Whamcloud - gitweb
67144360ae6feaf4dddd6f7ad8e0779f1ab5f8b3
[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.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/osd/osd_handler.c
33  *
34  * Top-level entry points into osd module
35  *
36  * Author: Nikita Danilov <nikita@clusterfs.com>
37  *         Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
38  */
39
40 #define DEBUG_SUBSYSTEM S_OSD
41
42 #include <linux/kallsyms.h>
43 #include <linux/module.h>
44 #include <linux/user_namespace.h>
45 #ifdef HAVE_UIDGID_HEADER
46 # include <linux/uidgid.h>
47 #endif
48
49 /* prerequisite for linux/xattr.h */
50 #include <linux/types.h>
51 /* prerequisite for linux/xattr.h */
52 #include <linux/fs.h>
53 /* XATTR_{REPLACE,CREATE} */
54 #include <linux/xattr.h>
55
56 #include <ldiskfs/ldiskfs.h>
57 #include <ldiskfs/xattr.h>
58 #include <ldiskfs/ldiskfs_extents.h>
59 #undef ENTRY
60 /*
61  * struct OBD_{ALLOC,FREE}*()
62  * OBD_FAIL_CHECK
63  */
64 #include <obd_support.h>
65 /* struct ptlrpc_thread */
66 #include <lustre_net.h>
67 #include <lustre_fid.h>
68 /* process_config */
69 #include <uapi/linux/lustre/lustre_param.h>
70
71 #include "osd_internal.h"
72 #include "osd_dynlocks.h"
73
74 /* llo_* api support */
75 #include <md_object.h>
76 #include <lustre_quota.h>
77
78 #include <lustre_linkea.h>
79
80 /* Maximum EA size is limited by LNET_MTU for remote objects */
81 #define OSD_MAX_EA_SIZE 1048364
82
83 int ldiskfs_pdo = 1;
84 module_param(ldiskfs_pdo, int, 0644);
85 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
86
87 int ldiskfs_track_declares_assert;
88 module_param(ldiskfs_track_declares_assert, int, 0644);
89 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
90
91 /* Slab to allocate dynlocks */
92 struct kmem_cache *dynlock_cachep;
93
94 /* Slab to allocate osd_it_ea */
95 struct kmem_cache *osd_itea_cachep;
96
97 static struct lu_kmem_descr ldiskfs_caches[] = {
98         {
99                 .ckd_cache = &dynlock_cachep,
100                 .ckd_name  = "dynlock_cache",
101                 .ckd_size  = sizeof(struct dynlock_handle)
102         },
103         {
104                 .ckd_cache = &osd_itea_cachep,
105                 .ckd_name  = "osd_itea_cache",
106                 .ckd_size  = sizeof(struct osd_it_ea)
107         },
108         {
109                 .ckd_cache = NULL
110         }
111 };
112
113 static const char dot[] = ".";
114 static const char dotdot[] = "..";
115
116 static const struct lu_object_operations      osd_lu_obj_ops;
117 static const struct dt_object_operations      osd_obj_ops;
118 static const struct dt_object_operations      osd_obj_otable_it_ops;
119 static const struct dt_index_operations       osd_index_iam_ops;
120 static const struct dt_index_operations       osd_index_ea_ops;
121
122 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
123                           const struct lu_fid *fid);
124 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
125                                                 struct osd_device *osd);
126
127 int osd_trans_declare_op2rb[] = {
128         [OSD_OT_ATTR_SET]       = OSD_OT_ATTR_SET,
129         [OSD_OT_PUNCH]          = OSD_OT_MAX,
130         [OSD_OT_XATTR_SET]      = OSD_OT_XATTR_SET,
131         [OSD_OT_CREATE]         = OSD_OT_DESTROY,
132         [OSD_OT_DESTROY]        = OSD_OT_CREATE,
133         [OSD_OT_REF_ADD]        = OSD_OT_REF_DEL,
134         [OSD_OT_REF_DEL]        = OSD_OT_REF_ADD,
135         [OSD_OT_WRITE]          = OSD_OT_WRITE,
136         [OSD_OT_INSERT]         = OSD_OT_DELETE,
137         [OSD_OT_DELETE]         = OSD_OT_INSERT,
138         [OSD_OT_QUOTA]          = OSD_OT_MAX,
139 };
140
141 static int osd_has_index(const struct osd_object *obj)
142 {
143         return obj->oo_dt.do_index_ops != NULL;
144 }
145
146 static int osd_object_invariant(const struct lu_object *l)
147 {
148         return osd_invariant(osd_obj(l));
149 }
150
151 /*
152  * Concurrency: doesn't matter
153  */
154 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
155 {
156         struct osd_thread_info *oti = osd_oti_get(env);
157
158         return oti->oti_w_locks > 0 && o->oo_owner == env;
159 }
160
161 /*
162  * Concurrency: doesn't access mutable data
163  */
164 static int osd_root_get(const struct lu_env *env,
165                         struct dt_device *dev, struct lu_fid *f)
166 {
167         lu_local_obj_fid(f, OSD_FS_ROOT_OID);
168         return 0;
169 }
170
171 /*
172  * the following set of functions are used to maintain per-thread
173  * cache of FID->ino mapping. this mechanism is needed to resolve
174  * FID to inode at dt_insert() which in turn stores ino in the
175  * directory entries to keep ldiskfs compatible with ext[34].
176  * due to locking-originated restrictions we can't lookup ino
177  * using LU cache (deadlock is possible). lookup using OI is quite
178  * expensive. so instead we maintain this cache and methods like
179  * dt_create() fill it. so in the majority of cases dt_insert() is
180  * able to find needed mapping in lockless manner.
181  */
182 static struct osd_idmap_cache *
183 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
184              const struct lu_fid *fid)
185 {
186         struct osd_thread_info *oti = osd_oti_get(env);
187         struct osd_idmap_cache *idc = oti->oti_ins_cache;
188         int i;
189
190         for (i = 0; i < oti->oti_ins_cache_used; i++) {
191                 if (!lu_fid_eq(&idc[i].oic_fid, fid))
192                         continue;
193                 if (idc[i].oic_dev != osd)
194                         continue;
195
196                 return idc + i;
197         }
198
199         return NULL;
200 }
201
202 static struct osd_idmap_cache *
203 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
204             const struct lu_fid *fid)
205 {
206         struct osd_thread_info *oti   = osd_oti_get(env);
207         struct osd_idmap_cache *idc;
208         int i;
209
210         if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
211                 i = oti->oti_ins_cache_size * 2;
212                 if (i == 0)
213                         i = OSD_INS_CACHE_SIZE;
214                 OBD_ALLOC(idc, sizeof(*idc) * i);
215                 if (idc == NULL)
216                         return ERR_PTR(-ENOMEM);
217                 if (oti->oti_ins_cache != NULL) {
218                         memcpy(idc, oti->oti_ins_cache,
219                                oti->oti_ins_cache_used * sizeof(*idc));
220                         OBD_FREE(oti->oti_ins_cache,
221                                  oti->oti_ins_cache_used * sizeof(*idc));
222                 }
223                 oti->oti_ins_cache = idc;
224                 oti->oti_ins_cache_size = i;
225         }
226
227         idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
228         idc->oic_fid = *fid;
229         idc->oic_dev = osd;
230         idc->oic_lid.oii_ino = 0;
231         idc->oic_lid.oii_gen = 0;
232         idc->oic_remote = 0;
233
234         return idc;
235 }
236
237 /*
238  * lookup mapping for the given fid in the cache, initialize a
239  * new one if not found. the initialization checks whether the
240  * object is local or remote. for local objects, OI is used to
241  * learn ino/generation. the function is used when the caller
242  * has no information about the object, e.g. at dt_insert().
243  */
244 static struct osd_idmap_cache *
245 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
246                      const struct lu_fid *fid)
247 {
248         struct osd_idmap_cache *idc;
249         int rc;
250
251         idc = osd_idc_find(env, osd, fid);
252         LASSERT(!IS_ERR(idc));
253         if (idc != NULL)
254                 return idc;
255
256         CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
257                osd->od_svname, PFID(fid));
258
259         /* new mapping is needed */
260         idc = osd_idc_add(env, osd, fid);
261         if (IS_ERR(idc)) {
262                 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
263                        osd->od_svname, PFID(fid), PTR_ERR(idc));
264                 return idc;
265         }
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         CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
314                osd->od_svname, PFID(fid));
315
316         /* new mapping is needed */
317         idc = osd_idc_add(env, osd, fid);
318         if (IS_ERR(idc)) {
319                 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
320                        osd->od_svname, PFID(fid), PTR_ERR(idc));
321                 return PTR_ERR(idc);
322         }
323
324         if (obj->oo_inode != NULL) {
325                 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
326                 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
327         }
328         return 0;
329 }
330
331 /*
332  * OSD object methods.
333  */
334
335 /*
336  * Concurrency: no concurrent access is possible that early in object
337  * life-cycle.
338  */
339 static struct lu_object *osd_object_alloc(const struct lu_env *env,
340                                           const struct lu_object_header *hdr,
341                                           struct lu_device *d)
342 {
343         struct osd_object *mo;
344
345         OBD_ALLOC_PTR(mo);
346         if (mo != NULL) {
347                 struct lu_object *l;
348                 struct lu_object_header *h;
349                 struct osd_device *o = osd_dev(d);
350
351                 l = &mo->oo_dt.do_lu;
352                 if (unlikely(o->od_in_init)) {
353                         OBD_ALLOC_PTR(h);
354                         if (!h) {
355                                 OBD_FREE_PTR(mo);
356                                 return NULL;
357                         }
358
359                         lu_object_header_init(h);
360                         lu_object_init(l, h, d);
361                         lu_object_add_top(h, l);
362                         mo->oo_header = h;
363                 } else {
364                         dt_object_init(&mo->oo_dt, NULL, d);
365                         mo->oo_header = NULL;
366                 }
367
368                 mo->oo_dt.do_ops = &osd_obj_ops;
369                 l->lo_ops = &osd_lu_obj_ops;
370                 init_rwsem(&mo->oo_sem);
371                 init_rwsem(&mo->oo_ext_idx_sem);
372                 spin_lock_init(&mo->oo_guard);
373                 INIT_LIST_HEAD(&mo->oo_xattr_list);
374                 return l;
375         }
376         return NULL;
377 }
378
379 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
380                 struct dentry *dentry, struct lustre_ost_attrs *loa)
381 {
382         int rc;
383
384         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
385                              (void *)loa, sizeof(*loa));
386         if (rc > 0) {
387                 struct lustre_mdt_attrs *lma = &loa->loa_lma;
388
389                 if (rc < sizeof(*lma))
390                         return -EINVAL;
391
392                 rc = 0;
393                 lustre_loa_swab(loa, true);
394                 /* Check LMA compatibility */
395                 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
396                         CWARN("%s: unsupported incompat LMA feature(s) %#x "
397                               "for fid = "DFID", ino = %lu\n",
398                               osd_ino2name(inode),
399                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
400                               PFID(&lma->lma_self_fid), inode->i_ino);
401                         rc = -EOPNOTSUPP;
402                 }
403         } else if (rc == 0) {
404                 rc = -ENODATA;
405         }
406
407         return rc;
408 }
409
410 /*
411  * retrieve object from backend ext fs.
412  **/
413 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
414                        struct osd_inode_id *id)
415 {
416         int rc;
417         struct inode *inode = NULL;
418
419         /*
420          * if we look for an inode withing a running
421          * transaction, then we risk to deadlock
422          * osd_dirent_check_repair() breaks this
423          */
424          /* LASSERT(current->journal_info == NULL); */
425
426         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
427         if (IS_ERR(inode)) {
428                 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
429                        id->oii_ino, PTR_ERR(inode));
430         } else if (id->oii_gen != OSD_OII_NOGEN &&
431                    inode->i_generation != id->oii_gen) {
432                 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
433                        "i_generation = %u\n",
434                        id->oii_ino, id->oii_gen, inode->i_generation);
435                 iput(inode);
436                 inode = ERR_PTR(-ESTALE);
437         } else if (inode->i_nlink == 0) {
438                 /*
439                  * due to parallel readdir and unlink,
440                  * we can have dead inode here.
441                  */
442                 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
443                 iput(inode);
444                 inode = ERR_PTR(-ESTALE);
445         } else if (is_bad_inode(inode)) {
446                 CWARN("%s: bad inode: ino = %u\n",
447                       osd_dev2name(dev), id->oii_ino);
448                 iput(inode);
449                 inode = ERR_PTR(-ENOENT);
450         } else if ((rc = osd_attach_jinode(inode))) {
451                 iput(inode);
452                 inode = ERR_PTR(rc);
453         } else {
454                 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
455                 if (id->oii_gen == OSD_OII_NOGEN)
456                         osd_id_gen(id, inode->i_ino, inode->i_generation);
457
458                 /*
459                  * Do not update file c/mtime in ldiskfs.
460                  * NB: we don't have any lock to protect this because we don't
461                  * have reference on osd_object now, but contention with
462                  * another lookup + attr_set can't happen in the tiny window
463                  * between if (...) and set S_NOCMTIME.
464                  */
465                 if (!(inode->i_flags & S_NOCMTIME))
466                         inode->i_flags |= S_NOCMTIME;
467         }
468         return inode;
469 }
470
471 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
472                           handle_t *handle, struct dentry *child,
473                           struct inode *inode, struct htree_lock *hlock)
474 {
475         int rc, rc2;
476
477         rc = __ldiskfs_add_entry(handle, child, inode, hlock);
478         if (rc == -ENOBUFS || rc == -ENOSPC) {
479                 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
480                 struct inode *parent = child->d_parent->d_inode;
481                 struct lu_fid *fid = NULL;
482
483                 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
484                 if (!rc2) {
485                         fid = &loa->loa_lma.lma_self_fid;
486                 } else if (rc2 == -ENODATA) {
487                         if (unlikely(parent == inode->i_sb->s_root->d_inode)) {
488                                 fid = &info->oti_fid3;
489                                 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
490                         } else if (!osd->od_is_ost && osd->od_index == 0) {
491                                 fid = &info->oti_fid3;
492                                 lu_igif_build(fid, parent->i_ino,
493                                               parent->i_generation);
494                         }
495                 }
496
497                 if (fid != NULL)
498                         CWARN("%s: directory (inode: %lu, FID: "DFID") %s "
499                               "maximum entry limit\n",
500                               osd_name(osd), parent->i_ino, PFID(fid),
501                               rc == -ENOSPC ? "has reached" : "is approaching");
502                 else
503                         CWARN("%s: directory (inode: %lu, FID: unknown) %s "
504                               "maximum entry limit\n",
505                               osd_name(osd), parent->i_ino,
506                               rc == -ENOSPC ? "has reached" : "is approaching");
507
508                 /* ignore such error now */
509                 if (rc == -ENOBUFS)
510                         rc = 0;
511         }
512
513         return rc;
514 }
515
516
517 struct inode *
518 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
519              struct osd_inode_id *id, struct lu_fid *fid)
520 {
521         struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
522         struct inode *inode;
523         int rc;
524
525         inode = osd_iget(info, dev, id);
526         if (IS_ERR(inode))
527                 return inode;
528
529         rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
530         if (!rc) {
531                 *fid = loa->loa_lma.lma_self_fid;
532         } else if (rc == -ENODATA) {
533                 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
534                         lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
535                 else
536                         lu_igif_build(fid, inode->i_ino, inode->i_generation);
537         } else {
538                 iput(inode);
539                 inode = ERR_PTR(rc);
540         }
541         return inode;
542 }
543
544 static struct inode *osd_iget_check(struct osd_thread_info *info,
545                                     struct osd_device *dev,
546                                     const struct lu_fid *fid,
547                                     struct osd_inode_id *id,
548                                     bool trusted)
549 {
550         struct inode *inode;
551         int rc = 0;
552
553         ENTRY;
554
555         /*
556          * The cached OI mapping is trustable. If we cannot locate the inode
557          * via the cached OI mapping, then return the failure to the caller
558          * directly without further OI checking.
