4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_open.c
38 * Lustre Metadata Target (mdt) open/close file handling
40 * Author: Huang Hua <huanghua@clusterfs.com>
43 #define DEBUG_SUBSYSTEM S_MDS
45 #include <lustre_acl.h>
46 #include <lustre_mds.h>
47 #include "mdt_internal.h"
49 /* we do nothing because we do not have refcount now */
50 static void mdt_mfd_get(void *mfdp)
54 static struct portals_handle_ops mfd_handle_ops = {
55 .hop_addref = mdt_mfd_get,
59 /* Create a new mdt_file_data struct, initialize it,
60 * and insert it to global hash table */
61 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
63 struct mdt_file_data *mfd;
68 CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
69 mfd->mfd_handle.h_owner = med;
70 CFS_INIT_LIST_HEAD(&mfd->mfd_list);
71 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
78 * Find the mfd pointed to by handle in global hash table.
79 * In case of replay the handle is obsoleted
80 * but mfd can be found in mfd list by that handle.
81 * Callers need to be holding med_open_lock.
83 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
84 const struct lustre_handle *handle,
85 bool is_replay_or_resent)
87 struct mdt_file_data *mfd;
90 LASSERT(handle != NULL);
91 mfd = class_handle2object(handle->cookie, med);
92 /* during dw/setattr replay the mfd can be found by old handle */
93 if (mfd == NULL && is_replay_or_resent) {
94 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
95 if (mfd->mfd_old_handle.cookie == handle->cookie)
105 void mdt_mfd_free(struct mdt_file_data *mfd)
107 LASSERT(cfs_list_empty(&mfd->mfd_list));
108 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
111 static int mdt_create_data(struct mdt_thread_info *info,
112 struct mdt_object *p, struct mdt_object *o)
114 struct md_op_spec *spec = &info->mti_spec;
115 struct md_attr *ma = &info->mti_attr;
119 if (!md_should_create(spec->sp_cr_flags))
122 ma->ma_need = MA_INODE | MA_LOV;
124 mutex_lock(&o->mot_lov_mutex);
125 if (!(o->mot_flags & MOF_LOV_CREATED)) {
126 if (p != NULL && (fid_is_obf(mdt_object_fid(p)) ||
127 fid_is_dot_lustre(mdt_object_fid(p))))
128 GOTO(unlock, rc = -EPERM);
130 rc = mdo_create_data(info->mti_env,
131 p ? mdt_object_child(p) : NULL,
132 mdt_object_child(o), spec, ma);
134 rc = mdt_attr_get_complex(info, o, ma);
136 if (rc == 0 && ma->ma_valid & MA_LOV)
137 o->mot_flags |= MOF_LOV_CREATED;
140 mutex_unlock(&o->mot_lov_mutex);
144 static int mdt_ioepoch_opened(struct mdt_object *mo)
146 return mo->mot_ioepoch_count;
149 int mdt_object_is_som_enabled(struct mdt_object *mo)
151 return !mo->mot_ioepoch;
155 * Re-enable Size-on-MDS.
156 * Call under ->mot_ioepoch_mutex.
158 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
160 if (ioepoch == mo->mot_ioepoch) {
161 LASSERT(!mdt_ioepoch_opened(mo));
168 * Open the IOEpoch. It is allowed if @writecount is not negative.
169 * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
171 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
174 struct mdt_device *mdt = info->mti_mdt;
179 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
180 !S_ISREG(lu_object_attr(&o->mot_obj)))
183 mutex_lock(&o->mot_ioepoch_mutex);
184 if (mdt_ioepoch_opened(o)) {
185 /* Epoch continues even if there is no writers yet. */
186 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
187 o->mot_ioepoch, PFID(mdt_object_fid(o)));
189 /* XXX: ->mdt_ioepoch is not initialized at the mount */
190 spin_lock(&mdt->mdt_ioepoch_lock);
191 if (mdt->mdt_ioepoch < info->mti_replayepoch)
192 mdt->mdt_ioepoch = info->mti_replayepoch;
194 if (info->mti_replayepoch)
195 o->mot_ioepoch = info->mti_replayepoch;
196 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
197 o->mot_ioepoch = ++mdt->mdt_ioepoch;
199 o->mot_ioepoch = mdt->mdt_ioepoch;
201 spin_unlock(&mdt->mdt_ioepoch_lock);
203 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
204 o->mot_ioepoch, PFID(mdt_object_fid(o)));
206 o->mot_flags |= MOF_SOM_CREATED;
209 o->mot_ioepoch_count++;
210 mutex_unlock(&o->mot_ioepoch_mutex);
212 /* Cancel Size-on-MDS attributes cached on clients for the open case.
213 * In the truncate case, see mdt_reint_setattr(). */
214 if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
215 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
216 mdt_lock_reg_init(lh, LCK_EX);
217 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
220 mdt_object_unlock(info, o, lh, 1);
226 * Update SOM on-disk attributes.
227 * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
228 * mountid and attributes. If disabling, clean SOM xattr.
229 * Call under ->mot_ioepoch_mutex.
231 static int mdt_som_attr_set(struct mdt_thread_info *info,
232 struct mdt_object *obj, __u64 ioepoch, bool enable)
234 struct md_object *next = mdt_object_child(obj);
238 CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
239 " on "DFID".\n", enable ? "update" : "disabling",
240 ioepoch, PFID(mdt_object_fid(obj)));
243 struct lu_buf *buf = &info->mti_buf;
244 struct som_attrs *attrs;
245 struct md_attr *ma = &info->mti_attr;
246 struct lu_attr *la = &ma->ma_attr;
247 struct obd_device *obd = info->mti_mdt->mdt_lut.lut_obd;
249 attrs = (struct som_attrs *)info->mti_xattr_buf;
250 CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*attrs));
252 /* pack SOM attributes */
253 memset(attrs, 0, sizeof(*attrs));
254 attrs->som_ioepoch = ioepoch;
255 attrs->som_mountid = obd->u.obt.obt_mount_count;
256 if ((la->la_valid & LA_SIZE) != 0)
257 attrs->som_size = la->la_size;
258 if ((la->la_valid & LA_BLOCKS) != 0)
259 attrs->som_blocks = la->la_blocks;
260 lustre_som_swab(attrs);
262 /* update SOM attributes */
264 buf->lb_len = sizeof(*attrs);
265 rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_SOM, 0);
267 /* delete SOM attributes */
268 rc = mo_xattr_del(info->mti_env, next, XATTR_NAME_SOM);
274 /** Perform the eviction specific actions on ioepoch close. */
275 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
276 struct mdt_object *o)
280 mutex_lock(&o->mot_ioepoch_mutex);
281 CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
282 "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
283 o->mot_ioepoch_count);
284 o->mot_ioepoch_count--;
286 /* If eviction occured set MOF_SOM_RECOV,
287 * if no other epoch holders, disable SOM on disk. */
288 o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
289 if (!mdt_ioepoch_opened(o)) {
290 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
291 mdt_object_som_enable(o, o->mot_ioepoch);
293 mutex_unlock(&o->mot_ioepoch_mutex);
298 * Perform the replay specific actions on ioepoch close.
299 * Skip SOM attribute update if obtained and just forget about the inode state
300 * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
302 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
303 struct mdt_object *o)
305 int rc = MDT_IOEPOCH_CLOSED;
308 mutex_lock(&o->mot_ioepoch_mutex);
309 CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
310 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
311 o->mot_ioepoch_count--;
313 /* Get an info from the replayed request if client is supposed
314 * to send an Attibute Update, reconstruct @rc if so */
315 if (info->mti_ioepoch->flags & MF_SOM_AU)
316 rc = MDT_IOEPOCH_GETATTR;
318 if (!mdt_ioepoch_opened(o))
319 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
320 mutex_unlock(&o->mot_ioepoch_mutex);
326 * Regular file IOepoch close.
