4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_open.c
38 * Lustre Metadata Target (mdt) open/close file handling
40 * Author: Huang Hua <huanghua@clusterfs.com>
43 #define DEBUG_SUBSYSTEM S_MDS
45 #include <lustre_acl.h>
46 #include <lustre_mds.h>
47 #include "mdt_internal.h"
49 /* we do nothing because we do not have refcount now */
50 static void mdt_mfd_get(void *mfdp)
54 static struct portals_handle_ops mfd_handle_ops = {
55 .hop_addref = mdt_mfd_get,
59 /* Create a new mdt_file_data struct, initialize it,
60 * and insert it to global hash table */
61 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
63 struct mdt_file_data *mfd;
68 CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
69 mfd->mfd_handle.h_owner = med;
70 CFS_INIT_LIST_HEAD(&mfd->mfd_list);
71 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
78 * Find the mfd pointed to by handle in global hash table.
79 * In case of replay the handle is obsoleted
80 * but mfd can be found in mfd list by that handle.
81 * Callers need to be holding med_open_lock.
83 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
84 const struct lustre_handle *handle,
85 bool is_replay_or_resent)
87 struct mdt_file_data *mfd;
90 LASSERT(handle != NULL);
91 mfd = class_handle2object(handle->cookie, med);
92 /* during dw/setattr replay the mfd can be found by old handle */
93 if (mfd == NULL && is_replay_or_resent) {
94 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
95 if (mfd->mfd_old_handle.cookie == handle->cookie)
105 void mdt_mfd_free(struct mdt_file_data *mfd)
107 LASSERT(cfs_list_empty(&mfd->mfd_list));
108 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
111 static int mdt_create_data(struct mdt_thread_info *info,
112 struct mdt_object *p, struct mdt_object *o)
114 struct md_op_spec *spec = &info->mti_spec;
115 struct md_attr *ma = &info->mti_attr;
119 if (!md_should_create(spec->sp_cr_flags))
122 ma->ma_need = MA_INODE | MA_LOV;
124 mutex_lock(&o->mot_lov_mutex);
125 if (!(o->mot_flags & MOF_LOV_CREATED)) {
126 if (p != NULL && !fid_is_md_operative(mdt_object_fid(p)))
127 GOTO(unlock, rc = -EPERM);
129 rc = mdo_create_data(info->mti_env,
130 p ? mdt_object_child(p) : NULL,
131 mdt_object_child(o), spec, ma);
133 rc = mdt_attr_get_complex(info, o, ma);
135 if (rc == 0 && ma->ma_valid & MA_LOV)
136 o->mot_flags |= MOF_LOV_CREATED;
139 mutex_unlock(&o->mot_lov_mutex);
143 static int mdt_ioepoch_opened(struct mdt_object *mo)
145 return mo->mot_ioepoch_count;
148 int mdt_object_is_som_enabled(struct mdt_object *mo)
150 return !mo->mot_ioepoch;
154 * Re-enable Size-on-MDS.
155 * Call under ->mot_ioepoch_mutex.
157 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
159 if (ioepoch == mo->mot_ioepoch) {
160 LASSERT(!mdt_ioepoch_opened(mo));
167 * Open the IOEpoch. It is allowed if @writecount is not negative.
168 * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
170 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
173 struct mdt_device *mdt = info->mti_mdt;
178 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
179 !S_ISREG(lu_object_attr(&o->mot_obj)))
182 mutex_lock(&o->mot_ioepoch_mutex);
183 if (mdt_ioepoch_opened(o)) {
184 /* Epoch continues even if there is no writers yet. */
185 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
186 o->mot_ioepoch, PFID(mdt_object_fid(o)));
188 /* XXX: ->mdt_ioepoch is not initialized at the mount */
189 spin_lock(&mdt->mdt_ioepoch_lock);
190 if (mdt->mdt_ioepoch < info->mti_replayepoch)
191 mdt->mdt_ioepoch = info->mti_replayepoch;
193 if (info->mti_replayepoch)
194 o->mot_ioepoch = info->mti_replayepoch;
195 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
196 o->mot_ioepoch = ++mdt->mdt_ioepoch;
198 o->mot_ioepoch = mdt->mdt_ioepoch;
200 spin_unlock(&mdt->mdt_ioepoch_lock);
202 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
203 o->mot_ioepoch, PFID(mdt_object_fid(o)));
205 o->mot_flags |= MOF_SOM_CREATED;
208 o->mot_ioepoch_count++;
209 mutex_unlock(&o->mot_ioepoch_mutex);
211 /* Cancel Size-on-MDS attributes cached on clients for the open case.
212 * In the truncate case, see mdt_reint_setattr(). */
213 if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
214 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
215 mdt_lock_reg_init(lh, LCK_EX);
216 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
219 mdt_object_unlock(info, o, lh, 1);
225 * Update SOM on-disk attributes.
226 * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
227 * mountid and attributes. If disabling, clean SOM xattr.
228 * Call under ->mot_ioepoch_mutex.
230 static int mdt_som_attr_set(struct mdt_thread_info *info,
231 struct mdt_object *obj, __u64 ioepoch, bool enable)
233 struct md_object *next = mdt_object_child(obj);
237 CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
238 " on "DFID".\n", enable ? "update" : "disabling",
239 ioepoch, PFID(mdt_object_fid(obj)));
242 struct lu_buf *buf = &info->mti_buf;
243 struct som_attrs *attrs;
244 struct md_attr *ma = &info->mti_attr;
245 struct lu_attr *la = &ma->ma_attr;
246 struct obd_device *obd = info->mti_mdt->mdt_lut.lut_obd;
248 attrs = (struct som_attrs *)info->mti_xattr_buf;
249 CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*attrs));
251 /* pack SOM attributes */
252 memset(attrs, 0, sizeof(*attrs));
253 attrs->som_ioepoch = ioepoch;
254 attrs->som_mountid = obd->u.obt.obt_mount_count;
255 if ((la->la_valid & LA_SIZE) != 0)
256 attrs->som_size = la->la_size;
257 if ((la->la_valid & LA_BLOCKS) != 0)
258 attrs->som_blocks = la->la_blocks;
259 lustre_som_swab(attrs);
261 /* update SOM attributes */
263 buf->lb_len = sizeof(*attrs);
264 rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_SOM, 0);
266 /* delete SOM attributes */
267 rc = mo_xattr_del(info->mti_env, next, XATTR_NAME_SOM);
273 /** Perform the eviction specific actions on ioepoch close. */
274 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
275 struct mdt_object *o)
279 mutex_lock(&o->mot_ioepoch_mutex);
280 CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
281 "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
282 o->mot_ioepoch_count);
283 o->mot_ioepoch_count--;
285 /* If eviction occured set MOF_SOM_RECOV,
286 * if no other epoch holders, disable SOM on disk. */
287 o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
288 if (!mdt_ioepoch_opened(o)) {
289 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
290 mdt_object_som_enable(o, o->mot_ioepoch);
292 mutex_unlock(&o->mot_ioepoch_mutex);
297 * Perform the replay specific actions on ioepoch close.
298 * Skip SOM attribute update if obtained and just forget about the inode state
299 * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
301 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
302 struct mdt_object *o)
304 int rc = MDT_IOEPOCH_CLOSED;
307 mutex_lock(&o->mot_ioepoch_mutex);
308 CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
309 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
310 o->mot_ioepoch_count--;
312 /* Get an info from the replayed request if client is supposed
313 * to send an Attibute Update, reconstruct @rc if so */
314 if (info->mti_ioepoch->flags & MF_SOM_AU)
315 rc = MDT_IOEPOCH_GETATTR;
317 if (!mdt_ioepoch_opened(o))
318 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
319 mutex_unlock(&o->mot_ioepoch_mutex);
325 * Regular file IOepoch close.
326 * Closes the ioepoch, checks the object state, apply obtained attributes and
327 * re-enable SOM on the object, if possible. Also checks if the recovery is
328 * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
329 * to obtain SOM attributes under the server-side OST locks.
332 * MDT_IOEPOCH_CLOSED if ioepoch is closed.
