Whamcloud - gitweb
jbd2 spinlock hang fix for rhel5 and added error check
[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                 cfs_spin_lock(&exp->exp_lock);
111                 exp->exp_vbr_failed = 1;
112                 cfs_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         cfs_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                 cfs_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         cfs_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         cfs_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         cfs_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                 cfs_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 = cfs_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 (cfs_test_and_set_bit(cl_idx, bitmap)) {
337                         cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
338                                                         cl_idx);
339                         goto repeat;
340                 }
341         } else {
342                 if (cfs_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         cfs_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                                 cfs_spin_lock(&exp->exp_lock);
384                                 exp->exp_need_sync = 1;
385                                 cfs_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 (!cfs_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         cfs_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         cfs_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 (!cfs_test_and_clear_bit(fed->fed_lr_idx,
468                                     filter->fo_last_rcvd_slots)) {
469                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
470                        fed->fed_lr_idx);
471                 LBUG();
472         }
473         OBD_FREE_PTR(lcd);
474         RETURN(0);
475 free:
476         cfs_mutex_down(&fed->fed_lcd_lock);
477         fed->fed_lcd = NULL;
478         cfs_mutex_up(&fed->fed_lcd_lock);
479         OBD_FREE_PTR(lcd);
480
481         return 0;
482 }
483
484 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
485 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
486                                          struct filter_mod_data *fmd)
487 {
488         LASSERT_SPIN_LOCKED(&fed->fed_lock);
489         if (--fmd->fmd_refcount == 0) {
490                 /* XXX when we have persistent reservations and the handle
491                  * is stored herein we need to drop it here. */
492                 fed->fed_mod_count--;
493                 cfs_list_del(&fmd->fmd_list);
494                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
495         }
496 }
497
498 /* drop fmd reference, free it if last ref */
499 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
500 {
501         struct filter_export_data *fed;
502
503         if (fmd == NULL)
504                 return;
505
506         fed = &exp->exp_filter_data;
507         cfs_spin_lock(&fed->fed_lock);
508         filter_fmd_put_nolock(fed, fmd); /* caller reference */
509         cfs_spin_unlock(&fed->fed_lock);
510 }
511
512 /* expire entries from the end of the list if there are too many
513  * or they are too old */
514 static void filter_fmd_expire_nolock(struct filter_obd *filter,
515                                      struct filter_export_data *fed,
516                                      struct filter_mod_data *keep)
517 {
518         struct filter_mod_data *fmd, *tmp;
519
520         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
521                 if (fmd == keep)
522                         break;
523
524                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
525                     fed->fed_mod_count < filter->fo_fmd_max_num)
526                         break;
527
528                 cfs_list_del_init(&fmd->fmd_list);
529                 filter_fmd_put_nolock(fed, fmd); /* list reference */
530         }
531 }
532
533 void filter_fmd_expire(struct obd_export *exp)
534 {
535         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
536         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
537                                  &exp->exp_filter_data, NULL);
538         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
539 }
540
541 /* find specified objid, group in export fmd list.
542  * caller must hold fed_lock and take fmd reference itself */
543 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
544                                                 struct filter_export_data *fed,
545                                                 obd_id objid, obd_gr group)
546 {
547         struct filter_mod_data *found = NULL, *fmd;
548
549         LASSERT_SPIN_LOCKED(&fed->fed_lock);
550
551         cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
552                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
553                         found = fmd;
554                         cfs_list_del(&fmd->fmd_list);
555                         cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
556                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
557                         break;
558                 }
559         }
560
561         filter_fmd_expire_nolock(filter, fed, found);
562
563         return found;
564 }
565
566 /* Find fmd based on objid and group, or return NULL if not found. */
567 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
568                                         obd_id objid, obd_gr group)
569 {
570         struct filter_mod_data *fmd;
571
572         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
573         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
574                                      &exp->exp_filter_data, objid, group);
575         if (fmd)
576                 fmd->fmd_refcount++;    /* caller reference */
577         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
578
579         return fmd;
580 }
581
582 /* Find fmd based on objid and group, or create a new one if none is found.
583  * It is possible for this function to return NULL under memory pressure,
584  * or if objid = 0 is passed (which will only cause old entries to expire).
585  * Currently this is not fatal because any fmd state is transient and
586  * may also be freed when it gets sufficiently old. */
587 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
588                                        obd_id objid, obd_gr group)
589 {
590         struct filter_export_data *fed = &exp->exp_filter_data;
591         struct filter_mod_data *found = NULL, *fmd_new = NULL;
592
593         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
594
595         cfs_spin_lock(&fed->fed_lock);
596         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
597         if (fmd_new) {
598                 if (found == NULL) {
599                         cfs_list_add_tail(&fmd_new->fmd_list,
600                                           &fed->fed_mod_list);
601                         fmd_new->fmd_id = objid;
602                         fmd_new->fmd_gr = group;
603                         fmd_new->fmd_refcount++;   /* list reference */
604                         found = fmd_new;
605                         fed->fed_mod_count++;
606                 } else {
607                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
608                 }
609         }
610         if (found) {
611                 found->fmd_refcount++;          /* caller reference */
612                 found->fmd_expire = jiffies +
613                         exp->exp_obd->u.filter.fo_fmd_max_age;
614         }
615
616         cfs_spin_unlock(&fed->fed_lock);
617
618         return found;
619 }
620
621 #ifdef DO_FMD_DROP
622 /* drop fmd list reference so it will disappear when last reference is put.
