Whamcloud - gitweb
e225c5b9e349212bfde7e6ef26ea04a8c22e7294
[fs/lustre-release.git] / lustre / mdt / mdt_open.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
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.
9  *
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).
15  *
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
19  *
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
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2012, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/mdt/mdt_open.c
37  *
38  * Lustre Metadata Target (mdt) open/close file handling
39  *
40  * Author: Huang Hua <huanghua@clusterfs.com>
41  */
42
43 #define DEBUG_SUBSYSTEM S_MDS
44
45 #include <lustre_acl.h>
46 #include <lustre_mds.h>
47 #include "mdt_internal.h"
48
49 /* we do nothing because we do not have refcount now */
50 static void mdt_mfd_get(void *mfdp)
51 {
52 }
53
54 static struct portals_handle_ops mfd_handle_ops = {
55         .hop_addref = mdt_mfd_get,
56         .hop_free   = NULL,
57 };
58
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)
62 {
63         struct mdt_file_data *mfd;
64         ENTRY;
65
66         OBD_ALLOC_PTR(mfd);
67         if (mfd != NULL) {
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);
71         }
72         RETURN(mfd);
73 }
74
75 /*
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
79  */
80 struct mdt_file_data *mdt_handle2mfd(struct mdt_thread_info *info,
81                                      const struct lustre_handle *handle)
82 {
83         struct ptlrpc_request *req = mdt_info_req(info);
84         struct mdt_file_data  *mfd;
85         ENTRY;
86
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)
94                                 RETURN (mfd);
95                 }
96                 mfd = NULL;
97         }
98         RETURN (mfd);
99 }
100
101 /* free mfd */
102 void mdt_mfd_free(struct mdt_file_data *mfd)
103 {
104         LASSERT(cfs_list_empty(&mfd->mfd_list));
105         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
106 }
107
108 static int mdt_create_data(struct mdt_thread_info *info,
109                            struct mdt_object *p, struct mdt_object *o)
110 {
111         struct md_op_spec     *spec = &info->mti_spec;
112         struct md_attr        *ma   = &info->mti_attr;
113         int                    rc   = 0;
114         ENTRY;
115
116         if (!md_should_create(spec->sp_cr_flags))
117                 RETURN(0);
118
119         ma->ma_need = MA_INODE | MA_LOV;
120         ma->ma_valid = 0;
121         mutex_lock(&o->mot_lov_mutex);
122         if (!(o->mot_flags & MOF_LOV_CREATED)) {
123                 rc = mdo_create_data(info->mti_env,
124                                      p ? mdt_object_child(p) : NULL,
125                                      mdt_object_child(o), spec, ma);
126                 if (rc == 0)
127                         rc = mdt_attr_get_complex(info, o, ma);
128
129                 if (rc == 0 && ma->ma_valid & MA_LOV)
130                         o->mot_flags |= MOF_LOV_CREATED;
131         }
132         mutex_unlock(&o->mot_lov_mutex);
133         RETURN(rc);
134 }
135
136 static int mdt_ioepoch_opened(struct mdt_object *mo)
137 {
138         return mo->mot_ioepoch_count;
139 }
140
141 int mdt_object_is_som_enabled(struct mdt_object *mo)
142 {
143         return !mo->mot_ioepoch;
144 }
145
146 /**
147  * Re-enable Size-on-MDS.
148  * Call under ->mot_ioepoch_mutex.
149  */
150 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
151 {
152         if (ioepoch == mo->mot_ioepoch) {
153                 LASSERT(!mdt_ioepoch_opened(mo));
154                 mo->mot_ioepoch = 0;
155                 mo->mot_flags = 0;
156         }
157 }
158
159 /**
160  * Open the IOEpoch. It is allowed if @writecount is not negative.
161  * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
162  */
163 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
164                      int created)
165 {
166         struct mdt_device *mdt = info->mti_mdt;
167         int cancel = 0;
168         int rc = 0;
169         ENTRY;
170
171         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
172             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
173                 RETURN(0);
174
175         mutex_lock(&o->mot_ioepoch_mutex);
176         if (mdt_ioepoch_opened(o)) {
177                 /* Epoch continues even if there is no writers yet. */
178                 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
179                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
180         } else {
181                 /* XXX: ->mdt_ioepoch is not initialized at the mount */
182                 spin_lock(&mdt->mdt_ioepoch_lock);
183                 if (mdt->mdt_ioepoch < info->mti_replayepoch)
184                         mdt->mdt_ioepoch = info->mti_replayepoch;
185
186                 if (info->mti_replayepoch)
187                         o->mot_ioepoch = info->mti_replayepoch;
188                 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
189                         o->mot_ioepoch = ++mdt->mdt_ioepoch;
190                 else
191                         o->mot_ioepoch = mdt->mdt_ioepoch;
192
193                 spin_unlock(&mdt->mdt_ioepoch_lock);
194
195                 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
196                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
197                 if (created)
198                         o->mot_flags |= MOF_SOM_CREATED;
199                 cancel = 1;
200         }
201         o->mot_ioepoch_count++;
202         mutex_unlock(&o->mot_ioepoch_mutex);
203
204         /* Cancel Size-on-MDS attributes cached on clients for the open case.
205          * In the truncate case, see mdt_reint_setattr(). */
206         if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
207                 struct mdt_lock_handle  *lh = &info->mti_lh[MDT_LH_CHILD];
208                 mdt_lock_reg_init(lh, LCK_EX);
209                 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
210                                      MDT_LOCAL_LOCK);
211                 if (rc == 0)
212                         mdt_object_unlock(info, o, lh, 1);
213         }
214         RETURN(rc);
215 }
216
217 /**
218  * Update SOM on-disk attributes.
219  * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
220  * mountid and attributes. If disabling, clean SOM xattr.
221  * Call under ->mot_ioepoch_mutex.
222  */
223 static int mdt_som_attr_set(struct mdt_thread_info *info,
224                             struct mdt_object *obj, __u64 ioepoch, bool enable)
225 {
226         struct md_object        *next = mdt_object_child(obj);
227         int                      rc;
228         ENTRY;
229
230         CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
231                " on "DFID".\n", enable ? "update" : "disabling",
232                ioepoch, PFID(mdt_object_fid(obj)));
233
234         if (enable) {
235                 struct lu_buf           *buf = &info->mti_buf;
236                 struct som_attrs        *attrs;
237                 struct md_attr          *ma = &info->mti_attr;
238                 struct lu_attr          *la = &ma->ma_attr;
239                 struct obd_device       *obd = info->mti_mdt->mdt_lut.lut_obd;
240
241                 attrs = (struct som_attrs *)info->mti_xattr_buf;
242                 CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*attrs));
243
244                 /* pack SOM attributes */
245                 memset(attrs, 0, sizeof(*attrs));
246                 attrs->som_ioepoch = ioepoch;
247                 attrs->som_mountid = obd->u.obt.obt_mount_count;
248                 if ((la->la_valid & LA_SIZE) != 0)
249                         attrs->som_size = la->la_size;
250                 if ((la->la_valid & LA_BLOCKS) != 0)
251                         attrs->som_blocks = la->la_blocks;
252                 lustre_som_swab(attrs);
253
254                 /* update SOM attributes */
255                 buf->lb_buf = attrs;
256                 buf->lb_len = sizeof(*attrs);
257                 rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_SOM, 0);
258         } else {
259                 /* delete SOM attributes */
260                 rc = mo_xattr_del(info->mti_env, next, XATTR_NAME_SOM);
261         }
262
263         RETURN(rc);
264 }
265
266 /** Perform the eviction specific actions on ioepoch close. */
267 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
268                                                 struct mdt_object *o)
269 {
270         int rc = 0;
271
272         mutex_lock(&o->mot_ioepoch_mutex);
273         CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
274                "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
275                o->mot_ioepoch_count);
276         o->mot_ioepoch_count--;
277
278         /* If eviction occured set MOF_SOM_RECOV,
279          * if no other epoch holders, disable SOM on disk. */
280         o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
281         if (!mdt_ioepoch_opened(o)) {
282                 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
283                 mdt_object_som_enable(o, o->mot_ioepoch);
284         }
285         mutex_unlock(&o->mot_ioepoch_mutex);
286         RETURN(rc);
287 }
288
289 /**
290  * Perform the replay specific actions on ioepoch close.