559          */
560
561 again:
562         inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
563         if (IS_ERR(inode)) {
564                 rc = PTR_ERR(inode);
565                 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
566                         goto check_oi;
567
568                 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
569                        PFID(fid), id->oii_ino, rc);
570                 GOTO(put, rc);
571         }
572
573         if (is_bad_inode(inode)) {
574                 rc = -ENOENT;
575                 if (!trusted)
576                         goto check_oi;
577
578                 CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
579                        PFID(fid), id->oii_ino);
580                 GOTO(put, rc);
581         }
582
583         if (id->oii_gen != OSD_OII_NOGEN &&
584             inode->i_generation != id->oii_gen) {
585                 rc = -ESTALE;
586                 if (!trusted)
587                         goto check_oi;
588
589                 CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
590                        "oii_gen = %u, i_generation = %u\n", PFID(fid),
591                        id->oii_ino, id->oii_gen, inode->i_generation);
592                 GOTO(put, rc);
593         }
594
595         if (inode->i_nlink == 0) {
596                 rc = -ENOENT;
597                 if (!trusted)
598                         goto check_oi;
599
600                 CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
601                        PFID(fid), id->oii_ino);
602                 GOTO(put, rc);
603         }
604
605         ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
606
607 check_oi:
608         if (rc != 0) {
609                 __u32 saved_ino = id->oii_ino;
610                 __u32 saved_gen = id->oii_gen;
611
612                 LASSERT(!trusted);
613                 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
614
615                 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
616                 /*
617                  * XXX: There are four possible cases:
618                  *      1. rc = 0.
619                  *         Backup/restore caused the OI invalid.
620                  *      2. rc = 0.
621                  *         Someone unlinked the object but NOT removed
622                  *         the OI mapping, such as mount target device
623                  *         as ldiskfs, and modify something directly.
624                  *      3. rc = -ENOENT.
625                  *         Someone just removed the object between the
626                  *         former oi_lookup and the iget. It is normal.
627                  *      4. Other failure cases.
628                  *
629                  *      Generally, when the device is mounted, it will
630                  *      auto check whether the system is restored from
631                  *      file-level backup or not. We trust such detect
632                  *      to distinguish the 1st case from the 2nd case:
633                  *      if the OI files are consistent but may contain
634                  *      stale OI mappings because of case 2, if iget()
635                  *      returns -ENOENT or -ESTALE, then it should be
636                  *      the case 2.
637                  */
638                 if (rc != 0)
639                         /*
640                          * If the OI mapping was in OI file before the
641                          * osd_iget_check(), but now, it is disappear,
642                          * then it must be removed by race. That is a
643                          * normal race case.
644                          */
645                         GOTO(put, rc);
646
647                 /*
648                  * It is the OI scrub updated the OI mapping by race.
649                  * The new OI mapping must be valid.
650                  */
651                 if (saved_ino != id->oii_ino ||
652                     (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
653                         if (!IS_ERR(inode))
654                                 iput(inode);
655
656                         trusted = true;
657                         goto again;
658                 }
659
660                 if (IS_ERR(inode)) {
661                         if (dev->od_scrub.os_scrub.os_file.sf_flags &
662                             SF_INCONSISTENT)
663                                 /*
664                                  * It still can be the case 2, but we cannot
665                                  * distinguish it from the case 1. So return
666                                  * -EREMCHG to block current operation until
667                                  *  OI scrub rebuilt the OI mappings.
668                                  */
669                                 rc = -EREMCHG;
670                         else
671                                 rc = -ENOENT;
672
673                         GOTO(put, rc);
674                 }
675
676                 if (inode->i_generation == id->oii_gen)
677                         rc = -ENOENT;
678                 else
679                         rc = -EREMCHG;
680         } else {
681                 if (id->oii_gen == OSD_OII_NOGEN)
682                         osd_id_gen(id, inode->i_ino, inode->i_generation);
683
684                 /*
685                  * Do not update file c/mtime in ldiskfs.
686                  * NB: we don't have any lock to protect this because we don't
687                  * have reference on osd_object now, but contention with
688                  * another lookup + attr_set can't happen in the tiny window
689                  * between if (...) and set S_NOCMTIME.
690                  */
691                 if (!(inode->i_flags & S_NOCMTIME))
692                         inode->i_flags |= S_NOCMTIME;
693         }
694
695         GOTO(put, rc);
696
697 put:
698         if (rc != 0) {
699                 if (!IS_ERR(inode))
700                         iput(inode);
701
702                 inode = ERR_PTR(rc);
703         }
704
705         return inode;
706 }
707
708 /**
709  * \retval +v: new filter_fid does not contain self-fid
710  * \retval 0:  filter_fid_18_23, contains self-fid
711  * \retval -v: other failure cases
712  */
713 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
714                  struct dentry *dentry, struct lu_fid *fid)
715 {
716         struct filter_fid *ff = &info->oti_ff;
717         struct ost_id *ostid = &info->oti_ostid;
718         int rc;
719
720         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
721         if (rc == sizeof(struct filter_fid_18_23)) {
722                 struct filter_fid_18_23 *ff_old = (void *)ff;
723
724                 ostid_set_seq(ostid, le64_to_cpu(ff_old->ff_seq));
725                 rc = ostid_set_id(ostid, le64_to_cpu(ff_old->ff_objid));
726                 /*
727                  * XXX: use 0 as the index for compatibility, the caller will
728                  * handle index related issues when necessary.
729                  */
730                 if (!rc)
731                         ostid_to_fid(fid, ostid, 0);
732         } else if (rc >= (int)sizeof(struct filter_fid_24_29)) {
733                 rc = 1;
734         } else if (rc >= 0) {
735                 rc = -EINVAL;
736         }
737
738         return rc;
739 }
740
741 static int osd_lma_self_repair(struct osd_thread_info *info,
742                                struct osd_device *osd, struct inode *inode,
743                                const struct lu_fid *fid, __u32 compat)
744 {
745         handle_t *jh;
746         int rc;
747
748         LASSERT(current->journal_info == NULL);
749
750         jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
751                                   osd_dto_credits_noquota[DTO_XATTR_SET]);
752         if (IS_ERR(jh)) {
753                 rc = PTR_ERR(jh);
754                 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
755                       osd_name(osd), rc);
756                 return rc;
757         }
758
759         rc = osd_ea_fid_set(info, inode, fid, compat, 0);
760         if (rc != 0)
761                 CWARN("%s: cannot self repair the LMA: rc = %d\n",
762                       osd_name(osd), rc);
763         ldiskfs_journal_stop(jh);
764         return rc;
765 }
766
767 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
768 {
769         struct osd_thread_info *info = osd_oti_get(env);
770         struct osd_device *osd = osd_obj2dev(obj);
771         struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
772         struct lustre_mdt_attrs *lma = &loa->loa_lma;
773         struct inode *inode = obj->oo_inode;
774         struct dentry *dentry = &info->oti_obj_dentry;
775         struct lu_fid *fid = NULL;
776         const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
777         int rc;
778
779         ENTRY;
780
781         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
782                              (void *)loa, sizeof(*loa));
783         if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
784                 fid = &lma->lma_self_fid;
785                 rc = osd_get_idif(info, inode, dentry, fid);
786                 if (rc > 0 || (rc == -ENODATA && osd->od_index_in_idif)) {
787                         /*
788                          * For the given OST-object, if it has neither LMA nor
789                          * FID in XATTR_NAME_FID, then the given FID (which is
790                          * contained in the @obj, from client RPC for locating
791                          * the OST-object) is trusted. We use it to generate
792                          * the LMA.
793                          */
794                         osd_lma_self_repair(info, osd, inode, rfid,
795                                             LMAC_FID_ON_OST);
796                         RETURN(0);
797                 }
798         }
799
800         if (rc < 0)
801                 RETURN(rc);
802
803         if (rc > 0) {
804                 rc = 0;
805                 lustre_lma_swab(lma);
806                 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
807                              (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
808                               S_ISREG(inode->i_mode)))) {
809                         CWARN("%s: unsupported incompat LMA feature(s) %#x for "
810                               "fid = "DFID", ino = %lu\n", osd_name(osd),
811                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
812                               PFID(rfid), inode->i_ino);
813                         rc = -EOPNOTSUPP;
814                 } else {
815                         fid = &lma->lma_self_fid;
816                         if (lma->lma_compat & LMAC_STRIPE_INFO &&
817                             osd->od_is_ost)
818                                 obj->oo_pfid_in_lma = 1;
819                         if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
820                             !osd->od_is_ost)
821                                 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
822                 }
823         }
824
825         if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
826                 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
827                         struct ost_id   *oi   = &info->oti_ostid;
828                         struct lu_fid   *fid1 = &info->oti_fid3;
829                         __u32            idx  = fid_idif_ost_idx(rfid);
830
831                         /*
832                          * For old IDIF, the OST index is not part of the IDIF,
833                          * Means that different OSTs may have the same IDIFs.
834                          * Under such case, we need to make some compatible
835                          * check to make sure to trigger OI scrub properly.
836                          */
837                         if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
838                                 /* Given @rfid is new, LMA is old. */
839                                 fid_to_ostid(fid, oi);
840                                 ostid_to_fid(fid1, oi, idx);
841                                 if (lu_fid_eq(fid1, rfid)) {
842                                         if (osd->od_index_in_idif)
843                                                 osd_lma_self_repair(info, osd,
844                                                         inode, rfid,
845                                                         LMAC_FID_ON_OST);
846                                         RETURN(0);
847                                 }
848                         }
849                 }
850
851                 rc = -EREMCHG;
852         }
853
854         RETURN(rc);
855 }
856
857 struct osd_check_lmv_buf {
858 #ifdef HAVE_DIR_CONTEXT
859         /* please keep it as first member */
860         struct dir_context ctx;
861 #endif
862         struct osd_thread_info *oclb_info;
863         struct osd_device *oclb_dev;
864         struct osd_idmap_cache *oclb_oic;
865         int oclb_items;
866         bool oclb_found;
867 };
868
869 /**
870  * It is called internally by ->readdir() to filter out the
871  * local slave object's FID of the striped directory.
872  *
873  * \retval      1 found the local slave's FID
874  * \retval      0 continue to check next item
875  * \retval      -ve for failure
876  */
877 #ifdef HAVE_FILLDIR_USE_CTX
878 static int osd_stripe_dir_filldir(struct dir_context *buf,
879 #else
880 static int osd_stripe_dir_filldir(void *buf,
881 #endif
882                                   const char *name, int namelen,
883                                   loff_t offset, __u64 ino, unsigned int d_type)
884 {
885         struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
886         struct osd_thread_info *oti = oclb->oclb_info;
887         struct lu_fid *fid = &oti->oti_fid3;
888         struct osd_inode_id *id = &oti->oti_id3;
889         struct osd_device *dev = oclb->oclb_dev;
890         struct osd_idmap_cache *oic = oclb->oclb_oic;
891         struct inode *inode;
892
893         oclb->oclb_items++;
894
895         if (name[0] == '.')
896                 return 0;
897
898         fid_zero(fid);
899         sscanf(name + 1, SFID, RFID(fid));
900         if (!fid_is_sane(fid))
901                 return 0;
902
903         if (osd_remote_fid(oti->oti_env, dev, fid))
904                 return 0;
905
906         osd_id_gen(id, ino, OSD_OII_NOGEN);
907         inode = osd_iget(oti, dev, id);
908         if (IS_ERR(inode))
909                 return PTR_ERR(inode);
910
911         iput(inode);
912         osd_add_oi_cache(oti, dev, id, fid);
913         oic->oic_fid = *fid;
914         oic->oic_lid = *id;
915         oic->oic_dev = dev;
916         osd_oii_insert(dev, oic, true);
917         oclb->oclb_found = true;
918
919         return 1;
920 }
921
922 /*
923  * When lookup item under striped directory, we need to locate the master
924  * MDT-object of the striped directory firstly, then the client will send
925  * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
926  * and the item's name. If the system is restored from MDT file level backup,
927  * then before the OI scrub completely built the OI files, the OI mappings of
928  * the master MDT-object and slave MDT-object may be invalid. Usually, it is
929  * not a problem for the master MDT-object. Because when locate the master
930  * MDT-object, we will do name based lookup (for the striped directory itself)
931  * firstly, during such process we can setup the correct OI mapping for the
932  * master MDT-object. But it will be trouble for the slave MDT-object. Because
933  * the client will not trigger name based lookup on the MDT to locate the slave
934  * MDT-object before locating item under the striped directory, then when
935  * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
936  * is invalid and does not know what the right OI mapping is, then the MDT has
937  * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
938  * the OI file, related OI mapping is unknown yet, please try again later. And
939  * then client will re-try the RPC again and again until related OI mapping has
940  * been updated. That is quite inefficient.
941  *
942  * To resolve above trouble, we will handle it as the following two cases:
943  *
944  * 1) The slave MDT-object and the master MDT-object are on different MDTs.
945  *    It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
946  *    is linked under /REMOTE_PARENT_DIR with the name of its FID string.
947  *    We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
948  *    directly. Please check osd_fid_lookup().
949  *
950  * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
951  *    Under such case, during lookup the master MDT-object, we will lookup the
952  *    slave MDT-object via readdir against the master MDT-object, because the
953  *    slave MDT-objects information are stored as sub-directories with the name
954  *    "${FID}:${index}". Then when find the local slave MDT-object, its OI
955  *    mapping will be recorded. Then subsequent osd_fid_lookup() will know
956  *    the correct OI mapping for the slave MDT-object.
