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, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_open.c
38 * Lustre Metadata Target (mdt) open/close file handling
40 * Author: Huang Hua <huanghua@clusterfs.com>
43 #define DEBUG_SUBSYSTEM S_MDS
45 #include <lustre_acl.h>
46 #include <lustre_mds.h>
47 #include "mdt_internal.h"
48 #include <lustre_nodemap.h>
50 /* we do nothing because we do not have refcount now */
51 static void mdt_mfd_get(void *mfdp)
55 static struct portals_handle_ops mfd_handle_ops = {
56 .hop_addref = mdt_mfd_get,
60 /* Create a new mdt_file_data struct, initialize it,
61 * and insert it to global hash table */
62 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
64 struct mdt_file_data *mfd;
69 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
70 mfd->mfd_handle.h_owner = med;
71 INIT_LIST_HEAD(&mfd->mfd_list);
72 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
79 * Find the mfd pointed to by handle in global hash table.
80 * In case of replay the handle is obsoleted
81 * but mfd can be found in mfd list by that handle.
82 * Callers need to be holding med_open_lock.
84 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
85 const struct lustre_handle *handle,
86 bool is_replay_or_resent)
88 struct mdt_file_data *mfd;
91 LASSERT(handle != NULL);
92 mfd = class_handle2object(handle->cookie, med);
93 /* during dw/setattr replay the mfd can be found by old handle */
94 if (mfd == NULL && is_replay_or_resent) {
95 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
96 if (mfd->mfd_old_handle.cookie == handle->cookie)
106 void mdt_mfd_free(struct mdt_file_data *mfd)
108 LASSERT(list_empty(&mfd->mfd_list));
109 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
112 static int mdt_create_data(struct mdt_thread_info *info,
113 struct mdt_object *p, struct mdt_object *o)
115 struct md_op_spec *spec = &info->mti_spec;
116 struct md_attr *ma = &info->mti_attr;
120 if (!md_should_create(spec->sp_cr_flags))
123 ma->ma_need = MA_INODE | MA_LOV;
125 mutex_lock(&o->mot_lov_mutex);
126 if (!(o->mot_flags & MOF_LOV_CREATED)) {
127 rc = mdo_create_data(info->mti_env,
128 p ? mdt_object_child(p) : NULL,
129 mdt_object_child(o), spec, ma);
131 rc = mdt_attr_get_complex(info, o, ma);
133 if (rc == 0 && ma->ma_valid & MA_LOV)
134 o->mot_flags |= MOF_LOV_CREATED;
137 mutex_unlock(&o->mot_lov_mutex);
141 static int mdt_ioepoch_opened(struct mdt_object *mo)
143 return mo->mot_ioepoch_count;
146 int mdt_object_is_som_enabled(struct mdt_object *mo)
148 return !mo->mot_ioepoch;
152 * Re-enable Size-on-MDS.
153 * Call under ->mot_ioepoch_mutex.
155 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
157 if (ioepoch == mo->mot_ioepoch) {
158 LASSERT(!mdt_ioepoch_opened(mo));
165 * Open the IOEpoch. It is allowed if @writecount is not negative.
166 * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
168 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
171 struct mdt_device *mdt = info->mti_mdt;
176 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
177 !S_ISREG(lu_object_attr(&o->mot_obj)))
180 mutex_lock(&o->mot_ioepoch_mutex);
181 if (mdt_ioepoch_opened(o)) {
182 /* Epoch continues even if there is no writers yet. */
183 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
184 o->mot_ioepoch, PFID(mdt_object_fid(o)));
186 /* XXX: ->mdt_ioepoch is not initialized at the mount */
187 spin_lock(&mdt->mdt_ioepoch_lock);
188 if (mdt->mdt_ioepoch < info->mti_replayepoch)
189 mdt->mdt_ioepoch = info->mti_replayepoch;
191 if (info->mti_replayepoch)
192 o->mot_ioepoch = info->mti_replayepoch;
193 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
194 o->mot_ioepoch = ++mdt->mdt_ioepoch;
196 o->mot_ioepoch = mdt->mdt_ioepoch;
198 spin_unlock(&mdt->mdt_ioepoch_lock);
200 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
201 o->mot_ioepoch, PFID(mdt_object_fid(o)));
203 o->mot_flags |= MOF_SOM_CREATED;
206 o->mot_ioepoch_count++;
207 mutex_unlock(&o->mot_ioepoch_mutex);
209 /* Cancel Size-on-MDS attributes cached on clients for the open case.
210 * In the truncate case, see mdt_reint_setattr(). */
211 if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
212 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
213 mdt_lock_reg_init(lh, LCK_EX);
214 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
217 mdt_object_unlock(info, o, lh, 1);
223 * Update SOM on-disk attributes.
224 * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
225 * mountid and attributes. If disabling, clean SOM xattr.
226 * Call under ->mot_ioepoch_mutex.
228 static int mdt_som_attr_set(struct mdt_thread_info *info,
229 struct mdt_object *obj, __u64 ioepoch, bool enable)
231 struct md_object *next = mdt_object_child(obj);
235 CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
236 " on "DFID".\n", enable ? "update" : "disabling",
237 ioepoch, PFID(mdt_object_fid(obj)));
240 struct lu_buf *buf = &info->mti_buf;
241 struct som_attrs *attrs;
242 struct md_attr *ma = &info->mti_attr;
243 struct lu_attr *la = &ma->ma_attr;
244 struct obd_device *obd = info->mti_mdt->mdt_lut.lut_obd;
246 attrs = (struct som_attrs *)info->mti_xattr_buf;
247 CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*attrs));
249 /* pack SOM attributes */
250 memset(attrs, 0, sizeof(*attrs));
251 attrs->som_ioepoch = ioepoch;
252 attrs->som_mountid = obd->u.obt.obt_mount_count;
253 if ((la->la_valid & LA_SIZE) != 0)
254 attrs->som_size = la->la_size;
255 if ((la->la_valid & LA_BLOCKS) != 0)
256 attrs->som_blocks = la->la_blocks;
257 lustre_som_swab(attrs);
259 /* update SOM attributes */
261 buf->lb_len = sizeof(*attrs);
262 rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_SOM, 0);
264 /* delete SOM attributes */
265 rc = mo_xattr_del(info->mti_env, next, XATTR_NAME_SOM);
271 /** Perform the eviction specific actions on ioepoch close. */
272 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
273 struct mdt_object *o)
277 mutex_lock(&o->mot_ioepoch_mutex);
278 CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
279 "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
280 o->mot_ioepoch_count);
281 o->mot_ioepoch_count--;
283 /* If eviction occured set MOF_SOM_RECOV,
284 * if no other epoch holders, disable SOM on disk. */
285 o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
286 if (!mdt_ioepoch_opened(o)) {
287 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
288 mdt_object_som_enable(o, o->mot_ioepoch);
290 mutex_unlock(&o->mot_ioepoch_mutex);
295 * Perform the replay specific actions on ioepoch close.
296 * Skip SOM attribute update if obtained and just forget about the inode state
297 * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
299 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
300 struct mdt_object *o)
302 int rc = MDT_IOEPOCH_CLOSED;
305 mutex_lock(&o->mot_ioepoch_mutex);
306 CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
307 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
308 o->mot_ioepoch_count--;
310 /* Get an info from the replayed request if client is supposed
311 * to send an Attibute Update, reconstruct @rc if so */
312 if (info->mti_ioepoch->flags & MF_SOM_AU)
313 rc = MDT_IOEPOCH_GETATTR;
315 if (!mdt_ioepoch_opened(o))
316 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
317 mutex_unlock(&o->mot_ioepoch_mutex);
323 * Regular file IOepoch close.
324 * Closes the ioepoch, checks the object state, apply obtained attributes and
325 * re-enable SOM on the object, if possible. Also checks if the recovery is
326 * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
327 * to obtain SOM attributes under the server-side OST locks.
