Whamcloud - gitweb
Branch HEAD
[fs/lustre-release.git] / lustre / obdfilter / filter.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/obdfilter/filter.c
37  *
38  * Author: Peter Braam <braam@clusterfs.com>
39  * Author: Andreas Dilger <adilger@clusterfs.com>
40  */
41
42 /*
43  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44  *            (which need to get journal_lock, may block if journal full).
45  *
46  * Invariant: Call filter_start_transno() before any journal ops to avoid the
47  *            same deadlock problem.  We can (and want) to get rid of the
48  *            transno sem in favour of the dir/inode i_mutex to avoid single
49  *            threaded operation on the OST.
50  */
51
52 #define DEBUG_SUBSYSTEM S_FILTER
53
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
56 #endif
57 #include <linux/module.h>
58 #include <linux/fs.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
65
66 #include <obd_cksum.h>
67 #include <obd_class.h>
68 #include <obd_lov.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
79
80 #include "filter_internal.h"
81
82 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
83 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS1_N_BASE
84 static struct lvfs_callback_ops filter_lvfs_ops;
85 cfs_mem_cache_t *ll_fmd_cachep;
86
87 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
88                              void *cb_data, int error)
89 {
90         struct obd_export *exp = cb_data;
91         LASSERT(exp->exp_obd == obd);
92         obd_transno_commit_cb(obd, transno, exp, error);
93         atomic_dec(&exp->exp_cb_count);
94         class_export_put(exp);
95 }
96
97 int filter_version_get_check(struct obd_export *exp,
98                              struct obd_trans_info *oti, struct inode *inode)
99 {
100         __u64 curr_version;
101
102         if (inode == NULL || oti == NULL)
103                 RETURN(0);
104
105         curr_version = fsfilt_get_version(exp->exp_obd, inode);
106         if ((__s64)curr_version == -EOPNOTSUPP)
107                 RETURN(0);
108         /* VBR: version is checked always because costs nothing */
109         if (oti->oti_pre_version != 0 &&
110             oti->oti_pre_version != curr_version) {
111                 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
112                        oti->oti_pre_version, curr_version);
113                 spin_lock(&exp->exp_lock);
114                 exp->exp_vbr_failed = 1;
115                 spin_unlock(&exp->exp_lock);
116                 RETURN (-EOVERFLOW);
117         }
118         oti->oti_pre_version = curr_version;
119         RETURN(0);
120 }
121
122 /* Assumes caller has already pushed us into the kernel context. */
123 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
124                           struct obd_trans_info *oti, int rc, int force_sync)
125 {
126         struct filter_obd *filter = &exp->exp_obd->u.filter;
127         struct filter_export_data *fed = &exp->exp_filter_data;
128         struct lsd_client_data *lcd = fed->fed_lcd;
129         __u64 last_rcvd;
130         loff_t off;
131         int err, log_pri = D_RPCTRACE;
132
133         /* Propagate error code. */
134         if (rc)
135                 RETURN(rc);
136
137         if (!exp->exp_obd->obd_replayable || oti == NULL)
138                 RETURN(rc);
139
140         /* we don't allocate new transnos for replayed requests */
141         spin_lock(&filter->fo_translock);
142         if (oti->oti_transno == 0) {
143                 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
144                 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
145         } else {
146                 last_rcvd = oti->oti_transno;
147                 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
148                         filter->fo_fsd->lsd_last_transno =
149                                 cpu_to_le64(last_rcvd);
150         }
151         oti->oti_transno = last_rcvd;
152         if (last_rcvd <= le64_to_cpu(lcd->lcd_last_transno)) {
153                 spin_unlock(&filter->fo_translock);
154                 LBUG();
155         }
156         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
157         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
158         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
159         spin_unlock(&filter->fo_translock);
160
161         if (inode)
162                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
163
164         off = fed->fed_lr_off;
165         if (off <= 0) {
166                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
167                        fed->fed_lr_idx, fed->fed_lr_off);
168                 err = -EINVAL;
169         } else {
170                 class_export_get(exp); /* released when the cb is called */
171                 atomic_inc(&exp->exp_cb_count);
172                 if (!force_sync)
173                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
174                                                            last_rcvd,
175                                                            oti->oti_handle,
176                                                            filter_commit_cb,
177                                                            exp);
178
179                 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
180                                           lcd, sizeof(*lcd), &off,
181                                           force_sync | exp->exp_need_sync);
182                 if (force_sync)
183                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
184         }
185         if (err) {
186                 log_pri = D_ERROR;
187                 if (rc == 0)
188                         rc = err;
189         }
190
191         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
192                last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
193
194         RETURN(rc);
195 }
196
197 void f_dput(struct dentry *dentry)
198 {
199         /* Can't go inside filter_ddelete because it can block */
200         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
201                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
202         LASSERT(atomic_read(&dentry->d_count) > 0);
203
204         dput(dentry);
205 }
206
207 static void init_brw_stats(struct brw_stats *brw_stats)
208 {
209         int i;
210         for (i = 0; i < BRW_LAST; i++)
211                 spin_lock_init(&brw_stats->hist[i].oh_lock);
212 }
213
214 static int lprocfs_init_rw_stats(struct obd_device *obd,
215                                  struct lprocfs_stats **stats)
216 {
217         int num_stats;
218
219         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
220                                                         LPROC_FILTER_LAST - 1;
221         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
222         if (*stats == NULL)
223                 return -ENOMEM;
224
225         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
226         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
227                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
228         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
229                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
230
231         return(0);
232 }
233
234 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
235    plus the procfs overhead :( */
236 static int filter_export_stats_init(struct obd_device *obd,
237                                     struct obd_export *exp,
238                                     void *client_nid)
239 {
240         int rc, newnid = 0;
241         ENTRY;
242
243         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
244                 /* Self-export gets no proc entry */
245                 RETURN(0);
246
247         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
248         if (rc) {
249                 /* Mask error for already created
250                  * /proc entries */
251                 if (rc == -EALREADY)
252                         rc = 0;
253                 RETURN(rc);
254         }
255
256         if (newnid) {
257                 struct nid_stat *tmp = exp->exp_nid_stats;
258                 LASSERT(tmp != NULL);
259
260                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
261                 if (tmp->nid_brw_stats == NULL)
262                         RETURN(-ENOMEM);
263
264                 init_brw_stats(tmp->nid_brw_stats);
265                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
266                                         0644, &filter_per_nid_stats_fops,
267                                         exp->exp_nid_stats);
268                 if (rc)
269                         CWARN("Error adding the brw_stats file\n");
270
271                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
272                 if (rc)
273                         RETURN(rc);
274
275                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
276                                             tmp->nid_stats);
277                 if (rc)
278                         RETURN(rc);
279                 /* Always add in ldlm_stats */
280                 tmp->nid_ldlm_stats = 
281                         lprocfs_alloc_stats(LDLM_LAST_OPC - LDLM_FIRST_OPC,
282                                             LPROCFS_STATS_FLAG_NOPERCPU);
283                 if (tmp->nid_ldlm_stats == NULL)
284                         return -ENOMEM;
285
286                 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
287                 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
288                                             tmp->nid_ldlm_stats);
289                 if (rc)
290                         RETURN(rc);
291         }
292
293         RETURN(0);
294 }
295
296 /* Add client data to the FILTER.  We use a bitmap to locate a free space
297  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
298  * Otherwise, we have just read the data from the last_rcvd file and
299  * we know its offset. */
300 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
301                              int cl_idx)
302 {
303         struct filter_obd *filter = &obd->u.filter;
304         struct filter_export_data *fed = &exp->exp_filter_data;
305         unsigned long *bitmap = filter->fo_last_rcvd_slots;
306         int new_client = (cl_idx == -1);
307
308         ENTRY;
309
310         LASSERT(bitmap != NULL);
311         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
312
313         /* Self-export */
314         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
315                 RETURN(0);
316
317         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
318          * there's no need for extra complication here
319          */
320         if (new_client) {
321                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
322         repeat:
323                 if (cl_idx >= LR_MAX_CLIENTS) {
324                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
325                                cl_idx);
326                         RETURN(-EOVERFLOW);
327                 }
328                 if (test_and_set_bit(cl_idx, bitmap)) {
329                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
330                                                     cl_idx);
331                         goto repeat;
332                 }
333         } else {
334                 if (test_and_set_bit(cl_idx, bitmap)) {
335                         CERROR("FILTER client %d: bit already set in bitmap!\n",
336                                cl_idx);
337                         LBUG();
338                 }
339         }
340
341         fed->fed_lr_idx = cl_idx;
342         fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
343                 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
344         LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
345
346         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
347                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
348
349         if (new_client) {
350                 struct lvfs_run_ctxt saved;
351                 loff_t off = fed->fed_lr_off;
352                 int rc;
353                 void *handle;
354
355                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
356                        fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
357
358                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
359                 /* Transaction needed to fix bug 1403 */
360                 handle = fsfilt_start(obd,
361                                       filter->fo_rcvd_filp->f_dentry->d_inode,
362                                       FSFILT_OP_SETATTR, NULL);
363                 if (IS_ERR(handle)) {
364                         rc = PTR_ERR(handle);
365                         CERROR("unable to start transaction: rc %d\n", rc);
366                 } else {
367                         fed->fed_lcd->lcd_last_epoch =
368                                               filter->fo_fsd->lsd_start_epoch;
369                         exp->exp_last_request_time = cfs_time_current_sec();
370                         rc = fsfilt_add_journal_cb(obd, 0, handle,
371                                                    target_client_add_cb, exp);
372                         if (rc == 0) {
373                                 spin_lock(&exp->exp_lock);
374                                 exp->exp_need_sync = 1;
375                                 spin_unlock(&exp->exp_lock);
376                         }
377                         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
378                                                  fed->fed_lcd,
379                                                  sizeof(*fed->fed_lcd),
380                                                  &off, rc /* sync if no cb */);
381                         fsfilt_commit(obd,
382                                       filter->fo_rcvd_filp->f_dentry->d_inode,
383                                       handle, 0);
384                 }
385                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
386
387                 if (rc) {
388                         CERROR("error writing %s client idx %u: rc %d\n",
389                                LAST_RCVD, fed->fed_lr_idx, rc);
390                         RETURN(rc);
391                 }
392         }
393         RETURN(0);
394 }
395
396 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
397
398 static int filter_client_free(struct obd_export *exp)
399 {
400         struct filter_export_data *fed = &exp->exp_filter_data;
401         struct filter_obd *filter = &exp->exp_obd->u.filter;
402         struct obd_device *obd = exp->exp_obd;
403         struct lvfs_run_ctxt saved;
404         int rc;
405         loff_t off;
406         ENTRY;
407
408         if (fed->fed_lcd == NULL)
409                 RETURN(0);
410
411         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
412         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
413                 GOTO(free, 0);
414
415         LASSERT(filter->fo_last_rcvd_slots != NULL);
416
417         off = fed->fed_lr_off;
418
419         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
420                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
421
422         /* Don't clear fed_lr_idx here as it is likely also unset.  At worst
423          * we leak a client slot that will be cleaned on the next recovery. */
424         if (off <= 0) {
425                 CERROR("%s: client idx %d has med_off %lld\n",
426                        obd->obd_name, fed->fed_lr_idx, off);
427                 GOTO(free, rc = -EINVAL);
428         }
429
430         /* Clear the bit _after_ zeroing out the client so we don't
431            race with filter_client_add and zero out new clients.*/
432         if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
433                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
434                        fed->fed_lr_idx);
435                 LBUG();
436         }
437
438         if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
439                 /* Don't force sync on disconnect if aborting recovery,
440                  * or it does num_clients * num_osts.  b=17194 */
441                 int need_sync = exp->exp_need_sync &&
442                                  !(exp->exp_flags&OBD_OPT_ABORT_RECOV);
443                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
444                 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
445                                          sizeof(zero_lcd), &off, 0);
446
447                 /* Make sure the server's last_transno is up to date. Do this
448                  * after the client is freed so we know all the client's
449                  * transactions have been committed. */
450                 if (rc == 0)
451                         filter_update_server_data(obd, filter->fo_rcvd_filp,
452                                                   filter->fo_fsd, need_sync);
453                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
454
455                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
456                        "zero out client %s at idx %u/%llu in %s %ssync rc %d\n",
457                        fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
458                        LAST_RCVD, need_sync ? "" : "a", rc);
459         }
460
461         if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
462                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
463                        fed->fed_lr_idx);
464                 LBUG();
465         }
466
467         EXIT;
468 free:
469         OBD_FREE_PTR(fed->fed_lcd);
470         fed->fed_lcd = NULL;
471
472         return 0;
473 }
474
475 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
476 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
477                                          struct filter_mod_data *fmd)
478 {
479         LASSERT_SPIN_LOCKED(&fed->fed_lock);
480         if (--fmd->fmd_refcount == 0) {
481                 /* XXX when we have persistent reservations and the handle
482                  * is stored herein we need to drop it here. */
483                 fed->fed_mod_count--;
484                 list_del(&fmd->fmd_list);
485                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
486         }
487 }
488
489 /* drop fmd reference, free it if last ref */
490 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
491 {
492         struct filter_export_data *fed;
493
494         if (fmd == NULL)
495                 return;
496
497         fed = &exp->exp_filter_data;
498         spin_lock(&fed->fed_lock);
499         filter_fmd_put_nolock(fed, fmd); /* caller reference */
500         spin_unlock(&fed->fed_lock);
501 }
502
503 /* expire entries from the end of the list if there are too many
504  * or they are too old */
505 static void filter_fmd_expire_nolock(struct filter_obd *filter,
506                                      struct filter_export_data *fed,
507                                      struct filter_mod_data *keep)
508 {
509         struct filter_mod_data *fmd, *tmp;
510
511         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
512                 if (fmd == keep)
513                         break;
514
515                 if (time_before(jiffies, fmd->fmd_expire) &&
516                     fed->fed_mod_count < filter->fo_fmd_max_num)
517                         break;
518
519                 list_del_init(&fmd->fmd_list);
520                 filter_fmd_put_nolock(fed, fmd); /* list reference */
521         }
522 }
523
524 void filter_fmd_expire(struct obd_export *exp)
525 {
526         spin_lock(&exp->exp_filter_data.fed_lock);
527         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
528                                  &exp->exp_filter_data, NULL);
529         spin_unlock(&exp->exp_filter_data.fed_lock);
530 }
531
532 /* find specified objid, group in export fmd list.
