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, 2013, 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"
49 /* we do nothing because we do not have refcount now */
50 static void mdt_mfd_get(void *mfdp)
54 static struct portals_handle_ops mfd_handle_ops = {
55 .hop_addref = mdt_mfd_get,
59 /* Create a new mdt_file_data struct, initialize it,
60 * and insert it to global hash table */
61 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
63 struct mdt_file_data *mfd;
68 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
69 mfd->mfd_handle.h_owner = med;
70 INIT_LIST_HEAD(&mfd->mfd_list);
71 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
78 * Find the mfd pointed to by handle in global hash table.
79 * In case of replay the handle is obsoleted
80 * but mfd can be found in mfd list by that handle.
81 * Callers need to be holding med_open_lock.
83 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
84 const struct lustre_handle *handle,
85 bool is_replay_or_resent)
87 struct mdt_file_data *mfd;
90 LASSERT(handle != NULL);
91 mfd = class_handle2object(handle->cookie, med);
92 /* during dw/setattr replay the mfd can be found by old handle */
93 if (mfd == NULL && is_replay_or_resent) {
94 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
95 if (mfd->mfd_old_handle.cookie == handle->cookie)
105 void mdt_mfd_free(struct mdt_file_data *mfd)
107 LASSERT(list_empty(&mfd->mfd_list));
108 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
111 static int mdt_create_data(struct mdt_thread_info *info,
112 struct mdt_object *p, struct mdt_object *o)
114 struct md_op_spec *spec = &info->mti_spec;
115 struct md_attr *ma = &info->mti_attr;
119 if (!md_should_create(spec->sp_cr_flags))
122 ma->ma_need = MA_INODE | MA_LOV;
124 mutex_lock(&o->mot_lov_mutex);
125 if (!(o->mot_flags & MOF_LOV_CREATED)) {
126 rc = mdo_create_data(info->mti_env,
127 p ? mdt_object_child(p) : NULL,
128 mdt_object_child(o), spec, ma);
130 rc = mdt_attr_get_complex(info, o, ma);
132 if (rc == 0 && ma->ma_valid & MA_LOV)
133 o->mot_flags |= MOF_LOV_CREATED;
136 mutex_unlock(&o->mot_lov_mutex);
140 static int mdt_ioepoch_opened(struct mdt_object *mo)
142 return mo->mot_ioepoch_count;
145 int mdt_object_is_som_enabled(struct mdt_object *mo)
147 return !mo->mot_ioepoch;
151 * Re-enable Size-on-MDS.
152 * Call under ->mot_ioepoch_mutex.
154 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
156 if (ioepoch == mo->mot_ioepoch) {
157 LASSERT(!mdt_ioepoch_opened(mo));
164 * Open the IOEpoch. It is allowed if @writecount is not negative.
165 * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
167 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
170 struct mdt_device *mdt = info->mti_mdt;
175 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
176 !S_ISREG(lu_object_attr(&o->mot_obj)))
179 mutex_lock(&o->mot_ioepoch_mutex);
180 if (mdt_ioepoch_opened(o)) {
181 /* Epoch continues even if there is no writers yet. */
182 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
183 o->mot_ioepoch, PFID(mdt_object_fid(o)));
185 /* XXX: ->mdt_ioepoch is not initialized at the mount */
186 spin_lock(&mdt->mdt_ioepoch_lock);
187 if (mdt->mdt_ioepoch < info->mti_replayepoch)
188 mdt->mdt_ioepoch = info->mti_replayepoch;
190 if (info->mti_replayepoch)
191 o->mot_ioepoch = info->mti_replayepoch;
192 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
193 o->mot_ioepoch = ++mdt->mdt_ioepoch;
195 o->mot_ioepoch = mdt->mdt_ioepoch;
197 spin_unlock(&mdt->mdt_ioepoch_lock);
199 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
200 o->mot_ioepoch, PFID(mdt_object_fid(o)));
202 o->mot_flags |= MOF_SOM_CREATED;
205 o->mot_ioepoch_count++;
206 mutex_unlock(&o->mot_ioepoch_mutex);
208 /* Cancel Size-on-MDS attributes cached on clients for the open case.
209 * In the truncate case, see mdt_reint_setattr(). */
210 if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
211 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
212 mdt_lock_reg_init(lh, LCK_EX);
213 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
216 mdt_object_unlock(info, o, lh, 1);
222 * Update SOM on-disk attributes.
223 * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
224 * mountid and attributes. If disabling, clean SOM xattr.
225 * Call under ->mot_ioepoch_mutex.
227 static int mdt_som_attr_set(struct mdt_thread_info *info,
228 struct mdt_object *obj, __u64 ioepoch, bool enable)
230 struct md_object *next = mdt_object_child(obj);
234 CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
235 " on "DFID".\n", enable ? "update" : "disabling",
236 ioepoch, PFID(mdt_object_fid(obj)));
239 struct lu_buf *buf = &info->mti_buf;
240 struct som_attrs *attrs;
241 struct md_attr *ma = &info->mti_attr;
242 struct lu_attr *la = &ma->ma_attr;
243 struct obd_device *obd = info->mti_mdt->mdt_lut.lut_obd;
245 attrs = (struct som_attrs *)info->mti_xattr_buf;
246 CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*attrs));
248 /* pack SOM attributes */
249 memset(attrs, 0, sizeof(*attrs));
250 attrs->som_ioepoch = ioepoch;
251 attrs->som_mountid = obd->u.obt.obt_mount_count;
252 if ((la->la_valid & LA_SIZE) != 0)
253 attrs->som_size = la->la_size;
254 if ((la->la_valid & LA_BLOCKS) != 0)
255 attrs->som_blocks = la->la_blocks;
256 lustre_som_swab(attrs);
258 /* update SOM attributes */
260 buf->lb_len = sizeof(*attrs);
261 rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_SOM, 0);
263 /* delete SOM attributes */
264 rc = mo_xattr_del(info->mti_env, next, XATTR_NAME_SOM);
270 /** Perform the eviction specific actions on ioepoch close. */
271 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
272 struct mdt_object *o)
276 mutex_lock(&o->mot_ioepoch_mutex);
277 CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
278 "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
279 o->mot_ioepoch_count);
280 o->mot_ioepoch_count--;
282 /* If eviction occured set MOF_SOM_RECOV,
283 * if no other epoch holders, disable SOM on disk. */
284 o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
285 if (!mdt_ioepoch_opened(o)) {
286 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
287 mdt_object_som_enable(o, o->mot_ioepoch);
289 mutex_unlock(&o->mot_ioepoch_mutex);
294 * Perform the replay specific actions on ioepoch close.
295 * Skip SOM attribute update if obtained and just forget about the inode state
296 * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
298 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
299 struct mdt_object *o)
301 int rc = MDT_IOEPOCH_CLOSED;
304 mutex_lock(&o->mot_ioepoch_mutex);
305 CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
306 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
307 o->mot_ioepoch_count--;
309 /* Get an info from the replayed request if client is supposed
310 * to send an Attibute Update, reconstruct @rc if so */
311 if (info->mti_ioepoch->flags & MF_SOM_AU)
312 rc = MDT_IOEPOCH_GETATTR;
314 if (!mdt_ioepoch_opened(o))
315 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
316 mutex_unlock(&o->mot_ioepoch_mutex);
322 * Regular file IOepoch close.
323 * Closes the ioepoch, checks the object state, apply obtained attributes and
324 * re-enable SOM on the object, if possible. Also checks if the recovery is
325 * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
326 * to obtain SOM attributes under the server-side OST locks.
329 * MDT_IOEPOCH_CLOSED if ioepoch is closed.