330 * MDT_IOEPOCH_CLOSED if ioepoch is closed.
331 * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
333 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
334 struct mdt_object *o)
336 struct md_attr *tmp_ma;
340 int rc = 0, ret = MDT_IOEPOCH_CLOSED;
343 la = &info->mti_attr.ma_attr;
344 achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
346 mutex_lock(&o->mot_ioepoch_mutex);
347 o->mot_ioepoch_count--;
349 tmp_ma = &info->mti_u.som.attr;
350 tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
351 tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
352 tmp_ma->ma_som = &info->mti_u.som.data;
353 tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
354 tmp_ma->ma_valid = 0;
355 rc = mdt_attr_get_complex(info, o, tmp_ma);
359 /* Check the on-disk SOM state. */
360 if (o->mot_flags & MOF_SOM_RECOV)
362 else if (!(o->mot_flags & MOF_SOM_CREATED) &&
363 !(tmp_ma->ma_valid & MA_SOM))
366 CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
367 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
369 opened = mdt_ioepoch_opened(o);
371 * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
372 * Skip the check for file with no LOV or for unlink files.
374 if (!opened && tmp_ma->ma_valid & MA_LOV &&
375 !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
377 /* If some previous writer was evicted, re-ask the
378 * client for attributes. Even if attributes are
379 * provided, we cannot believe in them.
380 * Another use case is that there is no SOM cache on
381 * disk -- first access with SOM or there was an MDS
383 ret = MDT_IOEPOCH_GETATTR;
384 else if (o->mot_flags & MOF_SOM_CHANGE)
385 /* Some previous writer changed the attribute.
386 * Do not believe to the current Size-on-MDS
387 * update, re-ask client. */
388 ret = MDT_IOEPOCH_GETATTR;
389 else if (!(la->la_valid & LA_SIZE) && achange)
390 /* Attributes were changed by the last writer
391 * only but no Size-on-MDS update is received.*/
392 ret = MDT_IOEPOCH_GETATTR;
395 if (achange || ret == MDT_IOEPOCH_GETATTR)
396 o->mot_flags |= MOF_SOM_CHANGE;
398 /* If epoch ends and relable SOM attributes are obtained, update them.
399 * Create SOM ea for new files even if there is no attributes obtained
400 * (0-length file). */
401 if (ret == MDT_IOEPOCH_CLOSED && !opened) {
402 if (achange || o->mot_flags & MOF_SOM_CREATED) {
403 LASSERT(achange || !(la->la_valid & LA_SIZE));
404 rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
406 /* Avoid the following setattrs of these attributes,
407 * e.g. for atime update. */
408 info->mti_attr.ma_valid = 0;
410 mdt_object_som_enable(o, o->mot_ioepoch);
413 mutex_unlock(&o->mot_ioepoch_mutex);
414 /* If recovery is needed, tell the client to perform GETATTR under
416 if (ret == MDT_IOEPOCH_GETATTR && recovery) {
417 struct mdt_body *rep;
418 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
419 rep->mbo_valid |= OBD_MD_FLGETATTRLOCK;
425 mutex_unlock(&o->mot_ioepoch_mutex);
430 * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
431 * - a client closes the IOEpoch;
432 * - a client eviction occured.
434 * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
435 * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
436 * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
439 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
441 struct ptlrpc_request *req = mdt_info_req(info);
444 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
445 !S_ISREG(lu_object_attr(&o->mot_obj)))
448 LASSERT(o->mot_ioepoch_count);
449 LASSERT(info->mti_ioepoch == NULL ||
450 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
452 /* IOEpoch is closed only if client tells about it or eviction occures.
453 * In the replay case, always close the epoch. */
455 RETURN(mdt_ioepoch_close_on_eviction(info, o));
456 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
457 RETURN(mdt_ioepoch_close_on_replay(info, o));
458 if (info->mti_ioepoch && (info->mti_ioepoch->flags & MF_EPOCH_CLOSE))
459 RETURN(mdt_ioepoch_close_reg(info, o));
460 /* IO epoch is not closed. */
461 RETURN(MDT_IOEPOCH_OPENED);
465 * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
466 * for attribute update. It happens if:
467 * - SOM Attribute Update is obtained;
468 * - the client failed to obtain it and informs MDS about it;
469 * - a client eviction occured.
470 * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
473 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
475 struct ptlrpc_request *req = mdt_info_req(info);
477 int act = MDT_SOM_ENABLE;
481 LASSERT(!req || info->mti_ioepoch);
482 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
483 !S_ISREG(lu_object_attr(&o->mot_obj)))
486 /* No size whereas MF_SOM_CHANGE is set means client failed to
487 * obtain ost attributes, drop the SOM cache on disk if so. */
489 (info->mti_ioepoch &&
490 info->mti_ioepoch->flags & MF_SOM_CHANGE &&
491 !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
492 act = MDT_SOM_DISABLE;
494 mutex_lock(&o->mot_ioepoch_mutex);
495 /* Mark the object it is the recovery state if we failed to obtain
497 if (act == MDT_SOM_DISABLE)
498 o->mot_flags |= MOF_SOM_RECOV;
500 if (!mdt_ioepoch_opened(o)) {
501 ioepoch = info->mti_ioepoch ?
502 info->mti_ioepoch->ioepoch : o->mot_ioepoch;
505 && !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
506 rc = mdt_som_attr_set(info, o, ioepoch, act);
507 mdt_object_som_enable(o, ioepoch);
509 mutex_unlock(&o->mot_ioepoch_mutex);
513 int mdt_write_read(struct mdt_object *o)
517 mutex_lock(&o->mot_ioepoch_mutex);
518 rc = o->mot_writecount;
519 mutex_unlock(&o->mot_ioepoch_mutex);
523 int mdt_write_get(struct mdt_object *o)
527 mutex_lock(&o->mot_ioepoch_mutex);
528 if (o->mot_writecount < 0)
532 mutex_unlock(&o->mot_ioepoch_mutex);
536 void mdt_write_put(struct mdt_object *o)
539 mutex_lock(&o->mot_ioepoch_mutex);
541 mutex_unlock(&o->mot_ioepoch_mutex);
545 static int mdt_write_deny(struct mdt_object *o)
549 mutex_lock(&o->mot_ioepoch_mutex);
550 if (o->mot_writecount > 0)
554 mutex_unlock(&o->mot_ioepoch_mutex);
558 static void mdt_write_allow(struct mdt_object *o)
561 mutex_lock(&o->mot_ioepoch_mutex);
563 mutex_unlock(&o->mot_ioepoch_mutex);
567 /* there can be no real transaction so prepare the fake one */
568 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
570 struct mdt_device *mdt = info->mti_mdt;
571 struct ptlrpc_request *req = mdt_info_req(info);
572 struct tg_export_data *ted;
573 struct lsd_client_data *lcd;
576 /* transaction has occurred already */
577 if (lustre_msg_get_transno(req->rq_repmsg) != 0)
580 spin_lock(&mdt->mdt_lut.lut_translock);
582 if (info->mti_transno != 0) {
583 struct obd_export *exp = req->rq_export;
585 CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
586 mdt_obd_name(mdt), info->mti_transno,
587 libcfs_nid2str(exp->exp_connection->c_peer.nid),
589 spin_unlock(&mdt->mdt_lut.lut_translock);
592 } else if (info->mti_transno == 0) {
593 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
595 /* should be replay */
596 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
597 mdt->mdt_lut.lut_last_transno = info->mti_transno;
599 spin_unlock(&mdt->mdt_lut.lut_translock);
601 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
603 req->rq_export->exp_obd->obd_last_committed);
605 req->rq_transno = info->mti_transno;
606 lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
608 /* update lcd in memory only for resent cases */
609 ted = &req->rq_export->exp_target_data;
611 mutex_lock(&ted->ted_lcd_lock);
613 if (info->mti_transno < lcd->lcd_last_transno &&
614 info->mti_transno != 0) {
615 /* This should happen during replay. Do not update
616 * last rcvd info if replay req transno < last transno,
617 * otherwise the following resend(after replay) can not
618 * be checked correctly by xid */
619 mutex_unlock(&ted->ted_lcd_lock);
620 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
621 mdt_obd_name(mdt), info->mti_transno,
622 lcd->lcd_last_transno);
626 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
627 lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
628 if (info->mti_transno != 0)
629 lcd->lcd_last_close_transno = info->mti_transno;
