Whamcloud - gitweb
729d6dfc21dd69f1f672367021a83be808bb4fec
[fs/lustre-release.git] / lustre / obdfilter / filter.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  *  linux/fs/obdfilter/filter.c
5  *
6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
7  *   Author: Peter Braam <braam@clusterfs.com>
8  *   Author: Andreas Dilger <adilger@clusterfs.com>
9  *
10  *   This file is part of Lustre, http://www.lustre.org.
11  *
12  *   Lustre is free software; you can redistribute it and/or
13  *   modify it under the terms of version 2 of the GNU General Public
14  *   License as published by the Free Software Foundation.
15  *
16  *   Lustre is distributed in the hope that it will be useful,
17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
19  *   GNU General Public License for more details.
20  *
21  *   You should have received a copy of the GNU General Public License
22  *   along with Lustre; if not, write to the Free Software
23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24  */
25
26 /*
27  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
28  *            (which need to get journal_lock, may block if journal full).
29  *
30  * Invariant: Call filter_start_transno() before any journal ops to avoid the
31  *            same deadlock problem.  We can (and want) to get rid of the
32  *            transno sem in favour of the dir/inode i_mutex to avoid single
33  *            threaded operation on the OST.
34  */
35
36 #define DEBUG_SUBSYSTEM S_FILTER
37
38 #ifndef AUTOCONF_INCLUDED
39 #include <linux/config.h>
40 #endif
41 #include <linux/module.h>
42 #include <linux/fs.h>
43 #include <linux/dcache.h>
44 #include <linux/init.h>
45 #include <linux/version.h>
46 #include <linux/sched.h>
47 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
48 # include <linux/mount.h>
49 # include <linux/buffer_head.h>
50 #endif
51
52 #include <obd_class.h>
53 #include <obd_lov.h>
54 #include <lustre_dlm.h>
55 #include <lustre_fsfilt.h>
56 #include <lprocfs_status.h>
57 #include <lustre_log.h>
58 #include <lustre_commit_confd.h>
59 #include <libcfs/list.h>
60 #include <lustre_disk.h>
61 #include <lustre_quota.h>
62 #include <linux/slab.h>
63 #include <lustre_param.h>
64
65 #include "filter_internal.h"
66
67 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
68 #define FILTER_MIN_GROUPS 3
69 static struct lvfs_callback_ops filter_lvfs_ops;
70 cfs_mem_cache_t *ll_fmd_cachep;
71
72 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
73                              void *cb_data, int error)
74 {
75         obd_transno_commit_cb(obd, transno, error);
76 }
77
78 /* Assumes caller has already pushed us into the kernel context. */
79 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
80                           int rc, int force_sync)
81 {
82         struct filter_obd *filter = &exp->exp_obd->u.filter;
83         struct filter_export_data *fed = &exp->exp_filter_data;
84         struct filter_client_data *fcd = fed->fed_fcd;
85         __u64 last_rcvd;
86         loff_t off;
87         int err, log_pri = D_HA;
88
89         /* Propagate error code. */
90         if (rc)
91                 RETURN(rc);
92
93         if (!exp->exp_obd->obd_replayable || oti == NULL)
94                 RETURN(rc);
95
96         /* we don't allocate new transnos for replayed requests */
97         if (oti->oti_transno == 0) {
98                 spin_lock(&filter->fo_translock);
99                 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
100                 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
101                 spin_unlock(&filter->fo_translock);
102                 oti->oti_transno = last_rcvd;
103         } else {
104                 spin_lock(&filter->fo_translock);
105                 last_rcvd = oti->oti_transno;
106                 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
107                         filter->fo_fsd->lsd_last_transno =
108                                 cpu_to_le64(last_rcvd);
109                 spin_unlock(&filter->fo_translock);
110         }
111         fcd->fcd_last_rcvd = cpu_to_le64(last_rcvd);
112
113         /* could get xid from oti, if it's ever needed */
114         fcd->fcd_last_xid = 0;
115
116         off = fed->fed_lr_off;
117         if (off <= 0) {
118                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
119                        fed->fed_lr_idx, fed->fed_lr_off);
120                 err = -EINVAL;
121         } else {
122                 if (!force_sync)
123                         force_sync = fsfilt_add_journal_cb(exp->exp_obd, 
124                                                            last_rcvd,
125                                                            oti->oti_handle,
126                                                            filter_commit_cb,
127                                                            NULL);
128
129                 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
130                                           fcd, sizeof(*fcd), &off,
131                                           force_sync | exp->exp_need_sync);
132                 if (force_sync)
133                         filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
134         }
135         if (err) {
136                 log_pri = D_ERROR;
137                 if (rc == 0)
138                         rc = err;
139         }
140
141         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
142                last_rcvd, fcd->fcd_uuid, fed->fed_lr_idx, err);
143
144         RETURN(rc);
145 }
146
147 void f_dput(struct dentry *dentry)
148 {
149         /* Can't go inside filter_ddelete because it can block */
150         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
151                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
152         LASSERT(atomic_read(&dentry->d_count) > 0);
153
154         dput(dentry);
155 }
156
157 static void init_brw_stats(struct brw_stats *brw_stats)
158 {
159         int i;
160         for (i = 0; i < BRW_LAST; i++)
161                 spin_lock_init(&brw_stats->hist[i].oh_lock);
162 }
163
164 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
165    plus the procfs overhead :( */
166 static int filter_export_stats_init(struct obd_device *obd,
167                                     struct obd_export *exp)
168 {
169         struct filter_export_data *fed = &exp->exp_filter_data;
170         struct proc_dir_entry *brw_entry;
171         int rc, num_stats;
172         ENTRY;
173
174         init_brw_stats(&fed->fed_brw_stats);
175
176         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
177                 /* Self-export gets no proc entry */
178                 RETURN(0);
179
180         rc = lprocfs_exp_setup(exp);
181         if (rc)
182                 RETURN(rc);
183
184         /* Create a per export proc entry for brw_stats */
185         brw_entry = create_proc_entry("brw_stats", 0644, exp->exp_proc);
186         if (brw_entry == NULL)
187                RETURN(-ENOMEM);
188         brw_entry->proc_fops = &filter_per_export_stats_fops;
189         brw_entry->data = fed;
190
191         /* Create a per export proc entry for ops stats */
192         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
193                      LPROC_FILTER_LAST - 1;
194         exp->exp_ops_stats = lprocfs_alloc_stats(num_stats);
195         if (exp->exp_ops_stats == NULL)
196               RETURN(-ENOMEM);
197         lprocfs_init_ops_stats(LPROC_FILTER_LAST, exp->exp_ops_stats);
198         lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_READ_BYTES,
199                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
200         lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_WRITE_BYTES,
201                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
202         lprocfs_register_stats(exp->exp_proc, "stats", exp->exp_ops_stats);
203         RETURN(0);
204 }
205
206 /* Add client data to the FILTER.  We use a bitmap to locate a free space
207  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
208  * Otherwise, we have just read the data from the last_rcvd file and
209  * we know its offset. */
210 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
211                              int cl_idx)
212 {
213         struct filter_obd *filter = &obd->u.filter;
214         struct filter_export_data *fed = &exp->exp_filter_data;
215         unsigned long *bitmap = filter->fo_last_rcvd_slots;
216         int new_client = (cl_idx == -1);
217  
218         ENTRY;
219
220         LASSERT(bitmap != NULL);
221         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
222
223         /* Self-export */
224         if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid) == 0)
225                 RETURN(0);
226
227         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
228          * there's no need for extra complication here
229          */
230         if (new_client) {
231                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
232         repeat:
233                 if (cl_idx >= LR_MAX_CLIENTS) {
234                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
235                                cl_idx);
236                         RETURN(-EOVERFLOW);
237                 }
238                 if (test_and_set_bit(cl_idx, bitmap)) {
239                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
240                                                     cl_idx);
241                         goto repeat;
242                 }
243         } else {
244                 if (test_and_set_bit(cl_idx, bitmap)) {
245                         CERROR("FILTER client %d: bit already set in bitmap!\n",
246                                cl_idx);
247                         LBUG();
248                 }
249         }
250
251         fed->fed_lr_idx = cl_idx;
252         fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
253                 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
254         LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
255
256         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
257                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
258
259         if (new_client) {
260                 struct lvfs_run_ctxt saved;
261                 loff_t off = fed->fed_lr_off;
262                 int rc;
263                 void *handle;
264
265                 CDEBUG(D_INFO, "writing client fcd at idx %u (%llu) (len %u)\n",
266                        fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_fcd));
267
268                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
269                 /* Transaction needed to fix bug 1403 */
270                 handle = fsfilt_start(obd,
271                                       filter->fo_rcvd_filp->f_dentry->d_inode,
272                                       FSFILT_OP_SETATTR, NULL);
273                 if (IS_ERR(handle)) {
274                         rc = PTR_ERR(handle);
275                         CERROR("unable to start transaction: rc %d\n", rc);
276                 } else {
277                         rc = fsfilt_add_journal_cb(obd, 0, handle,
278                                                    target_client_add_cb, exp);
279                         if (rc == 0) {
280                                 spin_lock(&exp->exp_lock);
281                                 exp->exp_need_sync = 1;
282                                 spin_unlock(&exp->exp_lock);
283                         }
284                         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
285                                                  fed->fed_fcd,
286                                                  sizeof(*fed->fed_fcd),
287                                                  &off, rc /* sync if no cb */);
288                         fsfilt_commit(obd,
289                                       filter->fo_rcvd_filp->f_dentry->d_inode,
290                                       handle, 0);
291                 }
292                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
293
294                 if (rc) {
295                         CERROR("error writing %s client idx %u: rc %d\n",
296                                LAST_RCVD, fed->fed_lr_idx, rc);
297                         RETURN(rc);
298                 }
299         }
300         RETURN(0);
301 }
302
303 static int filter_client_free(struct obd_export *exp)
304 {
305         struct filter_export_data *fed = &exp->exp_filter_data;
306         struct filter_obd *filter = &exp->exp_obd->u.filter;
307         struct obd_device *obd = exp->exp_obd;
308         struct filter_client_data zero_fcd;
309         struct lvfs_run_ctxt saved;
310         int rc;
311         loff_t off;
312         ENTRY;
313
314         if (fed->fed_fcd == NULL)
315                 RETURN(0);
316
317         /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
318         if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid ) == 0)
319                 GOTO(free, 0);
320
321         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
322                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
323
324         LASSERT(filter->fo_last_rcvd_slots != NULL);
325
326         off = fed->fed_lr_off;
327
328         /* Don't clear fed_lr_idx here as it is likely also unset.  At worst
329          * we leak a client slot that will be cleaned on the next recovery. */
330         if (off <= 0) {
331                 CERROR("%s: client idx %d has med_off %lld\n",
332                        obd->obd_name, fed->fed_lr_idx, off);
333                 GOTO(free, rc = -EINVAL);
334         }
335
336         /* Clear the bit _after_ zeroing out the client so we don't
337            race with filter_client_add and zero out new clients.*/
338         if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
339                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
340                        fed->fed_lr_idx);
341                 LBUG();
342         }
343
344         if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
345                 memset(&zero_fcd, 0, sizeof zero_fcd);
346                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
347                 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_fcd,
348                                          sizeof(zero_fcd), &off,
349                                          (!exp->exp_libclient ||
350                                           exp->exp_need_sync));
351                 if (rc == 0)
352                         /* update server's transno */
353                         filter_update_server_data(obd, filter->fo_rcvd_filp,
354                                                   filter->fo_fsd,
355                                                   !exp->exp_libclient);
356                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
357
358                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
359                        "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
360                        fed->fed_fcd->fcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
361                        LAST_RCVD, rc);
362         }
363
364         if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
365                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
366                        fed->fed_lr_idx);
367                 LBUG();
368         }
369
370         EXIT;
371 free:
372         OBD_FREE(fed->fed_fcd, sizeof(*fed->fed_fcd));
373         fed->fed_fcd = NULL;
374
375         return 0;
376 }
377
378 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
379 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
380                                          struct filter_mod_data *fmd)
381 {
382         LASSERT_SPIN_LOCKED(&fed->fed_lock);
383         if (--fmd->fmd_refcount == 0) {
384                 /* XXX when we have persistent reservations and the handle
385                  * is stored herein we need to drop it here. */
386                 fed->fed_mod_count--;
387                 list_del(&fmd->fmd_list);
388                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
389         }
390 }
391
392 /* drop fmd reference, free it if last ref */
393 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
394 {
395         struct filter_export_data *fed;
396
397         if (fmd == NULL)
398                 return;
399
400         fed = &exp->exp_filter_data;
401         spin_lock(&fed->fed_lock);
402         filter_fmd_put_nolock(fed, fmd); /* caller reference */
403         spin_unlock(&fed->fed_lock);
404 }
405
406 /* expire entries from the end of the list if there are too many
407  * or they are too old */
408 static void filter_fmd_expire_nolock(struct filter_obd *filter,
409                                      struct filter_export_data *fed,
410                                      struct filter_mod_data *keep)
411 {
412         struct filter_mod_data *fmd, *tmp;
413
414         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
415                 if (fmd == keep)
416                         break;
417
418                 if (time_before(jiffies, fmd->fmd_expire) &&
419                     fed->fed_mod_count < filter->fo_fmd_max_num)
420                         break;
421
422                 list_del_init(&fmd->fmd_list);
423                 filter_fmd_put_nolock(fed, fmd); /* list reference */
424         }
425 }
426
427 void filter_fmd_expire(struct obd_export *exp)
428 {
429         spin_lock(&exp->exp_filter_data.fed_lock);
430         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
431                                  &exp->exp_filter_data, NULL);
432         spin_unlock(&exp->exp_filter_data.fed_lock);
433 }
434
435 /* find specified objid, group in export fmd list.
436  * caller must hold fed_lock and take fmd reference itself */
437 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
438                                                 struct filter_export_data *fed,
439                                                 obd_id objid, obd_gr group)
440 {
441         struct filter_mod_data *found = NULL, *fmd;
442
443         LASSERT_SPIN_LOCKED(&fed->fed_lock);
444
445         list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
446                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
447                         found = fmd;
448                         list_del(&fmd->fmd_list);
449                         list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
450                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
451                         break;
452                 }
453         }
454
455         filter_fmd_expire_nolock(filter, fed, found);
456
457         return found;
458 }
459
460 /* Find fmd based on objid and group, or return NULL if not found. */
461 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
462                                         obd_id objid, obd_gr group)
463 {
464         struct filter_mod_data *fmd;
465
466         spin_lock(&exp->exp_filter_data.fed_lock);
467         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
468                                      &exp->exp_filter_data, objid, group);
469         if (fmd)
470                 fmd->fmd_refcount++;    /* caller reference */
471         spin_unlock(&exp->exp_filter_data.fed_lock);
472
473         return fmd;
474 }
475
476 /* Find fmd based on objid and group, or create a new one if none is found.
477  * It is possible for this function to return NULL under memory pressure,
478  * or if objid = 0 is passed (which will only cause old entries to expire).