623  * This isn't so critical because it would in fact only affect the one client
624  * that is doing the unlink and at worst we have an stale entry referencing
625  * an object that should never be used again. */
626 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
627 {
628         struct filter_mod_data *found = NULL;
629
630         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
631         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
632         if (found) {
633                 cfs_list_del_init(&found->fmd_list);
634                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
635         }
636         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
637 }
638 #else
639 #define filter_fmd_drop(exp, objid, group)
640 #endif
641
642 /* remove all entries from fmd list */
643 static void filter_fmd_cleanup(struct obd_export *exp)
644 {
645         struct filter_export_data *fed = &exp->exp_filter_data;
646         struct filter_mod_data *fmd = NULL, *tmp;
647
648         cfs_spin_lock(&fed->fed_lock);
649         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
650                 cfs_list_del_init(&fmd->fmd_list);
651                 filter_fmd_put_nolock(fed, fmd);
652         }
653         cfs_spin_unlock(&fed->fed_lock);
654 }
655
656 static int filter_init_export(struct obd_export *exp)
657 {
658         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
659         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
660
661         cfs_spin_lock(&exp->exp_lock);
662         exp->exp_connecting = 1;
663         cfs_spin_unlock(&exp->exp_lock);
664
665         return ldlm_init_export(exp);
666 }
667
668 static int filter_free_server_data(struct filter_obd *filter)
669 {
670         OBD_FREE_PTR(filter->fo_fsd);
671         filter->fo_fsd = NULL;
672         OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
673         filter->fo_last_rcvd_slots = NULL;
674         return 0;
675 }
676
677 /* assumes caller is already in kernel ctxt */
678 int filter_update_server_data(struct obd_device *obd, struct file *filp,
679                               struct lr_server_data *fsd)
680 {
681         loff_t off = 0;
682         int rc;
683         ENTRY;
684
685         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
686         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
687                le64_to_cpu(fsd->lsd_last_transno));
688         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
689                le64_to_cpu(fsd->lsd_mount_count));
690
691         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, 0);
692         if (rc)
693                 CERROR("error writing lr_server_data: rc = %d\n", rc);
694
695         RETURN(rc);
696 }
697
698 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
699                              int force_sync)
700 {
701         struct filter_obd *filter = &obd->u.filter;
702         __u64 tmp;
703         loff_t off = 0;
704         int rc;
705         ENTRY;
706
707         if (filter->fo_last_objid_files[group] == NULL) {
708                 CERROR("Object group "LPU64" not fully setup; not updating "
709                        "last_objid\n", group);
710                 RETURN(-EINVAL);
711         }
712
713         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
714                obd->obd_name, group, filter->fo_last_objids[group]);
715
716         tmp = cpu_to_le64(filter->fo_last_objids[group]);
717         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
718                                  &tmp, sizeof(tmp), &off, force_sync);
719         if (rc)
720                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
721                        group, rc);
722         RETURN(rc);
723 }
724 extern int ost_handle(struct ptlrpc_request *req);
725 /* assumes caller has already in kernel ctxt */
726 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
727 {
728         struct filter_obd *filter = &obd->u.filter;
729         struct lr_server_data *fsd;
730         struct lsd_client_data *lcd = NULL;
731         struct inode *inode = filp->f_dentry->d_inode;
732         unsigned long last_rcvd_size = i_size_read(inode);
733         __u64 mount_count;
734         __u32 start_epoch;
735         int cl_idx;
736         loff_t off = 0;
737         int rc;
738
739         /* ensure padding in the struct is the correct size */
740         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
741                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
742         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
743                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
744
745         OBD_ALLOC(fsd, sizeof(*fsd));
746         if (!fsd)
747                 RETURN(-ENOMEM);
748         filter->fo_fsd = fsd;
749
750         OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
751         if (filter->fo_last_rcvd_slots == NULL) {
752                 OBD_FREE(fsd, sizeof(*fsd));
753                 RETURN(-ENOMEM);
754         }
755
756         if (last_rcvd_size == 0) {
757                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
758
759                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
760                 fsd->lsd_last_transno = 0;
761                 mount_count = fsd->lsd_mount_count = 0;
762                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
763                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
764                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
765                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
766                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
767                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
768                  * MDS first connects and assigns the OST index number. */
769                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
770                                                         OBD_INCOMPAT_OST);
771         } else {
772                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
773                 if (rc) {
774                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
775                                LAST_RCVD, rc);
776                         GOTO(err_fsd, rc);
777                 }
778                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
779                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
780                                            "using the wrong disk %s. Were the "
781                                            "/dev/ assignments rearranged?\n",
782                                            obd->obd_uuid.uuid, fsd->lsd_uuid);
783                         GOTO(err_fsd, rc = -EINVAL);
784                 }
785                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
786                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
787                 /* COMPAT_146 */
788                 /* Assume old last_rcvd format unless I_C_LR is set */
789                 if (!(fsd->lsd_feature_incompat &
790                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
791                         fsd->lsd_last_transno = fsd->lsd_compat14;
792                 /* end COMPAT_146 */
793                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
