Whamcloud - gitweb
LU-5423 llite: pack suppgid to MDS correctly
[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, 2014, 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 #include <lustre_nodemap.h>
49
50 /* we do nothing because we do not have refcount now */
51 static void mdt_mfd_get(void *mfdp)
52 {
53 }
54
55 static struct portals_handle_ops mfd_handle_ops = {
56         .hop_addref = mdt_mfd_get,
57         .hop_free   = NULL,
58 };
59
60 /* Create a new mdt_file_data struct, initialize it,
61  * and insert it to global hash table */
62 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
63 {
64         struct mdt_file_data *mfd;
65         ENTRY;
66
67         OBD_ALLOC_PTR(mfd);
68         if (mfd != NULL) {
69                 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
70                 mfd->mfd_handle.h_owner = med;
71                 INIT_LIST_HEAD(&mfd->mfd_list);
72                 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
73         }
74
75         RETURN(mfd);
76 }
77
78 /*
79  * Find the mfd pointed to by handle in global hash table.
80  * In case of replay the handle is obsoleted
81  * but mfd can be found in mfd list by that handle.
82  * Callers need to be holding med_open_lock.
83  */
84 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
85                                      const struct lustre_handle *handle,
86                                      bool is_replay_or_resent)
87 {
88         struct mdt_file_data   *mfd;
89         ENTRY;
90
91         LASSERT(handle != NULL);
92         mfd = class_handle2object(handle->cookie, med);
93         /* during dw/setattr replay the mfd can be found by old handle */
94         if (mfd == NULL && is_replay_or_resent) {
95                 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
96                         if (mfd->mfd_old_handle.cookie == handle->cookie)
97                                 RETURN(mfd);
98                 }
99                 mfd = NULL;
100         }
101
102         RETURN(mfd);
103 }
104
105 /* free mfd */
106 void mdt_mfd_free(struct mdt_file_data *mfd)
107 {
108         LASSERT(list_empty(&mfd->mfd_list));
109         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
110 }
111
112 static int mdt_create_data(struct mdt_thread_info *info,
113                            struct mdt_object *p, struct mdt_object *o)
114 {
115         struct md_op_spec     *spec = &info->mti_spec;
116         struct md_attr        *ma   = &info->mti_attr;
117         int                    rc   = 0;
118         ENTRY;
119
120         if (!md_should_create(spec->sp_cr_flags))
121                 RETURN(0);
122
123         ma->ma_need = MA_INODE | MA_LOV;
124         ma->ma_valid = 0;
125         mutex_lock(&o->mot_lov_mutex);
126         if (!(o->mot_flags & MOF_LOV_CREATED)) {
127                 rc = mdo_create_data(info->mti_env,
128                                      p ? mdt_object_child(p) : NULL,
129                                      mdt_object_child(o), spec, ma);
130                 if (rc == 0)
131                         rc = mdt_attr_get_complex(info, o, ma);
132
133                 if (rc == 0 && ma->ma_valid & MA_LOV)
134                         o->mot_flags |= MOF_LOV_CREATED;
135         }
136
137         mutex_unlock(&o->mot_lov_mutex);
138         RETURN(rc);
139 }
140
141 static int mdt_ioepoch_opened(struct mdt_object *mo)
142 {
143         return mo->mot_ioepoch_count;
144 }
145
146 int mdt_object_is_som_enabled(struct mdt_object *mo)
147 {
148         return !mo->mot_ioepoch;
149 }
150
151 /**
152  * Re-enable Size-on-MDS.
153  * Call under ->mot_ioepoch_mutex.
154  */
155 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
156 {
157         if (ioepoch == mo->mot_ioepoch) {
158                 LASSERT(!mdt_ioepoch_opened(mo));
159                 mo->mot_ioepoch = 0;
160                 mo->mot_flags = 0;
161         }
162 }
163
164 /**
165  * Open the IOEpoch. It is allowed if @writecount is not negative.
166  * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
167  */
168 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
169                      int created)
170 {
171         struct mdt_device *mdt = info->mti_mdt;
172         int cancel = 0;
173         int rc = 0;
174         ENTRY;
175
176         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
177             !S_ISREG(lu_object_attr(&o->mot_obj)))
178                 RETURN(0);
179
180         mutex_lock(&o->mot_ioepoch_mutex);
181         if (mdt_ioepoch_opened(o)) {
182                 /* Epoch continues even if there is no writers yet. */
183                 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
184                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
185         } else {
186                 /* XXX: ->mdt_ioepoch is not initialized at the mount */
187                 spin_lock(&mdt->mdt_ioepoch_lock);
188                 if (mdt->mdt_ioepoch < info->mti_replayepoch)
189                         mdt->mdt_ioepoch = info->mti_replayepoch;
190
191                 if (info->mti_replayepoch)
192                         o->mot_ioepoch = info->mti_replayepoch;
193                 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
194                         o->mot_ioepoch = ++mdt->mdt_ioepoch;
195                 else
196                         o->mot_ioepoch = mdt->mdt_ioepoch;
197
198                 spin_unlock(&mdt->mdt_ioepoch_lock);
199
200                 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
201                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
202                 if (created)
203                         o->mot_flags |= MOF_SOM_CREATED;
204                 cancel = 1;
205         }
206         o->mot_ioepoch_count++;
207         mutex_unlock(&o->mot_ioepoch_mutex);
208
209         /* Cancel Size-on-MDS attributes cached on clients for the open case.
210          * In the truncate case, see mdt_reint_setattr(). */
211         if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
212                 struct mdt_lock_handle  *lh = &info->mti_lh[MDT_LH_CHILD];
213                 mdt_lock_reg_init(lh, LCK_EX);
214                 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
215                                      MDT_LOCAL_LOCK);
216                 if (rc == 0)
217                         mdt_object_unlock(info, o, lh, 1);
218         }
219         RETURN(rc);
220 }
221
222 /**
223  * Update SOM on-disk attributes.
224  * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
225  * mountid and attributes. If disabling, clean SOM xattr.
226  * Call under ->mot_ioepoch_mutex.
227  */
228 static int mdt_som_attr_set(struct mdt_thread_info *info,
229                             struct mdt_object *obj, __u64 ioepoch, bool enable)
230 {
231         struct md_object        *next = mdt_object_child(obj);
232         int                      rc;
233         ENTRY;
234
235         CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
236                " on "DFID".\n", enable ? "update" : "disabling",
237                ioepoch, PFID(mdt_object_fid(obj)));
238
239         if (enable) {
240                 struct lu_buf           *buf = &info->mti_buf;
241                 struct som_attrs        *attrs;
242                 struct md_attr          *ma = &info->mti_attr;
243                 struct lu_attr          *la = &ma->ma_attr;
244                 struct obd_device       *obd = info->mti_mdt->mdt_lut.lut_obd;
245
246                 attrs = (struct som_attrs *)info->mti_xattr_buf;
247                 CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*attrs));
248
249                 /* pack SOM attributes */
250                 memset(attrs, 0, sizeof(*attrs));
251                 attrs->som_ioepoch = ioepoch;
252                 attrs->som_mountid = obd->u.obt.obt_mount_count;
253                 if ((la->la_valid & LA_SIZE) != 0)
254                         attrs->som_size = la->la_size;
255                 if ((la->la_valid & LA_BLOCKS) != 0)
256                         attrs->som_blocks = la->la_blocks;
257                 lustre_som_swab(attrs);
258
259                 /* update SOM attributes */
260                 buf->lb_buf = attrs;
261                 buf->lb_len = sizeof(*attrs);
262                 rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_SOM, 0);
263         } else {
264                 /* delete SOM attributes */
265                 rc = mo_xattr_del(info->mti_env, next, XATTR_NAME_SOM);
266         }
267
268         RETURN(rc);
269 }
270
271 /** Perform the eviction specific actions on ioepoch close. */
272 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
273                                                 struct mdt_object *o)
274 {
275         int rc = 0;
276
277         mutex_lock(&o->mot_ioepoch_mutex);
278         CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
279                "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
280                o->mot_ioepoch_count);
281         o->mot_ioepoch_count--;
282
283         /* If eviction occured set MOF_SOM_RECOV,
284          * if no other epoch holders, disable SOM on disk. */
285         o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
286         if (!mdt_ioepoch_opened(o)) {
287                 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
288                 mdt_object_som_enable(o, o->mot_ioepoch);
289         }
290         mutex_unlock(&o->mot_ioepoch_mutex);
291         RETURN(rc);
292 }
293
294 /**
295  * Perform the replay specific actions on ioepoch close.
296  * Skip SOM attribute update if obtained and just forget about the inode state
297  * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
298  */
299 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
300                                               struct mdt_object *o)
301 {
302         int rc = MDT_IOEPOCH_CLOSED;
303         ENTRY;
304
305         mutex_lock(&o->mot_ioepoch_mutex);
306         CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
307                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
308         o->mot_ioepoch_count--;
309
310         /* Get an info from the replayed request if client is supposed
311          * to send an Attibute Update, reconstruct @rc if so */
312         if (info->mti_ioepoch->flags & MF_SOM_AU)
313                 rc = MDT_IOEPOCH_GETATTR;
314
315         if (!mdt_ioepoch_opened(o))
316                 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
317         mutex_unlock(&o->mot_ioepoch_mutex);
318
319         RETURN(rc);
320 }
321
322 /**
323  * Regular file IOepoch close.
324  * Closes the ioepoch, checks the object state, apply obtained attributes and
325  * re-enable SOM on the object, if possible. Also checks if the recovery is
326  * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
327  * to obtain SOM attributes under the server-side OST locks.
328  *
329  * Return value:
330  * MDT_IOEPOCH_CLOSED if ioepoch is closed.
