1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
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>
44 # define EXPORT_SYMTAB
46 #define DEBUG_SUBSYSTEM S_MDS
48 #include <lustre_acl.h>
49 #include <lustre_mds.h>
50 #include "mdt_internal.h"
52 /* we do nothing because we do not have refcount now */
53 static void mdt_mfd_get(void *mfdp)
57 /* Create a new mdt_file_data struct, initialize it,
58 * and insert it to global hash table */
59 struct mdt_file_data *mdt_mfd_new(void)
61 struct mdt_file_data *mfd;
66 CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
67 CFS_INIT_LIST_HEAD(&mfd->mfd_list);
68 class_handle_hash(&mfd->mfd_handle, mdt_mfd_get);
74 * Find the mfd pointed to by handle in global hash table.
75 * In case of replay the handle is obsoleted
76 * but mfd can be found in mfd list by that handle
78 struct mdt_file_data *mdt_handle2mfd(struct mdt_thread_info *info,
79 const struct lustre_handle *handle)
81 struct ptlrpc_request *req = mdt_info_req(info);
82 struct mdt_file_data *mfd;
85 LASSERT(handle != NULL);
86 mfd = class_handle2object(handle->cookie);
87 /* during dw/setattr replay the mfd can be found by old handle */
88 if (mfd == NULL && req_is_replay(req)) {
89 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
90 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
91 if (mfd->mfd_old_handle.cookie == handle->cookie)
100 void mdt_mfd_free(struct mdt_file_data *mfd)
102 LASSERT(cfs_list_empty(&mfd->mfd_list));
103 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
106 static int mdt_create_data(struct mdt_thread_info *info,
107 struct mdt_object *p, struct mdt_object *o)
109 struct md_op_spec *spec = &info->mti_spec;
110 struct md_attr *ma = &info->mti_attr;
114 if (!md_should_create(spec->sp_cr_flags))
117 ma->ma_need = MA_INODE | MA_LOV;
119 cfs_down(&o->mot_lov_sem);
120 if (!(o->mot_flags & MOF_LOV_CREATED)) {
121 rc = mdo_create_data(info->mti_env,
122 p ? mdt_object_child(p) : NULL,
123 mdt_object_child(o), spec, ma);
124 if (rc == 0 && ma->ma_valid & MA_LOV)
125 o->mot_flags |= MOF_LOV_CREATED;
127 cfs_up(&o->mot_lov_sem);
131 static int mdt_ioepoch_opened(struct mdt_object *mo)
133 return mo->mot_ioepoch_count;
136 int mdt_object_is_som_enabled(struct mdt_object *mo)
138 return !mo->mot_ioepoch;
142 * Re-enable Size-on-MDS.
143 * Call under ->mot_ioepoch_sem.
145 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
147 if (ioepoch == mo->mot_ioepoch) {
148 LASSERT(!mdt_ioepoch_opened(mo));
155 * Open the IOEpoch. It is allowed if @writecount is not negative.
156 * The epoch and writecount handling is performed under the mot_ioepoch_sem.
158 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
161 struct mdt_device *mdt = info->mti_mdt;
166 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
167 !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
170 cfs_down(&o->mot_ioepoch_sem);
171 if (mdt_ioepoch_opened(o)) {
172 /* Epoch continues even if there is no writers yet. */
173 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
174 o->mot_ioepoch, PFID(mdt_object_fid(o)));
176 /* XXX: ->mdt_ioepoch is not initialized at the mount */
177 cfs_spin_lock(&mdt->mdt_ioepoch_lock);
178 if (mdt->mdt_ioepoch < info->mti_replayepoch)
179 mdt->mdt_ioepoch = info->mti_replayepoch;
181 if (info->mti_replayepoch)
182 o->mot_ioepoch = info->mti_replayepoch;
183 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
184 o->mot_ioepoch = ++mdt->mdt_ioepoch;
186 o->mot_ioepoch = mdt->mdt_ioepoch;
188 cfs_spin_unlock(&mdt->mdt_ioepoch_lock);
190 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
191 o->mot_ioepoch, PFID(mdt_object_fid(o)));
193 o->mot_flags |= MOF_SOM_CREATED;
196 o->mot_ioepoch_count++;
197 cfs_up(&o->mot_ioepoch_sem);
199 /* Cancel Size-on-MDS attributes cached on clients for the open case.
200 * In the truncate case, see mdt_reint_setattr(). */
201 if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
202 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
203 mdt_lock_reg_init(lh, LCK_EX);
204 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
207 mdt_object_unlock(info, o, lh, 1);
213 * Update SOM on-disk attributes.
214 * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
215 * mountid and attributes. If disabling, zero IOEpoch id in lustre-ea.
216 * Call under ->mot_ioepoch_sem.
218 static int mdt_som_attr_set(struct mdt_thread_info *info,
219 struct mdt_object *obj, __u64 ioepoch, int enable)
221 struct md_attr *ma = &info->mti_attr;
225 CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
226 " on "DFID".\n", enable ? "update" : "disabling",
227 ioepoch, PFID(mdt_object_fid(obj)));
229 ma->ma_valid |= MA_SOM;
230 ma->ma_som = &info->mti_u.som.data;
232 struct mdt_device *mdt = info->mti_mdt;
233 struct lu_attr *la = &ma->ma_attr;
235 ma->ma_som->msd_ioepoch = ioepoch;
236 ma->ma_som->msd_size = la->la_valid & LA_SIZE ? la->la_size : 0;
237 ma->ma_som->msd_blocks = la->la_valid & LA_BLOCKS ?
