Whamcloud - gitweb
681897f70bf6ff82d220ea0b83e05d601d5657a5
[fs/lustre-release.git] / lustre / osd-ldiskfs / osd_handler.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, 2015, 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/osd/osd_handler.c
37  *
38  * Top-level entry points into osd module
39  *
40  * Author: Nikita Danilov <nikita@clusterfs.com>
41  *         Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
42  */
43
44 #define DEBUG_SUBSYSTEM S_OSD
45
46 #include <linux/module.h>
47 #include <linux/user_namespace.h>
48 #ifdef HAVE_UIDGID_HEADER
49 # include <linux/uidgid.h>
50 #endif
51
52 /* LUSTRE_VERSION_CODE */
53 #include <lustre_ver.h>
54 /* prerequisite for linux/xattr.h */
55 #include <linux/types.h>
56 /* prerequisite for linux/xattr.h */
57 #include <linux/fs.h>
58 /* XATTR_{REPLACE,CREATE} */
59 #include <linux/xattr.h>
60
61 #include <ldiskfs/ldiskfs.h>
62 #include <ldiskfs/xattr.h>
63 #include <ldiskfs/ldiskfs_extents.h>
64 #undef ENTRY
65 /*
66  * struct OBD_{ALLOC,FREE}*()
67  * OBD_FAIL_CHECK
68  */
69 #include <obd_support.h>
70 /* struct ptlrpc_thread */
71 #include <lustre_net.h>
72 #include <lustre_fid.h>
73 /* process_config */
74 #include <lustre_param.h>
75
76 #include "osd_internal.h"
77 #include "osd_dynlocks.h"
78
79 /* llo_* api support */
80 #include <md_object.h>
81 #include <lustre_quota.h>
82
83 #include <lustre_linkea.h>
84
85 int ldiskfs_pdo = 1;
86 module_param(ldiskfs_pdo, int, 0644);
87 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
88
89 int ldiskfs_track_declares_assert;
90 module_param(ldiskfs_track_declares_assert, int, 0644);
91 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
92
93 /* Slab to allocate dynlocks */
94 struct kmem_cache *dynlock_cachep;
95
96 /* Slab to allocate osd_it_ea */
97 struct kmem_cache *osd_itea_cachep;
98
99 static struct lu_kmem_descr ldiskfs_caches[] = {
100         {
101                 .ckd_cache = &dynlock_cachep,
102                 .ckd_name  = "dynlock_cache",
103                 .ckd_size  = sizeof(struct dynlock_handle)
104         },
105         {
106                 .ckd_cache = &osd_itea_cachep,
107                 .ckd_name  = "osd_itea_cache",
108                 .ckd_size  = sizeof(struct osd_it_ea)
109         },
110         {
111                 .ckd_cache = NULL
112         }
113 };
114
115 static const char dot[] = ".";
116 static const char dotdot[] = "..";
117 static const char remote_obj_dir[] = "REM_OBJ_DIR";
118
119 static const struct lu_object_operations      osd_lu_obj_ops;
120 static const struct dt_object_operations      osd_obj_ops;
121 static const struct dt_object_operations      osd_obj_ea_ops;
122 static const struct dt_object_operations      osd_obj_otable_it_ops;
123 static const struct dt_index_operations       osd_index_iam_ops;
124 static const struct dt_index_operations       osd_index_ea_ops;
125
126 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
127                           const struct lu_fid *fid);
128 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
129                                                 struct osd_device *osd);
130
131 int osd_trans_declare_op2rb[] = {
132         [OSD_OT_ATTR_SET]       = OSD_OT_ATTR_SET,
133         [OSD_OT_PUNCH]          = OSD_OT_MAX,
134         [OSD_OT_XATTR_SET]      = OSD_OT_XATTR_SET,
135         [OSD_OT_CREATE]         = OSD_OT_DESTROY,
136         [OSD_OT_DESTROY]        = OSD_OT_CREATE,
137         [OSD_OT_REF_ADD]        = OSD_OT_REF_DEL,
138         [OSD_OT_REF_DEL]        = OSD_OT_REF_ADD,
139         [OSD_OT_WRITE]          = OSD_OT_WRITE,
140         [OSD_OT_INSERT]         = OSD_OT_DELETE,
141         [OSD_OT_DELETE]         = OSD_OT_INSERT,
142         [OSD_OT_QUOTA]          = OSD_OT_MAX,
143 };
144
145 static int osd_has_index(const struct osd_object *obj)
146 {
147         return obj->oo_dt.do_index_ops != NULL;
148 }
149
150 static int osd_object_invariant(const struct lu_object *l)
151 {
152         return osd_invariant(osd_obj(l));
153 }
154
155 /*
156  * Concurrency: doesn't matter
157  */
158
159 /*
160  * Concurrency: doesn't matter
161  */
162 static int osd_write_locked(const struct lu_env *env, struct osd_object *o)
163 {
164         struct osd_thread_info *oti = osd_oti_get(env);
165         return oti->oti_w_locks > 0 && o->oo_owner == env;
166 }
167
168 /*
169  * Concurrency: doesn't access mutable data
170  */
171 static int osd_root_get(const struct lu_env *env,
172                         struct dt_device *dev, struct lu_fid *f)
173 {
174         lu_local_obj_fid(f, OSD_FS_ROOT_OID);
175         return 0;
176 }
177
178 /*
179  * the following set of functions are used to maintain per-thread
180  * cache of FID->ino mapping. this mechanism is needed to resolve
181  * FID to inode at dt_insert() which in turn stores ino in the
182  * directory entries to keep ldiskfs compatible with ext[34].
183  * due to locking-originated restrictions we can't lookup ino
184  * using LU cache (deadlock is possible). lookup using OI is quite
185  * expensive. so instead we maintain this cache and methods like
186  * dt_create() fill it. so in the majority of cases dt_insert() is
187  * able to find needed mapping in lockless manner.
188  */
189 static struct osd_idmap_cache *
190 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
191              const struct lu_fid *fid)
192 {
193         struct osd_thread_info  *oti   = osd_oti_get(env);
194         struct osd_idmap_cache  *idc    = oti->oti_ins_cache;
195         int i;
196         for (i = 0; i < oti->oti_ins_cache_used; i++) {
197                 if (!lu_fid_eq(&idc[i].oic_fid, fid))
198                         continue;
199                 if (idc[i].oic_dev != osd)
200                         continue;
201
202                 return idc + i;
203         }
204
205         return NULL;
206 }
207
208 static struct osd_idmap_cache *
209 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
210             const struct lu_fid *fid)
211 {
212         struct osd_thread_info  *oti   = osd_oti_get(env);
213         struct osd_idmap_cache  *idc;
214         int i;
215
216         if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
217                 i = oti->oti_ins_cache_size * 2;
218                 if (i == 0)
219                         i = OSD_INS_CACHE_SIZE;
220                 OBD_ALLOC(idc, sizeof(*idc) * i);
221                 if (idc == NULL)
222                         return ERR_PTR(-ENOMEM);
223                 if (oti->oti_ins_cache != NULL) {
224                         memcpy(idc, oti->oti_ins_cache,
225                                oti->oti_ins_cache_used * sizeof(*idc));
226                         OBD_FREE(oti->oti_ins_cache,
227                                  oti->oti_ins_cache_used * sizeof(*idc));
228                 }
229                 oti->oti_ins_cache = idc;
230                 oti->oti_ins_cache_size = i;
231         }
232
233         idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
234         idc->oic_fid = *fid;
235         idc->oic_dev = osd;
236         idc->oic_lid.oii_ino = 0;
237         idc->oic_lid.oii_gen = 0;
238         idc->oic_remote = 0;
239
240         return idc;
241 }
242
243 /*
244  * lookup mapping for the given fid in the cache, initialize a
245  * new one if not found. the initialization checks whether the
246  * object is local or remote. for local objects, OI is used to
247  * learn ino/generation. the function is used when the caller
248  * has no information about the object, e.g. at dt_insert().
249  */
250 static struct osd_idmap_cache *
251 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
252                      const struct lu_fid *fid)
253 {
254         struct osd_idmap_cache *idc;
255         int rc;
256
257         idc = osd_idc_find(env, osd, fid);
258         LASSERT(!IS_ERR(idc));
259         if (idc != NULL)
260                 return idc;
261
262         /* new mapping is needed */
263         idc = osd_idc_add(env, osd, fid);
264         if (IS_ERR(idc))
265                 return idc;
266
267         /* initialize it */
268         rc = osd_remote_fid(env, osd, fid);
269         if (unlikely(rc < 0))
270                 return ERR_PTR(rc);
271
272         if (rc == 0) {
273                 /* the object is local, lookup in OI */
274                 /* XXX: probably cheaper to lookup in LU first? */
275                 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
276                                    &idc->oic_lid, 0);
277                 if (unlikely(rc < 0)) {
278                         CERROR("can't lookup: rc = %d\n", rc);
279                         return ERR_PTR(rc);
280                 }
281         } else {
282                 /* the object is remote */
283                 idc->oic_remote = 1;
284         }
285
286         return idc;
287 }
288
289 /*
290  * lookup mapping for given FID and fill it from the given object.
291  * the object is lolcal by definition.
292  */
293 static int osd_idc_find_and_init(const struct lu_env *env,
294                                  struct osd_device *osd,
295                                  struct osd_object *obj)
296 {
297         const struct lu_fid     *fid = lu_object_fid(&obj->oo_dt.do_lu);
298         struct osd_idmap_cache  *idc;
299
300         idc = osd_idc_find(env, osd, fid);
301         LASSERT(!IS_ERR(idc));
302         if (idc != NULL) {
303                 if (obj->oo_inode == NULL)
304                         return 0;
305                 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
306                         LASSERT(idc->oic_lid.oii_ino == 0);
307                         idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
308                         idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
309                 }
310                 return 0;
311         }
312
313         /* new mapping is needed */
314         idc = osd_idc_add(env, osd, fid);
315         if (IS_ERR(idc))
316                 return PTR_ERR(idc);
317
318         if (obj->oo_inode != NULL) {
319                 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
320                 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
321         }
322         return 0;
323 }
324
325 /*
326  * OSD object methods.
327  */
328
329 /*
330  * Concurrency: no concurrent access is possible that early in object
331  * life-cycle.
332  */
333 static struct lu_object *osd_object_alloc(const struct lu_env *env,
334                                           const struct lu_object_header *hdr,
335                                           struct lu_device *d)
336 {
337         struct osd_object *mo;
338
339         OBD_ALLOC_PTR(mo);
340         if (mo != NULL) {
341                 struct lu_object *l;
342
343                 l = &mo->oo_dt.do_lu;
344                 dt_object_init(&mo->oo_dt, NULL, d);
345                 mo->oo_dt.do_ops = &osd_obj_ea_ops;
346                 l->lo_ops = &osd_lu_obj_ops;
347                 init_rwsem(&mo->oo_sem);
348                 init_rwsem(&mo->oo_ext_idx_sem);
349                 spin_lock_init(&mo->oo_guard);
350                 return l;
351         } else {
352                 return NULL;
353         }
354 }
355
356 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
357                 struct dentry *dentry, struct lustre_mdt_attrs *lma)
358 {
359         int rc;
360
361         CLASSERT(LMA_OLD_SIZE >= sizeof(*lma));
362         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
363                              info->oti_mdt_attrs_old, LMA_OLD_SIZE);
364         if (rc > 0) {
365                 if ((void *)lma != (void *)info->oti_mdt_attrs_old)
366                         memcpy(lma, info->oti_mdt_attrs_old, sizeof(*lma));
367                 rc = 0;
368                 lustre_lma_swab(lma);
369                 /* Check LMA compatibility */
370                 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
371                         CWARN("%.16s: unsupported incompat LMA feature(s) %#x "
372                               "for fid = "DFID", ino = %lu\n",
373                               LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
374                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
375                               PFID(&lma->lma_self_fid), inode->i_ino);
376                         rc = -EOPNOTSUPP;
377                 }
378         } else if (rc == 0) {
379                 rc = -ENODATA;
380         }
381
382         return rc;
383 }
384
385 /*
386  * retrieve object from backend ext fs.
387  **/
388 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
389                        struct osd_inode_id *id)
390 {
391         struct inode *inode = NULL;
392
393         /* if we look for an inode withing a running
394          * transaction, then we risk to deadlock */
395         /* osd_dirent_check_repair() breaks this */
396         /*LASSERT(current->journal_info == NULL);*/
397
398         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
399         if (IS_ERR(inode)) {
400                 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
401                        id->oii_ino, PTR_ERR(inode));
402         } else if (id->oii_gen != OSD_OII_NOGEN &&
403                    inode->i_generation != id->oii_gen) {
404                 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
405                        "i_generation = %u\n",
406                        id->oii_ino, id->oii_gen, inode->i_generation);
407                 iput(inode);
408                 inode = ERR_PTR(-ESTALE);
409         } else if (inode->i_nlink == 0) {
410                 /* due to parallel readdir and unlink,
411                 * we can have dead inode here. */
412                 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
413                 iput(inode);
414                 inode = ERR_PTR(-ESTALE);
415         } else if (is_bad_inode(inode)) {
416                 CWARN("%.16s: bad inode: ino = %u\n",
417                 LDISKFS_SB(osd_sb(dev))->s_es->s_volume_name, id->oii_ino);
418                 iput(inode);
419                 inode = ERR_PTR(-ENOENT);
420         } else {
421                 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
422                 if (id->oii_gen == OSD_OII_NOGEN)
423                         osd_id_gen(id, inode->i_ino, inode->i_generation);
424
425                 /* Do not update file c/mtime in ldiskfs.