957  */
958 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
959                          struct inode *inode, struct osd_idmap_cache *oic)
960 {
961         struct lu_buf *buf = &oti->oti_big_buf;
962         struct dentry *dentry = &oti->oti_obj_dentry;
963         struct file *filp = &oti->oti_file;
964         const struct file_operations *fops;
965         struct lmv_mds_md_v1 *lmv1;
966         struct osd_check_lmv_buf oclb = {
967 #ifdef HAVE_DIR_CONTEXT
968                 .ctx.actor = osd_stripe_dir_filldir,
969 #endif
970                 .oclb_info = oti,
971                 .oclb_dev = dev,
972                 .oclb_oic = oic,
973                 .oclb_found = false,
974         };
975         int rc = 0;
976
977         ENTRY;
978
979 again:
980         rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
981                              buf->lb_len);
982         if (rc == -ERANGE) {
983                 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
984                 if (rc > 0) {
985                         lu_buf_realloc(buf, rc);
986                         if (buf->lb_buf == NULL)
987                                 GOTO(out, rc = -ENOMEM);
988
989                         goto again;
990                 }
991         }
992
993         if (unlikely(rc == 0 || rc == -ENODATA))
994                 GOTO(out, rc = 0);
995
996         if (rc < 0)
997                 GOTO(out, rc);
998
999         if (unlikely(buf->lb_buf == NULL)) {
1000                 lu_buf_realloc(buf, rc);
1001                 if (buf->lb_buf == NULL)
1002                         GOTO(out, rc = -ENOMEM);
1003
1004                 goto again;
1005         }
1006
1007         lmv1 = buf->lb_buf;
1008         if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
1009                 GOTO(out, rc = 0);
1010
1011         fops = inode->i_fop;
1012         dentry->d_inode = inode;
1013         dentry->d_sb = inode->i_sb;
1014         filp->f_pos = 0;
1015         filp->f_path.dentry = dentry;
1016         filp->f_mode = FMODE_64BITHASH;
1017         filp->f_mapping = inode->i_mapping;
1018         filp->f_op = fops;
1019         filp->private_data = NULL;
1020         set_file_inode(filp, inode);
1021
1022         do {
1023                 oclb.oclb_items = 0;
1024 #ifdef HAVE_DIR_CONTEXT
1025                 oclb.ctx.pos = filp->f_pos;
1026 #ifdef HAVE_ITERATE_SHARED
1027                 rc = fops->iterate_shared(filp, &oclb.ctx);
1028 #else
1029                 rc = fops->iterate(filp, &oclb.ctx);
1030 #endif
1031                 filp->f_pos = oclb.ctx.pos;
1032 #else
1033                 rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
1034 #endif
1035         } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
1036                  filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
1037         fops->release(inode, filp);
1038
1039 out:
1040         if (rc < 0)
1041                 CDEBUG(D_LFSCK, "%s: fail to check LMV EA, inode = %lu/%u,"
1042                        DFID": rc = %d\n", osd_ino2name(inode),
1043                        inode->i_ino, inode->i_generation,
1044                        PFID(&oic->oic_fid), rc);
1045         else
1046                 rc = 0;
1047
1048         RETURN(rc);
1049 }
1050
1051 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
1052                           const struct lu_fid *fid,
1053                           const struct lu_object_conf *conf)
1054 {
1055         struct osd_thread_info *info;
1056         struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
1057         struct osd_device *dev;
1058         struct osd_idmap_cache *oic;
1059         struct osd_inode_id *id;
1060         struct inode *inode = NULL;
1061         struct lustre_scrub *scrub;
1062         struct scrub_file *sf;
1063         __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1064         __u32 saved_ino;
1065         __u32 saved_gen;
1066         int result = 0;
1067         int rc1 = 0;
1068         bool remote = false;
1069         bool trusted = true;
1070         bool updated = false;
1071         bool checked = false;
1072
1073         ENTRY;
1074
1075         LINVRNT(osd_invariant(obj));
1076         LASSERT(obj->oo_inode == NULL);
1077         LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
1078
1079         dev = osd_dev(ldev);
1080         scrub = &dev->od_scrub.os_scrub;
1081         sf = &scrub->os_file;
1082         info = osd_oti_get(env);
1083         LASSERT(info);
1084         oic = &info->oti_cache;
1085
1086         if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1087                 RETURN(-ENOENT);
1088
1089         /*
1090          * For the object is created as locking anchor, or for the object to
1091          * be created on disk. No need to osd_oi_lookup() at here because FID
1092          * shouldn't never be re-used, if it's really a duplicate FID from
1093          * unexpected reason, we should be able to detect it later by calling
1094          * do_create->osd_oi_insert().
1095          */
1096         if (conf && conf->loc_flags & LOC_F_NEW)
1097                 GOTO(out, result = 0);
1098
1099         /* Search order: 1. per-thread cache. */
1100         if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1101                 id = &oic->oic_lid;
1102                 goto iget;
1103         }
1104
1105         id = &info->oti_id;
1106         if (!list_empty(&scrub->os_inconsistent_items)) {
1107                 /* Search order: 2. OI scrub pending list. */
1108                 result = osd_oii_lookup(dev, fid, id);
1109                 if (!result)
1110                         goto iget;
1111         }
1112
1113         /*
1114          * The OI mapping in the OI file can be updated by the OI scrub
1115          * when we locate the inode via FID. So it may be not trustable.
1116          */
1117         trusted = false;
1118
1119         /* Search order: 3. OI files. */
1120         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1121         if (result == -ENOENT) {
1122                 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1123                     fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1124                     !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
1125                                       sf->sf_oi_bitmap))
1126                         GOTO(out, result = 0);
1127
1128                 goto trigger;
1129         }
1130
1131         if (result)
1132                 GOTO(out, result);
1133
1134 iget:
1135         obj->oo_inode = NULL;
1136         /* for later passes through checks, not true on first pass */
1137         if (!IS_ERR_OR_NULL(inode))
1138                 iput(inode);
1139
1140         inode = osd_iget_check(info, dev, fid, id, trusted);
1141         if (!IS_ERR(inode)) {
1142                 obj->oo_inode = inode;
1143                 result = 0;
1144                 if (remote)
1145                         goto trigger;
1146
1147                 goto check_lma;
1148         }
1149
1150         result = PTR_ERR(inode);
1151         if (result == -ENOENT || result == -ESTALE)
1152                 GOTO(out, result = 0);
1153
1154         if (result != -EREMCHG)
1155                 GOTO(out, result);
1156
1157 trigger:
1158         /*
1159          * We still have chance to get the valid inode: for the
1160          * object which is referenced by remote name entry, the
1161          * object on the local MDT will be linked under the dir
1162          * of "/REMOTE_PARENT_DIR" with its FID string as name.
1163          *
1164          * We do not know whether the object for the given FID
1165          * is referenced by some remote name entry or not, and
1166          * especially for DNE II, a multiple-linked object may
1167          * have many name entries reside on many MDTs.
1168          *
1169          * To simplify the operation, OSD will not distinguish
1170          * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1171          * only happened for the RPC from other MDT during the
1172          * OI scrub, or for the client side RPC with FID only,
1173          * such as FID to path, or from old connected client.
1174          */
1175         if (!remote) {
1176                 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1177                 if (!rc1) {
1178                         remote = true;
1179                         trusted = true;
1180                         flags |= SS_AUTO_PARTIAL;
1181                         flags &= ~SS_AUTO_FULL;
1182                         goto iget;
1183                 }
1184         }
1185
1186         if (thread_is_running(&scrub->os_thread)) {
1187                 if (scrub->os_partial_scan && !scrub->os_in_join)
1188                         goto join;
1189
1190                 if (IS_ERR_OR_NULL(inode) || result)
1191                         GOTO(out, result = -EINPROGRESS);
1192
1193                 LASSERT(remote);
1194                 LASSERT(obj->oo_inode == inode);
1195
1196                 osd_add_oi_cache(info, dev, id, fid);
1197                 osd_oii_insert(dev, oic, true);
1198                 goto found;
1199         }
1200
1201         if (dev->od_auto_scrub_interval == AS_NEVER) {
1202                 if (!remote)
1203                         GOTO(out, result = -EREMCHG);
1204
1205                 LASSERT(!result);
1206                 LASSERT(obj->oo_inode == inode);
1207
1208                 osd_add_oi_cache(info, dev, id, fid);
1209                 goto found;
1210         }
1211
1212 join:
1213         rc1 = osd_scrub_start(env, dev, flags);
1214         LCONSOLE_WARN("%s: trigger OI scrub by RPC for the " DFID" with flags "
1215                       "0x%x, rc = %d\n", osd_name(dev), PFID(fid), flags, rc1);
1216         if (rc1 && rc1 != -EALREADY)
1217                 GOTO(out, result = -EREMCHG);
1218
1219         if (IS_ERR_OR_NULL(inode) || result)
1220                 GOTO(out, result = -EINPROGRESS);
1221
1222         LASSERT(remote);
1223         LASSERT(obj->oo_inode == inode);
1224
1225         osd_add_oi_cache(info, dev, id, fid);
1226         osd_oii_insert(dev, oic, true);
1227         goto found;
1228
1229 check_lma:
1230         checked = true;
1231         if (unlikely(obj->oo_header))
1232                 goto found;
1233
1234         result = osd_check_lma(env, obj);
1235         if (!result)
1236                 goto found;
1237
1238         LASSERTF(id->oii_ino == inode->i_ino &&
1239                  id->oii_gen == inode->i_generation,
1240                  "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1241                  PFID(fid), id->oii_ino, id->oii_gen,
1242                  inode->i_ino, inode->i_generation);
1243
1244         saved_ino = inode->i_ino;
1245         saved_gen = inode->i_generation;
1246
1247         if (unlikely(result == -ENODATA)) {
1248                 /*
1249                  * If the OI scrub updated the OI mapping by race, it
1250                  * must be valid. Trust the inode that has no LMA EA.
1251                  */
1252                 if (updated)
1253                         goto found;
1254
1255                 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1256                 if (!result) {
1257                         /*
1258                          * The OI mapping is still there, the inode is still
1259                          * valid. It is just becaues the inode has no LMA EA.
1260                          */
1261                         if (saved_ino == id->oii_ino &&
1262                             saved_gen == id->oii_gen)
1263                                 goto found;
1264
1265                         /*
1266                          * It is the OI scrub updated the OI mapping by race.
1267                          * The new OI mapping must be valid.
1268                          */
1269                         trusted = true;
1270                         updated = true;
1271                         goto iget;
1272                 }
1273
1274                 /*
1275                  * "result == -ENOENT" means that the OI mappinghas been
1276                  * removed by race, so the inode belongs to other object.
1277                  *
1278                  * Others error can be returned  directly.
1279                  */
1280                 if (result == -ENOENT) {
1281                         LASSERT(trusted);
1282
1283                         obj->oo_inode = NULL;
1284                         result = 0;
1285                 }
1286         }
1287
1288         if (result != -EREMCHG)
1289                 GOTO(out, result);
1290
1291         LASSERT(!updated);
1292
1293         result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1294         /*
1295          * "result == -ENOENT" means the cached OI mapping has been removed
1296          * from the OI file by race, above inode belongs to other object.
1297          */
1298         if (result == -ENOENT) {
1299                 LASSERT(trusted);
1300
1301                 obj->oo_inode = NULL;
1302                 GOTO(out, result = 0);
1303         }
1304
1305         if (result)
1306                 GOTO(out, result);
1307
1308         if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1309                 result = -EREMCHG;
1310                 goto trigger;
1311         }
1312
1313         /*
1314          * It is the OI scrub updated the OI mapping by race.
1315          * The new OI mapping must be valid.
1316          */
1317         trusted = true;
1318         updated = true;
1319         goto iget;
1320
1321 found:
1322         if (!checked) {
1323                 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
1324                 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
1325
1326                 result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
1327                 if (!result) {
1328                         if (lma->lma_compat & LMAC_STRIPE_INFO &&
1329                             dev->od_is_ost)
1330                                 obj->oo_pfid_in_lma = 1;
1331                         if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
1332                             !dev->od_is_ost)
1333                                 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
1334                 } else if (result != -ENODATA) {
1335                         GOTO(out, result);
1336                 }
1337         }
1338
1339         obj->oo_compat_dot_created = 1;
1340         obj->oo_compat_dotdot_created = 1;
1341
1342         if (S_ISDIR(inode->i_mode) &&
1343             (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1344                 osd_check_lmv(info, dev, inode, oic);
1345
1346         result = osd_attach_jinode(inode);
1347         if (result)
1348                 GOTO(out, result);
1349
1350         if (!ldiskfs_pdo)
1351                 GOTO(out, result = 0);
1352
1353         LASSERT(!obj->oo_hl_head);
1354         obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1355
1356         GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1357
1358 out:
1359         if (result || !obj->oo_inode) {
1360                 if (!IS_ERR_OR_NULL(inode))
1361                         iput(inode);
1362
1363                 obj->oo_inode = NULL;
1364                 if (trusted)
1365                         fid_zero(&oic->oic_fid);
1366         }
1367
1368         LINVRNT(osd_invariant(obj));
1369         return result;
1370 }
1371
1372 /*
1373  * Concurrency: shouldn't matter.
1374  */
1375 static void osd_object_init0(struct osd_object *obj)
1376 {
1377         LASSERT(obj->oo_inode != NULL);
1378         obj->oo_dt.do_body_ops = &osd_body_ops;
1379         obj->oo_dt.do_lu.lo_header->loh_attr |=
1380                 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1381 }
1382
1383 /*
1384  * Concurrency: no concurrent access is possible that early in object
1385  * life-cycle.
1386  */
1387 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1388                            const struct lu_object_conf *conf)
1389 {
1390         struct osd_object *obj = osd_obj(l);
1391         int result;
1392
1393         LINVRNT(osd_invariant(obj));
1394
1395         if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1396                 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1397                 l->lo_header->loh_attr |= LOHA_EXISTS;
1398                 return 0;
1399         }
1400
1401         result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1402         obj->oo_dt.do_body_ops = &osd_body_ops_new;
1403         if (result == 0 && obj->oo_inode != NULL) {
1404                 struct osd_thread_info *oti = osd_oti_get(env);
1405                 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1406
1407                 osd_object_init0(obj);
1408                 if (unlikely(obj->oo_header))
1409                         return 0;
1410
1411                 result = osd_get_lma(oti, obj->oo_inode,
1412                                      &oti->oti_obj_dentry, loa);
1413                 if (!result) {
1414                         /*
1415                          * Convert LMAI flags to lustre LMA flags
1416                          * and cache it to oo_lma_flags
1417                          */
1418                         obj->oo_lma_flags =
1419                                 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1420                 } else if (result == -ENODATA) {
1421                         result = 0;
1422                 }
1423         }
1424
1425         LINVRNT(osd_invariant(obj));
1426         return result;
1427 }
1428
1429 /*
1430  * The first part of oxe_buf is xattr name, and is '\0' terminated.
1431  * The left part is for value, binary mode.
