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 static const char mfd_open_handle_owner[] = "mdt";
49 /* Create a new mdt_file_data struct, initialize it,
50 * and insert it to global hash table */
51 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
53 struct mdt_file_data *mfd;
58 refcount_set(&mfd->mfd_open_handle.h_ref, 1);
59 INIT_HLIST_NODE(&mfd->mfd_open_handle.h_link);
61 INIT_LIST_HEAD(&mfd->mfd_list);
62 class_handle_hash(&mfd->mfd_open_handle, mfd_open_handle_owner);
69 * Find the mfd pointed to by handle in global hash table.
70 * In case of replay the handle is obsoleted
71 * but mfd can be found in mfd list by that handle.
72 * Callers need to be holding med_open_lock.
74 struct mdt_file_data *mdt_open_handle2mfd(struct mdt_export_data *med,
75 const struct lustre_handle *open_handle,
76 bool is_replay_or_resent)
78 struct mdt_file_data *mfd;
81 LASSERT(open_handle != NULL);
82 mfd = class_handle2object(open_handle->cookie, mfd_open_handle_owner);
84 refcount_dec(&mfd->mfd_open_handle.h_ref);
86 /* during dw/setattr replay the mfd can be found by old handle */
87 if ((!mfd || mfd->mfd_owner != med) && is_replay_or_resent) {
88 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
89 if (mfd->mfd_open_handle_old.cookie ==
100 void mdt_mfd_free(struct mdt_file_data *mfd)
102 LASSERT(refcount_read(&mfd->mfd_open_handle.h_ref) == 1);
103 LASSERT(list_empty(&mfd->mfd_list));
104 OBD_FREE_PRE(mfd, sizeof(*mfd), "rcu");
105 kfree_rcu(mfd, mfd_open_handle.h_rcu);
108 static int mdt_create_data(struct mdt_thread_info *info,
109 struct mdt_object *p, struct mdt_object *o)
111 struct md_op_spec *spec = &info->mti_spec;
112 struct md_attr *ma = &info->mti_attr;
116 if (!md_should_create(spec->sp_cr_flags))
119 ma->ma_need = MA_INODE | MA_LOV;
121 mutex_lock(&o->mot_lov_mutex);
122 if (!o->mot_lov_created) {
123 rc = mdo_create_data(info->mti_env,
124 p ? mdt_object_child(p) : NULL,
125 mdt_object_child(o), spec, ma);
127 rc = mdt_attr_get_complex(info, o, ma);
129 if (rc == 0 && ma->ma_valid & MA_LOV)
130 o->mot_lov_created = 1;
133 mutex_unlock(&o->mot_lov_mutex);
137 int mdt_write_read(struct mdt_object *o)
141 spin_lock(&o->mot_write_lock);
142 rc = o->mot_write_count;
143 spin_unlock(&o->mot_write_lock);
147 int mdt_write_get(struct mdt_object *o)
151 spin_lock(&o->mot_write_lock);
152 if (o->mot_write_count < 0)
155 o->mot_write_count++;
156 spin_unlock(&o->mot_write_lock);
161 void mdt_write_put(struct mdt_object *o)
164 spin_lock(&o->mot_write_lock);
165 o->mot_write_count--;
166 spin_unlock(&o->mot_write_lock);
170 static int mdt_write_deny(struct mdt_object *o)
174 spin_lock(&o->mot_write_lock);
175 if (o->mot_write_count > 0)
178 o->mot_write_count--;
179 spin_unlock(&o->mot_write_lock);
183 static void mdt_write_allow(struct mdt_object *o)
186 spin_lock(&o->mot_write_lock);
187 o->mot_write_count++;
188 spin_unlock(&o->mot_write_lock);
192 /* there can be no real transaction so prepare the fake one */
193 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
195 struct mdt_device *mdt = info->mti_mdt;
196 struct ptlrpc_request *req = mdt_info_req(info);
197 struct tg_export_data *ted;
198 struct lsd_client_data *lcd;
201 if (mdt_rdonly(req->rq_export))
204 /* transaction has occurred already */
205 if (lustre_msg_get_transno(req->rq_repmsg) != 0)
208 if (tgt_is_multimodrpcs_client(req->rq_export)) {
211 /* generate an empty transaction to get a transno
213 th = dt_trans_create(info->mti_env, mdt->mdt_bottom);
215 rc = dt_trans_start(info->mti_env, mdt->mdt_bottom, th);
216 dt_trans_stop(info->mti_env, mdt->mdt_bottom, th);
221 spin_lock(&mdt->mdt_lut.lut_translock);
223 if (info->mti_transno != 0) {
224 struct obd_export *exp = req->rq_export;
226 CERROR("%s: replay trans %llu NID %s: rc = %d\n",
227 mdt_obd_name(mdt), info->mti_transno,
228 obd_export_nid2str(exp), rc);
229 spin_unlock(&mdt->mdt_lut.lut_translock);
232 } else if (info->mti_transno == 0) {
233 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
235 /* should be replay */
236 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
237 mdt->mdt_lut.lut_last_transno = info->mti_transno;
239 spin_unlock(&mdt->mdt_lut.lut_translock);
241 CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
243 req->rq_export->exp_obd->obd_last_committed);
245 req->rq_transno = info->mti_transno;
246 lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
248 /* update lcd in memory only for resent cases */
249 ted = &req->rq_export->exp_target_data;
251 mutex_lock(&ted->ted_lcd_lock);
253 if (info->mti_transno < lcd->lcd_last_transno &&
254 info->mti_transno != 0) {
255 /* This should happen during replay. Do not update
256 * last rcvd info if replay req transno < last transno,
257 * otherwise the following resend(after replay) can not
258 * be checked correctly by xid */
259 mutex_unlock(&ted->ted_lcd_lock);
260 CDEBUG(D_HA, "%s: transno = %llu < last_transno = %llu\n",
261 mdt_obd_name(mdt), info->mti_transno,
262 lcd->lcd_last_transno);
266 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
267 if (info->mti_transno != 0)
268 lcd->lcd_last_close_transno = info->mti_transno;
269 lcd->lcd_last_close_xid = req->rq_xid;
270 lcd->lcd_last_close_result = rc;
272 /* VBR: save versions in last_rcvd for reconstruct. */
273 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
275 lcd->lcd_pre_versions[0] = pre_versions[0];
276 lcd->lcd_pre_versions[1] = pre_versions[1];
277 lcd->lcd_pre_versions[2] = pre_versions[2];
278 lcd->lcd_pre_versions[3] = pre_versions[3];
280 if (info->mti_transno != 0)
281 lcd->lcd_last_transno = info->mti_transno;
283 lcd->lcd_last_xid = req->rq_xid;
284 lcd->lcd_last_result = rc;
285 lcd->lcd_last_data = info->mti_opdata;
287 mutex_unlock(&ted->ted_lcd_lock);
292 void mdt_mfd_set_mode(struct mdt_file_data *mfd, u64 open_flags)
294 LASSERT(mfd != NULL);
296 CDEBUG(D_DENTRY, DFID " Change mfd open_flags %#llo -> %#llo.\n",
297 PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_open_flags,
300 mfd->mfd_open_flags = open_flags;
304 * prep ma_lmm/ma_lmv for md_attr from reply
306 static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
307 struct mdt_object *obj,
310 if (ma->ma_lmv || ma->ma_lmm) {
311 CDEBUG(D_INFO, DFID " %s already set.\n",
312 PFID(mdt_object_fid(obj)),
313 ma->ma_lmv ? (ma->ma_lmm ? "ma_lmv and ma_lmm"
319 if (S_ISDIR(obj->mot_header.loh_attr)) {
320 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
322 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
325 if (ma->ma_lmv_size > 0)
326 ma->ma_need |= MA_LMV;
328 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
330 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
333 if (ma->ma_lmm_size > 0)
334 ma->ma_need |= MA_LOV;
338 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
339 struct mdt_object *o, u64 open_flags, int created,
340 struct ldlm_reply *rep)
342 struct ptlrpc_request *req = mdt_info_req(info);
343 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
344 struct mdt_file_data *mfd;
345 struct md_attr *ma = &info->mti_attr;
346 struct lu_attr *la = &ma->ma_attr;
347 struct mdt_body *repbody;
352 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
354 isreg = S_ISREG(la->la_mode);
355 isdir = S_ISDIR(la->la_mode);
356 if (isreg && !(ma->ma_valid & MA_LOV) &&
357 !(open_flags & MDS_OPEN_RELEASE)) {
359 * No EA, check whether it is will set regEA and dirEA since in
360 * above attr get, these size might be zero, so reset it, to
361 * retrieve the MD after create obj.