533  * caller must hold fed_lock and take fmd reference itself */
534 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
535                                                 struct filter_export_data *fed,
536                                                 obd_id objid, obd_gr group)
537 {
538         struct filter_mod_data *found = NULL, *fmd;
539
540         LASSERT_SPIN_LOCKED(&fed->fed_lock);
541
542         list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
543                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
544                         found = fmd;
545                         list_del(&fmd->fmd_list);
546                         list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
547                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
548                         break;
549                 }
550         }
551
552         filter_fmd_expire_nolock(filter, fed, found);
553
554         return found;
555 }
556
557 /* Find fmd based on objid and group, or return NULL if not found. */
558 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
559                                         obd_id objid, obd_gr group)
560 {
561         struct filter_mod_data *fmd;
562
563         spin_lock(&exp->exp_filter_data.fed_lock);
564         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
565                                      &exp->exp_filter_data, objid, group);
566         if (fmd)
567                 fmd->fmd_refcount++;    /* caller reference */
568         spin_unlock(&exp->exp_filter_data.fed_lock);
569
570         return fmd;
571 }
572
573 /* Find fmd based on objid and group, or create a new one if none is found.
574  * It is possible for this function to return NULL under memory pressure,
575  * or if objid = 0 is passed (which will only cause old entries to expire).
576  * Currently this is not fatal because any fmd state is transient and
577  * may also be freed when it gets sufficiently old. */
578 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
579                                        obd_id objid, obd_gr group)
580 {
581         struct filter_export_data *fed = &exp->exp_filter_data;
582         struct filter_mod_data *found = NULL, *fmd_new = NULL;
583
584         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
585
586         spin_lock(&fed->fed_lock);
587         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
588         if (fmd_new) {
589                 if (found == NULL) {
590                         list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
591                         fmd_new->fmd_id = objid;
592                         fmd_new->fmd_gr = group;
593                         fmd_new->fmd_refcount++;   /* list reference */
594                         found = fmd_new;
595                         fed->fed_mod_count++;
596                 } else {
597                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
598                 }
599         }
600         if (found) {
601                 found->fmd_refcount++;          /* caller reference */
602                 found->fmd_expire = jiffies +
603                         exp->exp_obd->u.filter.fo_fmd_max_age;
604         }
605
606         spin_unlock(&fed->fed_lock);
607
608         return found;
609 }
610
611 #ifdef DO_FMD_DROP
612 /* drop fmd list reference so it will disappear when last reference is put.
613  * This isn't so critical because it would in fact only affect the one client
614  * that is doing the unlink and at worst we have an stale entry referencing
615  * an object that should never be used again. */
616 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
617 {
618         struct filter_mod_data *found = NULL;
619
620         spin_lock(&exp->exp_filter_data.fed_lock);
621         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
622         if (found) {
623                 list_del_init(&found->fmd_list);
624                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
625         }
626         spin_unlock(&exp->exp_filter_data.fed_lock);
627 }
628 #else
629 #define filter_fmd_drop(exp, objid, group)
630 #endif
631
632 /* remove all entries from fmd list */
633 static void filter_fmd_cleanup(struct obd_export *exp)
634 {
635         struct filter_export_data *fed = &exp->exp_filter_data;
636         struct filter_mod_data *fmd = NULL, *tmp;
637
638         spin_lock(&fed->fed_lock);
639         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
640                 list_del_init(&fmd->fmd_list);
641                 filter_fmd_put_nolock(fed, fmd);
642         }
643         spin_unlock(&fed->fed_lock);
644 }
645
646 static int filter_init_export(struct obd_export *exp)
647 {
648         spin_lock_init(&exp->exp_filter_data.fed_lock);
649         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
650
651         spin_lock(&exp->exp_lock);
652         exp->exp_connecting = 1;
653         spin_unlock(&exp->exp_lock);
654
655         return ldlm_init_export(exp);
656 }
657
658 static int filter_free_server_data(struct filter_obd *filter)
659 {
660         OBD_FREE_PTR(filter->fo_fsd);
661         filter->fo_fsd = NULL;
662         OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
663         filter->fo_last_rcvd_slots = NULL;
664         return 0;
665 }
666
667 /* assumes caller is already in kernel ctxt */
668 int filter_update_server_data(struct obd_device *obd, struct file *filp,
669                               struct lr_server_data *fsd, int force_sync)
670 {
671         loff_t off = 0;
672         int rc;
673         ENTRY;
674
675         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
676         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
677                le64_to_cpu(fsd->lsd_last_transno));
678         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
679                le64_to_cpu(fsd->lsd_mount_count));
680
681         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
682         if (rc)
683                 CERROR("error writing lr_server_data: rc = %d\n", rc);
684
685         RETURN(rc);
686 }
687
688 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
689                              int force_sync)
690 {
691         struct filter_obd *filter = &obd->u.filter;
692         __u64 tmp;
693         loff_t off = 0;
694         int rc;
695         ENTRY;
696
697         if (filter->fo_last_objid_files[group] == NULL) {
698                 CERROR("Object group "LPU64" not fully setup; not updating "
699                        "last_objid\n", group);
700                 RETURN(-EINVAL);
701         }
702
703         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
704                obd->obd_name, group, filter->fo_last_objids[group]);
705
706         tmp = cpu_to_le64(filter->fo_last_objids[group]);
707         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
708                                  &tmp, sizeof(tmp), &off, force_sync);
709         if (rc)
710                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
711                        group, rc);
712         RETURN(rc);
713 }
714 extern int ost_handle(struct ptlrpc_request *req);
715 /* assumes caller has already in kernel ctxt */
716 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
717 {
718         struct filter_obd *filter = &obd->u.filter;
719         struct lr_server_data *fsd;
720         struct lsd_client_data *lcd = NULL;
721         struct inode *inode = filp->f_dentry->d_inode;
722         unsigned long last_rcvd_size = i_size_read(inode);
723         __u64 mount_count;
724         __u32 start_epoch;
725         int cl_idx;
726         loff_t off = 0;
727         int rc;
728
729         /* ensure padding in the struct is the correct size */
730         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
731                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
732         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
733                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
734
735         OBD_ALLOC(fsd, sizeof(*fsd));
736         if (!fsd)
737                 RETURN(-ENOMEM);
738         filter->fo_fsd = fsd;
739
740         OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
741         if (filter->fo_last_rcvd_slots == NULL) {
742                 OBD_FREE(fsd, sizeof(*fsd));
743                 RETURN(-ENOMEM);
744         }
745
746         if (last_rcvd_size == 0) {
747                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
748
749                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
750                 fsd->lsd_last_transno = 0;
751                 mount_count = fsd->lsd_mount_count = 0;
752                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
753                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
754                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
755                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
756                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
757                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
758         } else {
759                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
760                 if (rc) {
761                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
762                                LAST_RCVD, rc);
763                         GOTO(err_fsd, rc);
764                 }
765                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
766                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
767                                            "using the wrong disk %s. Were the "
768                                            "/dev/ assignments rearranged?\n",
769                                            obd->obd_uuid.uuid, fsd->lsd_uuid);
770                         GOTO(err_fsd, rc = -EINVAL);
771                 }
772                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
773                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
774                 /* COMPAT_146 */
775                 /* Assume old last_rcvd format unless I_C_LR is set */
776                 if (!(fsd->lsd_feature_incompat &
777                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
778                         fsd->lsd_last_transno = fsd->lsd_compat14;
779                 /* end COMPAT_146 */
780         }
781
782         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
783                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
784                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
785                        ~FILTER_INCOMPAT_SUPP);
786                 GOTO(err_fsd, rc = -EINVAL);
787         }
788         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
789                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
790                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
791                        ~FILTER_ROCOMPAT_SUPP);
792                 /* Do something like remount filesystem read-only */
793                 GOTO(err_fsd, rc = -EINVAL);
794         }
795
796         start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
797
798         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
799                obd->obd_name, start_epoch);
800         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
801                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
802         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
803                obd->obd_name, mount_count + 1);
804         CDEBUG(D_INODE, "%s: server data size: %u\n",
805                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
806         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
807                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
808         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
809                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
810         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
811                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
812         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
813                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
814                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
815                 le16_to_cpu(fsd->lsd_client_size));
816
817         if (!obd->obd_replayable) {
818                 CWARN("%s: recovery support OFF\n", obd->obd_name);
819                 GOTO(out, rc = 0);
820         }
821
822         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
823              off < last_rcvd_size; cl_idx++) {
824                 __u64 last_rcvd;
825                 struct obd_export *exp;
826                 struct filter_export_data *fed;
827
828                 if (!lcd) {
829                         OBD_ALLOC_PTR(lcd);
830                         if (!lcd)
831                                 GOTO(err_client, rc = -ENOMEM);
832                 }
833
834                 /* Don't assume off is incremented properly by
835                  * fsfilt_read_record(), in case sizeof(*lcd)
836                  * isn't the same as fsd->lsd_client_size.  */
837                 off = le32_to_cpu(fsd->lsd_client_start) +
838                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
839                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
840                 if (rc) {
841                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
842                                LAST_RCVD, cl_idx, off, rc);
843                         break; /* read error shouldn't cause startup to fail */
844                 }
845
846                 if (lcd->lcd_uuid[0] == '\0') {
847                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
848                                cl_idx);
849                         continue;
850                 }
851
852                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
853
854                 /* These exports are cleaned up by filter_disconnect(), so they
855                  * need to be set up like real exports as filter_connect() does.