479  * Currently this is not fatal because any fmd state is transient and
480  * may also be freed when it gets sufficiently old. */
481 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
482                                        obd_id objid, obd_gr group)
483 {
484         struct filter_export_data *fed = &exp->exp_filter_data;
485         struct filter_mod_data *found = NULL, *fmd_new = NULL;
486
487         OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO, sizeof(*fmd_new));
488
489         spin_lock(&fed->fed_lock);
490         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
491         if (fmd_new) {
492                 if (found == NULL) {
493                         list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
494                         fmd_new->fmd_id = objid;
495                         fmd_new->fmd_gr = group;
496                         fmd_new->fmd_refcount++;   /* list reference */
497                         found = fmd_new;
498                         fed->fed_mod_count++;
499                 } else {
500                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
501                 }
502         }
503         if (found) {
504                 found->fmd_refcount++;          /* caller reference */
505                 found->fmd_expire = jiffies +
506                         exp->exp_obd->u.filter.fo_fmd_max_age;
507         }
508
509         spin_unlock(&fed->fed_lock);
510
511         return found;
512 }
513
514 #ifdef DO_FMD_DROP
515 /* drop fmd list reference so it will disappear when last reference is put.
516  * This isn't so critical because it would in fact only affect the one client
517  * that is doing the unlink and at worst we have an stale entry referencing
518  * an object that should never be used again. */
519 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
520 {
521         struct filter_mod_data *found = NULL;
522
523         spin_lock(&exp->exp_filter_data.fed_lock);
524         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
525         if (found) {
526                 list_del_init(&found->fmd_list);
527                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
528         }
529         spin_unlock(&exp->exp_filter_data.fed_lock);
530 }
531 #else
532 #define filter_fmd_drop(exp, objid, group)
533 #endif
534
535 /* remove all entries from fmd list */
536 static void filter_fmd_cleanup(struct obd_export *exp)
537 {
538         struct filter_export_data *fed = &exp->exp_filter_data;
539         struct filter_mod_data *fmd = NULL, *tmp;
540
541         spin_lock(&fed->fed_lock);
542         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
543                 list_del_init(&fmd->fmd_list);
544                 filter_fmd_put_nolock(fed, fmd);
545         }
546         spin_unlock(&fed->fed_lock);
547 }
548
549 static int filter_init_export(struct obd_export *exp)
550 {
551         spin_lock_init(&exp->exp_filter_data.fed_lock);
552         INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
553         
554         spin_lock(&exp->exp_lock);
555         exp->exp_connecting = 1;
556         spin_unlock(&exp->exp_lock);
557
558         return 0;
559 }
560
561 static int filter_free_server_data(struct filter_obd *filter)
562 {
563         OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
564         filter->fo_fsd = NULL;
565         OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
566         filter->fo_last_rcvd_slots = NULL;
567         return 0;
568 }
569
570 /* assumes caller is already in kernel ctxt */
571 int filter_update_server_data(struct obd_device *obd, struct file *filp,
572                               struct lr_server_data *fsd, int force_sync)
573 {
574         loff_t off = 0;
575         int rc;
576         ENTRY;
577
578         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
579         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
580                le64_to_cpu(fsd->lsd_last_transno));
581         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
582                le64_to_cpu(fsd->lsd_mount_count));
583
584         fsd->lsd_compat14 = fsd->lsd_last_transno;
585         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
586         if (rc)
587                 CERROR("error writing lr_server_data: rc = %d\n", rc);
588
589         RETURN(rc);
590 }
591
592 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
593                              int force_sync)
594 {
595         struct filter_obd *filter = &obd->u.filter;
596         __u64 tmp;
597         loff_t off = 0;
598         int rc;
599         ENTRY;
600
601         if (filter->fo_last_objid_files[group] == NULL) {
602                 CERROR("Object group "LPU64" not fully setup; not updating "
603                        "last_objid\n", group);
604                 RETURN(-EINVAL);
605         }
606
607         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
608                obd->obd_name, group, filter->fo_last_objids[group]);
609
610         tmp = cpu_to_le64(filter->fo_last_objids[group]);
611         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
612                                  &tmp, sizeof(tmp), &off, force_sync);
613         if (rc)
614                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
615                        group, rc);
616         RETURN(rc);
617 }
618 extern int ost_handle(struct ptlrpc_request *req);
619 /* assumes caller has already in kernel ctxt */
620 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
621 {
622         struct filter_obd *filter = &obd->u.filter;
623         struct lr_server_data *fsd;
624         struct filter_client_data *fcd = NULL;
625         struct inode *inode = filp->f_dentry->d_inode;
626         unsigned long last_rcvd_size = i_size_read(inode);
627         __u64 mount_count;
628         int cl_idx;
629         loff_t off = 0;
630         int rc;
631
632         /* ensure padding in the struct is the correct size */
633         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
634                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
635         CLASSERT (offsetof(struct filter_client_data, fcd_padding) +
636                  sizeof(fcd->fcd_padding) == LR_CLIENT_SIZE);
637
638         OBD_ALLOC(fsd, sizeof(*fsd));
639         if (!fsd)
640                 RETURN(-ENOMEM);
641         filter->fo_fsd = fsd;
642
643         OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
644         if (filter->fo_last_rcvd_slots == NULL) {
645                 OBD_FREE(fsd, sizeof(*fsd));
646                 RETURN(-ENOMEM);
647         }
648
649         if (last_rcvd_size == 0) {
650                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
651
652                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
653                 fsd->lsd_last_transno = 0;
654                 mount_count = fsd->lsd_mount_count = 0;
655                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
656                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
657                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
658                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
659                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
660                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
661         } else {
662                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
663                 if (rc) {
664                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
665                                LAST_RCVD, rc);
666                         GOTO(err_fsd, rc);
667                 }
668                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
669                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
670                                            "using the wrong disk %s. Were the "
671                                            "/dev/ assignments rearranged?\n",
672                                            obd->obd_uuid.uuid, fsd->lsd_uuid);
673                         GOTO(err_fsd, rc = -EINVAL);
674                 }
675                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
676                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
677                 /* COMPAT_146 */
678                 /* Assume old last_rcvd format unless I_C_LR is set */
679                 if (!(fsd->lsd_feature_incompat &
680                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
681                         fsd->lsd_last_transno = fsd->lsd_compat14;
682                 /* end COMPAT_146 */
683         }
684
685         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
686                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
687                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
688                        ~FILTER_INCOMPAT_SUPP);
689                 GOTO(err_fsd, rc = -EINVAL);
690         }
691         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
692                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
693                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
694                        ~FILTER_ROCOMPAT_SUPP);
695                 /* Do something like remount filesystem read-only */
696                 GOTO(err_fsd, rc = -EINVAL);
697         }
698
699         CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
700                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
701         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
702                obd->obd_name, mount_count + 1);
703         CDEBUG(D_INODE, "%s: server data size: %u\n",
704                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
705         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
706                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
707         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
708                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
709         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
710                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
711         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
712                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
713                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
714                 le16_to_cpu(fsd->lsd_client_size));
715
716         if (!obd->obd_replayable) {
717                 CWARN("%s: recovery support OFF\n", obd->obd_name);
718                 GOTO(out, rc = 0);
719         }
720
721         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
722              off < last_rcvd_size; cl_idx++) {
723                 __u64 last_rcvd;
724                 struct obd_export *exp;
725                 struct filter_export_data *fed;
726
727                 if (!fcd) {
728                         OBD_ALLOC(fcd, sizeof(*fcd));
729                         if (!fcd)
730                                 GOTO(err_client, rc = -ENOMEM);
731                 }
732
733                 /* Don't assume off is incremented properly by
734                  * fsfilt_read_record(), in case sizeof(*fcd)
735                  * isn't the same as fsd->lsd_client_size.  */
736                 off = le32_to_cpu(fsd->lsd_client_start) +
737                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
738                 rc = fsfilt_read_record(obd, filp, fcd, sizeof(*fcd), &off);
739                 if (rc) {
740                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
741                                LAST_RCVD, cl_idx, off, rc);
742                         break; /* read error shouldn't cause startup to fail */
743                 }
744
745                 if (fcd->fcd_uuid[0] == '\0') {
746                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
747                                cl_idx);
748                         continue;
749                 }
750
751                 last_rcvd = le64_to_cpu(fcd->fcd_last_rcvd);
752
753                 /* These exports are cleaned up by filter_disconnect(), so they
754                  * need to be set up like real exports as filter_connect() does.
755                  */
756                 exp = class_new_export(obd, (struct obd_uuid *)fcd->fcd_uuid);
757
758                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
759                        " srv lr: "LPU64" fcd_group %d\n", fcd->fcd_uuid, cl_idx,
760                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno),
761                        le32_to_cpu(fcd->fcd_group));
762                 if (IS_ERR(exp)) {
763                         if (PTR_ERR(exp) == -EALREADY) {
764                                 /* export already exists, zero out this one */
765                                 CERROR("Zeroing out duplicate export due to "
766                                        "bug 10479.\n");
767                                 fcd->fcd_uuid[0] = '\0';
768                         } else {
769                                 GOTO(err_client, rc = PTR_ERR(exp));
770                         }
771                 } else {
772                         fed = &exp->exp_filter_data;
773                         fed->fed_fcd = fcd;
774                         fed->fed_group = le32_to_cpu(fcd->fcd_group);
775                         filter_export_stats_init(obd, exp);
776                         rc = filter_client_add(obd, exp, cl_idx);
777                         /* can't fail for existing client */
778                         LASSERTF(rc == 0, "rc = %d\n", rc);
779
780                         fcd = NULL;
781                         spin_lock(&exp->exp_lock);
782                         exp->exp_connecting = 0;
783                         exp->exp_in_recovery = 0;
784                         spin_unlock(&exp->exp_lock);
785                         obd->obd_max_recoverable_clients++;
786                         class_export_put(exp);
787                 }
788
789                 /* Need to check last_rcvd even for duplicated exports. */
790                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
791                        cl_idx, last_rcvd);
792
793                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
794                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
795         }
796
797         if (fcd)
798                 OBD_FREE(fcd, sizeof(*fcd));
799
800         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
801
802         target_recovery_init(obd, ost_handle);
803
804 out:
805         filter->fo_mount_count = mount_count + 1;
806         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
807
808         /* save it, so mount count and last_transno is current */
809         rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
810         if (rc)
811                 GOTO(err_client, rc);
812
813         RETURN(0);
814
815 err_client:
816         target_recovery_fini(obd);
817 err_fsd:
818         filter_free_server_data(filter);
819         RETURN(rc);
820 }
821
822 static int filter_cleanup_groups(struct obd_device *obd)
823 {
824         struct filter_obd *filter = &obd->u.filter;
825         struct file *filp;
826         struct dentry *dentry;
827         int i, j;
828         ENTRY;
829
830         if (filter->fo_dentry_O_groups != NULL) {
831                 for (i = 0; i < filter->fo_group_count; i++) {
832                         dentry = filter->fo_dentry_O_groups[i];
833                         if (dentry != NULL)
834                                 f_dput(dentry);
835                 }
836                 OBD_FREE(filter->fo_dentry_O_groups,
837                          filter->fo_group_count *
838                          sizeof(*filter->fo_dentry_O_groups));
839                 filter->fo_dentry_O_groups = NULL;
840         }
841         if (filter->fo_last_objid_files != NULL) {
842                 for (i = 0; i < filter->fo_group_count; i++) {
843                         filp = filter->fo_last_objid_files[i];
844                         if (filp != NULL)
845                                 filp_close(filp, 0);
846                 }
847                 OBD_FREE(filter->fo_last_objid_files,
848                          filter->fo_group_count *
849                          sizeof(*filter->fo_last_objid_files));
850                 filter->fo_last_objid_files = NULL;
851         }
852         if (filter->fo_dentry_O_sub != NULL) {
853                 for (i = 0; i < filter->fo_group_count; i++) {
854                         for (j = 0; j < filter->fo_subdir_count; j++) {
855                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
856                                 if (dentry != NULL)
857                                         f_dput(dentry);
858                         }
859                 }
860                 OBD_FREE(filter->fo_dentry_O_sub,
861                          filter->fo_group_count *
862                          sizeof(*filter->fo_dentry_O_sub));
863                 filter->fo_dentry_O_sub = NULL;
864         }
865         if (filter->fo_last_objids != NULL) {
866                 OBD_FREE(filter->fo_last_objids,
867                          filter->fo_group_count *
868                          sizeof(*filter->fo_last_objids));
869                 filter->fo_last_objids = NULL;
870         }
871         if (filter->fo_dentry_O != NULL) {
872                 f_dput(filter->fo_dentry_O);
873                 filter->fo_dentry_O = NULL;
874         }
875         RETURN(0);
876 }
877
878 static int filter_update_last_group(struct obd_device *obd, int group)
879 {
880         struct filter_obd *filter = &obd->u.filter;
881         struct file *filp = NULL;
882         int last_group = 0, rc;
883         loff_t off = 0;
884         ENTRY;
885
886         if (group <= filter->fo_committed_group)
887                 RETURN(0);
888
889         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
890         if (IS_ERR(filp)) {
891                 rc = PTR_ERR(filp);
892                 filp = NULL;
893                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
894                 GOTO(cleanup, rc);
895         }
896
897         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
898         if (rc) {
899                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
900                 GOTO(cleanup, rc);
901         }
902         LASSERT(off == 0 || last_group >= FILTER_MIN_GROUPS);
903         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
904                obd->obd_name, last_group, group);
905
906         off = 0;
907         last_group = group;
908         /* must be sync: bXXXX */
909         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
910         if (rc) {
911                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
912                 GOTO(cleanup, rc);
913         }
914
915         filter->fo_committed_group = group;
916 cleanup:
917         if (filp)
918                 filp_close(filp, 0);
919         RETURN(rc);
920 }
921
922 static int filter_read_group_internal(struct obd_device *obd, int group,
923                                       int create)
924 {
925         struct filter_obd *filter = &obd->u.filter;
926         __u64 *new_objids = NULL;
927         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
928         struct dentry **new_groups = NULL;
929         struct file **new_files = NULL;
930         struct dentry *dentry;
