Whamcloud - gitweb
b=21571 stacksize and locking fixes for loadgen patch from umka
[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 static struct lvfs_callback_ops filter_lvfs_ops;
83 cfs_mem_cache_t *ll_fmd_cachep;
84
85 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
86                              void *cb_data, int error)
87 {
88         struct obd_export *exp = cb_data;
89         LASSERT(exp->exp_obd == obd);
90         obd_transno_commit_cb(obd, transno, exp, error);
91         class_export_cb_put(exp);
92 }
93
94 int filter_version_get_check(struct obd_export *exp,
95                              struct obd_trans_info *oti, struct inode *inode)
96 {
97         __u64 curr_version;
98
99         if (inode == NULL || oti == NULL)
100                 RETURN(0);
101
102         curr_version = fsfilt_get_version(exp->exp_obd, inode);
103         if ((__s64)curr_version == -EOPNOTSUPP)
104                 RETURN(0);
105         /* VBR: version is checked always because costs nothing */
106         if (oti->oti_pre_version != 0 &&
107             oti->oti_pre_version != curr_version) {
108                 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
109                        oti->oti_pre_version, curr_version);
110                 spin_lock(&exp->exp_lock);
111                 exp->exp_vbr_failed = 1;
112                 spin_unlock(&exp->exp_lock);
113                 RETURN (-EOVERFLOW);
114         }
115         oti->oti_pre_version = curr_version;
116         RETURN(0);
117 }
118
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121                           struct obd_trans_info *oti, int rc, int force_sync)
122 {
123         struct filter_obd *filter = &exp->exp_obd->u.filter;
124         struct filter_export_data *fed = &exp->exp_filter_data;
125         struct lsd_client_data *lcd;
126         __u64 last_rcvd;
127         loff_t off;
128         int err, log_pri = D_RPCTRACE;
129
130         /* Propagate error code. */
131         if (rc)
132                 RETURN(rc);
133
134         if (!exp->exp_obd->obd_replayable || oti == NULL)
135                 RETURN(rc);
136
137         mutex_down(&fed->fed_lcd_lock);
138         lcd = fed->fed_lcd;
139         /* if the export has already been disconnected, we have no last_rcvd slot,
140          * update server data with latest transno then */
141         if (lcd == NULL) {
142                 mutex_up(&fed->fed_lcd_lock);
143                 CWARN("commit transaction for disconnected client %s: rc %d\n",
144                       exp->exp_client_uuid.uuid, rc);
145                 err = filter_update_server_data(exp->exp_obd,
146                                                 filter->fo_rcvd_filp,
147                                                 filter->fo_fsd);
148                 RETURN(err);
149         }
150
151         /* we don't allocate new transnos for replayed requests */
152         spin_lock(&filter->fo_translock);
153         if (oti->oti_transno == 0) {
154                 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
155                 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
156         } else {
157                 last_rcvd = oti->oti_transno;
158                 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
159                         filter->fo_fsd->lsd_last_transno =
160                                 cpu_to_le64(last_rcvd);
161         }
162         oti->oti_transno = last_rcvd;
163
164         LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
165         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
166         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
167         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
168         spin_unlock(&filter->fo_translock);
169
170         if (inode)
171                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
172
173         off = fed->fed_lr_off;
174         if (off <= 0) {
175                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
176                        fed->fed_lr_idx, fed->fed_lr_off);
177                 err = -EINVAL;
178         } else {
179                 class_export_cb_get(exp); /* released when the cb is called */
180                 if (!force_sync)
181                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
182                                                            last_rcvd,
183                                                            oti->oti_handle,
184                                                            filter_commit_cb,
185                                                            exp);
186
187                 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
188                                           lcd, sizeof(*lcd), &off,
189                                           force_sync | exp->exp_need_sync);
190                 if (force_sync)
191                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
192         }
193         if (err) {
194                 log_pri = D_ERROR;
195                 if (rc == 0)
196                         rc = err;
197         }
198
199         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
200                last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
201         mutex_up(&fed->fed_lcd_lock);
202         RETURN(rc);
203 }
204
205 void f_dput(struct dentry *dentry)
206 {
207         /* Can't go inside filter_ddelete because it can block */
208         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
209                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
210         LASSERT(atomic_read(&dentry->d_count) > 0);
211
212         dput(dentry);
213 }
214
215 static void init_brw_stats(struct brw_stats *brw_stats)
216 {
217         int i;
218         for (i = 0; i < BRW_LAST; i++)
219                 spin_lock_init(&brw_stats->hist[i].oh_lock);
220 }
221
222 static int lprocfs_init_rw_stats(struct obd_device *obd,
223                                  struct lprocfs_stats **stats)
224 {
225         int num_stats;
226
227         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
228                                                         LPROC_FILTER_LAST - 1;
229         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
230         if (*stats == NULL)
231                 return -ENOMEM;
232
233         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
234         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
235                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
236         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
237                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
238
239         return(0);
240 }
241
242 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
243    plus the procfs overhead :( */
244 static int filter_export_stats_init(struct obd_device *obd,
245                                     struct obd_export *exp,
246                                     void *client_nid)
247 {
248         int rc, newnid = 0;
249         ENTRY;
250
251         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
252                 /* Self-export gets no proc entry */
253                 RETURN(0);
254
255         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
256         if (rc) {
257                 /* Mask error for already created
258                  * /proc entries */
259                 if (rc == -EALREADY)
260                         rc = 0;
261                 RETURN(rc);
262         }
263
264         if (newnid) {
265                 struct nid_stat *tmp = exp->exp_nid_stats;
266                 LASSERT(tmp != NULL);
267
268                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
269                 if (tmp->nid_brw_stats == NULL)
270                         RETURN(-ENOMEM);
271
272                 init_brw_stats(tmp->nid_brw_stats);
273                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
274                                         0644, &filter_per_nid_stats_fops,
275                                         exp->exp_nid_stats);
276                 if (rc)
277                         CWARN("Error adding the brw_stats file\n");
278
279                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
280                 if (rc)
281                         RETURN(rc);
282
283                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
284                                             tmp->nid_stats);
285                 if (rc)
286                         RETURN(rc);
287                 /* Always add in ldlm_stats */
288                 tmp->nid_ldlm_stats = 
289                         lprocfs_alloc_stats(LDLM_LAST_OPC - LDLM_FIRST_OPC,
290                                             LPROCFS_STATS_FLAG_NOPERCPU);
291                 if (tmp->nid_ldlm_stats == NULL)
292                         return -ENOMEM;
293
294                 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
295                 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
296                                             tmp->nid_ldlm_stats);
297                 if (rc)
298                         RETURN(rc);
299         }
300
301         RETURN(0);
302 }
303
304 /* Add client data to the FILTER.  We use a bitmap to locate a free space
305  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
306  * Otherwise, we have just read the data from the last_rcvd file and
307  * we know its offset. */
308 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
309                              int cl_idx)
310 {
311         struct filter_obd *filter = &obd->u.filter;
312         struct filter_export_data *fed = &exp->exp_filter_data;
313         unsigned long *bitmap = filter->fo_last_rcvd_slots;
314         int new_client = (cl_idx == -1);
315
316         ENTRY;
317
318         LASSERT(bitmap != NULL);
319         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
320
321         /* Self-export */
322         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
323                 RETURN(0);
324
325         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
326          * there's no need for extra complication here
327          */
328         if (new_client) {
329                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
330         repeat:
331                 if (cl_idx >= LR_MAX_CLIENTS) {
332                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
333                                cl_idx);
334                         RETURN(-EOVERFLOW);
335                 }
336                 if (test_and_set_bit(cl_idx, bitmap)) {
337                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
338                                                     cl_idx);
339                         goto repeat;
340                 }
341         } else {
342                 if (test_and_set_bit(cl_idx, bitmap)) {
343                         CERROR("FILTER client %d: bit already set in bitmap!\n",
344                                cl_idx);
345                         LBUG();
346                 }
347         }
348
349         fed->fed_lr_idx = cl_idx;
350         fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
351                 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
352         init_mutex(&fed->fed_lcd_lock);
353         LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
354
355         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
356                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
357
358         if (new_client) {
359                 struct lvfs_run_ctxt saved;
360                 loff_t off = fed->fed_lr_off;
361                 int rc;
362                 void *handle;
363
364                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
365                        fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
366
367                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
368                 /* Transaction needed to fix bug 1403 */
369                 handle = fsfilt_start(obd,
370                                       filter->fo_rcvd_filp->f_dentry->d_inode,
371                                       FSFILT_OP_SETATTR, NULL);
372                 if (IS_ERR(handle)) {
373                         rc = PTR_ERR(handle);
374                         CERROR("unable to start transaction: rc %d\n", rc);
375                 } else {
376                         fed->fed_lcd->lcd_last_epoch =
377                                               filter->fo_fsd->lsd_start_epoch;
378                         exp->exp_last_request_time = cfs_time_current_sec();
379                         rc = fsfilt_add_journal_cb(obd, 0, handle,
380                                                    target_client_add_cb,
381                                                    class_export_cb_get(exp));
382                         if (rc == 0) {
383                                 spin_lock(&exp->exp_lock);
384                                 exp->exp_need_sync = 1;
385                                 spin_unlock(&exp->exp_lock);
386                         }
387                         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
388                                                  fed->fed_lcd,
389                                                  sizeof(*fed->fed_lcd),
390                                                  &off, rc /* sync if no cb */);
391                         fsfilt_commit(obd,
392                                       filter->fo_rcvd_filp->f_dentry->d_inode,
393                                       handle, 0);
394                 }
395                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
396
397                 if (rc) {
398                         CERROR("error writing %s client idx %u: rc %d\n",
399                                LAST_RCVD, fed->fed_lr_idx, rc);
400                         RETURN(rc);
401                 }
402         }
403         RETURN(0);
404 }
405
406 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
407
408 static int filter_client_free(struct obd_export *exp)
409 {
410         struct filter_export_data *fed = &exp->exp_filter_data;
411         struct filter_obd *filter = &exp->exp_obd->u.filter;
412         struct obd_device *obd = exp->exp_obd;
413         struct lvfs_run_ctxt saved;
414         struct lsd_client_data *lcd = fed->fed_lcd;
415         int rc;
416         loff_t off;
417         ENTRY;
418
419         if (lcd == NULL)
420                 RETURN(0);
421
422         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
423         if (strcmp(lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
424                 GOTO(free, 0);
425
426         LASSERT(filter->fo_last_rcvd_slots != NULL);
427
428         off = fed->fed_lr_off;
429
430         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
431                fed->fed_lr_idx, fed->fed_lr_off, lcd->lcd_uuid);
432
433         /* Don't clear fed_lr_idx here as it is likely also unset.  At worst
434          * we leak a client slot that will be cleaned on the next recovery. */
435         if (off <= 0) {
436                 CERROR("%s: client idx %d has med_off %lld\n",
437                        obd->obd_name, fed->fed_lr_idx, off);
438                 GOTO(free, rc = -EINVAL);
439         }
440
441         /* Clear the bit _after_ zeroing out the client so we don't
442            race with filter_client_add and zero out new clients.*/
443         if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
444                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
445                        fed->fed_lr_idx);
446                 LBUG();
447         }
448
449         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
450         /* Make sure the server's last_transno is up to date.