330 * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
332 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
333 struct mdt_object *o)
335 struct md_attr *tmp_ma;
339 int rc = 0, ret = MDT_IOEPOCH_CLOSED;
342 la = &info->mti_attr.ma_attr;
343 achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
345 mutex_lock(&o->mot_ioepoch_mutex);
346 o->mot_ioepoch_count--;
348 tmp_ma = &info->mti_u.som.attr;
349 tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
350 tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
351 tmp_ma->ma_som = &info->mti_u.som.data;
352 tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
353 tmp_ma->ma_valid = 0;
354 rc = mdt_attr_get_complex(info, o, tmp_ma);
358 /* Check the on-disk SOM state. */
359 if (o->mot_flags & MOF_SOM_RECOV)
361 else if (!(o->mot_flags & MOF_SOM_CREATED) &&
362 !(tmp_ma->ma_valid & MA_SOM))
365 CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
366 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
368 opened = mdt_ioepoch_opened(o);
370 * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
371 * Skip the check for file with no LOV or for unlink files.
373 if (!opened && tmp_ma->ma_valid & MA_LOV &&
374 !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
376 /* If some previous writer was evicted, re-ask the
377 * client for attributes. Even if attributes are
378 * provided, we cannot believe in them.
379 * Another use case is that there is no SOM cache on
380 * disk -- first access with SOM or there was an MDS
382 ret = MDT_IOEPOCH_GETATTR;
383 else if (o->mot_flags & MOF_SOM_CHANGE)
384 /* Some previous writer changed the attribute.
385 * Do not believe to the current Size-on-MDS
386 * update, re-ask client. */
387 ret = MDT_IOEPOCH_GETATTR;
388 else if (!(la->la_valid & LA_SIZE) && achange)
389 /* Attributes were changed by the last writer
390 * only but no Size-on-MDS update is received.*/
391 ret = MDT_IOEPOCH_GETATTR;
394 if (achange || ret == MDT_IOEPOCH_GETATTR)
395 o->mot_flags |= MOF_SOM_CHANGE;
397 /* If epoch ends and relable SOM attributes are obtained, update them.
398 * Create SOM ea for new files even if there is no attributes obtained
399 * (0-length file). */
400 if (ret == MDT_IOEPOCH_CLOSED && !opened) {
401 if (achange || o->mot_flags & MOF_SOM_CREATED) {
402 LASSERT(achange || !(la->la_valid & LA_SIZE));
403 rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
405 /* Avoid the following setattrs of these attributes,
406 * e.g. for atime update. */
407 info->mti_attr.ma_valid = 0;
409 mdt_object_som_enable(o, o->mot_ioepoch);
412 mutex_unlock(&o->mot_ioepoch_mutex);
413 /* If recovery is needed, tell the client to perform GETATTR under
415 if (ret == MDT_IOEPOCH_GETATTR && recovery) {
416 struct mdt_body *rep;
417 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
418 rep->mbo_valid |= OBD_MD_FLGETATTRLOCK;
424 mutex_unlock(&o->mot_ioepoch_mutex);
429 * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
430 * - a client closes the IOEpoch;
431 * - a client eviction occured.
433 * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
434 * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
435 * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
438 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
440 struct ptlrpc_request *req = mdt_info_req(info);
443 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
444 !S_ISREG(lu_object_attr(&o->mot_obj)))
447 LASSERT(o->mot_ioepoch_count);
448 LASSERT(info->mti_ioepoch == NULL ||
449 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
451 /* IOEpoch is closed only if client tells about it or eviction occures.
452 * In the replay case, always close the epoch. */
454 RETURN(mdt_ioepoch_close_on_eviction(info, o));
455 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
456 RETURN(mdt_ioepoch_close_on_replay(info, o));
457 if (info->mti_ioepoch && (info->mti_ioepoch->flags & MF_EPOCH_CLOSE))
458 RETURN(mdt_ioepoch_close_reg(info, o));
459 /* IO epoch is not closed. */
460 RETURN(MDT_IOEPOCH_OPENED);
464 * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
465 * for attribute update. It happens if:
466 * - SOM Attribute Update is obtained;
467 * - the client failed to obtain it and informs MDS about it;
468 * - a client eviction occured.
469 * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
472 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
474 struct ptlrpc_request *req = mdt_info_req(info);
476 int act = MDT_SOM_ENABLE;
480 LASSERT(!req || info->mti_ioepoch);
481 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
482 !S_ISREG(lu_object_attr(&o->mot_obj)))
485 /* No size whereas MF_SOM_CHANGE is set means client failed to
486 * obtain ost attributes, drop the SOM cache on disk if so. */
488 (info->mti_ioepoch &&
489 info->mti_ioepoch->flags & MF_SOM_CHANGE &&
490 !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
491 act = MDT_SOM_DISABLE;
493 mutex_lock(&o->mot_ioepoch_mutex);
494 /* Mark the object it is the recovery state if we failed to obtain
496 if (act == MDT_SOM_DISABLE)
497 o->mot_flags |= MOF_SOM_RECOV;
499 if (!mdt_ioepoch_opened(o)) {
500 ioepoch = info->mti_ioepoch ?
501 info->mti_ioepoch->ioepoch : o->mot_ioepoch;
504 && !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
505 rc = mdt_som_attr_set(info, o, ioepoch, act);
506 mdt_object_som_enable(o, ioepoch);
508 mutex_unlock(&o->mot_ioepoch_mutex);
512 int mdt_write_read(struct mdt_object *o)
516 mutex_lock(&o->mot_ioepoch_mutex);
517 rc = o->mot_writecount;
518 mutex_unlock(&o->mot_ioepoch_mutex);
522 int mdt_write_get(struct mdt_object *o)
526 mutex_lock(&o->mot_ioepoch_mutex);
527 if (o->mot_writecount < 0)
531 mutex_unlock(&o->mot_ioepoch_mutex);
535 void mdt_write_put(struct mdt_object *o)
538 mutex_lock(&o->mot_ioepoch_mutex);
540 mutex_unlock(&o->mot_ioepoch_mutex);
544 static int mdt_write_deny(struct mdt_object *o)
548 mutex_lock(&o->mot_ioepoch_mutex);
549 if (o->mot_writecount > 0)
553 mutex_unlock(&o->mot_ioepoch_mutex);
557 static void mdt_write_allow(struct mdt_object *o)
560 mutex_lock(&o->mot_ioepoch_mutex);
562 mutex_unlock(&o->mot_ioepoch_mutex);
566 /* there can be no real transaction so prepare the fake one */
567 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
569 struct mdt_device *mdt = info->mti_mdt;
570 struct ptlrpc_request *req = mdt_info_req(info);
571 struct tg_export_data *ted;
572 struct lsd_client_data *lcd;
575 /* transaction has occurred already */
576 if (lustre_msg_get_transno(req->rq_repmsg) != 0)
579 spin_lock(&mdt->mdt_lut.lut_translock);
581 if (info->mti_transno != 0) {
582 struct obd_export *exp = req->rq_export;
584 CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
585 mdt_obd_name(mdt), info->mti_transno,
586 libcfs_nid2str(exp->exp_connection->c_peer.nid),
588 spin_unlock(&mdt->mdt_lut.lut_translock);
591 } else if (info->mti_transno == 0) {
592 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
594 /* should be replay */
595 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
596 mdt->mdt_lut.lut_last_transno = info->mti_transno;
598 spin_unlock(&mdt->mdt_lut.lut_translock);
600 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
602 req->rq_export->exp_obd->obd_last_committed);
604 req->rq_transno = info->mti_transno;
605 lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
607 /* update lcd in memory only for resent cases */
608 ted = &req->rq_export->exp_target_data;
610 mutex_lock(&ted->ted_lcd_lock);
612 if (info->mti_transno < lcd->lcd_last_transno &&
613 info->mti_transno != 0) {
614 /* This should happen during replay. Do not update
615 * last rcvd info if replay req transno < last transno,
616 * otherwise the following resend(after replay) can not
617 * be checked correctly by xid */
618 mutex_unlock(&ted->ted_lcd_lock);
619 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
620 mdt_obd_name(mdt), info->mti_transno,
621 lcd->lcd_last_transno);
625 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
626 lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
627 if (info->mti_transno != 0)
628 lcd->lcd_last_close_transno = info->mti_transno;
629 lcd->lcd_last_close_xid = req->rq_xid;
630 lcd->lcd_last_close_result = rc;
632 /* VBR: save versions in last_rcvd for reconstruct. */
633 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
635 lcd->lcd_pre_versions[0] = pre_versions[0];
636 lcd->lcd_pre_versions[1] = pre_versions[1];
637 lcd->lcd_pre_versions[2] = pre_versions[2];
638 lcd->lcd_pre_versions[3] = pre_versions[3];
640 if (info->mti_transno != 0)
641 lcd->lcd_last_transno = info->mti_transno;
643 lcd->lcd_last_xid = req->rq_xid;
644 lcd->lcd_last_result = rc;
645 lcd->lcd_last_data = info->mti_opdata;
647 mutex_unlock(&ted->ted_lcd_lock);
652 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
654 LASSERT(mfd != NULL);
656 CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n",
657 PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
659 mfd->mfd_mode = mode;
663 * prep ma_lmm/ma_lmv for md_attr from reply
665 void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
666 struct mdt_object *obj,
669 LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
670 if (S_ISDIR(obj->mot_header.loh_attr)) {
671 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
673 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
676 if (ma->ma_lmv_size > 0)
677 ma->ma_need |= MA_LMV;
679 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
681 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
684 if (ma->ma_lmm_size > 0)
685 ma->ma_need |= MA_LOV;
689 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
690 struct mdt_object *o, __u64 flags, int created,
691 struct ldlm_reply *rep)
693 struct ptlrpc_request *req = mdt_info_req(info);
694 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
695 struct mdt_file_data *mfd;
696 struct md_attr *ma = &info->mti_attr;
697 struct lu_attr *la = &ma->ma_attr;
698 struct mdt_body *repbody;
699 int rc = 0, isdir, isreg;
702 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
704 isreg = S_ISREG(la->la_mode);
705 isdir = S_ISDIR(la->la_mode);
706 if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
708 * No EA, check whether it is will set regEA and dirEA since in
709 * above attr get, these size might be zero, so reset it, to
710 * retrieve the MD after create obj.