1432  */
1433 struct osd_xattr_entry {
1434         struct list_head        oxe_list;
1435         size_t                  oxe_len;
1436         size_t                  oxe_namelen;
1437         bool                    oxe_exist;
1438         struct rcu_head         oxe_rcu;
1439         char                    oxe_buf[0];
1440 };
1441
1442 static int osd_oxc_get(struct osd_object *obj, const char *name,
1443                        struct lu_buf *buf)
1444 {
1445         struct osd_xattr_entry *tmp;
1446         struct osd_xattr_entry *oxe = NULL;
1447         size_t namelen = strlen(name);
1448         int rc;
1449
1450         ENTRY;
1451
1452         rcu_read_lock();
1453         list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1454                 if (namelen == tmp->oxe_namelen &&
1455                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1456                         oxe = tmp;
1457                         break;
1458                 }
1459         }
1460
1461         if (oxe == NULL)
1462                 GOTO(out, rc = -ENOENT);
1463
1464         if (!oxe->oxe_exist)
1465                 GOTO(out, rc = -ENODATA);
1466
1467         /* vallen */
1468         rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1469         LASSERT(rc > 0);
1470
1471         if (buf->lb_buf == NULL)
1472                 GOTO(out, rc);
1473
1474         if (buf->lb_len < rc)
1475                 GOTO(out, rc = -ERANGE);
1476
1477         memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1478         EXIT;
1479 out:
1480         rcu_read_unlock();
1481
1482         return rc;
1483 }
1484
1485 static void osd_oxc_free(struct rcu_head *head)
1486 {
1487         struct osd_xattr_entry *oxe;
1488
1489         oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1490         OBD_FREE(oxe, oxe->oxe_len);
1491 }
1492
1493 static void osd_oxc_add(struct osd_object *obj, const char *name,
1494                         const char *buf, int buflen)
1495 {
1496         struct osd_xattr_entry *oxe;
1497         struct osd_xattr_entry *old = NULL;
1498         struct osd_xattr_entry *tmp;
1499         size_t namelen = strlen(name);
1500         size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1501
1502         OBD_ALLOC(oxe, len);
1503         if (oxe == NULL)
1504                 return;
1505
1506         INIT_LIST_HEAD(&oxe->oxe_list);
1507         oxe->oxe_len = len;
1508         oxe->oxe_namelen = namelen;
1509         memcpy(oxe->oxe_buf, name, namelen);
1510         if (buflen > 0) {
1511                 LASSERT(buf != NULL);
1512                 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1513                 oxe->oxe_exist = true;
1514         } else {
1515                 oxe->oxe_exist = false;
1516         }
1517
1518         /* this should be rarely called, just remove old and add new */
1519         spin_lock(&obj->oo_guard);
1520         list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1521                 if (namelen == tmp->oxe_namelen &&
1522                     strncmp(name, tmp->oxe_buf, namelen) == 0) {
1523                         old = tmp;
1524                         break;
1525                 }
1526         }
1527         if (old != NULL) {
1528                 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1529                 call_rcu(&old->oxe_rcu, osd_oxc_free);
1530         } else {
1531                 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1532         }
1533         spin_unlock(&obj->oo_guard);
1534 }
1535
1536 static void osd_oxc_del(struct osd_object *obj, const char *name)
1537 {
1538         struct osd_xattr_entry *oxe;
1539         size_t namelen = strlen(name);
1540
1541         spin_lock(&obj->oo_guard);
1542         list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1543                 if (namelen == oxe->oxe_namelen &&
1544                     strncmp(name, oxe->oxe_buf, namelen) == 0) {
1545                         list_del_rcu(&oxe->oxe_list);
1546                         call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1547                         break;
1548                 }
1549         }
1550         spin_unlock(&obj->oo_guard);
1551 }
1552
1553 static void osd_oxc_fini(struct osd_object *obj)
1554 {
1555         struct osd_xattr_entry *oxe, *next;
1556
1557         list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1558                 list_del(&oxe->oxe_list);
1559                 OBD_FREE(oxe, oxe->oxe_len);
1560         }
1561 }
1562
1563 /*
1564  * Concurrency: no concurrent access is possible that late in object
1565  * life-cycle.
1566  */
1567 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1568 {
1569         struct osd_object *obj = osd_obj(l);
1570         struct lu_object_header *h = obj->oo_header;
1571
1572         LINVRNT(osd_invariant(obj));
1573
1574         osd_oxc_fini(obj);
1575         dt_object_fini(&obj->oo_dt);
1576         if (obj->oo_hl_head != NULL)
1577                 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1578         OBD_FREE_PTR(obj);
1579         if (unlikely(h)) {
1580                 lu_object_header_fini(h);
1581                 OBD_FREE_PTR(h);
1582         }
1583 }
1584
1585 /*
1586  * Concurrency: no concurrent access is possible that late in object
1587  * life-cycle.
1588  */
1589 static void osd_index_fini(struct osd_object *o)
1590 {
1591         struct iam_container *bag;
1592
1593         if (o->oo_dir != NULL) {
1594                 bag = &o->oo_dir->od_container;
1595                 if (o->oo_inode != NULL) {
1596                         if (bag->ic_object == o->oo_inode)
1597                                 iam_container_fini(bag);
1598                 }
1599                 OBD_FREE_PTR(o->oo_dir);
1600                 o->oo_dir = NULL;
1601         }
1602 }
1603
1604 /*
1605  * Concurrency: no concurrent access is possible that late in object
1606  * life-cycle (for all existing callers, that is. New callers have to provide
1607  * their own locking.)
1608  */
1609 static int osd_inode_unlinked(const struct inode *inode)
1610 {
1611         return inode->i_nlink == 0;
1612 }
1613
1614 enum {
1615         OSD_TXN_OI_DELETE_CREDITS    = 20,
1616         OSD_TXN_INODE_DELETE_CREDITS = 20
1617 };
1618
1619 /*
1620  * Journal
1621  */
1622
1623 #if OSD_THANDLE_STATS
1624 /**
1625  * Set time when the handle is allocated
1626  */
1627 static void osd_th_alloced(struct osd_thandle *oth)
1628 {
1629         oth->oth_alloced = ktime_get();
1630 }
1631
1632 /**
1633  * Set time when the handle started
1634  */
1635 static void osd_th_started(struct osd_thandle *oth)
1636 {
1637         oth->oth_started = ktime_get();
1638 }
1639
1640 /**
1641  * Check whether the we deal with this handle for too long.
1642  */
1643 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1644                                 ktime_t alloced, ktime_t started,
1645                                 ktime_t closed)
1646 {
1647         ktime_t now = ktime_get();
1648
1649         LASSERT(dev != NULL);
1650
1651         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1652                             ktime_us_delta(started, alloced));
1653         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1654                             ktime_us_delta(closed, started));
1655         lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1656                             ktime_us_delta(now, closed));
1657
1658         if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1659                 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1660                                 oth, now, alloced, started, closed);
1661                 libcfs_debug_dumpstack(NULL);
1662         }
1663 }
1664
1665 #define OSD_CHECK_SLOW_TH(oth, dev, expr)                               \
1666 {                                                                       \
1667         ktime_t __closed = ktime_get();                                 \
1668         ktime_t __alloced = oth->oth_alloced;                           \
1669         ktime_t __started = oth->oth_started;                           \
1670                                                                         \
1671         expr;                                                           \
1672         __osd_th_check_slow(oth, dev, __alloced, __started, __closed);  \
1673 }
1674
1675 #else /* OSD_THANDLE_STATS */
1676
1677 #define osd_th_alloced(h)                  do {} while(0)
1678 #define osd_th_started(h)                  do {} while(0)
1679 #define OSD_CHECK_SLOW_TH(oth, dev, expr)  expr
1680
1681 #endif /* OSD_THANDLE_STATS */
1682
1683 /*
1684  * Concurrency: doesn't access mutable data.
1685  */
1686 static int osd_param_is_not_sane(const struct osd_device *dev,
1687                                  const struct thandle *th)
1688 {
1689         struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1690
1691         return oh->ot_credits > osd_transaction_size(dev);
1692 }
1693
1694 /*
1695  * Concurrency: shouldn't matter.
1696  */
1697 static void osd_trans_commit_cb(struct super_block *sb,
1698                                 struct ldiskfs_journal_cb_entry *jcb, int error)
1699 {
1700         struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1701         struct thandle *th = &oh->ot_super;
1702         struct lu_device *lud = &th->th_dev->dd_lu_dev;
1703         struct dt_txn_commit_cb *dcb, *tmp;
1704
1705         LASSERT(oh->ot_handle == NULL);
1706
1707         if (error)
1708                 CERROR("transaction @0x%p commit error: %d\n", th, error);
1709
1710         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DELAY_TRANS, 40);
1711         /* call per-transaction callbacks if any */
1712         list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1713                                  dcb_linkage) {
1714                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1715                          "commit callback entry: magic=%x name='%s'\n",
1716                          dcb->dcb_magic, dcb->dcb_name);
1717                 list_del_init(&dcb->dcb_linkage);
1718                 dcb->dcb_func(NULL, th, dcb, error);
1719         }
1720
1721         lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1722         lu_device_put(lud);
1723         th->th_dev = NULL;
1724
1725         OBD_FREE_PTR(oh);
1726 }
1727
1728 #ifndef HAVE_SB_START_WRITE
1729 # define sb_start_write(sb) do {} while (0)
1730 # define sb_end_write(sb) do {} while (0)
1731 #endif
1732
1733 static struct thandle *osd_trans_create(const struct lu_env *env,
1734                                         struct dt_device *d)
1735 {
1736         struct osd_thread_info *oti = osd_oti_get(env);
1737         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1738         struct osd_thandle *oh;
1739         struct thandle *th;
1740
1741         ENTRY;
1742
1743         if (d->dd_rdonly) {
1744                 CERROR("%s: someone try to start transaction under "
1745                        "readonly mode, should be disabled.\n",
1746                        osd_name(osd_dt_dev(d)));
1747                 dump_stack();
1748                 RETURN(ERR_PTR(-EROFS));
1749         }
1750
1751         /* on pending IO in this thread should left from prev. request */
1752         LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1753
1754         sb_start_write(osd_sb(osd_dt_dev(d)));
1755
1756         OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1757         if (!oh) {
1758                 sb_end_write(osd_sb(osd_dt_dev(d)));
1759                 RETURN(ERR_PTR(-ENOMEM));
1760         }
1761
1762         oh->ot_quota_trans = &oti->oti_quota_trans;
1763         memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1764         th = &oh->ot_super;
1765         th->th_dev = d;
1766         th->th_result = 0;
1767         oh->ot_credits = 0;
1768         INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1769         INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1770         INIT_LIST_HEAD(&oh->ot_trunc_locks);
1771         osd_th_alloced(oh);
1772
1773         memset(oti->oti_declare_ops, 0,
1774                sizeof(oti->oti_declare_ops));
1775         memset(oti->oti_declare_ops_cred, 0,
1776                sizeof(oti->oti_declare_ops_cred));
1777         memset(oti->oti_declare_ops_used, 0,
1778                sizeof(oti->oti_declare_ops_used));
1779
1780         oti->oti_ins_cache_depth++;
1781
1782         RETURN(th);
1783 }
1784
1785 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1786 {
1787         struct osd_thread_info *oti = osd_oti_get(env);
1788         struct osd_thandle *oh;
1789
1790         oh = container_of0(th, struct osd_thandle, ot_super);
1791         LASSERT(oh != NULL);
1792
1793         CWARN("  create: %u/%u/%u, destroy: %u/%u/%u\n",
1794               oti->oti_declare_ops[OSD_OT_CREATE],
1795               oti->oti_declare_ops_cred[OSD_OT_CREATE],
1796               oti->oti_declare_ops_used[OSD_OT_CREATE],
1797               oti->oti_declare_ops[OSD_OT_DESTROY],
1798               oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1799               oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1800         CWARN("  attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1801               oti->oti_declare_ops[OSD_OT_ATTR_SET],
1802               oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1803               oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1804               oti->oti_declare_ops[OSD_OT_XATTR_SET],
1805               oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1806               oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1807         CWARN("  write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1808               oti->oti_declare_ops[OSD_OT_WRITE],
1809               oti->oti_declare_ops_cred[OSD_OT_WRITE],
1810               oti->oti_declare_ops_used[OSD_OT_WRITE],
1811               oti->oti_declare_ops[OSD_OT_PUNCH],
1812               oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1813               oti->oti_declare_ops_used[OSD_OT_PUNCH],
1814               oti->oti_declare_ops[OSD_OT_QUOTA],
1815               oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1816               oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1817         CWARN("  insert: %u/%u/%u, delete: %u/%u/%u\n",
1818               oti->oti_declare_ops[OSD_OT_INSERT],
1819               oti->oti_declare_ops_cred[OSD_OT_INSERT],
1820               oti->oti_declare_ops_used[OSD_OT_INSERT],
1821               oti->oti_declare_ops[OSD_OT_DELETE],
1822               oti->oti_declare_ops_cred[OSD_OT_DELETE],
1823               oti->oti_declare_ops_used[OSD_OT_DELETE]);
1824         CWARN("  ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1825               oti->oti_declare_ops[OSD_OT_REF_ADD],
1826               oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1827               oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1828               oti->oti_declare_ops[OSD_OT_REF_DEL],
1829               oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1830               oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1831 }
1832
1833 /*
1834  * Concurrency: shouldn't matter.
1835  */
1836 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1837                            struct thandle *th)
1838 {
1839         struct osd_thread_info *oti = osd_oti_get(env);
1840         struct osd_device *dev = osd_dt_dev(d);
1841         handle_t *jh;
1842         struct osd_thandle *oh;
1843         int rc;
1844
1845         ENTRY;
1846
1847         LASSERT(current->journal_info == NULL);
1848
1849         oh = container_of0(th, struct osd_thandle, ot_super);
1850         LASSERT(oh != NULL);
1851         LASSERT(oh->ot_handle == NULL);
1852
1853         rc = dt_txn_hook_start(env, d, th);
1854         if (rc != 0)
1855                 GOTO(out, rc);
1856
1857         if (unlikely(osd_param_is_not_sane(dev, th))) {
1858                 static unsigned long last_printed;
1859                 static int last_credits;
1860
1861                 /*
1862                  * don't make noise on a tiny testing systems
1863                  * actual credits misuse will be caught anyway
1864                  */
1865                 if (last_credits != oh->ot_credits &&
1866                     time_after(jiffies, last_printed +
1867                                msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1868                     osd_transaction_size(dev) > 512) {
1869                         CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1870                               oh->ot_credits, osd_transaction_size(dev));
1871                         osd_trans_dump_creds(env, th);
1872                         libcfs_debug_dumpstack(NULL);
1873                         last_credits = oh->ot_credits;
1874                         last_printed = jiffies;
1875                 }
1876                 /*
1877                  * XXX Limit the credits to 'max_transaction_buffers', and
1878                  *     let the underlying filesystem to catch the error if
1879                  *     we really need so many credits.