426                  * NB: we don't have any lock to protect this because we don't
427                  * have reference on osd_object now, but contention with
428                  * another lookup + attr_set can't happen in the tiny window
429                  * between if (...) and set S_NOCMTIME. */
430                 if (!(inode->i_flags & S_NOCMTIME))
431                         inode->i_flags |= S_NOCMTIME;
432         }
433         return inode;
434 }
435
436 int osd_ldiskfs_add_entry(struct osd_thread_info *info,
437                           handle_t *handle, struct dentry *child,
438                           struct inode *inode, struct htree_lock *hlock)
439 {
440         int rc, rc2;
441
442         rc = __ldiskfs_add_entry(handle, child, inode, hlock);
443         if (rc == -ENOBUFS || rc == -ENOSPC) {
444                 char fidbuf[FID_LEN + 1];
445                 struct lustre_mdt_attrs lma;
446                 struct lu_fid fid = { };
447                 char *errstr;
448                 struct dentry *p_dentry = child->d_parent;
449
450                 rc2 = osd_get_lma(info, p_dentry->d_inode, p_dentry,
451                                  &lma);
452                 if (rc2 == 0) {
453                         fid = lma.lma_self_fid;
454                         snprintf(fidbuf, sizeof(fidbuf), DFID, PFID(&fid));
455                 } else if (rc2 == -ENODATA) {
456                         if (unlikely(p_dentry->d_inode ==
457                                      inode->i_sb->s_root->d_inode))
458                                 lu_local_obj_fid(&fid, OSD_FS_ROOT_OID);
459                         else if (info->oti_dev && !info->oti_dev->od_is_ost &&
460                                  fid_seq_is_mdt0(fid_seq(&fid)))
461                                 lu_igif_build(&fid, p_dentry->d_inode->i_ino,
462                                               p_dentry->d_inode->i_generation);
463                         snprintf(fidbuf, sizeof(fidbuf), DFID, PFID(&fid));
464                 } else {
465                         snprintf(fidbuf, FID_LEN, "%s", "unknown");
466                 }
467
468                 if (rc == -ENOSPC)
469                         errstr = "has reached";
470                 else
471                         errstr = "is approaching";
472                 CWARN("%.16s: directory (inode: %lu FID: %s) %s maximum entry limit\n",
473                         LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
474                         p_dentry->d_inode->i_ino, fidbuf, errstr);
475                 /* ignore such error now */
476                 if (rc == -ENOBUFS)
477                         rc = 0;
478         }
479         return rc;
480 }
481
482
483 static struct inode *
484 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
485              struct osd_inode_id *id, struct lu_fid *fid)
486 {
487         struct lustre_mdt_attrs *lma   = &info->oti_mdt_attrs;
488         struct inode            *inode;
489         int                      rc;
490
491         inode = osd_iget(info, dev, id);
492         if (IS_ERR(inode))
493                 return inode;
494
495         rc = osd_get_lma(info, inode, &info->oti_obj_dentry, lma);
496         if (rc == 0) {
497                 *fid = lma->lma_self_fid;
498         } else if (rc == -ENODATA) {
499                 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
500                         lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
501                 else
502                         lu_igif_build(fid, inode->i_ino, inode->i_generation);
503         } else {
504                 iput(inode);
505                 inode = ERR_PTR(rc);
506         }
507         return inode;
508 }
509
510 static struct inode *osd_iget_check(struct osd_thread_info *info,
511                                     struct osd_device *dev,
512                                     const struct lu_fid *fid,
513                                     struct osd_inode_id *id,
514                                     bool cached)
515 {
516         struct inode    *inode;
517         int              rc     = 0;
518         ENTRY;
519
520         /* The cached OI mapping is trustable. If we cannot locate the inode
521          * via the cached OI mapping, then return the failure to the caller
522          * directly without further OI checking. */
523
524         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
525         if (IS_ERR(inode)) {
526                 rc = PTR_ERR(inode);
527                 if (cached || (rc != -ENOENT && rc != -ESTALE)) {
528                         CDEBUG(D_INODE, "no inode: ino = %u, rc = %d\n",
529                                id->oii_ino, rc);
530
531                         GOTO(put, rc);
532                 }
533
534                 goto check_oi;
535         }
536
537         if (is_bad_inode(inode)) {
538                 rc = -ENOENT;
539                 if (cached) {
540                         CDEBUG(D_INODE, "bad inode: ino = %u\n", id->oii_ino);
541
542                         GOTO(put, rc);
543                 }
544
545                 goto check_oi;
546         }
547
548         if (id->oii_gen != OSD_OII_NOGEN &&
549             inode->i_generation != id->oii_gen) {
550                 rc = -ESTALE;
551                 if (cached) {
552                         CDEBUG(D_INODE, "unmatched inode: ino = %u, "
553                                "oii_gen = %u, i_generation = %u\n",
554                                id->oii_ino, id->oii_gen, inode->i_generation);
555
556                         GOTO(put, rc);
557                 }
558
559                 goto check_oi;
560         }
561
562         if (inode->i_nlink == 0) {
563                 rc = -ENOENT;
564                 if (cached) {
565                         CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
566
567                         GOTO(put, rc);
568                 }
569
570                 goto check_oi;
571         }
572
573         ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
574
575 check_oi:
576         if (rc != 0) {
577                 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
578
579                 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
580                 /* XXX: There are four possible cases:
581                  *      1. rc = 0.
582                  *         Backup/restore caused the OI invalid.
583                  *      2. rc = 0.
584                  *         Someone unlinked the object but NOT removed
585                  *         the OI mapping, such as mount target device
586                  *         as ldiskfs, and modify something directly.
587                  *      3. rc = -ENOENT.
588                  *         Someone just removed the object between the
589                  *         former oi_lookup and the iget. It is normal.
590                  *      4. Other failure cases.
591                  *
592                  *      Generally, when the device is mounted, it will
593                  *      auto check whether the system is restored from
594                  *      file-level backup or not. We trust such detect
595                  *      to distinguish the 1st case from the 2nd case. */
596                 if (rc == 0) {
597                         if (!IS_ERR(inode) && inode->i_generation != 0 &&
598                             inode->i_generation == id->oii_gen)
599                                 /* "id->oii_gen != OSD_OII_NOGEN" is for
600                                  * "@cached == false" case. */
601                                 rc = -ENOENT;
602                         else
603                                 rc = -EREMCHG;
604                 } else {
605                         /* If the OI mapping was in OI file before the
606                          * osd_iget_check(), but now, it is disappear,
607                          * then it must be removed by race. That is a
608                          * normal race case. */
609                 }
610         } else {
611                 if (id->oii_gen == OSD_OII_NOGEN)
612                         osd_id_gen(id, inode->i_ino, inode->i_generation);
613
614                 /* Do not update file c/mtime in ldiskfs.
615                  * NB: we don't have any lock to protect this because we don't
616                  * have reference on osd_object now, but contention with
617                  * another lookup + attr_set can't happen in the tiny window
618                  * between if (...) and set S_NOCMTIME. */
619                 if (!(inode->i_flags & S_NOCMTIME))
620                         inode->i_flags |= S_NOCMTIME;
621         }
622
623         GOTO(put, rc);
624
625 put:
626         if (rc != 0) {
627                 if (!IS_ERR(inode))
628                         iput(inode);
629
630                 inode = ERR_PTR(rc);
631         }
632
633         return inode;
634 }
635
636 /**
637  * \retval +v: new filter_fid, does not contain self-fid
638  * \retval 0:  filter_fid_old, contains self-fid
639  * \retval -v: other failure cases
640  */
641 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
642                  struct dentry *dentry, struct lu_fid *fid)
643 {
644         struct filter_fid_old   *ff     = &info->oti_ff;
645         struct ost_id           *ostid  = &info->oti_ostid;
646         int                      rc;
647
648         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
649         if (rc == sizeof(*ff)) {
650                 rc = 0;
651                 ostid_set_seq(ostid, le64_to_cpu(ff->ff_seq));
652                 ostid_set_id(ostid, le64_to_cpu(ff->ff_objid));
653                 /* XXX: use 0 as the index for compatibility, the caller will
654                  *      handle index related issues when necessarry. */
655                 ostid_to_fid(fid, ostid, 0);
656         } else if (rc == sizeof(struct filter_fid)) {
657                 rc = 1;
658         } else if (rc >= 0) {
659                 rc = -EINVAL;
660         }
661
662         return rc;
663 }
664
665 static int osd_lma_self_repair(struct osd_thread_info *info,
666                                struct osd_device *osd, struct inode *inode,
667                                const struct lu_fid *fid, __u32 compat)
668 {
669         handle_t *jh;
670         int       rc;
671
672         LASSERT(current->journal_info == NULL);
673
674         jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
675                                   osd_dto_credits_noquota[DTO_XATTR_SET]);
676         if (IS_ERR(jh)) {
677                 rc = PTR_ERR(jh);
678                 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
679                       osd_name(osd), rc);
680                 return rc;
681         }
682
683         rc = osd_ea_fid_set(info, inode, fid, compat, 0);
684         if (rc != 0)
685                 CWARN("%s: cannot self repair the LMA: rc = %d\n",
686                       osd_name(osd), rc);
687         ldiskfs_journal_stop(jh);
688         return rc;
689 }
690
691 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
692 {
693         struct osd_thread_info  *info   = osd_oti_get(env);
694         struct osd_device       *osd    = osd_obj2dev(obj);
695         struct lustre_mdt_attrs *lma    = &info->oti_mdt_attrs;
696         struct inode            *inode  = obj->oo_inode;
697         struct dentry           *dentry = &info->oti_obj_dentry;
698         struct lu_fid           *fid    = NULL;
699         const struct lu_fid     *rfid   = lu_object_fid(&obj->oo_dt.do_lu);
700         int                      rc;
701         ENTRY;
702
703         CLASSERT(LMA_OLD_SIZE >= sizeof(*lma));
704         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
705                              info->oti_mdt_attrs_old, LMA_OLD_SIZE);
706         if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
707                 fid = &lma->lma_self_fid;
708                 rc = osd_get_idif(info, inode, dentry, fid);
709                 if ((rc > 0) || (rc == -ENODATA && osd->od_index_in_idif)) {
710                         /* For the given OST-object, if it has neither LMA nor
711                          * FID in XATTR_NAME_FID, then the given FID (which is
712                          * contained in the @obj, from client RPC for locating
713                          * the OST-object) is trusted. We use it to generate
714                          * the LMA. */
715                         osd_lma_self_repair(info, osd, inode, rfid,
716                                             LMAC_FID_ON_OST);
717                         RETURN(0);
718                 }
719         }
720
721         if (rc < 0)
722                 RETURN(rc);
723
724         if (rc > 0) {
725                 rc = 0;
726                 lustre_lma_swab(lma);
727                 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
728                              CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT))) {
729                         CWARN("%s: unsupported incompat LMA feature(s) %#x for "
730                               "fid = "DFID", ino = %lu\n", osd_name(osd),
731                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
732                               PFID(rfid), inode->i_ino);
733                         rc = -EOPNOTSUPP;
734                 } else {
735                         fid = &lma->lma_self_fid;
736                 }
737         }
738
739         if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
740                 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
741                         struct ost_id   *oi   = &info->oti_ostid;
742                         struct lu_fid   *fid1 = &info->oti_fid3;
743                         __u32            idx  = fid_idif_ost_idx(rfid);
744
745                         /* For old IDIF, the OST index is not part of the IDIF,
746                          * Means that different OSTs may have the same IDIFs.
747                          * Under such case, we need to make some compatible
748                          * check to make sure to trigger OI scrub properly. */
749                         if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
750                                 /* Given @rfid is new, LMA is old. */
751                                 fid_to_ostid(fid, oi);
752                                 ostid_to_fid(fid1, oi, idx);
753                                 if (lu_fid_eq(fid1, rfid)) {
754                                         if (osd->od_index_in_idif)
755                                                 osd_lma_self_repair(info, osd,
756                                                         inode, rfid,
757                                                         LMAC_FID_ON_OST);
758                                         RETURN(0);
759                                 }
760                         }
761                 }
762
763                 rc = -EREMCHG;
764         }
765
766         RETURN(rc);
767 }
768
769 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
770                           const struct lu_fid *fid,
771                           const struct lu_object_conf *conf)
772 {
773         struct osd_thread_info *info;
774         struct lu_device       *ldev   = obj->oo_dt.do_lu.lo_dev;
775         struct osd_device      *dev;
776         struct osd_idmap_cache *oic;
777         struct osd_inode_id    *id;
778         struct inode           *inode;
779         struct osd_scrub       *scrub;
780         struct scrub_file      *sf;
781         int                     result;
782         int                     rc1     = 0;
783         bool                    cached  = true;
784         bool                    remote  = false;
785         ENTRY;
786
787         LINVRNT(osd_invariant(obj));
788         LASSERT(obj->oo_inode == NULL);
789         LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
790
791         dev = osd_dev(ldev);
792         scrub = &dev->od_scrub;
793         sf = &scrub->os_file;
794         info = osd_oti_get(env);
795         LASSERT(info);
796         oic = &info->oti_cache;
797
798         if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
799                 RETURN(-ENOENT);
800
801         /* For the object is created as locking anchor, or for the object to
802          * be created on disk. No need to osd_oi_lookup() at here because FID
803          * shouldn't never be re-used, if it's really a duplicate FID from
804          * unexpected reason, we should be able to detect it later by calling
805          * do_create->osd_oi_insert(). */
806         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
807                 GOTO(out, result = 0);
808
809         /* Search order: 1. per-thread cache. */
810         if (lu_fid_eq(fid, &oic->oic_fid) &&
811             likely(oic->oic_dev == dev)) {
812                 id = &oic->oic_lid;
813                 goto iget;
814         }
815
816         id = &info->oti_id;
817         if (!list_empty(&scrub->os_inconsistent_items)) {
818                 /* Search order: 2. OI scrub pending list. */
819                 result = osd_oii_lookup(dev, fid, id);
820                 if (result == 0)
821                         goto iget;
822         }
823
824         cached = false;
825         /* Search order: 3. OI files. */
826         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
827         if (result == -ENOENT) {
828                 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
829                     fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
830                     !ldiskfs_test_bit(osd_oi_fid2idx(dev,fid),
831                                       sf->sf_oi_bitmap))
832                         GOTO(out, result = 0);
833
834                 goto trigger;
835         }
836
837         if (result != 0)
838                 GOTO(out, result);
839
840 iget:
841         inode = osd_iget_check(info, dev, fid, id, cached);
842         if (IS_ERR(inode)) {
843                 result = PTR_ERR(inode);
844                 if (result == -ENOENT || result == -ESTALE)
845                         GOTO(out, result = -ENOENT);
846
847                 if (result == -EREMCHG) {
848
849 trigger:
850                         /* We still have chance to get the valid inode: for the
851                          * object which is referenced by remote name entry, the
852                          * object on the local MDT will be linked under the dir
853                          * of "/REMOTE_PARENT_DIR" with its FID string as name.
854                          *
855                          * We do not know whether the object for the given FID
856                          * is referenced by some remote name entry or not, and
857                          * especially for DNE II, a multiple-linked object may
858                          * have many name entries reside on many MDTs.
859                          *
860                          * To simplify the operation, OSD will not distinguish
861                          * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
862                          * only happened for the RPC from other MDT during the
863                          * OI scrub, or for the client side RPC with FID only,
864                          * such as FID to path, or from old connected client. */
865                         if (!remote &&
866                             !fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
867                                 rc1 = osd_lookup_in_remote_parent(info, dev,
868                                                                   fid, id);
869                                 if (rc1 == 0) {
870                                         remote = true;
871                                         cached = true;
872                                         goto iget;
873                                 }
874                         }
875
876                         if (thread_is_running(&scrub->os_thread)) {
877                                 if (remote) {
878                                         osd_add_oi_cache(info, dev, id, fid);
879                                         osd_oii_insert(dev, oic, true);
880                                 } else {
881                                         result = -EINPROGRESS;
882                                 }
883                         } else if (!dev->od_noscrub) {
884                                 __u32 flags = SS_CLEAR_DRYRUN |
885                                               SS_CLEAR_FAILOUT;
886
887                                 flags |= (remote ? SS_AUTO_PARTIAL :
888                                                    SS_AUTO_FULL);
889                                 rc1 = osd_scrub_start(dev, flags);
890                                 LCONSOLE_WARN("%.16s: trigger OI scrub by RPC "
891                                               "for the "DFID" with flags 0x%x,"
892                                               " rc = %d\n", osd_name(dev),
893                                               PFID(fid), flags, rc1);
894                                 if (rc1 == 0 || rc1 == -EALREADY) {
895                                         result = -EINPROGRESS;
896                                         if (remote) {
897                                                 osd_add_oi_cache(info, dev, id,
898                                                                  fid);
899                                                 osd_oii_insert(dev, oic, true);
900                                         }
901                                 } else {
902                                         result = -EREMCHG;
903                                 }
904                         } else {
905                                 result = -EREMCHG;
906                         }
907                 }
908
909                 GOTO(out, result);
910         } else if (remote) {
911                 result = 0;
912                 goto trigger;
913         }
914
915         obj->oo_inode = inode;
916         LASSERT(obj->oo_inode->i_sb == osd_sb(dev));
917
918         result = osd_check_lma(env, obj);
919         if (result != 0) {
920                 if (result == -ENODATA) {
921                         if (cached) {
922                                 result = osd_oi_lookup(info, dev, fid, id,
923                                                        OI_CHECK_FLD);
924                                 if (result != 0) {
925                                         /* result == -ENOENT means that the OI
926                                          * mapping has been removed by race,
927                                          * the target inode belongs to other
928                                          * object.
