Whamcloud - gitweb
b=14608
[fs/lustre-release.git] / lustre / obdfilter / filter.c
1 /*
2  * -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
3  * vim:expandtab:shiftwidth=8:tabstop=8:
4  *
5  * GPL HEADER START
6  *
7  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License version 2 only,
11  * as published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License version 2 for more details (a copy is included
17  * in the LICENSE file that accompanied this code).
18  *
19  * You should have received a copy of the GNU General Public License
20  * version 2 along with this program; If not, see [sun.com URL with a
21  * copy of GPLv2].
22  *
23  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
24  * CA 95054 USA or visit www.sun.com if you need additional information or
25  * have any questions.
26  *
27  * GPL HEADER END
28  */
29 /*
30  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
31  * Use is subject to license terms.
32  */
33 /*
34  * This file is part of Lustre, http://www.lustre.org/
35  * Lustre is a trademark of Sun Microsystems, Inc.
36  *
37  * lustre/obdfilter/filter.c
38  *
39  * Author: Peter Braam <braam@clusterfs.com>
40  * Author: Andreas Dilger <adilger@clusterfs.com>
41  */
42
43 /*
44  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
45  *            (which need to get journal_lock, may block if journal full).
46  *
47  * Invariant: Call filter_start_transno() before any journal ops to avoid the
48  *            same deadlock problem.  We can (and want) to get rid of the
49  *            transno sem in favour of the dir/inode i_mutex to avoid single
50  *            threaded operation on the OST.
51  */
52
53 #define DEBUG_SUBSYSTEM S_FILTER
54
55 #ifndef AUTOCONF_INCLUDED
56 #include <linux/config.h>
57 #endif
58 #include <linux/module.h>
59 #include <linux/fs.h>
60 #include <linux/dcache.h>
61 #include <linux/init.h>
62 #include <linux/version.h>
63 #include <linux/sched.h>
64 #include <linux/mount.h>
65 #include <linux/buffer_head.h>
66
67 #include <obd_cksum.h>
68 #include <obd_class.h>
69 #include <obd_lov.h>
70 #include <lustre_dlm.h>
71 #include <lustre_fsfilt.h>
72 #include <lprocfs_status.h>
73 #include <lustre_log.h>
74 #include <libcfs/list.h>
75 #include <lustre_disk.h>
76 #include <lustre_quota.h>
77 #include <linux/slab.h>
78 #include <lustre_param.h>
79
80 #include "filter_internal.h"
81
82 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
83 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS0
84 static struct lvfs_callback_ops filter_lvfs_ops;
85 cfs_mem_cache_t *ll_fmd_cachep;
86
87 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
88                              void *cb_data, int error)
89 {
90         obd_transno_commit_cb(obd, transno, error);
91 }
92
93 /* Assumes caller has already pushed us into the kernel context. */
94 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
95                           int rc, int force_sync)
96 {
97         struct filter_obd *filter = &exp->exp_obd->u.filter;
98         struct filter_export_data *fed = &exp->exp_filter_data;
99         struct lsd_client_data *lcd = fed->fed_lcd;
100         __u64 last_rcvd;
101         loff_t off;
102         int err, log_pri = D_RPCTRACE;
103
104         /* Propagate error code. */
105         if (rc)
106                 RETURN(rc);
107
108         if (!exp->exp_obd->obd_replayable || oti == NULL)
109                 RETURN(rc);
110
111         /* we don't allocate new transnos for replayed requests */
112         if (oti->oti_transno == 0) {
113                 spin_lock(&filter->fo_translock);
114                 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
115                 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
116                 spin_unlock(&filter->fo_translock);
117                 oti->oti_transno = last_rcvd;
118         } else {
119                 spin_lock(&filter->fo_translock);
120                 last_rcvd = oti->oti_transno;
121                 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
122                         filter->fo_fsd->lsd_last_transno =
123                                 cpu_to_le64(last_rcvd);
124                 spin_unlock(&filter->fo_translock);
125         }
126         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
127
128         /* could get xid from oti, if it's ever needed */
129         lcd->lcd_last_xid = 0;
130
131         off = fed->fed_lr_off;
132         if (off <= 0) {
133                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
134                        fed->fed_lr_idx, fed->fed_lr_off);
135                 err = -EINVAL;
136         } else {
137                 if (!force_sync)
138                         force_sync = fsfilt_add_journal_cb(exp->exp_obd, 
139                                                            last_rcvd,
140                                                            oti->oti_handle,
141                                                            filter_commit_cb,
142                                                            NULL);
143
144                 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
145                                           lcd, sizeof(*lcd), &off,
146                                           force_sync | exp->exp_need_sync);
147                 if (force_sync)
148                         filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
149         }
150         if (err) {
151                 log_pri = D_ERROR;
152                 if (rc == 0)
153                         rc = err;
154         }
155
156         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
157                last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
158
159         RETURN(rc);
160 }
161
162 void f_dput(struct dentry *dentry)
163 {
164         /* Can't go inside filter_ddelete because it can block */
165         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
166                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
167         LASSERT(atomic_read(&dentry->d_count) > 0);
168
169         dput(dentry);
170 }
171
172 static void init_brw_stats(struct brw_stats *brw_stats)
173 {
174         int i;
175         for (i = 0; i < BRW_LAST; i++)
176                 spin_lock_init(&brw_stats->hist[i].oh_lock);
177 }
178
179 static int lprocfs_init_rw_stats(struct obd_device *obd,
180                                  struct lprocfs_stats **stats)
181 {
182         int num_stats;
183
184         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
185                                                         LPROC_FILTER_LAST - 1;
186         *stats = lprocfs_alloc_stats(num_stats, 0);
187         if (*stats == NULL)
188                 return -ENOMEM;
189
190         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
191         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
192                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
193         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
194                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
195
196         return(0);
197 }
198
199 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
200    plus the procfs overhead :( */
201 static int filter_export_stats_init(struct obd_device *obd,
202                                     struct obd_export *exp,
203                                     void *client_nid)
204 {
205         struct filter_export_data *fed = &exp->exp_filter_data;
206         int rc, newnid = 0;
207         ENTRY;
208
209         init_brw_stats(&fed->fed_brw_stats);
210
211         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
212                 /* Self-export gets no proc entry */
213                 RETURN(0);
214
215         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
216         if (rc)
217                 RETURN(rc);
218
219         if (newnid) {
220                 struct nid_stat *tmp = exp->exp_nid_stats;
221                 LASSERT(tmp != NULL);
222
223                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
224                 if (tmp->nid_brw_stats == NULL)
225                         RETURN(-ENOMEM);
226
227                 init_brw_stats(tmp->nid_brw_stats);
228                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
229                                         0644, &filter_per_nid_stats_fops,
230                                         exp->exp_nid_stats);
231                 if (rc)
232                         CWARN("Error adding the brw_stats file\n");
233
234                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
235                 if (rc)
236                         RETURN(rc);
237
238                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
239                                             tmp->nid_stats);
240                 if (rc)
241                         RETURN(rc);
242         }
243
244         RETURN(0);
245 }
246
247 /* Add client data to the FILTER.  We use a bitmap to locate a free space
248  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
249  * Otherwise, we have just read the data from the last_rcvd file and
250  * we know its offset. */
251 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
252                              int cl_idx)
253 {
254         struct filter_obd *filter = &obd->u.filter;
255         struct filter_export_data *fed = &exp->exp_filter_data;
256         unsigned long *bitmap = filter->fo_last_rcvd_slots;
257         int new_client = (cl_idx == -1);
258  
259         ENTRY;
260
261         LASSERT(bitmap != NULL);
262         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
263
264         /* Self-export */
265         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
266                 RETURN(0);
267
268         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
269          * there's no need for extra complication here
270          */
271         if (new_client) {
272                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
273         repeat:
274                 if (cl_idx >= LR_MAX_CLIENTS) {
275                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
276                                cl_idx);
277                         RETURN(-EOVERFLOW);
278                 }
279                 if (test_and_set_bit(cl_idx, bitmap)) {
280                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
281                                                     cl_idx);
282                         goto repeat;
283                 }
284         } else {
285                 if (test_and_set_bit(cl_idx, bitmap)) {
286                         CERROR("FILTER client %d: bit already set in bitmap!\n",
287                                cl_idx);
288                         LBUG();
289                 }
290         }
291
292         fed->fed_lr_idx = cl_idx;
293         fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
294                 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
295         LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
296
297         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
298                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
299
300         if (new_client) {
301                 struct lvfs_run_ctxt saved;
302                 loff_t off = fed->fed_lr_off;
303                 int rc;
304                 void *handle;
305
306                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
307                        fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
308
309                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
310                 /* Transaction needed to fix bug 1403 */
311                 handle = fsfilt_start(obd,
312                                       filter->fo_rcvd_filp->f_dentry->d_inode,
313                                       FSFILT_OP_SETATTR, NULL);
314                 if (IS_ERR(handle)) {
315                         rc = PTR_ERR(handle);
316                         CERROR("unable to start transaction: rc %d\n", rc);
317                 } else {
318                         rc = fsfilt_add_journal_cb(obd, 0, handle,
319                                                    target_client_add_cb, exp);
320                         if (rc == 0) {
321                                 spin_lock(&exp->exp_lock);
322                                 exp->exp_need_sync = 1;
323                                 spin_unlock(&exp->exp_lock);
324                         }
325                         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
326                                                  fed->fed_lcd,
327                                                  sizeof(*fed->fed_lcd),
328                                                  &off, rc /* sync if no cb */);
329                         fsfilt_commit(obd,
330                                       filter->fo_rcvd_filp->f_dentry->d_inode,
331                                       handle, 0);
332                 }
333                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
334
335                 if (rc) {
336                         CERROR("error writing %s client idx %u: rc %d\n",
337                                LAST_RCVD, fed->fed_lr_idx, rc);
338                         RETURN(rc);
339                 }
340         }
341         RETURN(0);
342 }
343
344 static int filter_client_free(struct obd_export *exp)
345 {
346         struct filter_export_data *fed = &exp->exp_filter_data;
347         struct filter_obd *filter = &exp->exp_obd->u.filter;
348         struct obd_device *obd = exp->exp_obd;
349         struct lsd_client_data zero_lcd;
350         struct lvfs_run_ctxt saved;
351         int rc;
352         loff_t off;
353         ENTRY;
354
355         if (fed->fed_lcd == NULL)
356                 RETURN(0);
357
358         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
359         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
360                 GOTO(free, 0);
361
362         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
363                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
364
365         LASSERT(filter->fo_last_rcvd_slots != NULL);
366
367         off = fed->fed_lr_off;
368
369         /* Don't clear fed_lr_idx here as it is likely also unset.  At worst
370          * we leak a client slot that will be cleaned on the next recovery. */
371         if (off <= 0) {
372                 CERROR("%s: client idx %d has med_off %lld\n",
373                        obd->obd_name, fed->fed_lr_idx, off);
374                 GOTO(free, rc = -EINVAL);
375         }
376
377         /* Clear the bit _after_ zeroing out the client so we don't
378            race with filter_client_add and zero out new clients.*/
379         if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
380                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
381                        fed->fed_lr_idx);
382                 LBUG();
383         }
384
385         if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
386                 memset(&zero_lcd, 0, sizeof zero_lcd);
387                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
388                 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
389                                          sizeof(zero_lcd), &off,
390                                          (!exp->exp_libclient ||
391                                           exp->exp_need_sync));
392                 if (rc == 0)
393                         /* update server's transno */
394                         filter_update_server_data(obd, filter->fo_rcvd_filp,
395                                                   filter->fo_fsd,
396                                                   !exp->exp_libclient);
397                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
398
399                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
400                        "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
401                        fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
402                        LAST_RCVD, rc);
403         }
404
405         if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
406                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
407                        fed->fed_lr_idx);
408                 LBUG();
409         }
410
411         EXIT;
412 free:
413         OBD_FREE_PTR(fed->fed_lcd);
414         fed->fed_lcd = NULL;
415
416         return 0;
417 }
418
419 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
420 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
421                                          struct filter_mod_data *fmd)
422 {
423         LASSERT_SPIN_LOCKED(&fed->fed_lock);
424         if (--fmd->fmd_refcount == 0) {
425                 /* XXX when we have persistent reservations and the handle
426                  * is stored herein we need to drop it here. */
427                 fed->fed_mod_count--;
428                 list_del(&fmd->fmd_list);
429                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
430         }
431 }
432
433 /* drop fmd reference, free it if last ref */
434 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
435 {
436         struct filter_export_data *fed;
437
438         if (fmd == NULL)
439                 return;
440
441         fed = &exp->exp_filter_data;
442         spin_lock(&fed->fed_lock);
443         filter_fmd_put_nolock(fed, fmd); /* caller reference */
444         spin_unlock(&fed->fed_lock);
445 }
446
447 /* expire entries from the end of the list if there are too many
448  * or they are too old */
449 static void filter_fmd_expire_nolock(struct filter_obd *filter,
450                                      struct filter_export_data *fed,
451                                      struct filter_mod_data *keep)
452 {
453         struct filter_mod_data *fmd, *tmp;
454
455         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
456                 if (fmd == keep)
457                         break;
458
459                 if (time_before(jiffies, fmd->fmd_expire) &&
460                     fed->fed_mod_count < filter->fo_fmd_max_num)
461                         break;
462
463                 list_del_init(&fmd->fmd_list);
464                 filter_fmd_put_nolock(fed, fmd); /* list reference */
465         }
466 }
467
468 void filter_fmd_expire(struct obd_export *exp)
469 {
470         spin_lock(&exp->exp_filter_data.fed_lock);
471         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
472                                  &exp->exp_filter_data, NULL);
473         spin_unlock(&exp->exp_filter_data.fed_lock);
474 }
475
476 /* find specified objid, group in export fmd list.