794                  * MDS first connects and assigns the OST index number. */
795                 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
796                                                          OBD_INCOMPAT_OST);
797         }
798
799         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
800                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
801                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
802                        ~FILTER_INCOMPAT_SUPP);
803                 GOTO(err_fsd, rc = -EINVAL);
804         }
805         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
806                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
807                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
808                        ~FILTER_ROCOMPAT_SUPP);
809                 /* Do something like remount filesystem read-only */
810                 GOTO(err_fsd, rc = -EINVAL);
811         }
812
813         start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
814
815         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
816                obd->obd_name, start_epoch);
817         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
818                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
819         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
820                obd->obd_name, mount_count + 1);
821         CDEBUG(D_INODE, "%s: server data size: %u\n",
822                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
823         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
824                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
825         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
826                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
827         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
828                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
829         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
830                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
831                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
832                 le16_to_cpu(fsd->lsd_client_size));
833
834         if (!obd->obd_replayable) {
835                 CWARN("%s: recovery support OFF\n", obd->obd_name);
836                 GOTO(out, rc = 0);
837         }
838
839         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
840              off < last_rcvd_size; cl_idx++) {
841                 __u64 last_rcvd;
842                 struct obd_export *exp;
843                 struct filter_export_data *fed;
844
845                 if (!lcd) {
846                         OBD_ALLOC_PTR(lcd);
847                         if (!lcd)
848                                 GOTO(err_client, rc = -ENOMEM);
849                 }
850
851                 /* Don't assume off is incremented properly by
852                  * fsfilt_read_record(), in case sizeof(*lcd)
853                  * isn't the same as fsd->lsd_client_size.  */
854                 off = le32_to_cpu(fsd->lsd_client_start) +
855                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
856                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
857                 if (rc) {
858                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
859                                LAST_RCVD, cl_idx, off, rc);
860                         break; /* read error shouldn't cause startup to fail */
861                 }
862
863                 if (lcd->lcd_uuid[0] == '\0') {
864                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
865                                cl_idx);
866                         continue;
867                 }
868
869                 check_lcd(obd->obd_name, cl_idx, lcd);
870
871                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
872
873                 /* These exports are cleaned up by filter_disconnect(), so they
874                  * need to be set up like real exports as filter_connect() does.
875                  */
876                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
877
878                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
879                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
880                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
881                 if (IS_ERR(exp)) {
882                         if (PTR_ERR(exp) == -EALREADY) {
883                                 /* export already exists, zero out this one */
884                                 CERROR("Zeroing out duplicate export due to "
885                                        "bug 10479.\n");
886                                 lcd->lcd_uuid[0] = '\0';
887                         } else {
888                                 GOTO(err_client, rc = PTR_ERR(exp));
889                         }
890                 } else {
891                         fed = &exp->exp_filter_data;
892                         fed->fed_lcd = lcd;
893                         fed->fed_group = 0; /* will be assigned at connect */
894                         filter_export_stats_init(obd, exp, NULL);
895                         rc = filter_client_add(obd, exp, cl_idx);
896                         /* can't fail for existing client */
897                         LASSERTF(rc == 0, "rc = %d\n", rc);
898
899                         /* VBR: set export last committed */
900                         exp->exp_last_committed = last_rcvd;
901                         cfs_spin_lock(&exp->exp_lock);
902                         exp->exp_connecting = 0;
903                         exp->exp_in_recovery = 0;
904                         cfs_spin_unlock(&exp->exp_lock);
905                         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
906                         obd->obd_max_recoverable_clients++;
907                         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
908                         lcd = NULL;
909                         class_export_put(exp);
910                 }
911
912                 /* Need to check last_rcvd even for duplicated exports. */
913                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
914                        cl_idx, last_rcvd);
915
916                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
917                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
918         }
919
920         if (lcd)
921                 OBD_FREE_PTR(lcd);
922
923         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
924 out:
925         filter->fo_mount_count = mount_count + 1;
926         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
927
928         /* save it, so mount count and last_transno is current */
929         rc = filter_update_server_data(obd, filp, filter->fo_fsd);
930         if (rc)
931                 GOTO(err_client, rc);
932
933         RETURN(0);
934
935 err_client:
936         class_disconnect_exports(obd);
937 err_fsd:
938         filter_free_server_data(filter);
939         RETURN(rc);
940 }
941
942 static int filter_cleanup_groups(struct obd_device *obd)
943 {
944         struct filter_obd *filter = &obd->u.filter;
945         struct file *filp;
946         struct dentry *dentry;
