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_obf(mdt_object_fid(p)) ||
127 fid_is_dot_lustre(mdt_object_fid(p))))
128 GOTO(unlock, rc = -EPERM);
130 rc = mdo_create_data(info->mti_env,
131 p ? mdt_object_child(p) : NULL,
132 mdt_object_child(o), spec, ma);
134 rc = mdt_attr_get_complex(info, o, ma);
136 if (rc == 0 && ma->ma_valid & MA_LOV)
137 o->mot_flags |= MOF_LOV_CREATED;
140 mutex_unlock(&o->mot_lov_mutex);
144 static int mdt_ioepoch_opened(struct mdt_object *mo)
146 return mo->mot_ioepoch_count;
149 int mdt_object_is_som_enabled(struct mdt_object *mo)
151 return !mo->mot_ioepoch;
155 * Re-enable Size-on-MDS.
156 * Call under ->mot_ioepoch_mutex.
158 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
160 if (ioepoch == mo->mot_ioepoch) {
161 LASSERT(!mdt_ioepoch_opened(mo));
168 * Open the IOEpoch. It is allowed if @writecount is not negative.
169 * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
171 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
174 struct mdt_device *mdt = info->mti_mdt;
179 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
180 !S_ISREG(lu_object_attr(&o->mot_obj)))
183 mutex_lock(&o->mot_ioepoch_mutex);
184 if (mdt_ioepoch_opened(o)) {
185 /* Epoch continues even if there is no writers yet. */
186 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
187 o->mot_ioepoch, PFID(mdt_object_fid(o)));
189 /* XXX: ->mdt_ioepoch is not initialized at the mount */
190 spin_lock(&mdt->mdt_ioepoch_lock);
191 if (mdt->mdt_ioepoch < info->mti_replayepoch)
192 mdt->mdt_ioepoch = info->mti_replayepoch;
194 if (info->mti_replayepoch)
195 o->mot_ioepoch = info->mti_replayepoch;
196 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
197 o->mot_ioepoch = ++mdt->mdt_ioepoch;
199 o->mot_ioepoch = mdt->mdt_ioepoch;
201 spin_unlock(&mdt->mdt_ioepoch_lock);
203 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
204 o->mot_ioepoch, PFID(mdt_object_fid(o)));
206 o->mot_flags |= MOF_SOM_CREATED;
209 o->mot_ioepoch_count++;
210 mutex_unlock(&o->mot_ioepoch_mutex);
212 /* Cancel Size-on-MDS attributes cached on clients for the open case.
213 * In the truncate case, see mdt_reint_setattr(). */
214 if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
215 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
216 mdt_lock_reg_init(lh, LCK_EX);
217 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
220 mdt_object_unlock(info, o, lh, 1);
226 * Update SOM on-disk attributes.
227 * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
228 * mountid and attributes. If disabling, clean SOM xattr.
229 * Call under ->mot_ioepoch_mutex.
231 static int mdt_som_attr_set(struct mdt_thread_info *info,
232 struct mdt_object *obj, __u64 ioepoch, bool enable)
234 struct md_object *next = mdt_object_child(obj);
238 CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
239 " on "DFID".\n", enable ? "update" : "disabling",
240 ioepoch, PFID(mdt_object_fid(obj)));
243 struct lu_buf *buf = &info->mti_buf;
244 struct som_attrs *attrs;
245 struct md_attr *ma = &info->mti_attr;
246 struct lu_attr *la = &ma->ma_attr;
247 struct obd_device *obd = info->mti_mdt->mdt_lut.lut_obd;
249 attrs = (struct som_attrs *)info->mti_xattr_buf;
250 CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*attrs));
252 /* pack SOM attributes */
253 memset(attrs, 0, sizeof(*attrs));
254 attrs->som_ioepoch = ioepoch;
255 attrs->som_mountid = obd->u.obt.obt_mount_count;
256 if ((la->la_valid & LA_SIZE) != 0)
257 attrs->som_size = la->la_size;
258 if ((la->la_valid & LA_BLOCKS) != 0)
259 attrs->som_blocks = la->la_blocks;
260 lustre_som_swab(attrs);
262 /* update SOM attributes */
264 buf->lb_len = sizeof(*attrs);
265 rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_SOM, 0);
267 /* delete SOM attributes */
268 rc = mo_xattr_del(info->mti_env, next, XATTR_NAME_SOM);
274 /** Perform the eviction specific actions on ioepoch close. */
275 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
276 struct mdt_object *o)
280 mutex_lock(&o->mot_ioepoch_mutex);
281 CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
282 "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
283 o->mot_ioepoch_count);
284 o->mot_ioepoch_count--;
286 /* If eviction occured set MOF_SOM_RECOV,
287 * if no other epoch holders, disable SOM on disk. */
288 o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
289 if (!mdt_ioepoch_opened(o)) {
290 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
291 mdt_object_som_enable(o, o->mot_ioepoch);
293 mutex_unlock(&o->mot_ioepoch_mutex);
298 * Perform the replay specific actions on ioepoch close.
299 * Skip SOM attribute update if obtained and just forget about the inode state
300 * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
302 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
303 struct mdt_object *o)
305 int rc = MDT_IOEPOCH_CLOSED;
308 mutex_lock(&o->mot_ioepoch_mutex);
309 CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
310 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
311 o->mot_ioepoch_count--;
313 /* Get an info from the replayed request if client is supposed
314 * to send an Attibute Update, reconstruct @rc if so */
315 if (info->mti_ioepoch->flags & MF_SOM_AU)
316 rc = MDT_IOEPOCH_GETATTR;
318 if (!mdt_ioepoch_opened(o))
319 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
320 mutex_unlock(&o->mot_ioepoch_mutex);
326 * Regular file IOepoch close.
327 * Closes the ioepoch, checks the object state, apply obtained attributes and
328 * re-enable SOM on the object, if possible. Also checks if the recovery is
329 * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
330 * to obtain SOM attributes under the server-side OST locks.
333 * MDT_IOEPOCH_CLOSED if ioepoch is closed.
334 * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
336 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
337 struct mdt_object *o)
339 struct md_attr *tmp_ma;
343 int rc = 0, ret = MDT_IOEPOCH_CLOSED;
346 la = &info->mti_attr.ma_attr;
347 achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
349 mutex_lock(&o->mot_ioepoch_mutex);
350 o->mot_ioepoch_count--;
352 tmp_ma = &info->mti_u.som.attr;
353 tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
354 tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
355 tmp_ma->ma_som = &info->mti_u.som.data;
356 tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
357 tmp_ma->ma_valid = 0;
358 rc = mdt_attr_get_complex(info, o, tmp_ma);
362 /* Check the on-disk SOM state. */
363 if (o->mot_flags & MOF_SOM_RECOV)
365 else if (!(o->mot_flags & MOF_SOM_CREATED) &&
366 !(tmp_ma->ma_valid & MA_SOM))
369 CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
370 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
372 opened = mdt_ioepoch_opened(o);
374 * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
375 * Skip the check for file with no LOV or for unlink files.
377 if (!opened && tmp_ma->ma_valid & MA_LOV &&
378 !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
380 /* If some previous writer was evicted, re-ask the
381 * client for attributes. Even if attributes are
382 * provided, we cannot believe in them.
383 * Another use case is that there is no SOM cache on
384 * disk -- first access with SOM or there was an MDS
386 ret = MDT_IOEPOCH_GETATTR;
387 else if (o->mot_flags & MOF_SOM_CHANGE)
388 /* Some previous writer changed the attribute.