331  * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
332  */
333 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
334                                         struct mdt_object *o)
335 {
336         struct md_attr *tmp_ma;
337         struct lu_attr *la;
338         int achange, opened;
339         int recovery = 0;
340         int rc = 0, ret = MDT_IOEPOCH_CLOSED;
341         ENTRY;
342
343         la = &info->mti_attr.ma_attr;
344         achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
345
346         mutex_lock(&o->mot_ioepoch_mutex);
347         o->mot_ioepoch_count--;
348
349         tmp_ma = &info->mti_u.som.attr;
350         tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
351         tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
352         tmp_ma->ma_som = &info->mti_u.som.data;
353         tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
354         tmp_ma->ma_valid = 0;
355         rc = mdt_attr_get_complex(info, o, tmp_ma);
356         if (rc)
357                 GOTO(error_up, rc);
358
359         /* Check the on-disk SOM state. */
360         if (o->mot_flags & MOF_SOM_RECOV)
361                 recovery = 1;
362         else if (!(o->mot_flags & MOF_SOM_CREATED) &&
363                  !(tmp_ma->ma_valid & MA_SOM))
364                 recovery = 1;
365
366         CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
367                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
368
369         opened = mdt_ioepoch_opened(o);
370         /**
371          * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
372          * Skip the check for file with no LOV  or for unlink files.
373          */
374         if (!opened && tmp_ma->ma_valid & MA_LOV &&
375             !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
376                 if (recovery)
377                         /* If some previous writer was evicted, re-ask the
378                          * client for attributes. Even if attributes are
379                          * provided, we cannot believe in them.
380                          * Another use case is that there is no SOM cache on
381                          * disk -- first access with SOM or there was an MDS
382                          * failure. */
383                         ret = MDT_IOEPOCH_GETATTR;
384                 else if (o->mot_flags & MOF_SOM_CHANGE)
385                         /* Some previous writer changed the attribute.
386                          * Do not believe to the current Size-on-MDS
387                          * update, re-ask client. */
388                         ret = MDT_IOEPOCH_GETATTR;
389                 else if (!(la->la_valid & LA_SIZE) && achange)
390                         /* Attributes were changed by the last writer
391                          * only but no Size-on-MDS update is received.*/
392                         ret = MDT_IOEPOCH_GETATTR;
393         }
394
395         if (achange || ret == MDT_IOEPOCH_GETATTR)
396                 o->mot_flags |= MOF_SOM_CHANGE;
397
398         /* If epoch ends and relable SOM attributes are obtained, update them.
399          * Create SOM ea for new files even if there is no attributes obtained
400          * (0-length file). */
401         if (ret == MDT_IOEPOCH_CLOSED && !opened) {
402                 if (achange || o->mot_flags & MOF_SOM_CREATED) {
403                         LASSERT(achange || !(la->la_valid & LA_SIZE));
404                         rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
405                                               MDT_SOM_ENABLE);
406                         /* Avoid the following setattrs of these attributes,
407                          * e.g. for atime update. */
408                         info->mti_attr.ma_valid = 0;
409                 }
410                 mdt_object_som_enable(o, o->mot_ioepoch);
411         }
412
413         mutex_unlock(&o->mot_ioepoch_mutex);
414         /* If recovery is needed, tell the client to perform GETATTR under
415          * the lock. */
416         if (ret == MDT_IOEPOCH_GETATTR && recovery) {
417                 struct mdt_body *rep;
418                 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
419                 rep->mbo_valid |= OBD_MD_FLGETATTRLOCK;
420         }
421
422         RETURN(rc ? : ret);
423
424 error_up:
425         mutex_unlock(&o->mot_ioepoch_mutex);
426         return rc;
427 }
428
429 /**
430  * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
431  * - a client closes the IOEpoch;
432  * - a client eviction occured.
433  * Return values:
434  * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
435  * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
436  * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
437  * update is needed.
438  */
439 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
440 {
441         struct ptlrpc_request *req = mdt_info_req(info);
442         ENTRY;
443
444         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
445             !S_ISREG(lu_object_attr(&o->mot_obj)))
446                 RETURN(0);
447
448         LASSERT(o->mot_ioepoch_count);
449         LASSERT(info->mti_ioepoch == NULL ||
450                 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
451
452         /* IOEpoch is closed only if client tells about it or eviction occures.
453          * In the replay case, always close the epoch. */
454         if (req == NULL)
455                 RETURN(mdt_ioepoch_close_on_eviction(info, o));
456         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
457                 RETURN(mdt_ioepoch_close_on_replay(info, o));
458         if (info->mti_ioepoch && (info->mti_ioepoch->flags & MF_EPOCH_CLOSE))
459                 RETURN(mdt_ioepoch_close_reg(info, o));
460         /* IO epoch is not closed. */
461         RETURN(MDT_IOEPOCH_OPENED);
462 }
463
464 /**
465  * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
466  * for attribute update. It happens if:
467  * - SOM Attribute Update is obtained;
468  * - the client failed to obtain it and informs MDS about it;
469  * - a client eviction occured.
470  * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
471  * cache otherwise.
472  */
473 static int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
474 {
475         struct ptlrpc_request   *req = mdt_info_req(info);
476         __u64                    ioepoch = 0;
477         int                      act = MDT_SOM_ENABLE;
478         int                      rc = 0;
479         ENTRY;
480
481         LASSERT(!req || info->mti_ioepoch);
482         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
483             !S_ISREG(lu_object_attr(&o->mot_obj)))
484                 RETURN(0);
485
486         /* No size whereas MF_SOM_CHANGE is set means client failed to
487          * obtain ost attributes, drop the SOM cache on disk if so. */
488         if (!req ||
489             (info->mti_ioepoch &&
490              info->mti_ioepoch->flags & MF_SOM_CHANGE &&
491              !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
492                 act = MDT_SOM_DISABLE;
493
494         mutex_lock(&o->mot_ioepoch_mutex);
495         /* Mark the object it is the recovery state if we failed to obtain
496          * SOM attributes. */
497         if (act == MDT_SOM_DISABLE)
498                 o->mot_flags |= MOF_SOM_RECOV;
499
500         if (!mdt_ioepoch_opened(o)) {
501                 ioepoch =  info->mti_ioepoch ?
502                         info->mti_ioepoch->ioepoch : o->mot_ioepoch;
503
504                 if (req != NULL
505                     && !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
506                         rc = mdt_som_attr_set(info, o, ioepoch, act);
507                 mdt_object_som_enable(o, ioepoch);
508         }
509         mutex_unlock(&o->mot_ioepoch_mutex);
510         RETURN(rc);
511 }
512
513 int mdt_write_read(struct mdt_object *o)
514 {
515         int rc = 0;
516         ENTRY;
517         mutex_lock(&o->mot_ioepoch_mutex);
518         rc = o->mot_writecount;
519         mutex_unlock(&o->mot_ioepoch_mutex);
520         RETURN(rc);
521 }
522
523 int mdt_write_get(struct mdt_object *o)
524 {
525         int rc = 0;
526         ENTRY;
527         mutex_lock(&o->mot_ioepoch_mutex);
528         if (o->mot_writecount < 0)
529                 rc = -ETXTBSY;
530         else
531                 o->mot_writecount++;
532         mutex_unlock(&o->mot_ioepoch_mutex);
533         RETURN(rc);
534 }
535
536 void mdt_write_put(struct mdt_object *o)
537 {
538         ENTRY;
539         mutex_lock(&o->mot_ioepoch_mutex);
540         o->mot_writecount--;
541         mutex_unlock(&o->mot_ioepoch_mutex);
542         EXIT;
543 }
544
545 static int mdt_write_deny(struct mdt_object *o)
546 {
547         int rc = 0;
548         ENTRY;
549         mutex_lock(&o->mot_ioepoch_mutex);
550         if (o->mot_writecount > 0)
551                 rc = -ETXTBSY;
552         else
553                 o->mot_writecount--;
554         mutex_unlock(&o->mot_ioepoch_mutex);
555         RETURN(rc);
556 }
557
558 static void mdt_write_allow(struct mdt_object *o)
559 {
560         ENTRY;
561         mutex_lock(&o->mot_ioepoch_mutex);
562         o->mot_writecount++;
563         mutex_unlock(&o->mot_ioepoch_mutex);
564         EXIT;
565 }
566
567 /* there can be no real transaction so prepare the fake one */
568 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
569 {
570         struct mdt_device      *mdt = info->mti_mdt;
571         struct ptlrpc_request  *req = mdt_info_req(info);
572         struct tg_export_data  *ted;
573         struct lsd_client_data *lcd;
574
575         ENTRY;
576         /* transaction has occurred already */
577         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
578                 RETURN_EXIT;
579
580         spin_lock(&mdt->mdt_lut.lut_translock);
581         if (rc != 0) {
582                 if (info->mti_transno != 0) {
583                         struct obd_export *exp = req->rq_export;
584
585                         CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
586                                mdt_obd_name(mdt), info->mti_transno,
587                                libcfs_nid2str(exp->exp_connection->c_peer.nid),
588                                rc);
589                         spin_unlock(&mdt->mdt_lut.lut_translock);
590                         RETURN_EXIT;
591                 }
592         } else if (info->mti_transno == 0) {
593                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
594         } else {
595                 /* should be replay */
596                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
597                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
598         }
599         spin_unlock(&mdt->mdt_lut.lut_translock);
600
601         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
602                info->mti_transno,
603                req->rq_export->exp_obd->obd_last_committed);
604
605         req->rq_transno = info->mti_transno;
606         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
607
608         /* update lcd in memory only for resent cases */
609         ted = &req->rq_export->exp_target_data;
610         LASSERT(ted);
611         mutex_lock(&ted->ted_lcd_lock);
612         lcd = ted->ted_lcd;
613         if (info->mti_transno < lcd->lcd_last_transno &&
614             info->mti_transno != 0) {
615                 /* This should happen during replay. Do not update
616                  * last rcvd info if replay req transno < last transno,
617                  * otherwise the following resend(after replay) can not
618                  * be checked correctly by xid */
619                 mutex_unlock(&ted->ted_lcd_lock);
620                 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
621                        mdt_obd_name(mdt), info->mti_transno,
622                        lcd->lcd_last_transno);
623                 RETURN_EXIT;
624         }
625
626         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
627             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
628                 if (info->mti_transno != 0)
629                         lcd->lcd_last_close_transno = info->mti_transno;
630                 lcd->lcd_last_close_xid = req->rq_xid;
631                 lcd->lcd_last_close_result = rc;
632         } else {
633                 /* VBR: save versions in last_rcvd for reconstruct. */
634                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
635                 if (pre_versions) {
636                         lcd->lcd_pre_versions[0] = pre_versions[0];
637                         lcd->lcd_pre_versions[1] = pre_versions[1];
638                         lcd->lcd_pre_versions[2] = pre_versions[2];
639                         lcd->lcd_pre_versions[3] = pre_versions[3];
640                 }
641                 if (info->mti_transno != 0)
642                         lcd->lcd_last_transno = info->mti_transno;
643
644                 lcd->lcd_last_xid = req->rq_xid;
645                 lcd->lcd_last_result = rc;
646                 lcd->lcd_last_data = info->mti_opdata;
647         }
648         mutex_unlock(&ted->ted_lcd_lock);
649
650         EXIT;
651 }
652
653 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
654 {
655         LASSERT(mfd != NULL);
656
657         CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n",
658                PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
659
660         mfd->mfd_mode = mode;
661 }
662
663 /**
664  * prep ma_lmm/ma_lmv for md_attr from reply
665  */
666 static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
667                                      struct mdt_object *obj,
668                                      struct md_attr *ma)
669 {
670         LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
671         if (S_ISDIR(obj->mot_header.loh_attr)) {
672                 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
673                                                     &RMF_MDT_MD);
674                 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
675                                                        &RMF_MDT_MD,
676                                                        RCL_SERVER);
677                 if (ma->ma_lmv_size > 0)
678                         ma->ma_need |= MA_LMV;
679         } else {
680                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
681                                                     &RMF_MDT_MD);
682                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
683                                                        &RMF_MDT_MD,
684                                                        RCL_SERVER);
685                 if (ma->ma_lmm_size > 0)
686                         ma->ma_need |= MA_LOV;
687         }
688 }
689
690 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
691                         struct mdt_object *o, __u64 flags, int created,
692                         struct ldlm_reply *rep)
693 {
694         struct ptlrpc_request   *req = mdt_info_req(info);
695         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
696         struct mdt_file_data    *mfd;
697         struct md_attr          *ma  = &info->mti_attr;
698         struct lu_attr          *la  = &ma->ma_attr;
699         struct mdt_body         *repbody;
700         int                      rc = 0, isdir, isreg;
701         ENTRY;
702
703         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
704
705         isreg = S_ISREG(la->la_mode);
706         isdir = S_ISDIR(la->la_mode);
707         if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
708                 /*
709                  * No EA, check whether it is will set regEA and dirEA since in
710                  * above attr get, these size might be zero, so reset it, to
711                  * retrieve the MD after create obj.