712 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
715 /* in replay case, p == NULL */
716 rc = mdt_create_data(info, p, o);
720 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
721 mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
724 CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
725 ma->ma_valid, ma->ma_lmm_size);
727 if (ma->ma_valid & MA_LOV) {
728 LASSERT(ma->ma_lmm_size != 0);
729 repbody->mbo_eadatasize = ma->ma_lmm_size;
731 repbody->mbo_valid |= OBD_MD_FLDIREA;
733 repbody->mbo_valid |= OBD_MD_FLEASIZE;
736 if (ma->ma_valid & MA_LMV) {
737 LASSERT(ma->ma_lmv_size != 0);
738 repbody->mbo_eadatasize = ma->ma_lmv_size;
740 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
743 if (flags & FMODE_WRITE) {
744 rc = mdt_write_get(o);
746 mdt_ioepoch_open(info, o, created);
747 repbody->mbo_ioepoch = o->mot_ioepoch;
749 } else if (flags & MDS_FMODE_EXEC) {
750 rc = mdt_write_deny(o);
755 rc = mo_open(info->mti_env, mdt_object_child(o),
756 created ? flags | MDS_OPEN_CREATED : flags);
760 mfd = mdt_mfd_new(med);
762 GOTO(err_out, rc = -ENOMEM);
765 * Keep a reference on this object for this open, and is
766 * released by mdt_mfd_close().
768 mdt_object_get(info->mti_env, o);
770 mfd->mfd_xid = req->rq_xid;
773 * @flags is always not zero. At least it should be FMODE_READ,
774 * FMODE_WRITE or MDS_FMODE_EXEC.
779 mdt_mfd_set_mode(mfd, flags);
781 atomic_inc(&o->mot_open_count);
782 if (flags & MDS_OPEN_LEASE)
783 atomic_inc(&o->mot_lease_count);
786 if (req_is_replay(req)) {
787 struct mdt_file_data *old_mfd;
788 /* Check wheather old cookie already exist in
789 * the list, becasue when do recovery, client
790 * might be disconnected from server, and
791 * restart replay, so there maybe some orphan
792 * mfd here, we should remove them */
793 LASSERT(info->mti_rr.rr_handle != NULL);
794 spin_lock(&med->med_open_lock);
795 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
796 if (old_mfd != NULL) {
797 CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
798 "cookie = "LPX64"\n", mfd,
799 PFID(mdt_object_fid(mfd->mfd_object)),
800 info->mti_rr.rr_handle->cookie);
801 class_handle_unhash(&old_mfd->mfd_handle);
802 list_del_init(&old_mfd->mfd_list);
803 spin_unlock(&med->med_open_lock);
804 /* no attr update for that close */
806 ma->ma_valid |= MA_FLAGS;
807 ma->ma_attr_flags |= MDS_RECOV_OPEN;
808 mdt_mfd_close(info, old_mfd);
809 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
810 ma->ma_valid &= ~MA_FLAGS;
812 spin_unlock(&med->med_open_lock);
813 CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
814 "cookie = "LPX64"\n",
815 PFID(mdt_object_fid(mfd->mfd_object)),
816 info->mti_rr.rr_handle->cookie);
819 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
820 info->mti_rr.rr_handle->cookie);
822 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
825 repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
827 if (req->rq_export->exp_disconnected) {
828 spin_lock(&med->med_open_lock);
829 class_handle_unhash(&mfd->mfd_handle);
830 list_del_init(&mfd->mfd_list);
831 spin_unlock(&med->med_open_lock);
832 mdt_mfd_close(info, mfd);
834 spin_lock(&med->med_open_lock);
835 list_add(&mfd->mfd_list, &med->med_open_head);
836 spin_unlock(&med->med_open_lock);
839 mdt_empty_transno(info, rc);
844 if (flags & FMODE_WRITE)
845 /* XXX We also need to close io epoch here.
846 * See LU-1220 - green */
848 else if (flags & MDS_FMODE_EXEC)
854 int mdt_finish_open(struct mdt_thread_info *info,
855 struct mdt_object *p, struct mdt_object *o,
856 __u64 flags, int created, struct ldlm_reply *rep)
858 struct ptlrpc_request *req = mdt_info_req(info);
859 struct obd_export *exp = req->rq_export;
860 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
861 struct md_attr *ma = &info->mti_attr;
862 struct lu_attr *la = &ma->ma_attr;
863 struct mdt_file_data *mfd;
864 struct mdt_body *repbody;
866 int isreg, isdir, islnk;
870 LASSERT(ma->ma_valid & MA_INODE);
872 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
874 isreg = S_ISREG(la->la_mode);
875 isdir = S_ISDIR(la->la_mode);
876 islnk = S_ISLNK(la->la_mode);
877 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
879 /* LU-2275, simulate broken behaviour (esp. prevalent in
880 * pre-2.4 servers where a very strange reply is sent on error
881 * that looks like it was actually almost succesful and a failure at the
883 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
884 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
888 if (flags & MDS_OPEN_LOCK)
889 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
894 if (exp_connect_rmtclient(exp)) {
895 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
897 rc = mdt_pack_remote_perm(info, o, buf);
899 repbody->mbo_valid &= ~OBD_MD_FLRMTPERM;
900 repbody->mbo_aclsize = 0;
902 repbody->mbo_valid |= OBD_MD_FLRMTPERM;
903 repbody->mbo_aclsize = sizeof(struct mdt_remote_perm);
906 #ifdef CONFIG_FS_POSIX_ACL
907 else if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
908 const struct lu_env *env = info->mti_env;
909 struct md_object *next = mdt_object_child(o);
910 struct lu_buf *buf = &info->mti_buf;
912 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
913 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
915 if (buf->lb_len > 0) {
916 rc = mo_xattr_get(env, next, buf,
917 XATTR_NAME_ACL_ACCESS);
919 if (rc == -ENODATA) {
920 repbody->mbo_aclsize = 0;
921 repbody->mbo_valid |= OBD_MD_FLACL;
923 } else if (rc == -EOPNOTSUPP) {
926 CERROR("got acl size: %d\n", rc);
929 repbody->mbo_aclsize = rc;
930 repbody->mbo_valid |= OBD_MD_FLACL;
937 if (info->mti_mdt->mdt_lut.lut_mds_capa &&
938 exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) {
939 struct lustre_capa *capa;
941 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
943 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
944 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
947 repbody->mbo_valid |= OBD_MD_FLMDSCAPA;
949 if (info->mti_mdt->mdt_lut.lut_oss_capa &&
950 exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA &&
951 S_ISREG(lu_object_attr(&o->mot_obj))) {
952 struct lustre_capa *capa;
954 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
956 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
957 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
960 repbody->mbo_valid |= OBD_MD_FLOSSCAPA;
964 * If we are following a symlink, don't open; and do not return open
965 * handle for special nodes as client required.