1880                  *
1881                  *     This should be removed when we can calculate the
1882                  *     credits precisely.
1883                  */
1884                 oh->ot_credits = osd_transaction_size(dev);
1885         } else if (ldiskfs_track_declares_assert != 0) {
1886                 /*
1887                  * reserve few credits to prevent an assertion in JBD
1888                  * our debugging mechanism will be able to detected
1889                  * overuse. this can help to debug single-update
1890                  * transactions
1891                  */
1892                 oh->ot_credits += 10;
1893                 if (unlikely(osd_param_is_not_sane(dev, th)))
1894                         oh->ot_credits = osd_transaction_size(dev);
1895         }
1896
1897         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
1898                 GOTO(out, rc = -EIO);
1899
1900         /*
1901          * XXX temporary stuff. Some abstraction layer should
1902          * be used.
1903          */
1904         jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1905         osd_th_started(oh);
1906         if (!IS_ERR(jh)) {
1907                 oh->ot_handle = jh;
1908                 LASSERT(oti->oti_txns == 0);
1909
1910                 lu_device_get(&d->dd_lu_dev);
1911                 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1912                               "osd-tx", th);
1913                 oti->oti_txns++;
1914                 rc = 0;
1915         } else {
1916                 rc = PTR_ERR(jh);
1917         }
1918 out:
1919         RETURN(rc);
1920 }
1921
1922 static int osd_seq_exists(const struct lu_env *env,
1923                           struct osd_device *osd, u64 seq)
1924 {
1925         struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1926         struct seq_server_site *ss = osd_seq_site(osd);
1927         int rc;
1928
1929         ENTRY;
1930
1931         LASSERT(ss != NULL);
1932         LASSERT(ss->ss_server_fld != NULL);
1933
1934         rc = osd_fld_lookup(env, osd, seq, range);
1935         if (rc != 0) {
1936                 if (rc != -ENOENT)
1937                         CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1938                                osd_name(osd), seq, rc);
1939                 RETURN(0);
1940         }
1941
1942         RETURN(ss->ss_node_id == range->lsr_index);
1943 }
1944
1945 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1946 {
1947         struct dt_txn_commit_cb *dcb;
1948         struct dt_txn_commit_cb *tmp;
1949
1950         /* call per-transaction stop callbacks if any */
1951         list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1952                                  dcb_linkage) {
1953                 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1954                          "commit callback entry: magic=%x name='%s'\n",
1955                          dcb->dcb_magic, dcb->dcb_name);
1956                 list_del_init(&dcb->dcb_linkage);
1957                 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1958         }
1959 }
1960
1961 /*
1962  * Concurrency: shouldn't matter.
1963  */
1964 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1965                           struct thandle *th)
1966 {
1967         struct osd_thread_info *oti = osd_oti_get(env);
1968         struct osd_thandle *oh;
1969         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1970         struct osd_device *osd = osd_dt_dev(th->th_dev);
1971         struct qsd_instance *qsd = osd_def_qsd(osd);
1972         struct lquota_trans *qtrans;
1973         struct list_head truncates = LIST_HEAD_INIT(truncates);
1974         int rc = 0, remove_agents = 0;
1975
1976         ENTRY;
1977
1978         oh = container_of0(th, struct osd_thandle, ot_super);
1979
1980         remove_agents = oh->ot_remove_agents;
1981
1982         qtrans = oh->ot_quota_trans;
1983         oh->ot_quota_trans = NULL;
1984
1985         /* move locks to local list, stop tx, execute truncates */
1986         list_splice(&oh->ot_trunc_locks, &truncates);
1987
1988         if (oh->ot_handle != NULL) {
1989                 int rc2;
1990
1991                 handle_t *hdl = oh->ot_handle;
1992
1993                 /*
1994                  * add commit callback
1995                  * notice we don't do this in osd_trans_start()
1996                  * as underlying transaction can change during truncate
1997                  */
1998                 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1999                                              &oh->ot_jcb);
2000
2001                 LASSERT(oti->oti_txns == 1);
2002                 oti->oti_txns--;
2003
2004                 rc = dt_txn_hook_stop(env, th);
2005                 if (rc != 0)
2006                         CERROR("%s: failed in transaction hook: rc = %d\n",
2007                                osd_name(osd), rc);
2008
2009                 osd_trans_stop_cb(oh, rc);
2010                 /* hook functions might modify th_sync */
2011                 hdl->h_sync = th->th_sync;
2012
2013                 oh->ot_handle = NULL;
2014                 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2015                 if (rc2 != 0)
2016                         CERROR("%s: failed to stop transaction: rc = %d\n",
2017                                osd_name(osd), rc2);
2018                 if (!rc)
2019                         rc = rc2;
2020
2021                 osd_process_truncates(&truncates);
2022         } else {
2023                 osd_trans_stop_cb(oh, th->th_result);
2024                 OBD_FREE_PTR(oh);
2025         }
2026
2027         osd_trunc_unlock_all(&truncates);
2028
2029         /* inform the quota slave device that the transaction is stopping */
2030         qsd_op_end(env, qsd, qtrans);
2031
2032         /*
2033          * as we want IO to journal and data IO be concurrent, we don't block
2034          * awaiting data IO completion in osd_do_bio(), instead we wait here
2035          * once transaction is submitted to the journal. all reqular requests
2036          * don't do direct IO (except read/write), thus this wait_event becomes
2037          * no-op for them.
2038          *
2039          * IMPORTANT: we have to wait till any IO submited by the thread is
2040          * completed otherwise iobuf may be corrupted by different request
2041          */
2042         wait_event(iobuf->dr_wait,
2043                        atomic_read(&iobuf->dr_numreqs) == 0);
2044         osd_fini_iobuf(osd, iobuf);
2045         if (!rc)
2046                 rc = iobuf->dr_error;
2047
2048         if (unlikely(remove_agents != 0))
2049                 osd_process_scheduled_agent_removals(env, osd);
2050
2051         oti->oti_ins_cache_depth--;
2052         /* reset OI cache for safety */
2053         if (oti->oti_ins_cache_depth == 0)
2054                 oti->oti_ins_cache_used = 0;
2055
2056         sb_end_write(osd_sb(osd));
2057
2058         RETURN(rc);
2059 }
2060
2061 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2062 {
2063         struct osd_thandle *oh = container_of0(th, struct osd_thandle,
2064                                                ot_super);
2065
2066         LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2067         LASSERT(&dcb->dcb_func != NULL);
2068         if (dcb->dcb_flags & DCB_TRANS_STOP)
2069                 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2070         else
2071                 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2072
2073         return 0;
2074 }
2075
2076 /*
2077  * Called just before object is freed. Releases all resources except for
2078  * object itself (that is released by osd_object_free()).
2079  *
2080  * Concurrency: no concurrent access is possible that late in object
2081  * life-cycle.
2082  */
2083 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2084 {
2085         struct osd_object *obj = osd_obj(l);
2086         struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2087         struct inode *inode = obj->oo_inode;
2088         __u64 projid;
2089         qid_t uid;
2090         qid_t gid;
2091
2092         LINVRNT(osd_invariant(obj));
2093
2094         /*
2095          * If object is unlinked remove fid->ino mapping from object index.
2096          */
2097
2098         osd_index_fini(obj);
2099
2100         if (!inode)
2101                 return;
2102
2103         uid = i_uid_read(inode);
2104         gid = i_gid_read(inode);
2105         projid = i_projid_read(inode);
2106
2107         obj->oo_inode = NULL;
2108         iput(inode);
2109
2110         /* do not rebalance quota if the caller needs to release memory
2111          * otherwise qsd_refresh_usage() may went into a new ldiskfs
2112          * transaction and risk to deadlock - LU-12178 */
2113         if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
2114                 return;
2115
2116         if (!obj->oo_header && qsd) {
2117                 struct osd_thread_info *info = osd_oti_get(env);
2118                 struct lquota_id_info *qi = &info->oti_qi;
2119
2120                 /* Release granted quota to master if necessary */
2121                 qi->lqi_id.qid_uid = uid;
2122                 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2123
2124                 qi->lqi_id.qid_uid = gid;
2125                 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2126
2127                 qi->lqi_id.qid_uid = projid;
2128                 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2129         }
2130 }
2131
2132 /*
2133  * Concurrency: ->loo_object_release() is called under site spin-lock.
2134  */
2135 static void osd_object_release(const struct lu_env *env,
2136                                struct lu_object *l)
2137 {
2138         struct osd_object *o = osd_obj(l);
2139
2140         /*
2141          * nobody should be releasing a non-destroyed object with nlink=0
2142          * the API allows this, but ldiskfs doesn't like and then report
2143          * this inode as deleted
2144          */
2145         LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2146                   o->oo_inode->i_nlink == 0));
2147 }
2148
2149 /*
2150  * Concurrency: shouldn't matter.
2151  */
2152 static int osd_object_print(const struct lu_env *env, void *cookie,
2153                             lu_printer_t p, const struct lu_object *l)
2154 {
2155         struct osd_object *o = osd_obj(l);
2156         struct iam_descr *d;
2157
2158         if (o->oo_dir != NULL)
2159                 d = o->oo_dir->od_container.ic_descr;
2160         else
2161                 d = NULL;
2162         return (*p)(env, cookie,
2163                     LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2164                     o, o->oo_inode,
2165                     o->oo_inode ? o->oo_inode->i_ino : 0UL,
2166                     o->oo_inode ? o->oo_inode->i_generation : 0,
2167                     d ? d->id_ops->id_name : "plain");
2168 }
2169
2170 /*
2171  * Concurrency: shouldn't matter.
2172  */
2173 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2174                 struct obd_statfs *sfs)
2175 {
2176         struct osd_device *osd = osd_dt_dev(d);
2177         struct super_block *sb = osd_sb(osd);
2178         struct kstatfs *ksfs;
2179         __u64 reserved;
2180         int result = 0;
2181
2182         if (unlikely(osd->od_mnt == NULL))
2183                 return -EINPROGRESS;
2184
2185         /* osd_lproc.c call this without env, allocate ksfs for that case */
2186         if (unlikely(env == NULL)) {
2187                 OBD_ALLOC_PTR(ksfs);
2188                 if (ksfs == NULL)
2189                         return -ENOMEM;
2190         } else {
2191                 ksfs = &osd_oti_get(env)->oti_ksfs;
2192         }
2193
2194         result = sb->s_op->statfs(sb->s_root, ksfs);
2195         if (result)
2196                 goto out;
2197
2198         statfs_pack(sfs, ksfs);
2199         if (unlikely(sb->s_flags & MS_RDONLY))
2200                 sfs->os_state |= OS_STATE_READONLY;
2201         if (ldiskfs_has_feature_extents(sb))
2202                 sfs->os_maxbytes = sb->s_maxbytes;
2203         else
2204                 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2205
2206         /*
2207          * Reserve some space so to avoid fragmenting the filesystem too much.
2208          * Fragmentation not only impacts performance, but can also increase
2209          * metadata overhead significantly, causing grant calculation to be
2210          * wrong.
2211          *
2212          * Reserve 0.78% of total space, at least 8MB for small filesystems.
2213          */
2214         CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2215         reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2216         if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2217                 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2218
2219         sfs->os_blocks -= reserved;
2220         sfs->os_bfree  -= min(reserved, sfs->os_bfree);
2221         sfs->os_bavail -= min(reserved, sfs->os_bavail);
2222
2223 out:
2224         if (unlikely(env == NULL))
2225                 OBD_FREE_PTR(ksfs);
2226         return result;
2227 }
2228
2229 /**
2230  * Estimate space needed for file creations. We assume the largest filename
2231  * which is 2^64 - 1, hence a filename of 20 chars.
2232  * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2233  */
2234 #ifdef __LDISKFS_DIR_REC_LEN
2235 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2236 #else
2237 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2238 #endif
2239
2240 /*
2241  * Concurrency: doesn't access mutable data.
2242  */
2243 static void osd_conf_get(const struct lu_env *env,
2244                          const struct dt_device *dev,
2245                          struct dt_device_param *param)
2246 {
2247         struct osd_device *d = osd_dt_dev(dev);
2248         struct super_block *sb = osd_sb(d);
2249         struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2250         const char *name;
2251         int ea_overhead;
2252
2253         /*
2254          * XXX should be taken from not-yet-existing fs abstraction layer.
2255          */
2256         param->ddp_max_name_len = LDISKFS_NAME_LEN;
2257         param->ddp_max_nlink    = LDISKFS_LINK_MAX;
2258         param->ddp_symlink_max  = sb->s_blocksize;
2259         param->ddp_mount_type   = LDD_MT_LDISKFS;
2260         if (ldiskfs_has_feature_extents(sb))
2261                 param->ddp_maxbytes = sb->s_maxbytes;
2262         else
2263                 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2264         /*
2265          * inode are statically allocated, so per-inode space consumption
2266          * is the space consumed by the directory entry
2267          */
2268         param->ddp_inodespace     = PER_OBJ_USAGE;
2269         /*
2270          * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2271          * is 128MB) which is unlikely to be hit in real life. Report a smaller
2272          * maximum length to not under-count the actual number of extents
2273          * needed for writing a file if there are sub-optimal block allocations.
2274          */
2275         param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
2276         /* worst-case extent insertion metadata overhead */
2277         param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2278         param->ddp_mntopts = 0;
2279         if (test_opt(sb, XATTR_USER))
2280                 param->ddp_mntopts |= MNTOPT_USERXATTR;
2281         if (test_opt(sb, POSIX_ACL))
2282                 param->ddp_mntopts |= MNTOPT_ACL;
2283
2284         /*
2285          * LOD might calculate the max stripe count based on max_ea_size,
2286          * so we need take account in the overhead as well,
2287          * xattr_header + magic + xattr_entry_head
2288          */
2289         ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2290                       LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2291
2292 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2293         if (ldiskfs_has_feature_ea_inode(sb))
2294                 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2295                                                                 ea_overhead;
2296         else
2297 #endif
2298                 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2299
2300         if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE - ea_overhead)
2301                 param->ddp_max_ea_size = OBD_MAX_EA_SIZE - ea_overhead;
2302
2303         /*
2304          * Preferred RPC size for efficient disk IO.  4MB shows good
2305          * all-around performance for ldiskfs, but use bigalloc chunk size
2306          * by default if larger.