712                  */
713                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
714                                                        &RMF_MDT_MD,
715                                                        RCL_SERVER);
716                 /* in replay case, p == NULL */
717                 rc = mdt_create_data(info, p, o);
718                 if (rc)
719                         RETURN(rc);
720
721                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
722                         mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
723         }
724
725         CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
726                ma->ma_valid, ma->ma_lmm_size);
727
728         if (ma->ma_valid & MA_LOV) {
729                 LASSERT(ma->ma_lmm_size != 0);
730                 repbody->mbo_eadatasize = ma->ma_lmm_size;
731                 if (isdir)
732                         repbody->mbo_valid |= OBD_MD_FLDIREA;
733                 else
734                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
735         }
736
737         if (ma->ma_valid & MA_LMV) {
738                 LASSERT(ma->ma_lmv_size != 0);
739                 repbody->mbo_eadatasize = ma->ma_lmv_size;
740                 LASSERT(isdir);
741                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
742         }
743
744         if (flags & FMODE_WRITE) {
745                 rc = mdt_write_get(o);
746                 if (rc == 0) {
747                         mdt_ioepoch_open(info, o, created);
748                         repbody->mbo_ioepoch = o->mot_ioepoch;
749                 }
750         } else if (flags & MDS_FMODE_EXEC) {
751                 rc = mdt_write_deny(o);
752         }
753         if (rc)
754                 RETURN(rc);
755
756         rc = mo_open(info->mti_env, mdt_object_child(o),
757                      created ? flags | MDS_OPEN_CREATED : flags);
758         if (rc != 0) {
759                 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
760                  * client does not know which suppgid should be sent to the MDS,
761                  * or some other(s) changed the target file's GID after this RPC
762                  * sent to the MDS with the suppgid as the original GID, then we
763                  * give the client another chance to send the right suppgid. */
764                 if (rc == -EACCES &&
765                     allow_client_chgrp(info, lu_ucred(info->mti_env)))
766                         mdt_set_disposition(info, rep, DISP_OPEN_DENY);
767
768                 GOTO(err_out, rc);
769         }
770
771         mfd = mdt_mfd_new(med);
772         if (mfd == NULL)
773                 GOTO(err_out, rc = -ENOMEM);
774
775         /*
776          * Keep a reference on this object for this open, and is
777          * released by mdt_mfd_close().
778          */
779         mdt_object_get(info->mti_env, o);
780         mfd->mfd_object = o;
781         mfd->mfd_xid = req->rq_xid;
782
783         /*
784          * @flags is always not zero. At least it should be FMODE_READ,
785          * FMODE_WRITE or MDS_FMODE_EXEC.
786          */
787         LASSERT(flags != 0);
788
789         /* Open handling. */
790         mdt_mfd_set_mode(mfd, flags);
791
792         atomic_inc(&o->mot_open_count);
793         if (flags & MDS_OPEN_LEASE)
794                 atomic_inc(&o->mot_lease_count);
795
796         /* replay handle */
797         if (req_is_replay(req)) {
798                 struct mdt_file_data *old_mfd;
799                 /* Check wheather old cookie already exist in
800                  * the list, becasue when do recovery, client
801                  * might be disconnected from server, and
802                  * restart replay, so there maybe some orphan
803                  * mfd here, we should remove them */
804                 LASSERT(info->mti_rr.rr_handle != NULL);
805                 spin_lock(&med->med_open_lock);
806                 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
807                 if (old_mfd != NULL) {
808                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
809                                "cookie = "LPX64"\n", mfd,
810                                PFID(mdt_object_fid(mfd->mfd_object)),
811                                info->mti_rr.rr_handle->cookie);
812                         class_handle_unhash(&old_mfd->mfd_handle);
813                         list_del_init(&old_mfd->mfd_list);
814                         spin_unlock(&med->med_open_lock);
815                         /* no attr update for that close */
816                         la->la_valid = 0;
817                         ma->ma_valid |= MA_FLAGS;
818                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
819                         mdt_mfd_close(info, old_mfd);
820                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
821                         ma->ma_valid &= ~MA_FLAGS;
822                 } else {
823                         spin_unlock(&med->med_open_lock);
824                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
825                                "cookie = "LPX64"\n",
826                                PFID(mdt_object_fid(mfd->mfd_object)),
827                                info->mti_rr.rr_handle->cookie);
828                 }
829
830                 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
831                        info->mti_rr.rr_handle->cookie);
832
833                 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
834         }
835
836         repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
837
838         if (req->rq_export->exp_disconnected) {
839                 spin_lock(&med->med_open_lock);
840                 class_handle_unhash(&mfd->mfd_handle);
841                 list_del_init(&mfd->mfd_list);
842                 spin_unlock(&med->med_open_lock);
843                 mdt_mfd_close(info, mfd);
844         } else {
845                 spin_lock(&med->med_open_lock);
846                 list_add(&mfd->mfd_list, &med->med_open_head);
847                 spin_unlock(&med->med_open_lock);
848         }
849
850         mdt_empty_transno(info, rc);
851
852         RETURN(rc);
853
854 err_out:
855         if (flags & FMODE_WRITE)
856                 /* XXX We also need to close io epoch here.
857                  * See LU-1220 - green */
858                 mdt_write_put(o);
859         else if (flags & MDS_FMODE_EXEC)
860                 mdt_write_allow(o);
861
862         return rc;
863 }
864
865 static int mdt_finish_open(struct mdt_thread_info *info,
866                            struct mdt_object *p, struct mdt_object *o,
867                            __u64 flags, int created, struct ldlm_reply *rep)
868 {
869         struct ptlrpc_request   *req = mdt_info_req(info);
870         struct obd_export       *exp = req->rq_export;
871         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
872         struct md_attr          *ma  = &info->mti_attr;
873         struct lu_attr          *la  = &ma->ma_attr;
874         struct mdt_file_data    *mfd;
875         struct mdt_body         *repbody;
876         int                      rc = 0;
877         int                      isreg, isdir, islnk;
878         struct list_head        *t;
879         ENTRY;
880
881         LASSERT(ma->ma_valid & MA_INODE);
882
883         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
884
885         isreg = S_ISREG(la->la_mode);
886         isdir = S_ISDIR(la->la_mode);
887         islnk = S_ISLNK(la->la_mode);
888         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
889
890         /* LU-2275, simulate broken behaviour (esp. prevalent in
891          * pre-2.4 servers where a very strange reply is sent on error
892          * that looks like it was actually almost succesful and a failure at the
893          * same time */
894         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
895                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
896                                                DISP_LOOKUP_NEG |
897                                                DISP_LOOKUP_POS);
898
899                 if (flags & MDS_OPEN_LOCK)
900                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
901
902                 RETURN(-ENOENT);
903         }
904
905         if (exp_connect_rmtclient(exp)) {
906                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
907
908                 rc = mdt_pack_remote_perm(info, o, buf);
909                 if (rc) {
910                         repbody->mbo_valid &= ~OBD_MD_FLRMTPERM;
911                         repbody->mbo_aclsize = 0;
912                 } else {
913                         repbody->mbo_valid |= OBD_MD_FLRMTPERM;
914                         repbody->mbo_aclsize = sizeof(struct mdt_remote_perm);
915                 }
916         }
917 #ifdef CONFIG_FS_POSIX_ACL
918         else if (exp_connect_flags(exp) & OBD_CONNECT_ACL)
919                 rc = mdt_pack_acl2body(info, repbody, o,
920                                        exp->exp_target_data.ted_nodemap);
921 #endif
922
923         if (info->mti_mdt->mdt_lut.lut_mds_capa &&
924             exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) {
925                 struct lustre_capa *capa;
926
927                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
928                 LASSERT(capa);
929                 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
930                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
931                 if (rc)
932                         RETURN(rc);
933                 repbody->mbo_valid |= OBD_MD_FLMDSCAPA;
934         }
935         if (info->mti_mdt->mdt_lut.lut_oss_capa &&
936             exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA &&
937             S_ISREG(lu_object_attr(&o->mot_obj))) {
938                 struct lustre_capa *capa;
939
940                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
941                 LASSERT(capa);
942                 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
943                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
944                 if (rc)
945                         RETURN(rc);
946                 repbody->mbo_valid |= OBD_MD_FLOSSCAPA;
947         }
948
949         /*
950          * If we are following a symlink, don't open; and do not return open
951          * handle for special nodes as client required.