931         struct file *filp;
932         int old_count = filter->fo_group_count, rc, stage = 0, i;
933         char name[25];
934         __u64 last_objid;
935         loff_t off = 0;
936         int len = group + 1;
937
938         snprintf(name, 24, "%d", group);
939         name[24] = '\0';
940
941         if (!create) {
942                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
943                                            strlen(name));
944                 if (IS_ERR(dentry)) {
945                         CERROR("Cannot lookup expected object group %d: %ld\n",
946                                group, PTR_ERR(dentry));
947                         RETURN(PTR_ERR(dentry));
948                 }
949         } else {
950                 dentry = simple_mkdir(filter->fo_dentry_O, name, 0700, 1);
951                 if (IS_ERR(dentry)) {
952                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
953                                PTR_ERR(dentry));
954                         RETURN(PTR_ERR(dentry));
955                 }
956         }
957         stage = 1;
958
959         snprintf(name, 24, "O/%d/LAST_ID", group);
960         name[24] = '\0';
961         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
962         if (IS_ERR(filp)) {
963                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
964                 GOTO(cleanup, rc = PTR_ERR(filp));
965         }
966         stage = 2;
967
968         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
969         if (rc) {
970                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
971                 GOTO(cleanup, rc);
972         }
973
974         if (filter->fo_subdir_count) {
975                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
976                 if (tmp_subdirs == NULL)
977                         GOTO(cleanup, rc = -ENOMEM);
978                 stage = 3;
979
980                 for (i = 0; i < filter->fo_subdir_count; i++) {
981                         char dir[20];
982                         snprintf(dir, sizeof(dir), "d%u", i);
983
984                         tmp_subdirs->dentry[i] = simple_mkdir(dentry, dir, 0700, 1);
985                         if (IS_ERR(tmp_subdirs->dentry[i])) {
986                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
987                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
988                                        group, dir, rc);
989                                 GOTO(cleanup, rc);
990                         }
991
992                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
993                                tmp_subdirs->dentry[i]);
994                 }
995         }
996
997         /* 'group' is an index; we need an array of length 'group + 1' */
998         if (group + 1 > old_count) {
999                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1000                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1001                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1002                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1003                 stage = 4;
1004                 if (new_objids == NULL || new_subdirs == NULL ||
1005                     new_groups == NULL || new_files == NULL)
1006                         GOTO(cleanup, rc = -ENOMEM);
1007
1008                 if (old_count) {
1009                         memcpy(new_objids, filter->fo_last_objids,
1010                                old_count * sizeof(*new_objids));
1011                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1012                                old_count * sizeof(*new_subdirs));
1013                         memcpy(new_groups, filter->fo_dentry_O_groups,
1014                                old_count * sizeof(*new_groups));
1015                         memcpy(new_files, filter->fo_last_objid_files,
1016                                old_count * sizeof(*new_files));
1017
1018                         OBD_FREE(filter->fo_last_objids,
1019                                  old_count * sizeof(*new_objids));
1020                         OBD_FREE(filter->fo_dentry_O_sub,
1021                                  old_count * sizeof(*new_subdirs));
1022                         OBD_FREE(filter->fo_dentry_O_groups,
1023                                  old_count * sizeof(*new_groups));
1024                         OBD_FREE(filter->fo_last_objid_files,
1025                                  old_count * sizeof(*new_files));
1026                 }
1027                 filter->fo_last_objids = new_objids;
1028                 filter->fo_dentry_O_sub = new_subdirs;
1029                 filter->fo_dentry_O_groups = new_groups;
1030                 filter->fo_last_objid_files = new_files;
1031                 filter->fo_group_count = len;
1032         }
1033
1034         filter->fo_dentry_O_groups[group] = dentry;
1035         filter->fo_last_objid_files[group] = filp;
1036         if (filter->fo_subdir_count) {
1037                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1038                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1039         }
1040
1041         filter_update_last_group(obd, group);
1042
1043         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1044                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1045                 rc = filter_update_last_objid(obd, group, 1);
1046                 RETURN(rc);
1047         }
1048
1049         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1050         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1051                obd->obd_name, group, last_objid);
1052         RETURN(0);
1053  cleanup:
1054         switch (stage) {
1055         case 4:
1056                 if (new_objids != NULL)
1057                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1058                 if (new_subdirs != NULL)
1059                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1060                 if (new_groups != NULL)
1061                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1062                 if (new_files != NULL)
1063                         OBD_FREE(new_files, len * sizeof(*new_files));
1064         case 3:
1065                 if (filter->fo_subdir_count) {
1066                         for (i = 0; i < filter->fo_subdir_count; i++) {
1067                                 if (tmp_subdirs->dentry[i] != NULL)
1068                                         dput(tmp_subdirs->dentry[i]);
1069                         }
1070                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1071                 }
1072         case 2:
1073                 filp_close(filp, 0);
1074         case 1:
1075                 dput(dentry);
1076         }
1077         RETURN(rc);
1078 }
1079
1080 static int filter_read_groups(struct obd_device *obd, int last_group,
1081                               int create)
1082 {
1083         struct filter_obd *filter = &obd->u.filter;
1084         int old_count, group, rc = 0;
1085
1086         down(&filter->fo_init_lock);
1087         old_count = filter->fo_group_count;
1088         for (group = old_count; group <= last_group; group++) {
1089                 if (group == 0)
1090                         continue; /* no group zero */
1091
1092                 rc = filter_read_group_internal(obd, group, create);
1093                 if (rc != 0)
1094                         break;
1095         }
1096         up(&filter->fo_init_lock);
1097         return rc;
1098 }
1099
1100 /* FIXME: object groups */
1101 static int filter_prep_groups(struct obd_device *obd)
1102 {
1103         struct filter_obd *filter = &obd->u.filter;
1104         struct dentry *dentry, *O_dentry;
1105         struct file *filp;
1106         int    last_group, rc = 0, cleanup_phase = 0;
1107         loff_t off = 0;
1108         ENTRY;
1109
1110         O_dentry = simple_mkdir(current->fs->pwd, "O", 0700, 1);
1111         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1112         if (IS_ERR(O_dentry)) {
1113                 rc = PTR_ERR(O_dentry);
1114                 CERROR("cannot open/create O: rc = %d\n", rc);
1115                 GOTO(cleanup, rc);
1116         }
1117         filter->fo_dentry_O = O_dentry;
1118         cleanup_phase = 1; /* O_dentry */
1119
1120         /* Lookup "R" to tell if we're on an old OST FS and need to convert
1121          * from O/R/<dir>/<objid> to O/0/<dir>/<objid>.  This can be removed
1122          * some time post 1.0 when all old-style OSTs have converted along
1123          * with the init_objid hack. */
1124         dentry = ll_lookup_one_len("R", O_dentry, 1);
1125         if (IS_ERR(dentry))
1126                 GOTO(cleanup, rc = PTR_ERR(dentry));
1127         if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1128                 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1129                 ENTRY;
1130
1131                 CWARN("converting OST to new object layout\n");
1132                 if (IS_ERR(O0_dentry)) {
1133                         rc = PTR_ERR(O0_dentry);
1134                         CERROR("error looking up O/0: rc %d\n", rc);
1135                         GOTO(cleanup_R, rc);
1136                 }
1137
1138                 if (O0_dentry->d_inode) {
1139                         CERROR("Both O/R and O/0 exist. Fix manually.\n");
1140                         GOTO(cleanup_O0, rc = -EEXIST);
1141                 }
1142
1143                 LOCK_INODE_MUTEX(O_dentry->d_inode);
1144                 rc = vfs_rename(O_dentry->d_inode, dentry,
1145                                 O_dentry->d_inode, O0_dentry);
1146                 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1147
1148                 if (rc) {
1149                         CERROR("error renaming O/R to O/0: rc %d\n", rc);
1150                         GOTO(cleanup_O0, rc);
1151                 }
1152                 filter->fo_fsd->lsd_feature_incompat |=
1153                         cpu_to_le32(OBD_INCOMPAT_GROUPS);
1154                 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1155                                                filter->fo_fsd, 1);
1156                 GOTO(cleanup_O0, rc);
1157
1158         cleanup_O0:
1159                 f_dput(O0_dentry);
1160         cleanup_R:
1161                 f_dput(dentry);
1162                 if (rc)
1163                         GOTO(cleanup, rc);
1164         } else {
1165                 f_dput(dentry);
1166         }
1167
1168         cleanup_phase = 2; /* groups */
1169
1170         /* we have to initialize all groups before first connections from
1171          * clients because they may send create/destroy for any group -bzzz */
1172         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1173         if (IS_ERR(filp)) {
1174                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1175                 GOTO(cleanup, rc = PTR_ERR(filp));
1176         }
1177         cleanup_phase = 3; /* filp */
1178
1179         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1180         if (rc) {
1181                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1182                 GOTO(cleanup, rc);
1183         }
1184         if (off == 0) {
1185                 last_group = FILTER_MIN_GROUPS;
1186         } else {
1187                 LASSERT(last_group >= FILTER_MIN_GROUPS);
1188         }
1189
1190         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1191               FILTER_MIN_GROUPS, last_group);
1192         filter->fo_committed_group = last_group;
1193         rc = filter_read_groups(obd, last_group, 1);
1194         if (rc)
1195                 GOTO(cleanup, rc);
1196
1197         filp_close(filp, 0);
1198         RETURN(0);
1199
1200  cleanup:
1201         switch (cleanup_phase) {
1202         case 3:
1203                 filp_close(filp, 0);
1204         case 2:
1205                 filter_cleanup_groups(obd);
1206         case 1:
1207                 f_dput(filter->fo_dentry_O);
1208                 filter->fo_dentry_O = NULL;
1209         default:
1210                 break;
1211         }
1212         return rc;
1213
1214 }
1215
1216 /* setup the object store with correct subdirectories */
1217 static int filter_prep(struct obd_device *obd)
1218 {
1219         struct lvfs_run_ctxt saved;
1220         struct filter_obd *filter = &obd->u.filter;
1221         struct file *file;
1222         struct inode *inode;
1223         int rc = 0;
1224         ENTRY;
1225
1226         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1227         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1228         if (!file || IS_ERR(file)) {
1229                 rc = PTR_ERR(file);
1230                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1231                        LAST_RCVD, rc);
1232                 GOTO(out, rc);
1233         }
1234         filter->fo_rcvd_filp = file;
1235         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1236                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1237                        file->f_dentry->d_inode->i_mode);
1238                 GOTO(err_filp, rc = -ENOENT);
1239         }
1240
1241         inode = file->f_dentry->d_parent->d_inode;
1242         /* We use i_op->unlink directly in filter_vfs_unlink() */
1243         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1244                 CERROR("%s: filesystem does not support create/unlink ops\n",
1245                        obd->obd_name);
1246                 GOTO(err_filp, rc = -EOPNOTSUPP);
1247         }
1248
1249         rc = filter_init_server_data(obd, file);
1250         if (rc) {
1251                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1252                 GOTO(err_filp, rc);
1253         }
1254         /* open and create health check io file*/
1255         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1256         if (IS_ERR(file)) {
1257                 rc = PTR_ERR(file);
1258                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1259                        HEALTH_CHECK, rc);
1260                 GOTO(err_server_data, rc);
1261         }
1262         filter->fo_health_check_filp = file;
1263         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1264                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1265                        file->f_dentry->d_inode->i_mode);
1266                 GOTO(err_health_check, rc = -ENOENT);
1267         }
1268         rc = lvfs_check_io_health(obd, file);
1269         if (rc)
1270                 GOTO(err_health_check, rc);
1271
1272         rc = filter_prep_groups(obd);
1273         if (rc)
1274                 GOTO(err_health_check, rc);
1275 out:
1276         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1277
1278         return(rc);
1279
1280 err_health_check:
1281         if (filp_close(filter->fo_health_check_filp, 0))
1282                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1283         filter->fo_health_check_filp = NULL;
1284 err_server_data:
1285         target_recovery_fini(obd);
1286         filter_free_server_data(filter);
1287 err_filp:
1288         if (filp_close(filter->fo_rcvd_filp, 0))
1289                 CERROR("can't close %s after error\n", LAST_RCVD);
1290         filter->fo_rcvd_filp = NULL;
1291         goto out;
1292 }
1293
1294 /* cleanup the filter: write last used object id to status file */
1295 static void filter_post(struct obd_device *obd)
1296 {
1297         struct lvfs_run_ctxt saved;
1298         struct filter_obd *filter = &obd->u.filter;
1299         int rc, i;
1300
1301         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1302          * best to start a transaction with h_sync, because we removed this
1303          * from lastobjid */
1304
1305         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1306         rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1307                                        filter->fo_fsd, 0);
1308         if (rc)
1309                 CERROR("error writing server data: rc = %d\n", rc);
1310
1311         for (i = 1; i < filter->fo_group_count; i++) {
1312                 rc = filter_update_last_objid(obd, i,
1313                                 (i == filter->fo_group_count - 1));
1314                 if (rc)
1315                         CERROR("error writing group %d lastobjid: rc = %d\n",
1316                                i, rc);
1317         }
1318
1319         rc = filp_close(filter->fo_rcvd_filp, 0);
1320         filter->fo_rcvd_filp = NULL;
1321         if (rc)
1322                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1323
1324         rc = filp_close(filter->fo_health_check_filp, 0);
1325         filter->fo_health_check_filp = NULL;
1326         if (rc)
1327                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1328
1329         filter_cleanup_groups(obd);
1330         filter_free_server_data(filter);
1331         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1332
1333         filter_free_capa_keys(filter);
1334         cleanup_capa_hash(filter->fo_capa_hash);
1335 }
1336
1337 static void filter_set_last_id(struct filter_obd *filter,
1338                                obd_id id, obd_gr group)
1339 {
1340         LASSERT(filter->fo_fsd != NULL);
1341         LASSERT(group <= filter->fo_group_count);
1342
1343         spin_lock(&filter->fo_objidlock);
1344         filter->fo_last_objids[group] = id;
1345         spin_unlock(&filter->fo_objidlock);
1346 }
1347
1348 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1349 {
1350         obd_id id;
1351         LASSERT(filter->fo_fsd != NULL);
1352         LASSERT(group <= filter->fo_group_count);
1353
1354         /* FIXME: object groups */
1355         spin_lock(&filter->fo_objidlock);
1356         id = filter->fo_last_objids[group];
1357         spin_unlock(&filter->fo_objidlock);
1358
1359         return id;
1360 }
1361
1362 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1363 {
1364         LOCK_INODE_MUTEX(dparent->d_inode);
1365         return 0;
1366 }
1367
1368 /* We never dget the object parent, so DON'T dput it either */
1369 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1370 {
1371         struct filter_obd *filter = &obd->u.filter;
1372         struct filter_subdirs *subdirs;
1373         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1374
1375         if (group > 0 || filter->fo_subdir_count == 0)
1376                 return filter->fo_dentry_O_groups[group];
1377
1378         subdirs = &filter->fo_dentry_O_sub[group];
1379         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1380 }
1381
1382 /* We never dget the object parent, so DON'T dput it either */
1383 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1384                                   obd_id objid)
1385 {
1386         unsigned long now = jiffies;
1387         struct dentry *dparent = filter_parent(obd, group, objid);
1388         int rc;
1389
1390         if (IS_ERR(dparent))
1391                 return dparent;
1392
1393         rc = filter_lock_dentry(obd, dparent);
1394         fsfilt_check_slow(obd, now, obd_timeout, "parent lock");
1395         return rc ? ERR_PTR(rc) : dparent;
1396 }
1397
1398 /* We never dget the object parent, so DON'T dput it either */
1399 static void filter_parent_unlock(struct dentry *dparent)
1400 {
1401         UNLOCK_INODE_MUTEX(dparent->d_inode);
1402 }
1403
1404 /* How to get files, dentries, inodes from object id's.