630 lcd->lcd_last_close_xid = req->rq_xid;
631 lcd->lcd_last_close_result = rc;
633 /* VBR: save versions in last_rcvd for reconstruct. */
634 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
636 lcd->lcd_pre_versions[0] = pre_versions[0];
637 lcd->lcd_pre_versions[1] = pre_versions[1];
638 lcd->lcd_pre_versions[2] = pre_versions[2];
639 lcd->lcd_pre_versions[3] = pre_versions[3];
641 if (info->mti_transno != 0)
642 lcd->lcd_last_transno = info->mti_transno;
644 lcd->lcd_last_xid = req->rq_xid;
645 lcd->lcd_last_result = rc;
646 lcd->lcd_last_data = info->mti_opdata;
648 mutex_unlock(&ted->ted_lcd_lock);
653 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
655 LASSERT(mfd != NULL);
657 CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n",
658 PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
660 mfd->mfd_mode = mode;
664 * prep ma_lmm/ma_lmv for md_attr from reply
666 void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
667 struct mdt_object *obj,
670 LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
671 if (S_ISDIR(obj->mot_header.loh_attr)) {
672 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
674 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
677 if (ma->ma_lmv_size > 0)
678 ma->ma_need |= MA_LMV;
680 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
682 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
685 if (ma->ma_lmm_size > 0)
686 ma->ma_need |= MA_LOV;
690 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
691 struct mdt_object *o, __u64 flags, int created,
692 struct ldlm_reply *rep)
694 struct ptlrpc_request *req = mdt_info_req(info);
695 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
696 struct mdt_file_data *mfd;
697 struct md_attr *ma = &info->mti_attr;
698 struct lu_attr *la = &ma->ma_attr;
699 struct mdt_body *repbody;
700 int rc = 0, isdir, isreg;
703 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
705 isreg = S_ISREG(la->la_mode);
706 isdir = S_ISDIR(la->la_mode);
707 if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
709 * No EA, check whether it is will set regEA and dirEA since in
710 * above attr get, these size might be zero, so reset it, to
711 * retrieve the MD after create obj.
713 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
716 /* in replay case, p == NULL */
717 rc = mdt_create_data(info, p, o);
721 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
722 mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
725 CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
726 ma->ma_valid, ma->ma_lmm_size);
728 if (ma->ma_valid & MA_LOV) {
729 LASSERT(ma->ma_lmm_size != 0);
730 repbody->mbo_eadatasize = ma->ma_lmm_size;
732 repbody->mbo_valid |= OBD_MD_FLDIREA;
734 repbody->mbo_valid |= OBD_MD_FLEASIZE;
737 if (ma->ma_valid & MA_LMV) {
738 LASSERT(ma->ma_lmv_size != 0);
739 repbody->mbo_eadatasize = ma->ma_lmv_size;
741 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
744 if (flags & FMODE_WRITE) {
745 rc = mdt_write_get(o);
747 mdt_ioepoch_open(info, o, created);
748 repbody->mbo_ioepoch = o->mot_ioepoch;
750 } else if (flags & MDS_FMODE_EXEC) {
751 rc = mdt_write_deny(o);
756 rc = mo_open(info->mti_env, mdt_object_child(o),
757 created ? flags | MDS_OPEN_CREATED : flags);
761 mfd = mdt_mfd_new(med);
763 GOTO(err_out, rc = -ENOMEM);
766 * Keep a reference on this object for this open, and is
767 * released by mdt_mfd_close().
769 mdt_object_get(info->mti_env, o);
771 mfd->mfd_xid = req->rq_xid;
774 * @flags is always not zero. At least it should be FMODE_READ,
775 * FMODE_WRITE or MDS_FMODE_EXEC.
780 mdt_mfd_set_mode(mfd, flags);
782 atomic_inc(&o->mot_open_count);
783 if (flags & MDS_OPEN_LEASE)
784 atomic_inc(&o->mot_lease_count);
787 if (req_is_replay(req)) {
788 struct mdt_file_data *old_mfd;
789 /* Check wheather old cookie already exist in
790 * the list, becasue when do recovery, client
791 * might be disconnected from server, and
792 * restart replay, so there maybe some orphan
793 * mfd here, we should remove them */
794 LASSERT(info->mti_rr.rr_handle != NULL);
795 spin_lock(&med->med_open_lock);
796 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
797 if (old_mfd != NULL) {
798 CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
799 "cookie = "LPX64"\n", mfd,
800 PFID(mdt_object_fid(mfd->mfd_object)),
801 info->mti_rr.rr_handle->cookie);
802 class_handle_unhash(&old_mfd->mfd_handle);
803 list_del_init(&old_mfd->mfd_list);
804 spin_unlock(&med->med_open_lock);
805 /* no attr update for that close */
807 ma->ma_valid |= MA_FLAGS;
808 ma->ma_attr_flags |= MDS_RECOV_OPEN;
809 mdt_mfd_close(info, old_mfd);
810 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
811 ma->ma_valid &= ~MA_FLAGS;
813 spin_unlock(&med->med_open_lock);
814 CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
815 "cookie = "LPX64"\n",
816 PFID(mdt_object_fid(mfd->mfd_object)),
817 info->mti_rr.rr_handle->cookie);
820 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
821 info->mti_rr.rr_handle->cookie);
823 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
826 repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
828 if (req->rq_export->exp_disconnected) {
829 spin_lock(&med->med_open_lock);
830 class_handle_unhash(&mfd->mfd_handle);
831 list_del_init(&mfd->mfd_list);
832 spin_unlock(&med->med_open_lock);
833 mdt_mfd_close(info, mfd);
835 spin_lock(&med->med_open_lock);
836 list_add(&mfd->mfd_list, &med->med_open_head);
837 spin_unlock(&med->med_open_lock);
840 mdt_empty_transno(info, rc);
845 if (flags & FMODE_WRITE)
846 /* XXX We also need to close io epoch here.
847 * See LU-1220 - green */
849 else if (flags & MDS_FMODE_EXEC)
855 int mdt_finish_open(struct mdt_thread_info *info,
856 struct mdt_object *p, struct mdt_object *o,
857 __u64 flags, int created, struct ldlm_reply *rep)
859 struct ptlrpc_request *req = mdt_info_req(info);
860 struct obd_export *exp = req->rq_export;
861 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
862 struct md_attr *ma = &info->mti_attr;
863 struct lu_attr *la = &ma->ma_attr;
864 struct mdt_file_data *mfd;
865 struct mdt_body *repbody;
867 int isreg, isdir, islnk;
871 LASSERT(ma->ma_valid & MA_INODE);
873 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
875 isreg = S_ISREG(la->la_mode);
876 isdir = S_ISDIR(la->la_mode);
877 islnk = S_ISLNK(la->la_mode);
878 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
880 /* LU-2275, simulate broken behaviour (esp. prevalent in
881 * pre-2.4 servers where a very strange reply is sent on error
882 * that looks like it was actually almost succesful and a failure at the
884 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
885 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
889 if (flags & MDS_OPEN_LOCK)
890 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
895 if (exp_connect_rmtclient(exp)) {
896 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
898 rc = mdt_pack_remote_perm(info, o, buf);
900 repbody->mbo_valid &= ~OBD_MD_FLRMTPERM;
901 repbody->mbo_aclsize = 0;
903 repbody->mbo_valid |= OBD_MD_FLRMTPERM;
904 repbody->mbo_aclsize = sizeof(struct mdt_remote_perm);
907 #ifdef CONFIG_FS_POSIX_ACL
908 else if (exp_connect_flags(exp) & OBD_CONNECT_ACL)
909 rc = mdt_pack_acl2body(info, repbody, o,
910 exp->exp_target_data.ted_nodemap);
913 if (info->mti_mdt->mdt_lut.lut_mds_capa &&
914 exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) {
915 struct lustre_capa *capa;
917 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
919 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
920 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
923 repbody->mbo_valid |= OBD_MD_FLMDSCAPA;
925 if (info->mti_mdt->mdt_lut.lut_oss_capa &&
926 exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA &&
927 S_ISREG(lu_object_attr(&o->mot_obj))) {
928 struct lustre_capa *capa;
930 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
932 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
933 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
936 repbody->mbo_valid |= OBD_MD_FLOSSCAPA;
940 * If we are following a symlink, don't open; and do not return open
941 * handle for special nodes as client required.