477  * caller must hold fed_lock and take fmd reference itself */
478 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
479                                                 struct filter_export_data *fed,
480                                                 obd_id objid, obd_gr group)
481 {
482         struct filter_mod_data *found = NULL, *fmd;
483
484         LASSERT_SPIN_LOCKED(&fed->fed_lock);
485
486         list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
487                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
488                         found = fmd;
489                         list_del(&fmd->fmd_list);
490                         list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
491                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
492                         break;
493                 }
494         }
495
496         filter_fmd_expire_nolock(filter, fed, found);
497
498         return found;
499 }
500
501 /* Find fmd based on objid and group, or return NULL if not found. */
502 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
503                                         obd_id objid, obd_gr group)
504 {
505         struct filter_mod_data *fmd;
506
507         spin_lock(&exp->exp_filter_data.fed_lock);
508         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
509                                      &exp->exp_filter_data, objid, group);
510         if (fmd)
511                 fmd->fmd_refcount++;    /* caller reference */
512         spin_unlock(&exp->exp_filter_data.fed_lock);
513
514         return fmd;
515 }
516
517 /* Find fmd based on objid and group, or create a new one if none is found.
518  * It is possible for this function to return NULL under memory pressure,
519  * or if objid = 0 is passed (which will only cause old entries to expire).
520  * Currently this is not fatal because any fmd state is transient and
521  * may also be freed when it gets sufficiently old. */
522 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
523                                        obd_id objid, obd_gr group)
524 {
525         struct filter_export_data *fed = &exp->exp_filter_data;
526         struct filter_mod_data *found = NULL, *fmd_new = NULL;
527
528         OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO, sizeof(*fmd_new));
529
530         spin_lock(&fed->fed_lock);
531         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
532         if (fmd_new) {
533                 if (found == NULL) {
534                         list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
535                         fmd_new->fmd_id = objid;
536                         fmd_new->fmd_gr = group;
537                         fmd_new->fmd_refcount++;   /* list reference */
538                         found = fmd_new;
539                         fed->fed_mod_count++;
540                 } else {
541                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
542                 }
543         }
544         if (found) {
545                 found->fmd_refcount++;          /* caller reference */
546                 found->fmd_expire = jiffies +
547                         exp->exp_obd->u.filter.fo_fmd_max_age;
548         }
549
550         spin_unlock(&fed->fed_lock);
551
552         return found;
553 }
554
555 #ifdef DO_FMD_DROP
556 /* drop fmd list reference so it will disappear when last reference is put.
557  * This isn't so critical because it would in fact only affect the one client
558  * that is doing the unlink and at worst we have an stale entry referencing
559  * an object that should never be used again. */
560 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
561 {
562         struct filter_mod_data *found = NULL;
563
564         spin_lock(&exp->exp_filter_data.fed_lock);
565         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
566         if (found) {
567                 list_del_init(&found->fmd_list);
568                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
569         }
570         spin_unlock(&exp->exp_filter_data.fed_lock);
571 }
572 #else
573 #define filter_fmd_drop(exp, objid, group)
574 #endif
575
576 /* remove all entries from fmd list */
577 static void filter_fmd_cleanup(struct obd_export *exp)
578 {
579         struct filter_export_data *fed = &exp->exp_filter_data;
580         struct filter_mod_data *fmd = NULL, *tmp;
581
582         spin_lock(&fed->fed_lock);
583         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
584                 list_del_init(&fmd->fmd_list);
585                 filter_fmd_put_nolock(fed, fmd);
586         }
587         spin_unlock(&fed->fed_lock);
588 }
589
590 static int filter_init_export(struct obd_export *exp)
591 {
592         spin_lock_init(&exp->exp_filter_data.fed_lock);
593         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
594
595         spin_lock(&exp->exp_lock);
596         exp->exp_connecting = 1;
597         spin_unlock(&exp->exp_lock);
598
599         return 0;
600 }
601
602 static int filter_free_server_data(struct filter_obd *filter)
603 {
604         OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
605         filter->fo_fsd = NULL;
606         OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
607         filter->fo_last_rcvd_slots = NULL;
608         return 0;
609 }
610
611 /* assumes caller is already in kernel ctxt */
612 int filter_update_server_data(struct obd_device *obd, struct file *filp,
613                               struct lr_server_data *fsd, int force_sync)
614 {
615         loff_t off = 0;
616         int rc;
617         ENTRY;
618
619         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
620         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
621                le64_to_cpu(fsd->lsd_last_transno));
622         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
623                le64_to_cpu(fsd->lsd_mount_count));
624
625         fsd->lsd_compat14 = fsd->lsd_last_transno;
626         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
627         if (rc)
628                 CERROR("error writing lr_server_data: rc = %d\n", rc);
629
630         RETURN(rc);
631 }
632
633 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
634                              int force_sync)
635 {
636         struct filter_obd *filter = &obd->u.filter;
637         __u64 tmp;
638         loff_t off = 0;
639         int rc;
640         ENTRY;
641
642         if (filter->fo_last_objid_files[group] == NULL) {
643                 CERROR("Object group "LPU64" not fully setup; not updating "
644                        "last_objid\n", group);
645                 RETURN(-EINVAL);
646         }
647
648         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
649                obd->obd_name, group, filter->fo_last_objids[group]);
650
651         tmp = cpu_to_le64(filter->fo_last_objids[group]);
652         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
653                                  &tmp, sizeof(tmp), &off, force_sync);
654         if (rc)
655                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
656                        group, rc);
657         RETURN(rc);
658 }
659 extern int ost_handle(struct ptlrpc_request *req);
660 /* assumes caller has already in kernel ctxt */
661 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
662 {
663         struct filter_obd *filter = &obd->u.filter;
664         struct lr_server_data *fsd;
665         struct lsd_client_data *lcd = NULL;
666         struct inode *inode = filp->f_dentry->d_inode;
667         unsigned long last_rcvd_size = i_size_read(inode);
668         __u64 mount_count;
669         int cl_idx;
670         loff_t off = 0;
671         int rc;
672
673         /* ensure padding in the struct is the correct size */
674         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
675                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
676         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
677                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
678
679         OBD_ALLOC(fsd, sizeof(*fsd));
680         if (!fsd)
681                 RETURN(-ENOMEM);
682         filter->fo_fsd = fsd;
683
684         OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
685         if (filter->fo_last_rcvd_slots == NULL) {
686                 OBD_FREE(fsd, sizeof(*fsd));
687                 RETURN(-ENOMEM);
688         }
689
690         if (last_rcvd_size == 0) {
691                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
692
693                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
694                 fsd->lsd_last_transno = 0;
695                 mount_count = fsd->lsd_mount_count = 0;
696                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
697                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
698                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
699                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
700                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
701                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
702         } else {
703                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
704                 if (rc) {
705                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
706                                LAST_RCVD, rc);
707                         GOTO(err_fsd, rc);
708                 }
709                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
710                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
711                                            "using the wrong disk %s. Were the "
712                                            "/dev/ assignments rearranged?\n",
713                                            obd->obd_uuid.uuid, fsd->lsd_uuid);
714                         GOTO(err_fsd, rc = -EINVAL);
715                 }
716                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
717                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
718                 /* COMPAT_146 */
719                 /* Assume old last_rcvd format unless I_C_LR is set */
720                 if (!(fsd->lsd_feature_incompat &
721                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
722                         fsd->lsd_last_transno = fsd->lsd_compat14;
723                 /* end COMPAT_146 */
724         }
725
726         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
727                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
728                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
729                        ~FILTER_INCOMPAT_SUPP);
730                 GOTO(err_fsd, rc = -EINVAL);
731         }
732         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
733                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
734                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
735                        ~FILTER_ROCOMPAT_SUPP);
736                 /* Do something like remount filesystem read-only */
737                 GOTO(err_fsd, rc = -EINVAL);
738         }
739
740         CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
741                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
742         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
743                obd->obd_name, mount_count + 1);
744         CDEBUG(D_INODE, "%s: server data size: %u\n",
745                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
746         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
747                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
748         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
749                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
750         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
751                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
752         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
753                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
754                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
755                 le16_to_cpu(fsd->lsd_client_size));
756
757         if (!obd->obd_replayable) {
758                 CWARN("%s: recovery support OFF\n", obd->obd_name);
759                 GOTO(out, rc = 0);
760         }
761
762         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
763              off < last_rcvd_size; cl_idx++) {
764                 __u64 last_rcvd;
765                 struct obd_export *exp;
766                 struct filter_export_data *fed;
767
768                 if (!lcd) {
769                         OBD_ALLOC_PTR(lcd);
770                         if (!lcd)
771                                 GOTO(err_client, rc = -ENOMEM);
772                 }
773
774                 /* Don't assume off is incremented properly by
775                  * fsfilt_read_record(), in case sizeof(*lcd)
776                  * isn't the same as fsd->lsd_client_size.  */
777                 off = le32_to_cpu(fsd->lsd_client_start) +
778                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
779                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
780                 if (rc) {
781                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
782                                LAST_RCVD, cl_idx, off, rc);
783                         break; /* read error shouldn't cause startup to fail */
784                 }
785
786                 if (lcd->lcd_uuid[0] == '\0') {
787                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
788                                cl_idx);
789                         continue;
790                 }
791
792                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
793
794                 /* These exports are cleaned up by filter_disconnect(), so they
795                  * need to be set up like real exports as filter_connect() does.