389 * Do not believe to the current Size-on-MDS
390 * update, re-ask client. */
391 ret = MDT_IOEPOCH_GETATTR;
392 else if (!(la->la_valid & LA_SIZE) && achange)
393 /* Attributes were changed by the last writer
394 * only but no Size-on-MDS update is received.*/
395 ret = MDT_IOEPOCH_GETATTR;
398 if (achange || ret == MDT_IOEPOCH_GETATTR)
399 o->mot_flags |= MOF_SOM_CHANGE;
401 /* If epoch ends and relable SOM attributes are obtained, update them.
402 * Create SOM ea for new files even if there is no attributes obtained
403 * (0-length file). */
404 if (ret == MDT_IOEPOCH_CLOSED && !opened) {
405 if (achange || o->mot_flags & MOF_SOM_CREATED) {
406 LASSERT(achange || !(la->la_valid & LA_SIZE));
407 rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
409 /* Avoid the following setattrs of these attributes,
410 * e.g. for atime update. */
411 info->mti_attr.ma_valid = 0;
413 mdt_object_som_enable(o, o->mot_ioepoch);
416 mutex_unlock(&o->mot_ioepoch_mutex);
417 /* If recovery is needed, tell the client to perform GETATTR under
419 if (ret == MDT_IOEPOCH_GETATTR && recovery) {
420 struct mdt_body *rep;
421 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
422 rep->valid |= OBD_MD_FLGETATTRLOCK;
428 mutex_unlock(&o->mot_ioepoch_mutex);
433 * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
434 * - a client closes the IOEpoch;
435 * - a client eviction occured.
437 * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
438 * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
439 * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
442 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
444 struct ptlrpc_request *req = mdt_info_req(info);
447 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
448 !S_ISREG(lu_object_attr(&o->mot_obj)))
451 LASSERT(o->mot_ioepoch_count);
452 LASSERT(info->mti_ioepoch == NULL ||
453 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
455 /* IOEpoch is closed only if client tells about it or eviction occures.
456 * In the replay case, always close the epoch. */
458 RETURN(mdt_ioepoch_close_on_eviction(info, o));
459 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
460 RETURN(mdt_ioepoch_close_on_replay(info, o));
461 if (info->mti_ioepoch && (info->mti_ioepoch->flags & MF_EPOCH_CLOSE))
462 RETURN(mdt_ioepoch_close_reg(info, o));
463 /* IO epoch is not closed. */
464 RETURN(MDT_IOEPOCH_OPENED);
468 * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
469 * for attribute update. It happens if:
470 * - SOM Attribute Update is obtained;
471 * - the client failed to obtain it and informs MDS about it;
472 * - a client eviction occured.
473 * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
476 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
478 struct ptlrpc_request *req = mdt_info_req(info);
480 int act = MDT_SOM_ENABLE;
484 LASSERT(!req || info->mti_ioepoch);
485 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
486 !S_ISREG(lu_object_attr(&o->mot_obj)))
489 /* No size whereas MF_SOM_CHANGE is set means client failed to
490 * obtain ost attributes, drop the SOM cache on disk if so. */
492 (info->mti_ioepoch &&
493 info->mti_ioepoch->flags & MF_SOM_CHANGE &&
494 !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
495 act = MDT_SOM_DISABLE;
497 mutex_lock(&o->mot_ioepoch_mutex);
498 /* Mark the object it is the recovery state if we failed to obtain
500 if (act == MDT_SOM_DISABLE)
501 o->mot_flags |= MOF_SOM_RECOV;
503 if (!mdt_ioepoch_opened(o)) {
504 ioepoch = info->mti_ioepoch ?
505 info->mti_ioepoch->ioepoch : o->mot_ioepoch;
508 && !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
509 rc = mdt_som_attr_set(info, o, ioepoch, act);
510 mdt_object_som_enable(o, ioepoch);
512 mutex_unlock(&o->mot_ioepoch_mutex);
516 int mdt_write_read(struct mdt_object *o)
520 mutex_lock(&o->mot_ioepoch_mutex);
521 rc = o->mot_writecount;
522 mutex_unlock(&o->mot_ioepoch_mutex);
526 int mdt_write_get(struct mdt_object *o)
530 mutex_lock(&o->mot_ioepoch_mutex);
531 if (o->mot_writecount < 0)
535 mutex_unlock(&o->mot_ioepoch_mutex);
539 void mdt_write_put(struct mdt_object *o)
542 mutex_lock(&o->mot_ioepoch_mutex);
544 mutex_unlock(&o->mot_ioepoch_mutex);
548 static int mdt_write_deny(struct mdt_object *o)
552 mutex_lock(&o->mot_ioepoch_mutex);
553 if (o->mot_writecount > 0)
557 mutex_unlock(&o->mot_ioepoch_mutex);
561 static void mdt_write_allow(struct mdt_object *o)
564 mutex_lock(&o->mot_ioepoch_mutex);
566 mutex_unlock(&o->mot_ioepoch_mutex);
570 /* there can be no real transaction so prepare the fake one */
571 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
573 struct mdt_device *mdt = info->mti_mdt;
574 struct ptlrpc_request *req = mdt_info_req(info);
575 struct tg_export_data *ted;
576 struct lsd_client_data *lcd;
579 /* transaction has occurred already */
580 if (lustre_msg_get_transno(req->rq_repmsg) != 0)
583 spin_lock(&mdt->mdt_lut.lut_translock);
585 if (info->mti_transno != 0) {
586 struct obd_export *exp = req->rq_export;
588 CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
589 mdt_obd_name(mdt), info->mti_transno,
590 libcfs_nid2str(exp->exp_connection->c_peer.nid),
592 spin_unlock(&mdt->mdt_lut.lut_translock);
595 } else if (info->mti_transno == 0) {
596 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
598 /* should be replay */
599 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
600 mdt->mdt_lut.lut_last_transno = info->mti_transno;
602 spin_unlock(&mdt->mdt_lut.lut_translock);
604 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
606 req->rq_export->exp_obd->obd_last_committed);
608 req->rq_transno = info->mti_transno;
609 lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
611 /* update lcd in memory only for resent cases */
612 ted = &req->rq_export->exp_target_data;
614 mutex_lock(&ted->ted_lcd_lock);
616 if (info->mti_transno < lcd->lcd_last_transno &&
617 info->mti_transno != 0) {
618 /* This should happen during replay. Do not update
619 * last rcvd info if replay req transno < last transno,
620 * otherwise the following resend(after replay) can not
621 * be checked correctly by xid */
622 mutex_unlock(&ted->ted_lcd_lock);
623 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
624 mdt_obd_name(mdt), info->mti_transno,
625 lcd->lcd_last_transno);
629 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
630 lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
631 if (info->mti_transno != 0)
632 lcd->lcd_last_close_transno = info->mti_transno;
633 lcd->lcd_last_close_xid = req->rq_xid;
634 lcd->lcd_last_close_result = rc;
636 /* VBR: save versions in last_rcvd for reconstruct. */
637 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
639 lcd->lcd_pre_versions[0] = pre_versions[0];
640 lcd->lcd_pre_versions[1] = pre_versions[1];
641 lcd->lcd_pre_versions[2] = pre_versions[2];
642 lcd->lcd_pre_versions[3] = pre_versions[3];
644 if (info->mti_transno != 0)
645 lcd->lcd_last_transno = info->mti_transno;
647 lcd->lcd_last_xid = req->rq_xid;
648 lcd->lcd_last_result = rc;
649 lcd->lcd_last_data = info->mti_opdata;
651 mutex_unlock(&ted->ted_lcd_lock);
656 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
658 LASSERT(mfd != NULL);
660 CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n",
661 PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
663 mfd->mfd_mode = mode;
666 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
667 struct mdt_object *o, __u64 flags, int created)
669 struct ptlrpc_request *req = mdt_info_req(info);
670 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
671 struct mdt_file_data *mfd;
672 struct md_attr *ma = &info->mti_attr;
673 struct lu_attr *la = &ma->ma_attr;
674 struct mdt_body *repbody;
675 int rc = 0, isdir, isreg;
678 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
680 isreg = S_ISREG(la->la_mode);
681 isdir = S_ISDIR(la->la_mode);
682 if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
684 * No EA, check whether it is will set regEA and dirEA since in
685 * above attr get, these size might be zero, so reset it, to
686 * retrieve the MD after create obj.