363 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
366 /* in replay case, p == NULL */
367 rc = mdt_create_data(info, p, o);
371 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
372 mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
375 CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n",
376 ma->ma_valid, ma->ma_lmm_size);
378 if (ma->ma_valid & MA_LOV) {
379 LASSERT(ma->ma_lmm_size != 0);
380 repbody->mbo_eadatasize = ma->ma_lmm_size;
382 repbody->mbo_valid |= OBD_MD_FLDIREA;
384 repbody->mbo_valid |= OBD_MD_FLEASIZE;
387 if (ma->ma_valid & MA_LMV) {
388 LASSERT(ma->ma_lmv_size != 0);
389 repbody->mbo_eadatasize = ma->ma_lmv_size;
391 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
394 if (open_flags & MDS_FMODE_WRITE)
395 rc = mdt_write_get(o);
396 else if (open_flags & MDS_FMODE_EXEC)
397 rc = mdt_write_deny(o);
402 rc = mo_open(info->mti_env, mdt_object_child(o),
403 created ? open_flags | MDS_OPEN_CREATED : open_flags,
406 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
407 * client does not know which suppgid should be sent to the MDS,
408 * or some other(s) changed the target file's GID after this RPC
409 * sent to the MDS with the suppgid as the original GID, then we
410 * give the client another chance to send the right suppgid. */
412 allow_client_chgrp(info, lu_ucred(info->mti_env)))
413 mdt_set_disposition(info, rep, DISP_OPEN_DENY);
418 mfd = mdt_mfd_new(med);
420 GOTO(err_out, rc = -ENOMEM);
423 * Keep a reference on this object for this open, and is
424 * released by mdt_mfd_close().
426 mdt_object_get(info->mti_env, o);
428 mfd->mfd_xid = req->rq_xid;
431 * @open_flags is always not zero. At least it should be FMODE_READ,
432 * FMODE_WRITE or MDS_FMODE_EXEC.
434 LASSERT(open_flags != 0);
437 mdt_mfd_set_mode(mfd, open_flags);
439 atomic_inc(&o->mot_open_count);
440 if (open_flags & MDS_OPEN_LEASE)
441 atomic_inc(&o->mot_lease_count);
444 if (req_is_replay(req)) {
445 struct mdt_file_data *old_mfd;
446 /* Check wheather old cookie already exist in
447 * the list, becasue when do recovery, client
448 * might be disconnected from server, and
449 * restart replay, so there maybe some orphan
450 * mfd here, we should remove them */
451 LASSERT(info->mti_rr.rr_open_handle != NULL);
452 spin_lock(&med->med_open_lock);
453 old_mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
455 if (old_mfd != NULL) {
456 CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
457 "cookie = %#llx\n", mfd,
458 PFID(mdt_object_fid(mfd->mfd_object)),
459 info->mti_rr.rr_open_handle->cookie);
460 class_handle_unhash(&old_mfd->mfd_open_handle);
461 list_del_init(&old_mfd->mfd_list);
462 spin_unlock(&med->med_open_lock);
463 /* no attr update for that close */
465 ma->ma_valid |= MA_FLAGS;
466 ma->ma_attr_flags |= MDS_RECOV_OPEN;
467 mdt_mfd_close(info, old_mfd);
468 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
469 ma->ma_valid &= ~MA_FLAGS;
471 spin_unlock(&med->med_open_lock);
472 CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
474 PFID(mdt_object_fid(mfd->mfd_object)),
475 info->mti_rr.rr_open_handle->cookie);
478 CDEBUG(D_HA, "Store old cookie %#llx in new mfd\n",
479 info->mti_rr.rr_open_handle->cookie);
481 mfd->mfd_open_handle_old = *info->mti_rr.rr_open_handle;
484 repbody->mbo_open_handle.cookie = mfd->mfd_open_handle.h_cookie;
486 if (req->rq_export->exp_disconnected) {
487 spin_lock(&med->med_open_lock);
488 class_handle_unhash(&mfd->mfd_open_handle);
489 list_del_init(&mfd->mfd_list);
490 spin_unlock(&med->med_open_lock);
491 mdt_mfd_close(info, mfd);
493 spin_lock(&med->med_open_lock);
494 list_add_tail(&mfd->mfd_list, &med->med_open_head);
495 spin_unlock(&med->med_open_lock);
498 mdt_empty_transno(info, rc);
503 if (open_flags & MDS_FMODE_WRITE)
505 else if (open_flags & MDS_FMODE_EXEC)
511 static int mdt_finish_open(struct mdt_thread_info *info,
512 struct mdt_object *p, struct mdt_object *o,
513 u64 open_flags, int created,
514 struct ldlm_reply *rep)
516 struct ptlrpc_request *req = mdt_info_req(info);
517 struct obd_export *exp = req->rq_export;
518 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
519 struct md_attr *ma = &info->mti_attr;
520 struct lu_attr *la = &ma->ma_attr;
521 struct mdt_file_data *mfd;
522 struct mdt_body *repbody;
524 int isreg, isdir, islnk;
528 LASSERT(ma->ma_valid & MA_INODE);
530 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
532 isreg = S_ISREG(la->la_mode);
533 isdir = S_ISDIR(la->la_mode);
534 islnk = S_ISLNK(la->la_mode);
535 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
537 /* compatibility check for 2.10 clients when it tries to open mirrored
538 * files. 2.10 clients don't verify overlapping components so they
539 * would read and write mirrored files just as if they were normal
540 * PFL files, which will cause the problem that sycned mirrors actually
541 * contain different data.
542 * Older clients are not a concern here because they don't even
543 * understand PFL layout. */
544 if (isreg && !exp_connect_flr(exp) && ma->ma_valid & MA_LOV &&
545 mdt_lmm_is_flr(ma->ma_lmm)) {
546 /* LU-10286: for simplicity clients who don't understand
547 * mirrored layout(with connect flag OBD_CONNECT2_FLR) won't
548 * be able to open mirrored files */
552 /* Overstriped files can crash older clients */
553 if (isreg && !exp_connect_overstriping(exp) &&
554 mdt_lmm_is_overstriping(ma->ma_lmm))
557 /* LU-2275, simulate broken behaviour (esp. prevalent in
558 * pre-2.4 servers where a very strange reply is sent on error
559 * that looks like it was actually almost successful and a
560 * failure at the same time.) */
561 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
562 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
566 if (open_flags & MDS_OPEN_LOCK)
567 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
572 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
573 if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
574 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
576 RETURN(PTR_ERR(nodemap));
578 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
579 nodemap_putref(nodemap);
586 * If we are following a symlink, don't open; and do not return open
587 * handle for special nodes as client required.
589 if (islnk || (!isreg && !isdir &&
590 (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
591 lustre_msg_set_transno(req->rq_repmsg, 0);
596 * We need to return the existing object's fid back, so it is done here,
597 * after preparing the reply.
599 if (!created && (open_flags & MDS_OPEN_EXCL) &&
600 (open_flags & MDS_OPEN_CREAT))
603 /* This can't be done earlier, we need to return reply body */
605 if (open_flags & (MDS_OPEN_CREAT | MDS_FMODE_WRITE)) {
606 /* We are trying to create or write an existing dir. */
609 } else if (open_flags & MDS_OPEN_DIRECTORY)
612 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
613 OBD_FAIL_MDS_LDLM_REPLY_NET | OBD_FAIL_ONCE))
617 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
618 spin_lock(&med->med_open_lock);
619 list_for_each(t, &med->med_open_head) {
620 mfd = list_entry(t, struct mdt_file_data, mfd_list);
621 if (mfd->mfd_xid == req->rq_xid)
625 spin_unlock(&med->med_open_lock);
628 repbody->mbo_open_handle.cookie =
629 mfd->mfd_open_handle.h_cookie;
630 /* set repbody->ea_size for resent case */
631 if (ma->ma_valid & MA_LOV) {
632 LASSERT(ma->ma_lmm_size != 0);
633 repbody->mbo_eadatasize = ma->ma_lmm_size;
635 repbody->mbo_valid |= OBD_MD_FLDIREA;
637 repbody->mbo_valid |= OBD_MD_FLEASIZE;
639 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
644 rc = mdt_mfd_open(info, p, o, open_flags, created, rep);
646 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
651 void mdt_reconstruct_open(struct mdt_thread_info *info,
652 struct mdt_lock_handle *lhc)
654 const struct lu_env *env = info->mti_env;
655 struct mdt_device *mdt = info->mti_mdt;
656 struct req_capsule *pill = info->mti_pill;
657 struct ptlrpc_request *req = mdt_info_req(info);
658 struct md_attr *ma = &info->mti_attr;
659 struct mdt_reint_record *rr = &info->mti_rr;
660 u64 open_flags = info->mti_spec.sp_cr_flags;
661 struct ldlm_reply *ldlm_rep;
662 struct mdt_object *parent;
663 struct mdt_object *child;
664 struct mdt_body *repbody;
669 LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
670 ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
671 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
673 ma->ma_need = MA_INODE | MA_HSM;
676 opdata = mdt_req_from_lrd(req, info->mti_reply_data);
677 mdt_set_disposition(info, ldlm_rep, opdata);
679 CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n",
680 ldlm_rep->lock_policy_res1, req->rq_status);
682 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
684 /* We did not create successfully, return error to client. */
685 GOTO(out, rc = req->rq_status);
687 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
688 struct obd_export *exp = req->rq_export;
690 * We failed after creation, but we do not know in which step
691 * we failed. So try to check the child object.