239 ma->ma_som->msd_mountid = mdt->mdt_lut.lut_obd->u.obt.obt_mount_count;
240 ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
242 ma->ma_som->msd_ioepoch = IOEPOCH_INVAL;
243 ma->ma_attr.la_valid &= LA_ATIME;
246 /* Since we have opened the file, it is unnecessary
247 * to check permission when close it. Between the "open"
248 * and "close", maybe someone has changed the file mode
249 * or flags, or the file created mode do not permit wirte,
250 * and so on. Just set MDS_PERM_BYPASS for all the cases. */
251 ma->ma_attr_flags |= MDS_PERM_BYPASS | MDS_SOM;
253 rc = mdt_attr_set(info, obj, ma, 0);
257 /** Perform the eviction specific actions on ioepoch close. */
258 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
259 struct mdt_object *o)
263 cfs_down(&o->mot_ioepoch_sem);
264 CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
265 "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
266 o->mot_ioepoch_count);
267 o->mot_ioepoch_count--;
269 /* If eviction occured set MOF_SOM_RECOV,
270 * if no other epoch holders, disable SOM on disk. */
271 o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
272 if (!mdt_ioepoch_opened(o)) {
273 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
274 mdt_object_som_enable(o, o->mot_ioepoch);
276 cfs_up(&o->mot_ioepoch_sem);
281 * Perform the replay specific actions on ioepoch close.
282 * Skip SOM attribute update if obtained and just forget about the inode state
283 * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
285 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
286 struct mdt_object *o)
288 int rc = MDT_IOEPOCH_CLOSED;
291 cfs_down(&o->mot_ioepoch_sem);
292 CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
293 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
294 o->mot_ioepoch_count--;
296 /* Get an info from the replayed request if client is supposed
297 * to send an Attibute Update, reconstruct @rc if so */
298 if (info->mti_ioepoch->flags & MF_SOM_AU)
299 rc = MDT_IOEPOCH_GETATTR;
301 if (!mdt_ioepoch_opened(o))
302 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
303 cfs_up(&o->mot_ioepoch_sem);
309 * Regular file IOepoch close.
310 * Closes the ioepoch, checks the object state, apply obtained attributes and
311 * re-enable SOM on the object, if possible. Also checks if the recovery is
312 * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
313 * to obtain SOM attributes under the server-side OST locks.
316 * MDT_IOEPOCH_CLOSED if ioepoch is closed.
317 * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
319 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
320 struct mdt_object *o)
322 struct md_attr *tmp_ma;
326 int rc = 0, ret = MDT_IOEPOCH_CLOSED;
329 la = &info->mti_attr.ma_attr;
330 achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
332 cfs_down(&o->mot_ioepoch_sem);
333 o->mot_ioepoch_count--;
335 tmp_ma = &info->mti_u.som.attr;
336 tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
337 tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
338 tmp_ma->ma_som = &info->mti_u.som.data;
339 tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
340 tmp_ma->ma_valid = 0;
341 rc = mo_attr_get(info->mti_env, mdt_object_child(o), tmp_ma);
345 /* Check the on-disk SOM state. */
346 if (o->mot_flags & MOF_SOM_RECOV)
348 else if (!(o->mot_flags & MOF_SOM_CREATED) &&
349 !(tmp_ma->ma_valid & MA_SOM))
352 CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
353 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
355 opened = mdt_ioepoch_opened(o);
357 * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
358 * Skip the check for file with no LOV or for unlink files.
360 if (!opened && tmp_ma->ma_valid & MA_LOV &&
361 !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
363 /* If some previous writer was evicted, re-ask the
364 * client for attributes. Even if attributes are
365 * provided, we cannot believe in them.
366 * Another use case is that there is no SOM cache on
367 * disk -- first access with SOM or there was an MDS
369 ret = MDT_IOEPOCH_GETATTR;
370 else if (o->mot_flags & MOF_SOM_CHANGE)
371 /* Some previous writer changed the attribute.
372 * Do not believe to the current Size-on-MDS
373 * update, re-ask client. */
374 ret = MDT_IOEPOCH_GETATTR;
375 else if (!(la->la_valid & LA_SIZE) && achange)
376 /* Attributes were changed by the last writer
377 * only but no Size-on-MDS update is received.*/
378 ret = MDT_IOEPOCH_GETATTR;
381 if (achange || ret == MDT_IOEPOCH_GETATTR)
382 o->mot_flags |= MOF_SOM_CHANGE;
384 /* If epoch ends and relable SOM attributes are obtained, update them.
385 * Create SOM ea for new files even if there is no attributes obtained
386 * (0-length file). */
387 if (ret == MDT_IOEPOCH_CLOSED && !opened) {
388 if (achange || o->mot_flags & MOF_SOM_CREATED) {
389 LASSERT(achange || !(la->la_valid & LA_SIZE));
390 rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
392 /* Avoid the following setattrs of these attributes,
393 * e.g. for atime update. */
394 info->mti_attr.ma_valid = 0;
396 mdt_object_som_enable(o, o->mot_ioepoch);
399 cfs_up(&o->mot_ioepoch_sem);
400 /* If recovery is needed, tell the client to perform GETATTR under
402 if (ret == MDT_IOEPOCH_GETATTR && recovery) {
403 struct mdt_body *rep;
404 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
405 rep->valid |= OBD_MD_FLGETATTRLOCK;
411 cfs_up(&o->mot_ioepoch_sem);
416 * Close IOEpoch (opened file or FMODE_EPOCH state). It happens if:
417 * - a client closes the IOEpoch;
418 * - a client eviction occured.
420 * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
421 * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
422 * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
425 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
427 struct ptlrpc_request *req = mdt_info_req(info);
430 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
431 !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
434 LASSERT(o->mot_ioepoch_count);
435 LASSERT(info->mti_ioepoch == NULL ||
436 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
438 /* IOEpoch is closed only if client tells about it or eviction occures.