796                  */
797                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
798
799                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
800                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
801                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
802                 if (IS_ERR(exp)) {
803                         if (PTR_ERR(exp) == -EALREADY) {
804                                 /* export already exists, zero out this one */
805                                 CERROR("Zeroing out duplicate export due to "
806                                        "bug 10479.\n");
807                                 lcd->lcd_uuid[0] = '\0';
808                         } else {
809                                 GOTO(err_client, rc = PTR_ERR(exp));
810                         }
811                 } else {
812                         fed = &exp->exp_filter_data;
813                         fed->fed_lcd = lcd;
814                         fed->fed_group = 0; /* will be assigned at connect */
815                         filter_export_stats_init(obd, exp, NULL);
816                         rc = filter_client_add(obd, exp, cl_idx);
817                         /* can't fail for existing client */
818                         LASSERTF(rc == 0, "rc = %d\n", rc);
819
820                         lcd = NULL;
821                         spin_lock(&exp->exp_lock);
822                         exp->exp_connecting = 0;
823                         exp->exp_in_recovery = 0;
824                         spin_unlock(&exp->exp_lock);
825                         obd->obd_max_recoverable_clients++;
826                         class_export_put(exp);
827                 }
828
829                 /* Need to check last_rcvd even for duplicated exports. */
830                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
831                        cl_idx, last_rcvd);
832
833                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
834                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
835         }
836
837         if (lcd)
838                 OBD_FREE_PTR(lcd);
839
840         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
841
842         target_recovery_init(obd, ost_handle);
843
844 out:
845         filter->fo_mount_count = mount_count + 1;
846         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
847
848         /* save it, so mount count and last_transno is current */
849         rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
850         if (rc)
851                 GOTO(err_client, rc);
852
853         RETURN(0);
854
855 err_client:
856         target_recovery_fini(obd);
857 err_fsd:
858         filter_free_server_data(filter);
859         RETURN(rc);
860 }
861
862 static int filter_cleanup_groups(struct obd_device *obd)
863 {
864         struct filter_obd *filter = &obd->u.filter;
865         struct file *filp;
866         struct dentry *dentry;
867         int i, j;
868         ENTRY;
869
870         if (filter->fo_dentry_O_groups != NULL) {
871                 for (i = 0; i < filter->fo_group_count; i++) {
872                         dentry = filter->fo_dentry_O_groups[i];
873                         if (dentry != NULL)
874                                 f_dput(dentry);
875                 }
876                 OBD_FREE(filter->fo_dentry_O_groups,
877                          filter->fo_group_count *
878                          sizeof(*filter->fo_dentry_O_groups));
879                 filter->fo_dentry_O_groups = NULL;
880         }
881         if (filter->fo_last_objid_files != NULL) {
882                 for (i = 0; i < filter->fo_group_count; i++) {
883                         filp = filter->fo_last_objid_files[i];
884                         if (filp != NULL)
885                                 filp_close(filp, 0);
886                 }
887                 OBD_FREE(filter->fo_last_objid_files,
888                          filter->fo_group_count *
889                          sizeof(*filter->fo_last_objid_files));
890                 filter->fo_last_objid_files = NULL;
891         }
892         if (filter->fo_dentry_O_sub != NULL) {
893                 for (i = 0; i < filter->fo_group_count; i++) {
894                         for (j = 0; j < filter->fo_subdir_count; j++) {
895                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
896                                 if (dentry != NULL)
897                                         f_dput(dentry);
898                         }
899                 }
900                 OBD_FREE(filter->fo_dentry_O_sub,
901                          filter->fo_group_count *
902                          sizeof(*filter->fo_dentry_O_sub));
903                 filter->fo_dentry_O_sub = NULL;
904         }
905         if (filter->fo_last_objids != NULL) {
906                 OBD_FREE(filter->fo_last_objids,
907                          filter->fo_group_count *
908                          sizeof(*filter->fo_last_objids));
909                 filter->fo_last_objids = NULL;
910         }
911         if (filter->fo_dentry_O != NULL) {
912                 f_dput(filter->fo_dentry_O);
913                 filter->fo_dentry_O = NULL;
914         }
915         RETURN(0);
916 }
917
918 static int filter_update_last_group(struct obd_device *obd, int group)
919 {
920         struct filter_obd *filter = &obd->u.filter;
921         struct file *filp = NULL;
922         int last_group = 0, rc;
923         loff_t off = 0;
924         ENTRY;
925
926         if (group <= filter->fo_committed_group)
927                 RETURN(0);
928
929         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
930         if (IS_ERR(filp)) {
931                 rc = PTR_ERR(filp);
932                 filp = NULL;
933                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
934                 GOTO(cleanup, rc);
935         }
936
937         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
938         if (rc) {
939                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
940                 GOTO(cleanup, rc);
941         }
942         LASSERT(off == 0 || last_group >= FILTER_MIN_GROUPS);
943         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
944                obd->obd_name, last_group, group);
945
946         off = 0;
947         last_group = group;
948         /* must be sync: bXXXX */
949         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
950         if (rc) {
951                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
952                 GOTO(cleanup, rc);
953         }
954
955         filter->fo_committed_group = group;
956 cleanup:
957         if (filp)
958                 filp_close(filp, 0);
959         RETURN(rc);
960 }
961
962 static int filter_read_group_internal(struct obd_device *obd, int group,
963                                       int create)
964 {
965         struct filter_obd *filter = &obd->u.filter;
966         __u64 *new_objids = NULL;
967         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
968         struct dentry **new_groups = NULL;
969         struct file **new_files = NULL;
970         struct dentry *dentry;
971         struct file *filp;
972         int old_count = filter->fo_group_count, rc, stage = 0, i;
973         char name[25];
974         __u64 last_objid;
975         loff_t off = 0;
976         int len = group + 1;
977
978         snprintf(name, 24, "%d", group);
979         name[24] = '\0';
980
981         if (!create) {
982                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
983                                            strlen(name));
984                 if (IS_ERR(dentry)) {
985                         CERROR("Cannot lookup expected object group %d: %ld\n",
986                                group, PTR_ERR(dentry));
987                         RETURN(PTR_ERR(dentry));
988                 }
989         } else {
990                 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt, 
991                                       name, 0700, 1);
992                 if (IS_ERR(dentry)) {
993                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
994                                PTR_ERR(dentry));
995                         RETURN(PTR_ERR(dentry));
996                 }
997         }
998         stage = 1;
999
1000         snprintf(name, 24, "O/%d/LAST_ID", group);
1001         name[24] = '\0';
1002         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1003         if (IS_ERR(filp)) {
1004                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1005                 GOTO(cleanup, rc = PTR_ERR(filp));
1006         }
1007         stage = 2;
1008
1009         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1010         if (rc) {
1011                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1012                 GOTO(cleanup, rc);
1013         }
1014
1015         if (filter->fo_subdir_count) {
1016                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1017                 if (tmp_subdirs == NULL)
1018                         GOTO(cleanup, rc = -ENOMEM);
1019                 stage = 3;
1020
1021                 for (i = 0; i < filter->fo_subdir_count; i++) {
1022                         char dir[20];
1023                         snprintf(dir, sizeof(dir), "d%u", i);
1024
1025                         tmp_subdirs->dentry[i] = simple_mkdir(dentry, 
1026                                                               filter->fo_vfsmnt,
1027                                                               dir, 0700, 1);
1028                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1029                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1030                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1031                                        group, dir, rc);
1032                                 GOTO(cleanup, rc);
1033                         }
1034
1035                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1036                                tmp_subdirs->dentry[i]);
1037                 }
1038         }
1039
1040         /* 'group' is an index; we need an array of length 'group + 1' */
1041         if (group + 1 > old_count) {
1042                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1043                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1044                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1045                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1046                 stage = 4;
1047                 if (new_objids == NULL || new_subdirs == NULL ||
1048                     new_groups == NULL || new_files == NULL)
1049                         GOTO(cleanup, rc = -ENOMEM);
1050
1051                 if (old_count) {
1052                         memcpy(new_objids, filter->fo_last_objids,
1053                                old_count * sizeof(*new_objids));
1054                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1055                                old_count * sizeof(*new_subdirs));
1056                         memcpy(new_groups, filter->fo_dentry_O_groups,
1057                                old_count * sizeof(*new_groups));
1058                         memcpy(new_files, filter->fo_last_objid_files,
1059                                old_count * sizeof(*new_files));
1060
1061                         OBD_FREE(filter->fo_last_objids,
1062                                  old_count * sizeof(*new_objids));
1063                         OBD_FREE(filter->fo_dentry_O_sub,
1064                                  old_count * sizeof(*new_subdirs));
1065                         OBD_FREE(filter->fo_dentry_O_groups,
1066                                  old_count * sizeof(*new_groups));
1067                         OBD_FREE(filter->fo_last_objid_files,
1068                                  old_count * sizeof(*new_files));
1069                 }
1070                 filter->fo_last_objids = new_objids;
1071                 filter->fo_dentry_O_sub = new_subdirs;
1072                 filter->fo_dentry_O_groups = new_groups;
1073                 filter->fo_last_objid_files = new_files;
1074                 filter->fo_group_count = len;
1075         }
1076
1077         filter->fo_dentry_O_groups[group] = dentry;
1078         filter->fo_last_objid_files[group] = filp;
1079         if (filter->fo_subdir_count) {
1080                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1081                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1082         }
1083
1084         filter_update_last_group(obd, group);
1085
1086         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1087                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1088                 rc = filter_update_last_objid(obd, group, 1);
1089                 RETURN(rc);
1090         }
1091
1092         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1093         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1094                obd->obd_name, group, last_objid);
1095         RETURN(0);
1096  cleanup:
1097         switch (stage) {
1098         case 4:
1099                 if (new_objids != NULL)
1100                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1101                 if (new_subdirs != NULL)
1102                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1103                 if (new_groups != NULL)
1104                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1105                 if (new_files != NULL)
1106                         OBD_FREE(new_files, len * sizeof(*new_files));
1107         case 3:
1108                 if (filter->fo_subdir_count) {
1109                         for (i = 0; i < filter->fo_subdir_count; i++) {
1110                                 if (tmp_subdirs->dentry[i] != NULL)
1111                                         dput(tmp_subdirs->dentry[i]);
1112                         }
1113                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1114                 }
1115         case 2:
1116                 filp_close(filp, 0);
1117         case 1:
1118                 dput(dentry);
1119         }
1120         RETURN(rc);
1121 }
1122
1123 static int filter_read_groups(struct obd_device *obd, int last_group,
1124                               int create)
1125 {
1126         struct filter_obd *filter = &obd->u.filter;
1127         int old_count, group, rc = 0;
1128
1129         down(&filter->fo_init_lock);
1130         old_count = filter->fo_group_count;
1131         for (group = old_count; group <= last_group; group++) {
1132                 if (group == 0)
1133                         continue; /* no group zero */
1134
1135                 rc = filter_read_group_internal(obd, group, create);
1136                 if (rc != 0)
1137                         break;
1138         }
1139         up(&filter->fo_init_lock);
1140         return rc;
1141 }
1142
1143 /* FIXME: object groups */
1144 static int filter_prep_groups(struct obd_device *obd)
1145 {
1146         struct filter_obd *filter = &obd->u.filter;
1147         struct dentry *dentry, *O_dentry;
1148         struct file *filp;
1149         int    last_group, rc = 0, cleanup_phase = 0;
1150         loff_t off = 0;
1151         ENTRY;
1152
1153         O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt, 
1154                                 "O", 0700, 1);
1155         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1156         if (IS_ERR(O_dentry)) {
1157                 rc = PTR_ERR(O_dentry);
1158                 CERROR("cannot open/create O: rc = %d\n", rc);
1159                 GOTO(cleanup, rc);
1160         }
1161         filter->fo_dentry_O = O_dentry;
1162         cleanup_phase = 1; /* O_dentry */
1163
1164         /* Lookup "R" to tell if we're on an old OST FS and need to convert
1165          * from O/R/<dir>/<objid> to O/0/<dir>/<objid>.  This can be removed
1166          * some time post 1.0 when all old-style OSTs have converted along
1167          * with the init_objid hack. */
1168         dentry = ll_lookup_one_len("R", O_dentry, 1);
1169         if (IS_ERR(dentry))
1170                 GOTO(cleanup, rc = PTR_ERR(dentry));
1171         if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1172                 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1173                 ENTRY;
1174
1175                 CWARN("converting OST to new object layout\n");
1176                 if (IS_ERR(O0_dentry)) {
1177                         rc = PTR_ERR(O0_dentry);
1178                         CERROR("error looking up O/0: rc %d\n", rc);
1179                         GOTO(cleanup_R, rc);
1180                 }
1181
1182                 if (O0_dentry->d_inode) {
1183                         CERROR("Both O/R and O/0 exist. Fix manually.\n");
1184                         GOTO(cleanup_O0, rc = -EEXIST);
1185                 }
1186
1187                 LOCK_INODE_MUTEX(O_dentry->d_inode);
1188                 rc = ll_vfs_rename(O_dentry->d_inode, dentry, filter->fo_vfsmnt,
1189                                    O_dentry->d_inode, O0_dentry,
1190                                    filter->fo_vfsmnt);
1191                 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1192
1193                 if (rc) {
1194                         CERROR("error renaming O/R to O/0: rc %d\n", rc);
1195                         GOTO(cleanup_O0, rc);
1196                 }