943 if (islnk || (!isreg && !isdir &&
944 (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
945 lustre_msg_set_transno(req->rq_repmsg, 0);
950 * We need to return the existing object's fid back, so it is done here,
951 * after preparing the reply.
953 if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
956 /* This can't be done earlier, we need to return reply body */
958 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
959 /* We are trying to create or write an existing dir. */
962 } else if (flags & MDS_OPEN_DIRECTORY)
965 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
966 OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
971 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
972 spin_lock(&med->med_open_lock);
973 list_for_each(t, &med->med_open_head) {
974 mfd = list_entry(t, struct mdt_file_data, mfd_list);
975 if (mfd->mfd_xid == req->rq_xid)
979 spin_unlock(&med->med_open_lock);
982 repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
983 /* set repbody->ea_size for resent case */
984 if (ma->ma_valid & MA_LOV) {
985 LASSERT(ma->ma_lmm_size != 0);
986 repbody->mbo_eadatasize = ma->ma_lmm_size;
988 repbody->mbo_valid |= OBD_MD_FLDIREA;
990 repbody->mbo_valid |= OBD_MD_FLEASIZE;
992 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
997 rc = mdt_mfd_open(info, p, o, flags, created, rep);
999 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
1004 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
1005 struct lsd_client_data *lcd);
1007 void mdt_reconstruct_open(struct mdt_thread_info *info,
1008 struct mdt_lock_handle *lhc)
1010 const struct lu_env *env = info->mti_env;
1011 struct mdt_device *mdt = info->mti_mdt;
1012 struct req_capsule *pill = info->mti_pill;
1013 struct ptlrpc_request *req = mdt_info_req(info);
1014 struct tg_export_data *ted = &req->rq_export->exp_target_data;
1015 struct lsd_client_data *lcd = ted->ted_lcd;
1016 struct md_attr *ma = &info->mti_attr;
1017 struct mdt_reint_record *rr = &info->mti_rr;
1018 __u64 flags = info->mti_spec.sp_cr_flags;
1019 struct ldlm_reply *ldlm_rep;
1020 struct mdt_object *parent;
1021 struct mdt_object *child;
1022 struct mdt_body *repbody;
1026 LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1027 ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1028 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1030 ma->ma_need = MA_INODE | MA_HSM;
1033 mdt_req_from_lcd(req, lcd);
1034 mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1036 CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1037 ldlm_rep->lock_policy_res1, req->rq_status);
1039 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1040 req->rq_status != 0)
1041 /* We did not create successfully, return error to client. */
1042 GOTO(out, rc = req->rq_status);
1044 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1045 struct obd_export *exp = req->rq_export;
1047 * We failed after creation, but we do not know in which step
1048 * we failed. So try to check the child object.
1050 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1051 if (IS_ERR(parent)) {
1052 rc = PTR_ERR(parent);
1053 LCONSOLE_WARN("Parent "DFID" lookup error %d."
1054 " Evicting client %s with export %s.\n",
1055 PFID(rr->rr_fid1), rc,
1056 obd_uuid2str(&exp->exp_client_uuid),
1057 obd_export_nid2str(exp));
1058 mdt_export_evict(exp);
1062 child = mdt_object_find(env, mdt, rr->rr_fid2);
1063 if (IS_ERR(child)) {
1064 rc = PTR_ERR(child);
1065 LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
1066 "evicting client %s with export %s\n",
1067 PFID(rr->rr_fid2), rc,
1068 obd_uuid2str(&exp->exp_client_uuid),
1069 obd_export_nid2str(exp));
1070 mdt_object_put(env, parent);
1071 mdt_export_evict(exp);
1075 if (unlikely(mdt_object_remote(child))) {
1076 /* the child object was created on remote server */
1077 if (!mdt_is_dne_client(exp)) {
1078 /* Return -EIO for old client */
1079 mdt_object_put(env, parent);
1080 mdt_object_put(env, child);
1081 GOTO(out, rc = -EIO);
1083 repbody->mbo_fid1 = *rr->rr_fid2;
1084 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1087 if (mdt_object_exists(child)) {
1088 mdt_set_capainfo(info, 1, rr->rr_fid2,
1090 mdt_prep_ma_buf_from_rep(info, child, ma);
1091 rc = mdt_attr_get_complex(info, child, ma);
1093 rc = mdt_finish_open(info, parent,
1097 /* the child does not exist, we should do
1099 mdt_object_put(env, parent);
1100 mdt_object_put(env, child);
1101 GOTO(regular_open, 0);
1104 mdt_object_put(env, parent);
1105 mdt_object_put(env, child);
1109 /* We did not try to create, so we are a pure open */
1110 rc = mdt_reint_open(info, lhc);
1115 req->rq_status = rc;
1116 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1117 LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1120 int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
1122 __u64 flags = info->mti_spec.sp_cr_flags;
1123 struct mdt_reint_record *rr = &info->mti_rr;
1124 struct md_attr *ma = &info->mti_attr;
1125 struct mdt_object *o;
1129 o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1131 RETURN(rc = PTR_ERR(o));
1133 if (unlikely(mdt_object_remote(o))) {
1134 /* the child object was created on remote server */
1135 struct mdt_body *repbody;
1137 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1140 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1141 repbody->mbo_fid1 = *rr->rr_fid2;
1142 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1145 if (mdt_object_exists(o)) {
1146 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1149 mdt_prep_ma_buf_from_rep(info, o, ma);
1150 rc = mdt_attr_get_complex(info, o, ma);
1152 rc = mdt_finish_open(info, NULL, o, flags, 0,
1159 mdt_object_put(info->mti_env, o);
1163 /* lock object for open */
1164 static int mdt_object_open_lock(struct mdt_thread_info *info,
1165 struct mdt_object *obj,
1166 struct mdt_lock_handle *lhc,
1169 struct md_attr *ma = &info->mti_attr;
1170 __u64 open_flags = info->mti_spec.sp_cr_flags;
1171 ldlm_mode_t lm = LCK_CR;
1172 bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
1173 bool try_layout = false;
1174 bool create_layout = false;
1179 mdt_lock_handle_init(lhc);
1181 if (req_is_replay(mdt_info_req(info)))
1184 if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
1185 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
1186 md_should_create(open_flags))
1187 create_layout = true;
1188 if (exp_connect_layout(info->mti_exp) && !create_layout &&
1189 ma->ma_need & MA_LOV)
1194 /* lease open, acquire write mode of open sem */
1195 down_write(&obj->mot_open_sem);
1197 /* Lease exists and ask for new lease */
1198 if (atomic_read(&obj->mot_lease_count) > 0) {
1199 /* only exclusive open is supported, so lease
1200 * are conflicted to each other */
1201 GOTO(out, rc = -EBUSY);
1204 /* Lease must be with open lock */
1205 if (!(open_flags & MDS_OPEN_LOCK)) {
1206 CERROR("Request lease for file:"DFID ", but open lock "
1207 "is missed, open_flags = "LPO64".\n",
1208 PFID(mdt_object_fid(obj)), open_flags);
1209 GOTO(out, rc = -EPROTO);
1212 /* XXX: only exclusive open is supported. */
1214 *ibits = MDS_INODELOCK_OPEN;
1216 /* never grant LCK_EX layout lock to client */
1218 } else { /* normal open */
1219 /* normal open holds read mode of open sem */
1220 down_read(&obj->mot_open_sem);
1222 if (open_flags & MDS_OPEN_LOCK) {
1223 if (open_flags & FMODE_WRITE)
1225 else if (open_flags & MDS_FMODE_EXEC)
1230 *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
1231 } else if (atomic_read(&obj->mot_lease_count) > 0) {
1232 if (open_flags & FMODE_WRITE)
1238 *ibits = MDS_INODELOCK_OPEN;
1241 lhc = &info->mti_lh[MDT_LH_LOCAL];
1243 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
1244 PFID(mdt_object_fid(obj)),
1245 atomic_read(&obj->mot_open_count), lm);
1248 mdt_lock_reg_init(lhc, lm);
1250 /* one problem to return layout lock on open is that it may result
1251 * in too many layout locks cached on the client side. */
1252 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
1253 /* return lookup lock to validate inode at the client side,
1254 * this is pretty important otherwise mdt will return layout
1255 * lock for each open.