688 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
691 /* in replay case, p == NULL */
692 rc = mdt_create_data(info, p, o);
697 CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
698 ma->ma_valid, ma->ma_lmm_size);
700 if (ma->ma_valid & MA_LOV) {
701 LASSERT(ma->ma_lmm_size != 0);
702 repbody->eadatasize = ma->ma_lmm_size;
704 repbody->valid |= OBD_MD_FLDIREA;
706 repbody->valid |= OBD_MD_FLEASIZE;
709 if (flags & FMODE_WRITE) {
710 rc = mdt_write_get(o);
712 mdt_ioepoch_open(info, o, created);
713 repbody->ioepoch = o->mot_ioepoch;
715 } else if (flags & MDS_FMODE_EXEC) {
716 rc = mdt_write_deny(o);
721 rc = mo_open(info->mti_env, mdt_object_child(o),
722 created ? flags | MDS_OPEN_CREATED : flags);
726 mfd = mdt_mfd_new(med);
728 GOTO(err_out, rc = -ENOMEM);
731 * Keep a reference on this object for this open, and is
732 * released by mdt_mfd_close().
734 mdt_object_get(info->mti_env, o);
736 mfd->mfd_xid = req->rq_xid;
739 * @flags is always not zero. At least it should be FMODE_READ,
740 * FMODE_WRITE or MDS_FMODE_EXEC.
745 mdt_mfd_set_mode(mfd, flags);
747 atomic_inc(&o->mot_open_count);
748 if (flags & MDS_OPEN_LEASE)
749 atomic_inc(&o->mot_lease_count);
752 if (req_is_replay(req)) {
753 struct mdt_file_data *old_mfd;
754 /* Check wheather old cookie already exist in
755 * the list, becasue when do recovery, client
756 * might be disconnected from server, and
757 * restart replay, so there maybe some orphan
758 * mfd here, we should remove them */
759 LASSERT(info->mti_rr.rr_handle != NULL);
760 spin_lock(&med->med_open_lock);
761 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
762 if (old_mfd != NULL) {
763 CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
764 "cookie = "LPX64"\n", mfd,
765 PFID(mdt_object_fid(mfd->mfd_object)),
766 info->mti_rr.rr_handle->cookie);
767 class_handle_unhash(&old_mfd->mfd_handle);
768 cfs_list_del_init(&old_mfd->mfd_list);
769 spin_unlock(&med->med_open_lock);
770 /* no attr update for that close */
772 ma->ma_valid |= MA_FLAGS;
773 ma->ma_attr_flags |= MDS_RECOV_OPEN;
774 mdt_mfd_close(info, old_mfd);
775 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
776 ma->ma_valid &= ~MA_FLAGS;
778 spin_unlock(&med->med_open_lock);
779 CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
780 "cookie = "LPX64"\n",
781 PFID(mdt_object_fid(mfd->mfd_object)),
782 info->mti_rr.rr_handle->cookie);
785 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
786 info->mti_rr.rr_handle->cookie);
788 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
791 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
793 if (req->rq_export->exp_disconnected) {
794 spin_lock(&med->med_open_lock);
795 class_handle_unhash(&mfd->mfd_handle);
796 cfs_list_del_init(&mfd->mfd_list);
797 spin_unlock(&med->med_open_lock);
798 mdt_mfd_close(info, mfd);
800 spin_lock(&med->med_open_lock);
801 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
802 spin_unlock(&med->med_open_lock);
805 mdt_empty_transno(info, rc);
810 if (flags & FMODE_WRITE)
811 /* XXX We also need to close io epoch here.
812 * See LU-1220 - green */
814 else if (flags & FMODE_EXEC)
819 int mdt_finish_open(struct mdt_thread_info *info,
820 struct mdt_object *p, struct mdt_object *o,
821 __u64 flags, int created, struct ldlm_reply *rep)
823 struct ptlrpc_request *req = mdt_info_req(info);
824 struct obd_export *exp = req->rq_export;
825 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
826 struct md_attr *ma = &info->mti_attr;
827 struct lu_attr *la = &ma->ma_attr;
828 struct mdt_file_data *mfd;
829 struct mdt_body *repbody;
831 int isreg, isdir, islnk;
835 LASSERT(ma->ma_valid & MA_INODE);
837 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
839 isreg = S_ISREG(la->la_mode);
840 isdir = S_ISDIR(la->la_mode);
841 islnk = S_ISLNK(la->la_mode);
842 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
844 /* LU-2275, simulate broken behaviour (esp. prevalent in
845 * pre-2.4 servers where a very strange reply is sent on error
846 * that looks like it was actually almost succesful and a failure at the
848 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
849 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
853 if (flags & MDS_OPEN_LOCK)
854 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
859 if (exp_connect_rmtclient(exp)) {
860 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
862 rc = mdt_pack_remote_perm(info, o, buf);
864 repbody->valid &= ~OBD_MD_FLRMTPERM;
865 repbody->aclsize = 0;
867 repbody->valid |= OBD_MD_FLRMTPERM;
868 repbody->aclsize = sizeof(struct mdt_remote_perm);
871 #ifdef CONFIG_FS_POSIX_ACL
872 else if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
873 const struct lu_env *env = info->mti_env;
874 struct md_object *next = mdt_object_child(o);
875 struct lu_buf *buf = &info->mti_buf;
877 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
878 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
880 if (buf->lb_len > 0) {
881 rc = mo_xattr_get(env, next, buf,
882 XATTR_NAME_ACL_ACCESS);
884 if (rc == -ENODATA) {
885 repbody->aclsize = 0;
886 repbody->valid |= OBD_MD_FLACL;
888 } else if (rc == -EOPNOTSUPP) {
891 CERROR("got acl size: %d\n", rc);
894 repbody->aclsize = rc;
895 repbody->valid |= OBD_MD_FLACL;
902 if (info->mti_mdt->mdt_lut.lut_mds_capa &&
903 exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) {
904 struct lustre_capa *capa;
906 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
908 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
909 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
912 repbody->valid |= OBD_MD_FLMDSCAPA;
914 if (info->mti_mdt->mdt_lut.lut_oss_capa &&
915 exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA &&
916 S_ISREG(lu_object_attr(&o->mot_obj))) {
917 struct lustre_capa *capa;
919 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
921 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
922 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
925 repbody->valid |= OBD_MD_FLOSSCAPA;
929 * If we are following a symlink, don't open; and do not return open
930 * handle for special nodes as client required.
932 if (islnk || (!isreg && !isdir &&
933 (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
934 lustre_msg_set_transno(req->rq_repmsg, 0);
939 * We need to return the existing object's fid back, so it is done here,
940 * after preparing the reply.