856                  */
857                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
858
859                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
860                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
861                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
862                 if (IS_ERR(exp)) {
863                         if (PTR_ERR(exp) == -EALREADY) {
864                                 /* export already exists, zero out this one */
865                                 CERROR("Zeroing out duplicate export due to "
866                                        "bug 10479.\n");
867                                 lcd->lcd_uuid[0] = '\0';
868                         } else {
869                                 GOTO(err_client, rc = PTR_ERR(exp));
870                         }
871                 } else {
872                         fed = &exp->exp_filter_data;
873                         fed->fed_lcd = lcd;
874                         fed->fed_group = 0; /* will be assigned at connect */
875                         filter_export_stats_init(obd, exp, NULL);
876                         rc = filter_client_add(obd, exp, cl_idx);
877                         /* can't fail for existing client */
878                         LASSERTF(rc == 0, "rc = %d\n", rc);
879
880                         /* VBR: set export last committed */
881                         exp->exp_last_committed = last_rcvd;
882                         spin_lock(&exp->exp_lock);
883                         exp->exp_connecting = 0;
884                         exp->exp_in_recovery = 0;
885                         spin_unlock(&exp->exp_lock);
886                         spin_lock_bh(&obd->obd_processing_task_lock);
887                         obd->obd_max_recoverable_clients++;
888                         spin_unlock_bh(&obd->obd_processing_task_lock);
889                         lcd = NULL;
890                         class_export_put(exp);
891                 }
892
893                 /* Need to check last_rcvd even for duplicated exports. */
894                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
895                        cl_idx, last_rcvd);
896
897                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
898                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
899         }
900
901         if (lcd)
902                 OBD_FREE_PTR(lcd);
903
904         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
905
906         target_recovery_init(&filter->fo_lut, ost_handle);
907
908 out:
909         filter->fo_mount_count = mount_count + 1;
910         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
911
912         /* save it, so mount count and last_transno is current */
913         rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
914         if (rc)
915                 GOTO(err_client, rc);
916
917         RETURN(0);
918
919 err_client:
920         target_recovery_fini(obd);
921 err_fsd:
922         filter_free_server_data(filter);
923         RETURN(rc);
924 }
925
926 static int filter_cleanup_groups(struct obd_device *obd)
927 {
928         struct filter_obd *filter = &obd->u.filter;
929         struct file *filp;
930         struct dentry *dentry;
931         int i, j;
932         ENTRY;
933
934         if (filter->fo_dentry_O_groups != NULL) {
935                 for (i = 0; i < filter->fo_group_count; i++) {
936                         dentry = filter->fo_dentry_O_groups[i];
937                         if (dentry != NULL)
938                                 f_dput(dentry);
939                 }
940                 OBD_FREE(filter->fo_dentry_O_groups,
941                          filter->fo_group_count *
942                          sizeof(*filter->fo_dentry_O_groups));
943                 filter->fo_dentry_O_groups = NULL;
944         }
945         if (filter->fo_last_objid_files != NULL) {
946                 for (i = 0; i < filter->fo_group_count; i++) {
947                         filp = filter->fo_last_objid_files[i];
948                         if (filp != NULL)
949                                 filp_close(filp, 0);
950                 }
951                 OBD_FREE(filter->fo_last_objid_files,
952                          filter->fo_group_count *
953                          sizeof(*filter->fo_last_objid_files));
954                 filter->fo_last_objid_files = NULL;
955         }
956         if (filter->fo_dentry_O_sub != NULL) {
957                 for (i = 0; i < filter->fo_group_count; i++) {
958                         for (j = 0; j < filter->fo_subdir_count; j++) {
959                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
960                                 if (dentry != NULL)
961                                         f_dput(dentry);
962                         }
963                 }
964                 OBD_FREE(filter->fo_dentry_O_sub,
965                          filter->fo_group_count *
966                          sizeof(*filter->fo_dentry_O_sub));
967                 filter->fo_dentry_O_sub = NULL;
968         }
969         if (filter->fo_last_objids != NULL) {
970                 OBD_FREE(filter->fo_last_objids,
971                          filter->fo_group_count *
972                          sizeof(*filter->fo_last_objids));
973                 filter->fo_last_objids = NULL;
974         }
975         if (filter->fo_dentry_O != NULL) {
976                 f_dput(filter->fo_dentry_O);
977                 filter->fo_dentry_O = NULL;
978         }
979         RETURN(0);
980 }
981
982 static int filter_update_last_group(struct obd_device *obd, int group)
983 {
984         struct filter_obd *filter = &obd->u.filter;
985         struct file *filp = NULL;
986         int last_group = 0, rc;
987         loff_t off = 0;
988         ENTRY;
989
990         if (group <= filter->fo_committed_group)
991                 RETURN(0);
992
993         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
994         if (IS_ERR(filp)) {
995                 rc = PTR_ERR(filp);
996                 filp = NULL;
997                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
998                 GOTO(cleanup, rc);
999         }
1000
1001         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1002         if (rc) {
1003                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1004                 GOTO(cleanup, rc);
1005         }
1006         LASSERTF(off == 0 || CHECK_MDS_GROUP(last_group),
1007                  "off = %llu and last_group = %d\n", off, last_group);
1008
1009         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1010                obd->obd_name, last_group, group);
1011
1012         off = 0;
1013         last_group = group;
1014         /* must be sync: bXXXX */
1015         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1016         if (rc) {
1017                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1018                 GOTO(cleanup, rc);
1019         }
1020
1021         filter->fo_committed_group = group;
1022 cleanup:
1023         if (filp)
1024                 filp_close(filp, 0);
1025         RETURN(rc);
1026 }
1027
1028 static int filter_read_group_internal(struct obd_device *obd, int group,
1029                                       int create)
1030 {
1031         struct filter_obd *filter = &obd->u.filter;
1032         __u64 *new_objids = NULL;
1033         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1034         struct dentry **new_groups = NULL;
1035         struct file **new_files = NULL;
1036         struct dentry *dentry;
1037         struct file *filp;
1038         int old_count = filter->fo_group_count, rc, stage = 0, i;
1039         char name[25];
1040         __u64 last_objid;
1041         loff_t off = 0;
1042         int len = group + 1;
1043
1044         snprintf(name, 24, "%d", group);
1045         name[24] = '\0';
1046
1047         if (!create) {
1048                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1049                                            strlen(name));
1050                 if (IS_ERR(dentry)) {
1051                         CERROR("Cannot lookup expected object group %d: %ld\n",
1052                                group, PTR_ERR(dentry));
1053                         RETURN(PTR_ERR(dentry));
1054                 }
1055         } else {
1056                 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1057                                       name, 0700, 1);
1058                 if (IS_ERR(dentry)) {
1059                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1060                                PTR_ERR(dentry));
1061                         RETURN(PTR_ERR(dentry));
1062                 }
1063         }
1064         stage = 1;
1065
1066         snprintf(name, 24, "O/%d/LAST_ID", group);
1067         name[24] = '\0';
1068         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1069         if (IS_ERR(filp)) {
1070                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1071                 GOTO(cleanup, rc = PTR_ERR(filp));
1072         }
1073         stage = 2;
1074
1075         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1076         if (rc) {
1077                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1078                 GOTO(cleanup, rc);
1079         }
1080
1081         if (filter->fo_subdir_count) {
1082                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1083                 if (tmp_subdirs == NULL)
1084                         GOTO(cleanup, rc = -ENOMEM);
1085                 stage = 3;
1086
1087                 for (i = 0; i < filter->fo_subdir_count; i++) {
1088                         char dir[20];
1089                         snprintf(dir, sizeof(dir), "d%u", i);
1090
1091                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1092                                                               filter->fo_vfsmnt,
1093                                                               dir, 0700, 1);
1094                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1095                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1096                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1097                                        group, dir, rc);
1098                                 GOTO(cleanup, rc);
1099                         }
1100
1101                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1102                                tmp_subdirs->dentry[i]);
1103                 }
1104         }
1105
1106         /* 'group' is an index; we need an array of length 'group + 1' */
1107         if (group + 1 > old_count) {
1108                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1109                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1110                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1111                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1112                 stage = 4;
1113                 if (new_objids == NULL || new_subdirs == NULL ||
1114                     new_groups == NULL || new_files == NULL)
1115                         GOTO(cleanup, rc = -ENOMEM);
1116
1117                 if (old_count) {
1118                         memcpy(new_objids, filter->fo_last_objids,
1119                                old_count * sizeof(*new_objids));
1120                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1121                                old_count * sizeof(*new_subdirs));
1122                         memcpy(new_groups, filter->fo_dentry_O_groups,
1123                                old_count * sizeof(*new_groups));
1124                         memcpy(new_files, filter->fo_last_objid_files,
1125                                old_count * sizeof(*new_files));
1126
1127                         OBD_FREE(filter->fo_last_objids,
1128                                  old_count * sizeof(*new_objids));
1129                         OBD_FREE(filter->fo_dentry_O_sub,
1130                                  old_count * sizeof(*new_subdirs));
1131                         OBD_FREE(filter->fo_dentry_O_groups,
1132                                  old_count * sizeof(*new_groups));
1133                         OBD_FREE(filter->fo_last_objid_files,
1134                                  old_count * sizeof(*new_files));
1135                 }
1136                 filter->fo_last_objids = new_objids;
1137                 filter->fo_dentry_O_sub = new_subdirs;
1138                 filter->fo_dentry_O_groups = new_groups;
1139                 filter->fo_last_objid_files = new_files;
1140                 filter->fo_group_count = len;
1141         }
1142
1143         filter->fo_dentry_O_groups[group] = dentry;
1144         filter->fo_last_objid_files[group] = filp;
1145         if (filter->fo_subdir_count) {
1146                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1147                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1148         }
1149
1150         filter_update_last_group(obd, group);
1151
1152         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1153                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1154                 rc = filter_update_last_objid(obd, group, 1);
1155                 RETURN(rc);
1156         }
1157
1158         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1159         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1160                obd->obd_name, group, last_objid);
1161         RETURN(0);
1162  cleanup:
1163         switch (stage) {
1164         case 4:
1165                 if (new_objids != NULL)
1166                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1167                 if (new_subdirs != NULL)
1168                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1169                 if (new_groups != NULL)
1170                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1171                 if (new_files != NULL)
1172                         OBD_FREE(new_files, len * sizeof(*new_files));
1173         case 3:
1174                 if (filter->fo_subdir_count) {
1175                         for (i = 0; i < filter->fo_subdir_count; i++) {
1176                                 if (tmp_subdirs->dentry[i] != NULL)
1177                                         dput(tmp_subdirs->dentry[i]);
1178                         }
1179                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1180                 }
1181         case 2:
1182                 filp_close(filp, 0);
1183         case 1:
1184                 dput(dentry);
1185         }
1186         RETURN(rc);
1187 }
1188
1189 static int filter_read_groups(struct obd_device *obd, int last_group,
1190                               int create)
1191 {
1192         struct filter_obd *filter = &obd->u.filter;
1193         int old_count, group, rc = 0;
1194
1195         down(&filter->fo_init_lock);
1196         old_count = filter->fo_group_count;
1197         for (group = old_count; group <= last_group; group++) {
1198
1199                 rc = filter_read_group_internal(obd, group, create);
1200                 if (rc != 0)
1201                         break;
1202         }
1203         up(&filter->fo_init_lock);
1204         return rc;
1205 }
1206
1207 /* FIXME: object groups */
1208 static int filter_prep_groups(struct obd_device *obd)
1209 {
1210         struct filter_obd *filter = &obd->u.filter;
1211         struct dentry *dentry, *O_dentry;
1212         struct file *filp;
1213         int    last_group, rc = 0, cleanup_phase = 0;
1214         loff_t off = 0;
1215         ENTRY;
1216
1217         O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1218                                 "O", 0700, 1);
1219         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1220         if (IS_ERR(O_dentry)) {
1221                 rc = PTR_ERR(O_dentry);
1222                 CERROR("cannot open/create O: rc = %d\n", rc);
1223                 GOTO(cleanup, rc);
1224         }
1225         filter->fo_dentry_O = O_dentry;
1226         cleanup_phase = 1; /* O_dentry */
1227
1228         /* Lookup "R" to tell if we're on an old OST FS and need to convert
1229          * from O/R/<dir>/<objid> to O/0/<dir>/<objid>.  This can be removed
1230          * some time post 1.0 when all old-style OSTs have converted along
1231          * with the init_objid hack. */
1232         dentry = ll_lookup_one_len("R", O_dentry, 1);
1233         if (IS_ERR(dentry))
1234                 GOTO(cleanup, rc = PTR_ERR(dentry));
1235         if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1236                 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1237                 ENTRY;
1238
1239                 CWARN("converting OST to new object layout\n");
1240                 if (IS_ERR(O0_dentry)) {
