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_RCU(&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 open_flags)
296 LASSERT(mfd != NULL);
298 CDEBUG(D_DENTRY, DFID " Change mfd open_flags %#llo -> %#llo.\n",
299 PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_open_flags,
302 mfd->mfd_open_flags = open_flags;
306 * prep ma_lmm/ma_lmv for md_attr from reply
308 static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
309 struct mdt_object *obj,
312 LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
313 if (S_ISDIR(obj->mot_header.loh_attr)) {
314 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
316 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
319 if (ma->ma_lmv_size > 0)
320 ma->ma_need |= MA_LMV;
322 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
324 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
327 if (ma->ma_lmm_size > 0)
328 ma->ma_need |= MA_LOV;
332 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
333 struct mdt_object *o, u64 open_flags, int created,
334 struct ldlm_reply *rep)
336 struct ptlrpc_request *req = mdt_info_req(info);
337 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
338 struct mdt_file_data *mfd;
339 struct md_attr *ma = &info->mti_attr;
340 struct lu_attr *la = &ma->ma_attr;
341 struct mdt_body *repbody;
346 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
348 isreg = S_ISREG(la->la_mode);
349 isdir = S_ISDIR(la->la_mode);
350 if (isreg && !(ma->ma_valid & MA_LOV) &&
351 !(open_flags & MDS_OPEN_RELEASE)) {
353 * No EA, check whether it is will set regEA and dirEA since in
354 * above attr get, these size might be zero, so reset it, to
355 * retrieve the MD after create obj.
357 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
360 /* in replay case, p == NULL */
361 rc = mdt_create_data(info, p, o);
365 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
366 mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
369 CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n",
370 ma->ma_valid, ma->ma_lmm_size);
372 if (ma->ma_valid & MA_LOV) {
373 LASSERT(ma->ma_lmm_size != 0);
374 repbody->mbo_eadatasize = ma->ma_lmm_size;
376 repbody->mbo_valid |= OBD_MD_FLDIREA;
378 repbody->mbo_valid |= OBD_MD_FLEASIZE;
381 if (ma->ma_valid & MA_LMV) {
382 LASSERT(ma->ma_lmv_size != 0);
383 repbody->mbo_eadatasize = ma->ma_lmv_size;
385 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
388 if (open_flags & MDS_FMODE_WRITE)
389 rc = mdt_write_get(o);
390 else if (open_flags & MDS_FMODE_EXEC)
391 rc = mdt_write_deny(o);
396 rc = mo_open(info->mti_env, mdt_object_child(o),
397 created ? open_flags | MDS_OPEN_CREATED : open_flags);
399 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
400 * client does not know which suppgid should be sent to the MDS,
401 * or some other(s) changed the target file's GID after this RPC
402 * sent to the MDS with the suppgid as the original GID, then we
403 * give the client another chance to send the right suppgid. */
405 allow_client_chgrp(info, lu_ucred(info->mti_env)))
406 mdt_set_disposition(info, rep, DISP_OPEN_DENY);
411 mfd = mdt_mfd_new(med);
413 GOTO(err_out, rc = -ENOMEM);
416 * Keep a reference on this object for this open, and is
417 * released by mdt_mfd_close().
419 mdt_object_get(info->mti_env, o);
421 mfd->mfd_xid = req->rq_xid;
424 * @open_flags is always not zero. At least it should be FMODE_READ,
425 * FMODE_WRITE or MDS_FMODE_EXEC.
427 LASSERT(open_flags != 0);
430 mdt_mfd_set_mode(mfd, open_flags);
432 atomic_inc(&o->mot_open_count);
433 if (open_flags & MDS_OPEN_LEASE)
434 atomic_inc(&o->mot_lease_count);
437 if (req_is_replay(req)) {
438 struct mdt_file_data *old_mfd;
439 /* Check wheather old cookie already exist in
440 * the list, becasue when do recovery, client
441 * might be disconnected from server, and
442 * restart replay, so there maybe some orphan
443 * mfd here, we should remove them */
444 LASSERT(info->mti_rr.rr_open_handle != NULL);
445 spin_lock(&med->med_open_lock);
446 old_mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
448 if (old_mfd != NULL) {
449 CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
450 "cookie = %#llx\n", mfd,
451 PFID(mdt_object_fid(mfd->mfd_object)),
452 info->mti_rr.rr_open_handle->cookie);
453 class_handle_unhash(&old_mfd->mfd_open_handle);
454 list_del_init(&old_mfd->mfd_list);
455 spin_unlock(&med->med_open_lock);
456 /* no attr update for that close */
458 ma->ma_valid |= MA_FLAGS;
459 ma->ma_attr_flags |= MDS_RECOV_OPEN;
460 mdt_mfd_close(info, old_mfd);
461 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
462 ma->ma_valid &= ~MA_FLAGS;
464 spin_unlock(&med->med_open_lock);
465 CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
467 PFID(mdt_object_fid(mfd->mfd_object)),
468 info->mti_rr.rr_open_handle->cookie);
471 CDEBUG(D_HA, "Store old cookie %#llx in new mfd\n",
472 info->mti_rr.rr_open_handle->cookie);
474 mfd->mfd_open_handle_old = *info->mti_rr.rr_open_handle;
477 repbody->mbo_open_handle.cookie = mfd->mfd_open_handle.h_cookie;
479 if (req->rq_export->exp_disconnected) {
480 spin_lock(&med->med_open_lock);
481 class_handle_unhash(&mfd->mfd_open_handle);
482 list_del_init(&mfd->mfd_list);
483 spin_unlock(&med->med_open_lock);
484 mdt_mfd_close(info, mfd);
486 spin_lock(&med->med_open_lock);
487 list_add_tail(&mfd->mfd_list, &med->med_open_head);
488 spin_unlock(&med->med_open_lock);
491 mdt_empty_transno(info, rc);
496 if (open_flags & MDS_FMODE_WRITE)
498 else if (open_flags & MDS_FMODE_EXEC)
504 static int mdt_finish_open(struct mdt_thread_info *info,
505 struct mdt_object *p, struct mdt_object *o,
506 u64 open_flags, int created,
507 struct ldlm_reply *rep)
509 struct ptlrpc_request *req = mdt_info_req(info);
510 struct obd_export *exp = req->rq_export;
511 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
512 struct md_attr *ma = &info->mti_attr;
513 struct lu_attr *la = &ma->ma_attr;
514 struct mdt_file_data *mfd;
515 struct mdt_body *repbody;
517 int isreg, isdir, islnk;
521 LASSERT(ma->ma_valid & MA_INODE);
523 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
525 isreg = S_ISREG(la->la_mode);
526 isdir = S_ISDIR(la->la_mode);
527 islnk = S_ISLNK(la->la_mode);
528 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
530 /* compatibility check for 2.10 clients when it tries to open mirrored
531 * files. 2.10 clients don't verify overlapping components so they
532 * would read and write mirrored files just as if they were normal
533 * PFL files, which will cause the problem that sycned mirrors actually
534 * contain different data.