333 * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
335 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
336 struct mdt_object *o)
338 struct md_attr *tmp_ma;
342 int rc = 0, ret = MDT_IOEPOCH_CLOSED;
345 la = &info->mti_attr.ma_attr;
346 achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
348 mutex_lock(&o->mot_ioepoch_mutex);
349 o->mot_ioepoch_count--;
351 tmp_ma = &info->mti_u.som.attr;
352 tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
353 tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
354 tmp_ma->ma_som = &info->mti_u.som.data;
355 tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
356 tmp_ma->ma_valid = 0;
357 rc = mdt_attr_get_complex(info, o, tmp_ma);
361 /* Check the on-disk SOM state. */
362 if (o->mot_flags & MOF_SOM_RECOV)
364 else if (!(o->mot_flags & MOF_SOM_CREATED) &&
365 !(tmp_ma->ma_valid & MA_SOM))
368 CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
369 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
371 opened = mdt_ioepoch_opened(o);
373 * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
374 * Skip the check for file with no LOV or for unlink files.
376 if (!opened && tmp_ma->ma_valid & MA_LOV &&
377 !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
379 /* If some previous writer was evicted, re-ask the
380 * client for attributes. Even if attributes are
381 * provided, we cannot believe in them.
382 * Another use case is that there is no SOM cache on
383 * disk -- first access with SOM or there was an MDS
385 ret = MDT_IOEPOCH_GETATTR;
386 else if (o->mot_flags & MOF_SOM_CHANGE)
387 /* Some previous writer changed the attribute.
388 * Do not believe to the current Size-on-MDS
389 * update, re-ask client. */
390 ret = MDT_IOEPOCH_GETATTR;
391 else if (!(la->la_valid & LA_SIZE) && achange)
392 /* Attributes were changed by the last writer
393 * only but no Size-on-MDS update is received.*/
394 ret = MDT_IOEPOCH_GETATTR;
397 if (achange || ret == MDT_IOEPOCH_GETATTR)
398 o->mot_flags |= MOF_SOM_CHANGE;
400 /* If epoch ends and relable SOM attributes are obtained, update them.
401 * Create SOM ea for new files even if there is no attributes obtained
402 * (0-length file). */
403 if (ret == MDT_IOEPOCH_CLOSED && !opened) {
404 if (achange || o->mot_flags & MOF_SOM_CREATED) {
405 LASSERT(achange || !(la->la_valid & LA_SIZE));
406 rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
408 /* Avoid the following setattrs of these attributes,
409 * e.g. for atime update. */
410 info->mti_attr.ma_valid = 0;
412 mdt_object_som_enable(o, o->mot_ioepoch);
415 mutex_unlock(&o->mot_ioepoch_mutex);
416 /* If recovery is needed, tell the client to perform GETATTR under
418 if (ret == MDT_IOEPOCH_GETATTR && recovery) {
419 struct mdt_body *rep;
420 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
421 rep->valid |= OBD_MD_FLGETATTRLOCK;
427 mutex_unlock(&o->mot_ioepoch_mutex);
432 * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
433 * - a client closes the IOEpoch;
434 * - a client eviction occured.
436 * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
437 * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
438 * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
441 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
443 struct ptlrpc_request *req = mdt_info_req(info);
446 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
447 !S_ISREG(lu_object_attr(&o->mot_obj)))
450 LASSERT(o->mot_ioepoch_count);
451 LASSERT(info->mti_ioepoch == NULL ||
452 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
454 /* IOEpoch is closed only if client tells about it or eviction occures.
455 * In the replay case, always close the epoch. */
457 RETURN(mdt_ioepoch_close_on_eviction(info, o));
458 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
459 RETURN(mdt_ioepoch_close_on_replay(info, o));
460 if (info->mti_ioepoch && (info->mti_ioepoch->flags & MF_EPOCH_CLOSE))
461 RETURN(mdt_ioepoch_close_reg(info, o));
462 /* IO epoch is not closed. */
463 RETURN(MDT_IOEPOCH_OPENED);
467 * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
468 * for attribute update. It happens if:
469 * - SOM Attribute Update is obtained;
470 * - the client failed to obtain it and informs MDS about it;
471 * - a client eviction occured.
472 * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
475 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
477 struct ptlrpc_request *req = mdt_info_req(info);
479 int act = MDT_SOM_ENABLE;
483 LASSERT(!req || info->mti_ioepoch);
484 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
485 !S_ISREG(lu_object_attr(&o->mot_obj)))
488 /* No size whereas MF_SOM_CHANGE is set means client failed to
489 * obtain ost attributes, drop the SOM cache on disk if so. */
491 (info->mti_ioepoch &&
492 info->mti_ioepoch->flags & MF_SOM_CHANGE &&
493 !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
494 act = MDT_SOM_DISABLE;
496 mutex_lock(&o->mot_ioepoch_mutex);
497 /* Mark the object it is the recovery state if we failed to obtain
499 if (act == MDT_SOM_DISABLE)
500 o->mot_flags |= MOF_SOM_RECOV;
502 if (!mdt_ioepoch_opened(o)) {
503 ioepoch = info->mti_ioepoch ?
504 info->mti_ioepoch->ioepoch : o->mot_ioepoch;
507 && !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
508 rc = mdt_som_attr_set(info, o, ioepoch, act);
509 mdt_object_som_enable(o, ioepoch);
511 mutex_unlock(&o->mot_ioepoch_mutex);
515 int mdt_write_read(struct mdt_object *o)
519 mutex_lock(&o->mot_ioepoch_mutex);
520 rc = o->mot_writecount;
521 mutex_unlock(&o->mot_ioepoch_mutex);
525 int mdt_write_get(struct mdt_object *o)
529 mutex_lock(&o->mot_ioepoch_mutex);
530 if (o->mot_writecount < 0)
534 mutex_unlock(&o->mot_ioepoch_mutex);
538 void mdt_write_put(struct mdt_object *o)
541 mutex_lock(&o->mot_ioepoch_mutex);
543 mutex_unlock(&o->mot_ioepoch_mutex);
547 static int mdt_write_deny(struct mdt_object *o)
551 mutex_lock(&o->mot_ioepoch_mutex);
552 if (o->mot_writecount > 0)
556 mutex_unlock(&o->mot_ioepoch_mutex);
560 static void mdt_write_allow(struct mdt_object *o)
563 mutex_lock(&o->mot_ioepoch_mutex);
565 mutex_unlock(&o->mot_ioepoch_mutex);
569 /* there can be no real transaction so prepare the fake one */
570 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
572 struct mdt_device *mdt = info->mti_mdt;
573 struct ptlrpc_request *req = mdt_info_req(info);
574 struct tg_export_data *ted;
575 struct lsd_client_data *lcd;
578 /* transaction has occurred already */
579 if (lustre_msg_get_transno(req->rq_repmsg) != 0)
582 spin_lock(&mdt->mdt_lut.lut_translock);
584 if (info->mti_transno != 0) {
585 struct obd_export *exp = req->rq_export;
587 CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
588 mdt_obd_name(mdt), info->mti_transno,
589 libcfs_nid2str(exp->exp_connection->c_peer.nid),
591 spin_unlock(&mdt->mdt_lut.lut_translock);
594 } else if (info->mti_transno == 0) {
595 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
597 /* should be replay */
598 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
599 mdt->mdt_lut.lut_last_transno = info->mti_transno;
601 spin_unlock(&mdt->mdt_lut.lut_translock);
603 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
605 req->rq_export->exp_obd->obd_last_committed);
607 req->rq_transno = info->mti_transno;
608 lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
610 /* update lcd in memory only for resent cases */
611 ted = &req->rq_export->exp_target_data;
613 mutex_lock(&ted->ted_lcd_lock);
615 if (info->mti_transno < lcd->lcd_last_transno &&
616 info->mti_transno != 0) {
617 /* This should happen during replay. Do not update
618 * last rcvd info if replay req transno < last transno,
619 * otherwise the following resend(after replay) can not
620 * be checked correctly by xid */
621 mutex_unlock(&ted->ted_lcd_lock);
622 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
623 mdt_obd_name(mdt), info->mti_transno,
624 lcd->lcd_last_transno);
628 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
629 lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
630 if (info->mti_transno != 0)
631 lcd->lcd_last_close_transno = info->mti_transno;
632 lcd->lcd_last_close_xid = req->rq_xid;
633 lcd->lcd_last_close_result = rc;
635 /* VBR: save versions in last_rcvd for reconstruct. */
636 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
638 lcd->lcd_pre_versions[0] = pre_versions[0];
639 lcd->lcd_pre_versions[1] = pre_versions[1];
640 lcd->lcd_pre_versions[2] = pre_versions[2];
641 lcd->lcd_pre_versions[3] = pre_versions[3];
643 if (info->mti_transno != 0)
644 lcd->lcd_last_transno = info->mti_transno;
646 lcd->lcd_last_xid = req->rq_xid;
647 lcd->lcd_last_result = rc;
648 lcd->lcd_last_data = info->mti_opdata;
650 mutex_unlock(&ted->ted_lcd_lock);
655 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
657 LASSERT(mfd != NULL);
659 CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n",
660 PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
662 mfd->mfd_mode = mode;
666 * prep ma_lmm/ma_lmv for md_attr from reply
668 void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
669 struct mdt_object *obj,
672 LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
673 if (S_ISDIR(obj->mot_header.loh_attr)) {
674 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
676 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
679 if (ma->ma_lmv_size > 0)
680 ma->ma_need |= MA_LMV;
682 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
684 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
687 if (ma->ma_lmm_size > 0)
688 ma->ma_need |= MA_LOV;
692 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
693 struct mdt_object *o, __u64 flags, int created,
694 struct ldlm_reply *rep)
696 struct ptlrpc_request *req = mdt_info_req(info);
697 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
698 struct mdt_file_data *mfd;
699 struct md_attr *ma = &info->mti_attr;
700 struct lu_attr *la = &ma->ma_attr;
701 struct mdt_body *repbody;
702 int rc = 0, isdir, isreg;
705 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
707 isreg = S_ISREG(la->la_mode);
708 isdir = S_ISDIR(la->la_mode);
709 if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
711 * No EA, check whether it is will set regEA and dirEA since in
712 * above attr get, these size might be zero, so reset it, to
713 * retrieve the MD after create obj.