967 if (islnk || (!isreg && !isdir &&
968 (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
969 lustre_msg_set_transno(req->rq_repmsg, 0);
974 * We need to return the existing object's fid back, so it is done here,
975 * after preparing the reply.
977 if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
980 /* This can't be done earlier, we need to return reply body */
982 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
983 /* We are trying to create or write an existing dir. */
986 } else if (flags & MDS_OPEN_DIRECTORY)
989 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
990 OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
995 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
996 spin_lock(&med->med_open_lock);
997 list_for_each(t, &med->med_open_head) {
998 mfd = list_entry(t, struct mdt_file_data, mfd_list);
999 if (mfd->mfd_xid == req->rq_xid)
1003 spin_unlock(&med->med_open_lock);
1006 repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
1007 /* set repbody->ea_size for resent case */
1008 if (ma->ma_valid & MA_LOV) {
1009 LASSERT(ma->ma_lmm_size != 0);
1010 repbody->mbo_eadatasize = ma->ma_lmm_size;
1012 repbody->mbo_valid |= OBD_MD_FLDIREA;
1014 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1016 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
1021 rc = mdt_mfd_open(info, p, o, flags, created, rep);
1023 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
1028 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
1029 struct lsd_client_data *lcd);
1031 void mdt_reconstruct_open(struct mdt_thread_info *info,
1032 struct mdt_lock_handle *lhc)
1034 const struct lu_env *env = info->mti_env;
1035 struct mdt_device *mdt = info->mti_mdt;
1036 struct req_capsule *pill = info->mti_pill;
1037 struct ptlrpc_request *req = mdt_info_req(info);
1038 struct tg_export_data *ted = &req->rq_export->exp_target_data;
1039 struct lsd_client_data *lcd = ted->ted_lcd;
1040 struct md_attr *ma = &info->mti_attr;
1041 struct mdt_reint_record *rr = &info->mti_rr;
1042 __u64 flags = info->mti_spec.sp_cr_flags;
1043 struct ldlm_reply *ldlm_rep;
1044 struct mdt_object *parent;
1045 struct mdt_object *child;
1046 struct mdt_body *repbody;
1050 LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1051 ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1052 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1054 ma->ma_need = MA_INODE | MA_HSM;
1057 mdt_req_from_lcd(req, lcd);
1058 mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1060 CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1061 ldlm_rep->lock_policy_res1, req->rq_status);
1063 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1064 req->rq_status != 0)
1065 /* We did not create successfully, return error to client. */
1066 GOTO(out, rc = req->rq_status);
1068 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1069 struct obd_export *exp = req->rq_export;
1071 * We failed after creation, but we do not know in which step
1072 * we failed. So try to check the child object.
1074 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1075 if (IS_ERR(parent)) {
1076 rc = PTR_ERR(parent);
1077 LCONSOLE_WARN("Parent "DFID" lookup error %d."
1078 " Evicting client %s with export %s.\n",
1079 PFID(rr->rr_fid1), rc,
1080 obd_uuid2str(&exp->exp_client_uuid),
1081 obd_export_nid2str(exp));
1082 mdt_export_evict(exp);
1086 child = mdt_object_find(env, mdt, rr->rr_fid2);
1087 if (IS_ERR(child)) {
1088 rc = PTR_ERR(child);
1089 LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
1090 "evicting client %s with export %s\n",
1091 PFID(rr->rr_fid2), rc,
1092 obd_uuid2str(&exp->exp_client_uuid),
1093 obd_export_nid2str(exp));
1094 mdt_object_put(env, parent);
1095 mdt_export_evict(exp);
1099 if (unlikely(mdt_object_remote(child))) {
1100 /* the child object was created on remote server */
1101 if (!mdt_is_dne_client(exp)) {
1102 /* Return -EIO for old client */
1103 mdt_object_put(env, parent);
1104 mdt_object_put(env, child);
1105 GOTO(out, rc = -EIO);
1107 repbody->mbo_fid1 = *rr->rr_fid2;
1108 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1111 if (mdt_object_exists(child)) {
1112 mdt_set_capainfo(info, 1, rr->rr_fid2,
1114 mdt_prep_ma_buf_from_rep(info, child, ma);
1115 rc = mdt_attr_get_complex(info, child, ma);
1117 rc = mdt_finish_open(info, parent,
1121 /* the child does not exist, we should do
1123 mdt_object_put(env, parent);
1124 mdt_object_put(env, child);
1125 GOTO(regular_open, 0);
1128 mdt_object_put(env, parent);
1129 mdt_object_put(env, child);
1133 /* We did not try to create, so we are a pure open */
1134 rc = mdt_reint_open(info, lhc);
1139 req->rq_status = rc;
1140 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1141 LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1144 int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
1146 __u64 flags = info->mti_spec.sp_cr_flags;
1147 struct mdt_reint_record *rr = &info->mti_rr;
1148 struct md_attr *ma = &info->mti_attr;
1149 struct mdt_object *o;
1153 o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1155 RETURN(rc = PTR_ERR(o));
1157 if (unlikely(mdt_object_remote(o))) {
1158 /* the child object was created on remote server */
1159 struct mdt_body *repbody;
1161 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1164 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1165 repbody->mbo_fid1 = *rr->rr_fid2;
1166 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1169 if (mdt_object_exists(o)) {
1170 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1173 mdt_prep_ma_buf_from_rep(info, o, ma);
1174 rc = mdt_attr_get_complex(info, o, ma);
1176 rc = mdt_finish_open(info, NULL, o, flags, 0,
1183 mdt_object_put(info->mti_env, o);
1187 /* lock object for open */
1188 static int mdt_object_open_lock(struct mdt_thread_info *info,
1189 struct mdt_object *obj,
1190 struct mdt_lock_handle *lhc,
1193 struct md_attr *ma = &info->mti_attr;
1194 __u64 open_flags = info->mti_spec.sp_cr_flags;
1195 ldlm_mode_t lm = LCK_CR;
1196 bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
1197 bool try_layout = false;
1198 bool create_layout = false;
1203 mdt_lock_handle_init(lhc);
1205 if (req_is_replay(mdt_info_req(info)))
1208 if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
1209 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
1210 md_should_create(open_flags))
1211 create_layout = true;
1212 if (exp_connect_layout(info->mti_exp) && !create_layout &&
1213 ma->ma_need & MA_LOV)
1218 /* lease open, acquire write mode of open sem */
1219 down_write(&obj->mot_open_sem);
1221 /* Lease exists and ask for new lease */
1222 if (atomic_read(&obj->mot_lease_count) > 0) {
1223 /* only exclusive open is supported, so lease
1224 * are conflicted to each other */
1225 GOTO(out, rc = -EBUSY);
1228 /* Lease must be with open lock */
1229 if (!(open_flags & MDS_OPEN_LOCK)) {
1230 CERROR("Request lease for file:"DFID ", but open lock "
1231 "is missed, open_flags = "LPO64".\n",
1232 PFID(mdt_object_fid(obj)), open_flags);
1233 GOTO(out, rc = -EPROTO);
1236 /* XXX: only exclusive open is supported. */
1238 *ibits = MDS_INODELOCK_OPEN;
1240 /* never grant LCK_EX layout lock to client */
1242 } else { /* normal open */
1243 /* normal open holds read mode of open sem */
1244 down_read(&obj->mot_open_sem);
1246 if (open_flags & MDS_OPEN_LOCK) {
1247 if (open_flags & FMODE_WRITE)
1249 else if (open_flags & MDS_FMODE_EXEC)
1254 *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
1255 } else if (atomic_read(&obj->mot_lease_count) > 0) {
1256 if (open_flags & FMODE_WRITE)
1262 *ibits = MDS_INODELOCK_OPEN;
1265 lhc = &info->mti_lh[MDT_LH_LOCAL];
1267 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
1268 PFID(mdt_object_fid(obj)),
1269 atomic_read(&obj->mot_open_count), lm);
1272 mdt_lock_reg_init(lhc, lm);
1274 /* one problem to return layout lock on open is that it may result
1275 * in too many layout locks cached on the client side. */
1276 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
1277 /* return lookup lock to validate inode at the client side,
1278 * this is pretty important otherwise mdt will return layout
1279 * lock for each open.