327 * Closes the ioepoch, checks the object state, apply obtained attributes and
328 * re-enable SOM on the object, if possible. Also checks if the recovery is
329 * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
330 * to obtain SOM attributes under the server-side OST locks.
333 * MDT_IOEPOCH_CLOSED if ioepoch is closed.
334 * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
336 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
337 struct mdt_object *o)
339 struct md_attr *tmp_ma;
343 int rc = 0, ret = MDT_IOEPOCH_CLOSED;
346 la = &info->mti_attr.ma_attr;
347 achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
349 mutex_lock(&o->mot_ioepoch_mutex);
350 o->mot_ioepoch_count--;
352 tmp_ma = &info->mti_u.som.attr;
353 tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
354 tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
355 tmp_ma->ma_som = &info->mti_u.som.data;
356 tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
357 tmp_ma->ma_valid = 0;
358 rc = mdt_attr_get_complex(info, o, tmp_ma);
362 /* Check the on-disk SOM state. */
363 if (o->mot_flags & MOF_SOM_RECOV)
365 else if (!(o->mot_flags & MOF_SOM_CREATED) &&
366 !(tmp_ma->ma_valid & MA_SOM))
369 CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
370 o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
372 opened = mdt_ioepoch_opened(o);
374 * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
375 * Skip the check for file with no LOV or for unlink files.
377 if (!opened && tmp_ma->ma_valid & MA_LOV &&
378 !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
380 /* If some previous writer was evicted, re-ask the
381 * client for attributes. Even if attributes are
382 * provided, we cannot believe in them.
383 * Another use case is that there is no SOM cache on
384 * disk -- first access with SOM or there was an MDS
386 ret = MDT_IOEPOCH_GETATTR;
387 else if (o->mot_flags & MOF_SOM_CHANGE)
388 /* Some previous writer changed the attribute.
389 * Do not believe to the current Size-on-MDS
390 * update, re-ask client. */
391 ret = MDT_IOEPOCH_GETATTR;
392 else if (!(la->la_valid & LA_SIZE) && achange)
393 /* Attributes were changed by the last writer
394 * only but no Size-on-MDS update is received.*/
395 ret = MDT_IOEPOCH_GETATTR;
398 if (achange || ret == MDT_IOEPOCH_GETATTR)
399 o->mot_flags |= MOF_SOM_CHANGE;
401 /* If epoch ends and relable SOM attributes are obtained, update them.
402 * Create SOM ea for new files even if there is no attributes obtained
403 * (0-length file). */
404 if (ret == MDT_IOEPOCH_CLOSED && !opened) {
405 if (achange || o->mot_flags & MOF_SOM_CREATED) {
406 LASSERT(achange || !(la->la_valid & LA_SIZE));
407 rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
409 /* Avoid the following setattrs of these attributes,
410 * e.g. for atime update. */
411 info->mti_attr.ma_valid = 0;
413 mdt_object_som_enable(o, o->mot_ioepoch);
416 mutex_unlock(&o->mot_ioepoch_mutex);
417 /* If recovery is needed, tell the client to perform GETATTR under
419 if (ret == MDT_IOEPOCH_GETATTR && recovery) {
420 struct mdt_body *rep;
421 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
422 rep->valid |= OBD_MD_FLGETATTRLOCK;
428 mutex_unlock(&o->mot_ioepoch_mutex);
433 * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
434 * - a client closes the IOEpoch;
435 * - a client eviction occured.
437 * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
438 * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
439 * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
442 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
444 struct ptlrpc_request *req = mdt_info_req(info);
447 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
448 !S_ISREG(lu_object_attr(&o->mot_obj)))
451 LASSERT(o->mot_ioepoch_count);
452 LASSERT(info->mti_ioepoch == NULL ||
453 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
455 /* IOEpoch is closed only if client tells about it or eviction occures.
456 * In the replay case, always close the epoch. */
458 RETURN(mdt_ioepoch_close_on_eviction(info, o));
459 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
460 RETURN(mdt_ioepoch_close_on_replay(info, o));
461 if (info->mti_ioepoch && (info->mti_ioepoch->flags & MF_EPOCH_CLOSE))
462 RETURN(mdt_ioepoch_close_reg(info, o));
463 /* IO epoch is not closed. */
464 RETURN(MDT_IOEPOCH_OPENED);
468 * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
469 * for attribute update. It happens if:
470 * - SOM Attribute Update is obtained;
471 * - the client failed to obtain it and informs MDS about it;
472 * - a client eviction occured.
473 * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
476 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
478 struct ptlrpc_request *req = mdt_info_req(info);
480 int act = MDT_SOM_ENABLE;
484 LASSERT(!req || info->mti_ioepoch);
485 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
486 !S_ISREG(lu_object_attr(&o->mot_obj)))
489 /* No size whereas MF_SOM_CHANGE is set means client failed to
490 * obtain ost attributes, drop the SOM cache on disk if so. */
492 (info->mti_ioepoch &&
493 info->mti_ioepoch->flags & MF_SOM_CHANGE &&
494 !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
495 act = MDT_SOM_DISABLE;
497 mutex_lock(&o->mot_ioepoch_mutex);
498 /* Mark the object it is the recovery state if we failed to obtain
500 if (act == MDT_SOM_DISABLE)
501 o->mot_flags |= MOF_SOM_RECOV;
503 if (!mdt_ioepoch_opened(o)) {
504 ioepoch = info->mti_ioepoch ?
505 info->mti_ioepoch->ioepoch : o->mot_ioepoch;
508 && !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
509 rc = mdt_som_attr_set(info, o, ioepoch, act);
510 mdt_object_som_enable(o, ioepoch);
512 mutex_unlock(&o->mot_ioepoch_mutex);
516 int mdt_write_read(struct mdt_object *o)
520 mutex_lock(&o->mot_ioepoch_mutex);
521 rc = o->mot_writecount;
522 mutex_unlock(&o->mot_ioepoch_mutex);
526 int mdt_write_get(struct mdt_object *o)
530 mutex_lock(&o->mot_ioepoch_mutex);
531 if (o->mot_writecount < 0)
535 mutex_unlock(&o->mot_ioepoch_mutex);
539 void mdt_write_put(struct mdt_object *o)
542 mutex_lock(&o->mot_ioepoch_mutex);
544 mutex_unlock(&o->mot_ioepoch_mutex);
548 static int mdt_write_deny(struct mdt_object *o)
552 mutex_lock(&o->mot_ioepoch_mutex);
553 if (o->mot_writecount > 0)
557 mutex_unlock(&o->mot_ioepoch_mutex);
561 static void mdt_write_allow(struct mdt_object *o)
564 mutex_lock(&o->mot_ioepoch_mutex);
566 mutex_unlock(&o->mot_ioepoch_mutex);
570 /* there can be no real transaction so prepare the fake one */
571 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
573 struct mdt_device *mdt = info->mti_mdt;
574 struct ptlrpc_request *req = mdt_info_req(info);
575 struct tg_export_data *ted;
576 struct lsd_client_data *lcd;
579 /* transaction has occurred already */
580 if (lustre_msg_get_transno(req->rq_repmsg) != 0)
583 spin_lock(&mdt->mdt_lut.lut_translock);
585 if (info->mti_transno != 0) {
586 struct obd_export *exp = req->rq_export;
588 CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
589 mdt_obd_name(mdt), info->mti_transno,
590 libcfs_nid2str(exp->exp_connection->c_peer.nid),
592 spin_unlock(&mdt->mdt_lut.lut_translock);
595 } else if (info->mti_transno == 0) {
596 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
598 /* should be replay */
599 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
600 mdt->mdt_lut.lut_last_transno = info->mti_transno;
602 spin_unlock(&mdt->mdt_lut.lut_translock);
604 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
606 req->rq_export->exp_obd->obd_last_committed);
608 req->rq_transno = info->mti_transno;
609 lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
611 /* update lcd in memory only for resent cases */
612 ted = &req->rq_export->exp_target_data;
614 mutex_lock(&ted->ted_lcd_lock);
616 if (info->mti_transno < lcd->lcd_last_transno &&
617 info->mti_transno != 0) {
618 /* This should happen during replay. Do not update
619 * last rcvd info if replay req transno < last transno,
620 * otherwise the following resend(after replay) can not
621 * be checked correctly by xid */
622 mutex_unlock(&ted->ted_lcd_lock);
623 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
624 mdt_obd_name(mdt), info->mti_transno,
625 lcd->lcd_last_transno);
629 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
630 lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
631 if (info->mti_transno != 0)
632 lcd->lcd_last_close_transno = info->mti_transno;
633 lcd->lcd_last_close_xid = req->rq_xid;
634 lcd->lcd_last_close_result = rc;
636 /* VBR: save versions in last_rcvd for reconstruct. */
637 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
639 lcd->lcd_pre_versions[0] = pre_versions[0];
640 lcd->lcd_pre_versions[1] = pre_versions[1];
641 lcd->lcd_pre_versions[2] = pre_versions[2];
642 lcd->lcd_pre_versions[3] = pre_versions[3];
644 if (info->mti_transno != 0)
645 lcd->lcd_last_transno = info->mti_transno;
647 lcd->lcd_last_xid = req->rq_xid;
648 lcd->lcd_last_result = rc;
649 lcd->lcd_last_data = info->mti_opdata;
651 mutex_unlock(&ted->ted_lcd_lock);
656 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
658 LASSERT(mfd != NULL);
660 CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n",
661 PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
663 mfd->mfd_mode = mode;
666 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
667 struct mdt_object *o, __u64 flags, int created,
668 struct ldlm_reply *rep)
670 struct ptlrpc_request *req = mdt_info_req(info);
671 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
672 struct mdt_file_data *mfd;
673 struct md_attr *ma = &info->mti_attr;
674 struct lu_attr *la = &ma->ma_attr;
675 struct mdt_body *repbody;
676 int rc = 0, isdir, isreg;
679 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
681 isreg = S_ISREG(la->la_mode);
682 isdir = S_ISDIR(la->la_mode);
683 if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
685 * No EA, check whether it is will set regEA and dirEA since in
686 * above attr get, these size might be zero, so reset it, to
687 * retrieve the MD after create obj.
