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, 2012, Whamcloud, Inc.
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(void)
63 struct mdt_file_data *mfd;
68 CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
69 CFS_INIT_LIST_HEAD(&mfd->mfd_list);
70 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
76 * Find the mfd pointed to by handle in global hash table.
77 * In case of replay the handle is obsoleted
78 * but mfd can be found in mfd list by that handle
80 struct mdt_file_data *mdt_handle2mfd(struct mdt_thread_info *info,
81 const struct lustre_handle *handle)
83 struct ptlrpc_request *req = mdt_info_req(info);
84 struct mdt_file_data *mfd;
87 LASSERT(handle != NULL);
88 mfd = class_handle2object(handle->cookie);
89 /* during dw/setattr replay the mfd can be found by old handle */
90 if (mfd == NULL && req_is_replay(req)) {
91 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
92 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
93 if (mfd->mfd_old_handle.cookie == handle->cookie)
102 void mdt_mfd_free(struct mdt_file_data *mfd)
104 LASSERT(cfs_list_empty(&mfd->mfd_list));
105 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
108 static int mdt_create_data(struct mdt_thread_info *info,
109 struct mdt_object *p, struct mdt_object *o)
111 struct md_op_spec *spec = &info->mti_spec;
112 struct md_attr *ma = &info->mti_attr;
116 if (!md_should_create(spec->sp_cr_flags))
119 ma->ma_need = MA_INODE | MA_LOV;
121 cfs_mutex_lock(&o->mot_lov_mutex);
122 if (!(o->mot_flags & MOF_LOV_CREATED)) {
123 rc = mdo_create_data(info->mti_env,
124 p ? mdt_object_child(p) : NULL,
125 mdt_object_child(o), spec, ma);
127 rc = mdt_attr_get_complex(info, o, ma);
129 if (rc == 0 && ma->ma_valid & MA_LOV)
130 o->mot_flags |= MOF_LOV_CREATED;
132 cfs_mutex_unlock(&o->mot_lov_mutex);
136 static int mdt_ioepoch_opened(struct mdt_object *mo)
138 return mo->mot_ioepoch_count;
141 int mdt_object_is_som_enabled(struct mdt_object *mo)
143 return !mo->mot_ioepoch;
147 * Re-enable Size-on-MDS.
148 * Call under ->mot_ioepoch_mutex.
150 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
152 if (ioepoch == mo->mot_ioepoch) {
153 LASSERT(!mdt_ioepoch_opened(mo));
160 * Open the IOEpoch. It is allowed if @writecount is not negative.
161 * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
163 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
166 struct mdt_device *mdt = info->mti_mdt;
171 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
172 !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
175 cfs_mutex_lock(&o->mot_ioepoch_mutex);
176 if (mdt_ioepoch_opened(o)) {
177 /* Epoch continues even if there is no writers yet. */
178 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
179 o->mot_ioepoch, PFID(mdt_object_fid(o)));
181 /* XXX: ->mdt_ioepoch is not initialized at the mount */
182 cfs_spin_lock(&mdt->mdt_ioepoch_lock);
183 if (mdt->mdt_ioepoch < info->mti_replayepoch)
184 mdt->mdt_ioepoch = info->mti_replayepoch;
186 if (info->mti_replayepoch)
187 o->mot_ioepoch = info->mti_replayepoch;
188 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
189 o->mot_ioepoch = ++mdt->mdt_ioepoch;
191 o->mot_ioepoch = mdt->mdt_ioepoch;
193 cfs_spin_unlock(&mdt->mdt_ioepoch_lock);
195 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
196 o->mot_ioepoch, PFID(mdt_object_fid(o)));
198 o->mot_flags |= MOF_SOM_CREATED;
201 o->mot_ioepoch_count++;
202 cfs_mutex_unlock(&o->mot_ioepoch_mutex);
204 /* Cancel Size-on-MDS attributes cached on clients for the open case.
205 * In the truncate case, see mdt_reint_setattr(). */
206 if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
207 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
208 mdt_lock_reg_init(lh, LCK_EX);
209 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
212 mdt_object_unlock(info, o, lh, 1);
218 * Update SOM on-disk attributes.
219 * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
220 * mountid and attributes. If disabling, zero IOEpoch id in lustre-ea.
221 * Call under ->mot_ioepoch_mutex.
223 static int mdt_som_attr_set(struct mdt_thread_info *info,
224 struct mdt_object *obj, __u64 ioepoch, int enable)
226 struct lustre_mdt_attrs *lma;
227 struct md_attr *ma = &info->mti_attr;
228 struct lu_buf *buf = &info->mti_buf;
229 struct md_object *next = mdt_object_child(obj);
230 struct mdt_device *mdt = info->mti_mdt;
231 struct lu_attr *la = &ma->ma_attr;
235 CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
236 " on "DFID".\n", enable ? "update" : "disabling",
237 ioepoch, PFID(mdt_object_fid(obj)));
239 lma = (struct lustre_mdt_attrs *) info->mti_xattr_buf;
240 CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*lma));
243 buf->lb_len = sizeof(info->mti_xattr_buf);
244 rc = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_LMA);
246 lustre_lma_swab(lma);
247 } else if (rc == -ENODATA) {
248 memset(lma, 0, sizeof(*lma));
254 memcpy(&lma->lma_self_fid, mdt_object_fid(obj), sizeof(lma->lma_self_fid));
257 lma->lma_ioepoch = ioepoch;
258 lma->lma_som_size = la->la_valid & LA_SIZE ? la->la_size : 0;
259 lma->lma_som_blocks = la->la_valid & LA_BLOCKS ? la->la_blocks : 0;
260 lma->lma_som_mountid = mdt->mdt_lut.lut_obd->u.obt.obt_mount_count;
262 lma->lma_compat |= LMAC_SOM;
264 lma->lma_compat &= ~LMAC_SOM;
266 rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_LMA, 0);
271 /** Perform the eviction specific actions on ioepoch close. */
272 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
273 struct mdt_object *o)
277 cfs_mutex_lock(&o->mot_ioepoch_mutex);
278 CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
279 "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
280 o->mot_ioepoch_count);
281 o->mot_ioepoch_count--;
283 /* If eviction occured set MOF_SOM_RECOV,
284 * if no other epoch holders, disable SOM on disk. */
285 o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
286 if (!mdt_ioepoch_opened(o)) {
287 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
288 mdt_object_som_enable(o, o->mot_ioepoch);
290 cfs_mutex_unlock(&o->mot_ioepoch_mutex);
295 * Perform the replay specific actions on ioepoch close.
