd_list));
107 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_open_handle);
110 static int mdt_create_data(struct mdt_thread_info *info,
111 struct mdt_object *p, struct mdt_object *o)
113 struct md_op_spec *spec = &info->mti_spec;
114 struct md_attr *ma = &info->mti_attr;
118 if (!md_should_create(spec->sp_cr_flags))
121 ma->ma_need = MA_INODE | MA_LOV;
123 mutex_lock(&o->mot_lov_mutex);
124 if (!o->mot_lov_created) {
125 rc = mdo_create_data(info->mti_env,
126 p ? mdt_object_child(p) : NULL,
127 mdt_object_child(o), spec, ma);
129 rc = mdt_attr_get_complex(info, o, ma);
131 if (rc == 0 && ma->ma_valid & MA_LOV)
132 o->mot_lov_created = 1;
135 mutex_unlock(&o->mot_lov_mutex);
139 int mdt_write_read(struct mdt_object *o)
143 spin_lock(&o->mot_write_lock);
144 rc = o->mot_write_count;
145 spin_unlock(&o->mot_write_lock);
149 int mdt_write_get(struct mdt_object *o)
153 spin_lock(&o->mot_write_lock);
154 if (o->mot_write_count < 0)
157 o->mot_write_count++;
158 spin_unlock(&o->mot_write_lock);
163 void mdt_write_put(struct mdt_object *o)
166 spin_lock(&o->mot_write_lock);
167 o->mot_write_count--;
168 spin_unlock(&o->mot_write_lock);
172 static int mdt_write_deny(struct mdt_object *o)
176 spin_lock(&o->mot_write_lock);
177 if (o->mot_write_count > 0)
180 o->mot_write_count--;
181 spin_unlock(&o->mot_write_lock);
185 static void mdt_write_allow(struct mdt_object *o)
188 spin_lock(&o->mot_write_lock);
189 o->mot_write_count++;
190 spin_unlock(&o->mot_write_lock);
194 /* there can be no real transaction so prepare the fake one */
195 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
197 struct mdt_device *mdt = info->mti_mdt;
198 struct ptlrpc_request *req = mdt_info_req(info);
199 struct tg_export_data *ted;
200 struct lsd_client_data *lcd;
203 if (mdt_rdonly(req->rq_export))
206 /* transaction has occurred already */
207 if (lustre_msg_get_transno(req->rq_repmsg) != 0)
210 if (tgt_is_multimodrpcs_client(req->rq_export)) {
213 /* generate an empty transaction to get a transno
215 th = dt_trans_create(info->mti_env, mdt->mdt_bottom);
217 rc = dt_trans_start(info->mti_env, mdt->mdt_bottom, th);
218 dt_trans_stop(info->mti_env, mdt->mdt_bottom, th);
223 spin_lock(&mdt->mdt_lut.lut_translock);
225 if (info->mti_transno != 0) {
226 struct obd_export *exp = req->rq_export;
228 CERROR("%s: replay trans %llu NID %s: rc = %d\n",
229 mdt_obd_name(mdt), info->mti_transno,
230 obd_export_nid2str(exp), rc);
231 spin_unlock(&mdt->mdt_lut.lut_translock);
234 } else if (info->mti_transno == 0) {
235 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
237 /* should be replay */
238 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
239 mdt->mdt_lut.lut_last_transno = info->mti_transno;
241 spin_unlock(&mdt->mdt_lut.lut_translock);
243 CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
245 req->rq_export->exp_obd->obd_last_committed);
247 req->rq_transno = info->mti_transno;
248 lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
250 /* update lcd in memory only for resent cases */
251 ted = &req->rq_export->exp_target_data;
253 mutex_lock(&ted->ted_lcd_lock);
255 if (info->mti_transno < lcd->lcd_last_transno &&
256 info->mti_transno != 0) {
257 /* This should happen during replay. Do not update
258 * last rcvd info if replay req transno < last transno,
259 * otherwise the following resend(after replay) can not
260 * be checked correctly by xid */
261 mutex_unlock(&ted->ted_lcd_lock);
262 CDEBUG(D_HA, "%s: transno = %llu < last_transno = %llu\n",
263 mdt_obd_name(mdt), info->mti_transno,
264 lcd->lcd_last_transno);
268 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
269 if (info->mti_transno != 0)
270 lcd->lcd_last_close_transno = info->mti_transno;
271 lcd->lcd_last_close_xid = req->rq_xid;
272 lcd->lcd_last_close_result = rc;
274 /* VBR: save versions in last_rcvd for reconstruct. */
275 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
277 lcd->lcd_pre_versions[0] = pre_versions[0];
278 lcd->lcd_pre_versions[1] = pre_versions[1];
279 lcd->lcd_pre_versions[2] = pre_versions[2];
280 lcd->lcd_pre_versions[3] = pre_versions[3];
282 if (info->mti_transno != 0)
283 lcd->lcd_last_transno = info->mti_transno;
285 lcd->lcd_last_xid = req->rq_xid;
286 lcd->lcd_last_result = rc;
287 lcd->lcd_last_data = info->mti_opdata;
289 mutex_unlock(&ted->ted_lcd_lock);
294 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
296 LASSERT(mfd != NULL);
298 CDEBUG(D_DENTRY, DFID " Change mfd mode %#llo -> %#llo.\n",
299 PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
301 mfd->mfd_mode = mode;
305 * prep ma_lmm/ma_lmv for md_attr from reply
307 static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
308 struct mdt_object *obj,
311 LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
312 if (S_ISDIR(obj->mot_header.loh_attr)) {
313 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
315 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
318 if (ma->ma_lmv_size > 0)
319 ma->ma_need |= MA_LMV;
321 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
323 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
326 if (ma->ma_lmm_size > 0)
327 ma->ma_need |= MA_LOV;
331 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
332 struct mdt_object *o, __u64 flags, int created,
333 struct ldlm_reply *rep)
335 struct ptlrpc_request *req = mdt_info_req(info);
336 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
337 struct mdt_file_data *mfd;
338 struct md_attr *ma = &info->mti_attr;
339 struct lu_attr *la = &ma->ma_attr;
340 struct mdt_body *repbody;
341 int rc = 0, isdir, isreg;
344 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
346 isreg = S_ISREG(la->la_mode);
347 isdir = S_ISDIR(la->la_mode);
348 if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
350 * No EA, check whether it is will set regEA and dirEA since in
351 * above attr get, these size might be zero, so reset it, to
352 * retrieve the MD after create obj.
354 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
357 /* in replay case, p == NULL */
358 rc = mdt_create_data(info, p, o);
362 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
363 mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
366 CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n",
367 ma->ma_valid, ma->ma_lmm_size);
369 if (ma->ma_valid & MA_LOV) {
370 LASSERT(ma->ma_lmm_size != 0);
371 repbody->mbo_eadatasize = ma->ma_lmm_size;
373 repbody->mbo_valid |= OBD_MD_FLDIREA;
375 repbody->mbo_valid |= OBD_MD_FLEASIZE;
378 if (ma->ma_valid & MA_LMV) {
379 LASSERT(ma->ma_lmv_size != 0);
380 repbody->mbo_eadatasize = ma->ma_lmv_size;
382 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
385 if (flags & MDS_FMODE_WRITE)
386 rc = mdt_write_get(o);
387 else if (flags & MDS_FMODE_EXEC)
388 rc = mdt_write_deny(o);
393 rc = mo_open(info->mti_env, mdt_object_child(o),
394 created ? flags | MDS_OPEN_CREATED : flags);
396 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
397 * client does not know which suppgid should be sent to the MDS,
398 * or some other(s) changed the target file's GID after this RPC
399 * sent to the MDS with the suppgid as the original GID, then we
400 * give the client another chance to send the right suppgid. */
402 allow_client_chgrp(info, lu_ucred(info->mti_env)))
403 mdt_set_disposition(info, rep, DISP_OPEN_DENY);
408 mfd = mdt_mfd_new(med);
410 GOTO(err_out, rc = -ENOMEM);
413 * Keep a reference on this object for this open, and is
414 * released by mdt_mfd_close().
