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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_open.c
38 * Lustre Metadata Target (mdt) open/close file handling
40 * Author: Huang Hua <huanghua@clusterfs.com>
43 #define DEBUG_SUBSYSTEM S_MDS
45 #include <lustre_acl.h>
46 #include <lustre_mds.h>
47 #include "mdt_internal.h"
48 #include <lustre_nodemap.h>
50 /* we do nothing because we do not have refcount now */
51 static void mdt_mfd_get(void *mfdp)
55 static struct portals_handle_ops mfd_handle_ops = {
56 .hop_addref = mdt_mfd_get,
60 /* Create a new mdt_file_data struct, initialize it,
61 * and insert it to global hash table */
62 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
64 struct mdt_file_data *mfd;
69 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
70 mfd->mfd_handle.h_owner = med;
71 INIT_LIST_HEAD(&mfd->mfd_list);
72 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
79 * Find the mfd pointed to by handle in global hash table.
80 * In case of replay the handle is obsoleted
81 * but mfd can be found in mfd list by that handle.
82 * Callers need to be holding med_open_lock.
84 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
85 const struct lustre_handle *handle,
86 bool is_replay_or_resent)
88 struct mdt_file_data *mfd;
91 LASSERT(handle != NULL);
92 mfd = class_handle2object(handle->cookie, med);
93 /* during dw/setattr replay the mfd can be found by old handle */
94 if (mfd == NULL && is_replay_or_resent) {
95 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
96 if (mfd->mfd_old_handle.cookie == handle->cookie)
106 void mdt_mfd_free(struct mdt_file_data *mfd)
108 LASSERT(list_empty(&mfd->mfd_list));
109 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
112 static int mdt_create_data(struct mdt_thread_info *info,
113 struct mdt_object *p, struct mdt_object *o)
115 struct md_op_spec *spec = &info->mti_spec;
116 struct md_attr *ma = &info->mti_attr;
120 if (!md_should_create(spec->sp_cr_flags))
123 ma->ma_need = MA_INODE | MA_LOV;
125 mutex_lock(&o->mot_lov_mutex);
126 if (!(o->mot_flags & MOF_LOV_CREATED)) {
127 rc = mdo_create_data(info->mti_env,
128 p ? mdt_object_child(p) : NULL,
129 mdt_object_child(o), spec, ma);
131 rc = mdt_attr_get_complex(info, o, ma);
133 if (rc == 0 && ma->ma_valid & MA_LOV)
134 o->mot_flags |= MOF_LOV_CREATED;
137 mutex_unlock(&o->mot_lov_mutex);
141 int mdt_write_read(struct mdt_object *o)
145 spin_lock(&o->mot_write_lock);
146 rc = o->mot_write_count;
147 spin_unlock(&o->mot_write_lock);
151 int mdt_write_get(struct mdt_object *o)
155 spin_lock(&o->mot_write_lock);
156 if (o->mot_write_count < 0)
159 o->mot_write_count++;
160 spin_unlock(&o->mot_write_lock);
165 void mdt_write_put(struct mdt_object *o)
168 spin_lock(&o->mot_write_lock);
169 o->mot_write_count--;
170 spin_unlock(&o->mot_write_lock);
174 static int mdt_write_deny(struct mdt_object *o)
178 spin_lock(&o->mot_write_lock);
179 if (o->mot_write_count > 0)
182 o->mot_write_count--;
183 spin_unlock(&o->mot_write_lock);
187 static void mdt_write_allow(struct mdt_object *o)
190 spin_lock(&o->mot_write_lock);
191 o->mot_write_count++;
192 spin_unlock(&o->mot_write_lock);
196 /* there can be no real transaction so prepare the fake one */
197 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
199 struct mdt_device *mdt = info->mti_mdt;
200 struct ptlrpc_request *req = mdt_info_req(info);
201 struct tg_export_data *ted;
202 struct lsd_client_data *lcd;
205 /* transaction has occurred already */
206 if (lustre_msg_get_transno(req->rq_repmsg) != 0)
209 if (tgt_is_multimodrpcs_client(req->rq_export)) {
212 /* generate an empty transaction to get a transno
214 th = dt_trans_create(info->mti_env, mdt->mdt_bottom);
216 rc = dt_trans_start(info->mti_env, mdt->mdt_bottom, th);
217 dt_trans_stop(info->mti_env, mdt->mdt_bottom, th);
222 spin_lock(&mdt->mdt_lut.lut_translock);
224 if (info->mti_transno != 0) {
225 struct obd_export *exp = req->rq_export;
227 CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
228 mdt_obd_name(mdt), info->mti_transno,
229 libcfs_nid2str(exp->exp_connection->c_peer.nid),
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 = "LPU64", last_committed = "LPU64"\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 = "LPU64" < last_transno = "LPU64"\n",
263 mdt_obd_name(mdt), info->mti_transno,
264 lcd->lcd_last_transno);
268 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
269 if (info->mti_transno != 0)
270 lcd->lcd_last_close_transno = info->mti_transno;
271 lcd->lcd_last_close_xid = req->rq_xid;
272 lcd->lcd_last_close_result = rc;
274 /* VBR: save versions in last_rcvd for reconstruct. */
275 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
277 lcd->lcd_pre_versions[0] = pre_versions[0];
278 lcd->lcd_pre_versions[1] = pre_versions[1];
279 lcd->lcd_pre_versions[2] = pre_versions[2];
280 lcd->lcd_pre_versions[3] = pre_versions[3];
282 if (info->mti_transno != 0)
283 lcd->lcd_last_transno = info->mti_transno;
285 lcd->lcd_last_xid = req->rq_xid;
286 lcd->lcd_last_result = rc;
287 lcd->lcd_last_data = info->mti_opdata;
289 mutex_unlock(&ted->ted_lcd_lock);
294 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
296 LASSERT(mfd != NULL);
298 CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n",
299 PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
301 mfd->mfd_mode = mode;
305 * prep ma_lmm/ma_lmv for md_attr from reply
307 static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
308 struct mdt_object *obj,
311 LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
312 if (S_ISDIR(obj->mot_header.loh_attr)) {
313 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
315 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
318 if (ma->ma_lmv_size > 0)
319 ma->ma_need |= MA_LMV;
321 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
323 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
326 if (ma->ma_lmm_size > 0)
327 ma->ma_need |= MA_LOV;
331 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
332 struct mdt_object *o, __u64 flags, int created,
333 struct ldlm_reply *rep)
335 struct ptlrpc_request *req = mdt_info_req(info);
336 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
337 struct mdt_file_data *mfd;
338 struct md_attr *ma = &info->mti_attr;
339 struct lu_attr *la = &ma->ma_attr;
340 struct mdt_body *repbody;
341 int rc = 0, isdir, isreg;
344 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
346 isreg = S_ISREG(la->la_mode);
347 isdir = S_ISDIR(la->la_mode);
348 if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
350 * No EA, check whether it is will set regEA and dirEA since in
351 * above attr get, these size might be zero, so reset it, to
352 * retrieve the MD after create obj.