1280 * However this is a double-edged sword because changing
1281 * permission will revoke huge # of LOOKUP locks. */
1282 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
1283 if (!mdt_object_lock_try(info, obj, lhc, *ibits,
1285 *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
1287 rc = mdt_object_lock(info, obj, lhc, *ibits,
1290 } else if (*ibits != 0) {
1291 rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK);
1294 CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
1295 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
1296 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
1298 /* will change layout, revoke layout locks by enqueuing EX lock. */
1299 if (rc == 0 && create_layout) {
1300 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
1302 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
1303 ", open_flags = "LPO64"\n",
1304 PFID(mdt_object_fid(obj)), open_flags);
1306 /* We cannot enqueue another lock for the same resource we
1307 * already have a lock for, due to mechanics of waiting list
1308 * iterating in ldlm, see LU-3601.
1309 * As such we'll drop the open lock we just got above here,
1310 * it's ok not to have this open lock as it's main purpose is to
1311 * flush unused cached client open handles. */
1312 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1313 mdt_object_unlock(info, obj, lhc, 1);
1315 LASSERT(!try_layout);
1316 mdt_lock_handle_init(ll);
1317 mdt_lock_reg_init(ll, LCK_EX);
1318 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
1321 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
1324 /* Check if there is any other open handles after acquiring
1325 * open lock. At this point, caching open handles have been revoked
1327 * XXX: Now only exclusive open is supported. Need to check the
1328 * type of open for generic lease support. */
1329 if (rc == 0 && acq_lease) {
1330 struct ptlrpc_request *req = mdt_info_req(info);
1331 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1332 struct mdt_file_data *mfd;
1333 bool is_replay_or_resent;
1336 /* For lease: application can open a file and then apply lease,
1337 * @handle contains original open handle in that case.
1338 * In recovery, open REQ will be replayed and the lease REQ may
1339 * be resent that means the open handle is already stale, so we
1340 * need to fix it up here by finding new handle. */
1341 is_replay_or_resent = req_is_replay(req) ||
1342 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
1344 /* if the request is _not_ a replay request, rr_handle
1345 * may be used to hold an openhandle which is issuing the
1346 * lease request, so that this openhandle doesn't count. */
1347 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
1348 is_replay_or_resent);
1352 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
1353 PFID(mdt_object_fid(obj)),
1354 atomic_read(&obj->mot_open_count), open_count);
1356 if (atomic_read(&obj->mot_open_count) > open_count)
1357 GOTO(out, rc = -EBUSY);
1365 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1366 struct mdt_object *obj,
1367 struct mdt_lock_handle *lhc,
1368 __u64 ibits, int rc)
1370 __u64 open_flags = info->mti_spec.sp_cr_flags;
1371 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1374 if (req_is_replay(mdt_info_req(info)))
1377 /* Release local lock - the lock put in MDT_LH_LOCAL will never
1378 * return to client side. */
1379 if (lustre_handle_is_used(&ll->mlh_reg_lh))
1380 mdt_object_unlock(info, obj, ll, 1);
1382 ll = &info->mti_lh[MDT_LH_LAYOUT];
1383 /* Release local layout lock, layout was created */
1384 if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1385 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1386 mdt_object_unlock(info, obj, ll, 1);
1389 if (open_flags & MDS_OPEN_LEASE)
1390 up_write(&obj->mot_open_sem);
1392 up_read(&obj->mot_open_sem);
1394 /* Cross-ref case, the lock should be returned to the client */
1395 if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1398 if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
1399 /* for the open request, the lock will only return to client
1400 * if open or layout lock is granted. */
1404 if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1405 struct ldlm_reply *ldlm_rep;
1407 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1408 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1409 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1410 mdt_object_unlock(info, obj, lhc, 1);
1416 * Check release is permitted for the current HSM flags.
1418 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1420 if (!(ma->ma_valid & MA_HSM))
1423 if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1426 if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1432 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1433 struct mdt_lock_handle *lhc)
1435 const struct lu_env *env = info->mti_env;
1436 struct mdt_device *mdt = info->mti_mdt;
1437 __u64 flags = info->mti_spec.sp_cr_flags;
1438 struct mdt_reint_record *rr = &info->mti_rr;
1439 struct md_attr *ma = &info->mti_attr;
1440 struct mdt_object *parent= NULL;
1441 struct mdt_object *o;
1443 int object_locked = 0;
1447 if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1448 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1449 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1450 if (IS_ERR(parent)) {
1451 CDEBUG(D_INODE, "Fail to find parent "DFID
1452 " for anonymous created %ld, try to"
1453 " use server-side parent.\n",
1454 PFID(rr->rr_fid1), PTR_ERR(parent));
1459 ma->ma_need |= MA_PFID;
1462 o = mdt_object_find(env, mdt, rr->rr_fid2);
1464 GOTO(out_parent_put, rc = PTR_ERR(o));
1466 if (mdt_object_remote(o)) {
1467 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1468 mdt_obd_name(info->mti_mdt),
1470 GOTO(out, rc = -EREMOTE);
1471 } else if (!mdt_object_exists(o)) {
1472 mdt_set_disposition(info, rep,
1476 GOTO(out, rc = -ENOENT);
1479 mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1481 mdt_prep_ma_buf_from_rep(info, o, ma);
1482 if (flags & MDS_OPEN_RELEASE)
1483 ma->ma_need |= MA_HSM;
1484 rc = mdt_attr_get_complex(info, o, ma);
1488 /* If a release request, check file flags are fine and ask for an
1489 * exclusive open access. */
1490 if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1491 GOTO(out, rc = -EPERM);
1493 rc = mdt_check_resent_lock(info, o, lhc);
1496 } else if (rc > 0) {
1497 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1500 GOTO(out_unlock, rc);
1503 if (ma->ma_valid & MA_PFID) {
1504 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1505 if (IS_ERR(parent)) {
1506 CDEBUG(D_INODE, "Fail to find parent "DFID
1507 " for anonymous created %ld, try to"
1508 " use system default.\n",
1509 PFID(&ma->ma_pfid), PTR_ERR(parent));
1514 rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1516 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1517 if (flags & MDS_OPEN_LOCK)
1518 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1519 if (flags & MDS_OPEN_LEASE)
1520 mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1522 GOTO(out_unlock, rc);
1526 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1528 mdt_object_put(env, o);
1531 mdt_object_put(env, parent);
1535 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1536 static int mdt_cross_open(struct mdt_thread_info *info,
1537 const struct lu_fid *parent_fid,
1538 const struct lu_fid *fid,
1539 struct ldlm_reply *rep, __u32 flags)
1541 struct md_attr *ma = &info->mti_attr;
1542 struct mdt_object *o;
1546 o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1548 RETURN(rc = PTR_ERR(o));
1550 if (mdt_object_remote(o)) {
1551 /* Something is wrong here, the object is on another MDS! */
1552 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1553 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1554 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1556 "Object isn't on this server! FLD error?\n");
1559 if (mdt_object_exists(o)) {
1560 /* Do permission check for cross-open. */
1561 rc = mo_permission(info->mti_env, NULL,
1562 mdt_object_child(o),
1563 NULL, flags | MDS_OPEN_CROSS);
1567 mdt_prep_ma_buf_from_rep(info, o, ma);
1568 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1569 rc = mdt_attr_get_complex(info, o, ma);
1573 rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1576 * Something is wrong here. lookup was positive but
1577 * there is no object!