1256 * However this is a double-edged sword because changing
1257 * permission will revoke huge # of LOOKUP locks. */
1258 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
1259 if (!mdt_object_lock_try(info, obj, lhc, *ibits,
1261 *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
1263 rc = mdt_object_lock(info, obj, lhc, *ibits,
1266 } else if (*ibits != 0) {
1267 rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK);
1270 CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
1271 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
1272 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
1274 /* will change layout, revoke layout locks by enqueuing EX lock. */
1275 if (rc == 0 && create_layout) {
1276 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
1278 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
1279 ", open_flags = "LPO64"\n",
1280 PFID(mdt_object_fid(obj)), open_flags);
1282 /* We cannot enqueue another lock for the same resource we
1283 * already have a lock for, due to mechanics of waiting list
1284 * iterating in ldlm, see LU-3601.
1285 * As such we'll drop the open lock we just got above here,
1286 * it's ok not to have this open lock as it's main purpose is to
1287 * flush unused cached client open handles. */
1288 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1289 mdt_object_unlock(info, obj, lhc, 1);
1291 LASSERT(!try_layout);
1292 mdt_lock_handle_init(ll);
1293 mdt_lock_reg_init(ll, LCK_EX);
1294 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
1297 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
1300 /* Check if there is any other open handles after acquiring
1301 * open lock. At this point, caching open handles have been revoked
1303 * XXX: Now only exclusive open is supported. Need to check the
1304 * type of open for generic lease support. */
1305 if (rc == 0 && acq_lease) {
1306 struct ptlrpc_request *req = mdt_info_req(info);
1307 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1308 struct mdt_file_data *mfd;
1309 bool is_replay_or_resent;
1312 /* For lease: application can open a file and then apply lease,
1313 * @handle contains original open handle in that case.
1314 * In recovery, open REQ will be replayed and the lease REQ may
1315 * be resent that means the open handle is already stale, so we
1316 * need to fix it up here by finding new handle. */
1317 is_replay_or_resent = req_is_replay(req) ||
1318 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
1320 /* if the request is _not_ a replay request, rr_handle
1321 * may be used to hold an openhandle which is issuing the
1322 * lease request, so that this openhandle doesn't count. */
1323 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
1324 is_replay_or_resent);
1328 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
1329 PFID(mdt_object_fid(obj)),
1330 atomic_read(&obj->mot_open_count), open_count);
1332 if (atomic_read(&obj->mot_open_count) > open_count)
1333 GOTO(out, rc = -EBUSY);
1341 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1342 struct mdt_object *obj,
1343 struct mdt_lock_handle *lhc,
1344 __u64 ibits, int rc)
1346 __u64 open_flags = info->mti_spec.sp_cr_flags;
1347 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1350 if (req_is_replay(mdt_info_req(info)))
1353 /* Release local lock - the lock put in MDT_LH_LOCAL will never
1354 * return to client side. */
1355 if (lustre_handle_is_used(&ll->mlh_reg_lh))
1356 mdt_object_unlock(info, obj, ll, 1);
1358 ll = &info->mti_lh[MDT_LH_LAYOUT];
1359 /* Release local layout lock, layout was created */
1360 if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1361 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1362 mdt_object_unlock(info, obj, ll, 1);
1365 if (open_flags & MDS_OPEN_LEASE)
1366 up_write(&obj->mot_open_sem);
1368 up_read(&obj->mot_open_sem);
1370 /* Cross-ref case, the lock should be returned to the client */
1371 if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1374 if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
1375 /* for the open request, the lock will only return to client
1376 * if open or layout lock is granted. */
1380 if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1381 struct ldlm_reply *ldlm_rep;
1383 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1384 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1385 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1386 mdt_object_unlock(info, obj, lhc, 1);
1392 * Check release is permitted for the current HSM flags.
1394 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1396 if (!(ma->ma_valid & MA_HSM))
1399 if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1402 if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1408 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1409 struct mdt_lock_handle *lhc)
1411 const struct lu_env *env = info->mti_env;
1412 struct mdt_device *mdt = info->mti_mdt;
1413 __u64 flags = info->mti_spec.sp_cr_flags;
1414 struct mdt_reint_record *rr = &info->mti_rr;
1415 struct md_attr *ma = &info->mti_attr;
1416 struct mdt_object *parent= NULL;
1417 struct mdt_object *o;
1419 int object_locked = 0;
1423 if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1424 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1425 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1426 if (IS_ERR(parent)) {
1427 CDEBUG(D_INODE, "Fail to find parent "DFID
1428 " for anonymous created %ld, try to"
1429 " use server-side parent.\n",
1430 PFID(rr->rr_fid1), PTR_ERR(parent));
1435 ma->ma_need |= MA_PFID;
1438 o = mdt_object_find(env, mdt, rr->rr_fid2);
1440 GOTO(out_parent_put, rc = PTR_ERR(o));
1442 if (mdt_object_remote(o)) {
1443 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1444 mdt_obd_name(info->mti_mdt),
1446 GOTO(out, rc = -EREMOTE);
1447 } else if (!mdt_object_exists(o)) {
1448 mdt_set_disposition(info, rep,
1452 GOTO(out, rc = -ENOENT);
1455 mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1457 mdt_prep_ma_buf_from_rep(info, o, ma);
1458 if (flags & MDS_OPEN_RELEASE)
1459 ma->ma_need |= MA_HSM;
1460 rc = mdt_attr_get_complex(info, o, ma);
1464 /* If a release request, check file flags are fine and ask for an
1465 * exclusive open access. */
1466 if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1467 GOTO(out, rc = -EPERM);
1469 rc = mdt_check_resent_lock(info, o, lhc);
1472 } else if (rc > 0) {
1473 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1476 GOTO(out_unlock, rc);
1479 if (ma->ma_valid & MA_PFID) {
1480 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1481 if (IS_ERR(parent)) {
1482 CDEBUG(D_INODE, "Fail to find parent "DFID
1483 " for anonymous created %ld, try to"
1484 " use system default.\n",
1485 PFID(&ma->ma_pfid), PTR_ERR(parent));
1490 rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1492 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1493 if (flags & MDS_OPEN_LOCK)
1494 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1495 if (flags & MDS_OPEN_LEASE)
1496 mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1498 GOTO(out_unlock, rc);
1502 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1504 mdt_object_put(env, o);
1507 mdt_object_put(env, parent);
1511 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1512 static int mdt_cross_open(struct mdt_thread_info *info,
1513 const struct lu_fid *parent_fid,
1514 const struct lu_fid *fid,
1515 struct ldlm_reply *rep, __u32 flags)
1517 struct md_attr *ma = &info->mti_attr;
1518 struct mdt_object *o;
1522 o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1524 RETURN(rc = PTR_ERR(o));
1526 if (mdt_object_remote(o)) {
1527 /* Something is wrong here, the object is on another MDS! */
1528 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1529 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1530 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1532 "Object isn't on this server! FLD error?\n");
1535 if (mdt_object_exists(o)) {
1536 /* Do permission check for cross-open. */
1537 rc = mo_permission(info->mti_env, NULL,
1538 mdt_object_child(o),
1539 NULL, flags | MDS_OPEN_CROSS);
1543 mdt_prep_ma_buf_from_rep(info, o, ma);
1544 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1545 rc = mdt_attr_get_complex(info, o, ma);
1549 rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1552 * Something is wrong here. lookup was positive but
1553 * there is no object!