1241                         rc = PTR_ERR(O0_dentry);
1242                         CERROR("error looking up O/0: rc %d\n", rc);
1243                         GOTO(cleanup_R, rc);
1244                 }
1245
1246                 if (O0_dentry->d_inode) {
1247                         CERROR("Both O/R and O/0 exist. Fix manually.\n");
1248                         GOTO(cleanup_O0, rc = -EEXIST);
1249                 }
1250
1251                 LOCK_INODE_MUTEX(O_dentry->d_inode);
1252                 rc = ll_vfs_rename(O_dentry->d_inode, dentry, filter->fo_vfsmnt,
1253                                    O_dentry->d_inode, O0_dentry,
1254                                    filter->fo_vfsmnt);
1255                 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1256
1257                 if (rc) {
1258                         CERROR("error renaming O/R to O/0: rc %d\n", rc);
1259                         GOTO(cleanup_O0, rc);
1260                 }
1261                 filter->fo_fsd->lsd_feature_incompat |=
1262                         cpu_to_le32(OBD_INCOMPAT_GROUPS);
1263                 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1264                                                filter->fo_fsd, 1);
1265                 GOTO(cleanup_O0, rc);
1266
1267         cleanup_O0:
1268                 f_dput(O0_dentry);
1269         cleanup_R:
1270                 f_dput(dentry);
1271                 if (rc)
1272                         GOTO(cleanup, rc);
1273         } else {
1274                 f_dput(dentry);
1275         }
1276
1277         cleanup_phase = 2; /* groups */
1278
1279         /* we have to initialize all groups before first connections from
1280          * clients because they may send create/destroy for any group -bzzz */
1281         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1282         if (IS_ERR(filp)) {
1283                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1284                 GOTO(cleanup, rc = PTR_ERR(filp));
1285         }
1286         cleanup_phase = 3; /* filp */
1287
1288         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1289         if (rc) {
1290                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1291                 GOTO(cleanup, rc);
1292         }
1293         if (off == 0) {
1294                 last_group = FILTER_MIN_GROUPS;
1295         } else {
1296                 LASSERT_MDS_GROUP(last_group);
1297         }
1298
1299         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1300               FILTER_MIN_GROUPS, last_group);
1301         filter->fo_committed_group = last_group;
1302         rc = filter_read_groups(obd, last_group, 1);
1303         if (rc)
1304                 GOTO(cleanup, rc);
1305
1306         filp_close(filp, 0);
1307         RETURN(0);
1308
1309  cleanup:
1310         switch (cleanup_phase) {
1311         case 3:
1312                 filp_close(filp, 0);
1313         case 2:
1314                 filter_cleanup_groups(obd);
1315         case 1:
1316                 f_dput(filter->fo_dentry_O);
1317                 filter->fo_dentry_O = NULL;
1318         default:
1319                 break;
1320         }
1321         return rc;
1322
1323 }
1324
1325 /* setup the object store with correct subdirectories */
1326 static int filter_prep(struct obd_device *obd)
1327 {
1328         struct lvfs_run_ctxt saved;
1329         struct filter_obd *filter = &obd->u.filter;
1330         struct file *file;
1331         struct inode *inode;
1332         int rc = 0;
1333         ENTRY;
1334
1335         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1336         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1337         if (!file || IS_ERR(file)) {
1338                 rc = PTR_ERR(file);
1339                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1340                        LAST_RCVD, rc);
1341                 GOTO(out, rc);
1342         }
1343         filter->fo_rcvd_filp = file;
1344         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1345                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1346                        file->f_dentry->d_inode->i_mode);
1347                 GOTO(err_filp, rc = -ENOENT);
1348         }
1349
1350         inode = file->f_dentry->d_parent->d_inode;
1351         /* We use i_op->unlink directly in filter_vfs_unlink() */
1352         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1353                 CERROR("%s: filesystem does not support create/unlink ops\n",
1354                        obd->obd_name);
1355                 GOTO(err_filp, rc = -EOPNOTSUPP);
1356         }
1357
1358         /** lu_target has very limited use in filter now */
1359         lut_init(NULL, &filter->fo_lut, obd, NULL);
1360
1361         rc = filter_init_server_data(obd, file);
1362         if (rc) {
1363                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1364                 GOTO(err_filp, rc);
1365         }
1366         /* open and create health check io file*/
1367         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1368         if (IS_ERR(file)) {
1369                 rc = PTR_ERR(file);
1370                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1371                        HEALTH_CHECK, rc);
1372                 GOTO(err_server_data, rc);
1373         }
1374         filter->fo_health_check_filp = file;
1375         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1376                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1377                        file->f_dentry->d_inode->i_mode);
1378                 GOTO(err_health_check, rc = -ENOENT);
1379         }
1380         rc = lvfs_check_io_health(obd, file);
1381         if (rc)
1382                 GOTO(err_health_check, rc);
1383
1384         rc = filter_prep_groups(obd);
1385         if (rc)
1386                 GOTO(err_health_check, rc);
1387 out:
1388         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1389
1390         return(rc);
1391
1392 err_health_check:
1393         if (filp_close(filter->fo_health_check_filp, 0))
1394                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1395         filter->fo_health_check_filp = NULL;
1396 err_server_data:
1397         target_recovery_fini(obd);
1398         filter_free_server_data(filter);
1399 err_filp:
1400         if (filp_close(filter->fo_rcvd_filp, 0))
1401                 CERROR("can't close %s after error\n", LAST_RCVD);
1402         filter->fo_rcvd_filp = NULL;
1403         goto out;
1404 }
1405
1406 /* cleanup the filter: write last used object id to status file */
1407 static void filter_post(struct obd_device *obd)
1408 {
1409         struct lvfs_run_ctxt saved;
1410         struct filter_obd *filter = &obd->u.filter;
1411         int rc, i;
1412
1413         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1414          * best to start a transaction with h_sync, because we removed this
1415          * from lastobjid */
1416
1417         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1418         rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1419                                        filter->fo_fsd, 0);
1420         if (rc)
1421                 CERROR("error writing server data: rc = %d\n", rc);
1422
1423         for (i = 0; i < filter->fo_group_count; i++) {
1424                 rc = filter_update_last_objid(obd, i,
1425                                 (i == filter->fo_group_count - 1));
1426                 if (rc)
1427                         CERROR("error writing group %d lastobjid: rc = %d\n",
1428                                i, rc);
1429         }
1430
1431         rc = filp_close(filter->fo_rcvd_filp, 0);
1432         filter->fo_rcvd_filp = NULL;
1433         if (rc)
1434                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1435
1436         rc = filp_close(filter->fo_health_check_filp, 0);
1437         filter->fo_health_check_filp = NULL;
1438         if (rc)
1439                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1440
1441         filter_cleanup_groups(obd);
1442         filter_free_server_data(filter);
1443         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1444
1445         filter_free_capa_keys(filter);
1446         cleanup_capa_hash(filter->fo_capa_hash);
1447 }
1448
1449 static void filter_set_last_id(struct filter_obd *filter,
1450                                obd_id id, obd_gr group)
1451 {
1452         LASSERT(filter->fo_fsd != NULL);
1453         LASSERT(group <= filter->fo_group_count);
1454
1455         spin_lock(&filter->fo_objidlock);
1456         filter->fo_last_objids[group] = id;
1457         spin_unlock(&filter->fo_objidlock);
1458 }
1459
1460 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1461 {
1462         obd_id id;
1463         LASSERT(filter->fo_fsd != NULL);
1464         LASSERT(group <= filter->fo_group_count);
1465
1466         /* FIXME: object groups */
1467         spin_lock(&filter->fo_objidlock);
1468         id = filter->fo_last_objids[group];
1469         spin_unlock(&filter->fo_objidlock);
1470         return id;
1471 }
1472
1473 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1474 {
1475         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1476         return 0;
1477 }
1478
1479 /* We never dget the object parent, so DON'T dput it either */
1480 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1481 {
1482         struct filter_obd *filter = &obd->u.filter;
1483         struct filter_subdirs *subdirs;
1484         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1485
1486         if ((group > FILTER_GROUP_MDS0 && group < FILTER_GROUP_MDS1_N_BASE) ||
1487              filter->fo_subdir_count == 0)
1488                 return filter->fo_dentry_O_groups[group];
1489
1490         subdirs = &filter->fo_dentry_O_sub[group];
1491         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1492 }
1493
1494 /* We never dget the object parent, so DON'T dput it either */
1495 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1496                                   obd_id objid)
1497 {
1498         unsigned long now = jiffies;
1499         struct dentry *dparent = filter_parent(obd, group, objid);
1500         int rc;
1501
1502         if (IS_ERR(dparent))
1503                 return dparent;
1504         if (dparent == NULL)
1505                 return ERR_PTR(-ENOENT);
1506
1507         rc = filter_lock_dentry(obd, dparent);
1508         fsfilt_check_slow(obd, now, "parent lock");
1509         return rc ? ERR_PTR(rc) : dparent;
1510 }
1511
1512 /* We never dget the object parent, so DON'T dput it either */
1513 static void filter_parent_unlock(struct dentry *dparent)
1514 {
1515         UNLOCK_INODE_MUTEX(dparent->d_inode);
1516 }
1517
1518 /* How to get files, dentries, inodes from object id's.
1519  *
1520  * If dir_dentry is passed, the caller has already locked the parent
1521  * appropriately for this operation (normally a write lock).  If
1522  * dir_dentry is NULL, we do a read lock while we do the lookup to
1523  * avoid races with create/destroy and such changing the directory
1524  * internal to the filesystem code. */
1525 struct dentry *filter_fid2dentry(struct obd_device *obd,
1526                                  struct dentry *dir_dentry,
1527                                  obd_gr group, obd_id id)
1528 {
1529         struct dentry *dparent = dir_dentry;
1530         struct dentry *dchild;
1531         char name[32];
1532         int len;
1533         ENTRY;
1534
1535         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1536             obd->u.filter.fo_destroys_in_progress == 0) {
1537                 /* don't fail lookups for orphan recovery, it causes
1538                  * later LBUGs when objects still exist during precreate */
1539                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1540                 RETURN(ERR_PTR(-ENOENT));
1541         }
1542         if (id == 0) {
1543                 CERROR("fatal: invalid object id 0\n");
1544                 RETURN(ERR_PTR(-ESTALE));
1545         }
1546
1547         len = sprintf(name, LPU64, id);
1548         if (dir_dentry == NULL) {
1549                 dparent = filter_parent_lock(obd, group, id);
1550                 if (IS_ERR(dparent)) {
1551                         CERROR("%s: error getting object "LPU64":"LPU64
1552                                " parent: rc %ld\n", obd->obd_name,
1553                                id, group, PTR_ERR(dparent));
1554                         RETURN(dparent);
1555                 }
1556         }
1557         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1558                dparent->d_name.len, dparent->d_name.name, name);
1559         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1560         if (dir_dentry == NULL)
1561                 filter_parent_unlock(dparent);
1562         if (IS_ERR(dchild)) {
1563                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1564                        PTR_ERR(dchild));
1565                 RETURN(dchild);
1566         }
1567
1568         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1569                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1570                        obd->obd_name, id, dchild->d_inode->i_ino);
1571                 f_dput(dchild);
1572                 RETURN(ERR_PTR(-ENOENT));
1573         }
1574
1575         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1576                name, dchild, atomic_read(&dchild->d_count));
1577
1578         LASSERT(atomic_read(&dchild->d_count) > 0);
1579
1580         RETURN(dchild);
1581 }
1582
1583 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1584                                   obd_id group, struct lustre_handle *lockh)
1585 {
1586         int flags = LDLM_AST_DISCARD_DATA, rc;
1587         struct ldlm_res_id res_id;
1588         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1589         ENTRY;
1590
1591         osc_build_res_name(objid, group, &res_id);
1592         /* Tell the clients that the object is gone now and that they should
1593          * throw away any cached pages. */
1594         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1595                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1596                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1597                                     NULL, lockh);
1598         if (rc != ELDLM_OK)
1599                 lockh->cookie = 0;
1600         RETURN(rc);
1601 }
1602
1603 static void filter_fini_destroy(struct obd_device *obd,
1604                                 struct lustre_handle *lockh)
1605 {
1606         if (lockh->cookie)
1607                 ldlm_lock_decref(lockh, LCK_PW);
1608 }
1609
1610 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1611  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1612  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1613  * reverse.  Caller must take i_sem before starting the transaction and we
1614  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1615 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1616                       struct vfsmount *mnt)
1617 {
1618         int rc;
1619         ENTRY;
1620
1621         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1622          * itself we already hold dir->i_mutex for child create/unlink ops */
1623         LASSERT(dentry->d_inode != NULL);
1624         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1625         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1626
1627
1628         /* may_delete() */
1629         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1630                 GOTO(out, rc = -ENOENT);
1631
1632         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1633         if (rc)
1634                 GOTO(out, rc);
1635
1636         if (IS_APPEND(dir))
1637                 GOTO(out, rc = -EPERM);
1638
1639         /* check_sticky() */
1640         if ((dentry->d_inode->i_uid != current->fsuid &&
1641              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1642             IS_IMMUTABLE(dentry->d_inode))
1643                 GOTO(out, rc = -EPERM);
1644
1645         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1646          *       that was done because of journal_start (which is already done
1647          *       here) or some other ordering issue. */
1648         DQUOT_INIT(dir);
1649
1650         rc = ll_security_inode_unlink(dir, dentry, mnt);
1651         if (rc)
1652                 GOTO(out, rc);
1653
1654         rc = dir->i_op->unlink(dir, dentry);
1655 out:
1656         /* need to drop i_mutex before we lose inode reference */
1657         UNLOCK_INODE_MUTEX(dentry->d_inode);
1658         if (rc == 0)
1659                 d_delete(dentry);
1660
1661         RETURN(rc);
1662 }
1663
1664 /* Caller must hold LCK_PW on parent and push us into kernel context.
