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, 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(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 mutex_lock(&o->mot_lov_mutex);
122 if (!(o->mot_flags & MOF_LOV_CREATED)) {
123 if (p != NULL && (fid_is_obf(mdt_object_fid(p)) ||
124 fid_is_dot_lustre(mdt_object_fid(p))))
125 GOTO(unlock, rc = -EPERM);
127 rc = mdo_create_data(info->mti_env,
128 p ? mdt_object_child(p) : NULL,
129 mdt_object_child(o), spec, ma);
131 rc = mdt_attr_get_complex(info, o, ma);
133 if (rc == 0 && ma->ma_valid & MA_LOV)
134 o->mot_flags |= MOF_LOV_CREATED;
137 mutex_unlock(&o->mot_lov_mutex);
141 static int mdt_ioepoch_opened(struct mdt_object *mo)
143 return mo->mot_ioepoch_count;
146 int mdt_object_is_som_enabled(struct mdt_object *mo)
148 return !mo->mot_ioepoch;
152 * Re-enable Size-on-MDS.
153 * Call under ->mot_ioepoch_mutex.
155 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
157 if (ioepoch == mo->mot_ioepoch) {
158 LASSERT(!mdt_ioepoch_opened(mo));
165 * Open the IOEpoch. It is allowed if @writecount is not negative.
166 * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
168 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
171 struct mdt_device *mdt = info->mti_mdt;
176 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
177 !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
180 mutex_lock(&o->mot_ioepoch_mutex);
181 if (mdt_ioepoch_opened(o)) {
182 /* Epoch continues even if there is no writers yet. */
183 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
184 o->mot_ioepoch, PFID(mdt_object_fid(o)));
186 /* XXX: ->mdt_ioepoch is not initialized at the mount */
187 spin_lock(&mdt->mdt_ioepoch_lock);
188 if (mdt->mdt_ioepoch < info->mti_replayepoch)
189 mdt->mdt_ioepoch = info->mti_replayepoch;
191 if (info->mti_replayepoch)
192 o->mot_ioepoch = info->mti_replayepoch;
193 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
194 o->mot_ioepoch = ++mdt->mdt_ioepoch;
196 o->mot_ioepoch = mdt->mdt_ioepoch;
198 spin_unlock(&mdt->mdt_ioepoch_lock);
200 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
201 o->mot_ioepoch, PFID(mdt_object_fid(o)));
203 o->mot_flags |= MOF_SOM_CREATED;
206 o->mot_ioepoch_count++;
207 mutex_unlock(&o->mot_ioepoch_mutex);
209 /* Cancel Size-on-MDS attributes cached on clients for the open case.
210 * In the truncate case, see mdt_reint_setattr(). */
211 if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
212 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
213 mdt_lock_reg_init(lh, LCK_EX);
214 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
217 mdt_object_unlock(info, o, lh, 1);
223 * Update SOM on-disk attributes.
224 * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
225 * mountid and attributes. If disabling, clean SOM xattr.
226 * Call under ->mot_ioepoch_mutex.
228 static int mdt_som_attr_set(struct mdt_thread_info *info,
229 struct mdt_object *obj, __u64 ioepoch, bool enable)
231 struct md_object *next = mdt_object_child(obj);
235 CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
236 " on "DFID".\n", enable ? "update" : "disabling",
237 ioepoch, PFID(mdt_object_fid(obj)));
240 struct lu_buf *buf = &info->mti_buf;
241 struct som_attrs *attrs;
242 struct md_attr *ma = &info->mti_attr;
243 struct lu_attr *la = &ma->ma_attr;
244 struct obd_device *obd = info->mti_mdt->mdt_lut.lut_obd;
246 attrs = (struct som_attrs *)info->mti_xattr_buf;
247 CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*attrs));
249 /* pack SOM attributes */
250 memset(attrs, 0, sizeof(*attrs));
251 attrs->som_ioepoch = ioepoch;
252 attrs->som_mountid = obd->u.obt.obt_mount_count;
253 if ((la->la_valid & LA_SIZE) != 0)
254 attrs->som_size = la->la_size;
255 if ((la->la_valid & LA_BLOCKS) != 0)
256 attrs->som_blocks = la->la_blocks;
257 lustre_som_swab(attrs);
259 /* update SOM attributes */
261 buf->lb_len = sizeof(*attrs);
262 rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_SOM, 0);
264 /* delete SOM attributes */
265 rc = mo_xattr_del(info->mti_env, next, XATTR_NAME_SOM);
271 /** Perform the eviction specific actions on ioepoch close. */
272 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
273 struct mdt_object *o)
277 mutex_lock(&o->mot_ioepoch_mutex);
278 CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
279 "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
280 o->mot_ioepoch_count);
281 o->mot_ioepoch_count--;
283 /* If eviction occured set MOF_SOM_RECOV,
284 * if no other epoch holders, disable SOM on disk. */
285 o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
286 if (!mdt_ioepoch_opened(o)) {
287 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
288 mdt_object_som_enable(o, o->mot_ioepoch);
290 mutex_unlock(&o->mot_ioepoch_mutex);
295 * Perform the replay specific actions on ioepoch close.
296 * Skip SOM attribute update if obtained and just forget about the inode state
297 * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
299 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
300 struct mdt_object *o)
302 int rc = MDT_IOEPOCH_CLOSED;
305 mutex_lock(&o->mot_ioepoch_mutex);
306 CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
307 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
308 o->mot_ioepoch_count--;
310 /* Get an info from the replayed request if client is supposed
311 * to send an Attibute Update, reconstruct @rc if so */
312 if (info->mti_ioepoch->flags & MF_SOM_AU)
313 rc = MDT_IOEPOCH_GETATTR;
315 if (!mdt_ioepoch_opened(o))
316 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
317 mutex_unlock(&o->mot_ioepoch_mutex);
323 * Regular file IOepoch close.