451          * This should be done before zeroing client slot so last_transno will
452          * be in server data or in client data in case of failure */
453         filter_update_server_data(obd, filter->fo_rcvd_filp, filter->fo_fsd);
454
455         mutex_down(&fed->fed_lcd_lock);
456         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
457                                  sizeof(zero_lcd), &off, 0);
458         fed->fed_lcd = NULL;
459         mutex_up(&fed->fed_lcd_lock);
460         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
461
462         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
463                "zero out client %s at idx %u/%llu in %s, rc %d\n",
464                lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
465                LAST_RCVD, rc);
466
467         if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
468                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
469                        fed->fed_lr_idx);
470                 LBUG();
471         }
472         OBD_FREE_PTR(lcd);
473         RETURN(0);
474 free:
475         mutex_down(&fed->fed_lcd_lock);
476         fed->fed_lcd = NULL;
477         mutex_up(&fed->fed_lcd_lock);
478         OBD_FREE_PTR(lcd);
479
480         return 0;
481 }
482
483 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
484 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
485                                          struct filter_mod_data *fmd)
486 {
487         LASSERT_SPIN_LOCKED(&fed->fed_lock);
488         if (--fmd->fmd_refcount == 0) {
489                 /* XXX when we have persistent reservations and the handle
490                  * is stored herein we need to drop it here. */
491                 fed->fed_mod_count--;
492                 list_del(&fmd->fmd_list);
493                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
494         }
495 }
496
497 /* drop fmd reference, free it if last ref */
498 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
499 {
500         struct filter_export_data *fed;
501
502         if (fmd == NULL)
503                 return;
504
505         fed = &exp->exp_filter_data;
506         spin_lock(&fed->fed_lock);
507         filter_fmd_put_nolock(fed, fmd); /* caller reference */
508         spin_unlock(&fed->fed_lock);
509 }
510
511 /* expire entries from the end of the list if there are too many
512  * or they are too old */
513 static void filter_fmd_expire_nolock(struct filter_obd *filter,
514                                      struct filter_export_data *fed,
515                                      struct filter_mod_data *keep)
516 {
517         struct filter_mod_data *fmd, *tmp;
518
519         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
520                 if (fmd == keep)
521                         break;
522
523                 if (time_before(jiffies, fmd->fmd_expire) &&
524                     fed->fed_mod_count < filter->fo_fmd_max_num)
525                         break;
526
527                 list_del_init(&fmd->fmd_list);
528                 filter_fmd_put_nolock(fed, fmd); /* list reference */
529         }
530 }
531
532 void filter_fmd_expire(struct obd_export *exp)
533 {
534         spin_lock(&exp->exp_filter_data.fed_lock);
535         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
536                                  &exp->exp_filter_data, NULL);
537         spin_unlock(&exp->exp_filter_data.fed_lock);
538 }
539
540 /* find specified objid, group in export fmd list.
541  * caller must hold fed_lock and take fmd reference itself */
542 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
543                                                 struct filter_export_data *fed,
544                                                 obd_id objid, obd_gr group)
545 {
546         struct filter_mod_data *found = NULL, *fmd;
547
548         LASSERT_SPIN_LOCKED(&fed->fed_lock);
549
550         list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
551                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
552                         found = fmd;
553                         list_del(&fmd->fmd_list);
554                         list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
555                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
556                         break;
557                 }
558         }
559
560         filter_fmd_expire_nolock(filter, fed, found);
561
562         return found;
563 }
564
565 /* Find fmd based on objid and group, or return NULL if not found. */
566 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
567                                         obd_id objid, obd_gr group)
568 {
569         struct filter_mod_data *fmd;
570
571         spin_lock(&exp->exp_filter_data.fed_lock);
572         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
573                                      &exp->exp_filter_data, objid, group);
574         if (fmd)
575                 fmd->fmd_refcount++;    /* caller reference */
576         spin_unlock(&exp->exp_filter_data.fed_lock);
577
578         return fmd;
579 }
580
581 /* Find fmd based on objid and group, or create a new one if none is found.
582  * It is possible for this function to return NULL under memory pressure,
583  * or if objid = 0 is passed (which will only cause old entries to expire).
584  * Currently this is not fatal because any fmd state is transient and
585  * may also be freed when it gets sufficiently old. */
586 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
587                                        obd_id objid, obd_gr group)
588 {
589         struct filter_export_data *fed = &exp->exp_filter_data;
590         struct filter_mod_data *found = NULL, *fmd_new = NULL;
591
592         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
593
594         spin_lock(&fed->fed_lock);
595         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
596         if (fmd_new) {
597                 if (found == NULL) {
598                         list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
599                         fmd_new->fmd_id = objid;
600                         fmd_new->fmd_gr = group;
601                         fmd_new->fmd_refcount++;   /* list reference */
602                         found = fmd_new;
603                         fed->fed_mod_count++;
604                 } else {
605                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
606                 }
607         }
608         if (found) {
609                 found->fmd_refcount++;          /* caller reference */
610                 found->fmd_expire = jiffies +
611                         exp->exp_obd->u.filter.fo_fmd_max_age;
612         }
613
614         spin_unlock(&fed->fed_lock);
615
616         return found;
617 }
618
619 #ifdef DO_FMD_DROP
620 /* drop fmd list reference so it will disappear when last reference is put.
621  * This isn't so critical because it would in fact only affect the one client
622  * that is doing the unlink and at worst we have an stale entry referencing
623  * an object that should never be used again. */
624 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
625 {
626         struct filter_mod_data *found = NULL;
627
628         spin_lock(&exp->exp_filter_data.fed_lock);
629         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
630         if (found) {
631                 list_del_init(&found->fmd_list);
632                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
633         }
634         spin_unlock(&exp->exp_filter_data.fed_lock);
635 }
636 #else
637 #define filter_fmd_drop(exp, objid, group)
638 #endif
639
640 /* remove all entries from fmd list */
641 static void filter_fmd_cleanup(struct obd_export *exp)
642 {
643         struct filter_export_data *fed = &exp->exp_filter_data;
644         struct filter_mod_data *fmd = NULL, *tmp;
645
646         spin_lock(&fed->fed_lock);
647         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
648                 list_del_init(&fmd->fmd_list);
649                 filter_fmd_put_nolock(fed, fmd);
650         }
651         spin_unlock(&fed->fed_lock);
652 }
653
654 static int filter_init_export(struct obd_export *exp)
655 {
656         spin_lock_init(&exp->exp_filter_data.fed_lock);
657         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
658
659         spin_lock(&exp->exp_lock);
660         exp->exp_connecting = 1;
661         spin_unlock(&exp->exp_lock);
662
663         return ldlm_init_export(exp);
664 }
665
666 static int filter_free_server_data(struct filter_obd *filter)
667 {
668         OBD_FREE_PTR(filter->fo_fsd);
669         filter->fo_fsd = NULL;
670         OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
671         filter->fo_last_rcvd_slots = NULL;
672         return 0;
673 }
674
675 /* assumes caller is already in kernel ctxt */
676 int filter_update_server_data(struct obd_device *obd, struct file *filp,
677                               struct lr_server_data *fsd)
678 {
679         loff_t off = 0;
680         int rc;
681         ENTRY;
682
683         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
684         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
685                le64_to_cpu(fsd->lsd_last_transno));
686         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
687                le64_to_cpu(fsd->lsd_mount_count));
688
689         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, 0);
690         if (rc)
691                 CERROR("error writing lr_server_data: rc = %d\n", rc);
692
693         RETURN(rc);
694 }
695
696 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
697                              int force_sync)
698 {
699         struct filter_obd *filter = &obd->u.filter;
700         __u64 tmp;
701         loff_t off = 0;
702         int rc;
703         ENTRY;
704
705         if (filter->fo_last_objid_files[group] == NULL) {
706                 CERROR("Object group "LPU64" not fully setup; not updating "
707                        "last_objid\n", group);
708                 RETURN(-EINVAL);
709         }
710
711         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
712                obd->obd_name, group, filter->fo_last_objids[group]);
713
714         tmp = cpu_to_le64(filter->fo_last_objids[group]);
715         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
716                                  &tmp, sizeof(tmp), &off, force_sync);
717         if (rc)
718                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
719                        group, rc);
720         RETURN(rc);
721 }
722 extern int ost_handle(struct ptlrpc_request *req);
723 /* assumes caller has already in kernel ctxt */
724 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
725 {
726         struct filter_obd *filter = &obd->u.filter;
727         struct lr_server_data *fsd;
728         struct lsd_client_data *lcd = NULL;
729         struct inode *inode = filp->f_dentry->d_inode;
730         unsigned long last_rcvd_size = i_size_read(inode);
731         __u64 mount_count;
732         __u32 start_epoch;
733         int cl_idx;
734         loff_t off = 0;
735         int rc;
736
737         /* ensure padding in the struct is the correct size */
738         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
739                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
740         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
741                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
742
743         OBD_ALLOC(fsd, sizeof(*fsd));
744         if (!fsd)
745                 RETURN(-ENOMEM);
746         filter->fo_fsd = fsd;
747
748         OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
749         if (filter->fo_last_rcvd_slots == NULL) {
750                 OBD_FREE(fsd, sizeof(*fsd));
751                 RETURN(-ENOMEM);
752         }
753
754         if (last_rcvd_size == 0) {
755                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
756
757                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
758                 fsd->lsd_last_transno = 0;
759                 mount_count = fsd->lsd_mount_count = 0;
760                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
761                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
762                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
763                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
764                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
765                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
766                  * MDS first connects and assigns the OST index number. */
767                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
768                                                         OBD_INCOMPAT_OST);
769         } else {
770                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
771                 if (rc) {
772                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
773                                LAST_RCVD, rc);
774                         GOTO(err_fsd, rc);
775                 }
776                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
777                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
778                                            "using the wrong disk %s. Were the "
779                                            "/dev/ assignments rearranged?\n",
780                                            obd->obd_uuid.uuid, fsd->lsd_uuid);
781                         GOTO(err_fsd, rc = -EINVAL);
782                 }
783                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
784                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
785                 /* COMPAT_146 */
786                 /* Assume old last_rcvd format unless I_C_LR is set */
787                 if (!(fsd->lsd_feature_incompat &
788                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
789                         fsd->lsd_last_transno = fsd->lsd_compat14;
790                 /* end COMPAT_146 */
791                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
792                  * MDS first connects and assigns the OST index number. */
793                 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
794                                                          OBD_INCOMPAT_OST);
795         }
796
797         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
798                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
799                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
800                        ~FILTER_INCOMPAT_SUPP);
801                 GOTO(err_fsd, rc = -EINVAL);
802         }
803         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
804                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
805                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
806                        ~FILTER_ROCOMPAT_SUPP);
807                 /* Do something like remount filesystem read-only */
808                 GOTO(err_fsd, rc = -EINVAL);
809         }
810
811         start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
812
813         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
814                obd->obd_name, start_epoch);
815         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
816                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
817         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
818                obd->obd_name, mount_count + 1);
819         CDEBUG(D_INODE, "%s: server data size: %u\n",
820                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
821         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
822                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
823         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
824                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
825         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
826                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
827         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
828                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
829                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
830                 le16_to_cpu(fsd->lsd_client_size));
831
832         if (!obd->obd_replayable) {
833                 CWARN("%s: recovery support OFF\n", obd->obd_name);
834                 GOTO(out, rc = 0);
835         }
836
837         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
838              off < last_rcvd_size; cl_idx++) {
839                 __u64 last_rcvd;
840                 struct obd_export *exp;
841                 struct filter_export_data *fed;
842
843                 if (!lcd) {
844                         OBD_ALLOC_PTR(lcd);
845                         if (!lcd)
846                                 GOTO(err_client, rc = -ENOMEM);
847                 }
848
849                 /* Don't assume off is incremented properly by
850                  * fsfilt_read_record(), in case sizeof(*lcd)
851                  * isn't the same as fsd->lsd_client_size.  */
852                 off = le32_to_cpu(fsd->lsd_client_start) +
853                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
854                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
855                 if (rc) {
856                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
857                                LAST_RCVD, cl_idx, off, rc);
858                         break; /* read error shouldn't cause startup to fail */
859                 }
860
861                 if (lcd->lcd_uuid[0] == '\0') {
862                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
863                                cl_idx);
864                         continue;
865                 }
866
867                 check_lcd(obd->obd_name, cl_idx, lcd);
868
869                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
870
871                 /* These exports are cleaned up by filter_disconnect(), so they
872                  * need to be set up like real exports as filter_connect() does.