952          */
953         if (islnk || (!isreg && !isdir &&
954             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
955                 lustre_msg_set_transno(req->rq_repmsg, 0);
956                 RETURN(0);
957         }
958
959         /*
960          * We need to return the existing object's fid back, so it is done here,
961          * after preparing the reply.
962          */
963         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
964                 RETURN(-EEXIST);
965
966         /* This can't be done earlier, we need to return reply body */
967         if (isdir) {
968                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
969                         /* We are trying to create or write an existing dir. */
970                         RETURN(-EISDIR);
971                 }
972         } else if (flags & MDS_OPEN_DIRECTORY)
973                 RETURN(-ENOTDIR);
974
975         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
976                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
977                 RETURN(-EAGAIN);
978         }
979
980         mfd = NULL;
981         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
982                 spin_lock(&med->med_open_lock);
983                 list_for_each(t, &med->med_open_head) {
984                         mfd = list_entry(t, struct mdt_file_data, mfd_list);
985                         if (mfd->mfd_xid == req->rq_xid)
986                                 break;
987                         mfd = NULL;
988                 }
989                 spin_unlock(&med->med_open_lock);
990
991                 if (mfd != NULL) {
992                         repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
993                         /* set repbody->ea_size for resent case */
994                         if (ma->ma_valid & MA_LOV) {
995                                 LASSERT(ma->ma_lmm_size != 0);
996                                 repbody->mbo_eadatasize = ma->ma_lmm_size;
997                                 if (isdir)
998                                         repbody->mbo_valid |= OBD_MD_FLDIREA;
999                                 else
1000                                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
1001                         }
1002                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
1003                         RETURN(0);
1004                 }
1005         }
1006
1007         rc = mdt_mfd_open(info, p, o, flags, created, rep);
1008         if (!rc)
1009                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
1010
1011         RETURN(rc);
1012 }
1013
1014 void mdt_reconstruct_open(struct mdt_thread_info *info,
1015                           struct mdt_lock_handle *lhc)
1016 {
1017         const struct lu_env *env = info->mti_env;
1018         struct mdt_device       *mdt  = info->mti_mdt;
1019         struct req_capsule      *pill = info->mti_pill;
1020         struct ptlrpc_request   *req  = mdt_info_req(info);
1021         struct tg_export_data   *ted  = &req->rq_export->exp_target_data;
1022         struct lsd_client_data  *lcd  = ted->ted_lcd;
1023         struct md_attr          *ma   = &info->mti_attr;
1024         struct mdt_reint_record *rr   = &info->mti_rr;
1025         __u64                   flags = info->mti_spec.sp_cr_flags;
1026         struct ldlm_reply       *ldlm_rep;
1027         struct mdt_object       *parent;
1028         struct mdt_object       *child;
1029         struct mdt_body         *repbody;
1030         int                      rc;
1031         ENTRY;
1032
1033         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1034         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1035         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1036
1037         ma->ma_need = MA_INODE | MA_HSM;
1038         ma->ma_valid = 0;
1039
1040         mdt_req_from_lcd(req, lcd);
1041         mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1042
1043         CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1044                ldlm_rep->lock_policy_res1, req->rq_status);
1045
1046         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1047             req->rq_status != 0)
1048                 /* We did not create successfully, return error to client. */
1049                 GOTO(out, rc = req->rq_status);
1050
1051         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1052                 struct obd_export *exp = req->rq_export;
1053                 /*
1054                  * We failed after creation, but we do not know in which step
1055                  * we failed. So try to check the child object.
1056                  */
1057                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1058                 if (IS_ERR(parent)) {
1059                         rc = PTR_ERR(parent);
1060                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
1061                                       " Evicting client %s with export %s.\n",
1062                                       PFID(rr->rr_fid1), rc,
1063                                       obd_uuid2str(&exp->exp_client_uuid),
1064                                       obd_export_nid2str(exp));
1065                         mdt_export_evict(exp);
1066                         RETURN_EXIT;
1067                 }
1068
1069                 child = mdt_object_find(env, mdt, rr->rr_fid2);
1070                 if (IS_ERR(child)) {
1071                         rc = PTR_ERR(child);
1072                         LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
1073                                       "evicting client %s with export %s\n",
1074                                       PFID(rr->rr_fid2), rc,
1075                                       obd_uuid2str(&exp->exp_client_uuid),
1076                                       obd_export_nid2str(exp));
1077                         mdt_object_put(env, parent);
1078                         mdt_export_evict(exp);
1079                         RETURN_EXIT;
1080                 }
1081
1082                 if (unlikely(mdt_object_remote(child))) {
1083                         /* the child object was created on remote server */
1084                         if (!mdt_is_dne_client(exp)) {
1085                                 /* Return -EIO for old client */
1086                                 mdt_object_put(env, parent);
1087                                 mdt_object_put(env, child);
1088                                 GOTO(out, rc = -EIO);
1089                         }
1090                         repbody->mbo_fid1 = *rr->rr_fid2;
1091                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1092                         rc = 0;
1093                 } else {
1094                         if (mdt_object_exists(child)) {
1095                                 mdt_set_capainfo(info, 1, rr->rr_fid2,
1096                                                  BYPASS_CAPA);
1097                                 mdt_prep_ma_buf_from_rep(info, child, ma);
1098                                 rc = mdt_attr_get_complex(info, child, ma);
1099                                 if (rc == 0)
1100                                         rc = mdt_finish_open(info, parent,
1101                                                              child, flags,
1102                                                              1, ldlm_rep);
1103                         } else {
1104                                 /* the child does not exist, we should do
1105                                  * regular open */
1106                                 mdt_object_put(env, parent);
1107                                 mdt_object_put(env, child);
1108                                 GOTO(regular_open, 0);
1109                         }
1110                 }
1111                 mdt_object_put(env, parent);
1112                 mdt_object_put(env, child);
1113                 GOTO(out, rc);
1114         } else {
1115 regular_open:
1116                 /* We did not try to create, so we are a pure open */
1117                 rc = mdt_reint_open(info, lhc);
1118         }
1119
1120         EXIT;
1121 out:
1122         req->rq_status = rc;
1123         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1124         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1125 }
1126
1127 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
1128 {
1129         __u64                    flags = info->mti_spec.sp_cr_flags;
1130         struct mdt_reint_record *rr = &info->mti_rr;
1131         struct md_attr          *ma = &info->mti_attr;
1132         struct mdt_object       *o;
1133         int                      rc;
1134         ENTRY;
1135
1136         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1137         if (IS_ERR(o))
1138                 RETURN(rc = PTR_ERR(o));
1139
1140         if (unlikely(mdt_object_remote(o))) {
1141                 /* the child object was created on remote server */
1142                 struct mdt_body *repbody;
1143
1144                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1145                                                 DISP_LOOKUP_EXECD |
1146                                                 DISP_LOOKUP_POS));
1147                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1148                 repbody->mbo_fid1 = *rr->rr_fid2;
1149                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1150                 rc = 0;
1151         } else {
1152                 if (mdt_object_exists(o)) {
1153                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1154                                                         DISP_LOOKUP_EXECD |
1155                                                         DISP_LOOKUP_POS));
1156                         mdt_prep_ma_buf_from_rep(info, o, ma);
1157                         rc = mdt_attr_get_complex(info, o, ma);
1158                         if (rc == 0)
1159                                 rc = mdt_finish_open(info, NULL, o, flags, 0,
1160                                                      rep);
1161                 } else {
1162                         rc = -ENOENT;
1163                 }
1164         }
1165
1166         mdt_object_put(info->mti_env, o);
1167         RETURN(rc);
1168 }
1169
1170 /* lock object for open */
1171 static int mdt_object_open_lock(struct mdt_thread_info *info,
1172                                 struct mdt_object *obj,
1173                                 struct mdt_lock_handle *lhc,
1174                                 __u64 *ibits)
1175 {
1176         struct md_attr  *ma = &info->mti_attr;
1177         __u64            open_flags = info->mti_spec.sp_cr_flags;
1178         ldlm_mode_t      lm = LCK_CR;
1179         bool             acq_lease = !!(open_flags & MDS_OPEN_LEASE);
1180         bool             try_layout = false;
1181         bool             create_layout = false;
1182         int              rc = 0;
1183         ENTRY;
1184
1185         *ibits = 0;
1186         mdt_lock_handle_init(lhc);
1187
1188         if (req_is_replay(mdt_info_req(info)))
1189                 RETURN(0);
1190
1191         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
1192                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
1193                     md_should_create(open_flags))
1194                         create_layout = true;
1195                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
1196                     ma->ma_need & MA_LOV)
1197                         try_layout = true;
1198         }
1199
1200         if (acq_lease) {
1201                 /* lease open, acquire write mode of open sem */
1202                 down_write(&obj->mot_open_sem);
1203
1204                 /* Lease exists and ask for new lease */
1205                 if (atomic_read(&obj->mot_lease_count) > 0) {
1206                         /* only exclusive open is supported, so lease
1207                          * are conflicted to each other */
1208                         GOTO(out, rc = -EBUSY);
1209                 }
1210
1211                 /* Lease must be with open lock */
1212                 if (!(open_flags & MDS_OPEN_LOCK)) {
1213                         CERROR("Request lease for file:"DFID ", but open lock "
1214                                 "is missed, open_flags = "LPO64".\n",
1215                                 PFID(mdt_object_fid(obj)), open_flags);
1216                         GOTO(out, rc = -EPROTO);
1217                 }
1218
1219                 /* XXX: only exclusive open is supported. */
1220                 lm = LCK_EX;
1221                 *ibits = MDS_INODELOCK_OPEN;
1222
1223                 /* never grant LCK_EX layout lock to client */
1224                 try_layout = false;
1225         } else { /* normal open */
1226                 /* normal open holds read mode of open sem */
1227                 down_read(&obj->mot_open_sem);
1228
1229                 if (open_flags & MDS_OPEN_LOCK) {
1230                         if (open_flags & FMODE_WRITE)
1231                                 lm = LCK_CW;
1232                         else if (open_flags & MDS_FMODE_EXEC)
1233                                 lm = LCK_PR;
1234                         else
1235                                 lm = LCK_CR;
1236
1237                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
1238                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
1239                         if (open_flags & FMODE_WRITE)
1240                                 lm = LCK_CW;
1241                         else
1242                                 lm = LCK_CR;
1243
1244                         /* revoke lease */
1245                         *ibits = MDS_INODELOCK_OPEN;
1246                         try_layout = false;
1247
1248                         lhc = &info->mti_lh[MDT_LH_LOCAL];
1249                 }
1250                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
1251                         PFID(mdt_object_fid(obj)),
1252                         atomic_read(&obj->mot_open_count), lm);
1253         }
1254
1255         mdt_lock_reg_init(lhc, lm);
1256
1257         /* one problem to return layout lock on open is that it may result
1258          * in too many layout locks cached on the client side. */
1259         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
1260                 /* return lookup lock to validate inode at the client side,
1261                  * this is pretty important otherwise mdt will return layout
1262                  * lock for each open.