324 * Closes the ioepoch, checks the object state, apply obtained attributes and
325 * re-enable SOM on the object, if possible. Also checks if the recovery is
326 * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
327 * to obtain SOM attributes under the server-side OST locks.
330 * MDT_IOEPOCH_CLOSED if ioepoch is closed.
331 * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
333 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
334 struct mdt_object *o)
336 struct md_attr *tmp_ma;
340 int rc = 0, ret = MDT_IOEPOCH_CLOSED;
343 la = &info->mti_attr.ma_attr;
344 achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
346 mutex_lock(&o->mot_ioepoch_mutex);
347 o->mot_ioepoch_count--;
349 tmp_ma = &info->mti_u.som.attr;
350 tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
351 tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
352 tmp_ma->ma_som = &info->mti_u.som.data;
353 tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
354 tmp_ma->ma_valid = 0;
355 rc = mdt_attr_get_complex(info, o, tmp_ma);
359 /* Check the on-disk SOM state. */
360 if (o->mot_flags & MOF_SOM_RECOV)
362 else if (!(o->mot_flags & MOF_SOM_CREATED) &&
363 !(tmp_ma->ma_valid & MA_SOM))
366 CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
367 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
369 opened = mdt_ioepoch_opened(o);
371 * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
372 * Skip the check for file with no LOV or for unlink files.
374 if (!opened && tmp_ma->ma_valid & MA_LOV &&
375 !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
377 /* If some previous writer was evicted, re-ask the
378 * client for attributes. Even if attributes are
379 * provided, we cannot believe in them.
380 * Another use case is that there is no SOM cache on
381 * disk -- first access with SOM or there was an MDS
383 ret = MDT_IOEPOCH_GETATTR;
384 else if (o->mot_flags & MOF_SOM_CHANGE)
385 /* Some previous writer changed the attribute.
386 * Do not believe to the current Size-on-MDS
387 * update, re-ask client. */
388 ret = MDT_IOEPOCH_GETATTR;
389 else if (!(la->la_valid & LA_SIZE) && achange)
390 /* Attributes were changed by the last writer
391 * only but no Size-on-MDS update is received.*/
392 ret = MDT_IOEPOCH_GETATTR;
395 if (achange || ret == MDT_IOEPOCH_GETATTR)
396 o->mot_flags |= MOF_SOM_CHANGE;
398 /* If epoch ends and relable SOM attributes are obtained, update them.
399 * Create SOM ea for new files even if there is no attributes obtained
400 * (0-length file). */
401 if (ret == MDT_IOEPOCH_CLOSED && !opened) {
402 if (achange || o->mot_flags & MOF_SOM_CREATED) {
403 LASSERT(achange || !(la->la_valid & LA_SIZE));
404 rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
406 /* Avoid the following setattrs of these attributes,
407 * e.g. for atime update. */
408 info->mti_attr.ma_valid = 0;
410 mdt_object_som_enable(o, o->mot_ioepoch);
413 mutex_unlock(&o->mot_ioepoch_mutex);
414 /* If recovery is needed, tell the client to perform GETATTR under
416 if (ret == MDT_IOEPOCH_GETATTR && recovery) {
417 struct mdt_body *rep;
418 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
419 rep->valid |= OBD_MD_FLGETATTRLOCK;
425 mutex_unlock(&o->mot_ioepoch_mutex);
430 * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
431 * - a client closes the IOEpoch;
432 * - a client eviction occured.
434 * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
435 * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
436 * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
439 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
441 struct ptlrpc_request *req = mdt_info_req(info);
444 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
445 !S_ISREG(lu_object_attr(&o->mot_obj.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 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 mutex_unlock(&o->mot_ioepoch_mutex);
512 int mdt_write_read(struct mdt_object *o)
516 mutex_lock(&o->mot_ioepoch_mutex);
517 rc = o->mot_writecount;
518 mutex_unlock(&o->mot_ioepoch_mutex);
522 int mdt_write_get(struct mdt_object *o)
526 mutex_lock(&o->mot_ioepoch_mutex);
527 if (o->mot_writecount < 0)
531 mutex_unlock(&o->mot_ioepoch_mutex);
535 void mdt_write_put(struct mdt_object *o)
538 mutex_lock(&o->mot_ioepoch_mutex);
540 mutex_unlock(&o->mot_ioepoch_mutex);
544 static int mdt_write_deny(struct mdt_object *o)
548 mutex_lock(&o->mot_ioepoch_mutex);
549 if (o->mot_writecount > 0)
553 mutex_unlock(&o->mot_ioepoch_mutex);
557 static void mdt_write_allow(struct mdt_object *o)
560 mutex_lock(&o->mot_ioepoch_mutex);
562 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 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 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 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 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 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 /* if file is released, we can't deny write because we must
713 * restore (write) it to access it.*/
714 if ((ma->ma_valid & MA_HSM) &&
715 (ma->ma_hsm.mh_flags & HS_RELEASED))
718 rc = mdt_write_deny(o);
723 rc = mo_open(info->mti_env, mdt_object_child(o),
724 created ? flags | MDS_OPEN_CREATED : flags);
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);
737 * @flags is always not zero. At least it should be FMODE_READ,
738 * FMODE_WRITE or MDS_FMODE_EXEC.