1579 CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1580 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1585 mdt_object_put(info->mti_env, o);
1589 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1591 struct mdt_device *mdt = info->mti_mdt;
1592 struct ptlrpc_request *req = mdt_info_req(info);
1593 struct mdt_object *parent;
1594 struct mdt_object *child;
1595 struct mdt_lock_handle *lh;
1596 struct ldlm_reply *ldlm_rep;
1597 struct mdt_body *repbody;
1598 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1599 struct md_attr *ma = &info->mti_attr;
1600 __u64 create_flags = info->mti_spec.sp_cr_flags;
1602 struct mdt_reint_record *rr = &info->mti_rr;
1605 int object_locked = 0;
1609 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1610 (obd_timeout + 1) / 4);
1612 mdt_counter_incr(req, LPROC_MDT_OPEN);
1613 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1615 ma->ma_need = MA_INODE;
1618 LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1619 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1621 if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1622 CERROR("file join is not supported anymore.\n");
1623 GOTO(out, result = err_serious(-EOPNOTSUPP));
1625 msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1627 if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1628 info->mti_spec.u.sp_ea.eadata == NULL)
1629 GOTO(out, result = err_serious(-EINVAL));
1631 CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1632 "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1633 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1634 PFID(rr->rr_fid2), create_flags,
1635 ma->ma_attr.la_mode, msg_flags);
1637 if (info->mti_cross_ref) {
1638 /* This is cross-ref open */
1639 mdt_set_disposition(info, ldlm_rep,
1640 (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1642 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1643 ldlm_rep, create_flags);
1645 } else if (req_is_replay(req) ||
1646 (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1647 /* This is a replay request or from liblustre with ea. */
1648 result = mdt_open_by_fid(info, ldlm_rep);
1650 if (result != -ENOENT) {
1651 if (req->rq_export->exp_libclient &&
1652 create_flags & MDS_OPEN_HAS_EA)
1653 GOTO(out, result = 0);
1656 /* We didn't find the correct object, so we need to re-create it
1657 * via a regular replay. */
1658 if (!(create_flags & MDS_OPEN_CREAT)) {
1659 DEBUG_REQ(D_ERROR, req,
1660 "OPEN & CREAT not in open replay/by_fid.");
1661 GOTO(out, result = -EFAULT);
1663 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1664 } else if (create_flags & (MDS_OPEN_BY_FID | MDS_OPEN_LOCK)) {
1666 * MDS_OPEN_LOCK is checked for backward compatibility with 2.1
1669 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1671 CDEBUG(D_INFO, "no object for "DFID": %d\n",
1672 PFID(rr->rr_fid2), result);
1676 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1677 GOTO(out, result = err_serious(-ENOMEM));
1679 mdt_set_disposition(info, ldlm_rep,
1680 (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1682 if (!lu_name_is_valid(&rr->rr_name))
1683 GOTO(out, result = -EPROTO);
1685 lh = &info->mti_lh[MDT_LH_PARENT];
1686 mdt_lock_pdo_init(lh,
1687 (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1690 parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1691 MDS_INODELOCK_UPDATE);
1693 GOTO(out, result = PTR_ERR(parent));
1695 /* get and check version of parent */
1696 result = mdt_version_get_check(info, parent, 0);
1698 GOTO(out_parent, result);
1700 fid_zero(child_fid);
1702 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1703 &rr->rr_name, child_fid, &info->mti_spec);
1705 LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1706 "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1707 PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1710 if (result != 0 && result != -ENOENT && result != -ESTALE)
1711 GOTO(out_parent, result);
1713 if (result == -ENOENT || result == -ESTALE) {
1714 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1715 if (result == -ESTALE) {
1717 * -ESTALE means the parent is a dead(unlinked) dir, so
1718 * it should return -ENOENT to in accordance with the
1719 * original mds implementaion.
1721 GOTO(out_parent, result = -ENOENT);
1723 if (!(create_flags & MDS_OPEN_CREAT))
1724 GOTO(out_parent, result);
1725 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1726 GOTO(out_parent, result = -EROFS);
1727 *child_fid = *info->mti_rr.rr_fid2;
1728 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1730 /* In the function below, .hs_keycmp resolves to
1731 * lu_obj_hop_keycmp() */
1732 /* coverity[overrun-buffer-val] */
1733 child = mdt_object_new(info->mti_env, mdt, child_fid);
1736 * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1737 * return FID back in that case.
1739 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1740 child = mdt_object_find(info->mti_env, mdt, child_fid);
1743 GOTO(out_parent, result = PTR_ERR(child));
1745 /** check version of child */
1746 rc = mdt_version_get_check(info, child, 1);
1748 GOTO(out_child, result = rc);
1750 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1751 if (result == -ENOENT) {
1752 /* Create under OBF and .lustre is not permitted */
1753 if (!fid_is_md_operative(rr->rr_fid1))
1754 GOTO(out_child, result = -EPERM);
1756 /* save versions in reply */
1757 mdt_version_get_save(info, parent, 0);
1758 mdt_version_get_save(info, child, 1);
1760 /* version of child will be changed */
1761 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1763 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1764 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1766 /* Let lower layers know what is lock mode on directory. */
1767 info->mti_spec.sp_cr_mode =
1768 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1771 * Do not perform lookup sanity check. We know that name does
1774 info->mti_spec.sp_cr_lookup = 0;
1775 info->mti_spec.sp_feat = &dt_directory_features;
1777 result = mdo_create(info->mti_env,
1778 mdt_object_child(parent),
1780 mdt_object_child(child),
1783 if (result == -ERESTART) {
1784 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1785 GOTO(out_child, result);
1787 mdt_prep_ma_buf_from_rep(info, child, ma);
1788 /* XXX: we should call this once, see few lines below */
1790 result = mdt_attr_get_complex(info, child, ma);
1793 GOTO(out_child, result);
1798 * The object is on remote node, return its FID for remote open.
1800 if (mdt_object_remote(child)) {
1802 * Check if this lock already was sent to client and
1803 * this is resent case. For resent case do not take lock
1804 * again, use what is already granted.