1405  *
1406  * If dir_dentry is passed, the caller has already locked the parent
1407  * appropriately for this operation (normally a write lock).  If
1408  * dir_dentry is NULL, we do a read lock while we do the lookup to
1409  * avoid races with create/destroy and such changing the directory
1410  * internal to the filesystem code. */
1411 struct dentry *filter_fid2dentry(struct obd_device *obd,
1412                                  struct dentry *dir_dentry,
1413                                  obd_gr group, obd_id id)
1414 {
1415         struct dentry *dparent = dir_dentry;
1416         struct dentry *dchild;
1417         char name[32];
1418         int len;
1419         ENTRY;
1420
1421         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1422             obd->u.filter.fo_destroys_in_progress == 0) {
1423                 /* don't fail lookups for orphan recovery, it causes
1424                  * later LBUGs when objects still exist during precreate */
1425                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1426                 RETURN(ERR_PTR(-ENOENT));
1427         }
1428         if (id == 0) {
1429                 CERROR("fatal: invalid object id 0\n");
1430                 RETURN(ERR_PTR(-ESTALE));
1431         }
1432
1433         len = sprintf(name, LPU64, id);
1434         if (dir_dentry == NULL) {
1435                 dparent = filter_parent_lock(obd, group, id);
1436                 if (IS_ERR(dparent)) {
1437                         CERROR("%s: error getting object "LPU64":"LPU64
1438                                " parent: rc %ld\n", obd->obd_name,
1439                                id, group, PTR_ERR(dparent));
1440                         RETURN(dparent);
1441                 }
1442         }
1443         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1444                dparent->d_name.len, dparent->d_name.name, name);
1445         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1446         if (dir_dentry == NULL)
1447                 filter_parent_unlock(dparent);
1448         if (IS_ERR(dchild)) {
1449                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1450                        PTR_ERR(dchild));
1451                 RETURN(dchild);
1452         }
1453
1454         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1455                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1456                        obd->obd_name, id, dchild->d_inode->i_ino);
1457                 f_dput(dchild);
1458                 RETURN(ERR_PTR(-ENOENT));
1459         }
1460
1461         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1462                name, dchild, atomic_read(&dchild->d_count));
1463
1464         LASSERT(atomic_read(&dchild->d_count) > 0);
1465
1466         RETURN(dchild);
1467 }
1468
1469 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1470                                   obd_id group)
1471 {
1472         struct lustre_handle lockh;
1473         int flags = LDLM_AST_DISCARD_DATA, rc;
1474         struct ldlm_res_id res_id = { .name = { objid, 0, group, 0} };
1475         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1476
1477         ENTRY;
1478         /* Tell the clients that the object is gone now and that they should
1479          * throw away any cached pages. */
1480         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1481                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1482                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1483                                     &lockh);
1484
1485         /* We only care about the side-effects, just drop the lock. */
1486         if (rc == ELDLM_OK)
1487                 ldlm_lock_decref(&lockh, LCK_PW);
1488
1489         RETURN(rc);
1490 }
1491
1492 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1493  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1494  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1495  * reverse.  Caller must take i_sem before starting the transaction and we
1496  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1497 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry)
1498 {
1499         int rc;
1500         ENTRY;
1501
1502         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1503          * itself we already hold dir->i_mutex for child create/unlink ops */
1504         LASSERT(dentry->d_inode != NULL);
1505         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1506         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1507
1508
1509         /* may_delete() */
1510         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1511                 GOTO(out, rc = -ENOENT);
1512
1513         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1514         if (rc)
1515                 GOTO(out, rc);
1516
1517         if (IS_APPEND(dir))
1518                 GOTO(out, rc = -EPERM);
1519
1520         /* check_sticky() */
1521         if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1522             IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1523                 GOTO(out, rc = -EPERM);
1524
1525         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1526          *       that was done because of journal_start (which is already done
1527          *       here) or some other ordering issue. */
1528         DQUOT_INIT(dir);
1529
1530 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1531         rc = security_inode_unlink(dir, dentry);
1532         if (rc)
1533                 GOTO(out, rc);
1534 #endif
1535
1536         rc = dir->i_op->unlink(dir, dentry);
1537 out:
1538         /* need to drop i_mutex before we lose inode reference */
1539         UNLOCK_INODE_MUTEX(dentry->d_inode);
1540         if (rc == 0)
1541                 d_delete(dentry);
1542
1543         RETURN(rc);
1544 }
1545
1546 /* Caller must hold LCK_PW on parent and push us into kernel context.
1547  * Caller must hold child i_mutex, we drop it always.
1548  * Caller is also required to ensure that dchild->d_inode exists. */
1549 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1550                                    obd_gr group, struct dentry *dparent,
1551                                    struct dentry *dchild)
1552 {
1553         struct inode *inode = dchild->d_inode;
1554         int rc;
1555
1556         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1557                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1558                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1559                        (unsigned long)inode->i_nlink,
1560                        atomic_read(&inode->i_count));
1561         }
1562
1563         rc = filter_vfs_unlink(dparent->d_inode, dchild);
1564         if (rc)
1565                 CERROR("error unlinking objid %.*s: rc %d\n",
1566                        dchild->d_name.len, dchild->d_name.name, rc);
1567         return(rc);
1568 }
1569
1570 static int filter_intent_policy(struct ldlm_namespace *ns,
1571                                 struct ldlm_lock **lockp, void *req_cookie,
1572                                 ldlm_mode_t mode, int flags, void *data)
1573 {
1574         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1575         struct ptlrpc_request *req = req_cookie;
1576         struct ldlm_lock *lock = *lockp, *l = NULL;
1577         struct ldlm_resource *res = lock->l_resource;
1578         ldlm_processing_policy policy;
1579         struct ost_lvb *res_lvb, *reply_lvb;
1580         struct ldlm_reply *rep;
1581         struct list_head *tmp;
1582         ldlm_error_t err;
1583         int rc, tmpflags = 0, only_liblustre = 0;
1584         int repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1585                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1586                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1587         ENTRY;
1588
1589         policy = ldlm_get_processing_policy(res);
1590         LASSERT(policy != NULL);
1591         LASSERT(req != NULL);
1592
1593         rc = lustre_pack_reply(req, 3, repsize, NULL);
1594         if (rc)
1595                 RETURN(req->rq_status = rc);
1596
1597         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1598         LASSERT(rep != NULL);
1599
1600         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1601                                    sizeof(*reply_lvb));
1602         LASSERT(reply_lvb != NULL);
1603
1604         //fixup_handle_for_resent_req(req, lock, &lockh);
1605
1606         /* If we grant any lock at all, it will be a whole-file read lock.
1607          * Call the extent policy function to see if our request can be
1608          * granted, or is blocked. */
1609         lock->l_policy_data.l_extent.start = 0;
1610         lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1611         lock->l_req_mode = LCK_PR;
1612
1613         LASSERT(ns == res->lr_namespace);
1614         lock_res(res);
1615         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1616         check_res_locked(res);
1617
1618         /* FIXME: we should change the policy function slightly, to not make
1619          * this list at all, since we just turn around and free it */
1620         while (!list_empty(&rpc_list)) {
1621                 struct ldlm_lock *wlock =
1622                         list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1623                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1624                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1625                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1626                 list_del_init(&wlock->l_cp_ast);
1627                 LDLM_LOCK_PUT(wlock);
1628         }
1629
1630         /* The lock met with no resistance; we're finished. */
1631         if (rc == LDLM_ITER_CONTINUE) {
1632                 /* do not grant locks to the liblustre clients: they cannot
1633                  * handle ASTs robustly.  We need to do this while still
1634                  * holding ns_lock to avoid the lock remaining on the res_link
1635                  * list (and potentially being added to l_pending_list by an
1636                  * AST) when we are going to drop this lock ASAP. */
1637                 if (lock->l_export->exp_libclient ||
1638                     OBD_FAIL_CHECK(OBD_FAIL_LDLM_GLIMPSE)) {
1639                         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1640                         ldlm_resource_unlink_lock(lock);
1641                         err = ELDLM_LOCK_ABORTED;
1642                 } else {
1643                         err = ELDLM_LOCK_REPLACED;
1644                 }
1645                 unlock_res(res);
1646                 RETURN(err);
1647         }
1648
1649         /* Do not grant any lock, but instead send GL callbacks.  The extent
1650          * policy nicely created a list of all PW locks for us.  We will choose
1651          * the highest of those which are larger than the size in the LVB, if
1652          * any, and perform a glimpse callback. */
1653         res_lvb = res->lr_lvb_data;
1654         LASSERT(res_lvb != NULL);
1655         *reply_lvb = *res_lvb;
1656
1657         list_for_each(tmp, &res->lr_granted) {
1658                 struct ldlm_lock *tmplock =
1659                         list_entry(tmp, struct ldlm_lock, l_res_link);
1660
1661                 if (tmplock->l_granted_mode == LCK_PR)
1662                         continue;
1663                 /*
1664                  * ->ns_lock guarantees that no new locks are granted, and,
1665                  * therefore, that res->lr_lvb_data cannot increase beyond the
1666                  * end of already granted lock. As a result, it is safe to
1667                  * check against "stale" reply_lvb->lvb_size value without
1668                  * res->lr_lvb_sem.
1669                  */
1670                 if (tmplock->l_policy_data.l_extent.end <= reply_lvb->lvb_size)
1671                         continue;
1672
1673                 /* Don't send glimpse ASTs to liblustre clients.  They aren't
1674                  * listening for them, and they do entirely synchronous I/O
1675                  * anyways. */
1676                 if (tmplock->l_export == NULL ||
1677                     tmplock->l_export->exp_libclient == 1) {
1678                         only_liblustre = 1;
1679                         continue;
1680                 }
1681
1682                 if (l == NULL) {
1683                         l = LDLM_LOCK_GET(tmplock);
1684                         continue;
1685                 }
1686
1687                 if (l->l_policy_data.l_extent.start >
1688                     tmplock->l_policy_data.l_extent.start)
1689                         continue;
1690
1691                 LDLM_LOCK_PUT(l);
1692                 l = LDLM_LOCK_GET(tmplock);
1693         }
1694         unlock_res(res);
1695
1696         /* There were no PW locks beyond the size in the LVB; finished. */
1697         if (l == NULL) {
1698                 if (only_liblustre) {
1699                         /* If we discovered a liblustre client with a PW lock,
1700                          * however, the LVB may be out of date!  The LVB is
1701                          * updated only on glimpse (which we don't do for
1702                          * liblustre clients) and cancel (which the client
1703                          * obviously has not yet done).  So if it has written
1704                          * data but kept the lock, the LVB is stale and needs
1705                          * to be updated from disk.
1706                          *
1707                          * Of course, this will all disappear when we switch to
1708                          * taking liblustre locks on the OST. */
1709                         ldlm_res_lvbo_update(res, NULL, 0, 1);
1710                 }
1711                 RETURN(ELDLM_LOCK_ABORTED);
1712         }
1713
1714         /*
1715          * This check is for lock taken in filter_prepare_destroy() that does
1716          * not have l_glimpse_ast set. So the logic is: if there is a lock
1717          * with no l_glimpse_ast set, this object is being destroyed already.
1718          *
1719          * Hence, if you are grabbing DLM locks on the server, always set
1720          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1721          */
1722         if (l->l_glimpse_ast == NULL) {
1723                 /* We are racing with unlink(); just return -ENOENT */
1724                 rep->lock_policy_res1 = -ENOENT;
1725                 goto out;
1726         }
1727
1728         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1729         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1730         /* Update the LVB from disk if the AST failed (this is a legal race) */
1731         /*
1732          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1733          * sending ast is not handled. This can result in lost client writes.
1734          */
1735         if (rc != 0)
1736                 ldlm_res_lvbo_update(res, NULL, 0, 1);
1737
1738         lock_res(res);
1739         *reply_lvb = *res_lvb;
1740         unlock_res(res);
1741
1742  out:
1743         LDLM_LOCK_PUT(l);
1744
1745         RETURN(ELDLM_LOCK_ABORTED);
1746 }
1747
1748 /*
1749  * per-obd_device iobuf pool.
1750  *
1751  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1752  * allocations in write-path has to be minimized (see bug 5137).
1753  *
1754  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1755  * OST threads (see ost_thread_{init,done}()).
1756  *
1757  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1758  * at the OST layer there are only (potentially) multiple obd_device of type
1759  * unknown at the time of OST thread creation.
1760  *
1761  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1762  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1763  * it's very own iobuf.
1764  *
1765  * Functions below
1766  *
1767  *     filter_kiobuf_pool_init()
1768  *
1769  *     filter_kiobuf_pool_done()
1770  *
1771  *     filter_iobuf_get()
1772  *
1773  * operate on this array. They are "generic" in a sense that they don't depend
1774  * on actual type of iobuf's (the latter depending on Linux kernel version).
1775  */
1776
1777 /*
1778  * destroy pool created by filter_iobuf_pool_init
1779  */
1780 static void filter_iobuf_pool_done(struct filter_obd *filter)
1781 {
1782         struct filter_iobuf **pool;
1783         int i;
1784
1785         ENTRY;
1786
1787         pool = filter->fo_iobuf_pool;
1788         if (pool != NULL) {
1789                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1790                         if (pool[i] != NULL)
1791                                 filter_free_iobuf(pool[i]);
1792                 }
1793                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1794                 filter->fo_iobuf_pool = NULL;
1795         }
1796         EXIT;
1797 }
1798
1799 /*
1800  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1801  */
1802 static int filter_iobuf_pool_init(struct filter_obd *filter)
1803 {
1804         void **pool;
1805
1806         ENTRY;
1807
1808
1809         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1810                       GFP_KERNEL);
1811         if (filter->fo_iobuf_pool == NULL)
1812                 RETURN(-ENOMEM);
1813
1814         filter->fo_iobuf_count = OSS_THREADS_MAX;
1815
1816         RETURN(0);
1817 }
1818
1819 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1820  * many threads there will be so we allocate a large empty array and only
1821  * fill in those slots that are actually in use.
1822  * If we haven't allocated a pool entry for this thread before, do so now. */
1823 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1824 {
1825         int thread_id                    = oti ? oti->oti_thread_id : -1;
1826         struct filter_iobuf  *pool       = NULL;
1827         struct filter_iobuf **pool_place = NULL;
1828
1829         if (thread_id >= 0) {
1830                 LASSERT(thread_id < filter->fo_iobuf_count);
1831                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1832         }
1833
1834         if (unlikely(pool == NULL)) {
1835                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1836                                           PTLRPC_MAX_BRW_PAGES);
1837                 if (pool_place != NULL)
1838                         *pool_place = pool;
1839         }
1840
1841         return pool;
1842 }
1843
1844 /* mount the file system (secretly).  lustre_cfg parameters are:
1845  * 1 = device
1846  * 2 = fstype
1847  * 3 = flags: failover=f, failout=n
1848  * 4 = mount options
1849  */
1850 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1851                         void *option)
1852 {
1853         struct filter_obd *filter = &obd->u.filter;
1854         struct vfsmount *mnt;
1855         struct lustre_mount_info *lmi;
1856         struct obd_uuid uuid;
1857         __u8 *uuid_ptr;
1858         char *str, *label;
1859         char ns_name[48];
1860         int rc, i;
1861         ENTRY;
1862
1863         if (lcfg->lcfg_bufcount < 3 ||
1864             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1865             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1866                 RETURN(-EINVAL);
1867
1868         lmi = server_get_mount(obd->obd_name);
1869         if (lmi) {
1870                 /* We already mounted in lustre_fill_super.