416 mdt_object_get(info->mti_env, o);
418 mfd->mfd_xid = req->rq_xid;
421 * @flags is always not zero. At least it should be FMODE_READ,
422 * FMODE_WRITE or MDS_FMODE_EXEC.
427 mdt_mfd_set_mode(mfd, flags);
429 atomic_inc(&o->mot_open_count);
430 if (flags & MDS_OPEN_LEASE)
431 atomic_inc(&o->mot_lease_count);
434 if (req_is_replay(req)) {
435 struct mdt_file_data *old_mfd;
436 /* Check wheather old cookie already exist in
437 * the list, becasue when do recovery, client
438 * might be disconnected from server, and
439 * restart replay, so there maybe some orphan
440 * mfd here, we should remove them */
441 LASSERT(info->mti_rr.rr_open_handle != NULL);
442 spin_lock(&med->med_open_lock);
443 old_mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
445 if (old_mfd != NULL) {
446 CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
447 "cookie = %#llx\n", mfd,
448 PFID(mdt_object_fid(mfd->mfd_object)),
449 info->mti_rr.rr_open_handle->cookie);
450 class_handle_unhash(&old_mfd->mfd_open_handle);
451 list_del_init(&old_mfd->mfd_list);
452 spin_unlock(&med->med_open_lock);
453 /* no attr update for that close */
455 ma->ma_valid |= MA_FLAGS;
456 ma->ma_attr_flags |= MDS_RECOV_OPEN;
457 mdt_mfd_close(info, old_mfd);
458 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
459 ma->ma_valid &= ~MA_FLAGS;
461 spin_unlock(&med->med_open_lock);
462 CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
464 PFID(mdt_object_fid(mfd->mfd_object)),
465 info->mti_rr.rr_open_handle->cookie);
468 CDEBUG(D_HA, "Store old cookie %#llx in new mfd\n",
469 info->mti_rr.rr_open_handle->cookie);
471 mfd->mfd_open_handle_old = *info->mti_rr.rr_open_handle;
474 repbody->mbo_open_handle.cookie = mfd->mfd_open_handle.h_cookie;
476 if (req->rq_export->exp_disconnected) {
477 spin_lock(&med->med_open_lock);
478 class_handle_unhash(&mfd->mfd_open_handle);
479 list_del_init(&mfd->mfd_list);
480 spin_unlock(&med->med_open_lock);
481 mdt_mfd_close(info, mfd);
483 spin_lock(&med->med_open_lock);
484 list_add_tail(&mfd->mfd_list, &med->med_open_head);
485 spin_unlock(&med->med_open_lock);
488 mdt_empty_transno(info, rc);
493 if (flags & MDS_FMODE_WRITE)
495 else if (flags & MDS_FMODE_EXEC)
501 static int mdt_finish_open(struct mdt_thread_info *info,
502 struct mdt_object *p, struct mdt_object *o,
503 __u64 flags, int created, struct ldlm_reply *rep)
505 struct ptlrpc_request *req = mdt_info_req(info);
506 struct obd_export *exp = req->rq_export;
507 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
508 struct md_attr *ma = &info->mti_attr;
509 struct lu_attr *la = &ma->ma_attr;
510 struct mdt_file_data *mfd;
511 struct mdt_body *repbody;
513 int isreg, isdir, islnk;
517 LASSERT(ma->ma_valid & MA_INODE);
519 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
521 isreg = S_ISREG(la->la_mode);
522 isdir = S_ISDIR(la->la_mode);
523 islnk = S_ISLNK(la->la_mode);
524 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
526 /* compatibility check for 2.10 clients when it tries to open mirrored
527 * files. 2.10 clients don't verify overlapping components so they
528 * would read and write mirrored files just as if they were normal
529 * PFL files, which will cause the problem that sycned mirrors actually
530 * contain different data.
531 * Older clients are not a concern here because they don't even
532 * understand PFL layout. */
533 if (isreg && !exp_connect_flr(exp) && ma->ma_valid & MA_LOV &&
534 mdt_lmm_is_flr(ma->ma_lmm)) {
535 /* LU-10286: for simplicity clients who don't understand
536 * mirrored layout(with connect flag OBD_CONNECT2_FLR) won't
537 * be able to open mirrored files */
541 /* LU-2275, simulate broken behaviour (esp. prevalent in
542 * pre-2.4 servers where a very strange reply is sent on error
543 * that looks like it was actually almost successful and a
544 * failure at the same time.) */
545 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
546 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
550 if (flags & MDS_OPEN_LOCK)
551 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
556 #ifdef CONFIG_FS_POSIX_ACL
557 if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
558 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
560 RETURN(PTR_ERR(nodemap));
562 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
563 nodemap_putref(nodemap);
570 * If we are following a symlink, don't open; and do not return open
571 * handle for special nodes as client required.
573 if (islnk || (!isreg && !isdir &&
574 (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
575 lustre_msg_set_transno(req->rq_repmsg, 0);
580 * We need to return the existing object's fid back, so it is done here,
581 * after preparing the reply.
583 if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
586 /* This can't be done earlier, we need to return reply body */
588 if (flags & (MDS_OPEN_CREAT | MDS_FMODE_WRITE)) {
589 /* We are trying to create or write an existing dir. */
592 } else if (flags & MDS_OPEN_DIRECTORY)
595 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
596 OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
601 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
602 spin_lock(&med->med_open_lock);
603 list_for_each(t, &med->med_open_head) {
604 mfd = list_entry(t, struct mdt_file_data, mfd_list);
605 if (mfd->mfd_xid == req->rq_xid)
609 spin_unlock(&med->med_open_lock);
612 repbody->mbo_open_handle.cookie =
613 mfd->mfd_open_handle.h_cookie;
614 /* set repbody->ea_size for resent case */
615 if (ma->ma_valid & MA_LOV) {
616 LASSERT(ma->ma_lmm_size != 0);
617 repbody->mbo_eadatasize = ma->ma_lmm_size;
619 repbody->mbo_valid |= OBD_MD_FLDIREA;
621 repbody->mbo_valid |= OBD_MD_FLEASIZE;
623 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
628 rc = mdt_mfd_open(info, p, o, flags, created, rep);
630 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
635 void mdt_reconstruct_open(struct mdt_thread_info *info,
636 struct mdt_lock_handle *lhc)
638 const struct lu_env *env = info->mti_env;
639 struct mdt_device *mdt = info->mti_mdt;
640 struct req_capsule *pill = info->mti_pill;
641 struct ptlrpc_request *req = mdt_info_req(info);
642 struct md_attr *ma = &info->mti_attr;
643 struct mdt_reint_record *rr = &info->mti_rr;
644 __u64 flags = info->mti_spec.sp_cr_flags;
645 struct ldlm_reply *ldlm_rep;
646 struct mdt_object *parent;
647 struct mdt_object *child;
648 struct mdt_body *repbody;
653 LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
654 ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
655 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
657 ma->ma_need = MA_INODE | MA_HSM;
660 opdata = mdt_req_from_lrd(req, info->mti_reply_data);
661 mdt_set_disposition(info, ldlm_rep, opdata);
663 CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n",
664 ldlm_rep->lock_policy_res1, req->rq_status);
666 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
668 /* We did not create successfully, return error to client. */
669 GOTO(out, rc = req->rq_status);
671 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
672 struct obd_export *exp = req->rq_export;
674 * We failed after creation, but we do not know in which step
675 * we failed. So try to check the child object.