1263                  * However this is a double-edged sword because changing
1264                  * permission will revoke huge # of LOOKUP locks. */
1265                 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
1266                 if (!mdt_object_lock_try(info, obj, lhc, *ibits,
1267                                          MDT_CROSS_LOCK)) {
1268                         *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
1269                         if (*ibits != 0)
1270                                 rc = mdt_object_lock(info, obj, lhc, *ibits,
1271                                                 MDT_CROSS_LOCK);
1272                 }
1273         } else if (*ibits != 0) {
1274                 rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK);
1275         }
1276
1277         CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
1278                 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
1279                 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
1280
1281         /* will change layout, revoke layout locks by enqueuing EX lock. */
1282         if (rc == 0 && create_layout) {
1283                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
1284
1285                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
1286                         ", open_flags = "LPO64"\n",
1287                         PFID(mdt_object_fid(obj)), open_flags);
1288
1289                 /* We cannot enqueue another lock for the same resource we
1290                  * already have a lock for, due to mechanics of waiting list
1291                  * iterating in ldlm, see LU-3601.
1292                  * As such we'll drop the open lock we just got above here,
1293                  * it's ok not to have this open lock as it's main purpose is to
1294                  * flush unused cached client open handles. */
1295                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1296                         mdt_object_unlock(info, obj, lhc, 1);
1297
1298                 LASSERT(!try_layout);
1299                 mdt_lock_handle_init(ll);
1300                 mdt_lock_reg_init(ll, LCK_EX);
1301                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
1302                                         MDT_LOCAL_LOCK);
1303
1304                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
1305         }
1306
1307         /* Check if there is any other open handles after acquiring
1308          * open lock. At this point, caching open handles have been revoked
1309          * by open lock.
1310          * XXX: Now only exclusive open is supported. Need to check the
1311          * type of open for generic lease support. */
1312         if (rc == 0 && acq_lease) {
1313                 struct ptlrpc_request *req = mdt_info_req(info);
1314                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1315                 struct mdt_file_data *mfd;
1316                 bool is_replay_or_resent;
1317                 int open_count = 0;
1318
1319                 /* For lease: application can open a file and then apply lease,
1320                  * @handle contains original open handle in that case.
1321                  * In recovery, open REQ will be replayed and the lease REQ may
1322                  * be resent that means the open handle is already stale, so we
1323                  * need to fix it up here by finding new handle. */
1324                 is_replay_or_resent = req_is_replay(req) ||
1325                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
1326
1327                 /* if the request is _not_ a replay request, rr_handle
1328                  * may be used to hold an openhandle which is issuing the
1329                  * lease request, so that this openhandle doesn't count. */
1330                 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
1331                                      is_replay_or_resent);
1332                 if (mfd != NULL)
1333                         ++open_count;
1334
1335                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
1336                         PFID(mdt_object_fid(obj)),
1337                         atomic_read(&obj->mot_open_count), open_count);
1338
1339                 if (atomic_read(&obj->mot_open_count) > open_count)
1340                         GOTO(out, rc = -EBUSY);
1341         }
1342         GOTO(out, rc);
1343
1344 out:
1345         RETURN(rc);
1346 }
1347
1348 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1349                                    struct mdt_object *obj,
1350                                    struct mdt_lock_handle *lhc,
1351                                    __u64 ibits, int rc)
1352 {
1353         __u64 open_flags = info->mti_spec.sp_cr_flags;
1354         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1355         ENTRY;
1356
1357         if (req_is_replay(mdt_info_req(info)))
1358                 RETURN_EXIT;
1359
1360         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1361          * return to client side. */
1362         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1363                 mdt_object_unlock(info, obj, ll, 1);
1364
1365         ll = &info->mti_lh[MDT_LH_LAYOUT];
1366         /* Release local layout lock, layout was created */
1367         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1368                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1369                 mdt_object_unlock(info, obj, ll, 1);
1370         }
1371
1372         if (open_flags & MDS_OPEN_LEASE)
1373                 up_write(&obj->mot_open_sem);
1374         else
1375                 up_read(&obj->mot_open_sem);
1376
1377         /* Cross-ref case, the lock should be returned to the client */
1378         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1379                 RETURN_EXIT;
1380
1381         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
1382                 /* for the open request, the lock will only return to client
1383                  * if open or layout lock is granted. */
1384                 rc = 1;
1385         }
1386
1387         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1388                 struct ldlm_reply       *ldlm_rep;
1389
1390                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1391                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1392                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1393                         mdt_object_unlock(info, obj, lhc, 1);
1394         }
1395         RETURN_EXIT;
1396 }
1397
1398 /**
1399  * Check release is permitted for the current HSM flags.
1400  */
1401 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1402 {
1403         if (!(ma->ma_valid & MA_HSM))
1404                 return false;
1405
1406         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1407                 return false;
1408
1409         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1410                 return false;
1411
1412         return true;
1413 }
1414
1415 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1416                                 struct ldlm_reply *rep,
1417                                 struct mdt_lock_handle *lhc)
1418 {
1419         const struct lu_env     *env   = info->mti_env;
1420         struct mdt_device       *mdt   = info->mti_mdt;
1421         __u64                    flags = info->mti_spec.sp_cr_flags;
1422         struct mdt_reint_record *rr    = &info->mti_rr;
1423         struct md_attr          *ma    = &info->mti_attr;
1424         struct mdt_object       *parent= NULL;
1425         struct mdt_object       *o;
1426         int                      rc;
1427         int                      object_locked = 0;
1428         __u64                    ibits = 0;
1429         ENTRY;
1430
1431         if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1432                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1433                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1434                         if (IS_ERR(parent)) {
1435                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1436                                        " for anonymous created %ld, try to"
1437                                        " use server-side parent.\n",
1438                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1439                                 parent = NULL;
1440                         }
1441                 }
1442                 if (parent == NULL)
1443                         ma->ma_need |= MA_PFID;
1444         }
1445
1446         o = mdt_object_find(env, mdt, rr->rr_fid2);
1447         if (IS_ERR(o))
1448                 GOTO(out_parent_put, rc = PTR_ERR(o));
1449
1450         if (mdt_object_remote(o)) {
1451                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1452                        mdt_obd_name(info->mti_mdt),
1453                        PFID(rr->rr_fid2));
1454                 GOTO(out, rc = -EREMOTE);
1455         } else if (!mdt_object_exists(o)) {
1456                 mdt_set_disposition(info, rep,
1457                                     DISP_IT_EXECD |
1458                                     DISP_LOOKUP_EXECD |
1459                                     DISP_LOOKUP_NEG);
1460                 GOTO(out, rc = -ENOENT);
1461         }
1462
1463         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1464
1465         mdt_prep_ma_buf_from_rep(info, o, ma);
1466         if (flags & MDS_OPEN_RELEASE)
1467                 ma->ma_need |= MA_HSM;
1468         rc = mdt_attr_get_complex(info, o, ma);
1469         if (rc)
1470                 GOTO(out, rc);
1471
1472         /* If a release request, check file flags are fine and ask for an
1473          * exclusive open access. */
1474         if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1475                 GOTO(out, rc = -EPERM);
1476
1477         rc = mdt_check_resent_lock(info, o, lhc);
1478         if (rc < 0) {
1479                 GOTO(out, rc);
1480         } else if (rc > 0) {
1481                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1482                 object_locked = 1;
1483                 if (rc)
1484                         GOTO(out_unlock, rc);
1485         }
1486
1487         if (ma->ma_valid & MA_PFID) {
1488                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1489                 if (IS_ERR(parent)) {
1490                         CDEBUG(D_INODE, "Fail to find parent "DFID
1491                                " for anonymous created %ld, try to"
1492                                " use system default.\n",
1493                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1494                         parent = NULL;
1495                 }
1496         }
1497
1498         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1499         if (!rc) {
1500                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1501                 if (flags & MDS_OPEN_LOCK)
1502                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1503                 if (flags & MDS_OPEN_LEASE)
1504                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1505         }
1506         GOTO(out_unlock, rc);
1507
1508 out_unlock:
1509         if (object_locked)
1510                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1511 out:
1512         mdt_object_put(env, o);
1513 out_parent_put:
1514         if (parent != NULL)
1515                 mdt_object_put(env, parent);
1516         return rc;
1517 }
1518
1519 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1520 static int mdt_cross_open(struct mdt_thread_info *info,
1521                           const struct lu_fid *parent_fid,
1522                           const struct lu_fid *fid,
1523                           struct ldlm_reply *rep, __u32 flags)
1524 {
1525         struct md_attr    *ma = &info->mti_attr;
1526         struct mdt_object *o;
1527         int                rc;
1528         ENTRY;
1529
1530         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1531         if (IS_ERR(o))
1532                 RETURN(rc = PTR_ERR(o));
1533
1534         if (mdt_object_remote(o)) {
1535                 /* Something is wrong here, the object is on another MDS! */
1536                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1537                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1538                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1539                                 &o->mot_obj,
1540                                 "Object isn't on this server! FLD error?\n");
1541                 rc = -EFAULT;
1542         } else {
1543                 if (mdt_object_exists(o)) {
1544                         /* Do permission check for cross-open. */
1545                         rc = mo_permission(info->mti_env, NULL,
1546                                            mdt_object_child(o),
1547                                            NULL, flags | MDS_OPEN_CROSS);
1548                         if (rc)
1549                                 goto out;
1550
1551                         mdt_prep_ma_buf_from_rep(info, o, ma);
1552                         mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1553                         rc = mdt_attr_get_complex(info, o, ma);
1554                         if (rc != 0)
1555                                 GOTO(out, rc);
1556
1557                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1558                 } else {
1559                         /*
1560                          * Something is wrong here. lookup was positive but
1561                          * there is no object!