354 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
357 /* in replay case, p == NULL */
358 rc = mdt_create_data(info, p, o);
362 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
363 mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
366 CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
367 ma->ma_valid, ma->ma_lmm_size);
369 if (ma->ma_valid & MA_LOV) {
370 LASSERT(ma->ma_lmm_size != 0);
371 repbody->mbo_eadatasize = ma->ma_lmm_size;
373 repbody->mbo_valid |= OBD_MD_FLDIREA;
375 repbody->mbo_valid |= OBD_MD_FLEASIZE;
378 if (ma->ma_valid & MA_LMV) {
379 LASSERT(ma->ma_lmv_size != 0);
380 repbody->mbo_eadatasize = ma->ma_lmv_size;
382 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
385 if (flags & FMODE_WRITE)
386 rc = mdt_write_get(o);
387 else if (flags & MDS_FMODE_EXEC)
388 rc = mdt_write_deny(o);
393 rc = mo_open(info->mti_env, mdt_object_child(o),
394 created ? flags | MDS_OPEN_CREATED : flags);
396 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
397 * client does not know which suppgid should be sent to the MDS,
398 * or some other(s) changed the target file's GID after this RPC
399 * sent to the MDS with the suppgid as the original GID, then we
400 * give the client another chance to send the right suppgid. */
402 allow_client_chgrp(info, lu_ucred(info->mti_env)))
403 mdt_set_disposition(info, rep, DISP_OPEN_DENY);
408 mfd = mdt_mfd_new(med);
410 GOTO(err_out, rc = -ENOMEM);
413 * Keep a reference on this object for this open, and is
414 * released by mdt_mfd_close().
416 mdt_object_get(info->mti_env, o);
418 mfd->mfd_xid = req->rq_xid;
421 * @flags is always not zero. At least it should be FMODE_READ,
422 * FMODE_WRITE or MDS_FMODE_EXEC.
427 mdt_mfd_set_mode(mfd, flags);
429 atomic_inc(&o->mot_open_count);
430 if (flags & MDS_OPEN_LEASE)
431 atomic_inc(&o->mot_lease_count);
434 if (req_is_replay(req)) {
435 struct mdt_file_data *old_mfd;
436 /* Check wheather old cookie already exist in
437 * the list, becasue when do recovery, client
438 * might be disconnected from server, and
439 * restart replay, so there maybe some orphan
440 * mfd here, we should remove them */
441 LASSERT(info->mti_rr.rr_handle != NULL);
442 spin_lock(&med->med_open_lock);
443 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
444 if (old_mfd != NULL) {
445 CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
446 "cookie = "LPX64"\n", mfd,
447 PFID(mdt_object_fid(mfd->mfd_object)),
448 info->mti_rr.rr_handle->cookie);
449 class_handle_unhash(&old_mfd->mfd_handle);
450 list_del_init(&old_mfd->mfd_list);
451 spin_unlock(&med->med_open_lock);
452 /* no attr update for that close */
454 ma->ma_valid |= MA_FLAGS;
455 ma->ma_attr_flags |= MDS_RECOV_OPEN;
456 mdt_mfd_close(info, old_mfd);
457 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
458 ma->ma_valid &= ~MA_FLAGS;
460 spin_unlock(&med->med_open_lock);
461 CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
462 "cookie = "LPX64"\n",
463 PFID(mdt_object_fid(mfd->mfd_object)),
464 info->mti_rr.rr_handle->cookie);
467 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
468 info->mti_rr.rr_handle->cookie);
470 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
473 repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
475 if (req->rq_export->exp_disconnected) {
476 spin_lock(&med->med_open_lock);
477 class_handle_unhash(&mfd->mfd_handle);
478 list_del_init(&mfd->mfd_list);
479 spin_unlock(&med->med_open_lock);
480 mdt_mfd_close(info, mfd);
482 spin_lock(&med->med_open_lock);
483 list_add(&mfd->mfd_list, &med->med_open_head);
484 spin_unlock(&med->med_open_lock);
487 mdt_empty_transno(info, rc);
492 if (flags & FMODE_WRITE)
494 else if (flags & MDS_FMODE_EXEC)
500 static int mdt_finish_open(struct mdt_thread_info *info,
501 struct mdt_object *p, struct mdt_object *o,
502 __u64 flags, int created, struct ldlm_reply *rep)
504 struct ptlrpc_request *req = mdt_info_req(info);
505 struct obd_export *exp = req->rq_export;
506 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
507 struct md_attr *ma = &info->mti_attr;
508 struct lu_attr *la = &ma->ma_attr;
509 struct mdt_file_data *mfd;
510 struct mdt_body *repbody;
512 int isreg, isdir, islnk;
516 LASSERT(ma->ma_valid & MA_INODE);
518 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
520 isreg = S_ISREG(la->la_mode);
521 isdir = S_ISDIR(la->la_mode);
522 islnk = S_ISLNK(la->la_mode);
523 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
525 /* LU-2275, simulate broken behaviour (esp. prevalent in
526 * pre-2.4 servers where a very strange reply is sent on error
527 * that looks like it was actually almost successful and a
528 * failure at the same time.) */
529 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
530 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
534 if (flags & MDS_OPEN_LOCK)
535 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
540 if (exp_connect_rmtclient(exp)) {
541 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
543 rc = mdt_pack_remote_perm(info, o, buf);
545 repbody->mbo_valid &= ~OBD_MD_FLRMTPERM;
546 repbody->mbo_aclsize = 0;
548 repbody->mbo_valid |= OBD_MD_FLRMTPERM;
549 repbody->mbo_aclsize = sizeof(struct mdt_remote_perm);
552 #ifdef CONFIG_FS_POSIX_ACL
553 else if (exp_connect_flags(exp) & OBD_CONNECT_ACL)
554 rc = mdt_pack_acl2body(info, repbody, o,
555 exp->exp_target_data.ted_nodemap);
559 * If we are following a symlink, don't open; and do not return open
560 * handle for special nodes as client required.