693 parent = mdt_object_find(env, mdt, rr->rr_fid1);
694 if (IS_ERR(parent)) {
695 rc = PTR_ERR(parent);
696 LCONSOLE_WARN("Parent "DFID" lookup error %d."
697 " Evicting client %s with export %s.\n",
698 PFID(rr->rr_fid1), rc,
699 obd_uuid2str(&exp->exp_client_uuid),
700 obd_export_nid2str(exp));
701 mdt_export_evict(exp);
705 child = mdt_object_find(env, mdt, rr->rr_fid2);
708 LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
709 "evicting client %s with export %s\n",
710 PFID(rr->rr_fid2), rc,
711 obd_uuid2str(&exp->exp_client_uuid),
712 obd_export_nid2str(exp));
713 mdt_object_put(env, parent);
714 mdt_export_evict(exp);
718 if (unlikely(mdt_object_remote(child))) {
719 mdt_object_put(env, parent);
720 mdt_object_put(env, child);
721 /* the child object was created on remote server */
722 if (!mdt_is_dne_client(exp))
723 /* Return -EIO for old client */
724 GOTO(out, rc = -EIO);
725 repbody->mbo_fid1 = *rr->rr_fid2;
726 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
729 if (mdt_object_exists(child)) {
730 mdt_prep_ma_buf_from_rep(info, child, ma);
731 rc = mdt_attr_get_complex(info, child, ma);
733 rc = mdt_finish_open(info, parent, child,
734 open_flags, 1, ldlm_rep);
735 mdt_object_put(env, parent);
736 mdt_object_put(env, child);
738 mdt_pack_size2body(info, rr->rr_fid2,
742 /* the child does not exist, we should do regular open */
743 mdt_object_put(env, parent);
744 mdt_object_put(env, child);
746 /* We did not try to create, so we are a pure open */
747 rc = mdt_reint_open(info, lhc);
751 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
752 LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
755 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
757 u64 open_flags = info->mti_spec.sp_cr_flags;
758 struct mdt_reint_record *rr = &info->mti_rr;
759 struct md_attr *ma = &info->mti_attr;
760 struct mdt_object *o;
764 o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
766 RETURN(rc = PTR_ERR(o));
768 if (unlikely(mdt_object_remote(o))) {
769 /* the child object was created on remote server */
770 struct mdt_body *repbody;
772 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
775 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
776 repbody->mbo_fid1 = *rr->rr_fid2;
777 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
780 if (mdt_object_exists(o)) {
781 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
784 mdt_prep_ma_buf_from_rep(info, o, ma);
785 rc = mdt_attr_get_complex(info, o, ma);
787 rc = mdt_finish_open(info, NULL, o, open_flags,
794 mdt_object_put(info->mti_env, o);
798 /* lock object for open */
799 static int mdt_object_open_lock(struct mdt_thread_info *info,
800 struct mdt_object *obj,
801 struct mdt_lock_handle *lhc,
804 struct md_attr *ma = &info->mti_attr;
805 __u64 open_flags = info->mti_spec.sp_cr_flags;
807 enum ldlm_mode lm = LCK_CR;
808 bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
809 bool try_layout = false;
810 bool create_layout = false;
812 __u32 dom_stripe = 0;
813 unsigned int dom_only = 0;
814 unsigned int dom_lock = 0;
819 mdt_lock_handle_init(lhc);
821 if (req_is_replay(mdt_info_req(info)))
824 if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
825 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
826 md_should_create(open_flags))
827 create_layout = true;
828 if (exp_connect_layout(info->mti_exp) && !create_layout &&
829 ma->ma_need & MA_LOV)
832 /* DoM files can take IO lock at OPEN when it makes sense,
833 * check if file has DoM stripe and ask for lock if client
834 * no lock on that resource yet.
836 if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
837 dom_stripe = mdt_lmm_dom_entry_check(ma->ma_lmm,
839 /* If only DOM stripe is being used then we can expect IO
840 * to it after OPEN and will return corresponding DOM ibit
841 * using default strategy from mdt_opts.mo_dom_lock.
842 * Otherwise trylock mode is used always and DOM ibit will
843 * be returned optionally.
846 !mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
847 dom_lock = !dom_only ? TRYLOCK_DOM_ON_OPEN :
848 info->mti_mdt->mdt_opts.mo_dom_lock;
852 /* lease open, acquire write mode of open sem */
853 down_write(&obj->mot_open_sem);
855 /* Lease exists and ask for new lease */
856 if (atomic_read(&obj->mot_lease_count) > 0) {
857 /* only exclusive open is supported, so lease
858 * are conflicted to each other */
859 GOTO(out, rc = -EBUSY);
862 /* Lease must be with open lock */
863 if (!(open_flags & MDS_OPEN_LOCK)) {
864 CERROR("%s: Request lease for file:"DFID ", but open lock "
865 "is missed, open_flags = %#llo : rc = %d\n",
866 mdt_obd_name(info->mti_mdt),
867 PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
868 GOTO(out, rc = -EPROTO);
871 /* XXX: only exclusive open is supported. */
873 *ibits = MDS_INODELOCK_OPEN;
875 /* never grant LCK_EX layout lock to client */
877 } else { /* normal open */
878 /* normal open holds read mode of open sem */
879 down_read(&obj->mot_open_sem);
881 if (open_flags & MDS_OPEN_LOCK) {
882 if (open_flags & MDS_FMODE_WRITE)
884 else if (open_flags & MDS_FMODE_EXEC)
889 *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
890 } else if (atomic_read(&obj->mot_lease_count) > 0) {
891 if (open_flags & MDS_FMODE_WRITE)
897 *ibits = MDS_INODELOCK_OPEN;
900 lhc = &info->mti_lh[MDT_LH_LOCAL];
901 } else if (dom_lock) {
902 lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR;
903 trybits |= MDS_INODELOCK_DOM | MDS_INODELOCK_LAYOUT;
906 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
907 PFID(mdt_object_fid(obj)),
908 atomic_read(&obj->mot_lease_count), lm);
911 mdt_lock_reg_init(lhc, lm);
913 /* Return lookup lock to validate inode at the client side.
914 * This is pretty important otherwise MDT will return layout
915 * lock for each open.
916 * However this is a double-edged sword because changing
917 * permission will revoke a huge number of LOOKUP locks.
919 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
920 if (!(*ibits & MDS_INODELOCK_LOOKUP))
921 trybits |= MDS_INODELOCK_LOOKUP;
922 trybits |= MDS_INODELOCK_LAYOUT;
925 if (*ibits | trybits)
926 rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, false);
928 CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
929 ", open_flags = %#llo, try_layout = %d : rc = %d\n",
930 mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
931 *ibits, trybits, open_flags, try_layout, rc);
933 /* will change layout, revoke layout locks by enqueuing EX lock. */
934 if (rc == 0 && create_layout) {
935 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
937 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
938 ", open_flags = %#llo\n",
939 PFID(mdt_object_fid(obj)), open_flags);
941 /* We cannot enqueue another lock for the same resource we
942 * already have a lock for, due to mechanics of waiting list
943 * iterating in ldlm, see LU-3601.