929                                          *
930                                          * Others error also can be returned
931                                          * directly. */
932                                         iput(inode);
933                                         obj->oo_inode = NULL;
934                                         GOTO(out, result);
935                                 } else {
936                                         /* result == 0 means the cached OI
937                                          * mapping is still in the OI file,
938                                          * the target the inode is valid. */
939                                 }
940                         } else {
941                                 /* The current OI mapping is from the OI file,
942                                  * since the inode has been found via
943                                  * osd_iget_check(), no need recheck OI. */
944                         }
945
946                         goto found;
947                 }
948
949                 iput(inode);
950                 obj->oo_inode = NULL;
951                 if (result != -EREMCHG)
952                         GOTO(out, result);
953
954                 if (cached) {
955                         result = osd_oi_lookup(info, dev, fid, id,
956                                                OI_CHECK_FLD);
957                         /* result == -ENOENT means the cached OI mapping
958                          * has been removed from the OI file by race,
959                          * above target inode belongs to other object.
960                          *
961                          * Others error also can be returned directly. */
962                         if (result != 0)
963                                 GOTO(out, result);
964
965                         /* result == 0, goto trigger */
966                 } else {
967                         /* The current OI mapping is from the OI file,
968                          * since the inode has been found via
969                          * osd_iget_check(), no need recheck OI. */
970                 }
971
972                 goto trigger;
973         }
974
975 found:
976         obj->oo_compat_dot_created = 1;
977         obj->oo_compat_dotdot_created = 1;
978
979         if (!S_ISDIR(inode->i_mode) || !ldiskfs_pdo) /* done */
980                 GOTO(out, result = 0);
981
982         LASSERT(obj->oo_hl_head == NULL);
983         obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
984         if (obj->oo_hl_head == NULL) {
985                 obj->oo_inode = NULL;
986                 iput(inode);
987                 GOTO(out, result = -ENOMEM);
988         }
989         GOTO(out, result = 0);
990
991 out:
992         if (result != 0 && cached)
993                 fid_zero(&oic->oic_fid);
994
995         LINVRNT(osd_invariant(obj));
996         return result;
997 }
998
999 /*
1000  * Concurrency: shouldn't matter.
1001  */
1002 static void osd_object_init0(struct osd_object *obj)
1003 {
1004         LASSERT(obj->oo_inode != NULL);
1005         obj->oo_dt.do_body_ops = &osd_body_ops;
1006         obj->oo_dt.do_lu.lo_header->loh_attr |=
1007                 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1008 }
1009
1010 /*
1011  * Concurrency: no concurrent access is possible that early in object
1012  * life-cycle.
1013  */
1014 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1015                            const struct lu_object_conf *conf)
1016 {
1017         struct osd_object *obj = osd_obj(l);
1018         int result;
1019
1020         LINVRNT(osd_invariant(obj));
1021
1022         if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1023                 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1024                 l->lo_header->loh_attr |= LOHA_EXISTS;
1025                 return 0;
1026         }
1027
1028         result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1029         obj->oo_dt.do_body_ops = &osd_body_ops_new;
1030         if (result == 0 && obj->oo_inode != NULL) {
1031                 struct osd_thread_info *oti = osd_oti_get(env);
1032                 struct lustre_mdt_attrs *lma = &oti->oti_mdt_attrs;
1033
1034                 osd_object_init0(obj);
1035                 result = osd_get_lma(oti, obj->oo_inode,
1036                                      &oti->oti_obj_dentry, lma);
1037                 if (result == 0) {
1038                         /* Convert LMAI flags to lustre LMA flags
1039                          * and cache it to oo_lma_flags */
1040                         obj->oo_lma_flags =
1041                                 lma_to_lustre_flags(lma->lma_incompat);
1042                 } else if (result == -ENODATA) {
1043                         result = 0;
1044                 }
1045         }
1046
1047         LINVRNT(osd_invariant(obj));
1048         return result;
1049 }
1050
1051 /*
1052  * Concurrency: no concurrent access is possible that late in object
1053  * life-cycle.
1054  */
1055 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1056 {
1057         struct osd_object *obj = osd_obj(l);
1058
1059         LINVRNT(osd_invariant(obj));
1060
1061         dt_object_fini(&obj->oo_dt);
1062         if (obj->oo_hl_head != NULL)
1063                 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1064         OBD_FREE_PTR(obj);
1065 }
1066
1067 /*
1068  * Concurrency: no concurrent access is possible that late in object
1069  * life-cycle.
1070  */
1071 static void osd_index_fini(struct osd_object *o)
1072 {
1073         struct iam_container *bag;
1074
1075         if (o->oo_dir != NULL) {
1076                 bag = &o->oo_dir->od_container;
1077                 if (o->oo_inode != NULL) {
1078                         if (bag->ic_object == o->oo_inode)
1079                                 iam_container_fini(bag);
1080                 }
1081                 OBD_FREE_PTR(o->oo_dir);
1082                 o->oo_dir = NULL;
1083         }
1084 }
1085
1086 /*
1087  * Concurrency: no concurrent access is possible that late in object
1088  * life-cycle (for all existing callers, that is. New callers have to provide
1089  * their own locking.)
1090  */
1091 static int osd_inode_unlinked(const struct inode *inode)
1092 {
1093         return inode->i_nlink == 0;
1094 }
1095
1096 enum {
1097         OSD_TXN_OI_DELETE_CREDITS    = 20,
1098         OSD_TXN_INODE_DELETE_CREDITS = 20
1099 };
1100
1101 /*
1102  * Journal
1103  */
1104
1105 #if OSD_THANDLE_STATS
1106 /**
1107  * Set time when the handle is allocated
1108  */
1109 static void osd_th_alloced(struct osd_thandle *oth)
1110 {
1111         oth->oth_alloced = cfs_time_current();
1112 }
1113
1114 /**
1115  * Set time when the handle started
1116  */
1117 static void osd_th_started(struct osd_thandle *oth)
1118 {
1119         oth->oth_started = cfs_time_current();
1120 }
1121
1122 /**
1123  * Helper function to convert time interval to microseconds packed in
1124  * long int.
1125  */
1126 static long interval_to_usec(cfs_time_t start, cfs_time_t end)
1127 {
1128         struct timeval val;
1129
1130         cfs_duration_usec(cfs_time_sub(end, start), &val);
1131         return val.tv_sec * 1000000 + val.tv_usec;
1132 }
1133
1134 /**
1135  * Check whether the we deal with this handle for too long.
1136  */
1137 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1138                                 cfs_time_t alloced, cfs_time_t started,
1139                                 cfs_time_t closed)
1140 {
1141         cfs_time_t now = cfs_time_current();
1142
1143         LASSERT(dev != NULL);
1144
1145         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1146                             interval_to_usec(alloced, started));
1147         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1148                             interval_to_usec(started, closed));
1149         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1150                             interval_to_usec(closed, now));
1151
1152         if (cfs_time_before(cfs_time_add(alloced, cfs_time_seconds(30)), now)) {
1153                 CWARN("transaction handle %p was open for too long: "
1154                       "now "CFS_TIME_T" ,"
1155                       "alloced "CFS_TIME_T" ,"
1156                       "started "CFS_TIME_T" ,"
1157                       "closed "CFS_TIME_T"\n",
1158                       oth, now, alloced, started, closed);
1159                 libcfs_debug_dumpstack(NULL);
1160         }
1161 }
1162
1163 #define OSD_CHECK_SLOW_TH(oth, dev, expr)                               \
1164 {                                                                       \
1165         cfs_time_t __closed = cfs_time_current();                       \
1166         cfs_time_t __alloced = oth->oth_alloced;                        \
1167         cfs_time_t __started = oth->oth_started;                        \
1168                                                                         \
1169         expr;                                                           \
1170         __osd_th_check_slow(oth, dev, __alloced, __started, __closed);  \
1171 }
1172
1173 #else /* OSD_THANDLE_STATS */
1174
1175 #define osd_th_alloced(h)                  do {} while(0)
1176 #define osd_th_started(h)                  do {} while(0)
1177 #define OSD_CHECK_SLOW_TH(oth, dev, expr)  expr
1178
1179 #endif /* OSD_THANDLE_STATS */
1180
1181 /*
1182  * Concurrency: doesn't access mutable data.
1183  */
1184 static int osd_param_is_not_sane(const struct osd_device *dev,
1185                                  const struct thandle *th)
1186 {
1187         struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1188
1189         return oh->ot_credits > osd_transaction_size(dev);
1190 }
1191
1192 /*
1193  * Concurrency: shouldn't matter.
1194  */
1195 static void osd_trans_commit_cb(struct super_block *sb,
1196                                 struct ldiskfs_journal_cb_entry *jcb, int error)
1197 {
1198         struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1199         struct thandle     *th  = &oh->ot_super;
1200         struct lu_device   *lud = &th->th_dev->dd_lu_dev;
1201         struct dt_txn_commit_cb *dcb, *tmp;
1202
1203         LASSERT(oh->ot_handle == NULL);
1204
1205         if (error)
1206                 CERROR("transaction @0x%p commit error: %d\n", th, error);
1207
1208         dt_txn_hook_commit(th);
1209
1210         /* call per-transaction callbacks if any */
1211         list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1212                                  dcb_linkage) {
1213                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1214                          "commit callback entry: magic=%x name='%s'\n",
1215                          dcb->dcb_magic, dcb->dcb_name);
1216                 list_del_init(&dcb->dcb_linkage);
1217                 dcb->dcb_func(NULL, th, dcb, error);
1218         }
1219
1220         lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1221         lu_device_put(lud);
1222         th->th_dev = NULL;
1223
1224         lu_context_exit(&th->th_ctx);
1225         lu_context_fini(&th->th_ctx);
1226         OBD_FREE_PTR(oh);
1227 }
1228
1229 #ifndef HAVE_SB_START_WRITE
1230 # define sb_start_write(sb) do {} while (0)
1231 # define sb_end_write(sb) do {} while (0)
1232 #endif
1233
1234 static struct thandle *osd_trans_create(const struct lu_env *env,
1235                                         struct dt_device *d)
1236 {
1237         struct osd_thread_info  *oti = osd_oti_get(env);
1238         struct osd_iobuf        *iobuf = &oti->oti_iobuf;
1239         struct osd_thandle      *oh;
1240         struct thandle          *th;
1241         ENTRY;
1242
1243         /* on pending IO in this thread should left from prev. request */
1244         LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1245
1246         sb_start_write(osd_sb(osd_dt_dev(d)));
1247
1248         OBD_ALLOC_GFP(oh, sizeof *oh, GFP_NOFS);
1249         if (oh != NULL) {
1250                 oh->ot_quota_trans = &oti->oti_quota_trans;
1251                 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1252                 th = &oh->ot_super;
1253                 th->th_dev = d;
1254                 th->th_result = 0;
1255                 th->th_tags = LCT_TX_HANDLE;
1256                 oh->ot_credits = 0;
1257                 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1258                 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1259                 osd_th_alloced(oh);
1260
1261                 memset(oti->oti_declare_ops, 0,
1262                        sizeof(oti->oti_declare_ops));
1263                 memset(oti->oti_declare_ops_cred, 0,
1264                        sizeof(oti->oti_declare_ops_cred));
1265                 memset(oti->oti_declare_ops_used, 0,
1266                        sizeof(oti->oti_declare_ops_used));
1267         } else {
1268                 sb_end_write(osd_sb(osd_dt_dev(d)));
1269                 th = ERR_PTR(-ENOMEM);
1270         }
1271         RETURN(th);
1272 }
1273
1274 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1275 {
1276         struct osd_thread_info  *oti = osd_oti_get(env);
1277         struct osd_thandle      *oh;
1278
1279         oh = container_of0(th, struct osd_thandle, ot_super);
1280         LASSERT(oh != NULL);
1281
1282         CWARN("  create: %u/%u/%u, destroy: %u/%u/%u\n",
1283               oti->oti_declare_ops[OSD_OT_CREATE],
1284               oti->oti_declare_ops_cred[OSD_OT_CREATE],
1285               oti->oti_declare_ops_used[OSD_OT_CREATE],
1286               oti->oti_declare_ops[OSD_OT_DESTROY],
1287               oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1288               oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1289         CWARN("  attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1290               oti->oti_declare_ops[OSD_OT_ATTR_SET],
1291               oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1292               oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1293               oti->oti_declare_ops[OSD_OT_XATTR_SET],
1294               oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1295               oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1296         CWARN("  write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1297               oti->oti_declare_ops[OSD_OT_WRITE],
1298               oti->oti_declare_ops_cred[OSD_OT_WRITE],
1299               oti->oti_declare_ops_used[OSD_OT_WRITE],
1300               oti->oti_declare_ops[OSD_OT_PUNCH],
1301               oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1302               oti->oti_declare_ops_used[OSD_OT_PUNCH],
1303               oti->oti_declare_ops[OSD_OT_QUOTA],
1304               oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1305               oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1306         CWARN("  insert: %u/%u/%u, delete: %u/%u/%u\n",
1307               oti->oti_declare_ops[OSD_OT_INSERT],
1308               oti->oti_declare_ops_cred[OSD_OT_INSERT],
1309               oti->oti_declare_ops_used[OSD_OT_INSERT],
1310               oti->oti_declare_ops[OSD_OT_DELETE],
1311               oti->oti_declare_ops_cred[OSD_OT_DELETE],
1312               oti->oti_declare_ops_used[OSD_OT_DELETE]);
1313         CWARN("  ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1314               oti->oti_declare_ops[OSD_OT_REF_ADD],
1315               oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1316               oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1317               oti->oti_declare_ops[OSD_OT_REF_DEL],
1318               oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1319               oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1320 }
1321
1322 /*
1323  * Concurrency: shouldn't matter.
1324  */
1325 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1326                            struct thandle *th)
1327 {
1328         struct osd_thread_info *oti = osd_oti_get(env);
1329         struct osd_device  *dev = osd_dt_dev(d);
1330         handle_t           *jh;
1331         struct osd_thandle *oh;
1332         int rc;
1333
1334         ENTRY;
1335
1336         LASSERT(current->journal_info == NULL);
1337
1338         oh = container_of0(th, struct osd_thandle, ot_super);
1339         LASSERT(oh != NULL);
1340         LASSERT(oh->ot_handle == NULL);
1341
1342         rc = dt_txn_hook_start(env, d, th);
1343         if (rc != 0)
1344                 GOTO(out, rc);
1345
1346         if (unlikely(osd_param_is_not_sane(dev, th))) {
1347                 static unsigned long last_printed;
1348                 static int last_credits;
1349
1350                 /* don't make noise on a tiny testing systems
1351                  * actual credits misuse will be caught anyway */
1352                 if (last_credits != oh->ot_credits &&
1353                     time_after(jiffies, last_printed +
1354                                msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1355                     osd_transaction_size(dev) > 512) {
1356                         osd_trans_dump_creds(env, th);
1357                         libcfs_debug_dumpstack(NULL);
1358                         last_credits = oh->ot_credits;
1359                         last_printed = jiffies;
1360                 }
1361                 /* XXX Limit the credits to 'max_transaction_buffers', and
1362                  *     let the underlying filesystem to catch the error if
1363                  *     we really need so many credits.