296 * Skip SOM attribute update if obtained and just forget about the inode state
297 * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
299 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
300 struct mdt_object *o)
302 int rc = MDT_IOEPOCH_CLOSED;
305 cfs_mutex_lock(&o->mot_ioepoch_mutex);
306 CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
307 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
308 o->mot_ioepoch_count--;
310 /* Get an info from the replayed request if client is supposed
311 * to send an Attibute Update, reconstruct @rc if so */
312 if (info->mti_ioepoch->flags & MF_SOM_AU)
313 rc = MDT_IOEPOCH_GETATTR;
315 if (!mdt_ioepoch_opened(o))
316 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
317 cfs_mutex_unlock(&o->mot_ioepoch_mutex);
323 * Regular file IOepoch close.
324 * Closes the ioepoch, checks the object state, apply obtained attributes and
325 * re-enable SOM on the object, if possible. Also checks if the recovery is
326 * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
327 * to obtain SOM attributes under the server-side OST locks.
330 * MDT_IOEPOCH_CLOSED if ioepoch is closed.
331 * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
333 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
334 struct mdt_object *o)
336 struct md_attr *tmp_ma;
340 int rc = 0, ret = MDT_IOEPOCH_CLOSED;
343 la = &info->mti_attr.ma_attr;
344 achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
346 cfs_mutex_lock(&o->mot_ioepoch_mutex);
347 o->mot_ioepoch_count--;
349 tmp_ma = &info->mti_u.som.attr;
350 tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
351 tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
352 tmp_ma->ma_som = &info->mti_u.som.data;
353 tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
354 tmp_ma->ma_valid = 0;
355 rc = mdt_attr_get_complex(info, o, tmp_ma);
359 /* Check the on-disk SOM state. */
360 if (o->mot_flags & MOF_SOM_RECOV)
362 else if (!(o->mot_flags & MOF_SOM_CREATED) &&
363 !(tmp_ma->ma_valid & MA_SOM))
366 CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
367 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
369 opened = mdt_ioepoch_opened(o);
371 * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
372 * Skip the check for file with no LOV or for unlink files.
374 if (!opened && tmp_ma->ma_valid & MA_LOV &&
375 !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
377 /* If some previous writer was evicted, re-ask the
378 * client for attributes. Even if attributes are
379 * provided, we cannot believe in them.
380 * Another use case is that there is no SOM cache on
381 * disk -- first access with SOM or there was an MDS
383 ret = MDT_IOEPOCH_GETATTR;
384 else if (o->mot_flags & MOF_SOM_CHANGE)
385 /* Some previous writer changed the attribute.
386 * Do not believe to the current Size-on-MDS
387 * update, re-ask client. */
388 ret = MDT_IOEPOCH_GETATTR;
389 else if (!(la->la_valid & LA_SIZE) && achange)
390 /* Attributes were changed by the last writer
391 * only but no Size-on-MDS update is received.*/
392 ret = MDT_IOEPOCH_GETATTR;
395 if (achange || ret == MDT_IOEPOCH_GETATTR)
396 o->mot_flags |= MOF_SOM_CHANGE;
398 /* If epoch ends and relable SOM attributes are obtained, update them.
399 * Create SOM ea for new files even if there is no attributes obtained
400 * (0-length file). */
401 if (ret == MDT_IOEPOCH_CLOSED && !opened) {
402 if (achange || o->mot_flags & MOF_SOM_CREATED) {
403 LASSERT(achange || !(la->la_valid & LA_SIZE));
404 rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
406 /* Avoid the following setattrs of these attributes,
407 * e.g. for atime update. */
408 info->mti_attr.ma_valid = 0;
410 mdt_object_som_enable(o, o->mot_ioepoch);
413 cfs_mutex_unlock(&o->mot_ioepoch_mutex);
414 /* If recovery is needed, tell the client to perform GETATTR under
416 if (ret == MDT_IOEPOCH_GETATTR && recovery) {
417 struct mdt_body *rep;
418 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
419 rep->valid |= OBD_MD_FLGETATTRLOCK;
425 cfs_mutex_unlock(&o->mot_ioepoch_mutex);
430 * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
431 * - a client closes the IOEpoch;
432 * - a client eviction occured.
434 * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
435 * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
436 * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
439 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
441 struct ptlrpc_request *req = mdt_info_req(info);
444 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
445 !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
448 LASSERT(o->mot_ioepoch_count);
449 LASSERT(info->mti_ioepoch == NULL ||
450 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
452 /* IOEpoch is closed only if client tells about it or eviction occures.
453 * In the replay case, always close the epoch. */
455 RETURN(mdt_ioepoch_close_on_eviction(info, o));
456 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
457 RETURN(mdt_ioepoch_close_on_replay(info, o));
458 if (info->mti_ioepoch->flags & MF_EPOCH_CLOSE)
459 RETURN(mdt_ioepoch_close_reg(info, o));
460 /* IO epoch is not closed. */
461 RETURN(MDT_IOEPOCH_OPENED);
465 * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
466 * for attribute update. It happens if:
467 * - SOM Attribute Update is obtained;
468 * - the client failed to obtain it and informs MDS about it;
469 * - a client eviction occured.