944 * As such we'll drop the open lock we just got above here,
945 * it's ok not to have this open lock as it's main purpose is to
946 * flush unused cached client open handles. */
947 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
948 mdt_object_unlock(info, obj, lhc, 1);
950 LASSERT(!try_layout);
951 mdt_lock_handle_init(ll);
952 mdt_lock_reg_init(ll, LCK_EX);
953 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
955 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
958 /* Check if there is any other open handles after acquiring
959 * open lock. At this point, caching open handles have been revoked
961 * XXX: Now only exclusive open is supported. Need to check the
962 * type of open for generic lease support. */
963 if (rc == 0 && acq_lease) {
964 struct ptlrpc_request *req = mdt_info_req(info);
965 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
966 struct mdt_file_data *mfd;
967 bool is_replay_or_resent;
970 /* For lease: application can open a file and then apply lease,
971 * @handle contains original open handle in that case.
972 * In recovery, open REQ will be replayed and the lease REQ may
973 * be resent that means the open handle is already stale, so we
974 * need to fix it up here by finding new handle. */
975 is_replay_or_resent = req_is_replay(req) ||
976 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
978 /* if the request is _not_ a replay request, rr_open_handle
979 * may be used to hold an open file handle which is issuing the
980 * lease request, so that this openhandle doesn't count. */
981 mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
982 is_replay_or_resent);
986 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
987 PFID(mdt_object_fid(obj)),
988 atomic_read(&obj->mot_open_count), open_count);
990 if (atomic_read(&obj->mot_open_count) > open_count)
991 GOTO(out, rc = -EBUSY);
999 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1000 struct mdt_object *obj,
1001 struct mdt_lock_handle *lhc,
1002 __u64 ibits, int rc)
1004 __u64 open_flags = info->mti_spec.sp_cr_flags;
1005 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1008 if (req_is_replay(mdt_info_req(info)))
1011 /* Release local lock - the lock put in MDT_LH_LOCAL will never
1012 * return to client side. */
1013 if (lustre_handle_is_used(&ll->mlh_reg_lh))
1014 mdt_object_unlock(info, obj, ll, 1);
1016 ll = &info->mti_lh[MDT_LH_LAYOUT];
1017 /* Release local layout lock, layout was created */
1018 if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1019 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1020 mdt_object_unlock(info, obj, ll, 1);
1023 if (open_flags & MDS_OPEN_LEASE)
1024 up_write(&obj->mot_open_sem);
1026 up_read(&obj->mot_open_sem);
1028 /* Cross-ref case, the lock should be returned to the client */
1029 if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1032 if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1033 !(ibits & MDS_INODELOCK_DOM)) {
1034 /* for the open request, the lock will only return to client
1035 * if open or layout lock is granted. */
1039 if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1040 struct ldlm_reply *ldlm_rep;
1042 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1043 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1044 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1045 mdt_object_unlock(info, obj, lhc, 1);
1051 * Check release is permitted for the current HSM flags.
1053 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1055 if (!(ma->ma_valid & MA_HSM))
1058 if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1061 if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1067 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1068 struct ldlm_reply *rep,
1069 struct mdt_lock_handle *lhc)
1071 const struct lu_env *env = info->mti_env;
1072 struct mdt_device *mdt = info->mti_mdt;
1073 u64 open_flags = info->mti_spec.sp_cr_flags;
1074 struct mdt_reint_record *rr = &info->mti_rr;
1075 struct md_attr *ma = &info->mti_attr;
1076 struct mdt_object *parent = NULL;
1077 struct mdt_object *o;
1078 bool object_locked = false;
1083 if (md_should_create(open_flags)) {
1084 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1085 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1086 if (IS_ERR(parent)) {
1087 CDEBUG(D_INODE, "Fail to find parent "DFID
1088 " for anonymous created %ld, try to"
1089 " use server-side parent.\n",
1090 PFID(rr->rr_fid1), PTR_ERR(parent));
1095 ma->ma_need |= MA_PFID;
1098 o = mdt_object_find(env, mdt, rr->rr_fid2);
1100 GOTO(out_parent_put, rc = PTR_ERR(o));
1102 if (mdt_object_remote(o)) {
1103 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1104 mdt_obd_name(info->mti_mdt),
1106 GOTO(out, rc = -EREMOTE);
1107 } else if (!mdt_object_exists(o)) {
1108 mdt_set_disposition(info, rep,
1112 GOTO(out, rc = -ENOENT);
1115 mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1117 mdt_prep_ma_buf_from_rep(info, o, ma);
1118 if (open_flags & MDS_OPEN_RELEASE)
1119 ma->ma_need |= MA_HSM;
1120 rc = mdt_attr_get_complex(info, o, ma);
1124 /* We should not change file's existing LOV EA */
1125 if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1126 open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1127 GOTO(out, rc = -EEXIST);
1129 /* If a release request, check file open flags are fine and ask for an
1130 * exclusive open access. */
1131 if (open_flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1132 GOTO(out, rc = -EPERM);
1134 rc = mdt_check_resent_lock(info, o, lhc);
1137 } else if (rc > 0) {
1138 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1139 object_locked = true;
1141 GOTO(out_unlock, rc);
1144 if (ma->ma_valid & MA_PFID) {
1145 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1146 if (IS_ERR(parent)) {
1147 CDEBUG(D_INODE, "Fail to find parent "DFID
1148 " for anonymous created %ld, try to"
1149 " use system default.\n",
1150 PFID(&ma->ma_pfid), PTR_ERR(parent));
1155 rc = mdt_finish_open(info, parent, o, open_flags, 0, rep);
1157 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1158 if (open_flags & MDS_OPEN_LOCK)
1159 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1160 if (open_flags & MDS_OPEN_LEASE)
1161 mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1163 GOTO(out_unlock, rc);
1167 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1169 mdt_object_put(env, o);
1171 mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
1174 mdt_object_put(env, parent);
1178 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1179 static int mdt_cross_open(struct mdt_thread_info *info,
1180 const struct lu_fid *parent_fid,
1181 const struct lu_fid *fid,
1182 struct ldlm_reply *rep, u64 open_flags)
1184 struct md_attr *ma = &info->mti_attr;
1185 struct mdt_object *o;
1189 o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1191 RETURN(rc = PTR_ERR(o));
1193 if (mdt_object_remote(o)) {
1194 /* Something is wrong here, the object is on another MDS! */
1195 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1196 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1197 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1199 "Object isn't on this server! FLD error?");
1202 if (mdt_object_exists(o)) {
1205 /* Do permission check for cross-open after converting
1206 * MDS_OPEN_* flags to MAY_* permission mask.
1208 mask = mds_accmode(open_flags);
1210 rc = mo_permission(info->mti_env, NULL,
1211 mdt_object_child(o), NULL, mask);
1215 mdt_prep_ma_buf_from_rep(info, o, ma);
1216 rc = mdt_attr_get_complex(info, o, ma);
1220 rc = mdt_pack_secctx_in_reply(info, o);
1224 rc = mdt_pack_encctx_in_reply(info, o);
1228 rc = mdt_finish_open(info, NULL, o, open_flags, 0, rep);
1231 * Something is wrong here. lookup was positive but
1232 * there is no object!
1234 CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1235 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1240 mdt_object_put(info->mti_env, o);
1246 * find root object and take its xattr lock if it's on remote MDT, later create
1247 * may use fs default striping (which is stored in root xattr).