1665  * Caller must hold child i_mutex, we drop it always.
1666  * Caller is also required to ensure that dchild->d_inode exists. */
1667 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1668                                    obd_gr group, struct dentry *dparent,
1669                                    struct dentry *dchild)
1670 {
1671         struct inode *inode = dchild->d_inode;
1672         struct filter_obd *filter = &obd->u.filter;
1673         int rc;
1674
1675         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1676                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1677                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1678                        (unsigned long)inode->i_nlink,
1679                        atomic_read(&inode->i_count));
1680         }
1681
1682         rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1683         if (rc)
1684                 CERROR("error unlinking objid %.*s: rc %d\n",
1685                        dchild->d_name.len, dchild->d_name.name, rc);
1686         return(rc);
1687 }
1688
1689 struct filter_intent_args {
1690         struct ldlm_lock **victim;
1691         __u64 size;
1692         int *liblustre;
1693 };
1694
1695 static enum interval_iter filter_intent_cb(struct interval_node *n,
1696                                            void *args)
1697 {
1698         struct ldlm_interval *node = (struct ldlm_interval *)n;
1699         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1700         __u64 size = arg->size;
1701         struct ldlm_lock **v = arg->victim;
1702         struct ldlm_lock *lck;
1703
1704         /* If the interval is lower than the current file size,
1705          * just break. */
1706         if (interval_high(n) <= size)
1707                 return INTERVAL_ITER_STOP;
1708
1709         list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1710                 /* Don't send glimpse ASTs to liblustre clients.
1711                  * They aren't listening for them, and they do
1712                  * entirely synchronous I/O anyways. */
1713                 if (lck->l_export == NULL ||
1714                     lck->l_export->exp_libclient == 1)
1715                         continue;
1716
1717                 if (*arg->liblustre)
1718                         *arg->liblustre = 0;
1719
1720                 if (*v == NULL) {
1721                         *v = LDLM_LOCK_GET(lck);
1722                 } else if ((*v)->l_policy_data.l_extent.start <
1723                            lck->l_policy_data.l_extent.start) {
1724                         LDLM_LOCK_RELEASE(*v);
1725                         *v = LDLM_LOCK_GET(lck);
1726                 }
1727
1728                 /* the same policy group - every lock has the
1729                  * same extent, so needn't do it any more */
1730                 break;
1731         }
1732
1733         return INTERVAL_ITER_CONT;
1734 }
1735
1736 static int filter_intent_policy(struct ldlm_namespace *ns,
1737                                 struct ldlm_lock **lockp, void *req_cookie,
1738                                 ldlm_mode_t mode, int flags, void *data)
1739 {
1740         CFS_LIST_HEAD(rpc_list);
1741         struct ptlrpc_request *req = req_cookie;
1742         struct ldlm_lock *lock = *lockp, *l = NULL;
1743         struct ldlm_resource *res = lock->l_resource;
1744         ldlm_processing_policy policy;
1745         struct ost_lvb *res_lvb, *reply_lvb;
1746         struct ldlm_reply *rep;
1747         ldlm_error_t err;
1748         int idx, rc, tmpflags = 0, only_liblustre = 1;
1749         struct ldlm_interval_tree *tree;
1750         struct filter_intent_args arg;
1751         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1752                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1753                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1754         ENTRY;
1755
1756         policy = ldlm_get_processing_policy(res);
1757         LASSERT(policy != NULL);
1758         LASSERT(req != NULL);
1759
1760         rc = lustre_pack_reply(req, 3, repsize, NULL);
1761         if (rc)
1762                 RETURN(req->rq_status = rc);
1763
1764         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1765         LASSERT(rep != NULL);
1766
1767         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1768                                    sizeof(*reply_lvb));
1769         LASSERT(reply_lvb != NULL);
1770
1771         //fixup_handle_for_resent_req(req, lock, &lockh);
1772
1773         /* Call the extent policy function to see if our request can be
1774          * granted, or is blocked.
1775          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1776          * lock, and should not be granted if the lock will be blocked.
1777          */
1778
1779         LASSERT(ns == res->lr_namespace);
1780         lock_res(res);
1781         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1782         check_res_locked(res);
1783
1784         /* FIXME: we should change the policy function slightly, to not make
1785          * this list at all, since we just turn around and free it */
1786         while (!list_empty(&rpc_list)) {
1787                 struct ldlm_lock *wlock =
1788                         list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1789                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1790                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1791                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1792                 list_del_init(&wlock->l_cp_ast);
1793                 LDLM_LOCK_RELEASE(wlock);
1794         }
1795
1796         /* The lock met with no resistance; we're finished. */
1797         if (rc == LDLM_ITER_CONTINUE) {
1798                 /* do not grant locks to the liblustre clients: they cannot
1799                  * handle ASTs robustly.  We need to do this while still
1800                  * holding ns_lock to avoid the lock remaining on the res_link
1801                  * list (and potentially being added to l_pending_list by an
1802                  * AST) when we are going to drop this lock ASAP. */
1803                 if (lock->l_export->exp_libclient ||
1804                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1805                         ldlm_resource_unlink_lock(lock);
1806                         err = ELDLM_LOCK_ABORTED;
1807                 } else {
1808                         err = ELDLM_LOCK_REPLACED;
1809                 }
1810                 unlock_res(res);
1811                 RETURN(err);
1812         }
1813
1814         /* Do not grant any lock, but instead send GL callbacks.  The extent
1815          * policy nicely created a list of all PW locks for us.  We will choose
1816          * the highest of those which are larger than the size in the LVB, if
1817          * any, and perform a glimpse callback. */
1818         res_lvb = res->lr_lvb_data;
1819         LASSERT(res_lvb != NULL);
1820         *reply_lvb = *res_lvb;
1821
1822         /*
1823          * ->ns_lock guarantees that no new locks are granted, and,
1824          * therefore, that res->lr_lvb_data cannot increase beyond the
1825          * end of already granted lock. As a result, it is safe to
1826          * check against "stale" reply_lvb->lvb_size value without
1827          * res->lr_lvb_sem.
1828          */
1829         arg.size = reply_lvb->lvb_size;
1830         arg.victim = &l;
1831         arg.liblustre = &only_liblustre;
1832         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1833                 tree = &res->lr_itree[idx];
1834                 if (tree->lit_mode == LCK_PR)
1835                         continue;
1836
1837                 interval_iterate_reverse(tree->lit_root,
1838                                          filter_intent_cb, &arg);
1839         }
1840         unlock_res(res);
1841
1842         /* There were no PW locks beyond the size in the LVB; finished. */
1843         if (l == NULL) {
1844                 if (only_liblustre) {
1845                         /* If we discovered a liblustre client with a PW lock,
1846                          * however, the LVB may be out of date!  The LVB is
1847                          * updated only on glimpse (which we don't do for
1848                          * liblustre clients) and cancel (which the client
1849                          * obviously has not yet done).  So if it has written
1850                          * data but kept the lock, the LVB is stale and needs
1851                          * to be updated from disk.
1852                          *
1853                          * Of course, this will all disappear when we switch to
1854                          * taking liblustre locks on the OST. */
1855                         ldlm_res_lvbo_update(res, NULL, 0, 1);
1856                 }
1857                 RETURN(ELDLM_LOCK_ABORTED);
1858         }
1859
1860         /*
1861          * This check is for lock taken in filter_prepare_destroy() that does
1862          * not have l_glimpse_ast set. So the logic is: if there is a lock
1863          * with no l_glimpse_ast set, this object is being destroyed already.
1864          *
1865          * Hence, if you are grabbing DLM locks on the server, always set
1866          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1867          */
1868         if (l->l_glimpse_ast == NULL) {
1869                 /* We are racing with unlink(); just return -ENOENT */
1870                 rep->lock_policy_res1 = -ENOENT;
1871                 goto out;
1872         }
1873
1874         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1875         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1876         /* Update the LVB from disk if the AST failed (this is a legal race) */
1877         /*
1878          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1879          * sending ast is not handled. This can result in lost client writes.
1880          */
1881         if (rc != 0)
1882                 ldlm_res_lvbo_update(res, NULL, 0, 1);
1883
1884         lock_res(res);
1885         *reply_lvb = *res_lvb;
1886         unlock_res(res);
1887
1888  out:
1889         LDLM_LOCK_RELEASE(l);
1890
1891         RETURN(ELDLM_LOCK_ABORTED);
1892 }
1893
1894 /*
1895  * per-obd_device iobuf pool.
1896  *
1897  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1898  * allocations in write-path has to be minimized (see bug 5137).
1899  *
1900  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1901  * OST threads (see ost_thread_{init,done}()).
1902  *
1903  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1904  * at the OST layer there are only (potentially) multiple obd_device of type
1905  * unknown at the time of OST thread creation.
1906  *
1907  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1908  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1909  * it's very own iobuf.
1910  *
1911  * Functions below
1912  *
1913  *     filter_kiobuf_pool_init()
1914  *
1915  *     filter_kiobuf_pool_done()
1916  *
1917  *     filter_iobuf_get()
1918  *
1919  * operate on this array. They are "generic" in a sense that they don't depend
1920  * on actual type of iobuf's (the latter depending on Linux kernel version).
1921  */
1922
1923 /*
1924  * destroy pool created by filter_iobuf_pool_init
1925  */
1926 static void filter_iobuf_pool_done(struct filter_obd *filter)
1927 {
1928         struct filter_iobuf **pool;
1929         int i;
1930
1931         ENTRY;
1932
1933         pool = filter->fo_iobuf_pool;
1934         if (pool != NULL) {
1935                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1936                         if (pool[i] != NULL)
1937                                 filter_free_iobuf(pool[i]);
1938                 }
1939                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1940                 filter->fo_iobuf_pool = NULL;
1941         }
1942         EXIT;
1943 }
1944
1945 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1946 {
1947         struct filter_obd       *filter = &obd->u.filter;
1948         struct sptlrpc_rule_set  tmp_rset;
1949         int                      rc;
1950
1951         sptlrpc_rule_set_init(&tmp_rset);
1952         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1953         if (rc) {
1954                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1955                        obd->obd_name, rc);
1956                 return rc;
1957         }
1958
1959         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1960
1961         write_lock(&filter->fo_sptlrpc_lock);
1962         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1963         filter->fo_sptlrpc_rset = tmp_rset;
1964         write_unlock(&filter->fo_sptlrpc_lock);
1965
1966         return 0;
1967 }
1968
1969 /*
1970  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1971  */
1972 static int filter_iobuf_pool_init(struct filter_obd *filter)
1973 {
1974         void **pool;
1975
1976         ENTRY;
1977
1978
1979         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1980                       GFP_KERNEL);
1981         if (filter->fo_iobuf_pool == NULL)
1982                 RETURN(-ENOMEM);
1983
1984         filter->fo_iobuf_count = OSS_THREADS_MAX;
1985
1986         RETURN(0);
1987 }
1988
1989 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1990  * many threads there will be so we allocate a large empty array and only
1991  * fill in those slots that are actually in use.
1992  * If we haven't allocated a pool entry for this thread before, do so now. */
1993 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1994 {
1995         int thread_id                    = (oti && oti->oti_thread) ?
1996                                            oti->oti_thread->t_id : -1;
1997         struct filter_iobuf  *pool       = NULL;
1998         struct filter_iobuf **pool_place = NULL;
1999
2000         if (thread_id >= 0) {
2001                 LASSERT(thread_id < filter->fo_iobuf_count);
2002                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
2003         }
2004
2005         if (unlikely(pool == NULL)) {
2006                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
2007                                           PTLRPC_MAX_BRW_PAGES);
2008                 if (pool_place != NULL)
2009                         *pool_place = pool;
2010         }
2011
2012         return pool;
2013 }
2014
2015 /* mount the file system (secretly).  lustre_cfg parameters are:
2016  * 1 = device
2017  * 2 = fstype
2018  * 3 = flags: failover=f, failout=n
2019  * 4 = mount options
2020  */
2021 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
2022                         void *option)
2023 {
2024         struct filter_obd *filter = &obd->u.filter;
2025         struct vfsmount *mnt;
2026         struct lustre_mount_info *lmi;
2027         struct obd_uuid uuid;
2028         __u8 *uuid_ptr;
2029         char *str, *label;
2030         char ns_name[48];
2031         request_queue_t *q;
2032         int rc, i;
2033         ENTRY;
2034
2035         if (lcfg->lcfg_bufcount < 3 ||
2036             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
2037             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
2038                 RETURN(-EINVAL);
2039
2040         lmi = server_get_mount(obd->obd_name);
2041         if (lmi) {
2042                 /* We already mounted in lustre_fill_super.