1555 CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1556 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1561 mdt_object_put(info->mti_env, o);
1565 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1567 struct mdt_device *mdt = info->mti_mdt;
1568 struct ptlrpc_request *req = mdt_info_req(info);
1569 struct mdt_object *parent;
1570 struct mdt_object *child;
1571 struct mdt_lock_handle *lh;
1572 struct ldlm_reply *ldlm_rep;
1573 struct mdt_body *repbody;
1574 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1575 struct md_attr *ma = &info->mti_attr;
1576 __u64 create_flags = info->mti_spec.sp_cr_flags;
1578 struct mdt_reint_record *rr = &info->mti_rr;
1581 int object_locked = 0;
1585 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1586 (obd_timeout + 1) / 4);
1588 mdt_counter_incr(req, LPROC_MDT_OPEN);
1589 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1591 ma->ma_need = MA_INODE;
1594 LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1595 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1597 if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1598 CERROR("file join is not supported anymore.\n");
1599 GOTO(out, result = err_serious(-EOPNOTSUPP));
1601 msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1603 if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1604 info->mti_spec.u.sp_ea.eadata == NULL)
1605 GOTO(out, result = err_serious(-EINVAL));
1607 CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1608 "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1609 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1610 PFID(rr->rr_fid2), create_flags,
1611 ma->ma_attr.la_mode, msg_flags);
1613 if (info->mti_cross_ref) {
1614 /* This is cross-ref open */
1615 mdt_set_disposition(info, ldlm_rep,
1616 (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1618 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1619 ldlm_rep, create_flags);
1621 } else if (req_is_replay(req) ||
1622 (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1623 /* This is a replay request or from liblustre with ea. */
1624 result = mdt_open_by_fid(info, ldlm_rep);
1626 if (result != -ENOENT) {
1627 if (req->rq_export->exp_libclient &&
1628 create_flags & MDS_OPEN_HAS_EA)
1629 GOTO(out, result = 0);
1632 /* We didn't find the correct object, so we need to re-create it
1633 * via a regular replay. */
1634 if (!(create_flags & MDS_OPEN_CREAT)) {
1635 DEBUG_REQ(D_ERROR, req,
1636 "OPEN & CREAT not in open replay/by_fid.");
1637 GOTO(out, result = -EFAULT);
1639 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1640 } else if (create_flags & (MDS_OPEN_BY_FID | MDS_OPEN_LOCK)) {
1642 * MDS_OPEN_LOCK is checked for backward compatibility with 2.1
1645 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1647 CDEBUG(D_INFO, "no object for "DFID": %d\n",
1648 PFID(rr->rr_fid2), result);
1652 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1653 GOTO(out, result = err_serious(-ENOMEM));
1655 mdt_set_disposition(info, ldlm_rep,
1656 (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1658 if (!lu_name_is_valid(&rr->rr_name))
1659 GOTO(out, result = -EPROTO);
1661 lh = &info->mti_lh[MDT_LH_PARENT];
1662 mdt_lock_pdo_init(lh,
1663 (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1666 parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1667 MDS_INODELOCK_UPDATE);
1669 GOTO(out, result = PTR_ERR(parent));
1671 /* get and check version of parent */
1672 result = mdt_version_get_check(info, parent, 0);
1674 GOTO(out_parent, result);
1676 fid_zero(child_fid);
1678 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1679 &rr->rr_name, child_fid, &info->mti_spec);
1681 LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1682 "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1683 PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1686 if (result != 0 && result != -ENOENT && result != -ESTALE)
1687 GOTO(out_parent, result);
1689 if (result == -ENOENT || result == -ESTALE) {
1690 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1691 if (result == -ESTALE) {
1693 * -ESTALE means the parent is a dead(unlinked) dir, so
1694 * it should return -ENOENT to in accordance with the
1695 * original mds implementaion.
1697 GOTO(out_parent, result = -ENOENT);
1699 if (!(create_flags & MDS_OPEN_CREAT))
1700 GOTO(out_parent, result);
1701 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1702 GOTO(out_parent, result = -EROFS);
1703 *child_fid = *info->mti_rr.rr_fid2;
1704 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1706 /* In the function below, .hs_keycmp resolves to
1707 * lu_obj_hop_keycmp() */
1708 /* coverity[overrun-buffer-val] */
1709 child = mdt_object_new(info->mti_env, mdt, child_fid);
1712 * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1713 * return FID back in that case.
1715 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1716 child = mdt_object_find(info->mti_env, mdt, child_fid);
1719 GOTO(out_parent, result = PTR_ERR(child));
1721 /** check version of child */
1722 rc = mdt_version_get_check(info, child, 1);
1724 GOTO(out_child, result = rc);
1726 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1727 if (result == -ENOENT) {
1728 /* Create under OBF and .lustre is not permitted */
1729 if (!fid_is_md_operative(rr->rr_fid1))
1730 GOTO(out_child, result = -EPERM);
1732 /* save versions in reply */
1733 mdt_version_get_save(info, parent, 0);
1734 mdt_version_get_save(info, child, 1);
1736 /* version of child will be changed */
1737 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1739 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1740 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1742 /* Let lower layers know what is lock mode on directory. */
1743 info->mti_spec.sp_cr_mode =
1744 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1747 * Do not perform lookup sanity check. We know that name does
1750 info->mti_spec.sp_cr_lookup = 0;
1751 info->mti_spec.sp_feat = &dt_directory_features;
1753 result = mdo_create(info->mti_env,
1754 mdt_object_child(parent),
1756 mdt_object_child(child),
1759 if (result == -ERESTART) {
1760 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1761 GOTO(out_child, result);
1763 mdt_prep_ma_buf_from_rep(info, child, ma);
1764 /* XXX: we should call this once, see few lines below */
1766 result = mdt_attr_get_complex(info, child, ma);
1769 GOTO(out_child, result);
1774 * The object is on remote node, return its FID for remote open.
1776 if (mdt_object_remote(child)) {
1778 * Check if this lock already was sent to client and
1779 * this is resent case. For resent case do not take lock
1780 * again, use what is already granted.