439 * In the replay case, always close the epoch. */
441 RETURN(mdt_ioepoch_close_on_eviction(info, o));
442 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
443 RETURN(mdt_ioepoch_close_on_replay(info, o));
444 if (info->mti_ioepoch->flags & MF_EPOCH_CLOSE)
445 RETURN(mdt_ioepoch_close_reg(info, o));
446 /* IO epoch is not closed. */
447 RETURN(MDT_IOEPOCH_OPENED);
451 * Close FMODE_SOM state, when IOEpoch is already closed and we are waiting for
452 * attribute update. It happens if:
453 * - SOM Attribute Update is obtained;
454 * - the client failed to obtain it and informs MDS about it;
455 * - a client eviction occured.
456 * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
459 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
461 struct ptlrpc_request *req = mdt_info_req(info);
463 int act = MDT_SOM_ENABLE;
467 LASSERT(!req || info->mti_ioepoch);
468 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
469 !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
472 /* No size whereas MF_SOM_CHANGE is set means client failed to
473 * obtain ost attributes, drop the SOM cache on disk if so. */
475 (info->mti_ioepoch &&
476 info->mti_ioepoch->flags & MF_SOM_CHANGE &&
477 !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
478 act = MDT_SOM_DISABLE;
480 cfs_down(&o->mot_ioepoch_sem);
481 /* Mark the object it is the recovery state if we failed to obtain
483 if (act == MDT_SOM_DISABLE)
484 o->mot_flags |= MOF_SOM_RECOV;
486 if (!mdt_ioepoch_opened(o)) {
487 ioepoch = info->mti_ioepoch ?
488 info->mti_ioepoch->ioepoch : o->mot_ioepoch;
490 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
491 rc = mdt_som_attr_set(info, o, ioepoch, act);
492 mdt_object_som_enable(o, ioepoch);
494 cfs_up(&o->mot_ioepoch_sem);
498 int mdt_write_read(struct mdt_object *o)
502 cfs_down(&o->mot_ioepoch_sem);
503 rc = o->mot_writecount;
504 cfs_up(&o->mot_ioepoch_sem);
508 int mdt_write_get(struct mdt_object *o)
512 cfs_down(&o->mot_ioepoch_sem);
513 if (o->mot_writecount < 0)
517 cfs_up(&o->mot_ioepoch_sem);
521 void mdt_write_put(struct mdt_object *o)
524 cfs_down(&o->mot_ioepoch_sem);
526 cfs_up(&o->mot_ioepoch_sem);
530 static int mdt_write_deny(struct mdt_object *o)
534 cfs_down(&o->mot_ioepoch_sem);
535 if (o->mot_writecount > 0)
539 cfs_up(&o->mot_ioepoch_sem);
543 static void mdt_write_allow(struct mdt_object *o)
546 cfs_down(&o->mot_ioepoch_sem);
548 cfs_up(&o->mot_ioepoch_sem);
552 /* there can be no real transaction so prepare the fake one */
553 static void mdt_empty_transno(struct mdt_thread_info* info)
555 struct mdt_device *mdt = info->mti_mdt;
556 struct ptlrpc_request *req = mdt_info_req(info);
559 /* transaction has occurred already */
560 if (lustre_msg_get_transno(req->rq_repmsg) != 0)
563 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
564 if (info->mti_transno == 0) {
565 info->mti_transno = ++ mdt->mdt_lut.lut_last_transno;
567 /* should be replay */
568 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
569 mdt->mdt_lut.lut_last_transno = info->mti_transno;
571 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
573 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
575 req->rq_export->exp_obd->obd_last_committed);
577 req->rq_transno = info->mti_transno;
578 lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
582 void mdt_mfd_set_mode(struct mdt_file_data *mfd, int mode)
584 LASSERT(mfd != NULL);
586 CDEBUG(D_HA, "Change mfd %p mode 0x%x->0x%x\n",
587 mfd, (unsigned int)mfd->mfd_mode, (unsigned int)mode);
589 mfd->mfd_mode = mode;
592 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
593 struct mdt_object *o, __u64 flags, int created)
595 struct ptlrpc_request *req = mdt_info_req(info);
596 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
597 struct mdt_file_data *mfd;
598 struct md_attr *ma = &info->mti_attr;
599 struct lu_attr *la = &ma->ma_attr;
600 struct mdt_body *repbody;
601 int rc = 0, isdir, isreg;
604 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
606 isreg = S_ISREG(la->la_mode);
607 isdir = S_ISDIR(la->la_mode);
608 if (isreg && !(ma->ma_valid & MA_LOV)) {
610 * No EA, check whether it is will set regEA and dirEA since in
611 * above attr get, these size might be zero, so reset it, to
612 * retrieve the MD after create obj.
614 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
617 /* in replay case, p == NULL */
618 rc = mdt_create_data(info, p, o);
623 CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
624 ma->ma_valid, ma->ma_lmm_size);
626 if (ma->ma_valid & MA_LOV) {
627 LASSERT(ma->ma_lmm_size != 0);
628 repbody->eadatasize = ma->ma_lmm_size;
630 repbody->valid |= OBD_MD_FLDIREA;
632 repbody->valid |= OBD_MD_FLEASIZE;
635 if (flags & FMODE_WRITE) {
636 rc = mdt_write_get(o);
638 mdt_ioepoch_open(info, o, created);
639 repbody->ioepoch = o->mot_ioepoch;
641 } else if (flags & MDS_FMODE_EXEC) {
642 rc = mdt_write_deny(o);
647 rc = mo_open(info->mti_env, mdt_object_child(o),
648 created ? flags | MDS_OPEN_CREATED : flags);
655 * Keep a reference on this object for this open, and is
656 * released by mdt_mfd_close().