470 * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
473 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
475 struct ptlrpc_request *req = mdt_info_req(info);
477 int act = MDT_SOM_ENABLE;
481 LASSERT(!req || info->mti_ioepoch);
482 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
483 !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
486 /* No size whereas MF_SOM_CHANGE is set means client failed to
487 * obtain ost attributes, drop the SOM cache on disk if so. */
489 (info->mti_ioepoch &&
490 info->mti_ioepoch->flags & MF_SOM_CHANGE &&
491 !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
492 act = MDT_SOM_DISABLE;
494 cfs_mutex_lock(&o->mot_ioepoch_mutex);
495 /* Mark the object it is the recovery state if we failed to obtain
497 if (act == MDT_SOM_DISABLE)
498 o->mot_flags |= MOF_SOM_RECOV;
500 if (!mdt_ioepoch_opened(o)) {
501 ioepoch = info->mti_ioepoch ?
502 info->mti_ioepoch->ioepoch : o->mot_ioepoch;
504 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
505 rc = mdt_som_attr_set(info, o, ioepoch, act);
506 mdt_object_som_enable(o, ioepoch);
508 cfs_mutex_unlock(&o->mot_ioepoch_mutex);
512 int mdt_write_read(struct mdt_object *o)
516 cfs_mutex_lock(&o->mot_ioepoch_mutex);
517 rc = o->mot_writecount;
518 cfs_mutex_unlock(&o->mot_ioepoch_mutex);
522 int mdt_write_get(struct mdt_object *o)
526 cfs_mutex_lock(&o->mot_ioepoch_mutex);
527 if (o->mot_writecount < 0)
531 cfs_mutex_unlock(&o->mot_ioepoch_mutex);
535 void mdt_write_put(struct mdt_object *o)
538 cfs_mutex_lock(&o->mot_ioepoch_mutex);
540 cfs_mutex_unlock(&o->mot_ioepoch_mutex);
544 static int mdt_write_deny(struct mdt_object *o)
548 cfs_mutex_lock(&o->mot_ioepoch_mutex);
549 if (o->mot_writecount > 0)
553 cfs_mutex_unlock(&o->mot_ioepoch_mutex);
557 static void mdt_write_allow(struct mdt_object *o)
560 cfs_mutex_lock(&o->mot_ioepoch_mutex);
562 cfs_mutex_unlock(&o->mot_ioepoch_mutex);
566 /* there can be no real transaction so prepare the fake one */
567 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
569 struct mdt_device *mdt = info->mti_mdt;
570 struct ptlrpc_request *req = mdt_info_req(info);
571 struct tg_export_data *ted;
572 struct lsd_client_data *lcd;
575 /* transaction has occurred already */
576 if (lustre_msg_get_transno(req->rq_repmsg) != 0)
579 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
581 if (info->mti_transno != 0) {
582 struct obd_export *exp = req->rq_export;
584 CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
585 mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name,
587 libcfs_nid2str(exp->exp_connection->c_peer.nid),
591 } else if (info->mti_transno == 0) {
592 info->mti_transno = ++ mdt->mdt_lut.lut_last_transno;
594 /* should be replay */
595 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
596 mdt->mdt_lut.lut_last_transno = info->mti_transno;
598 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
600 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
602 req->rq_export->exp_obd->obd_last_committed);
604 req->rq_transno = info->mti_transno;
605 lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
607 /* update lcd in memory only for resent cases */
608 ted = &req->rq_export->exp_target_data;
610 cfs_mutex_lock(&ted->ted_lcd_lock);
612 if (info->mti_transno < lcd->lcd_last_transno &&
613 info->mti_transno != 0) {
614 /* This should happen during replay. Do not update
615 * last rcvd info if replay req transno < last transno,
616 * otherwise the following resend(after replay) can not
617 * be checked correctly by xid */
618 cfs_mutex_unlock(&ted->ted_lcd_lock);
619 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
620 mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name,
621 info->mti_transno, lcd->lcd_last_transno);
625 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
626 lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
627 if (info->mti_transno != 0)
628 lcd->lcd_last_close_transno = info->mti_transno;
629 lcd->lcd_last_close_xid = req->rq_xid;
630 lcd->lcd_last_close_result = rc;
632 /* VBR: save versions in last_rcvd for reconstruct. */
633 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
635 lcd->lcd_pre_versions[0] = pre_versions[0];
636 lcd->lcd_pre_versions[1] = pre_versions[1];
637 lcd->lcd_pre_versions[2] = pre_versions[2];
638 lcd->lcd_pre_versions[3] = pre_versions[3];
640 if (info->mti_transno != 0)
641 lcd->lcd_last_transno = info->mti_transno;
643 lcd->lcd_last_xid = req->rq_xid;
644 lcd->lcd_last_result = rc;
645 lcd->lcd_last_data = info->mti_opdata;
647 cfs_mutex_unlock(&ted->ted_lcd_lock);
652 void mdt_mfd_set_mode(struct mdt_file_data *mfd, int mode)
654 LASSERT(mfd != NULL);
656 CDEBUG(D_HA, "Change mfd %p mode 0x%x->0x%x\n",
657 mfd, (unsigned int)mfd->mfd_mode, (unsigned int)mode);
659 mfd->mfd_mode = mode;
662 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
663 struct mdt_object *o, __u64 flags, int created)
665 struct ptlrpc_request *req = mdt_info_req(info);
666 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
667 struct mdt_file_data *mfd;
668 struct md_attr *ma = &info->mti_attr;
669 struct lu_attr *la = &ma->ma_attr;
670 struct mdt_body *repbody;
671 int rc = 0, isdir, isreg;
674 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
676 isreg = S_ISREG(la->la_mode);
677 isdir = S_ISDIR(la->la_mode);
678 if (isreg && !(ma->ma_valid & MA_LOV)) {
680 * No EA, check whether it is will set regEA and dirEA since in
681 * above attr get, these size might be zero, so reset it, to
682 * retrieve the MD after create obj.