677 parent = mdt_object_find(env, mdt, rr->rr_fid1);
678 if (IS_ERR(parent)) {
679 rc = PTR_ERR(parent);
680 LCONSOLE_WARN("Parent "DFID" lookup error %d."
681 " Evicting client %s with export %s.\n",
682 PFID(rr->rr_fid1), rc,
683 obd_uuid2str(&exp->exp_client_uuid),
684 obd_export_nid2str(exp));
685 mdt_export_evict(exp);
689 child = mdt_object_find(env, mdt, rr->rr_fid2);
692 LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
693 "evicting client %s with export %s\n",
694 PFID(rr->rr_fid2), rc,
695 obd_uuid2str(&exp->exp_client_uuid),
696 obd_export_nid2str(exp));
697 mdt_object_put(env, parent);
698 mdt_export_evict(exp);
702 if (unlikely(mdt_object_remote(child))) {
703 /* the child object was created on remote server */
704 if (!mdt_is_dne_client(exp)) {
705 /* Return -EIO for old client */
706 mdt_object_put(env, parent);
707 mdt_object_put(env, child);
708 GOTO(out, rc = -EIO);
710 repbody->mbo_fid1 = *rr->rr_fid2;
711 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
714 if (mdt_object_exists(child)) {
715 mdt_prep_ma_buf_from_rep(info, child, ma);
716 rc = mdt_attr_get_complex(info, child, ma);
718 rc = mdt_finish_open(info, parent,
722 /* the child does not exist, we should do
724 mdt_object_put(env, parent);
725 mdt_object_put(env, child);
726 GOTO(regular_open, 0);
729 mdt_object_put(env, parent);
730 mdt_object_put(env, child);
734 /* We did not try to create, so we are a pure open */
735 rc = mdt_reint_open(info, lhc);
741 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
742 LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
745 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
747 __u64 flags = info->mti_spec.sp_cr_flags;
748 struct mdt_reint_record *rr = &info->mti_rr;
749 struct md_attr *ma = &info->mti_attr;
750 struct mdt_object *o;
754 o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
756 RETURN(rc = PTR_ERR(o));
758 if (unlikely(mdt_object_remote(o))) {
759 /* the child object was created on remote server */
760 struct mdt_body *repbody;
762 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
765 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
766 repbody->mbo_fid1 = *rr->rr_fid2;
767 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
770 if (mdt_object_exists(o)) {
771 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
774 mdt_prep_ma_buf_from_rep(info, o, ma);
775 rc = mdt_attr_get_complex(info, o, ma);
777 rc = mdt_finish_open(info, NULL, o, flags, 0,
784 mdt_object_put(info->mti_env, o);
788 /* lock object for open */
789 static int mdt_object_open_lock(struct mdt_thread_info *info,
790 struct mdt_object *obj,
791 struct mdt_lock_handle *lhc,
794 struct md_attr *ma = &info->mti_attr;
795 __u64 open_flags = info->mti_spec.sp_cr_flags;
797 enum ldlm_mode lm = LCK_CR;
798 bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
799 bool try_layout = false;
800 bool create_layout = false;
802 int dom_stripes = LMM_NO_DOM;
803 bool dom_lock = false;
808 mdt_lock_handle_init(lhc);
810 if (req_is_replay(mdt_info_req(info)))
813 if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
814 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
815 md_should_create(open_flags))
816 create_layout = true;
817 if (exp_connect_layout(info->mti_exp) && !create_layout &&
818 ma->ma_need & MA_LOV)
821 /* DoM files can have just MDT stripe or combined MDT + OST
823 * - In the first case the open for read/write will do IO to
824 * the MDT stripe and it makes sense to take IO lock in
825 * advance along with OPEN even if it is blocking lock.
826 * - In the second case it is just size of MDT stripe and it
827 * is quite unlikely that client will write into it, though
828 * it may read it. So IO lock will be taken optionally if it
829 * is non-blocking one.
831 if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
832 dom_stripes = mdt_lmm_dom_entry(ma->ma_lmm);
834 if (dom_stripes == LMM_DOM_ONLY &&
835 info->mti_mdt->mdt_opts.mo_dom_lock > 0 &&
836 !mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
841 /* lease open, acquire write mode of open sem */
842 down_write(&obj->mot_open_sem);
844 /* Lease exists and ask for new lease */
845 if (atomic_read(&obj->mot_lease_count) > 0) {
846 /* only exclusive open is supported, so lease
847 * are conflicted to each other */
848 GOTO(out, rc = -EBUSY);
851 /* Lease must be with open lock */
852 if (!(open_flags & MDS_OPEN_LOCK)) {
853 CERROR("%s: Request lease for file:"DFID ", but open lock "
854 "is missed, open_flags = %#llo : rc = %d\n",
855 mdt_obd_name(info->mti_mdt),
856 PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
857 GOTO(out, rc = -EPROTO);
860 /* XXX: only exclusive open is supported. */
862 *ibits = MDS_INODELOCK_OPEN;
864 /* never grant LCK_EX layout lock to client */
866 } else { /* normal open */
867 /* normal open holds read mode of open sem */
868 down_read(&obj->mot_open_sem);
870 if (open_flags & MDS_OPEN_LOCK) {
871 if (open_flags & MDS_FMODE_WRITE)
873 else if (open_flags & MDS_FMODE_EXEC)
878 *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
879 } else if (atomic_read(&obj->mot_lease_count) > 0) {
880 if (open_flags & MDS_FMODE_WRITE)
886 *ibits = MDS_INODELOCK_OPEN;
889 lhc = &info->mti_lh[MDT_LH_LOCAL];
890 } else if (dom_lock) {
891 lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR;
892 if (info->mti_mdt->mdt_opts.mo_dom_lock ==
893 TRYLOCK_DOM_ON_OPEN) {
894 trybits |= MDS_INODELOCK_DOM |
895 MDS_INODELOCK_LAYOUT;
897 /* mo_dom_lock == ALWAYS_DOM_LOCK_ON_OPEN */
898 *ibits = MDS_INODELOCK_DOM;
899 if (info->mti_mdt->mdt_opts.mo_dom_read_open) {
900 trybits |= MDS_INODELOCK_LAYOUT;
905 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
906 PFID(mdt_object_fid(obj)),
907 atomic_read(&obj->mot_open_count), lm);
910 mdt_lock_reg_init(lhc, lm);
912 /* one problem to return layout lock on open is that it may result
913 * in too many layout locks cached on the client side. */
914 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
915 trybits |= MDS_INODELOCK_LAYOUT;
918 if (*ibits | trybits)
919 rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, false);
921 CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
922 ", open_flags = %#llo, try_layout = %d : rc = %d\n",
923 mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
924 *ibits, trybits, open_flags, try_layout, rc);
926 /* will change layout, revoke layout locks by enqueuing EX lock. */
927 if (rc == 0 && create_layout) {
928 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
930 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
931 ", open_flags = %#llo\n",
932 PFID(mdt_object_fid(obj)), open_flags);
934 /* We cannot enqueue another lock for the same resource we
935 * already have a lock for, due to mechanics of waiting list
936 * iterating in ldlm, see LU-3601.