947         int i, j;
948         ENTRY;
949
950         if (filter->fo_dentry_O_groups != NULL) {
951                 for (i = 0; i < filter->fo_group_count; i++) {
952                         dentry = filter->fo_dentry_O_groups[i];
953                         if (dentry != NULL)
954                                 f_dput(dentry);
955                 }
956                 OBD_FREE(filter->fo_dentry_O_groups,
957                          filter->fo_group_count *
958                          sizeof(*filter->fo_dentry_O_groups));
959                 filter->fo_dentry_O_groups = NULL;
960         }
961         if (filter->fo_last_objid_files != NULL) {
962                 for (i = 0; i < filter->fo_group_count; i++) {
963                         filp = filter->fo_last_objid_files[i];
964                         if (filp != NULL)
965                                 filp_close(filp, 0);
966                 }
967                 OBD_FREE(filter->fo_last_objid_files,
968                          filter->fo_group_count *
969                          sizeof(*filter->fo_last_objid_files));
970                 filter->fo_last_objid_files = NULL;
971         }
972         if (filter->fo_dentry_O_sub != NULL) {
973                 for (i = 0; i < filter->fo_group_count; i++) {
974                         for (j = 0; j < filter->fo_subdir_count; j++) {
975                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
976                                 if (dentry != NULL)
977                                         f_dput(dentry);
978                         }
979                 }
980                 OBD_FREE(filter->fo_dentry_O_sub,
981                          filter->fo_group_count *
982                          sizeof(*filter->fo_dentry_O_sub));
983                 filter->fo_dentry_O_sub = NULL;
984         }
985         if (filter->fo_last_objids != NULL) {
986                 OBD_FREE(filter->fo_last_objids,
987                          filter->fo_group_count *
988                          sizeof(*filter->fo_last_objids));
989                 filter->fo_last_objids = NULL;
990         }
991         if (filter->fo_dentry_O != NULL) {
992                 f_dput(filter->fo_dentry_O);
993                 filter->fo_dentry_O = NULL;
994         }
995         RETURN(0);
996 }
997
998 static int filter_update_last_group(struct obd_device *obd, int group)
999 {
1000         struct filter_obd *filter = &obd->u.filter;
1001         struct file *filp = NULL;
1002         int last_group = 0, rc;
1003         loff_t off = 0;
1004         ENTRY;
1005
1006         if (group <= filter->fo_committed_group)
1007                 RETURN(0);
1008
1009         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1010         if (IS_ERR(filp)) {
1011                 rc = PTR_ERR(filp);
1012                 filp = NULL;
1013                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1014                 GOTO(cleanup, rc);
1015         }
1016
1017         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1018         if (rc) {
1019                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1020                 GOTO(cleanup, rc);
1021         }
1022
1023         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1024                obd->obd_name, last_group, group);
1025
1026         off = 0;
1027         last_group = group;
1028         /* must be sync: bXXXX */
1029         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1030         if (rc) {
1031                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1032                 GOTO(cleanup, rc);
1033         }
1034
1035         filter->fo_committed_group = group;
1036 cleanup:
1037         if (filp)
1038                 filp_close(filp, 0);
1039         RETURN(rc);
1040 }
1041
1042 static int filter_read_group_internal(struct obd_device *obd, int group,
1043                                       int create)
1044 {
1045         struct filter_obd *filter = &obd->u.filter;
1046         __u64 *new_objids = NULL;
1047         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1048         struct dentry **new_groups = NULL;
1049         struct file **new_files = NULL;
1050         struct dentry *dentry;
1051         struct file *filp;
1052         int old_count = filter->fo_group_count, rc, stage = 0, i;
1053         char name[25];
1054         __u64 last_objid;
1055         loff_t off = 0;
1056         int len = group + 1;
1057
1058         snprintf(name, 24, "%d", group);
1059         name[24] = '\0';
1060
1061         if (!create) {
1062                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1063                                            strlen(name));
1064                 if (IS_ERR(dentry)) {
1065                         CERROR("Cannot lookup expected object group %d: %ld\n",
1066                                group, PTR_ERR(dentry));
1067                         RETURN(PTR_ERR(dentry));
1068                 }
1069         } else {
1070                 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1071                                       name, 0700, 1);
1072                 if (IS_ERR(dentry)) {
1073                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1074                                PTR_ERR(dentry));
1075                         RETURN(PTR_ERR(dentry));
1076                 }
1077         }
1078         stage = 1;
1079
1080         snprintf(name, 24, "O/%d/LAST_ID", group);
1081         name[24] = '\0';
1082         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1083         if (IS_ERR(filp)) {
1084                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1085                 GOTO(cleanup, rc = PTR_ERR(filp));
1086         }
1087         stage = 2;
1088
1089         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1090         if (rc) {
1091                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1092                 GOTO(cleanup, rc);
1093         }
1094
1095         if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1096                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1097                 if (tmp_subdirs == NULL)
1098                         GOTO(cleanup, rc = -ENOMEM);
1099                 stage = 3;
1100
1101                 for (i = 0; i < filter->fo_subdir_count; i++) {
1102                         char dir[20];
1103                         snprintf(dir, sizeof(dir), "d%u", i);
1104
1105                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1106                                                               filter->fo_vfsmnt,
1107                                                               dir, 0700, 1);
1108                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1109                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1110                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1111                                        group, dir, rc);
1112                                 GOTO(cleanup, rc);
1113                         }
1114
1115                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1116                                tmp_subdirs->dentry[i]);
1117                 }
1118         }
1119
1120         /* 'group' is an index; we need an array of length 'group + 1' */
1121         if (group + 1 > old_count) {