684 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
687 /* in replay case, p == NULL */
688 rc = mdt_create_data(info, p, o);
693 CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
694 ma->ma_valid, ma->ma_lmm_size);
696 if (ma->ma_valid & MA_LOV) {
697 LASSERT(ma->ma_lmm_size != 0);
698 repbody->eadatasize = ma->ma_lmm_size;
700 repbody->valid |= OBD_MD_FLDIREA;
702 repbody->valid |= OBD_MD_FLEASIZE;
705 if (flags & FMODE_WRITE) {
706 rc = mdt_write_get(o);
708 mdt_ioepoch_open(info, o, created);
709 repbody->ioepoch = o->mot_ioepoch;
711 } else if (flags & MDS_FMODE_EXEC) {
712 rc = mdt_write_deny(o);
717 rc = mo_open(info->mti_env, mdt_object_child(o),
718 created ? flags | MDS_OPEN_CREATED : flags);
725 * Keep a reference on this object for this open, and is
726 * released by mdt_mfd_close().
728 mdt_object_get(info->mti_env, o);
731 * @flags is always not zero. At least it should be FMODE_READ,
732 * FMODE_WRITE or MDS_FMODE_EXEC.
737 mdt_mfd_set_mode(mfd, flags);
740 mfd->mfd_xid = req->rq_xid;
743 if (req_is_replay(req)) {
744 struct mdt_file_data *old_mfd;
745 /* Check wheather old cookie already exist in
746 * the list, becasue when do recovery, client
747 * might be disconnected from server, and
748 * restart replay, so there maybe some orphan
749 * mfd here, we should remove them */
750 LASSERT(info->mti_rr.rr_handle != NULL);
751 old_mfd = mdt_handle2mfd(info, info->mti_rr.rr_handle);
753 CDEBUG(D_HA, "del orph mfd %p fid=("DFID") "
754 "cookie=" LPX64"\n", mfd,
755 PFID(mdt_object_fid(mfd->mfd_object)),
756 info->mti_rr.rr_handle->cookie);
757 cfs_spin_lock(&med->med_open_lock);
758 class_handle_unhash(&old_mfd->mfd_handle);
759 cfs_list_del_init(&old_mfd->mfd_list);
760 cfs_spin_unlock(&med->med_open_lock);
761 /* no attr update for that close */
763 ma->ma_valid |= MA_FLAGS;
764 ma->ma_attr_flags |= MDS_RECOV_OPEN;
765 mdt_mfd_close(info, old_mfd);
766 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
767 ma->ma_valid &= ~MA_FLAGS;
769 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
770 info->mti_rr.rr_handle->cookie);
771 mfd->mfd_old_handle.cookie =
772 info->mti_rr.rr_handle->cookie;
774 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
776 if (req->rq_export->exp_disconnected) {
777 cfs_spin_lock(&med->med_open_lock);
778 class_handle_unhash(&mfd->mfd_handle);
779 cfs_list_del_init(&mfd->mfd_list);
780 cfs_spin_unlock(&med->med_open_lock);
781 mdt_mfd_close(info, mfd);
783 cfs_spin_lock(&med->med_open_lock);
784 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
785 cfs_spin_unlock(&med->med_open_lock);
788 mdt_empty_transno(info, rc);
790 GOTO(err_out, rc = -ENOMEM);
796 if (flags & FMODE_WRITE)
797 /* XXX We also need to close io epoch here.
798 * See LU-1220 - green */
800 else if (flags & FMODE_EXEC)
805 int mdt_finish_open(struct mdt_thread_info *info,
806 struct mdt_object *p, struct mdt_object *o,
807 __u64 flags, int created, struct ldlm_reply *rep)
809 struct ptlrpc_request *req = mdt_info_req(info);
810 struct obd_export *exp = req->rq_export;
811 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
812 struct md_attr *ma = &info->mti_attr;
813 struct lu_attr *la = &ma->ma_attr;
814 struct mdt_file_data *mfd;
815 struct mdt_body *repbody;
817 int isreg, isdir, islnk;
821 LASSERT(ma->ma_valid & MA_INODE);
823 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
825 isreg = S_ISREG(la->la_mode);
826 isdir = S_ISDIR(la->la_mode);
827 islnk = S_ISLNK(la->la_mode);
828 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
830 if (exp_connect_rmtclient(exp)) {
831 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
833 rc = mdt_pack_remote_perm(info, o, buf);
835 repbody->valid &= ~OBD_MD_FLRMTPERM;
836 repbody->aclsize = 0;
838 repbody->valid |= OBD_MD_FLRMTPERM;
839 repbody->aclsize = sizeof(struct mdt_remote_perm);
842 #ifdef CONFIG_FS_POSIX_ACL
843 else if (exp->exp_connect_flags & OBD_CONNECT_ACL) {
844 const struct lu_env *env = info->mti_env;
845 struct md_object *next = mdt_object_child(o);
846 struct lu_buf *buf = &info->mti_buf;
848 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
849 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
851 if (buf->lb_len > 0) {
852 rc = mo_xattr_get(env, next, buf,
853 XATTR_NAME_ACL_ACCESS);
855 if (rc == -ENODATA) {
856 repbody->aclsize = 0;
857 repbody->valid |= OBD_MD_FLACL;
859 } else if (rc == -EOPNOTSUPP) {
862 CERROR("got acl size: %d\n", rc);
865 repbody->aclsize = rc;
866 repbody->valid |= OBD_MD_FLACL;
873 if (info->mti_mdt->mdt_opts.mo_mds_capa &&
874 exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) {
875 struct lustre_capa *capa;
877 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
879 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
880 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
883 repbody->valid |= OBD_MD_FLMDSCAPA;
885 if (info->mti_mdt->mdt_opts.mo_oss_capa &&
886 exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA &&
887 S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) {
888 struct lustre_capa *capa;
890 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
892 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
893 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
896 repbody->valid |= OBD_MD_FLOSSCAPA;
900 * If we are following a symlink, don't open; and do not return open
901 * handle for special nodes as client required.