2307          */
2308 #if defined(LDISKFS_CLUSTER_SIZE)
2309         if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2310                 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2311         else
2312 #endif
2313                 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2314
2315         param->ddp_t10_cksum_type = 0;
2316         if (bi) {
2317                 unsigned short interval = blk_integrity_interval(bi);
2318                 name = blk_integrity_name(bi);
2319                 /*
2320                  * Expected values:
2321                  * T10-DIF-TYPE1-CRC
2322                  * T10-DIF-TYPE3-CRC
2323                  * T10-DIF-TYPE1-IP
2324                  * T10-DIF-TYPE3-IP
2325                  */
2326                 if (strncmp(name, "T10-DIF-TYPE",
2327                             sizeof("T10-DIF-TYPE") - 1) == 0) {
2328                         /* also skip "1/3-" at end */
2329                         const int type_off = sizeof("T10-DIF-TYPE.");
2330                         char type_number = name[type_off - 2];
2331
2332                         if (interval != 512 && interval != 4096) {
2333                                 CERROR("%s: unsupported T10PI sector size %u\n",
2334                                        d->od_svname, interval);
2335                         } else if (type_number != '1' && type_number != '3') {
2336                                 CERROR("%s: unsupported T10PI type %s\n",
2337                                        d->od_svname, name);
2338                         } else if (strcmp(name + type_off, "CRC") == 0) {
2339                                 d->od_t10_type = type_number == '1' ?
2340                                         OSD_T10_TYPE1_CRC : OSD_T10_TYPE3_CRC;
2341                                 param->ddp_t10_cksum_type = interval == 512 ?
2342                                         OBD_CKSUM_T10CRC512 :
2343                                         OBD_CKSUM_T10CRC4K;
2344                         } else if (strcmp(name + type_off, "IP") == 0) {
2345                                 d->od_t10_type = type_number == '1' ?
2346                                         OSD_T10_TYPE1_IP : OSD_T10_TYPE3_IP;
2347                                 param->ddp_t10_cksum_type = interval == 512 ?
2348                                         OBD_CKSUM_T10IP512 :
2349                                         OBD_CKSUM_T10IP4K;
2350                         } else {
2351                                 CERROR("%s: unsupported checksum type of "
2352                                        "T10PI type '%s'",
2353                                        d->od_svname, name);
2354                         }
2355
2356                 } else {
2357                         CERROR("%s: unsupported T10PI type '%s'",
2358                                d->od_svname, name);
2359                 }
2360         }
2361 }
2362
2363 static struct super_block *osd_mnt_sb_get(const struct dt_device *d)
2364 {
2365         return osd_sb(osd_dt_dev(d));
2366 }
2367
2368 /*
2369  * Concurrency: shouldn't matter.
2370  */
2371 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2372 {
2373         int rc;
2374         struct super_block *s = osd_sb(osd_dt_dev(d));
2375         ENTRY;
2376
2377         down_read(&s->s_umount);
2378         rc = s->s_op->sync_fs(s, 1);
2379         up_read(&s->s_umount);
2380
2381         CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2382                rc);
2383
2384         return rc;
2385 }
2386
2387 /**
2388  * Start commit for OSD device.
2389  *
2390  * An implementation of dt_commit_async method for OSD device.
2391  * Asychronously starts underlayng fs sync and thereby a transaction
2392  * commit.
2393  *
2394  * \param env environment
2395  * \param d dt device
2396  *
2397  * \see dt_device_operations
2398  */
2399 static int osd_commit_async(const struct lu_env *env,
2400                             struct dt_device *d)
2401 {
2402         struct super_block *s = osd_sb(osd_dt_dev(d));
2403         int rc;
2404
2405         ENTRY;
2406
2407         CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2408         down_read(&s->s_umount);
2409         rc = s->s_op->sync_fs(s, 0);
2410         up_read(&s->s_umount);
2411
2412         RETURN(rc);
2413 }
2414
2415 /* Our own copy of the set readonly functions if present, or NU if not. */
2416 static int (*priv_dev_set_rdonly)(struct block_device *bdev);
2417 static int (*priv_dev_check_rdonly)(struct block_device *bdev);
2418 /* static int (*priv_dev_clear_rdonly)(struct block_device *bdev); */
2419
2420 /*
2421  * Concurrency: shouldn't matter.
2422  */
2423 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2424 {
2425         struct super_block *sb = osd_sb(osd_dt_dev(d));
2426         struct block_device *dev = sb->s_bdev;
2427         int rc = -EOPNOTSUPP;
2428
2429         ENTRY;
2430
2431         if (priv_dev_set_rdonly) {
2432                 struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2433
2434                 rc = 0;
2435                 CERROR("*** setting %s read-only ***\n",
2436                        osd_dt_dev(d)->od_svname);
2437
2438                 if (sb->s_op->freeze_fs) {
2439                         rc = sb->s_op->freeze_fs(sb);
2440                         if (rc)
2441                                 goto out;
2442                 }
2443
2444                 if (jdev && (jdev != dev)) {
2445                         CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2446                                (long)jdev);
2447                         priv_dev_set_rdonly(jdev);
2448                 }
2449                 CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2450                 priv_dev_set_rdonly(dev);
2451
2452                 if (sb->s_op->unfreeze_fs)
2453                         sb->s_op->unfreeze_fs(sb);
2454         }
2455
2456 out:
2457         if (rc)
2458                 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2459                        osd_dt_dev(d)->od_svname, (long)dev, rc);
2460
2461         RETURN(rc);
2462 }
2463
2464 /**
2465  * Note: we do not count into QUOTA here.
2466  * If we mount with --data_journal we may need more.
2467  */
2468 const int osd_dto_credits_noquota[DTO_NR] = {
2469         /**
2470          * Insert.
2471          * INDEX_EXTRA_TRANS_BLOCKS(8) +
2472          * SINGLEDATA_TRANS_BLOCKS(8)
2473          * XXX Note: maybe iam need more, since iam have more level than
2474          *           EXT3 htree.
2475          */
2476         [DTO_INDEX_INSERT]  = 16,
2477         /**
2478          * Delete
2479          * just modify a single entry, probably merge few within a block
2480          */
2481         [DTO_INDEX_DELETE]  = 1,
2482         /**
2483          * Used for OI scrub
2484          */
2485         [DTO_INDEX_UPDATE]  = 16,
2486         /**
2487          * 4(inode, inode bits, groups, GDT)
2488          *   notice: OI updates are counted separately with DTO_INDEX_INSERT
2489          */
2490         [DTO_OBJECT_CREATE] = 4,
2491         /**
2492          * 4(inode, inode bits, groups, GDT)
2493          *   notice: OI updates are counted separately with DTO_INDEX_DELETE
2494          */
2495         [DTO_OBJECT_DELETE] = 4,
2496         /**
2497          * Attr set credits (inode)
2498          */
2499         [DTO_ATTR_SET_BASE] = 1,
2500         /**
2501          * Xattr set. The same as xattr of EXT3.
2502          * DATA_TRANS_BLOCKS(14)
2503          * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2504          * are also counted in. Do not know why?
2505          */
2506         [DTO_XATTR_SET]     = 14,
2507         /**
2508          * credits for inode change during write.
2509          */
2510         [DTO_WRITE_BASE]    = 3,
2511         /**
2512          * credits for single block write.
2513          */
2514         [DTO_WRITE_BLOCK]   = 14,
2515         /**
2516          * Attr set credits for chown.
2517          * This is extra credits for setattr, and it is null without quota
2518          */
2519         [DTO_ATTR_SET_CHOWN] = 0
2520 };
2521
2522 static const struct dt_device_operations osd_dt_ops = {
2523         .dt_root_get       = osd_root_get,
2524         .dt_statfs         = osd_statfs,
2525         .dt_trans_create   = osd_trans_create,
2526         .dt_trans_start    = osd_trans_start,
2527         .dt_trans_stop     = osd_trans_stop,
2528         .dt_trans_cb_add   = osd_trans_cb_add,
2529         .dt_conf_get       = osd_conf_get,
2530         .dt_mnt_sb_get     = osd_mnt_sb_get,
2531         .dt_sync           = osd_sync,
2532         .dt_ro             = osd_ro,
2533         .dt_commit_async   = osd_commit_async,
2534 };
2535
2536 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2537                           unsigned int role)
2538 {
2539         struct osd_object *obj = osd_dt_obj(dt);
2540         struct osd_thread_info *oti = osd_oti_get(env);
2541
2542         LINVRNT(osd_invariant(obj));
2543
2544         LASSERT(obj->oo_owner != env);
2545         down_read_nested(&obj->oo_sem, role);
2546
2547         LASSERT(obj->oo_owner == NULL);
2548         oti->oti_r_locks++;
2549 }
2550
2551 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2552                            unsigned int role)
2553 {
2554         struct osd_object *obj = osd_dt_obj(dt);
2555         struct osd_thread_info *oti = osd_oti_get(env);
2556
2557         LINVRNT(osd_invariant(obj));
2558
2559         LASSERT(obj->oo_owner != env);
2560         down_write_nested(&obj->oo_sem, role);
2561
2562         LASSERT(obj->oo_owner == NULL);
2563         obj->oo_owner = env;
2564         oti->oti_w_locks++;
2565 }
2566
2567 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2568 {
2569         struct osd_object *obj = osd_dt_obj(dt);
2570         struct osd_thread_info *oti = osd_oti_get(env);
2571
2572         LINVRNT(osd_invariant(obj));
2573
2574         LASSERT(oti->oti_r_locks > 0);
2575         oti->oti_r_locks--;
2576         up_read(&obj->oo_sem);
2577 }
2578
2579 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2580 {
2581         struct osd_object *obj = osd_dt_obj(dt);
2582         struct osd_thread_info *oti = osd_oti_get(env);
2583
2584         LINVRNT(osd_invariant(obj));
2585
2586         LASSERT(obj->oo_owner == env);
2587         LASSERT(oti->oti_w_locks > 0);
2588         oti->oti_w_locks--;
2589         obj->oo_owner = NULL;
2590         up_write(&obj->oo_sem);
2591 }
2592
2593 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2594 {
2595         struct osd_object *obj = osd_dt_obj(dt);
2596
2597         LINVRNT(osd_invariant(obj));
2598
2599         return obj->oo_owner == env;
2600 }
2601
2602 static struct timespec *osd_inode_time(const struct lu_env *env,
2603                                        struct inode *inode, __u64 seconds)
2604 {
2605         struct osd_thread_info *oti = osd_oti_get(env);
2606         struct timespec *t = &oti->oti_time;
2607
2608         t->tv_sec = seconds;
2609         t->tv_nsec = 0;
2610         *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2611         return t;
2612 }
2613
2614 static void osd_inode_getattr(const struct lu_env *env,
2615                               struct inode *inode, struct lu_attr *attr)
2616 {
2617         attr->la_valid  |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2618                            LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2619                            LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2620                            LA_BLKSIZE | LA_TYPE;
2621
2622         attr->la_atime = inode->i_atime.tv_sec;
2623         attr->la_mtime = inode->i_mtime.tv_sec;
2624         attr->la_ctime = inode->i_ctime.tv_sec;
2625         attr->la_mode    = inode->i_mode;
2626         attr->la_size    = i_size_read(inode);
2627         attr->la_blocks  = inode->i_blocks;
2628         attr->la_uid     = i_uid_read(inode);
2629         attr->la_gid     = i_gid_read(inode);
2630         attr->la_projid  = i_projid_read(inode);
2631         attr->la_flags   = ll_inode_to_ext_flags(inode->i_flags);
2632         attr->la_nlink   = inode->i_nlink;
2633         attr->la_rdev    = inode->i_rdev;
2634         attr->la_blksize = 1 << inode->i_blkbits;
2635         attr->la_blkbits = inode->i_blkbits;
2636         /*
2637          * Ext4 did not transfer inherit flags from raw inode
2638          * to inode flags, and ext4 internally test raw inode
2639          * @i_flags directly. Instead of patching ext4, we do it here.
2640          */
2641         if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2642                 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2643 }
2644
2645 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2646                         struct lu_attr *attr)
2647 {
2648         struct osd_object *obj = osd_dt_obj(dt);
2649
2650         if (unlikely(!dt_object_exists(dt)))
2651                 return -ENOENT;
2652         if (unlikely(obj->oo_destroyed))
2653                 return -ENOENT;
2654
2655         LASSERT(!dt_object_remote(dt));
2656         LINVRNT(osd_invariant(obj));
2657
2658         spin_lock(&obj->oo_guard);
2659         osd_inode_getattr(env, obj->oo_inode, attr);
2660         if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2661                 attr->la_valid |= LA_FLAGS;
2662                 attr->la_flags |= LUSTRE_ORPHAN_FL;
2663         }
2664         spin_unlock(&obj->oo_guard);
2665
2666         return 0;
2667 }
2668
2669 static int osd_declare_attr_qid(const struct lu_env *env,
2670                                 struct osd_object *obj,
2671                                 struct osd_thandle *oh, long long bspace,
2672                                 qid_t old_id, qid_t new_id, bool enforce,
2673                                 unsigned int type, bool ignore_edquot)
2674 {
2675         int rc;
2676         struct osd_thread_info *info = osd_oti_get(env);
2677         struct lquota_id_info  *qi = &info->oti_qi;
2678
2679         qi->lqi_type = type;
2680         /* inode accounting */
2681         qi->lqi_is_blk = false;
2682
2683         /* one more inode for the new id ... */
2684         qi->lqi_id.qid_uid = new_id;
2685         qi->lqi_space      = 1;
2686         /* Reserve credits for the new id */
2687         rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2688         if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2689                 rc = 0;
2690         if (rc)
2691                 RETURN(rc);
2692
2693         /* and one less inode for the current id */
2694         qi->lqi_id.qid_uid = old_id;
2695         qi->lqi_space = -1;
2696         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2697         if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2698                 rc = 0;
2699         if (rc)
2700                 RETURN(rc);
2701
2702         /* block accounting */
2703         qi->lqi_is_blk = true;
2704
2705         /* more blocks for the new id ... */
2706         qi->lqi_id.qid_uid = new_id;
2707         qi->lqi_space      = bspace;
2708         /*
2709          * Credits for the new uid has been reserved, re-use "obj"
2710          * to save credit reservation.
2711          */
2712         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2713         if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2714                 rc = 0;
2715         if (rc)
2716                 RETURN(rc);
2717
2718         /* and finally less blocks for the current uid */
2719         qi->lqi_id.qid_uid = old_id;
2720         qi->lqi_space      = -bspace;
2721         rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2722         if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2723                 rc = 0;
2724
2725         RETURN(rc);
2726 }
2727
2728 static int osd_declare_attr_set(const struct lu_env *env,
2729                                 struct dt_object *dt,
2730                                 const struct lu_attr *attr,
2731                                 struct thandle *handle)
2732 {
2733         struct osd_thandle *oh;
2734         struct osd_object *obj;
2735         qid_t uid;
2736         qid_t gid;
2737         long long bspace;
2738         int rc = 0;
2739         bool enforce;
2740
2741         ENTRY;
2742
2743         LASSERT(dt != NULL);
2744         LASSERT(handle != NULL);
2745
2746         obj = osd_dt_obj(dt);
2747         LASSERT(osd_invariant(obj));
2748
2749         oh = container_of0(handle, struct osd_thandle, ot_super);
2750         LASSERT(oh->ot_handle == NULL);
2751
2752         osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2753                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2754
2755         osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2756                              osd_dto_credits_noquota[DTO_XATTR_SET]);
2757
2758         if (attr == NULL || obj->oo_inode == NULL)
2759                 RETURN(rc);
2760
2761         bspace   = obj->oo_inode->i_blocks << 9;
2762         bspace   = toqb(bspace);
2763
2764         /*
2765          * Changing ownership is always preformed by super user, it should not
2766          * fail with EDQUOT unless required explicitly.