1122                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1123                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1124                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1125                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1126                 stage = 4;
1127                 if (new_objids == NULL || new_subdirs == NULL ||
1128                     new_groups == NULL || new_files == NULL)
1129                         GOTO(cleanup, rc = -ENOMEM);
1130
1131                 if (old_count) {
1132                         memcpy(new_objids, filter->fo_last_objids,
1133                                old_count * sizeof(*new_objids));
1134                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1135                                old_count * sizeof(*new_subdirs));
1136                         memcpy(new_groups, filter->fo_dentry_O_groups,
1137                                old_count * sizeof(*new_groups));
1138                         memcpy(new_files, filter->fo_last_objid_files,
1139                                old_count * sizeof(*new_files));
1140
1141                         OBD_FREE(filter->fo_last_objids,
1142                                  old_count * sizeof(*new_objids));
1143                         OBD_FREE(filter->fo_dentry_O_sub,
1144                                  old_count * sizeof(*new_subdirs));
1145                         OBD_FREE(filter->fo_dentry_O_groups,
1146                                  old_count * sizeof(*new_groups));
1147                         OBD_FREE(filter->fo_last_objid_files,
1148                                  old_count * sizeof(*new_files));
1149                 }
1150                 filter->fo_last_objids = new_objids;
1151                 filter->fo_dentry_O_sub = new_subdirs;
1152                 filter->fo_dentry_O_groups = new_groups;
1153                 filter->fo_last_objid_files = new_files;
1154                 filter->fo_group_count = len;
1155         }
1156
1157         filter->fo_dentry_O_groups[group] = dentry;
1158         filter->fo_last_objid_files[group] = filp;
1159         if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1160                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1161                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1162         }
1163
1164         filter_update_last_group(obd, group);
1165
1166         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1167                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1168                 rc = filter_update_last_objid(obd, group, 1);
1169                 RETURN(rc);
1170         }
1171
1172         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1173         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1174                obd->obd_name, group, last_objid);
1175         RETURN(0);
1176  cleanup:
1177         switch (stage) {
1178         case 4:
1179                 if (new_objids != NULL)
1180                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1181                 if (new_subdirs != NULL)
1182                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1183                 if (new_groups != NULL)
1184                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1185                 if (new_files != NULL)
1186                         OBD_FREE(new_files, len * sizeof(*new_files));
1187         case 3:
1188                 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1189                         for (i = 0; i < filter->fo_subdir_count; i++) {
1190                                 if (tmp_subdirs->dentry[i] != NULL)
1191                                         dput(tmp_subdirs->dentry[i]);
1192                         }
1193                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1194                 }
1195         case 2:
1196                 filp_close(filp, 0);
1197         case 1:
1198                 dput(dentry);
1199         }
1200         RETURN(rc);
1201 }
1202
1203 static int filter_read_groups(struct obd_device *obd, int last_group,
1204                               int create)
1205 {
1206         struct filter_obd *filter = &obd->u.filter;
1207         int old_count, group, rc = 0;
1208
1209         cfs_down(&filter->fo_init_lock);
1210         old_count = filter->fo_group_count;
1211         for (group = old_count; group <= last_group; group++) {
1212                 rc = filter_read_group_internal(obd, group, create);
1213                 if (rc != 0)
1214                         break;
1215         }
1216         cfs_up(&filter->fo_init_lock);
1217         return rc;
1218 }
1219
1220 /* FIXME: object groups */
1221 static int filter_prep_groups(struct obd_device *obd)
1222 {
1223         struct filter_obd *filter = &obd->u.filter;
1224         struct dentry *O_dentry;
1225         struct file *filp;
1226         int    last_group, rc = 0, cleanup_phase = 0;
1227         loff_t off = 0;
1228         ENTRY;
1229
1230         O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1231                                 "O", 0700, 1);
1232         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1233         if (IS_ERR(O_dentry)) {
1234                 rc = PTR_ERR(O_dentry);
1235                 CERROR("cannot open/create O: rc = %d\n", rc);
1236                 GOTO(cleanup, rc);
1237         }
1238         filter->fo_dentry_O = O_dentry;
1239         cleanup_phase = 1; /* O_dentry */
1240
1241         /* we have to initialize all groups before first connections from
1242          * clients because they may send create/destroy for any group -bzzz */
1243         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1244         if (IS_ERR(filp)) {
1245                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1246                 GOTO(cleanup, rc = PTR_ERR(filp));
1247         }
1248         cleanup_phase = 2; /* filp */
1249
1250         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1251         if (rc) {
1252                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1253                 GOTO(cleanup, rc);
1254         }
1255
1256         if (off == 0)
1257                 last_group = FILTER_GROUP_MDS0;
1258
1259         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1260               FILTER_GROUP_MDS0, last_group);
1261         filter->fo_committed_group = last_group;
1262         rc = filter_read_groups(obd, last_group, 1);
1263         if (rc)
1264                 GOTO(cleanup, rc);
1265
1266         filp_close(filp, 0);
1267         RETURN(0);
1268
1269  cleanup:
1270         switch (cleanup_phase) {
1271         case 2:
1272                 filp_close(filp, 0);
1273         case 1:
1274                 filter_cleanup_groups(obd);
1275                 f_dput(filter->fo_dentry_O);
1276                 filter->fo_dentry_O = NULL;
1277         default:
1278                 break;
1279         }
1280         return rc;
1281
1282 }
1283
1284 /* setup the object store with correct subdirectories */
1285 static int filter_prep(struct obd_device *obd)
1286 {
1287         struct lvfs_run_ctxt saved;
1288         struct filter_obd *filter = &obd->u.filter;
1289         struct file *file;
1290         struct inode *inode;
1291         int rc = 0;
1292         ENTRY;
1293