689 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
692 /* in replay case, p == NULL */
693 rc = mdt_create_data(info, p, o);
697 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
698 mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
701 CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
702 ma->ma_valid, ma->ma_lmm_size);
704 if (ma->ma_valid & MA_LOV) {
705 LASSERT(ma->ma_lmm_size != 0);
706 repbody->eadatasize = ma->ma_lmm_size;
708 repbody->valid |= OBD_MD_FLDIREA;
710 repbody->valid |= OBD_MD_FLEASIZE;
713 if (flags & FMODE_WRITE) {
714 rc = mdt_write_get(o);
716 mdt_ioepoch_open(info, o, created);
717 repbody->ioepoch = o->mot_ioepoch;
719 } else if (flags & MDS_FMODE_EXEC) {
720 rc = mdt_write_deny(o);
725 rc = mo_open(info->mti_env, mdt_object_child(o),
726 created ? flags | MDS_OPEN_CREATED : flags);
730 mfd = mdt_mfd_new(med);
732 GOTO(err_out, rc = -ENOMEM);
735 * Keep a reference on this object for this open, and is
736 * released by mdt_mfd_close().
738 mdt_object_get(info->mti_env, o);
740 mfd->mfd_xid = req->rq_xid;
743 * @flags is always not zero. At least it should be FMODE_READ,
744 * FMODE_WRITE or MDS_FMODE_EXEC.
749 mdt_mfd_set_mode(mfd, flags);
751 atomic_inc(&o->mot_open_count);
752 if (flags & MDS_OPEN_LEASE)
753 atomic_inc(&o->mot_lease_count);
756 if (req_is_replay(req)) {
757 struct mdt_file_data *old_mfd;
758 /* Check wheather old cookie already exist in
759 * the list, becasue when do recovery, client
760 * might be disconnected from server, and
761 * restart replay, so there maybe some orphan
762 * mfd here, we should remove them */
763 LASSERT(info->mti_rr.rr_handle != NULL);
764 spin_lock(&med->med_open_lock);
765 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
766 if (old_mfd != NULL) {
767 CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
768 "cookie = "LPX64"\n", mfd,
769 PFID(mdt_object_fid(mfd->mfd_object)),
770 info->mti_rr.rr_handle->cookie);
771 class_handle_unhash(&old_mfd->mfd_handle);
772 cfs_list_del_init(&old_mfd->mfd_list);
773 spin_unlock(&med->med_open_lock);
774 /* no attr update for that close */
776 ma->ma_valid |= MA_FLAGS;
777 ma->ma_attr_flags |= MDS_RECOV_OPEN;
778 mdt_mfd_close(info, old_mfd);
779 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
780 ma->ma_valid &= ~MA_FLAGS;
782 spin_unlock(&med->med_open_lock);
783 CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
784 "cookie = "LPX64"\n",
785 PFID(mdt_object_fid(mfd->mfd_object)),
786 info->mti_rr.rr_handle->cookie);
789 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
790 info->mti_rr.rr_handle->cookie);
792 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
795 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
797 if (req->rq_export->exp_disconnected) {
798 spin_lock(&med->med_open_lock);
799 class_handle_unhash(&mfd->mfd_handle);
800 cfs_list_del_init(&mfd->mfd_list);
801 spin_unlock(&med->med_open_lock);
802 mdt_mfd_close(info, mfd);
804 spin_lock(&med->med_open_lock);
805 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
806 spin_unlock(&med->med_open_lock);
809 mdt_empty_transno(info, rc);
814 if (flags & FMODE_WRITE)
815 /* XXX We also need to close io epoch here.
816 * See LU-1220 - green */
818 else if (flags & FMODE_EXEC)
823 int mdt_finish_open(struct mdt_thread_info *info,
824 struct mdt_object *p, struct mdt_object *o,
825 __u64 flags, int created, struct ldlm_reply *rep)
827 struct ptlrpc_request *req = mdt_info_req(info);
828 struct obd_export *exp = req->rq_export;
829 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
830 struct md_attr *ma = &info->mti_attr;
831 struct lu_attr *la = &ma->ma_attr;
832 struct mdt_file_data *mfd;
833 struct mdt_body *repbody;
835 int isreg, isdir, islnk;
839 LASSERT(ma->ma_valid & MA_INODE);
841 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
843 isreg = S_ISREG(la->la_mode);
844 isdir = S_ISDIR(la->la_mode);
845 islnk = S_ISLNK(la->la_mode);
846 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
848 /* LU-2275, simulate broken behaviour (esp. prevalent in
849 * pre-2.4 servers where a very strange reply is sent on error
850 * that looks like it was actually almost succesful and a failure at the
852 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
853 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
857 if (flags & MDS_OPEN_LOCK)
858 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
863 if (exp_connect_rmtclient(exp)) {
864 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
866 rc = mdt_pack_remote_perm(info, o, buf);
868 repbody->valid &= ~OBD_MD_FLRMTPERM;
869 repbody->aclsize = 0;
871 repbody->valid |= OBD_MD_FLRMTPERM;
872 repbody->aclsize = sizeof(struct mdt_remote_perm);
875 #ifdef CONFIG_FS_POSIX_ACL
876 else if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
877 const struct lu_env *env = info->mti_env;
878 struct md_object *next = mdt_object_child(o);
879 struct lu_buf *buf = &info->mti_buf;
881 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
882 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
884 if (buf->lb_len > 0) {
885 rc = mo_xattr_get(env, next, buf,
886 XATTR_NAME_ACL_ACCESS);
888 if (rc == -ENODATA) {
889 repbody->aclsize = 0;
890 repbody->valid |= OBD_MD_FLACL;
892 } else if (rc == -EOPNOTSUPP) {
895 CERROR("got acl size: %d\n", rc);
898 repbody->aclsize = rc;
899 repbody->valid |= OBD_MD_FLACL;
906 if (info->mti_mdt->mdt_lut.lut_mds_capa &&
907 exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) {
908 struct lustre_capa *capa;
910 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
912 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
913 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
916 repbody->valid |= OBD_MD_FLMDSCAPA;
918 if (info->mti_mdt->mdt_lut.lut_oss_capa &&
919 exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA &&
920 S_ISREG(lu_object_attr(&o->mot_obj))) {
921 struct lustre_capa *capa;
923 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
925 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
926 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
929 repbody->valid |= OBD_MD_FLOSSCAPA;
933 * If we are following a symlink, don't open; and do not return open
934 * handle for special nodes as client required.