1364                  *
1365                  *     This should be removed when we can calculate the
1366                  *     credits precisely. */
1367                 oh->ot_credits = osd_transaction_size(dev);
1368         }
1369
1370         /*
1371          * XXX temporary stuff. Some abstraction layer should
1372          * be used.
1373          */
1374         jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1375         osd_th_started(oh);
1376         if (!IS_ERR(jh)) {
1377                 oh->ot_handle = jh;
1378                 LASSERT(oti->oti_txns == 0);
1379                 lu_context_init(&th->th_ctx, th->th_tags);
1380                 lu_context_enter(&th->th_ctx);
1381
1382                 lu_device_get(&d->dd_lu_dev);
1383                 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1384                               "osd-tx", th);
1385                 oti->oti_txns++;
1386                 rc = 0;
1387         } else {
1388                 rc = PTR_ERR(jh);
1389         }
1390 out:
1391         RETURN(rc);
1392 }
1393
1394 static int osd_seq_exists(const struct lu_env *env,
1395                           struct osd_device *osd, u64 seq)
1396 {
1397         struct lu_seq_range     *range = &osd_oti_get(env)->oti_seq_range;
1398         struct seq_server_site  *ss = osd_seq_site(osd);
1399         int                     rc;
1400         ENTRY;
1401
1402         LASSERT(ss != NULL);
1403         LASSERT(ss->ss_server_fld != NULL);
1404
1405         rc = osd_fld_lookup(env, osd, seq, range);
1406         if (rc != 0) {
1407                 if (rc != -ENOENT)
1408                         CERROR("%s: can't lookup FLD sequence "LPX64
1409                                ": rc = %d\n", osd_name(osd), seq, rc);
1410                 RETURN(0);
1411         }
1412
1413         RETURN(ss->ss_node_id == range->lsr_index);
1414 }
1415
1416 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1417 {
1418         struct dt_txn_commit_cb *dcb;
1419         struct dt_txn_commit_cb *tmp;
1420
1421         /* call per-transaction stop callbacks if any */
1422         list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1423                                  dcb_linkage) {
1424                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1425                          "commit callback entry: magic=%x name='%s'\n",
1426                          dcb->dcb_magic, dcb->dcb_name);
1427                 list_del_init(&dcb->dcb_linkage);
1428                 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1429         }
1430 }
1431
1432 /*
1433  * Concurrency: shouldn't matter.
1434  */
1435 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1436                           struct thandle *th)
1437 {
1438         int                     rc = 0, remove_agents = 0;
1439         struct osd_thandle     *oh;
1440         struct osd_thread_info *oti = osd_oti_get(env);
1441         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1442         struct osd_device      *osd = osd_dt_dev(th->th_dev);
1443         struct qsd_instance    *qsd = osd->od_quota_slave;
1444         struct lquota_trans    *qtrans;
1445         ENTRY;
1446
1447         oh = container_of0(th, struct osd_thandle, ot_super);
1448
1449         /* reset OI cache for safety */
1450         oti->oti_ins_cache_used = 0;
1451
1452         remove_agents = oh->ot_remove_agents;
1453
1454         qtrans = oh->ot_quota_trans;
1455         oh->ot_quota_trans = NULL;
1456
1457         if (oh->ot_handle != NULL) {
1458                 handle_t *hdl = oh->ot_handle;
1459
1460                 /*
1461                  * add commit callback
1462                  * notice we don't do this in osd_trans_start()
1463                  * as underlying transaction can change during truncate
1464                  */
1465                 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1466                                          &oh->ot_jcb);
1467
1468                 LASSERT(oti->oti_txns == 1);
1469                 oti->oti_txns--;
1470
1471                 rc = dt_txn_hook_stop(env, th);
1472                 if (rc != 0)
1473                         CERROR("%s: failed in transaction hook: rc = %d\n",
1474                                osd_name(osd), rc);
1475
1476                 osd_trans_stop_cb(oh, rc);
1477                 /* hook functions might modify th_sync */
1478                 hdl->h_sync = th->th_sync;
1479
1480                 oh->ot_handle = NULL;
1481                 OSD_CHECK_SLOW_TH(oh, osd, rc = ldiskfs_journal_stop(hdl));
1482                 if (rc != 0)
1483                         CERROR("%s: failed to stop transaction: rc = %d\n",
1484                                osd_name(osd), rc);
1485         } else {
1486                 osd_trans_stop_cb(oh, th->th_result);
1487                 OBD_FREE_PTR(oh);
1488         }
1489
1490         /* inform the quota slave device that the transaction is stopping */
1491         qsd_op_end(env, qsd, qtrans);
1492
1493         /* as we want IO to journal and data IO be concurrent, we don't block
1494          * awaiting data IO completion in osd_do_bio(), instead we wait here
1495          * once transaction is submitted to the journal. all reqular requests
1496          * don't do direct IO (except read/write), thus this wait_event becomes
1497          * no-op for them.
1498          *
1499          * IMPORTANT: we have to wait till any IO submited by the thread is
1500          * completed otherwise iobuf may be corrupted by different request
1501          */
1502         wait_event(iobuf->dr_wait,
1503                        atomic_read(&iobuf->dr_numreqs) == 0);
1504         osd_fini_iobuf(osd, iobuf);
1505         if (!rc)
1506                 rc = iobuf->dr_error;
1507
1508         if (unlikely(remove_agents != 0))
1509                 osd_process_scheduled_agent_removals(env, osd);
1510
1511         sb_end_write(osd_sb(osd));
1512
1513         RETURN(rc);
1514 }
1515
1516 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
1517 {
1518         struct osd_thandle *oh = container_of0(th, struct osd_thandle,
1519                                                ot_super);
1520
1521         LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
1522         LASSERT(&dcb->dcb_func != NULL);
1523         if (dcb->dcb_flags & DCB_TRANS_STOP)
1524                 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
1525         else
1526                 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
1527
1528         return 0;
1529 }
1530
1531 /*
1532  * Called just before object is freed. Releases all resources except for
1533  * object itself (that is released by osd_object_free()).
1534  *
1535  * Concurrency: no concurrent access is possible that late in object
1536  * life-cycle.
1537  */
1538 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
1539 {
1540         struct osd_object *obj   = osd_obj(l);
1541         struct inode      *inode = obj->oo_inode;
1542
1543         LINVRNT(osd_invariant(obj));
1544
1545         /*
1546          * If object is unlinked remove fid->ino mapping from object index.
1547          */
1548
1549         osd_index_fini(obj);
1550         if (inode != NULL) {
1551                 struct qsd_instance     *qsd = osd_obj2dev(obj)->od_quota_slave;
1552                 qid_t                    uid = i_uid_read(inode);
1553                 qid_t                    gid = i_gid_read(inode);
1554
1555                 iput(inode);
1556                 obj->oo_inode = NULL;
1557
1558                 if (qsd != NULL) {
1559                         struct osd_thread_info  *info = osd_oti_get(env);
1560                         struct lquota_id_info   *qi = &info->oti_qi;
1561
1562                         /* Release granted quota to master if necessary */
1563                         qi->lqi_id.qid_uid = uid;
1564                         qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
1565
1566                         qi->lqi_id.qid_uid = gid;
1567                         qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
1568                 }
1569         }
1570 }
1571
1572 /*
1573  * Concurrency: ->loo_object_release() is called under site spin-lock.
1574  */
1575 static void osd_object_release(const struct lu_env *env,
1576                                struct lu_object *l)
1577 {
1578         struct osd_object *o = osd_obj(l);
1579         /* nobody should be releasing a non-destroyed object with nlink=0
1580          * the API allows this, but ldiskfs doesn't like and then report
1581          * this inode as deleted */
1582         if (unlikely(!o->oo_destroyed && o->oo_inode && o->oo_inode->i_nlink == 0))
1583                 LBUG();
1584 }
1585
1586 /*
1587  * Concurrency: shouldn't matter.
1588  */
1589 static int osd_object_print(const struct lu_env *env, void *cookie,
1590                             lu_printer_t p, const struct lu_object *l)
1591 {
1592         struct osd_object *o = osd_obj(l);
1593         struct iam_descr  *d;
1594
1595         if (o->oo_dir != NULL)
1596                 d = o->oo_dir->od_container.ic_descr;
1597         else
1598                 d = NULL;
1599         return (*p)(env, cookie,
1600                     LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
1601                     o, o->oo_inode,
1602                     o->oo_inode ? o->oo_inode->i_ino : 0UL,
1603                     o->oo_inode ? o->oo_inode->i_generation : 0,
1604                     d ? d->id_ops->id_name : "plain");
1605 }
1606
1607 /*
1608  * Concurrency: shouldn't matter.
1609  */
1610 int osd_statfs(const struct lu_env *env, struct dt_device *d,
1611                struct obd_statfs *sfs)
1612 {
1613         struct osd_device       *osd = osd_dt_dev(d);
1614         struct super_block      *sb = osd_sb(osd);
1615         struct kstatfs          *ksfs;
1616         __u64                    reserved;
1617         int                      result = 0;
1618
1619         if (unlikely(osd->od_mnt == NULL))
1620                 return -EINPROGRESS;
1621
1622         /* osd_lproc.c call this without env, allocate ksfs for that case */
1623         if (unlikely(env == NULL)) {
1624                 OBD_ALLOC_PTR(ksfs);
1625                 if (ksfs == NULL)
1626                         return -ENOMEM;
1627         } else {
1628                 ksfs = &osd_oti_get(env)->oti_ksfs;
1629         }
1630
1631         result = sb->s_op->statfs(sb->s_root, ksfs);
1632         if (result)
1633                 goto out;
1634
1635         statfs_pack(sfs, ksfs);
1636         if (unlikely(sb->s_flags & MS_RDONLY))
1637                 sfs->os_state = OS_STATE_READONLY;
1638         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1639                                          LDISKFS_FEATURE_INCOMPAT_EXTENTS))
1640                 sfs->os_maxbytes = sb->s_maxbytes;
1641         else
1642                 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
1643
1644         /*
1645          * Reserve some space so to avoid fragmenting the filesystem too much.
1646          * Fragmentation not only impacts performance, but can also increase
1647          * metadata overhead significantly, causing grant calculation to be
1648          * wrong.
1649          *
1650          * Reserve 0.78% of total space, at least 8MB for small filesystems.
1651          */
1652         CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
1653         reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
1654         if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
1655                 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
1656
1657         sfs->os_blocks -= reserved;
1658         sfs->os_bfree  -= min(reserved, sfs->os_bfree);
1659         sfs->os_bavail -= min(reserved, sfs->os_bavail);
1660
1661 out:
1662         if (unlikely(env == NULL))
1663                 OBD_FREE_PTR(ksfs);
1664         return result;
1665 }
1666
1667 /**
1668  * Estimate space needed for file creations. We assume the largest filename
1669  * which is 2^64 - 1, hence a filename of 20 chars.
1670  * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
1671  */
1672 #ifdef __LDISKFS_DIR_REC_LEN
1673 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
1674 #else
1675 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
1676 #endif
1677
1678 /*
1679  * Concurrency: doesn't access mutable data.
1680  */
1681 static void osd_conf_get(const struct lu_env *env,
1682                          const struct dt_device *dev,
1683                          struct dt_device_param *param)
1684 {
1685         struct super_block *sb = osd_sb(osd_dt_dev(dev));
1686         int                ea_overhead;
1687
1688         /*
1689          * XXX should be taken from not-yet-existing fs abstraction layer.
1690          */
1691         param->ddp_max_name_len = LDISKFS_NAME_LEN;
1692         param->ddp_max_nlink    = LDISKFS_LINK_MAX;
1693         param->ddp_symlink_max  = sb->s_blocksize;
1694         param->ddp_mount_type     = LDD_MT_LDISKFS;
1695         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EXTENTS))
1696                 param->ddp_maxbytes = sb->s_maxbytes;
1697         else
1698                 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
1699         /* inode are statically allocated, so per-inode space consumption
1700          * is the space consumed by the directory entry */
1701         param->ddp_inodespace     = PER_OBJ_USAGE;
1702         /* EXT_INIT_MAX_LEN is the theoretical maximum extent size  (32k blocks
1703          * = 128MB) which is unlikely to be hit in real life. Report a smaller
1704          * maximum length to not under count the actual number of extents
1705          * needed for writing a file. */
1706         param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 2;
1707         /* worst-case extent insertion metadata overhead */
1708         param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
1709         param->ddp_mntopts      = 0;
1710         if (test_opt(sb, XATTR_USER))
1711                 param->ddp_mntopts |= MNTOPT_USERXATTR;
1712         if (test_opt(sb, POSIX_ACL))
1713                 param->ddp_mntopts |= MNTOPT_ACL;
1714
1715         /* LOD might calculate the max stripe count based on max_ea_size,
1716          * so we need take account in the overhead as well,
1717          * xattr_header + magic + xattr_entry_head */
1718         ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
1719                       LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
1720
1721 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
1722         if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EA_INODE))
1723                 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
1724                                                                 ea_overhead;
1725         else
1726 #endif
1727                 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
1728 }
1729
1730 /*
1731  * Concurrency: shouldn't matter.
1732  */
1733 static int osd_sync(const struct lu_env *env, struct dt_device *d)
1734 {
1735         int rc;
1736
1737         CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_LDISKFS_NAME);
1738
1739         rc = ldiskfs_force_commit(osd_sb(osd_dt_dev(d)));
1740
1741         CDEBUG(D_CACHE, "synced OSD %s: rc = %d\n",
1742                LUSTRE_OSD_LDISKFS_NAME, rc);
1743
1744         return rc;
1745 }
1746
1747 /**
1748  * Start commit for OSD device.
1749  *
1750  * An implementation of dt_commit_async method for OSD device.
1751  * Asychronously starts underlayng fs sync and thereby a transaction
1752  * commit.
1753  *
1754  * \param env environment
1755  * \param d dt device
1756  *
1757  * \see dt_device_operations
1758  */
1759 static int osd_commit_async(const struct lu_env *env,
1760                             struct dt_device *d)
1761 {
1762         struct super_block *s = osd_sb(osd_dt_dev(d));
1763         ENTRY;
1764
1765         CDEBUG(D_HA, "async commit OSD %s\n", LUSTRE_OSD_LDISKFS_NAME);
1766         RETURN(s->s_op->sync_fs(s, 0));
1767 }
1768
1769 /*
1770  * Concurrency: shouldn't matter.
1771  */
1772
1773 static int osd_ro(const struct lu_env *env, struct dt_device *d)
1774 {
1775         struct super_block *sb = osd_sb(osd_dt_dev(d));
1776         struct block_device *dev = sb->s_bdev;
1777 #ifdef HAVE_DEV_SET_RDONLY
1778         struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
1779         int rc = 0;
1780 #else
1781         int rc = -EOPNOTSUPP;
1782 #endif
1783         ENTRY;
1784
1785 #ifdef HAVE_DEV_SET_RDONLY
1786         CERROR("*** setting %s read-only ***\n", osd_dt_dev(d)->od_svname);
1787
1788         if (jdev && (jdev != dev)) {
1789                 CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
1790                        (long)jdev);
1791                 dev_set_rdonly(jdev);
1792         }
1793         CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
1794         dev_set_rdonly(dev);
1795 #else
1796         CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
1797                osd_dt_dev(d)->od_svname, (long)dev, rc);
1798 #endif
1799         RETURN(rc);
1800 }
1801
1802 /**
1803  * Note: we do not count into QUOTA here.
1804  * If we mount with --data_journal we may need more.
1805  */
1806 const int osd_dto_credits_noquota[DTO_NR] = {
1807         /**
1808          * Insert.