562 if (islnk || (!isreg && !isdir &&
563 (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
564 lustre_msg_set_transno(req->rq_repmsg, 0);
569 * We need to return the existing object's fid back, so it is done here,
570 * after preparing the reply.
572 if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
575 /* This can't be done earlier, we need to return reply body */
577 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
578 /* We are trying to create or write an existing dir. */
581 } else if (flags & MDS_OPEN_DIRECTORY)
584 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
585 OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
590 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
591 spin_lock(&med->med_open_lock);
592 list_for_each(t, &med->med_open_head) {
593 mfd = list_entry(t, struct mdt_file_data, mfd_list);
594 if (mfd->mfd_xid == req->rq_xid)
598 spin_unlock(&med->med_open_lock);
601 repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
602 /* set repbody->ea_size for resent case */
603 if (ma->ma_valid & MA_LOV) {
604 LASSERT(ma->ma_lmm_size != 0);
605 repbody->mbo_eadatasize = ma->ma_lmm_size;
607 repbody->mbo_valid |= OBD_MD_FLDIREA;
609 repbody->mbo_valid |= OBD_MD_FLEASIZE;
611 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
616 rc = mdt_mfd_open(info, p, o, flags, created, rep);
618 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
623 void mdt_reconstruct_open(struct mdt_thread_info *info,
624 struct mdt_lock_handle *lhc)
626 const struct lu_env *env = info->mti_env;
627 struct mdt_device *mdt = info->mti_mdt;
628 struct req_capsule *pill = info->mti_pill;
629 struct ptlrpc_request *req = mdt_info_req(info);
630 struct md_attr *ma = &info->mti_attr;
631 struct mdt_reint_record *rr = &info->mti_rr;
632 __u64 flags = info->mti_spec.sp_cr_flags;
633 struct ldlm_reply *ldlm_rep;
634 struct mdt_object *parent;
635 struct mdt_object *child;
636 struct mdt_body *repbody;
641 LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
642 ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
643 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
645 ma->ma_need = MA_INODE | MA_HSM;
648 opdata = mdt_req_from_lrd(req, info->mti_reply_data);
649 mdt_set_disposition(info, ldlm_rep, opdata);
651 CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
652 ldlm_rep->lock_policy_res1, req->rq_status);
654 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
656 /* We did not create successfully, return error to client. */
657 GOTO(out, rc = req->rq_status);
659 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
660 struct obd_export *exp = req->rq_export;
662 * We failed after creation, but we do not know in which step
663 * we failed. So try to check the child object.
665 parent = mdt_object_find(env, mdt, rr->rr_fid1);
666 if (IS_ERR(parent)) {
667 rc = PTR_ERR(parent);
668 LCONSOLE_WARN("Parent "DFID" lookup error %d."
669 " Evicting client %s with export %s.\n",
670 PFID(rr->rr_fid1), rc,
671 obd_uuid2str(&exp->exp_client_uuid),
672 obd_export_nid2str(exp));
673 mdt_export_evict(exp);
677 child = mdt_object_find(env, mdt, rr->rr_fid2);
680 LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
681 "evicting client %s with export %s\n",
682 PFID(rr->rr_fid2), rc,
683 obd_uuid2str(&exp->exp_client_uuid),
684 obd_export_nid2str(exp));
685 mdt_object_put(env, parent);
686 mdt_export_evict(exp);
690 if (unlikely(mdt_object_remote(child))) {
691 /* the child object was created on remote server */
692 if (!mdt_is_dne_client(exp)) {
693 /* Return -EIO for old client */
694 mdt_object_put(env, parent);
695 mdt_object_put(env, child);
696 GOTO(out, rc = -EIO);
698 repbody->mbo_fid1 = *rr->rr_fid2;
699 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
702 if (mdt_object_exists(child)) {
703 mdt_prep_ma_buf_from_rep(info, child, ma);
704 rc = mdt_attr_get_complex(info, child, ma);
706 rc = mdt_finish_open(info, parent,
710 /* the child does not exist, we should do
712 mdt_object_put(env, parent);
713 mdt_object_put(env, child);
714 GOTO(regular_open, 0);
717 mdt_object_put(env, parent);
718 mdt_object_put(env, child);
722 /* We did not try to create, so we are a pure open */
723 rc = mdt_reint_open(info, lhc);
729 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
730 LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
733 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
735 __u64 flags = info->mti_spec.sp_cr_flags;
736 struct mdt_reint_record *rr = &info->mti_rr;
737 struct md_attr *ma = &info->mti_attr;
738 struct mdt_object *o;
742 o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
744 RETURN(rc = PTR_ERR(o));
746 if (unlikely(mdt_object_remote(o))) {
747 /* the child object was created on remote server */
748 struct mdt_body *repbody;
750 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
753 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
754 repbody->mbo_fid1 = *rr->rr_fid2;
755 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
758 if (mdt_object_exists(o)) {
759 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
762 mdt_prep_ma_buf_from_rep(info, o, ma);
763 rc = mdt_attr_get_complex(info, o, ma);
765 rc = mdt_finish_open(info, NULL, o, flags, 0,
772 mdt_object_put(info->mti_env, o);
776 /* lock object for open */
777 static int mdt_object_open_lock(struct mdt_thread_info *info,
778 struct mdt_object *obj,
779 struct mdt_lock_handle *lhc,
782 struct md_attr *ma = &info->mti_attr;
783 __u64 open_flags = info->mti_spec.sp_cr_flags;
784 enum ldlm_mode lm = LCK_CR;
785 bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
786 bool try_layout = false;
787 bool create_layout = false;
792 mdt_lock_handle_init(lhc);
794 if (req_is_replay(mdt_info_req(info)))
797 if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
798 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
799 md_should_create(open_flags))
800 create_layout = true;
801 if (exp_connect_layout(info->mti_exp) && !create_layout &&
802 ma->ma_need & MA_LOV)
807 /* lease open, acquire write mode of open sem */
808 down_write(&obj->mot_open_sem);
810 /* Lease exists and ask for new lease */
811 if (atomic_read(&obj->mot_lease_count) > 0) {
812 /* only exclusive open is supported, so lease
813 * are conflicted to each other */
814 GOTO(out, rc = -EBUSY);
817 /* Lease must be with open lock */
818 if (!(open_flags & MDS_OPEN_LOCK)) {
819 CERROR("Request lease for file:"DFID ", but open lock "
820 "is missed, open_flags = "LPO64".\n",
821 PFID(mdt_object_fid(obj)), open_flags);
822 GOTO(out, rc = -EPROTO);
825 /* XXX: only exclusive open is supported. */
827 *ibits = MDS_INODELOCK_OPEN;
829 /* never grant LCK_EX layout lock to client */
831 } else { /* normal open */
832 /* normal open holds read mode of open sem */
833 down_read(&obj->mot_open_sem);
835 if (open_flags & MDS_OPEN_LOCK) {
836 if (open_flags & FMODE_WRITE)
838 else if (open_flags & MDS_FMODE_EXEC)
843 *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
844 } else if (atomic_read(&obj->mot_lease_count) > 0) {
845 if (open_flags & FMODE_WRITE)
851 *ibits = MDS_INODELOCK_OPEN;
854 lhc = &info->mti_lh[MDT_LH_LOCAL];
856 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
857 PFID(mdt_object_fid(obj)),
858 atomic_read(&obj->mot_open_count), lm);
861 mdt_lock_reg_init(lhc, lm);
863 /* one problem to return layout lock on open is that it may result
864 * in too many layout locks cached on the client side. */
865 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
866 /* return lookup lock to validate inode at the client side,
867 * this is pretty important otherwise mdt will return layout
868 * lock for each open.