936 if (islnk || (!isreg && !isdir &&
937 (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
938 lustre_msg_set_transno(req->rq_repmsg, 0);
943 * We need to return the existing object's fid back, so it is done here,
944 * after preparing the reply.
946 if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
949 /* This can't be done earlier, we need to return reply body */
951 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
952 /* We are trying to create or write an existing dir. */
955 } else if (flags & MDS_OPEN_DIRECTORY)
958 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
959 OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
964 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
965 spin_lock(&med->med_open_lock);
966 cfs_list_for_each(t, &med->med_open_head) {
967 mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
968 if (mfd->mfd_xid == req->rq_xid)
972 spin_unlock(&med->med_open_lock);
975 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
976 /*set repbody->ea_size for resent case*/
977 if (ma->ma_valid & MA_LOV) {
978 LASSERT(ma->ma_lmm_size != 0);
979 repbody->eadatasize = ma->ma_lmm_size;
981 repbody->valid |= OBD_MD_FLDIREA;
983 repbody->valid |= OBD_MD_FLEASIZE;
985 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
990 rc = mdt_mfd_open(info, p, o, flags, created, rep);
992 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
997 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
998 struct lsd_client_data *lcd);
1000 void mdt_reconstruct_open(struct mdt_thread_info *info,
1001 struct mdt_lock_handle *lhc)
1003 const struct lu_env *env = info->mti_env;
1004 struct mdt_device *mdt = info->mti_mdt;
1005 struct req_capsule *pill = info->mti_pill;
1006 struct ptlrpc_request *req = mdt_info_req(info);
1007 struct tg_export_data *ted = &req->rq_export->exp_target_data;
1008 struct lsd_client_data *lcd = ted->ted_lcd;
1009 struct md_attr *ma = &info->mti_attr;
1010 struct mdt_reint_record *rr = &info->mti_rr;
1011 __u64 flags = info->mti_spec.sp_cr_flags;
1012 struct ldlm_reply *ldlm_rep;
1013 struct mdt_object *parent;
1014 struct mdt_object *child;
1015 struct mdt_body *repbody;
1019 LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1020 ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1021 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1023 ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
1024 ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
1026 ma->ma_need = MA_INODE | MA_HSM;
1027 if (ma->ma_lmm_size > 0)
1028 ma->ma_need |= MA_LOV;
1032 mdt_req_from_lcd(req, lcd);
1033 mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1035 CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1036 ldlm_rep->lock_policy_res1, req->rq_status);
1038 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1039 req->rq_status != 0)
1040 /* We did not create successfully, return error to client. */
1041 GOTO(out, rc = req->rq_status);
1043 if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1044 struct obd_export *exp = req->rq_export;
1046 * We failed after creation, but we do not know in which step
1047 * we failed. So try to check the child object.
1049 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1050 if (IS_ERR(parent)) {
1051 rc = PTR_ERR(parent);
1052 LCONSOLE_WARN("Parent "DFID" lookup error %d."
1053 " Evicting client %s with export %s.\n",
1054 PFID(rr->rr_fid1), rc,
1055 obd_uuid2str(&exp->exp_client_uuid),
1056 obd_export_nid2str(exp));
1057 mdt_export_evict(exp);
1060 child = mdt_object_find(env, mdt, rr->rr_fid2);
1061 if (IS_ERR(child)) {
1062 rc = PTR_ERR(child);
1063 LCONSOLE_WARN("Child "DFID" lookup error %d."
1064 " Evicting client %s with export %s.\n",
1065 PFID(mdt_object_fid(child)), rc,
1066 obd_uuid2str(&exp->exp_client_uuid),
1067 obd_export_nid2str(exp));
1068 mdt_object_put(env, parent);
1069 mdt_export_evict(exp);
1073 if (unlikely(mdt_object_remote(child))) {
1074 /* the child object was created on remote server */
1075 if (!mdt_is_dne_client(exp)) {
1076 /* Return -EIO for old client */
1077 mdt_object_put(env, parent);
1078 mdt_object_put(env, child);
1079 GOTO(out, rc = -EIO);
1081 repbody->fid1 = *rr->rr_fid2;
1082 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1085 if (mdt_object_exists(child)) {
1086 mdt_set_capainfo(info, 1, rr->rr_fid2,
1088 rc = mdt_attr_get_complex(info, child, ma);
1090 rc = mdt_finish_open(info, parent,
1094 /* the child does not exist, we should do
1096 mdt_object_put(env, parent);
1097 mdt_object_put(env, child);
1098 GOTO(regular_open, 0);
1101 mdt_object_put(env, parent);
1102 mdt_object_put(env, child);
1106 /* We did not try to create, so we are a pure open */
1107 rc = mdt_reint_open(info, lhc);
1112 req->rq_status = rc;
1113 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1114 LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1117 int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
1119 __u64 flags = info->mti_spec.sp_cr_flags;
1120 struct mdt_reint_record *rr = &info->mti_rr;
1121 struct md_attr *ma = &info->mti_attr;
1122 struct mdt_object *o;
1126 o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1128 RETURN(rc = PTR_ERR(o));
1130 if (unlikely(mdt_object_remote(o))) {
1131 /* the child object was created on remote server */
1132 struct mdt_body *repbody;
1134 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1137 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1138 repbody->fid1 = *rr->rr_fid2;
1139 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1142 if (mdt_object_exists(o)) {
1143 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1147 rc = mdt_attr_get_complex(info, o, ma);
1149 rc = mdt_finish_open(info, NULL, o, flags, 0,
1156 mdt_object_put(info->mti_env, o);
1160 /* lock object for open */
1161 static int mdt_object_open_lock(struct mdt_thread_info *info,
1162 struct mdt_object *obj,
1163 struct mdt_lock_handle *lhc,
1166 struct md_attr *ma = &info->mti_attr;
1167 __u64 open_flags = info->mti_spec.sp_cr_flags;
1168 ldlm_mode_t lm = LCK_CR;
1169 bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
1170 bool try_layout = false;
1171 bool create_layout = false;
1176 mdt_lock_handle_init(lhc);
1178 if (req_is_replay(mdt_info_req(info)))
1181 if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
1182 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
1183 md_should_create(open_flags))
1184 create_layout = true;
1185 if (exp_connect_layout(info->mti_exp) && !create_layout &&
1186 ma->ma_need & MA_LOV)
1191 /* lease open, acquire write mode of open sem */
1192 down_write(&obj->mot_open_sem);
1194 /* Lease exists and ask for new lease */
1195 if (atomic_read(&obj->mot_lease_count) > 0) {
1196 /* only exclusive open is supported, so lease
1197 * are conflicted to each other */
1198 GOTO(out, rc = -EBUSY);
1201 /* Lease must be with open lock */
1202 if (!(open_flags & MDS_OPEN_LOCK)) {
1203 CERROR("Request lease for file:"DFID ", but open lock "
1204 "is missed, open_flags = "LPO64".\n",
1205 PFID(mdt_object_fid(obj)), open_flags);
1206 GOTO(out, rc = -EPROTO);
1209 /* XXX: only exclusive open is supported. */
1211 *ibits = MDS_INODELOCK_OPEN;
1213 /* never grant LCK_EX layout lock to client */
1215 } else { /* normal open */
1216 /* normal open holds read mode of open sem */
1217 down_read(&obj->mot_open_sem);
1219 if (open_flags & MDS_OPEN_LOCK) {
1220 if (open_flags & FMODE_WRITE)
1222 else if (open_flags & MDS_FMODE_EXEC)
1227 *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
1228 } else if (atomic_read(&obj->mot_lease_count) > 0) {
1229 if (open_flags & FMODE_WRITE)
1235 *ibits = MDS_INODELOCK_OPEN;
1238 lhc = &info->mti_lh[MDT_LH_LOCAL];
1240 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
1241 PFID(mdt_object_fid(obj)),
1242 atomic_read(&obj->mot_open_count), lm);
1245 mdt_lock_reg_init(lhc, lm);
1247 /* one problem to return layout lock on open is that it may result
1248 * in too many layout locks cached on the client side. */
1249 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
1250 /* return lookup lock to validate inode at the client side,
1251 * this is pretty important otherwise mdt will return layout
1252 * lock for each open.