1197                 filter->fo_fsd->lsd_feature_incompat |=
1198                         cpu_to_le32(OBD_INCOMPAT_GROUPS);
1199                 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1200                                                filter->fo_fsd, 1);
1201                 GOTO(cleanup_O0, rc);
1202
1203         cleanup_O0:
1204                 f_dput(O0_dentry);
1205         cleanup_R:
1206                 f_dput(dentry);
1207                 if (rc)
1208                         GOTO(cleanup, rc);
1209         } else {
1210                 f_dput(dentry);
1211         }
1212
1213         cleanup_phase = 2; /* groups */
1214
1215         /* we have to initialize all groups before first connections from
1216          * clients because they may send create/destroy for any group -bzzz */
1217         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1218         if (IS_ERR(filp)) {
1219                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1220                 GOTO(cleanup, rc = PTR_ERR(filp));
1221         }
1222         cleanup_phase = 3; /* filp */
1223
1224         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1225         if (rc) {
1226                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1227                 GOTO(cleanup, rc);
1228         }
1229         if (off == 0) {
1230                 last_group = FILTER_MIN_GROUPS;
1231         } else {
1232                 LASSERT(last_group >= FILTER_MIN_GROUPS);
1233         }
1234
1235         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1236               FILTER_MIN_GROUPS, last_group);
1237         filter->fo_committed_group = last_group;
1238         rc = filter_read_groups(obd, last_group, 1);
1239         if (rc)
1240                 GOTO(cleanup, rc);
1241
1242         filp_close(filp, 0);
1243         RETURN(0);
1244
1245  cleanup:
1246         switch (cleanup_phase) {
1247         case 3:
1248                 filp_close(filp, 0);
1249         case 2:
1250                 filter_cleanup_groups(obd);
1251         case 1:
1252                 f_dput(filter->fo_dentry_O);
1253                 filter->fo_dentry_O = NULL;
1254         default:
1255                 break;
1256         }
1257         return rc;
1258
1259 }
1260
1261 /* setup the object store with correct subdirectories */
1262 static int filter_prep(struct obd_device *obd)
1263 {
1264         struct lvfs_run_ctxt saved;
1265         struct filter_obd *filter = &obd->u.filter;
1266         struct file *file;
1267         struct inode *inode;
1268         int rc = 0;
1269         ENTRY;
1270
1271         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1272         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1273         if (!file || IS_ERR(file)) {
1274                 rc = PTR_ERR(file);
1275                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1276                        LAST_RCVD, rc);
1277                 GOTO(out, rc);
1278         }
1279         filter->fo_rcvd_filp = file;
1280         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1281                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1282                        file->f_dentry->d_inode->i_mode);
1283                 GOTO(err_filp, rc = -ENOENT);
1284         }
1285
1286         inode = file->f_dentry->d_parent->d_inode;
1287         /* We use i_op->unlink directly in filter_vfs_unlink() */
1288         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1289                 CERROR("%s: filesystem does not support create/unlink ops\n",
1290                        obd->obd_name);
1291                 GOTO(err_filp, rc = -EOPNOTSUPP);
1292         }
1293
1294         rc = filter_init_server_data(obd, file);
1295         if (rc) {
1296                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1297                 GOTO(err_filp, rc);
1298         }
1299         /* open and create health check io file*/
1300         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1301         if (IS_ERR(file)) {
1302                 rc = PTR_ERR(file);
1303                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1304                        HEALTH_CHECK, rc);
1305                 GOTO(err_server_data, rc);
1306         }
1307         filter->fo_health_check_filp = file;
1308         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1309                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1310                        file->f_dentry->d_inode->i_mode);
1311                 GOTO(err_health_check, rc = -ENOENT);
1312         }
1313         rc = lvfs_check_io_health(obd, file);
1314         if (rc)
1315                 GOTO(err_health_check, rc);
1316
1317         rc = filter_prep_groups(obd);
1318         if (rc)
1319                 GOTO(err_health_check, rc);
1320 out:
1321         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1322
1323         return(rc);
1324
1325 err_health_check:
1326         if (filp_close(filter->fo_health_check_filp, 0))
1327                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1328         filter->fo_health_check_filp = NULL;
1329 err_server_data:
1330         target_recovery_fini(obd);
1331         filter_free_server_data(filter);
1332 err_filp:
1333         if (filp_close(filter->fo_rcvd_filp, 0))
1334                 CERROR("can't close %s after error\n", LAST_RCVD);
1335         filter->fo_rcvd_filp = NULL;
1336         goto out;
1337 }
1338
1339 /* cleanup the filter: write last used object id to status file */
1340 static void filter_post(struct obd_device *obd)
1341 {
1342         struct lvfs_run_ctxt saved;
1343         struct filter_obd *filter = &obd->u.filter;
1344         int rc, i;
1345
1346         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1347          * best to start a transaction with h_sync, because we removed this
1348          * from lastobjid */
1349
1350         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1351         rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1352                                        filter->fo_fsd, 0);
1353         if (rc)
1354                 CERROR("error writing server data: rc = %d\n", rc);
1355
1356         for (i = 1; i < filter->fo_group_count; i++) {
1357                 rc = filter_update_last_objid(obd, i,
1358                                 (i == filter->fo_group_count - 1));
1359                 if (rc)
1360                         CERROR("error writing group %d lastobjid: rc = %d\n",
1361                                i, rc);
1362         }
1363
1364         rc = filp_close(filter->fo_rcvd_filp, 0);
1365         filter->fo_rcvd_filp = NULL;
1366         if (rc)
1367                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1368
1369         rc = filp_close(filter->fo_health_check_filp, 0);
1370         filter->fo_health_check_filp = NULL;
1371         if (rc)
1372                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1373
1374         filter_cleanup_groups(obd);
1375         filter_free_server_data(filter);
1376         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1377
1378         filter_free_capa_keys(filter);
1379         cleanup_capa_hash(filter->fo_capa_hash);
1380 }
1381
1382 static void filter_set_last_id(struct filter_obd *filter,
1383                                obd_id id, obd_gr group)
1384 {
1385         LASSERT(filter->fo_fsd != NULL);
1386         LASSERT(group <= filter->fo_group_count);
1387
1388         spin_lock(&filter->fo_objidlock);
1389         filter->fo_last_objids[group] = id;
1390         spin_unlock(&filter->fo_objidlock);
1391 }
1392
1393 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1394 {
1395         obd_id id;
1396         LASSERT(filter->fo_fsd != NULL);
1397         LASSERT(group <= filter->fo_group_count);
1398
1399         /* FIXME: object groups */
1400         spin_lock(&filter->fo_objidlock);
1401         id = filter->fo_last_objids[group];
1402         spin_unlock(&filter->fo_objidlock);
1403
1404         return id;
1405 }
1406
1407 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1408 {
1409         LOCK_INODE_MUTEX(dparent->d_inode);
1410         return 0;
1411 }
1412
1413 /* We never dget the object parent, so DON'T dput it either */
1414 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1415 {
1416         struct filter_obd *filter = &obd->u.filter;
1417         struct filter_subdirs *subdirs;
1418         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1419
1420         if (group > 0 || filter->fo_subdir_count == 0)
1421                 return filter->fo_dentry_O_groups[group];
1422
1423         subdirs = &filter->fo_dentry_O_sub[group];
1424         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1425 }
1426
1427 /* We never dget the object parent, so DON'T dput it either */
1428 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1429                                   obd_id objid)
1430 {
1431         unsigned long now = jiffies;
1432         struct dentry *dparent = filter_parent(obd, group, objid);
1433         int rc;
1434
1435         if (IS_ERR(dparent))
1436                 return dparent;
1437         if (dparent == NULL)
1438                 return ERR_PTR(-ENOENT);
1439
1440         rc = filter_lock_dentry(obd, dparent);
1441         fsfilt_check_slow(obd, now, "parent lock");
1442         return rc ? ERR_PTR(rc) : dparent;
1443 }
1444
1445 /* We never dget the object parent, so DON'T dput it either */
1446 static void filter_parent_unlock(struct dentry *dparent)
1447 {
1448         UNLOCK_INODE_MUTEX(dparent->d_inode);
1449 }
1450
1451 /* How to get files, dentries, inodes from object id's.
1452  *
1453  * If dir_dentry is passed, the caller has already locked the parent
1454  * appropriately for this operation (normally a write lock).  If
1455  * dir_dentry is NULL, we do a read lock while we do the lookup to
1456  * avoid races with create/destroy and such changing the directory
1457  * internal to the filesystem code. */
1458 struct dentry *filter_fid2dentry(struct obd_device *obd,
1459                                  struct dentry *dir_dentry,
1460                                  obd_gr group, obd_id id)
1461 {
1462         struct dentry *dparent = dir_dentry;
1463         struct dentry *dchild;
1464         char name[32];
1465         int len;
1466         ENTRY;
1467
1468         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1469             obd->u.filter.fo_destroys_in_progress == 0) {
1470                 /* don't fail lookups for orphan recovery, it causes
1471                  * later LBUGs when objects still exist during precreate */
1472                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1473                 RETURN(ERR_PTR(-ENOENT));
1474         }
1475         if (id == 0) {
1476                 CERROR("fatal: invalid object id 0\n");
1477                 RETURN(ERR_PTR(-ESTALE));
1478         }
1479
1480         len = sprintf(name, LPU64, id);
1481         if (dir_dentry == NULL) {
1482                 dparent = filter_parent_lock(obd, group, id);
1483                 if (IS_ERR(dparent)) {
1484                         CERROR("%s: error getting object "LPU64":"LPU64
1485                                " parent: rc %ld\n", obd->obd_name,
1486                                id, group, PTR_ERR(dparent));
1487                         RETURN(dparent);
1488                 }
1489         }
1490         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1491                dparent->d_name.len, dparent->d_name.name, name);
1492         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1493         if (dir_dentry == NULL)
1494                 filter_parent_unlock(dparent);
1495         if (IS_ERR(dchild)) {
1496                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1497                        PTR_ERR(dchild));
1498                 RETURN(dchild);
1499         }
1500
1501         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1502                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1503                        obd->obd_name, id, dchild->d_inode->i_ino);
1504                 f_dput(dchild);
1505                 RETURN(ERR_PTR(-ENOENT));
1506         }
1507
1508         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1509                name, dchild, atomic_read(&dchild->d_count));
1510
1511         LASSERT(atomic_read(&dchild->d_count) > 0);
1512
1513         RETURN(dchild);
1514 }
1515
1516 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1517                                   obd_id group)
1518 {
1519         struct lustre_handle lockh;
1520         int flags = LDLM_AST_DISCARD_DATA, rc;
1521         struct ldlm_res_id res_id;
1522         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1523         ENTRY;
1524
1525         osc_build_res_name(objid, group, &res_id);
1526         /* Tell the clients that the object is gone now and that they should
1527          * throw away any cached pages. */
1528         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1529                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1530                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1531                                     &lockh);
1532
1533         /* We only care about the side-effects, just drop the lock. */
1534         if (rc == ELDLM_OK)
1535                 ldlm_lock_decref(&lockh, LCK_PW);
1536
1537         RETURN(rc);
1538 }
1539
1540 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1541  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1542  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1543  * reverse.  Caller must take i_sem before starting the transaction and we
1544  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1545 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1546                       struct vfsmount *mnt)
1547 {
1548         int rc;
1549         ENTRY;
1550
1551         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1552          * itself we already hold dir->i_mutex for child create/unlink ops */
1553         LASSERT(dentry->d_inode != NULL);
1554         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1555         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1556
1557
1558         /* may_delete() */
1559         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1560                 GOTO(out, rc = -ENOENT);
1561
1562         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1563         if (rc)
1564                 GOTO(out, rc);
1565
1566         if (IS_APPEND(dir))
1567                 GOTO(out, rc = -EPERM);
1568
1569         /* check_sticky() */
1570         if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1571             IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1572                 GOTO(out, rc = -EPERM);
1573
1574         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1575          *       that was done because of journal_start (which is already done
1576          *       here) or some other ordering issue. */
1577         DQUOT_INIT(dir);
1578
1579         rc = ll_security_inode_unlink(dir, dentry, mnt);
1580         if (rc)
1581                 GOTO(out, rc);
1582
1583         rc = dir->i_op->unlink(dir, dentry);
1584 out:
1585         /* need to drop i_mutex before we lose inode reference */
1586         UNLOCK_INODE_MUTEX(dentry->d_inode);
1587         if (rc == 0)
1588                 d_delete(dentry);
1589
1590         RETURN(rc);
1591 }
1592
1593 /* Caller must hold LCK_PW on parent and push us into kernel context.
1594  * Caller must hold child i_mutex, we drop it always.
1595  * Caller is also required to ensure that dchild->d_inode exists. */
1596 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1597                                    obd_gr group, struct dentry *dparent,
1598                                    struct dentry *dchild)
1599 {
1600         struct inode *inode = dchild->d_inode;
1601         struct filter_obd *filter = &obd->u.filter;
1602         int rc;
1603
1604         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1605                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1606                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1607                        (unsigned long)inode->i_nlink,
1608                        atomic_read(&inode->i_count));
1609         }
1610
1611         rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1612         if (rc)
1613                 CERROR("error unlinking objid %.*s: rc %d\n",
1614                        dchild->d_name.len, dchild->d_name.name, rc);
1615         return(rc);
1616 }
1617
1618 struct filter_intent_args {
1619         struct ldlm_lock **victim;
1620         __u64 size;
1621         int *liblustre;
1622 };
1623
1624 static enum interval_iter filter_intent_cb(struct interval_node *n,
1625                                            void *args)
1626 {
1627         struct ldlm_interval *node = (struct ldlm_interval *)n;
1628         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1629         __u64 size = arg->size;
1630         struct ldlm_lock **v = arg->victim;
1631         struct ldlm_lock *lck;
1632
1633         /* If the interval is lower than the current file size,
1634          * just break. */
1635         if (interval_high(n) <= size)
1636                 return INTERVAL_ITER_STOP;
1637
1638         list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1639                 /* Don't send glimpse ASTs to liblustre clients.