291  * Skip SOM attribute update if obtained and just forget about the inode state
292  * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
293  */
294 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
295                                               struct mdt_object *o)
296 {
297         int rc = MDT_IOEPOCH_CLOSED;
298         ENTRY;
299
300         mutex_lock(&o->mot_ioepoch_mutex);
301         CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
302                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
303         o->mot_ioepoch_count--;
304
305         /* Get an info from the replayed request if client is supposed
306          * to send an Attibute Update, reconstruct @rc if so */
307         if (info->mti_ioepoch->flags & MF_SOM_AU)
308                 rc = MDT_IOEPOCH_GETATTR;
309
310         if (!mdt_ioepoch_opened(o))
311                 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
312         mutex_unlock(&o->mot_ioepoch_mutex);
313
314         RETURN(rc);
315 }
316
317 /**
318  * Regular file IOepoch close.
319  * Closes the ioepoch, checks the object state, apply obtained attributes and
320  * re-enable SOM on the object, if possible. Also checks if the recovery is
321  * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
322  * to obtain SOM attributes under the server-side OST locks.
323  *
324  * Return value:
325  * MDT_IOEPOCH_CLOSED if ioepoch is closed.
326  * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
327  */
328 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
329                                         struct mdt_object *o)
330 {
331         struct md_attr *tmp_ma;
332         struct lu_attr *la;
333         int achange, opened;
334         int recovery = 0;
335         int rc = 0, ret = MDT_IOEPOCH_CLOSED;
336         ENTRY;
337
338         la = &info->mti_attr.ma_attr;
339         achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
340
341         mutex_lock(&o->mot_ioepoch_mutex);
342         o->mot_ioepoch_count--;
343
344         tmp_ma = &info->mti_u.som.attr;
345         tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
346         tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
347         tmp_ma->ma_som = &info->mti_u.som.data;
348         tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
349         tmp_ma->ma_valid = 0;
350         rc = mdt_attr_get_complex(info, o, tmp_ma);
351         if (rc)
352                 GOTO(error_up, rc);
353
354         /* Check the on-disk SOM state. */
355         if (o->mot_flags & MOF_SOM_RECOV)
356                 recovery = 1;
357         else if (!(o->mot_flags & MOF_SOM_CREATED) &&
358                  !(tmp_ma->ma_valid & MA_SOM))
359                 recovery = 1;
360
361         CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
362                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
363
364         opened = mdt_ioepoch_opened(o);
365         /**
366          * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
367          * Skip the check for file with no LOV  or for unlink files.
368          */
369         if (!opened && tmp_ma->ma_valid & MA_LOV &&
370             !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
371                 if (recovery)
372                         /* If some previous writer was evicted, re-ask the
373                          * client for attributes. Even if attributes are
374                          * provided, we cannot believe in them.
375                          * Another use case is that there is no SOM cache on
376                          * disk -- first access with SOM or there was an MDS
377                          * failure. */
378                         ret = MDT_IOEPOCH_GETATTR;
379                 else if (o->mot_flags & MOF_SOM_CHANGE)
380                         /* Some previous writer changed the attribute.
381                          * Do not believe to the current Size-on-MDS
382                          * update, re-ask client. */
383                         ret = MDT_IOEPOCH_GETATTR;
384                 else if (!(la->la_valid & LA_SIZE) && achange)
385                         /* Attributes were changed by the last writer
386                          * only but no Size-on-MDS update is received.*/
387                         ret = MDT_IOEPOCH_GETATTR;
388         }
389
390         if (achange || ret == MDT_IOEPOCH_GETATTR)
391                 o->mot_flags |= MOF_SOM_CHANGE;
392
393         /* If epoch ends and relable SOM attributes are obtained, update them.
394          * Create SOM ea for new files even if there is no attributes obtained
395          * (0-length file). */
396         if (ret == MDT_IOEPOCH_CLOSED && !opened) {
397                 if (achange || o->mot_flags & MOF_SOM_CREATED) {
398                         LASSERT(achange || !(la->la_valid & LA_SIZE));
399                         rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
400                                               MDT_SOM_ENABLE);
401                         /* Avoid the following setattrs of these attributes,
402                          * e.g. for atime update. */
403                         info->mti_attr.ma_valid = 0;
404                 }
405                 mdt_object_som_enable(o, o->mot_ioepoch);
406         }
407
408         mutex_unlock(&o->mot_ioepoch_mutex);
409         /* If recovery is needed, tell the client to perform GETATTR under
410          * the lock. */
411         if (ret == MDT_IOEPOCH_GETATTR && recovery) {
412                 struct mdt_body *rep;
413                 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
414                 rep->valid |= OBD_MD_FLGETATTRLOCK;
415         }
416
417         RETURN(rc ? : ret);
418
419 error_up:
420         mutex_unlock(&o->mot_ioepoch_mutex);
421         return rc;
422 }
423
424 /**
425  * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
426  * - a client closes the IOEpoch;
427  * - a client eviction occured.
428  * Return values:
429  * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
430  * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
431  * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
432  * update is needed.
433  */
434 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
435 {
436         struct ptlrpc_request *req = mdt_info_req(info);
437         ENTRY;
438
439         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
440             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
441                 RETURN(0);
442
443         LASSERT(o->mot_ioepoch_count);
444         LASSERT(info->mti_ioepoch == NULL ||
445                 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
446
447         /* IOEpoch is closed only if client tells about it or eviction occures.
448          * In the replay case, always close the epoch. */
449         if (req == NULL)
450                 RETURN(mdt_ioepoch_close_on_eviction(info, o));
451         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
452                 RETURN(mdt_ioepoch_close_on_replay(info, o));
453         if (info->mti_ioepoch->flags & MF_EPOCH_CLOSE)
454                 RETURN(mdt_ioepoch_close_reg(info, o));
455         /* IO epoch is not closed. */
456         RETURN(MDT_IOEPOCH_OPENED);
457 }
458
459 /**
460  * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
461  * for attribute update. It happens if:
462  * - SOM Attribute Update is obtained;
463  * - the client failed to obtain it and informs MDS about it;
464  * - a client eviction occured.
465  * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
466  * cache otherwise.
467  */
468 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
469 {
470         struct ptlrpc_request *req = mdt_info_req(info);
471         __u64 ioepoch = 0;
472         int act = MDT_SOM_ENABLE;
473         int rc = 0;
474         ENTRY;
475
476         LASSERT(!req || info->mti_ioepoch);
477         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
478             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
479                 RETURN(0);
480
481         /* No size whereas MF_SOM_CHANGE is set means client failed to
482          * obtain ost attributes, drop the SOM cache on disk if so. */
483         if (!req ||
484             (info->mti_ioepoch &&
485              info->mti_ioepoch->flags & MF_SOM_CHANGE &&
486              !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
487                 act = MDT_SOM_DISABLE;
488
489         mutex_lock(&o->mot_ioepoch_mutex);
490         /* Mark the object it is the recovery state if we failed to obtain
491          * SOM attributes. */
492         if (act == MDT_SOM_DISABLE)
493                 o->mot_flags |= MOF_SOM_RECOV;
494
495         if (!mdt_ioepoch_opened(o)) {
496                 ioepoch =  info->mti_ioepoch ?