1871                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1872                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1873                 mnt = lmi->lmi_mnt;
1874                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1875         } else {
1876                 /* old path - used by lctl */
1877                 CERROR("Using old MDS mount method\n");
1878                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1879                                     MS_NOATIME|MS_NODIRATIME,
1880                                     lustre_cfg_string(lcfg, 1), option);
1881                 if (IS_ERR(mnt)) {
1882                         rc = PTR_ERR(mnt);
1883                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1884                                            lustre_cfg_string(lcfg, 1), rc);
1885                         RETURN(rc);
1886                 }
1887
1888                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1889         }
1890         if (IS_ERR(obd->obd_fsops))
1891                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1892
1893         rc = filter_iobuf_pool_init(filter);
1894         if (rc != 0)
1895                 GOTO(err_ops, rc);
1896
1897         LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1898
1899         /* failover is the default */
1900         obd->obd_replayable = 1;
1901
1902         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1903                 str = lustre_cfg_string(lcfg, 3);
1904                 if (strchr(str, 'n')) {
1905                         CWARN("%s: recovery disabled\n", obd->obd_name);
1906                         obd->obd_replayable = 0;
1907                 }
1908         }
1909
1910         filter->fo_vfsmnt = mnt;
1911         obd->u.obt.obt_sb = mnt->mnt_sb;
1912         filter->fo_fstype = mnt->mnt_sb->s_type->name;
1913         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1914
1915         fsfilt_setup(obd, obd->u.obt.obt_sb);
1916
1917         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1918         obd->obd_lvfs_ctxt.pwdmnt = mnt;
1919         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1920         obd->obd_lvfs_ctxt.fs = get_ds();
1921         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
1922
1923         sema_init(&filter->fo_init_lock, 1);
1924         filter->fo_committed_group = 0;
1925
1926         rc = filter_prep(obd);
1927         if (rc)
1928                 GOTO(err_ops, rc);
1929
1930         filter->fo_destroys_in_progress = 0;
1931         for (i = 0; i < 32; i++)
1932                 sema_init(&filter->fo_create_locks[i], 1);
1933
1934         spin_lock_init(&filter->fo_translock);
1935         spin_lock_init(&filter->fo_objidlock);
1936         INIT_LIST_HEAD(&filter->fo_export_list);
1937         sema_init(&filter->fo_alloc_lock, 1);
1938         init_brw_stats(&filter->fo_filter_stats);
1939         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1940         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
1941         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
1942
1943         INIT_LIST_HEAD(&filter->fo_llog_list);
1944         spin_lock_init(&filter->fo_llog_list_lock);
1945
1946         filter->fo_fl_oss_capa = 0;
1947         INIT_LIST_HEAD(&filter->fo_capa_keys);
1948         filter->fo_capa_hash = init_capa_hash();
1949         if (filter->fo_capa_hash == NULL)
1950                 GOTO(err_ops, rc = -ENOMEM);
1951
1952         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
1953         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER,
1954                                                 LDLM_NAMESPACE_GREEDY);
1955         if (obd->obd_namespace == NULL)
1956                 GOTO(err_post, rc = -ENOMEM);
1957         obd->obd_namespace->ns_lvbp = obd;
1958         obd->obd_namespace->ns_lvbo = &filter_lvbo;
1959         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
1960
1961         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1962                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1963
1964         rc = llog_cat_initialize(obd, NULL, 1, NULL);
1965         if (rc) {
1966                 CERROR("failed to setup llogging subsystems\n");
1967                 GOTO(err_post, rc);
1968         }
1969
1970         rc = lquota_setup(filter_quota_interface_ref, obd);
1971         if (rc)
1972                 GOTO(err_post, rc);
1973
1974         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
1975         if (uuid_ptr != NULL) {
1976                 class_uuid_unparse(uuid_ptr, &uuid);
1977                 str = uuid.uuid;
1978         } else {
1979                 str = "no UUID";
1980         }
1981
1982         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
1983
1984         if (obd->obd_recovering) {
1985                 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
1986                               "recovery until %d %s reconnect, or if no clients"
1987                               " reconnect for %d:%.02d; during that time new "
1988                               "clients will not be allowed to connect. "
1989                               "Recovery progress can be monitored by watching "
1990                               "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
1991                               obd->obd_name, lustre_cfg_string(lcfg, 1),
1992                               label ?: "", label ? "/" : "", str,
1993                               obd->obd_max_recoverable_clients,
1994                               (obd->obd_max_recoverable_clients == 1)
1995                               ? "client" : "clients",
1996                               (int)(OBD_RECOVERY_TIMEOUT) / 60,
1997                               (int)(OBD_RECOVERY_TIMEOUT) % 60,
1998                               obd->obd_name);
1999         } else {
2000                 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2001                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2002                               label ?: "", label ? "/" : "", str,
2003                               obd->obd_replayable ? "enabled" : "disabled");
2004         }
2005
2006         RETURN(0);
2007
2008 err_post:
2009         filter_post(obd);
2010 err_ops:
2011         fsfilt_put_ops(obd->obd_fsops);
2012         filter_iobuf_pool_done(filter);
2013 err_mntput:
2014         server_put_mount(obd->obd_name, mnt);
2015         obd->u.obt.obt_sb = 0;
2016         return rc;
2017 }
2018
2019 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2020 {
2021         struct lprocfs_static_vars lvars;
2022         unsigned long page;
2023         int rc;
2024
2025         CLASSERT(offsetof(struct obd_device, u.obt) ==
2026                  offsetof(struct obd_device, u.filter.fo_obt));
2027
2028         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2029                 RETURN(-EINVAL);
2030
2031         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2032         page = get_zeroed_page(GFP_KERNEL);
2033         if (!page)
2034                 RETURN(-ENOMEM);
2035
2036         /* lprocfs must be setup before the filter so state can be safely added
2037          * to /proc incrementally as the filter is setup */
2038         lprocfs_init_vars(filter, &lvars);
2039         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2040             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2041                 /* Init obdfilter private stats here */
2042                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2043                                      LPROCFS_CNTR_AVGMINMAX,
2044                                      "read_bytes", "bytes");
2045                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2046                                      LPROCFS_CNTR_AVGMINMAX,
2047                                      "write_bytes", "bytes");
2048
2049                 lproc_filter_attach_seqstat(obd);
2050                 obd->obd_proc_exports = proc_mkdir("exports",
2051                                                    obd->obd_proc_entry);
2052         }
2053
2054         memcpy((void *)page, lustre_cfg_buf(lcfg, 4),
2055                LUSTRE_CFG_BUFLEN(lcfg, 4));
2056         rc = filter_common_setup(obd, lcfg, (void *)page);
2057         free_page(page);
2058
2059         if (rc) {
2060                 lprocfs_obd_cleanup(obd);
2061                 lprocfs_free_obd_stats(obd);
2062         }
2063
2064         return rc;
2065 }
2066
2067 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
2068 static struct llog_operations filter_size_orig_logops = {
2069         lop_setup: llog_obd_origin_setup,
2070         lop_cleanup: llog_obd_origin_cleanup,
2071         lop_add: llog_obd_origin_add
2072 };
2073
2074 static int filter_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
2075                             struct obd_device *tgt, int count,
2076                             struct llog_catid *catid,
2077                             struct obd_uuid *uuid)
2078 {
2079         struct llog_ctxt *ctxt;
2080         int rc;
2081         ENTRY;
2082
2083         filter_mds_ost_repl_logops = llog_client_ops;
2084         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2085         filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
2086         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2087
2088         rc = llog_setup(obd, llogs, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2089                         &filter_mds_ost_repl_logops);
2090         if (rc)
2091                 RETURN(rc);
2092
2093         /* FIXME - assign unlink_cb for filter's recovery */
2094         if (!llogs)
2095                 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2096         else
2097                 ctxt = llog_get_context_from_llogs(llogs, LLOG_MDS_OST_REPL_CTXT);
2098
2099         LASSERT(ctxt != NULL);
2100         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2101
2102         rc = llog_setup(obd, llogs, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2103                         &filter_size_orig_logops);
2104         RETURN(rc);
2105 }
2106
2107 static int filter_group_llog_cleanup(struct llog_ctxt *ctxt)
2108 {
2109         int rc = 0;
2110         ENTRY;
2111
2112         if (CTXTP(ctxt, cleanup))
2113                 rc = CTXTP(ctxt, cleanup)(ctxt);
2114
2115         if (ctxt->loc_exp)
2116                 class_export_put(ctxt->loc_exp);
2117         OBD_FREE(ctxt, sizeof(*ctxt));
2118
2119         RETURN(rc);
2120 }
2121
2122 static int filter_group_llog_finish(struct obd_llogs *llogs)
2123 {
2124         struct llog_ctxt *ctxt;
2125         int rc = 0, rc2 = 0;
2126         ENTRY;
2127
2128         ctxt = llog_get_context_from_llogs(llogs, LLOG_MDS_OST_REPL_CTXT);
2129         if (ctxt)
2130                 rc = filter_group_llog_cleanup(ctxt);
2131
2132         ctxt = llog_get_context_from_llogs(llogs, LLOG_SIZE_ORIG_CTXT);
2133         if (ctxt)
2134                 rc2 = filter_group_llog_cleanup(ctxt);
2135         if (!rc)
2136                 rc = rc2;
2137
2138         RETURN(rc);
2139 }
2140
2141 static int filter_llog_finish(struct obd_device *obd, int count)
2142 {
2143         struct llog_ctxt *ctxt;
2144         int rc = 0, rc2 = 0;
2145         ENTRY;
2146
2147         ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2148         if (ctxt)
2149                 rc = llog_cleanup(ctxt);
2150
2151         ctxt = llog_get_context(obd, LLOG_SIZE_ORIG_CTXT);
2152         if (ctxt)
2153                 rc2 = llog_cleanup(ctxt);
2154         if (!rc)
2155                 rc = rc2;
2156
2157         RETURN(rc);
2158 }
2159
2160 struct obd_llogs *filter_grab_llog_for_group(struct obd_device *obd, int group,
2161                                              struct obd_export *export)
2162 {
2163         struct filter_group_llog *fglog, *nlog;
2164         struct filter_obd *filter;
2165         struct llog_ctxt *ctxt;
2166         struct list_head *cur;
2167         int rc;
2168
2169         filter = &obd->u.filter;
2170
2171         spin_lock(&filter->fo_llog_list_lock);
2172         list_for_each(cur, &filter->fo_llog_list) {
2173                 fglog = list_entry(cur, struct filter_group_llog, list);
2174                 if (fglog->group == group) {
2175                         if (!(fglog->exp == NULL || fglog->exp == export || export == NULL))
2176                                 CWARN("%s: export for group %d changes: 0x%p -> 0x%p\n",
2177                                       obd->obd_name, group, fglog->exp, export);
2178                         spin_unlock(&filter->fo_llog_list_lock);
2179                         goto init;
2180                 }
2181         }
2182         spin_unlock(&filter->fo_llog_list_lock);
2183
2184         if (export == NULL)
2185                 RETURN(NULL);
2186
2187         OBD_ALLOC_PTR(fglog);
2188         if (fglog == NULL)
2189                 RETURN(NULL);
2190         fglog->group = group;
2191
2192         OBD_ALLOC_PTR(fglog->llogs);
2193         if (fglog->llogs == NULL) {
2194                 OBD_FREE_PTR(fglog);
2195                 RETURN(NULL);
2196         }
2197
2198         spin_lock(&filter->fo_llog_list_lock);
2199         list_for_each(cur, &filter->fo_llog_list) {
2200                 nlog = list_entry(cur, struct filter_group_llog, list);
2201                 LASSERT(nlog->group != group);
2202         }
2203         list_add(&fglog->list, &filter->fo_llog_list);
2204         spin_unlock(&filter->fo_llog_list_lock);
2205
2206         rc = llog_cat_initialize(obd, fglog->llogs, 1, NULL);
2207         if (rc) {
2208                 OBD_FREE_PTR(fglog->llogs);
2209                 OBD_FREE_PTR(fglog);
2210                 RETURN(NULL);
2211         }
2212
2213 init:
2214         if (export) {
2215                 fglog->exp = export;
2216                 ctxt = llog_get_context_from_llogs(fglog->llogs,
2217                                                LLOG_MDS_OST_REPL_CTXT);
2218                 LASSERT(ctxt != NULL);
2219
2220                 llog_receptor_accept(ctxt, export->exp_imp_reverse);
2221         }
2222         CDEBUG(D_OTHER, "%s: new llog 0x%p for group %u\n",
2223                obd->obd_name, fglog->llogs, group);
2224
2225         RETURN(fglog->llogs);
2226 }
2227
2228 static int filter_llog_connect(struct obd_export *exp,
2229                                struct llogd_conn_body *body)
2230 {
2231         struct obd_device *obd = exp->exp_obd;
2232         struct llog_ctxt *ctxt;
2233         struct obd_llogs *llog;
2234         int rc;
2235         ENTRY;
2236
2237         CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2238                 (unsigned) body->lgdc_logid.lgl_ogr,
2239                 (unsigned) body->lgdc_logid.lgl_oid,
2240                 (unsigned) body->lgdc_logid.lgl_ogen);
2241
2242         llog = filter_grab_llog_for_group(obd, body->lgdc_logid.lgl_ogr, exp);
2243         LASSERT(llog != NULL);
2244         ctxt = llog_get_context_from_llogs(llog, body->lgdc_ctxt_idx);
2245         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2246                  body->lgdc_ctxt_idx);
2247         rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2248                           &body->lgdc_gen, NULL);
2249         if (rc != 0)
2250                 CERROR("failed to connect rc %d idx %d\n", rc,
2251                                 body->lgdc_ctxt_idx);
2252
2253         RETURN(rc);
2254 }
2255
2256 static int filter_llog_preclean (struct obd_device *obd)
2257 {
2258         struct filter_group_llog *log;
2259         struct filter_obd *filter;
2260         int rc = 0;
2261         ENTRY;
2262
2263         filter = &obd->u.filter;
2264         spin_lock(&filter->fo_llog_list_lock);
2265         while (!list_empty(&filter->fo_llog_list)) {
2266                 log = list_entry(filter->fo_llog_list.next,
2267                                  struct filter_group_llog, list);
2268                 list_del(&log->list);
2269                 spin_unlock(&filter->fo_llog_list_lock);
2270
2271                 rc = filter_group_llog_finish(log->llogs);
2272                 if (rc)
2273                         CERROR("failed to cleanup llogging subsystem for %u\n",
2274                                 log->group);
2275                 OBD_FREE_PTR(log->llogs);
2276                 OBD_FREE_PTR(log);
2277                 spin_lock(&filter->fo_llog_list_lock);
2278         }
2279         spin_unlock(&filter->fo_llog_list_lock);
2280
2281         rc = obd_llog_finish(obd, 0);
2282         if (rc)
2283                 CERROR("failed to cleanup llogging subsystem\n");
2284
2285         RETURN(rc);
2286 }
2287
2288 static int filter_precleanup(struct obd_device *obd,
2289                              enum obd_cleanup_stage stage)
2290 {
2291         int rc = 0;
2292         ENTRY;
2293
2294         switch(stage) {
2295         case OBD_CLEANUP_EARLY:
2296                 break;
2297         case OBD_CLEANUP_EXPORTS:
2298                 target_cleanup_recovery(obd);
2299                 break;
2300         case OBD_CLEANUP_SELF_EXP:
2301                 rc = filter_llog_preclean(obd);
2302                 break;
2303         case OBD_CLEANUP_OBD:
2304                 break;
2305         }
2306         RETURN(rc);
2307 }
2308
2309 static int filter_cleanup(struct obd_device *obd)
2310 {
2311         struct filter_obd *filter = &obd->u.filter;
2312         ENTRY;
2313
2314         if (obd->obd_fail)
2315                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2316                               "will be preserved.\n", obd->obd_name);
2317
2318         if (!list_empty(&obd->obd_exports)) {
2319                 CERROR("%s: still has clients!\n", obd->obd_name);
2320                 class_disconnect_exports(obd);
2321                 if (!list_empty(&obd->obd_exports)) {
2322                         CERROR("still has exports after forced cleanup?\n");
2323                         RETURN(-EBUSY);
2324                 }
2325         }
2326
2327         lprocfs_obd_cleanup(obd);
2328         lprocfs_free_obd_stats(obd);
2329         lquota_cleanup(filter_quota_interface_ref, obd);
2330
2331         /* Stop recovery before namespace cleanup. */
2332         target_stop_recovery_thread(obd);
2333         target_cleanup_recovery(obd);
2334
2335         ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2336
2337         if (obd->u.obt.obt_sb == NULL)
2338                 RETURN(0);
2339
2340         filter_post(obd);
2341
2342         shrink_dcache_parent(obd->u.obt.obt_sb->s_root);
2343
2344         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2345
2346         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2347         obd->u.obt.obt_sb = NULL;
2348
2349         fsfilt_put_ops(obd->obd_fsops);
2350
2351         filter_iobuf_pool_done(filter);
2352
2353         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2354
2355         RETURN(0);
2356 }
2357
2358 static int filter_connect_internal(struct obd_export *exp,
2359                                    struct obd_connect_data *data)
2360 {
2361         if (!data)
2362                 RETURN(0);
2363
2364         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2365                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2366                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2367                data->ocd_connect_flags, data->ocd_version,
2368                data->ocd_grant, data->ocd_index);
2369