1640                  * They aren't listening for them, and they do
1641                  * entirely synchronous I/O anyways. */
1642                 if (lck->l_export == NULL ||
1643                     lck->l_export->exp_libclient == 1)
1644                         continue;
1645
1646                 if (*arg->liblustre)
1647                         *arg->liblustre = 0;
1648
1649                 if (*v == NULL) {
1650                         *v = LDLM_LOCK_GET(lck);
1651                 } else if ((*v)->l_policy_data.l_extent.start <
1652                            lck->l_policy_data.l_extent.start) {
1653                         LDLM_LOCK_PUT(*v);
1654                         *v = LDLM_LOCK_GET(lck);
1655                 }
1656
1657                 /* the same policy group - every lock has the
1658                  * same extent, so needn't do it any more */
1659                 break;
1660         }
1661
1662         return INTERVAL_ITER_CONT;
1663 }
1664
1665 static int filter_intent_policy(struct ldlm_namespace *ns,
1666                                 struct ldlm_lock **lockp, void *req_cookie,
1667                                 ldlm_mode_t mode, int flags, void *data)
1668 {
1669         CFS_LIST_HEAD(rpc_list);
1670         struct ptlrpc_request *req = req_cookie;
1671         struct ldlm_lock *lock = *lockp, *l = NULL;
1672         struct ldlm_resource *res = lock->l_resource;
1673         ldlm_processing_policy policy;
1674         struct ost_lvb *res_lvb, *reply_lvb;
1675         struct ldlm_reply *rep;
1676         ldlm_error_t err;
1677         int idx, rc, tmpflags = 0, only_liblustre = 1;
1678         struct ldlm_interval_tree *tree;
1679         struct filter_intent_args arg;
1680         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1681                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1682                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1683         ENTRY;
1684
1685         policy = ldlm_get_processing_policy(res);
1686         LASSERT(policy != NULL);
1687         LASSERT(req != NULL);
1688
1689         rc = lustre_pack_reply(req, 3, repsize, NULL);
1690         if (rc)
1691                 RETURN(req->rq_status = rc);
1692
1693         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1694         LASSERT(rep != NULL);
1695
1696         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1697                                    sizeof(*reply_lvb));
1698         LASSERT(reply_lvb != NULL);
1699
1700         //fixup_handle_for_resent_req(req, lock, &lockh);
1701
1702         /* If we grant any lock at all, it will be a whole-file read lock.
1703          * Call the extent policy function to see if our request can be
1704          * granted, or is blocked. 
1705          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse lock
1706          */
1707         lock->l_policy_data.l_extent.start = 0;
1708         lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1709         lock->l_req_mode = LCK_PR;
1710
1711         LASSERT(ns == res->lr_namespace);
1712         lock_res(res);
1713         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1714         check_res_locked(res);
1715
1716         /* FIXME: we should change the policy function slightly, to not make
1717          * this list at all, since we just turn around and free it */
1718         while (!list_empty(&rpc_list)) {
1719                 struct ldlm_lock *wlock =
1720                         list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1721                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1722                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1723                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1724                 list_del_init(&wlock->l_cp_ast);
1725                 LDLM_LOCK_PUT(wlock);
1726         }
1727
1728         /* The lock met with no resistance; we're finished. */
1729         if (rc == LDLM_ITER_CONTINUE) {
1730                 /* do not grant locks to the liblustre clients: they cannot
1731                  * handle ASTs robustly.  We need to do this while still
1732                  * holding ns_lock to avoid the lock remaining on the res_link
1733                  * list (and potentially being added to l_pending_list by an
1734                  * AST) when we are going to drop this lock ASAP. */
1735                 if (lock->l_export->exp_libclient ||
1736                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1737                         ldlm_resource_unlink_lock(lock);
1738                         err = ELDLM_LOCK_ABORTED;
1739                 } else {
1740                         err = ELDLM_LOCK_REPLACED;
1741                 }
1742                 unlock_res(res);
1743                 RETURN(err);
1744         }
1745
1746         /* Do not grant any lock, but instead send GL callbacks.  The extent
1747          * policy nicely created a list of all PW locks for us.  We will choose
1748          * the highest of those which are larger than the size in the LVB, if
1749          * any, and perform a glimpse callback. */
1750         res_lvb = res->lr_lvb_data;
1751         LASSERT(res_lvb != NULL);
1752         *reply_lvb = *res_lvb;
1753
1754         /*
1755          * ->ns_lock guarantees that no new locks are granted, and,
1756          * therefore, that res->lr_lvb_data cannot increase beyond the
1757          * end of already granted lock. As a result, it is safe to
1758          * check against "stale" reply_lvb->lvb_size value without
1759          * res->lr_lvb_sem.
1760          */
1761         arg.size = reply_lvb->lvb_size;
1762         arg.victim = &l;
1763         arg.liblustre = &only_liblustre;
1764         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1765                 tree = &res->lr_itree[idx];
1766                 if (tree->lit_mode == LCK_PR)
1767                         continue;
1768
1769                 interval_iterate_reverse(tree->lit_root, 
1770                                          filter_intent_cb, &arg);
1771         }
1772         unlock_res(res);
1773
1774         /* There were no PW locks beyond the size in the LVB; finished. */
1775         if (l == NULL) {
1776                 if (only_liblustre) {
1777                         /* If we discovered a liblustre client with a PW lock,
1778                          * however, the LVB may be out of date!  The LVB is
1779                          * updated only on glimpse (which we don't do for
1780                          * liblustre clients) and cancel (which the client
1781                          * obviously has not yet done).  So if it has written
1782                          * data but kept the lock, the LVB is stale and needs
1783                          * to be updated from disk.
1784                          *
1785                          * Of course, this will all disappear when we switch to
1786                          * taking liblustre locks on the OST. */
1787                         ldlm_res_lvbo_update(res, NULL, 0, 1);
1788                 }
1789                 RETURN(ELDLM_LOCK_ABORTED);
1790         }
1791
1792         /*
1793          * This check is for lock taken in filter_prepare_destroy() that does
1794          * not have l_glimpse_ast set. So the logic is: if there is a lock
1795          * with no l_glimpse_ast set, this object is being destroyed already.
1796          *
1797          * Hence, if you are grabbing DLM locks on the server, always set
1798          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1799          */
1800         if (l->l_glimpse_ast == NULL) {
1801                 /* We are racing with unlink(); just return -ENOENT */
1802                 rep->lock_policy_res1 = -ENOENT;
1803                 goto out;
1804         }
1805
1806         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1807         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1808         /* Update the LVB from disk if the AST failed (this is a legal race) */
1809         /*
1810          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1811          * sending ast is not handled. This can result in lost client writes.
1812          */
1813         if (rc != 0)
1814                 ldlm_res_lvbo_update(res, NULL, 0, 1);
1815
1816         lock_res(res);
1817         *reply_lvb = *res_lvb;
1818         unlock_res(res);
1819
1820  out:
1821         LDLM_LOCK_PUT(l);
1822
1823         RETURN(ELDLM_LOCK_ABORTED);
1824 }
1825
1826 /*
1827  * per-obd_device iobuf pool.
1828  *
1829  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1830  * allocations in write-path has to be minimized (see bug 5137).
1831  *
1832  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1833  * OST threads (see ost_thread_{init,done}()).
1834  *
1835  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1836  * at the OST layer there are only (potentially) multiple obd_device of type
1837  * unknown at the time of OST thread creation.
1838  *
1839  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1840  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1841  * it's very own iobuf.
1842  *
1843  * Functions below
1844  *
1845  *     filter_kiobuf_pool_init()
1846  *
1847  *     filter_kiobuf_pool_done()
1848  *
1849  *     filter_iobuf_get()
1850  *
1851  * operate on this array. They are "generic" in a sense that they don't depend
1852  * on actual type of iobuf's (the latter depending on Linux kernel version).
1853  */
1854
1855 /*
1856  * destroy pool created by filter_iobuf_pool_init
1857  */
1858 static void filter_iobuf_pool_done(struct filter_obd *filter)
1859 {
1860         struct filter_iobuf **pool;
1861         int i;
1862
1863         ENTRY;
1864
1865         pool = filter->fo_iobuf_pool;
1866         if (pool != NULL) {
1867                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1868                         if (pool[i] != NULL)
1869                                 filter_free_iobuf(pool[i]);
1870                 }
1871                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1872                 filter->fo_iobuf_pool = NULL;
1873         }
1874         EXIT;
1875 }
1876
1877 /*
1878  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1879  */
1880 static int filter_iobuf_pool_init(struct filter_obd *filter)
1881 {
1882         void **pool;
1883
1884         ENTRY;
1885
1886
1887         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1888                       GFP_KERNEL);
1889         if (filter->fo_iobuf_pool == NULL)
1890                 RETURN(-ENOMEM);
1891
1892         filter->fo_iobuf_count = OSS_THREADS_MAX;
1893
1894         RETURN(0);
1895 }
1896
1897 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1898  * many threads there will be so we allocate a large empty array and only
1899  * fill in those slots that are actually in use.
1900  * If we haven't allocated a pool entry for this thread before, do so now. */
1901 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1902 {
1903         int thread_id                    = oti ? oti->oti_thread_id : -1;
1904         struct filter_iobuf  *pool       = NULL;
1905         struct filter_iobuf **pool_place = NULL;
1906
1907         if (thread_id >= 0) {
1908                 LASSERT(thread_id < filter->fo_iobuf_count);
1909                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1910         }
1911
1912         if (unlikely(pool == NULL)) {
1913                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1914                                           PTLRPC_MAX_BRW_PAGES);
1915                 if (pool_place != NULL)
1916                         *pool_place = pool;
1917         }
1918
1919         return pool;
1920 }
1921
1922 /* mount the file system (secretly).  lustre_cfg parameters are:
1923  * 1 = device
1924  * 2 = fstype
1925  * 3 = flags: failover=f, failout=n
1926  * 4 = mount options
1927  */
1928 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1929                         void *option)
1930 {
1931         struct filter_obd *filter = &obd->u.filter;
1932         struct vfsmount *mnt;
1933         struct lustre_mount_info *lmi;
1934         struct obd_uuid uuid;
1935         __u8 *uuid_ptr;
1936         char *str, *label;
1937         char ns_name[48];
1938         request_queue_t *q;
1939         int rc, i;
1940         ENTRY;
1941
1942         if (lcfg->lcfg_bufcount < 3 ||
1943             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1944             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1945                 RETURN(-EINVAL);
1946
1947         lmi = server_get_mount(obd->obd_name);
1948         if (lmi) {
1949                 /* We already mounted in lustre_fill_super.