1253 * However this is a double-edged sword because changing
1254 * permission will revoke huge # of LOOKUP locks. */
1255 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
1256 if (!mdt_object_lock_try(info, obj, lhc, *ibits,
1258 *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
1260 rc = mdt_object_lock(info, obj, lhc, *ibits,
1263 } else if (*ibits != 0) {
1264 rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK);
1267 CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
1268 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
1269 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
1271 /* will change layout, revoke layout locks by enqueuing EX lock. */
1272 if (rc == 0 && create_layout) {
1273 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
1275 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
1276 ", open_flags = "LPO64"\n",
1277 PFID(mdt_object_fid(obj)), open_flags);
1279 LASSERT(!try_layout);
1280 mdt_lock_handle_init(ll);
1281 mdt_lock_reg_init(ll, LCK_EX);
1282 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
1285 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
1288 /* Check if there is any other open handles after acquiring
1289 * open lock. At this point, caching open handles have been revoked
1291 * XXX: Now only exclusive open is supported. Need to check the
1292 * type of open for generic lease support. */
1293 if (rc == 0 && acq_lease) {
1294 struct ptlrpc_request *req = mdt_info_req(info);
1295 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1296 struct mdt_file_data *mfd;
1297 bool is_replay_or_resent;
1300 /* For lease: application can open a file and then apply lease,
1301 * @handle contains original open handle in that case.
1302 * In recovery, open REQ will be replayed and the lease REQ may
1303 * be resent that means the open handle is already stale, so we
1304 * need to fix it up here by finding new handle. */
1305 is_replay_or_resent = req_is_replay(req) ||
1306 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
1308 /* if the request is _not_ a replay request, rr_handle
1309 * may be used to hold an openhandle which is issuing the
1310 * lease request, so that this openhandle doesn't count. */
1311 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
1312 is_replay_or_resent);
1316 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
1317 PFID(mdt_object_fid(obj)),
1318 atomic_read(&obj->mot_open_count), open_count);
1320 if (atomic_read(&obj->mot_open_count) > open_count)
1321 GOTO(out, rc = -EBUSY);
1329 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1330 struct mdt_object *obj,
1331 struct mdt_lock_handle *lhc,
1332 __u64 ibits, int rc)
1334 __u64 open_flags = info->mti_spec.sp_cr_flags;
1335 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1338 if (req_is_replay(mdt_info_req(info)))
1341 /* Release local lock - the lock put in MDT_LH_LOCAL will never
1342 * return to client side. */
1343 if (lustre_handle_is_used(&ll->mlh_reg_lh))
1344 mdt_object_unlock(info, obj, ll, 1);
1346 ll = &info->mti_lh[MDT_LH_LAYOUT];
1347 /* Release local layout lock, layout was created */
1348 if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1349 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1350 mdt_object_unlock(info, obj, ll, 1);
1353 if (open_flags & MDS_OPEN_LEASE)
1354 up_write(&obj->mot_open_sem);
1356 up_read(&obj->mot_open_sem);
1358 /* Cross-ref case, the lock should be returned to the client */
1359 if (ibits == 0 || rc == -EREMOTE)
1362 if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
1363 /* for the open request, the lock will only return to client
1364 * if open or layout lock is granted. */
1369 struct ldlm_reply *ldlm_rep;
1371 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1372 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1373 mdt_object_unlock(info, obj, lhc, 1);
1379 * Check release is permitted for the current HSM flags.
1381 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1383 if (!(ma->ma_valid & MA_HSM))
1386 if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1389 if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1395 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1396 struct mdt_lock_handle *lhc)
1398 const struct lu_env *env = info->mti_env;
1399 struct mdt_device *mdt = info->mti_mdt;
1400 __u64 flags = info->mti_spec.sp_cr_flags;
1401 struct mdt_reint_record *rr = &info->mti_rr;
1402 struct md_attr *ma = &info->mti_attr;
1403 struct mdt_object *parent= NULL;
1404 struct mdt_object *o;
1409 if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1410 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1411 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1412 if (IS_ERR(parent)) {
1413 CDEBUG(D_INODE, "Fail to find parent "DFID
1414 " for anonymous created %ld, try to"
1415 " use server-side parent.\n",
1416 PFID(rr->rr_fid1), PTR_ERR(parent));
1421 ma->ma_need |= MA_PFID;
1424 o = mdt_object_find(env, mdt, rr->rr_fid2);
1426 RETURN(rc = PTR_ERR(o));
1428 if (mdt_object_remote(o)) {
1429 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1430 mdt_obd_name(info->mti_mdt),
1432 GOTO(out, rc = -EREMOTE);
1433 } else if (!mdt_object_exists(o)) {
1434 mdt_set_disposition(info, rep,
1438 GOTO(out, rc = -ENOENT);
1441 mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1443 if (flags & MDS_OPEN_RELEASE)
1444 ma->ma_need |= MA_HSM;
1445 rc = mdt_attr_get_complex(info, o, ma);
1449 /* If a release request, check file flags are fine and ask for an
1450 * exclusive open access. */
1451 if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1452 GOTO(out, rc = -EPERM);
1454 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1456 GOTO(out_unlock, rc);
1458 if (ma->ma_valid & MA_PFID) {
1459 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1460 if (IS_ERR(parent)) {
1461 CDEBUG(D_INODE, "Fail to find parent "DFID
1462 " for anonymous created %ld, try to"
1463 " use system default.\n",
1464 PFID(&ma->ma_pfid), PTR_ERR(parent));
1469 rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1471 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1472 if (flags & MDS_OPEN_LOCK)
1473 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1474 if (flags & MDS_OPEN_LEASE)
1475 mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1477 GOTO(out_unlock, rc);
1480 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1482 mdt_object_put(env, o);
1484 mdt_object_put(env, parent);
1488 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1489 static int mdt_cross_open(struct mdt_thread_info *info,
1490 const struct lu_fid *parent_fid,
1491 const struct lu_fid *fid,
1492 struct ldlm_reply *rep, __u32 flags)
1494 struct md_attr *ma = &info->mti_attr;
1495 struct mdt_object *o;
1499 o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1501 RETURN(rc = PTR_ERR(o));
1503 if (mdt_object_remote(o)) {
1504 /* Something is wrong here, the object is on another MDS! */
1505 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1506 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1507 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1509 "Object isn't on this server! FLD error?\n");
1512 if (mdt_object_exists(o)) {
1513 /* Do permission check for cross-open. */
1514 rc = mo_permission(info->mti_env, NULL,
1515 mdt_object_child(o),
1516 NULL, flags | MDS_OPEN_CROSS);
1520 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1521 rc = mdt_attr_get_complex(info, o, ma);
1525 /* Do not create lov object if the fid is opened
1527 if (S_ISREG(ma->ma_attr.la_mode) &&
1528 !(ma->ma_valid & MA_LOV) && (flags & FMODE_WRITE) &&
1529 fid_is_obf(parent_fid))
1530 GOTO(out, rc = -EPERM);
1532 rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1535 * Something is wrong here. lookup was positive but
1536 * there is no object!