658 mdt_object_get(info->mti_env, o);
661 * @flags is always not zero. At least it should be FMODE_READ,
662 * FMODE_WRITE or FMODE_EXEC.
667 mdt_mfd_set_mode(mfd, flags);
670 mfd->mfd_xid = req->rq_xid;
673 if (req_is_replay(req)) {
674 struct mdt_file_data *old_mfd;
675 /* Check wheather old cookie already exist in
676 * the list, becasue when do recovery, client
677 * might be disconnected from server, and
678 * restart replay, so there maybe some orphan
679 * mfd here, we should remove them */
680 LASSERT(info->mti_rr.rr_handle != NULL);
681 old_mfd = mdt_handle2mfd(info, info->mti_rr.rr_handle);
683 CDEBUG(D_HA, "del orph mfd %p fid=("DFID") "
684 "cookie=" LPX64"\n", mfd,
685 PFID(mdt_object_fid(mfd->mfd_object)),
686 info->mti_rr.rr_handle->cookie);
687 cfs_spin_lock(&med->med_open_lock);
688 class_handle_unhash(&old_mfd->mfd_handle);
689 cfs_list_del_init(&old_mfd->mfd_list);
690 cfs_spin_unlock(&med->med_open_lock);
691 /* no attr update for that close */
693 mdt_mfd_close(info, old_mfd);
695 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
696 info->mti_rr.rr_handle->cookie);
697 mfd->mfd_old_handle.cookie =
698 info->mti_rr.rr_handle->cookie;
700 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
702 if (req->rq_export->exp_disconnected) {
703 cfs_spin_lock(&med->med_open_lock);
704 class_handle_unhash(&mfd->mfd_handle);
705 cfs_list_del_init(&mfd->mfd_list);
706 cfs_spin_unlock(&med->med_open_lock);
707 mdt_mfd_close(info, mfd);
709 cfs_spin_lock(&med->med_open_lock);
710 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
711 cfs_spin_unlock(&med->med_open_lock);
714 mdt_empty_transno(info);
722 static int mdt_finish_open(struct mdt_thread_info *info,
723 struct mdt_object *p, struct mdt_object *o,
724 __u64 flags, int created, struct ldlm_reply *rep)
726 struct ptlrpc_request *req = mdt_info_req(info);
727 struct obd_export *exp = req->rq_export;
728 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
729 struct md_attr *ma = &info->mti_attr;
730 struct lu_attr *la = &ma->ma_attr;
731 struct mdt_file_data *mfd;
732 struct mdt_body *repbody;
734 int isreg, isdir, islnk;
738 LASSERT(ma->ma_valid & MA_INODE);
740 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
742 isreg = S_ISREG(la->la_mode);
743 isdir = S_ISDIR(la->la_mode);
744 islnk = S_ISLNK(la->la_mode);
745 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
747 if (exp_connect_rmtclient(exp)) {
748 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
750 rc = mdt_pack_remote_perm(info, o, buf);
752 repbody->valid &= ~OBD_MD_FLRMTPERM;
753 repbody->aclsize = 0;
755 repbody->valid |= OBD_MD_FLRMTPERM;
756 repbody->aclsize = sizeof(struct mdt_remote_perm);
759 #ifdef CONFIG_FS_POSIX_ACL
760 else if (exp->exp_connect_flags & OBD_CONNECT_ACL) {
761 const struct lu_env *env = info->mti_env;
762 struct md_object *next = mdt_object_child(o);
763 struct lu_buf *buf = &info->mti_buf;
765 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
766 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
768 if (buf->lb_len > 0) {
769 rc = mo_xattr_get(env, next, buf,
770 XATTR_NAME_ACL_ACCESS);
772 if (rc == -ENODATA) {
773 repbody->aclsize = 0;
774 repbody->valid |= OBD_MD_FLACL;
776 } else if (rc == -EOPNOTSUPP) {
779 CERROR("got acl size: %d\n", rc);
782 repbody->aclsize = rc;
783 repbody->valid |= OBD_MD_FLACL;
790 if (info->mti_mdt->mdt_opts.mo_mds_capa &&
791 exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) {
792 struct lustre_capa *capa;
794 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
796 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
797 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
800 repbody->valid |= OBD_MD_FLMDSCAPA;
802 if (info->mti_mdt->mdt_opts.mo_oss_capa &&
803 exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA &&
804 S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) {
805 struct lustre_capa *capa;
807 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
809 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
810 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
813 repbody->valid |= OBD_MD_FLOSSCAPA;
817 * If we are following a symlink, don't open; and do not return open
818 * handle for special nodes as client required.
820 if (islnk || (!isreg && !isdir &&
821 (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH))) {
822 lustre_msg_set_transno(req->rq_repmsg, 0);
826 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
829 * We need to return the existing object's fid back, so it is done here,
830 * after preparing the reply.