903 if (islnk || (!isreg && !isdir &&
904 (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH))) {
905 lustre_msg_set_transno(req->rq_repmsg, 0);
909 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
912 * We need to return the existing object's fid back, so it is done here,
913 * after preparing the reply.
915 if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
918 /* This can't be done earlier, we need to return reply body */
920 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
921 /* We are trying to create or write an existing dir. */
924 } else if (flags & MDS_OPEN_DIRECTORY)
927 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
928 OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
933 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
934 cfs_spin_lock(&med->med_open_lock);
935 cfs_list_for_each(t, &med->med_open_head) {
936 mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
937 if (mfd->mfd_xid == req->rq_xid) {
942 cfs_spin_unlock(&med->med_open_lock);
945 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
946 /*set repbody->ea_size for resent case*/
947 if (ma->ma_valid & MA_LOV) {
948 LASSERT(ma->ma_lmm_size != 0);
949 repbody->eadatasize = ma->ma_lmm_size;
951 repbody->valid |= OBD_MD_FLDIREA;
953 repbody->valid |= OBD_MD_FLEASIZE;
959 rc = mdt_mfd_open(info, p, o, flags, created);
963 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
964 struct lsd_client_data *lcd);
966 void mdt_reconstruct_open(struct mdt_thread_info *info,
967 struct mdt_lock_handle *lhc)
969 const struct lu_env *env = info->mti_env;
970 struct mdt_device *mdt = info->mti_mdt;
971 struct req_capsule *pill = info->mti_pill;
972 struct ptlrpc_request *req = mdt_info_req(info);
973 struct tg_export_data *ted = &req->rq_export->exp_target_data;
974 struct lsd_client_data *lcd = ted->ted_lcd;
975 struct md_attr *ma = &info->mti_attr;
976 struct mdt_reint_record *rr = &info->mti_rr;
977 __u32 flags = info->mti_spec.sp_cr_flags;
978 struct ldlm_reply *ldlm_rep;
979 struct mdt_object *parent;
980 struct mdt_object *child;
981 struct mdt_body *repbody;
985 LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
986 ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
987 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
989 ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
990 ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
992 ma->ma_need = MA_INODE;
993 if (ma->ma_lmm_size > 0)
994 ma->ma_need |= MA_LOV;
998 mdt_req_from_lcd(req, lcd);
999 mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1001 CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1002 ldlm_rep->lock_policy_res1, req->rq_status);
1004 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1005 req->rq_status != 0)
1006 /* We did not create successfully, return error to client. */
1007 GOTO(out, rc = req->rq_status);
1009 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1010 struct obd_export *exp = req->rq_export;
1012 * We failed after creation, but we do not know in which step
1013 * we failed. So try to check the child object.
1015 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1016 if (IS_ERR(parent)) {
1017 rc = PTR_ERR(parent);
1018 LCONSOLE_WARN("Parent "DFID" lookup error %d."
1019 " Evicting client %s with export %s.\n",
1020 PFID(rr->rr_fid1), rc,
1021 obd_uuid2str(&exp->exp_client_uuid),
1022 obd_export_nid2str(exp));
1023 mdt_export_evict(exp);
1026 child = mdt_object_find(env, mdt, rr->rr_fid2);
1027 if (IS_ERR(child)) {
1028 rc = PTR_ERR(child);
1029 LCONSOLE_WARN("Child "DFID" lookup error %d."
1030 " Evicting client %s with export %s.\n",
1031 PFID(mdt_object_fid(child)), rc,
1032 obd_uuid2str(&exp->exp_client_uuid),
1033 obd_export_nid2str(exp));
1034 mdt_object_put(env, parent);
1035 mdt_export_evict(exp);
1038 rc = mdt_object_exists(child);
1041 mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA);
1042 rc = mdt_attr_get_complex(info, child, ma);
1044 rc = mdt_finish_open(info, parent, child,
1045 flags, 1, ldlm_rep);
1046 } else if (rc < 0) {
1047 /* the child object was created on remote server */
1048 repbody->fid1 = *rr->rr_fid2;
1049 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1051 } else if (rc == 0) {
1052 /* the child does not exist, we should do regular open */
1053 mdt_object_put(env, parent);
1054 mdt_object_put(env, child);
1055 GOTO(regular_open, 0);
1057 mdt_object_put(env, parent);
1058 mdt_object_put(env, child);
1062 /* We did not try to create, so we are a pure open */
1063 rc = mdt_reint_open(info, lhc);
1068 req->rq_status = rc;
1069 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1070 LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1073 int mdt_open_by_fid(struct mdt_thread_info* info,
1074 struct ldlm_reply *rep)
1076 __u32 flags = info->mti_spec.sp_cr_flags;
1077 struct mdt_reint_record *rr = &info->mti_rr;
1078 struct md_attr *ma = &info->mti_attr;
1079 struct mdt_object *o;
1083 o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1085 RETURN(rc = PTR_ERR(o));
1087 rc = mdt_object_exists(o);
1089 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1093 rc = mdt_attr_get_complex(info, o, ma);
1095 rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1096 } else if (rc == 0) {
1099 /* the child object was created on remote server */
1100 struct mdt_body *repbody;
1101 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1102 repbody->fid1 = *rr->rr_fid2;
1103 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1107 mdt_object_put(info->mti_env, o);
1111 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1112 struct mdt_lock_handle *lhc)
1114 const struct lu_env *env = info->mti_env;
1115 struct mdt_device *mdt = info->mti_mdt;
1116 __u32 flags = info->mti_spec.sp_cr_flags;
1117 struct mdt_reint_record *rr = &info->mti_rr;
1118 struct md_attr *ma = &info->mti_attr;
1119 struct mdt_object *parent= NULL;
1120 struct mdt_object *o;
1125 if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1126 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1127 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1128 if (IS_ERR(parent)) {
1129 CDEBUG(D_INODE, "Fail to find parent "DFID
1130 " for anonymous created %ld, try to"
1131 " use server-side parent.\n",
1132 PFID(rr->rr_fid1), PTR_ERR(parent));
1137 ma->ma_need |= MA_PFID;
1140 o = mdt_object_find(env, mdt, rr->rr_fid2);
1142 RETURN(rc = PTR_ERR(o));
1144 rc = mdt_object_exists(o);
1146 mdt_set_disposition(info, rep, (DISP_LOOKUP_EXECD |