1294         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1295         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1296         if (!file || IS_ERR(file)) {
1297                 rc = PTR_ERR(file);
1298                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1299                        LAST_RCVD, rc);
1300                 GOTO(out, rc);
1301         }
1302         filter->fo_rcvd_filp = file;
1303         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1304                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1305                        file->f_dentry->d_inode->i_mode);
1306                 GOTO(err_filp, rc = -ENOENT);
1307         }
1308
1309         inode = file->f_dentry->d_parent->d_inode;
1310         /* We use i_op->unlink directly in filter_vfs_unlink() */
1311         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1312                 CERROR("%s: filesystem does not support create/unlink ops\n",
1313                        obd->obd_name);
1314                 GOTO(err_filp, rc = -EOPNOTSUPP);
1315         }
1316
1317         /** lu_target has very limited use in filter now */
1318         lut_init(NULL, &filter->fo_lut, obd, NULL);
1319
1320         rc = filter_init_server_data(obd, file);
1321         if (rc) {
1322                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1323                 GOTO(err_filp, rc);
1324         }
1325
1326         target_recovery_init(&filter->fo_lut, ost_handle);
1327
1328         /* open and create health check io file*/
1329         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1330         if (IS_ERR(file)) {
1331                 rc = PTR_ERR(file);
1332                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1333                        HEALTH_CHECK, rc);
1334                 GOTO(err_server_data, rc);
1335         }
1336         filter->fo_obt.obt_health_check_filp = file;
1337         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1338                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1339                        file->f_dentry->d_inode->i_mode);
1340                 GOTO(err_health_check, rc = -ENOENT);
1341         }
1342         rc = lvfs_check_io_health(obd, file);
1343         if (rc)
1344                 GOTO(err_health_check, rc);
1345
1346         rc = filter_prep_groups(obd);
1347         if (rc)
1348                 GOTO(err_health_check, rc);
1349 out:
1350         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1351
1352         return(rc);
1353
1354 err_health_check:
1355         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1356                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1357         filter->fo_obt.obt_health_check_filp = NULL;
1358 err_server_data:
1359         target_recovery_fini(obd);
1360         filter_free_server_data(filter);
1361 err_filp:
1362         if (filp_close(filter->fo_rcvd_filp, 0))
1363                 CERROR("can't close %s after error\n", LAST_RCVD);
1364         filter->fo_rcvd_filp = NULL;
1365         goto out;
1366 }
1367
1368 /* cleanup the filter: write last used object id to status file */
1369 static void filter_post(struct obd_device *obd)
1370 {
1371         struct lvfs_run_ctxt saved;
1372         struct filter_obd *filter = &obd->u.filter;
1373         int rc, i;
1374
1375         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1376          * best to start a transaction with h_sync, because we removed this
1377          * from lastobjid */
1378
1379         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1380         rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1381                                        filter->fo_fsd);
1382         if (rc)
1383                 CERROR("error writing server data: rc = %d\n", rc);
1384
1385         for (i = 0; i < filter->fo_group_count; i++) {
1386                 rc = filter_update_last_objid(obd, i,
1387                                 (i == filter->fo_group_count - 1));
1388                 if (rc)
1389                         CERROR("error writing group %d lastobjid: rc = %d\n",
1390                                i, rc);
1391         }
1392
1393         rc = filp_close(filter->fo_rcvd_filp, 0);
1394         filter->fo_rcvd_filp = NULL;
1395         if (rc)
1396                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1397
1398         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1399         filter->fo_obt.obt_health_check_filp = NULL;
1400         if (rc)
1401                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1402
1403         filter_cleanup_groups(obd);
1404         filter_free_server_data(filter);
1405         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1406
1407         filter_free_capa_keys(filter);
1408         cleanup_capa_hash(filter->fo_capa_hash);
1409 }
1410
1411 static void filter_set_last_id(struct filter_obd *filter,
1412                                obd_id id, obd_gr group)
1413 {
1414         LASSERT(filter->fo_fsd != NULL);
1415         LASSERT(group <= filter->fo_group_count);
1416
1417         cfs_spin_lock(&filter->fo_objidlock);
1418         filter->fo_last_objids[group] = id;
1419         cfs_spin_unlock(&filter->fo_objidlock);
1420 }
1421
1422 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1423 {
1424         obd_id id;
1425         LASSERT(filter->fo_fsd != NULL);
1426         LASSERT(group <= filter->fo_group_count);
1427
1428         /* FIXME: object groups */
1429         cfs_spin_lock(&filter->fo_objidlock);
1430         id = filter->fo_last_objids[group];
1431         cfs_spin_unlock(&filter->fo_objidlock);
1432         return id;
1433 }
1434
1435 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1436 {
1437         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1438         return 0;
1439 }
1440
1441 /* We never dget the object parent, so DON'T dput it either */
1442 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1443 {
1444         struct filter_obd *filter = &obd->u.filter;
1445         struct filter_subdirs *subdirs;
1446         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1447
1448         if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1449                 return filter->fo_dentry_O_groups[group];
1450
1451         subdirs = &filter->fo_dentry_O_sub[group];
1452         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1453 }
1454
1455 /* We never dget the object parent, so DON'T dput it either */
1456 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1457                                   obd_id objid)
1458 {
1459         unsigned long now = jiffies;
1460         struct dentry *dparent = filter_parent(obd, group, objid);
1461         int rc;
1462
1463         if (IS_ERR(dparent))
1464                 return dparent;
1465         if (dparent == NULL)
1466                 return ERR_PTR(-ENOENT);
1467
1468         rc = filter_lock_dentry(obd, dparent);
1469         fsfilt_check_slow(obd, now, "parent lock");
1470         return rc ? ERR_PTR(rc) : dparent;
1471 }
1472
1473 /* We never dget the object parent, so DON'T dput it either */
1474 static void filter_parent_unlock(struct dentry *dparent)
1475 {
1476         UNLOCK_INODE_MUTEX(dparent->d_inode);
1477 }
1478
1479 /* How to get files, dentries, inodes from object id's.