1562                          */
1563                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1564                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1565                         rc = -EFAULT;
1566                 }
1567         }
1568 out:
1569         mdt_object_put(info->mti_env, o);
1570         RETURN(rc);
1571 }
1572
1573 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1574 {
1575         struct mdt_device       *mdt = info->mti_mdt;
1576         struct ptlrpc_request   *req = mdt_info_req(info);
1577         struct mdt_object       *parent;
1578         struct mdt_object       *child;
1579         struct mdt_lock_handle  *lh;
1580         struct ldlm_reply       *ldlm_rep;
1581         struct mdt_body         *repbody;
1582         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1583         struct md_attr          *ma = &info->mti_attr;
1584         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1585         __u64                    ibits = 0;
1586         struct mdt_reint_record *rr = &info->mti_rr;
1587         int                      result, rc;
1588         int                      created = 0;
1589         int                      object_locked = 0;
1590         __u32                    msg_flags;
1591         ENTRY;
1592
1593         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1594                                (obd_timeout + 1) / 4);
1595
1596         mdt_counter_incr(req, LPROC_MDT_OPEN);
1597         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1598
1599         ma->ma_need = MA_INODE;
1600         ma->ma_valid = 0;
1601
1602         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1603         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1604
1605         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1606                 CERROR("file join is not supported anymore.\n");
1607                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1608         }
1609         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1610
1611         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1612             info->mti_spec.u.sp_ea.eadata == NULL)
1613                 GOTO(out, result = err_serious(-EINVAL));
1614
1615         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1616                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1617                PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1618                PFID(rr->rr_fid2), create_flags,
1619                ma->ma_attr.la_mode, msg_flags);
1620
1621         if (info->mti_cross_ref) {
1622                 /* This is cross-ref open */
1623                 mdt_set_disposition(info, ldlm_rep,
1624                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1625                              DISP_LOOKUP_POS));
1626                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1627                                         ldlm_rep, create_flags);
1628                 GOTO(out, result);
1629         } else if (req_is_replay(req) ||
1630             (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1631                 /* This is a replay request or from liblustre with ea. */
1632                 result = mdt_open_by_fid(info, ldlm_rep);
1633
1634                 if (result != -ENOENT) {
1635                         if (req->rq_export->exp_libclient &&
1636                             create_flags & MDS_OPEN_HAS_EA)
1637                                 GOTO(out, result = 0);
1638                         GOTO(out, result);
1639                 }
1640                 /* We didn't find the correct object, so we need to re-create it
1641                  * via a regular replay. */
1642                 if (!(create_flags & MDS_OPEN_CREAT)) {
1643                         DEBUG_REQ(D_ERROR, req,
1644                                   "OPEN & CREAT not in open replay/by_fid.");
1645                         GOTO(out, result = -EFAULT);
1646                 }
1647                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1648         } else if (create_flags & (MDS_OPEN_BY_FID | MDS_OPEN_LOCK)) {
1649                 /*
1650                  * MDS_OPEN_LOCK is checked for backward compatibility with 2.1
1651                  * client.
1652                  */
1653                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1654                 if (result < 0)
1655                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1656                                PFID(rr->rr_fid2), result);
1657                 GOTO(out, result);
1658         }
1659
1660         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1661                 GOTO(out, result = err_serious(-ENOMEM));
1662
1663         mdt_set_disposition(info, ldlm_rep,
1664                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1665
1666         if (!lu_name_is_valid(&rr->rr_name))
1667                 GOTO(out, result = -EPROTO);
1668
1669         lh = &info->mti_lh[MDT_LH_PARENT];
1670         mdt_lock_pdo_init(lh,
1671                           (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1672                           &rr->rr_name);
1673
1674         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1675                                       MDS_INODELOCK_UPDATE);
1676         if (IS_ERR(parent))
1677                 GOTO(out, result = PTR_ERR(parent));
1678
1679         /* get and check version of parent */
1680         result = mdt_version_get_check(info, parent, 0);
1681         if (result)
1682                 GOTO(out_parent, result);
1683
1684         fid_zero(child_fid);
1685
1686         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1687                             &rr->rr_name, child_fid, &info->mti_spec);
1688
1689         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1690                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1691                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1692                  PFID(child_fid));
1693
1694         if (result != 0 && result != -ENOENT && result != -ESTALE)
1695                 GOTO(out_parent, result);
1696
1697         if (result == -ENOENT || result == -ESTALE) {
1698                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1699                 if (result == -ESTALE) {
1700                         /*
1701                          * -ESTALE means the parent is a dead(unlinked) dir, so
1702                          * it should return -ENOENT to in accordance with the
1703                          * original mds implementaion.
1704                          */
1705                         GOTO(out_parent, result = -ENOENT);
1706                 }
1707                 if (!(create_flags & MDS_OPEN_CREAT))
1708                         GOTO(out_parent, result);
1709                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1710                         GOTO(out_parent, result = -EROFS);
1711                 *child_fid = *info->mti_rr.rr_fid2;
1712                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1713                          PFID(child_fid));
1714                 /* In the function below, .hs_keycmp resolves to
1715                  * lu_obj_hop_keycmp() */
1716                 /* coverity[overrun-buffer-val] */
1717                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1718         } else {
1719                 /*
1720                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1721                  * return FID back in that case.
1722                  */
1723                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1724                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1725         }
1726         if (IS_ERR(child))
1727                 GOTO(out_parent, result = PTR_ERR(child));
1728
1729         /** check version of child  */
1730         rc = mdt_version_get_check(info, child, 1);
1731         if (rc)
1732                 GOTO(out_child, result = rc);
1733
1734         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1735         if (result == -ENOENT) {
1736                 /* Create under OBF and .lustre is not permitted */
1737                 if (!fid_is_md_operative(rr->rr_fid1))
1738                         GOTO(out_child, result = -EPERM);
1739
1740                 /* save versions in reply */
1741                 mdt_version_get_save(info, parent, 0);
1742                 mdt_version_get_save(info, child, 1);
1743
1744                 /* version of child will be changed */
1745                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1746
1747                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1748                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1749
1750                 /* Let lower layers know what is lock mode on directory. */
1751                 info->mti_spec.sp_cr_mode =
1752                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1753
1754                 /*
1755                  * Do not perform lookup sanity check. We know that name does
1756                  * not exist.
1757                  */
1758                 info->mti_spec.sp_cr_lookup = 0;
1759                 info->mti_spec.sp_feat = &dt_directory_features;
1760
1761                 result = mdo_create(info->mti_env,
1762                                     mdt_object_child(parent),
1763                                     &rr->rr_name,
1764                                     mdt_object_child(child),
1765                                     &info->mti_spec,
1766                                     &info->mti_attr);
1767                 if (result == -ERESTART) {
1768                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1769                         GOTO(out_child, result);
1770                 } else {
1771                         mdt_prep_ma_buf_from_rep(info, child, ma);
1772                         /* XXX: we should call this once, see few lines below */
1773                         if (result == 0)
1774                                 result = mdt_attr_get_complex(info, child, ma);
1775
1776                         if (result != 0)
1777                                 GOTO(out_child, result);
1778                 }
1779                 created = 1;
1780         } else {
1781                 /*
1782                  * The object is on remote node, return its FID for remote open.
1783                  */
1784                 if (mdt_object_remote(child)) {
1785                         /*
1786                          * Check if this lock already was sent to client and
1787                          * this is resent case. For resent case do not take lock
1788                          * again, use what is already granted.