715 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
718 /* in replay case, p == NULL */
719 rc = mdt_create_data(info, p, o);
723 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
724 mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
727 CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
728 ma->ma_valid, ma->ma_lmm_size);
730 if (ma->ma_valid & MA_LOV) {
731 LASSERT(ma->ma_lmm_size != 0);
732 repbody->eadatasize = ma->ma_lmm_size;
734 repbody->valid |= OBD_MD_FLDIREA;
736 repbody->valid |= OBD_MD_FLEASIZE;
739 if (ma->ma_valid & MA_LMV) {
740 LASSERT(ma->ma_lmv_size != 0);
741 repbody->eadatasize = ma->ma_lmv_size;
743 repbody->valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
746 if (flags & FMODE_WRITE) {
747 rc = mdt_write_get(o);
749 mdt_ioepoch_open(info, o, created);
750 repbody->ioepoch = o->mot_ioepoch;
752 } else if (flags & MDS_FMODE_EXEC) {
753 rc = mdt_write_deny(o);
758 rc = mo_open(info->mti_env, mdt_object_child(o),
759 created ? flags | MDS_OPEN_CREATED : flags);
763 mfd = mdt_mfd_new(med);
765 GOTO(err_out, rc = -ENOMEM);
768 * Keep a reference on this object for this open, and is
769 * released by mdt_mfd_close().
771 mdt_object_get(info->mti_env, o);
773 mfd->mfd_xid = req->rq_xid;
776 * @flags is always not zero. At least it should be FMODE_READ,
777 * FMODE_WRITE or MDS_FMODE_EXEC.
782 mdt_mfd_set_mode(mfd, flags);
784 atomic_inc(&o->mot_open_count);
785 if (flags & MDS_OPEN_LEASE)
786 atomic_inc(&o->mot_lease_count);
789 if (req_is_replay(req)) {
790 struct mdt_file_data *old_mfd;
791 /* Check wheather old cookie already exist in
792 * the list, becasue when do recovery, client
793 * might be disconnected from server, and
794 * restart replay, so there maybe some orphan
795 * mfd here, we should remove them */
796 LASSERT(info->mti_rr.rr_handle != NULL);
797 spin_lock(&med->med_open_lock);
798 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
799 if (old_mfd != NULL) {
800 CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
801 "cookie = "LPX64"\n", mfd,
802 PFID(mdt_object_fid(mfd->mfd_object)),
803 info->mti_rr.rr_handle->cookie);
804 class_handle_unhash(&old_mfd->mfd_handle);
805 cfs_list_del_init(&old_mfd->mfd_list);
806 spin_unlock(&med->med_open_lock);
807 /* no attr update for that close */
809 ma->ma_valid |= MA_FLAGS;
810 ma->ma_attr_flags |= MDS_RECOV_OPEN;
811 mdt_mfd_close(info, old_mfd);
812 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
813 ma->ma_valid &= ~MA_FLAGS;
815 spin_unlock(&med->med_open_lock);
816 CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
817 "cookie = "LPX64"\n",
818 PFID(mdt_object_fid(mfd->mfd_object)),
819 info->mti_rr.rr_handle->cookie);
822 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
823 info->mti_rr.rr_handle->cookie);
825 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
828 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
830 if (req->rq_export->exp_disconnected) {
831 spin_lock(&med->med_open_lock);
832 class_handle_unhash(&mfd->mfd_handle);
833 cfs_list_del_init(&mfd->mfd_list);
834 spin_unlock(&med->med_open_lock);
835 mdt_mfd_close(info, mfd);
837 spin_lock(&med->med_open_lock);
838 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
839 spin_unlock(&med->med_open_lock);
842 mdt_empty_transno(info, rc);
847 if (flags & FMODE_WRITE)
848 /* XXX We also need to close io epoch here.
849 * See LU-1220 - green */
851 else if (flags & MDS_FMODE_EXEC)
857 int mdt_finish_open(struct mdt_thread_info *info,
858 struct mdt_object *p, struct mdt_object *o,
859 __u64 flags, int created, struct ldlm_reply *rep)
861 struct ptlrpc_request *req = mdt_info_req(info);
862 struct obd_export *exp = req->rq_export;
863 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
864 struct md_attr *ma = &info->mti_attr;
865 struct lu_attr *la = &ma->ma_attr;
866 struct mdt_file_data *mfd;
867 struct mdt_body *repbody;
869 int isreg, isdir, islnk;
873 LASSERT(ma->ma_valid & MA_INODE);
875 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
877 isreg = S_ISREG(la->la_mode);
878 isdir = S_ISDIR(la->la_mode);
879 islnk = S_ISLNK(la->la_mode);
880 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
882 /* LU-2275, simulate broken behaviour (esp. prevalent in
883 * pre-2.4 servers where a very strange reply is sent on error
884 * that looks like it was actually almost succesful and a failure at the
886 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
887 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
891 if (flags & MDS_OPEN_LOCK)
892 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
897 if (exp_connect_rmtclient(exp)) {
898 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
900 rc = mdt_pack_remote_perm(info, o, buf);
902 repbody->valid &= ~OBD_MD_FLRMTPERM;
903 repbody->aclsize = 0;
905 repbody->valid |= OBD_MD_FLRMTPERM;
906 repbody->aclsize = sizeof(struct mdt_remote_perm);
909 #ifdef CONFIG_FS_POSIX_ACL
910 else if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
911 const struct lu_env *env = info->mti_env;
912 struct md_object *next = mdt_object_child(o);
913 struct lu_buf *buf = &info->mti_buf;
915 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
916 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
918 if (buf->lb_len > 0) {
919 rc = mo_xattr_get(env, next, buf,
920 XATTR_NAME_ACL_ACCESS);
922 if (rc == -ENODATA) {
923 repbody->aclsize = 0;
924 repbody->valid |= OBD_MD_FLACL;
926 } else if (rc == -EOPNOTSUPP) {
929 CERROR("got acl size: %d\n", rc);
932 repbody->aclsize = rc;
933 repbody->valid |= OBD_MD_FLACL;
940 if (info->mti_mdt->mdt_lut.lut_mds_capa &&
941 exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) {
942 struct lustre_capa *capa;
944 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
946 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
947 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
950 repbody->valid |= OBD_MD_FLMDSCAPA;
952 if (info->mti_mdt->mdt_lut.lut_oss_capa &&
953 exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA &&
954 S_ISREG(lu_object_attr(&o->mot_obj))) {
955 struct lustre_capa *capa;
957 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
959 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
960 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
963 repbody->valid |= OBD_MD_FLOSSCAPA;
967 * If we are following a symlink, don't open; and do not return open
968 * handle for special nodes as client required.