2370         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2371         exp->exp_connect_flags = data->ocd_connect_flags;
2372         data->ocd_version = LUSTRE_VERSION_CODE;
2373
2374         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2375                 struct filter_export_data *fed = &exp->exp_filter_data;
2376                 obd_size left, want;
2377
2378                 spin_lock(&exp->exp_obd->obd_osfs_lock);
2379                 left = filter_grant_space_left(exp);
2380                 want = data->ocd_grant;
2381                 filter_grant(exp, fed->fed_grant, want, left);
2382                 data->ocd_grant = fed->fed_grant;
2383                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2384
2385                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2386                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2387                        exp->exp_client_uuid.uuid, exp,
2388                        data->ocd_grant, want, left);
2389         }
2390
2391         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2392                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2393                 struct lr_server_data *lsd = filter->fo_fsd;
2394                 int index = le32_to_cpu(lsd->lsd_ost_index);
2395
2396                 if (!(lsd->lsd_feature_compat &
2397                       cpu_to_le32(OBD_COMPAT_OST))) {
2398                         /* this will only happen on the first connect */
2399                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2400                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2401                         filter_update_server_data(exp->exp_obd,
2402                                                   filter->fo_rcvd_filp, lsd, 1);
2403                 } else if (index != data->ocd_index) {
2404                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2405                                            " %u doesn't match actual OST index"
2406                                            " %u in last_rcvd file, bad "
2407                                            "configuration?\n",
2408                                            obd_export_nid2str(exp), index,
2409                                            data->ocd_index);
2410                         RETURN(-EBADF);
2411                 }
2412         }
2413
2414         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2415                 data->ocd_brw_size = 65536;
2416         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2417                 data->ocd_brw_size = min(data->ocd_brw_size,
2418                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2419                 LASSERT(data->ocd_brw_size);
2420         }
2421
2422         /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2423          * FIXME: We don't strictly need the COMPAT flag for that,
2424          * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2425          * FIXME: We needed it for the index, as index 0 is valid. */
2426
2427         RETURN(0);
2428 }
2429
2430 static int filter_reconnect(struct obd_export *exp, struct obd_device *obd,
2431                             struct obd_uuid *cluuid,
2432                             struct obd_connect_data *data)
2433 {
2434         int rc;
2435         ENTRY;
2436
2437         if (exp == NULL || obd == NULL || cluuid == NULL)
2438                 RETURN(-EINVAL);
2439
2440         rc = filter_connect_internal(exp, data);
2441
2442         RETURN(rc);
2443 }
2444
2445 /* nearly identical to mds_connect */
2446 static int filter_connect(const struct lu_env *env,
2447                           struct lustre_handle *conn, struct obd_device *obd,
2448                           struct obd_uuid *cluuid,
2449                           struct obd_connect_data *data)
2450 {
2451         struct lvfs_run_ctxt saved;
2452         struct obd_export *exp;
2453         struct filter_export_data *fed;
2454         struct filter_client_data *fcd = NULL;
2455         __u32 group;
2456         int rc;
2457         ENTRY;
2458
2459         if (conn == NULL || obd == NULL || cluuid == NULL)
2460                 RETURN(-EINVAL);
2461
2462         rc = class_connect(conn, obd, cluuid);
2463         if (rc)
2464                 RETURN(rc);
2465         exp = class_conn2export(conn);
2466         LASSERT(exp != NULL);
2467
2468         fed = &exp->exp_filter_data;
2469
2470         rc = filter_connect_internal(exp, data);
2471         if (rc)
2472                 GOTO(cleanup, rc);
2473
2474         filter_export_stats_init(obd, exp);
2475         group = data->ocd_group;
2476         if (obd->obd_replayable) {
2477                 OBD_ALLOC(fcd, sizeof(*fcd));
2478                 if (!fcd) {
2479                         CERROR("filter: out of memory for client data\n");
2480                         GOTO(cleanup, rc = -ENOMEM);
2481                 }
2482
2483                 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
2484                 fed->fed_fcd = fcd;
2485                 fed->fed_fcd->fcd_group = group;
2486                 rc = filter_client_add(obd, exp, -1);
2487                 if (rc)
2488                         GOTO(cleanup, rc);
2489         }
2490         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2491                obd->obd_name, exp->exp_handle.h_cookie, group);
2492         if (group == 0)
2493                 GOTO(cleanup, rc);
2494
2495         if (fed->fed_group != 0 && fed->fed_group != group) {
2496                 CERROR("!!! This export (nid %s) used object group %d "
2497                        "earlier; now it's trying to use group %d!  This could "
2498                        "be a bug in the MDS.  Tell CFS.\n",
2499                        obd_export_nid2str(exp), fed->fed_group, group);
2500                 GOTO(cleanup, rc = -EPROTO);
2501         }
2502         fed->fed_group = group;
2503
2504         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2505         rc = filter_read_groups(obd, group, 1);
2506         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2507         if (rc != 0) {
2508                 CERROR("can't read group %u\n", group);
2509                 GOTO(cleanup, rc);
2510         }
2511
2512         GOTO(cleanup, rc);
2513
2514 cleanup:
2515         if (rc) {
2516                 if (fcd) {
2517                         OBD_FREE(fcd, sizeof(*fcd));
2518                         fed->fed_fcd = NULL;
2519                 }
2520                 class_disconnect(exp);
2521         } else {
2522                 class_export_put(exp);
2523         }
2524
2525         RETURN(rc);
2526 }
2527
2528 /* Do extra sanity checks for grant accounting.  We do this at connect,
2529  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2530  * always get rid of it or turn it off when we know accounting is good. */
2531 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2532 {
2533         struct filter_export_data *fed;
2534         struct obd_export *exp;
2535         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2536         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2537         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2538
2539         if (list_empty(&obd->obd_exports))
2540                 return;
2541
2542         /* We don't want to do this for large machines that do lots of
2543            mounts or unmounts.  It burns... */
2544         if (obd->obd_num_exports > 100)
2545                 return;
2546
2547         spin_lock(&obd->obd_osfs_lock);
2548         spin_lock(&obd->obd_dev_lock);
2549         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2550                 int error = 0;
2551                 fed = &exp->exp_filter_data;
2552                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2553                     fed->fed_dirty < 0)
2554                         error = 1;
2555                 if (maxsize > 0) { /* we may not have done a statfs yet */
2556                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2557                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2558                                  exp->exp_client_uuid.uuid, exp,
2559                                  fed->fed_grant, fed->fed_pending, maxsize);
2560                         LASSERTF(fed->fed_dirty <= maxsize,
2561                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2562                                  exp->exp_client_uuid.uuid, exp,
2563                                  fed->fed_dirty, maxsize);
2564                 }
2565                 if (error)
2566                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2567                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2568                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2569                 else
2570                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2571                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2572                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2573                 tot_granted += fed->fed_grant + fed->fed_pending;
2574                 tot_pending += fed->fed_pending;
2575                 tot_dirty += fed->fed_dirty;
2576         }
2577         fo_tot_granted = obd->u.filter.fo_tot_granted;
2578         fo_tot_pending = obd->u.filter.fo_tot_pending;
2579         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2580         spin_unlock(&obd->obd_dev_lock);
2581         spin_unlock(&obd->obd_osfs_lock);
2582
2583         /* Do these assertions outside the spinlocks so we don't kill system */
2584         if (tot_granted != fo_tot_granted)
2585                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2586                        func, tot_granted, fo_tot_granted);
2587         if (tot_pending != fo_tot_pending)
2588                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2589                        func, tot_pending, fo_tot_pending);
2590         if (tot_dirty != fo_tot_dirty)
2591                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2592                        func, tot_dirty, fo_tot_dirty);
2593         if (tot_pending > tot_granted)
2594                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2595                        func, tot_pending, tot_granted);
2596         if (tot_granted > maxsize)
2597                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2598                        func, tot_granted, maxsize);
2599         if (tot_dirty > maxsize)
2600                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2601                        func, tot_dirty, maxsize);
2602 }
2603
2604 /* Remove this client from the grant accounting totals.  We also remove
2605  * the export from the obd device under the osfs and dev locks to ensure
2606  * that the filter_grant_sanity_check() calculations are always valid.
2607  * The client should do something similar when it invalidates its import. */
2608 static void filter_grant_discard(struct obd_export *exp)
2609 {
2610         struct obd_device *obd = exp->exp_obd;
2611         struct filter_obd *filter = &obd->u.filter;
2612         struct filter_export_data *fed = &exp->exp_filter_data;
2613
2614         spin_lock(&obd->obd_osfs_lock);
2615         spin_lock(&obd->obd_dev_lock);
2616         list_del_init(&exp->exp_obd_chain);
2617         spin_unlock(&obd->obd_dev_lock);
2618
2619         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2620                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2621                  obd->obd_name, filter->fo_tot_granted,
2622                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2623         filter->fo_tot_granted -= fed->fed_grant;
2624         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2625                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2626                  obd->obd_name, filter->fo_tot_pending,
2627                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2628         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2629         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2630                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2631                  obd->obd_name, filter->fo_tot_dirty,
2632                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2633         filter->fo_tot_dirty -= fed->fed_dirty;
2634         fed->fed_dirty = 0;
2635         fed->fed_grant = 0;
2636
2637         spin_unlock(&obd->obd_osfs_lock);
2638 }
2639
2640 static int filter_destroy_export(struct obd_export *exp)
2641 {
2642         ENTRY;
2643
2644         if (exp->exp_filter_data.fed_pending)
2645                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2646                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2647                        exp, exp->exp_filter_data.fed_pending);
2648
2649         target_destroy_export(exp);
2650
2651         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2652                 RETURN(0);
2653
2654         lprocfs_exp_cleanup(exp);
2655
2656         if (exp->exp_obd->obd_replayable)
2657                 filter_client_free(exp);
2658         else
2659                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2660
2661         filter_grant_discard(exp);
2662         filter_fmd_cleanup(exp);
2663
2664         if (!(exp->exp_flags & OBD_OPT_FORCE))
2665                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2666
2667         RETURN(0);
2668 }
2669
2670 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2671 {
2672         struct filter_group_llog *fglog, *nlog;
2673         struct filter_obd *filter;
2674         int worked = 0, group;
2675         struct llog_ctxt *ctxt;
2676         ENTRY;
2677
2678         filter = &obd->u.filter;
2679
2680         /* we can't sync log holding spinlock. also, we do not want to get
2681          * into livelock. so we do following: loop over MDS's exports in
2682          * group order and skip already synced llogs -bzzz */
2683         do {
2684                 /* look for group with min. number, but > worked */
2685                 fglog = NULL;
2686                 group = 1 << 30;
2687                 spin_lock(&filter->fo_llog_list_lock);
2688                 list_for_each_entry(nlog, &filter->fo_llog_list, list) {
2689                         if (nlog->group <= worked) {
2690                                 /* this group is already synced */
2691                                 continue;
2692                         }
2693                         if (group < nlog->group) {
2694                                 /* we have group with smaller number to sync */
2695                                 continue;
2696                         }
2697                         /* store current minimal group */
2698                         fglog = nlog;
2699                         group = nlog->group;
2700                 }
2701                 spin_unlock(&filter->fo_llog_list_lock);
2702
2703                 if (fglog == NULL)
2704                         break;
2705
2706                 worked = fglog->group;
2707                 if (fglog->exp && (dexp == fglog->exp || dexp == NULL)) {
2708                         ctxt = llog_get_context_from_llogs(fglog->llogs,
2709                                                 LLOG_MDS_OST_REPL_CTXT);
2710                         LASSERT(ctxt != NULL);
2711                         llog_sync(ctxt, fglog->exp);
2712                 }
2713         } while (fglog != NULL);
2714 }
2715
2716 /* also incredibly similar to mds_disconnect */
2717 static int filter_disconnect(struct obd_export *exp)
2718 {
2719         struct obd_device *obd = exp->exp_obd;
2720         int rc;
2721         ENTRY;
2722
2723         LASSERT(exp);
2724         class_export_get(exp);
2725
2726         if (!(exp->exp_flags & OBD_OPT_FORCE))
2727                 filter_grant_sanity_check(obd, __FUNCTION__);
2728         filter_grant_discard(exp);
2729
2730         /* Disconnect early so that clients can't keep using export */
2731         rc = class_disconnect(exp);
2732         if (exp->exp_obd->obd_namespace != NULL)
2733                 ldlm_cancel_locks_for_export(exp);
2734
2735         fsfilt_sync(obd, obd->u.obt.obt_sb);
2736
2737         /* flush any remaining cancel messages out to the target */
2738         filter_sync_llogs(obd, exp);
2739         class_export_put(exp);
2740         RETURN(rc);
2741 }
2742
2743 /* reverse import is changed, sync all cancels */
2744 static void filter_revimp_update(struct obd_export *exp)
2745 {
2746         ENTRY;
2747
2748         LASSERT(exp);
2749         class_export_get(exp);
2750
2751         /* flush any remaining cancel messages out to the target */
2752         filter_sync_llogs(exp->exp_obd, exp);
2753         class_export_put(exp);
2754         EXIT;
2755 }
2756
2757 static int filter_ping(struct obd_export *exp)
2758 {
2759         filter_fmd_expire(exp);
2760
2761         return 0;
2762 }
2763
2764 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2765                                   const char *what, int quiet)
2766 {
2767         struct dentry *dchild = NULL;
2768         obd_gr group = 0;
2769
2770         if (oa->o_valid & OBD_MD_FLGROUP)
2771                 group = oa->o_gr;
2772
2773         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
2774
2775         if (IS_ERR(dchild)) {
2776                 CERROR("%s error looking up object: "LPU64"\n",
2777                        what, oa->o_id);
2778                 RETURN(dchild);
2779         }
2780
2781         if (dchild->d_inode == NULL) {
2782                 if (!quiet)
2783                         CERROR("%s: %s on non-existent object: "LPU64"\n",
2784                                obd->obd_name, what, oa->o_id);
2785                 f_dput(dchild);
2786                 RETURN(ERR_PTR(-ENOENT));
2787         }
2788
2789         return dchild;
2790 }
2791
2792 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2793 {
2794         struct dentry *dentry = NULL;
2795         struct obd_device *obd;
2796         int rc = 0;
2797         ENTRY;
2798
2799         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
2800                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
2801         if (rc)
2802                 RETURN(rc);
2803
2804         obd = class_exp2obd(exp);
2805         if (obd == NULL) {
2806                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
2807                 RETURN(-EINVAL);
2808         }
2809
2810         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
2811         if (IS_ERR(dentry))
2812                 RETURN(PTR_ERR(dentry));
2813
2814         /* Limit the valid bits in the return data to what we actually use */
2815         oinfo->oi_oa->o_valid = OBD_MD_FLID;
2816         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
2817
2818         f_dput(dentry);
2819         RETURN(rc);
2820 }
2821
2822 /* this should be enabled/disabled in condition to enabled/disabled large
2823  * inodes (fast EAs) in backing store FS. */
2824 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
2825                         void *handle, struct obdo *oa)
2826 {
2827         struct obd_device *obd = exp->exp_obd;
2828         int rc = 0;
2829         ENTRY;
2830
2831         if (oa->o_valid & OBD_MD_FLFID) {
2832                 struct filter_fid ff;
2833
2834                 if (!(oa->o_valid & OBD_MD_FLGROUP))
2835                         oa->o_gr = 0;
2836                 /* packing fid and converting it to LE for storing into EA.