937 * As such we'll drop the open lock we just got above here,
938 * it's ok not to have this open lock as it's main purpose is to
939 * flush unused cached client open handles. */
940 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
941 mdt_object_unlock(info, obj, lhc, 1);
943 LASSERT(!try_layout);
944 mdt_lock_handle_init(ll);
945 mdt_lock_reg_init(ll, LCK_EX);
946 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
948 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
951 /* Check if there is any other open handles after acquiring
952 * open lock. At this point, caching open handles have been revoked
954 * XXX: Now only exclusive open is supported. Need to check the
955 * type of open for generic lease support. */
956 if (rc == 0 && acq_lease) {
957 struct ptlrpc_request *req = mdt_info_req(info);
958 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
959 struct mdt_file_data *mfd;
960 bool is_replay_or_resent;
963 /* For lease: application can open a file and then apply lease,
964 * @handle contains original open handle in that case.
965 * In recovery, open REQ will be replayed and the lease REQ may
966 * be resent that means the open handle is already stale, so we
967 * need to fix it up here by finding new handle. */
968 is_replay_or_resent = req_is_replay(req) ||
969 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
971 /* if the request is _not_ a replay request, rr_open_handle
972 * may be used to hold an open file handle which is issuing the
973 * lease request, so that this openhandle doesn't count. */
974 mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
975 is_replay_or_resent);
979 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
980 PFID(mdt_object_fid(obj)),
981 atomic_read(&obj->mot_open_count), open_count);
983 if (atomic_read(&obj->mot_open_count) > open_count)
984 GOTO(out, rc = -EBUSY);
992 static void mdt_object_open_unlock(struct mdt_thread_info *info,
993 struct mdt_object *obj,
994 struct mdt_lock_handle *lhc,
997 __u64 open_flags = info->mti_spec.sp_cr_flags;
998 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1001 if (req_is_replay(mdt_info_req(info)))
1004 /* Release local lock - the lock put in MDT_LH_LOCAL will never
1005 * return to client side. */
1006 if (lustre_handle_is_used(&ll->mlh_reg_lh))
1007 mdt_object_unlock(info, obj, ll, 1);
1009 ll = &info->mti_lh[MDT_LH_LAYOUT];
1010 /* Release local layout lock, layout was created */
1011 if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1012 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1013 mdt_object_unlock(info, obj, ll, 1);
1016 if (open_flags & MDS_OPEN_LEASE)
1017 up_write(&obj->mot_open_sem);
1019 up_read(&obj->mot_open_sem);
1021 /* Cross-ref case, the lock should be returned to the client */
1022 if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1025 if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1026 !(ibits & MDS_INODELOCK_DOM)) {
1027 /* for the open request, the lock will only return to client
1028 * if open or layout lock is granted. */
1032 if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1033 struct ldlm_reply *ldlm_rep;
1035 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1036 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1037 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1038 mdt_object_unlock(info, obj, lhc, 1);
1044 * Check release is permitted for the current HSM flags.
1046 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1048 if (!(ma->ma_valid & MA_HSM))
1051 if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1054 if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1060 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1061 struct ldlm_reply *rep,
1062 struct mdt_lock_handle *lhc)
1064 const struct lu_env *env = info->mti_env;
1065 struct mdt_device *mdt = info->mti_mdt;
1066 __u64 flags = info->mti_spec.sp_cr_flags;
1067 struct mdt_reint_record *rr = &info->mti_rr;
1068 struct md_attr *ma = &info->mti_attr;
1069 struct mdt_object *parent= NULL;
1070 struct mdt_object *o;
1072 bool object_locked = false;
1076 if (md_should_create(flags)) {
1077 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1078 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1079 if (IS_ERR(parent)) {
1080 CDEBUG(D_INODE, "Fail to find parent "DFID
1081 " for anonymous created %ld, try to"
1082 " use server-side parent.\n",
1083 PFID(rr->rr_fid1), PTR_ERR(parent));
1088 ma->ma_need |= MA_PFID;
1091 o = mdt_object_find(env, mdt, rr->rr_fid2);
1093 GOTO(out_parent_put, rc = PTR_ERR(o));
1095 if (mdt_object_remote(o)) {
1096 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1097 mdt_obd_name(info->mti_mdt),
1099 GOTO(out, rc = -EREMOTE);
1100 } else if (!mdt_object_exists(o)) {
1101 mdt_set_disposition(info, rep,
1105 GOTO(out, rc = -ENOENT);
1108 mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1110 mdt_prep_ma_buf_from_rep(info, o, ma);
1111 if (flags & MDS_OPEN_RELEASE)
1112 ma->ma_need |= MA_HSM;
1113 rc = mdt_attr_get_complex(info, o, ma);
1117 /* We should not change file's existing LOV EA */
1118 if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1119 flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1120 GOTO(out, rc = -EEXIST);
1122 /* If a release request, check file flags are fine and ask for an
1123 * exclusive open access. */
1124 if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1125 GOTO(out, rc = -EPERM);
1127 rc = mdt_check_resent_lock(info, o, lhc);
1130 } else if (rc > 0) {
1131 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1132 object_locked = true;
1134 GOTO(out_unlock, rc);
1137 if (ma->ma_valid & MA_PFID) {
1138 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1139 if (IS_ERR(parent)) {
1140 CDEBUG(D_INODE, "Fail to find parent "DFID
1141 " for anonymous created %ld, try to"
1142 " use system default.\n",
1143 PFID(&ma->ma_pfid), PTR_ERR(parent));
1148 rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1150 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1151 if (flags & MDS_OPEN_LOCK)
1152 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1153 if (flags & MDS_OPEN_LEASE)
1154 mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1156 GOTO(out_unlock, rc);
1160 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1162 mdt_object_put(env, o);
1164 mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
1167 mdt_object_put(env, parent);
1171 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1172 static int mdt_cross_open(struct mdt_thread_info *info,
1173 const struct lu_fid *parent_fid,
1174 const struct lu_fid *fid,
1175 struct ldlm_reply *rep, __u32 flags)
1177 struct md_attr *ma = &info->mti_attr;
1178 struct mdt_object *o;
1182 o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1184 RETURN(rc = PTR_ERR(o));
1186 if (mdt_object_remote(o)) {
1187 /* Something is wrong here, the object is on another MDS! */
1188 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1189 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1190 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1192 "Object isn't on this server! FLD error?");
1195 if (mdt_object_exists(o)) {
1196 /* Do permission check for cross-open. */
1197 rc = mo_permission(info->mti_env, NULL,
1198 mdt_object_child(o),
1199 NULL, flags | MDS_OPEN_CROSS);
1203 mdt_prep_ma_buf_from_rep(info, o, ma);
1204 rc = mdt_attr_get_complex(info, o, ma);
1208 rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1211 * Something is wrong here. lookup was positive but
1212 * there is no object!
1214 CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1215 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1220 mdt_object_put(info->mti_env, o);
1225 * find root object and take its xattr lock if it's on remote MDT, later create
1226 * may use fs default striping (which is stored in root xattr).