535 * Older clients are not a concern here because they don't even
536 * understand PFL layout. */
537 if (isreg && !exp_connect_flr(exp) && ma->ma_valid & MA_LOV &&
538 mdt_lmm_is_flr(ma->ma_lmm)) {
539 /* LU-10286: for simplicity clients who don't understand
540 * mirrored layout(with connect flag OBD_CONNECT2_FLR) won't
541 * be able to open mirrored files */
545 /* Overstriped files can crash older clients */
546 if (isreg && !exp_connect_overstriping(exp) &&
547 mdt_lmm_is_overstriping(ma->ma_lmm))
550 /* LU-2275, simulate broken behaviour (esp. prevalent in
551 * pre-2.4 servers where a very strange reply is sent on error
552 * that looks like it was actually almost successful and a
553 * failure at the same time.) */
554 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
555 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
559 if (open_flags & MDS_OPEN_LOCK)
560 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
565 #ifdef CONFIG_FS_POSIX_ACL
566 if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
567 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
569 RETURN(PTR_ERR(nodemap));
571 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
572 nodemap_putref(nodemap);
579 * If we are following a symlink, don't open; and do not return open
580 * handle for special nodes as client required.
582 if (islnk || (!isreg && !isdir &&
583 (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
584 lustre_msg_set_transno(req->rq_repmsg, 0);
589 * We need to return the existing object's fid back, so it is done here,
590 * after preparing the reply.
592 if (!created && (open_flags & MDS_OPEN_EXCL) &&
593 (open_flags & MDS_OPEN_CREAT))
596 /* This can't be done earlier, we need to return reply body */
598 if (open_flags & (MDS_OPEN_CREAT | MDS_FMODE_WRITE)) {
599 /* We are trying to create or write an existing dir. */
602 } else if (open_flags & MDS_OPEN_DIRECTORY)
605 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
606 OBD_FAIL_MDS_LDLM_REPLY_NET | OBD_FAIL_ONCE))
610 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
611 spin_lock(&med->med_open_lock);
612 list_for_each(t, &med->med_open_head) {
613 mfd = list_entry(t, struct mdt_file_data, mfd_list);
614 if (mfd->mfd_xid == req->rq_xid)
618 spin_unlock(&med->med_open_lock);
621 repbody->mbo_open_handle.cookie =
622 mfd->mfd_open_handle.h_cookie;
623 /* set repbody->ea_size for resent case */
624 if (ma->ma_valid & MA_LOV) {
625 LASSERT(ma->ma_lmm_size != 0);
626 repbody->mbo_eadatasize = ma->ma_lmm_size;
628 repbody->mbo_valid |= OBD_MD_FLDIREA;
630 repbody->mbo_valid |= OBD_MD_FLEASIZE;
632 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
637 rc = mdt_mfd_open(info, p, o, open_flags, created, rep);
639 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
644 void mdt_reconstruct_open(struct mdt_thread_info *info,
645 struct mdt_lock_handle *lhc)
647 const struct lu_env *env = info->mti_env;
648 struct mdt_device *mdt = info->mti_mdt;
649 struct req_capsule *pill = info->mti_pill;
650 struct ptlrpc_request *req = mdt_info_req(info);
651 struct md_attr *ma = &info->mti_attr;
652 struct mdt_reint_record *rr = &info->mti_rr;
653 u64 open_flags = info->mti_spec.sp_cr_flags;
654 struct ldlm_reply *ldlm_rep;
655 struct mdt_object *parent;
656 struct mdt_object *child;
657 struct mdt_body *repbody;
662 LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
663 ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
664 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
666 ma->ma_need = MA_INODE | MA_HSM;
669 opdata = mdt_req_from_lrd(req, info->mti_reply_data);
670 mdt_set_disposition(info, ldlm_rep, opdata);
672 CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n",
673 ldlm_rep->lock_policy_res1, req->rq_status);
675 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
677 /* We did not create successfully, return error to client. */
678 GOTO(out, rc = req->rq_status);
680 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
681 struct obd_export *exp = req->rq_export;
683 * We failed after creation, but we do not know in which step
684 * we failed. So try to check the child object.
686 parent = mdt_object_find(env, mdt, rr->rr_fid1);
687 if (IS_ERR(parent)) {
688 rc = PTR_ERR(parent);
689 LCONSOLE_WARN("Parent "DFID" lookup error %d."
690 " Evicting client %s with export %s.\n",
691 PFID(rr->rr_fid1), rc,
692 obd_uuid2str(&exp->exp_client_uuid),
693 obd_export_nid2str(exp));
694 mdt_export_evict(exp);
698 child = mdt_object_find(env, mdt, rr->rr_fid2);
701 LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
702 "evicting client %s with export %s\n",
703 PFID(rr->rr_fid2), rc,
704 obd_uuid2str(&exp->exp_client_uuid),
705 obd_export_nid2str(exp));
706 mdt_object_put(env, parent);
707 mdt_export_evict(exp);
711 if (unlikely(mdt_object_remote(child))) {
712 mdt_object_put(env, parent);
713 mdt_object_put(env, child);
714 /* the child object was created on remote server */
715 if (!mdt_is_dne_client(exp))
716 /* Return -EIO for old client */
717 GOTO(out, rc = -EIO);
718 repbody->mbo_fid1 = *rr->rr_fid2;
719 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
722 if (mdt_object_exists(child)) {
723 mdt_prep_ma_buf_from_rep(info, child, ma);
724 rc = mdt_attr_get_complex(info, child, ma);
726 rc = mdt_finish_open(info, parent, child,
727 open_flags, 1, ldlm_rep);
728 mdt_object_put(env, parent);
729 mdt_object_put(env, child);
731 mdt_pack_size2body(info, rr->rr_fid2,
735 /* the child does not exist, we should do regular open */
736 mdt_object_put(env, parent);
737 mdt_object_put(env, child);
739 /* We did not try to create, so we are a pure open */
740 rc = mdt_reint_open(info, lhc);
744 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
745 LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
748 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
750 u64 open_flags = info->mti_spec.sp_cr_flags;
751 struct mdt_reint_record *rr = &info->mti_rr;
752 struct md_attr *ma = &info->mti_attr;
753 struct mdt_object *o;
757 o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
759 RETURN(rc = PTR_ERR(o));
761 if (unlikely(mdt_object_remote(o))) {
762 /* the child object was created on remote server */
763 struct mdt_body *repbody;
765 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
768 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
769 repbody->mbo_fid1 = *rr->rr_fid2;
770 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
773 if (mdt_object_exists(o)) {
774 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
777 mdt_prep_ma_buf_from_rep(info, o, ma);
778 rc = mdt_attr_get_complex(info, o, ma);
780 rc = mdt_finish_open(info, NULL, o, open_flags,
787 mdt_object_put(info->mti_env, o);
791 /* lock object for open */
792 static int mdt_object_open_lock(struct mdt_thread_info *info,
793 struct mdt_object *obj,
794 struct mdt_lock_handle *lhc,
797 struct md_attr *ma = &info->mti_attr;
798 __u64 open_flags = info->mti_spec.sp_cr_flags;
800 enum ldlm_mode lm = LCK_CR;
801 bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
802 bool try_layout = false;
803 bool create_layout = false;
805 int dom_stripes = LMM_NO_DOM;
806 bool dom_lock = false;
811 mdt_lock_handle_init(lhc);
813 if (req_is_replay(mdt_info_req(info)))
816 if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
817 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
818 md_should_create(open_flags))
819 create_layout = true;
820 if (exp_connect_layout(info->mti_exp) && !create_layout &&
821 ma->ma_need & MA_LOV)
824 /* DoM files can have just MDT stripe or combined MDT + OST
826 * - In the first case the open for read/write will do IO to
827 * the MDT stripe and it makes sense to take IO lock in
828 * advance along with OPEN even if it is blocking lock.