2837                  * Here ->o_stripe_idx should be filled by LOV and rest of
2838                  * fields - by client. */
2839                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
2840                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
2841                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
2842                 ff.ff_objid = cpu_to_le64(oa->o_id);
2843                 ff.ff_group = cpu_to_le64(oa->o_gr);
2844
2845                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
2846                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
2847                        oa->o_generation, oa->o_id, oa->o_gr);
2848
2849                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
2850                 if (rc)
2851                         CERROR("store fid in object failed! rc: %d\n", rc);
2852         } else {
2853                 CDEBUG(D_HA, "OSS object without fid info!\n");
2854         }
2855
2856         RETURN(rc);
2857 }
2858
2859 /* this is called from filter_truncate() until we have filter_punch() */
2860 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
2861                             struct obdo *oa, struct obd_trans_info *oti)
2862 {
2863         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
2864         struct llog_cookie *fcc = NULL;
2865         struct filter_obd *filter;
2866         int rc, err, locked = 0, sync = 0;
2867         unsigned int ia_valid;
2868         struct inode *inode;
2869         struct iattr iattr;
2870         void *handle;
2871         ENTRY;
2872
2873         LASSERT(dentry != NULL);
2874         LASSERT(!IS_ERR(dentry));
2875
2876         inode = dentry->d_inode;
2877         LASSERT(inode != NULL);
2878
2879         filter = &exp->exp_obd->u.filter;
2880         iattr_from_obdo(&iattr, oa, oa->o_valid);
2881         ia_valid = iattr.ia_valid;
2882
2883         if (oa->o_valid & OBD_MD_FLCOOKIE) {
2884                 OBD_ALLOC(fcc, sizeof(*fcc));
2885                 if (fcc != NULL)
2886                         memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
2887         }
2888
2889         if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
2890                 DQUOT_INIT(inode);
2891                 LOCK_INODE_MUTEX(inode);
2892                 locked = 1;
2893         }
2894
2895         /* If the inode still has SUID+SGID bits set (see filter_precreate())
2896          * then we will accept the UID+GID sent by the client during write for
2897          * initializing the ownership of this inode.  We only allow this to
2898          * happen once so clear these bits in setattr. In 2.6 kernels it is
2899          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
2900          * the flags that are actually being set. */
2901         if (ia_valid & (ATTR_UID | ATTR_GID)) {
2902                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
2903                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
2904
2905                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
2906                         if (!(ia_valid & ATTR_MODE)) {
2907                                 iattr.ia_mode = inode->i_mode;
2908                                 iattr.ia_valid |= ATTR_MODE;
2909                         }
2910                         iattr.ia_mode &= ~S_ISUID;
2911                 }
2912                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
2913                         if (!(iattr.ia_valid & ATTR_MODE)) {
2914                                 iattr.ia_mode = inode->i_mode;
2915                                 iattr.ia_valid |= ATTR_MODE;
2916                         }
2917                         iattr.ia_mode &= ~S_ISGID;
2918                 }
2919
2920                 orig_ids[USRQUOTA] = inode->i_uid;
2921                 orig_ids[GRPQUOTA] = inode->i_gid;
2922                 handle = fsfilt_start_log(exp->exp_obd, inode,
2923                                           FSFILT_OP_SETATTR, oti, 1);
2924                 if (IS_ERR(handle))
2925                         GOTO(out_unlock, rc = PTR_ERR(handle));
2926
2927                 /* update inode EA only once when inode is suid bit marked. As
2928                  * on 2.6.x UID and GID may be set separately, we check here
2929                  * only one of them to avoid double setting. */
2930                 if (inode->i_mode & S_ISUID)
2931                         filter_update_fidea(exp, inode, handle, oa);
2932         } else {
2933                 handle = fsfilt_start(exp->exp_obd, inode,
2934                                       FSFILT_OP_SETATTR, oti);
2935                 if (IS_ERR(handle))
2936                         GOTO(out_unlock, rc = PTR_ERR(handle));
2937         }
2938         if (oa->o_valid & OBD_MD_FLFLAGS) {
2939                 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
2940                                       EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
2941         } else {
2942                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
2943                 if (fcc != NULL)
2944                         /* set cancel cookie callback function */
2945                         sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
2946                                                      filter_cancel_cookies_cb,
2947                                                      fcc);
2948         }
2949
2950         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
2951                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
2952
2953         /* The truncate might have used up our transaction credits.  Make
2954          * sure we have one left for the last_rcvd update. */
2955         err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
2956
2957         rc = filter_finish_transno(exp, oti, rc, sync);
2958         if (sync) {
2959                 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
2960                 fcc = NULL;
2961         }
2962
2963         err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
2964         if (err) {
2965                 CERROR("error on commit, err = %d\n", err);
2966                 if (!rc)
2967                         rc = err;
2968         } else {
2969                 fcc = NULL;
2970         }
2971
2972         if (locked) {
2973                 /* Let's flush truncated page on disk immediately, then we can
2974                  * avoid need to search for page aliases before directio writes
2975                  * and this sort of stuff at expense of somewhat slower
2976                  * truncates not on a page boundary. I believe this is the only
2977                  * place in filter code that can lead to pages getting to
2978                  * pagecache so far. */
2979                 filter_clear_truncated_page(inode);
2980                 UNLOCK_INODE_MUTEX(inode);
2981                 locked = 0;
2982         }
2983
2984         EXIT;
2985 out_unlock:
2986         if (locked)
2987                 UNLOCK_INODE_MUTEX(inode);
2988
2989         if (fcc)
2990                 OBD_FREE(fcc, sizeof(*fcc));
2991
2992         /* trigger quota release */
2993         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
2994                 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
2995                 int rc2 = lquota_adjust(filter_quota_interface_ref,
2996                                         exp->exp_obd, cur_ids,
2997                                         orig_ids, rc, FSFILT_OP_SETATTR);
2998                 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
2999                        "filter adjust qunit. (rc:%d)\n", rc2);
3000         }
3001         return rc;
3002 }
3003
3004 /* this is called from filter_truncate() until we have filter_punch() */
3005 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3006                    struct obd_trans_info *oti)
3007 {
3008         struct ldlm_res_id res_id = { .name = { oinfo->oi_oa->o_id, 0,
3009                                                 oinfo->oi_oa->o_gr, 0 } };
3010         struct filter_mod_data *fmd;
3011         struct lvfs_run_ctxt saved;
3012         struct filter_obd *filter;
3013         struct ldlm_resource *res;
3014         struct dentry *dentry;
3015         int rc;
3016         ENTRY;
3017
3018         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3019                               oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
3020         if (rc)
3021                 RETURN(rc);
3022
3023         dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
3024                                     __FUNCTION__, 1);
3025         if (IS_ERR(dentry))
3026                 RETURN(PTR_ERR(dentry));
3027
3028         filter = &exp->exp_obd->u.filter;
3029         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3030         lock_kernel();
3031
3032         if (oinfo->oi_oa->o_valid &
3033             (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3034                 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
3035                 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3036                         fmd->fmd_mactime_xid = oti->oti_xid;
3037                 filter_fmd_put(exp, fmd);
3038         }
3039
3040         /* setting objects attributes (including owner/group) */
3041         rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
3042         if (rc)
3043                 GOTO(out_unlock, rc);
3044
3045         res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3046                                 &res_id, LDLM_EXTENT, 0);
3047
3048         if (res != NULL) {
3049                 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3050                 ldlm_resource_putref(res);
3051         }
3052
3053         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3054
3055         /* Quota release need uid/gid info */
3056         obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
3057                         FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3058
3059         EXIT;
3060 out_unlock:
3061         unlock_kernel();
3062         f_dput(dentry);
3063         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3064         return rc;
3065 }
3066
3067 /* XXX identical to osc_unpackmd */
3068 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3069                            struct lov_mds_md *lmm, int lmm_bytes)
3070 {
3071         int lsm_size;
3072         ENTRY;
3073
3074         if (lmm != NULL) {
3075                 if (lmm_bytes < sizeof (*lmm)) {
3076                         CERROR("lov_mds_md too small: %d, need %d\n",
3077                                lmm_bytes, (int)sizeof(*lmm));
3078                         RETURN(-EINVAL);
3079                 }
3080                 /* XXX LOV_MAGIC etc check? */
3081
3082                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3083                         CERROR("lov_mds_md: zero lmm_object_id\n");
3084                         RETURN(-EINVAL);
3085                 }
3086         }
3087
3088         lsm_size = lov_stripe_md_size(1);
3089         if (lsmp == NULL)
3090                 RETURN(lsm_size);
3091
3092         if (*lsmp != NULL && lmm == NULL) {
3093                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3094                 OBD_FREE(*lsmp, lsm_size);
3095                 *lsmp = NULL;
3096                 RETURN(0);
3097         }
3098
3099         if (*lsmp == NULL) {
3100                 OBD_ALLOC(*lsmp, lsm_size);
3101                 if (*lsmp == NULL)
3102                         RETURN(-ENOMEM);
3103
3104                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3105                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3106                         OBD_FREE(*lsmp, lsm_size);
3107                         RETURN(-ENOMEM);
3108                 }
3109                 loi_init((*lsmp)->lsm_oinfo[0]);
3110         }
3111
3112         if (lmm != NULL) {
3113                 /* XXX zero *lsmp? */
3114                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3115                 LASSERT((*lsmp)->lsm_object_id);
3116         }
3117
3118         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3119
3120         RETURN(lsm_size);
3121 }
3122
3123 /* caller must hold fo_create_locks[oa->o_gr] */
3124 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3125                                       struct filter_obd *filter)
3126 {
3127         struct obdo doa; /* XXX obdo on stack */
3128         obd_id last, id;
3129         int rc;
3130         ENTRY;
3131
3132         LASSERT(oa);
3133         LASSERT(oa->o_gr != 0);
3134         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3135         LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3136
3137         memset(&doa, 0, sizeof(doa));
3138
3139         doa.o_valid |= OBD_MD_FLGROUP;
3140         doa.o_gr = oa->o_gr;
3141         doa.o_mode = S_IFREG;
3142
3143         if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3144                 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3145                        exp->exp_obd->obd_name, doa.o_gr);
3146                 RETURN(0);
3147         }
3148
3149         last = filter_last_id(filter, doa.o_gr);
3150
3151         CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3152                exp->exp_obd->obd_name, oa->o_id + 1, last);
3153                
3154         for (id = last; id > oa->o_id; id--) {
3155                 doa.o_id = id;
3156                 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
3157                 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3158                         CEMERG("error destroying precreate objid "LPU64": %d\n",
3159                                id, rc);
3160                 filter_set_last_id(filter, id - 1, doa.o_gr);
3161                 /* update last_id on disk periodically so that if we restart
3162                  * we don't need to re-scan all of the just-deleted objects. */
3163                 if ((id & 511) == 0)
3164                         filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3165         }
3166
3167         CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3168                exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3169
3170         rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3171         clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3172
3173         RETURN(rc);
3174 }
3175
3176 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3177                             obd_gr group, int *num);
3178 /* returns a negative error or a nonnegative number of files to create */
3179 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3180                                    obd_gr group, struct obd_trans_info *oti)
3181 {
3182         struct obd_device *obd = exp->exp_obd;
3183         struct filter_obd *filter = &obd->u.filter;
3184         int diff, rc;
3185         ENTRY;
3186
3187         /* delete orphans request */
3188         if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3189                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3190                         CERROR("%s: dropping old orphan cleanup request\n",
3191                                obd->obd_name);
3192                         RETURN(0);
3193                 }
3194                 /* This causes inflight precreates to abort and drop lock */
3195                 set_bit(group, &filter->fo_destroys_in_progress);
3196                 down(&filter->fo_create_locks[group]);
3197                 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3198                         CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3199                                exp->exp_obd->obd_name, group);
3200                         up(&filter->fo_create_locks[group]);
3201                         RETURN(0);
3202                 }
3203                 diff = oa->o_id - filter_last_id(filter, group);
3204                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3205                        filter_last_id(filter, group), diff);
3206
3207                 if (-diff > OST_MAX_PRECREATE) {
3208                         CERROR("%s: ignoring bogus orphan destroy request: "
3209                                "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3210                                oa->o_id, filter_last_id(filter, group));
3211                         /* FIXME: should reset precreate_next_id on MDS */
3212                         GOTO(out, rc = -EINVAL);
3213                 }
3214                 if (diff < 0) {
3215                         rc = filter_destroy_precreated(exp, oa, filter);
3216                         if (rc)
3217                                 CERROR("%s: unable to write lastobjid, but "
3218                                        "orphans were deleted\n", obd->obd_name);
3219                         GOTO(out, rc);
3220                 } else {
3221                         /* XXX: Used by MDS for the first time! */
3222                         clear_bit(group, &filter->fo_destroys_in_progress);
3223                 }
3224         } else {
3225                 down(&filter->fo_create_locks[group]);
3226                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3227                         CERROR("%s: dropping old precreate request\n",
3228                                obd->obd_name);
3229                         GOTO(out, rc = 0);
3230                 }
3231                 /* only precreate if group == 0 and o_id is specfied */
3232                 if (group < FILTER_GROUP_MDS0 || oa->o_id == 0)
3233                         diff = 1;
3234                 else
3235                         diff = oa->o_id - filter_last_id(filter, group);
3236                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3237                        filter_last_id(filter, group), diff);
3238
3239                 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3240                          oa->o_id, filter_last_id(filter, group), diff);
3241         }
3242
3243         if (diff > 0) {
3244                 oa->o_id = filter_last_id(&obd->u.filter, group);
3245                 rc = filter_precreate(obd, oa, group, &diff);
3246                 oa->o_id = filter_last_id(&obd->u.filter, group);
3247                 oa->o_gr = group;
3248                 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3249                 GOTO(out, rc);
3250         }
3251         /* else diff == 0 */
3252         GOTO(out, rc = 0);
3253 out:
3254         up(&filter->fo_create_locks[group]);
3255         return rc;
3256 }
3257
3258 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3259                          __u64 max_age)
3260 {
3261         struct filter_obd *filter = &obd->u.filter;
3262         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3263         int rc;
3264         ENTRY;
3265
3266         /* at least try to account for cached pages.  its still racey and
3267          * might be under-reporting if clients haven't announced their
3268          * caches with brw recently */
3269         spin_lock(&obd->obd_osfs_lock);
3270         rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3271         memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3272         spin_unlock(&obd->obd_osfs_lock);
3273
3274         CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3275                " pending "LPU64" free "LPU64" avail "LPU64"\n",
3276                filter->fo_tot_dirty, filter->fo_tot_granted,
3277                filter->fo_tot_pending,
3278                osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3279
3280         filter_grant_sanity_check(obd, __FUNCTION__);
3281
3282         osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3283                                ((filter->fo_tot_dirty + filter->fo_tot_pending +
3284                                  osfs->os_bsize - 1) >> blockbits));
3285
3286         /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3287          * stop creating files on MDS if OST is not good shape to create
3288          * objects.*/
3289         osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3290                 EROFS : 0;
3291         RETURN(rc);
3292 }
3293
3294 /* We rely on the fact that only one thread will be creating files in a given
3295  * group at a time, which is why we don't need an atomic filter_get_new_id.