1789                          */
1790                         LASSERT(lhc != NULL);
1791
1792                         rc = mdt_check_resent_lock(info, child, lhc);
1793                         if (rc < 0) {
1794                                 GOTO(out_child, result = rc);
1795                         } else if (rc > 0) {
1796                                 mdt_lock_handle_init(lhc);
1797                                 mdt_lock_reg_init(lhc, LCK_PR);
1798
1799                                 rc = mdt_object_lock(info, child, lhc,
1800                                                      MDS_INODELOCK_LOOKUP,
1801                                                      MDT_CROSS_LOCK);
1802                         }
1803                         repbody->mbo_fid1 = *mdt_object_fid(child);
1804                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1805                         if (rc != 0)
1806                                 result = rc;
1807                         else
1808                                 result = -MDT_EREMOTE_OPEN;
1809                         GOTO(out_child, result);
1810                 } else if (mdt_object_exists(child)) {
1811                         /* We have to get attr & LOV EA & HSM for this
1812                          * object. */
1813                         mdt_prep_ma_buf_from_rep(info, child, ma);
1814                         ma->ma_need |= MA_HSM;
1815                         result = mdt_attr_get_complex(info, child, ma);
1816                         if (result != 0)
1817                                 GOTO(out_child, result);
1818                 } else {
1819                         /* Object does not exist. Likely FS corruption. */
1820                         CERROR("%s: name '"DNAME"' present, but FID "
1821                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1822                                PNAME(&rr->rr_name), PFID(child_fid));
1823                         GOTO(out_child, result = -EIO);
1824                 }
1825         }
1826
1827         rc = mdt_check_resent_lock(info, child, lhc);
1828         if (rc < 0) {
1829                 GOTO(out_child, result = rc);
1830         } else if (rc == 0) {
1831                 /* the open lock might already be gotten in
1832                  * ldlm_handle_enqueue() */
1833                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1834                 if (create_flags & MDS_OPEN_LOCK)
1835                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1836         } else {
1837                 /* get openlock if this isn't replay and client requested it */
1838                 if (!req_is_replay(req)) {
1839                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1840                         object_locked = 1;
1841                         if (rc != 0)
1842                                 GOTO(out_child_unlock, result = rc);
1843                         else if (create_flags & MDS_OPEN_LOCK)
1844                                 mdt_set_disposition(info, ldlm_rep,
1845                                                     DISP_OPEN_LOCK);
1846                 }
1847         }
1848         /* Try to open it now. */
1849         rc = mdt_finish_open(info, parent, child, create_flags,
1850                              created, ldlm_rep);
1851         if (rc) {
1852                 result = rc;
1853                 /* openlock will be released if mdt_finish_open failed */
1854                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1855
1856                 if (created && create_flags & MDS_OPEN_VOLATILE) {
1857                         CERROR("%s: cannot open volatile file "DFID", orphan "
1858                                "file will be left in PENDING directory until "
1859                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1860                                PFID(mdt_object_fid(child)), rc);
1861                         GOTO(out_child_unlock, result);
1862                 }
1863
1864                 if (created) {
1865                         ma->ma_need = 0;
1866                         ma->ma_valid = 0;
1867                         rc = mdo_unlink(info->mti_env,
1868                                         mdt_object_child(parent),
1869                                         mdt_object_child(child),
1870                                         &rr->rr_name,
1871                                         &info->mti_attr, 0);
1872                         if (rc != 0)
1873                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1874                                        mdt_obd_name(info->mti_mdt),
1875                                        PFID(mdt_object_fid(child)), rc);
1876                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1877                 }
1878         }
1879         EXIT;
1880 out_child_unlock:
1881         if (object_locked)
1882                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1883 out_child:
1884         mdt_object_put(info->mti_env, child);
1885 out_parent:
1886         mdt_object_unlock_put(info, parent, lh, result || !created);
1887 out:
1888         if (result)
1889                 lustre_msg_set_transno(req->rq_repmsg, 0);
1890         return result;
1891 }
1892
1893 /**
1894  * Create an orphan object use local root.
1895  */
1896 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1897                                           struct mdt_device *mdt,
1898                                           const struct lu_fid *fid,
1899                                           struct md_attr *attr, fmode_t fmode)
1900 {
1901         const struct lu_env *env = info->mti_env;
1902         struct md_op_spec *spec = &info->mti_spec;
1903         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1904         struct mdt_object *obj = NULL;
1905         struct mdt_object *local_root;
1906         static const struct lu_name lname = {
1907                 .ln_name = "i_am_nobody",
1908                 .ln_namelen = sizeof("i_am_nobody") - 1,
1909         };
1910         struct lu_ucred *uc;
1911         cfs_cap_t uc_cap_save;
1912         int rc;
1913         ENTRY;
1914
1915         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1916         if (rc != 0)
1917                 RETURN(ERR_PTR(rc));
1918
1919         local_root = mdt_object_find(env, mdt, local_root_fid);
1920         if (IS_ERR(local_root))
1921                 RETURN(local_root);
1922
1923         obj = mdt_object_new(env, mdt, fid);
1924         if (IS_ERR(obj))
1925                 GOTO(out, rc = PTR_ERR(obj));
1926
1927         spec->sp_cr_lookup = 0;
1928         spec->sp_feat = &dt_directory_features;
1929         spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1930         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1931         if (attr->ma_valid & MA_LOV) {
1932                 spec->u.sp_ea.eadata = attr->ma_lmm;
1933                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1934                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1935         } else {
1936                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1937         }
1938
1939         uc = lu_ucred(env);
1940         uc_cap_save = uc->uc_cap;
1941         uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1942         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1943                         mdt_object_child(obj), spec, attr);
1944         uc->uc_cap = uc_cap_save;
1945         if (rc < 0) {
1946                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1947                        mdt_obd_name(mdt), PFID(fid), rc);
1948                 GOTO(out, rc);
1949         }
1950
1951         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1952         if (rc < 0)
1953                 CERROR("%s: cannot open volatile file "DFID", orphan "
1954                        "file will be left in PENDING directory until "
1955                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1956                        PFID(fid), rc);
1957         GOTO(out, rc);
1958
1959 out:
1960         if (rc < 0) {
1961                 if (!IS_ERR(obj))
1962                         mdt_object_put(env, obj);
1963                 obj = ERR_PTR(rc);
1964         }
1965         mdt_object_put(env, local_root);
1966         return obj;
1967 }
1968
1969 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1970                            struct md_attr *ma)
1971 {
1972         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1973         struct close_data      *data;
1974         struct ldlm_lock       *lease;
1975         struct mdt_object      *orphan;
1976         struct md_attr         *orp_ma;
1977         struct lu_buf          *buf;
1978         bool                    lease_broken;
1979         int                     rc;
1980         int                     rc2;
1981         ENTRY;
1982
1983         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1984                 RETURN(-EROFS);
1985
1986         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1987         if (data == NULL)
1988                 RETURN(-EPROTO);
1989
1990         lease = ldlm_handle2lock(&data->cd_handle);
1991         if (lease == NULL)
1992                 RETURN(-ESTALE);
1993
1994         /* try to hold open_sem so that nobody else can open the file */
1995         if (!down_write_trylock(&o->mot_open_sem)) {
1996                 ldlm_lock_cancel(lease);
1997                 GOTO(out_reprocess, rc = -EBUSY);
1998         }
1999
2000         /* Check if the lease open lease has already canceled */
2001         lock_res_and_lock(lease);
2002         lease_broken = ldlm_is_cancel(lease);
2003         unlock_res_and_lock(lease);
2004
2005         LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2006                    PFID(mdt_object_fid(o)), lease_broken);
2007
2008         /* Cancel server side lease. Client side counterpart should
2009          * have been cancelled. It's okay to cancel it now as we've
2010          * held mot_open_sem. */
2011         ldlm_lock_cancel(lease);
2012
2013         if (lease_broken) /* don't perform release task */
2014                 GOTO(out_unlock, rc = -ESTALE);
2015
2016         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2017                 GOTO(out_unlock, rc = -EINVAL);
2018
2019         /* ma_need was set before but it seems fine to change it in order to
2020          * avoid modifying the one from RPC */
2021         ma->ma_need = MA_HSM;
2022         rc = mdt_attr_get_complex(info, o, ma);
2023         if (rc != 0)
2024                 GOTO(out_unlock, rc);
2025
2026         if (!mdt_hsm_release_allow(ma))
2027                 GOTO(out_unlock, rc = -EPERM);
2028
2029         /* already released? */
2030         if (ma->ma_hsm.mh_flags & HS_RELEASED)
2031                 GOTO(out_unlock, rc = 0);
2032
2033         /* Compare on-disk and packed data_version */
2034         if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
2035                 CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
2036                        " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
2037                        data->cd_data_version, ma->ma_hsm.mh_arch_ver);
2038                 GOTO(out_unlock, rc = -EPERM);
2039         }
2040
2041         ma->ma_valid = MA_INODE;
2042         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
2043         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
2044         if (rc < 0)
2045                 GOTO(out_unlock, rc);
2046
2047         ma->ma_need = MA_INODE | MA_LOV;
2048         rc = mdt_attr_get_complex(info, o, ma);
2049         if (rc < 0)
2050                 GOTO(out_unlock, rc);
2051
2052         if (!(ma->ma_valid & MA_LOV)) {
2053                 /* Even empty file are released */
2054                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2055                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2056                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2057                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2058                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2059         } else {
2060                 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
2061                  * ma_lmm as lov_user_md, then it will be confused by union of
2062                  * layout_gen and stripe_offset. */
2063                 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
2064                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2065                 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
2066                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
2067                 else
2068                         GOTO(out_unlock, rc = -EINVAL);
2069         }
2070
2071         /* Set file as released */
2072         ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
2073
2074         /* Hopefully it's not used in this call path */
2075         orp_ma = &info->mti_u.som.attr;
2076         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2077         orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
2078         orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
2079         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2080         orp_ma->ma_lmm = ma->ma_lmm;
2081         orp_ma->ma_lmm_size = ma->ma_lmm_size;
2082         orp_ma->ma_valid = MA_INODE | MA_LOV;
2083         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2084                                  FMODE_WRITE);
2085         if (IS_ERR(orphan)) {
2086                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2087                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2088                        PTR_ERR(orphan));
2089                 GOTO(out_unlock, rc = PTR_ERR(orphan));
2090         }
2091
2092         /* Set up HSM attribute for orphan object */
2093         CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
2094         buf = &info->mti_buf;
2095         buf->lb_buf = info->mti_xattr_buf;
2096         buf->lb_len = sizeof(struct hsm_attrs);
2097         ma->ma_hsm.mh_flags |= HS_RELEASED;
2098         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2099         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2100
2101         mdt_lock_reg_init(lh, LCK_EX);
2102         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2103                              MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
2104         if (rc != 0)
2105                 GOTO(out_close, rc);
2106
2107         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2108                           XATTR_NAME_HSM, 0);
2109
2110         if (rc == 0)
2111                 /* Swap layout with orphan object */
2112                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2113                                      mdt_object_child(orphan),
2114                                      SWAP_LAYOUTS_MDS_HSM);
2115
2116         /* Release exclusive LL */
2117         mdt_object_unlock(info, o, lh, 1);
2118
2119         EXIT;
2120
2121 out_close:
2122         /* Close orphan object anyway */
2123         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2124                        FMODE_WRITE);
2125         if (rc2 < 0)
2126                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2127                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2128         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2129                         "object closed\n");
2130         mdt_object_put(info->mti_env, orphan);
2131
2132 out_unlock:
2133         up_write(&o->mot_open_sem);
2134
2135         if (rc == 0) { /* already released */
2136                 struct mdt_body *repbody;
2137                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2138                 LASSERT(repbody != NULL);
2139                 repbody->mbo_valid |= OBD_MD_FLRELEASED;
2140         }
2141
2142 out_reprocess:
2143         ldlm_reprocess_all(lease->l_resource);
2144         LDLM_LOCK_PUT(lease);
2145
2146         ma->ma_valid = 0;
2147         ma->ma_need = 0;
2148
2149         return rc;
2150 }
2151
2152 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
2153                                       MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
2154
2155 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2156 {
2157         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
2158 }
2159
2160 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2161 {
2162         struct mdt_object *o = mfd->mfd_object;
2163         struct md_object *next = mdt_object_child(o);
2164         struct md_attr *ma = &info->mti_attr;
2165         int ret = MDT_IOEPOCH_CLOSED;
2166         int rc = 0;
2167         __u64 mode;
2168         ENTRY;
2169
2170         mode = mfd->mfd_mode;
2171
2172         if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
2173                 rc = mdt_hsm_release(info, o, ma);
2174                 if (rc < 0) {
2175                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2176                                 mdt_obd_name(info->mti_mdt),
2177                                 PFID(mdt_object_fid(o)), rc);
2178                         /* continue to close even error occurred. */
2179                 }
2180         }
2181
2182         if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
2183                 mdt_write_put(o);
2184                 ret = mdt_ioepoch_close(info, o);
2185         } else if (mode & MDS_FMODE_EXEC) {
2186                 mdt_write_allow(o);
2187         } else if (mode & MDS_FMODE_EPOCH) {
2188                 ret = mdt_ioepoch_close(info, o);
2189         } else if (mode & MDS_FMODE_SOM) {
2190                 ret = mdt_som_au_close(info, o);
2191         }
2192
2193         /* Update atime on close only. */
2194         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
2195             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
2196                 /* Set the atime only. */
2197                 ma->ma_valid = MA_INODE;
2198                 ma->ma_attr.la_valid = LA_ATIME;
2199                 rc = mo_attr_set(info->mti_env, next, ma);
2200         }
2201
2202         /* If file data is modified, add the dirty flag. */
2203         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2204                 rc = mdt_add_dirty_flag(info, o, ma);
2205
2206         ma->ma_need |= MA_INODE;
2207         ma->ma_valid &= ~MA_INODE;
2208
2209         if (!MFD_CLOSED(mode))
2210                 rc = mo_close(info->mti_env, next, ma, mode);
2211
2212         if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
2213                 struct mdt_export_data *med;
2214
2215                 /* The IOepoch is still opened or SOM update is needed.