1228 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1229 struct mdt_device *mdt)
1231 struct mdt_object *md_root = mdt->mdt_md_root;
1232 struct lustre_handle lhroot;
1235 if (md_root == NULL) {
1236 lu_root_fid(&info->mti_tmp_fid1);
1237 md_root = mdt_object_find(info->mti_env, mdt,
1238 &info->mti_tmp_fid1);
1239 if (IS_ERR(md_root))
1240 return PTR_ERR(md_root);
1242 spin_lock(&mdt->mdt_lock);
1243 if (mdt->mdt_md_root != NULL) {
1244 spin_unlock(&mdt->mdt_lock);
1246 LASSERTF(mdt->mdt_md_root == md_root,
1247 "Different root object ("
1248 DFID") instances, %p, %p\n",
1249 PFID(&info->mti_tmp_fid1),
1250 mdt->mdt_md_root, md_root);
1251 LASSERT(atomic_read(
1252 &md_root->mot_obj.lo_header->loh_ref) > 1);
1254 mdt_object_put(info->mti_env, md_root);
1256 mdt->mdt_md_root = md_root;
1257 spin_unlock(&mdt->mdt_lock);
1261 if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1264 rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root),
1265 &lhroot, LCK_PR, MDS_INODELOCK_XATTR,
1270 md_root->mot_cache_attr = 1;
1272 /* don't cancel this lock, so that we know the cached xattr is valid. */
1273 ldlm_lock_decref(&lhroot, LCK_PR);
1278 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1280 struct mdt_device *mdt = info->mti_mdt;
1281 struct ptlrpc_request *req = mdt_info_req(info);
1282 struct mdt_object *parent;
1283 struct mdt_object *child;
1284 struct mdt_lock_handle *lh;
1285 struct ldlm_reply *ldlm_rep;
1286 struct mdt_body *repbody;
1287 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1288 struct md_attr *ma = &info->mti_attr;
1289 __u64 create_flags = info->mti_spec.sp_cr_flags;
1291 struct mdt_reint_record *rr = &info->mti_rr;
1294 int object_locked = 0;
1298 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1299 (obd_timeout + 1) / 4);
1301 mdt_counter_incr(req, LPROC_MDT_OPEN);
1302 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1304 ma->ma_need = MA_INODE;
1307 LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1308 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1310 if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1311 CERROR("file join is not supported anymore.\n");
1312 GOTO(out, result = err_serious(-EOPNOTSUPP));
1314 msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1316 if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1317 info->mti_spec.u.sp_ea.eadata == NULL)
1318 GOTO(out, result = err_serious(-EINVAL));
1320 if (create_flags & MDS_FMODE_WRITE &&
1321 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1322 GOTO(out, result = -EROFS);
1324 CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1325 "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1326 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1327 PFID(rr->rr_fid2), create_flags,
1328 ma->ma_attr.la_mode, msg_flags);
1330 if (info->mti_cross_ref) {
1331 /* This is cross-ref open */
1332 mdt_set_disposition(info, ldlm_rep,
1333 (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1335 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1336 ldlm_rep, create_flags);
1338 } else if (req_is_replay(req)) {
1339 result = mdt_open_by_fid(info, ldlm_rep);
1341 if (result != -ENOENT)
1344 /* We didn't find the correct object, so we need to re-create it
1345 * via a regular replay. */
1346 if (!(create_flags & MDS_OPEN_CREAT)) {
1347 DEBUG_REQ(D_ERROR, req,
1348 "OPEN & CREAT not in open replay/by_fid.");
1349 GOTO(out, result = -EFAULT);
1351 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1352 } else if (create_flags & (MDS_OPEN_BY_FID | MDS_OPEN_LOCK)) {
1354 * MDS_OPEN_LOCK is checked for backward compatibility with 2.1
1357 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1359 CDEBUG(D_INFO, "no object for "DFID": %d\n",
1360 PFID(rr->rr_fid2), result);
1364 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1365 GOTO(out, result = err_serious(-ENOMEM));
1367 mdt_set_disposition(info, ldlm_rep,
1368 (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1370 if (!lu_name_is_valid(&rr->rr_name))
1371 GOTO(out, result = -EPROTO);
1373 result = mdt_lock_root_xattr(info, mdt);
1378 lh = &info->mti_lh[MDT_LH_PARENT];
1379 mdt_lock_pdo_init(lh,
1380 (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1383 parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1385 GOTO(out, result = PTR_ERR(parent));
1387 result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1389 mdt_object_put(info->mti_env, parent);
1393 /* get and check version of parent */
1394 result = mdt_version_get_check(info, parent, 0);
1396 GOTO(out_parent, result);
1398 fid_zero(child_fid);
1400 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1401 &rr->rr_name, child_fid, &info->mti_spec);
1403 LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1404 "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1405 PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1408 if (result != 0 && result != -ENOENT && result != -ESTALE)
1409 GOTO(out_parent, result);
1411 if (result == -ENOENT || result == -ESTALE) {
1412 /* If the object is dead, let's check if the object
1413 * is being migrated to a new object */
1414 if (result == -ESTALE) {
1415 struct lu_buf lmv_buf;
1417 lmv_buf.lb_buf = info->mti_xattr_buf;
1418 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
1419 rc = mo_xattr_get(info->mti_env,
1420 mdt_object_child(parent),
1421 &lmv_buf, XATTR_NAME_LMV);
1423 struct lmv_mds_md_v1 *lmv;
1425 lmv = lmv_buf.lb_buf;
1426 if (le32_to_cpu(lmv->lmv_hash_type) &
1427 LMV_HASH_FLAG_MIGRATION) {
1428 /* Get the new parent FID and retry */
1429 mdt_object_unlock_put(info, parent,
1431 mdt_lock_handle_init(lh);
1433 (struct lu_fid *)rr->rr_fid1,
1434 &lmv->lmv_stripe_fids[1]);
1440 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1441 if (result == -ESTALE) {
1443 * -ESTALE means the parent is a dead(unlinked) dir, so
1444 * it should return -ENOENT to in accordance with the
1445 * original mds implementaion.
1447 GOTO(out_parent, result = -ENOENT);
1450 if (!(create_flags & MDS_OPEN_CREAT))
1451 GOTO(out_parent, result);
1452 if (mdt_rdonly(req->rq_export))
1453 GOTO(out_parent, result = -EROFS);
1454 *child_fid = *info->mti_rr.rr_fid2;
1455 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1457 /* In the function below, .hs_keycmp resolves to
1458 * lu_obj_hop_keycmp() */
1459 /* coverity[overrun-buffer-val] */
1460 child = mdt_object_new(info->mti_env, mdt, child_fid);
1463 * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1464 * return FID back in that case.
1466 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1467 child = mdt_object_find(info->mti_env, mdt, child_fid);
1470 GOTO(out_parent, result = PTR_ERR(child));
1472 /** check version of child */
1473 rc = mdt_version_get_check(info, child, 1);
1475 GOTO(out_child, result = rc);
1477 if (result == -ENOENT) {
1478 /* Create under OBF and .lustre is not permitted */
1479 if (!fid_is_md_operative(rr->rr_fid1))
1480 GOTO(out_child, result = -EPERM);
1482 /* save versions in reply */
1483 mdt_version_get_save(info, parent, 0);
1484 mdt_version_get_save(info, child, 1);
1486 /* version of child will be changed */
1487 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1489 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1490 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1492 /* Let lower layers know what is lock mode on directory. */
1493 info->mti_spec.sp_cr_mode =
1494 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1496 /* Don't do lookup sanity check. We know name doesn't exist. */
1497 info->mti_spec.sp_cr_lookup = 0;
1498 info->mti_spec.sp_feat = &dt_directory_features;
1500 result = mdo_create(info->mti_env, mdt_object_child(parent),
1501 &rr->rr_name, mdt_object_child(child),
1502 &info->mti_spec, &info->mti_attr);
1503 if (result == -ERESTART) {
1504 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1505 GOTO(out_child, result);
1507 mdt_prep_ma_buf_from_rep(info, child, ma);
1508 /* XXX: we should call this once, see few lines below */
1510 result = mdt_attr_get_complex(info, child, ma);
1513 GOTO(out_child, result);
1516 mdt_counter_incr(req, LPROC_MDT_MKNOD);
1519 * The object is on remote node, return its FID for remote open.