869 * However this is a double-edged sword because changing
870 * permission will revoke huge # of LOOKUP locks. */
871 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
872 if (!mdt_object_lock_try(info, obj, lhc, *ibits)) {
873 *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
875 rc = mdt_object_lock(info, obj, lhc, *ibits);
877 } else if (*ibits != 0) {
878 rc = mdt_object_lock(info, obj, lhc, *ibits);
881 CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
882 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
883 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
885 /* will change layout, revoke layout locks by enqueuing EX lock. */
886 if (rc == 0 && create_layout) {
887 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
889 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
890 ", open_flags = "LPO64"\n",
891 PFID(mdt_object_fid(obj)), open_flags);
893 /* We cannot enqueue another lock for the same resource we
894 * already have a lock for, due to mechanics of waiting list
895 * iterating in ldlm, see LU-3601.
896 * As such we'll drop the open lock we just got above here,
897 * it's ok not to have this open lock as it's main purpose is to
898 * flush unused cached client open handles. */
899 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
900 mdt_object_unlock(info, obj, lhc, 1);
902 LASSERT(!try_layout);
903 mdt_lock_handle_init(ll);
904 mdt_lock_reg_init(ll, LCK_EX);
905 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
907 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
910 /* Check if there is any other open handles after acquiring
911 * open lock. At this point, caching open handles have been revoked
913 * XXX: Now only exclusive open is supported. Need to check the
914 * type of open for generic lease support. */
915 if (rc == 0 && acq_lease) {
916 struct ptlrpc_request *req = mdt_info_req(info);
917 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
918 struct mdt_file_data *mfd;
919 bool is_replay_or_resent;
922 /* For lease: application can open a file and then apply lease,
923 * @handle contains original open handle in that case.
924 * In recovery, open REQ will be replayed and the lease REQ may
925 * be resent that means the open handle is already stale, so we
926 * need to fix it up here by finding new handle. */
927 is_replay_or_resent = req_is_replay(req) ||
928 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
930 /* if the request is _not_ a replay request, rr_handle
931 * may be used to hold an openhandle which is issuing the
932 * lease request, so that this openhandle doesn't count. */
933 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
934 is_replay_or_resent);
938 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
939 PFID(mdt_object_fid(obj)),
940 atomic_read(&obj->mot_open_count), open_count);
942 if (atomic_read(&obj->mot_open_count) > open_count)
943 GOTO(out, rc = -EBUSY);
951 static void mdt_object_open_unlock(struct mdt_thread_info *info,
952 struct mdt_object *obj,
953 struct mdt_lock_handle *lhc,
956 __u64 open_flags = info->mti_spec.sp_cr_flags;
957 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
960 if (req_is_replay(mdt_info_req(info)))
963 /* Release local lock - the lock put in MDT_LH_LOCAL will never
964 * return to client side. */
965 if (lustre_handle_is_used(&ll->mlh_reg_lh))
966 mdt_object_unlock(info, obj, ll, 1);
968 ll = &info->mti_lh[MDT_LH_LAYOUT];
969 /* Release local layout lock, layout was created */
970 if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
971 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
972 mdt_object_unlock(info, obj, ll, 1);
975 if (open_flags & MDS_OPEN_LEASE)
976 up_write(&obj->mot_open_sem);
978 up_read(&obj->mot_open_sem);
980 /* Cross-ref case, the lock should be returned to the client */
981 if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
984 if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
985 /* for the open request, the lock will only return to client
986 * if open or layout lock is granted. */
990 if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
991 struct ldlm_reply *ldlm_rep;
993 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
994 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
995 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
996 mdt_object_unlock(info, obj, lhc, 1);
1002 * Check release is permitted for the current HSM flags.