1249 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1250 struct mdt_device *mdt)
1252 struct mdt_object *md_root = mdt->mdt_md_root;
1253 struct lustre_handle lhroot;
1256 if (md_root == NULL) {
1257 lu_root_fid(&info->mti_tmp_fid1);
1258 md_root = mdt_object_find(info->mti_env, mdt,
1259 &info->mti_tmp_fid1);
1260 if (IS_ERR(md_root))
1261 return PTR_ERR(md_root);
1263 spin_lock(&mdt->mdt_lock);
1264 if (mdt->mdt_md_root != NULL) {
1265 spin_unlock(&mdt->mdt_lock);
1267 LASSERTF(mdt->mdt_md_root == md_root,
1268 "Different root object ("
1269 DFID") instances, %p, %p\n",
1270 PFID(&info->mti_tmp_fid1),
1271 mdt->mdt_md_root, md_root);
1272 LASSERT(atomic_read(
1273 &md_root->mot_obj.lo_header->loh_ref) > 1);
1275 mdt_object_put(info->mti_env, md_root);
1277 mdt->mdt_md_root = md_root;
1278 spin_unlock(&mdt->mdt_lock);
1282 if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1285 rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root),
1286 &lhroot, LCK_PR, MDS_INODELOCK_XATTR,
1291 md_root->mot_cache_attr = 1;
1293 /* don't cancel this lock, so that we know the cached xattr is valid. */
1294 ldlm_lock_decref(&lhroot, LCK_PR);
1299 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1301 struct mdt_device *mdt = info->mti_mdt;
1302 struct ptlrpc_request *req = mdt_info_req(info);
1303 struct mdt_object *parent;
1304 struct mdt_object *child;
1305 struct mdt_lock_handle *lh;
1306 struct ldlm_reply *ldlm_rep;
1307 struct mdt_body *repbody;
1308 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1309 struct md_attr *ma = &info->mti_attr;
1310 u64 open_flags = info->mti_spec.sp_cr_flags;
1312 struct mdt_reint_record *rr = &info->mti_rr;
1315 int object_locked = 0;
1316 enum ldlm_mode lock_mode = LCK_PR;
1318 ktime_t kstart = ktime_get();
1321 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1322 (obd_timeout + 1) / 4);
1324 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1326 ma->ma_need = MA_INODE;
1329 LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1330 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1332 if (unlikely(open_flags & MDS_OPEN_JOIN_FILE)) {
1333 CERROR("file join is not supported anymore.\n");
1334 GOTO(out, result = err_serious(-EOPNOTSUPP));
1336 msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1338 if ((open_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1339 info->mti_spec.u.sp_ea.eadata == NULL)
1340 GOTO(out, result = err_serious(-EINVAL));
1342 if (open_flags & MDS_FMODE_WRITE &&
1343 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1344 GOTO(out, result = -EROFS);
1346 CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1347 "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1348 PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2),
1349 open_flags, ma->ma_attr.la_mode, msg_flags);
1351 if (info->mti_cross_ref) {
1352 /* This is cross-ref open */
1353 mdt_set_disposition(info, ldlm_rep,
1354 (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1356 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1357 ldlm_rep, open_flags);
1359 } else if (req_is_replay(req)) {
1360 result = mdt_open_by_fid(info, ldlm_rep);
1362 if (result != -ENOENT)
1365 /* We didn't find the correct object, so we need to re-create it
1366 * via a regular replay. */
1367 if (!(open_flags & MDS_OPEN_CREAT)) {
1368 DEBUG_REQ(D_ERROR, req,
1369 "OPEN & CREAT not in open replay/by_fid");
1370 GOTO(out, result = -EFAULT);
1372 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1373 } else if (open_flags & MDS_OPEN_BY_FID) {
1374 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1376 CDEBUG(D_INFO, "no object for "DFID": %d\n",
1377 PFID(rr->rr_fid2), result);
1381 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1382 GOTO(out, result = err_serious(-ENOMEM));
1384 mdt_set_disposition(info, ldlm_rep,
1385 (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1387 if (!lu_name_is_valid(&rr->rr_name))
1388 GOTO(out, result = -EPROTO);
1390 result = mdt_lock_root_xattr(info, mdt);
1394 parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1396 GOTO(out, result = PTR_ERR(parent));
1398 /* get and check version of parent */
1399 result = mdt_version_get_check(info, parent, 0);
1401 mdt_object_put(info->mti_env, parent);
1405 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
1407 lh = &info->mti_lh[MDT_LH_PARENT];
1408 mdt_lock_pdo_init(lh, lock_mode, &rr->rr_name);
1410 result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1412 mdt_object_put(info->mti_env, parent);
1415 fid_zero(child_fid);
1418 if ((open_flags & MDS_OPEN_VOLATILE) == 0)
1419 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1420 &rr->rr_name, child_fid, &info->mti_spec);
1422 LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1423 "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1424 PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1427 if (result != 0 && result != -ENOENT)
1428 GOTO(out_parent, result);
1430 OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1432 if (result == -ENOENT) {
1433 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1434 if (!(open_flags & MDS_OPEN_CREAT))
1435 GOTO(out_parent, result);
1436 if (mdt_rdonly(req->rq_export))
1437 GOTO(out_parent, result = -EROFS);
1439 if (lock_mode == LCK_PR) {
1440 /* first pass: get write lock and restart */
1441 mdt_object_unlock(info, parent, lh, 1);
1442 mdt_clear_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1443 mdt_lock_handle_init(lh);
1448 *child_fid = *info->mti_rr.rr_fid2;
1449 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1451 /* In the function below, .hs_keycmp resolves to
1452 * lu_obj_hop_keycmp() */
1453 /* coverity[overrun-buffer-val] */
1454 child = mdt_object_new(info->mti_env, mdt, child_fid);
1457 * Check for O_EXCL is moved to the mdt_finish_open(),
1458 * we need to return FID back in that case.
1460 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1461 child = mdt_object_find(info->mti_env, mdt, child_fid);
1464 GOTO(out_parent, result = PTR_ERR(child));
1466 /** check version of child */
1467 rc = mdt_version_get_check(info, child, 1);
1469 GOTO(out_child, result = rc);
1471 if (result == -ENOENT) {
1472 /* Create under OBF and .lustre is not permitted */
1473 if (!fid_is_md_operative(rr->rr_fid1) &&
1474 (open_flags & MDS_OPEN_VOLATILE) == 0)
1475 GOTO(out_child, result = -EPERM);
1477 /* save versions in reply */
1478 mdt_version_get_save(info, parent, 0);
1479 mdt_version_get_save(info, child, 1);
1481 /* version of child will be changed */
1482 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1484 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1485 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1487 /* Let lower layers know what is lock mode on directory. */
1488 info->mti_spec.sp_cr_mode =
1489 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1491 /* Don't do lookup sanity check. We know name doesn't exist. */
1492 info->mti_spec.sp_cr_lookup = 0;
1493 info->mti_spec.sp_feat = &dt_directory_features;
1495 result = mdo_create(info->mti_env, mdt_object_child(parent),
1496 &rr->rr_name, mdt_object_child(child),
1497 &info->mti_spec, &info->mti_attr);
1498 if (result == -ERESTART) {
1499 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1500 GOTO(out_child, result);
1502 mdt_prep_ma_buf_from_rep(info, child, ma);
1503 /* XXX: we should call this once, see few lines below */
1505 result = mdt_attr_get_complex(info, child, ma);
1508 GOTO(out_child, result);
1511 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1512 ktime_us_delta(ktime_get(), kstart));
1515 * The object is on remote node, return its FID for remote open.
1517 if (mdt_object_remote(child)) {
1519 * Check if this lock already was sent to client and
1520 * this is resent case. For resent case do not take lock
1521 * again, use what is already granted.
1523 LASSERT(lhc != NULL);
1525 rc = mdt_check_resent_lock(info, child, lhc);
1527 GOTO(out_child, result = rc);
1528 } else if (rc > 0) {
1529 mdt_lock_handle_init(lhc);
1530 mdt_lock_reg_init(lhc, LCK_PR);
1532 rc = mdt_object_lock(info, child, lhc,
1533 MDS_INODELOCK_LOOKUP);
1535 repbody->mbo_fid1 = *mdt_object_fid(child);
1536 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1540 result = -MDT_EREMOTE_OPEN;
1541 GOTO(out_child, result);
1542 } else if (mdt_object_exists(child)) {
1543 /* Check early for MDS_OPEN_DIRECTORY/O_DIRECTORY to
1544 * avoid opening regular files from lfs getstripe
1545 * since doing so breaks the leases used by lfs
1546 * mirror. See LU-13693. */
1547 if (open_flags & MDS_OPEN_DIRECTORY &&
1548 S_ISREG(lu_object_attr(&child->mot_obj)))
1549 GOTO(out_child, result = -ENOTDIR);
1551 /* We have to get attr & LOV EA & HSM for this
1553 mdt_prep_ma_buf_from_rep(info, child, ma);
1554 ma->ma_need |= MA_HSM;
1555 result = mdt_attr_get_complex(info, child, ma);
1557 GOTO(out_child, result);
1559 /* Object does not exist. Likely FS corruption. */
1560 CERROR("%s: name '"DNAME"' present, but FID "
1561 DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1562 PNAME(&rr->rr_name), PFID(child_fid));
1563 GOTO(out_child, result = -EIO);
1567 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1568 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1570 rc = mdt_pack_secctx_in_reply(info, child);
1572 GOTO(out_child, result = rc);
1574 rc = mdt_pack_encctx_in_reply(info, child);
1576 GOTO(out_child, result = rc);
1578 rc = mdt_check_resent_lock(info, child, lhc);
1580 GOTO(out_child, result = rc);
1581 } else if (rc == 0) {
1582 /* the open lock might already be gotten in
1583 * ldlm_handle_enqueue() */
1584 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1585 if (open_flags & MDS_OPEN_LOCK)
1586 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1588 /* get openlock if this isn't replay and client requested it */
1589 if (!req_is_replay(req)) {
1590 rc = mdt_object_open_lock(info, child, lhc, &ibits);
1593 GOTO(out_child_unlock, result = rc);
1594 else if (open_flags & MDS_OPEN_LOCK)
1595 mdt_set_disposition(info, ldlm_rep,
1599 /* Try to open it now. */
1600 rc = mdt_finish_open(info, parent, child, open_flags,
1604 /* openlock will be released if mdt_finish_open() failed */
1605 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1607 if (created && (open_flags & MDS_OPEN_VOLATILE)) {
1608 CERROR("%s: cannot open volatile file "DFID", orphan "
1609 "file will be left in PENDING directory until "
1610 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1611 PFID(mdt_object_fid(child)), rc);
1612 GOTO(out_child_unlock, result);
1618 rc = mdo_unlink(info->mti_env,
1619 mdt_object_child(parent),
1620 mdt_object_child(child),
1622 &info->mti_attr, 0);
1624 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1625 mdt_obd_name(info->mti_mdt),
1626 PFID(mdt_object_fid(child)), rc);
1627 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1631 mdt_counter_incr(req, LPROC_MDT_OPEN,
1632 ktime_us_delta(ktime_get(), kstart));
1637 mdt_object_open_unlock(info, child, lhc, ibits, result);
1639 mdt_object_put(info->mti_env, child);
1641 mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh);
1643 mdt_object_unlock_put(info, parent, lh, result || !created);
1646 lustre_msg_set_transno(req->rq_repmsg, 0);
1651 * Create an orphan object use local root.