970 if (islnk || (!isreg && !isdir &&
971 (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
972 lustre_msg_set_transno(req->rq_repmsg, 0);
977 * We need to return the existing object's fid back, so it is done here,
978 * after preparing the reply.
980 if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
983 /* This can't be done earlier, we need to return reply body */
985 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
986 /* We are trying to create or write an existing dir. */
989 } else if (flags & MDS_OPEN_DIRECTORY)
992 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
993 OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
998 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
999 spin_lock(&med->med_open_lock);
1000 cfs_list_for_each(t, &med->med_open_head) {
1001 mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
1002 if (mfd->mfd_xid == req->rq_xid)
1006 spin_unlock(&med->med_open_lock);
1009 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
1010 /*set repbody->ea_size for resent case*/
1011 if (ma->ma_valid & MA_LOV) {
1012 LASSERT(ma->ma_lmm_size != 0);
1013 repbody->eadatasize = ma->ma_lmm_size;
1015 repbody->valid |= OBD_MD_FLDIREA;
1017 repbody->valid |= OBD_MD_FLEASIZE;
1019 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
1024 rc = mdt_mfd_open(info, p, o, flags, created, rep);
1026 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
1031 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
1032 struct lsd_client_data *lcd);
1034 void mdt_reconstruct_open(struct mdt_thread_info *info,
1035 struct mdt_lock_handle *lhc)
1037 const struct lu_env *env = info->mti_env;
1038 struct mdt_device *mdt = info->mti_mdt;
1039 struct req_capsule *pill = info->mti_pill;
1040 struct ptlrpc_request *req = mdt_info_req(info);
1041 struct tg_export_data *ted = &req->rq_export->exp_target_data;
1042 struct lsd_client_data *lcd = ted->ted_lcd;
1043 struct md_attr *ma = &info->mti_attr;
1044 struct mdt_reint_record *rr = &info->mti_rr;
1045 __u64 flags = info->mti_spec.sp_cr_flags;
1046 struct ldlm_reply *ldlm_rep;
1047 struct mdt_object *parent;
1048 struct mdt_object *child;
1049 struct mdt_body *repbody;
1053 LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1054 ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1055 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1057 ma->ma_need = MA_INODE | MA_HSM;
1060 mdt_req_from_lcd(req, lcd);
1061 mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1063 CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1064 ldlm_rep->lock_policy_res1, req->rq_status);
1066 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1067 req->rq_status != 0)
1068 /* We did not create successfully, return error to client. */
1069 GOTO(out, rc = req->rq_status);
1071 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1072 struct obd_export *exp = req->rq_export;
1074 * We failed after creation, but we do not know in which step
1075 * we failed. So try to check the child object.
1077 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1078 if (IS_ERR(parent)) {
1079 rc = PTR_ERR(parent);
1080 LCONSOLE_WARN("Parent "DFID" lookup error %d."
1081 " Evicting client %s with export %s.\n",
1082 PFID(rr->rr_fid1), rc,
1083 obd_uuid2str(&exp->exp_client_uuid),
1084 obd_export_nid2str(exp));
1085 mdt_export_evict(exp);
1089 child = mdt_object_find(env, mdt, rr->rr_fid2);
1090 if (IS_ERR(child)) {
1091 rc = PTR_ERR(child);
1092 LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
1093 "evicting client %s with export %s\n",
1094 PFID(rr->rr_fid2), rc,
1095 obd_uuid2str(&exp->exp_client_uuid),
1096 obd_export_nid2str(exp));
1097 mdt_object_put(env, parent);
1098 mdt_export_evict(exp);
1102 if (unlikely(mdt_object_remote(child))) {
1103 /* the child object was created on remote server */
1104 if (!mdt_is_dne_client(exp)) {
1105 /* Return -EIO for old client */
1106 mdt_object_put(env, parent);
1107 mdt_object_put(env, child);
1108 GOTO(out, rc = -EIO);
1110 repbody->fid1 = *rr->rr_fid2;
1111 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1114 if (mdt_object_exists(child)) {
1115 mdt_set_capainfo(info, 1, rr->rr_fid2,
1117 mdt_prep_ma_buf_from_rep(info, child, ma);
1118 rc = mdt_attr_get_complex(info, child, ma);
1120 rc = mdt_finish_open(info, parent,
1124 /* the child does not exist, we should do
1126 mdt_object_put(env, parent);
1127 mdt_object_put(env, child);
1128 GOTO(regular_open, 0);
1131 mdt_object_put(env, parent);
1132 mdt_object_put(env, child);
1136 /* We did not try to create, so we are a pure open */
1137 rc = mdt_reint_open(info, lhc);
1142 req->rq_status = rc;
1143 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1144 LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1147 int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
1149 __u64 flags = info->mti_spec.sp_cr_flags;
1150 struct mdt_reint_record *rr = &info->mti_rr;
1151 struct md_attr *ma = &info->mti_attr;
1152 struct mdt_object *o;
1156 o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1158 RETURN(rc = PTR_ERR(o));
1160 if (unlikely(mdt_object_remote(o))) {
1161 /* the child object was created on remote server */
1162 struct mdt_body *repbody;
1164 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1167 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1168 repbody->fid1 = *rr->rr_fid2;
1169 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1172 if (mdt_object_exists(o)) {
1173 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1176 mdt_prep_ma_buf_from_rep(info, o, ma);
1177 rc = mdt_attr_get_complex(info, o, ma);
1179 rc = mdt_finish_open(info, NULL, o, flags, 0,
1186 mdt_object_put(info->mti_env, o);
1190 /* lock object for open */
1191 static int mdt_object_open_lock(struct mdt_thread_info *info,
1192 struct mdt_object *obj,
1193 struct mdt_lock_handle *lhc,
1196 struct md_attr *ma = &info->mti_attr;
1197 __u64 open_flags = info->mti_spec.sp_cr_flags;
1198 ldlm_mode_t lm = LCK_CR;
1199 bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
1200 bool try_layout = false;
1201 bool create_layout = false;
1206 mdt_lock_handle_init(lhc);
1208 if (req_is_replay(mdt_info_req(info)))
1211 if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
1212 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
1213 md_should_create(open_flags))
1214 create_layout = true;
1215 if (exp_connect_layout(info->mti_exp) && !create_layout &&
1216 ma->ma_need & MA_LOV)
1221 /* lease open, acquire write mode of open sem */
1222 down_write(&obj->mot_open_sem);
1224 /* Lease exists and ask for new lease */
1225 if (atomic_read(&obj->mot_lease_count) > 0) {
1226 /* only exclusive open is supported, so lease
1227 * are conflicted to each other */
1228 GOTO(out, rc = -EBUSY);
1231 /* Lease must be with open lock */
1232 if (!(open_flags & MDS_OPEN_LOCK)) {
1233 CERROR("Request lease for file:"DFID ", but open lock "
1234 "is missed, open_flags = "LPO64".\n",
1235 PFID(mdt_object_fid(obj)), open_flags);
1236 GOTO(out, rc = -EPROTO);
1239 /* XXX: only exclusive open is supported. */
1241 *ibits = MDS_INODELOCK_OPEN;
1243 /* never grant LCK_EX layout lock to client */
1245 } else { /* normal open */
1246 /* normal open holds read mode of open sem */
1247 down_read(&obj->mot_open_sem);
1249 if (open_flags & FMODE_WRITE)
1251 else if (open_flags & MDS_FMODE_EXEC)
1256 if (open_flags & MDS_OPEN_LOCK) {
1257 *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
1258 } else if (open_flags & (FMODE_WRITE | MDS_FMODE_EXEC) ||
1259 atomic_read(&obj->mot_lease_count) > 0) {
1260 /* We need to flush conflicting locks or revoke a lease.