942 if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
945 /* This can't be done earlier, we need to return reply body */
947 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
948 /* We are trying to create or write an existing dir. */
951 } else if (flags & MDS_OPEN_DIRECTORY)
954 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
955 OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
960 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
961 spin_lock(&med->med_open_lock);
962 cfs_list_for_each(t, &med->med_open_head) {
963 mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
964 if (mfd->mfd_xid == req->rq_xid)
968 spin_unlock(&med->med_open_lock);
971 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
972 /*set repbody->ea_size for resent case*/
973 if (ma->ma_valid & MA_LOV) {
974 LASSERT(ma->ma_lmm_size != 0);
975 repbody->eadatasize = ma->ma_lmm_size;
977 repbody->valid |= OBD_MD_FLDIREA;
979 repbody->valid |= OBD_MD_FLEASIZE;
981 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
986 rc = mdt_mfd_open(info, p, o, flags, created);
988 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
993 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
994 struct lsd_client_data *lcd);
996 void mdt_reconstruct_open(struct mdt_thread_info *info,
997 struct mdt_lock_handle *lhc)
999 const struct lu_env *env = info->mti_env;
1000 struct mdt_device *mdt = info->mti_mdt;
1001 struct req_capsule *pill = info->mti_pill;
1002 struct ptlrpc_request *req = mdt_info_req(info);
1003 struct tg_export_data *ted = &req->rq_export->exp_target_data;
1004 struct lsd_client_data *lcd = ted->ted_lcd;
1005 struct md_attr *ma = &info->mti_attr;
1006 struct mdt_reint_record *rr = &info->mti_rr;
1007 __u64 flags = info->mti_spec.sp_cr_flags;
1008 struct ldlm_reply *ldlm_rep;
1009 struct mdt_object *parent;
1010 struct mdt_object *child;
1011 struct mdt_body *repbody;
1015 LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1016 ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1017 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1019 ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
1020 ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
1022 ma->ma_need = MA_INODE | MA_HSM;
1023 if (ma->ma_lmm_size > 0)
1024 ma->ma_need |= MA_LOV;
1028 mdt_req_from_lcd(req, lcd);
1029 mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1031 CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1032 ldlm_rep->lock_policy_res1, req->rq_status);
1034 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1035 req->rq_status != 0)
1036 /* We did not create successfully, return error to client. */
1037 GOTO(out, rc = req->rq_status);
1039 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1040 struct obd_export *exp = req->rq_export;
1042 * We failed after creation, but we do not know in which step
1043 * we failed. So try to check the child object.
1045 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1046 if (IS_ERR(parent)) {
1047 rc = PTR_ERR(parent);
1048 LCONSOLE_WARN("Parent "DFID" lookup error %d."
1049 " Evicting client %s with export %s.\n",
1050 PFID(rr->rr_fid1), rc,
1051 obd_uuid2str(&exp->exp_client_uuid),
1052 obd_export_nid2str(exp));
1053 mdt_export_evict(exp);
1056 child = mdt_object_find(env, mdt, rr->rr_fid2);
1057 if (IS_ERR(child)) {
1058 rc = PTR_ERR(child);
1059 LCONSOLE_WARN("Child "DFID" lookup error %d."
1060 " Evicting client %s with export %s.\n",
1061 PFID(mdt_object_fid(child)), rc,
1062 obd_uuid2str(&exp->exp_client_uuid),
1063 obd_export_nid2str(exp));
1064 mdt_object_put(env, parent);
1065 mdt_export_evict(exp);
1069 if (unlikely(mdt_object_remote(child))) {
1070 /* the child object was created on remote server */
1071 if (!mdt_is_dne_client(exp)) {
1072 /* Return -EIO for old client */
1073 mdt_object_put(env, parent);
1074 mdt_object_put(env, child);
1075 GOTO(out, rc = -EIO);
1077 repbody->fid1 = *rr->rr_fid2;
1078 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1081 if (mdt_object_exists(child)) {
1082 mdt_set_capainfo(info, 1, rr->rr_fid2,
1084 rc = mdt_attr_get_complex(info, child, ma);
1086 rc = mdt_finish_open(info, parent,
1090 /* the child does not exist, we should do
1092 mdt_object_put(env, parent);
1093 mdt_object_put(env, child);
1094 GOTO(regular_open, 0);
1097 mdt_object_put(env, parent);
1098 mdt_object_put(env, child);
1102 /* We did not try to create, so we are a pure open */
1103 rc = mdt_reint_open(info, lhc);
1108 req->rq_status = rc;
1109 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1110 LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1113 int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
1115 __u64 flags = info->mti_spec.sp_cr_flags;
1116 struct mdt_reint_record *rr = &info->mti_rr;
1117 struct md_attr *ma = &info->mti_attr;
1118 struct mdt_object *o;
1122 o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1124 RETURN(rc = PTR_ERR(o));
1126 if (unlikely(mdt_object_remote(o))) {
1127 /* the child object was created on remote server */
1128 struct mdt_body *repbody;
1130 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1133 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1134 repbody->fid1 = *rr->rr_fid2;
1135 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1138 if (mdt_object_exists(o)) {
1139 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1143 rc = mdt_attr_get_complex(info, o, ma);
1145 rc = mdt_finish_open(info, NULL, o, flags, 0,
1152 mdt_object_put(info->mti_env, o);
1156 /* lock object for open */
1157 static int mdt_object_open_lock(struct mdt_thread_info *info,
1158 struct mdt_object *obj,
1159 struct mdt_lock_handle *lhc,
1162 struct md_attr *ma = &info->mti_attr;
1163 __u64 open_flags = info->mti_spec.sp_cr_flags;
1164 ldlm_mode_t lm = LCK_CR;
1165 bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
1166 bool try_layout = false;
1167 bool create_layout = false;
1172 mdt_lock_handle_init(lhc);
1174 if (req_is_replay(mdt_info_req(info)))
1177 if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
1178 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
1179 md_should_create(open_flags))
1180 create_layout = true;
1181 if (exp_connect_layout(info->mti_exp) && !create_layout &&
1182 ma->ma_need & MA_LOV)
1187 /* lease open, acquire write mode of open sem */
1188 down_write(&obj->mot_open_sem);
1190 /* Lease exists and ask for new lease */
1191 if (atomic_read(&obj->mot_lease_count) > 0) {
1192 /* only exclusive open is supported, so lease
1193 * are conflicted to each other */
1194 GOTO(out, rc = -EBUSY);
1197 /* Lease must be with open lock */
1198 if (!(open_flags & MDS_OPEN_LOCK)) {
1199 CERROR("Request lease for file:"DFID ", but open lock "
1200 "is missed, open_flags = "LPO64".\n",
1201 PFID(mdt_object_fid(obj)), open_flags);
1202 GOTO(out, rc = -EPROTO);
1205 /* XXX: only exclusive open is supported. */
1207 *ibits = MDS_INODELOCK_OPEN;
1209 /* never grant LCK_EX layout lock to client */
1211 } else { /* normal open */
1212 /* normal open holds read mode of open sem */
1213 down_read(&obj->mot_open_sem);
1215 if (open_flags & MDS_OPEN_LOCK) {
1216 if (open_flags & FMODE_WRITE)
1218 else if (open_flags & MDS_FMODE_EXEC)
1223 *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
1224 } else if (atomic_read(&obj->mot_lease_count) > 0) {
1225 if (open_flags & FMODE_WRITE)
1231 *ibits = MDS_INODELOCK_OPEN;
1234 lhc = &info->mti_lh[MDT_LH_LOCAL];
1236 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
1237 PFID(mdt_object_fid(obj)),
1238 atomic_read(&obj->mot_open_count), lm);
1241 mdt_lock_reg_init(lhc, lm);
1243 /* one problem to return layout lock on open is that it may result
1244 * in too many layout locks cached on the client side. */
1245 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
1246 /* return lookup lock to validate inode at the client side,
1247 * this is pretty important otherwise mdt will return layout
1248 * lock for each open.