1950                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1951                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1952                 mnt = lmi->lmi_mnt;
1953                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1954         } else {
1955                 /* old path - used by lctl */
1956                 CERROR("Using old MDS mount method\n");
1957                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1958                                     MS_NOATIME|MS_NODIRATIME,
1959                                     lustre_cfg_string(lcfg, 1), option);
1960                 if (IS_ERR(mnt)) {
1961                         rc = PTR_ERR(mnt);
1962                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1963                                            lustre_cfg_string(lcfg, 1), rc);
1964                         RETURN(rc);
1965                 }
1966
1967                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1968         }
1969         if (IS_ERR(obd->obd_fsops))
1970                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1971
1972         rc = filter_iobuf_pool_init(filter);
1973         if (rc != 0)
1974                 GOTO(err_ops, rc);
1975
1976         LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1977
1978         /* failover is the default */
1979         obd->obd_replayable = 1;
1980
1981         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1982                 str = lustre_cfg_string(lcfg, 3);
1983                 if (strchr(str, 'n')) {
1984                         CWARN("%s: recovery disabled\n", obd->obd_name);
1985                         obd->obd_replayable = 0;
1986                 }
1987         }
1988
1989         filter->fo_vfsmnt = mnt;
1990         obd->u.obt.obt_sb = mnt->mnt_sb;
1991         filter->fo_fstype = mnt->mnt_sb->s_type->name;
1992         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1993
1994         fsfilt_setup(obd, obd->u.obt.obt_sb);
1995
1996         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1997         obd->obd_lvfs_ctxt.pwdmnt = mnt;
1998         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1999         obd->obd_lvfs_ctxt.fs = get_ds();
2000         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2001
2002         init_mutex(&filter->fo_init_lock);
2003         filter->fo_committed_group = 0;
2004
2005         rc = filter_prep(obd);
2006         if (rc)
2007                 GOTO(err_ops, rc);
2008
2009         filter->fo_destroys_in_progress = 0;
2010         for (i = 0; i < 32; i++)
2011                 sema_init(&filter->fo_create_locks[i], 1);
2012
2013         spin_lock_init(&filter->fo_translock);
2014         spin_lock_init(&filter->fo_objidlock);
2015         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2016         sema_init(&filter->fo_alloc_lock, 1);
2017         init_brw_stats(&filter->fo_filter_stats);
2018         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2019         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2020         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2021
2022         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2023         spin_lock_init(&filter->fo_llog_list_lock);
2024
2025         filter->fo_sptlrpc_lock = RW_LOCK_UNLOCKED;
2026         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2027
2028         filter->fo_fl_oss_capa = 0;
2029         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2030         filter->fo_capa_hash = init_capa_hash();
2031         if (filter->fo_capa_hash == NULL)
2032                 GOTO(err_ops, rc = -ENOMEM);
2033
2034         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2035         obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2036                                                 LDLM_NAMESPACE_GREEDY);
2037         if (obd->obd_namespace == NULL)
2038                 GOTO(err_post, rc = -ENOMEM);
2039         obd->obd_namespace->ns_lvbp = obd;
2040         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2041         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2042
2043         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2044                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2045
2046         rc = llog_cat_initialize(obd, &obd->obd_olg, 1, NULL);
2047         if (rc) {
2048                 CERROR("failed to setup llogging subsystems\n");
2049                 GOTO(err_post, rc);
2050         }
2051
2052         rc = lquota_setup(filter_quota_interface_ref, obd);
2053         if (rc)
2054                 GOTO(err_post, rc);
2055
2056         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2057         if (q->max_sectors < q->max_hw_sectors &&
2058             q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2059                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2060                               "for larger I/O requests: max_sectors = %u "
2061                               "could be up to max_hw_sectors=%u\n",
2062                               obd->obd_name, mnt->mnt_sb->s_id,
2063                               q->max_sectors, q->max_hw_sectors);
2064
2065         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2066         if (uuid_ptr != NULL) {
2067                 class_uuid_unparse(uuid_ptr, &uuid);
2068                 str = uuid.uuid;
2069         } else {
2070                 str = "no UUID";
2071         }
2072
2073         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2074
2075         if (obd->obd_recovering) {
2076                 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2077                               "recovery for at least %d:%.02d, or until %d "
2078                               "client%s reconnect. During this time new clients"
2079                               " will not be allowed to connect. "
2080                               "Recovery progress can be monitored by watching "
2081                               "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
2082                               obd->obd_name, lustre_cfg_string(lcfg, 1),
2083                               label ?: "", label ? "/" : "", str,
2084                               obd->obd_recovery_timeout / 60,
2085                               obd->obd_recovery_timeout % 60,
2086                               obd->obd_max_recoverable_clients,
2087                               (obd->obd_max_recoverable_clients == 1) ? "":"s",
2088                               obd->obd_name);
2089         } else {
2090                 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2091                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2092                               label ?: "", label ? "/" : "", str,
2093                               obd->obd_replayable ? "enabled" : "disabled");
2094         }
2095
2096         RETURN(0);
2097
2098 err_post:
2099         filter_post(obd);
2100 err_ops:
2101         fsfilt_put_ops(obd->obd_fsops);
2102         filter_iobuf_pool_done(filter);
2103 err_mntput:
2104         server_put_mount(obd->obd_name, mnt);
2105         obd->u.obt.obt_sb = 0;
2106         return rc;
2107 }
2108
2109 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2110 {
2111         struct lprocfs_static_vars lvars;
2112         unsigned long addr;
2113         struct page *page;
2114         int rc;
2115
2116         CLASSERT(offsetof(struct obd_device, u.obt) ==
2117                  offsetof(struct obd_device, u.filter.fo_obt));
2118
2119         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2120                 RETURN(-EINVAL);
2121
2122         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2123         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2124         if (!page)
2125                 RETURN(-ENOMEM);
2126         addr = (unsigned long)cfs_page_address(page);
2127         clear_page((void *)addr);
2128
2129         /* lprocfs must be setup before the filter so state can be safely added
2130          * to /proc incrementally as the filter is setup */
2131         lprocfs_filter_init_vars(&lvars);
2132         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2133             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2134                 /* Init obdfilter private stats here */
2135                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2136                                      LPROCFS_CNTR_AVGMINMAX,
2137                                      "read_bytes", "bytes");
2138                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2139                                      LPROCFS_CNTR_AVGMINMAX,
2140                                      "write_bytes", "bytes");
2141
2142                 lproc_filter_attach_seqstat(obd);
2143                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2144                                                         obd->obd_proc_entry,
2145                                                         NULL, NULL);
2146                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2147                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2148                         CERROR("error %d setting up lprocfs for %s\n",
2149                                rc, "exports");
2150                         obd->obd_proc_exports_entry = NULL;
2151                 }
2152         }
2153         if (obd->obd_proc_exports_entry)
2154                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2155                                    lprocfs_nid_stats_clear_read,
2156                                    lprocfs_nid_stats_clear_write, obd);
2157
2158         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2159                LUSTRE_CFG_BUFLEN(lcfg, 4));
2160         rc = filter_common_setup(obd, lcfg, (void *)addr);
2161         OBD_PAGE_FREE(page);
2162
2163         if (rc) {
2164                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2165                 lprocfs_free_per_client_stats(obd);
2166                 lprocfs_free_obd_stats(obd);
2167                 lprocfs_obd_cleanup(obd);
2168         }
2169
2170         return rc;
2171 }
2172
2173 static struct llog_operations filter_mds_ost_repl_logops;
2174
2175 static struct llog_operations filter_size_orig_logops = {
2176         .lop_setup   = llog_obd_origin_setup,
2177         .lop_cleanup = llog_obd_origin_cleanup,
2178         .lop_add     = llog_obd_origin_add
2179 };
2180
2181 static int filter_olg_fini(struct obd_llog_group *olg)
2182 {
2183         struct llog_ctxt *ctxt;
2184         int rc = 0, rc2 = 0;
2185         ENTRY;
2186
2187         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2188         if (ctxt)
2189                 rc = llog_cleanup(ctxt);
2190
2191         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2192         if (ctxt)
2193                 rc2 = llog_cleanup(ctxt);
2194         if (!rc)
2195                 rc = rc2;
2196
2197         RETURN(rc);
2198 }
2199
2200 static int 
2201 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2202                 struct obd_device *tgt)
2203 {
2204         int rc;
2205         ENTRY;
2206
2207         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2208                         &filter_mds_ost_repl_logops);
2209         if (rc)
2210                 GOTO(cleanup, rc);
2211
2212         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2213                         &filter_size_orig_logops);
2214         if (rc)
2215                 GOTO(cleanup, rc);
2216         EXIT;
2217 cleanup:
2218         if (rc)
2219                 filter_olg_fini(olg);
2220         return rc;
2221 }
2222
2223 /**
2224  * Init the default olg, which is embeded in the obd_device, for filter.
2225  */
2226 static int 
2227 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2228                         struct obd_device *tgt)
2229 {
2230         struct filter_obd *filter = &obd->u.filter;
2231         struct llog_ctxt *ctxt;
2232         int rc;
2233         ENTRY;
2234
2235         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2236         if (!filter->fo_lcm)
2237                 RETURN(-ENOMEM);
2238
2239         filter_mds_ost_repl_logops = llog_client_ops;
2240         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2241         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2242         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2243
2244         rc = filter_olg_init(obd, olg, tgt);
2245         if (rc)
2246                 GOTO(cleanup_lcm, rc);
2247
2248         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2249         if (!ctxt) {
2250                 CERROR("Can't get ctxt for %p:%x\n", olg, 
2251                        LLOG_MDS_OST_REPL_CTXT);
2252                 GOTO(cleanup_olg, rc = -ENODEV);
2253         }
2254         ctxt->loc_lcm = filter->fo_lcm;
2255         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2256         llog_ctxt_put(ctxt);
2257
2258         RETURN(0);
2259 cleanup_olg:
2260         filter_olg_fini(olg);
2261 cleanup_lcm:
2262         llog_recov_thread_fini(filter->fo_lcm, 1);
2263         filter->fo_lcm = NULL;
2264         return rc;
2265 }
2266                      
2267 static int 
2268 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2269                  struct obd_device *tgt, int count, struct llog_catid *catid,
2270                  struct obd_uuid *uuid)
2271 {
2272         struct filter_obd *filter = &obd->u.filter;
2273         struct llog_ctxt *ctxt;
2274         int rc;
2275         ENTRY;
2276
2277         LASSERT(olg != NULL);
2278         if (olg == &obd->obd_olg)
2279                 return filter_default_olg_init(obd, olg, tgt);
2280
2281         LASSERT(filter->fo_lcm != NULL);
2282         rc = filter_olg_init(obd, olg, tgt);
2283         if (rc)
2284                 RETURN(rc);
2285         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2286         if (!ctxt) {
2287                 CERROR("Can't get ctxt for %p:%x\n", olg, 
2288                        LLOG_MDS_OST_REPL_CTXT);
2289                 filter_olg_fini(olg);
2290                 RETURN(-ENODEV);
2291         }
2292         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2293         ctxt->loc_lcm = filter->fo_lcm;
2294         llog_ctxt_put(ctxt);
2295         RETURN(rc);
2296 }
2297
2298 static int filter_llog_finish(struct obd_device *obd, int count)
2299 {
2300         struct filter_obd *filter = &obd->u.filter;
2301         struct llog_ctxt *ctxt;
2302         ENTRY;
2303         
2304         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2305         LASSERT(ctxt != NULL);
2306         mutex_down(&ctxt->loc_sem);
2307         if (ctxt->loc_imp) {
2308                 /*
2309                  * Balance class_import_get() in llog_receptor_accept(). This
2310                  * is safe to do here, as llog is already synchronized and its
2311                  * import may go.
2312                  */
2313                 class_import_put(ctxt->loc_imp);
2314                 ctxt->loc_imp = NULL;
2315         }
2316         mutex_up(&ctxt->loc_sem);
2317         llog_ctxt_put(ctxt);
2318
2319         if (filter->fo_lcm) {
2320                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2321                 filter->fo_lcm = NULL;
2322         }
2323         RETURN(filter_olg_fini(&obd->obd_olg));
2324 }
2325
2326 /**
2327  * Find the group llog according to group index in the llog group list.
2328  */
2329 static struct obd_llog_group *
2330 filter_find_olg_internal(struct filter_obd *filter, int group)
2331 {
2332         struct obd_llog_group *olg;
2333
2334         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2335         list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2336                 if (olg->olg_group == group)
2337                         RETURN(olg);
2338         }
2339         RETURN(NULL);
2340 }
2341
2342 /**
2343  * Find the group llog according to group index on the filter
2344  */
2345 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2346 {
2347         struct obd_llog_group *olg = NULL;
2348         struct filter_obd *filter;
2349
2350         filter = &obd->u.filter;
2351
2352         if (group == OBD_LLOG_GROUP)
2353                 RETURN(&obd->obd_olg);
2354
2355         spin_lock(&filter->fo_llog_list_lock);
2356         olg = filter_find_olg_internal(filter, group);
2357         spin_unlock(&filter->fo_llog_list_lock);
2358
2359         RETURN(olg); 
2360 }
2361 /**
2362  * Find the llog_group of the filter according to the group. If it can not
2363  * find, create the llog_group, which only happens when mds is being synced 
2364  * with OST.