743 mdt_mfd_set_mode(mfd, flags);
746 mfd->mfd_xid = req->rq_xid;
749 if (req_is_replay(req)) {
750 struct mdt_file_data *old_mfd;
751 /* Check wheather old cookie already exist in
752 * the list, becasue when do recovery, client
753 * might be disconnected from server, and
754 * restart replay, so there maybe some orphan
755 * mfd here, we should remove them */
756 LASSERT(info->mti_rr.rr_handle != NULL);
757 old_mfd = mdt_handle2mfd(info, info->mti_rr.rr_handle);
759 CDEBUG(D_HA, "del orph mfd %p fid=("DFID") "
760 "cookie=" LPX64"\n", mfd,
761 PFID(mdt_object_fid(mfd->mfd_object)),
762 info->mti_rr.rr_handle->cookie);
763 spin_lock(&med->med_open_lock);
764 class_handle_unhash(&old_mfd->mfd_handle);
765 cfs_list_del_init(&old_mfd->mfd_list);
766 spin_unlock(&med->med_open_lock);
767 /* no attr update for that close */
769 ma->ma_valid |= MA_FLAGS;
770 ma->ma_attr_flags |= MDS_RECOV_OPEN;
771 mdt_mfd_close(info, old_mfd);
772 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
773 ma->ma_valid &= ~MA_FLAGS;
775 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
776 info->mti_rr.rr_handle->cookie);
777 mfd->mfd_old_handle.cookie =
778 info->mti_rr.rr_handle->cookie;
780 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
782 if (req->rq_export->exp_disconnected) {
783 spin_lock(&med->med_open_lock);
784 class_handle_unhash(&mfd->mfd_handle);
785 cfs_list_del_init(&mfd->mfd_list);
786 spin_unlock(&med->med_open_lock);
787 mdt_mfd_close(info, mfd);
789 spin_lock(&med->med_open_lock);
790 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
791 spin_unlock(&med->med_open_lock);
794 mdt_empty_transno(info, rc);
796 GOTO(err_out, rc = -ENOMEM);
802 if (flags & FMODE_WRITE)
803 /* XXX We also need to close io epoch here.
804 * See LU-1220 - green */
806 else if (flags & FMODE_EXEC)
811 int mdt_finish_open(struct mdt_thread_info *info,
812 struct mdt_object *p, struct mdt_object *o,
813 __u64 flags, int created, struct ldlm_reply *rep)
815 struct ptlrpc_request *req = mdt_info_req(info);
816 struct obd_export *exp = req->rq_export;
817 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
818 struct md_attr *ma = &info->mti_attr;
819 struct lu_attr *la = &ma->ma_attr;
820 struct mdt_file_data *mfd;
821 struct mdt_body *repbody;
823 int isreg, isdir, islnk;
827 LASSERT(ma->ma_valid & MA_INODE);
829 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
831 isreg = S_ISREG(la->la_mode);
832 isdir = S_ISDIR(la->la_mode);
833 islnk = S_ISLNK(la->la_mode);
834 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
836 /* LU-2275, simulate broken behaviour (esp. prevalent in
837 * pre-2.4 servers where a very strange reply is sent on error
838 * that looks like it was actually almost succesful and a failure at the
840 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
841 mdt_set_disposition(info, rep, DISP_OPEN_LOCK | \
842 DISP_OPEN_OPEN | DISP_LOOKUP_NEG | \
847 if (exp_connect_rmtclient(exp)) {
848 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
850 rc = mdt_pack_remote_perm(info, o, buf);
852 repbody->valid &= ~OBD_MD_FLRMTPERM;
853 repbody->aclsize = 0;
855 repbody->valid |= OBD_MD_FLRMTPERM;
856 repbody->aclsize = sizeof(struct mdt_remote_perm);
859 #ifdef CONFIG_FS_POSIX_ACL
860 else if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
861 const struct lu_env *env = info->mti_env;
862 struct md_object *next = mdt_object_child(o);
863 struct lu_buf *buf = &info->mti_buf;
865 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
866 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
868 if (buf->lb_len > 0) {
869 rc = mo_xattr_get(env, next, buf,
870 XATTR_NAME_ACL_ACCESS);
872 if (rc == -ENODATA) {
873 repbody->aclsize = 0;
874 repbody->valid |= OBD_MD_FLACL;
876 } else if (rc == -EOPNOTSUPP) {
879 CERROR("got acl size: %d\n", rc);
882 repbody->aclsize = rc;
883 repbody->valid |= OBD_MD_FLACL;
890 if (info->mti_mdt->mdt_opts.mo_mds_capa &&
891 exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) {
892 struct lustre_capa *capa;
894 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
896 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
897 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
900 repbody->valid |= OBD_MD_FLMDSCAPA;
902 if (info->mti_mdt->mdt_opts.mo_oss_capa &&
903 exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA &&
904 S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) {
905 struct lustre_capa *capa;
907 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
909 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
910 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
913 repbody->valid |= OBD_MD_FLOSSCAPA;
917 * If we are following a symlink, don't open; and do not return open
918 * handle for special nodes as client required.
920 if (islnk || (!isreg && !isdir &&
921 (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
922 lustre_msg_set_transno(req->rq_repmsg, 0);
927 * We need to return the existing object's fid back, so it is done here,
928 * after preparing the reply.