873                  */
874                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
875
876                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
877                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
878                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
879                 if (IS_ERR(exp)) {
880                         if (PTR_ERR(exp) == -EALREADY) {
881                                 /* export already exists, zero out this one */
882                                 CERROR("Zeroing out duplicate export due to "
883                                        "bug 10479.\n");
884                                 lcd->lcd_uuid[0] = '\0';
885                         } else {
886                                 GOTO(err_client, rc = PTR_ERR(exp));
887                         }
888                 } else {
889                         fed = &exp->exp_filter_data;
890                         fed->fed_lcd = lcd;
891                         fed->fed_group = 0; /* will be assigned at connect */
892                         filter_export_stats_init(obd, exp, NULL);
893                         rc = filter_client_add(obd, exp, cl_idx);
894                         /* can't fail for existing client */
895                         LASSERTF(rc == 0, "rc = %d\n", rc);
896
897                         /* VBR: set export last committed */
898                         exp->exp_last_committed = last_rcvd;
899                         spin_lock(&exp->exp_lock);
900                         exp->exp_connecting = 0;
901                         exp->exp_in_recovery = 0;
902                         spin_unlock(&exp->exp_lock);
903                         spin_lock_bh(&obd->obd_processing_task_lock);
904                         obd->obd_max_recoverable_clients++;
905                         spin_unlock_bh(&obd->obd_processing_task_lock);
906                         lcd = NULL;
907                         class_export_put(exp);
908                 }
909
910                 /* Need to check last_rcvd even for duplicated exports. */
911                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
912                        cl_idx, last_rcvd);
913
914                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
915                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
916         }
917
918         if (lcd)
919                 OBD_FREE_PTR(lcd);
920
921         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
922 out:
923         filter->fo_mount_count = mount_count + 1;
924         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
925
926         /* save it, so mount count and last_transno is current */
927         rc = filter_update_server_data(obd, filp, filter->fo_fsd);
928         if (rc)
929                 GOTO(err_client, rc);
930
931         RETURN(0);
932
933 err_client:
934         class_disconnect_exports(obd);
935 err_fsd:
936         filter_free_server_data(filter);
937         RETURN(rc);
938 }
939
940 static int filter_cleanup_groups(struct obd_device *obd)
941 {
942         struct filter_obd *filter = &obd->u.filter;
943         struct file *filp;
944         struct dentry *dentry;
945         int i, j;
946         ENTRY;
947
948         if (filter->fo_dentry_O_groups != NULL) {
949                 for (i = 0; i < filter->fo_group_count; i++) {
950                         dentry = filter->fo_dentry_O_groups[i];
951                         if (dentry != NULL)
952                                 f_dput(dentry);
953                 }
954                 OBD_FREE(filter->fo_dentry_O_groups,
955                          filter->fo_group_count *
956                          sizeof(*filter->fo_dentry_O_groups));
957                 filter->fo_dentry_O_groups = NULL;
958         }
959         if (filter->fo_last_objid_files != NULL) {
960                 for (i = 0; i < filter->fo_group_count; i++) {
961                         filp = filter->fo_last_objid_files[i];
962                         if (filp != NULL)
963                                 filp_close(filp, 0);
964                 }
965                 OBD_FREE(filter->fo_last_objid_files,
966                          filter->fo_group_count *
967                          sizeof(*filter->fo_last_objid_files));
968                 filter->fo_last_objid_files = NULL;
969         }
970         if (filter->fo_dentry_O_sub != NULL) {
971                 for (i = 0; i < filter->fo_group_count; i++) {
972                         for (j = 0; j < filter->fo_subdir_count; j++) {
973                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
974                                 if (dentry != NULL)
975                                         f_dput(dentry);
976                         }
977                 }
978                 OBD_FREE(filter->fo_dentry_O_sub,
979                          filter->fo_group_count *
980                          sizeof(*filter->fo_dentry_O_sub));
981                 filter->fo_dentry_O_sub = NULL;
982         }
983         if (filter->fo_last_objids != NULL) {
984                 OBD_FREE(filter->fo_last_objids,
985                          filter->fo_group_count *
986                          sizeof(*filter->fo_last_objids));
987                 filter->fo_last_objids = NULL;
988         }
989         if (filter->fo_dentry_O != NULL) {
990                 f_dput(filter->fo_dentry_O);
991                 filter->fo_dentry_O = NULL;
992         }
993         RETURN(0);
994 }
995
996 static int filter_update_last_group(struct obd_device *obd, int group)
997 {
998         struct filter_obd *filter = &obd->u.filter;
999         struct file *filp = NULL;
1000         int last_group = 0, rc;
1001         loff_t off = 0;
1002         ENTRY;
1003
1004         if (group <= filter->fo_committed_group)
1005                 RETURN(0);
1006
1007         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1008         if (IS_ERR(filp)) {
1009                 rc = PTR_ERR(filp);
1010                 filp = NULL;
1011                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1012                 GOTO(cleanup, rc);
1013         }
1014
1015         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1016         if (rc) {
1017                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1018                 GOTO(cleanup, rc);
1019         }
1020
1021         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1022                obd->obd_name, last_group, group);
1023
1024         off = 0;
1025         last_group = group;
1026         /* must be sync: bXXXX */
1027         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1028         if (rc) {
1029                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1030                 GOTO(cleanup, rc);
1031         }
1032
1033         filter->fo_committed_group = group;
1034 cleanup:
1035         if (filp)
1036                 filp_close(filp, 0);
1037         RETURN(rc);
1038 }
1039
1040 static int filter_read_group_internal(struct obd_device *obd, int group,
1041                                       int create)
1042 {
1043         struct filter_obd *filter = &obd->u.filter;
1044         __u64 *new_objids = NULL;
1045         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1046         struct dentry **new_groups = NULL;
1047         struct file **new_files = NULL;
1048         struct dentry *dentry;
1049         struct file *filp;
1050         int old_count = filter->fo_group_count, rc, stage = 0, i;
1051         char name[25];
1052         __u64 last_objid;
1053         loff_t off = 0;
1054         int len = group + 1;
1055
1056         snprintf(name, 24, "%d", group);
1057         name[24] = '\0';
1058
1059         if (!create) {
1060                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1061                                            strlen(name));
1062                 if (IS_ERR(dentry)) {
1063                         CERROR("Cannot lookup expected object group %d: %ld\n",
1064                                group, PTR_ERR(dentry));
1065                         RETURN(PTR_ERR(dentry));
1066                 }
1067         } else {
1068                 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1069                                       name, 0700, 1);
1070                 if (IS_ERR(dentry)) {
1071                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1072                                PTR_ERR(dentry));
1073                         RETURN(PTR_ERR(dentry));
1074                 }
1075         }
1076         stage = 1;
1077
1078         snprintf(name, 24, "O/%d/LAST_ID", group);
1079         name[24] = '\0';
1080         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1081         if (IS_ERR(filp)) {
1082                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1083                 GOTO(cleanup, rc = PTR_ERR(filp));
1084         }
1085         stage = 2;
1086
1087         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1088         if (rc) {
1089                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1090                 GOTO(cleanup, rc);
1091         }
1092
1093         if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1094                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1095                 if (tmp_subdirs == NULL)
1096                         GOTO(cleanup, rc = -ENOMEM);
1097                 stage = 3;
1098
1099                 for (i = 0; i < filter->fo_subdir_count; i++) {
1100                         char dir[20];
1101                         snprintf(dir, sizeof(dir), "d%u", i);
1102
1103                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1104                                                               filter->fo_vfsmnt,
1105                                                               dir, 0700, 1);
1106                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1107                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1108                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1109                                        group, dir, rc);
1110                                 GOTO(cleanup, rc);
1111                         }
1112
1113                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1114                                tmp_subdirs->dentry[i]);
1115                 }
1116         }
1117
1118         /* 'group' is an index; we need an array of length 'group + 1' */
1119         if (group + 1 > old_count) {
1120                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1121                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1122                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1123                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1124                 stage = 4;
1125                 if (new_objids == NULL || new_subdirs == NULL ||
1126                     new_groups == NULL || new_files == NULL)
1127                         GOTO(cleanup, rc = -ENOMEM);
1128
1129                 if (old_count) {
1130                         memcpy(new_objids, filter->fo_last_objids,
1131                                old_count * sizeof(*new_objids));
1132                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1133                                old_count * sizeof(*new_subdirs));
1134                         memcpy(new_groups, filter->fo_dentry_O_groups,
1135                                old_count * sizeof(*new_groups));
1136                         memcpy(new_files, filter->fo_last_objid_files,
1137                                old_count * sizeof(*new_files));
1138
1139                         OBD_FREE(filter->fo_last_objids,
1140                                  old_count * sizeof(*new_objids));
1141                         OBD_FREE(filter->fo_dentry_O_sub,
1142                                  old_count * sizeof(*new_subdirs));
1143                         OBD_FREE(filter->fo_dentry_O_groups,
1144                                  old_count * sizeof(*new_groups));
1145                         OBD_FREE(filter->fo_last_objid_files,
1146                                  old_count * sizeof(*new_files));
1147                 }
1148                 filter->fo_last_objids = new_objids;
1149                 filter->fo_dentry_O_sub = new_subdirs;
1150                 filter->fo_dentry_O_groups = new_groups;
1151                 filter->fo_last_objid_files = new_files;
1152                 filter->fo_group_count = len;
1153         }
1154
1155         filter->fo_dentry_O_groups[group] = dentry;
1156         filter->fo_last_objid_files[group] = filp;
1157         if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1158                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1159                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1160         }
1161
1162         filter_update_last_group(obd, group);
1163
1164         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1165                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1166                 rc = filter_update_last_objid(obd, group, 1);
1167                 RETURN(rc);
1168         }
1169
1170         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1171         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1172                obd->obd_name, group, last_objid);
1173         RETURN(0);
1174  cleanup:
1175         switch (stage) {
1176         case 4:
1177                 if (new_objids != NULL)
1178                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1179                 if (new_subdirs != NULL)
1180                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1181                 if (new_groups != NULL)
1182                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1183                 if (new_files != NULL)
1184                         OBD_FREE(new_files, len * sizeof(*new_files));
1185         case 3:
1186                 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1187                         for (i = 0; i < filter->fo_subdir_count; i++) {
1188                                 if (tmp_subdirs->dentry[i] != NULL)
1189                                         dput(tmp_subdirs->dentry[i]);
1190                         }
1191                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1192                 }
1193         case 2:
1194                 filp_close(filp, 0);
1195         case 1:
1196                 dput(dentry);
1197         }
1198         RETURN(rc);
1199 }
1200
1201 static int filter_read_groups(struct obd_device *obd, int last_group,
1202                               int create)
1203 {
1204         struct filter_obd *filter = &obd->u.filter;
1205         int old_count, group, rc = 0;
1206
1207         down(&filter->fo_init_lock);
1208         old_count = filter->fo_group_count;
1209         for (group = old_count; group <= last_group; group++) {
1210                 rc = filter_read_group_internal(obd, group, create);
1211                 if (rc != 0)
1212                         break;
1213         }
1214         up(&filter->fo_init_lock);
1215         return rc;
1216 }
1217
1218 /* FIXME: object groups */
1219 static int filter_prep_groups(struct obd_device *obd)
1220 {
1221         struct filter_obd *filter = &obd->u.filter;
1222         struct dentry *O_dentry;
1223         struct file *filp;
1224         int    last_group, rc = 0, cleanup_phase = 0;
1225         loff_t off = 0;
1226         ENTRY;
1227
1228         O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1229                                 "O", 0700, 1);
1230         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1231         if (IS_ERR(O_dentry)) {
1232                 rc = PTR_ERR(O_dentry);
1233                 CERROR("cannot open/create O: rc = %d\n", rc);
1234                 GOTO(cleanup, rc);
1235         }
1236         filter->fo_dentry_O = O_dentry;
1237         cleanup_phase = 1; /* O_dentry */
1238
1239         /* we have to initialize all groups before first connections from
1240          * clients because they may send create/destroy for any group -bzzz */
1241         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1242         if (IS_ERR(filp)) {
1243                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1244                 GOTO(cleanup, rc = PTR_ERR(filp));
1245         }
1246         cleanup_phase = 2; /* filp */
1247
1248         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1249         if (rc) {