2043                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2044                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2045                 mnt = lmi->lmi_mnt;
2046                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2047         } else {
2048                 /* old path - used by lctl */
2049                 CERROR("Using old MDS mount method\n");
2050                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2051                                     MS_NOATIME|MS_NODIRATIME,
2052                                     lustre_cfg_string(lcfg, 1), option);
2053                 if (IS_ERR(mnt)) {
2054                         rc = PTR_ERR(mnt);
2055                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2056                                            lustre_cfg_string(lcfg, 1), rc);
2057                         RETURN(rc);
2058                 }
2059
2060                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2061         }
2062         if (IS_ERR(obd->obd_fsops))
2063                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2064
2065         rc = filter_iobuf_pool_init(filter);
2066         if (rc != 0)
2067                 GOTO(err_ops, rc);
2068
2069         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2070                 CERROR("%s: Underlying device is marked as read-only. "
2071                        "Setup failed\n", obd->obd_name);
2072                 GOTO(err_ops, rc = -EROFS);
2073         }
2074
2075         /* failover is the default */
2076         obd->obd_replayable = 1;
2077
2078         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2079                 str = lustre_cfg_string(lcfg, 3);
2080                 if (strchr(str, 'n')) {
2081                         CWARN("%s: recovery disabled\n", obd->obd_name);
2082                         obd->obd_replayable = 0;
2083                 }
2084         }
2085
2086         filter->fo_vfsmnt = mnt;
2087         obd->u.obt.obt_sb = mnt->mnt_sb;
2088         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2089         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2090
2091         fsfilt_setup(obd, obd->u.obt.obt_sb);
2092
2093         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2094         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2095         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2096         obd->obd_lvfs_ctxt.fs = get_ds();
2097         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2098
2099         init_mutex(&filter->fo_init_lock);
2100         filter->fo_committed_group = 0;
2101         filter->fo_destroys_in_progress = 0;
2102         for (i = 0; i < 32; i++)
2103                 sema_init(&filter->fo_create_locks[i], 1);
2104
2105         spin_lock_init(&filter->fo_translock);
2106         spin_lock_init(&filter->fo_objidlock);
2107         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2108         sema_init(&filter->fo_alloc_lock, 1);
2109         init_brw_stats(&filter->fo_filter_stats);
2110         filter->fo_read_cache = 1; /* enable read-only cache by default */
2111         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2112         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2113         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2114         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2115
2116         rc = filter_prep(obd);
2117         if (rc)
2118                 GOTO(err_ops, rc);
2119
2120         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2121         spin_lock_init(&filter->fo_llog_list_lock);
2122
2123         filter->fo_fl_oss_capa = 1;
2124
2125         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2126         filter->fo_capa_hash = init_capa_hash();
2127         if (filter->fo_capa_hash == NULL)
2128                 GOTO(err_post, rc = -ENOMEM);
2129
2130         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2131         obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2132                                                 LDLM_NAMESPACE_GREEDY);
2133         if (obd->obd_namespace == NULL)
2134                 GOTO(err_post, rc = -ENOMEM);
2135         obd->obd_namespace->ns_lvbp = obd;
2136         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2137         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2138
2139         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2140                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2141
2142         rc = obd_llog_init(obd, &obd->obd_olg, obd, 1, NULL, NULL);
2143         if (rc) {
2144                 CERROR("failed to setup llogging subsystems\n");
2145                 GOTO(err_post, rc);
2146         }
2147
2148         rwlock_init(&filter->fo_sptlrpc_lock);
2149         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2150         /* do this after llog being initialized */
2151         filter_adapt_sptlrpc_conf(obd, 1);
2152
2153         rc = lquota_setup(filter_quota_interface_ref, obd);
2154         if (rc)
2155                 GOTO(err_post, rc);
2156
2157         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2158         if (q->max_sectors < q->max_hw_sectors &&
2159             q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2160                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2161                               "for larger I/O requests: max_sectors = %u "
2162                               "could be up to max_hw_sectors=%u\n",
2163                               obd->obd_name, mnt->mnt_sb->s_id,
2164                               q->max_sectors, q->max_hw_sectors);
2165
2166         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2167         if (uuid_ptr != NULL) {
2168                 class_uuid_unparse(uuid_ptr, &uuid);
2169                 str = uuid.uuid;
2170         } else {
2171                 str = "no UUID";
2172         }
2173
2174         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2175
2176         if (obd->obd_recovering) {
2177                 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2178                               "recovery for at least %d:%.02d, or until %d "
2179                               "client%s reconnect%s.\n",
2180                               obd->obd_name, lustre_cfg_string(lcfg, 1),
2181                               label ?: "", label ? "/" : "", str,
2182                               obd->obd_recovery_timeout / 60,
2183                               obd->obd_recovery_timeout % 60,
2184                               obd->obd_max_recoverable_clients,
2185                               (obd->obd_max_recoverable_clients == 1) ? "":"s",
2186                               (obd->obd_max_recoverable_clients == 1) ? "s":"");
2187         } else {
2188                 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2189                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2190                               label ?: "", label ? "/" : "", str,
2191                               obd->obd_replayable ? "enabled" : "disabled");
2192         }
2193
2194         RETURN(0);
2195
2196 err_post:
2197         filter_post(obd);
2198 err_ops:
2199         fsfilt_put_ops(obd->obd_fsops);
2200         filter_iobuf_pool_done(filter);
2201 err_mntput:
2202         server_put_mount(obd->obd_name, mnt);
2203         obd->u.obt.obt_sb = 0;
2204         return rc;
2205 }
2206
2207 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2208 {
2209         struct lprocfs_static_vars lvars;
2210         unsigned long addr;
2211         struct page *page;
2212         int rc;
2213
2214         CLASSERT(offsetof(struct obd_device, u.obt) ==
2215                  offsetof(struct obd_device, u.filter.fo_obt));
2216
2217         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2218                 RETURN(-EINVAL);
2219
2220         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2221         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2222         if (!page)
2223                 RETURN(-ENOMEM);
2224         addr = (unsigned long)cfs_page_address(page);
2225         clear_page((void *)addr);
2226
2227         /* lprocfs must be setup before the filter so state can be safely added
2228          * to /proc incrementally as the filter is setup */
2229         lprocfs_filter_init_vars(&lvars);
2230         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2231             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2232                 /* Init obdfilter private stats here */
2233                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2234                                      LPROCFS_CNTR_AVGMINMAX,
2235                                      "read_bytes", "bytes");
2236                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2237                                      LPROCFS_CNTR_AVGMINMAX,
2238                                      "write_bytes", "bytes");
2239                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2240                                      LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2241                                      "get_page", "usec");
2242                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2243                                      LPROCFS_CNTR_AVGMINMAX,
2244                                      "get_page_failures", "num");
2245                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2246                                      LPROCFS_CNTR_AVGMINMAX,
2247                                      "cache_access", "pages");
2248                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2249                                      LPROCFS_CNTR_AVGMINMAX,
2250                                      "cache_hit", "pages");
2251                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2252                                      LPROCFS_CNTR_AVGMINMAX,
2253                                      "cache_miss", "pages");
2254
2255                 lproc_filter_attach_seqstat(obd);
2256                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2257                                                         obd->obd_proc_entry,
2258                                                         NULL, NULL);
2259                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2260                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2261                         CERROR("error %d setting up lprocfs for %s\n",
2262                                rc, "exports");
2263                         obd->obd_proc_exports_entry = NULL;
2264                 }
2265         }
2266         if (obd->obd_proc_exports_entry)
2267                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2268                                    lprocfs_nid_stats_clear_read,
2269                                    lprocfs_nid_stats_clear_write, obd, NULL);
2270
2271         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2272                LUSTRE_CFG_BUFLEN(lcfg, 4));
2273         rc = filter_common_setup(obd, lcfg, (void *)addr);
2274         OBD_PAGE_FREE(page);
2275
2276         if (rc) {
2277                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2278                 lprocfs_free_per_client_stats(obd);
2279                 lprocfs_free_obd_stats(obd);
2280                 lprocfs_obd_cleanup(obd);
2281         }
2282
2283         return rc;
2284 }
2285
2286 static struct llog_operations filter_mds_ost_repl_logops;
2287
2288 static struct llog_operations filter_size_orig_logops = {
2289         .lop_setup   = llog_obd_origin_setup,
2290         .lop_cleanup = llog_obd_origin_cleanup,
2291         .lop_add     = llog_obd_origin_add
2292 };
2293
2294 static int filter_olg_fini(struct obd_llog_group *olg)
2295 {
2296         struct llog_ctxt *ctxt;
2297         int rc = 0, rc2 = 0;
2298         ENTRY;
2299
2300         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2301         if (ctxt)
2302                 rc = llog_cleanup(ctxt);
2303
2304         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2305         if (ctxt) {
2306                 rc2 = llog_cleanup(ctxt);
2307                 if (!rc)
2308                         rc = rc2;
2309         }
2310
2311         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2312         if (ctxt) {
2313                 rc2 = llog_cleanup(ctxt);
2314                 if (!rc)
2315                         rc = rc2;
2316         }
2317
2318         RETURN(rc);
2319 }
2320
2321 static int
2322 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2323                 struct obd_device *tgt)
2324 {
2325         int rc;
2326         ENTRY;
2327
2328         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2329                         &filter_mds_ost_repl_logops);
2330         if (rc)
2331                 GOTO(cleanup, rc);
2332
2333         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2334                         &filter_size_orig_logops);
2335         if (rc)
2336                 GOTO(cleanup, rc);
2337         EXIT;
2338 cleanup:
2339         if (rc)
2340                 filter_olg_fini(olg);
2341         return rc;
2342 }
2343
2344 /**
2345  * Init the default olg, which is embeded in the obd_device, for filter.
2346  */
2347 static int
2348 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2349                         struct obd_device *tgt)
2350 {
2351         struct filter_obd *filter = &obd->u.filter;
2352         struct llog_ctxt *ctxt;
2353         int rc;
2354         ENTRY;
2355
2356         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2357         if (!filter->fo_lcm)
2358                 RETURN(-ENOMEM);
2359
2360         filter_mds_ost_repl_logops = llog_client_ops;
2361         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2362         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2363         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2364
2365         rc = filter_olg_init(obd, olg, tgt);
2366         if (rc)
2367                 GOTO(cleanup_lcm, rc);
2368
2369         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2370                         &llog_lvfs_ops);
2371         if (rc)
2372                 GOTO(cleanup_olg, rc);
2373
2374         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2375         if (!ctxt) {
2376                 CERROR("Can't get ctxt for %p:%x\n", olg,
2377                        LLOG_MDS_OST_REPL_CTXT);
2378                 GOTO(cleanup_olg, rc = -ENODEV);
2379         }
2380         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2381         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2382         llog_ctxt_put(ctxt);
2383
2384         RETURN(0);
2385 cleanup_olg:
2386         filter_olg_fini(olg);
2387 cleanup_lcm:
2388         llog_recov_thread_fini(filter->fo_lcm, 1);
2389         filter->fo_lcm = NULL;
2390         return rc;
2391 }
2392
2393 static int
2394 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2395                  struct obd_device *tgt, int count, struct llog_catid *catid,
2396                  struct obd_uuid *uuid)
2397 {
2398         struct filter_obd *filter = &obd->u.filter;
2399         struct llog_ctxt *ctxt;
2400         int rc;
2401         ENTRY;
2402
2403         LASSERT(olg != NULL);
2404         if (olg == &obd->obd_olg)
2405                 return filter_default_olg_init(obd, olg, tgt);
2406
2407         LASSERT(filter->fo_lcm != NULL);
2408         rc = filter_olg_init(obd, olg, tgt);
2409         if (rc)
2410                 RETURN(rc);
2411         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2412         if (!ctxt) {
2413                 CERROR("Can't get ctxt for %p:%x\n", olg,
2414                        LLOG_MDS_OST_REPL_CTXT);
2415                 filter_olg_fini(olg);
2416                 RETURN(-ENODEV);
2417         }
2418         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2419         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2420         llog_ctxt_put(ctxt);
2421         RETURN(rc);
2422 }
2423
2424 static int filter_llog_finish(struct obd_device *obd, int count)
2425 {
2426         struct filter_obd *filter = &obd->u.filter;
2427         struct llog_ctxt *ctxt;
2428         ENTRY;
2429
2430         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2431         if (ctxt) {
2432                 /*
2433                  * Make sure that no cached llcds left in recov_thread.