930 if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
933 /* This can't be done earlier, we need to return reply body */
935 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
936 /* We are trying to create or write an existing dir. */
939 } else if (flags & MDS_OPEN_DIRECTORY)
942 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
943 OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
948 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
949 spin_lock(&med->med_open_lock);
950 cfs_list_for_each(t, &med->med_open_head) {
951 mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
952 if (mfd->mfd_xid == req->rq_xid)
956 spin_unlock(&med->med_open_lock);
959 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
960 /*set repbody->ea_size for resent case*/
961 if (ma->ma_valid & MA_LOV) {
962 LASSERT(ma->ma_lmm_size != 0);
963 repbody->eadatasize = ma->ma_lmm_size;
965 repbody->valid |= OBD_MD_FLDIREA;
967 repbody->valid |= OBD_MD_FLEASIZE;
969 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
974 rc = mdt_mfd_open(info, p, o, flags, created);
976 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
981 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
982 struct lsd_client_data *lcd);
984 void mdt_reconstruct_open(struct mdt_thread_info *info,
985 struct mdt_lock_handle *lhc)
987 const struct lu_env *env = info->mti_env;
988 struct mdt_device *mdt = info->mti_mdt;
989 struct req_capsule *pill = info->mti_pill;
990 struct ptlrpc_request *req = mdt_info_req(info);
991 struct tg_export_data *ted = &req->rq_export->exp_target_data;
992 struct lsd_client_data *lcd = ted->ted_lcd;
993 struct md_attr *ma = &info->mti_attr;
994 struct mdt_reint_record *rr = &info->mti_rr;
995 __u32 flags = info->mti_spec.sp_cr_flags;
996 struct ldlm_reply *ldlm_rep;
997 struct mdt_object *parent;
998 struct mdt_object *child;
999 struct mdt_body *repbody;
1003 LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1004 ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1005 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1007 ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
1008 ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
1010 ma->ma_need = MA_INODE | MA_HSM;
1011 if (ma->ma_lmm_size > 0)
1012 ma->ma_need |= MA_LOV;
1016 mdt_req_from_lcd(req, lcd);
1017 mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1019 CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1020 ldlm_rep->lock_policy_res1, req->rq_status);
1022 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1023 req->rq_status != 0)
1024 /* We did not create successfully, return error to client. */
1025 GOTO(out, rc = req->rq_status);
1027 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1028 struct obd_export *exp = req->rq_export;
1030 * We failed after creation, but we do not know in which step
1031 * we failed. So try to check the child object.
1033 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1034 if (IS_ERR(parent)) {
1035 rc = PTR_ERR(parent);
1036 LCONSOLE_WARN("Parent "DFID" lookup error %d."
1037 " Evicting client %s with export %s.\n",
1038 PFID(rr->rr_fid1), rc,
1039 obd_uuid2str(&exp->exp_client_uuid),
1040 obd_export_nid2str(exp));
1041 mdt_export_evict(exp);
1044 child = mdt_object_find(env, mdt, rr->rr_fid2);
1045 if (IS_ERR(child)) {
1046 rc = PTR_ERR(child);
1047 LCONSOLE_WARN("Child "DFID" lookup error %d."
1048 " Evicting client %s with export %s.\n",
1049 PFID(mdt_object_fid(child)), rc,
1050 obd_uuid2str(&exp->exp_client_uuid),
1051 obd_export_nid2str(exp));
1052 mdt_object_put(env, parent);
1053 mdt_export_evict(exp);
1057 if (unlikely(mdt_object_remote(child))) {
1058 /* the child object was created on remote server */
1059 if (!mdt_is_dne_client(exp)) {
1060 /* Return -EIO for old client */
1061 mdt_object_put(env, parent);
1062 mdt_object_put(env, child);
1063 GOTO(out, rc = -EIO);
1065 repbody->fid1 = *rr->rr_fid2;
1066 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1069 if (mdt_object_exists(child)) {
1070 mdt_set_capainfo(info, 1, rr->rr_fid2,
1072 rc = mdt_attr_get_complex(info, child, ma);
1074 rc = mdt_finish_open(info, parent,
1078 /* the child does not exist, we should do
1080 mdt_object_put(env, parent);
1081 mdt_object_put(env, child);
1082 GOTO(regular_open, 0);
1085 mdt_object_put(env, parent);
1086 mdt_object_put(env, child);
1090 /* We did not try to create, so we are a pure open */
1091 rc = mdt_reint_open(info, lhc);
1096 req->rq_status = rc;
1097 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1098 LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1101 int mdt_open_by_fid(struct mdt_thread_info* info,
1102 struct ldlm_reply *rep)
1104 __u32 flags = info->mti_spec.sp_cr_flags;
1105 struct mdt_reint_record *rr = &info->mti_rr;
1106 struct md_attr *ma = &info->mti_attr;
1107 struct mdt_object *o;
1111 o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1113 RETURN(rc = PTR_ERR(o));
1115 if (unlikely(mdt_object_remote(o))) {
1116 /* the child object was created on remote server */
1117 struct mdt_body *repbody;
1119 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1122 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1123 repbody->fid1 = *rr->rr_fid2;
1124 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1127 if (mdt_object_exists(o)) {
1128 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1132 rc = mdt_attr_get_complex(info, o, ma);
1134 rc = mdt_finish_open(info, NULL, o, flags, 0,
1141 mdt_object_put(info->mti_env, o);
1145 /* lock object for open */
1146 static int mdt_object_open_lock(struct mdt_thread_info *info,
1147 struct mdt_object *obj,
1148 struct mdt_lock_handle *lhc,
1151 struct md_attr *ma = &info->mti_attr;
1152 __u64 open_flags = info->mti_spec.sp_cr_flags;
1153 ldlm_mode_t lm = LCK_CR;
1154 bool try_layout = false;
1155 bool create_layout = false;
1160 if (open_flags & MDS_OPEN_LOCK) {
1161 if (open_flags & FMODE_WRITE)
1163 /* if file is released, we can't deny write because we must
1164 * restore (write) it to access it. */
1165 else if ((open_flags & MDS_FMODE_EXEC) &&
1166 !((ma->ma_valid & MA_HSM) &&
1167 (ma->ma_hsm.mh_flags & HS_RELEASED)))
1172 *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
1175 if (S_ISREG(lu_object_attr(&obj->mot_obj.mo_lu))) {
1176 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
1177 md_should_create(open_flags))
1178 create_layout = true;
1179 if (exp_connect_layout(info->mti_exp) && !create_layout &&
1180 ma->ma_need & MA_LOV)
1184 mdt_lock_handle_init(lhc);
1185 mdt_lock_reg_init(lhc, lm);
1187 /* one problem to return layout lock on open is that it may result
1188 * in too many layout locks cached on the client side. */
1189 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
1190 /* return lookup lock to validate inode at the client side,
1191 * this is pretty important otherwise mdt will return layout
1192 * lock for each open.