1148 GOTO(out, rc = -ENOENT);
1149 } else if (rc < 0) {
1150 CERROR("NFS remote open shouldn't happen.\n");
1153 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1157 if (flags & FMODE_WRITE)
1159 else if (flags & MDS_FMODE_EXEC)
1164 mdt_lock_handle_init(lhc);
1165 mdt_lock_reg_init(lhc, lm);
1166 rc = mdt_object_lock(info, o, lhc,
1167 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1172 rc = mdt_attr_get_complex(info, o, ma);
1176 if (ma->ma_valid & MA_PFID) {
1177 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1178 if (IS_ERR(parent)) {
1179 CDEBUG(D_INODE, "Fail to find parent "DFID
1180 " for anonymous created %ld, try to"
1181 " use system default.\n",
1182 PFID(&ma->ma_pfid), PTR_ERR(parent));
1187 if (flags & MDS_OPEN_LOCK)
1188 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1189 rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1191 if (!(flags & MDS_OPEN_LOCK) || rc)
1192 mdt_object_unlock(info, o, lhc, 1);
1196 mdt_object_put(env, o);
1198 mdt_object_put(env, parent);
1202 int mdt_pin(struct mdt_thread_info* info)
1205 RETURN(err_serious(-EOPNOTSUPP));
1208 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1209 int mdt_cross_open(struct mdt_thread_info* info,
1210 const struct lu_fid *fid,
1211 struct ldlm_reply *rep, __u32 flags)
1213 struct md_attr *ma = &info->mti_attr;
1214 struct mdt_object *o;
1218 o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1220 RETURN(rc = PTR_ERR(o));
1222 rc = mdt_object_exists(o);
1224 /* Do permission check for cross-open. */
1225 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o),
1226 NULL, flags | MDS_OPEN_CROSS);
1230 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1231 rc = mdt_attr_get_complex(info, o, ma);
1233 rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1234 } else if (rc == 0) {
1236 * Something is wrong here. lookup was positive but there is
1239 CERROR("Cross-ref object doesn't exist!\n");
1242 /* Something is wrong here, the object is on another MDS! */
1243 CERROR("The object isn't on this server! FLD error?\n");
1244 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1246 "Object isn't on this server! FLD error?\n");
1252 mdt_object_put(info->mti_env, o);
1256 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1258 struct mdt_device *mdt = info->mti_mdt;
1259 struct ptlrpc_request *req = mdt_info_req(info);
1260 struct mdt_object *parent;
1261 struct mdt_object *child;
1262 struct mdt_lock_handle *lh;
1263 struct ldlm_reply *ldlm_rep;
1264 struct mdt_body *repbody;
1265 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1266 struct md_attr *ma = &info->mti_attr;
1267 __u64 create_flags = info->mti_spec.sp_cr_flags;
1268 struct mdt_reint_record *rr = &info->mti_rr;
1269 struct lu_name *lname;
1275 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1276 (obd_timeout + 1) / 4);
1278 mdt_counter_incr(req, LPROC_MDT_OPEN);
1279 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1281 ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1282 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1284 ma->ma_need = MA_INODE;
1285 if (ma->ma_lmm_size > 0)
1286 ma->ma_need |= MA_LOV;
1290 LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1291 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1293 if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1294 CERROR("file join is not supported anymore.\n");
1295 GOTO(out, result = err_serious(-EOPNOTSUPP));
1297 msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1299 if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1300 info->mti_spec.u.sp_ea.eadata == NULL)
1301 GOTO(out, result = err_serious(-EINVAL));
1303 CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1304 "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1305 PFID(rr->rr_fid1), rr->rr_name,
1306 PFID(rr->rr_fid2), create_flags,
1307 ma->ma_attr.la_mode, msg_flags);
1309 if (req_is_replay(req) ||
1310 (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1311 /* This is a replay request or from liblustre with ea. */
1312 result = mdt_open_by_fid(info, ldlm_rep);
1314 if (result != -ENOENT) {
1315 if (req->rq_export->exp_libclient &&
1316 create_flags & MDS_OPEN_HAS_EA)
1317 GOTO(out, result = 0);
1320 /* We didn't find the correct object, so we need to re-create it
1321 * via a regular replay. */
1322 if (!(create_flags & MDS_OPEN_CREAT)) {
1323 DEBUG_REQ(D_ERROR, req,
1324 "OPEN & CREAT not in open replay/by_fid.");
1325 GOTO(out, result = -EFAULT);
1327 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1328 } else if ((rr->rr_namelen == 0 && !info->mti_cross_ref &&
1329 create_flags & MDS_OPEN_LOCK) ||
1330 (create_flags & MDS_OPEN_BY_FID)) {
1331 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1332 if (result != -ENOENT && !(create_flags & MDS_OPEN_CREAT))
1334 if (unlikely(rr->rr_namelen == 0))
1335 GOTO(out, result = -EINVAL);
1336 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1339 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1340 GOTO(out, result = err_serious(-ENOMEM));
1342 mdt_set_disposition(info, ldlm_rep,
1343 (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1345 if (info->mti_cross_ref) {
1346 /* This is cross-ref open */
1347 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1348 result = mdt_cross_open(info, rr->rr_fid1, ldlm_rep,
1353 lh = &info->mti_lh[MDT_LH_PARENT];
1354 mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1355 LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1357 parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1358 MDS_INODELOCK_UPDATE);
1360 GOTO(out, result = PTR_ERR(parent));
1362 /* get and check version of parent */
1363 result = mdt_version_get_check(info, parent, 0);
1365 GOTO(out_parent, result);
1367 fid_zero(child_fid);
1369 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1370 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1371 lname, child_fid, &info->mti_spec);
1372 LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1373 "looking for "DFID"/%s, result fid="DFID"\n",
1374 PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1376 if (result != 0 && result != -ENOENT && result != -ESTALE)
1377 GOTO(out_parent, result);
1379 if (result == -ENOENT || result == -ESTALE) {
1380 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1381 if (result == -ESTALE) {
1383 * -ESTALE means the parent is a dead(unlinked) dir, so
1384 * it should return -ENOENT to in accordance with the
1385 * original mds implementaion.