1782 LASSERT(lhc != NULL);
1784 rc = mdt_check_resent_lock(info, child, lhc);
1786 GOTO(out_child, result = rc);
1787 } else if (rc > 0) {
1788 mdt_lock_handle_init(lhc);
1789 mdt_lock_reg_init(lhc, LCK_PR);
1791 rc = mdt_object_lock(info, child, lhc,
1792 MDS_INODELOCK_LOOKUP,
1795 repbody->mbo_fid1 = *mdt_object_fid(child);
1796 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1800 result = -MDT_EREMOTE_OPEN;
1801 GOTO(out_child, result);
1802 } else if (mdt_object_exists(child)) {
1803 /* We have to get attr & LOV EA & HSM for this
1805 mdt_prep_ma_buf_from_rep(info, child, ma);
1806 ma->ma_need |= MA_HSM;
1807 result = mdt_attr_get_complex(info, child, ma);
1809 GOTO(out_child, result);
1811 /* Object does not exist. Likely FS corruption. */
1812 CERROR("%s: name '"DNAME"' present, but FID "
1813 DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1814 PNAME(&rr->rr_name), PFID(child_fid));
1815 GOTO(out_child, result = -EIO);
1819 rc = mdt_check_resent_lock(info, child, lhc);
1821 GOTO(out_child, result = rc);
1822 } else if (rc == 0) {
1823 /* the open lock might already be gotten in
1824 * ldlm_handle_enqueue() */
1825 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1826 if (create_flags & MDS_OPEN_LOCK)
1827 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1829 /* get openlock if this isn't replay and client requested it */
1830 if (!req_is_replay(req)) {
1831 rc = mdt_object_open_lock(info, child, lhc, &ibits);
1834 GOTO(out_child_unlock, result = rc);
1835 else if (create_flags & MDS_OPEN_LOCK)
1836 mdt_set_disposition(info, ldlm_rep,
1840 /* Try to open it now. */
1841 rc = mdt_finish_open(info, parent, child, create_flags,
1845 /* openlock will be released if mdt_finish_open failed */
1846 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1848 if (created && create_flags & MDS_OPEN_VOLATILE) {
1849 CERROR("%s: cannot open volatile file "DFID", orphan "
1850 "file will be left in PENDING directory until "
1851 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1852 PFID(mdt_object_fid(child)), rc);
1853 GOTO(out_child_unlock, result);
1859 rc = mdo_unlink(info->mti_env,
1860 mdt_object_child(parent),
1861 mdt_object_child(child),
1863 &info->mti_attr, 0);
1865 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1866 mdt_obd_name(info->mti_mdt),
1867 PFID(mdt_object_fid(child)), rc);
1868 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1874 mdt_object_open_unlock(info, child, lhc, ibits, result);
1876 mdt_object_put(info->mti_env, child);
1878 mdt_object_unlock_put(info, parent, lh, result || !created);
1881 lustre_msg_set_transno(req->rq_repmsg, 0);
1886 * Create an orphan object use local root.
1888 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1889 struct mdt_device *mdt,
1890 const struct lu_fid *fid,
1891 struct md_attr *attr, fmode_t fmode)
1893 const struct lu_env *env = info->mti_env;
1894 struct md_op_spec *spec = &info->mti_spec;
1895 struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1896 struct mdt_object *obj = NULL;
1897 struct mdt_object *local_root;
1898 static const struct lu_name lname = {
1899 .ln_name = "i_am_nobody",
1900 .ln_namelen = sizeof("i_am_nobody") - 1,
1902 struct lu_ucred *uc;
1903 cfs_cap_t uc_cap_save;
1907 rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1909 RETURN(ERR_PTR(rc));
1911 local_root = mdt_object_find(env, mdt, local_root_fid);
1912 if (IS_ERR(local_root))
1915 obj = mdt_object_new(env, mdt, fid);
1917 GOTO(out, rc = PTR_ERR(obj));
1919 spec->sp_cr_lookup = 0;
1920 spec->sp_feat = &dt_directory_features;
1921 spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1922 spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1923 if (attr->ma_valid & MA_LOV) {
1924 spec->u.sp_ea.eadata = attr->ma_lmm;
1925 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1926 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1928 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1932 uc_cap_save = uc->uc_cap;
1933 uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1934 rc = mdo_create(env, mdt_object_child(local_root), &lname,
1935 mdt_object_child(obj), spec, attr);
1936 uc->uc_cap = uc_cap_save;
1938 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1939 mdt_obd_name(mdt), PFID(fid), rc);
1943 rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1945 CERROR("%s: cannot open volatile file "DFID", orphan "
1946 "file will be left in PENDING directory until "
1947 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1954 mdt_object_put(env, obj);
1957 mdt_object_put(env, local_root);
1961 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1964 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1965 struct close_data *data;
1966 struct ldlm_lock *lease;
1967 struct mdt_object *orphan;
1968 struct md_attr *orp_ma;
1975 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1978 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1982 lease = ldlm_handle2lock(&data->cd_handle);
1986 /* try to hold open_sem so that nobody else can open the file */
1987 if (!down_write_trylock(&o->mot_open_sem)) {
1988 ldlm_lock_cancel(lease);
1989 GOTO(out_reprocess, rc = -EBUSY);
1992 /* Check if the lease open lease has already canceled */
1993 lock_res_and_lock(lease);
1994 lease_broken = ldlm_is_cancel(lease);
1995 unlock_res_and_lock(lease);
1997 LDLM_DEBUG(lease, DFID " lease broken? %d\n",
1998 PFID(mdt_object_fid(o)), lease_broken);
2000 /* Cancel server side lease. Client side counterpart should
2001 * have been cancelled. It's okay to cancel it now as we've
2002 * held mot_open_sem. */
2003 ldlm_lock_cancel(lease);
2005 if (lease_broken) /* don't perform release task */
2006 GOTO(out_unlock, rc = -ESTALE);
2008 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2009 GOTO(out_unlock, rc = -EINVAL);
2011 /* ma_need was set before but it seems fine to change it in order to
2012 * avoid modifying the one from RPC */
2013 ma->ma_need = MA_HSM;
2014 rc = mdt_attr_get_complex(info, o, ma);
2016 GOTO(out_unlock, rc);
2018 if (!mdt_hsm_release_allow(ma))
2019 GOTO(out_unlock, rc = -EPERM);
2021 /* already released? */
2022 if (ma->ma_hsm.mh_flags & HS_RELEASED)
2023 GOTO(out_unlock, rc = 0);
2025 /* Compare on-disk and packed data_version */
2026 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
2027 CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
2028 " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
2029 data->cd_data_version, ma->ma_hsm.mh_arch_ver);
2030 GOTO(out_unlock, rc = -EPERM);
2033 ma->ma_valid = MA_INODE;
2034 ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
2035 rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
2037 GOTO(out_unlock, rc);
2039 ma->ma_need = MA_INODE | MA_LOV;
2040 rc = mdt_attr_get_complex(info, o, ma);
2042 GOTO(out_unlock, rc);
2044 if (!(ma->ma_valid & MA_LOV)) {
2045 /* Even empty file are released */
2046 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2047 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2048 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2049 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2050 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2052 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
2053 * ma_lmm as lov_user_md, then it will be confused by union of
2054 * layout_gen and stripe_offset. */
2055 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
2056 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2057 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
2058 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
2060 GOTO(out_unlock, rc = -EINVAL);
2063 /* Set file as released */
2064 ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
2066 /* Hopefully it's not used in this call path */
2067 orp_ma = &info->mti_u.som.attr;
2068 orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2069 orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
2070 orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
2071 orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2072 orp_ma->ma_lmm = ma->ma_lmm;
2073 orp_ma->ma_lmm_size = ma->ma_lmm_size;
2074 orp_ma->ma_valid = MA_INODE | MA_LOV;
2075 orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2077 if (IS_ERR(orphan)) {
2078 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2079 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2081 GOTO(out_unlock, rc = PTR_ERR(orphan));
2084 /* Set up HSM attribute for orphan object */
2085 CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
2086 buf = &info->mti_buf;
2087 buf->lb_buf = info->mti_xattr_buf;
2088 buf->lb_len = sizeof(struct hsm_attrs);
2089 ma->ma_hsm.mh_flags |= HS_RELEASED;
2090 lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2091 ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2093 mdt_lock_reg_init(lh, LCK_EX);
2094 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2095 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
2097 GOTO(out_close, rc);
2099 rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2103 /* Swap layout with orphan object */
2104 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2105 mdt_object_child(orphan),
2106 SWAP_LAYOUTS_MDS_HSM);
2108 /* Release exclusive LL */
2109 mdt_object_unlock(info, o, lh, 1);
2114 /* Close orphan object anyway */
2115 rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2118 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2119 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2120 LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2122 mdt_object_put(info->mti_env, orphan);
2125 up_write(&o->mot_open_sem);
2127 if (rc == 0) { /* already released */
2128 struct mdt_body *repbody;
2129 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2130 LASSERT(repbody != NULL);
2131 repbody->mbo_valid |= OBD_MD_FLRELEASED;
2135 ldlm_reprocess_all(lease->l_resource);
2136 LDLM_LOCK_PUT(lease);
2144 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
2145 MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
2147 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2149 return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
2152 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2154 struct mdt_object *o = mfd->mfd_object;
2155 struct md_object *next = mdt_object_child(o);
2156 struct md_attr *ma = &info->mti_attr;
2157 int ret = MDT_IOEPOCH_CLOSED;
2162 mode = mfd->mfd_mode;
2164 if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
2165 rc = mdt_hsm_release(info, o, ma);
2167 CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2168 mdt_obd_name(info->mti_mdt),
2169 PFID(mdt_object_fid(o)), rc);
2170 /* continue to close even error occurred. */
2174 if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
2176 ret = mdt_ioepoch_close(info, o);
2177 } else if (mode & MDS_FMODE_EXEC) {
2179 } else if (mode & MDS_FMODE_EPOCH) {
2180 ret = mdt_ioepoch_close(info, o);
2181 } else if (mode & MDS_FMODE_SOM) {
2182 ret = mdt_som_au_close(info, o);
2185 /* Update atime on close only. */
2186 if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
2187 && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
2188 /* Set the atime only. */
2189 ma->ma_valid = MA_INODE;
2190 ma->ma_attr.la_valid = LA_ATIME;
2191 rc = mo_attr_set(info->mti_env, next, ma);
2194 /* If file data is modified, add the dirty flag. */
2195 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2196 rc = mdt_add_dirty_flag(info, o, ma);
2198 ma->ma_need |= MA_INODE;
2199 ma->ma_valid &= ~MA_INODE;
2201 if (!MFD_CLOSED(mode))
2202 rc = mo_close(info->mti_env, next, ma, mode);
2204 if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
2205 struct mdt_export_data *med;
2207 /* The IOepoch is still opened or SOM update is needed.