2434                  * We actually do sync in disconnect time, but disconnect
2435                  * may not come being marked rq_no_resend = 1.
2436                  */
2437                 llog_sync(ctxt, NULL);
2438
2439                 /*
2440                  * Balance class_import_get() in llog_receptor_accept().
2441                  * This is safe to do, as llog is already synchronized
2442                  * and its import may go.
2443                  */
2444                 mutex_down(&ctxt->loc_sem);
2445                 if (ctxt->loc_imp) {
2446                         class_import_put(ctxt->loc_imp);
2447                         ctxt->loc_imp = NULL;
2448                 }
2449                 mutex_up(&ctxt->loc_sem);
2450                 llog_ctxt_put(ctxt);
2451         }
2452
2453         if (filter->fo_lcm) {
2454                 mutex_down(&ctxt->loc_sem);
2455                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2456                 filter->fo_lcm = NULL;
2457                 mutex_up(&ctxt->loc_sem);
2458         }
2459         RETURN(filter_olg_fini(&obd->obd_olg));
2460 }
2461
2462 /**
2463  * Find the group llog according to group index in the llog group list.
2464  */
2465 static struct obd_llog_group *
2466 filter_find_olg_internal(struct filter_obd *filter, int group)
2467 {
2468         struct obd_llog_group *olg;
2469
2470         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2471         list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2472                 if (olg->olg_group == group)
2473                         RETURN(olg);
2474         }
2475         RETURN(NULL);
2476 }
2477
2478 /**
2479  * Find the group llog according to group index on the filter
2480  */
2481 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2482 {
2483         struct obd_llog_group *olg = NULL;
2484         struct filter_obd *filter;
2485
2486         filter = &obd->u.filter;
2487
2488         if (group == FILTER_GROUP_LLOG)
2489                 RETURN(&obd->obd_olg);
2490
2491         spin_lock(&filter->fo_llog_list_lock);
2492         olg = filter_find_olg_internal(filter, group);
2493         spin_unlock(&filter->fo_llog_list_lock);
2494
2495         RETURN(olg);
2496 }
2497 /**
2498  * Find the llog_group of the filter according to the group. If it can not
2499  * find, create the llog_group, which only happens when mds is being synced
2500  * with OST.
2501  */
2502 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2503 {
2504         struct obd_llog_group *olg = NULL;
2505         struct filter_obd *filter;
2506         int rc;
2507
2508         filter = &obd->u.filter;
2509
2510         if (group == FILTER_GROUP_LLOG)
2511                 RETURN(&obd->obd_olg);
2512
2513         spin_lock(&filter->fo_llog_list_lock);
2514         olg = filter_find_olg_internal(filter, group);
2515         if (olg) {
2516                 if (olg->olg_initializing) {
2517                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2518                 } else {
2519                         GOTO(out_unlock, olg);
2520                 }
2521         }
2522         OBD_ALLOC_PTR(olg);
2523         if (olg == NULL)
2524                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2525
2526         llog_group_init(olg, group);
2527         list_add(&olg->olg_list, &filter->fo_llog_list);
2528         olg->olg_initializing = 1;
2529         spin_unlock(&filter->fo_llog_list_lock);
2530
2531         rc = llog_cat_initialize(obd, olg, 1, NULL);
2532         if (rc) {
2533                spin_lock(&filter->fo_llog_list_lock);
2534                list_del(&olg->olg_list);
2535                spin_unlock(&filter->fo_llog_list_lock);
2536                OBD_FREE_PTR(olg);
2537                GOTO(out, olg = ERR_PTR(-ENOMEM));
2538         }
2539         spin_lock(&filter->fo_llog_list_lock);
2540         olg->olg_initializing = 0;
2541         spin_unlock(&filter->fo_llog_list_lock);
2542         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2543               obd->obd_name, group, olg);
2544 out:
2545         RETURN(olg);
2546
2547 out_unlock:
2548         spin_unlock(&filter->fo_llog_list_lock);
2549         GOTO(out, olg);
2550 }
2551
2552 static int filter_llog_connect(struct obd_export *exp,
2553                                struct llogd_conn_body *body)
2554 {
2555         struct obd_device *obd = exp->exp_obd;
2556         struct llog_ctxt *ctxt;
2557         struct obd_llog_group *olg;
2558         int rc;
2559         ENTRY;
2560
2561         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2562                obd->obd_name, body->lgdc_logid.lgl_oid,
2563                body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2564
2565         olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2566         if (!olg) {
2567                 CERROR(" %s: can not find olg of group %d\n",
2568                        obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2569                 RETURN(-ENOENT);
2570         }
2571         llog_group_set_export(olg, exp);
2572
2573         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2574         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2575                  body->lgdc_ctxt_idx);
2576
2577         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2578               obd->obd_name, body->lgdc_logid.lgl_oid,
2579               body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2580
2581         spin_lock_bh(&obd->obd_processing_task_lock);
2582         obd->u.filter.fo_mds_ost_sync = 1;
2583         spin_unlock_bh(&obd->obd_processing_task_lock);
2584         rc = llog_connect(ctxt, &body->lgdc_logid,
2585                           &body->lgdc_gen, NULL);
2586         llog_ctxt_put(ctxt);
2587         if (rc != 0)
2588                 CERROR("failed to connect rc %d idx %d\n", rc,
2589                                 body->lgdc_ctxt_idx);
2590
2591         RETURN(rc);
2592 }
2593
2594 static int filter_llog_preclean(struct obd_device *obd)
2595 {
2596         struct obd_llog_group *olg, *tmp;
2597         struct filter_obd *filter;
2598         struct list_head  remove_list;
2599         int rc = 0;
2600         ENTRY;
2601
2602         rc = obd_llog_finish(obd, 0);
2603         if (rc)
2604                 CERROR("failed to cleanup llogging subsystem\n");
2605
2606         filter = &obd->u.filter;
2607         CFS_INIT_LIST_HEAD(&remove_list);
2608
2609         spin_lock(&filter->fo_llog_list_lock);
2610         while (!list_empty(&filter->fo_llog_list)) {
2611                 olg = list_entry(filter->fo_llog_list.next,
2612                                  struct obd_llog_group, olg_list);
2613                 list_del(&olg->olg_list);
2614                 list_add(&olg->olg_list, &remove_list);
2615         }
2616         spin_unlock(&filter->fo_llog_list_lock);
2617
2618         list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2619                 list_del_init(&olg->olg_list);
2620                 rc = filter_olg_fini(olg);
2621                 if (rc)
2622                         CERROR("failed to cleanup llogging subsystem for %u\n",
2623                                olg->olg_group);
2624                 OBD_FREE_PTR(olg);
2625         }
2626
2627         RETURN(rc);
2628 }
2629
2630 static int filter_precleanup(struct obd_device *obd,
2631                              enum obd_cleanup_stage stage)
2632 {
2633         int rc = 0;
2634         ENTRY;
2635
2636         switch(stage) {
2637         case OBD_CLEANUP_EARLY:
2638                 break;
2639         case OBD_CLEANUP_EXPORTS:
2640                 /* Stop recovery before namespace cleanup. */
2641                 target_recovery_fini(obd);
2642                 rc = filter_llog_preclean(obd);
2643                 break;
2644         }
2645         RETURN(rc);
2646 }
2647
2648 static int filter_cleanup(struct obd_device *obd)
2649 {
2650         struct filter_obd *filter = &obd->u.filter;
2651         ENTRY;
2652
2653         if (obd->obd_fail)
2654                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2655                               "will be preserved.\n", obd->obd_name);
2656
2657         obd_exports_barrier(obd);
2658         obd_zombie_barrier();
2659
2660         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2661         lprocfs_free_per_client_stats(obd);
2662         lprocfs_free_obd_stats(obd);
2663         lprocfs_obd_cleanup(obd);
2664         lquota_cleanup(filter_quota_interface_ref, obd);
2665
2666         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2667         obd->obd_namespace = NULL;
2668
2669         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2670
2671         if (obd->u.obt.obt_sb == NULL)
2672                 RETURN(0);
2673
2674         filter_post(obd);
2675
2676         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2677         shrink_dcache_sb(obd->u.obt.obt_sb);
2678
2679         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2680         obd->u.obt.obt_sb = NULL;
2681
2682         fsfilt_put_ops(obd->obd_fsops);
2683
2684         filter_iobuf_pool_done(filter);
2685
2686         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2687
2688         RETURN(0);
2689 }
2690
2691 static int filter_connect_internal(struct obd_export *exp,
2692                                    struct obd_connect_data *data)
2693 {
2694         struct filter_export_data *fed = &exp->exp_filter_data;
2695
2696         if (!data)
2697                 RETURN(0);
2698
2699         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2700                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2701                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2702                data->ocd_connect_flags, data->ocd_version,
2703                data->ocd_grant, data->ocd_index);
2704
2705         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2706                 CWARN("!!! This export (nid %s) used object group %d "
2707                        "earlier; now it's trying to use group %d!  This could "
2708                        "be a bug in the MDS. Please report to "
2709                        "http://bugzilla.lustre.org/\n",
2710                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2711                 RETURN(-EPROTO);
2712         }
2713         fed->fed_group = data->ocd_group;
2714
2715         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2716         exp->exp_connect_flags = data->ocd_connect_flags;
2717         data->ocd_version = LUSTRE_VERSION_CODE;
2718
2719         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2720         if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2721                   data->ocd_connect_flags & OBD_CONNECT_MDS))
2722                 RETURN(-EPROTO);
2723
2724         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2725                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2726                 obd_size left, want;
2727
2728                 spin_lock(&exp->exp_obd->obd_osfs_lock);
2729                 left = filter_grant_space_left(exp);
2730                 want = data->ocd_grant;
2731                 filter_grant(exp, fed->fed_grant, want, left);
2732                 data->ocd_grant = fed->fed_grant;
2733                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2734
2735                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2736                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2737                        exp->exp_client_uuid.uuid, exp,
2738                        data->ocd_grant, want, left);
2739                 
2740                 filter->fo_tot_granted_clients ++;
2741         }
2742
2743         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2744                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2745                 struct lr_server_data *lsd = filter->fo_fsd;
2746                 int index = le32_to_cpu(lsd->lsd_ost_index);
2747
2748                 if (!(lsd->lsd_feature_compat &
2749                       cpu_to_le32(OBD_COMPAT_OST))) {
2750                         /* this will only happen on the first connect */
2751                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2752                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2753                         filter_update_server_data(exp->exp_obd,
2754                                                   filter->fo_rcvd_filp, lsd, 1);
2755                 } else if (index != data->ocd_index) {
2756                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2757                                            " %u doesn't match actual OST index"
2758                                            " %u in last_rcvd file, bad "
2759                                            "configuration?\n",
2760                                            obd_export_nid2str(exp), index,
2761                                            data->ocd_index);
2762                         RETURN(-EBADF);
2763                 }
2764                 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2765                  * FIXME: We don't strictly need the COMPAT flag for that,
2766                  * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2767                  * FIXME: We needed it for the index, as index 0 is valid. */
2768         }
2769
2770         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2771                 data->ocd_brw_size = 65536;
2772         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2773                 data->ocd_brw_size = min(data->ocd_brw_size,
2774                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2775                 LASSERT(data->ocd_brw_size);
2776         }
2777
2778         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2779                 __u32 cksum_types = data->ocd_cksum_types;
2780
2781                 /* The client set in ocd_cksum_types the checksum types it
2782                  * supports. We have to mask off the algorithms that we don't
2783                  * support */
2784                 if (cksum_types & OBD_CKSUM_ALL)
2785                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2786                 else
2787                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2788
2789                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2790                                    "%x\n", exp->exp_obd->obd_name,
2791                                    obd_export_nid2str(exp), cksum_types,
2792                                    data->ocd_cksum_types);
2793         } else {
2794                 /* This client does not support OBD_CONNECT_CKSUM
2795                  * fall back to CRC32 */
2796                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2797                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2798                                    exp->exp_obd->obd_name,
2799                                    obd_export_nid2str(exp));
2800         }
2801
2802         RETURN(0);
2803 }
2804
2805 static int filter_reconnect(const struct lu_env *env,
2806                             struct obd_export *exp, struct obd_device *obd,
2807                             struct obd_uuid *cluuid,
2808                             struct obd_connect_data *data,
2809                             void *localdata)
2810 {
2811         int rc;
2812         ENTRY;
2813
2814         if (exp == NULL || obd == NULL || cluuid == NULL)