829 * - In the second case it is just size of MDT stripe and it
830 * is quite unlikely that client will write into it, though
831 * it may read it. So IO lock will be taken optionally if it
832 * is non-blocking one.
834 if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
835 dom_stripes = mdt_lmm_dom_entry(ma->ma_lmm);
837 if (dom_stripes == LMM_DOM_ONLY &&
838 info->mti_mdt->mdt_opts.mo_dom_lock > 0 &&
839 !mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
844 /* lease open, acquire write mode of open sem */
845 down_write(&obj->mot_open_sem);
847 /* Lease exists and ask for new lease */
848 if (atomic_read(&obj->mot_lease_count) > 0) {
849 /* only exclusive open is supported, so lease
850 * are conflicted to each other */
851 GOTO(out, rc = -EBUSY);
854 /* Lease must be with open lock */
855 if (!(open_flags & MDS_OPEN_LOCK)) {
856 CERROR("%s: Request lease for file:"DFID ", but open lock "
857 "is missed, open_flags = %#llo : rc = %d\n",
858 mdt_obd_name(info->mti_mdt),
859 PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
860 GOTO(out, rc = -EPROTO);
863 /* XXX: only exclusive open is supported. */
865 *ibits = MDS_INODELOCK_OPEN;
867 /* never grant LCK_EX layout lock to client */
869 } else { /* normal open */
870 /* normal open holds read mode of open sem */
871 down_read(&obj->mot_open_sem);
873 if (open_flags & MDS_OPEN_LOCK) {
874 if (open_flags & MDS_FMODE_WRITE)
876 else if (open_flags & MDS_FMODE_EXEC)
881 *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
882 } else if (atomic_read(&obj->mot_lease_count) > 0) {
883 if (open_flags & MDS_FMODE_WRITE)
889 *ibits = MDS_INODELOCK_OPEN;
892 lhc = &info->mti_lh[MDT_LH_LOCAL];
893 } else if (dom_lock) {
894 lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR;
895 if (info->mti_mdt->mdt_opts.mo_dom_lock ==
896 TRYLOCK_DOM_ON_OPEN) {
897 trybits |= MDS_INODELOCK_DOM |
898 MDS_INODELOCK_LAYOUT;
900 /* mo_dom_lock == ALWAYS_DOM_LOCK_ON_OPEN */
901 *ibits = MDS_INODELOCK_DOM;
902 if (info->mti_mdt->mdt_opts.mo_dom_read_open) {
903 trybits |= MDS_INODELOCK_LAYOUT;
908 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
909 PFID(mdt_object_fid(obj)),
910 atomic_read(&obj->mot_lease_count), lm);
913 mdt_lock_reg_init(lhc, lm);
915 /* Return lookup lock to validate inode at the client side.
916 * This is pretty important otherwise MDT will return layout
917 * lock for each open.
918 * However this is a double-edged sword because changing
919 * permission will revoke a huge number of LOOKUP locks.
921 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
922 if (!(*ibits & MDS_INODELOCK_LOOKUP))
923 trybits |= MDS_INODELOCK_LOOKUP;
924 trybits |= MDS_INODELOCK_LAYOUT;
927 if (*ibits | trybits)
928 rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, false);
930 CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
931 ", open_flags = %#llo, try_layout = %d : rc = %d\n",
932 mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
933 *ibits, trybits, open_flags, try_layout, rc);
935 /* will change layout, revoke layout locks by enqueuing EX lock. */
936 if (rc == 0 && create_layout) {
937 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
939 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
940 ", open_flags = %#llo\n",
941 PFID(mdt_object_fid(obj)), open_flags);
943 /* We cannot enqueue another lock for the same resource we
944 * already have a lock for, due to mechanics of waiting list
945 * iterating in ldlm, see LU-3601.
946 * As such we'll drop the open lock we just got above here,
947 * it's ok not to have this open lock as it's main purpose is to
948 * flush unused cached client open handles. */
949 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
950 mdt_object_unlock(info, obj, lhc, 1);
952 LASSERT(!try_layout);
953 mdt_lock_handle_init(ll);
954 mdt_lock_reg_init(ll, LCK_EX);
955 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
957 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
960 /* Check if there is any other open handles after acquiring
961 * open lock. At this point, caching open handles have been revoked
963 * XXX: Now only exclusive open is supported. Need to check the
964 * type of open for generic lease support. */
965 if (rc == 0 && acq_lease) {
966 struct ptlrpc_request *req = mdt_info_req(info);
967 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
968 struct mdt_file_data *mfd;
969 bool is_replay_or_resent;
972 /* For lease: application can open a file and then apply lease,
973 * @handle contains original open handle in that case.
974 * In recovery, open REQ will be replayed and the lease REQ may
975 * be resent that means the open handle is already stale, so we
976 * need to fix it up here by finding new handle. */
977 is_replay_or_resent = req_is_replay(req) ||
978 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
980 /* if the request is _not_ a replay request, rr_open_handle
981 * may be used to hold an open file handle which is issuing the
982 * lease request, so that this openhandle doesn't count. */
983 mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
984 is_replay_or_resent);
988 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
989 PFID(mdt_object_fid(obj)),
990 atomic_read(&obj->mot_open_count), open_count);
992 if (atomic_read(&obj->mot_open_count) > open_count)
993 GOTO(out, rc = -EBUSY);
1001 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1002 struct mdt_object *obj,
1003 struct mdt_lock_handle *lhc,
1004 __u64 ibits, int rc)
1006 __u64 open_flags = info->mti_spec.sp_cr_flags;
1007 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1010 if (req_is_replay(mdt_info_req(info)))
1013 /* Release local lock - the lock put in MDT_LH_LOCAL will never
1014 * return to client side. */
1015 if (lustre_handle_is_used(&ll->mlh_reg_lh))
1016 mdt_object_unlock(info, obj, ll, 1);
1018 ll = &info->mti_lh[MDT_LH_LAYOUT];
1019 /* Release local layout lock, layout was created */
1020 if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1021 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1022 mdt_object_unlock(info, obj, ll, 1);
1025 if (open_flags & MDS_OPEN_LEASE)
1026 up_write(&obj->mot_open_sem);
1028 up_read(&obj->mot_open_sem);
1030 /* Cross-ref case, the lock should be returned to the client */
1031 if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1034 if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1035 !(ibits & MDS_INODELOCK_DOM)) {
1036 /* for the open request, the lock will only return to client
1037 * if open or layout lock is granted. */
1041 if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1042 struct ldlm_reply *ldlm_rep;
1044 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1045 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1046 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1047 mdt_object_unlock(info, obj, lhc, 1);
1053 * Check release is permitted for the current HSM flags.