497                         info->mti_ioepoch->ioepoch : o->mot_ioepoch;
498
499                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
500                         rc = mdt_som_attr_set(info, o, ioepoch, act);
501                 mdt_object_som_enable(o, ioepoch);
502         }
503         mutex_unlock(&o->mot_ioepoch_mutex);
504         RETURN(rc);
505 }
506
507 int mdt_write_read(struct mdt_object *o)
508 {
509         int rc = 0;
510         ENTRY;
511         mutex_lock(&o->mot_ioepoch_mutex);
512         rc = o->mot_writecount;
513         mutex_unlock(&o->mot_ioepoch_mutex);
514         RETURN(rc);
515 }
516
517 int mdt_write_get(struct mdt_object *o)
518 {
519         int rc = 0;
520         ENTRY;
521         mutex_lock(&o->mot_ioepoch_mutex);
522         if (o->mot_writecount < 0)
523                 rc = -ETXTBSY;
524         else
525                 o->mot_writecount++;
526         mutex_unlock(&o->mot_ioepoch_mutex);
527         RETURN(rc);
528 }
529
530 void mdt_write_put(struct mdt_object *o)
531 {
532         ENTRY;
533         mutex_lock(&o->mot_ioepoch_mutex);
534         o->mot_writecount--;
535         mutex_unlock(&o->mot_ioepoch_mutex);
536         EXIT;
537 }
538
539 static int mdt_write_deny(struct mdt_object *o)
540 {
541         int rc = 0;
542         ENTRY;
543         mutex_lock(&o->mot_ioepoch_mutex);
544         if (o->mot_writecount > 0)
545                 rc = -ETXTBSY;
546         else
547                 o->mot_writecount--;
548         mutex_unlock(&o->mot_ioepoch_mutex);
549         RETURN(rc);
550 }
551
552 static void mdt_write_allow(struct mdt_object *o)
553 {
554         ENTRY;
555         mutex_lock(&o->mot_ioepoch_mutex);
556         o->mot_writecount++;
557         mutex_unlock(&o->mot_ioepoch_mutex);
558         EXIT;
559 }
560
561 /* there can be no real transaction so prepare the fake one */
562 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
563 {
564         struct mdt_device      *mdt = info->mti_mdt;
565         struct ptlrpc_request  *req = mdt_info_req(info);
566         struct tg_export_data  *ted;
567         struct lsd_client_data *lcd;
568
569         ENTRY;
570         /* transaction has occurred already */
571         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
572                 RETURN_EXIT;
573
574         spin_lock(&mdt->mdt_lut.lut_translock);
575         if (rc != 0) {
576                 if (info->mti_transno != 0) {
577                         struct obd_export *exp = req->rq_export;
578
579                         CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
580                                 mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name,
581                                 info->mti_transno,
582                                 libcfs_nid2str(exp->exp_connection->c_peer.nid),
583                                 rc);
584                         RETURN_EXIT;
585                 }
586         } else if (info->mti_transno == 0) {
587                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
588         } else {
589                 /* should be replay */
590                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
591                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
592         }
593         spin_unlock(&mdt->mdt_lut.lut_translock);
594
595         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
596                         info->mti_transno,
597                         req->rq_export->exp_obd->obd_last_committed);
598
599         req->rq_transno = info->mti_transno;
600         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
601
602         /* update lcd in memory only for resent cases */
603         ted = &req->rq_export->exp_target_data;
604         LASSERT(ted);
605         mutex_lock(&ted->ted_lcd_lock);
606         lcd = ted->ted_lcd;
607         if (info->mti_transno < lcd->lcd_last_transno &&
608             info->mti_transno != 0) {
609                 /* This should happen during replay. Do not update
610                  * last rcvd info if replay req transno < last transno,
611                  * otherwise the following resend(after replay) can not
612                  * be checked correctly by xid */
613                 mutex_unlock(&ted->ted_lcd_lock);
614                 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
615                         mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name,
616                         info->mti_transno, lcd->lcd_last_transno);
617                 RETURN_EXIT;
618         }
619
620         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
621             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
622                 if (info->mti_transno != 0)
623                         lcd->lcd_last_close_transno = info->mti_transno;
624                 lcd->lcd_last_close_xid = req->rq_xid;
625                 lcd->lcd_last_close_result = rc;
626         } else {
627                 /* VBR: save versions in last_rcvd for reconstruct. */
628                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
629                 if (pre_versions) {
630                         lcd->lcd_pre_versions[0] = pre_versions[0];
631                         lcd->lcd_pre_versions[1] = pre_versions[1];
632                         lcd->lcd_pre_versions[2] = pre_versions[2];
633                         lcd->lcd_pre_versions[3] = pre_versions[3];
634                 }
635                 if (info->mti_transno != 0)
636                         lcd->lcd_last_transno = info->mti_transno;
637
638                 lcd->lcd_last_xid = req->rq_xid;
639                 lcd->lcd_last_result = rc;
640                 lcd->lcd_last_data = info->mti_opdata;
641         }
642         mutex_unlock(&ted->ted_lcd_lock);
643
644         EXIT;
645 }
646
647 void mdt_mfd_set_mode(struct mdt_file_data *mfd, int mode)
648 {
649         LASSERT(mfd != NULL);
650
651         CDEBUG(D_HA, "Change mfd %p mode 0x%x->0x%x\n",
652                mfd, (unsigned int)mfd->mfd_mode, (unsigned int)mode);
653
654         mfd->mfd_mode = mode;
655 }
656
657 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
658                         struct mdt_object *o, __u64 flags, int created)
659 {
660         struct ptlrpc_request   *req = mdt_info_req(info);
661         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
662         struct mdt_file_data    *mfd;
663         struct md_attr          *ma  = &info->mti_attr;
664         struct lu_attr          *la  = &ma->ma_attr;
665         struct mdt_body         *repbody;
666         int                      rc = 0, isdir, isreg;
667         ENTRY;
668
669         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
670
671         isreg = S_ISREG(la->la_mode);
672         isdir = S_ISDIR(la->la_mode);
673         if (isreg && !(ma->ma_valid & MA_LOV)) {
674                 /*
675                  * No EA, check whether it is will set regEA and dirEA since in
676                  * above attr get, these size might be zero, so reset it, to
677                  * retrieve the MD after create obj.
678                  */
679                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
680                                                        &RMF_MDT_MD,
681                                                        RCL_SERVER);
682                 /* in replay case, p == NULL */
683                 rc = mdt_create_data(info, p, o);
684                 if (rc)
685                         RETURN(rc);
686         }
687
688         CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
689                ma->ma_valid, ma->ma_lmm_size);
690
691         if (ma->ma_valid & MA_LOV) {
692                 LASSERT(ma->ma_lmm_size != 0);
693                 repbody->eadatasize = ma->ma_lmm_size;
694                 if (isdir)
695                         repbody->valid |= OBD_MD_FLDIREA;
696                 else
697                         repbody->valid |= OBD_MD_FLEASIZE;
698         }
699
700         if (flags & FMODE_WRITE) {
701                 rc = mdt_write_get(o);
702                 if (rc == 0) {
703                         mdt_ioepoch_open(info, o, created);
704                         repbody->ioepoch = o->mot_ioepoch;
705                 }
706         } else if (flags & MDS_FMODE_EXEC) {
707                 rc = mdt_write_deny(o);
708         }
709         if (rc)
710                 RETURN(rc);
711
712         rc = mo_open(info->mti_env, mdt_object_child(o),
713                      created ? flags | MDS_OPEN_CREATED : flags);
714         if (rc)
715                 GOTO(err_out, rc);
716
717         mfd = mdt_mfd_new();
718         if (mfd != NULL) {
719                 /*
720                  * Keep a reference on this object for this open, and is
721                  * released by mdt_mfd_close().