1653 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1654 struct mdt_device *mdt,
1655 const struct lu_fid *fid,
1656 struct md_attr *attr, fmode_t fmode)
1658 const struct lu_env *env = info->mti_env;
1659 struct md_op_spec *spec = &info->mti_spec;
1660 struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1661 struct mdt_object *obj = NULL;
1662 struct mdt_object *local_root;
1663 static const struct lu_name lname = {
1664 .ln_name = "i_am_nobody",
1665 .ln_namelen = sizeof("i_am_nobody") - 1,
1667 struct lu_ucred *uc;
1668 cfs_cap_t uc_cap_save;
1672 rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1674 RETURN(ERR_PTR(rc));
1676 local_root = mdt_object_find(env, mdt, local_root_fid);
1677 if (IS_ERR(local_root))
1680 obj = mdt_object_new(env, mdt, fid);
1682 GOTO(out, rc = PTR_ERR(obj));
1684 spec->sp_cr_lookup = 0;
1685 spec->sp_feat = &dt_directory_features;
1686 spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1687 spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1688 if (attr->ma_valid & MA_LOV) {
1689 spec->u.sp_ea.eadata = attr->ma_lmm;
1690 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1691 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1693 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1697 uc_cap_save = uc->uc_cap;
1698 uc->uc_cap |= BIT(CAP_DAC_OVERRIDE);
1699 rc = mdo_create(env, mdt_object_child(local_root), &lname,
1700 mdt_object_child(obj), spec, attr);
1701 uc->uc_cap = uc_cap_save;
1703 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1704 mdt_obd_name(mdt), PFID(fid), rc);
1708 rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED, spec);
1710 CERROR("%s: cannot open volatile file "DFID", orphan "
1711 "file will be left in PENDING directory until "
1712 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1719 mdt_object_put(env, obj);
1722 mdt_object_put(env, local_root);
1726 /* XXX Look into layout in MDT layer. */
1727 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1729 struct lov_comp_md_v1 *comp_v1;
1730 struct lov_mds_md *v1;
1734 if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1735 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1737 if (comp_v1->lcm_entry_count == 0)
1740 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1741 off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1742 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1743 v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1746 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1751 static inline int mdt_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1753 struct lov_comp_md_v1 *comp_v1;
1755 if (le32_to_cpu(lmm->lmm_magic == LOV_MAGIC_COMP_V1)) {
1756 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1757 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1758 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1759 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1760 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1767 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1770 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1771 struct lu_ucred *uc = mdt_ucred(info);
1772 struct close_data *data;
1773 struct ldlm_lock *lease;
1774 struct mdt_object *orphan;
1775 struct md_attr *orp_ma;
1783 if (mdt_rdonly(info->mti_exp))
1786 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1790 lease = ldlm_handle2lock(&data->cd_handle);
1794 /* try to hold open_sem so that nobody else can open the file */
1795 if (!down_write_trylock(&o->mot_open_sem)) {
1796 ldlm_lock_cancel(lease);
1797 GOTO(out_reprocess, rc = -EBUSY);
1800 /* Check if the lease open lease has already canceled */
1801 lock_res_and_lock(lease);
1802 lease_broken = ldlm_is_cancel(lease);
1803 unlock_res_and_lock(lease);
1805 LDLM_DEBUG(lease, DFID " lease broken? %d",
1806 PFID(mdt_object_fid(o)), lease_broken);
1808 /* Cancel server side lease. Client side counterpart should
1809 * have been cancelled. It's okay to cancel it now as we've
1810 * held mot_open_sem. */
1811 ldlm_lock_cancel(lease);
1813 if (lease_broken) /* don't perform release task */
1814 GOTO(out_unlock, rc = -ESTALE);
1816 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1817 GOTO(out_unlock, rc = -EINVAL);
1819 /* ma_need was set before but it seems fine to change it in order to
1820 * avoid modifying the one from RPC */
1821 ma->ma_need = MA_HSM;
1822 rc = mdt_attr_get_complex(info, o, ma);
1824 GOTO(out_unlock, rc);
1826 if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
1827 if (ma->ma_valid & MA_HSM) {
1828 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1829 GOTO(out_unlock, rc = -EALREADY);
1831 if (ma->ma_hsm.mh_arch_id != data->cd_archive_id)
1833 DFID" archive id diff: %llu:%u\n",
1834 PFID(mdt_object_fid(o)),
1835 ma->ma_hsm.mh_arch_id,
1836 data->cd_archive_id);
1838 if (!(ma->ma_hsm.mh_flags & HS_DIRTY) &&
1839 ma->ma_hsm.mh_arch_ver == data->cd_data_version) {
1841 DFID" data version matches: packed=%llu "
1842 "and on-disk=%llu\n",
1843 PFID(mdt_object_fid(o)),
1844 data->cd_data_version,
1845 ma->ma_hsm.mh_arch_ver);
1846 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1849 if (ma->ma_hsm.mh_flags & HS_DIRTY)
1850 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1852 /* Set up HSM attribte for PCC archived object */
1853 BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1854 sizeof(info->mti_xattr_buf));
1855 buf = &info->mti_buf;
1856 buf->lb_buf = info->mti_xattr_buf;
1857 buf->lb_len = sizeof(struct hsm_attrs);
1858 memset(&ma->ma_hsm, 0, sizeof(ma->ma_hsm));
1859 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1860 ma->ma_hsm.mh_arch_id = data->cd_archive_id;
1861 ma->ma_hsm.mh_arch_ver = data->cd_data_version;
1862 lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1864 rc = mo_xattr_set(info->mti_env, mdt_object_child(o),
1865 buf, XATTR_NAME_HSM, 0);
1867 GOTO(out_unlock, rc);
1870 if (!mdt_hsm_release_allow(ma))
1871 GOTO(out_unlock, rc = -EPERM);
1873 /* already released? */
1874 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1875 GOTO(out_unlock, rc = 0);
1877 /* Compare on-disk and packed data_version */
1878 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1879 CDEBUG(D_HSM, DFID" data_version mismatches: "
1880 "packed=%llu and on-disk=%llu\n",
1881 PFID(mdt_object_fid(o)),
1882 data->cd_data_version,
1883 ma->ma_hsm.mh_arch_ver);
1884 GOTO(out_unlock, rc = -EPERM);
1888 ma->ma_valid = MA_INODE;
1889 ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1890 rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1892 GOTO(out_unlock, rc);
1894 mutex_lock(&o->mot_som_mutex);
1895 rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size,
1896 ma->ma_attr.la_blocks);
1897 mutex_unlock(&o->mot_som_mutex);
1900 "%s: File "DFID" SOM update failed: rc = %d\n",
1901 mdt_obd_name(info->mti_mdt),
1902 PFID(mdt_object_fid(o)), rc2);
1905 ma->ma_need = MA_INODE | MA_LOV;
1906 rc = mdt_attr_get_complex(info, o, ma);
1908 GOTO(out_unlock, rc);
1910 if (!(ma->ma_valid & MA_LOV)) {
1911 /* Even empty file are released */
1912 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1913 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
1914 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1915 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1916 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1918 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
1919 * ma_lmm as lov_user_md, then it will be confused by union of
1920 * layout_gen and stripe_offset. */
1921 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
1923 ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1925 GOTO(out_unlock, rc = -EINVAL);
1928 /* Set file as released. */
1929 rc = mdt_hsm_set_released(ma->ma_lmm);
1931 GOTO(out_unlock, rc);
1933 orp_ma = &info->mti_u.hsm.attr;
1934 orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1935 /* We use root ownership to bypass potential quota