1055 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1057 if (!(ma->ma_valid & MA_HSM))
1060 if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1063 if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1069 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1070 struct ldlm_reply *rep,
1071 struct mdt_lock_handle *lhc)
1073 const struct lu_env *env = info->mti_env;
1074 struct mdt_device *mdt = info->mti_mdt;
1075 u64 open_flags = info->mti_spec.sp_cr_flags;
1076 struct mdt_reint_record *rr = &info->mti_rr;
1077 struct md_attr *ma = &info->mti_attr;
1078 struct mdt_object *parent = NULL;
1079 struct mdt_object *o;
1080 bool object_locked = false;
1085 if (md_should_create(open_flags)) {
1086 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1087 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1088 if (IS_ERR(parent)) {
1089 CDEBUG(D_INODE, "Fail to find parent "DFID
1090 " for anonymous created %ld, try to"
1091 " use server-side parent.\n",
1092 PFID(rr->rr_fid1), PTR_ERR(parent));
1097 ma->ma_need |= MA_PFID;
1100 o = mdt_object_find(env, mdt, rr->rr_fid2);
1102 GOTO(out_parent_put, rc = PTR_ERR(o));
1104 if (mdt_object_remote(o)) {
1105 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1106 mdt_obd_name(info->mti_mdt),
1108 GOTO(out, rc = -EREMOTE);
1109 } else if (!mdt_object_exists(o)) {
1110 mdt_set_disposition(info, rep,
1114 GOTO(out, rc = -ENOENT);
1117 mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1119 mdt_prep_ma_buf_from_rep(info, o, ma);
1120 if (open_flags & MDS_OPEN_RELEASE)
1121 ma->ma_need |= MA_HSM;
1122 rc = mdt_attr_get_complex(info, o, ma);
1126 /* We should not change file's existing LOV EA */
1127 if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1128 open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1129 GOTO(out, rc = -EEXIST);
1131 /* If a release request, check file open flags are fine and ask for an
1132 * exclusive open access. */
1133 if (open_flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1134 GOTO(out, rc = -EPERM);
1136 rc = mdt_check_resent_lock(info, o, lhc);
1139 } else if (rc > 0) {
1140 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1141 object_locked = true;
1143 GOTO(out_unlock, rc);
1146 if (ma->ma_valid & MA_PFID) {
1147 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1148 if (IS_ERR(parent)) {
1149 CDEBUG(D_INODE, "Fail to find parent "DFID
1150 " for anonymous created %ld, try to"
1151 " use system default.\n",
1152 PFID(&ma->ma_pfid), PTR_ERR(parent));
1157 rc = mdt_finish_open(info, parent, o, open_flags, 0, rep);
1159 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1160 if (open_flags & MDS_OPEN_LOCK)
1161 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1162 if (open_flags & MDS_OPEN_LEASE)
1163 mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1165 GOTO(out_unlock, rc);
1169 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1171 mdt_object_put(env, o);
1173 mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
1176 mdt_object_put(env, parent);
1180 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1181 static int mdt_cross_open(struct mdt_thread_info *info,
1182 const struct lu_fid *parent_fid,
1183 const struct lu_fid *fid,
1184 struct ldlm_reply *rep, u64 open_flags)
1186 struct md_attr *ma = &info->mti_attr;
1187 struct mdt_object *o;
1191 o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1193 RETURN(rc = PTR_ERR(o));
1195 if (mdt_object_remote(o)) {
1196 /* Something is wrong here, the object is on another MDS! */
1197 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1198 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1199 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1201 "Object isn't on this server! FLD error?");
1204 if (mdt_object_exists(o)) {
1207 /* Do permission check for cross-open after converting
1208 * MDS_OPEN_* flags to MAY_* permission mask.
1210 mask = mds_accmode(open_flags);
1212 rc = mo_permission(info->mti_env, NULL,
1213 mdt_object_child(o), NULL, mask);
1217 mdt_prep_ma_buf_from_rep(info, o, ma);
1218 rc = mdt_attr_get_complex(info, o, ma);
1222 mdt_pack_secctx_in_reply(info, o);
1224 rc = mdt_finish_open(info, NULL, o, open_flags, 0, rep);
1227 * Something is wrong here. lookup was positive but
1228 * there is no object!
1230 CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1231 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1236 mdt_object_put(info->mti_env, o);
1242 * find root object and take its xattr lock if it's on remote MDT, later create
1243 * may use fs default striping (which is stored in root xattr).
1245 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1246 struct mdt_device *mdt)
1248 struct mdt_object *md_root = mdt->mdt_md_root;
1249 struct lustre_handle lhroot;
1252 if (md_root == NULL) {
1253 lu_root_fid(&info->mti_tmp_fid1);
1254 md_root = mdt_object_find(info->mti_env, mdt,
1255 &info->mti_tmp_fid1);
1256 if (IS_ERR(md_root))
1257 return PTR_ERR(md_root);
1259 spin_lock(&mdt->mdt_lock);
1260 if (mdt->mdt_md_root != NULL) {
1261 spin_unlock(&mdt->mdt_lock);
1263 LASSERTF(mdt->mdt_md_root == md_root,
1264 "Different root object ("
1265 DFID") instances, %p, %p\n",
1266 PFID(&info->mti_tmp_fid1),
1267 mdt->mdt_md_root, md_root);
1268 LASSERT(atomic_read(
1269 &md_root->mot_obj.lo_header->loh_ref) > 1);
1271 mdt_object_put(info->mti_env, md_root);
1273 mdt->mdt_md_root = md_root;
1274 spin_unlock(&mdt->mdt_lock);
1278 if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1281 rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root),
1282 &lhroot, LCK_PR, MDS_INODELOCK_XATTR,
1287 md_root->mot_cache_attr = 1;
1289 /* don't cancel this lock, so that we know the cached xattr is valid. */
1290 ldlm_lock_decref(&lhroot, LCK_PR);
1295 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1297 struct mdt_device *mdt = info->mti_mdt;
1298 struct ptlrpc_request *req = mdt_info_req(info);
1299 struct mdt_object *parent;
1300 struct mdt_object *child;
1301 struct mdt_lock_handle *lh;
1302 struct ldlm_reply *ldlm_rep;
1303 struct mdt_body *repbody;
1304 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1305 struct md_attr *ma = &info->mti_attr;
1306 u64 open_flags = info->mti_spec.sp_cr_flags;
1308 struct mdt_reint_record *rr = &info->mti_rr;
1311 int object_locked = 0;
1315 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1316 (obd_timeout + 1) / 4);
1318 mdt_counter_incr(req, LPROC_MDT_OPEN);
1319 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1321 ma->ma_need = MA_INODE;
1324 LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1325 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1327 if (unlikely(open_flags & MDS_OPEN_JOIN_FILE)) {
1328 CERROR("file join is not supported anymore.\n");
1329 GOTO(out, result = err_serious(-EOPNOTSUPP));
1331 msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1333 if ((open_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1334 info->mti_spec.u.sp_ea.eadata == NULL)
1335 GOTO(out, result = err_serious(-EINVAL));
1337 if (open_flags & MDS_FMODE_WRITE &&
1338 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1339 GOTO(out, result = -EROFS);
1341 CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1342 "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1343 PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2),
1344 open_flags, ma->ma_attr.la_mode, msg_flags);
1346 if (info->mti_cross_ref) {
1347 /* This is cross-ref open */
1348 mdt_set_disposition(info, ldlm_rep,
1349 (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1351 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1352 ldlm_rep, open_flags);
1354 } else if (req_is_replay(req)) {
1355 result = mdt_open_by_fid(info, ldlm_rep);
1357 if (result != -ENOENT)
1360 /* We didn't find the correct object, so we need to re-create it
1361 * via a regular replay. */
1362 if (!(open_flags & MDS_OPEN_CREAT)) {