1250                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1251                 GOTO(cleanup, rc);
1252         }
1253
1254         if (off == 0)
1255                 last_group = FILTER_GROUP_MDS0;
1256
1257         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1258               FILTER_GROUP_MDS0, last_group);
1259         filter->fo_committed_group = last_group;
1260         rc = filter_read_groups(obd, last_group, 1);
1261         if (rc)
1262                 GOTO(cleanup, rc);
1263
1264         filp_close(filp, 0);
1265         RETURN(0);
1266
1267  cleanup:
1268         switch (cleanup_phase) {
1269         case 2:
1270                 filp_close(filp, 0);
1271         case 1:
1272                 filter_cleanup_groups(obd);
1273                 f_dput(filter->fo_dentry_O);
1274                 filter->fo_dentry_O = NULL;
1275         default:
1276                 break;
1277         }
1278         return rc;
1279
1280 }
1281
1282 /* setup the object store with correct subdirectories */
1283 static int filter_prep(struct obd_device *obd)
1284 {
1285         struct lvfs_run_ctxt saved;
1286         struct filter_obd *filter = &obd->u.filter;
1287         struct file *file;
1288         struct inode *inode;
1289         int rc = 0;
1290         ENTRY;
1291
1292         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1293         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1294         if (!file || IS_ERR(file)) {
1295                 rc = PTR_ERR(file);
1296                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1297                        LAST_RCVD, rc);
1298                 GOTO(out, rc);
1299         }
1300         filter->fo_rcvd_filp = file;
1301         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1302                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1303                        file->f_dentry->d_inode->i_mode);
1304                 GOTO(err_filp, rc = -ENOENT);
1305         }
1306
1307         inode = file->f_dentry->d_parent->d_inode;
1308         /* We use i_op->unlink directly in filter_vfs_unlink() */
1309         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1310                 CERROR("%s: filesystem does not support create/unlink ops\n",
1311                        obd->obd_name);
1312                 GOTO(err_filp, rc = -EOPNOTSUPP);
1313         }
1314
1315         /** lu_target has very limited use in filter now */
1316         lut_init(NULL, &filter->fo_lut, obd, NULL);
1317
1318         rc = filter_init_server_data(obd, file);
1319         if (rc) {
1320                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1321                 GOTO(err_filp, rc);
1322         }
1323
1324         target_recovery_init(&filter->fo_lut, ost_handle);
1325
1326         /* open and create health check io file*/
1327         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1328         if (IS_ERR(file)) {
1329                 rc = PTR_ERR(file);
1330                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1331                        HEALTH_CHECK, rc);
1332                 GOTO(err_server_data, rc);
1333         }
1334         filter->fo_obt.obt_health_check_filp = file;
1335         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1336                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1337                        file->f_dentry->d_inode->i_mode);
1338                 GOTO(err_health_check, rc = -ENOENT);
1339         }
1340         rc = lvfs_check_io_health(obd, file);
1341         if (rc)
1342                 GOTO(err_health_check, rc);
1343
1344         rc = filter_prep_groups(obd);
1345         if (rc)
1346                 GOTO(err_health_check, rc);
1347 out:
1348         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1349
1350         return(rc);
1351
1352 err_health_check:
1353         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1354                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1355         filter->fo_obt.obt_health_check_filp = NULL;
1356 err_server_data:
1357         target_recovery_fini(obd);
1358         filter_free_server_data(filter);
1359 err_filp:
1360         if (filp_close(filter->fo_rcvd_filp, 0))
1361                 CERROR("can't close %s after error\n", LAST_RCVD);
1362         filter->fo_rcvd_filp = NULL;
1363         goto out;
1364 }
1365
1366 /* cleanup the filter: write last used object id to status file */
1367 static void filter_post(struct obd_device *obd)
1368 {
1369         struct lvfs_run_ctxt saved;
1370         struct filter_obd *filter = &obd->u.filter;
1371         int rc, i;
1372
1373         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1374          * best to start a transaction with h_sync, because we removed this
1375          * from lastobjid */
1376
1377         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1378         rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1379                                        filter->fo_fsd);
1380         if (rc)
1381                 CERROR("error writing server data: rc = %d\n", rc);
1382
1383         for (i = 0; i < filter->fo_group_count; i++) {
1384                 rc = filter_update_last_objid(obd, i,
1385                                 (i == filter->fo_group_count - 1));
1386                 if (rc)
1387                         CERROR("error writing group %d lastobjid: rc = %d\n",
1388                                i, rc);
1389         }
1390
1391         rc = filp_close(filter->fo_rcvd_filp, 0);
1392         filter->fo_rcvd_filp = NULL;
1393         if (rc)
1394                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1395
1396         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1397         filter->fo_obt.obt_health_check_filp = NULL;
1398         if (rc)
1399                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1400
1401         filter_cleanup_groups(obd);
1402         filter_free_server_data(filter);
1403         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1404
1405         filter_free_capa_keys(filter);
1406         cleanup_capa_hash(filter->fo_capa_hash);
1407 }
1408
1409 static void filter_set_last_id(struct filter_obd *filter,
1410                                obd_id id, obd_gr group)
1411 {
1412         LASSERT(filter->fo_fsd != NULL);
1413         LASSERT(group <= filter->fo_group_count);
1414
1415         spin_lock(&filter->fo_objidlock);
1416         filter->fo_last_objids[group] = id;
1417         spin_unlock(&filter->fo_objidlock);
1418 }
1419
1420 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1421 {
1422         obd_id id;
1423         LASSERT(filter->fo_fsd != NULL);
1424         LASSERT(group <= filter->fo_group_count);
1425
1426         /* FIXME: object groups */
1427         spin_lock(&filter->fo_objidlock);
1428         id = filter->fo_last_objids[group];
1429         spin_unlock(&filter->fo_objidlock);
1430         return id;
1431 }
1432
1433 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1434 {
1435         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1436         return 0;
1437 }
1438
1439 /* We never dget the object parent, so DON'T dput it either */
1440 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1441 {
1442         struct filter_obd *filter = &obd->u.filter;
1443         struct filter_subdirs *subdirs;
1444         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1445
1446         if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1447                 return filter->fo_dentry_O_groups[group];
1448
1449         subdirs = &filter->fo_dentry_O_sub[group];
1450         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1451 }
1452
1453 /* We never dget the object parent, so DON'T dput it either */
1454 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1455                                   obd_id objid)
1456 {
1457         unsigned long now = jiffies;
1458         struct dentry *dparent = filter_parent(obd, group, objid);
1459         int rc;
1460
1461         if (IS_ERR(dparent))
1462                 return dparent;
1463         if (dparent == NULL)
1464                 return ERR_PTR(-ENOENT);
1465
1466         rc = filter_lock_dentry(obd, dparent);
1467         fsfilt_check_slow(obd, now, "parent lock");
1468         return rc ? ERR_PTR(rc) : dparent;
1469 }
1470
1471 /* We never dget the object parent, so DON'T dput it either */
1472 static void filter_parent_unlock(struct dentry *dparent)
1473 {
1474         UNLOCK_INODE_MUTEX(dparent->d_inode);
1475 }
1476
1477 /* How to get files, dentries, inodes from object id's.
1478  *
1479  * If dir_dentry is passed, the caller has already locked the parent
1480  * appropriately for this operation (normally a write lock).  If
1481  * dir_dentry is NULL, we do a read lock while we do the lookup to
1482  * avoid races with create/destroy and such changing the directory
1483  * internal to the filesystem code. */
1484 struct dentry *filter_fid2dentry(struct obd_device *obd,
1485                                  struct dentry *dir_dentry,
1486                                  obd_gr group, obd_id id)
1487 {
1488         struct dentry *dparent = dir_dentry;
1489         struct dentry *dchild;
1490         char name[32];
1491         int len;
1492         ENTRY;
1493
1494         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1495             obd->u.filter.fo_destroys_in_progress == 0) {
1496                 /* don't fail lookups for orphan recovery, it causes
1497                  * later LBUGs when objects still exist during precreate */
1498                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1499                 RETURN(ERR_PTR(-ENOENT));
1500         }
1501         if (id == 0) {
1502                 CERROR("fatal: invalid object id 0\n");
1503                 RETURN(ERR_PTR(-ESTALE));
1504         }
1505
1506         len = sprintf(name, LPU64, id);
1507         if (dir_dentry == NULL) {
1508                 dparent = filter_parent_lock(obd, group, id);
1509                 if (IS_ERR(dparent)) {
1510                         CERROR("%s: error getting object "LPU64":"LPU64
1511                                " parent: rc %ld\n", obd->obd_name,
1512                                id, group, PTR_ERR(dparent));
1513                         RETURN(dparent);
1514                 }
1515         }
1516         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1517                dparent->d_name.len, dparent->d_name.name, name);
1518         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1519         if (dir_dentry == NULL)
1520                 filter_parent_unlock(dparent);
1521         if (IS_ERR(dchild)) {
1522                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1523                        PTR_ERR(dchild));
1524                 RETURN(dchild);
1525         }
1526
1527         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1528                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1529                        obd->obd_name, id, dchild->d_inode->i_ino);
1530                 f_dput(dchild);
1531                 RETURN(ERR_PTR(-ENOENT));
1532         }
1533
1534         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1535                name, dchild, atomic_read(&dchild->d_count));
1536
1537         LASSERT(atomic_read(&dchild->d_count) > 0);
1538
1539         RETURN(dchild);
1540 }
1541
1542 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1543                                   obd_id group, struct lustre_handle *lockh)
1544 {
1545         int flags = LDLM_AST_DISCARD_DATA, rc;
1546         struct ldlm_res_id res_id;
1547         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1548         ENTRY;
1549
1550         osc_build_res_name(objid, group, &res_id);
1551         /* Tell the clients that the object is gone now and that they should
1552          * throw away any cached pages. */
1553         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1554                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1555                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1556                                     lockh);
1557         if (rc != ELDLM_OK)
1558                 lockh->cookie = 0;
1559         RETURN(rc);
1560 }
1561
1562 static void filter_fini_destroy(struct obd_device *obd,
1563                                 struct lustre_handle *lockh)
1564 {
1565         if (lockh->cookie)
1566                 ldlm_lock_decref(lockh, LCK_PW);
1567 }
1568
1569 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1570  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1571  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1572  * reverse.  Caller must take i_sem before starting the transaction and we
1573  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1574 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1575                       struct vfsmount *mnt)
1576 {
1577         int rc;
1578         ENTRY;
1579
1580         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1581          * itself we already hold dir->i_mutex for child create/unlink ops */
1582         LASSERT(dentry->d_inode != NULL);
1583         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1584         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1585
1586
1587         /* may_delete() */
1588         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1589                 GOTO(out, rc = -ENOENT);
1590
1591         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1592         if (rc)
1593                 GOTO(out, rc);
1594
1595         if (IS_APPEND(dir))
1596                 GOTO(out, rc = -EPERM);
1597
1598         /* check_sticky() */
1599         if ((dentry->d_inode->i_uid != current->fsuid &&
1600              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1601             IS_IMMUTABLE(dentry->d_inode))
1602                 GOTO(out, rc = -EPERM);
1603
1604         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1605          *       that was done because of journal_start (which is already done
1606          *       here) or some other ordering issue. */
1607         DQUOT_INIT(dir);
1608
1609         rc = ll_security_inode_unlink(dir, dentry, mnt);
1610         if (rc)
1611                 GOTO(out, rc);
1612
1613         rc = dir->i_op->unlink(dir, dentry);
1614 out:
1615         /* need to drop i_mutex before we lose inode reference */
1616         UNLOCK_INODE_MUTEX(dentry->d_inode);
1617         if (rc == 0)
1618                 d_delete(dentry);
1619
1620         RETURN(rc);
1621 }
1622
1623 /* Caller must hold LCK_PW on parent and push us into kernel context.
1624  * Caller must hold child i_mutex, we drop it always.
1625  * Caller is also required to ensure that dchild->d_inode exists. */
1626 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1627                                    obd_gr group, struct dentry *dparent,
1628                                    struct dentry *dchild)
1629 {
1630         struct inode *inode = dchild->d_inode;
1631         struct filter_obd *filter = &obd->u.filter;
1632         int rc;
1633
1634         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1635                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1636                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1637                        (unsigned long)inode->i_nlink,
1638                        atomic_read(&inode->i_count));
1639         }
1640
1641         rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1642         if (rc)
1643                 CERROR("error unlinking objid %.*s: rc %d\n",
1644                        dchild->d_name.len, dchild->d_name.name, rc);
1645         return(rc);
1646 }
1647
1648 struct filter_intent_args {
1649         struct ldlm_lock **victim;
1650         __u64 size;
1651         int *liblustre;
1652 };
1653
1654 static enum interval_iter filter_intent_cb(struct interval_node *n,
1655                                            void *args)
1656 {
1657         struct ldlm_interval *node = (struct ldlm_interval *)n;
1658         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1659         __u64 size = arg->size;
1660         struct ldlm_lock **v = arg->victim;
1661         struct ldlm_lock *lck;
1662
1663         /* If the interval is lower than the current file size,
1664          * just break. */
1665         if (interval_high(n) <= size)
1666                 return INTERVAL_ITER_STOP;
1667
1668         list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1669                 /* Don't send glimpse ASTs to liblustre clients.