2365  */
2366 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2367 {
2368         struct obd_llog_group *olg = NULL;
2369         struct filter_obd *filter;
2370         int rc;
2371
2372         filter = &obd->u.filter;
2373
2374         if (group == OBD_LLOG_GROUP)
2375                 RETURN(&obd->obd_olg);
2376
2377         spin_lock(&filter->fo_llog_list_lock);
2378         olg = filter_find_olg_internal(filter, group);
2379         if (olg) {
2380                 if (olg->olg_initializing) {
2381                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2382                 } else {
2383                         GOTO(out_unlock, olg);
2384                 }
2385         }
2386         OBD_ALLOC_PTR(olg);
2387         if (olg == NULL)
2388                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2389
2390         llog_group_init(olg, group);
2391         list_add(&olg->olg_list, &filter->fo_llog_list);
2392         olg->olg_initializing = 1;
2393         spin_unlock(&filter->fo_llog_list_lock);
2394
2395         rc = llog_cat_initialize(obd, olg, 1, NULL);
2396         if (rc) {
2397                spin_lock(&filter->fo_llog_list_lock);
2398                list_del(&olg->olg_list);
2399                spin_unlock(&filter->fo_llog_list_lock);
2400                OBD_FREE_PTR(olg);
2401                GOTO(out, olg = ERR_PTR(-ENOMEM));
2402         }
2403         spin_lock(&filter->fo_llog_list_lock);
2404         olg->olg_initializing = 0;
2405         spin_unlock(&filter->fo_llog_list_lock);
2406         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2407               obd->obd_name, group, olg);
2408 out:
2409         RETURN(olg);
2410
2411 out_unlock:
2412         spin_unlock(&filter->fo_llog_list_lock);
2413         GOTO(out, olg);
2414 }
2415
2416 static int filter_llog_connect(struct obd_export *exp,
2417                                struct llogd_conn_body *body)
2418 {
2419         struct obd_device *obd = exp->exp_obd;
2420         struct llog_ctxt *ctxt;
2421         struct obd_llog_group *olg;
2422         int rc;
2423         ENTRY;
2424
2425         CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2426                 (unsigned) body->lgdc_logid.lgl_ogr,
2427                 (unsigned) body->lgdc_logid.lgl_oid,
2428                 (unsigned) body->lgdc_logid.lgl_ogen);
2429
2430         olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2431         if (!olg) {
2432                 CERROR(" %s: can not find olg of group %d\n",
2433                        obd->obd_name, (int)body->lgdc_logid.lgl_ogr); 
2434                 RETURN(-ENOENT);
2435         }
2436         llog_group_set_export(olg, exp);
2437
2438         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2439         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2440                  body->lgdc_ctxt_idx);
2441         rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2442                           &body->lgdc_gen, NULL);
2443         llog_ctxt_put(ctxt);
2444         if (rc != 0)
2445                 CERROR("failed to connect rc %d idx %d\n", rc,
2446                                 body->lgdc_ctxt_idx);
2447
2448         RETURN(rc);
2449 }
2450
2451 static int filter_llog_preclean(struct obd_device *obd)
2452 {
2453         struct obd_llog_group *olg, *tmp;
2454         struct filter_obd *filter;
2455         struct list_head  remove_list;
2456         int rc = 0;
2457         ENTRY;
2458
2459         rc = obd_llog_finish(obd, 0);
2460         if (rc)
2461                 CERROR("failed to cleanup llogging subsystem\n");
2462
2463         filter = &obd->u.filter;
2464         CFS_INIT_LIST_HEAD(&remove_list);
2465
2466         spin_lock(&filter->fo_llog_list_lock);
2467         while (!list_empty(&filter->fo_llog_list)) {
2468                 olg = list_entry(filter->fo_llog_list.next,
2469                                  struct obd_llog_group, olg_list);
2470                 list_del(&olg->olg_list);
2471                 list_add(&olg->olg_list, &remove_list);
2472         }
2473         spin_unlock(&filter->fo_llog_list_lock);
2474         
2475         list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {  
2476                 list_del_init(&olg->olg_list);
2477                 rc = filter_olg_fini(olg);
2478                 if (rc)
2479                         CERROR("failed to cleanup llogging subsystem for %u\n",
2480                                olg->olg_group);
2481                 OBD_FREE_PTR(olg);
2482         }
2483
2484         RETURN(rc);
2485 }
2486
2487 static int filter_precleanup(struct obd_device *obd,
2488                              enum obd_cleanup_stage stage)
2489 {
2490         int rc = 0;
2491         ENTRY;
2492
2493         switch(stage) {
2494         case OBD_CLEANUP_EARLY:
2495                 break;
2496         case OBD_CLEANUP_EXPORTS:
2497                 target_cleanup_recovery(obd);
2498                 rc = filter_llog_preclean(obd);
2499                 break;
2500         }
2501         RETURN(rc);
2502 }
2503
2504 static int filter_cleanup(struct obd_device *obd)
2505 {
2506         struct filter_obd *filter = &obd->u.filter;
2507         ENTRY;
2508
2509         if (obd->obd_fail)
2510                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2511                               "will be preserved.\n", obd->obd_name);
2512
2513         if (!list_empty(&obd->obd_exports)) {
2514                 CERROR("%s: still has clients!\n", obd->obd_name);
2515                 class_disconnect_exports(obd);
2516                 if (!list_empty(&obd->obd_exports)) {
2517                         CERROR("still has exports after forced cleanup?\n");
2518                         RETURN(-EBUSY);
2519                 }
2520         }
2521
2522         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2523         lprocfs_free_per_client_stats(obd);
2524         lprocfs_free_obd_stats(obd);
2525         lprocfs_obd_cleanup(obd);
2526         lquota_cleanup(filter_quota_interface_ref, obd);
2527
2528         /* Stop recovery before namespace cleanup. */
2529         target_stop_recovery_thread(obd);
2530         target_cleanup_recovery(obd);
2531
2532         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2533         obd->obd_namespace = NULL;
2534
2535         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2536
2537         if (obd->u.obt.obt_sb == NULL)
2538                 RETURN(0);
2539
2540         filter_post(obd);
2541
2542         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2543         shrink_dcache_sb(obd->u.obt.obt_sb);
2544
2545         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2546         obd->u.obt.obt_sb = NULL;
2547
2548         fsfilt_put_ops(obd->obd_fsops);
2549
2550         filter_iobuf_pool_done(filter);
2551
2552         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2553
2554         RETURN(0);
2555 }
2556
2557 static int filter_connect_internal(struct obd_export *exp,
2558                                    struct obd_connect_data *data)
2559 {
2560         struct filter_export_data *fed = &exp->exp_filter_data;
2561
2562         if (!data)
2563                 RETURN(0);
2564
2565         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2566                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2567                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2568                data->ocd_connect_flags, data->ocd_version,
2569                data->ocd_grant, data->ocd_index);
2570
2571         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2572                 CWARN("!!! This export (nid %s) used object group %d "
2573                        "earlier; now it's trying to use group %d!  This could "
2574                        "be a bug in the MDS. Please report to "
2575                        "http://bugzilla.lustre.org/\n",
2576                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2577                 RETURN(-EPROTO);
2578         }
2579         fed->fed_group = data->ocd_group;
2580
2581         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2582         exp->exp_connect_flags = data->ocd_connect_flags;
2583         data->ocd_version = LUSTRE_VERSION_CODE;
2584
2585         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2586                 obd_size left, want;
2587
2588                 spin_lock(&exp->exp_obd->obd_osfs_lock);
2589                 left = filter_grant_space_left(exp);
2590                 want = data->ocd_grant;
2591                 filter_grant(exp, fed->fed_grant, want, left);
2592                 data->ocd_grant = fed->fed_grant;
2593                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2594
2595                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2596                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2597                        exp->exp_client_uuid.uuid, exp,
2598                        data->ocd_grant, want, left);
2599         }
2600
2601         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2602                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2603                 struct lr_server_data *lsd = filter->fo_fsd;
2604                 int index = le32_to_cpu(lsd->lsd_ost_index);
2605
2606                 if (!(lsd->lsd_feature_compat &
2607                       cpu_to_le32(OBD_COMPAT_OST))) {
2608                         /* this will only happen on the first connect */
2609                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2610                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2611                         filter_update_server_data(exp->exp_obd,
2612                                                   filter->fo_rcvd_filp, lsd, 1);
2613                 } else if (index != data->ocd_index) {
2614                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2615                                            " %u doesn't match actual OST index"
2616                                            " %u in last_rcvd file, bad "
2617                                            "configuration?\n",
2618                                            obd_export_nid2str(exp), index,
2619                                            data->ocd_index);
2620                         RETURN(-EBADF);
2621                 }
2622         }
2623
2624         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2625                 data->ocd_brw_size = 65536;
2626         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2627                 data->ocd_brw_size = min(data->ocd_brw_size,
2628                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2629                 LASSERT(data->ocd_brw_size);
2630         }
2631
2632         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2633                 __u32 cksum_types = data->ocd_cksum_types;
2634
2635                 /* The client set in ocd_cksum_types the checksum types it
2636                  * supports. We have to mask off the algorithms that we don't
2637                  * support */
2638                 if (cksum_types & OBD_CKSUM_ALL)
2639                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2640                 else
2641                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2642
2643                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2644                                    "%x\n", exp->exp_obd->obd_name,
2645                                    obd_export_nid2str(exp), cksum_types,
2646                                    data->ocd_cksum_types);
2647         } else {
2648                 /* This client does not support OBD_CONNECT_CKSUM
2649                  * fall back to CRC32 */
2650                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2651                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2652                                    exp->exp_obd->obd_name,
2653                                    obd_export_nid2str(exp));
2654         }
2655
2656         /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2657          * FIXME: We don't strictly need the COMPAT flag for that,
2658          * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2659          * FIXME: We needed it for the index, as index 0 is valid. */
2660
2661         RETURN(0);
2662 }
2663
2664 static int filter_reconnect(const struct lu_env *env,
2665                             struct obd_export *exp, struct obd_device *obd,
2666                             struct obd_uuid *cluuid,
2667                             struct obd_connect_data *data)
2668 {
2669         int rc;
2670         ENTRY;
2671
2672         if (exp == NULL || obd == NULL || cluuid == NULL)
2673                 RETURN(-EINVAL);
2674
2675         rc = filter_connect_internal(exp, data);
2676
2677         RETURN(rc);
2678 }
2679
2680 /* nearly identical to mds_connect */
2681 static int filter_connect(const struct lu_env *env,
2682                           struct lustre_handle *conn, struct obd_device *obd,
2683                           struct obd_uuid *cluuid,
2684                           struct obd_connect_data *data, void *localdata)
2685 {
2686         struct lvfs_run_ctxt saved;
2687         struct obd_export *exp;
2688         struct filter_export_data *fed;
2689         struct lsd_client_data *lcd = NULL;
2690         __u32 group;
2691         int rc;
2692         ENTRY;
2693
2694         if (conn == NULL || obd == NULL || cluuid == NULL)
2695                 RETURN(-EINVAL);
2696
2697         rc = class_connect(conn, obd, cluuid);
2698         if (rc)
2699                 RETURN(rc);
2700         exp = class_conn2export(conn);
2701         LASSERT(exp != NULL);
2702
2703         fed = &exp->exp_filter_data;
2704
2705         rc = filter_connect_internal(exp, data);
2706         if (rc)
2707                 GOTO(cleanup, rc);
2708
2709         filter_export_stats_init(obd, exp, localdata);
2710         if (obd->obd_replayable) {
2711                 OBD_ALLOC(lcd, sizeof(*lcd));
2712                 if (!lcd) {
2713                         CERROR("filter: out of memory for client data\n");
2714                         GOTO(cleanup, rc = -ENOMEM);
2715                 }
2716
2717                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2718                 fed->fed_lcd = lcd;
2719                 rc = filter_client_add(obd, exp, -1);
2720                 if (rc)
2721                         GOTO(cleanup, rc);
2722         }
2723
2724         group = data->ocd_group;
2725         if (group == 0)
2726                 GOTO(cleanup, rc);
2727
2728         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2729               obd->obd_name, exp->exp_handle.h_cookie, group);
2730
2731         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2732         rc = filter_read_groups(obd, group, 1);
2733         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2734         if (rc != 0) {
2735                 CERROR("can't read group %u\n", group);
2736                 GOTO(cleanup, rc);
2737         }
2738
2739         GOTO(cleanup, rc);
2740
2741 cleanup:
2742         if (rc) {
2743                 if (lcd) {
2744                         OBD_FREE_PTR(lcd);
2745                         fed->fed_lcd = NULL;
2746                 }
2747                 class_disconnect(exp);
2748         } else {
2749                 class_export_put(exp);
2750         }
2751
2752         RETURN(rc);
2753 }
2754
2755 /* Do extra sanity checks for grant accounting.  We do this at connect,
2756  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2757  * always get rid of it or turn it off when we know accounting is good. */
2758 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2759 {
2760         struct filter_export_data *fed;
2761         struct obd_export *exp;
2762         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2763         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2764         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2765
2766         if (list_empty(&obd->obd_exports))
2767                 return;
2768
2769         /* We don't want to do this for large machines that do lots of
2770            mounts or unmounts.  It burns... */
2771         if (obd->obd_num_exports > 100)
2772                 return;
2773
2774         spin_lock(&obd->obd_osfs_lock);
2775         spin_lock(&obd->obd_dev_lock);
2776         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2777                 int error = 0;
2778                 fed = &exp->exp_filter_data;
2779                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2780                     fed->fed_dirty < 0)
2781                         error = 1;
2782                 if (maxsize > 0) { /* we may not have done a statfs yet */
2783                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2784                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2785                                  exp->exp_client_uuid.uuid, exp,
2786                                  fed->fed_grant, fed->fed_pending, maxsize);
2787                         LASSERTF(fed->fed_dirty <= maxsize,
2788                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2789                                  exp->exp_client_uuid.uuid, exp,
2790                                  fed->fed_dirty, maxsize);
2791                 }
2792                 if (error)
2793                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2794                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2795                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2796                 else
2797                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2798                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2799                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2800                 tot_granted += fed->fed_grant + fed->fed_pending;
2801                 tot_pending += fed->fed_pending;
2802                 tot_dirty += fed->fed_dirty;
2803         }
2804         fo_tot_granted = obd->u.filter.fo_tot_granted;
2805         fo_tot_pending = obd->u.filter.fo_tot_pending;
2806         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2807         spin_unlock(&obd->obd_dev_lock);
2808         spin_unlock(&obd->obd_osfs_lock);
2809
2810         /* Do these assertions outside the spinlocks so we don't kill system */
2811         if (tot_granted != fo_tot_granted)
2812                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2813                        func, tot_granted, fo_tot_granted);
2814         if (tot_pending != fo_tot_pending)
2815                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2816                        func, tot_pending, fo_tot_pending);
2817         if (tot_dirty != fo_tot_dirty)
2818                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2819                        func, tot_dirty, fo_tot_dirty);
2820         if (tot_pending > tot_granted)
2821                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2822                        func, tot_pending, tot_granted);
2823         if (tot_granted > maxsize)
2824                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2825                        func, tot_granted, maxsize);
2826         if (tot_dirty > maxsize)
2827                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2828                        func, tot_dirty, maxsize);
2829 }
2830
2831 /* Remove this client from the grant accounting totals.  We also remove
2832  * the export from the obd device under the osfs and dev locks to ensure
2833  * that the filter_grant_sanity_check() calculations are always valid.