1363 DEBUG_REQ(D_ERROR, req,
1364 "OPEN & CREAT not in open replay/by_fid.");
1365 GOTO(out, result = -EFAULT);
1367 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1368 } else if (open_flags & (MDS_OPEN_BY_FID | MDS_OPEN_LOCK)) {
1370 * MDS_OPEN_LOCK is checked for backward compatibility with 2.1
1373 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1375 CDEBUG(D_INFO, "no object for "DFID": %d\n",
1376 PFID(rr->rr_fid2), result);
1380 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1381 GOTO(out, result = err_serious(-ENOMEM));
1383 mdt_set_disposition(info, ldlm_rep,
1384 (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1386 if (!lu_name_is_valid(&rr->rr_name))
1387 GOTO(out, result = -EPROTO);
1389 result = mdt_lock_root_xattr(info, mdt);
1394 lh = &info->mti_lh[MDT_LH_PARENT];
1395 mdt_lock_pdo_init(lh, (open_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1398 parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1400 GOTO(out, result = PTR_ERR(parent));
1402 result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1404 mdt_object_put(info->mti_env, parent);
1408 /* get and check version of parent */
1409 result = mdt_version_get_check(info, parent, 0);
1411 GOTO(out_parent, result);
1413 fid_zero(child_fid);
1415 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1416 &rr->rr_name, child_fid, &info->mti_spec);
1418 LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1419 "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1420 PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1423 if (result != 0 && result != -ENOENT && result != -ESTALE)
1424 GOTO(out_parent, result);
1426 if (result == -ENOENT || result == -ESTALE) {
1427 /* If the object is dead, let's check if the object
1428 * is being migrated to a new object */
1429 if (result == -ESTALE) {
1430 struct lu_buf lmv_buf;
1432 lmv_buf.lb_buf = info->mti_xattr_buf;
1433 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
1434 rc = mo_xattr_get(info->mti_env,
1435 mdt_object_child(parent),
1436 &lmv_buf, XATTR_NAME_LMV);
1438 struct lmv_mds_md_v1 *lmv;
1440 lmv = lmv_buf.lb_buf;
1441 if (le32_to_cpu(lmv->lmv_hash_type) &
1442 LMV_HASH_FLAG_MIGRATION) {
1443 /* Get the new parent FID and retry */
1444 mdt_object_unlock_put(info, parent,
1446 mdt_lock_handle_init(lh);
1448 (struct lu_fid *)rr->rr_fid1,
1449 &lmv->lmv_stripe_fids[1]);
1455 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1456 if (result == -ESTALE) {
1458 * -ESTALE means the parent is a dead(unlinked) dir, so
1459 * it should return -ENOENT to in accordance with the
1460 * original mds implementaion.
1462 GOTO(out_parent, result = -ENOENT);
1465 if (!(open_flags & MDS_OPEN_CREAT))
1466 GOTO(out_parent, result);
1467 if (mdt_rdonly(req->rq_export))
1468 GOTO(out_parent, result = -EROFS);
1469 *child_fid = *info->mti_rr.rr_fid2;
1470 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1472 /* In the function below, .hs_keycmp resolves to
1473 * lu_obj_hop_keycmp() */
1474 /* coverity[overrun-buffer-val] */
1475 child = mdt_object_new(info->mti_env, mdt, child_fid);
1478 * Check for O_EXCL is moved to the mdt_finish_open(),
1479 * we need to return FID back in that case.
1481 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1482 child = mdt_object_find(info->mti_env, mdt, child_fid);
1485 GOTO(out_parent, result = PTR_ERR(child));
1487 /** check version of child */
1488 rc = mdt_version_get_check(info, child, 1);
1490 GOTO(out_child, result = rc);
1492 if (result == -ENOENT) {
1493 /* Create under OBF and .lustre is not permitted */
1494 if (!fid_is_md_operative(rr->rr_fid1))
1495 GOTO(out_child, result = -EPERM);
1497 /* save versions in reply */
1498 mdt_version_get_save(info, parent, 0);
1499 mdt_version_get_save(info, child, 1);
1501 /* version of child will be changed */
1502 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1504 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1505 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1507 /* Let lower layers know what is lock mode on directory. */
1508 info->mti_spec.sp_cr_mode =
1509 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1511 /* Don't do lookup sanity check. We know name doesn't exist. */
1512 info->mti_spec.sp_cr_lookup = 0;
1513 info->mti_spec.sp_feat = &dt_directory_features;
1515 result = mdo_create(info->mti_env, mdt_object_child(parent),
1516 &rr->rr_name, mdt_object_child(child),
1517 &info->mti_spec, &info->mti_attr);
1518 if (result == -ERESTART) {
1519 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1520 GOTO(out_child, result);
1522 mdt_prep_ma_buf_from_rep(info, child, ma);
1523 /* XXX: we should call this once, see few lines below */
1525 result = mdt_attr_get_complex(info, child, ma);
1528 GOTO(out_child, result);
1531 mdt_counter_incr(req, LPROC_MDT_MKNOD);
1534 * The object is on remote node, return its FID for remote open.
1536 if (mdt_object_remote(child)) {
1538 * Check if this lock already was sent to client and
1539 * this is resent case. For resent case do not take lock
1540 * again, use what is already granted.
1542 LASSERT(lhc != NULL);
1544 rc = mdt_check_resent_lock(info, child, lhc);
1546 GOTO(out_child, result = rc);
1547 } else if (rc > 0) {
1548 mdt_lock_handle_init(lhc);
1549 mdt_lock_reg_init(lhc, LCK_PR);
1551 rc = mdt_object_lock(info, child, lhc,
1552 MDS_INODELOCK_LOOKUP);
1554 repbody->mbo_fid1 = *mdt_object_fid(child);
1555 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1559 result = -MDT_EREMOTE_OPEN;
1560 GOTO(out_child, result);
1561 } else if (mdt_object_exists(child)) {
1562 /* We have to get attr & LOV EA & HSM for this
1564 mdt_prep_ma_buf_from_rep(info, child, ma);
1565 ma->ma_need |= MA_HSM;
1566 result = mdt_attr_get_complex(info, child, ma);
1568 GOTO(out_child, result);
1570 /* Object does not exist. Likely FS corruption. */
1571 CERROR("%s: name '"DNAME"' present, but FID "
1572 DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1573 PNAME(&rr->rr_name), PFID(child_fid));
1574 GOTO(out_child, result = -EIO);
1578 mdt_pack_secctx_in_reply(info, child);
1580 rc = mdt_check_resent_lock(info, child, lhc);
1582 GOTO(out_child, result = rc);
1583 } else if (rc == 0) {
1584 /* the open lock might already be gotten in
1585 * ldlm_handle_enqueue() */
1586 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1587 if (open_flags & MDS_OPEN_LOCK)
1588 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1590 /* get openlock if this isn't replay and client requested it */
1591 if (!req_is_replay(req)) {
1592 rc = mdt_object_open_lock(info, child, lhc, &ibits);
1595 GOTO(out_child_unlock, result = rc);
1596 else if (open_flags & MDS_OPEN_LOCK)
1597 mdt_set_disposition(info, ldlm_rep,
1601 /* Try to open it now. */
1602 rc = mdt_finish_open(info, parent, child, open_flags,
1606 /* openlock will be released if mdt_finish_open() failed */
1607 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1609 if (created && (open_flags & MDS_OPEN_VOLATILE)) {
1610 CERROR("%s: cannot open volatile file "DFID", orphan "
1611 "file will be left in PENDING directory until "
1612 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1613 PFID(mdt_object_fid(child)), rc);
1614 GOTO(out_child_unlock, result);
1620 rc = mdo_unlink(info->mti_env,
1621 mdt_object_child(parent),
1622 mdt_object_child(child),
1624 &info->mti_attr, 0);
1626 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1627 mdt_obd_name(info->mti_mdt),
1628 PFID(mdt_object_fid(child)), rc);
1629 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1635 mdt_object_open_unlock(info, child, lhc, ibits, result);
1637 mdt_object_put(info->mti_env, child);
1639 mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh);
1641 mdt_object_unlock_put(info, parent, lh, result || !created);
1644 lustre_msg_set_transno(req->rq_repmsg, 0);
1649 * Create an orphan object use local root.