3296  * Even if we had that atomic function, the following race would exist:
3297  *
3298  * thread 1: gets id x from filter_next_id
3299  * thread 2: gets id (x + 1) from filter_next_id
3300  * thread 2: creates object (x + 1)
3301  * thread 1: tries to create object x, gets -ENOSPC
3302  *
3303  * Caller must hold fo_create_locks[group]
3304  */
3305 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3306                             obd_gr group, int *num)
3307 {
3308         struct dentry *dchild = NULL, *dparent = NULL;
3309         struct filter_obd *filter;
3310         struct obd_statfs *osfs;
3311         int err = 0, rc = 0, recreate_obj = 0, i;
3312         unsigned long enough_time = jiffies + min(obd_timeout * HZ / 4, 10U*HZ);
3313         obd_id next_id;
3314         void *handle = NULL;
3315         ENTRY;
3316
3317         filter = &obd->u.filter;
3318
3319         LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3320
3321         OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3322
3323         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3324             (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3325                 recreate_obj = 1;
3326         } else {
3327                 OBD_ALLOC(osfs, sizeof(*osfs));
3328                 if (osfs == NULL)
3329                         RETURN(-ENOMEM);
3330                 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ);
3331                 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3332                         CDEBUG(D_HA,"%s: not enough space for create "LPU64"\n",
3333                                obd->obd_name, osfs->os_bavail <<
3334                                filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3335                         *num = 0;
3336                         rc = -ENOSPC;
3337                 }
3338                 OBD_FREE(osfs, sizeof(*osfs));
3339                 if (rc)
3340                         RETURN(rc);
3341         }
3342
3343         CDEBUG(D_HA, "%s: precreating %d objects in group "LPU64" at "LPU64"\n",
3344                obd->obd_name, *num, group, oa->o_id);
3345
3346         for (i = 0; i < *num && err == 0; i++) {
3347                 int cleanup_phase = 0;
3348
3349                 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3350                         CWARN("%s: create aborted by destroy\n",
3351                               obd->obd_name);
3352                         rc = -EAGAIN;
3353                         break;
3354                 }
3355                 
3356                 if (recreate_obj) {
3357                         __u64 last_id;
3358                         next_id = oa->o_id;
3359                         last_id = filter_last_id(filter, group);
3360                         if (next_id > last_id) {
3361                                 CERROR("Error: Trying to recreate obj greater"
3362                                        "than last id "LPD64" > "LPD64"\n",
3363                                        next_id, last_id);
3364                                 GOTO(cleanup, rc = -EINVAL);
3365                         }
3366                 } else
3367                         next_id = filter_last_id(filter, group) + 1;
3368
3369                 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
3370
3371                 dparent = filter_parent_lock(obd, group, next_id);
3372                 if (IS_ERR(dparent))
3373                         GOTO(cleanup, rc = PTR_ERR(dparent));
3374                 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3375
3376                 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3377                 if (IS_ERR(dchild))
3378                         GOTO(cleanup, rc = PTR_ERR(dchild));
3379                 cleanup_phase = 2;  /* f_dput(dchild) */
3380
3381                 if (dchild->d_inode != NULL) {
3382                         /* This would only happen if lastobjid was bad on disk*/
3383                         /* Could also happen if recreating missing obj but
3384                          * already exists
3385                          */
3386                         if (recreate_obj) {
3387                                 CERROR("%s: recreating existing object %.*s?\n",
3388                                        obd->obd_name, dchild->d_name.len,
3389                                        dchild->d_name.name);
3390                         } else {
3391                                 CERROR("%s: Serious error: objid %.*s already "
3392                                        "exists; is this filesystem corrupt?\n",
3393                                        obd->obd_name, dchild->d_name.len,
3394                                        dchild->d_name.name);
3395                                 LBUG();
3396                         }
3397                         GOTO(cleanup, rc = -EEXIST);
3398                 }
3399
3400                 handle = fsfilt_start_log(obd, dparent->d_inode,
3401                                           FSFILT_OP_CREATE, NULL, 1);
3402                 if (IS_ERR(handle))
3403                         GOTO(cleanup, rc = PTR_ERR(handle));
3404                 cleanup_phase = 3;
3405
3406                 /* We mark object SUID+SGID to flag it for accepting UID+GID
3407                  * from client on first write.  Currently the permission bits
3408                  * on the OST are never used, so this is OK. */
3409                 rc = ll_vfs_create(dparent->d_inode, dchild,
3410                                    S_IFREG |  S_ISUID | S_ISGID | 0666, NULL);
3411                 if (rc) {
3412                         CERROR("create failed rc = %d\n", rc);
3413                         GOTO(cleanup, rc);
3414                 }
3415
3416                 if (!recreate_obj) {
3417                         filter_set_last_id(filter, next_id, group);
3418                         err = filter_update_last_objid(obd, group, 0);
3419                         if (err)
3420                                 CERROR("unable to write lastobjid "
3421                                        "but file created\n");
3422                 }
3423
3424         cleanup:
3425                 switch(cleanup_phase) {
3426                 case 3:
3427                         err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3428                         if (err) {
3429                                 CERROR("error on commit, err = %d\n", err);
3430                                 if (!rc)
3431                                         rc = err;
3432                         }
3433                 case 2:
3434                         f_dput(dchild);
3435                 case 1:
3436                         filter_parent_unlock(dparent);
3437                 case 0:
3438                         break;
3439                 }
3440
3441                 if (rc)
3442                         break;
3443                 if (time_after(jiffies, enough_time)) {
3444                         CDEBUG(D_HA, "%s: precreate slow - want %d got %d \n",
3445                                obd->obd_name, *num, i);
3446                         break;
3447                 }
3448         }
3449         *num = i;
3450
3451         CDEBUG(D_HA, "%s: created %d objects for group "LPU64": "LPU64"\n",
3452                obd->obd_name, i, group, filter->fo_last_objids[group]);
3453
3454         RETURN(rc);
3455 }
3456
3457 static int filter_create(struct obd_export *exp, struct obdo *oa,
3458                          struct lov_stripe_md **ea, struct obd_trans_info *oti)
3459 {
3460         struct filter_export_data *fed;
3461         struct obd_device *obd = NULL;
3462         struct filter_obd *filter;
3463         struct lvfs_run_ctxt saved;
3464         struct lov_stripe_md *lsm = NULL;
3465         int rc = 0, diff, group = oa->o_gr;
3466         ENTRY;
3467
3468         if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
3469                 CERROR("!!! nid %s sent invalid object group %d\n",
3470                         obd_export_nid2str(exp), group);
3471                 RETURN(-EINVAL);
3472         }
3473
3474         obd = exp->exp_obd;
3475         fed = &exp->exp_filter_data;
3476         filter = &obd->u.filter;
3477
3478         if (fed->fed_group != group) {
3479                 CERROR("!!! this export (nid %s) used object group %d "
3480                         "earlier; now it's trying to use group %d!  This could "
3481                         "be a bug in the MDS.  Tell CFS.\n",
3482                         obd_export_nid2str(exp), fed->fed_group, group);
3483                 RETURN(-ENOTUNIQ);
3484         }
3485
3486         CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
3487                oa->o_gr, oa->o_id);
3488         if (ea != NULL) {
3489                 lsm = *ea;
3490                 if (lsm == NULL) {
3491                         rc = obd_alloc_memmd(exp, &lsm);
3492                         if (rc < 0)
3493                                 RETURN(rc);
3494                 }
3495         }
3496
3497         obd = exp->exp_obd;
3498         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3499
3500         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3501             (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3502                 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3503                         CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3504                                oa->o_id, filter_last_id(filter,
3505                                                         oa->o_gr));
3506                         rc = -EINVAL;
3507                 } else {
3508                         diff = 1;
3509                         down(&filter->fo_create_locks[oa->o_gr]);
3510                         rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3511                         up(&filter->fo_create_locks[oa->o_gr]);
3512                 }
3513         } else {
3514                 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3515         }
3516
3517         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3518         if (rc && ea != NULL && *ea != lsm) {
3519                 obd_free_memmd(exp, &lsm);
3520         } else if (rc == 0 && ea != NULL) {
3521                 /* XXX LOV STACKING: the lsm that is passed to us from
3522                  * LOV does not have valid lsm_oinfo data structs, so
3523                  * don't go touching that.  This needs to be fixed in a
3524                  * big way. */
3525                 lsm->lsm_object_id = oa->o_id;
3526                 *ea = lsm;
3527         }
3528
3529         RETURN(rc);
3530 }
3531
3532 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3533                    struct lov_stripe_md *md, struct obd_trans_info *oti,
3534                    struct obd_export *md_exp)
3535 {
3536         unsigned int qcids[MAXQUOTAS] = {0, 0};
3537         struct obd_device *obd;
3538         struct filter_obd *filter;
3539         struct dentry *dchild = NULL, *dparent = NULL;
3540         struct lvfs_run_ctxt saved;
3541         void *handle = NULL;
3542         struct llog_cookie *fcc = NULL;
3543         int rc, rc2, cleanup_phase = 0, sync = 0;
3544         struct iattr iattr;
3545         ENTRY;
3546
3547         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3548
3549         obd = exp->exp_obd;
3550         filter = &obd->u.filter;
3551
3552         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3553         cleanup_phase = 1;
3554
3555         dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3556         if (IS_ERR(dchild))
3557                 GOTO(cleanup, rc = PTR_ERR(dchild));
3558         cleanup_phase = 2;
3559
3560         if (dchild->d_inode == NULL) {
3561                 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3562                        oa->o_id);
3563                 /* If object already gone, cancel cookie right now */
3564                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3565                         fcc = obdo_logcookie(oa);
3566                         llog_cancel(llog_get_context(obd, fcc->lgc_subsys + 1),
3567                                     NULL, 1, fcc, 0);
3568                         fcc = NULL; /* we didn't allocate fcc, don't free it */
3569                 }
3570                 GOTO(cleanup, rc = -ENOENT);
3571         }
3572
3573         filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3574
3575         /* Our MDC connection is established by the MDS to us */
3576         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3577                 OBD_ALLOC(fcc, sizeof(*fcc));
3578                 if (fcc != NULL)
3579                         memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3580         }
3581         DQUOT_INIT(dchild->d_inode);
3582
3583         /* we're gonna truncate it first in order to avoid possible deadlock:
3584          *      P1                      P2
3585          * open trasaction      open transaction
3586          * down(i_zombie)       down(i_zombie)
3587          *                      restart transaction
3588          * (see BUG 4180) -bzzz
3589          */
3590         LOCK_INODE_MUTEX(dchild->d_inode);
3591         handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3592                                   NULL, 1);
3593         if (IS_ERR(handle)) {
3594                 UNLOCK_INODE_MUTEX(dchild->d_inode);
3595                 GOTO(cleanup, rc = PTR_ERR(handle));
3596         }
3597
3598         iattr.ia_valid = ATTR_SIZE;
3599         iattr.ia_size = 0;
3600         rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3601         rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3602         UNLOCK_INODE_MUTEX(dchild->d_inode);
3603         if (rc)
3604                 GOTO(cleanup, rc);
3605         if (rc2)
3606                 GOTO(cleanup, rc = rc2);
3607
3608         /* We don't actually need to lock the parent until we are unlinking
3609          * here, and not while truncating above.  That avoids holding the
3610          * parent lock for a long time during truncate, which can block other
3611          * threads from doing anything to objects in that directory. bug 7171 */
3612         dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3613         if (IS_ERR(dparent))
3614                 GOTO(cleanup, rc = PTR_ERR(dparent));
3615         cleanup_phase = 3; /* filter_parent_unlock */
3616
3617         LOCK_INODE_MUTEX(dchild->d_inode);
3618         handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3619         if (IS_ERR(handle)) {
3620                 UNLOCK_INODE_MUTEX(dchild->d_inode);
3621                 GOTO(cleanup, rc = PTR_ERR(handle));
3622         }
3623         cleanup_phase = 4; /* fsfilt_commit */
3624
3625         /* Quota release need uid/gid of inode */
3626         obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3627
3628         filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3629
3630         /* this drops dchild->d_inode->i_mutex unconditionally */
3631         rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3632
3633         EXIT;
3634 cleanup:
3635         switch(cleanup_phase) {
3636         case 4:
3637                 if (fcc != NULL)
3638                         sync = fsfilt_add_journal_cb(obd, 0, oti ?
3639                                                      oti->oti_handle : handle,
3640                                                      filter_cancel_cookies_cb,
3641                                                      fcc);
3642                 /* If add_journal_cb failed, then filter_finish_transno
3643                  * will commit the handle and we will do a sync 
3644                  * on commit. then we call callback directly to free 
3645                  * the fcc. 
3646                  */
3647                 rc = filter_finish_transno(exp, oti, rc, sync);
3648                 if (sync) {
3649                         filter_cancel_cookies_cb(obd, 0, fcc, rc); 
3650                         fcc = NULL;
3651                 }
3652                 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3653                 if (rc2) {
3654                         CERROR("error on commit, err = %d\n", rc2);
3655                         if (!rc)
3656                                 rc = rc2;