1249 * However this is a double-edged sword because changing
1250 * permission will revoke huge # of LOOKUP locks. */
1251 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
1252 if (!mdt_object_lock_try(info, obj, lhc, *ibits,
1254 *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
1256 rc = mdt_object_lock(info, obj, lhc, *ibits,
1259 } else if (*ibits != 0) {
1260 rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK);
1263 CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
1264 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
1265 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
1267 /* will change layout, revoke layout locks by enqueuing EX lock. */
1268 if (rc == 0 && create_layout) {
1269 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
1271 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
1272 ", open_flags = "LPO64"\n",
1273 PFID(mdt_object_fid(obj)), open_flags);
1275 LASSERT(!try_layout);
1276 mdt_lock_handle_init(ll);
1277 mdt_lock_reg_init(ll, LCK_EX);
1278 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
1281 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
1284 /* Check if there is any other open handles after acquiring
1285 * open lock. At this point, caching open handles have been revoked
1287 * XXX: Now only exclusive open is supported. Need to check the
1288 * type of open for generic lease support. */
1289 if (rc == 0 && acq_lease) {
1290 struct ptlrpc_request *req = mdt_info_req(info);
1291 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1292 struct mdt_file_data *mfd;
1293 bool is_replay_or_resent;
1296 /* For lease: application can open a file and then apply lease,
1297 * @handle contains original open handle in that case.
1298 * In recovery, open REQ will be replayed and the lease REQ may
1299 * be resent that means the open handle is already stale, so we
1300 * need to fix it up here by finding new handle. */
1301 is_replay_or_resent = req_is_replay(req) ||
1302 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
1304 /* if the request is _not_ a replay request, rr_handle
1305 * may be used to hold an openhandle which is issuing the
1306 * lease request, so that this openhandle doesn't count. */
1307 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
1308 is_replay_or_resent);
1312 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
1313 PFID(mdt_object_fid(obj)),
1314 atomic_read(&obj->mot_open_count), open_count);
1316 if (atomic_read(&obj->mot_open_count) > open_count)
1317 GOTO(out, rc = -EBUSY);
1325 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1326 struct mdt_object *obj,
1327 struct mdt_lock_handle *lhc,
1328 __u64 ibits, int rc)
1330 __u64 open_flags = info->mti_spec.sp_cr_flags;
1331 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1334 if (req_is_replay(mdt_info_req(info)))
1337 /* Release local lock - the lock put in MDT_LH_LOCAL will never
1338 * return to client side. */
1339 if (lustre_handle_is_used(&ll->mlh_reg_lh))
1340 mdt_object_unlock(info, obj, ll, 1);
1342 ll = &info->mti_lh[MDT_LH_LAYOUT];
1343 /* Release local layout lock, layout was created */
1344 if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1345 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1346 mdt_object_unlock(info, obj, ll, 1);
1349 if (open_flags & MDS_OPEN_LEASE)
1350 up_write(&obj->mot_open_sem);
1352 up_read(&obj->mot_open_sem);
1354 /* Cross-ref case, the lock should be returned to the client */
1355 if (ibits == 0 || rc == -EREMOTE)
1358 if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
1359 /* for the open request, the lock will only return to client
1360 * if open or layout lock is granted. */
1365 struct ldlm_reply *ldlm_rep;
1367 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1368 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1369 mdt_object_unlock(info, obj, lhc, 1);
1375 * Check release is permitted for the current HSM flags.
1377 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1379 if (!(ma->ma_valid & MA_HSM))
1382 if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1385 if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1391 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1392 struct mdt_lock_handle *lhc)
1394 const struct lu_env *env = info->mti_env;
1395 struct mdt_device *mdt = info->mti_mdt;
1396 __u64 flags = info->mti_spec.sp_cr_flags;
1397 struct mdt_reint_record *rr = &info->mti_rr;
1398 struct md_attr *ma = &info->mti_attr;
1399 struct mdt_object *parent= NULL;
1400 struct mdt_object *o;
1405 if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1406 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1407 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1408 if (IS_ERR(parent)) {
1409 CDEBUG(D_INODE, "Fail to find parent "DFID
1410 " for anonymous created %ld, try to"
1411 " use server-side parent.\n",
1412 PFID(rr->rr_fid1), PTR_ERR(parent));
1417 ma->ma_need |= MA_PFID;
1420 o = mdt_object_find(env, mdt, rr->rr_fid2);
1422 RETURN(rc = PTR_ERR(o));
1424 if (mdt_object_remote(o)) {
1425 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1426 mdt_obd_name(info->mti_mdt),
1428 GOTO(out, rc = -EREMOTE);
1429 } else if (!mdt_object_exists(o)) {
1430 mdt_set_disposition(info, rep,
1434 GOTO(out, rc = -ENOENT);
1437 mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1439 if (flags & MDS_OPEN_RELEASE)
1440 ma->ma_need |= MA_HSM;
1441 rc = mdt_attr_get_complex(info, o, ma);
1445 /* If a release request, check file flags are fine and ask for an
1446 * exclusive open access. */
1447 if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1448 GOTO(out, rc = -EPERM);
1450 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1452 GOTO(out_unlock, rc);
1454 if (ma->ma_valid & MA_PFID) {
1455 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1456 if (IS_ERR(parent)) {
1457 CDEBUG(D_INODE, "Fail to find parent "DFID
1458 " for anonymous created %ld, try to"
1459 " use system default.\n",
1460 PFID(&ma->ma_pfid), PTR_ERR(parent));
1465 rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1467 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1468 if (flags & MDS_OPEN_LOCK)
1469 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1470 if (flags & MDS_OPEN_LEASE)
1471 mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1473 GOTO(out_unlock, rc);
1476 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1478 mdt_object_put(env, o);
1480 mdt_object_put(env, parent);
1484 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1485 static int mdt_cross_open(struct mdt_thread_info *info,
1486 const struct lu_fid *parent_fid,
1487 const struct lu_fid *fid,
1488 struct ldlm_reply *rep, __u32 flags)
1490 struct md_attr *ma = &info->mti_attr;
1491 struct mdt_object *o;
1495 o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1497 RETURN(rc = PTR_ERR(o));
1499 if (mdt_object_remote(o)) {
1500 /* Something is wrong here, the object is on another MDS! */
1501 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1502 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1503 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1505 "Object isn't on this server! FLD error?\n");
1508 if (mdt_object_exists(o)) {
1509 /* Do permission check for cross-open. */
1510 rc = mo_permission(info->mti_env, NULL,
1511 mdt_object_child(o),
1512 NULL, flags | MDS_OPEN_CROSS);
1516 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1517 rc = mdt_attr_get_complex(info, o, ma);
1521 /* Do not create lov object if the fid is opened
1523 if (S_ISREG(ma->ma_attr.la_mode) &&
1524 !(ma->ma_valid & MA_LOV) && (flags & FMODE_WRITE) &&
1525 fid_is_obf(parent_fid))
1526 GOTO(out, rc = -EPERM);
1528 rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1531 * Something is wrong here. lookup was positive but
1532 * there is no object!