1387 GOTO(out_parent, result = -ENOENT);
1389 if (!(create_flags & MDS_OPEN_CREAT))
1390 GOTO(out_parent, result);
1391 *child_fid = *info->mti_rr.rr_fid2;
1392 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1394 /* In the function below, .hs_keycmp resolves to
1395 * lu_obj_hop_keycmp() */
1396 /* coverity[overrun-buffer-val] */
1397 child = mdt_object_new(info->mti_env, mdt, child_fid);
1400 * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1401 * return FID back in that case.
1403 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1404 child = mdt_object_find(info->mti_env, mdt, child_fid);
1407 GOTO(out_parent, result = PTR_ERR(child));
1409 /** check version of child */
1410 rc = mdt_version_get_check(info, child, 1);
1412 GOTO(out_child, result = rc);
1414 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1415 if (result == -ENOENT) {
1416 if (mdt_object_obf(parent))
1417 GOTO(out_child, result = -EPERM);
1419 /* save versions in reply */
1420 mdt_version_get_save(info, parent, 0);
1421 mdt_version_get_save(info, child, 1);
1423 /* version of child will be changed */
1424 info->mti_mos = child;
1426 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1427 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1429 /* Let lower layers know what is lock mode on directory. */
1430 info->mti_spec.sp_cr_mode =
1431 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1434 * Do not perform lookup sanity check. We know that name does
1437 info->mti_spec.sp_cr_lookup = 0;
1438 info->mti_spec.sp_feat = &dt_directory_features;
1440 result = mdo_create(info->mti_env,
1441 mdt_object_child(parent),
1443 mdt_object_child(child),
1446 if (result == -ERESTART) {
1447 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1448 GOTO(out_child, result);
1451 /* XXX: we should call this once, see few lines below */
1453 result = mdt_attr_get_complex(info, child, ma);
1456 GOTO(out_child, result);
1460 /* We have to get attr & lov ea for this object */
1461 result = mdt_attr_get_complex(info, child, ma);
1463 * The object is on remote node, return its FID for remote open.
1465 if (result == -EREMOTE) {
1467 * Check if this lock already was sent to client and
1468 * this is resent case. For resent case do not take lock
1469 * again, use what is already granted.
1471 LASSERT(lhc != NULL);
1473 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1474 struct ldlm_lock *lock;
1476 LASSERT(msg_flags & MSG_RESENT);
1478 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1480 CERROR("Invalid lock handle "LPX64"\n",
1481 lhc->mlh_reg_lh.cookie);
1484 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1485 &lock->l_resource->lr_name));
1486 LDLM_LOCK_PUT(lock);
1489 mdt_lock_handle_init(lhc);
1490 mdt_lock_reg_init(lhc, LCK_PR);
1492 rc = mdt_object_lock(info, child, lhc,
1493 MDS_INODELOCK_LOOKUP,
1496 repbody->fid1 = *mdt_object_fid(child);
1497 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1500 GOTO(out_child, result);
1504 LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1506 /* get openlock if this is not replay and if a client requested it */
1507 if (!req_is_replay(req) && create_flags & MDS_OPEN_LOCK) {
1510 if (create_flags & FMODE_WRITE)
1512 else if (create_flags & MDS_FMODE_EXEC)
1516 mdt_lock_handle_init(lhc);
1517 mdt_lock_reg_init(lhc, lm);
1518 rc = mdt_object_lock(info, child, lhc,
1519 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1523 GOTO(out_child, result);
1526 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1530 /* Try to open it now. */
1531 rc = mdt_finish_open(info, parent, child, create_flags,
1535 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1536 /* openlock was acquired and mdt_finish_open failed -
1537 drop the openlock */
1538 mdt_object_unlock(info, child, lhc, 1);
1542 ma->ma_cookie_size = 0;
1543 rc = mdo_unlink(info->mti_env,
1544 mdt_object_child(parent),
1545 mdt_object_child(child),
1549 CERROR("Error in cleanup of open\n");
1554 mdt_object_put(info->mti_env, child);
1556 mdt_object_unlock_put(info, parent, lh, result || !created);
1558 if (result && result != -EREMOTE)
1559 lustre_msg_set_transno(req->rq_repmsg, 0);
1563 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
1564 MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
1566 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1568 return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1571 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1573 struct mdt_object *o = mfd->mfd_object;
1574 struct md_object *next = mdt_object_child(o);
1575 struct md_attr *ma = &info->mti_attr;
1576 int ret = MDT_IOEPOCH_CLOSED;
1581 mode = mfd->mfd_mode;
1583 if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
1585 ret = mdt_ioepoch_close(info, o);
1586 } else if (mode & MDS_FMODE_EXEC) {
1588 } else if (mode & MDS_FMODE_EPOCH) {
1589 ret = mdt_ioepoch_close(info, o);
1590 } else if (mode & MDS_FMODE_SOM) {
1591 ret = mdt_som_au_close(info, o);
1594 /* Update atime on close only. */
1595 if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1596 && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1597 /* Set the atime only. */
1598 ma->ma_valid = MA_INODE;
1599 ma->ma_attr.la_valid = LA_ATIME;
1600 rc = mo_attr_set(info->mti_env, next, ma);
1603 ma->ma_need |= MA_INODE;
1604 ma->ma_valid &= ~MA_INODE;
1606 if (!MFD_CLOSED(mode))
1607 rc = mo_close(info->mti_env, next, ma, mode);
1609 if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
1610 struct mdt_export_data *med;
1612 /* The IOepoch is still opened or SOM update is needed.