1521 if (mdt_object_remote(child)) {
1523 * Check if this lock already was sent to client and
1524 * this is resent case. For resent case do not take lock
1525 * again, use what is already granted.
1527 LASSERT(lhc != NULL);
1529 rc = mdt_check_resent_lock(info, child, lhc);
1531 GOTO(out_child, result = rc);
1532 } else if (rc > 0) {
1533 mdt_lock_handle_init(lhc);
1534 mdt_lock_reg_init(lhc, LCK_PR);
1536 rc = mdt_object_lock(info, child, lhc,
1537 MDS_INODELOCK_LOOKUP);
1539 repbody->mbo_fid1 = *mdt_object_fid(child);
1540 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1544 result = -MDT_EREMOTE_OPEN;
1545 GOTO(out_child, result);
1546 } else if (mdt_object_exists(child)) {
1547 /* We have to get attr & LOV EA & HSM for this
1549 mdt_prep_ma_buf_from_rep(info, child, ma);
1550 ma->ma_need |= MA_HSM;
1551 result = mdt_attr_get_complex(info, child, ma);
1553 GOTO(out_child, result);
1555 /* Object does not exist. Likely FS corruption. */
1556 CERROR("%s: name '"DNAME"' present, but FID "
1557 DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1558 PNAME(&rr->rr_name), PFID(child_fid));
1559 GOTO(out_child, result = -EIO);
1563 rc = mdt_check_resent_lock(info, child, lhc);
1565 GOTO(out_child, result = rc);
1566 } else if (rc == 0) {
1567 /* the open lock might already be gotten in
1568 * ldlm_handle_enqueue() */
1569 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1570 if (create_flags & MDS_OPEN_LOCK)
1571 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1573 /* get openlock if this isn't replay and client requested it */
1574 if (!req_is_replay(req)) {
1575 rc = mdt_object_open_lock(info, child, lhc, &ibits);
1578 GOTO(out_child_unlock, result = rc);
1579 else if (create_flags & MDS_OPEN_LOCK)
1580 mdt_set_disposition(info, ldlm_rep,
1584 /* Try to open it now. */
1585 rc = mdt_finish_open(info, parent, child, create_flags,
1589 /* openlock will be released if mdt_finish_open failed */
1590 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1592 if (created && create_flags & MDS_OPEN_VOLATILE) {
1593 CERROR("%s: cannot open volatile file "DFID", orphan "
1594 "file will be left in PENDING directory until "
1595 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1596 PFID(mdt_object_fid(child)), rc);
1597 GOTO(out_child_unlock, result);
1603 rc = mdo_unlink(info->mti_env,
1604 mdt_object_child(parent),
1605 mdt_object_child(child),
1607 &info->mti_attr, 0);
1609 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1610 mdt_obd_name(info->mti_mdt),
1611 PFID(mdt_object_fid(child)), rc);
1612 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1618 mdt_object_open_unlock(info, child, lhc, ibits, result);
1620 mdt_object_put(info->mti_env, child);
1622 mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh);
1624 mdt_object_unlock_put(info, parent, lh, result || !created);
1627 lustre_msg_set_transno(req->rq_repmsg, 0);
1632 * Create an orphan object use local root.
1634 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1635 struct mdt_device *mdt,
1636 const struct lu_fid *fid,
1637 struct md_attr *attr, fmode_t fmode)
1639 const struct lu_env *env = info->mti_env;
1640 struct md_op_spec *spec = &info->mti_spec;
1641 struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1642 struct mdt_object *obj = NULL;
1643 struct mdt_object *local_root;
1644 static const struct lu_name lname = {
1645 .ln_name = "i_am_nobody",
1646 .ln_namelen = sizeof("i_am_nobody") - 1,
1648 struct lu_ucred *uc;
1649 cfs_cap_t uc_cap_save;
1653 rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1655 RETURN(ERR_PTR(rc));
1657 local_root = mdt_object_find(env, mdt, local_root_fid);
1658 if (IS_ERR(local_root))
1661 obj = mdt_object_new(env, mdt, fid);
1663 GOTO(out, rc = PTR_ERR(obj));
1665 spec->sp_cr_lookup = 0;
1666 spec->sp_feat = &dt_directory_features;
1667 spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1668 spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1669 if (attr->ma_valid & MA_LOV) {
1670 spec->u.sp_ea.eadata = attr->ma_lmm;
1671 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1672 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1674 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1678 uc_cap_save = uc->uc_cap;
1679 uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1680 rc = mdo_create(env, mdt_object_child(local_root), &lname,
1681 mdt_object_child(obj), spec, attr);
1682 uc->uc_cap = uc_cap_save;
1684 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1685 mdt_obd_name(mdt), PFID(fid), rc);
1689 rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1691 CERROR("%s: cannot open volatile file "DFID", orphan "
1692 "file will be left in PENDING directory until "
1693 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1700 mdt_object_put(env, obj);
1703 mdt_object_put(env, local_root);
1707 /* XXX Look into layout in MDT layer. */
1708 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1710 struct lov_comp_md_v1 *comp_v1;
1711 struct lov_mds_md *v1;
1715 if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1716 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1718 if (comp_v1->lcm_entry_count == 0)
1721 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1722 off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1723 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1724 v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1727 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1732 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1735 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1736 struct lu_ucred *uc = mdt_ucred(info);
1737 struct close_data *data;
1738 struct ldlm_lock *lease;
1739 struct mdt_object *orphan;
1740 struct md_attr *orp_ma;
1748 if (mdt_rdonly(info->mti_exp))
1751 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1755 lease = ldlm_handle2lock(&data->cd_handle);
1759 /* try to hold open_sem so that nobody else can open the file */
1760 if (!down_write_trylock(&o->mot_open_sem)) {
1761 ldlm_lock_cancel(lease);
1762 GOTO(out_reprocess, rc = -EBUSY);
1765 /* Check if the lease open lease has already canceled */
1766 lock_res_and_lock(lease);
1767 lease_broken = ldlm_is_cancel(lease);
1768 unlock_res_and_lock(lease);
1770 LDLM_DEBUG(lease, DFID " lease broken? %d",
1771 PFID(mdt_object_fid(o)), lease_broken);
1773 /* Cancel server side lease. Client side counterpart should
1774 * have been cancelled. It's okay to cancel it now as we've
1775 * held mot_open_sem. */
1776 ldlm_lock_cancel(lease);
1778 if (lease_broken) /* don't perform release task */
1779 GOTO(out_unlock, rc = -ESTALE);
1781 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1782 GOTO(out_unlock, rc = -EINVAL);
1784 /* ma_need was set before but it seems fine to change it in order to
1785 * avoid modifying the one from RPC */
1786 ma->ma_need = MA_HSM;
1787 rc = mdt_attr_get_complex(info, o, ma);
1789 GOTO(out_unlock, rc);
1791 if (!mdt_hsm_release_allow(ma))
1792 GOTO(out_unlock, rc = -EPERM);
1794 /* already released? */
1795 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1796 GOTO(out_unlock, rc = 0);
1798 /* Compare on-disk and packed data_version */
1799 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1800 CDEBUG(D_HSM, DFID" data_version mismatches: packed=%llu"
1801 " and on-disk=%llu\n", PFID(mdt_object_fid(o)),
1802 data->cd_data_version, ma->ma_hsm.mh_arch_ver);
1803 GOTO(out_unlock, rc = -EPERM);
1806 ma->ma_valid = MA_INODE;
1807 ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1808 rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1810 GOTO(out_unlock, rc);
1812 mutex_lock(&o->mot_som_mutex);