2216                  * Put mfd back into the list. */
2217                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
2218                 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
2219                                       MDS_FMODE_EPOCH : MDS_FMODE_SOM);
2220
2221                 LASSERT(mdt_info_req(info));
2222                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
2223                 spin_lock(&med->med_open_lock);
2224                 list_add(&mfd->mfd_list, &med->med_open_head);
2225                 class_handle_hash_back(&mfd->mfd_handle);
2226                 spin_unlock(&med->med_open_lock);
2227
2228                 if (ret == MDT_IOEPOCH_OPENED) {
2229                         ret = 0;
2230                 } else {
2231                         ret = -EAGAIN;
2232                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
2233                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
2234                 }
2235         } else {
2236                 /* adjust open and lease count */
2237                 if (mode & MDS_OPEN_LEASE) {
2238                         LASSERT(atomic_read(&o->mot_lease_count) > 0);
2239                         atomic_dec(&o->mot_lease_count);
2240                 }
2241                 LASSERT(atomic_read(&o->mot_open_count) > 0);
2242                 atomic_dec(&o->mot_open_count);
2243
2244                 mdt_mfd_free(mfd);
2245                 mdt_object_put(info->mti_env, o);
2246         }
2247
2248         RETURN(rc ? rc : ret);
2249 }
2250
2251 int mdt_close(struct tgt_session_info *tsi)
2252 {
2253         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2254         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2255         struct mdt_export_data *med;
2256         struct mdt_file_data   *mfd;
2257         struct mdt_object      *o;
2258         struct md_attr         *ma = &info->mti_attr;
2259         struct mdt_body        *repbody = NULL;
2260         int rc, ret = 0;
2261         ENTRY;
2262
2263         mdt_counter_incr(req, LPROC_MDT_CLOSE);
2264         /* Close may come with the Size-on-MDS update. Unpack it. */
2265         rc = mdt_close_unpack(info);
2266         if (rc)
2267                 GOTO(out, rc = err_serious(rc));
2268
2269         LASSERT(info->mti_ioepoch);
2270
2271         /* These fields are no longer used and are left for compatibility.
2272          * size is always zero */
2273         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2274                              0);
2275         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2276                              0);
2277         rc = req_capsule_server_pack(info->mti_pill);
2278         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2279                 mdt_client_compatibility(info);
2280                 if (rc == 0)
2281                         mdt_fix_reply(info);
2282                 mdt_exit_ucred(info);
2283                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2284         }
2285
2286         /* Continue to close handle even if we can not pack reply */
2287         if (rc == 0) {
2288                 repbody = req_capsule_server_get(info->mti_pill,
2289                                                  &RMF_MDT_BODY);
2290                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2291                                                     &RMF_MDT_MD);
2292                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2293                                                        &RMF_MDT_MD,
2294                                                        RCL_SERVER);
2295                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
2296                 repbody->mbo_eadatasize = 0;
2297                 repbody->mbo_aclsize = 0;
2298         } else {
2299                 rc = err_serious(rc);
2300         }
2301
2302         med = &req->rq_export->exp_mdt_data;
2303         spin_lock(&med->med_open_lock);
2304         mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2305                              req_is_replay(req));
2306         if (mdt_mfd_closed(mfd)) {
2307                 spin_unlock(&med->med_open_lock);
2308                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2309                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
2310                        info->mti_ioepoch->handle.cookie);
2311                 /** not serious error since bug 3633 */
2312                 rc = -ESTALE;
2313         } else {
2314                 class_handle_unhash(&mfd->mfd_handle);
2315                 list_del_init(&mfd->mfd_list);
2316                 spin_unlock(&med->med_open_lock);
2317
2318                 /* Do not lose object before last unlink. */
2319                 o = mfd->mfd_object;
2320                 mdt_object_get(info->mti_env, o);
2321                 ret = mdt_mfd_close(info, mfd);
2322                 if (repbody != NULL)
2323                         rc = mdt_handle_last_unlink(info, o, ma);
2324                 mdt_empty_transno(info, rc);
2325                 mdt_object_put(info->mti_env, o);
2326         }
2327         if (repbody != NULL) {
2328                 mdt_client_compatibility(info);
2329                 rc = mdt_fix_reply(info);
2330         }
2331
2332         mdt_exit_ucred(info);
2333         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2334                 GOTO(out, rc = err_serious(-ENOMEM));
2335
2336         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2337                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2338                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2339 out:
2340         mdt_thread_info_fini(info);
2341         RETURN(rc ? rc : ret);
2342 }
2343
2344 /**
2345  * DONE_WRITING rpc handler.
2346  *
2347  * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
2348  * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
2349  * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
2350  * skip attributes and reconstruct the reply here.
2351  */
2352 int mdt_done_writing(struct tgt_session_info *tsi)
2353 {
2354         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2355         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2356         struct mdt_body         *repbody = NULL;
2357         struct mdt_export_data  *med;
2358         struct mdt_file_data    *mfd;
2359         int rc;
2360         ENTRY;
2361
2362         rc = req_capsule_server_pack(tsi->tsi_pill);
2363         if (rc)
2364                 GOTO(out, rc = err_serious(rc));
2365
2366         repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2367         repbody->mbo_eadatasize = 0;
2368         repbody->mbo_aclsize = 0;
2369
2370         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
2371         rc = mdt_close_unpack(info);
2372         if (rc)
2373                 GOTO(out, rc = err_serious(rc));
2374
2375         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2376                 mdt_exit_ucred(info);
2377                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2378         }
2379
2380         med = &info->mti_exp->exp_mdt_data;
2381         spin_lock(&med->med_open_lock);
2382         mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2383                              req_is_replay(req));
2384         if (mfd == NULL) {
2385                 spin_unlock(&med->med_open_lock);
2386                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
2387                        ": cookie = "LPX64" ioepoch = "LPU64"\n",
2388                        PFID(info->mti_rr.rr_fid1),
2389                        info->mti_ioepoch->handle.cookie,
2390                        info->mti_ioepoch->ioepoch);
2391                 /* If this is a replay, reconstruct the transno. */
2392                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2393                         rc = info->mti_ioepoch->flags & MF_SOM_AU ?
2394                              -EAGAIN : 0;
2395                         mdt_empty_transno(info, rc);
2396                 } else
2397                         rc = -ESTALE;
2398                 GOTO(error_ucred, rc);
2399         }
2400
2401         LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
2402                 mfd->mfd_mode == MDS_FMODE_TRUNC);
2403         class_handle_unhash(&mfd->mfd_handle);
2404         list_del_init(&mfd->mfd_list);
2405         spin_unlock(&med->med_open_lock);
2406
2407         /* Set EPOCH CLOSE flag if not set by client. */
2408         info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
2409         info->mti_attr.ma_valid = 0;
2410
2411         info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
2412         OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2413         if (info->mti_attr.ma_lmm == NULL)
2414                 GOTO(error_ucred, rc = -ENOMEM);
2415
2416         rc = mdt_mfd_close(info, mfd);
2417
2418         OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2419         mdt_empty_transno(info, rc);
2420 error_ucred:
2421         mdt_exit_ucred(info);
2422 out:
2423         mdt_thread_info_fini(info);
2424         RETURN(rc);
2425 }