1809          * INDEX_EXTRA_TRANS_BLOCKS(8) +
1810          * SINGLEDATA_TRANS_BLOCKS(8)
1811          * XXX Note: maybe iam need more, since iam have more level than
1812          *           EXT3 htree.
1813          */
1814         [DTO_INDEX_INSERT]  = 16,
1815         /**
1816          * Delete
1817          * just modify a single entry, probably merge few within a block
1818          */
1819         [DTO_INDEX_DELETE]  = 1,
1820         /**
1821          * Used for OI scrub
1822          */
1823         [DTO_INDEX_UPDATE]  = 16,
1824         /**
1825          * 4(inode, inode bits, groups, GDT)
1826          *   notice: OI updates are counted separately with DTO_INDEX_INSERT
1827          */
1828         [DTO_OBJECT_CREATE] = 4,
1829         /**
1830          * 4(inode, inode bits, groups, GDT)
1831          *   notice: OI updates are counted separately with DTO_INDEX_DELETE
1832          */
1833         [DTO_OBJECT_DELETE] = 4,
1834         /**
1835          * Attr set credits (inode)
1836          */
1837         [DTO_ATTR_SET_BASE] = 1,
1838         /**
1839          * Xattr set. The same as xattr of EXT3.
1840          * DATA_TRANS_BLOCKS(14)
1841          * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
1842          * are also counted in. Do not know why?
1843          */
1844         [DTO_XATTR_SET]     = 14,
1845         /**
1846          * credits for inode change during write.
1847          */
1848         [DTO_WRITE_BASE]    = 3,
1849         /**
1850          * credits for single block write.
1851          */
1852         [DTO_WRITE_BLOCK]   = 14,
1853         /**
1854          * Attr set credits for chown.
1855          * This is extra credits for setattr, and it is null without quota
1856          */
1857         [DTO_ATTR_SET_CHOWN] = 0
1858 };
1859
1860 static const struct dt_device_operations osd_dt_ops = {
1861         .dt_root_get       = osd_root_get,
1862         .dt_statfs         = osd_statfs,
1863         .dt_trans_create   = osd_trans_create,
1864         .dt_trans_start    = osd_trans_start,
1865         .dt_trans_stop     = osd_trans_stop,
1866         .dt_trans_cb_add   = osd_trans_cb_add,
1867         .dt_conf_get       = osd_conf_get,
1868         .dt_sync           = osd_sync,
1869         .dt_ro             = osd_ro,
1870         .dt_commit_async   = osd_commit_async,
1871 };
1872
1873 static void osd_object_read_lock(const struct lu_env *env,
1874                                  struct dt_object *dt, unsigned role)
1875 {
1876         struct osd_object *obj = osd_dt_obj(dt);
1877         struct osd_thread_info *oti = osd_oti_get(env);
1878
1879         LINVRNT(osd_invariant(obj));
1880
1881         LASSERT(obj->oo_owner != env);
1882         down_read_nested(&obj->oo_sem, role);
1883
1884         LASSERT(obj->oo_owner == NULL);
1885         oti->oti_r_locks++;
1886 }
1887
1888 static void osd_object_write_lock(const struct lu_env *env,
1889                                   struct dt_object *dt, unsigned role)
1890 {
1891         struct osd_object *obj = osd_dt_obj(dt);
1892         struct osd_thread_info *oti = osd_oti_get(env);
1893
1894         LINVRNT(osd_invariant(obj));
1895
1896         LASSERT(obj->oo_owner != env);
1897         down_write_nested(&obj->oo_sem, role);
1898
1899         LASSERT(obj->oo_owner == NULL);
1900         obj->oo_owner = env;
1901         oti->oti_w_locks++;
1902 }
1903
1904 static void osd_object_read_unlock(const struct lu_env *env,
1905                                    struct dt_object *dt)
1906 {
1907         struct osd_object *obj = osd_dt_obj(dt);
1908         struct osd_thread_info *oti = osd_oti_get(env);
1909
1910         LINVRNT(osd_invariant(obj));
1911
1912         LASSERT(oti->oti_r_locks > 0);
1913         oti->oti_r_locks--;
1914         up_read(&obj->oo_sem);
1915 }
1916
1917 static void osd_object_write_unlock(const struct lu_env *env,
1918                                     struct dt_object *dt)
1919 {
1920         struct osd_object *obj = osd_dt_obj(dt);
1921         struct osd_thread_info *oti = osd_oti_get(env);
1922
1923         LINVRNT(osd_invariant(obj));
1924
1925         LASSERT(obj->oo_owner == env);
1926         LASSERT(oti->oti_w_locks > 0);
1927         oti->oti_w_locks--;
1928         obj->oo_owner = NULL;
1929         up_write(&obj->oo_sem);
1930 }
1931
1932 static int osd_object_write_locked(const struct lu_env *env,
1933                                    struct dt_object *dt)
1934 {
1935         struct osd_object *obj = osd_dt_obj(dt);
1936
1937         LINVRNT(osd_invariant(obj));
1938
1939         return obj->oo_owner == env;
1940 }
1941
1942 static struct timespec *osd_inode_time(const struct lu_env *env,
1943                                        struct inode *inode, __u64 seconds)
1944 {
1945         struct osd_thread_info  *oti = osd_oti_get(env);
1946         struct timespec         *t   = &oti->oti_time;
1947
1948         t->tv_sec = seconds;
1949         t->tv_nsec = 0;
1950         *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
1951         return t;
1952 }
1953
1954 static void osd_inode_getattr(const struct lu_env *env,
1955                               struct inode *inode, struct lu_attr *attr)
1956 {
1957         attr->la_valid  |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
1958                            LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
1959                            LA_FLAGS | LA_NLINK | LA_RDEV | LA_BLKSIZE |
1960                            LA_TYPE;
1961
1962         attr->la_atime   = LTIME_S(inode->i_atime);
1963         attr->la_mtime   = LTIME_S(inode->i_mtime);
1964         attr->la_ctime   = LTIME_S(inode->i_ctime);
1965         attr->la_mode    = inode->i_mode;
1966         attr->la_size    = i_size_read(inode);
1967         attr->la_blocks  = inode->i_blocks;
1968         attr->la_uid     = i_uid_read(inode);
1969         attr->la_gid     = i_gid_read(inode);
1970         attr->la_flags   = ll_inode_to_ext_flags(inode->i_flags);
1971         attr->la_nlink   = inode->i_nlink;
1972         attr->la_rdev    = inode->i_rdev;
1973         attr->la_blksize = 1 << inode->i_blkbits;
1974         attr->la_blkbits = inode->i_blkbits;
1975 }
1976
1977 static int osd_attr_get(const struct lu_env *env,
1978                         struct dt_object *dt,
1979                         struct lu_attr *attr)
1980 {
1981         struct osd_object *obj = osd_dt_obj(dt);
1982
1983         if (unlikely(!dt_object_exists(dt)))
1984                 return -ENOENT;
1985         if (unlikely(obj->oo_destroyed))
1986                 return -ENOENT;
1987
1988         LASSERT(!dt_object_remote(dt));
1989         LINVRNT(osd_invariant(obj));
1990
1991         spin_lock(&obj->oo_guard);
1992         osd_inode_getattr(env, obj->oo_inode, attr);
1993         if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
1994                 attr->la_flags |= LUSTRE_ORPHAN_FL;
1995         spin_unlock(&obj->oo_guard);
1996
1997         return 0;
1998 }
1999
2000 static int osd_declare_attr_set(const struct lu_env *env,
2001                                 struct dt_object *dt,
2002                                 const struct lu_attr *attr,
2003                                 struct thandle *handle)
2004 {
2005         struct osd_thandle     *oh;
2006         struct osd_object      *obj;
2007         struct osd_thread_info *info = osd_oti_get(env);
2008         struct lquota_id_info  *qi = &info->oti_qi;
2009         qid_t                   uid;
2010         qid_t                   gid;
2011         long long               bspace;
2012         int                     rc = 0;
2013         bool                    enforce;
2014         ENTRY;
2015
2016         LASSERT(dt != NULL);
2017         LASSERT(handle != NULL);
2018
2019         obj = osd_dt_obj(dt);
2020         LASSERT(osd_invariant(obj));
2021
2022         oh = container_of0(handle, struct osd_thandle, ot_super);
2023         LASSERT(oh->ot_handle == NULL);
2024
2025         osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2026                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2027
2028         osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2029                              osd_dto_credits_noquota[DTO_XATTR_SET]);
2030
2031         if (attr == NULL || obj->oo_inode == NULL)
2032                 RETURN(rc);
2033
2034         bspace   = obj->oo_inode->i_blocks;
2035         bspace <<= obj->oo_inode->i_sb->s_blocksize_bits;
2036         bspace   = toqb(bspace);
2037
2038         /* Changing ownership is always preformed by super user, it should not
2039          * fail with EDQUOT.
2040          *
2041          * We still need to call the osd_declare_qid() to calculate the journal
2042          * credits for updating quota accounting files and to trigger quota
2043          * space adjustment once the operation is completed.*/
2044         if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2045                 /* USERQUOTA */
2046                 uid = i_uid_read(obj->oo_inode);
2047                 qi->lqi_type = USRQUOTA;
2048                 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2049                 /* inode accounting */
2050                 qi->lqi_is_blk = false;
2051
2052                 /* one more inode for the new uid ... */
2053                 qi->lqi_id.qid_uid = attr->la_uid;
2054                 qi->lqi_space      = 1;
2055                 /* Reserve credits for the new uid */
2056                 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2057                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2058                         rc = 0;
2059                 if (rc)
2060                         RETURN(rc);
2061
2062                 /* and one less inode for the current uid */
2063                 qi->lqi_id.qid_uid = uid;
2064                 qi->lqi_space      = -1;
2065                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2066                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2067                         rc = 0;
2068                 if (rc)
2069                         RETURN(rc);
2070
2071                 /* block accounting */
2072                 qi->lqi_is_blk = true;
2073
2074                 /* more blocks for the new uid ... */
2075                 qi->lqi_id.qid_uid = attr->la_uid;
2076                 qi->lqi_space      = bspace;
2077                 /*
2078                  * Credits for the new uid has been reserved, re-use "obj"
2079                  * to save credit reservation.
2080                  */
2081                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2082                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2083                         rc = 0;
2084                 if (rc)
2085                         RETURN(rc);
2086
2087                 /* and finally less blocks for the current uid */
2088                 qi->lqi_id.qid_uid = uid;
2089                 qi->lqi_space      = -bspace;
2090                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2091                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2092                         rc = 0;
2093                 if (rc)
2094                         RETURN(rc);
2095
2096                 /* GROUP QUOTA */
2097                 gid = i_gid_read(obj->oo_inode);
2098                 qi->lqi_type = GRPQUOTA;
2099                 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2100
2101                 /* inode accounting */
2102                 qi->lqi_is_blk = false;
2103
2104                 /* one more inode for the new gid ... */
2105                 qi->lqi_id.qid_gid = attr->la_gid;
2106                 qi->lqi_space      = 1;
2107                 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2108                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2109                         rc = 0;
2110                 if (rc)
2111                         RETURN(rc);
2112
2113                 /* and one less inode for the current gid */
2114                 qi->lqi_id.qid_gid = gid;
2115                 qi->lqi_space      = -1;
2116                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2117                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2118                         rc = 0;
2119                 if (rc)
2120                         RETURN(rc);
2121
2122                 /* block accounting */
2123                 qi->lqi_is_blk = true;
2124
2125                 /* more blocks for the new gid ... */
2126                 qi->lqi_id.qid_gid = attr->la_gid;
2127                 qi->lqi_space      = bspace;
2128                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2129                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2130                         rc = 0;
2131                 if (rc)
2132                         RETURN(rc);
2133
2134                 /* and finally less blocks for the current gid */
2135                 qi->lqi_id.qid_gid = gid;
2136                 qi->lqi_space      = -bspace;
2137                 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2138                 if (rc == -EDQUOT || rc == -EINPROGRESS)
2139                         rc = 0;
2140                 if (rc)
2141                         RETURN(rc);
2142         }
2143
2144         RETURN(rc);
2145 }
2146
2147 static int osd_inode_setattr(const struct lu_env *env,
2148                              struct inode *inode, const struct lu_attr *attr)
2149 {
2150         __u64 bits = attr->la_valid;
2151
2152         /* Only allow set size for regular file */
2153         if (!S_ISREG(inode->i_mode))
2154                 bits &= ~(LA_SIZE | LA_BLOCKS);
2155
2156         if (bits == 0)
2157                 return 0;
2158
2159         if (bits & LA_ATIME)
2160                 inode->i_atime  = *osd_inode_time(env, inode, attr->la_atime);
2161         if (bits & LA_CTIME)
2162                 inode->i_ctime  = *osd_inode_time(env, inode, attr->la_ctime);
2163         if (bits & LA_MTIME)
2164                 inode->i_mtime  = *osd_inode_time(env, inode, attr->la_mtime);
2165         if (bits & LA_SIZE) {
2166                 LDISKFS_I(inode)->i_disksize = attr->la_size;
2167                 i_size_write(inode, attr->la_size);
2168         }
2169
2170 #if 0
2171         /* OSD should not change "i_blocks" which is used by quota.