2767          *
2768          * We still need to call the osd_declare_qid() to calculate the journal
2769          * credits for updating quota accounting files and to trigger quota
2770          * space adjustment once the operation is completed.
2771          */
2772         if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2773                 bool ignore_edquot = !(attr->la_flags & LUSTRE_SET_SYNC_FL);
2774
2775                 if (!ignore_edquot)
2776                         CDEBUG(D_QUOTA, "%s: enforce quota on UID %u, GID %u"
2777                                "(the quota space is %lld)\n",
2778                                obj->oo_inode->i_sb->s_id, attr->la_uid,
2779                                attr->la_gid, bspace);
2780
2781                 /* USERQUOTA */
2782                 uid = i_uid_read(obj->oo_inode);
2783                 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2784                 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2785                                           attr->la_uid, enforce, USRQUOTA,
2786                                           true);
2787                 if (rc)
2788                         RETURN(rc);
2789
2790                 gid = i_gid_read(obj->oo_inode);
2791                 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2792                 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2793                                           i_gid_read(obj->oo_inode),
2794                                           attr->la_gid, enforce, GRPQUOTA,
2795                                           ignore_edquot);
2796                 if (rc)
2797                         RETURN(rc);
2798
2799         }
2800 #ifdef HAVE_PROJECT_QUOTA
2801         if (attr->la_valid & LA_PROJID) {
2802                 __u32 projid = i_projid_read(obj->oo_inode);
2803
2804                 enforce = (attr->la_valid & LA_PROJID) &&
2805                                         (attr->la_projid != projid);
2806                 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2807                                           (qid_t)projid, (qid_t)attr->la_projid,
2808                                           enforce, PRJQUOTA, true);
2809                 if (rc)
2810                         RETURN(rc);
2811         }
2812 #endif
2813         RETURN(rc);
2814 }
2815
2816 static int osd_inode_setattr(const struct lu_env *env,
2817                              struct inode *inode, const struct lu_attr *attr)
2818 {
2819         __u64 bits = attr->la_valid;
2820
2821         /* Only allow set size for regular file */
2822         if (!S_ISREG(inode->i_mode))
2823                 bits &= ~(LA_SIZE | LA_BLOCKS);
2824
2825         if (bits == 0)
2826                 return 0;
2827
2828         if (bits & LA_ATIME)
2829                 inode->i_atime  = *osd_inode_time(env, inode, attr->la_atime);
2830         if (bits & LA_CTIME)
2831                 inode->i_ctime  = *osd_inode_time(env, inode, attr->la_ctime);
2832         if (bits & LA_MTIME)
2833                 inode->i_mtime  = *osd_inode_time(env, inode, attr->la_mtime);
2834         if (bits & LA_SIZE) {
2835                 spin_lock(&inode->i_lock);
2836                 LDISKFS_I(inode)->i_disksize = attr->la_size;
2837                 i_size_write(inode, attr->la_size);
2838                 spin_unlock(&inode->i_lock);
2839         }
2840
2841         /*
2842          * OSD should not change "i_blocks" which is used by quota.
2843          * "i_blocks" should be changed by ldiskfs only.
2844          */
2845         if (bits & LA_MODE)
2846                 inode->i_mode = (inode->i_mode & S_IFMT) |
2847                                 (attr->la_mode & ~S_IFMT);
2848         if (bits & LA_UID)
2849                 i_uid_write(inode, attr->la_uid);
2850         if (bits & LA_GID)
2851                 i_gid_write(inode, attr->la_gid);
2852         if (bits & LA_PROJID)
2853                 i_projid_write(inode, attr->la_projid);
2854         if (bits & LA_NLINK)
2855                 set_nlink(inode, attr->la_nlink);
2856         if (bits & LA_RDEV)
2857                 inode->i_rdev = attr->la_rdev;
2858
2859         if (bits & LA_FLAGS) {
2860                 /* always keep S_NOCMTIME */
2861                 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2862                                  S_NOCMTIME;
2863                 /*
2864                  * Ext4 did not transfer inherit flags from
2865                  * @inode->i_flags to raw inode i_flags when writing
2866                  * flags, we do it explictly here.
2867                  */
2868                 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2869                         LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2870                 else
2871                         LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2872         }
2873         return 0;
2874 }
2875
2876 #ifdef HAVE_PROJECT_QUOTA
2877 static int osd_transfer_project(struct inode *inode, __u32 projid)
2878 {
2879         struct super_block *sb = inode->i_sb;
2880         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2881         int err;
2882         kprojid_t kprojid;
2883         struct ldiskfs_iloc iloc;
2884         struct ldiskfs_inode *raw_inode;
2885         struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
2886
2887         if (!ldiskfs_has_feature_project(sb)) {
2888                 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
2889                         == LDISKFS_DEF_PROJID);
2890                 if (projid != LDISKFS_DEF_PROJID)
2891                         return -EOPNOTSUPP;
2892                 else
2893                         return 0;
2894         }
2895
2896         if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
2897                 return -EOPNOTSUPP;
2898
2899         kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
2900         if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
2901                 return 0;
2902
2903         err = ldiskfs_get_inode_loc(inode, &iloc);
2904         if (err)
2905                 return err;
2906
2907         raw_inode = ldiskfs_raw_inode(&iloc);
2908         if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
2909                 err = -EOVERFLOW;
2910                 brelse(iloc.bh);
2911                 return err;
2912         }
2913         brelse(iloc.bh);
2914
2915         dquot_initialize(inode);
2916         transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
2917         if (transfer_to[PRJQUOTA]) {
2918                 err = __dquot_transfer(inode, transfer_to);
2919                 dqput(transfer_to[PRJQUOTA]);
2920                 if (err)
2921                         return err;
2922         }
2923
2924         return err;
2925 }
2926 #endif
2927
2928 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2929 {
2930         int rc;
2931
2932         if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2933             (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2934                 struct iattr iattr;
2935
2936                 ll_vfs_dq_init(inode);
2937                 iattr.ia_valid = 0;
2938                 if (attr->la_valid & LA_UID)
2939                         iattr.ia_valid |= ATTR_UID;
2940                 if (attr->la_valid & LA_GID)
2941                         iattr.ia_valid |= ATTR_GID;
2942                 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2943                 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2944
2945                 rc = ll_vfs_dq_transfer(inode, &iattr);
2946                 if (rc) {
2947                         CERROR("%s: quota transfer failed: rc = %d. Is quota "
2948                                "enforcement enabled on the ldiskfs "
2949                                "filesystem?\n", inode->i_sb->s_id, rc);
2950                         return rc;
2951                 }
2952         }
2953
2954         /* Handle project id transfer here properly */
2955         if (attr->la_valid & LA_PROJID &&
2956             attr->la_projid != i_projid_read(inode)) {
2957 #ifdef HAVE_PROJECT_QUOTA
2958                 rc = osd_transfer_project(inode, attr->la_projid);
2959 #else
2960                 rc = -ENOTSUPP;
2961 #endif
2962                 if (rc) {
2963                         CERROR("%s: quota transfer failed: rc = %d. Is project "
2964                                "enforcement enabled on the ldiskfs "
2965                                "filesystem?\n", inode->i_sb->s_id, rc);
2966                         return rc;
2967                 }
2968         }
2969         return 0;
2970 }
2971
2972 static int osd_attr_set(const struct lu_env *env,
2973                         struct dt_object *dt,
2974                         const struct lu_attr *attr,
2975                         struct thandle *handle)
2976 {
2977         struct osd_object *obj = osd_dt_obj(dt);
2978         struct inode *inode;
2979         int rc;
2980
2981         if (!dt_object_exists(dt))
2982                 return -ENOENT;
2983
2984         LASSERT(handle != NULL);
2985         LASSERT(!dt_object_remote(dt));
2986         LASSERT(osd_invariant(obj));
2987
2988         osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2989
2990         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
2991             !osd_obj2dev(obj)->od_is_ost) {
2992                 struct osd_thread_info *oti = osd_oti_get(env);
2993                 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
2994                 struct lu_fid *fid1 = &oti->oti_fid;
2995                 struct osd_inode_id *id = &oti->oti_id;
2996                 struct iam_path_descr *ipd;
2997                 struct iam_container *bag;
2998                 struct osd_thandle *oh;
2999                 int rc;
3000
3001                 fid_cpu_to_be(fid1, fid0);
3002                 memset(id, 1, sizeof(*id));
3003                 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3004                                   fid0)->oi_dir.od_container;
3005                 ipd = osd_idx_ipd_get(env, bag);
3006                 if (unlikely(ipd == NULL))
3007                         RETURN(-ENOMEM);
3008
3009                 oh = container_of0(handle, struct osd_thandle, ot_super);
3010                 rc = iam_update(oh->ot_handle, bag,
3011                                 (const struct iam_key *)fid1,
3012                                 (const struct iam_rec *)id, ipd);
3013                 osd_ipd_put(env, bag, ipd);
3014                 return(rc > 0 ? 0 : rc);
3015         }
3016
3017         inode = obj->oo_inode;
3018
3019         rc = osd_quota_transfer(inode, attr);
3020         if (rc)
3021                 return rc;
3022
3023         spin_lock(&obj->oo_guard);
3024         rc = osd_inode_setattr(env, inode, attr);
3025         spin_unlock(&obj->oo_guard);
3026         if (rc != 0)
3027                 GOTO(out, rc);
3028
3029         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3030
3031         if (!(attr->la_valid & LA_FLAGS))
3032                 GOTO(out, rc);
3033
3034         /* Let's check if there are extra flags need to be set into LMA */
3035         if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3036                 struct osd_thread_info *info = osd_oti_get(env);
3037                 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3038
3039                 LASSERT(!obj->oo_pfid_in_lma);
3040
3041                 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3042                                  &info->oti_ost_attrs);
3043                 if (rc)
3044                         GOTO(out, rc);
3045
3046                 lma->lma_incompat |=
3047                         lustre_to_lma_flags(attr->la_flags);
3048                 lustre_lma_swab(lma);
3049                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3050                                      lma, sizeof(*lma), XATTR_REPLACE);
3051                 if (rc != 0) {
3052                         struct osd_device *osd = osd_obj2dev(obj);
3053
3054                         CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3055                               osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3056                               lma->lma_incompat, rc);
3057                 } else {
3058                         obj->oo_lma_flags =
3059                                 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3060                 }
3061                 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3062         }
3063 out:
3064         osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3065
3066         return rc;
3067 }
3068
3069 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3070                                            struct osd_object *obj,
3071                                            const char *name, const int namelen)
3072 {
3073         return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3074 }
3075
3076 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3077                       umode_t mode, struct dt_allocation_hint *hint,
3078                       struct thandle *th, struct lu_attr *attr)
3079 {
3080         int result;
3081         struct osd_device *osd = osd_obj2dev(obj);
3082         struct osd_thandle *oth;
3083         struct dt_object *parent = NULL;
3084         struct inode *inode;
3085         uid_t owner[2] = {0, 0};
3086
3087         if (attr->la_valid & LA_UID)
3088                 owner[0] = attr->la_uid;
3089         if (attr->la_valid & LA_GID)
3090                 owner[1] = attr->la_gid;
3091
3092         LINVRNT(osd_invariant(obj));
3093         LASSERT(obj->oo_inode == NULL);
3094         LASSERT(obj->oo_hl_head == NULL);
3095
3096         if (S_ISDIR(mode) && ldiskfs_pdo) {
3097                 obj->oo_hl_head =
3098                         ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3099                 if (obj->oo_hl_head == NULL)
3100                         return -ENOMEM;
3101         }
3102
3103         oth = container_of(th, struct osd_thandle, ot_super);
3104         LASSERT(oth->ot_handle->h_transaction != NULL);
3105
3106         if (hint != NULL && hint->dah_parent != NULL &&
3107             !dt_object_remote(hint->dah_parent))
3108                 parent = hint->dah_parent;
3109
3110         inode = ldiskfs_create_inode(oth->ot_handle,
3111                                      parent ? osd_dt_obj(parent)->oo_inode :
3112                                               osd_sb(osd)->s_root->d_inode,
3113                                      mode, owner);
3114         if (!IS_ERR(inode)) {
3115                 /* Do not update file c/mtime in ldiskfs. */
3116                 inode->i_flags |= S_NOCMTIME;
3117
3118                 /*
3119                  * For new created object, it must be consistent,
3120                  * and it is unnecessary to scrub against it.
3121                  */
3122                 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3123
3124                 obj->oo_inode = inode;
3125                 result = 0;
3126         } else {
3127                 if (obj->oo_hl_head != NULL) {
3128                         ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3129                         obj->oo_hl_head = NULL;
3130                 }
3131                 result = PTR_ERR(inode);
3132         }
3133         LINVRNT(osd_invariant(obj));
3134         return result;
3135 }
3136
3137 enum {
3138         OSD_NAME_LEN = 255
3139 };
3140
3141 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3142                      struct lu_attr *attr,
3143                      struct dt_allocation_hint *hint,
3144                      struct dt_object_format *dof,
3145                      struct thandle *th)
3146 {
3147         int result;
3148         struct osd_thandle *oth;
3149         __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3150
3151         LASSERT(S_ISDIR(attr->la_mode));
3152
3153         oth = container_of(th, struct osd_thandle, ot_super);
3154         LASSERT(oth->ot_handle->h_transaction != NULL);
3155         result = osd_mkfile(info, obj, mode, hint, th, attr);
3156
3157         return result;
3158 }
3159
3160 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3161                         struct lu_attr *attr,
3162                         struct dt_allocation_hint *hint,
3163                         struct dt_object_format *dof,
3164                         struct thandle *th)
3165 {
3166         int result;
3167         struct osd_thandle *oth;
3168         const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3169
3170         __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3171
3172         LASSERT(S_ISREG(attr->la_mode));
3173
3174         oth = container_of(th, struct osd_thandle, ot_super);
3175         LASSERT(oth->ot_handle->h_transaction != NULL);
3176
3177         result = osd_mkfile(info, obj, mode, hint, th, attr);
3178         if (result == 0) {
3179                 LASSERT(obj->oo_inode != NULL);
3180                 if (feat->dif_flags & DT_IND_VARKEY)
3181                         result = iam_lvar_create(obj->oo_inode,
3182                                                  feat->dif_keysize_max,
3183                                                  feat->dif_ptrsize,
3184                                                  feat->dif_recsize_max,
3185                                                  oth->ot_handle);
3186                 else
3187                         result = iam_lfix_create(obj->oo_inode,
3188                                                  feat->dif_keysize_max,
3189                                                  feat->dif_ptrsize,
3190                                                  feat->dif_recsize_max,
3191                                                  oth->ot_handle);
3192         }
3193         return result;
3194 }
3195
3196 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3197                      struct lu_attr *attr,
3198                      struct dt_allocation_hint *hint,
3199                      struct dt_object_format *dof,
3200                      struct thandle *th)
3201 {
3202         LASSERT(S_ISREG(attr->la_mode));
3203         return osd_mkfile(info, obj, (attr->la_mode &
3204                          (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3205                           attr);
3206 }
3207
3208 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3209                      struct lu_attr *attr,
3210                      struct dt_allocation_hint *hint,
3211                      struct dt_object_format *dof,
3212                      struct thandle *th)
3213 {
3214         LASSERT(S_ISLNK(attr->la_mode));
3215         return osd_mkfile(info, obj, (attr->la_mode &
3216                          (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3217                           attr);
3218 }
3219
3220 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3221                      struct lu_attr *attr,
3222                      struct dt_allocation_hint *hint,
3223                      struct dt_object_format *dof,
3224                      struct thandle *th)
3225 {
3226         umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3227         int result;
3228
3229         LINVRNT(osd_invariant(obj));
3230         LASSERT(obj->oo_inode == NULL);
3231         LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3232                 S_ISFIFO(mode) || S_ISSOCK(mode));
3233
3234         result = osd_mkfile(info, obj, mode, hint, th, attr);
3235         if (result == 0) {
3236                 LASSERT(obj->oo_inode != NULL);
3237                 /*
3238                  * This inode should be marked dirty for i_rdev.  Currently
3239                  * that is done in the osd_attr_init().