1004 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1006 if (!(ma->ma_valid & MA_HSM))
1009 if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1012 if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1018 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1019 struct ldlm_reply *rep,
1020 struct mdt_lock_handle *lhc)
1022 const struct lu_env *env = info->mti_env;
1023 struct mdt_device *mdt = info->mti_mdt;
1024 __u64 flags = info->mti_spec.sp_cr_flags;
1025 struct mdt_reint_record *rr = &info->mti_rr;
1026 struct md_attr *ma = &info->mti_attr;
1027 struct mdt_object *parent= NULL;
1028 struct mdt_object *o;
1030 bool object_locked = false;
1034 if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1035 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1036 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1037 if (IS_ERR(parent)) {
1038 CDEBUG(D_INODE, "Fail to find parent "DFID
1039 " for anonymous created %ld, try to"
1040 " use server-side parent.\n",
1041 PFID(rr->rr_fid1), PTR_ERR(parent));
1046 ma->ma_need |= MA_PFID;
1049 o = mdt_object_find(env, mdt, rr->rr_fid2);
1051 GOTO(out_parent_put, rc = PTR_ERR(o));
1053 if (mdt_object_remote(o)) {
1054 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1055 mdt_obd_name(info->mti_mdt),
1057 GOTO(out, rc = -EREMOTE);
1058 } else if (!mdt_object_exists(o)) {
1059 mdt_set_disposition(info, rep,
1063 GOTO(out, rc = -ENOENT);
1066 mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1068 mdt_prep_ma_buf_from_rep(info, o, ma);
1069 if (flags & MDS_OPEN_RELEASE)
1070 ma->ma_need |= MA_HSM;
1071 rc = mdt_attr_get_complex(info, o, ma);
1075 /* We should not change file's existing LOV EA */
1076 if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1077 flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1078 GOTO(out, rc = -EEXIST);
1080 /* If a release request, check file flags are fine and ask for an
1081 * exclusive open access. */
1082 if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1083 GOTO(out, rc = -EPERM);
1085 rc = mdt_check_resent_lock(info, o, lhc);
1088 } else if (rc > 0) {
1089 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1090 object_locked = true;
1092 GOTO(out_unlock, rc);
1095 if (ma->ma_valid & MA_PFID) {
1096 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1097 if (IS_ERR(parent)) {
1098 CDEBUG(D_INODE, "Fail to find parent "DFID
1099 " for anonymous created %ld, try to"
1100 " use system default.\n",
1101 PFID(&ma->ma_pfid), PTR_ERR(parent));
1106 rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1108 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1109 if (flags & MDS_OPEN_LOCK)
1110 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1111 if (flags & MDS_OPEN_LEASE)
1112 mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1114 GOTO(out_unlock, rc);
1118 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1120 mdt_object_put(env, o);
1123 mdt_object_put(env, parent);
1127 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1128 static int mdt_cross_open(struct mdt_thread_info *info,
1129 const struct lu_fid *parent_fid,
1130 const struct lu_fid *fid,
1131 struct ldlm_reply *rep, __u32 flags)
1133 struct md_attr *ma = &info->mti_attr;
1134 struct mdt_object *o;
1138 o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1140 RETURN(rc = PTR_ERR(o));
1142 if (mdt_object_remote(o)) {
1143 /* Something is wrong here, the object is on another MDS! */
1144 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1145 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1146 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1148 "Object isn't on this server! FLD error?\n");
1151 if (mdt_object_exists(o)) {
1152 /* Do permission check for cross-open. */
1153 rc = mo_permission(info->mti_env, NULL,
1154 mdt_object_child(o),
1155 NULL, flags | MDS_OPEN_CROSS);
1159 mdt_prep_ma_buf_from_rep(info, o, ma);
1160 rc = mdt_attr_get_complex(info, o, ma);
1164 rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1167 * Something is wrong here. lookup was positive but
1168 * there is no object!
1170 CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1171 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1176 mdt_object_put(info->mti_env, o);
1180 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1182 struct mdt_device *mdt = info->mti_mdt;
1183 struct ptlrpc_request *req = mdt_info_req(info);
1184 struct mdt_object *parent;
1185 struct mdt_object *child;
1186 struct mdt_lock_handle *lh;
1187 struct ldlm_reply *ldlm_rep;
1188 struct mdt_body *repbody;
1189 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1190 struct md_attr *ma = &info->mti_attr;
1191 __u64 create_flags = info->mti_spec.sp_cr_flags;
1193 struct mdt_reint_record *rr = &info->mti_rr;
1196 int object_locked = 0;
1200 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1201 (obd_timeout + 1) / 4);
1203 mdt_counter_incr(req, LPROC_MDT_OPEN);
1204 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1206 ma->ma_need = MA_INODE;
1209 LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1210 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1212 if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1213 CERROR("file join is not supported anymore.\n");
1214 GOTO(out, result = err_serious(-EOPNOTSUPP));
1216 msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1218 if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1219 info->mti_spec.u.sp_ea.eadata == NULL)
1220 GOTO(out, result = err_serious(-EINVAL));
1222 CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1223 "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1224 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1225 PFID(rr->rr_fid2), create_flags,
1226 ma->ma_attr.la_mode, msg_flags);
1228 if (info->mti_cross_ref) {
1229 /* This is cross-ref open */
1230 mdt_set_disposition(info, ldlm_rep,
1231 (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1233 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1234 ldlm_rep, create_flags);
1236 } else if (req_is_replay(req) ||
1237 (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1238 /* This is a replay request or from liblustre with ea. */
1239 result = mdt_open_by_fid(info, ldlm_rep);
1241 if (result != -ENOENT) {
1242 if (req->rq_export->exp_libclient &&
1243 create_flags & MDS_OPEN_HAS_EA)
1244 GOTO(out, result = 0);
1247 /* We didn't find the correct object, so we need to re-create it
1248 * via a regular replay. */
1249 if (!(create_flags & MDS_OPEN_CREAT)) {
1250 DEBUG_REQ(D_ERROR, req,
1251 "OPEN & CREAT not in open replay/by_fid.");
1252 GOTO(out, result = -EFAULT);
1254 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1255 } else if (create_flags & (MDS_OPEN_BY_FID | MDS_OPEN_LOCK)) {
1257 * MDS_OPEN_LOCK is checked for backward compatibility with 2.1
1260 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1262 CDEBUG(D_INFO, "no object for "DFID": %d\n",
1263 PFID(rr->rr_fid2), result);
1267 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1268 GOTO(out, result = err_serious(-ENOMEM));
1270 mdt_set_disposition(info, ldlm_rep,
1271 (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1273 if (!lu_name_is_valid(&rr->rr_name))
1274 GOTO(out, result = -EPROTO);
1277 lh = &info->mti_lh[MDT_LH_PARENT];
1278 mdt_lock_pdo_init(lh,
1279 (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1282 parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1284 GOTO(out, result = PTR_ERR(parent));
1286 result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1288 mdt_object_put(info->mti_env, parent);
1292 /* get and check version of parent */
1293 result = mdt_version_get_check(info, parent, 0);
1295 GOTO(out_parent, result);
1297 fid_zero(child_fid);
1299 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1300 &rr->rr_name, child_fid, &info->mti_spec);
1302 LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1303 "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1304 PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1307 if (result != 0 && result != -ENOENT && result != -ESTALE)
1308 GOTO(out_parent, result);
1310 if (result == -ENOENT || result == -ESTALE) {
1311 /* If the object is dead, let's check if the object
1312 * is being migrated to a new object */
1313 if (result == -ESTALE) {
1314 struct lu_buf lmv_buf;
1316 lmv_buf.lb_buf = info->mti_xattr_buf;
1317 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
1318 rc = mo_xattr_get(info->mti_env,
1319 mdt_object_child(parent),
1320 &lmv_buf, XATTR_NAME_LMV);
1322 struct lmv_mds_md_v1 *lmv;
1324 lmv = lmv_buf.lb_buf;
1325 if (le32_to_cpu(lmv->lmv_hash_type) &
1326 LMV_HASH_FLAG_MIGRATION) {
1327 /* Get the new parent FID and retry */
1328 mdt_object_unlock_put(info, parent,
1330 mdt_lock_handle_init(lh);
1332 (struct lu_fid *)rr->rr_fid1,
1333 &lmv->lmv_stripe_fids[1]);
1339 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1340 if (result == -ESTALE) {
1342 * -ESTALE means the parent is a dead(unlinked) dir, so
1343 * it should return -ENOENT to in accordance with the
1344 * original mds implementaion.