2172          * "i_blocks" should be changed by ldiskfs only. */
2173         if (bits & LA_BLOCKS)
2174                 inode->i_blocks = attr->la_blocks;
2175 #endif
2176         if (bits & LA_MODE)
2177                 inode->i_mode = (inode->i_mode & S_IFMT) |
2178                                 (attr->la_mode & ~S_IFMT);
2179         if (bits & LA_UID)
2180                 i_uid_write(inode, attr->la_uid);
2181         if (bits & LA_GID)
2182                 i_gid_write(inode, attr->la_gid);
2183         if (bits & LA_NLINK)
2184                 set_nlink(inode, attr->la_nlink);
2185         if (bits & LA_RDEV)
2186                 inode->i_rdev = attr->la_rdev;
2187
2188         if (bits & LA_FLAGS) {
2189                 /* always keep S_NOCMTIME */
2190                 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2191                                  S_NOCMTIME;
2192         }
2193         return 0;
2194 }
2195
2196 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2197 {
2198         if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2199             (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2200                 struct iattr    iattr;
2201                 int             rc;
2202
2203                 ll_vfs_dq_init(inode);
2204                 iattr.ia_valid = 0;
2205                 if (attr->la_valid & LA_UID)
2206                         iattr.ia_valid |= ATTR_UID;
2207                 if (attr->la_valid & LA_GID)
2208                         iattr.ia_valid |= ATTR_GID;
2209                 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2210                 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2211
2212                 rc = ll_vfs_dq_transfer(inode, &iattr);
2213                 if (rc) {
2214                         CERROR("%s: quota transfer failed: rc = %d. Is quota "
2215                                "enforcement enabled on the ldiskfs "
2216                                "filesystem?\n", inode->i_sb->s_id, rc);
2217                         return rc;
2218                 }
2219         }
2220         return 0;
2221 }
2222
2223 static int osd_attr_set(const struct lu_env *env,
2224                         struct dt_object *dt,
2225                         const struct lu_attr *attr,
2226                         struct thandle *handle)
2227 {
2228         struct osd_object *obj = osd_dt_obj(dt);
2229         struct inode      *inode;
2230         int rc;
2231
2232         if (!dt_object_exists(dt))
2233                 return -ENOENT;
2234
2235         LASSERT(handle != NULL);
2236         LASSERT(!dt_object_remote(dt));
2237         LASSERT(osd_invariant(obj));
2238
2239         osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2240
2241         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING)) {
2242                 struct osd_thread_info  *oti  = osd_oti_get(env);
2243                 const struct lu_fid     *fid0 = lu_object_fid(&dt->do_lu);
2244                 struct lu_fid           *fid1 = &oti->oti_fid;
2245                 struct osd_inode_id     *id   = &oti->oti_id;
2246                 struct iam_path_descr   *ipd;
2247                 struct iam_container    *bag;
2248                 struct osd_thandle      *oh;
2249                 int                      rc;
2250
2251                 fid_cpu_to_be(fid1, fid0);
2252                 memset(id, 1, sizeof(*id));
2253                 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2254                                   fid0)->oi_dir.od_container;
2255                 ipd = osd_idx_ipd_get(env, bag);
2256                 if (unlikely(ipd == NULL))
2257                         RETURN(-ENOMEM);
2258
2259                 oh = container_of0(handle, struct osd_thandle, ot_super);
2260                 rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)fid1,
2261                                 (const struct iam_rec *)id, ipd);
2262                 osd_ipd_put(env, bag, ipd);
2263                 return(rc > 0 ? 0 : rc);
2264         }
2265
2266         inode = obj->oo_inode;
2267
2268         rc = osd_quota_transfer(inode, attr);
2269         if (rc)
2270                 return rc;
2271
2272         spin_lock(&obj->oo_guard);
2273         rc = osd_inode_setattr(env, inode, attr);
2274         spin_unlock(&obj->oo_guard);
2275         if (rc != 0)
2276                 GOTO(out, rc);
2277
2278         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2279
2280         if (!(attr->la_valid & LA_FLAGS))
2281                 GOTO(out, rc);
2282
2283         /* Let's check if there are extra flags need to be set into LMA */
2284         if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
2285                 struct osd_thread_info *info = osd_oti_get(env);
2286                 struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
2287
2288                 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, lma);
2289                 if (rc != 0)
2290                         GOTO(out, rc);
2291
2292                 lma->lma_incompat |=
2293                         lustre_to_lma_flags(attr->la_flags);
2294                 lustre_lma_swab(lma);
2295                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
2296                                      lma, sizeof(*lma), XATTR_REPLACE);
2297                 if (rc != 0) {
2298                         struct osd_device *osd = osd_obj2dev(obj);
2299
2300                         CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
2301                               osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
2302                               lma->lma_incompat, rc);
2303                 } else {
2304                         obj->oo_lma_flags =
2305                                 attr->la_flags & LUSTRE_LMA_FL_MASKS;
2306                 }
2307                 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
2308         }
2309 out:
2310         osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
2311
2312         return rc;
2313 }
2314
2315 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
2316                                            struct osd_object *obj,
2317                                            const char *name, const int namelen)
2318 {
2319         return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
2320 }
2321
2322 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
2323                       umode_t mode, struct dt_allocation_hint *hint,
2324                       struct thandle *th)
2325 {
2326         int result;
2327         struct osd_device  *osd = osd_obj2dev(obj);
2328         struct osd_thandle *oth;
2329         struct dt_object   *parent = NULL;
2330         struct inode       *inode;
2331
2332         LINVRNT(osd_invariant(obj));
2333         LASSERT(obj->oo_inode == NULL);
2334         LASSERT(obj->oo_hl_head == NULL);
2335
2336         if (S_ISDIR(mode) && ldiskfs_pdo) {
2337                 obj->oo_hl_head =ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
2338                 if (obj->oo_hl_head == NULL)
2339                         return -ENOMEM;
2340         }
2341
2342         oth = container_of(th, struct osd_thandle, ot_super);
2343         LASSERT(oth->ot_handle->h_transaction != NULL);
2344
2345         if (hint != NULL && hint->dah_parent != NULL &&
2346             !dt_object_remote(hint->dah_parent))
2347                 parent = hint->dah_parent;
2348
2349         inode = ldiskfs_create_inode(oth->ot_handle,
2350                                      parent ? osd_dt_obj(parent)->oo_inode :
2351                                               osd_sb(osd)->s_root->d_inode,
2352                                      mode);
2353         if (!IS_ERR(inode)) {
2354                 /* Do not update file c/mtime in ldiskfs. */
2355                 inode->i_flags |= S_NOCMTIME;
2356
2357                 /* For new created object, it must be consistent,
2358                  * and it is unnecessary to scrub against it. */
2359                 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
2360
2361                 obj->oo_inode = inode;
2362                 result = 0;
2363         } else {
2364                 if (obj->oo_hl_head != NULL) {
2365                         ldiskfs_htree_lock_head_free(obj->oo_hl_head);
2366                         obj->oo_hl_head = NULL;
2367                 }
2368                 result = PTR_ERR(inode);
2369         }
2370         LINVRNT(osd_invariant(obj));
2371         return result;
2372 }
2373
2374 enum {
2375         OSD_NAME_LEN = 255
2376 };
2377
2378 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
2379                      struct lu_attr *attr,
2380                      struct dt_allocation_hint *hint,
2381                      struct dt_object_format *dof,
2382                      struct thandle *th)
2383 {
2384         int result;
2385         struct osd_thandle *oth;
2386         __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX));
2387
2388         LASSERT(S_ISDIR(attr->la_mode));
2389
2390         oth = container_of(th, struct osd_thandle, ot_super);
2391         LASSERT(oth->ot_handle->h_transaction != NULL);
2392         result = osd_mkfile(info, obj, mode, hint, th);
2393
2394         return result;
2395 }
2396
2397 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
2398                         struct lu_attr *attr,
2399                         struct dt_allocation_hint *hint,
2400                         struct dt_object_format *dof,
2401                         struct thandle *th)
2402 {
2403         int result;
2404         struct osd_thandle *oth;
2405         const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
2406
2407         __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
2408
2409         LASSERT(S_ISREG(attr->la_mode));
2410
2411         oth = container_of(th, struct osd_thandle, ot_super);
2412         LASSERT(oth->ot_handle->h_transaction != NULL);
2413
2414         result = osd_mkfile(info, obj, mode, hint, th);
2415         if (result == 0) {
2416                 LASSERT(obj->oo_inode != NULL);
2417                 if (feat->dif_flags & DT_IND_VARKEY)
2418                         result = iam_lvar_create(obj->oo_inode,
2419                                                  feat->dif_keysize_max,
2420                                                  feat->dif_ptrsize,
2421                                                  feat->dif_recsize_max,
2422                                                  oth->ot_handle);
2423                 else
2424                         result = iam_lfix_create(obj->oo_inode,
2425                                                  feat->dif_keysize_max,
2426                                                  feat->dif_ptrsize,
2427                                                  feat->dif_recsize_max,
2428                                                  oth->ot_handle);
2429
2430         }
2431         return result;
2432 }
2433
2434 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
2435                      struct lu_attr *attr,
2436                      struct dt_allocation_hint *hint,
2437                      struct dt_object_format *dof,
2438                      struct thandle *th)
2439 {
2440         LASSERT(S_ISREG(attr->la_mode));
2441         return osd_mkfile(info, obj, (attr->la_mode &
2442                                (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2443 }
2444
2445 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
2446                      struct lu_attr *attr,
2447                      struct dt_allocation_hint *hint,
2448                      struct dt_object_format *dof,
2449                      struct thandle *th)
2450 {
2451         LASSERT(S_ISLNK(attr->la_mode));
2452         return osd_mkfile(info, obj, (attr->la_mode &
2453                               (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2454 }
2455
2456 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
2457                      struct lu_attr *attr,
2458                      struct dt_allocation_hint *hint,
2459                      struct dt_object_format *dof,
2460                      struct thandle *th)
2461 {
2462         umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
2463         int result;
2464
2465         LINVRNT(osd_invariant(obj));
2466         LASSERT(obj->oo_inode == NULL);
2467         LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
2468                 S_ISFIFO(mode) || S_ISSOCK(mode));
2469
2470         result = osd_mkfile(info, obj, mode, hint, th);
2471         if (result == 0) {
2472                 LASSERT(obj->oo_inode != NULL);
2473                 /*
2474                  * This inode should be marked dirty for i_rdev.  Currently
2475                  * that is done in the osd_attr_init().
2476                  */
2477                 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
2478                                    attr->la_rdev);
2479         }
2480         LINVRNT(osd_invariant(obj));
2481         return result;
2482 }
2483
2484 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
2485                               struct lu_attr *,
2486                               struct dt_allocation_hint *hint,
2487                               struct dt_object_format *dof,
2488                               struct thandle *);
2489
2490 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
2491 {
2492         osd_obj_type_f result;
2493
2494         switch (type) {
2495         case DFT_DIR:
2496                 result = osd_mkdir;
2497                 break;
2498         case DFT_REGULAR:
2499                 result = osd_mkreg;
2500                 break;
2501         case DFT_SYM:
2502                 result = osd_mksym;
2503                 break;
2504         case DFT_NODE:
2505                 result = osd_mknod;
2506                 break;
2507         case DFT_INDEX:
2508                 result = osd_mk_index;
2509                 break;
2510
2511         default:
2512                 LBUG();
2513                 break;
2514         }
2515         return result;
2516 }
2517
2518
2519 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
2520                         struct dt_object *parent, struct dt_object *child,
2521                         umode_t child_mode)
2522 {
2523         LASSERT(ah);
2524
2525         ah->dah_parent = parent;
2526         ah->dah_mode = child_mode;
2527
2528         if (parent != NULL && !dt_object_remote(parent)) {
2529                 /* will help to find FID->ino at dt_insert("..") */
2530                 struct osd_object *pobj = osd_dt_obj(parent);
2531                 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
2532         }
2533 }
2534
2535 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
2536                           struct lu_attr *attr, struct dt_object_format *dof)
2537 {
2538         struct inode   *inode = obj->oo_inode;
2539         __u64           valid = attr->la_valid;
2540         int             result;
2541
2542         attr->la_valid &= ~(LA_TYPE | LA_MODE);
2543
2544         if (dof->dof_type != DFT_NODE)
2545                 attr->la_valid &= ~LA_RDEV;
2546         if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
2547                 attr->la_valid &= ~LA_ATIME;
2548         if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
2549                 attr->la_valid &= ~LA_CTIME;
2550         if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
2551                 attr->la_valid &= ~LA_MTIME;
2552
2553         result = osd_quota_transfer(inode, attr);
2554         if (result)
2555                 return;
2556
2557         if (attr->la_valid != 0) {
2558                 result = osd_inode_setattr(info->oti_env, inode, attr);
2559                 /*
2560                  * The osd_inode_setattr() should always succeed here.  The
2561                  * only error that could be returned is EDQUOT when we are
2562                  * trying to change the UID or GID of the inode. However, this
2563                  * should not happen since quota enforcement is no longer
2564                  * enabled on ldiskfs (lquota takes care of it).
2565                  */
2566                 LASSERTF(result == 0, "%d\n", result);
2567                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2568         }
2569
2570         attr->la_valid = valid;
2571 }
2572
2573 /**
2574  * Helper function for osd_object_create()
2575  *
2576  * \retval 0, on success
2577  */
2578 static int __osd_object_create(struct osd_thread_info *info,
2579                                struct osd_object *obj, struct lu_attr *attr,
2580                                struct dt_allocation_hint *hint,
2581                                struct dt_object_format *dof,
2582                                struct thandle *th)
2583 {
2584         int     result;
2585         __u32   umask;
2586
2587         osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
2588
2589         /* we drop umask so that permissions we pass are not affected */
2590         umask = current->fs->umask;
2591         current->fs->umask = 0;
2592
2593         result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
2594                                                   th);
2595         if (likely(obj->oo_inode != NULL)) {
2596                 LASSERT(obj->oo_inode->i_state & I_NEW);
2597
2598                 /* Unlock the inode before attr initialization to avoid
2599                  * unnecessary dqget operations. LU-6378 */
2600                 unlock_new_inode(obj->oo_inode);
2601         }
2602
2603         if (likely(result == 0)) {
2604                 osd_attr_init(info, obj, attr, dof);
2605                 osd_object_init0(obj);
2606         }
2607
2608         /* restore previous umask value */
2609         current->fs->umask = umask;
2610
2611         osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
2612
2613         return result;
2614 }
2615
2616 /**
2617  * Helper function for osd_object_create()
2618  *
2619  * \retval 0, on success
2620  */
2621 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
2622                            const struct lu_fid *fid, struct thandle *th)
2623 {
2624         struct osd_thread_info *info = osd_oti_get(env);
2625         struct osd_inode_id    *id   = &info->oti_id;
2626         struct osd_device      *osd  = osd_obj2dev(obj);
2627         struct osd_thandle     *oh;
2628         int                     rc;
2629
2630         LASSERT(obj->oo_inode != NULL);
2631
2632         oh = container_of0(th, struct osd_thandle, ot_super);
2633         LASSERT(oh->ot_handle);
2634         osd_trans_exec_op(env, th, OSD_OT_INSERT);
2635
2636         osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
2637         rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
2638                            OI_CHECK_FLD, NULL);
2639         osd_trans_exec_check(env, th, OSD_OT_INSERT);
2640
2641         return rc;
2642 }
2643
2644 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
2645                    u64 seq, struct lu_seq_range *range)
2646 {
2647         struct seq_server_site  *ss = osd_seq_site(osd);
2648
2649         if (fid_seq_is_idif(seq)) {
2650                 fld_range_set_ost(range);
2651                 range->lsr_index = idif_ost_idx(seq);
2652                 return 0;
2653         }
2654
2655         if (!fid_seq_in_fldb(seq)) {
2656                 fld_range_set_mdt(range);
2657                 if (ss != NULL)
2658                         /* FIXME: If ss is NULL, it suppose not get lsr_index
2659                          * at all */
2660                         range->lsr_index = ss->ss_node_id;
2661                 return 0;
2662         }
2663
2664         LASSERT(ss != NULL);
2665         fld_range_set_any(range);
2666         /* OSD will only do local fld lookup */
2667         return fld_local_lookup(env, ss->ss_server_fld, seq, range);
2668 }
2669
2670 /*
2671  * Concurrency: no external locking is necessary.