1670                  * They aren't listening for them, and they do
1671                  * entirely synchronous I/O anyways. */
1672                 if (lck->l_export == NULL ||
1673                     lck->l_export->exp_libclient == 1)
1674                         continue;
1675
1676                 if (*arg->liblustre)
1677                         *arg->liblustre = 0;
1678
1679                 if (*v == NULL) {
1680                         *v = LDLM_LOCK_GET(lck);
1681                 } else if ((*v)->l_policy_data.l_extent.start <
1682                            lck->l_policy_data.l_extent.start) {
1683                         LDLM_LOCK_RELEASE(*v);
1684                         *v = LDLM_LOCK_GET(lck);
1685                 }
1686
1687                 /* the same policy group - every lock has the
1688                  * same extent, so needn't do it any more */
1689                 break;
1690         }
1691
1692         return INTERVAL_ITER_CONT;
1693 }
1694
1695 static int filter_intent_policy(struct ldlm_namespace *ns,
1696                                 struct ldlm_lock **lockp, void *req_cookie,
1697                                 ldlm_mode_t mode, int flags, void *data)
1698 {
1699         CFS_LIST_HEAD(rpc_list);
1700         struct ptlrpc_request *req = req_cookie;
1701         struct ldlm_lock *lock = *lockp, *l = NULL;
1702         struct ldlm_resource *res = lock->l_resource;
1703         ldlm_processing_policy policy;
1704         struct ost_lvb *res_lvb, *reply_lvb;
1705         struct ldlm_reply *rep;
1706         ldlm_error_t err;
1707         int idx, rc, tmpflags = 0, only_liblustre = 1;
1708         struct ldlm_interval_tree *tree;
1709         struct filter_intent_args arg;
1710         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1711                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1712                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1713         ENTRY;
1714
1715         policy = ldlm_get_processing_policy(res);
1716         LASSERT(policy != NULL);
1717         LASSERT(req != NULL);
1718
1719         rc = lustre_pack_reply(req, 3, repsize, NULL);
1720         if (rc)
1721                 RETURN(req->rq_status = rc);
1722
1723         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1724         LASSERT(rep != NULL);
1725
1726         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1727                                    sizeof(*reply_lvb));
1728         LASSERT(reply_lvb != NULL);
1729
1730         //fixup_handle_for_resent_req(req, lock, &lockh);
1731
1732         /* Call the extent policy function to see if our request can be
1733          * granted, or is blocked.
1734          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1735          * lock, and should not be granted if the lock will be blocked.
1736          */
1737
1738         LASSERT(ns == res->lr_namespace);
1739         lock_res(res);
1740         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1741         check_res_locked(res);
1742
1743         /* FIXME: we should change the policy function slightly, to not make
1744          * this list at all, since we just turn around and free it */
1745         while (!list_empty(&rpc_list)) {
1746                 struct ldlm_lock *wlock =
1747                         list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1748                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1749                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1750                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1751                 list_del_init(&wlock->l_cp_ast);
1752                 LDLM_LOCK_RELEASE(wlock);
1753         }
1754
1755         /* The lock met with no resistance; we're finished. */
1756         if (rc == LDLM_ITER_CONTINUE) {
1757                 /* do not grant locks to the liblustre clients: they cannot
1758                  * handle ASTs robustly.  We need to do this while still
1759                  * holding ns_lock to avoid the lock remaining on the res_link
1760                  * list (and potentially being added to l_pending_list by an
1761                  * AST) when we are going to drop this lock ASAP. */
1762                 if (lock->l_export->exp_libclient ||
1763                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1764                         ldlm_resource_unlink_lock(lock);
1765                         err = ELDLM_LOCK_ABORTED;
1766                 } else {
1767                         err = ELDLM_LOCK_REPLACED;
1768                 }
1769                 unlock_res(res);
1770                 RETURN(err);
1771         }
1772
1773         /* Do not grant any lock, but instead send GL callbacks.  The extent
1774          * policy nicely created a list of all PW locks for us.  We will choose
1775          * the highest of those which are larger than the size in the LVB, if
1776          * any, and perform a glimpse callback. */
1777         res_lvb = res->lr_lvb_data;
1778         LASSERT(res_lvb != NULL);
1779         *reply_lvb = *res_lvb;
1780
1781         /*
1782          * ->ns_lock guarantees that no new locks are granted, and,
1783          * therefore, that res->lr_lvb_data cannot increase beyond the
1784          * end of already granted lock. As a result, it is safe to
1785          * check against "stale" reply_lvb->lvb_size value without
1786          * res->lr_lvb_sem.
1787          */
1788         arg.size = reply_lvb->lvb_size;
1789         arg.victim = &l;
1790         arg.liblustre = &only_liblustre;
1791         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1792                 tree = &res->lr_itree[idx];
1793                 if (tree->lit_mode == LCK_PR)
1794                         continue;
1795
1796                 interval_iterate_reverse(tree->lit_root,
1797                                          filter_intent_cb, &arg);
1798         }
1799         unlock_res(res);
1800
1801         /* There were no PW locks beyond the size in the LVB; finished. */
1802         if (l == NULL) {
1803                 if (only_liblustre) {
1804                         /* If we discovered a liblustre client with a PW lock,
1805                          * however, the LVB may be out of date!  The LVB is
1806                          * updated only on glimpse (which we don't do for
1807                          * liblustre clients) and cancel (which the client
1808                          * obviously has not yet done).  So if it has written
1809                          * data but kept the lock, the LVB is stale and needs
1810                          * to be updated from disk.
1811                          *
1812                          * Of course, this will all disappear when we switch to
1813                          * taking liblustre locks on the OST. */
1814                         ldlm_res_lvbo_update(res, NULL, 1);
1815                 }
1816                 RETURN(ELDLM_LOCK_ABORTED);
1817         }
1818
1819         /*
1820          * This check is for lock taken in filter_prepare_destroy() that does
1821          * not have l_glimpse_ast set. So the logic is: if there is a lock
1822          * with no l_glimpse_ast set, this object is being destroyed already.
1823          *
1824          * Hence, if you are grabbing DLM locks on the server, always set
1825          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1826          */
1827         if (l->l_glimpse_ast == NULL) {
1828                 /* We are racing with unlink(); just return -ENOENT */
1829                 rep->lock_policy_res1 = -ENOENT;
1830                 goto out;
1831         }
1832
1833         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1834         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1835         /* Update the LVB from disk if the AST failed (this is a legal race) */
1836         /*
1837          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1838          * sending ast is not handled. This can result in lost client writes.
1839          */
1840         if (rc != 0)
1841                 ldlm_res_lvbo_update(res, NULL, 1);
1842
1843         lock_res(res);
1844         *reply_lvb = *res_lvb;
1845         unlock_res(res);
1846
1847  out:
1848         LDLM_LOCK_RELEASE(l);
1849
1850         RETURN(ELDLM_LOCK_ABORTED);
1851 }
1852
1853 /*
1854  * per-obd_device iobuf pool.
1855  *
1856  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1857  * allocations in write-path has to be minimized (see bug 5137).
1858  *
1859  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1860  * OST threads (see ost_thread_{init,done}()).
1861  *
1862  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1863  * at the OST layer there are only (potentially) multiple obd_device of type
1864  * unknown at the time of OST thread creation.
1865  *
1866  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1867  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1868  * it's very own iobuf.
1869  *
1870  * Functions below
1871  *
1872  *     filter_kiobuf_pool_init()
1873  *
1874  *     filter_kiobuf_pool_done()
1875  *
1876  *     filter_iobuf_get()
1877  *
1878  * operate on this array. They are "generic" in a sense that they don't depend
1879  * on actual type of iobuf's (the latter depending on Linux kernel version).
1880  */
1881
1882 /*
1883  * destroy pool created by filter_iobuf_pool_init
1884  */
1885 static void filter_iobuf_pool_done(struct filter_obd *filter)
1886 {
1887         struct filter_iobuf **pool;
1888         int i;
1889
1890         ENTRY;
1891
1892         pool = filter->fo_iobuf_pool;
1893         if (pool != NULL) {
1894                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1895                         if (pool[i] != NULL)
1896                                 filter_free_iobuf(pool[i]);
1897                 }
1898                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1899                 filter->fo_iobuf_pool = NULL;
1900         }
1901         EXIT;
1902 }
1903
1904 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1905 {
1906         struct filter_obd       *filter = &obd->u.filter;
1907         struct sptlrpc_rule_set  tmp_rset;
1908         int                      rc;
1909
1910         sptlrpc_rule_set_init(&tmp_rset);
1911         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1912         if (rc) {
1913                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1914                        obd->obd_name, rc);
1915                 return rc;
1916         }
1917
1918         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1919
1920         write_lock(&filter->fo_sptlrpc_lock);
1921         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1922         filter->fo_sptlrpc_rset = tmp_rset;
1923         write_unlock(&filter->fo_sptlrpc_lock);
1924
1925         return 0;
1926 }
1927
1928 /*
1929  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1930  */
1931 static int filter_iobuf_pool_init(struct filter_obd *filter)
1932 {
1933         void **pool;
1934
1935         ENTRY;
1936
1937
1938         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1939                       GFP_KERNEL);
1940         if (filter->fo_iobuf_pool == NULL)
1941                 RETURN(-ENOMEM);
1942
1943         filter->fo_iobuf_count = OSS_THREADS_MAX;
1944
1945         RETURN(0);
1946 }
1947
1948 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1949  * many threads there will be so we allocate a large empty array and only
1950  * fill in those slots that are actually in use.
1951  * If we haven't allocated a pool entry for this thread before, do so now. */
1952 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1953 {
1954         int thread_id                    = (oti && oti->oti_thread) ?
1955                                            oti->oti_thread->t_id : -1;
1956         struct filter_iobuf  *pool       = NULL;
1957         struct filter_iobuf **pool_place = NULL;
1958
1959         if (thread_id >= 0) {
1960                 LASSERT(thread_id < filter->fo_iobuf_count);
1961                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1962         }
1963
1964         if (unlikely(pool == NULL)) {
1965                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1966                                           PTLRPC_MAX_BRW_PAGES);
1967                 if (pool_place != NULL)
1968                         *pool_place = pool;
1969         }
1970
1971         return pool;
1972 }
1973
1974 /* mount the file system (secretly).  lustre_cfg parameters are:
1975  * 1 = device
1976  * 2 = fstype
1977  * 3 = flags: failover=f, failout=n
1978  * 4 = mount options
1979  */
1980 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1981                         void *option)
1982 {
1983         struct filter_obd *filter = &obd->u.filter;
1984         struct vfsmount *mnt;
1985         struct lustre_mount_info *lmi;
1986         struct obd_uuid uuid;
1987         __u8 *uuid_ptr;
1988         char *str, *label;
1989         char ns_name[48];
1990         request_queue_t *q;
1991         int rc, i;
1992         ENTRY;
1993
1994         if (lcfg->lcfg_bufcount < 3 ||
1995             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1996             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1997                 RETURN(-EINVAL);
1998
1999         lmi = server_get_mount(obd->obd_name);
2000         if (lmi) {
2001                 /* We already mounted in lustre_fill_super.