832 if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
835 /* This can't be done earlier, we need to return reply body */
837 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
838 /* We are trying to create or write an existing dir. */
841 } else if (flags & MDS_OPEN_DIRECTORY)
844 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
845 OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
850 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
851 cfs_spin_lock(&med->med_open_lock);
852 cfs_list_for_each(t, &med->med_open_head) {
853 mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
854 if (mfd->mfd_xid == req->rq_xid) {
859 cfs_spin_unlock(&med->med_open_lock);
862 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
863 /*set repbody->ea_size for resent case*/
864 if (ma->ma_valid & MA_LOV) {
865 LASSERT(ma->ma_lmm_size != 0);
866 repbody->eadatasize = ma->ma_lmm_size;
868 repbody->valid |= OBD_MD_FLDIREA;
870 repbody->valid |= OBD_MD_FLEASIZE;
876 rc = mdt_mfd_open(info, p, o, flags, created);
880 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
881 struct lsd_client_data *lcd);
883 void mdt_reconstruct_open(struct mdt_thread_info *info,
884 struct mdt_lock_handle *lhc)
886 const struct lu_env *env = info->mti_env;
887 struct mdt_device *mdt = info->mti_mdt;
888 struct req_capsule *pill = info->mti_pill;
889 struct ptlrpc_request *req = mdt_info_req(info);
890 struct tg_export_data *ted = &req->rq_export->exp_target_data;
891 struct lsd_client_data *lcd = ted->ted_lcd;
892 struct md_attr *ma = &info->mti_attr;
893 struct mdt_reint_record *rr = &info->mti_rr;
894 __u32 flags = info->mti_spec.sp_cr_flags;
895 struct ldlm_reply *ldlm_rep;
896 struct mdt_object *parent;
897 struct mdt_object *child;
898 struct mdt_body *repbody;
902 LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
903 ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
904 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
906 ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
907 ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
909 ma->ma_need = MA_INODE;
910 if (ma->ma_lmm_size > 0)
911 ma->ma_need |= MA_LOV;
915 mdt_req_from_lcd(req, lcd);
916 mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
918 CERROR("This is reconstruct open: disp="LPX64", result=%d\n",
919 ldlm_rep->lock_policy_res1, req->rq_status);
921 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
923 /* We did not create successfully, return error to client. */
924 GOTO(out, rc = req->rq_status);
926 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
927 struct obd_export *exp = req->rq_export;
929 * We failed after creation, but we do not know in which step
930 * we failed. So try to check the child object.
932 parent = mdt_object_find(env, mdt, rr->rr_fid1);
933 if (IS_ERR(parent)) {
934 rc = PTR_ERR(parent);
935 LCONSOLE_WARN("Parent "DFID" lookup error %d."
936 " Evicting client %s with export %s.\n",
937 PFID(rr->rr_fid1), rc,
938 obd_uuid2str(&exp->exp_client_uuid),
939 obd_export_nid2str(exp));
940 mdt_export_evict(exp);
943 child = mdt_object_find(env, mdt, rr->rr_fid2);
946 LCONSOLE_WARN("Child "DFID" lookup error %d."
947 " Evicting client %s with export %s.\n",
948 PFID(mdt_object_fid(child)), rc,
949 obd_uuid2str(&exp->exp_client_uuid),
950 obd_export_nid2str(exp));
951 mdt_object_put(env, parent);
952 mdt_export_evict(exp);
955 rc = mdt_object_exists(child);
957 struct md_object *next;
959 mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA);
960 next = mdt_object_child(child);
961 rc = mo_attr_get(env, next, ma);
963 rc = mdt_finish_open(info, parent, child,
966 /* the child object was created on remote server */
967 repbody->fid1 = *rr->rr_fid2;
968 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
970 } else if (rc == 0) {
971 /* the child does not exist, we should do regular open */
972 mdt_object_put(env, parent);
973 mdt_object_put(env, child);
974 GOTO(regular_open, 0);
976 mdt_object_put(env, parent);
977 mdt_object_put(env, child);
981 /* We did not try to create, so we are a pure open */
982 rc = mdt_reint_open(info, lhc);
988 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
989 LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
992 static int mdt_open_by_fid(struct mdt_thread_info* info,
993 struct ldlm_reply *rep)
995 const struct lu_env *env = info->mti_env;
996 __u32 flags = info->mti_spec.sp_cr_flags;
997 struct mdt_reint_record *rr = &info->mti_rr;
998 struct md_attr *ma = &info->mti_attr;
999 struct mdt_object *o;
1003 o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1005 RETURN(rc = PTR_ERR(o));
1007 rc = mdt_object_exists(o);
1009 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1013 rc = mo_attr_get(env, mdt_object_child(o), ma);
1015 rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1016 } else if (rc == 0) {
1019 /* the child object was created on remote server */
1020 struct mdt_body *repbody;
1021 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1022 repbody->fid1 = *rr->rr_fid2;
1023 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1027 mdt_object_put(info->mti_env, o);
1031 static int mdt_open_anon_by_fid(struct mdt_thread_info *info,
1032 struct ldlm_reply *rep,
1033 struct mdt_lock_handle *lhc)
1035 const struct lu_env *env = info->mti_env;
1036 struct mdt_device *mdt = info->mti_mdt;
1037 __u32 flags = info->mti_spec.sp_cr_flags;
1038 struct mdt_reint_record *rr = &info->mti_rr;
1039 struct md_attr *ma = &info->mti_attr;
1040 struct mdt_object *parent= NULL;
1041 struct mdt_object *o;
1046 if (md_should_create(flags)) {
1047 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1048 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1049 if (IS_ERR(parent)) {
1050 CDEBUG(D_INODE, "Fail to find parent "DFID
1051 " for anonymous created %ld, try to"
1052 " use server-side parent.\n",
1053 PFID(rr->rr_fid1), PTR_ERR(parent));
1058 ma->ma_need |= MA_PFID;
1061 o = mdt_object_find(env, mdt, rr->rr_fid2);
1063 RETURN(rc = PTR_ERR(o));
1065 rc = mdt_object_exists(o);
1067 mdt_set_disposition(info, rep, (DISP_LOOKUP_EXECD |
1069 GOTO(out, rc = -ENOENT);
1070 } else if (rc < 0) {
1071 CERROR("NFS remote open shouldn't happen.\n");