1261 * In either case there is no need to acquire a layout
1262 * lock since it won't be returned to the client. */
1264 *ibits = MDS_INODELOCK_OPEN;
1265 lhc = &info->mti_lh[MDT_LH_LOCAL];
1268 CDEBUG(D_INODE, "normal open FID = "DFID", open_count = %d, "
1270 PFID(mdt_object_fid(obj)),
1271 atomic_read(&obj->mot_open_count), lm);
1274 mdt_lock_reg_init(lhc, lm);
1276 /* one problem to return layout lock on open is that it may result
1277 * in too many layout locks cached on the client side. */
1278 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
1279 /* return lookup lock to validate inode at the client side,
1280 * this is pretty important otherwise mdt will return layout
1281 * lock for each open.
1282 * However this is a double-edged sword because changing
1283 * permission will revoke huge # of LOOKUP locks. */
1284 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
1285 if (!mdt_object_lock_try(info, obj, lhc, *ibits,
1287 *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
1289 rc = mdt_object_lock(info, obj, lhc, *ibits,
1292 } else if (*ibits != 0) {
1293 rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK);
1296 CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
1297 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
1298 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
1300 /* will change layout, revoke layout locks by enqueuing EX lock. */
1301 if (rc == 0 && create_layout) {
1302 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
1304 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
1305 ", open_flags = "LPO64"\n",
1306 PFID(mdt_object_fid(obj)), open_flags);
1308 /* We cannot enqueue another lock for the same resource we
1309 * already have a lock for, due to mechanics of waiting list
1310 * iterating in ldlm, see LU-3601.
1311 * As such we'll drop the open lock we just got above here,
1312 * it's ok not to have this open lock as it's main purpose is to
1313 * flush unused cached client open handles. */
1314 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1315 mdt_object_unlock(info, obj, lhc, 1);
1317 LASSERT(!try_layout);
1318 mdt_lock_handle_init(ll);
1319 mdt_lock_reg_init(ll, LCK_EX);
1320 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
1323 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
1326 /* Check if there is any other open handles after acquiring
1327 * open lock. At this point, caching open handles have been revoked
1329 * XXX: Now only exclusive open is supported. Need to check the
1330 * type of open for generic lease support. */
1331 if (rc == 0 && acq_lease) {
1332 struct ptlrpc_request *req = mdt_info_req(info);
1333 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1334 struct mdt_file_data *mfd;
1335 bool is_replay_or_resent;
1338 /* For lease: application can open a file and then apply lease,
1339 * @handle contains original open handle in that case.
1340 * In recovery, open REQ will be replayed and the lease REQ may
1341 * be resent that means the open handle is already stale, so we
1342 * need to fix it up here by finding new handle. */
1343 is_replay_or_resent = req_is_replay(req) ||
1344 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
1346 /* if the request is _not_ a replay request, rr_handle
1347 * may be used to hold an openhandle which is issuing the
1348 * lease request, so that this openhandle doesn't count. */
1349 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
1350 is_replay_or_resent);
1354 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
1355 PFID(mdt_object_fid(obj)),
1356 atomic_read(&obj->mot_open_count), open_count);
1358 if (atomic_read(&obj->mot_open_count) > open_count)
1359 GOTO(out, rc = -EBUSY);
1367 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1368 struct mdt_object *obj,
1369 struct mdt_lock_handle *lhc,
1370 __u64 ibits, int rc)
1372 __u64 open_flags = info->mti_spec.sp_cr_flags;
1373 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1376 if (req_is_replay(mdt_info_req(info)))
1379 /* Release local lock - the lock put in MDT_LH_LOCAL will never
1380 * return to client side. */
1381 if (lustre_handle_is_used(&ll->mlh_reg_lh))
1382 mdt_object_unlock(info, obj, ll, 1);
1384 ll = &info->mti_lh[MDT_LH_LAYOUT];
1385 /* Release local layout lock, layout was created */
1386 if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1387 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1388 mdt_object_unlock(info, obj, ll, 1);
1391 if (open_flags & MDS_OPEN_LEASE)
1392 up_write(&obj->mot_open_sem);
1394 up_read(&obj->mot_open_sem);
1396 /* Cross-ref case, the lock should be returned to the client */
1397 if (ibits == 0 || rc == -EREMOTE)
1400 if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
1401 /* for the open request, the lock will only return to client
1402 * if open or layout lock is granted. */
1406 if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1407 struct ldlm_reply *ldlm_rep;
1409 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1410 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1411 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1412 mdt_object_unlock(info, obj, lhc, 1);
1418 * Check release is permitted for the current HSM flags.
1420 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1422 if (!(ma->ma_valid & MA_HSM))
1425 if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1428 if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1434 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1435 struct mdt_lock_handle *lhc)
1437 const struct lu_env *env = info->mti_env;
1438 struct mdt_device *mdt = info->mti_mdt;
1439 __u64 flags = info->mti_spec.sp_cr_flags;
1440 struct mdt_reint_record *rr = &info->mti_rr;
1441 struct md_attr *ma = &info->mti_attr;
1442 struct mdt_object *parent= NULL;
1443 struct mdt_object *o;
1448 if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1449 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1450 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1451 if (IS_ERR(parent)) {
1452 CDEBUG(D_INODE, "Fail to find parent "DFID
1453 " for anonymous created %ld, try to"
1454 " use server-side parent.\n",
1455 PFID(rr->rr_fid1), PTR_ERR(parent));
1460 ma->ma_need |= MA_PFID;
1463 o = mdt_object_find(env, mdt, rr->rr_fid2);
1465 RETURN(rc = PTR_ERR(o));
1467 if (mdt_object_remote(o)) {
1468 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1469 mdt_obd_name(info->mti_mdt),
1471 GOTO(out, rc = -EREMOTE);
1472 } else if (!mdt_object_exists(o)) {
1473 mdt_set_disposition(info, rep,
1477 GOTO(out, rc = -ENOENT);
1480 mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1482 mdt_prep_ma_buf_from_rep(info, o, ma);
1483 if (flags & MDS_OPEN_RELEASE)
1484 ma->ma_need |= MA_HSM;
1485 rc = mdt_attr_get_complex(info, o, ma);
1489 /* If a release request, check file flags are fine and ask for an
1490 * exclusive open access. */
1491 if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1492 GOTO(out, rc = -EPERM);
1494 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1496 GOTO(out_unlock, rc);
1498 if (ma->ma_valid & MA_PFID) {
1499 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1500 if (IS_ERR(parent)) {
1501 CDEBUG(D_INODE, "Fail to find parent "DFID
1502 " for anonymous created %ld, try to"
1503 " use system default.\n",
1504 PFID(&ma->ma_pfid), PTR_ERR(parent));
1509 rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1511 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1512 if (flags & MDS_OPEN_LOCK)
1513 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1514 if (flags & MDS_OPEN_LEASE)
1515 mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1517 GOTO(out_unlock, rc);
1520 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1522 mdt_object_put(env, o);
1524 mdt_object_put(env, parent);
1528 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1529 static int mdt_cross_open(struct mdt_thread_info *info,
1530 const struct lu_fid *parent_fid,
1531 const struct lu_fid *fid,
1532 struct ldlm_reply *rep, __u32 flags)
1534 struct md_attr *ma = &info->mti_attr;
1535 struct mdt_object *o;
1539 o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1541 RETURN(rc = PTR_ERR(o));
1543 if (mdt_object_remote(o)) {
1544 /* Something is wrong here, the object is on another MDS! */
1545 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1546 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1547 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1549 "Object isn't on this server! FLD error?\n");
1552 if (mdt_object_exists(o)) {
1553 /* Do permission check for cross-open. */
1554 rc = mo_permission(info->mti_env, NULL,
1555 mdt_object_child(o),
1556 NULL, flags | MDS_OPEN_CROSS);
1560 mdt_prep_ma_buf_from_rep(info, o, ma);
1561 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1562 rc = mdt_attr_get_complex(info, o, ma);
1566 /* Do not create lov object if the fid is opened
1568 if (S_ISREG(ma->ma_attr.la_mode) &&
1569 !(ma->ma_valid & MA_LOV) && (flags & FMODE_WRITE) &&
1570 fid_is_obf(parent_fid))
1571 GOTO(out, rc = -EPERM);
1573 rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1576 * Something is wrong here. lookup was positive but
1577 * there is no object!