2002                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2003                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2004                 mnt = lmi->lmi_mnt;
2005                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2006         } else {
2007                 /* old path - used by lctl */
2008                 CERROR("Using old MDS mount method\n");
2009                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2010                                     MS_NOATIME|MS_NODIRATIME,
2011                                     lustre_cfg_string(lcfg, 1), option);
2012                 if (IS_ERR(mnt)) {
2013                         rc = PTR_ERR(mnt);
2014                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2015                                            lustre_cfg_string(lcfg, 1), rc);
2016                         RETURN(rc);
2017                 }
2018
2019                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2020         }
2021         if (IS_ERR(obd->obd_fsops))
2022                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2023
2024         rc = filter_iobuf_pool_init(filter);
2025         if (rc != 0)
2026                 GOTO(err_ops, rc);
2027
2028         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2029                 CERROR("%s: Underlying device is marked as read-only. "
2030                        "Setup failed\n", obd->obd_name);
2031                 GOTO(err_ops, rc = -EROFS);
2032         }
2033
2034         /* failover is the default */
2035         obd->obd_replayable = 1;
2036
2037         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2038                 str = lustre_cfg_string(lcfg, 3);
2039                 if (strchr(str, 'n')) {
2040                         CWARN("%s: recovery disabled\n", obd->obd_name);
2041                         obd->obd_replayable = 0;
2042                 }
2043         }
2044
2045         filter->fo_vfsmnt = mnt;
2046         obd->u.obt.obt_sb = mnt->mnt_sb;
2047         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2048         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2049
2050         fsfilt_setup(obd, obd->u.obt.obt_sb);
2051
2052         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2053         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2054         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2055         obd->obd_lvfs_ctxt.fs = get_ds();
2056         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2057
2058         init_mutex(&filter->fo_init_lock);
2059         filter->fo_committed_group = 0;
2060         filter->fo_destroys_in_progress = 0;
2061         for (i = 0; i < 32; i++)
2062                 sema_init(&filter->fo_create_locks[i], 1);
2063
2064         spin_lock_init(&filter->fo_translock);
2065         spin_lock_init(&filter->fo_objidlock);
2066         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2067         sema_init(&filter->fo_alloc_lock, 1);
2068         init_brw_stats(&filter->fo_filter_stats);
2069         filter->fo_read_cache = 1; /* enable read-only cache by default */
2070         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2071         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2072         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2073         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2074
2075         rc = filter_prep(obd);
2076         if (rc)
2077                 GOTO(err_ops, rc);
2078
2079         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2080         spin_lock_init(&filter->fo_llog_list_lock);
2081
2082         filter->fo_fl_oss_capa = 1;
2083
2084         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2085         filter->fo_capa_hash = init_capa_hash();
2086         if (filter->fo_capa_hash == NULL)
2087                 GOTO(err_post, rc = -ENOMEM);
2088
2089         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2090         obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2091                                                 LDLM_NAMESPACE_GREEDY);
2092         if (obd->obd_namespace == NULL)
2093                 GOTO(err_post, rc = -ENOMEM);
2094         obd->obd_namespace->ns_lvbp = obd;
2095         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2096         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2097
2098         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2099                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2100
2101         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2102         if (rc) {
2103                 CERROR("failed to setup llogging subsystems\n");
2104                 GOTO(err_post, rc);
2105         }
2106
2107         rwlock_init(&filter->fo_sptlrpc_lock);
2108         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2109         /* do this after llog being initialized */
2110         filter_adapt_sptlrpc_conf(obd, 1);
2111
2112         rc = lquota_setup(filter_quota_interface_ref, obd);
2113         if (rc)
2114                 GOTO(err_post, rc);
2115
2116         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2117         if (q->max_sectors < q->max_hw_sectors &&
2118             q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2119                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2120                               "for larger I/O requests: max_sectors = %u "
2121                               "could be up to max_hw_sectors=%u\n",
2122                               obd->obd_name, mnt->mnt_sb->s_id,
2123                               q->max_sectors, q->max_hw_sectors);
2124
2125         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2126         if (uuid_ptr != NULL) {
2127                 class_uuid_unparse(uuid_ptr, &uuid);
2128                 str = uuid.uuid;
2129         } else {
2130                 str = "no UUID";
2131         }
2132
2133         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2134
2135         if (obd->obd_recovering) {
2136                 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2137                               "recovery for at least %d:%.02d, or until %d "
2138                               "client%s reconnect%s.\n",
2139                               obd->obd_name, lustre_cfg_string(lcfg, 1),
2140                               label ?: "", label ? "/" : "", str,
2141                               obd->obd_recovery_timeout / 60,
2142                               obd->obd_recovery_timeout % 60,
2143                               obd->obd_max_recoverable_clients,
2144                               (obd->obd_max_recoverable_clients == 1) ? "":"s",
2145                               (obd->obd_max_recoverable_clients == 1) ? "s":"");
2146         } else {
2147                 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2148                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2149                               label ?: "", label ? "/" : "", str,
2150                               obd->obd_replayable ? "enabled" : "disabled");
2151         }
2152
2153         RETURN(0);
2154
2155 err_post:
2156         filter_post(obd);
2157 err_ops:
2158         fsfilt_put_ops(obd->obd_fsops);
2159         filter_iobuf_pool_done(filter);
2160 err_mntput:
2161         server_put_mount(obd->obd_name, mnt);
2162         obd->u.obt.obt_sb = 0;
2163         return rc;
2164 }
2165
2166 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2167 {
2168         struct lprocfs_static_vars lvars;
2169         unsigned long addr;
2170         struct page *page;
2171         int rc;
2172
2173         CLASSERT(offsetof(struct obd_device, u.obt) ==
2174                  offsetof(struct obd_device, u.filter.fo_obt));
2175
2176         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2177                 RETURN(-EINVAL);
2178
2179         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2180         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2181         if (!page)
2182                 RETURN(-ENOMEM);
2183         addr = (unsigned long)cfs_page_address(page);
2184         clear_page((void *)addr);
2185
2186         /* lprocfs must be setup before the filter so state can be safely added
2187          * to /proc incrementally as the filter is setup */
2188         lprocfs_filter_init_vars(&lvars);
2189         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2190             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2191                 /* Init obdfilter private stats here */
2192                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2193                                      LPROCFS_CNTR_AVGMINMAX,
2194                                      "read_bytes", "bytes");
2195                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2196                                      LPROCFS_CNTR_AVGMINMAX,
2197                                      "write_bytes", "bytes");
2198                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2199                                      LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2200                                      "get_page", "usec");
2201                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2202                                      LPROCFS_CNTR_AVGMINMAX,
2203                                      "get_page_failures", "num");
2204                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2205                                      LPROCFS_CNTR_AVGMINMAX,
2206                                      "cache_access", "pages");
2207                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2208                                      LPROCFS_CNTR_AVGMINMAX,
2209                                      "cache_hit", "pages");
2210                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2211                                      LPROCFS_CNTR_AVGMINMAX,
2212                                      "cache_miss", "pages");
2213
2214                 lproc_filter_attach_seqstat(obd);
2215                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2216                                                         obd->obd_proc_entry,
2217                                                         NULL, NULL);
2218                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2219                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2220                         CERROR("error %d setting up lprocfs for %s\n",
2221                                rc, "exports");
2222                         obd->obd_proc_exports_entry = NULL;
2223                 }
2224         }
2225         if (obd->obd_proc_exports_entry)
2226                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2227                                    lprocfs_nid_stats_clear_read,
2228                                    lprocfs_nid_stats_clear_write, obd, NULL);
2229
2230         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2231                LUSTRE_CFG_BUFLEN(lcfg, 4));
2232         rc = filter_common_setup(obd, lcfg, (void *)addr);
2233         OBD_PAGE_FREE(page);
2234
2235         if (rc) {
2236                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2237                 lprocfs_free_per_client_stats(obd);
2238                 lprocfs_free_obd_stats(obd);
2239                 lprocfs_obd_cleanup(obd);
2240         }
2241
2242         return rc;
2243 }
2244
2245 static struct llog_operations filter_mds_ost_repl_logops;
2246
2247 static struct llog_operations filter_size_orig_logops = {
2248         .lop_setup   = llog_obd_origin_setup,
2249         .lop_cleanup = llog_obd_origin_cleanup,
2250         .lop_add     = llog_obd_origin_add
2251 };
2252
2253 static int filter_olg_fini(struct obd_llog_group *olg)
2254 {
2255         struct llog_ctxt *ctxt;
2256         int rc = 0, rc2 = 0;
2257         ENTRY;
2258
2259         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2260         if (ctxt)
2261                 rc = llog_cleanup(ctxt);
2262
2263         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2264         if (ctxt) {
2265                 rc2 = llog_cleanup(ctxt);
2266                 if (!rc)
2267                         rc = rc2;
2268         }
2269
2270         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2271         if (ctxt) {
2272                 rc2 = llog_cleanup(ctxt);
2273                 if (!rc)
2274                         rc = rc2;
2275         }
2276
2277         RETURN(rc);
2278 }
2279
2280 static int
2281 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2282                 struct obd_device *tgt)
2283 {
2284         int rc;
2285         ENTRY;
2286
2287         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2288                         &filter_mds_ost_repl_logops);
2289         if (rc)
2290                 GOTO(cleanup, rc);
2291
2292         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2293                         &filter_size_orig_logops);
2294         if (rc)
2295                 GOTO(cleanup, rc);
2296         EXIT;
2297 cleanup:
2298         if (rc)
2299                 filter_olg_fini(olg);
2300         return rc;
2301 }
2302
2303 /**
2304  * Init the default olg, which is embeded in the obd_device, for filter.
2305  */
2306 static int
2307 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2308                         struct obd_device *tgt)
2309 {
2310         struct filter_obd *filter = &obd->u.filter;
2311         struct llog_ctxt *ctxt;
2312         int rc;
2313         ENTRY;
2314
2315         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2316         if (!filter->fo_lcm)
2317                 RETURN(-ENOMEM);
2318
2319         filter_mds_ost_repl_logops = llog_client_ops;
2320         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2321         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2322         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2323
2324         rc = filter_olg_init(obd, olg, tgt);
2325         if (rc)
2326                 GOTO(cleanup_lcm, rc);
2327
2328         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2329                         &llog_lvfs_ops);
2330         if (rc)
2331                 GOTO(cleanup_olg, rc);
2332
2333         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2334         if (!ctxt) {
2335                 CERROR("Can't get ctxt for %p:%x\n", olg,
2336                        LLOG_MDS_OST_REPL_CTXT);
2337                 GOTO(cleanup_olg, rc = -ENODEV);
2338         }
2339         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2340         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2341         llog_ctxt_put(ctxt);
2342
2343         RETURN(0);
2344 cleanup_olg:
2345         filter_olg_fini(olg);
2346 cleanup_lcm:
2347         llog_recov_thread_fini(filter->fo_lcm, 1);
2348         filter->fo_lcm = NULL;
2349         return rc;
2350 }
2351
2352 static int
2353 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2354                  struct obd_device *tgt, int *index)
2355 {
2356         struct filter_obd *filter = &obd->u.filter;
2357         struct llog_ctxt *ctxt;
2358         int rc;
2359         ENTRY;
2360
2361         LASSERT(olg != NULL);
2362         if (olg == &obd->obd_olg)
2363                 return filter_default_olg_init(obd, olg, tgt);
2364
2365         LASSERT(filter->fo_lcm != NULL);
2366         rc = filter_olg_init(obd, olg, tgt);
2367         if (rc)
2368                 RETURN(rc);
2369         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2370         if (!ctxt) {
2371                 CERROR("Can't get ctxt for %p:%x\n", olg,
2372                        LLOG_MDS_OST_REPL_CTXT);
2373                 filter_olg_fini(olg);
2374                 RETURN(-ENODEV);
2375         }
2376         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2377         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2378         llog_ctxt_put(ctxt);
2379         RETURN(rc);
2380 }
2381
2382 static int filter_llog_finish(struct obd_device *obd, int count)
2383 {
2384         struct filter_obd *filter = &obd->u.filter;
2385         struct llog_ctxt *ctxt;
2386         ENTRY;
2387
2388         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2389         if (ctxt) {
2390                 /*
2391                  * Make sure that no cached llcds left in recov_thread.
2392                  * We actually do sync in disconnect time, but disconnect
2393                  * may not come being marked rq_no_resend = 1.
2394                  */
2395                 llog_sync(ctxt, NULL);
2396
2397                 /*
2398                  * Balance class_import_get() in llog_receptor_accept().
2399                  * This is safe to do, as llog is already synchronized
2400                  * and its import may go.
2401                  */
2402                 mutex_down(&ctxt->loc_sem);
2403                 if (ctxt->loc_imp) {
2404                         class_import_put(ctxt->loc_imp);
2405                         ctxt->loc_imp = NULL;
2406                 }
2407                 mutex_up(&ctxt->loc_sem);
2408                 llog_ctxt_put(ctxt);
2409         }
2410
2411         if (filter->fo_lcm) {
2412                 mutex_down(&ctxt->loc_sem);
2413                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2414                 filter->fo_lcm = NULL;
2415                 mutex_up(&ctxt->loc_sem);
2416         }
2417         RETURN(filter_olg_fini(&obd->obd_olg));
2418 }
2419
2420 /**
2421  * Find the group llog according to group index in the llog group list.