1346 GOTO(out_parent, result = -ENOENT);
1349 if (!(create_flags & MDS_OPEN_CREAT))
1350 GOTO(out_parent, result);
1351 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1352 GOTO(out_parent, result = -EROFS);
1353 *child_fid = *info->mti_rr.rr_fid2;
1354 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1356 /* In the function below, .hs_keycmp resolves to
1357 * lu_obj_hop_keycmp() */
1358 /* coverity[overrun-buffer-val] */
1359 child = mdt_object_new(info->mti_env, mdt, child_fid);
1362 * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1363 * return FID back in that case.
1365 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1366 child = mdt_object_find(info->mti_env, mdt, child_fid);
1369 GOTO(out_parent, result = PTR_ERR(child));
1371 /** check version of child */
1372 rc = mdt_version_get_check(info, child, 1);
1374 GOTO(out_child, result = rc);
1376 if (result == -ENOENT) {
1377 /* Create under OBF and .lustre is not permitted */
1378 if (!fid_is_md_operative(rr->rr_fid1))
1379 GOTO(out_child, result = -EPERM);
1381 /* save versions in reply */
1382 mdt_version_get_save(info, parent, 0);
1383 mdt_version_get_save(info, child, 1);
1385 /* version of child will be changed */
1386 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1388 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1389 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1391 /* Let lower layers know what is lock mode on directory. */
1392 info->mti_spec.sp_cr_mode =
1393 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1396 * Do not perform lookup sanity check. We know that name does
1399 info->mti_spec.sp_cr_lookup = 0;
1400 info->mti_spec.sp_feat = &dt_directory_features;
1402 result = mdo_create(info->mti_env,
1403 mdt_object_child(parent),
1405 mdt_object_child(child),
1408 if (result == -ERESTART) {
1409 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1410 GOTO(out_child, result);
1412 mdt_prep_ma_buf_from_rep(info, child, ma);
1413 /* XXX: we should call this once, see few lines below */
1415 result = mdt_attr_get_complex(info, child, ma);
1418 GOTO(out_child, result);
1423 * The object is on remote node, return its FID for remote open.
1425 if (mdt_object_remote(child)) {
1427 * Check if this lock already was sent to client and
1428 * this is resent case. For resent case do not take lock
1429 * again, use what is already granted.
1431 LASSERT(lhc != NULL);
1433 rc = mdt_check_resent_lock(info, child, lhc);
1435 GOTO(out_child, result = rc);
1436 } else if (rc > 0) {
1437 mdt_lock_handle_init(lhc);
1438 mdt_lock_reg_init(lhc, LCK_PR);
1440 rc = mdt_object_lock(info, child, lhc,
1441 MDS_INODELOCK_LOOKUP);
1443 repbody->mbo_fid1 = *mdt_object_fid(child);
1444 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1448 result = -MDT_EREMOTE_OPEN;
1449 GOTO(out_child, result);
1450 } else if (mdt_object_exists(child)) {
1451 /* We have to get attr & LOV EA & HSM for this
1453 mdt_prep_ma_buf_from_rep(info, child, ma);
1454 ma->ma_need |= MA_HSM;
1455 result = mdt_attr_get_complex(info, child, ma);
1457 GOTO(out_child, result);
1459 /* Object does not exist. Likely FS corruption. */
1460 CERROR("%s: name '"DNAME"' present, but FID "
1461 DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1462 PNAME(&rr->rr_name), PFID(child_fid));
1463 GOTO(out_child, result = -EIO);
1467 rc = mdt_check_resent_lock(info, child, lhc);
1469 GOTO(out_child, result = rc);
1470 } else if (rc == 0) {
1471 /* the open lock might already be gotten in
1472 * ldlm_handle_enqueue() */
1473 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1474 if (create_flags & MDS_OPEN_LOCK)
1475 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1477 /* get openlock if this isn't replay and client requested it */
1478 if (!req_is_replay(req)) {
1479 rc = mdt_object_open_lock(info, child, lhc, &ibits);
1482 GOTO(out_child_unlock, result = rc);
1483 else if (create_flags & MDS_OPEN_LOCK)
1484 mdt_set_disposition(info, ldlm_rep,
1488 /* Try to open it now. */
1489 rc = mdt_finish_open(info, parent, child, create_flags,
1493 /* openlock will be released if mdt_finish_open failed */
1494 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1496 if (created && create_flags & MDS_OPEN_VOLATILE) {
1497 CERROR("%s: cannot open volatile file "DFID", orphan "
1498 "file will be left in PENDING directory until "
1499 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1500 PFID(mdt_object_fid(child)), rc);
1501 GOTO(out_child_unlock, result);
1507 rc = mdo_unlink(info->mti_env,
1508 mdt_object_child(parent),
1509 mdt_object_child(child),
1511 &info->mti_attr, 0);
1513 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1514 mdt_obd_name(info->mti_mdt),
1515 PFID(mdt_object_fid(child)), rc);
1516 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1522 mdt_object_open_unlock(info, child, lhc, ibits, result);
1524 mdt_object_put(info->mti_env, child);
1526 mdt_object_unlock_put(info, parent, lh, result || !created);
1529 lustre_msg_set_transno(req->rq_repmsg, 0);
1534 * Create an orphan object use local root.