1936 * restrictions on the user and group of the file. */
1937 orp_ma->ma_attr.la_uid = 0;
1938 orp_ma->ma_attr.la_gid = 0;
1939 orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1940 orp_ma->ma_lmm = ma->ma_lmm;
1941 orp_ma->ma_lmm_size = ma->ma_lmm_size;
1942 orp_ma->ma_valid = MA_INODE | MA_LOV;
1943 orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1945 if (IS_ERR(orphan)) {
1946 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1947 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1949 GOTO(out_unlock, rc = PTR_ERR(orphan));
1952 /* Set up HSM attribute for orphan object */
1953 BUILD_BUG_ON(sizeof(struct hsm_attrs) > sizeof(info->mti_xattr_buf));
1954 buf = &info->mti_buf;
1955 buf->lb_buf = info->mti_xattr_buf;
1956 buf->lb_len = sizeof(struct hsm_attrs);
1957 ma->ma_hsm.mh_flags |= HS_RELEASED;
1958 lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1959 ma->ma_hsm.mh_flags &= ~HS_RELEASED;
1961 mdt_lock_reg_init(lh, LCK_EX);
1962 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
1963 MDS_INODELOCK_XATTR);
1965 GOTO(out_close, rc);
1967 /* The orphan has root ownership so we need to raise
1968 * CAP_FOWNER to set the HSM attributes. */
1970 uc->uc_cap |= MD_CAP_TO_MASK(CAP_FOWNER);
1971 rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
1975 GOTO(out_layout_lock, rc);
1977 /* Swap layout with orphan objects. */
1978 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
1979 mdt_object_child(orphan),
1980 SWAP_LAYOUTS_MDS_HSM);
1982 if (!rc && ma->ma_attr_flags & MDS_PCC_ATTACH) {
1983 ma->ma_need = MA_LOV;
1984 rc = mdt_attr_get_complex(info, o, ma);
1990 /* Release exclusive LL */
1991 mdt_object_unlock(info, o, lh, 1);
1993 /* Close orphan object anyway */
1994 rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
1997 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
1998 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
1999 LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2001 mdt_object_put(info->mti_env, orphan);
2004 up_write(&o->mot_open_sem);
2006 /* already released */
2008 struct mdt_body *repbody;
2010 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2011 LASSERT(repbody != NULL);
2012 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2013 if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
2014 LASSERT(ma->ma_valid & MA_LOV);
2015 rc = mdt_get_lmm_gen(ma->ma_lmm,
2016 &repbody->mbo_layout_gen);
2018 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
2023 ldlm_reprocess_all(lease->l_resource, lease);
2024 LDLM_LOCK_PUT(lease);
2032 int mdt_close_handle_layouts(struct mdt_thread_info *info,
2033 struct mdt_object *o, struct md_attr *ma)
2035 struct mdt_lock_handle *lh1 = &info->mti_lh[MDT_LH_NEW];
2036 struct mdt_lock_handle *lh2 = &info->mti_lh[MDT_LH_OLD];
2037 struct close_data *data;
2038 struct ldlm_lock *lease;
2039 struct mdt_object *o1 = o, *o2 = NULL;
2041 bool swap_objects = false;
2045 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2048 if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
2051 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2055 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2058 rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
2061 * only MDS_CLOSE_LAYOUT_SPLIT use the same fid to indicate
2062 * mirror deletion, so we'd zero cd_fid, and keeps o2 be NULL.
2064 if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT))
2067 /* zero cd_fid to keeps o2 be NULL */
2068 fid_zero(&data->cd_fid);
2069 } else if (rc < 0) {
2070 /* Exchange o1 and o2, to enforce locking order */
2071 swap_objects = true;
2074 lease = ldlm_handle2lock(&data->cd_handle);
2078 if (!fid_is_zero(&data->cd_fid)) {
2079 o2 = mdt_object_find(info->mti_env, info->mti_mdt,
2082 GOTO(out_lease, rc = PTR_ERR(o2));
2084 if (!mdt_object_exists(o2))
2085 GOTO(out_obj, rc = -ENOENT);
2087 if (!S_ISREG(lu_object_attr(&o2->mot_obj)))
2088 GOTO(out_obj, rc = -EINVAL);
2094 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
2100 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2),
2106 /* try to hold open_sem so that nobody else can open the file */
2107 if (!down_write_trylock(&o->mot_open_sem)) {
2108 ldlm_lock_cancel(lease);
2109 GOTO(out_obj, rc = -EBUSY);
2112 /* Check if the lease open lease has already canceled */
2113 lock_res_and_lock(lease);
2114 lease_broken = ldlm_is_cancel(lease);
2115 unlock_res_and_lock(lease);
2117 LDLM_DEBUG(lease, DFID " lease broken? %d",
2118 PFID(mdt_object_fid(o)), lease_broken);
2120 /* Cancel server side lease. Client side counterpart should
2121 * have been cancelled. It's okay to cancel it now as we've
2122 * held mot_open_sem. */
2123 ldlm_lock_cancel(lease);
2126 GOTO(out_unlock_sem, rc = -ESTALE);
2128 mdt_lock_reg_init(lh1, LCK_EX);
2129 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
2130 MDS_INODELOCK_XATTR);
2132 GOTO(out_unlock_sem, rc);
2135 mdt_lock_reg_init(lh2, LCK_EX);
2136 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2137 MDS_INODELOCK_XATTR);
2139 GOTO(out_unlock1, rc);
2142 /* Swap layout with orphan object */
2143 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2144 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2145 mdt_object_child(o2), 0);
2146 } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2147 ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2148 struct lu_buf *buf = &info->mti_buf;
2149 struct md_rejig_data mrd;
2152 mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2154 if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)) {
2155 /* paranoid check again */
2157 ":only mirror split support NULL o2 object\n",
2158 PFID(mdt_object_fid(o)));
2159 GOTO(out_unlock1, rc = -EINVAL);
2162 /* set NULL mrd_obj for deleting mirror objects */
2166 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)
2167 mrd.mrd_mirror_id = data->cd_mirror_id;
2169 buf->lb_len = sizeof(mrd);
2171 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2173 ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2174 LU_XATTR_SPLIT : LU_XATTR_MERGE);
2175 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2178 mutex_lock(&o->mot_som_mutex);
2179 rc2 = mdt_set_som(info, o, SOM_FL_STRICT,
2180 ma->ma_attr.la_size,
2181 ma->ma_attr.la_blocks);
2182 mutex_unlock(&o->mot_som_mutex);
2184 CERROR(DFID": Setting i_blocks error: %d, "
2185 "i_blocks will be reported wrongly and "
2186 "can only be fixed in next resync\n",
2187 PFID(mdt_object_fid(o)), rc2);
2191 GOTO(out_unlock2, rc);
2196 /* Release exclusive LL */
2198 mdt_object_unlock(info, o2, lh2, 1);
2201 mdt_object_unlock(info, o1, lh1, 1);
2204 up_write(&o->mot_open_sem);
2206 /* already swapped */
2208 struct mdt_body *repbody;
2210 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2211 LASSERT(repbody != NULL);
2212 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2217 /* the 2nd object has been used, and swapped to o1 */
2218 mdt_object_put(info->mti_env, o1);
2220 /* the 2nd object has been used, and not swapped */
2221 mdt_object_put(info->mti_env, o2);
2223 ldlm_reprocess_all(lease->l_resource, lease);
2226 LDLM_LOCK_PUT(lease);
2236 static int mdt_close_resync_done(struct mdt_thread_info *info,
2237 struct mdt_object *o, struct md_attr *ma)
2239 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LOCAL];
2240 struct close_data *data;
2241 struct ldlm_lock *lease;
2242 struct md_layout_change layout = { 0 };
2243 __u32 *resync_ids = NULL;
2244 size_t resync_count = 0;
2250 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2253 if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2256 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2260 if (ptlrpc_req_need_swab(mdt_info_req(info)))
2261 lustre_swab_close_data_resync_done(&data->cd_resync);
2263 if (!fid_is_zero(&data->cd_fid))
2266 lease = ldlm_handle2lock(&data->cd_handle);