722                  */
723                 mdt_object_get(info->mti_env, o);
724
725                 /*
726                  * @flags is always not zero. At least it should be FMODE_READ,
727                  * FMODE_WRITE or MDS_FMODE_EXEC.
728                  */
729                 LASSERT(flags != 0);
730
731                 /* Open handling. */
732                 mdt_mfd_set_mode(mfd, flags);
733
734                 mfd->mfd_object = o;
735                 mfd->mfd_xid = req->rq_xid;
736
737                 /* replay handle */
738                 if (req_is_replay(req)) {
739                         struct mdt_file_data *old_mfd;
740                         /* Check wheather old cookie already exist in
741                          * the list, becasue when do recovery, client
742                          * might be disconnected from server, and
743                          * restart replay, so there maybe some orphan
744                          * mfd here, we should remove them */
745                         LASSERT(info->mti_rr.rr_handle != NULL);
746                         old_mfd = mdt_handle2mfd(info, info->mti_rr.rr_handle);
747                         if (old_mfd) {
748                                 CDEBUG(D_HA, "del orph mfd %p fid=("DFID") "
749                                        "cookie=" LPX64"\n", mfd,
750                                        PFID(mdt_object_fid(mfd->mfd_object)),
751                                        info->mti_rr.rr_handle->cookie);
752                                 spin_lock(&med->med_open_lock);
753                                 class_handle_unhash(&old_mfd->mfd_handle);
754                                 cfs_list_del_init(&old_mfd->mfd_list);
755                                 spin_unlock(&med->med_open_lock);
756                                 /* no attr update for that close */
757                                 la->la_valid = 0;
758                                 ma->ma_valid |= MA_FLAGS;
759                                 ma->ma_attr_flags |= MDS_RECOV_OPEN;
760                                 mdt_mfd_close(info, old_mfd);
761                                 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
762                                 ma->ma_valid &= ~MA_FLAGS;
763                         }
764                         CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
765                                info->mti_rr.rr_handle->cookie);
766                         mfd->mfd_old_handle.cookie =
767                                                 info->mti_rr.rr_handle->cookie;
768                 }
769                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
770
771                 if (req->rq_export->exp_disconnected) {
772                         spin_lock(&med->med_open_lock);
773                         class_handle_unhash(&mfd->mfd_handle);
774                         cfs_list_del_init(&mfd->mfd_list);
775                         spin_unlock(&med->med_open_lock);
776                         mdt_mfd_close(info, mfd);
777                 } else {
778                         spin_lock(&med->med_open_lock);
779                         cfs_list_add(&mfd->mfd_list, &med->med_open_head);
780                         spin_unlock(&med->med_open_lock);
781                 }
782
783                 mdt_empty_transno(info, rc);
784         } else {
785                 GOTO(err_out, rc = -ENOMEM);
786         }
787
788         RETURN(rc);
789
790 err_out:
791         if (flags & FMODE_WRITE)
792                         /* XXX We also need to close io epoch here.
793                          * See LU-1220 - green */
794                 mdt_write_put(o);
795         else if (flags & FMODE_EXEC)
796                 mdt_write_allow(o);
797         return rc;
798 }
799
800 int mdt_finish_open(struct mdt_thread_info *info,
801                     struct mdt_object *p, struct mdt_object *o,
802                     __u64 flags, int created, struct ldlm_reply *rep)
803 {
804         struct ptlrpc_request   *req = mdt_info_req(info);
805         struct obd_export       *exp = req->rq_export;
806         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
807         struct md_attr          *ma  = &info->mti_attr;
808         struct lu_attr          *la  = &ma->ma_attr;
809         struct mdt_file_data    *mfd;
810         struct mdt_body         *repbody;
811         int                      rc = 0;
812         int                      isreg, isdir, islnk;
813         cfs_list_t              *t;
814         ENTRY;
815
816         LASSERT(ma->ma_valid & MA_INODE);
817
818         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
819
820         isreg = S_ISREG(la->la_mode);
821         isdir = S_ISDIR(la->la_mode);
822         islnk = S_ISLNK(la->la_mode);
823         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
824
825         /* LU-2275, simulate broken behaviour (esp. prevalent in
826          * pre-2.4 servers where a very strange reply is sent on error
827          * that looks like it was actually almost succesful and a failure at the
828          * same time */
829         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
830                 mdt_set_disposition(info, rep, DISP_OPEN_LOCK | \
831                                     DISP_OPEN_OPEN | DISP_LOOKUP_NEG | \
832                                     DISP_LOOKUP_POS);
833                 RETURN(-ENOENT);
834         }
835
836         if (exp_connect_rmtclient(exp)) {
837                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
838
839                 rc = mdt_pack_remote_perm(info, o, buf);
840                 if (rc) {
841                         repbody->valid &= ~OBD_MD_FLRMTPERM;
842                         repbody->aclsize = 0;
843                 } else {
844                         repbody->valid |= OBD_MD_FLRMTPERM;
845                         repbody->aclsize = sizeof(struct mdt_remote_perm);
846                 }
847         }
848 #ifdef CONFIG_FS_POSIX_ACL
849         else if (exp->exp_connect_flags & OBD_CONNECT_ACL) {
850                 const struct lu_env *env = info->mti_env;
851                 struct md_object *next = mdt_object_child(o);
852                 struct lu_buf *buf = &info->mti_buf;
853
854                 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
855                 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
856                                                    RCL_SERVER);
857                 if (buf->lb_len > 0) {
858                         rc = mo_xattr_get(env, next, buf,
859                                           XATTR_NAME_ACL_ACCESS);
860                         if (rc < 0) {
861                                 if (rc == -ENODATA) {
862                                         repbody->aclsize = 0;
863                                         repbody->valid |= OBD_MD_FLACL;
864                                         rc = 0;
865                                 } else if (rc == -EOPNOTSUPP) {
866                                         rc = 0;
867                                 } else {
868                                         CERROR("got acl size: %d\n", rc);
869                                 }
870                         } else {
871                                 repbody->aclsize = rc;
872                                 repbody->valid |= OBD_MD_FLACL;
873                                 rc = 0;
874                         }
875                 }
876         }
877 #endif
878
879         if (info->mti_mdt->mdt_opts.mo_mds_capa &&
880             exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) {
881                 struct lustre_capa *capa;
882
883                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
884                 LASSERT(capa);
885                 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
886                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
887                 if (rc)
888                         RETURN(rc);
889                 repbody->valid |= OBD_MD_FLMDSCAPA;
890         }
891         if (info->mti_mdt->mdt_opts.mo_oss_capa &&
892             exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA &&
893             S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) {
894                 struct lustre_capa *capa;
895
896                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
897                 LASSERT(capa);
898                 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
899                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
900                 if (rc)
901                         RETURN(rc);
902                 repbody->valid |= OBD_MD_FLOSSCAPA;
903         }
904
905         /*
906          * If we are following a symlink, don't open; and do not return open
907          * handle for special nodes as client required.
908          */
909         if (islnk || (!isreg && !isdir &&
910             (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH))) {
911                 lustre_msg_set_transno(req->rq_repmsg, 0);
912                 RETURN(0);
913         }
914
915         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
916
917         /*
918          * We need to return the existing object's fid back, so it is done here,
919          * after preparing the reply.