1536 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1537 struct mdt_device *mdt,
1538 const struct lu_fid *fid,
1539 struct md_attr *attr, fmode_t fmode)
1541 const struct lu_env *env = info->mti_env;
1542 struct md_op_spec *spec = &info->mti_spec;
1543 struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1544 struct mdt_object *obj = NULL;
1545 struct mdt_object *local_root;
1546 static const struct lu_name lname = {
1547 .ln_name = "i_am_nobody",
1548 .ln_namelen = sizeof("i_am_nobody") - 1,
1550 struct lu_ucred *uc;
1551 cfs_cap_t uc_cap_save;
1555 rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1557 RETURN(ERR_PTR(rc));
1559 local_root = mdt_object_find(env, mdt, local_root_fid);
1560 if (IS_ERR(local_root))
1563 obj = mdt_object_new(env, mdt, fid);
1565 GOTO(out, rc = PTR_ERR(obj));
1567 spec->sp_cr_lookup = 0;
1568 spec->sp_feat = &dt_directory_features;
1569 spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1570 spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1571 if (attr->ma_valid & MA_LOV) {
1572 spec->u.sp_ea.eadata = attr->ma_lmm;
1573 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1574 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1576 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1580 uc_cap_save = uc->uc_cap;
1581 uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1582 rc = mdo_create(env, mdt_object_child(local_root), &lname,
1583 mdt_object_child(obj), spec, attr);
1584 uc->uc_cap = uc_cap_save;
1586 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1587 mdt_obd_name(mdt), PFID(fid), rc);
1591 rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1593 CERROR("%s: cannot open volatile file "DFID", orphan "
1594 "file will be left in PENDING directory until "
1595 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1602 mdt_object_put(env, obj);
1605 mdt_object_put(env, local_root);
1609 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1612 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1613 struct close_data *data;
1614 struct ldlm_lock *lease;
1615 struct mdt_object *orphan;
1616 struct md_attr *orp_ma;
1623 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1626 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1630 lease = ldlm_handle2lock(&data->cd_handle);
1634 /* try to hold open_sem so that nobody else can open the file */
1635 if (!down_write_trylock(&o->mot_open_sem)) {
1636 ldlm_lock_cancel(lease);
1637 GOTO(out_reprocess, rc = -EBUSY);
1640 /* Check if the lease open lease has already canceled */
1641 lock_res_and_lock(lease);
1642 lease_broken = ldlm_is_cancel(lease);
1643 unlock_res_and_lock(lease);
1645 LDLM_DEBUG(lease, DFID " lease broken? %d\n",
1646 PFID(mdt_object_fid(o)), lease_broken);
1648 /* Cancel server side lease. Client side counterpart should
1649 * have been cancelled. It's okay to cancel it now as we've
1650 * held mot_open_sem. */
1651 ldlm_lock_cancel(lease);
1653 if (lease_broken) /* don't perform release task */
1654 GOTO(out_unlock, rc = -ESTALE);
1656 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1657 GOTO(out_unlock, rc = -EINVAL);
1659 /* ma_need was set before but it seems fine to change it in order to
1660 * avoid modifying the one from RPC */
1661 ma->ma_need = MA_HSM;
1662 rc = mdt_attr_get_complex(info, o, ma);
1664 GOTO(out_unlock, rc);
1666 if (!mdt_hsm_release_allow(ma))
1667 GOTO(out_unlock, rc = -EPERM);
1669 /* already released? */
1670 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1671 GOTO(out_unlock, rc = 0);
1673 /* Compare on-disk and packed data_version */
1674 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1675 CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
1676 " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
1677 data->cd_data_version, ma->ma_hsm.mh_arch_ver);
1678 GOTO(out_unlock, rc = -EPERM);
1681 ma->ma_valid = MA_INODE;
1682 ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1683 rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1685 GOTO(out_unlock, rc);
1687 ma->ma_need = MA_INODE | MA_LOV;
1688 rc = mdt_attr_get_complex(info, o, ma);
1690 GOTO(out_unlock, rc);
1692 if (!(ma->ma_valid & MA_LOV)) {
1693 /* Even empty file are released */
1694 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1695 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
1696 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1697 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1698 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1700 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
1701 * ma_lmm as lov_user_md, then it will be confused by union of
1702 * layout_gen and stripe_offset. */
1703 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
1704 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
1705 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
1706 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
1708 GOTO(out_unlock, rc = -EINVAL);
1711 /* Set file as released */
1712 ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1714 orp_ma = &info->mti_u.hsm.attr;
1715 orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1716 orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
1717 orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
1718 orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1719 orp_ma->ma_lmm = ma->ma_lmm;
1720 orp_ma->ma_lmm_size = ma->ma_lmm_size;
1721 orp_ma->ma_valid = MA_INODE | MA_LOV;
1722 orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1724 if (IS_ERR(orphan)) {
1725 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1726 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1728 GOTO(out_unlock, rc = PTR_ERR(orphan));
1731 /* Set up HSM attribute for orphan object */
1732 CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
1733 buf = &info->mti_buf;
1734 buf->lb_buf = info->mti_xattr_buf;
1735 buf->lb_len = sizeof(struct hsm_attrs);
1736 ma->ma_hsm.mh_flags |= HS_RELEASED;
1737 lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1738 ma->ma_hsm.mh_flags &= ~HS_RELEASED;
1740 mdt_lock_reg_init(lh, LCK_EX);
1741 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
1742 MDS_INODELOCK_XATTR);
1744 GOTO(out_close, rc);
1746 rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
1750 /* Swap layout with orphan object */
1751 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
1752 mdt_object_child(orphan),
1753 SWAP_LAYOUTS_MDS_HSM);
1755 /* Release exclusive LL */
1756 mdt_object_unlock(info, o, lh, 1);
1761 /* Close orphan object anyway */
1762 rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
1765 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
1766 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
1767 LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
1769 mdt_object_put(info->mti_env, orphan);
1772 up_write(&o->mot_open_sem);
1774 /* already released */
1776 struct mdt_body *repbody;
1778 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1779 LASSERT(repbody != NULL);
1780 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1784 ldlm_reprocess_all(lease->l_resource);
1785 LDLM_LOCK_PUT(lease);
1793 int mdt_close_swap_layouts(struct mdt_thread_info *info,
1794 struct mdt_object *o, struct md_attr *ma)
1796 struct mdt_lock_handle *lh1 = &info->mti_lh[MDT_LH_NEW];
1797 struct mdt_lock_handle *lh2 = &info->mti_lh[MDT_LH_OLD];