1806 LASSERT(lhc != NULL);
1808 rc = mdt_check_resent_lock(info, child, lhc);
1810 GOTO(out_child, result = rc);
1811 } else if (rc > 0) {
1812 mdt_lock_handle_init(lhc);
1813 mdt_lock_reg_init(lhc, LCK_PR);
1815 rc = mdt_object_lock(info, child, lhc,
1816 MDS_INODELOCK_LOOKUP,
1819 repbody->mbo_fid1 = *mdt_object_fid(child);
1820 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1824 result = -MDT_EREMOTE_OPEN;
1825 GOTO(out_child, result);
1826 } else if (mdt_object_exists(child)) {
1827 /* We have to get attr & LOV EA & HSM for this
1829 mdt_prep_ma_buf_from_rep(info, child, ma);
1830 ma->ma_need |= MA_HSM;
1831 result = mdt_attr_get_complex(info, child, ma);
1833 GOTO(out_child, result);
1835 /* Object does not exist. Likely FS corruption. */
1836 CERROR("%s: name '"DNAME"' present, but FID "
1837 DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1838 PNAME(&rr->rr_name), PFID(child_fid));
1839 GOTO(out_child, result = -EIO);
1843 rc = mdt_check_resent_lock(info, child, lhc);
1845 GOTO(out_child, result = rc);
1846 } else if (rc == 0) {
1847 /* the open lock might already be gotten in
1848 * ldlm_handle_enqueue() */
1849 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1850 if (create_flags & MDS_OPEN_LOCK)
1851 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1853 /* get openlock if this isn't replay and client requested it */
1854 if (!req_is_replay(req)) {
1855 rc = mdt_object_open_lock(info, child, lhc, &ibits);
1858 GOTO(out_child_unlock, result = rc);
1859 else if (create_flags & MDS_OPEN_LOCK)
1860 mdt_set_disposition(info, ldlm_rep,
1864 /* Try to open it now. */
1865 rc = mdt_finish_open(info, parent, child, create_flags,
1869 /* openlock will be released if mdt_finish_open failed */
1870 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1872 if (created && create_flags & MDS_OPEN_VOLATILE) {
1873 CERROR("%s: cannot open volatile file "DFID", orphan "
1874 "file will be left in PENDING directory until "
1875 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1876 PFID(mdt_object_fid(child)), rc);
1877 GOTO(out_child_unlock, result);
1883 rc = mdo_unlink(info->mti_env,
1884 mdt_object_child(parent),
1885 mdt_object_child(child),
1887 &info->mti_attr, 0);
1889 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1890 mdt_obd_name(info->mti_mdt),
1891 PFID(mdt_object_fid(child)), rc);
1892 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1898 mdt_object_open_unlock(info, child, lhc, ibits, result);
1900 mdt_object_put(info->mti_env, child);
1902 mdt_object_unlock_put(info, parent, lh, result || !created);
1905 lustre_msg_set_transno(req->rq_repmsg, 0);
1910 * Create an orphan object use local root.
1912 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1913 struct mdt_device *mdt,
1914 const struct lu_fid *fid,
1915 struct md_attr *attr, fmode_t fmode)
1917 const struct lu_env *env = info->mti_env;
1918 struct md_op_spec *spec = &info->mti_spec;
1919 struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1920 struct mdt_object *obj = NULL;
1921 struct mdt_object *local_root;
1922 static const struct lu_name lname = {
1923 .ln_name = "i_am_nobody",
1924 .ln_namelen = sizeof("i_am_nobody") - 1,
1926 struct lu_ucred *uc;
1927 cfs_cap_t uc_cap_save;
1931 rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1933 RETURN(ERR_PTR(rc));
1935 local_root = mdt_object_find(env, mdt, local_root_fid);
1936 if (IS_ERR(local_root))
1939 obj = mdt_object_new(env, mdt, fid);
1941 GOTO(out, rc = PTR_ERR(obj));
1943 spec->sp_cr_lookup = 0;
1944 spec->sp_feat = &dt_directory_features;
1945 spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1946 spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1947 if (attr->ma_valid & MA_LOV) {
1948 spec->u.sp_ea.eadata = attr->ma_lmm;
1949 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1950 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1952 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1956 uc_cap_save = uc->uc_cap;
1957 uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1958 rc = mdo_create(env, mdt_object_child(local_root), &lname,
1959 mdt_object_child(obj), spec, attr);
1960 uc->uc_cap = uc_cap_save;
1962 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1963 mdt_obd_name(mdt), PFID(fid), rc);
1967 rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1969 CERROR("%s: cannot open volatile file "DFID", orphan "
1970 "file will be left in PENDING directory until "
1971 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1978 mdt_object_put(env, obj);
1981 mdt_object_put(env, local_root);
1985 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1988 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1989 struct close_data *data;
1990 struct ldlm_lock *lease;
1991 struct mdt_object *orphan;
1992 struct md_attr *orp_ma;
1999 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2002 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2006 lease = ldlm_handle2lock(&data->cd_handle);
2010 /* try to hold open_sem so that nobody else can open the file */
2011 if (!down_write_trylock(&o->mot_open_sem)) {
2012 ldlm_lock_cancel(lease);
2013 GOTO(out_reprocess, rc = -EBUSY);
2016 /* Check if the lease open lease has already canceled */
2017 lock_res_and_lock(lease);
2018 lease_broken = ldlm_is_cancel(lease);
2019 unlock_res_and_lock(lease);
2021 LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2022 PFID(mdt_object_fid(o)), lease_broken);
2024 /* Cancel server side lease. Client side counterpart should
2025 * have been cancelled. It's okay to cancel it now as we've
2026 * held mot_open_sem. */
2027 ldlm_lock_cancel(lease);
2029 if (lease_broken) /* don't perform release task */
2030 GOTO(out_unlock, rc = -ESTALE);
2032 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2033 GOTO(out_unlock, rc = -EINVAL);
2035 /* ma_need was set before but it seems fine to change it in order to
2036 * avoid modifying the one from RPC */
2037 ma->ma_need = MA_HSM;
2038 rc = mdt_attr_get_complex(info, o, ma);
2040 GOTO(out_unlock, rc);
2042 if (!mdt_hsm_release_allow(ma))
2043 GOTO(out_unlock, rc = -EPERM);
2045 /* already released? */
2046 if (ma->ma_hsm.mh_flags & HS_RELEASED)
2047 GOTO(out_unlock, rc = 0);
2049 /* Compare on-disk and packed data_version */
2050 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
2051 CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
2052 " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
2053 data->cd_data_version, ma->ma_hsm.mh_arch_ver);
2054 GOTO(out_unlock, rc = -EPERM);
2057 ma->ma_valid = MA_INODE;
2058 ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
2059 rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
2061 GOTO(out_unlock, rc);
2063 ma->ma_need = MA_INODE | MA_LOV;
2064 rc = mdt_attr_get_complex(info, o, ma);
2066 GOTO(out_unlock, rc);
2068 if (!(ma->ma_valid & MA_LOV)) {
2069 /* Even empty file are released */
2070 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2071 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2072 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2073 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2074 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2076 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
2077 * ma_lmm as lov_user_md, then it will be confused by union of
2078 * layout_gen and stripe_offset. */
2079 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
2080 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2081 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
2082 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
2084 GOTO(out_unlock, rc = -EINVAL);
2087 /* Set file as released */
2088 ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
2090 /* Hopefully it's not used in this call path */
2091 orp_ma = &info->mti_u.som.attr;
2092 orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2093 orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
2094 orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
2095 orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2096 orp_ma->ma_lmm = ma->ma_lmm;
2097 orp_ma->ma_lmm_size = ma->ma_lmm_size;
2098 orp_ma->ma_valid = MA_INODE | MA_LOV;
2099 orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2101 if (IS_ERR(orphan)) {
2102 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2103 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2105 GOTO(out_unlock, rc = PTR_ERR(orphan));
2108 /* Set up HSM attribute for orphan object */
2109 CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
2110 buf = &info->mti_buf;
2111 buf->lb_buf = info->mti_xattr_buf;
2112 buf->lb_len = sizeof(struct hsm_attrs);
2113 ma->ma_hsm.mh_flags |= HS_RELEASED;
2114 lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2115 ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2117 mdt_lock_reg_init(lh, LCK_EX);
2118 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2119 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
2121 GOTO(out_close, rc);
2123 rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2127 /* Swap layout with orphan object */
2128 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2129 mdt_object_child(orphan),
2130 SWAP_LAYOUTS_MDS_HSM);
2132 /* Release exclusive LL */
2133 mdt_object_unlock(info, o, lh, 1);
2138 /* Close orphan object anyway */
2139 rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2142 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2143 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2144 LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2146 mdt_object_put(info->mti_env, orphan);
2149 up_write(&o->mot_open_sem);
2151 if (rc == 0) { /* already released */
2152 struct mdt_body *repbody;
2153 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2154 LASSERT(repbody != NULL);
2155 repbody->mbo_valid |= OBD_MD_FLRELEASED;
2159 ldlm_reprocess_all(lease->l_resource);
2160 LDLM_LOCK_PUT(lease);
2168 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
2169 MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
2171 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2173 return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
2176 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2178 struct mdt_object *o = mfd->mfd_object;
2179 struct md_object *next = mdt_object_child(o);
2180 struct md_attr *ma = &info->mti_attr;
2181 int ret = MDT_IOEPOCH_CLOSED;
2186 mode = mfd->mfd_mode;
2188 if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
2189 rc = mdt_hsm_release(info, o, ma);
2191 CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2192 mdt_obd_name(info->mti_mdt),
2193 PFID(mdt_object_fid(o)), rc);
2194 /* continue to close even error occurred. */
2198 if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
2200 ret = mdt_ioepoch_close(info, o);
2201 } else if (mode & MDS_FMODE_EXEC) {
2203 } else if (mode & MDS_FMODE_EPOCH) {
2204 ret = mdt_ioepoch_close(info, o);
2205 } else if (mode & MDS_FMODE_SOM) {
2206 ret = mdt_som_au_close(info, o);
2209 /* Update atime on close only. */
2210 if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
2211 && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
2212 /* Set the atime only. */
2213 ma->ma_valid = MA_INODE;
2214 ma->ma_attr.la_valid = LA_ATIME;
2215 rc = mo_attr_set(info->mti_env, next, ma);
2218 /* If file data is modified, add the dirty flag. */
2219 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2220 rc = mdt_add_dirty_flag(info, o, ma);
2222 ma->ma_need |= MA_INODE;
2223 ma->ma_valid &= ~MA_INODE;
2225 if (!MFD_CLOSED(mode))
2226 rc = mo_close(info->mti_env, next, ma, mode);
2228 if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
2229 struct mdt_export_data *med;
2231 /* The IOepoch is still opened or SOM update is needed.