1579 CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1580 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1585 mdt_object_put(info->mti_env, o);
1589 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1591 struct mdt_device *mdt = info->mti_mdt;
1592 struct ptlrpc_request *req = mdt_info_req(info);
1593 struct mdt_object *parent;
1594 struct mdt_object *child;
1595 struct mdt_lock_handle *lh;
1596 struct ldlm_reply *ldlm_rep;
1597 struct mdt_body *repbody;
1598 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1599 struct md_attr *ma = &info->mti_attr;
1600 __u64 create_flags = info->mti_spec.sp_cr_flags;
1602 struct mdt_reint_record *rr = &info->mti_rr;
1608 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1609 (obd_timeout + 1) / 4);
1611 mdt_counter_incr(req, LPROC_MDT_OPEN);
1612 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1614 ma->ma_need = MA_INODE;
1617 LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1618 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1620 if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1621 CERROR("file join is not supported anymore.\n");
1622 GOTO(out, result = err_serious(-EOPNOTSUPP));
1624 msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1626 if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1627 info->mti_spec.u.sp_ea.eadata == NULL)
1628 GOTO(out, result = err_serious(-EINVAL));
1630 CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1631 "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1632 PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1633 PFID(rr->rr_fid2), create_flags,
1634 ma->ma_attr.la_mode, msg_flags);
1636 if (info->mti_cross_ref) {
1637 /* This is cross-ref open */
1638 mdt_set_disposition(info, ldlm_rep,
1639 (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1641 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1642 ldlm_rep, create_flags);
1644 } else if (req_is_replay(req) ||
1645 (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1646 /* This is a replay request or from liblustre with ea. */
1647 result = mdt_open_by_fid(info, ldlm_rep);
1649 if (result != -ENOENT) {
1650 if (req->rq_export->exp_libclient &&
1651 create_flags & MDS_OPEN_HAS_EA)
1652 GOTO(out, result = 0);
1655 /* We didn't find the correct object, so we need to re-create it
1656 * via a regular replay. */
1657 if (!(create_flags & MDS_OPEN_CREAT)) {
1658 DEBUG_REQ(D_ERROR, req,
1659 "OPEN & CREAT not in open replay/by_fid.");
1660 GOTO(out, result = -EFAULT);
1662 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1663 } else if ((!lu_name_is_valid(&rr->rr_name) &&
1664 (create_flags & MDS_OPEN_LOCK)) ||
1665 (create_flags & MDS_OPEN_BY_FID)) {
1666 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1667 /* If result is 0 then open by FID has found the file
1668 * and there is nothing left for us to do here. More
1669 * generally if it is anything other than -ENOENT or
1670 * -EREMOTE then we return that now. If -ENOENT and
1671 * MDS_OPEN_CREAT is set then we must create the file
1672 * below. If -EREMOTE then we need to return a LOOKUP
1673 * lock to the client, which we do below. Hence this
1674 * odd looking condition. See LU-2523. */
1675 if (!(result == -ENOENT && (create_flags & MDS_OPEN_CREAT)) &&
1679 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1682 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1683 GOTO(out, result = err_serious(-ENOMEM));
1685 mdt_set_disposition(info, ldlm_rep,
1686 (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1688 if (!lu_name_is_valid(&rr->rr_name))
1689 GOTO(out, result = -EPROTO);
1691 lh = &info->mti_lh[MDT_LH_PARENT];
1692 mdt_lock_pdo_init(lh,
1693 (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1696 parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1697 MDS_INODELOCK_UPDATE);
1699 GOTO(out, result = PTR_ERR(parent));
1701 /* get and check version of parent */
1702 result = mdt_version_get_check(info, parent, 0);
1704 GOTO(out_parent, result);
1706 fid_zero(child_fid);
1708 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1709 &rr->rr_name, child_fid, &info->mti_spec);
1711 LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1712 "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1713 PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1716 if (result != 0 && result != -ENOENT && result != -ESTALE)
1717 GOTO(out_parent, result);
1719 if (result == -ENOENT || result == -ESTALE) {
1720 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1721 if (result == -ESTALE) {
1723 * -ESTALE means the parent is a dead(unlinked) dir, so
1724 * it should return -ENOENT to in accordance with the
1725 * original mds implementaion.
1727 GOTO(out_parent, result = -ENOENT);
1729 if (!(create_flags & MDS_OPEN_CREAT))
1730 GOTO(out_parent, result);
1731 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1732 GOTO(out_parent, result = -EROFS);
1733 *child_fid = *info->mti_rr.rr_fid2;
1734 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1736 /* In the function below, .hs_keycmp resolves to
1737 * lu_obj_hop_keycmp() */
1738 /* coverity[overrun-buffer-val] */
1739 child = mdt_object_new(info->mti_env, mdt, child_fid);
1742 * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1743 * return FID back in that case.
1745 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1746 child = mdt_object_find(info->mti_env, mdt, child_fid);
1749 GOTO(out_parent, result = PTR_ERR(child));
1751 /** check version of child */
1752 rc = mdt_version_get_check(info, child, 1);
1754 GOTO(out_child, result = rc);
1756 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1757 if (result == -ENOENT) {
1758 /* Create under OBF and .lustre is not permitted */
1759 if (!fid_is_md_operative(rr->rr_fid1))
1760 GOTO(out_child, result = -EPERM);
1762 /* save versions in reply */
1763 mdt_version_get_save(info, parent, 0);
1764 mdt_version_get_save(info, child, 1);
1766 /* version of child will be changed */
1767 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1769 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1770 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1772 /* Let lower layers know what is lock mode on directory. */
1773 info->mti_spec.sp_cr_mode =
1774 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1777 * Do not perform lookup sanity check. We know that name does
1780 info->mti_spec.sp_cr_lookup = 0;
1781 info->mti_spec.sp_feat = &dt_directory_features;
1783 result = mdo_create(info->mti_env,
1784 mdt_object_child(parent),
1786 mdt_object_child(child),
1789 if (result == -ERESTART) {
1790 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1791 GOTO(out_child, result);
1793 mdt_prep_ma_buf_from_rep(info, child, ma);
1794 /* XXX: we should call this once, see few lines below */
1796 result = mdt_attr_get_complex(info, child, ma);
1799 GOTO(out_child, result);
1804 * The object is on remote node, return its FID for remote open.
1806 if (mdt_object_remote(child)) {
1808 * Check if this lock already was sent to client and
1809 * this is resent case. For resent case do not take lock
1810 * again, use what is already granted.