1798 struct close_data *data;
1799 struct ldlm_lock *lease;
1800 struct mdt_object *o1 = o, *o2;
1806 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1809 if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
1812 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1816 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1819 rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
1820 if (unlikely(rc == 0))
1823 /* Exchange o1 and o2, to enforce locking order */
1824 swap_objects = (rc < 0);
1826 lease = ldlm_handle2lock(&data->cd_handle);
1830 o2 = mdt_object_find(info->mti_env, info->mti_mdt, &data->cd_fid);
1832 GOTO(out_lease, rc = PTR_ERR(o2));
1834 if (!S_ISREG(lu_object_attr(&o2->mot_obj))) {
1835 swap_objects = false; /* not swapped yet */
1836 GOTO(out_obj, rc = -EINVAL);
1842 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1847 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1852 /* try to hold open_sem so that nobody else can open the file */
1853 if (!down_write_trylock(&o->mot_open_sem)) {
1854 ldlm_lock_cancel(lease);
1855 GOTO(out_obj, rc = -EBUSY);
1858 /* Check if the lease open lease has already canceled */
1859 lock_res_and_lock(lease);
1860 lease_broken = ldlm_is_cancel(lease);
1861 unlock_res_and_lock(lease);
1863 LDLM_DEBUG(lease, DFID " lease broken? %d\n",
1864 PFID(mdt_object_fid(o)), lease_broken);
1866 /* Cancel server side lease. Client side counterpart should
1867 * have been cancelled. It's okay to cancel it now as we've
1868 * held mot_open_sem. */
1869 ldlm_lock_cancel(lease);
1872 GOTO(out_unlock_sem, rc = -ESTALE);
1874 mdt_lock_reg_init(lh1, LCK_EX);
1875 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1876 MDS_INODELOCK_XATTR);
1878 GOTO(out_unlock_sem, rc);
1880 mdt_lock_reg_init(lh2, LCK_EX);
1881 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1882 MDS_INODELOCK_XATTR);
1884 GOTO(out_unlock1, rc);
1886 /* Swap layout with orphan object */
1887 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1888 mdt_object_child(o2), 0);
1890 GOTO(out_unlock2, rc);
1895 /* Release exclusive LL */
1896 mdt_object_unlock(info, o2, lh2, 1);
1899 mdt_object_unlock(info, o1, lh1, 1);
1902 up_write(&o->mot_open_sem);
1904 /* already swapped */
1906 struct mdt_body *repbody;
1908 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1909 LASSERT(repbody != NULL);
1910 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1914 mdt_object_put(info->mti_env, swap_objects ? o1 : o2);
1916 ldlm_reprocess_all(lease->l_resource);
1919 LDLM_LOCK_PUT(lease);
1929 #define MFD_CLOSED(mode) ((mode) == MDS_FMODE_CLOSED)
1930 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1932 return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1935 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1937 struct mdt_object *o = mfd->mfd_object;
1938 struct md_object *next = mdt_object_child(o);
1939 struct md_attr *ma = &info->mti_attr;
1944 mode = mfd->mfd_mode;
1946 if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
1947 rc = mdt_hsm_release(info, o, ma);
1949 CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
1950 mdt_obd_name(info->mti_mdt),
1951 PFID(mdt_object_fid(o)), rc);
1952 /* continue to close even error occurred. */
1956 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
1957 rc = mdt_close_swap_layouts(info, o, ma);
1960 "%s: cannot swap layout of "DFID": rc=%d\n",
1961 mdt_obd_name(info->mti_mdt),
1962 PFID(mdt_object_fid(o)), rc);
1963 /* continue to close even if error occurred. */
1967 if (mode & FMODE_WRITE)
1969 else if (mode & MDS_FMODE_EXEC)
1972 /* Update atime on close only. */
1973 if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1974 && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1975 /* Set the atime only. */
1976 ma->ma_valid = MA_INODE;
1977 ma->ma_attr.la_valid = LA_ATIME;
1978 rc = mo_attr_set(info->mti_env, next, ma);
1981 /* If file data is modified, add the dirty flag. */
1982 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
1983 rc = mdt_add_dirty_flag(info, o, ma);
1985 ma->ma_need |= MA_INODE;
1986 ma->ma_valid &= ~MA_INODE;
1988 if (!MFD_CLOSED(mode))
1989 rc = mo_close(info->mti_env, next, ma, mode);
1991 /* adjust open and lease count */
1992 if (mode & MDS_OPEN_LEASE) {
1993 LASSERT(atomic_read(&o->mot_lease_count) > 0);
1994 atomic_dec(&o->mot_lease_count);
1997 LASSERT(atomic_read(&o->mot_open_count) > 0);
1998 atomic_dec(&o->mot_open_count);
2000 mdt_object_put(info->mti_env, o);
2005 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2006 struct mdt_body *repbody)
2008 struct mdt_export_data *med;
2009 struct mdt_file_data *mfd;
2010 struct mdt_object *o;
2011 struct md_attr *ma = &info->mti_attr;
2016 med = &req->rq_export->exp_mdt_data;
2017 spin_lock(&med->med_open_lock);
2018 mfd = mdt_handle2mfd(med, &info->mti_close_handle, req_is_replay(req));
2019 if (mdt_mfd_closed(mfd)) {
2020 spin_unlock(&med->med_open_lock);
2021 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2022 ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
2023 info->mti_close_handle.cookie);
2024 /** not serious error since bug 3633 */
2027 class_handle_unhash(&mfd->mfd_handle);
2028 list_del_init(&mfd->mfd_list);
2029 spin_unlock(&med->med_open_lock);
2031 /* Do not lose object before last unlink. */
2032 o = mfd->mfd_object;
2033 mdt_object_get(info->mti_env, o);
2034 ret = mdt_mfd_close(info, mfd);
2035 rc = mdt_handle_last_unlink(info, o, ma);
2036 mdt_object_put(info->mti_env, o);
2039 RETURN(rc ? rc : ret);
2042 int mdt_close(struct tgt_session_info *tsi)
2044 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2045 struct ptlrpc_request *req = tgt_ses_req(tsi);
2046 struct md_attr *ma = &info->mti_attr;
2047 struct mdt_body *repbody = NULL;
2051 mdt_counter_incr(req, LPROC_MDT_CLOSE);
2052 /* Close may come with the Size-on-MDS update. Unpack it. */
2053 rc = mdt_close_unpack(info);
2055 GOTO(out, rc = err_serious(rc));
2057 /* These fields are no longer used and are left for compatibility.
2058 * size is always zero */
2059 req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2061 req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2063 rc = req_capsule_server_pack(info->mti_pill);
2064 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2065 mdt_client_compatibility(info);
2067 mdt_fix_reply(info);
2068 mdt_exit_ucred(info);
2069 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2072 /* Continue to close handle even if we can not pack reply */
2074 repbody = req_capsule_server_get(info->mti_pill,
2076 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2078 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2081 ma->ma_need = MA_INODE | MA_LOV;
2082 repbody->mbo_eadatasize = 0;
2083 repbody->mbo_aclsize = 0;
2085 rc = err_serious(rc);
2088 rc = mdt_close_internal(info, req, repbody);
2090 mdt_empty_transno(info, rc);
2092 if (repbody != NULL) {
2093 mdt_client_compatibility(info);
2094 rc = mdt_fix_reply(info);
2097 mdt_exit_ucred(info);
2098 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2099 GOTO(out, rc = err_serious(-ENOMEM));
2101 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2102 OBD_FAIL_MDS_CLOSE_NET_REP))
2103 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2105 mdt_thread_info_fini(info);
2106 RETURN(rc ? rc : ret);