2815                 RETURN(-EINVAL);
2816
2817         rc = filter_connect_internal(exp, data);
2818         if (rc == 0)
2819                 filter_export_stats_init(obd, exp, localdata);
2820
2821         RETURN(rc);
2822 }
2823
2824 /* nearly identical to mds_connect */
2825 static int filter_connect(const struct lu_env *env,
2826                           struct obd_export **exp, struct obd_device *obd,
2827                           struct obd_uuid *cluuid,
2828                           struct obd_connect_data *data, void *localdata)
2829 {
2830         struct lvfs_run_ctxt saved;
2831         struct lustre_handle conn = { 0 };
2832         struct obd_export *lexp;
2833         struct filter_export_data *fed;
2834         struct lsd_client_data *lcd = NULL;
2835         __u32 group;
2836         int rc;
2837         ENTRY;
2838
2839         if (exp == NULL || obd == NULL || cluuid == NULL)
2840                 RETURN(-EINVAL);
2841
2842         rc = class_connect(&conn, obd, cluuid);
2843         if (rc)
2844                 RETURN(rc);
2845         lexp = class_conn2export(&conn);
2846         LASSERT(lexp != NULL);
2847
2848         fed = &lexp->exp_filter_data;
2849
2850         rc = filter_connect_internal(lexp, data);
2851         if (rc)
2852                 GOTO(cleanup, rc);
2853
2854         filter_export_stats_init(obd, lexp, localdata);
2855         if (obd->obd_replayable) {
2856                 OBD_ALLOC(lcd, sizeof(*lcd));
2857                 if (!lcd) {
2858                         CERROR("filter: out of memory for client data\n");
2859                         GOTO(cleanup, rc = -ENOMEM);
2860                 }
2861
2862                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2863                 fed->fed_lcd = lcd;
2864                 rc = filter_client_add(obd, lexp, -1);
2865                 if (rc)
2866                         GOTO(cleanup, rc);
2867         }
2868
2869         group = data->ocd_group;
2870
2871         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2872               obd->obd_name, lexp->exp_handle.h_cookie, group);
2873
2874         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2875         rc = filter_read_groups(obd, group, 1);
2876         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2877         if (rc != 0) {
2878                 CERROR("can't read group %u\n", group);
2879                 GOTO(cleanup, rc);
2880         }
2881
2882         GOTO(cleanup, rc);
2883
2884 cleanup:
2885         if (rc) {
2886                 if (lcd) {
2887                         OBD_FREE_PTR(lcd);
2888                         fed->fed_lcd = NULL;
2889                 }
2890                 class_disconnect(lexp);
2891                 *exp = NULL;
2892         } else {
2893                 *exp = lexp;
2894         }
2895
2896         RETURN(rc);
2897 }
2898
2899 /* Do extra sanity checks for grant accounting.  We do this at connect,
2900  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2901  * always get rid of it or turn it off when we know accounting is good. */
2902 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2903 {
2904         struct filter_export_data *fed;
2905         struct obd_export *exp;
2906         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2907         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2908         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2909
2910         if (list_empty(&obd->obd_exports))
2911                 return;
2912
2913         /* We don't want to do this for large machines that do lots of
2914            mounts or unmounts.  It burns... */
2915         if (obd->obd_num_exports > 100)
2916                 return;
2917
2918         spin_lock(&obd->obd_osfs_lock);
2919         spin_lock(&obd->obd_dev_lock);
2920         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2921                 int error = 0;
2922                 fed = &exp->exp_filter_data;
2923                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2924                     fed->fed_dirty < 0)
2925                         error = 1;
2926                 if (maxsize > 0) { /* we may not have done a statfs yet */
2927                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2928                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2929                                  exp->exp_client_uuid.uuid, exp,
2930                                  fed->fed_grant, fed->fed_pending, maxsize);
2931                         LASSERTF(fed->fed_dirty <= maxsize,
2932                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2933                                  exp->exp_client_uuid.uuid, exp,
2934                                  fed->fed_dirty, maxsize);
2935                 }
2936                 if (error)
2937                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2938                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2939                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2940                 else
2941                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2942                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2943                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2944                 tot_granted += fed->fed_grant + fed->fed_pending;
2945                 tot_pending += fed->fed_pending;
2946                 tot_dirty += fed->fed_dirty;
2947         }
2948         fo_tot_granted = obd->u.filter.fo_tot_granted;
2949         fo_tot_pending = obd->u.filter.fo_tot_pending;
2950         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2951         spin_unlock(&obd->obd_dev_lock);
2952         spin_unlock(&obd->obd_osfs_lock);
2953
2954         /* Do these assertions outside the spinlocks so we don't kill system */
2955         if (tot_granted != fo_tot_granted)
2956                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2957                        func, tot_granted, fo_tot_granted);
2958         if (tot_pending != fo_tot_pending)
2959                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2960                        func, tot_pending, fo_tot_pending);
2961         if (tot_dirty != fo_tot_dirty)
2962                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2963                        func, tot_dirty, fo_tot_dirty);
2964         if (tot_pending > tot_granted)
2965                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2966                        func, tot_pending, tot_granted);
2967         if (tot_granted > maxsize)
2968                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2969                        func, tot_granted, maxsize);
2970         if (tot_dirty > maxsize)
2971                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2972                        func, tot_dirty, maxsize);
2973 }
2974
2975 /* Remove this client from the grant accounting totals.  We also remove
2976  * the export from the obd device under the osfs and dev locks to ensure
2977  * that the filter_grant_sanity_check() calculations are always valid.
2978  * The client should do something similar when it invalidates its import. */
2979 static void filter_grant_discard(struct obd_export *exp)
2980 {
2981         struct obd_device *obd = exp->exp_obd;
2982         struct filter_obd *filter = &obd->u.filter;
2983         struct filter_export_data *fed = &exp->exp_filter_data;
2984
2985         spin_lock(&obd->obd_osfs_lock);
2986         spin_lock(&obd->obd_dev_lock);
2987         list_del_init(&exp->exp_obd_chain);
2988         spin_unlock(&obd->obd_dev_lock);
2989
2990         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2991                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2992                  obd->obd_name, filter->fo_tot_granted,
2993                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2994         filter->fo_tot_granted -= fed->fed_grant;
2995         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2996                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2997                  obd->obd_name, filter->fo_tot_pending,
2998                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2999         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
3000         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
3001                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
3002                  obd->obd_name, filter->fo_tot_dirty,
3003                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
3004         filter->fo_tot_dirty -= fed->fed_dirty;
3005         fed->fed_dirty = 0;
3006         fed->fed_grant = 0;
3007
3008         spin_unlock(&obd->obd_osfs_lock);
3009 }
3010
3011 static int filter_destroy_export(struct obd_export *exp)
3012 {
3013         ENTRY;
3014
3015         if (exp->exp_filter_data.fed_pending)
3016                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
3017                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
3018                        exp, exp->exp_filter_data.fed_pending);
3019
3020         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3021
3022         target_destroy_export(exp);
3023         ldlm_destroy_export(exp);
3024
3025         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
3026                 RETURN(0);
3027
3028
3029         if (exp->exp_obd->obd_replayable)
3030                 filter_client_free(exp);
3031         else
3032                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3033
3034         filter_grant_discard(exp);
3035         filter_fmd_cleanup(exp);
3036
3037         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3038                 struct filter_obd *filter = &exp->exp_obd->u.filter;
3039                 if (filter->fo_tot_granted_clients > 0)
3040                         filter->fo_tot_granted_clients --;
3041         }
3042
3043         if (!(exp->exp_flags & OBD_OPT_FORCE))
3044                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3045
3046         RETURN(0);
3047 }
3048
3049 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3050 {
3051         struct obd_llog_group *olg_min, *olg;
3052         struct filter_obd *filter;
3053         int worked = -1, group;
3054         struct llog_ctxt *ctxt;
3055         ENTRY;
3056
3057         filter = &obd->u.filter;
3058
3059         /* we can't sync log holding spinlock. also, we do not want to get
3060          * into livelock. so we do following: loop over MDS's exports in
3061          * group order and skip already synced llogs -bzzz */
3062         do {
3063                 /* look for group with min. number, but > worked */
3064                 olg_min = NULL;
3065                 group = 1 << 30;
3066                 spin_lock(&filter->fo_llog_list_lock);
3067                 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3068                         if (olg->olg_group <= worked) {
3069                                 /* this group is already synced */
3070                                 continue;
3071                         }
3072                         if (group < olg->olg_group) {
3073                                 /* we have group with smaller number to sync */
3074                                 continue;
3075                         }
3076                         /* store current minimal group */
3077                         olg_min = olg;
3078                         group = olg->olg_group;
3079                 }
3080                 spin_unlock(&filter->fo_llog_list_lock);
3081
3082                 if (olg_min == NULL)
3083                         break;
3084
3085                 worked = olg_min->olg_group;
3086                 if (olg_min->olg_exp &&
3087                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3088                         int err;
3089                         ctxt = llog_group_get_ctxt(olg_min,
3090                                                    LLOG_MDS_OST_REPL_CTXT);
3091                         if (ctxt) {
3092                                 err = llog_sync(ctxt, olg_min->olg_exp);
3093                                 llog_ctxt_put(ctxt);
3094                                 if (err) {
3095                                         CERROR("error flushing logs to MDS: "
3096                                                "rc %d\n", err);
3097                                 }
3098                         }
3099                 }
3100         } while (olg_min != NULL);
3101 }
3102
3103 /* Also incredibly similar to mds_disconnect */
3104 static int filter_disconnect(struct obd_export *exp)
3105 {
3106         struct obd_device *obd = exp->exp_obd;
3107         int rc;
3108         ENTRY;
3109
3110         LASSERT(exp);
3111         class_export_get(exp);
3112
3113         if (!(exp->exp_flags & OBD_OPT_FORCE))
3114                 filter_grant_sanity_check(obd, __FUNCTION__);
3115         filter_grant_discard(exp);
3116
3117         /* Flush any remaining cancel messages out to the target */
3118         filter_sync_llogs(obd, exp);
3119
3120         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3121
3122         /* Disconnect early so that clients can't keep using export */
3123         rc = class_disconnect(exp);
3124         if (exp->exp_obd->obd_namespace != NULL)
3125                 ldlm_cancel_locks_for_export(exp);
3126
3127         fsfilt_sync(obd, obd->u.obt.obt_sb);
3128
3129         lprocfs_exp_cleanup(exp);
3130
3131         class_export_put(exp);
3132         RETURN(rc);
3133 }
3134
3135 /* reverse import is changed, sync all cancels */
3136 static void filter_revimp_update(struct obd_export *exp)
3137 {
3138         ENTRY;
3139
3140         LASSERT(exp);
3141         class_export_get(exp);
3142
3143         /* flush any remaining cancel messages out to the target */
3144         filter_sync_llogs(exp->exp_obd, exp);
3145         class_export_put(exp);
3146         EXIT;
3147 }
3148
3149 static int filter_ping(struct obd_export *exp)
3150 {
3151         filter_fmd_expire(exp);
3152         return 0;
3153 }
3154
3155 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3156                                   const char *what, int quiet)
3157 {
3158         struct dentry *dchild = NULL;
3159         obd_gr group = 0;
3160
3161         if (oa->o_valid & OBD_MD_FLGROUP)
3162                 group = oa->o_gr;
3163
3164         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3165
3166         if (IS_ERR(dchild)) {
3167                 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3168                        what, group, oa->o_id);
3169                 RETURN(dchild);
3170         }
3171
3172         if (dchild->d_inode == NULL) {
3173                 if (!quiet)
3174                         CERROR("%s: %s on non-existent object: "LPU64"\n",
3175                                obd->obd_name, what, oa->o_id);
3176                 f_dput(dchild);
3177                 RETURN(ERR_PTR(-ENOENT));
3178         }
3179
3180         return dchild;
3181 }
3182
3183 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3184 {
3185         struct dentry *dentry = NULL;
3186         struct obd_device *obd;
3187         int rc = 0;
3188         ENTRY;
3189
3190         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3191                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3192         if (rc)
3193                 RETURN(rc);
3194
3195         obd = class_exp2obd(exp);
3196         if (obd == N