920          */
921         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
922                 RETURN(-EEXIST);
923
924         /* This can't be done earlier, we need to return reply body */
925         if (isdir) {
926                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
927                         /* We are trying to create or write an existing dir. */
928                         RETURN(-EISDIR);
929                 }
930         } else if (flags & MDS_OPEN_DIRECTORY)
931                 RETURN(-ENOTDIR);
932
933         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
934                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
935                 RETURN(-EAGAIN);
936         }
937
938         mfd = NULL;
939         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
940                 spin_lock(&med->med_open_lock);
941                 cfs_list_for_each(t, &med->med_open_head) {
942                         mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
943                         if (mfd->mfd_xid == req->rq_xid)
944                                 break;
945                         mfd = NULL;
946                 }
947                 spin_unlock(&med->med_open_lock);
948
949                 if (mfd != NULL) {
950                         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
951                         /*set repbody->ea_size for resent case*/
952                         if (ma->ma_valid & MA_LOV) {
953                                 LASSERT(ma->ma_lmm_size != 0);
954                                 repbody->eadatasize = ma->ma_lmm_size;
955                                 if (isdir)
956                                         repbody->valid |= OBD_MD_FLDIREA;
957                                 else
958                                         repbody->valid |= OBD_MD_FLEASIZE;
959                         }
960                         RETURN(0);
961                 }
962         }
963
964         rc = mdt_mfd_open(info, p, o, flags, created);
965         RETURN(rc);
966 }
967
968 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
969                              struct lsd_client_data *lcd);
970
971 void mdt_reconstruct_open(struct mdt_thread_info *info,
972                           struct mdt_lock_handle *lhc)
973 {
974         const struct lu_env *env = info->mti_env;
975         struct mdt_device       *mdt  = info->mti_mdt;
976         struct req_capsule      *pill = info->mti_pill;
977         struct ptlrpc_request   *req  = mdt_info_req(info);
978         struct tg_export_data   *ted  = &req->rq_export->exp_target_data;
979         struct lsd_client_data  *lcd  = ted->ted_lcd;
980         struct md_attr          *ma   = &info->mti_attr;
981         struct mdt_reint_record *rr   = &info->mti_rr;
982         __u32                   flags = info->mti_spec.sp_cr_flags;
983         struct ldlm_reply       *ldlm_rep;
984         struct mdt_object       *parent;
985         struct mdt_object       *child;
986         struct mdt_body         *repbody;
987         int                      rc;
988         ENTRY;
989
990         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
991         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
992         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
993
994         ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
995         ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
996                                                RCL_SERVER);
997         ma->ma_need = MA_INODE;
998         if (ma->ma_lmm_size > 0)
999                 ma->ma_need |= MA_LOV;
1000
1001         ma->ma_valid = 0;
1002
1003         mdt_req_from_lcd(req, lcd);
1004         mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1005
1006         CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1007                ldlm_rep->lock_policy_res1, req->rq_status);
1008
1009         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1010             req->rq_status != 0)
1011                 /* We did not create successfully, return error to client. */
1012                 GOTO(out, rc = req->rq_status);
1013
1014         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1015                 struct obd_export *exp = req->rq_export;
1016                 /*
1017                  * We failed after creation, but we do not know in which step
1018                  * we failed. So try to check the child object.
1019                  */
1020                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1021                 if (IS_ERR(parent)) {
1022                         rc = PTR_ERR(parent);
1023                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
1024                                       " Evicting client %s with export %s.\n",
1025                                       PFID(rr->rr_fid1), rc,
1026                                       obd_uuid2str(&exp->exp_client_uuid),
1027                                       obd_export_nid2str(exp));
1028                         mdt_export_evict(exp);
1029                         RETURN_EXIT;
1030                 }
1031                 child = mdt_object_find(env, mdt, rr->rr_fid2);
1032                 if (IS_ERR(child)) {
1033                         rc = PTR_ERR(child);
1034                         LCONSOLE_WARN("Child "DFID" lookup error %d."
1035                                       " Evicting client %s with export %s.\n",
1036                                       PFID(mdt_object_fid(child)), rc,
1037                                       obd_uuid2str(&exp->exp_client_uuid),
1038                                       obd_export_nid2str(exp));
1039                         mdt_object_put(env, parent);
1040                         mdt_export_evict(exp);
1041                         RETURN_EXIT;
1042                 }
1043                 rc = mdt_object_exists(child);
1044                 if (rc > 0) {
1045
1046                         mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA);
1047                         rc = mdt_attr_get_complex(info, child, ma);
1048                         if (rc == 0)
1049                               rc = mdt_finish_open(info, parent, child,
1050                                                    flags, 1, ldlm_rep);
1051                 } else if (rc < 0) {
1052                         /* the child object was created on remote server */
1053                         repbody->fid1 = *rr->rr_fid2;
1054                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1055                         rc = 0;
1056                 } else if (rc == 0) {
1057                         /* the child does not exist, we should do regular open */
1058                         mdt_object_put(env, parent);
1059                         mdt_object_put(env, child);
1060                         GOTO(regular_open, 0);
1061                 }
1062                 mdt_object_put(env, parent);
1063                 mdt_object_put(env, child);
1064                 GOTO(out, rc);
1065         } else {
1066 regular_open:
1067                 /* We did not try to create, so we are a pure open */
1068                 rc = mdt_reint_open(info, lhc);
1069         }
1070
1071         EXIT;
1072 out:
1073         req->rq_status = rc;
1074         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1075         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1076 }
1077
1078 int mdt_open_by_fid(struct mdt_thread_info* info,
1079                     struct ldlm_reply *rep)
1080 {
1081         __u32                    flags = info->mti_spec.sp_cr_flags;
1082         struct mdt_reint_record *rr = &info->mti_rr;
1083         struct md_attr          *ma = &info->mti_attr;
1084         struct mdt_object       *o;
1085         int                      rc;
1086         ENTRY;
1087
1088         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1089         if (IS_ERR(o))
1090                 RETURN(rc = PTR_ERR(o));
1091
1092         rc = mdt_object_exists(o);
1093         if (rc > 0) {
1094                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1095                                                 DISP_LOOKUP_EXECD |
1096                                                 DISP_LOOKUP_POS));
1097
1098                 rc = mdt_attr_get_complex(info, o, ma);
1099                 if (rc == 0)
1100                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1101         } else if (rc == 0) {
1102                 rc = -ENOENT;
1103         } else  {
1104                 /* the child object was created on remote server */
1105                 struct mdt_body *repbody;
1106                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1107                 repbody->fid1 = *rr->rr_fid2;
1108                 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1109                 rc = 0;
1110         }
1111
1112         mdt_object_put(info->mti_env, o);
1113         RETURN(rc);
1114 }
1115
1116 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1117                          struct mdt_lock_handle *lhc)
1118 {
1119         const struct lu_env     *env   = info->mti_env;
1120         struct mdt_device       *mdt   = info->mti_mdt;
1121         __u32                    flags = info->mti_spec.sp_cr_flags;
1122         struct mdt_reint_record *rr    = &info->mti_rr;
1123         struct md_attr          *ma    = &info->mti_attr;
1124         struct mdt_object       *parent= NULL;
1125         struct mdt_object       *o;
1126         int                      rc;
1127         ldlm_mode_t              lm;
1128         ENTRY;
1129
1130         if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1131                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1132                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1133                         if (IS_ERR(parent)) {
1134                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1135                                        " for anonymous created %ld, try to"
1136                                        " use server-side parent.\n",
1137                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1138                                 parent = NULL;
1139                         }
1140                 }
1141                 if (parent == NULL)
1142                         ma->ma_need |= MA_PFID;
1143         }
1144
1145         o = mdt_object_find(env, mdt, rr->rr_fid2);
1146         if (IS_ERR(o))
1147                 RETURN(rc = PTR_ERR(o));
1148
1149         rc = mdt_object_exists(o);