1812 LASSERT(lhc != NULL);
1814 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1815 struct ldlm_lock *lock;
1817 LASSERT(msg_flags & MSG_RESENT);
1819 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1821 CERROR("Invalid lock handle "LPX64"\n",
1822 lhc->mlh_reg_lh.cookie);
1825 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1826 &lock->l_resource->lr_name));
1827 LDLM_LOCK_PUT(lock);
1830 mdt_lock_handle_init(lhc);
1831 mdt_lock_reg_init(lhc, LCK_PR);
1833 rc = mdt_object_lock(info, child, lhc,
1834 MDS_INODELOCK_LOOKUP,
1837 repbody->fid1 = *mdt_object_fid(child);
1838 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1843 GOTO(out_child, result);
1844 } else if (mdt_object_exists(child)) {
1845 /* We have to get attr & LOV EA & HSM for this
1847 mdt_prep_ma_buf_from_rep(info, child, ma);
1848 ma->ma_need |= MA_HSM;
1849 result = mdt_attr_get_complex(info, child, ma);
1851 /* Object does not exist. Likely FS corruption. */
1852 CERROR("%s: name '"DNAME"' present, but FID "
1853 DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1854 PNAME(&rr->rr_name), PFID(child_fid));
1855 GOTO(out_child, result = -EIO);
1859 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1860 /* the open lock might already be gotten in
1861 * mdt_intent_fixup_resent */
1862 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1863 if (create_flags & MDS_OPEN_LOCK)
1864 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1866 /* get openlock if this isn't replay and client requested it */
1867 if (!req_is_replay(req)) {
1868 rc = mdt_object_open_lock(info, child, lhc, &ibits);
1870 GOTO(out_child_unlock, result = rc);
1871 else if (create_flags & MDS_OPEN_LOCK)
1872 mdt_set_disposition(info, ldlm_rep,
1876 /* Try to open it now. */
1877 rc = mdt_finish_open(info, parent, child, create_flags,
1881 /* openlock will be released if mdt_finish_open failed */
1882 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1884 if (created && create_flags & MDS_OPEN_VOLATILE) {
1885 CERROR("%s: cannot open volatile file "DFID", orphan "
1886 "file will be left in PENDING directory until "
1887 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1888 PFID(mdt_object_fid(child)), rc);
1889 GOTO(out_child_unlock, result);
1895 rc = mdo_unlink(info->mti_env,
1896 mdt_object_child(parent),
1897 mdt_object_child(child),
1899 &info->mti_attr, 0);
1901 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1902 mdt_obd_name(info->mti_mdt),
1903 PFID(mdt_object_fid(child)), rc);
1904 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1909 mdt_object_open_unlock(info, child, lhc, ibits, result);
1911 mdt_object_put(info->mti_env, child);
1913 mdt_object_unlock_put(info, parent, lh, result || !created);
1916 lustre_msg_set_transno(req->rq_repmsg, 0);
1921 * Create an orphan object use local root.
1923 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1924 struct mdt_device *mdt,
1925 const struct lu_fid *fid,
1926 struct md_attr *attr, fmode_t fmode)
1928 const struct lu_env *env = info->mti_env;
1929 struct md_op_spec *spec = &info->mti_spec;
1930 struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1931 struct mdt_object *obj = NULL;
1932 struct mdt_object *local_root;
1933 static const struct lu_name lname = {
1934 .ln_name = "i_am_nobody",
1935 .ln_namelen = sizeof("i_am_nobody") - 1,
1937 struct lu_ucred *uc;
1938 cfs_cap_t uc_cap_save;
1942 rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1944 RETURN(ERR_PTR(rc));
1946 local_root = mdt_object_find(env, mdt, local_root_fid);
1947 if (IS_ERR(local_root))
1950 obj = mdt_object_new(env, mdt, fid);
1952 GOTO(out, rc = PTR_ERR(obj));
1954 spec->sp_cr_lookup = 0;
1955 spec->sp_feat = &dt_directory_features;
1956 spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1957 spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1958 if (attr->ma_valid & MA_LOV) {
1959 spec->u.sp_ea.eadata = attr->ma_lmm;
1960 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1961 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1963 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1967 uc_cap_save = uc->uc_cap;
1968 uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1969 rc = mdo_create(env, mdt_object_child(local_root), &lname,
1970 mdt_object_child(obj), spec, attr);
1971 uc->uc_cap = uc_cap_save;
1973 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1974 mdt_obd_name(mdt), PFID(fid), rc);
1978 rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1980 CERROR("%s: cannot open volatile file "DFID", orphan "
1981 "file will be left in PENDING directory until "
1982 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1989 mdt_object_put(env, obj);
1992 mdt_object_put(env, local_root);
1996 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1999 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
2000 struct close_data *data;
2001 struct ldlm_lock *lease;
2002 struct mdt_object *orphan;
2003 struct md_attr *orp_ma;
2010 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2013 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2017 lease = ldlm_handle2lock(&data->cd_handle);
2021 /* try to hold open_sem so that nobody else can open the file */
2022 if (!down_write_trylock(&o->mot_open_sem)) {
2023 ldlm_lock_cancel(lease);
2024 GOTO(out_reprocess, rc = -EBUSY);
2027 /* Check if the lease open lease has already canceled */
2028 lock_res_and_lock(lease);
2029 lease_broken = ldlm_is_cancel(lease);
2030 unlock_res_and_lock(lease);
2032 LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2033 PFID(mdt_object_fid(o)), lease_broken);
2035 /* Cancel server side lease. Client side counterpart should
2036 * have been cancelled. It's okay to cancel it now as we've
2037 * held mot_open_sem. */
2038 ldlm_lock_cancel(lease);
2040 if (lease_broken) /* don't perform release task */
2041 GOTO(out_unlock, rc = -ESTALE);
2043 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2044 GOTO(out_unlock, rc = -EINVAL);
2046 /* ma_need was set before but it seems fine to change it in order to
2047 * avoid modifying the one from RPC */
2048 ma->ma_need = MA_HSM;
2049 rc = mdt_attr_get_complex(info, o, ma);
2051 GOTO(out_unlock, rc);
2053 if (!mdt_hsm_release_allow(ma))
2054 GOTO(out_unlock, rc = -EPERM);
2056 /* already released? */
2057 if (ma->ma_hsm.mh_flags & HS_RELEASED)
2058 GOTO(out_unlock, rc = 0);
2060 /* Compare on-disk and packed data_version */
2061 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
2062 CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
2063 " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
2064 data->cd_data_version, ma->ma_hsm.mh_arch_ver);
2065 GOTO(out_unlock, rc = -EPERM);
2068 ma->ma_valid = MA_INODE;
2069 ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
2070 rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
2072 GOTO(out_unlock, rc);
2074 ma->ma_need = MA_INODE | MA_LOV;
2075 rc = mdt_attr_get_complex(info, o, ma);
2077 GOTO(out_unlock, rc);
2079 if (!(ma->ma_valid & MA_LOV)) {
2080 /* Even empty file are released */
2081 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2082 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2083 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2084 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2085 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2087 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
2088 * ma_lmm as lov_user_md, then it will be confused by union of
2089 * layout_gen and stripe_offset. */
2090 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
2091 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2092 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
2093 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
2095 GOTO(out_unlock, rc = -EINVAL);
2098 /* Set file as released */
2099 ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
2101 /* Hopefully it's not used in this call path */
2102 orp_ma = &info->mti_u.som.attr;
2103 orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2104 orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
2105 orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
2106 orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2107 orp_ma->ma_lmm = ma->ma_lmm;
2108 orp_ma->ma_lmm_size = ma->ma_lmm_size;
2109 orp_ma->ma_valid = MA_INODE | MA_LOV;
2110 orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2112 if (IS_ERR(orphan)) {
2113 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2114 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2116 GOTO(out_unlock, rc = PTR_ERR(orphan));
2119 /* Set up HSM attribute for orphan object */
2120 CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
2121 buf = &info->mti_buf;
2122 buf->lb_buf = info->mti_xattr_buf;
2123 buf->lb_len = sizeof(struct hsm_attrs);
2124 ma->ma_hsm.mh_flags |= HS_RELEASED;
2125 lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2126 ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2128 mdt_lock_reg_init(lh, LCK_EX);
2129 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2130 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
2132 GOTO(out_close, rc);
2134 rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2138 /* Swap layout with orphan object */
2139 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2140 mdt_object_child(orphan),
2141 SWAP_LAYOUTS_MDS_HSM);
2143 /* Release exclusive LL */
2144 mdt_object_unlock(info, o, lh, 1);
2149 /* Close orphan object anyway */
2150 rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2153 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2154 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2155 LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2157 mdt_object_put(info->mti_env, orphan);
2160 up_write(&o->mot_open_sem);
2162 if (rc == 0) { /* already released */
2163 struct mdt_body *repbody;
2164 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2165 LASSERT(repbody != NULL);
2166 repbody->valid |= OBD_MD_FLRELEASED;
2170 ldlm_reprocess_all(lease->l_resource);
2171 LDLM_LOCK_PUT(lease);
2179 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
2180 MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
2182 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2184 return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
2187 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2189 struct mdt_object *o = mfd->mfd_object;
2190 struct md_object *next = mdt_object_child(o);
2191 struct md_attr *ma = &info->mti_attr;
2192 int ret = MDT_IOEPOCH_CLOSED;
2197 mode = mfd->mfd_mode;
2199 if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
2200 rc = mdt_hsm_release(info, o, ma);
2202 CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2203 mdt_obd_name(info->mti_mdt),
2204 PFID(mdt_object_fid(o)), rc);
2205 /* continue to close even error occurred. */
2209 if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
2211 ret = mdt_ioepoch_close(info, o);
2212 } else if (mode & MDS_FMODE_EXEC) {
2214 } else if (mode & MDS_FMODE_EPOCH) {
2215 ret = mdt_ioepoch_close(info, o);
2216 } else if (mode & MDS_FMODE_SOM) {
2217 ret = mdt_som_au_close(info, o);
2220 /* Update atime on close only. */
2221 if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
2222 && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
2223 /* Set the atime only. */
2224 ma->ma_valid = MA_INODE;
2225 ma->ma_attr.la_valid = LA_ATIME;
2226 rc = mo_attr_set(info->mti_env, next, ma);
2229 /* If file data is modified, add the dirty flag. */
2230 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2231 rc = mdt_add_dirty_flag(info, o, ma);
2233 ma->ma_need |= MA_INODE;
2234 ma->ma_valid &= ~MA_INODE;
2236 if (!MFD_CLOSED(mode))
2237 rc = mo_close(info->mti_env, next, ma, mode);
2239 if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
2240 struct mdt_export_data *med;
2242 /* The IOepoch is still opened or SOM update is needed.