1193 * However this is a double-edged sword because changing
1194 * permission will revoke huge # of LOOKUP locks. */
1195 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
1196 if (!mdt_object_lock_try(info, obj, lhc, *ibits,
1198 *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
1200 rc = mdt_object_lock(info, obj, lhc, *ibits,
1203 } else if (*ibits != 0) {
1204 rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK);
1207 CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
1208 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
1209 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
1211 /* will change layout, revoke layout locks by enqueuing EX lock. */
1212 if (rc == 0 && create_layout) {
1213 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
1215 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
1216 ", open_flags = "LPO64"\n",
1217 PFID(mdt_object_fid(obj)), open_flags);
1219 LASSERT(!try_layout);
1220 mdt_lock_handle_init(ll);
1221 mdt_lock_reg_init(ll, LCK_EX);
1222 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
1225 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
1231 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1232 struct mdt_object *obj,
1233 struct mdt_lock_handle *lhc,
1234 __u64 ibits, int rc)
1236 __u64 open_flags = info->mti_spec.sp_cr_flags;
1237 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
1239 /* Release local layout lock - the layout lock put in MDT_LH_LAYOUT
1240 * will never return to client side. */
1241 if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1242 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1243 mdt_object_unlock(info, obj, ll, 1);
1246 /* Cross-ref case, the lock should be returned to the client */
1247 if (ibits == 0 || rc == -EREMOTE)
1250 if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
1251 /* for the open request, the lock will only return to client
1252 * if open or layout lock is granted. */
1256 mdt_object_unlock(info, obj, lhc, 1);
1259 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1260 struct mdt_lock_handle *lhc)
1262 const struct lu_env *env = info->mti_env;
1263 struct mdt_device *mdt = info->mti_mdt;
1264 __u32 flags = info->mti_spec.sp_cr_flags;
1265 struct mdt_reint_record *rr = &info->mti_rr;
1266 struct md_attr *ma = &info->mti_attr;
1267 struct mdt_object *parent= NULL;
1268 struct mdt_object *o;
1273 if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1274 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1275 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1276 if (IS_ERR(parent)) {
1277 CDEBUG(D_INODE, "Fail to find parent "DFID
1278 " for anonymous created %ld, try to"
1279 " use server-side parent.\n",
1280 PFID(rr->rr_fid1), PTR_ERR(parent));
1285 ma->ma_need |= MA_PFID;
1288 o = mdt_object_find(env, mdt, rr->rr_fid2);
1290 RETURN(rc = PTR_ERR(o));
1292 if (mdt_object_remote(o)) {
1293 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1294 info->mti_mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name,
1296 GOTO(out, rc = -EREMOTE);
1297 } else if (!mdt_object_exists(o)) {
1298 mdt_set_disposition(info, rep,
1302 GOTO(out, rc = -ENOENT);
1305 mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1307 rc = mdt_attr_get_complex(info, o, ma);
1311 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1315 if (ma->ma_valid & MA_PFID) {
1316 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1317 if (IS_ERR(parent)) {
1318 CDEBUG(D_INODE, "Fail to find parent "DFID
1319 " for anonymous created %ld, try to"
1320 " use system default.\n",
1321 PFID(&ma->ma_pfid), PTR_ERR(parent));
1326 rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1328 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1329 if (flags & MDS_OPEN_LOCK)
1330 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1335 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1336 mdt_object_put(env, o);
1338 mdt_object_put(env, parent);
1342 int mdt_pin(struct mdt_thread_info* info)
1345 RETURN(err_serious(-EOPNOTSUPP));
1348 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1349 static int mdt_cross_open(struct mdt_thread_info *info,
1350 const struct lu_fid *parent_fid,
1351 const struct lu_fid *fid,
1352 struct ldlm_reply *rep, __u32 flags)
1354 struct md_attr *ma = &info->mti_attr;
1355 struct mdt_object *o;
1359 o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1361 RETURN(rc = PTR_ERR(o));
1363 if (mdt_object_remote(o)) {
1364 /* Something is wrong here, the object is on another MDS! */
1365 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1366 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1367 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1369 "Object isn't on this server! FLD error?\n");
1372 if (mdt_object_exists(o)) {
1373 /* Do permission check for cross-open. */
1374 rc = mo_permission(info->mti_env, NULL,
1375 mdt_object_child(o),
1376 NULL, flags | MDS_OPEN_CROSS);
1380 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1381 rc = mdt_attr_get_complex(info, o, ma);
1385 /* Do not create lov object if the fid is opened
1387 if (S_ISREG(ma->ma_attr.la_mode) &&
1388 !(ma->ma_valid & MA_LOV) && (flags & FMODE_WRITE) &&
1389 fid_is_obf(parent_fid))
1390 GOTO(out, rc = -EPERM);
1392 rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1395 * Something is wrong here. lookup was positive but
1396 * there is no object!