1150         if (rc == 0) {
1151                 mdt_set_disposition(info, rep, (DISP_LOOKUP_EXECD |
1152                                     DISP_LOOKUP_NEG));
1153                 GOTO(out, rc = -ENOENT);
1154         } else if (rc < 0) {
1155                 CERROR("NFS remote open shouldn't happen.\n");
1156                 GOTO(out, rc);
1157         }
1158         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1159                                         DISP_LOOKUP_EXECD |
1160                                         DISP_LOOKUP_POS));
1161
1162         if (flags & FMODE_WRITE)
1163                 lm = LCK_CW;
1164         else if (flags & MDS_FMODE_EXEC)
1165                 lm = LCK_PR;
1166         else
1167                 lm = LCK_CR;
1168
1169         mdt_lock_handle_init(lhc);
1170         mdt_lock_reg_init(lhc, lm);
1171         rc = mdt_object_lock(info, o, lhc,
1172                              MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1173                              MDT_CROSS_LOCK);
1174         if (rc)
1175                 GOTO(out, rc);
1176
1177         rc = mdt_attr_get_complex(info, o, ma);
1178         if (rc)
1179                 GOTO(out, rc);
1180
1181         if (ma->ma_valid & MA_PFID) {
1182                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1183                 if (IS_ERR(parent)) {
1184                         CDEBUG(D_INODE, "Fail to find parent "DFID
1185                                " for anonymous created %ld, try to"
1186                                " use system default.\n",
1187                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1188                         parent = NULL;
1189                 }
1190         }
1191
1192         if (flags & MDS_OPEN_LOCK)
1193                 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1194         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1195
1196         if (!(flags & MDS_OPEN_LOCK) || rc)
1197                 mdt_object_unlock(info, o, lhc, 1);
1198
1199         GOTO(out, rc);
1200 out:
1201         mdt_object_put(env, o);
1202         if (parent != NULL)
1203                 mdt_object_put(env, parent);
1204         return rc;
1205 }
1206
1207 int mdt_pin(struct mdt_thread_info* info)
1208 {
1209         ENTRY;
1210         RETURN(err_serious(-EOPNOTSUPP));
1211 }
1212
1213 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1214 int mdt_cross_open(struct mdt_thread_info* info,
1215                    const struct lu_fid *fid,
1216                    struct ldlm_reply *rep, __u32 flags)
1217 {
1218         struct md_attr    *ma = &info->mti_attr;
1219         struct mdt_object *o;
1220         int                rc;
1221         ENTRY;
1222
1223         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1224         if (IS_ERR(o))
1225                 RETURN(rc = PTR_ERR(o));
1226
1227         rc = mdt_object_exists(o);
1228         if (rc > 0) {
1229                 /* Do permission check for cross-open. */
1230                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o),
1231                                    NULL, flags | MDS_OPEN_CROSS);
1232                 if (rc)
1233                         goto out;
1234
1235                 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1236                 rc = mdt_attr_get_complex(info, o, ma);
1237                 if (rc == 0)
1238                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1239         } else if (rc == 0) {
1240                 /*
1241                  * Something is wrong here. lookup was positive but there is
1242                  * no object!
1243                  */
1244                 CERROR("Cross-ref object doesn't exist!\n");
1245                 rc = -EFAULT;
1246         } else  {
1247                 /* Something is wrong here, the object is on another MDS! */
1248                 CERROR("The object isn't on this server! FLD error?\n");
1249                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1250                                 &o->mot_obj.mo_lu,
1251                                 "Object isn't on this server! FLD error?\n");
1252
1253                 rc = -EFAULT;
1254         }
1255
1256 out:
1257         mdt_object_put(info->mti_env, o);
1258         RETURN(rc);
1259 }
1260
1261 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1262 {
1263         struct mdt_device       *mdt = info->mti_mdt;
1264         struct ptlrpc_request   *req = mdt_info_req(info);
1265         struct mdt_object       *parent;
1266         struct mdt_object       *child;
1267         struct mdt_lock_handle  *lh;
1268         struct ldlm_reply       *ldlm_rep;
1269         struct mdt_body         *repbody;
1270         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1271         struct md_attr          *ma = &info->mti_attr;
1272         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1273         struct mdt_reint_record *rr = &info->mti_rr;
1274         struct lu_name          *lname;
1275         int                      result, rc;
1276         int                      created = 0;
1277         __u32                    msg_flags;
1278         ENTRY;
1279
1280         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1281                                (obd_timeout + 1) / 4);
1282
1283         mdt_counter_incr(req, LPROC_MDT_OPEN);
1284         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1285
1286         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1287         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1288                                                RCL_SERVER);
1289         ma->ma_need = MA_INODE;
1290         if (ma->ma_lmm_size > 0)
1291                 ma->ma_need |= MA_LOV;
1292
1293         ma->ma_valid = 0;
1294
1295         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1296         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1297
1298         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1299                 CERROR("file join is not supported anymore.\n");
1300                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1301         }
1302         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1303
1304         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1305             info->mti_spec.u.sp_ea.eadata == NULL)
1306                 GOTO(out, result = err_serious(-EINVAL));
1307
1308         CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1309                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1310                PFID(rr->rr_fid1), rr->rr_name,
1311                PFID(rr->rr_fid2), create_flags,
1312                ma->ma_attr.la_mode, msg_flags);
1313
1314         if (req_is_replay(req) ||
1315             (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1316                 /* This is a replay request or from liblustre with ea. */
1317                 result = mdt_open_by_fid(info, ldlm_rep);
1318
1319                 if (result != -ENOENT) {
1320                         if (req->rq_export->exp_libclient &&
1321                             create_flags & MDS_OPEN_HAS_EA)
1322                                 GOTO(out, result = 0);
1323                         GOTO(out, result);
1324                 }
1325                 /* We didn't find the correct object, so we need to re-create it
1326                  * via a regular replay. */
1327                 if (!(create_flags & MDS_OPEN_CREAT)) {
1328                         DEBUG_REQ(D_ERROR, req,
1329                                   "OPEN & CREAT not in open replay/by_fid.");
1330                         GOTO(out, result = -EFAULT);
1331                 }
1332                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1333         } else if ((rr->rr_namelen == 0 && !info->mti_cross_ref &&
1334                     create_flags & MDS_OPEN_LOCK) ||
1335                    (create_flags & MDS_OPEN_BY_FID)) {
1336                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1337                 if (result != -ENOENT && !(create_flags & MDS_OPEN_CREAT))
1338                         GOTO(out, result);
1339                 if (unlikely(rr->rr_namelen == 0))
1340                         GOTO(out, result = -EINVAL);
1341                 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1342         }
1343
1344         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1345                 GOTO(out, result = err_serious(-ENOMEM));
1346
1347         mdt_set_disposition(info, ldlm_rep,
1348                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1349
1350         if (info->mti_cross_ref) {
1351                 /* This is cross-ref open */
1352                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1353                 result = mdt_cross_open(info, rr->rr_fid1, ldlm_rep,
1354                                         create_flags);
1355                 GOTO(out, result);
1356         }
1357
1358         lh = &info->mti_lh[MDT_LH_PARENT];
1359         mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1360                           LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1361
1362         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1363                                       MDS_INODELOCK_UPDATE);
1364         if (IS_ERR(parent))
1365                 GOTO(out, result = PTR_ERR(parent));
1366
1367         /* get and check version of parent */
1368         result = mdt_version_get_check(info, parent, 0);
1369         if (result)
1370                 GOTO(out_parent, result);
1371
1372         fid_zero(child_fid);
1373
1374         lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1375         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1376                             lname, child_fid, &info->mti_spec);
1377         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1378                  "looking for "DFID"/%s, result fid="DFID"\n",
1379                  PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1380
1381         if (result != 0 && result != -ENOENT && result != -ESTALE)
1382                 GOTO(out_parent, result);
1383
1384         if (result == -ENOENT || result == -ESTALE) {
1385                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1386                 if (result == -ESTALE) {
1387                         /*
1388                          * -ESTALE means the parent is a dead(unlinked) dir, so
1389                          * it should return -ENOENT to in accordance with the
1390                          * original mds implementaion.