2422  */
2423 static struct obd_llog_group *
2424 filter_find_olg_internal(struct filter_obd *filter, int group)
2425 {
2426         struct obd_llog_group *olg;
2427
2428         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2429         list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2430                 if (olg->olg_group == group)
2431                         RETURN(olg);
2432         }
2433         RETURN(NULL);
2434 }
2435
2436 /**
2437  * Find the group llog according to group index on the filter
2438  */
2439 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2440 {
2441         struct obd_llog_group *olg = NULL;
2442         struct filter_obd *filter;
2443
2444         filter = &obd->u.filter;
2445
2446         if (group == FILTER_GROUP_LLOG)
2447                 RETURN(&obd->obd_olg);
2448
2449         spin_lock(&filter->fo_llog_list_lock);
2450         olg = filter_find_olg_internal(filter, group);
2451         spin_unlock(&filter->fo_llog_list_lock);
2452
2453         RETURN(olg);
2454 }
2455 /**
2456  * Find the llog_group of the filter according to the group. If it can not
2457  * find, create the llog_group, which only happens when mds is being synced
2458  * with OST.
2459  */
2460 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2461 {
2462         struct obd_llog_group *olg = NULL;
2463         struct filter_obd *filter;
2464         int rc;
2465
2466         filter = &obd->u.filter;
2467
2468         if (group == FILTER_GROUP_LLOG)
2469                 RETURN(&obd->obd_olg);
2470
2471         spin_lock(&filter->fo_llog_list_lock);
2472         olg = filter_find_olg_internal(filter, group);
2473         if (olg) {
2474                 if (olg->olg_initializing) {
2475                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2476                 } else {
2477                         GOTO(out_unlock, olg);
2478                 }
2479         }
2480         OBD_ALLOC_PTR(olg);
2481         if (olg == NULL)
2482                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2483
2484         llog_group_init(olg, group);
2485         list_add(&olg->olg_list, &filter->fo_llog_list);
2486         olg->olg_initializing = 1;
2487         spin_unlock(&filter->fo_llog_list_lock);
2488
2489         rc = obd_llog_init(obd, olg, obd, NULL);
2490         if (rc) {
2491                spin_lock(&filter->fo_llog_list_lock);
2492                list_del(&olg->olg_list);
2493                spin_unlock(&filter->fo_llog_list_lock);
2494                OBD_FREE_PTR(olg);
2495                GOTO(out, olg = ERR_PTR(-ENOMEM));
2496         }
2497         spin_lock(&filter->fo_llog_list_lock);
2498         olg->olg_initializing = 0;
2499         spin_unlock(&filter->fo_llog_list_lock);
2500         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2501               obd->obd_name, group, olg);
2502 out:
2503         RETURN(olg);
2504
2505 out_unlock:
2506         spin_unlock(&filter->fo_llog_list_lock);
2507         GOTO(out, olg);
2508 }
2509
2510 static int filter_llog_connect(struct obd_export *exp,
2511                                struct llogd_conn_body *body)
2512 {
2513         struct obd_device *obd = exp->exp_obd;
2514         struct llog_ctxt *ctxt;
2515         struct obd_llog_group *olg;
2516         int rc;
2517         ENTRY;
2518
2519         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2520                obd->obd_name, body->lgdc_logid.lgl_oid,
2521                body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2522
2523         olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2524         if (!olg) {
2525                 CERROR(" %s: can not find olg of group %d\n",
2526                        obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2527                 RETURN(-ENOENT);
2528         }
2529         llog_group_set_export(olg, exp);
2530
2531         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2532         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2533                  body->lgdc_ctxt_idx);
2534
2535         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2536               obd->obd_name, body->lgdc_logid.lgl_oid,
2537               body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2538
2539         spin_lock_bh(&obd->obd_processing_task_lock);
2540         obd->u.filter.fo_mds_ost_sync = 1;
2541         spin_unlock_bh(&obd->obd_processing_task_lock);
2542         rc = llog_connect(ctxt, &body->lgdc_logid,
2543                           &body->lgdc_gen, NULL);
2544         llog_ctxt_put(ctxt);
2545         if (rc != 0)
2546                 CERROR("failed to connect rc %d idx %d\n", rc,
2547                                 body->lgdc_ctxt_idx);
2548
2549         RETURN(rc);
2550 }
2551
2552 static int filter_llog_preclean(struct obd_device *obd)
2553 {
2554         struct obd_llog_group *olg, *tmp;
2555         struct filter_obd *filter;
2556         struct list_head  remove_list;
2557         int rc = 0;
2558         ENTRY;
2559
2560         rc = obd_llog_finish(obd, 0);
2561         if (rc)
2562                 CERROR("failed to cleanup llogging subsystem\n");
2563
2564         filter = &obd->u.filter;
2565         CFS_INIT_LIST_HEAD(&remove_list);
2566
2567         spin_lock(&filter->fo_llog_list_lock);
2568         while (!list_empty(&filter->fo_llog_list)) {
2569                 olg = list_entry(filter->fo_llog_list.next,
2570                                  struct obd_llog_group, olg_list);
2571                 list_del(&olg->olg_list);
2572                 list_add(&olg->olg_list, &remove_list);
2573         }
2574         spin_unlock(&filter->fo_llog_list_lock);
2575
2576         list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2577                 list_del_init(&olg->olg_list);
2578                 rc = filter_olg_fini(olg);
2579                 if (rc)
2580                         CERROR("failed to cleanup llogging subsystem for %u\n",
2581                                olg->olg_group);
2582                 OBD_FREE_PTR(olg);
2583         }
2584
2585         RETURN(rc);
2586 }
2587
2588 static int filter_precleanup(struct obd_device *obd,
2589                              enum obd_cleanup_stage stage)
2590 {
2591         int rc = 0;
2592         ENTRY;
2593
2594         switch(stage) {
2595         case OBD_CLEANUP_EARLY:
2596                 break;
2597         case OBD_CLEANUP_EXPORTS:
2598                 /* Stop recovery before namespace cleanup. */
2599                 target_recovery_fini(obd);
2600                 rc = filter_llog_preclean(obd);
2601                 break;
2602         }
2603         RETURN(rc);
2604 }
2605
2606 static int filter_cleanup(struct obd_device *obd)
2607 {
2608         struct filter_obd *filter = &obd->u.filter;
2609         ENTRY;
2610
2611         if (obd->obd_fail)
2612                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2613                               "will be preserved.\n", obd->obd_name);
2614
2615         obd_exports_barrier(obd);
2616         obd_zombie_barrier();
2617
2618         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2619         lprocfs_free_per_client_stats(obd);
2620         lprocfs_free_obd_stats(obd);
2621         lprocfs_obd_cleanup(obd);
2622         lquota_cleanup(filter_quota_interface_ref, obd);
2623
2624         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2625         obd->obd_namespace = NULL;
2626
2627         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2628
2629         if (obd->u.obt.obt_sb == NULL)
2630                 RETURN(0);
2631
2632         filter_post(obd);
2633
2634         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2635         shrink_dcache_sb(obd->u.obt.obt_sb);
2636
2637         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2638         obd->u.obt.obt_sb = NULL;
2639
2640         fsfilt_put_ops(obd->obd_fsops);
2641
2642         filter_iobuf_pool_done(filter);
2643
2644         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2645
2646         RETURN(0);
2647 }
2648
2649 static int filter_connect_internal(struct obd_export *exp,
2650                                    struct obd_connect_data *data,
2651                                    int reconnect)
2652 {
2653         struct filter_export_data *fed = &exp->exp_filter_data;
2654
2655         if (!data)
2656                 RETURN(0);
2657
2658         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2659                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2660                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2661                data->ocd_connect_flags, data->ocd_version,
2662                data->ocd_grant, data->ocd_index);
2663
2664         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2665                 CWARN("!!! This export (nid %s) used object group %d "
2666                        "earlier; now it's trying to use group %d!  This could "
2667                        "be a bug in the MDS. Please report to "
2668                        "http://bugzilla.lustre.org/\n",
2669                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2670                 RETURN(-EPROTO);
2671         }
2672         fed->fed_group = data->ocd_group;
2673
2674         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2675         exp->exp_connect_flags = data->ocd_connect_flags;
2676         data->ocd_version = LUSTRE_VERSION_CODE;
2677
2678         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2679         if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2680                   data->ocd_connect_flags & OBD_CONNECT_MDS))
2681                 RETURN(-EPROTO);
2682
2683         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2684                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2685                 obd_size left, want;
2686
2687                 spin_lock(&exp->exp_obd->obd_osfs_lock);
2688                 left = filter_grant_space_left(exp);
2689                 want = data->ocd_grant;
2690                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2691                 data->ocd_grant = fed->fed_grant;
2692                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2693
2694                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2695                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2696                        exp->exp_client_uuid.uuid, exp,
2697                        data->ocd_grant, want, left);
2698
2699                 filter->fo_tot_granted_clients ++;
2700         }
2701
2702         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2703                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2704                 struct lr_server_data *lsd = filter->fo_fsd;
2705                 int index = le32_to_cpu(lsd->lsd_ost_index);
2706
2707                 if (!(lsd->lsd_feature_compat &
2708                       cpu_to_le32(OBD_COMPAT_OST))) {
2709                         /* this will only happen on the first connect */
2710                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2711                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2712                         /* sync is not needed here as filter_client_add will
2713                          * set exp_need_sync flag */
2714                         filter_update_server_data(exp->exp_obd,
2715                                                   filter->fo_rcvd_filp, lsd);
2716                 } else if (index != data->ocd_index) {
2717                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2718                                            " %u doesn't match actual OST index"
2719                                            " %u in last_rcvd file, bad "
2720                                            "configuration?\n",
2721                                            obd_export_nid2str(exp), index,
2722                                            data->ocd_index);
2723                         RETURN(-EBADF);
2724                 }
2725                 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2726                  * FIXME: We don't strictly need the COMPAT flag for that,
2727                  * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2728                  * FIXME: We needed it for the index, as index 0 is valid. */
2729         }
2730
2731         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2732                 data->ocd_brw_size = 65536;
2733         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2734                 data->ocd_brw_size = min(data->ocd_brw_size,
2735                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2736                 LASSERT(data->ocd_brw_size);
2737         }
2738
2739         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2740                 __u32 cksum_types = data->ocd_cksum_types;
2741
2742                 /* The client set in ocd_cksum_types the checksum types it
2743                  * supports. We have to mask off the algorithms that we don't
2744                  * support */
2745                 if (cksum_types & OBD_CKSUM_ALL)
2746                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2747                 else
2748                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2749
2750                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2751                                    "%x\n", exp->exp_obd->obd_name,
2752                                    obd_export_nid2str(exp), cksum_types,
2753                                    data->ocd_cksum_types);
2754         } else {
2755                 /* This client does not support OBD_CONNECT_CKSUM
2756                  * fall back to CRC32 */
2757                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2758                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2759                                    exp->exp_obd->obd_name,
2760                                    obd_export_nid2str(exp));
2761         }
2762
2763         RETURN(0);
2764 }
2765
2766 static int filter_reconnect(const struct lu_env *env,
2767                             struct obd_export *exp, struct obd_device *obd,
2768                             struct obd_uuid *cluuid,
2769                             struct obd_connect_data *data,
2770                             void *localdata)
2771 {
2772         int rc;
2773         ENTRY;
2774
2775         if (exp == NULL || obd == NULL || cluuid == NULL)
2776                 RETURN(-EINVAL);
2777
2778         rc = filter_connect_internal(exp, data, 1);
2779         if (rc == 0)
2780                 filter_export_stats_init(obd, exp, localdata);
2781
2782         RETURN(rc);
2783 }
2784
2785 /* nearly identical to mds_connect */
2786 static int filter_connect(const struct lu_env *env,
2787                           struct obd_export **exp, struct obd_device *obd,
2788                           struct obd_uuid *cluuid,
2789                           struct obd_connect_data *data, void *localdata)
2790 {
2791         struct lvfs_run_ctxt saved;
2792         struct lustre_handle conn = { 0 };
2793         struct obd_export *lexp;
2794         struct filter_export_data *fed;
2795         struct lsd_client_data *lcd = NULL;
2796         __u32 group;
2797         int rc;
2798         ENTRY;
2799
2800         if (exp == NULL || obd == NULL || cluuid == NULL)
2801                 RETURN(-EINVAL);
2802
2803         rc = class_connect(&conn, obd, cluuid);
2804         if (rc)
2805                 RETURN(rc);
2806         lexp = class_conn2export(&conn);
2807         LASSERT(lexp != NULL);
2808
2809         fed = &lexp->exp_filter_data;
2810
2811         rc = filter_connect_internal(lexp, data, 0);
2812         if (rc)
2813                 GOTO(cleanup, rc);
2814
2815         filter_export_stats_init(obd, lexp, localdata);
2816         if (obd->obd_replayable) {