1398 CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1399 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1404 mdt_object_put(info->mti_env, o);
1408 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1410 struct mdt_device *mdt = info->mti_mdt;
1411 struct ptlrpc_request *req = mdt_info_req(info);
1412 struct mdt_object *parent;
1413 struct mdt_object *child;
1414 struct mdt_lock_handle *lh;
1415 struct ldlm_reply *ldlm_rep;
1416 struct mdt_body *repbody;
1417 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1418 struct md_attr *ma = &info->mti_attr;
1419 __u64 create_flags = info->mti_spec.sp_cr_flags;
1421 struct mdt_reint_record *rr = &info->mti_rr;
1422 struct lu_name *lname;
1428 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1429 (obd_timeout + 1) / 4);
1431 mdt_counter_incr(req, LPROC_MDT_OPEN);
1432 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1434 ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1435 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1437 ma->ma_need = MA_INODE;
1438 if (ma->ma_lmm_size > 0)
1439 ma->ma_need |= MA_LOV;
1443 LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1444 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1446 if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1447 CERROR("file join is not supported anymore.\n");
1448 GOTO(out, result = err_serious(-EOPNOTSUPP));
1450 msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1452 if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1453 info->mti_spec.u.sp_ea.eadata == NULL)
1454 GOTO(out, result = err_serious(-EINVAL));
1456 CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1457 "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1458 PFID(rr->rr_fid1), rr->rr_name,
1459 PFID(rr->rr_fid2), create_flags,
1460 ma->ma_attr.la_mode, msg_flags);
1461 if (info->mti_cross_ref) {
1462 /* This is cross-ref open */
1463 mdt_set_disposition(info, ldlm_rep,
1464 (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1466 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1467 ldlm_rep, create_flags);
1469 } else if (req_is_replay(req) ||
1470 (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1471 /* This is a replay request or from liblustre with ea. */
1472 result = mdt_open_by_fid(info, ldlm_rep);
1474 if (result != -ENOENT) {
1475 if (req->rq_export->exp_libclient &&
1476 create_flags & MDS_OPEN_HAS_EA)
1477 GOTO(out, result = 0);
1480 /* We didn't find the correct object, so we need to re-create it
1481 * via a regular replay. */
1482 if (!(create_flags & MDS_OPEN_CREAT)) {
1483 DEBUG_REQ(D_ERROR, req,
1484 "OPEN & CREAT not in open replay/by_fid.");
1485 GOTO(out, result = -EFAULT);
1487 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1488 } else if ((rr->rr_namelen == 0 && create_flags & MDS_OPEN_LOCK) ||
1489 (create_flags & MDS_OPEN_BY_FID)) {
1490 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1491 /* If result is 0 then open by FID has found the file
1492 * and there is nothing left for us to do here. More
1493 * generally if it is anything other than -ENOENT or
1494 * -EREMOTE then we return that now. If -ENOENT and
1495 * MDS_OPEN_CREAT is set then we must create the file
1496 * below. If -EREMOTE then we need to return a LOOKUP
1497 * lock to the client, which we do below. Hence this
1498 * odd looking condition. See LU-2523. */
1499 if (!(result == -ENOENT && (create_flags & MDS_OPEN_CREAT)) &&
1503 if (unlikely(rr->rr_namelen == 0))
1504 GOTO(out, result = -EINVAL);
1506 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1509 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1510 GOTO(out, result = err_serious(-ENOMEM));
1512 mdt_set_disposition(info, ldlm_rep,
1513 (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1515 lh = &info->mti_lh[MDT_LH_PARENT];
1516 mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1517 LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1519 parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1520 MDS_INODELOCK_UPDATE);
1522 GOTO(out, result = PTR_ERR(parent));
1524 /* get and check version of parent */
1525 result = mdt_version_get_check(info, parent, 0);
1527 GOTO(out_parent, result);
1529 fid_zero(child_fid);
1531 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1532 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1533 lname, child_fid, &info->mti_spec);
1534 LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1535 "looking for "DFID"/%s, result fid="DFID"\n",
1536 PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1538 if (result != 0 && result != -ENOENT && result != -ESTALE)
1539 GOTO(out_parent, result);
1541 if (result == -ENOENT || result == -ESTALE) {
1542 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1543 if (result == -ESTALE) {
1545 * -ESTALE means the parent is a dead(unlinked) dir, so
1546 * it should return -ENOENT to in accordance with the
1547 * original mds implementaion.
1549 GOTO(out_parent, result = -ENOENT);
1551 if (!(create_flags & MDS_OPEN_CREAT))
1552 GOTO(out_parent, result);
1553 *child_fid = *info->mti_rr.rr_fid2;
1554 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1556 /* In the function below, .hs_keycmp resolves to
1557 * lu_obj_hop_keycmp() */
1558 /* coverity[overrun-buffer-val] */
1559 child = mdt_object_new(info->mti_env, mdt, child_fid);
1562 * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1563 * return FID back in that case.