1813 rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size,
1814 ma->ma_attr.la_blocks);
1815 mutex_unlock(&o->mot_som_mutex);
1818 "%s: File "DFID" SOM update failed: rc = %d\n",
1819 mdt_obd_name(info->mti_mdt),
1820 PFID(mdt_object_fid(o)), rc2);
1823 ma->ma_need = MA_INODE | MA_LOV;
1824 rc = mdt_attr_get_complex(info, o, ma);
1826 GOTO(out_unlock, rc);
1828 if (!(ma->ma_valid & MA_LOV)) {
1829 /* Even empty file are released */
1830 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1831 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
1832 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1833 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1834 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1836 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
1837 * ma_lmm as lov_user_md, then it will be confused by union of
1838 * layout_gen and stripe_offset. */
1839 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
1841 ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1843 GOTO(out_unlock, rc = -EINVAL);
1846 /* Set file as released. */
1847 rc = mdt_hsm_set_released(ma->ma_lmm);
1849 GOTO(out_unlock, rc);
1851 orp_ma = &info->mti_u.hsm.attr;
1852 orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1853 /* We use root ownership to bypass potential quota
1854 * restrictions on the user and group of the file. */
1855 orp_ma->ma_attr.la_uid = 0;
1856 orp_ma->ma_attr.la_gid = 0;
1857 orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1858 orp_ma->ma_lmm = ma->ma_lmm;
1859 orp_ma->ma_lmm_size = ma->ma_lmm_size;
1860 orp_ma->ma_valid = MA_INODE | MA_LOV;
1861 orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1863 if (IS_ERR(orphan)) {
1864 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1865 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1867 GOTO(out_unlock, rc = PTR_ERR(orphan));
1870 /* Set up HSM attribute for orphan object */
1871 CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
1872 buf = &info->mti_buf;
1873 buf->lb_buf = info->mti_xattr_buf;
1874 buf->lb_len = sizeof(struct hsm_attrs);
1875 ma->ma_hsm.mh_flags |= HS_RELEASED;
1876 lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1877 ma->ma_hsm.mh_flags &= ~HS_RELEASED;
1879 mdt_lock_reg_init(lh, LCK_EX);
1880 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
1881 MDS_INODELOCK_XATTR);
1883 GOTO(out_close, rc);
1885 /* The orphan has root ownership so we need to raise
1886 * CAP_FOWNER to set the HSM attributes. */
1888 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
1889 rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
1893 GOTO(out_layout_lock, rc);
1895 /* Swap layout with orphan objects. */
1896 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
1897 mdt_object_child(orphan),
1898 SWAP_LAYOUTS_MDS_HSM);
1902 /* Release exclusive LL */
1903 mdt_object_unlock(info, o, lh, 1);
1905 /* Close orphan object anyway */
1906 rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
1909 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
1910 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
1911 LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
1913 mdt_object_put(info->mti_env, orphan);
1916 up_write(&o->mot_open_sem);
1918 /* already released */
1920 struct mdt_body *repbody;
1922 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1923 LASSERT(repbody != NULL);
1924 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1928 ldlm_reprocess_all(lease->l_resource);
1929 LDLM_LOCK_PUT(lease);
1937 int mdt_close_handle_layouts(struct mdt_thread_info *info,
1938 struct mdt_object *o, struct md_attr *ma)
1940 struct mdt_lock_handle *lh1 = &info->mti_lh[MDT_LH_NEW];
1941 struct mdt_lock_handle *lh2 = &info->mti_lh[MDT_LH_OLD];
1942 struct close_data *data;
1943 struct ldlm_lock *lease;
1944 struct mdt_object *o1 = o, *o2;
1950 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1953 if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
1956 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1960 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1963 rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
1964 if (unlikely(rc == 0))
1967 /* Exchange o1 and o2, to enforce locking order */
1968 swap_objects = (rc < 0);
1970 lease = ldlm_handle2lock(&data->cd_handle);
1974 o2 = mdt_object_find(info->mti_env, info->mti_mdt, &data->cd_fid);
1976 GOTO(out_lease, rc = PTR_ERR(o2));
1978 if (!S_ISREG(lu_object_attr(&o2->mot_obj))) {
1979 swap_objects = false; /* not swapped yet */
1980 GOTO(out_obj, rc = -EINVAL);
1986 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1991 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1996 /* try to hold open_sem so that nobody else can open the file */
1997 if (!down_write_trylock(&o->mot_open_sem)) {
1998 ldlm_lock_cancel(lease);
1999 GOTO(out_obj, rc = -EBUSY);
2002 /* Check if the lease open lease has already canceled */
2003 lock_res_and_lock(lease);
2004 lease_broken = ldlm_is_cancel(lease);
2005 unlock_res_and_lock(lease);
2007 LDLM_DEBUG(lease, DFID " lease broken? %d",
2008 PFID(mdt_object_fid(o)), lease_broken);
2010 /* Cancel server side lease. Client side counterpart should
2011 * have been cancelled. It's okay to cancel it now as we've
2012 * held mot_open_sem. */
2013 ldlm_lock_cancel(lease);
2016 GOTO(out_unlock_sem, rc = -ESTALE);
2018 mdt_lock_reg_init(lh1, LCK_EX);
2019 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
2020 MDS_INODELOCK_XATTR);
2022 GOTO(out_unlock_sem, rc);
2024 mdt_lock_reg_init(lh2, LCK_EX);
2025 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2026 MDS_INODELOCK_XATTR);
2028 GOTO(out_unlock1, rc);
2030 /* Swap layout with orphan object */
2031 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2032 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2033 mdt_object_child(o2), 0);
2034 } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2035 ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2036 struct lu_buf *buf = &info->mti_buf;
2037 struct md_rejig_data mrd;
2039 mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2040 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)
2041 mrd.mrd_mirror_id = data->cd_mirror_id;
2043 buf->lb_len = sizeof(mrd);
2045 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2047 ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2048 LU_XATTR_SPLIT : LU_XATTR_MERGE);
2049 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2052 mutex_lock(&o->mot_som_mutex);
2053 rc2 = mdt_set_som(info, o, SOM_FL_STRICT,
2054 ma->ma_attr.la_size,
2055 ma->ma_attr.la_blocks);
2056 mutex_unlock(&o->mot_som_mutex);
2058 CERROR(DFID": Setting i_blocks error: %d, "
2059 "i_blocks will be reported wrongly and "
2060 "can only be fixed in next resync\n",
2061 PFID(mdt_object_fid(o)), rc2);
2065 GOTO(out_unlock2, rc);
2070 /* Release exclusive LL */
2071 mdt_object_unlock(info, o2, lh2, 1);
2074 mdt_object_unlock(info, o1, lh1, 1);
2077 up_write(&o->mot_open_sem);
2079 /* already swapped */
2081 struct mdt_body *repbody;
2083 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2084 LASSERT(repbody != NULL);
2085 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2089 mdt_object_put(info->mti_env, swap_objects ? o1 : o2);
2091 ldlm_reprocess_all(lease->l_resource);
2094 LDLM_LOCK_PUT(lease);
2104 static int mdt_close_resync_done(struct mdt_thread_info *info,
2105 struct mdt_object *o, struct md_attr *ma)
2107 struct close_data *data;
2108 struct ldlm_lock *lease;
2109 struct md_layout_change layout = { 0 };
2110 __u32 *resync_ids = NULL;
2111 size_t resync_count = 0;