2672  */
2673 static int osd_declare_object_create(const struct lu_env *env,
2674                                      struct dt_object *dt,
2675                                      struct lu_attr *attr,
2676                                      struct dt_allocation_hint *hint,
2677                                      struct dt_object_format *dof,
2678                                      struct thandle *handle)
2679 {
2680         struct osd_thandle      *oh;
2681         int                      rc;
2682         ENTRY;
2683
2684         LASSERT(handle != NULL);
2685
2686         oh = container_of0(handle, struct osd_thandle, ot_super);
2687         LASSERT(oh->ot_handle == NULL);
2688
2689         /* EA object consumes more credits than regular object: osd_mk_index
2690          * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
2691          * leaf_node, could involves the block, block bitmap, groups, GDT
2692          * change for each block, so add 4 * 2 credits in that case. */
2693         osd_trans_declare_op(env, oh, OSD_OT_CREATE,
2694                              osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
2695                              (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
2696         /* Reuse idle OI block may cause additional one OI block
2697          * to be changed. */
2698         osd_trans_declare_op(env, oh, OSD_OT_INSERT,
2699                              osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
2700
2701         if (!attr)
2702                 RETURN(0);
2703
2704         rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid, 1, oh,
2705                                    osd_dt_obj(dt), false, NULL, false);
2706         if (rc != 0)
2707                 RETURN(rc);
2708
2709         /* will help to find FID->ino mapping at dt_insert() */
2710         rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
2711                                    osd_dt_obj(dt));
2712
2713         RETURN(rc);
2714 }
2715
2716 static int osd_object_create(const struct lu_env *env, struct dt_object *dt,
2717                              struct lu_attr *attr,
2718                              struct dt_allocation_hint *hint,
2719                              struct dt_object_format *dof, struct thandle *th)
2720 {
2721         const struct lu_fid     *fid    = lu_object_fid(&dt->do_lu);
2722         struct osd_object       *obj    = osd_dt_obj(dt);
2723         struct osd_thread_info  *info   = osd_oti_get(env);
2724         int result;
2725         ENTRY;
2726
2727         if (dt_object_exists(dt))
2728                 return -EEXIST;
2729
2730         LINVRNT(osd_invariant(obj));
2731         LASSERT(!dt_object_remote(dt));
2732         LASSERT(osd_write_locked(env, obj));
2733         LASSERT(th != NULL);
2734
2735         if (unlikely(fid_is_acct(fid)))
2736                 /* Quota files can't be created from the kernel any more,
2737                  * 'tune2fs -O quota' will take care of creating them */
2738                 RETURN(-EPERM);
2739
2740         result = __osd_object_create(info, obj, attr, hint, dof, th);
2741         if (result == 0) {
2742                 result = __osd_oi_insert(env, obj, fid, th);
2743                 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
2744                         obj->oo_dt.do_body_ops = &osd_body_ops;
2745         }
2746         LASSERT(ergo(result == 0,
2747                 dt_object_exists(dt) && !dt_object_remote(dt)));
2748
2749         LASSERT(osd_invariant(obj));
2750         RETURN(result);
2751 }
2752
2753 /**
2754  * Called to destroy on-disk representation of the object
2755  *
2756  * Concurrency: must be locked
2757  */
2758 static int osd_declare_object_destroy(const struct lu_env *env,
2759                                       struct dt_object *dt,
2760                                       struct thandle *th)
2761 {
2762         struct osd_object  *obj = osd_dt_obj(dt);
2763         struct inode       *inode = obj->oo_inode;
2764         struct osd_thandle *oh;
2765         int                 rc;
2766         ENTRY;
2767
2768         if (inode == NULL)
2769                 RETURN(-ENOENT);
2770
2771         oh = container_of0(th, struct osd_thandle, ot_super);
2772         LASSERT(oh->ot_handle == NULL);
2773
2774         osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
2775                              osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
2776         /* Recycle idle OI leaf may cause additional three OI blocks
2777          * to be changed. */
2778         osd_trans_declare_op(env, oh, OSD_OT_DELETE,
2779                              osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
2780         /* one less inode */
2781         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2782                                    -1, oh, obj, false, NULL, false);
2783         if (rc)
2784                 RETURN(rc);
2785         /* data to be truncated */
2786         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2787                                    0, oh, obj, true, NULL, false);
2788         if (rc)
2789                 RETURN(rc);
2790
2791         /* will help to find FID->ino when this object is being
2792          * added to PENDING/ */
2793         rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
2794
2795         RETURN(rc);
2796 }
2797
2798 static int osd_object_destroy(const struct lu_env *env,
2799                               struct dt_object *dt,
2800                               struct thandle *th)
2801 {
2802         const struct lu_fid    *fid = lu_object_fid(&dt->do_lu);
2803         struct osd_object      *obj = osd_dt_obj(dt);
2804         struct inode           *inode = obj->oo_inode;
2805         struct osd_device      *osd = osd_obj2dev(obj);
2806         struct osd_thandle     *oh;
2807         int                     result;
2808         ENTRY;
2809
2810         oh = container_of0(th, struct osd_thandle, ot_super);
2811         LASSERT(oh->ot_handle);
2812         LASSERT(inode);
2813         LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
2814
2815         if (unlikely(fid_is_acct(fid)))
2816                 RETURN(-EPERM);
2817
2818         if (S_ISDIR(inode->i_mode)) {
2819                 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
2820                         inode->i_nlink == 2);
2821                 /* it will check/delete the inode from remote parent,
2822                  * how to optimize it? unlink performance impaction XXX */
2823                 result = osd_delete_from_remote_parent(env, osd, obj, oh);
2824                 if (result != 0 && result != -ENOENT) {
2825                         CERROR("%s: delete inode "DFID": rc = %d\n",
2826                                osd_name(osd), PFID(fid), result);
2827                 }
2828                 spin_lock(&obj->oo_guard);
2829                 clear_nlink(inode);
2830                 spin_unlock(&obj->oo_guard);
2831                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2832         }
2833
2834         osd_trans_exec_op(env, th, OSD_OT_DESTROY);
2835
2836         ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
2837         result = osd_oi_delete(osd_oti_get(env), osd, fid, oh->ot_handle,
2838                                OI_CHECK_FLD);
2839
2840         osd_trans_exec_check(env, th, OSD_OT_DESTROY);
2841         /* XXX: add to ext3 orphan list */
2842         /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
2843
2844         /* not needed in the cache anymore */
2845         set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
2846         obj->oo_destroyed = 1;
2847
2848         RETURN(0);
2849 }
2850
2851 /**
2852  * Put the fid into lustre_mdt_attrs, and then place the structure
2853  * inode's ea. This fid should not be altered during the life time
2854  * of the inode.
2855  *
2856  * \retval +ve, on success
2857  * \retval -ve, on error
2858  *
2859  * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
2860  */
2861 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
2862                    const struct lu_fid *fid, __u32 compat, __u32 incompat)
2863 {
2864         struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
2865         int                      rc;
2866         ENTRY;
2867
2868         if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
2869                 RETURN(0);
2870
2871         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
2872                 rc = -ENOMEM;
2873
2874         lustre_lma_init(lma, fid, compat, incompat);
2875         lustre_lma_swab(lma);
2876
2877         rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma, sizeof(*lma),
2878                              XATTR_CREATE);
2879         /* LMA may already exist, but we need to check that all the
2880          * desired compat/incompat flags have been added. */
2881         if (unlikely(rc == -EEXIST)) {
2882                 if (compat == 0 && incompat == 0)
2883                         RETURN(0);
2884
2885                 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
2886                                      XATTR_NAME_LMA, info->oti_mdt_attrs_old,
2887                                      LMA_OLD_SIZE);
2888                 if (rc <= 0)
2889                         RETURN(-EINVAL);
2890
2891                 lustre_lma_swab(lma);
2892                 if (!(~lma->lma_compat & compat) &&
2893                     !(~lma->lma_incompat & incompat))
2894                         RETURN(0);
2895
2896                 lma->lma_compat |= compat;
2897                 lma->lma_incompat |= incompat;
2898                 lustre_lma_swab(lma);
2899                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
2900                                      sizeof(*lma), XATTR_REPLACE);
2901         }
2902
2903         RETURN(rc);
2904 }
2905
2906 /**
2907  * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
2908  * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
2909  * To have compatilibility with 1.8 ldiskfs driver we need to have
2910  * magic number at start of fid data.
2911  * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
2912  * its inmemory API.
2913  */
2914 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
2915                                          const struct lu_fid *fid)
2916 {
2917         if (!fid_is_namespace_visible(fid) ||
2918             OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
2919                 param->edp_magic = 0;
2920                 return;
2921         }
2922
2923         param->edp_magic = LDISKFS_LUFID_MAGIC;
2924         param->edp_len =  sizeof(struct lu_fid) + 1;
2925         fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
2926 }
2927
2928 /**
2929  * Try to read the fid from inode ea into dt_rec.
2930  *
2931  * \param fid object fid.
2932  *
2933  * \retval 0 on success
2934  */
2935 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
2936                           __u32 ino, struct lu_fid *fid,
2937                           struct osd_inode_id *id)
2938 {
2939         struct osd_thread_info *info  = osd_oti_get(env);
2940         struct inode           *inode;
2941         ENTRY;
2942
2943         osd_id_gen(id, ino, OSD_OII_NOGEN);
2944         inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
2945         if (IS_ERR(inode))
2946                 RETURN(PTR_ERR(inode));
2947
2948         iput(inode);
2949         RETURN(0);
2950 }
2951
2952 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
2953                                         struct inode *dir,
2954                                         struct inode *parent_dir,
2955                                         const struct lu_fid *dot_fid,
2956                                         const struct lu_fid *dot_dot_fid,
2957                                         struct osd_thandle *oth)
2958 {
2959         struct ldiskfs_dentry_param *dot_ldp;
2960         struct ldiskfs_dentry_param *dot_dot_ldp;
2961         __u32 saved_nlink = dir->i_nlink;
2962         int rc;
2963
2964         dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
2965         osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
2966
2967         dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
2968         dot_ldp->edp_magic = 0;
2969
2970         rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
2971                                     dir, dot_ldp, dot_dot_ldp);
2972         /* The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
2973          * the subseqent ref_add() will increase the dir->i_nlink
2974          * as 3. That is incorrect for new created directory.
2975          *
2976          * It looks like hack, because we want to make the OSD API
2977          * to be order-independent for new created directory object
2978          * between dt_insert(..) and ref_add() operations.
2979          *
2980          * Here, we only restore the in-RAM dir-inode's nlink attr,
2981          * becuase if the nlink attr is not 2, then there will be
2982          * ref_add() called following the dt_insert(..), such call
2983          * will make both the in-RAM and on-disk dir-inode's nlink
2984          * attr to be set as 2. LU-7447 */
2985         set_nlink(dir, saved_nlink);
2986         return rc;
2987 }
2988
2989 /**
2990  * Create an local agent inode for remote entry
2991  */
2992 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
2993                                                   struct osd_device *osd,
2994                                                   struct osd_object *pobj,
2995                                                   const struct lu_fid *fid,
2996                                                   __u32 type,
2997                                                   struct thandle *th)
2998 {
2999         struct osd_thread_info  *info = osd_oti_get(env);
3000         struct inode            *local;
3001         struct osd_thandle      *oh;
3002         int                     rc;
3003         ENTRY;
3004
3005         LASSERT(th);
3006         oh = container_of(th, struct osd_thandle, ot_super);
3007         LASSERT(oh->ot_handle->h_transaction != NULL);
3008
3009         local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
3010         if (IS_ERR(local)) {
3011                 CERROR("%s: create local error %d\n", osd_name(osd),
3012                        (int)PTR_ERR(local));
3013                 RETURN(local);
3014         }
3015
3016         ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3017         unlock_new_inode(local);
3018
3019         /* Set special LMA flag for local agent inode */
3020         rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3021         if (rc != 0) {
3022                 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3023                        osd_name(osd), PFID(fid), rc);
3024                 RETURN(ERR_PTR(rc));
3025         }
3026
3027         if (!S_ISDIR(type))
3028                 RETURN(local);
3029
3030         rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3031                                          lu_object_fid(&pobj->oo_dt.do_lu),
3032                                          fid, oh);
3033         if (rc != 0) {
3034                 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3035                         osd_name(osd), PFID(fid), rc);
3036                 RETURN(ERR_PTR(rc));
3037         }
3038
3039         RETURN(local);
3040 }
3041
3042 /**
3043  * when direntry is deleted, we have to take care of possible agent inode
3044  * referenced by that. unfortunately we can't do this at that point:
3045  * iget() within a running transaction leads to deadlock and we better do
3046  * not call that every delete declaration to save performance. so we put
3047  * a potention agent inode on a list and process that once the transaction
3048  * is over. Notice it's not any worse in terms of real orphans as regular
3049  * object destroy doesn't put inodes on the on-disk orphan list. this should
3050  * be addressed separately
3051  */
3052 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3053                                             struct osd_thandle *oh,
3054                                             __u32 ino)
3055 {
3056         struct osd_device      *osd = osd_dt_dev(oh->ot_super.th_dev);
3057         struct osd_obj_orphan *oor;
3058
3059         OBD_ALLOC_PTR(oor);
3060         if (oor == NULL)
3061                 return -ENOMEM;
3062
3063         oor->oor_ino = ino;
3064         oor->oor_env = (struct lu_env *)env;
3065         spin_lock(&osd->od_osfs_lock);
3066         list_add(&oor->oor_list, &osd->od_orphan_list);
3067         spin_unlock(&osd->od_osfs_lock);
3068
3069         oh->ot_remove_agents = 1;
3070
3071         return 0;
3072
3073 }
3074
3075 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3076                                                 struct osd_device *osd)
3077 {
3078         struct osd_thread_info *info = osd_oti_get(env);
3079         struct osd_obj_orphan *oor, *tmp;
3080         struct osd_inode_id id;
3081         struct list_head list;
3082         struct inode *inode;
3083         struct lu_fid fid;
3084         handle_t *jh;
3085         __u32 ino;
3086
3087         INIT_LIST_HEAD(&list);
3088
3089         spin_lock(&osd->od_osfs_lock);
3090         list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3091                 if (oor->oor_env == env) {
3092                         list_del(&oor->oor_list);
3093                         list_add(&oor->oor_list, &list);
3094                 }
3095         }
3096         spin_unlock(&osd->od_osfs_lock);
3097
3098         list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3099
3100                 ino = oor->oor_ino;
3101
3102                 list_del(&oor->oor_list);
3103                 OBD_FREE_PTR(oor);
3104
3105                 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3106                 inode = osd_iget_fid(info, osd, &id, &fid);
3107                 if (IS_ERR(inode))
3108                         continue;
3109
3110                 if (!osd_remote_fid(env, osd, &fid)) {
3111                         iput(inode);
3112                         continue;
3113                 }
3114
3115                 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
3116                 clear_nlink(inode);
3117                 mark_inode_dirty(inode);
3118                 ldiskfs_journal_stop(jh);
3119                 iput(inode);
3120         }
3121
3122         return 0;
3123 }
3124
3125 /**
3126  * OSD layer object create function for interoperability mode (b11826).
3127  * This is mostly similar to osd_object_create(). Only difference being, fid is
3128  * inserted into inode ea here.