1391                          */
1392                         GOTO(out_parent, result = -ENOENT);
1393                 }
1394                 if (!(create_flags & MDS_OPEN_CREAT))
1395                         GOTO(out_parent, result);
1396                 *child_fid = *info->mti_rr.rr_fid2;
1397                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1398                          PFID(child_fid));
1399                 /* In the function below, .hs_keycmp resolves to
1400                  * lu_obj_hop_keycmp() */
1401                 /* coverity[overrun-buffer-val] */
1402                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1403         } else {
1404                 /*
1405                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1406                  * return FID back in that case.
1407                  */
1408                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1409                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1410         }
1411         if (IS_ERR(child))
1412                 GOTO(out_parent, result = PTR_ERR(child));
1413
1414         /** check version of child  */
1415         rc = mdt_version_get_check(info, child, 1);
1416         if (rc)
1417                 GOTO(out_child, result = rc);
1418
1419         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1420         if (result == -ENOENT) {
1421                 if (mdt_object_obf(parent))
1422                         GOTO(out_child, result = -EPERM);
1423
1424                 /* save versions in reply */
1425                 mdt_version_get_save(info, parent, 0);
1426                 mdt_version_get_save(info, child, 1);
1427
1428                 /* version of child will be changed */
1429                 info->mti_mos = child;
1430
1431                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1432                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1433
1434                 /* Let lower layers know what is lock mode on directory. */
1435                 info->mti_spec.sp_cr_mode =
1436                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1437
1438                 /*
1439                  * Do not perform lookup sanity check. We know that name does
1440                  * not exist.
1441                  */
1442                 info->mti_spec.sp_cr_lookup = 0;
1443                 info->mti_spec.sp_feat = &dt_directory_features;
1444
1445                 result = mdo_create(info->mti_env,
1446                                     mdt_object_child(parent),
1447                                     lname,
1448                                     mdt_object_child(child),
1449                                     &info->mti_spec,
1450                                     &info->mti_attr);
1451                 if (result == -ERESTART) {
1452                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1453                         GOTO(out_child, result);
1454                 } else {
1455
1456                         /* XXX: we should call this once, see few lines below */
1457                         if (result == 0)
1458                                 result = mdt_attr_get_complex(info, child, ma);
1459
1460                         if (result != 0)
1461                                 GOTO(out_child, result);
1462                 }
1463                 created = 1;
1464         } else {
1465                 /* We have to get attr & lov ea for this object */
1466                 result = mdt_attr_get_complex(info, child, ma);
1467                 /*
1468                  * The object is on remote node, return its FID for remote open.
1469                  */
1470                 if (result == -EREMOTE) {
1471                         /*
1472                          * Check if this lock already was sent to client and
1473                          * this is resent case. For resent case do not take lock
1474                          * again, use what is already granted.
1475                          */
1476                         LASSERT(lhc != NULL);
1477
1478                         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1479                                 struct ldlm_lock *lock;
1480
1481                                 LASSERT(msg_flags & MSG_RESENT);
1482
1483                                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1484                                 if (!lock) {
1485                                         CERROR("Invalid lock handle "LPX64"\n",
1486                                                lhc->mlh_reg_lh.cookie);
1487                                         LBUG();
1488                                 }
1489                                 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1490                                                         &lock->l_resource->lr_name));
1491                                 LDLM_LOCK_PUT(lock);
1492                                 rc = 0;
1493                         } else {
1494                                 mdt_lock_handle_init(lhc);
1495                                 mdt_lock_reg_init(lhc, LCK_PR);
1496
1497                                 rc = mdt_object_lock(info, child, lhc,
1498                                                      MDS_INODELOCK_LOOKUP,
1499                                                      MDT_CROSS_LOCK);
1500                         }
1501                         repbody->fid1 = *mdt_object_fid(child);
1502                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1503                         if (rc != 0)
1504                                 result = rc;
1505                         GOTO(out_child, result);
1506                 }
1507         }
1508
1509         LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1510
1511         /* get openlock if this is not replay and if a client requested it */
1512         if (!req_is_replay(req) && create_flags & MDS_OPEN_LOCK) {
1513                 ldlm_mode_t lm;
1514
1515                 if (create_flags & FMODE_WRITE)
1516                         lm = LCK_CW;
1517                 else if (create_flags & MDS_FMODE_EXEC)
1518                         lm = LCK_PR;
1519                 else
1520                         lm = LCK_CR;
1521                 mdt_lock_handle_init(lhc);
1522                 mdt_lock_reg_init(lhc, lm);
1523                 rc = mdt_object_lock(info, child, lhc,
1524                                      MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1525                                      MDT_CROSS_LOCK);
1526                 if (rc) {
1527                         result = rc;
1528                         GOTO(out_child, result);
1529                 } else {
1530                         result = -EREMOTE;
1531                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1532                 }
1533         }
1534
1535         /* Try to open it now. */
1536         rc = mdt_finish_open(info, parent, child, create_flags,
1537                              created, ldlm_rep);
1538         if (rc) {
1539                 result = rc;
1540                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1541                         /* openlock was acquired and mdt_finish_open failed -
1542                            drop the openlock */
1543                         mdt_object_unlock(info, child, lhc, 1);
1544                 if (created) {
1545                         ma->ma_need = 0;
1546                         ma->ma_valid = 0;
1547                         ma->ma_cookie_size = 0;
1548                         rc = mdo_unlink(info->mti_env,
1549                                         mdt_object_child(parent),
1550                                         mdt_object_child(child),
1551                                         lname,
1552                                         &info->mti_attr);
1553                         if (rc != 0)
1554                                 CERROR("Error in cleanup of open\n");
1555                 }
1556         }
1557         EXIT;
1558 out_child:
1559         mdt_object_put(info->mti_env, child);
1560 out_parent:
1561         mdt_object_unlock_put(info, parent, lh, result || !created);
1562 out:
1563         if (result && result != -EREMOTE)
1564                 lustre_msg_set_transno(req->rq_repmsg, 0);
1565         return result;
1566 }
1567
1568 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
1569                                       MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
1570
1571 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1572 {
1573         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1574 }
1575
1576 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1577 {
1578         struct mdt_object *o = mfd->mfd_object;
1579         struct md_object *next = mdt_object_child(o);
1580         struct md_attr *ma = &info->mti_attr;
1581         int ret = MDT_IOEPOCH_CLOSED;
1582         int rc = 0;
1583         int mode;
1584         ENTRY;
1585
1586         mode = mfd->mfd_mode;
1587
1588         if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
1589                 mdt_write_put(o);
1590                 ret = mdt_ioepoch_close(info, o);
1591         } else if (mode & MDS_FMODE_EXEC) {
1592                 mdt_write_allow(o);
1593         } else if (mode & MDS_FMODE_EPOCH) {
1594                 ret = mdt_ioepoch_close(info, o);
1595         } else if (mode & MDS_FMODE_SOM) {
1596                 ret = mdt_som_au_close(info, o);
1597         }
1598
1599         /* Update atime on close only. */
1600         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1601             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1602                 /* Set the atime only. */
1603                 ma->ma_valid = MA_INODE;
1604                 ma->ma_attr.la_valid = LA_ATIME;
1605                 rc = mo_attr_set(info->mti_env, next, ma);
1606         }
1607
1608         ma->ma_need |= MA_INODE;
1609         ma->ma_valid &= ~MA_INODE;
1610
1611         if (!MFD_CLOSED(mode))
1612                 rc = mo_close(info->mti_env, next, ma, mode);
1613
1614         if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
1615                 struct mdt_export_data *med;
1616
1617                 /* The IOepoch is still opened or SOM update is needed.