1534 CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1535 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1540 mdt_object_put(info->mti_env, o);
1544 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1546 struct mdt_device *mdt = info->mti_mdt;
1547 struct ptlrpc_request *req = mdt_info_req(info);
1548 struct mdt_object *parent;
1549 struct mdt_object *child;
1550 struct mdt_lock_handle *lh;
1551 struct ldlm_reply *ldlm_rep;
1552 struct mdt_body *repbody;
1553 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1554 struct md_attr *ma = &info->mti_attr;
1555 __u64 create_flags = info->mti_spec.sp_cr_flags;
1557 struct mdt_reint_record *rr = &info->mti_rr;
1558 struct lu_name *lname;
1564 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1565 (obd_timeout + 1) / 4);
1567 mdt_counter_incr(req, LPROC_MDT_OPEN);
1568 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1570 ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1571 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1573 ma->ma_need = MA_INODE;
1574 if (ma->ma_lmm_size > 0)
1575 ma->ma_need |= MA_LOV;
1579 LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1580 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1582 if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1583 CERROR("file join is not supported anymore.\n");
1584 GOTO(out, result = err_serious(-EOPNOTSUPP));
1586 msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1588 if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1589 info->mti_spec.u.sp_ea.eadata == NULL)
1590 GOTO(out, result = err_serious(-EINVAL));
1592 CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1593 "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1594 PFID(rr->rr_fid1), rr->rr_name,
1595 PFID(rr->rr_fid2), create_flags,
1596 ma->ma_attr.la_mode, msg_flags);
1597 if (info->mti_cross_ref) {
1598 /* This is cross-ref open */
1599 mdt_set_disposition(info, ldlm_rep,
1600 (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1602 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1603 ldlm_rep, create_flags);
1605 } else if (req_is_replay(req) ||
1606 (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1607 /* This is a replay request or from liblustre with ea. */
1608 result = mdt_open_by_fid(info, ldlm_rep);
1610 if (result != -ENOENT) {
1611 if (req->rq_export->exp_libclient &&
1612 create_flags & MDS_OPEN_HAS_EA)
1613 GOTO(out, result = 0);
1616 /* We didn't find the correct object, so we need to re-create it
1617 * via a regular replay. */
1618 if (!(create_flags & MDS_OPEN_CREAT)) {
1619 DEBUG_REQ(D_ERROR, req,
1620 "OPEN & CREAT not in open replay/by_fid.");
1621 GOTO(out, result = -EFAULT);
1623 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1624 } else if ((rr->rr_namelen == 0 && create_flags & MDS_OPEN_LOCK) ||
1625 (create_flags & MDS_OPEN_BY_FID)) {
1626 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1627 /* If result is 0 then open by FID has found the file
1628 * and there is nothing left for us to do here. More
1629 * generally if it is anything other than -ENOENT or
1630 * -EREMOTE then we return that now. If -ENOENT and
1631 * MDS_OPEN_CREAT is set then we must create the file
1632 * below. If -EREMOTE then we need to return a LOOKUP
1633 * lock to the client, which we do below. Hence this
1634 * odd looking condition. See LU-2523. */
1635 if (!(result == -ENOENT && (create_flags & MDS_OPEN_CREAT)) &&
1639 if (unlikely(rr->rr_namelen == 0))
1640 GOTO(out, result = -EINVAL);
1642 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1645 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1646 GOTO(out, result = err_serious(-ENOMEM));
1648 mdt_set_disposition(info, ldlm_rep,
1649 (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1651 lh = &info->mti_lh[MDT_LH_PARENT];
1652 mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1653 LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1655 parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1656 MDS_INODELOCK_UPDATE);
1658 GOTO(out, result = PTR_ERR(parent));
1660 /* get and check version of parent */
1661 result = mdt_version_get_check(info, parent, 0);
1663 GOTO(out_parent, result);
1665 fid_zero(child_fid);
1667 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1668 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1669 lname, child_fid, &info->mti_spec);
1670 LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1671 "looking for "DFID"/%s, result fid="DFID"\n",
1672 PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1674 if (result != 0 && result != -ENOENT && result != -ESTALE)
1675 GOTO(out_parent, result);
1677 if (result == -ENOENT || result == -ESTALE) {
1678 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1679 if (result == -ESTALE) {
1681 * -ESTALE means the parent is a dead(unlinked) dir, so
1682 * it should return -ENOENT to in accordance with the
1683 * original mds implementaion.
1685 GOTO(out_parent, result = -ENOENT);
1687 if (!(create_flags & MDS_OPEN_CREAT))
1688 GOTO(out_parent, result);
1689 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1690 GOTO(out_parent, result = -EROFS);
1691 *child_fid = *info->mti_rr.rr_fid2;
1692 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1694 /* In the function below, .hs_keycmp resolves to
1695 * lu_obj_hop_keycmp() */
1696 /* coverity[overrun-buffer-val] */
1697 child = mdt_object_new(info->mti_env, mdt, child_fid);
1700 * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1701 * return FID back in that case.
1703 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1704 child = mdt_object_find(info->mti_env, mdt, child_fid);
1707 GOTO(out_parent, result = PTR_ERR(child));
1709 /** check version of child */
1710 rc = mdt_version_get_check(info, child, 1);
1712 GOTO(out_child, result = rc);
1714 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1715 if (result == -ENOENT) {
1716 /* Create under OBF and .lustre is not permitted */
1717 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
1718 GOTO(out_child, result = -EPERM);
1720 /* save versions in reply */
1721 mdt_version_get_save(info, parent, 0);
1722 mdt_version_get_save(info, child, 1);
1724 /* version of child will be changed */
1725 info->mti_mos = child;
1727 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1728 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1730 /* Let lower layers know what is lock mode on directory. */
1731 info->mti_spec.sp_cr_mode =
1732 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1735 * Do not perform lookup sanity check. We know that name does
1738 info->mti_spec.sp_cr_lookup = 0;
1739 info->mti_spec.sp_feat = &dt_directory_features;
1741 result = mdo_create(info->mti_env,
1742 mdt_object_child(parent),
1744 mdt_object_child(child),
1747 if (result == -ERESTART) {
1748 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1749 GOTO(out_child, result);
1752 /* XXX: we should call this once, see few lines below */
1754 result = mdt_attr_get_complex(info, child, ma);
1757 GOTO(out_child, result);
1762 * The object is on remote node, return its FID for remote open.
1764 if (mdt_object_remote(child)) {
1766 * Check if this lock already was sent to client and
1767 * this is resent case. For resent case do not take lock
1768 * again, use what is already granted.
1770 LASSERT(lhc != NULL);
1772 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1773 struct ldlm_lock *lock;
1775 LASSERT(msg_flags & MSG_RESENT);
1777 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1779 CERROR("Invalid lock handle "LPX64"\n",
1780 lhc->mlh_reg_lh.cookie);
1783 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1784 &lock->l_resource->lr_name));
1785 LDLM_LOCK_PUT(lock);
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->fid1 = *mdt_object_fid(child);
1796 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1801 GOTO(out_child, result);
1803 if (mdt_object_exists(child)) {
1804 /* We have to get attr & LOV EA & HSM for this
1806 ma->ma_need |= MA_HSM;
1807 result = mdt_attr_get_complex(info, child, ma);
1809 /*object non-exist!!!*/
1815 LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1817 /* get openlock if this is not replay and if a client requested it */
1818 if (!req_is_replay(req)) {
1819 rc = mdt_object_open_lock(info, child, lhc, &ibits);
1821 GOTO(out_child_unlock, result = rc);
1822 else if (create_flags & MDS_OPEN_LOCK)
1823 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1826 /* Try to open it now. */
1827 rc = mdt_finish_open(info, parent, child, create_flags,
1831 /* openlock will be released if mdt_finish_open failed */
1832 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1834 if (created && create_flags & MDS_OPEN_VOLATILE) {
1835 CERROR("%s: cannot open volatile file "DFID", orphan "
1836 "file will be left in PENDING directory until "
1837 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1838 PFID(mdt_object_fid(child)), rc);
1839 GOTO(out_child_unlock, result);
1845 ma->ma_cookie_size = 0;
1846 rc = mdo_unlink(info->mti_env,
1847 mdt_object_child(parent),
1848 mdt_object_child(child),
1850 &info->mti_attr, 0);
1852 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1853 mdt_obd_name(info->mti_mdt),
1854 PFID(mdt_object_fid(child)), rc);
1855 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1860 mdt_object_open_unlock(info, child, lhc, ibits, result);
1862 mdt_object_put(info->mti_env, child);
1864 mdt_object_unlock_put(info, parent, lh, result || !created);
1867 lustre_msg_set_transno(req->rq_repmsg, 0);
1872 * Create an orphan object use local root.