3129  *
3130  * \retval   0, on success
3131  * \retval -ve, on error
3132  */
3133 static int osd_object_ea_create(const struct lu_env *env, struct dt_object *dt,
3134                                 struct lu_attr *attr,
3135                                 struct dt_allocation_hint *hint,
3136                                 struct dt_object_format *dof,
3137                                 struct thandle *th)
3138 {
3139         const struct lu_fid     *fid    = lu_object_fid(&dt->do_lu);
3140         struct osd_object       *obj    = osd_dt_obj(dt);
3141         struct osd_thread_info  *info   = osd_oti_get(env);
3142         int                      result, on_ost = 0;
3143
3144         ENTRY;
3145
3146         if (dt_object_exists(dt))
3147                 RETURN(-EEXIST);
3148
3149         LASSERT(osd_invariant(obj));
3150         LASSERT(!dt_object_remote(dt));
3151         LASSERT(osd_write_locked(env, obj));
3152         LASSERT(th != NULL);
3153
3154         if (unlikely(fid_is_acct(fid)))
3155                 /* Quota files can't be created from the kernel any more,
3156                  * 'tune2fs -O quota' will take care of creating them */
3157                 RETURN(-EPERM);
3158
3159         result = __osd_object_create(info, obj, attr, hint, dof, th);
3160         if (result == 0) {
3161                 if (fid_is_idif(fid) &&
3162                     !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
3163                         struct lu_fid *tfid = &info->oti_fid;
3164                         struct ost_id *oi   = &info->oti_ostid;
3165
3166                         fid_to_ostid(fid, oi);
3167                         ostid_to_fid(tfid, oi, 0);
3168                         on_ost = 1;
3169                         result = osd_ea_fid_set(info, obj->oo_inode, tfid,
3170                                                 LMAC_FID_ON_OST, 0);
3171                 } else {
3172                         on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
3173                                                fid, OI_CHECK_FLD);
3174                         result = osd_ea_fid_set(info, obj->oo_inode, fid,
3175                                                 on_ost ? LMAC_FID_ON_OST : 0,
3176                                                 0);
3177                 }
3178                 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3179                         obj->oo_dt.do_body_ops = &osd_body_ops;
3180         }
3181
3182         if (result == 0)
3183                 result = __osd_oi_insert(env, obj, fid, th);
3184
3185         /* a small optimization - dt_insert() isn't usually applied
3186          * to OST objects, so we don't need to cache OI mapping for
3187          * OST objects */
3188         if (result == 0 && on_ost == 0) {
3189                 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3190                 result = osd_idc_find_and_init(env, osd, obj);
3191                 LASSERT(result == 0);
3192         }
3193
3194         LASSERT(ergo(result == 0,
3195                      dt_object_exists(dt) && !dt_object_remote(dt)));
3196         LINVRNT(osd_invariant(obj));
3197         RETURN(result);
3198 }
3199
3200 static int osd_declare_object_ref_add(const struct lu_env *env,
3201                                       struct dt_object *dt,
3202                                       struct thandle *handle)
3203 {
3204         struct osd_thandle       *oh;
3205
3206         /* it's possible that object doesn't exist yet */
3207         LASSERT(handle != NULL);
3208
3209         oh = container_of0(handle, struct osd_thandle, ot_super);
3210         LASSERT(oh->ot_handle == NULL);
3211
3212         osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
3213                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3214
3215         return 0;
3216 }
3217
3218 /*
3219  * Concurrency: @dt is write locked.
3220  */
3221 static int osd_object_ref_add(const struct lu_env *env,
3222                               struct dt_object *dt, struct thandle *th)
3223 {
3224         struct osd_object  *obj = osd_dt_obj(dt);
3225         struct inode       *inode = obj->oo_inode;
3226         struct osd_thandle *oh;
3227         int                 rc = 0;
3228
3229         if (!dt_object_exists(dt) || obj->oo_destroyed)
3230                 return -ENOENT;
3231
3232         LINVRNT(osd_invariant(obj));
3233         LASSERT(!dt_object_remote(dt));
3234         LASSERT(osd_write_locked(env, obj));
3235         LASSERT(th != NULL);
3236
3237         oh = container_of0(th, struct osd_thandle, ot_super);
3238         LASSERT(oh->ot_handle != NULL);
3239
3240         osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
3241
3242         CDEBUG(D_INODE, DFID" increase nlink %d\n",
3243                PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3244         /*
3245          * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
3246          * (65000) subdirectories by storing "1" in i_nlink if the link count
3247          * would otherwise overflow. Directory tranversal tools understand
3248          * that (st_nlink == 1) indicates that the filesystem dose not track
3249          * hard links count on the directory, and will not abort subdirectory
3250          * scanning early once (st_nlink - 2) subdirs have been found.
3251          *
3252          * This also has to properly handle the case of inodes with nlink == 0
3253          * in case they are being linked into the PENDING directory
3254          */
3255         spin_lock(&obj->oo_guard);
3256         if (unlikely(inode->i_nlink == 0))
3257                 /* inc_nlink from 0 may cause WARN_ON */
3258                 set_nlink(inode, 1);
3259         else {
3260                 ldiskfs_inc_count(oh->ot_handle, inode);
3261                 if (!S_ISDIR(inode->i_mode))
3262                         LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
3263         }
3264         spin_unlock(&obj->oo_guard);
3265
3266         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3267         LINVRNT(osd_invariant(obj));
3268
3269         osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
3270
3271         return rc;
3272 }
3273
3274 static int osd_declare_object_ref_del(const struct lu_env *env,
3275                                       struct dt_object *dt,
3276                                       struct thandle *handle)
3277 {
3278         struct osd_thandle *oh;
3279
3280         if (!dt_object_exists(dt))
3281                 return -ENOENT;
3282
3283         LASSERT(!dt_object_remote(dt));
3284         LASSERT(handle != NULL);
3285
3286         oh = container_of0(handle, struct osd_thandle, ot_super);
3287         LASSERT(oh->ot_handle == NULL);
3288
3289         osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
3290                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3291
3292         return 0;
3293 }
3294
3295 /*
3296  * Concurrency: @dt is write locked.
3297  */
3298 static int osd_object_ref_del(const struct lu_env *env, struct dt_object *dt,
3299                               struct thandle *th)
3300 {
3301         struct osd_object       *obj = osd_dt_obj(dt);
3302         struct inode            *inode = obj->oo_inode;
3303         struct osd_device       *osd = osd_dev(dt->do_lu.lo_dev);
3304         struct osd_thandle      *oh;
3305
3306         if (!dt_object_exists(dt))
3307                 return -ENOENT;
3308
3309         LINVRNT(osd_invariant(obj));
3310         LASSERT(!dt_object_remote(dt));
3311         LASSERT(osd_write_locked(env, obj));
3312         LASSERT(th != NULL);
3313
3314         oh = container_of0(th, struct osd_thandle, ot_super);
3315         LASSERT(oh->ot_handle != NULL);
3316
3317         osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
3318
3319         spin_lock(&obj->oo_guard);
3320         /* That can be result of upgrade from old Lustre version and
3321          * applied only to local files.  Just skip this ref_del call.
3322          * ext4_unlink() only treats this as a warning, don't LASSERT here.*/
3323         if (inode->i_nlink == 0) {
3324                 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
3325                              D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
3326                              ", maybe an upgraded file? (LU-3915)\n",
3327                              osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
3328                 spin_unlock(&obj->oo_guard);
3329                 return 0;
3330         }
3331
3332         CDEBUG(D_INODE, DFID" decrease nlink %d\n",
3333                PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3334
3335         ldiskfs_dec_count(oh->ot_handle, inode);
3336         spin_unlock(&obj->oo_guard);
3337
3338         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3339         LINVRNT(osd_invariant(obj));
3340
3341         osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
3342
3343         return 0;
3344 }
3345
3346 /*
3347  * Get the 64-bit version for an inode.
3348  */
3349 static int osd_object_version_get(const struct lu_env *env,
3350                                   struct dt_object *dt, dt_obj_version_t *ver)
3351 {
3352         struct inode *inode = osd_dt_obj(dt)->oo_inode;
3353
3354         CDEBUG(D_INODE, "Get version "LPX64" for inode %lu\n",
3355                LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3356         *ver = LDISKFS_I(inode)->i_fs_version;
3357         return 0;
3358 }
3359
3360 /*
3361  * Concurrency: @dt is read locked.
3362  */
3363 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
3364                          struct lu_buf *buf, const char *name)
3365 {
3366         struct osd_object      *obj    = osd_dt_obj(dt);
3367         struct inode           *inode  = obj->oo_inode;
3368         struct osd_thread_info *info   = osd_oti_get(env);
3369         struct dentry          *dentry = &info->oti_obj_dentry;
3370
3371         /* version get is not real XATTR but uses xattr API */
3372         if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3373                 /* for version we are just using xattr API but change inode
3374                  * field instead */
3375                 if (buf->lb_len == 0)
3376                         return sizeof(dt_obj_version_t);
3377
3378                 if (buf->lb_len < sizeof(dt_obj_version_t))
3379                         return -ERANGE;
3380
3381                 osd_object_version_get(env, dt, buf->lb_buf);
3382
3383                 return sizeof(dt_obj_version_t);
3384         }
3385
3386         if (!dt_object_exists(dt))
3387                 return -ENOENT;
3388
3389         LASSERT(!dt_object_remote(dt));
3390         LASSERT(inode->i_op != NULL);
3391         LASSERT(inode->i_op->getxattr != NULL);
3392
3393         return __osd_xattr_get(inode, dentry, name, buf->lb_buf, buf->lb_len);
3394 }
3395
3396
3397 static int osd_declare_xattr_set(const struct lu_env *env,
3398                                  struct dt_object *dt,
3399                                  const struct lu_buf *buf, const char *name,
3400                                  int fl, struct thandle *handle)
3401 {
3402         struct osd_thandle *oh;
3403         int credits;
3404         struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3405
3406         LASSERT(handle != NULL);
3407
3408         oh = container_of0(handle, struct osd_thandle, ot_super);
3409         LASSERT(oh->ot_handle == NULL);
3410
3411         if (strcmp(name, XATTR_NAME_LMA) == 0) {
3412                 /* For non-upgrading case, the LMA is set first and
3413                  * usually fit inode. But for upgrade case, the LMA
3414                  * may be in another separated EA block. */
3415                 if (!dt_object_exists(dt))
3416                         credits = 0;
3417                 else if (fl == LU_XATTR_REPLACE)
3418                         credits = 1;
3419                 else
3420                         goto upgrade;
3421         } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3422                 credits = 1;
3423         } else {
3424 upgrade:
3425                 credits = osd_dto_credits_noquota[DTO_XATTR_SET];
3426
3427                 if (buf != NULL) {
3428                         ssize_t buflen;
3429
3430                         if (buf->lb_buf == NULL && dt_object_exists(dt)) {
3431                                 /* learn xattr size from osd_xattr_get if
3432                                    attribute has not been read yet */
3433                                 buflen = __osd_xattr_get(
3434                                     osd_dt_obj(dt)->oo_inode,
3435                                     &osd_oti_get(env)->oti_obj_dentry,
3436                                     name, NULL, 0);
3437                                 if (buflen < 0)
3438                                         buflen = 0;
3439                         } else {
3440                                 buflen = buf->lb_len;
3441                         }
3442
3443                         if (buflen > sb->s_blocksize) {
3444                                 credits += osd_calc_bkmap_credits(
3445                                     sb, NULL, 0, -1,
3446                                     (buflen + sb->s_blocksize - 1) >>
3447                                     sb->s_blocksize_bits);
3448                         }
3449                 }
3450                 /*
3451                  * xattr set may involve inode quota change, reserve credits for
3452                  * dquot_initialize()
3453                  */
3454                 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
3455         }
3456
3457         osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
3458
3459         return 0;
3460 }
3461
3462 /*
3463  * Set the 64-bit version for object
3464  */
3465 static void osd_object_version_set(const struct lu_env *env,
3466                                    struct dt_object *dt,
3467                                    dt_obj_version_t *new_version)
3468 {
3469         struct inode *inode = osd_dt_obj(dt)->oo_inode;
3470
3471         CDEBUG(D_INODE, "Set version "LPX64" (old "LPX64") for inode %lu\n",
3472                *new_version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3473
3474         LDISKFS_I(inode)->i_fs_version = *new_version;
3475         /** Version is set after all inode operations are finished,
3476          *  so we should mark it dirty here */
3477         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3478 }
3479
3480 /*
3481  * Concurrency: @dt is write locked.
3482  */
3483 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
3484                          const struct lu_buf *buf, const char *name, int fl,
3485                          struct thandle *handle)
3486 {
3487         struct osd_object      *obj      = osd_dt_obj(dt);
3488         struct inode           *inode    = obj->oo_inode;
3489         struct osd_thread_info *info     = osd_oti_get(env);
3490         int                     fs_flags = 0;
3491         int                     rc;
3492         ENTRY;
3493
3494         LASSERT(handle != NULL);
3495
3496         /* version set is not real XATTR */
3497         if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3498                 /* for version we are just using xattr API but change inode
3499                  * field instead */
3500                 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
3501                 osd_object_version_set(env, dt, buf->lb_buf);
3502                 return sizeof(dt_obj_version_t);
3503         }
3504
3505         CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
3506                PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
3507
3508         osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3509         if (fl & LU_XATTR_REPLACE)
3510                 fs_flags |= XATTR_REPLACE;
3511
3512         if (fl & LU_XATTR_CREATE)
3513                 fs_flags |= XATTR_CREATE;
3514
3515         if (strcmp(name, XATTR_NAME_LMV) == 0) {
3516                 struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
3517                 int                      rc;
3518
3519                 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, lma);
3520                 if (rc != 0)
3521                         RETURN(rc);
3522
3523                 lma->lma_incompat |= LMAI_STRIPED;
3524                 lustre_lma_swab(lma);
3525                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3526                                      sizeof(*lma), XATTR_REPLACE);
3527                 if (rc != 0)
3528                         RETURN(rc);
3529         }
3530
3531         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LINKEA_OVERFLOW) &&
3532             strcmp(name, XATTR_NAME_LINK) == 0)
3533                 return -ENOSPC;
3534
3535         rc = __osd_xattr_set(info, inode, name, buf->lb_buf, buf->lb_len,
3536                                fs_flags);
3537         osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3538
3539         return rc;
3540 }
3541
3542 /*
3543  * Concurrency: @dt is read locked.
3544  */
3545 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
3546                           const struct lu_buf *buf)
3547 {
3548         struct osd_object      *obj    = osd_dt_obj(dt);
3549         struct inode           *inode  = obj->oo_inode;
3550         struct osd_thread_info *info   = osd_oti_get(env);
3551         struct dentry          *dentry = &info->oti_obj_dentry;
3552
3553         if (!dt_object_exists(dt))
3554                 return -ENOENT;
3555
3556         LASSERT(!dt_object_remote(dt));
3557         LASSERT(inode->i_op != NULL);
3558         LASSERT(inode->i_op->listxattr != NULL);
3559
3560         dentry->d_inode = inode;
3561         dentry->d_sb = inode->i_sb;
3562         return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
3563 }
3564
3565 static int osd_declare_xattr_del(const struct lu_env *env,
3566                                  struct dt_object *dt, const char *name,
3567                                  struct thandle *handle)
3568 {
3569         struct osd_thandle *oh;
3570         struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3571
3572         LASSERT(!dt_object_remote(dt));
3573         LASSERT(handle != NULL);
3574
3575         oh = container_of0(handle, struct osd_thandle, ot_super);
3576         LASSERT(oh->ot_handle == NULL);
3577
3578         osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
3579                              osd_dto_credits_noquota[DTO_XATTR_SET]);
3580         /*
3581          * xattr del may involve inode quota change, reserve credits for
3582          * dquot_initialize()
3583          */
3584         oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
3585
3586         return 0;
3587 }
3588
3589 /*
3590  * Concurrency: @dt is write locked.
3591  */
3592 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
3593                          const char *name, struct thandle *handle)
3594 {
3595         struct osd_object      *obj    = osd_dt_obj(dt);
3596         struct inode           *inode  = obj->oo_inode;
3597         struct osd_thread_info *info   = osd_oti_get(env);
3598         struct dentry          *dentry = &info->oti_obj_dentry;
3599         int                     rc;
3600
3601         if (!dt_object_exists(dt))
3602                 return -ENOENT;
3603
3604         LASSERT(!dt_object_remote(dt));
3605         LASSERT(inode->i_op != NULL);
3606         LASSERT(inode->i_op->removexattr != NULL);
3607         LASSERT(handle != NULL);
3608
3609         osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3610
3611         ll_vfs_dq_init(inode);
3612         dentry->d_inode = inode;
3613         dentry->d_sb = inode->i_sb;
3614         rc = inode->i_op->removexattr(dentry, name);
3615         osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3616         return rc;
3617 }
3618
3619 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
3620                            __u64 start, __u64 end)
3621 {
3622         struct osd_object       *obj    = osd_dt_obj(dt);
3623         struct inode            *inode  = obj->oo_inode;
3624         struct osd_thread_info  *info   = osd_oti_get(env);