1538 CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1539 mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1544 mdt_object_put(info->mti_env, o);
1548 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1550 struct mdt_device *mdt = info->mti_mdt;
1551 struct ptlrpc_request *req = mdt_info_req(info);
1552 struct mdt_object *parent;
1553 struct mdt_object *child;
1554 struct mdt_lock_handle *lh;
1555 struct ldlm_reply *ldlm_rep;
1556 struct mdt_body *repbody;
1557 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1558 struct md_attr *ma = &info->mti_attr;
1559 __u64 create_flags = info->mti_spec.sp_cr_flags;
1561 struct mdt_reint_record *rr = &info->mti_rr;
1562 struct lu_name *lname;
1568 OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1569 (obd_timeout + 1) / 4);
1571 mdt_counter_incr(req, LPROC_MDT_OPEN);
1572 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1574 ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1575 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1577 ma->ma_need = MA_INODE;
1578 if (ma->ma_lmm_size > 0)
1579 ma->ma_need |= MA_LOV;
1583 LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1584 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1586 if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1587 CERROR("file join is not supported anymore.\n");
1588 GOTO(out, result = err_serious(-EOPNOTSUPP));
1590 msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1592 if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1593 info->mti_spec.u.sp_ea.eadata == NULL)
1594 GOTO(out, result = err_serious(-EINVAL));
1596 CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1597 "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1598 PFID(rr->rr_fid1), rr->rr_name,
1599 PFID(rr->rr_fid2), create_flags,
1600 ma->ma_attr.la_mode, msg_flags);
1601 if (info->mti_cross_ref) {
1602 /* This is cross-ref open */
1603 mdt_set_disposition(info, ldlm_rep,
1604 (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1606 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1607 ldlm_rep, create_flags);
1609 } else if (req_is_replay(req) ||
1610 (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1611 /* This is a replay request or from liblustre with ea. */
1612 result = mdt_open_by_fid(info, ldlm_rep);
1614 if (result != -ENOENT) {
1615 if (req->rq_export->exp_libclient &&
1616 create_flags & MDS_OPEN_HAS_EA)
1617 GOTO(out, result = 0);
1620 /* We didn't find the correct object, so we need to re-create it
1621 * via a regular replay. */
1622 if (!(create_flags & MDS_OPEN_CREAT)) {
1623 DEBUG_REQ(D_ERROR, req,
1624 "OPEN & CREAT not in open replay/by_fid.");
1625 GOTO(out, result = -EFAULT);
1627 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1628 } else if ((rr->rr_namelen == 0 && create_flags & MDS_OPEN_LOCK) ||
1629 (create_flags & MDS_OPEN_BY_FID)) {
1630 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1631 /* If result is 0 then open by FID has found the file
1632 * and there is nothing left for us to do here. More
1633 * generally if it is anything other than -ENOENT or
1634 * -EREMOTE then we return that now. If -ENOENT and
1635 * MDS_OPEN_CREAT is set then we must create the file
1636 * below. If -EREMOTE then we need to return a LOOKUP
1637 * lock to the client, which we do below. Hence this
1638 * odd looking condition. See LU-2523. */
1639 if (!(result == -ENOENT && (create_flags & MDS_OPEN_CREAT)) &&
1643 if (unlikely(rr->rr_namelen == 0))
1644 GOTO(out, result = -EINVAL);
1646 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1649 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1650 GOTO(out, result = err_serious(-ENOMEM));
1652 mdt_set_disposition(info, ldlm_rep,
1653 (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1655 lh = &info->mti_lh[MDT_LH_PARENT];
1656 mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1657 LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1659 parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1660 MDS_INODELOCK_UPDATE);
1662 GOTO(out, result = PTR_ERR(parent));
1664 /* get and check version of parent */
1665 result = mdt_version_get_check(info, parent, 0);
1667 GOTO(out_parent, result);
1669 fid_zero(child_fid);
1671 lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1672 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1673 lname, child_fid, &info->mti_spec);
1674 LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1675 "looking for "DFID"/%s, result fid="DFID"\n",
1676 PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1678 if (result != 0 && result != -ENOENT && result != -ESTALE)
1679 GOTO(out_parent, result);
1681 if (result == -ENOENT || result == -ESTALE) {
1682 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1683 if (result == -ESTALE) {
1685 * -ESTALE means the parent is a dead(unlinked) dir, so
1686 * it should return -ENOENT to in accordance with the
1687 * original mds implementaion.
1689 GOTO(out_parent, result = -ENOENT);
1691 if (!(create_flags & MDS_OPEN_CREAT))
1692 GOTO(out_parent, result);
1693 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1694 GOTO(out_parent, result = -EROFS);
1695 *child_fid = *info->mti_rr.rr_fid2;
1696 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1698 /* In the function below, .hs_keycmp resolves to
1699 * lu_obj_hop_keycmp() */
1700 /* coverity[overrun-buffer-val] */
1701 child = mdt_object_new(info->mti_env, mdt, child_fid);
1704 * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1705 * return FID back in that case.
1707 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1708 child = mdt_object_find(info->mti_env, mdt, child_fid);
1711 GOTO(out_parent, result = PTR_ERR(child));
1713 /** check version of child */
1714 rc = mdt_version_get_check(info, child, 1);
1716 GOTO(out_child, result = rc);
1718 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1719 if (result == -ENOENT) {
1720 /* Create under OBF and .lustre is not permitted */
1721 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
1722 GOTO(out_child, result = -EPERM);
1724 /* save versions in reply */
1725 mdt_version_get_save(info, parent, 0);
1726 mdt_version_get_save(info, child, 1);
1728 /* version of child will be changed */
1729 info->mti_mos = child;
1731 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1732 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1734 /* Let lower layers know what is lock mode on directory. */
1735 info->mti_spec.sp_cr_mode =
1736 mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1739 * Do not perform lookup sanity check. We know that name does
1742 info->mti_spec.sp_cr_lookup = 0;
1743 info->mti_spec.sp_feat = &dt_directory_features;
1745 result = mdo_create(info->mti_env,
1746 mdt_object_child(parent),
1748 mdt_object_child(child),
1751 if (result == -ERESTART) {
1752 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1753 GOTO(out_child, result);
1756 /* XXX: we should call this once, see few lines below */
1758 result = mdt_attr_get_complex(info, child, ma);
1761 GOTO(out_child, result);
1766 * The object is on remote node, return its FID for remote open.
1768 if (mdt_object_remote(child)) {
1770 * Check if this lock already was sent to client and
1771 * this is resent case. For resent case do not take lock
1772 * again, use what is already granted.