1565 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1566 child = mdt_object_find(info->mti_env, mdt, child_fid);
1569 GOTO(out_parent, result = PTR_ERR(child));
1571 /** check version of child */
1572 rc = mdt_version_get_check(info, child, 1);
1574 GOTO(out_child, result = rc);
1576 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1577 if (result == -ENOENT) {
1578 /* Create under OBF and .lustre is not permitted */
1579 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
1580 GOTO(out_child, result = -EPERM);
1582 /* save versions in reply */
1583 mdt_version_get_save(info, parent, 0);
1584 mdt_version_get_save(info, child, 1);
1586 /* version of child will be changed */
1587 info->mti_mos = child;
1589 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1590 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1592 /* Let lower layers know what is lock mode on directory. */
1593 info->mti_spec.sp_cr_mode =
1594 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1597 * Do not perform lookup sanity check. We know that name does
1600 info->mti_spec.sp_cr_lookup = 0;
1601 info->mti_spec.sp_feat = &dt_directory_features;
1603 result = mdo_create(info->mti_env,
1604 mdt_object_child(parent),
1606 mdt_object_child(child),
1609 if (result == -ERESTART) {
1610 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1611 GOTO(out_child, result);
1614 /* XXX: we should call this once, see few lines below */
1616 result = mdt_attr_get_complex(info, child, ma);
1619 GOTO(out_child, result);
1624 * The object is on remote node, return its FID for remote open.
1626 if (mdt_object_remote(child)) {
1628 * Check if this lock already was sent to client and
1629 * this is resent case. For resent case do not take lock
1630 * again, use what is already granted.
1632 LASSERT(lhc != NULL);
1634 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1635 struct ldlm_lock *lock;
1637 LASSERT(msg_flags & MSG_RESENT);
1639 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1641 CERROR("Invalid lock handle "LPX64"\n",
1642 lhc->mlh_reg_lh.cookie);
1645 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1646 &lock->l_resource->lr_name));
1647 LDLM_LOCK_PUT(lock);
1650 mdt_lock_handle_init(lhc);
1651 mdt_lock_reg_init(lhc, LCK_PR);
1653 rc = mdt_object_lock(info, child, lhc,
1654 MDS_INODELOCK_LOOKUP,
1657 repbody->fid1 = *mdt_object_fid(child);
1658 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1663 GOTO(out_child, result);
1665 if (mdt_object_exists(child)) {
1666 /* We have to get attr & LOV EA & HSM for this
1668 ma->ma_need |= MA_HSM;
1669 result = mdt_attr_get_complex(info, child, ma);
1671 /*object non-exist!!!*/
1677 LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1679 /* get openlock if this is not replay and if a client requested it */
1680 if (!req_is_replay(req)) {
1681 rc = mdt_object_open_lock(info, child, lhc, &ibits);
1683 GOTO(out_child, result = rc);
1684 } else if (create_flags & MDS_OPEN_LOCK) {
1686 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1690 /* Try to open it now. */
1691 rc = mdt_finish_open(info, parent, child, create_flags,
1695 /* openlock will be released if mdt_finish_open failed */
1696 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1700 ma->ma_cookie_size = 0;
1701 rc = mdo_unlink(info->mti_env,
1702 mdt_object_child(parent),
1703 mdt_object_child(child),
1707 CERROR("Error in cleanup of open\n");
1708 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1713 mdt_object_open_unlock(info, child, lhc, ibits, result);
1714 mdt_object_put(info->mti_env, child);
1716 mdt_object_unlock_put(info, parent, lh, result || !created);
1718 if (result && result != -EREMOTE)
1719 lustre_msg_set_transno(req->rq_repmsg, 0);
1723 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
1724 MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
1726 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1728 return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1731 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1733 struct mdt_object *o = mfd->mfd_object;
1734 struct md_object *next = mdt_object_child(o);
1735 struct md_attr *ma = &info->mti_attr;
1736 int ret = MDT_IOEPOCH_CLOSED;
1741 mode = mfd->mfd_mode;
1743 if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
1745 ret = mdt_ioepoch_close(info, o);
1746 } else if (mode & MDS_FMODE_EXEC) {
1748 } else if (mode & MDS_FMODE_EPOCH) {
1749 ret = mdt_ioepoch_close(info, o);
1750 } else if (mode & MDS_FMODE_SOM) {
1751 ret = mdt_som_au_close(info, o);
1754 /* Update atime on close only. */
1755 if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1756 && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1757 /* Set the atime only. */
1758 ma->ma_valid = MA_INODE;
1759 ma->ma_attr.la_valid = LA_ATIME;
1760 rc = mo_attr_set(info->mti_env, next, ma);
1763 /* If file data is modified, add the dirty flag.
1765 * If MDS_CLOSE_CLEANUP is set, this file is being closed due to an
1766 * eviction, file could have been modified and now dirty
1767 * regarding to HSM archive, check this!
1768 * The logic here is to mark a file dirty if there's a chance it was
1769 * dirtied before the client was evicted, so that we don't have to wait
1770 * for a release attempt before finding out the file was actually dirty
1771 * and fail the release. Aggressively marking it dirty here will cause
1772 * the policy engine to attempt to re-archive it; when rearchiving, we
1773 * can compare the current version to the LMA data_version and make the
1774 * archive request into a noop if it's not actually dirty.
1776 if ((ma->ma_attr_flags & MDS_DATA_MODIFIED) ||
1777 ((ma->ma_attr_flags & MDS_CLOSE_CLEANUP) &&
1778 (mode & (FMODE_WRITE|MDS_FMODE_TRUNC))))
1779 rc = mdt_add_dirty_flag(info, o, ma);
1781 ma->ma_need |= MA_INODE;
1782 ma->ma_valid &= ~MA_INODE;
1784 if (!MFD_CLOSED(mode))
1785 rc = mo_close(info->mti_env, next, ma, mode);
1787 if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
1788 struct mdt_export_data *med;
1790 /* The IOepoch is still opened or SOM update is needed.