1874 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1875 struct mdt_device *mdt,
1876 const struct lu_fid *fid,
1877 struct md_attr *attr, fmode_t fmode)
1879 const struct lu_env *env = info->mti_env;
1880 struct md_op_spec *spec = &info->mti_spec;
1881 struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1882 struct mdt_object *obj = NULL;
1883 struct mdt_object *local_root;
1884 static const char name[] = "i_am_nobody";
1885 struct lu_name *lname;
1886 struct lu_ucred *uc;
1887 cfs_cap_t uc_cap_save;
1891 rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1893 RETURN(ERR_PTR(rc));
1895 local_root = mdt_object_find(env, mdt, local_root_fid);
1896 if (IS_ERR(local_root))
1899 obj = mdt_object_new(env, mdt, fid);
1901 GOTO(out, rc = PTR_ERR(obj));
1903 spec->sp_cr_lookup = 0;
1904 spec->sp_feat = &dt_directory_features;
1905 spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1906 spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1907 if (attr->ma_valid & MA_LOV) {
1908 spec->u.sp_ea.eadata = attr->ma_lmm;
1909 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1910 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1912 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1915 lname = mdt_name(env, (char *)name, sizeof(name) - 1);
1918 uc_cap_save = uc->uc_cap;
1919 uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1920 rc = mdo_create(env, mdt_object_child(local_root), lname,
1921 mdt_object_child(obj), spec, attr);
1922 uc->uc_cap = uc_cap_save;
1924 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1925 mdt_obd_name(mdt), PFID(fid), rc);
1929 rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1931 CERROR("%s: cannot open volatile file "DFID", orphan "
1932 "file will be left in PENDING directory until "
1933 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1940 mdt_object_put(env, obj);
1943 mdt_object_put(env, local_root);
1947 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1950 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1951 struct close_data *data;
1952 struct ldlm_lock *lease;
1953 struct mdt_object *orphan;
1954 struct md_attr *orp_ma;
1961 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1964 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1968 lease = ldlm_handle2lock(&data->cd_handle);
1972 /* try to hold open_sem so that nobody else can open the file */
1973 if (!down_write_trylock(&o->mot_open_sem)) {
1974 ldlm_lock_cancel(lease);
1975 GOTO(out_reprocess, rc = -EBUSY);
1978 /* Check if the lease open lease has already canceled */
1979 lock_res_and_lock(lease);
1980 lease_broken = ldlm_is_cancel(lease);
1981 unlock_res_and_lock(lease);
1983 LDLM_DEBUG(lease, DFID " lease broken? %d\n",
1984 PFID(mdt_object_fid(o)), lease_broken);
1986 /* Cancel server side lease. Client side counterpart should
1987 * have been cancelled. It's okay to cancel it now as we've
1988 * held mot_open_sem. */
1989 ldlm_lock_cancel(lease);
1991 if (lease_broken) /* don't perform release task */
1992 GOTO(out_unlock, rc = -ESTALE);
1994 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1995 GOTO(out_unlock, rc = -EINVAL);
1997 /* ma_need was set before but it seems fine to change it in order to
1998 * avoid modifying the one from RPC */
1999 ma->ma_need = MA_HSM;
2000 rc = mdt_attr_get_complex(info, o, ma);
2002 GOTO(out_unlock, rc);
2004 if (!mdt_hsm_release_allow(ma))
2005 GOTO(out_unlock, rc = -EPERM);
2007 /* already released? */
2008 if (ma->ma_hsm.mh_flags & HS_RELEASED)
2009 GOTO(out_unlock, rc = 0);
2011 /* Compare on-disk and packed data_version */
2012 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
2013 CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
2014 " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
2015 data->cd_data_version, ma->ma_hsm.mh_arch_ver);
2016 GOTO(out_unlock, rc = -EPERM);
2019 ma->ma_valid = MA_INODE;
2020 ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
2021 rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
2023 GOTO(out_unlock, rc);
2025 ma->ma_need = MA_INODE | MA_LOV;
2026 rc = mdt_attr_get_complex(info, o, ma);
2028 GOTO(out_unlock, rc);
2030 if (!(ma->ma_valid & MA_LOV)) {
2031 /* Even empty file are released */
2032 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2033 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2034 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2035 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2036 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2038 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
2039 * ma_lmm as lov_user_md, then it will be confused by union of
2040 * layout_gen and stripe_offset. */
2041 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
2042 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2043 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
2044 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
2046 GOTO(out_unlock, rc = -EINVAL);
2049 /* Set file as released */
2050 ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
2052 /* Hopefully it's not used in this call path */
2053 orp_ma = &info->mti_u.som.attr;
2054 orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2055 orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
2056 orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
2057 orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2058 orp_ma->ma_lmm = ma->ma_lmm;
2059 orp_ma->ma_lmm_size = ma->ma_lmm_size;
2060 orp_ma->ma_valid = MA_INODE | MA_LOV;
2061 orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2063 if (IS_ERR(orphan)) {
2064 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2065 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2067 GOTO(out_unlock, rc = PTR_ERR(orphan));
2070 /* Set up HSM attribute for orphan object */
2071 CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
2072 buf = &info->mti_buf;
2073 buf->lb_buf = info->mti_xattr_buf;
2074 buf->lb_len = sizeof(struct hsm_attrs);
2075 ma->ma_hsm.mh_flags |= HS_RELEASED;
2076 lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2077 ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2079 mdt_lock_reg_init(lh, LCK_EX);
2080 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2081 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
2083 GOTO(out_close, rc);
2085 rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2089 /* Swap layout with orphan object */
2090 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2091 mdt_object_child(orphan),
2092 SWAP_LAYOUTS_MDS_HSM);
2094 /* Release exclusive LL */
2095 mdt_object_unlock(info, o, lh, 1);
2100 /* Close orphan object anyway */
2101 rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2104 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2105 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2106 LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2108 mdt_object_put(info->mti_env, orphan);
2111 up_write(&o->mot_open_sem);
2113 if (rc == 0) { /* already released */
2114 struct mdt_body *repbody;
2115 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2116 LASSERT(repbody != NULL);
2117 repbody->valid |= OBD_MD_FLRELEASED;
2121 ldlm_reprocess_all(lease->l_resource);
2122 LDLM_LOCK_PUT(lease);
2130 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
2131 MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
2133 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2135 return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
2138 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2140 struct mdt_object *o = mfd->mfd_object;
2141 struct md_object *next = mdt_object_child(o);
2142 struct md_attr *ma = &info->mti_attr;
2143 int ret = MDT_IOEPOCH_CLOSED;
2148 mode = mfd->mfd_mode;
2150 if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
2151 rc = mdt_hsm_release(info, o, ma);
2153 CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2154 mdt_obd_name(info->mti_mdt),
2155 PFID(mdt_object_fid(o)), rc);
2156 /* continue to close even error occurred. */
2160 if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
2162 ret = mdt_ioepoch_close(info, o);
2163 } else if (mode & MDS_FMODE_EXEC) {
2165 } else if (mode & MDS_FMODE_EPOCH) {
2166 ret = mdt_ioepoch_close(info, o);
2167 } else if (mode & MDS_FMODE_SOM) {
2168 ret = mdt_som_au_close(info, o);
2171 /* Update atime on close only. */
2172 if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
2173 && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
2174 /* Set the atime only. */
2175 ma->ma_valid = MA_INODE;
2176 ma->ma_attr.la_valid = LA_ATIME;
2177 rc = mo_attr_set(info->mti_env, next, ma);
2180 /* If file data is modified, add the dirty flag. */
2181 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2182 rc = mdt_add_dirty_flag(info, o, ma);
2184 ma->ma_need |= MA_INODE;
2185 ma->ma_valid &= ~MA_INODE;
2187 if (!MFD_CLOSED(mode))
2188 rc = mo_close(info->mti_env, next, ma, mode);
2190 if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
2191 struct mdt_export_data *med;
2193 /* The IOepoch is still opened or SOM update is needed.