1774 LASSERT(lhc != NULL);
1776 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1777 struct ldlm_lock *lock;
1779 LASSERT(msg_flags & MSG_RESENT);
1781 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1783 CERROR("Invalid lock handle "LPX64"\n",
1784 lhc->mlh_reg_lh.cookie);
1787 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1788 &lock->l_resource->lr_name));
1789 LDLM_LOCK_PUT(lock);
1792 mdt_lock_handle_init(lhc);
1793 mdt_lock_reg_init(lhc, LCK_PR);
1795 rc = mdt_object_lock(info, child, lhc,
1796 MDS_INODELOCK_LOOKUP,
1799 repbody->fid1 = *mdt_object_fid(child);
1800 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1805 GOTO(out_child, result);
1807 if (mdt_object_exists(child)) {
1808 /* We have to get attr & LOV EA & HSM for this
1810 ma->ma_need |= MA_HSM;
1811 result = mdt_attr_get_complex(info, child, ma);
1813 /*object non-exist!!! Likely an fs corruption*/
1814 CERROR("%s: name %s present, but fid " DFID
1815 " invalid\n",mdt_obd_name(info->mti_mdt),
1816 rr->rr_name, PFID(child_fid));
1817 GOTO(out_child, result = -EIO);
1822 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1823 /* the open lock might already be gotten in
1824 * mdt_intent_fixup_resent */
1825 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1826 if (create_flags & MDS_OPEN_LOCK)
1827 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1829 /* get openlock if this isn't replay and client requested it */
1830 if (!req_is_replay(req)) {
1831 rc = mdt_object_open_lock(info, child, lhc, &ibits);
1833 GOTO(out_child_unlock, result = rc);
1834 else if (create_flags & MDS_OPEN_LOCK)
1835 mdt_set_disposition(info, ldlm_rep,
1839 /* Try to open it now. */
1840 rc = mdt_finish_open(info, parent, child, create_flags,
1844 /* openlock will be released if mdt_finish_open failed */
1845 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1847 if (created && create_flags & MDS_OPEN_VOLATILE) {
1848 CERROR("%s: cannot open volatile file "DFID", orphan "
1849 "file will be left in PENDING directory until "
1850 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1851 PFID(mdt_object_fid(child)), rc);
1852 GOTO(out_child_unlock, result);
1858 ma->ma_cookie_size = 0;
1859 rc = mdo_unlink(info->mti_env,
1860 mdt_object_child(parent),
1861 mdt_object_child(child),
1863 &info->mti_attr, 0);
1865 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1866 mdt_obd_name(info->mti_mdt),
1867 PFID(mdt_object_fid(child)), rc);
1868 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1873 mdt_object_open_unlock(info, child, lhc, ibits, result);
1875 mdt_object_put(info->mti_env, child);
1877 mdt_object_unlock_put(info, parent, lh, result || !created);
1880 lustre_msg_set_transno(req->rq_repmsg, 0);
1885 * Create an orphan object use local root.
1887 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1888 struct mdt_device *mdt,
1889 const struct lu_fid *fid,
1890 struct md_attr *attr, fmode_t fmode)
1892 const struct lu_env *env = info->mti_env;
1893 struct md_op_spec *spec = &info->mti_spec;
1894 struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1895 struct mdt_object *obj = NULL;
1896 struct mdt_object *local_root;
1897 static const char name[] = "i_am_nobody";
1898 struct lu_name *lname;
1899 struct lu_ucred *uc;
1900 cfs_cap_t uc_cap_save;
1904 rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1906 RETURN(ERR_PTR(rc));
1908 local_root = mdt_object_find(env, mdt, local_root_fid);
1909 if (IS_ERR(local_root))
1912 obj = mdt_object_new(env, mdt, fid);
1914 GOTO(out, rc = PTR_ERR(obj));
1916 spec->sp_cr_lookup = 0;
1917 spec->sp_feat = &dt_directory_features;
1918 spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1919 spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1920 if (attr->ma_valid & MA_LOV) {
1921 spec->u.sp_ea.eadata = attr->ma_lmm;
1922 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1923 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1925 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1928 lname = mdt_name(env, (char *)name, sizeof(name) - 1);
1931 uc_cap_save = uc->uc_cap;
1932 uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1933 rc = mdo_create(env, mdt_object_child(local_root), lname,
1934 mdt_object_child(obj), spec, attr);
1935 uc->uc_cap = uc_cap_save;
1937 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1938 mdt_obd_name(mdt), PFID(fid), rc);
1942 rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1944 CERROR("%s: cannot open volatile file "DFID", orphan "
1945 "file will be left in PENDING directory until "
1946 "next reboot, rc = %d\n", mdt_obd_name(mdt),
1953 mdt_object_put(env, obj);
1956 mdt_object_put(env, local_root);
1960 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1963 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1964 struct close_data *data;
1965 struct ldlm_lock *lease;
1966 struct mdt_object *orphan;
1967 struct md_attr *orp_ma;
1974 if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1977 data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1981 lease = ldlm_handle2lock(&data->cd_handle);
1985 /* try to hold open_sem so that nobody else can open the file */
1986 if (!down_write_trylock(&o->mot_open_sem)) {
1987 ldlm_lock_cancel(lease);
1988 GOTO(out_reprocess, rc = -EBUSY);
1991 /* Check if the lease open lease has already canceled */
1992 lock_res_and_lock(lease);
1993 lease_broken = ldlm_is_cancel(lease);
1994 unlock_res_and_lock(lease);
1996 LDLM_DEBUG(lease, DFID " lease broken? %d\n",
1997 PFID(mdt_object_fid(o)), lease_broken);
1999 /* Cancel server side lease. Client side counterpart should
2000 * have been cancelled. It's okay to cancel it now as we've
2001 * held mot_open_sem. */
2002 ldlm_lock_cancel(lease);
2004 if (lease_broken) /* don't perform release task */
2005 GOTO(out_unlock, rc = -ESTALE);
2007 if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2008 GOTO(out_unlock, rc = -EINVAL);
2010 /* ma_need was set before but it seems fine to change it in order to
2011 * avoid modifying the one from RPC */
2012 ma->ma_need = MA_HSM;
2013 rc = mdt_attr_get_complex(info, o, ma);
2015 GOTO(out_unlock, rc);
2017 if (!mdt_hsm_release_allow(ma))
2018 GOTO(out_unlock, rc = -EPERM);
2020 /* already released? */
2021 if (ma->ma_hsm.mh_flags & HS_RELEASED)
2022 GOTO(out_unlock, rc = 0);
2024 /* Compare on-disk and packed data_version */
2025 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
2026 CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
2027 " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
2028 data->cd_data_version, ma->ma_hsm.mh_arch_ver);
2029 GOTO(out_unlock, rc = -EPERM);
2032 ma->ma_valid = MA_INODE;
2033 ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
2034 rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
2036 GOTO(out_unlock, rc);
2038 ma->ma_need = MA_INODE | MA_LOV;
2039 rc = mdt_attr_get_complex(info, o, ma);
2041 GOTO(out_unlock, rc);
2043 if (!(ma->ma_valid & MA_LOV)) {
2044 /* Even empty file are released */
2045 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2046 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2047 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2048 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2049 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2051 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
2052 * ma_lmm as lov_user_md, then it will be confused by union of
2053 * layout_gen and stripe_offset. */
2054 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
2055 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2056 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
2057 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
2059 GOTO(out_unlock, rc = -EINVAL);
2062 /* Set file as released */
2063 ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
2065 /* Hopefully it's not used in this call path */
2066 orp_ma = &info->mti_u.som.attr;
2067 orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2068 orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
2069 orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
2070 orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2071 orp_ma->ma_lmm = ma->ma_lmm;
2072 orp_ma->ma_lmm_size = ma->ma_lmm_size;
2073 orp_ma->ma_valid = MA_INODE | MA_LOV;
2074 orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2076 if (IS_ERR(orphan)) {
2077 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2078 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2080 GOTO(out_unlock, rc = PTR_ERR(orphan));
2083 /* Set up HSM attribute for orphan object */
2084 CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
2085 buf = &info->mti_buf;
2086 buf->lb_buf = info->mti_xattr_buf;
2087 buf->lb_len = sizeof(struct hsm_attrs);
2088 ma->ma_hsm.mh_flags |= HS_RELEASED;
2089 lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2090 ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2092 mdt_lock_reg_init(lh, LCK_EX);
2093 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2094 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
2096 GOTO(out_close, rc);
2098 rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2102 /* Swap layout with orphan object */
2103 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2104 mdt_object_child(orphan),
2105 SWAP_LAYOUTS_MDS_HSM);
2107 /* Release exclusive LL */
2108 mdt_object_unlock(info, o, lh, 1);
2113 /* Close orphan object anyway */
2114 rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2117 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2118 mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2119 LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2121 mdt_object_put(info->mti_env, orphan);
2124 up_write(&o->mot_open_sem);
2126 if (rc == 0) { /* already released */
2127 struct mdt_body *repbody;
2128 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2129 LASSERT(repbody != NULL);
2130 repbody->valid |= OBD_MD_FLRELEASED;
2134 ldlm_reprocess_all(lease->l_resource);
2135 LDLM_LOCK_PUT(lease);
2143 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
2144 MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
2146 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2148 return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
2151 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2153 struct mdt_object *o = mfd->mfd_object;
2154 struct md_object *next = mdt_object_child(o);
2155 struct md_attr *ma = &info->mti_attr;
2156 int ret = MDT_IOEPOCH_CLOSED;
2161 mode = mfd->mfd_mode;
2163 if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
2164 rc = mdt_hsm_release(info, o, ma);
2166 CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2167 mdt_obd_name(info->mti_mdt),
2168 PFID(mdt_object_fid(o)), rc);
2169 /* continue to close even error occurred. */
2173 if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
2175 ret = mdt_ioepoch_close(info, o);
2176 } else if (mode & MDS_FMODE_EXEC) {
2178 } else if (mode & MDS_FMODE_EPOCH) {
2179 ret = mdt_ioepoch_close(info, o);
2180 } else if (mode & MDS_FMODE_SOM) {
2181 ret = mdt_som_au_close(info, o);
2184 /* Update atime on close only. */
2185 if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
2186 && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
2187 /* Set the atime only. */
2188 ma->ma_valid = MA_INODE;
2189 ma->ma_attr.la_valid = LA_ATIME;
2190 rc = mo_attr_set(info->mti_env, next, ma);
2193 /* If file data is modified, add the dirty flag. */
2194 if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2195 rc = mdt_add_dirty_flag(info, o, ma);
2197 ma->ma_need |= MA_INODE;
2198 ma->ma_valid &= ~MA_INODE;
2200 if (!MFD_CLOSED(mode))
2201 rc = mo_close(info->mti_env, next, ma, mode);
2203 if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
2204 struct mdt_export_data *med;
2206 /* The IOepoch is still opened or SOM update is needed.