1074 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1078 if (flags & FMODE_WRITE)
1080 else if (flags & MDS_FMODE_EXEC)
1085 mdt_lock_handle_init(lhc);
1086 mdt_lock_reg_init(lhc, lm);
1087 rc = mdt_object_lock(info, o, lhc,
1088 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1093 rc = mo_attr_get(env, mdt_object_child(o), ma);
1097 if (ma->ma_valid & MA_PFID) {
1098 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1099 if (IS_ERR(parent)) {
1100 CDEBUG(D_INODE, "Fail to find parent "DFID
1101 " for anonymous created %ld, try to"
1102 " use system default.\n",
1103 PFID(&ma->ma_pfid), PTR_ERR(parent));
1108 if (flags & MDS_OPEN_LOCK)
1109 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1110 rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1112 if (!(flags & MDS_OPEN_LOCK) || rc)
1113 mdt_object_unlock(info, o, lhc, 1);
1117 mdt_object_put(env, o);
1119 mdt_object_put(env, parent);
1123 int mdt_pin(struct mdt_thread_info* info)
1126 RETURN(err_serious(-EOPNOTSUPP));
1129 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1130 static int mdt_cross_open(struct mdt_thread_info* info,
1131 const struct lu_fid *fid,
1132 struct ldlm_reply *rep, __u32 flags)
1134 struct md_attr *ma = &info->mti_attr;
1135 struct mdt_object *o;
1139 o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1141 RETURN(rc = PTR_ERR(o));
1143 rc = mdt_object_exists(o);
1145 /* Do permission check for cross-open. */
1146 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o),
1147 NULL, flags | MDS_OPEN_CROSS);
1151 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1152 rc = mo_attr_get(info->mti_env, mdt_object_child(o), ma);
1154 rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1155 } else if (rc == 0) {
1157 * Something is wrong here. lookup was positive but there is
1160 CERROR("Cross-ref object doesn't exist!\n");
1163 /* Something is wrong here, the object is on another MDS! */
1164 CERROR("The object isn't on this server! FLD error?\n");
1165 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1167 "Object isn't on this server! FLD error?\n");
1173 mdt_object_put(info->mti_env, o);
1177 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1179 struct mdt_device *mdt = info->mti_mdt;
1180 struct ptlrpc_request *req = mdt_info_req(info);
1181 struct mdt_object *parent;
1182 struct mdt_object *child;
1183 struct mdt_lock_handle *lh;
1184 struct ldlm_reply *ldlm_rep;
1185 struct mdt_body *repbody;
1186 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1187 struct md_attr *ma = &info->mti_attr;
1188 __u64 create_flags = info->mti_spec.sp_cr_flags;
1189 struct mdt_reint_record *rr = &info->mti_rr;
1190 struct lu_name *lname;
1196 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1197 (obd_timeout + 1) / 4);
1199 mdt_counter_incr(req->rq_export, LPROC_MDT_OPEN);
1200 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1202 ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1203 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1205 ma->ma_need = MA_INODE;
1206 if (ma->ma_lmm_size > 0)
1207 ma->ma_need |= MA_LOV;
1211 LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1212 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1214 if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1215 CERROR("file join is not supported anymore.\n");
1216 GOTO(out, result = err_serious(-EOPNOTSUPP));
1218 msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1220 if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1221 info->mti_spec.u.sp_ea.eadata == NULL)
1222 GOTO(out, result = err_serious(-EINVAL));
1224 CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1225 "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1226 PFID(rr->rr_fid1), rr->rr_name,
1227 PFID(rr->rr_fid2), create_flags,
1228 ma->ma_attr.la_mode, msg_flags);
1230 if (req_is_replay(req) ||
1231 (req->rq_export->exp_libclient && create_flags&MDS_OPEN_HAS_EA)) {
1232 /* This is a replay request or from liblustre with ea. */
1233 result = mdt_open_by_fid(info, ldlm_rep);
1235 if (result != -ENOENT) {
1236 if (req->rq_export->exp_libclient &&
1237 create_flags&MDS_OPEN_HAS_EA)
1238 GOTO(out, result = 0);
1242 * We didn't find the correct object, so we need to re-create it
1243 * via a regular replay.
1245 if (!(create_flags & MDS_OPEN_CREAT)) {
1246 DEBUG_REQ(D_ERROR, req,
1247 "OPEN & CREAT not in open replay.");
1248 GOTO(out, result = -EFAULT);
1250 CDEBUG(D_INFO, "Open replay did find object, continue as "
1252 } else if (rr->rr_namelen == 0 && !info->mti_cross_ref &&
1253 create_flags & MDS_OPEN_LOCK) {
1254 result = mdt_open_anon_by_fid(info, ldlm_rep, lhc);
1258 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1259 GOTO(out, result = err_serious(-ENOMEM));
1261 mdt_set_disposition(info, ldlm_rep,
1262 (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1264 if (info->mti_cross_ref) {
1265 /* This is cross-ref open */
1266 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1267 result = mdt_cross_open(info, rr->rr_fid1, ldlm_rep,
1272 lh = &info->mti_lh[MDT_LH_PARENT];
1273 mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1274 LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1276 parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1277 MDS_INODELOCK_UPDATE);
1279 GOTO(out, result = PTR_ERR(parent));
1281 /* get and check version of parent */
1282 result = mdt_version_get_check(info, parent, 0);
1284 GOTO(out_parent, result);
1286 fid_zero(child_fid);
1288 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1289 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1290 lname, child_fid, &info->mti_spec);
1291 LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1292 "looking for "DFID"/%s, result fid="DFID"\n",
1293 PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1295 if (result != 0 && result != -ENOENT && result != -ESTALE)
1296 GOTO(out_parent, result);
1298 if (result == -ENOENT || result == -ESTALE) {
1299 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1300 if (result == -ESTALE) {
1302 * -ESTALE means the parent is a dead(unlinked) dir, so
1303 * it should return -ENOENT to in accordance with the
1304 * original mds implementaion.