1613 * Put mfd back into the list. */
1614 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
1615 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
1616 MDS_FMODE_EPOCH : MDS_FMODE_SOM);
1618 LASSERT(mdt_info_req(info));
1619 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
1620 cfs_spin_lock(&med->med_open_lock);
1621 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
1622 class_handle_hash_back(&mfd->mfd_handle);
1623 cfs_spin_unlock(&med->med_open_lock);
1625 if (ret == MDT_IOEPOCH_OPENED) {
1629 CDEBUG(D_INODE, "Size-on-MDS attribute update is "
1630 "needed on "DFID"\n", PFID(mdt_object_fid(o)));
1634 mdt_object_put(info->mti_env, o);
1637 RETURN(rc ? rc : ret);
1640 int mdt_close(struct mdt_thread_info *info)
1642 struct mdt_export_data *med;
1643 struct mdt_file_data *mfd;
1644 struct mdt_object *o;
1645 struct md_attr *ma = &info->mti_attr;
1646 struct mdt_body *repbody = NULL;
1647 struct ptlrpc_request *req = mdt_info_req(info);
1651 mdt_counter_incr(req, LPROC_MDT_CLOSE);
1652 /* Close may come with the Size-on-MDS update. Unpack it. */
1653 rc = mdt_close_unpack(info);
1655 RETURN(err_serious(rc));
1657 LASSERT(info->mti_ioepoch);
1659 req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
1660 info->mti_mdt->mdt_max_mdsize);
1661 req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
1662 info->mti_mdt->mdt_max_cookiesize);
1663 rc = req_capsule_server_pack(info->mti_pill);
1664 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
1665 mdt_client_compatibility(info);
1667 mdt_fix_reply(info);
1668 RETURN(lustre_msg_get_status(req->rq_repmsg));
1671 /* Continue to close handle even if we can not pack reply */
1673 repbody = req_capsule_server_get(info->mti_pill,
1675 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
1677 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
1680 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
1682 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
1685 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
1686 repbody->eadatasize = 0;
1687 repbody->aclsize = 0;
1689 rc = err_serious(rc);
1692 med = &req->rq_export->exp_mdt_data;
1693 cfs_spin_lock(&med->med_open_lock);
1694 mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1695 if (mdt_mfd_closed(mfd)) {
1696 cfs_spin_unlock(&med->med_open_lock);
1697 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
1698 ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
1699 info->mti_ioepoch->handle.cookie);
1700 /** not serious error since bug 3633 */
1703 class_handle_unhash(&mfd->mfd_handle);
1704 cfs_list_del_init(&mfd->mfd_list);
1705 cfs_spin_unlock(&med->med_open_lock);
1707 /* Do not lose object before last unlink. */
1708 o = mfd->mfd_object;
1709 mdt_object_get(info->mti_env, o);
1710 ret = mdt_mfd_close(info, mfd);
1711 if (repbody != NULL)
1712 rc = mdt_handle_last_unlink(info, o, ma);
1713 mdt_empty_transno(info, rc);
1714 mdt_object_put(info->mti_env, o);
1716 if (repbody != NULL) {
1717 mdt_client_compatibility(info);
1718 rc = mdt_fix_reply(info);
1721 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
1722 RETURN(err_serious(-ENOMEM));
1724 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
1725 OBD_FAIL_MDS_CLOSE_NET_REP))
1726 info->mti_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
1727 RETURN(rc ? rc : ret);
1731 * DONE_WRITING rpc handler.
1733 * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
1734 * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
1735 * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
1736 * skip attributes and reconstruct the reply here.
1738 int mdt_done_writing(struct mdt_thread_info *info)
1740 struct ptlrpc_request *req = mdt_info_req(info);
1741 struct mdt_body *repbody = NULL;
1742 struct mdt_export_data *med;
1743 struct mdt_file_data *mfd;
1747 rc = req_capsule_server_pack(info->mti_pill);
1749 RETURN(err_serious(rc));
1751 repbody = req_capsule_server_get(info->mti_pill,
1753 repbody->eadatasize = 0;
1754 repbody->aclsize = 0;
1756 /* Done Writing may come with the Size-on-MDS update. Unpack it. */
1757 rc = mdt_close_unpack(info);
1759 RETURN(err_serious(rc));
1761 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL))
1762 RETURN(lustre_msg_get_status(req->rq_repmsg));
1764 med = &info->mti_exp->exp_mdt_data;
1765 cfs_spin_lock(&med->med_open_lock);
1766 mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1768 cfs_spin_unlock(&med->med_open_lock);
1769 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
1770 ": cookie = "LPX64" ioepoch = "LPU64"\n",
1771 PFID(info->mti_rr.rr_fid1),
1772 info->mti_ioepoch->handle.cookie,
1773 info->mti_ioepoch->ioepoch);
1774 /* If this is a replay, reconstruct the transno. */
1775 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1776 rc = info->mti_ioepoch->flags & MF_SOM_AU ?
1778 mdt_empty_transno(info, rc);
1784 LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
1785 mfd->mfd_mode == MDS_FMODE_TRUNC);
1786 class_handle_unhash(&mfd->mfd_handle);
1787 cfs_list_del_init(&mfd->mfd_list);
1788 cfs_spin_unlock(&med->med_open_lock);
1790 /* Set EPOCH CLOSE flag if not set by client. */
1791 info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
1792 info->mti_attr.ma_valid = 0;
1794 info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
1795 OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1796 if (info->mti_attr.ma_lmm == NULL)
1799 rc = mdt_mfd_close(info, mfd);
1801 OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1802 mdt_empty_transno(info, rc);