2208 * Put mfd back into the list. */
2209 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
2210 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
2211 MDS_FMODE_EPOCH : MDS_FMODE_SOM);
2213 LASSERT(mdt_info_req(info));
2214 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
2215 spin_lock(&med->med_open_lock);
2216 list_add(&mfd->mfd_list, &med->med_open_head);
2217 class_handle_hash_back(&mfd->mfd_handle);
2218 spin_unlock(&med->med_open_lock);
2220 if (ret == MDT_IOEPOCH_OPENED) {
2224 CDEBUG(D_INODE, "Size-on-MDS attribute update is "
2225 "needed on "DFID"\n", PFID(mdt_object_fid(o)));
2228 /* adjust open and lease count */
2229 if (mode & MDS_OPEN_LEASE) {
2230 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2231 atomic_dec(&o->mot_lease_count);
2233 LASSERT(atomic_read(&o->mot_open_count) > 0);
2234 atomic_dec(&o->mot_open_count);
2237 mdt_object_put(info->mti_env, o);
2240 RETURN(rc ? rc : ret);
2243 int mdt_close(struct tgt_session_info *tsi)
2245 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2246 struct ptlrpc_request *req = tgt_ses_req(tsi);
2247 struct mdt_export_data *med;
2248 struct mdt_file_data *mfd;
2249 struct mdt_object *o;
2250 struct md_attr *ma = &info->mti_attr;
2251 struct mdt_body *repbody = NULL;
2255 mdt_counter_incr(req, LPROC_MDT_CLOSE);
2256 /* Close may come with the Size-on-MDS update. Unpack it. */
2257 rc = mdt_close_unpack(info);
2259 GOTO(out, rc = err_serious(rc));
2261 LASSERT(info->mti_ioepoch);
2263 /* These fields are no longer used and are left for compatibility.
2264 * size is always zero */
2265 req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2267 req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2269 rc = req_capsule_server_pack(info->mti_pill);
2270 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2271 mdt_client_compatibility(info);
2273 mdt_fix_reply(info);
2274 mdt_exit_ucred(info);
2275 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2278 /* Continue to close handle even if we can not pack reply */
2280 repbody = req_capsule_server_get(info->mti_pill,
2282 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2284 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2287 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
2288 repbody->mbo_eadatasize = 0;
2289 repbody->mbo_aclsize = 0;
2291 rc = err_serious(rc);
2294 med = &req->rq_export->exp_mdt_data;
2295 spin_lock(&med->med_open_lock);
2296 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2297 req_is_replay(req));
2298 if (mdt_mfd_closed(mfd)) {
2299 spin_unlock(&med->med_open_lock);
2300 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2301 ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
2302 info->mti_ioepoch->handle.cookie);
2303 /** not serious error since bug 3633 */
2306 class_handle_unhash(&mfd->mfd_handle);
2307 list_del_init(&mfd->mfd_list);
2308 spin_unlock(&med->med_open_lock);
2310 /* Do not lose object before last unlink. */
2311 o = mfd->mfd_object;
2312 mdt_object_get(info->mti_env, o);
2313 ret = mdt_mfd_close(info, mfd);
2314 if (repbody != NULL)
2315 rc = mdt_handle_last_unlink(info, o, ma);
2316 mdt_empty_transno(info, rc);
2317 mdt_object_put(info->mti_env, o);
2319 if (repbody != NULL) {
2320 mdt_client_compatibility(info);
2321 rc = mdt_fix_reply(info);
2324 mdt_exit_ucred(info);
2325 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2326 GOTO(out, rc = err_serious(-ENOMEM));
2328 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2329 OBD_FAIL_MDS_CLOSE_NET_REP))
2330 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2332 mdt_thread_info_fini(info);
2333 RETURN(rc ? rc : ret);
2337 * DONE_WRITING rpc handler.
2339 * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
2340 * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
2341 * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
2342 * skip attributes and reconstruct the reply here.
2344 int mdt_done_writing(struct tgt_session_info *tsi)
2346 struct ptlrpc_request *req = tgt_ses_req(tsi);
2347 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2348 struct mdt_body *repbody = NULL;
2349 struct mdt_export_data *med;
2350 struct mdt_file_data *mfd;
2354 rc = req_capsule_server_pack(tsi->tsi_pill);
2356 GOTO(out, rc = err_serious(rc));
2358 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2359 repbody->mbo_eadatasize = 0;
2360 repbody->mbo_aclsize = 0;
2362 /* Done Writing may come with the Size-on-MDS update. Unpack it. */
2363 rc = mdt_close_unpack(info);
2365 GOTO(out, rc = err_serious(rc));
2367 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2368 mdt_exit_ucred(info);
2369 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2372 med = &info->mti_exp->exp_mdt_data;
2373 spin_lock(&med->med_open_lock);
2374 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2375 req_is_replay(req));
2377 spin_unlock(&med->med_open_lock);
2378 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
2379 ": cookie = "LPX64" ioepoch = "LPU64"\n",
2380 PFID(info->mti_rr.rr_fid1),
2381 info->mti_ioepoch->handle.cookie,
2382 info->mti_ioepoch->ioepoch);
2383 /* If this is a replay, reconstruct the transno. */
2384 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2385 rc = info->mti_ioepoch->flags & MF_SOM_AU ?
2387 mdt_empty_transno(info, rc);
2390 GOTO(error_ucred, rc);
2393 LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
2394 mfd->mfd_mode == MDS_FMODE_TRUNC);
2395 class_handle_unhash(&mfd->mfd_handle);
2396 list_del_init(&mfd->mfd_list);
2397 spin_unlock(&med->med_open_lock);
2399 /* Set EPOCH CLOSE flag if not set by client. */
2400 info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
2401 info->mti_attr.ma_valid = 0;
2403 info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
2404 OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2405 if (info->mti_attr.ma_lmm == NULL)
2406 GOTO(error_ucred, rc = -ENOMEM);
2408 rc = mdt_mfd_close(info, mfd);
2410 OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2411 mdt_empty_transno(info, rc);
2413 mdt_exit_ucred(info);
2415 mdt_thread_info_fini(info);