2232 * Put mfd back into the list. */
2233 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
2234 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
2235 MDS_FMODE_EPOCH : MDS_FMODE_SOM);
2237 LASSERT(mdt_info_req(info));
2238 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
2239 spin_lock(&med->med_open_lock);
2240 list_add(&mfd->mfd_list, &med->med_open_head);
2241 class_handle_hash_back(&mfd->mfd_handle);
2242 spin_unlock(&med->med_open_lock);
2244 if (ret == MDT_IOEPOCH_OPENED) {
2248 CDEBUG(D_INODE, "Size-on-MDS attribute update is "
2249 "needed on "DFID"\n", PFID(mdt_object_fid(o)));
2252 /* adjust open and lease count */
2253 if (mode & MDS_OPEN_LEASE) {
2254 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2255 atomic_dec(&o->mot_lease_count);
2257 LASSERT(atomic_read(&o->mot_open_count) > 0);
2258 atomic_dec(&o->mot_open_count);
2261 mdt_object_put(info->mti_env, o);
2264 RETURN(rc ? rc : ret);
2267 int mdt_close(struct tgt_session_info *tsi)
2269 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2270 struct ptlrpc_request *req = tgt_ses_req(tsi);
2271 struct mdt_export_data *med;
2272 struct mdt_file_data *mfd;
2273 struct mdt_object *o;
2274 struct md_attr *ma = &info->mti_attr;
2275 struct mdt_body *repbody = NULL;
2279 mdt_counter_incr(req, LPROC_MDT_CLOSE);
2280 /* Close may come with the Size-on-MDS update. Unpack it. */
2281 rc = mdt_close_unpack(info);
2283 GOTO(out, rc = err_serious(rc));
2285 LASSERT(info->mti_ioepoch);
2287 /* These fields are no longer used and are left for compatibility.
2288 * size is always zero */
2289 req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2291 req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2293 rc = req_capsule_server_pack(info->mti_pill);
2294 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2295 mdt_client_compatibility(info);
2297 mdt_fix_reply(info);
2298 mdt_exit_ucred(info);
2299 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2302 /* Continue to close handle even if we can not pack reply */
2304 repbody = req_capsule_server_get(info->mti_pill,
2306 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2308 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2311 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
2312 repbody->mbo_eadatasize = 0;
2313 repbody->mbo_aclsize = 0;
2315 rc = err_serious(rc);
2318 med = &req->rq_export->exp_mdt_data;
2319 spin_lock(&med->med_open_lock);
2320 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2321 req_is_replay(req));
2322 if (mdt_mfd_closed(mfd)) {
2323 spin_unlock(&med->med_open_lock);
2324 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2325 ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
2326 info->mti_ioepoch->handle.cookie);
2327 /** not serious error since bug 3633 */
2330 class_handle_unhash(&mfd->mfd_handle);
2331 list_del_init(&mfd->mfd_list);
2332 spin_unlock(&med->med_open_lock);
2334 /* Do not lose object before last unlink. */
2335 o = mfd->mfd_object;
2336 mdt_object_get(info->mti_env, o);
2337 ret = mdt_mfd_close(info, mfd);
2338 if (repbody != NULL)
2339 rc = mdt_handle_last_unlink(info, o, ma);
2340 mdt_empty_transno(info, rc);
2341 mdt_object_put(info->mti_env, o);
2343 if (repbody != NULL) {
2344 mdt_client_compatibility(info);
2345 rc = mdt_fix_reply(info);
2348 mdt_exit_ucred(info);
2349 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2350 GOTO(out, rc = err_serious(-ENOMEM));
2352 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2353 OBD_FAIL_MDS_CLOSE_NET_REP))
2354 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2356 mdt_thread_info_fini(info);
2357 RETURN(rc ? rc : ret);
2361 * DONE_WRITING rpc handler.
2363 * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
2364 * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
2365 * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
2366 * skip attributes and reconstruct the reply here.
2368 int mdt_done_writing(struct tgt_session_info *tsi)
2370 struct ptlrpc_request *req = tgt_ses_req(tsi);
2371 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2372 struct mdt_body *repbody = NULL;
2373 struct mdt_export_data *med;
2374 struct mdt_file_data *mfd;
2378 rc = req_capsule_server_pack(tsi->tsi_pill);
2380 GOTO(out, rc = err_serious(rc));
2382 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2383 repbody->mbo_eadatasize = 0;
2384 repbody->mbo_aclsize = 0;
2386 /* Done Writing may come with the Size-on-MDS update. Unpack it. */
2387 rc = mdt_close_unpack(info);
2389 GOTO(out, rc = err_serious(rc));
2391 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2392 mdt_exit_ucred(info);
2393 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2396 med = &info->mti_exp->exp_mdt_data;
2397 spin_lock(&med->med_open_lock);
2398 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2399 req_is_replay(req));
2401 spin_unlock(&med->med_open_lock);
2402 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
2403 ": cookie = "LPX64" ioepoch = "LPU64"\n",
2404 PFID(info->mti_rr.rr_fid1),
2405 info->mti_ioepoch->handle.cookie,
2406 info->mti_ioepoch->ioepoch);
2407 /* If this is a replay, reconstruct the transno. */
2408 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2409 rc = info->mti_ioepoch->flags & MF_SOM_AU ?
2411 mdt_empty_transno(info, rc);
2414 GOTO(error_ucred, rc);
2417 LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
2418 mfd->mfd_mode == MDS_FMODE_TRUNC);
2419 class_handle_unhash(&mfd->mfd_handle);
2420 list_del_init(&mfd->mfd_list);
2421 spin_unlock(&med->med_open_lock);
2423 /* Set EPOCH CLOSE flag if not set by client. */
2424 info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
2425 info->mti_attr.ma_valid = 0;
2427 info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
2428 OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2429 if (info->mti_attr.ma_lmm == NULL)
2430 GOTO(error_ucred, rc = -ENOMEM);
2432 rc = mdt_mfd_close(info, mfd);
2434 OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2435 mdt_empty_transno(info, rc);
2437 mdt_exit_ucred(info);
2439 mdt_thread_info_fini(info);