2243 * Put mfd back into the list. */
2244 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
2245 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
2246 MDS_FMODE_EPOCH : MDS_FMODE_SOM);
2248 LASSERT(mdt_info_req(info));
2249 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
2250 spin_lock(&med->med_open_lock);
2251 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
2252 class_handle_hash_back(&mfd->mfd_handle);
2253 spin_unlock(&med->med_open_lock);
2255 if (ret == MDT_IOEPOCH_OPENED) {
2259 CDEBUG(D_INODE, "Size-on-MDS attribute update is "
2260 "needed on "DFID"\n", PFID(mdt_object_fid(o)));
2263 /* adjust open and lease count */
2264 if (mode & MDS_OPEN_LEASE) {
2265 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2266 atomic_dec(&o->mot_lease_count);
2268 LASSERT(atomic_read(&o->mot_open_count) > 0);
2269 atomic_dec(&o->mot_open_count);
2272 mdt_object_put(info->mti_env, o);
2275 RETURN(rc ? rc : ret);
2278 int mdt_close(struct tgt_session_info *tsi)
2280 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2281 struct ptlrpc_request *req = tgt_ses_req(tsi);
2282 struct mdt_export_data *med;
2283 struct mdt_file_data *mfd;
2284 struct mdt_object *o;
2285 struct md_attr *ma = &info->mti_attr;
2286 struct mdt_body *repbody = NULL;
2290 mdt_counter_incr(req, LPROC_MDT_CLOSE);
2291 /* Close may come with the Size-on-MDS update. Unpack it. */
2292 rc = mdt_close_unpack(info);
2294 GOTO(out, rc = err_serious(rc));
2296 LASSERT(info->mti_ioepoch);
2298 /* These fields are no longer used and are left for compatibility.
2299 * size is always zero */
2300 req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2302 req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2304 rc = req_capsule_server_pack(info->mti_pill);
2305 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2306 mdt_client_compatibility(info);
2308 mdt_fix_reply(info);
2309 mdt_exit_ucred(info);
2310 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2313 /* Continue to close handle even if we can not pack reply */
2315 repbody = req_capsule_server_get(info->mti_pill,
2317 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2319 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2322 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
2323 repbody->eadatasize = 0;
2324 repbody->aclsize = 0;
2326 rc = err_serious(rc);
2329 med = &req->rq_export->exp_mdt_data;
2330 spin_lock(&med->med_open_lock);
2331 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2332 req_is_replay(req));
2333 if (mdt_mfd_closed(mfd)) {
2334 spin_unlock(&med->med_open_lock);
2335 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2336 ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
2337 info->mti_ioepoch->handle.cookie);
2338 /** not serious error since bug 3633 */
2341 class_handle_unhash(&mfd->mfd_handle);
2342 cfs_list_del_init(&mfd->mfd_list);
2343 spin_unlock(&med->med_open_lock);
2345 /* Do not lose object before last unlink. */
2346 o = mfd->mfd_object;
2347 mdt_object_get(info->mti_env, o);
2348 ret = mdt_mfd_close(info, mfd);
2349 if (repbody != NULL)
2350 rc = mdt_handle_last_unlink(info, o, ma);
2351 mdt_empty_transno(info, rc);
2352 mdt_object_put(info->mti_env, o);
2354 if (repbody != NULL) {
2355 mdt_client_compatibility(info);
2356 rc = mdt_fix_reply(info);
2359 mdt_exit_ucred(info);
2360 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2361 GOTO(out, rc = err_serious(-ENOMEM));
2363 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2364 OBD_FAIL_MDS_CLOSE_NET_REP))
2365 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2367 mdt_thread_info_fini(info);
2368 RETURN(rc ? rc : ret);
2372 * DONE_WRITING rpc handler.
2374 * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
2375 * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
2376 * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
2377 * skip attributes and reconstruct the reply here.
2379 int mdt_done_writing(struct tgt_session_info *tsi)
2381 struct ptlrpc_request *req = tgt_ses_req(tsi);
2382 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2383 struct mdt_body *repbody = NULL;
2384 struct mdt_export_data *med;
2385 struct mdt_file_data *mfd;
2389 rc = req_capsule_server_pack(tsi->tsi_pill);
2391 GOTO(out, rc = err_serious(rc));
2393 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2394 repbody->eadatasize = 0;
2395 repbody->aclsize = 0;
2397 /* Done Writing may come with the Size-on-MDS update. Unpack it. */
2398 rc = mdt_close_unpack(info);
2400 GOTO(out, rc = err_serious(rc));
2402 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2403 mdt_exit_ucred(info);
2404 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2407 med = &info->mti_exp->exp_mdt_data;
2408 spin_lock(&med->med_open_lock);
2409 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2410 req_is_replay(req));
2412 spin_unlock(&med->med_open_lock);
2413 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
2414 ": cookie = "LPX64" ioepoch = "LPU64"\n",
2415 PFID(info->mti_rr.rr_fid1),
2416 info->mti_ioepoch->handle.cookie,
2417 info->mti_ioepoch->ioepoch);
2418 /* If this is a replay, reconstruct the transno. */
2419 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2420 rc = info->mti_ioepoch->flags & MF_SOM_AU ?
2422 mdt_empty_transno(info, rc);
2425 GOTO(error_ucred, rc);
2428 LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
2429 mfd->mfd_mode == MDS_FMODE_TRUNC);
2430 class_handle_unhash(&mfd->mfd_handle);
2431 cfs_list_del_init(&mfd->mfd_list);
2432 spin_unlock(&med->med_open_lock);
2434 /* Set EPOCH CLOSE flag if not set by client. */
2435 info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
2436 info->mti_attr.ma_valid = 0;
2438 info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
2439 OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2440 if (info->mti_attr.ma_lmm == NULL)
2441 GOTO(error_ucred, rc = -ENOMEM);
2443 rc = mdt_mfd_close(info, mfd);
2445 OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2446 mdt_empty_transno(info, rc);
2448 mdt_exit_ucred(info);
2450 mdt_thread_info_fini(info);