2207 * Put mfd back into the list. */
2208 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
2209 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
2210 MDS_FMODE_EPOCH : MDS_FMODE_SOM);
2212 LASSERT(mdt_info_req(info));
2213 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
2214 spin_lock(&med->med_open_lock);
2215 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
2216 class_handle_hash_back(&mfd->mfd_handle);
2217 spin_unlock(&med->med_open_lock);
2219 if (ret == MDT_IOEPOCH_OPENED) {
2223 CDEBUG(D_INODE, "Size-on-MDS attribute update is "
2224 "needed on "DFID"\n", PFID(mdt_object_fid(o)));
2227 /* adjust open and lease count */
2228 if (mode & MDS_OPEN_LEASE) {
2229 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2230 atomic_dec(&o->mot_lease_count);
2232 LASSERT(atomic_read(&o->mot_open_count) > 0);
2233 atomic_dec(&o->mot_open_count);
2236 mdt_object_put(info->mti_env, o);
2239 RETURN(rc ? rc : ret);
2242 int mdt_close(struct tgt_session_info *tsi)
2244 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2245 struct ptlrpc_request *req = tgt_ses_req(tsi);
2246 struct mdt_export_data *med;
2247 struct mdt_file_data *mfd;
2248 struct mdt_object *o;
2249 struct md_attr *ma = &info->mti_attr;
2250 struct mdt_body *repbody = NULL;
2254 mdt_counter_incr(req, LPROC_MDT_CLOSE);
2255 /* Close may come with the Size-on-MDS update. Unpack it. */
2256 rc = mdt_close_unpack(info);
2258 GOTO(out, rc = err_serious(rc));
2260 LASSERT(info->mti_ioepoch);
2262 req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2263 info->mti_mdt->mdt_max_mdsize);
2264 req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2265 info->mti_mdt->mdt_max_cookiesize);
2266 rc = req_capsule_server_pack(info->mti_pill);
2267 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2268 mdt_client_compatibility(info);
2270 mdt_fix_reply(info);
2271 mdt_exit_ucred(info);
2272 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2275 /* Continue to close handle even if we can not pack reply */
2277 repbody = req_capsule_server_get(info->mti_pill,
2279 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2281 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2284 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
2286 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
2289 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
2290 repbody->eadatasize = 0;
2291 repbody->aclsize = 0;
2293 rc = err_serious(rc);
2296 med = &req->rq_export->exp_mdt_data;
2297 spin_lock(&med->med_open_lock);
2298 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2299 req_is_replay(req));
2300 if (mdt_mfd_closed(mfd)) {
2301 spin_unlock(&med->med_open_lock);
2302 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2303 ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
2304 info->mti_ioepoch->handle.cookie);
2305 /** not serious error since bug 3633 */
2308 class_handle_unhash(&mfd->mfd_handle);
2309 cfs_list_del_init(&mfd->mfd_list);
2310 spin_unlock(&med->med_open_lock);
2312 /* Do not lose object before last unlink. */
2313 o = mfd->mfd_object;
2314 mdt_object_get(info->mti_env, o);
2315 ret = mdt_mfd_close(info, mfd);
2316 if (repbody != NULL)
2317 rc = mdt_handle_last_unlink(info, o, ma);
2318 mdt_empty_transno(info, rc);
2319 mdt_object_put(info->mti_env, o);
2321 if (repbody != NULL) {
2322 mdt_client_compatibility(info);
2323 rc = mdt_fix_reply(info);
2326 mdt_exit_ucred(info);
2327 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2328 GOTO(out, rc = err_serious(-ENOMEM));
2330 if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2331 OBD_FAIL_MDS_CLOSE_NET_REP))
2332 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2334 mdt_thread_info_fini(info);
2335 RETURN(rc ? rc : ret);
2339 * DONE_WRITING rpc handler.
2341 * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
2342 * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
2343 * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
2344 * skip attributes and reconstruct the reply here.
2346 int mdt_done_writing(struct tgt_session_info *tsi)
2348 struct ptlrpc_request *req = tgt_ses_req(tsi);
2349 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2350 struct mdt_body *repbody = NULL;
2351 struct mdt_export_data *med;
2352 struct mdt_file_data *mfd;
2356 rc = req_capsule_server_pack(tsi->tsi_pill);
2358 GOTO(out, rc = err_serious(rc));
2360 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2361 repbody->eadatasize = 0;
2362 repbody->aclsize = 0;
2364 /* Done Writing may come with the Size-on-MDS update. Unpack it. */
2365 rc = mdt_close_unpack(info);
2367 GOTO(out, rc = err_serious(rc));
2369 if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2370 mdt_exit_ucred(info);
2371 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2374 med = &info->mti_exp->exp_mdt_data;
2375 spin_lock(&med->med_open_lock);
2376 mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2377 req_is_replay(req));
2379 spin_unlock(&med->med_open_lock);
2380 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
2381 ": cookie = "LPX64" ioepoch = "LPU64"\n",
2382 PFID(info->mti_rr.rr_fid1),
2383 info->mti_ioepoch->handle.cookie,
2384 info->mti_ioepoch->ioepoch);
2385 /* If this is a replay, reconstruct the transno. */
2386 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2387 rc = info->mti_ioepoch->flags & MF_SOM_AU ?
2389 mdt_empty_transno(info, rc);
2392 GOTO(error_ucred, rc);
2395 LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
2396 mfd->mfd_mode == MDS_FMODE_TRUNC);
2397 class_handle_unhash(&mfd->mfd_handle);
2398 cfs_list_del_init(&mfd->mfd_list);
2399 spin_unlock(&med->med_open_lock);
2401 /* Set EPOCH CLOSE flag if not set by client. */
2402 info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
2403 info->mti_attr.ma_valid = 0;
2405 info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
2406 OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2407 if (info->mti_attr.ma_lmm == NULL)
2408 GOTO(error_ucred, rc = -ENOMEM);
2410 rc = mdt_mfd_close(info, mfd);
2412 OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2413 mdt_empty_transno(info, rc);
2415 mdt_exit_ucred(info);
2417 mdt_thread_info_fini(info);