2194 * Put mfd back into the list. */
2195 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
2196 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
2197 MDS_FMODE_EPOCH : MDS_FMODE_SOM);
2199 LASSERT(mdt_info_req(info));
2200 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
2201 spin_lock(&med->med_open_lock);
2202 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
2203 class_handle_hash_back(&mfd->mfd_handle);
2204 spin_unlock(&med->med_open_lock);
2206 if (ret == MDT_IOEPOCH_OPENED) {
2210 CDEBUG(D_INODE, "Size-on-MDS attribute update is "
2211 "needed on "DFID"\n", PFID(mdt_object_fid(o)));
2214 /* adjust open and lease count */
2215 if (mode & MDS_OPEN_LEASE) {
2216 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2217 atomic_dec(&o->mot_lease_count);
2219 LASSERT(atomic_read(&o->mot_open_count) > 0);
2220 atomic_dec(&o->mot_open_count);
2223 mdt_object_put(info->mti_env, o);
2226 RETURN(rc ? rc : ret);
2229 int mdt_close(struct tgt_session_info *tsi)
2231 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2232 struct ptlrpc_request *req = tgt_ses_req(tsi);
2233 struct mdt_export_data *med;
2234 struct mdt_file_data *mfd;
2235 struct mdt_object *o;
2236 struct md_attr *ma = &info->mti_attr;
2237 struct mdt_body *repbody = NULL;
2241 mdt_counter_incr(req, LPROC_MDT_CLOSE);
2242 /* Close may come with the Size-on-MDS update. Unpack it. */
2243 rc = mdt_close_unpack(info);
2245 GOTO(out, rc = err_serious(rc));
2247 LASSERT(info->mti_ioepoch);
2249 req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2250 info->mti_mdt->mdt_max_mdsize);
2251 req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2252 info->mti_mdt->mdt_max_cookiesize);
2253 rc = req_capsule_server_pack(info->mti_pill);
2254 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2255 mdt_client_compatibility(info);
2257 mdt_fix_reply(info);
2258 mdt_exit_ucred(info);
2259 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2262 /* Continue to close handle even if we can not pack reply */
2264 repbody = req_capsule_server_get(info->mti_pill,
2266 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2268 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2271 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
2273 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
2276 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
2277 repbody->eadatasize = 0;
2278 repbody->aclsize = 0;
2280 rc = err_serious(rc);
2283 med = &req->rq_export->exp_mdt_data;
2284 spin_lock(&med->med_open_lock);
2285 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2286 req_is_replay(req));
2287 if (mdt_mfd_closed(mfd)) {
2288 spin_unlock(&med->med_open_lock);
2289 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2290 ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
2291 info->mti_ioepoch->handle.cookie);
2292 /** not serious error since bug 3633 */
2295 class_handle_unhash(&mfd->mfd_handle);
2296 cfs_list_del_init(&mfd->mfd_list);
2297 spin_unlock(&med->med_open_lock);
2299 /* Do not lose object before last unlink. */
2300 o = mfd->mfd_object;
2301 mdt_object_get(info->mti_env, o);
2302 ret = mdt_mfd_close(info, mfd);
2303 if (repbody != NULL)
2304 rc = mdt_handle_last_unlink(info, o, ma);
2305 mdt_empty_transno(info, rc);
2306 mdt_object_put(info->mti_env, o);
2308 if (repbody != NULL) {
2309 mdt_client_compatibility(info);
2310 rc = mdt_fix_reply(info);
2313 mdt_exit_ucred(info);
2314 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2315 GOTO(out, rc = err_serious(-ENOMEM));
2317 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2318 OBD_FAIL_MDS_CLOSE_NET_REP))
2319 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2321 mdt_thread_info_fini(info);
2322 RETURN(rc ? rc : ret);
2326 * DONE_WRITING rpc handler.
2328 * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
2329 * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
2330 * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
2331 * skip attributes and reconstruct the reply here.
2333 int mdt_done_writing(struct tgt_session_info *tsi)
2335 struct ptlrpc_request *req = tgt_ses_req(tsi);
2336 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2337 struct mdt_body *repbody = NULL;
2338 struct mdt_export_data *med;
2339 struct mdt_file_data *mfd;
2343 rc = req_capsule_server_pack(tsi->tsi_pill);
2345 GOTO(out, rc = err_serious(rc));
2347 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2348 repbody->eadatasize = 0;
2349 repbody->aclsize = 0;
2351 /* Done Writing may come with the Size-on-MDS update. Unpack it. */
2352 rc = mdt_close_unpack(info);
2354 GOTO(out, rc = err_serious(rc));
2356 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2357 mdt_exit_ucred(info);
2358 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2361 med = &info->mti_exp->exp_mdt_data;
2362 spin_lock(&med->med_open_lock);
2363 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2364 req_is_replay(req));
2366 spin_unlock(&med->med_open_lock);
2367 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
2368 ": cookie = "LPX64" ioepoch = "LPU64"\n",
2369 PFID(info->mti_rr.rr_fid1),
2370 info->mti_ioepoch->handle.cookie,
2371 info->mti_ioepoch->ioepoch);
2372 /* If this is a replay, reconstruct the transno. */
2373 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2374 rc = info->mti_ioepoch->flags & MF_SOM_AU ?
2376 mdt_empty_transno(info, rc);
2379 GOTO(error_ucred, rc);
2382 LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
2383 mfd->mfd_mode == MDS_FMODE_TRUNC);
2384 class_handle_unhash(&mfd->mfd_handle);
2385 cfs_list_del_init(&mfd->mfd_list);
2386 spin_unlock(&med->med_open_lock);
2388 /* Set EPOCH CLOSE flag if not set by client. */
2389 info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
2390 info->mti_attr.ma_valid = 0;
2392 info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
2393 OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2394 if (info->mti_attr.ma_lmm == NULL)
2395 GOTO(error_ucred, rc = -ENOMEM);
2397 rc = mdt_mfd_close(info, mfd);
2399 OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2400 mdt_empty_transno(info, rc);
2402 mdt_exit_ucred(info);
2404 mdt_thread_info_fini(info);