1306 GOTO(out_parent, result = -ENOENT);
1308 if (!(create_flags & MDS_OPEN_CREAT))
1309 GOTO(out_parent, result);
1310 *child_fid = *info->mti_rr.rr_fid2;
1311 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1315 * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1316 * return FID back in that case.
1318 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1321 child = mdt_object_find(info->mti_env, mdt, child_fid);
1323 GOTO(out_parent, result = PTR_ERR(child));
1325 /** check version of child */
1326 rc = mdt_version_get_check(info, child, 1);
1328 GOTO(out_child, result = rc);
1330 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1331 if (result == -ENOENT) {
1332 /* save versions in reply */
1333 mdt_version_get_save(info, parent, 0);
1334 mdt_version_get_save(info, child, 1);
1336 /* version of child will be changed */
1337 info->mti_mos = child;
1339 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1340 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1342 /* Let lower layers know what is lock mode on directory. */
1343 info->mti_spec.sp_cr_mode =
1344 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1347 * Do not perform lookup sanity check. We know that name does
1350 info->mti_spec.sp_cr_lookup = 0;
1351 info->mti_spec.sp_feat = &dt_directory_features;
1353 result = mdo_create(info->mti_env,
1354 mdt_object_child(parent),
1356 mdt_object_child(child),
1359 if (result == -ERESTART) {
1360 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1361 GOTO(out_child, result);
1364 GOTO(out_child, result);
1368 /* We have to get attr & lov ea for this object */
1369 result = mo_attr_get(info->mti_env, mdt_object_child(child),
1372 * The object is on remote node, return its FID for remote open.
1374 if (result == -EREMOTE) {
1376 * Check if this lock already was sent to client and
1377 * this is resent case. For resent case do not take lock
1378 * again, use what is already granted.
1380 LASSERT(lhc != NULL);
1382 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1383 struct ldlm_lock *lock;
1385 LASSERT(msg_flags & MSG_RESENT);
1387 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1389 CERROR("Invalid lock handle "LPX64"\n",
1390 lhc->mlh_reg_lh.cookie);
1393 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1394 &lock->l_resource->lr_name));
1395 LDLM_LOCK_PUT(lock);
1398 mdt_lock_handle_init(lhc);
1399 mdt_lock_reg_init(lhc, LCK_PR);
1401 rc = mdt_object_lock(info, child, lhc,
1402 MDS_INODELOCK_LOOKUP,
1405 repbody->fid1 = *mdt_object_fid(child);
1406 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1409 GOTO(out_child, result);
1413 LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1415 /* get openlock if this is not replay and if a client requested it */
1416 if (!req_is_replay(req) && create_flags & MDS_OPEN_LOCK) {
1419 if (create_flags & FMODE_WRITE)
1421 else if (create_flags & MDS_FMODE_EXEC)
1425 mdt_lock_handle_init(lhc);
1426 mdt_lock_reg_init(lhc, lm);
1427 rc = mdt_object_lock(info, child, lhc,
1428 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1432 GOTO(out_child, result);
1435 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1439 /* Try to open it now. */
1440 rc = mdt_finish_open(info, parent, child, create_flags,
1444 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1445 /* openlock was acquired and mdt_finish_open failed -
1446 drop the openlock */
1447 mdt_object_unlock(info, child, lhc, 1);
1451 ma->ma_cookie_size = 0;
1452 info->mti_no_need_trans = 1;
1453 rc = mdo_unlink(info->mti_env,
1454 mdt_object_child(parent),
1455 mdt_object_child(child),
1459 CERROR("Error in cleanup of open\n");
1464 mdt_object_put(info->mti_env, child);
1466 mdt_object_unlock_put(info, parent, lh, result || !created);
1468 if (result && result != -EREMOTE)
1469 lustre_msg_set_transno(req->rq_repmsg, 0);
1473 #define MFD_CLOSED(mode) (((mode) & ~(FMODE_EPOCH | FMODE_SOM | \
1474 FMODE_TRUNC)) == FMODE_CLOSED)
1476 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1478 return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1481 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1483 struct mdt_object *o = mfd->mfd_object;
1484 struct md_object *next = mdt_object_child(o);
1485 struct md_attr *ma = &info->mti_attr;
1486 int ret = MDT_IOEPOCH_CLOSED;
1491 mode = mfd->mfd_mode;
1493 if ((mode & FMODE_WRITE) || (mode & FMODE_TRUNC)) {
1495 ret = mdt_ioepoch_close(info, o);
1496 } else if (mode & MDS_FMODE_EXEC) {
1498 } else if (mode & FMODE_EPOCH) {
1499 ret = mdt_ioepoch_close(info, o);
1500 } else if (mode & FMODE_SOM) {
1501 ret = mdt_som_au_close(info, o);
1504 /* Update atime on close only. */
1505 if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1506 && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1507 /* Set the atime only. */
1508 ma->ma_valid = MA_INODE;
1509 ma->ma_attr.la_valid = LA_ATIME;
1510 rc = mo_attr_set(info->mti_env, next, ma);
1513 ma->ma_need |= MA_INODE;
1514 ma->ma_valid &= ~MA_INODE;
1516 if (!MFD_CLOSED(mode))
1517 rc = mo_close(info->mti_env, next, ma);
1519 if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
1520 struct mdt_export_data *med;
1522 /* The IOepoch is still opened or SOM update is needed.