2834  * The client should do something similar when it invalidates its import. */
2835 static void filter_grant_discard(struct obd_export *exp)
2836 {
2837         struct obd_device *obd = exp->exp_obd;
2838         struct filter_obd *filter = &obd->u.filter;
2839         struct filter_export_data *fed = &exp->exp_filter_data;
2840
2841         spin_lock(&obd->obd_osfs_lock);
2842         spin_lock(&obd->obd_dev_lock);
2843         list_del_init(&exp->exp_obd_chain);
2844         spin_unlock(&obd->obd_dev_lock);
2845
2846         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2847                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2848                  obd->obd_name, filter->fo_tot_granted,
2849                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2850         filter->fo_tot_granted -= fed->fed_grant;
2851         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2852                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2853                  obd->obd_name, filter->fo_tot_pending,
2854                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2855         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2856         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2857                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2858                  obd->obd_name, filter->fo_tot_dirty,
2859                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2860         filter->fo_tot_dirty -= fed->fed_dirty;
2861         fed->fed_dirty = 0;
2862         fed->fed_grant = 0;
2863
2864         spin_unlock(&obd->obd_osfs_lock);
2865 }
2866
2867 static int filter_destroy_export(struct obd_export *exp)
2868 {
2869         ENTRY;
2870
2871         if (exp->exp_filter_data.fed_pending)
2872                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2873                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2874                        exp, exp->exp_filter_data.fed_pending);
2875
2876         /* Not ported yet the b1_6 quota functionality
2877          * lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2878          */
2879
2880         target_destroy_export(exp);
2881
2882         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2883                 RETURN(0);
2884
2885         lprocfs_exp_cleanup(exp);
2886
2887         if (exp->exp_obd->obd_replayable)
2888                 filter_client_free(exp);
2889         else
2890                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2891
2892         filter_grant_discard(exp);
2893         filter_fmd_cleanup(exp);
2894
2895         if (!(exp->exp_flags & OBD_OPT_FORCE))
2896                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2897
2898         RETURN(0);
2899 }
2900
2901 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2902 {
2903         struct obd_llog_group *olg_min, *olg;
2904         struct filter_obd *filter;
2905         int worked = 0, group;
2906         struct llog_ctxt *ctxt;
2907         ENTRY;
2908
2909         filter = &obd->u.filter;
2910
2911         /* we can't sync log holding spinlock. also, we do not want to get
2912          * into livelock. so we do following: loop over MDS's exports in
2913          * group order and skip already synced llogs -bzzz */
2914         do {
2915                 /* look for group with min. number, but > worked */
2916                 olg_min = NULL;
2917                 group = 1 << 30;
2918                 spin_lock(&filter->fo_llog_list_lock);
2919                 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2920                         if (olg->olg_group <= worked) {
2921                                 /* this group is already synced */
2922                                 continue;
2923                         }
2924                         if (group < olg->olg_group) {
2925                                 /* we have group with smaller number to sync */
2926                                 continue;
2927                         }
2928                         /* store current minimal group */
2929                         olg_min = olg;
2930                         group = olg->olg_group;
2931                 }
2932                 spin_unlock(&filter->fo_llog_list_lock);
2933
2934                 if (olg_min == NULL)
2935                         break;
2936
2937                 worked = olg_min->olg_group;
2938                 if (olg_min->olg_exp &&
2939                     (dexp == olg_min->olg_exp || dexp == NULL)) {
2940                         int err;
2941                         ctxt = llog_group_get_ctxt(olg_min,
2942                                                 LLOG_MDS_OST_REPL_CTXT);
2943                         LASSERT(ctxt != NULL);
2944                         err = llog_sync(ctxt, olg_min->olg_exp);
2945                         llog_ctxt_put(ctxt);
2946                         if (err)
2947                                 CERROR("error flushing logs to MDS: rc %d\n",
2948                                        err);                        
2949                 }
2950         } while (olg_min != NULL);
2951 }
2952
2953 /* also incredibly similar to mds_disconnect */
2954 static int filter_disconnect(struct obd_export *exp)
2955 {
2956         struct obd_device *obd = exp->exp_obd;
2957         int rc;
2958         ENTRY;
2959
2960         LASSERT(exp);
2961         class_export_get(exp);
2962
2963         if (!(exp->exp_flags & OBD_OPT_FORCE))
2964                 filter_grant_sanity_check(obd, __FUNCTION__);
2965         filter_grant_discard(exp);
2966
2967         /* Disconnect early so that clients can't keep using export */
2968         rc = class_disconnect(exp);
2969         if (exp->exp_obd->obd_namespace != NULL)
2970                 ldlm_cancel_locks_for_export(exp);
2971
2972         fsfilt_sync(obd, obd->u.obt.obt_sb);
2973
2974         /* flush any remaining cancel messages out to the target */
2975         filter_sync_llogs(obd, exp);
2976         class_export_put(exp);
2977         RETURN(rc);
2978 }
2979
2980 /* reverse import is changed, sync all cancels */
2981 static void filter_revimp_update(struct obd_export *exp)
2982 {
2983         ENTRY;
2984
2985         LASSERT(exp);
2986         class_export_get(exp);
2987
2988         /* flush any remaining cancel messages out to the target */
2989         filter_sync_llogs(exp->exp_obd, exp);
2990         class_export_put(exp);
2991         EXIT;
2992 }
2993
2994 static int filter_ping(struct obd_export *exp)
2995 {
2996         filter_fmd_expire(exp);
2997
2998         return 0;
2999 }
3000
3001 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3002                                   const char *what, int quiet)
3003 {
3004         struct dentry *dchild = NULL;
3005         obd_gr group = 0;
3006
3007         if (oa->o_valid & OBD_MD_FLGROUP)
3008                 group = oa->o_gr;
3009
3010         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3011
3012         if (IS_ERR(dchild)) {
3013                 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3014                        what, group, oa->o_id);
3015                 RETURN(dchild);
3016         }
3017
3018         if (dchild->d_inode == NULL) {
3019                 if (!quiet)
3020                         CERROR("%s: %s on non-existent object: "LPU64"\n",
3021                                obd->obd_name, what, oa->o_id);
3022                 f_dput(dchild);
3023                 RETURN(ERR_PTR(-ENOENT));
3024         }
3025
3026         return dchild;
3027 }
3028
3029 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3030 {
3031         struct dentry *dentry = NULL;
3032         struct obd_device *obd;
3033         int rc = 0;
3034         ENTRY;
3035
3036         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3037                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3038         if (rc)
3039                 RETURN(rc);
3040
3041         obd = class_exp2obd(exp);
3042         if (obd == NULL) {
3043                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3044                 RETURN(-EINVAL);
3045         }
3046
3047         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3048         if (IS_ERR(dentry))
3049                 RETURN(PTR_ERR(dentry));
3050
3051         /* Limit the valid bits in the return data to what we actually use */
3052         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3053         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3054
3055         f_dput(dentry);
3056         RETURN(rc);
3057 }
3058
3059 /* this should be enabled/disabled in condition to enabled/disabled large
3060  * inodes (fast EAs) in backing store FS. */
3061 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3062                         void *handle, struct obdo *oa)
3063 {
3064         struct obd_device *obd = exp->exp_obd;
3065         int rc = 0;
3066         ENTRY;
3067
3068         if (oa->o_valid & OBD_MD_FLFID) {
3069                 struct filter_fid ff;
3070
3071                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3072                         oa->o_gr = 0;
3073                 /* packing fid and converting it to LE for storing into EA.
3074                  * Here ->o_stripe_idx should be filled by LOV and rest of
3075                  * fields - by client. */
3076                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3077                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3078                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3079                 ff.ff_objid = cpu_to_le64(oa->o_id);
3080                 ff.ff_group = cpu_to_le64(oa->o_gr);
3081
3082                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3083                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3084                        oa->o_generation, oa->o_id, oa->o_gr);
3085
3086                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3087                 if (rc)
3088                         CERROR("store fid in object failed! rc: %d\n", rc);
3089         } else {
3090                 CDEBUG(D_HA, "OSS object without fid info!\n");
3091         }
3092
3093         RETURN(rc);
3094 }
3095
3096 /* this is called from filter_truncate() until we have filter_punch() */
3097 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3098                             struct obdo *oa, struct obd_trans_info *oti)
3099 {
3100         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3101         struct llog_cookie *fcc = NULL;
3102         struct filter_obd *filter;
3103         int rc, err, locked = 0, sync = 0;
3104         unsigned int ia_valid;
3105         struct inode *inode;
3106         struct iattr iattr;
3107         void *handle;
3108         ENTRY;
3109
3110         LASSERT(dentry != NULL);
3111         LASSERT(!IS_ERR(dentry));
3112
3113         inode = dentry->d_inode;
3114         LASSERT(inode != NULL);
3115
3116         filter = &exp->exp_obd->u.filter;
3117         iattr_from_obdo(&iattr, oa, oa->o_valid);
3118         ia_valid = iattr.ia_valid;
3119
3120         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3121                 OBD_ALLOC(fcc, sizeof(*fcc));
3122                 if (fcc != NULL)
3123                         memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3124         }
3125
3126         if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
3127                 DQUOT_INIT(inode);
3128                 LOCK_INODE_MUTEX(inode);
3129                 locked = 1;
3130         }
3131
3132         /* If the inode still has SUID+SGID bits set (see filter_precreate())
3133          * then we will accept the UID+GID sent by the client during write for
3134          * initializing the ownership of this inode.  We only allow this to
3135          * happen once so clear these bits in setattr. In 2.6 kernels it is
3136          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3137          * the flags that are actually being set. */
3138         if (ia_valid & (ATTR_UID | ATTR_GID)) {
3139                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3140                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3141
3142                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3143                         if (!(ia_valid & ATTR_MODE)) {
3144                                 iattr.ia_mode = inode->i_mode;
3145                                 iattr.ia_valid |= ATTR_MODE;
3146                         }
3147                         iattr.ia_mode &= ~S_ISUID;
3148                 }
3149                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3150                         if (!(iattr.ia_valid & ATTR_MODE)) {
3151                                 iattr.ia_mode = inode->i_mode;
3152                                 iattr.ia_valid |= ATTR_MODE;
3153                         }
3154                         iattr.ia_mode &= ~S_ISGID;
3155                 }
3156
3157                 orig_ids[USRQUOTA] = inode->i_uid;
3158                 orig_ids[GRPQUOTA] = inode->i_gid;
3159                 handle = fsfilt_start_log(exp->exp_obd, inode,
3160                                           FSFILT_OP_SETATTR, oti, 1);
3161                 if (IS_ERR(handle))
3162                         GOTO(out_unlock, rc = PTR_ERR(handle));
3163
3164                 /* update inode EA only once when inode is suid bit marked. As
3165                  * on 2.6.x UID and GID may be set separately, we check here
3166                  * only one of them to avoid double setting. */
3167                 if (inode->i_mode & S_ISUID)
3168                         filter_update_fidea(exp, inode, handle, oa);
3169         } else {
3170                 handle = fsfilt_start(exp->exp_obd, inode,
3171                                       FSFILT_OP_SETATTR, oti);
3172                 if (IS_ERR(handle))
3173                         GOTO(out_unlock, rc = PTR_ERR(handle));
3174         }
3175         if (oa->o_valid & OBD_MD_FLFLAGS) {
3176                 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3177                                       EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3178         } else {
3179                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3180                 if (fcc != NULL)
3181                         /* set cancel cookie callback function */
3182                         sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3183                                                      filter_cancel_cookies_cb,
3184                                                      fcc);
3185         }
3186
3187         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3188                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3189
3190         /* The truncate might have used up our transaction credits.  Make
3191          * sure we have one left for the last_rcvd update. */
3192         err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3193
3194         rc = filter_finish_transno(exp, oti, rc, sync);
3195         if (sync) {
3196                 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3197                 fcc = NULL;