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