1523 * Put mfd back into the list. */
1524 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
1525 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
1526 FMODE_EPOCH : FMODE_SOM);
1528 LASSERT(mdt_info_req(info));
1529 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
1530 cfs_spin_lock(&med->med_open_lock);
1531 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
1532 class_handle_hash_back(&mfd->mfd_handle);
1533 cfs_spin_unlock(&med->med_open_lock);
1535 if (ret == MDT_IOEPOCH_OPENED) {
1539 CDEBUG(D_INODE, "Size-on-MDS attribute update is "
1540 "needed on "DFID"\n", PFID(mdt_object_fid(o)));
1544 mdt_object_put(info->mti_env, o);
1547 RETURN(rc ? rc : ret);
1550 int mdt_close(struct mdt_thread_info *info)
1552 struct mdt_export_data *med;
1553 struct mdt_file_data *mfd;
1554 struct mdt_object *o;
1555 struct md_attr *ma = &info->mti_attr;
1556 struct mdt_body *repbody = NULL;
1557 struct ptlrpc_request *req = mdt_info_req(info);
1561 mdt_counter_incr(req->rq_export, LPROC_MDT_CLOSE);
1562 /* Close may come with the Size-on-MDS update. Unpack it. */
1563 rc = mdt_close_unpack(info);
1565 RETURN(err_serious(rc));
1567 LASSERT(info->mti_ioepoch);
1569 req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
1570 info->mti_mdt->mdt_max_mdsize);
1571 req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
1572 info->mti_mdt->mdt_max_cookiesize);
1573 rc = req_capsule_server_pack(info->mti_pill);
1574 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
1576 mdt_shrink_reply(info);
1577 RETURN(lustre_msg_get_status(req->rq_repmsg));
1580 /* Continue to close handle even if we can not pack reply */
1582 repbody = req_capsule_server_get(info->mti_pill,
1584 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
1586 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
1589 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
1591 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
1594 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
1595 repbody->eadatasize = 0;
1596 repbody->aclsize = 0;
1598 rc = err_serious(rc);
1601 med = &req->rq_export->exp_mdt_data;
1602 cfs_spin_lock(&med->med_open_lock);
1603 mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1604 if (mdt_mfd_closed(mfd)) {
1605 cfs_spin_unlock(&med->med_open_lock);
1606 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
1607 ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
1608 info->mti_ioepoch->handle.cookie);
1609 /** not serious error since bug 3633 */
1612 class_handle_unhash(&mfd->mfd_handle);
1613 cfs_list_del_init(&mfd->mfd_list);
1614 cfs_spin_unlock(&med->med_open_lock);
1616 /* Do not lose object before last unlink. */
1617 o = mfd->mfd_object;
1618 mdt_object_get(info->mti_env, o);
1619 ret = mdt_mfd_close(info, mfd);
1620 if (repbody != NULL)
1621 rc = mdt_handle_last_unlink(info, o, ma);
1622 mdt_empty_transno(info);
1623 mdt_object_put(info->mti_env, o);
1625 if (repbody != NULL)
1626 mdt_shrink_reply(info);
1628 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
1629 RETURN(err_serious(-ENOMEM));
1631 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
1632 OBD_FAIL_MDS_CLOSE_NET_REP))
1633 info->mti_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
1634 RETURN(rc ? rc : ret);
1638 * DONE_WRITING rpc handler.
1640 * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
1641 * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
1642 * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
1643 * skip attributes and reconstruct the reply here.
1645 int mdt_done_writing(struct mdt_thread_info *info)
1647 struct ptlrpc_request *req = mdt_info_req(info);
1648 struct mdt_body *repbody = NULL;
1649 struct mdt_export_data *med;
1650 struct mdt_file_data *mfd;
1654 rc = req_capsule_server_pack(info->mti_pill);
1656 RETURN(err_serious(rc));
1658 repbody = req_capsule_server_get(info->mti_pill,
1660 repbody->eadatasize = 0;
1661 repbody->aclsize = 0;
1663 /* Done Writing may come with the Size-on-MDS update. Unpack it. */
1664 rc = mdt_close_unpack(info);
1666 RETURN(err_serious(rc));
1668 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL))
1669 RETURN(lustre_msg_get_status(req->rq_repmsg));
1671 med = &info->mti_exp->exp_mdt_data;
1672 cfs_spin_lock(&med->med_open_lock);
1673 mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1675 cfs_spin_unlock(&med->med_open_lock);
1676 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
1677 ": cookie = "LPX64" ioepoch = "LPU64"\n",
1678 PFID(info->mti_rr.rr_fid1),
1679 info->mti_ioepoch->handle.cookie,
1680 info->mti_ioepoch->ioepoch);
1681 /* If this is a replay, reconstruct the transno. */
1682 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1683 mdt_empty_transno(info);
1684 RETURN(info->mti_ioepoch->flags & MF_SOM_AU ?
1690 LASSERT(mfd->mfd_mode == FMODE_EPOCH ||
1691 mfd->mfd_mode == FMODE_TRUNC);
1692 class_handle_unhash(&mfd->mfd_handle);
1693 cfs_list_del_init(&mfd->mfd_list);
1694 cfs_spin_unlock(&med->med_open_lock);
1696 /* Set EPOCH CLOSE flag if not set by client. */
1697 info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
1698 info->mti_attr.ma_valid = 0;
1700 info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
1701 OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1702 if (info->mti_attr.ma_lmm == NULL)
1705 rc = mdt_mfd_close(info, mfd);
1707 OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1708 mdt_empty_transno(info);