1480  *
1481  * If dir_dentry is passed, the caller has already locked the parent
1482  * appropriately for this operation (normally a write lock).  If
1483  * dir_dentry is NULL, we do a read lock while we do the lookup to
1484  * avoid races with create/destroy and such changing the directory
1485  * internal to the filesystem code. */
1486 struct dentry *filter_fid2dentry(struct obd_device *obd,
1487                                  struct dentry *dir_dentry,
1488                                  obd_gr group, obd_id id)
1489 {
1490         struct dentry *dparent = dir_dentry;
1491         struct dentry *dchild;
1492         char name[32];
1493         int len;
1494         ENTRY;
1495
1496         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1497             obd->u.filter.fo_destroys_in_progress == 0) {
1498                 /* don't fail lookups for orphan recovery, it causes
1499                  * later LBUGs when objects still exist during precreate */
1500                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1501                 RETURN(ERR_PTR(-ENOENT));
1502         }
1503         if (id == 0) {
1504                 CERROR("fatal: invalid object id 0\n");
1505                 RETURN(ERR_PTR(-ESTALE));
1506         }
1507
1508         len = sprintf(name, LPU64, id);
1509         if (dir_dentry == NULL) {
1510                 dparent = filter_parent_lock(obd, group, id);
1511                 if (IS_ERR(dparent)) {
1512                         CERROR("%s: error getting object "LPU64":"LPU64
1513                                " parent: rc %ld\n", obd->obd_name,
1514                                id, group, PTR_ERR(dparent));
1515                         RETURN(dparent);
1516                 }
1517         }
1518         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1519                dparent->d_name.len, dparent->d_name.name, name);
1520         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1521         if (dir_dentry == NULL)
1522                 filter_parent_unlock(dparent);
1523         if (IS_ERR(dchild)) {
1524                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1525                        PTR_ERR(dchild));
1526                 RETURN(dchild);
1527         }
1528
1529         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1530                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1531                        obd->obd_name, id, dchild->d_inode->i_ino);
1532                 f_dput(dchild);
1533                 RETURN(ERR_PTR(-ENOENT));
1534         }
1535
1536         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1537                name, dchild, atomic_read(&dchild->d_count));
1538
1539         LASSERT(atomic_read(&dchild->d_count) > 0);
1540
1541         RETURN(dchild);
1542 }
1543
1544 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1545                                   obd_id group, struct lustre_handle *lockh)
1546 {
1547         int flags = LDLM_AST_DISCARD_DATA, rc;
1548         struct ldlm_res_id res_id;
1549         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1550         ENTRY;
1551
1552         osc_build_res_name(objid, group, &res_id);
1553         /* Tell the clients that the object is gone now and that they should
1554          * throw away any cached pages. */
1555         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1556                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1557                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1558                                     lockh);
1559         if (rc != ELDLM_OK)
1560                 lockh->cookie = 0;
1561         RETURN(rc);
1562 }
1563
1564 static void filter_fini_destroy(struct obd_device *obd,
1565                                 struct lustre_handle *lockh)
1566 {
1567         if (lustre_handle_is_used(lockh))
1568                 ldlm_lock_decref(lockh, LCK_PW);
1569 }
1570
1571 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1572  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1573  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1574  * reverse.  Caller must take i_sem before starting the transaction and we
1575  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1576 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1577                       struct vfsmount *mnt)
1578 {
1579         int rc;
1580         ENTRY;
1581
1582         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1583          * itself we already hold dir->i_mutex for child create/unlink ops */
1584         LASSERT(dentry->d_inode != NULL);
1585         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1586         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1587
1588
1589         /* may_delete() */
1590         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1591                 GOTO(out, rc = -ENOENT);
1592
1593         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1594         if (rc)
1595                 GOTO(out, rc);
1596
1597         if (IS_APPEND(dir))
1598                 GOTO(out, rc = -EPERM);
1599
1600         /* check_sticky() */
1601         if ((dentry->d_inode->i_uid != current->fsuid &&
1602              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1603             IS_IMMUTABLE(dentry->d_inode))
1604                 GOTO(out, rc = -EPERM);
1605
1606         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1607          *       that was done because of journal_start (which is already done
1608          *       here) or some other ordering issue. */
1609         DQUOT_INIT(dir);
1610
1611         rc = ll_security_inode_unlink(dir, dentry, mnt);
1612         if (rc)
1613                 GOTO(out, rc);
1614
1615         rc = dir->i_op->unlink(dir, dentry);
1616 out:
1617         /* need to drop i_mutex before we lose inode reference */
1618         UNLOCK_INODE_MUTEX(dentry->d_inode);
1619         if (rc == 0)
1620                 d_delete(dentry);
1621
1622         RETURN(rc);
1623 }
1624
1625 /* Caller must hold LCK_PW on parent and push us into kernel context.
1626  * Caller must hold child i_mutex, we drop it always.
1627  * Caller is also required to ensure that dchild->d_inode exists. */
1628 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1629                                    obd_gr group, struct dentry *dparent,
1630                                    struct dentry *dchild)
1631 {
1632         struct inode *inode = dchild->d_inode;
1633         struct filter_obd *filter = &obd->u.filter;
1634         int rc;
1635
1636         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1637                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1638                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1639                        (unsigned long)inode->i_nlink,
1640                        atomic_read(&inode->i_count));
1641         }
1642
1643         rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1644         if (rc)
1645                 CERROR("error unlinking objid %.*s: rc %d\n",
1646                        dchild->d_name.len, dchild->d_name.name, rc);
1647         return(rc);
1648 }
1649
1650 struct filter_intent_args {
1651         struct ldlm_lock **victim;
1652         __u64 size;
1653         int *liblustre;
1654 };
1655
1656 static enum interval_iter filter_intent_cb(struct interval_node *n,
1657                                            void *args)
1658 {
1659         struct ldlm_interval *node = (struct ldlm_interval *)n;
1660         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1661         __u64 size = arg->size;
1662         struct ldlm_lock **v = arg->victim;
1663         struct ldlm_lock *lck;
1664
1665         /* If the interval is lower than the current file size,
1666          * just break. */
1667         if (interval_high(n) <= size)
1668                 return INTERVAL_ITER_STOP;
1669
1670         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1671                 /* Don't send glimpse ASTs to liblustre clients.