2817                 OBD_ALLOC(lcd, sizeof(*lcd));
2818                 if (!lcd) {
2819                         CERROR("filter: out of memory for client data\n");
2820                         GOTO(cleanup, rc = -ENOMEM);
2821                 }
2822
2823                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2824                 fed->fed_lcd = lcd;
2825                 rc = filter_client_add(obd, lexp, -1);
2826                 if (rc)
2827                         GOTO(cleanup, rc);
2828         }
2829
2830         group = data->ocd_group;
2831
2832         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2833               obd->obd_name, lexp->exp_handle.h_cookie, group);
2834
2835         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2836         rc = filter_read_groups(obd, group, 1);
2837         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2838         if (rc != 0) {
2839                 CERROR("can't read group %u\n", group);
2840                 GOTO(cleanup, rc);
2841         }
2842
2843         GOTO(cleanup, rc);
2844
2845 cleanup:
2846         if (rc) {
2847                 if (lcd) {
2848                         OBD_FREE_PTR(lcd);
2849                         fed->fed_lcd = NULL;
2850                 }
2851                 class_disconnect(lexp);
2852                 *exp = NULL;
2853         } else {
2854                 *exp = lexp;
2855         }
2856
2857         RETURN(rc);
2858 }
2859
2860 /* Do extra sanity checks for grant accounting.  We do this at connect,
2861  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2862  * always get rid of it or turn it off when we know accounting is good. */
2863 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2864 {
2865         struct filter_export_data *fed;
2866         struct obd_export *exp;
2867         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2868         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2869         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2870
2871         if (list_empty(&obd->obd_exports))
2872                 return;
2873
2874         /* We don't want to do this for large machines that do lots of
2875            mounts or unmounts.  It burns... */
2876         if (obd->obd_num_exports > 100)
2877                 return;
2878
2879         spin_lock(&obd->obd_osfs_lock);
2880         spin_lock(&obd->obd_dev_lock);
2881         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2882                 int error = 0;
2883                 fed = &exp->exp_filter_data;
2884                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2885                     fed->fed_dirty < 0)
2886                         error = 1;
2887                 if (maxsize > 0) { /* we may not have done a statfs yet */
2888                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2889                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2890                                  exp->exp_client_uuid.uuid, exp,
2891                                  fed->fed_grant, fed->fed_pending, maxsize);
2892                         LASSERTF(fed->fed_dirty <= maxsize,
2893                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2894                                  exp->exp_client_uuid.uuid, exp,
2895                                  fed->fed_dirty, maxsize);
2896                 }
2897                 if (error)
2898                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2899                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2900                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2901                 else
2902                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2903                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2904                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2905                 tot_granted += fed->fed_grant + fed->fed_pending;
2906                 tot_pending += fed->fed_pending;
2907                 tot_dirty += fed->fed_dirty;
2908         }
2909         fo_tot_granted = obd->u.filter.fo_tot_granted;
2910         fo_tot_pending = obd->u.filter.fo_tot_pending;
2911         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2912         spin_unlock(&obd->obd_dev_lock);
2913         spin_unlock(&obd->obd_osfs_lock);
2914
2915         /* Do these assertions outside the spinlocks so we don't kill system */
2916         if (tot_granted != fo_tot_granted)
2917                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2918                        func, tot_granted, fo_tot_granted);
2919         if (tot_pending != fo_tot_pending)
2920                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2921                        func, tot_pending, fo_tot_pending);
2922         if (tot_dirty != fo_tot_dirty)
2923                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2924                        func, tot_dirty, fo_tot_dirty);
2925         if (tot_pending > tot_granted)
2926                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2927                        func, tot_pending, tot_granted);
2928         if (tot_granted > maxsize)
2929                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2930                        func, tot_granted, maxsize);
2931         if (tot_dirty > maxsize)
2932                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2933                        func, tot_dirty, maxsize);
2934 }
2935
2936 /* Remove this client from the grant accounting totals.  We also remove
2937  * the export from the obd device under the osfs and dev locks to ensure
2938  * that the filter_grant_sanity_check() calculations are always valid.
2939  * The client should do something similar when it invalidates its import. */
2940 static void filter_grant_discard(struct obd_export *exp)
2941 {
2942         struct obd_device *obd = exp->exp_obd;
2943         struct filter_obd *filter = &obd->u.filter;
2944         struct filter_export_data *fed = &exp->exp_filter_data;
2945
2946         spin_lock(&obd->obd_osfs_lock);
2947         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2948                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2949                  obd->obd_name, filter->fo_tot_granted,
2950                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2951         filter->fo_tot_granted -= fed->fed_grant;
2952         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2953                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2954                  obd->obd_name, filter->fo_tot_pending,
2955                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2956         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2957         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2958                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2959                  obd->obd_name, filter->fo_tot_dirty,
2960                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2961         filter->fo_tot_dirty -= fed->fed_dirty;
2962         fed->fed_dirty = 0;
2963         fed->fed_grant = 0;
2964
2965         spin_unlock(&obd->obd_osfs_lock);
2966 }
2967
2968 static int filter_destroy_export(struct obd_export *exp)
2969 {
2970         ENTRY;
2971
2972         if (exp->exp_filter_data.fed_pending)
2973                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2974                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2975                        exp, exp->exp_filter_data.fed_pending);
2976
2977         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2978
2979         target_destroy_export(exp);
2980         ldlm_destroy_export(exp);
2981
2982         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2983                 RETURN(0);
2984
2985
2986         if (!exp->exp_obd->obd_replayable)
2987                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2988
2989         filter_grant_discard(exp);
2990         filter_fmd_cleanup(exp);
2991
2992         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2993                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2994                 if (filter->fo_tot_granted_clients > 0)
2995                         filter->fo_tot_granted_clients --;
2996         }
2997
2998         if (!(exp->exp_flags & OBD_OPT_FORCE))
2999                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3000
3001         RETURN(0);
3002 }
3003
3004 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3005 {
3006         struct obd_llog_group *olg_min, *olg;
3007         struct filter_obd *filter;
3008         int worked = -1, group;
3009         struct llog_ctxt *ctxt;
3010         ENTRY;
3011
3012         filter = &obd->u.filter;
3013
3014         /* we can't sync log holding spinlock. also, we do not want to get
3015          * into livelock. so we do following: loop over MDS's exports in
3016          * group order and skip already synced llogs -bzzz */
3017         do {
3018                 /* look for group with min. number, but > worked */
3019                 olg_min = NULL;
3020                 group = 1 << 30;
3021                 spin_lock(&filter->fo_llog_list_lock);
3022                 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3023                         if (olg->olg_group <= worked) {
3024                                 /* this group is already synced */
3025                                 continue;
3026                         }
3027                         if (group < olg->olg_group) {
3028                                 /* we have group with smaller number to sync */
3029                                 continue;
3030                         }
3031                         /* store current minimal group */
3032                         olg_min = olg;
3033                         group = olg->olg_group;
3034                 }
3035                 spin_unlock(&filter->fo_llog_list_lock);
3036
3037                 if (olg_min == NULL)
3038                         break;
3039
3040                 worked = olg_min->olg_group;
3041                 if (olg_min->olg_exp &&
3042                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3043                         int err;
3044                         ctxt = llog_group_get_ctxt(olg_min,
3045                                                    LLOG_MDS_OST_REPL_CTXT);
3046                         if (ctxt) {
3047                                 err = llog_sync(ctxt, olg_min->olg_exp);
3048                                 llog_ctxt_put(ctxt);
3049                                 if (err) {
3050                                         CERROR("error flushing logs to MDS: "
3051                                                "rc %d\n", err);
3052                                 }
3053                         }
3054                 }
3055         } while (olg_min != NULL);
3056 }
3057
3058 /* Also incredibly similar to mds_disconnect */
3059 static int filter_disconnect(struct obd_export *exp)
3060 {
3061         struct obd_device *obd = exp->exp_obd;
3062         int rc;
3063         ENTRY;
3064
3065         LASSERT(exp);
3066         class_export_get(exp);
3067
3068         if (!(exp->exp_flags & OBD_OPT_FORCE))
3069                 filter_grant_sanity_check(obd, __FUNCTION__);
3070         filter_grant_discard(exp);
3071
3072         /* Flush any remaining cancel messages out to the target */
3073         filter_sync_llogs(obd, exp);
3074
3075         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3076
3077         rc = server_disconnect_export(exp);
3078
3079         if (exp->exp_obd->obd_replayable)
3080                 filter_client_free(exp);
3081         else
3082                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3083
3084         class_export_put(exp);
3085         RETURN(rc);
3086 }
3087
3088 /* reverse import is changed, sync all cancels */
3089 static void filter_revimp_update(struct obd_export *exp)
3090 {
3091         ENTRY;
3092
3093         LASSERT(exp);
3094         class_export_get(exp);
3095
3096         /* flush any remaining cancel messages out to the target */
3097         filter_sync_llogs(exp->exp_obd, exp);
3098         class_export_put(exp);
3099         EXIT;
3100 }
3101
3102 static int filter_ping(struct obd_export *exp)
3103 {
3104         filter_fmd_expire(exp);
3105         return 0;
3106 }
3107
3108 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3109                                   const char *what, int quiet)
3110 {
3111         struct dentry *dchild = NULL;
3112         obd_gr group = 0;
3113
3114         if (oa->o_valid & OBD_MD_FLGROUP)
3115                 group = oa->o_gr;
3116
3117         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3118
3119         if (IS_ERR(dchild)) {
3120                 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3121                        what, group, oa->o_id);
3122                 RETURN(dchild);
3123         }
3124
3125         if (dchild->d_inode == NULL) {
3126                 if (!quiet)
3127                         CERROR("%s: %s on non-existent object: "LPU64"\n",
3128                                obd->obd_name, what, oa->o_id);
3129                 f_dput(dchild);
3130                 RETURN(ERR_PTR(-ENOENT));
3131         }
3132
3133         return dchild;
3134 }
3135
3136 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3137 {
3138         struct dentry *dentry = NULL;
3139         struct obd_device *obd;
3140         int rc = 0;
3141         ENTRY;
3142
3143         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3144         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3145                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3146         if (rc)
3147                 RETURN(rc);
3148
3149         obd = class_exp2obd(exp);
3150         if (obd == NULL) {
3151                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3152                 RETURN(-EINVAL);
3153         }
3154
3155         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3156         if (IS_ERR(dentry))
3157                 RETURN(PTR_ERR(dentry));
3158
3159         /* Limit the valid bits in the return data to what we actually use */
3160         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3161         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3162
3163         f_dput(dentry);
3164         RETURN(rc);
3165 }
3166
3167 /* this should be enabled/disabled in condition to enabled/disabled large
3168  * inodes (fast EAs) in backing store FS. */
3169 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3170                         void *handle, struct obdo *oa)
3171 {
3172         struct obd_device *obd = exp->exp_obd;
3173         int rc = 0;
3174         ENTRY;
3175
3176         if (oa->o_valid & OBD_MD_FLFID) {
3177                 struct filter_fid ff;
3178
3179                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3180                         oa->o_gr = 0;
3181                 /* packing fid and converting it to LE for storing into EA.
3182                  * Here ->o_stripe_idx should be filled by LOV and rest of
3183                  * fields - by client. */
3184                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3185                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3186                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3187                 ff.ff_objid = cpu_to_le64(oa->o_id);
3188                 ff.ff_group = cpu_to_le64(oa->o_gr);
3189
3190                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3191                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3192                        oa->o_generation, oa->o_id, oa->o_gr);
3193
3194                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3195                 if (rc)
3196                         CERROR("store fid in object failed! rc: %d\n", rc);