1651 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1652 struct mdt_device *mdt,
1653 const struct lu_fid *fid,
1654 struct md_attr *attr, fmode_t fmode)
1656 const struct lu_env *env = info->mti_env;
1657 struct md_op_spec *spec = &info->mti_spec;
1658 struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1659 struct mdt_object *obj = NULL;
1660 struct mdt_object *local_root;
1661 static const struct lu_name lname = {
1662 .ln_name = "i_am_nobody",
1663 .ln_namelen = sizeof("i_am_nobody") - 1,
1665 struct lu_ucred *uc;
1666 cfs_cap_t uc_cap_save;
1670 rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1672 RETURN(ERR_PTR(rc));
1674 local_root = mdt_object_find(env, mdt, local_root_fid);
1675 if (IS_ERR(local_root))
1678 obj = mdt_object_new(env, mdt, fid);
1680 GOTO(out, rc = PTR_ERR(obj));
1682 spec->sp_cr_lookup = 0;
1683 spec->sp_feat = &dt_directory_features;
1684 spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1685 spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1686 if (attr->ma_valid & MA_LOV) {
1687 spec->u.sp_ea.eadata = attr->ma_lmm;
1688 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1689 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1691 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1695 uc_cap_save = uc->uc_cap;
1696 uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1697 rc = mdo_create(env, mdt_object_child(local_root), &lname,
1698 mdt_object_child(obj), spec, attr);
1699 uc->uc_cap = uc_cap_save;
1701 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1702 mdt_obd_name(mdt), PFID(fid), rc);
1706 rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1708 CERROR("%s: cannot open volatile file "DFID", orphan "
1709 "file will be left in PENDING directory until "
1710 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1717 mdt_object_put(env, obj);
1720 mdt_object_put(env, local_root);
1724 /* XXX Look into layout in MDT layer. */
1725 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1727 struct lov_comp_md_v1 *comp_v1;
1728 struct lov_mds_md *v1;
1732 if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1733 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1735 if (comp_v1->lcm_entry_count == 0)
1738 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1739 off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1740 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1741 v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1744 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1749 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1752 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1753 struct lu_ucred *uc = mdt_ucred(info);
1754 struct close_data *data;
1755 struct ldlm_lock *lease;
1756 struct mdt_object *orphan;
1757 struct md_attr *orp_ma;
1765 if (mdt_rdonly(info->mti_exp))
1768 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1772 lease = ldlm_handle2lock(&data->cd_handle);
1776 /* try to hold open_sem so that nobody else can open the file */
1777 if (!down_write_trylock(&o->mot_open_sem)) {
1778 ldlm_lock_cancel(lease);
1779 GOTO(out_reprocess, rc = -EBUSY);
1782 /* Check if the lease open lease has already canceled */
1783 lock_res_and_lock(lease);
1784 lease_broken = ldlm_is_cancel(lease);
1785 unlock_res_and_lock(lease);
1787 LDLM_DEBUG(lease, DFID " lease broken? %d",
1788 PFID(mdt_object_fid(o)), lease_broken);
1790 /* Cancel server side lease. Client side counterpart should
1791 * have been cancelled. It's okay to cancel it now as we've
1792 * held mot_open_sem. */
1793 ldlm_lock_cancel(lease);
1795 if (lease_broken) /* don't perform release task */
1796 GOTO(out_unlock, rc = -ESTALE);
1798 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1799 GOTO(out_unlock, rc = -EINVAL);
1801 /* ma_need was set before but it seems fine to change it in order to
1802 * avoid modifying the one from RPC */
1803 ma->ma_need = MA_HSM;
1804 rc = mdt_attr_get_complex(info, o, ma);
1806 GOTO(out_unlock, rc);
1808 if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
1809 if (ma->ma_valid & MA_HSM) {
1810 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1811 GOTO(out_unlock, rc = -EALREADY);
1813 if (ma->ma_hsm.mh_arch_id != data->cd_archive_id)
1815 DFID" archive id diff: %llu:%u\n",
1816 PFID(mdt_object_fid(o)),
1817 ma->ma_hsm.mh_arch_id,
1818 data->cd_archive_id);
1820 if (!(ma->ma_hsm.mh_flags & HS_DIRTY) &&
1821 ma->ma_hsm.mh_arch_ver == data->cd_data_version) {
1823 DFID" data version matches: packed=%llu "
1824 "and on-disk=%llu\n",
1825 PFID(mdt_object_fid(o)),
1826 data->cd_data_version,
1827 ma->ma_hsm.mh_arch_ver);
1828 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1831 /* Set up HSM attribte for PCC archived object */
1832 CLASSERT(sizeof(struct hsm_attrs) <=
1833 sizeof(info->mti_xattr_buf));
1834 buf = &info->mti_buf;
1835 buf->lb_buf = info->mti_xattr_buf;
1836 buf->lb_len = sizeof(struct hsm_attrs);
1837 memset(&ma->ma_hsm, 0, sizeof(ma->ma_hsm));
1838 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1839 ma->ma_hsm.mh_arch_id = data->cd_archive_id;
1840 ma->ma_hsm.mh_arch_ver = data->cd_data_version;
1841 lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1843 rc = mo_xattr_set(info->mti_env, mdt_object_child(o),
1844 buf, XATTR_NAME_HSM, 0);
1846 GOTO(out_unlock, rc);
1849 if (!mdt_hsm_release_allow(ma))
1850 GOTO(out_unlock, rc = -EPERM);
1852 /* already released? */
1853 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1854 GOTO(out_unlock, rc = 0);
1856 /* Compare on-disk and packed data_version */
1857 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1858 CDEBUG(D_HSM, DFID" data_version mismatches: "
1859 "packed=%llu and on-disk=%llu\n",
1860 PFID(mdt_object_fid(o)),
1861 data->cd_data_version,
1862 ma->ma_hsm.mh_arch_ver);
1863 GOTO(out_unlock, rc = -EPERM);
1867 ma->ma_valid = MA_INODE;
1868 ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1869 rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1871 GOTO(out_unlock, rc);
1873 mutex_lock(&o->mot_som_mutex);
1874 rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size,
1875 ma->ma_attr.la_blocks);
1876 mutex_unlock(&o->mot_som_mutex);
1879 "%s: File "DFID" SOM update failed: rc = %d\n",
1880 mdt_obd_name(info->mti_mdt),
1881 PFID(mdt_object_fid(o)), rc2);
1884 ma->ma_need = MA_INODE | MA_LOV;
1885 rc = mdt_attr_get_complex(info, o, ma);
1887 GOTO(out_unlock, rc);
1889 if (!(ma->ma_valid & MA_LOV)) {
1890 /* Even empty file are released */
1891 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1892 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
1893 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1894 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1895 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1897 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
1898 * ma_lmm as lov_user_md, then it will be confused by union of
1899 * layout_gen and stripe_offset. */
1900 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
1902 ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1904 GOTO(out_unlock, rc = -EINVAL);
1907 /* Set file as released. */