1791 * Put mfd back into the list. */
1792 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
1793 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
1794 MDS_FMODE_EPOCH : MDS_FMODE_SOM);
1796 LASSERT(mdt_info_req(info));
1797 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
1798 spin_lock(&med->med_open_lock);
1799 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
1800 class_handle_hash_back(&mfd->mfd_handle);
1801 spin_unlock(&med->med_open_lock);
1803 if (ret == MDT_IOEPOCH_OPENED) {
1807 CDEBUG(D_INODE, "Size-on-MDS attribute update is "
1808 "needed on "DFID"\n", PFID(mdt_object_fid(o)));
1812 mdt_object_put(info->mti_env, o);
1815 RETURN(rc ? rc : ret);
1818 int mdt_close(struct mdt_thread_info *info)
1820 struct mdt_export_data *med;
1821 struct mdt_file_data *mfd;
1822 struct mdt_object *o;
1823 struct md_attr *ma = &info->mti_attr;
1824 struct mdt_body *repbody = NULL;
1825 struct ptlrpc_request *req = mdt_info_req(info);
1829 mdt_counter_incr(req, LPROC_MDT_CLOSE);
1830 /* Close may come with the Size-on-MDS update. Unpack it. */
1831 rc = mdt_close_unpack(info);
1833 RETURN(err_serious(rc));
1835 LASSERT(info->mti_ioepoch);
1837 req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
1838 info->mti_mdt->mdt_max_mdsize);
1839 req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
1840 info->mti_mdt->mdt_max_cookiesize);
1841 rc = req_capsule_server_pack(info->mti_pill);
1842 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
1843 mdt_client_compatibility(info);
1845 mdt_fix_reply(info);
1846 mdt_exit_ucred(info);
1847 RETURN(lustre_msg_get_status(req->rq_repmsg));
1850 /* Continue to close handle even if we can not pack reply */
1852 repbody = req_capsule_server_get(info->mti_pill,
1854 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
1856 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
1859 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
1861 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
1864 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
1865 repbody->eadatasize = 0;
1866 repbody->aclsize = 0;
1868 rc = err_serious(rc);
1871 med = &req->rq_export->exp_mdt_data;
1872 spin_lock(&med->med_open_lock);
1873 mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1874 if (mdt_mfd_closed(mfd)) {
1875 spin_unlock(&med->med_open_lock);
1876 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
1877 ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
1878 info->mti_ioepoch->handle.cookie);
1879 /** not serious error since bug 3633 */
1882 class_handle_unhash(&mfd->mfd_handle);
1883 cfs_list_del_init(&mfd->mfd_list);
1884 spin_unlock(&med->med_open_lock);
1886 /* Do not lose object before last unlink. */
1887 o = mfd->mfd_object;
1888 mdt_object_get(info->mti_env, o);
1889 ret = mdt_mfd_close(info, mfd);
1890 if (repbody != NULL)
1891 rc = mdt_handle_last_unlink(info, o, ma);
1892 mdt_empty_transno(info, rc);
1893 mdt_object_put(info->mti_env, o);
1895 if (repbody != NULL) {
1896 mdt_client_compatibility(info);
1897 rc = mdt_fix_reply(info);
1900 mdt_exit_ucred(info);
1901 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
1902 RETURN(err_serious(-ENOMEM));
1904 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
1905 OBD_FAIL_MDS_CLOSE_NET_REP))
1906 info->mti_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
1907 RETURN(rc ? rc : ret);
1911 * DONE_WRITING rpc handler.
1913 * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
1914 * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
1915 * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
1916 * skip attributes and reconstruct the reply here.
1918 int mdt_done_writing(struct mdt_thread_info *info)
1920 struct ptlrpc_request *req = mdt_info_req(info);
1921 struct mdt_body *repbody = NULL;
1922 struct mdt_export_data *med;
1923 struct mdt_file_data *mfd;
1927 rc = req_capsule_server_pack(info->mti_pill);
1929 RETURN(err_serious(rc));
1931 repbody = req_capsule_server_get(info->mti_pill,
1933 repbody->eadatasize = 0;
1934 repbody->aclsize = 0;
1936 /* Done Writing may come with the Size-on-MDS update. Unpack it. */
1937 rc = mdt_close_unpack(info);
1939 RETURN(err_serious(rc));
1941 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
1942 mdt_exit_ucred(info);
1943 RETURN(lustre_msg_get_status(req->rq_repmsg));
1946 med = &info->mti_exp->exp_mdt_data;
1947 spin_lock(&med->med_open_lock);
1948 mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1950 spin_unlock(&med->med_open_lock);
1951 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
1952 ": cookie = "LPX64" ioepoch = "LPU64"\n",
1953 PFID(info->mti_rr.rr_fid1),
1954 info->mti_ioepoch->handle.cookie,
1955 info->mti_ioepoch->ioepoch);
1956 /* If this is a replay, reconstruct the transno. */
1957 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1958 rc = info->mti_ioepoch->flags & MF_SOM_AU ?
1960 mdt_empty_transno(info, rc);
1963 GOTO(error_ucred, rc);
1966 LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
1967 mfd->mfd_mode == MDS_FMODE_TRUNC);
1968 class_handle_unhash(&mfd->mfd_handle);
1969 cfs_list_del_init(&mfd->mfd_list);
1970 spin_unlock(&med->med_open_lock);
1972 /* Set EPOCH CLOSE flag if not set by client. */
1973 info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
1974 info->mti_attr.ma_valid = 0;
1976 info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
1977 OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1978 if (info->mti_attr.ma_lmm == NULL)
1979 GOTO(error_ucred, rc = -ENOMEM);
1981 rc = mdt_mfd_close(info, mfd);
1983 OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1984 mdt_empty_transno(info, rc);
1986 mdt_exit_ucred(info);