2270 /* try to hold open_sem so that nobody else can open the file */
2271 if (!down_write_trylock(&o->mot_open_sem)) {
2272 ldlm_lock_cancel(lease);
2273 GOTO(out_reprocess, rc = -EBUSY);
2276 /* Check if the lease open lease has already canceled */
2277 lock_res_and_lock(lease);
2278 lease_broken = ldlm_is_cancel(lease);
2279 unlock_res_and_lock(lease);
2281 LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2282 PFID(mdt_object_fid(o)), lease_broken);
2284 /* Cancel server side lease. Client side counterpart should
2285 * have been cancelled. It's okay to cancel it now as we've
2286 * held mot_open_sem. */
2287 ldlm_lock_cancel(lease);
2289 if (lease_broken) /* don't perform release task */
2290 GOTO(out_unlock, rc = -ESTALE);
2292 resync_count = data->cd_resync.resync_count;
2294 if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2297 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2299 GOTO(out_unlock, rc = -EPROTO);
2301 OBD_ALLOC_PTR_ARRAY(resync_ids, resync_count);
2303 GOTO(out_unlock, rc = -ENOMEM);
2305 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2306 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2308 layout.mlc_resync_ids = resync_ids;
2310 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2313 layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2314 layout.mlc_resync_count = resync_count;
2315 if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2316 layout.mlc_som.lsa_valid = SOM_FL_STRICT;
2317 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2318 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2320 rc = mdt_layout_change(info, o, lhc, &layout);
2322 GOTO(out_unlock, rc);
2324 mdt_object_unlock(info, o, lhc, 0);
2329 up_write(&o->mot_open_sem);
2331 /* already released */
2333 struct mdt_body *repbody;
2335 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2336 LASSERT(repbody != NULL);
2337 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2341 OBD_FREE_PTR_ARRAY(resync_ids, resync_count);
2344 ldlm_reprocess_all(lease->l_resource, lease);
2345 LDLM_LOCK_PUT(lease);
2353 #define MFD_CLOSED(open_flags) ((open_flags) == MDS_FMODE_CLOSED)
2355 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2357 return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_open_flags));
2360 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2362 struct mdt_object *o = mfd->mfd_object;
2363 struct md_object *next = mdt_object_child(o);
2364 struct md_attr *ma = &info->mti_attr;
2365 struct lu_fid *ofid = &info->mti_tmp_fid1;
2372 open_flags = mfd->mfd_open_flags;
2373 intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2374 *ofid = *mdt_object_fid(o);
2376 /* the below message is checked in replay-single.sh test_46 */
2377 CDEBUG(D_INODE, "%s: %sclosing file handle "DFID" with intent: %llx\n",
2378 mdt_obd_name(info->mti_mdt),
2379 ma->ma_valid & MA_FORCE_LOG ? "force " : "", PFID(ofid), intent);
2382 case MDS_HSM_RELEASE: {
2383 rc = mdt_hsm_release(info, o, ma);
2385 CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2386 mdt_obd_name(info->mti_mdt),
2388 /* continue to close even error occurred. */
2392 case MDS_CLOSE_LAYOUT_MERGE:
2393 case MDS_CLOSE_LAYOUT_SPLIT:
2394 case MDS_CLOSE_LAYOUT_SWAP: {
2395 rc = mdt_close_handle_layouts(info, o, ma);
2398 "%s: cannot swap layout of "DFID": rc = %d\n",
2399 mdt_obd_name(info->mti_mdt),
2401 /* continue to close even if error occurred. */
2405 case MDS_CLOSE_RESYNC_DONE:
2406 rc = mdt_close_resync_done(info, o, ma);
2413 if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
2414 ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) {
2417 rc2 = mdt_lsom_update(info, o, false);
2420 "%s: File " DFID " LSOM failed: rc = %d\n",
2421 mdt_obd_name(info->mti_mdt),
2423 /* continue to close even if error occured. */
2426 if (open_flags & MDS_FMODE_WRITE)
2428 else if (open_flags & MDS_FMODE_EXEC)
2431 /* Update atime|mtime|ctime on close. */
2432 if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ ||
2433 open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
2434 (ma->ma_attr.la_valid & LA_ATIME ||
2435 ma->ma_attr.la_valid & LA_MTIME ||
2436 ma->ma_attr.la_valid & LA_CTIME)) {
2437 ma->ma_valid = MA_INODE;
2438 ma->ma_attr_flags |= MDS_CLOSE_UPDATE_TIMES;
2439 ma->ma_attr.la_valid &= (LA_ATIME | LA_MTIME | LA_CTIME);
2441 if (ma->ma_attr.la_valid & LA_MTIME) {
2442 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
2444 ma->ma_valid |= MA_PFID;
2447 rc = mo_attr_set(info->mti_env, next, ma);
2450 /* If file data is modified, add the dirty flag. */
2451 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2452 rc = mdt_add_dirty_flag(info, o, ma);
2454 ma->ma_need |= MA_INODE;
2455 ma->ma_valid &= ~MA_INODE;
2457 LASSERT(atomic_read(&o->mot_open_count) > 0);
2458 atomic_dec(&o->mot_open_count);
2459 mdt_handle_last_unlink(info, o, ma);
2461 if (!MFD_CLOSED(open_flags)) {
2462 rc = mo_close(info->mti_env, next, ma, open_flags);
2463 if (mdt_dom_check_for_discard(info, o))
2464 mdt_dom_discard_data(info, o);
2467 /* adjust open and lease count */
2468 if (open_flags & MDS_OPEN_LEASE) {
2469 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2470 atomic_dec(&o->mot_lease_count);
2474 mdt_object_put(info->mti_env, o);
2479 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2480 struct mdt_body *repbody)
2482 struct mdt_export_data *med;
2483 struct mdt_file_data *mfd;
2488 med = &req->rq_export->exp_mdt_data;
2489 spin_lock(&med->med_open_lock);
2490 mfd = mdt_open_handle2mfd(med, &info->mti_open_handle,
2491 req_is_replay(req));
2492 if (mdt_mfd_closed(mfd)) {
2493 spin_unlock(&med->med_open_lock);
2494 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2495 ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2496 info->mti_open_handle.cookie);
2497 /** not serious error since bug 3633 */
2500 class_handle_unhash(&mfd->mfd_open_handle);
2501 list_del_init(&mfd->mfd_list);
2502 spin_unlock(&med->med_open_lock);
2503 ret = mdt_mfd_close(info, mfd);
2506 RETURN(rc ? rc : ret);
2509 int mdt_close(struct tgt_session_info *tsi)
2511 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2512 struct ptlrpc_request *req = tgt_ses_req(tsi);
2513 struct md_attr *ma = &info->mti_attr;
2514 struct mdt_body *repbody = NULL;
2515 ktime_t kstart = ktime_get();
2519 /* Close may come with the Size-on-MDS update. Unpack it. */
2520 rc = mdt_close_unpack(info);
2522 GOTO(out, rc = err_serious(rc));
2524 /* These fields are no longer used and are left for compatibility.
2525 * size is always zero */
2526 req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2528 req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2530 rc = req_capsule_server_pack(info->mti_pill);
2531 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2532 mdt_client_compatibility(info);
2534 mdt_fix_reply(info);
2535 mdt_exit_ucred(info);
2536 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2539 /* Continue to close handle even if we can not pack reply */
2541 repbody = req_capsule_server_get(info->mti_pill,
2543 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2545 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2548 ma->ma_need = MA_INODE | MA_LOV;
2549 repbody->mbo_eadatasize = 0;
2550 repbody->mbo_aclsize = 0;
2552 rc = err_serious(rc);
2555 rc = mdt_close_internal(info, req, repbody);
2557 mdt_empty_transno(info, rc);
2559 if (repbody != NULL) {
2560 mdt_client_compatibility(info);
2561 rc = mdt_fix_reply(info);
2564 mdt_exit_ucred(info);
2565 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2566 GOTO(out, rc = err_serious(-ENOMEM));
2568 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2569 OBD_FAIL_MDS_CLOSE_NET_REP))
2570 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2572 mdt_thread_info_fini(info);
2574 mdt_counter_incr(req, LPROC_MDT_CLOSE,
2575 ktime_us_delta(ktime_get(), kstart));
2576 RETURN(rc ? rc : ret);