1908 rc = mdt_hsm_set_released(ma->ma_lmm);
1910 GOTO(out_unlock, rc);
1912 orp_ma = &info->mti_u.hsm.attr;
1913 orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1914 /* We use root ownership to bypass potential quota
1915 * restrictions on the user and group of the file. */
1916 orp_ma->ma_attr.la_uid = 0;
1917 orp_ma->ma_attr.la_gid = 0;
1918 orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1919 orp_ma->ma_lmm = ma->ma_lmm;
1920 orp_ma->ma_lmm_size = ma->ma_lmm_size;
1921 orp_ma->ma_valid = MA_INODE | MA_LOV;
1922 orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1924 if (IS_ERR(orphan)) {
1925 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1926 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1928 GOTO(out_unlock, rc = PTR_ERR(orphan));
1931 /* Set up HSM attribute for orphan object */
1932 CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
1933 buf = &info->mti_buf;
1934 buf->lb_buf = info->mti_xattr_buf;
1935 buf->lb_len = sizeof(struct hsm_attrs);
1936 ma->ma_hsm.mh_flags |= HS_RELEASED;
1937 lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1938 ma->ma_hsm.mh_flags &= ~HS_RELEASED;
1940 mdt_lock_reg_init(lh, LCK_EX);
1941 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
1942 MDS_INODELOCK_XATTR);
1944 GOTO(out_close, rc);
1946 /* The orphan has root ownership so we need to raise
1947 * CAP_FOWNER to set the HSM attributes. */
1949 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
1950 rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
1954 GOTO(out_layout_lock, rc);
1956 /* Swap layout with orphan objects. */
1957 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
1958 mdt_object_child(orphan),
1959 SWAP_LAYOUTS_MDS_HSM);
1963 /* Release exclusive LL */
1964 mdt_object_unlock(info, o, lh, 1);
1966 /* Close orphan object anyway */
1967 rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
1970 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
1971 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
1972 LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
1974 mdt_object_put(info->mti_env, orphan);
1977 up_write(&o->mot_open_sem);
1979 /* already released */
1981 struct mdt_body *repbody;
1983 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1984 LASSERT(repbody != NULL);
1985 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1989 ldlm_reprocess_all(lease->l_resource);
1990 LDLM_LOCK_PUT(lease);
1998 int mdt_close_handle_layouts(struct mdt_thread_info *info,
1999 struct mdt_object *o, struct md_attr *ma)
2001 struct mdt_lock_handle *lh1 = &info->mti_lh[MDT_LH_NEW];
2002 struct mdt_lock_handle *lh2 = &info->mti_lh[MDT_LH_OLD];
2003 struct close_data *data;
2004 struct ldlm_lock *lease;
2005 struct mdt_object *o1 = o, *o2;
2011 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2014 if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
2017 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2021 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2024 rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
2025 if (unlikely(rc == 0))
2028 /* Exchange o1 and o2, to enforce locking order */
2029 swap_objects = (rc < 0);
2031 lease = ldlm_handle2lock(&data->cd_handle);
2035 o2 = mdt_object_find(info->mti_env, info->mti_mdt, &data->cd_fid);
2037 GOTO(out_lease, rc = PTR_ERR(o2));
2039 if (!S_ISREG(lu_object_attr(&o2->mot_obj))) {
2040 swap_objects = false; /* not swapped yet */
2041 GOTO(out_obj, rc = -EINVAL);
2047 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
2052 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
2057 /* try to hold open_sem so that nobody else can open the file */
2058 if (!down_write_trylock(&o->mot_open_sem)) {
2059 ldlm_lock_cancel(lease);
2060 GOTO(out_obj, rc = -EBUSY);
2063 /* Check if the lease open lease has already canceled */
2064 lock_res_and_lock(lease);
2065 lease_broken = ldlm_is_cancel(lease);
2066 unlock_res_and_lock(lease);
2068 LDLM_DEBUG(lease, DFID " lease broken? %d",
2069 PFID(mdt_object_fid(o)), lease_broken);
2071 /* Cancel server side lease. Client side counterpart should
2072 * have been cancelled. It's okay to cancel it now as we've
2073 * held mot_open_sem. */
2074 ldlm_lock_cancel(lease);
2077 GOTO(out_unlock_sem, rc = -ESTALE);
2079 mdt_lock_reg_init(lh1, LCK_EX);
2080 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
2081 MDS_INODELOCK_XATTR);
2083 GOTO(out_unlock_sem, rc);
2085 mdt_lock_reg_init(lh2, LCK_EX);
2086 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2087 MDS_INODELOCK_XATTR);
2089 GOTO(out_unlock1, rc);
2091 /* Swap layout with orphan object */
2092 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2093 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2094 mdt_object_child(o2), 0);
2095 } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2096 ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2097 struct lu_buf *buf = &info->mti_buf;
2098 struct md_rejig_data mrd;
2100 mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2101 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)
2102 mrd.mrd_mirror_id = data->cd_mirror_id;
2104 buf->lb_len = sizeof(mrd);
2106 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2108 ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2109 LU_XATTR_SPLIT : LU_XATTR_MERGE);
2110 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2113 mutex_lock(&o->mot_som_mutex);
2114 rc2 = mdt_set_som(info, o, SOM_FL_STRICT,
2115 ma->ma_attr.la_size,
2116 ma->ma_attr.la_blocks);
2117 mutex_unlock(&o->mot_som_mutex);
2119 CERROR(DFID": Setting i_blocks error: %d, "
2120 "i_blocks will be reported wrongly and "
2121 "can only be fixed in next resync\n",
2122 PFID(mdt_object_fid(o)), rc2);
2126 GOTO(out_unlock2, rc);
2131 /* Release exclusive LL */
2132 mdt_object_unlock(info, o2, lh2, 1);
2135 mdt_object_unlock(info, o1, lh1, 1);
2138 up_write(&o->mot_open_sem);
2140 /* already swapped */
2142 struct mdt_body *repbody;
2144 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2145 LASSERT(repbody != NULL);
2146 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2150 mdt_object_put(info->mti_env, swap_objects ? o1 : o2);
2152 ldlm_reprocess_all(lease->l_resource);
2155 LDLM_LOCK_PUT(lease);
2165 static int mdt_close_resync_done(struct mdt_thread_info *info,
2166 struct mdt_object *o, struct md_attr *ma)
2168 struct close_data *data;
2169 struct ldlm_lock *lease;
2170 struct md_layout_change layout = { 0 };
2171 __u32 *resync_ids = NULL;
2172 size_t resync_count = 0;
2177 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2180 if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2183 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2187 if (ptlrpc_req_need_swab(mdt_info_req(info)))
2188 lustre_swab_close_data_resync_done(&data->cd_resync);
2190 if (!fid_is_zero(&data->cd_fid))
2193 lease = ldlm_handle2lock(&data->cd_handle);
2197 /* try to hold open_sem so that nobody else can open the file */
2198 if (!down_write_trylock(&o->mot_open_sem)) {
2199 ldlm_lock_cancel(lease);
2200 GOTO(out_reprocess, rc = -EBUSY);
2203 /* Check if the lease open lease has already canceled */