2116 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2119 if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2122 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2126 if (ptlrpc_req_need_swab(mdt_info_req(info)))
2127 lustre_swab_close_data_resync_done(&data->cd_resync);
2129 if (!fid_is_zero(&data->cd_fid))
2132 lease = ldlm_handle2lock(&data->cd_handle);
2136 /* try to hold open_sem so that nobody else can open the file */
2137 if (!down_write_trylock(&o->mot_open_sem)) {
2138 ldlm_lock_cancel(lease);
2139 GOTO(out_reprocess, rc = -EBUSY);
2142 /* Check if the lease open lease has already canceled */
2143 lock_res_and_lock(lease);
2144 lease_broken = ldlm_is_cancel(lease);
2145 unlock_res_and_lock(lease);
2147 LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2148 PFID(mdt_object_fid(o)), lease_broken);
2150 /* Cancel server side lease. Client side counterpart should
2151 * have been cancelled. It's okay to cancel it now as we've
2152 * held mot_open_sem. */
2153 ldlm_lock_cancel(lease);
2155 if (lease_broken) /* don't perform release task */
2156 GOTO(out_unlock, rc = -ESTALE);
2158 resync_count = data->cd_resync.resync_count;
2160 if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2163 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2165 GOTO(out_unlock, rc = -EPROTO);
2167 OBD_ALLOC(resync_ids, resync_count * sizeof(__u32));
2169 GOTO(out_unlock, rc = -ENOMEM);
2171 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2172 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2174 layout.mlc_resync_ids = resync_ids;
2176 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2179 layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2180 layout.mlc_resync_count = resync_count;
2181 if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2182 layout.mlc_som.lsa_valid = SOM_FL_STRICT;
2183 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2184 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2186 rc = mdt_layout_change(info, o, &layout);
2188 GOTO(out_unlock, rc);
2193 up_write(&o->mot_open_sem);
2195 /* already released */
2197 struct mdt_body *repbody;
2199 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2200 LASSERT(repbody != NULL);
2201 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2205 OBD_FREE(resync_ids, resync_count * sizeof(__u32));
2208 ldlm_reprocess_all(lease->l_resource);
2209 LDLM_LOCK_PUT(lease);
2217 #define MFD_CLOSED(mode) ((mode) == MDS_FMODE_CLOSED)
2218 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2220 return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
2223 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2225 struct mdt_object *o = mfd->mfd_object;
2226 struct md_object *next = mdt_object_child(o);
2227 struct md_attr *ma = &info->mti_attr;
2228 struct lu_fid *ofid = &info->mti_tmp_fid1;
2232 bool discard = false;
2236 mode = mfd->mfd_mode;
2237 intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2238 *ofid = *mdt_object_fid(o);
2240 CDEBUG(D_INODE, "%s: close file "DFID" with intent: %llx\n",
2241 mdt_obd_name(info->mti_mdt), PFID(ofid), intent);
2244 case MDS_HSM_RELEASE: {
2245 rc = mdt_hsm_release(info, o, ma);
2247 CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2248 mdt_obd_name(info->mti_mdt),
2250 /* continue to close even error occurred. */
2254 case MDS_CLOSE_LAYOUT_MERGE:
2255 case MDS_CLOSE_LAYOUT_SPLIT:
2256 case MDS_CLOSE_LAYOUT_SWAP: {
2257 rc = mdt_close_handle_layouts(info, o, ma);
2260 "%s: cannot swap layout of "DFID": rc = %d\n",
2261 mdt_obd_name(info->mti_mdt),
2263 /* continue to close even if error occurred. */
2267 case MDS_CLOSE_RESYNC_DONE:
2268 rc = mdt_close_resync_done(info, o, ma);
2275 if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
2276 ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) {
2279 rc2 = mdt_lsom_update(info, o, false);
2282 "%s: File " DFID " LSOM failed: rc = %d\n",
2283 mdt_obd_name(info->mti_mdt),
2285 /* continue to close even if error occured. */
2288 if (mode & MDS_FMODE_WRITE)
2290 else if (mode & MDS_FMODE_EXEC)
2293 /* Update atime on close only. */
2294 if ((mode & MDS_FMODE_EXEC || mode & MDS_FMODE_READ ||
2295 mode & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
2296 (ma->ma_attr.la_valid & LA_ATIME)) {
2297 /* Set the atime only. */
2298 ma->ma_valid = MA_INODE;
2299 ma->ma_attr.la_valid = LA_ATIME;
2300 rc = mo_attr_set(info->mti_env, next, ma);
2303 /* If file data is modified, add the dirty flag. */
2304 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2305 rc = mdt_add_dirty_flag(info, o, ma);
2307 ma->ma_need |= MA_INODE;
2308 ma->ma_valid &= ~MA_INODE;
2310 LASSERT(atomic_read(&o->mot_open_count) > 0);
2311 atomic_dec(&o->mot_open_count);
2312 mdt_handle_last_unlink(info, o, ma);
2314 if (!MFD_CLOSED(mode)) {
2315 rc = mo_close(info->mti_env, next, ma, mode);
2316 discard = mdt_dom_check_for_discard(info, o);
2319 /* adjust open and lease count */
2320 if (mode & MDS_OPEN_LEASE) {
2321 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2322 atomic_dec(&o->mot_lease_count);
2326 mdt_object_put(info->mti_env, o);
2329 mdt_dom_discard_data(info, ofid);
2334 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2335 struct mdt_body *repbody)
2337 struct mdt_export_data *med;
2338 struct mdt_file_data *mfd;
2343 med = &req->rq_export->exp_mdt_data;
2344 spin_lock(&med->med_open_lock);
2345 mfd = mdt_open_handle2mfd(med, &info->mti_open_handle,
2346 req_is_replay(req));
2347 if (mdt_mfd_closed(mfd)) {
2348 spin_unlock(&med->med_open_lock);
2349 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2350 ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2351 info->mti_open_handle.cookie);
2352 /** not serious error since bug 3633 */
2355 class_handle_unhash(&mfd->mfd_open_handle);
2356 list_del_init(&mfd->mfd_list);
2357 spin_unlock(&med->med_open_lock);
2358 ret = mdt_mfd_close(info, mfd);
2361 RETURN(rc ? rc : ret);
2364 int mdt_close(struct tgt_session_info *tsi)
2366 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2367 struct ptlrpc_request *req = tgt_ses_req(tsi);
2368 struct md_attr *ma = &info->mti_attr;
2369 struct mdt_body *repbody = NULL;
2373 mdt_counter_incr(req, LPROC_MDT_CLOSE);
2374 /* Close may come with the Size-on-MDS update. Unpack it. */
2375 rc = mdt_close_unpack(info);
2377 GOTO(out, rc = err_serious(rc));
2379 /* These fields are no longer used and are left for compatibility.
2380 * size is always zero */
2381 req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2383 req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2385 rc = req_capsule_server_pack(info->mti_pill);
2386 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2387 mdt_client_compatibility(info);
2389 mdt_fix_reply(info);
2390 mdt_exit_ucred(info);
2391 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2394 /* Continue to close handle even if we can not pack reply */
2396 repbody = req_capsule_server_get(info->mti_pill,
2398 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2400 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2403 ma->ma_need = MA_INODE | MA_LOV;
2404 repbody->mbo_eadatasize = 0;
2405 repbody->mbo_aclsize = 0;
2407 rc = err_serious(rc);
2410 rc = mdt_close_internal(info, req, repbody);
2412 mdt_empty_transno(info, rc);
2414 if (repbody != NULL) {
2415 mdt_client_compatibility(info);
2416 rc = mdt_fix_reply(info);
2419 mdt_exit_ucred(info);
2420 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2421 GOTO(out, rc = err_serious(-ENOMEM));
2423 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2424 OBD_FAIL_MDS_CLOSE_NET_REP))
2425 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2427 mdt_thread_info_fini(info);
2428 RETURN(rc ? rc : ret);