1618                  * Put mfd back into the list. */
1619                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
1620                 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
1621                                       MDS_FMODE_EPOCH : MDS_FMODE_SOM);
1622
1623                 LASSERT(mdt_info_req(info));
1624                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
1625                 spin_lock(&med->med_open_lock);
1626                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
1627                 class_handle_hash_back(&mfd->mfd_handle);
1628                 spin_unlock(&med->med_open_lock);
1629
1630                 if (ret == MDT_IOEPOCH_OPENED) {
1631                         ret = 0;
1632                 } else {
1633                         ret = -EAGAIN;
1634                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
1635                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
1636                 }
1637         } else {
1638                 mdt_mfd_free(mfd);
1639                 mdt_object_put(info->mti_env, o);
1640         }
1641
1642         RETURN(rc ? rc : ret);
1643 }
1644
1645 int mdt_close(struct mdt_thread_info *info)
1646 {
1647         struct mdt_export_data *med;
1648         struct mdt_file_data   *mfd;
1649         struct mdt_object      *o;
1650         struct md_attr         *ma = &info->mti_attr;
1651         struct mdt_body        *repbody = NULL;
1652         struct ptlrpc_request  *req = mdt_info_req(info);
1653         int rc, ret = 0;
1654         ENTRY;
1655
1656         mdt_counter_incr(req, LPROC_MDT_CLOSE);
1657         /* Close may come with the Size-on-MDS update. Unpack it. */
1658         rc = mdt_close_unpack(info);
1659         if (rc)
1660                 RETURN(err_serious(rc));
1661
1662         LASSERT(info->mti_ioepoch);
1663
1664         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
1665                              info->mti_mdt->mdt_max_mdsize);
1666         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
1667                              info->mti_mdt->mdt_max_cookiesize);
1668         rc = req_capsule_server_pack(info->mti_pill);
1669         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
1670                 mdt_client_compatibility(info);
1671                 if (rc == 0)
1672                         mdt_fix_reply(info);
1673                 mdt_exit_ucred(info);
1674                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1675         }
1676
1677         /* Continue to close handle even if we can not pack reply */
1678         if (rc == 0) {
1679                 repbody = req_capsule_server_get(info->mti_pill,
1680                                                  &RMF_MDT_BODY);
1681                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
1682                                                     &RMF_MDT_MD);
1683                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
1684                                                        &RMF_MDT_MD,
1685                                                        RCL_SERVER);
1686                 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
1687                                                        &RMF_LOGCOOKIES);
1688                 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
1689                                                           &RMF_LOGCOOKIES,
1690                                                           RCL_SERVER);
1691                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
1692                 repbody->eadatasize = 0;
1693                 repbody->aclsize = 0;
1694         } else {
1695                 rc = err_serious(rc);
1696         }
1697
1698         med = &req->rq_export->exp_mdt_data;
1699         spin_lock(&med->med_open_lock);
1700         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1701         if (mdt_mfd_closed(mfd)) {
1702                 spin_unlock(&med->med_open_lock);
1703                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
1704                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
1705                        info->mti_ioepoch->handle.cookie);
1706                 /** not serious error since bug 3633 */
1707                 rc = -ESTALE;
1708         } else {
1709                 class_handle_unhash(&mfd->mfd_handle);
1710                 cfs_list_del_init(&mfd->mfd_list);
1711                 spin_unlock(&med->med_open_lock);
1712
1713                 /* Do not lose object before last unlink. */
1714                 o = mfd->mfd_object;
1715                 mdt_object_get(info->mti_env, o);
1716                 ret = mdt_mfd_close(info, mfd);
1717                 if (repbody != NULL)
1718                         rc = mdt_handle_last_unlink(info, o, ma);
1719                 mdt_empty_transno(info, rc);
1720                 mdt_object_put(info->mti_env, o);
1721         }
1722         if (repbody != NULL) {
1723                 mdt_client_compatibility(info);
1724                 rc = mdt_fix_reply(info);
1725         }
1726
1727         mdt_exit_ucred(info);
1728         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
1729                 RETURN(err_serious(-ENOMEM));
1730
1731         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
1732                                  OBD_FAIL_MDS_CLOSE_NET_REP))
1733                 info->mti_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
1734         RETURN(rc ? rc : ret);
1735 }
1736
1737 /**
1738  * DONE_WRITING rpc handler.
1739  *
1740  * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
1741  * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
1742  * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
1743  * skip attributes and reconstruct the reply here.
1744  */
1745 int mdt_done_writing(struct mdt_thread_info *info)
1746 {
1747         struct ptlrpc_request   *req = mdt_info_req(info);
1748         struct mdt_body         *repbody = NULL;
1749         struct mdt_export_data  *med;
1750         struct mdt_file_data    *mfd;
1751         int rc;
1752         ENTRY;
1753
1754         rc = req_capsule_server_pack(info->mti_pill);
1755         if (rc)
1756                 RETURN(err_serious(rc));
1757
1758         repbody = req_capsule_server_get(info->mti_pill,
1759                                          &RMF_MDT_BODY);
1760         repbody->eadatasize = 0;
1761         repbody->aclsize = 0;
1762
1763         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
1764         rc = mdt_close_unpack(info);
1765         if (rc)
1766                 RETURN(err_serious(rc));
1767
1768         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
1769                 mdt_exit_ucred(info);
1770                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1771         }
1772
1773         med = &info->mti_exp->exp_mdt_data;
1774         spin_lock(&med->med_open_lock);
1775         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1776         if (mfd == NULL) {
1777                 spin_unlock(&med->med_open_lock);
1778                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
1779                        ": cookie = "LPX64" ioepoch = "LPU64"\n",
1780                        PFID(info->mti_rr.rr_fid1),
1781                        info->mti_ioepoch->handle.cookie,
1782                        info->mti_ioepoch->ioepoch);
1783                 /* If this is a replay, reconstruct the transno. */
1784                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1785                         rc = info->mti_ioepoch->flags & MF_SOM_AU ?
1786                              -EAGAIN : 0;
1787                         mdt_empty_transno(info, rc);
1788                 } else
1789                         rc = -ESTALE;
1790                 GOTO(error_ucred, rc);
1791         }
1792
1793         LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
1794                 mfd->mfd_mode == MDS_FMODE_TRUNC);
1795         class_handle_unhash(&mfd->mfd_handle);
1796         cfs_list_del_init(&mfd->mfd_list);
1797         spin_unlock(&med->med_open_lock);
1798
1799         /* Set EPOCH CLOSE flag if not set by client. */
1800         info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
1801         info->mti_attr.ma_valid = 0;
1802
1803         info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
1804         OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1805         if (info->mti_attr.ma_lmm == NULL)
1806                 GOTO(error_ucred, rc = -ENOMEM);
1807
1808         rc = mdt_mfd_close(info, mfd);
1809
1810         OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1811         mdt_empty_transno(info, rc);
1812 error_ucred:
1813         mdt_exit_ucred(info);
1814         RETURN(rc);
1815 }