2204 lock_res_and_lock(lease);
2205 lease_broken = ldlm_is_cancel(lease);
2206 unlock_res_and_lock(lease);
2208 LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2209 PFID(mdt_object_fid(o)), lease_broken);
2211 /* Cancel server side lease. Client side counterpart should
2212 * have been cancelled. It's okay to cancel it now as we've
2213 * held mot_open_sem. */
2214 ldlm_lock_cancel(lease);
2216 if (lease_broken) /* don't perform release task */
2217 GOTO(out_unlock, rc = -ESTALE);
2219 resync_count = data->cd_resync.resync_count;
2221 if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2224 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2226 GOTO(out_unlock, rc = -EPROTO);
2228 OBD_ALLOC(resync_ids, resync_count * sizeof(__u32));
2230 GOTO(out_unlock, rc = -ENOMEM);
2232 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2233 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2235 layout.mlc_resync_ids = resync_ids;
2237 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2240 layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2241 layout.mlc_resync_count = resync_count;
2242 if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2243 layout.mlc_som.lsa_valid = SOM_FL_STRICT;
2244 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2245 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2247 rc = mdt_layout_change(info, o, &layout);
2249 GOTO(out_unlock, rc);
2254 up_write(&o->mot_open_sem);
2256 /* already released */
2258 struct mdt_body *repbody;
2260 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2261 LASSERT(repbody != NULL);
2262 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2266 OBD_FREE(resync_ids, resync_count * sizeof(__u32));
2269 ldlm_reprocess_all(lease->l_resource);
2270 LDLM_LOCK_PUT(lease);
2278 #define MFD_CLOSED(open_flags) ((open_flags) == MDS_FMODE_CLOSED)
2280 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2282 return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_open_flags));
2285 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2287 struct mdt_object *o = mfd->mfd_object;
2288 struct md_object *next = mdt_object_child(o);
2289 struct md_attr *ma = &info->mti_attr;
2290 struct lu_fid *ofid = &info->mti_tmp_fid1;
2297 open_flags = mfd->mfd_open_flags;
2298 intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2299 *ofid = *mdt_object_fid(o);
2301 CDEBUG(D_INODE, "%s: close file "DFID" with intent: %llx\n",
2302 mdt_obd_name(info->mti_mdt), PFID(ofid), intent);
2305 case MDS_HSM_RELEASE: {
2306 rc = mdt_hsm_release(info, o, ma);
2308 CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2309 mdt_obd_name(info->mti_mdt),
2311 /* continue to close even error occurred. */
2315 case MDS_CLOSE_LAYOUT_MERGE:
2316 case MDS_CLOSE_LAYOUT_SPLIT:
2317 case MDS_CLOSE_LAYOUT_SWAP: {
2318 rc = mdt_close_handle_layouts(info, o, ma);
2321 "%s: cannot swap layout of "DFID": rc = %d\n",
2322 mdt_obd_name(info->mti_mdt),
2324 /* continue to close even if error occurred. */
2328 case MDS_CLOSE_RESYNC_DONE:
2329 rc = mdt_close_resync_done(info, o, ma);
2336 if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
2337 ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) {
2340 rc2 = mdt_lsom_update(info, o, false);
2343 "%s: File " DFID " LSOM failed: rc = %d\n",
2344 mdt_obd_name(info->mti_mdt),
2346 /* continue to close even if error occured. */
2349 if (open_flags & MDS_FMODE_WRITE)
2351 else if (open_flags & MDS_FMODE_EXEC)
2354 /* Update atime on close only. */
2355 if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ ||
2356 open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
2357 (ma->ma_attr.la_valid & LA_ATIME)) {
2358 /* Set the atime only. */
2359 ma->ma_valid = MA_INODE;
2360 ma->ma_attr.la_valid = LA_ATIME;
2361 rc = mo_attr_set(info->mti_env, next, ma);
2364 /* If file data is modified, add the dirty flag. */
2365 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2366 rc = mdt_add_dirty_flag(info, o, ma);
2368 ma->ma_need |= MA_INODE;
2369 ma->ma_valid &= ~MA_INODE;
2371 LASSERT(atomic_read(&o->mot_open_count) > 0);
2372 atomic_dec(&o->mot_open_count);
2373 mdt_handle_last_unlink(info, o, ma);
2375 if (!MFD_CLOSED(open_flags)) {
2376 rc = mo_close(info->mti_env, next, ma, open_flags);
2377 if (mdt_dom_check_for_discard(info, o))
2378 mdt_dom_discard_data(info, o);
2381 /* adjust open and lease count */
2382 if (open_flags & MDS_OPEN_LEASE) {
2383 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2384 atomic_dec(&o->mot_lease_count);
2388 mdt_object_put(info->mti_env, o);
2393 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2394 struct mdt_body *repbody)
2396 struct mdt_export_data *med;
2397 struct mdt_file_data *mfd;
2402 med = &req->rq_export->exp_mdt_data;
2403 spin_lock(&med->med_open_lock);
2404 mfd = mdt_open_handle2mfd(med, &info->mti_open_handle,
2405 req_is_replay(req));
2406 if (mdt_mfd_closed(mfd)) {
2407 spin_unlock(&med->med_open_lock);
2408 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2409 ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2410 info->mti_open_handle.cookie);
2411 /** not serious error since bug 3633 */
2414 class_handle_unhash(&mfd->mfd_open_handle);
2415 list_del_init(&mfd->mfd_list);
2416 spin_unlock(&med->med_open_lock);
2417 ret = mdt_mfd_close(info, mfd);
2420 RETURN(rc ? rc : ret);
2423 int mdt_close(struct tgt_session_info *tsi)
2425 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2426 struct ptlrpc_request *req = tgt_ses_req(tsi);
2427 struct md_attr *ma = &info->mti_attr;
2428 struct mdt_body *repbody = NULL;
2432 mdt_counter_incr(req, LPROC_MDT_CLOSE);
2433 /* Close may come with the Size-on-MDS update. Unpack it. */
2434 rc = mdt_close_unpack(info);
2436 GOTO(out, rc = err_serious(rc));
2438 /* These fields are no longer used and are left for compatibility.
2439 * size is always zero */
2440 req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2442 req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2444 rc = req_capsule_server_pack(info->mti_pill);
2445 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2446 mdt_client_compatibility(info);
2448 mdt_fix_reply(info);
2449 mdt_exit_ucred(info);
2450 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2453 /* Continue to close handle even if we can not pack reply */
2455 repbody = req_capsule_server_get(info->mti_pill,
2457 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2459 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2462 ma->ma_need = MA_INODE | MA_LOV;
2463 repbody->mbo_eadatasize = 0;
2464 repbody->mbo_aclsize = 0;
2466 rc = err_serious(rc);
2469 rc = mdt_close_internal(info, req, repbody);
2471 mdt_empty_transno(info, rc);
2473 if (repbody != NULL) {
2474 mdt_client_compatibility(info);
2475 rc = mdt_fix_reply(info);
2478 mdt_exit_ucred(info);
2479 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2480 GOTO(out, rc = err_serious(-ENOMEM));
2482 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2483 OBD_FAIL_MDS_CLOSE_NET_REP))
2484 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2486 mdt_thread_info_fini(info);
2487 RETURN(rc ? rc : ret);