3240                  */
3241                 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3242                                    attr->la_rdev);
3243         }
3244         LINVRNT(osd_invariant(obj));
3245         return result;
3246 }
3247
3248 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3249                               struct lu_attr *,
3250                               struct dt_allocation_hint *hint,
3251                               struct dt_object_format *dof,
3252                               struct thandle *);
3253
3254 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3255 {
3256         osd_obj_type_f result;
3257
3258         switch (type) {
3259         case DFT_DIR:
3260                 result = osd_mkdir;
3261                 break;
3262         case DFT_REGULAR:
3263                 result = osd_mkreg;
3264                 break;
3265         case DFT_SYM:
3266                 result = osd_mksym;
3267                 break;
3268         case DFT_NODE:
3269                 result = osd_mknod;
3270                 break;
3271         case DFT_INDEX:
3272                 result = osd_mk_index;
3273                 break;
3274
3275         default:
3276                 LBUG();
3277                 break;
3278         }
3279         return result;
3280 }
3281
3282
3283 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3284                         struct dt_object *parent, struct dt_object *child,
3285                         umode_t child_mode)
3286 {
3287         LASSERT(ah);
3288
3289         ah->dah_parent = parent;
3290         ah->dah_mode = child_mode;
3291
3292         if (parent != NULL && !dt_object_remote(parent)) {
3293                 /* will help to find FID->ino at dt_insert("..") */
3294                 struct osd_object *pobj = osd_dt_obj(parent);
3295
3296                 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3297         }
3298 }
3299
3300 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3301                           struct lu_attr *attr, struct dt_object_format *dof)
3302 {
3303         struct inode *inode = obj->oo_inode;
3304         __u64 valid = attr->la_valid;
3305         int result;
3306
3307         attr->la_valid &= ~(LA_TYPE | LA_MODE);
3308
3309         if (dof->dof_type != DFT_NODE)
3310                 attr->la_valid &= ~LA_RDEV;
3311         if ((valid & LA_ATIME) && (attr->la_atime == inode->i_atime.tv_sec))
3312                 attr->la_valid &= ~LA_ATIME;
3313         if ((valid & LA_CTIME) && (attr->la_ctime == inode->i_ctime.tv_sec))
3314                 attr->la_valid &= ~LA_CTIME;
3315         if ((valid & LA_MTIME) && (attr->la_mtime == inode->i_mtime.tv_sec))
3316                 attr->la_valid &= ~LA_MTIME;
3317
3318         result = osd_quota_transfer(inode, attr);
3319         if (result)
3320                 return;
3321
3322         if (attr->la_valid != 0) {
3323                 result = osd_inode_setattr(info->oti_env, inode, attr);
3324                 /*
3325                  * The osd_inode_setattr() should always succeed here.  The
3326                  * only error that could be returned is EDQUOT when we are
3327                  * trying to change the UID or GID of the inode. However, this
3328                  * should not happen since quota enforcement is no longer
3329                  * enabled on ldiskfs (lquota takes care of it).
3330                  */
3331                 LASSERTF(result == 0, "%d\n", result);
3332                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3333         }
3334
3335         attr->la_valid = valid;
3336 }
3337
3338 /**
3339  * Helper function for osd_create()
3340  *
3341  * \retval 0, on success
3342  */
3343 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3344                         struct lu_attr *attr, struct dt_allocation_hint *hint,
3345                         struct dt_object_format *dof, struct thandle *th)
3346 {
3347         int result;
3348         __u32 umask;
3349
3350         osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3351
3352         /* we drop umask so that permissions we pass are not affected */
3353         umask = current->fs->umask;
3354         current->fs->umask = 0;
3355
3356         result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3357                                                   th);
3358         if (likely(obj->oo_inode != NULL)) {
3359                 LASSERT(obj->oo_inode->i_state & I_NEW);
3360
3361                 /*
3362                  * Unlock the inode before attr initialization to avoid
3363                  * unnecessary dqget operations. LU-6378
3364                  */
3365                 unlock_new_inode(obj->oo_inode);
3366         }
3367
3368         if (likely(result == 0)) {
3369                 osd_attr_init(info, obj, attr, dof);
3370                 osd_object_init0(obj);
3371         }
3372
3373         /* restore previous umask value */
3374         current->fs->umask = umask;
3375
3376         osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3377
3378         return result;
3379 }
3380
3381 /**
3382  * Helper function for osd_create()
3383  *
3384  * \retval 0, on success
3385  */
3386 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3387                            const struct lu_fid *fid, struct thandle *th)
3388 {
3389         struct osd_thread_info *info = osd_oti_get(env);
3390         struct osd_inode_id    *id   = &info->oti_id;
3391         struct osd_device      *osd  = osd_obj2dev(obj);
3392         struct osd_thandle     *oh;
3393         int rc;
3394
3395         LASSERT(obj->oo_inode != NULL);
3396
3397         oh = container_of0(th, struct osd_thandle, ot_super);
3398         LASSERT(oh->ot_handle);
3399         osd_trans_exec_op(env, th, OSD_OT_INSERT);
3400
3401         osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3402         rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3403                            OI_CHECK_FLD, NULL);
3404         osd_trans_exec_check(env, th, OSD_OT_INSERT);
3405
3406         return rc;
3407 }
3408
3409 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3410                    u64 seq, struct lu_seq_range *range)
3411 {
3412         struct seq_server_site *ss = osd_seq_site(osd);
3413
3414         if (fid_seq_is_idif(seq)) {
3415                 fld_range_set_ost(range);
3416                 range->lsr_index = idif_ost_idx(seq);
3417                 return 0;
3418         }
3419
3420         if (!fid_seq_in_fldb(seq)) {
3421                 fld_range_set_mdt(range);
3422                 if (ss != NULL)
3423                         /*
3424                          * FIXME: If ss is NULL, it suppose not get lsr_index
3425                          * at all
3426                          */
3427                         range->lsr_index = ss->ss_node_id;
3428                 return 0;
3429         }
3430
3431         LASSERT(ss != NULL);
3432         fld_range_set_any(range);
3433         /* OSD will only do local fld lookup */
3434         return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3435 }
3436
3437 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3438                               struct lu_attr *attr,
3439                               struct dt_allocation_hint *hint,
3440                               struct dt_object_format *dof,
3441                               struct thandle *handle)
3442 {
3443         struct osd_thandle *oh;
3444         int rc;
3445
3446         ENTRY;
3447
3448         LASSERT(handle != NULL);
3449
3450         oh = container_of0(handle, struct osd_thandle, ot_super);
3451         LASSERT(oh->ot_handle == NULL);
3452
3453         /*
3454          * EA object consumes more credits than regular object: osd_mk_index
3455          * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3456          * leaf_node, could involves the block, block bitmap, groups, GDT
3457          * change for each block, so add 4 * 2 credits in that case.
3458          */
3459         osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3460                              osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3461                              (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3462         /*
3463          * Reuse idle OI block may cause additional one OI block
3464          * to be changed.
3465          */
3466         osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3467                              osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3468
3469         /* will help to find FID->ino mapping at dt_insert() */
3470         rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3471                                    osd_dt_obj(dt));
3472         if (rc != 0)
3473                 RETURN(rc);
3474
3475         if (!attr)
3476                 RETURN(0);
3477
3478         rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3479                                    attr->la_projid, 1, oh, osd_dt_obj(dt),
3480                                    NULL, OSD_QID_INODE);
3481         if (rc != 0)
3482                 RETURN(rc);
3483
3484         RETURN(rc);
3485 }
3486
3487 /**
3488  * Called to destroy on-disk representation of the object
3489  *
3490  * Concurrency: must be locked
3491  */
3492 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3493                                struct thandle *th)
3494 {
3495         struct osd_object *obj = osd_dt_obj(dt);
3496         struct inode *inode = obj->oo_inode;
3497         struct osd_thandle *oh;
3498         int rc;
3499
3500         ENTRY;
3501
3502         if (inode == NULL)
3503                 RETURN(-ENOENT);
3504
3505         oh = container_of0(th, struct osd_thandle, ot_super);
3506         LASSERT(oh->ot_handle == NULL);
3507
3508         osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3509                              osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3510
3511         /* For removing agent entry */
3512         if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3513                 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3514
3515         /*
3516          * Recycle idle OI leaf may cause additional three OI blocks
3517          * to be changed.
3518          */
3519         if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3520                 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3521                              osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3522         /* one less inode */
3523         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3524                                    i_projid_read(inode), -1, oh, obj, NULL,
3525                                    OSD_QID_INODE);
3526         if (rc)
3527                 RETURN(rc);
3528         /* data to be truncated */
3529         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3530                                    i_projid_read(inode), 0, oh, obj, NULL,
3531                                    OSD_QID_BLK);
3532         if (rc)
3533                 RETURN(rc);
3534
3535         /*
3536          * will help to find FID->ino when this object is being
3537          * added to PENDING
3538          */
3539         rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3540
3541         RETURN(rc);
3542 }
3543
3544 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3545                        struct thandle *th)
3546 {
3547         const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3548         struct osd_object *obj = osd_dt_obj(dt);
3549         struct inode *inode = obj->oo_inode;
3550         struct osd_device *osd = osd_obj2dev(obj);
3551         struct osd_thandle *oh;
3552         int result;
3553
3554         ENTRY;
3555
3556         oh = container_of0(th, struct osd_thandle, ot_super);
3557         LASSERT(oh->ot_handle);
3558         LASSERT(inode);
3559         LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3560
3561         if (unlikely(fid_is_acct(fid)))
3562                 RETURN(-EPERM);
3563
3564         if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
3565                 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
3566                 if (result != 0)
3567                         CERROR("%s: remove agent entry "DFID": rc = %d\n",
3568                                osd_name(osd), PFID(fid), result);
3569         }
3570
3571         if (S_ISDIR(inode->i_mode)) {
3572                 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
3573                         inode->i_nlink == 2);
3574
3575                 spin_lock(&obj->oo_guard);
3576                 clear_nlink(inode);
3577                 spin_unlock(&obj->oo_guard);
3578                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3579         }
3580
3581         osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3582
3583         ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3584
3585         if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3586                 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3587                                        oh->ot_handle, OI_CHECK_FLD);
3588
3589         osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3590         /* XXX: add to ext3 orphan list */
3591         /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3592
3593         /* not needed in the cache anymore */
3594         set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3595         obj->oo_destroyed = 1;
3596
3597         RETURN(0);
3598 }
3599
3600 /**
3601  * Put the fid into lustre_mdt_attrs, and then place the structure
3602  * inode's ea. This fid should not be altered during the life time
3603  * of the inode.
3604  *
3605  * \retval +ve, on success
3606  * \retval -ve, on error
3607  *
3608  * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3609  */
3610 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3611                    const struct lu_fid *fid, __u32 compat, __u32 incompat)
3612 {
3613         struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3614         struct lustre_mdt_attrs *lma = &loa->loa_lma;
3615         int rc;
3616
3617         ENTRY;
3618
3619         if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3620                 RETURN(0);
3621
3622         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3623                 rc = -ENOMEM;
3624
3625         lustre_loa_init(loa, fid, compat, incompat);
3626         lustre_loa_swab(loa, false);
3627
3628         /*
3629          * For the OST device with 256 bytes inode size by default,
3630          * the PFID EA will be stored together with LMA EA to avoid
3631          * performance trouble. Otherwise the PFID EA can be stored
3632          * independently. LU-8998
3633          */
3634         if ((compat & LMAC_FID_ON_OST) &&
3635             LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3636                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3637                                      sizeof(*loa), XATTR_CREATE);
3638         else
3639                 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3640                                      sizeof(*lma), XATTR_CREATE);
3641         /*
3642          * LMA may already exist, but we need to check that all the
3643          * desired compat/incompat flags have been added.
3644          */
3645         if (unlikely(rc == -EEXIST)) {
3646                 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3647                                      XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3648                 if (rc < 0)
3649                         RETURN(rc);
3650
3651                 if (rc < sizeof(*lma))
3652                         RETURN(-EINVAL);
3653
3654                 lustre_loa_swab(loa, true);
3655                 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3656                     ((compat == 0 && incompat == 0) ||
3657                      (!(~lma->lma_compat & compat) &&
3658                       !(~lma->lma_incompat & incompat))))
3659                         RETURN(0);
3660
3661                 lma->lma_self_fid = *fid;
3662                 lma->lma_compat |= compat;
3663                 lma->lma_incompat |= incompat;
3664                 if (rc == sizeof(*lma)) {
3665                         lustre_lma_swab(lma);
3666                         rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3667                                              sizeof(*lma), XATTR_REPLACE);
3668                 } else {
3669                         lustre_loa_swab(loa, false);
3670                         rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3671                                              sizeof(*loa), XATTR_REPLACE);
3672                 }
3673         }
3674
3675         RETURN(rc);
3676 }
3677
3678 /**
3679  * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3680  * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3681  * To have compatilibility with 1.8 ldiskfs driver we need to have
3682  * magic number at start of fid data.
3683  * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3684  * its inmemory API.
3685  */
3686