1672                  * They aren't listening for them, and they do
1673                  * entirely synchronous I/O anyways. */
1674                 if (lck->l_export == NULL ||
1675                     lck->l_export->exp_libclient == 1)
1676                         continue;
1677
1678                 if (*arg->liblustre)
1679                         *arg->liblustre = 0;
1680
1681                 if (*v == NULL) {
1682                         *v = LDLM_LOCK_GET(lck);
1683                 } else if ((*v)->l_policy_data.l_extent.start <
1684                            lck->l_policy_data.l_extent.start) {
1685                         LDLM_LOCK_RELEASE(*v);
1686                         *v = LDLM_LOCK_GET(lck);
1687                 }
1688
1689                 /* the same policy group - every lock has the
1690                  * same extent, so needn't do it any more */
1691                 break;
1692         }
1693
1694         return INTERVAL_ITER_CONT;
1695 }
1696
1697 static int filter_intent_policy(struct ldlm_namespace *ns,
1698                                 struct ldlm_lock **lockp, void *req_cookie,
1699                                 ldlm_mode_t mode, int flags, void *data)
1700 {
1701         CFS_LIST_HEAD(rpc_list);
1702         struct ptlrpc_request *req = req_cookie;
1703         struct ldlm_lock *lock = *lockp, *l = NULL;
1704         struct ldlm_resource *res = lock->l_resource;
1705         ldlm_processing_policy policy;
1706         struct ost_lvb *res_lvb, *reply_lvb;
1707         struct ldlm_reply *rep;
1708         ldlm_error_t err;
1709         int idx, rc, tmpflags = 0, only_liblustre = 1;
1710         struct ldlm_interval_tree *tree;
1711         struct filter_intent_args arg;
1712         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1713                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1714                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1715         ENTRY;
1716
1717         policy = ldlm_get_processing_policy(res);
1718         LASSERT(policy != NULL);
1719         LASSERT(req != NULL);
1720
1721         rc = lustre_pack_reply(req, 3, repsize, NULL);
1722         if (rc)
1723                 RETURN(req->rq_status = rc);
1724
1725         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1726         LASSERT(rep != NULL);
1727
1728         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1729                                    sizeof(*reply_lvb));
1730         LASSERT(reply_lvb != NULL);
1731
1732         //fixup_handle_for_resent_req(req, lock, &lockh);
1733
1734         /* Call the extent policy function to see if our request can be
1735          * granted, or is blocked.
1736          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1737          * lock, and should not be granted if the lock will be blocked.
1738          */
1739
1740         LASSERT(ns == res->lr_namespace);
1741         lock_res(res);
1742         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1743         check_res_locked(res);
1744
1745         /* FIXME: we should change the policy function slightly, to not make
1746          * this list at all, since we just turn around and free it */
1747         while (!cfs_list_empty(&rpc_list)) {
1748                 struct ldlm_lock *wlock =
1749                         cfs_list_entry(rpc_list.next, struct ldlm_lock,
1750                                        l_cp_ast);
1751                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1752                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1753                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1754                 cfs_list_del_init(&wlock->l_cp_ast);
1755                 LDLM_LOCK_RELEASE(wlock);
1756         }
1757
1758         /* The lock met with no resistance; we're finished. */
1759         if (rc == LDLM_ITER_CONTINUE) {
1760                 /* do not grant locks to the liblustre clients: they cannot
1761                  * handle ASTs robustly.  We need to do this while still
1762                  * holding ns_lock to avoid the lock remaining on the res_link
1763                  * list (and potentially being added to l_pending_list by an
1764                  * AST) when we are going to drop this lock ASAP. */
1765                 if (lock->l_export->exp_libclient ||
1766                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1767                         ldlm_resource_unlink_lock(lock);
1768                         err = ELDLM_LOCK_ABORTED;
1769                 } else {
1770                         err = ELDLM_LOCK_REPLACED;
1771                 }
1772                 unlock_res(res);
1773                 RETURN(err);
1774         }
1775
1776         /* Do not grant any lock, but instead send GL callbacks.  The extent
1777          * policy nicely created a list of all PW locks for us.  We will choose
1778          * the highest of those which are larger than the size in the LVB, if
1779          * any, and perform a glimpse callback. */
1780         res_lvb = res->lr_lvb_data;
1781         LASSERT(res_lvb != NULL);
1782         *reply_lvb = *res_lvb;
1783
1784         /*
1785          * ->ns_lock guarantees that no new locks are granted, and,
1786          * therefore, that res->lr_lvb_data cannot increase beyond the
1787          * end of already granted lock. As a result, it is safe to
1788          * check against "stale" reply_lvb->lvb_size value without
1789          * res->lr_lvb_sem.
1790          */
1791         arg.size = reply_lvb->lvb_size;
1792         arg.victim = &l;
1793         arg.liblustre = &only_liblustre;
1794         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1795                 tree = &res->lr_itree[idx];
1796                 if (tree->lit_mode == LCK_PR)
1797                         continue;
1798
1799                 interval_iterate_reverse(tree->lit_root,
1800                                          filter_intent_cb, &arg);
1801         }
1802         unlock_res(res);
1803
1804         /* There were no PW locks beyond the size in the LVB; finished. */
1805         if (l == NULL) {
1806                 if (only_liblustre) {
1807                         /* If we discovered a liblustre client with a PW lock,
1808                          * however, the LVB may be out of date!  The LVB is
1809                          * updated only on glimpse (which we don't do for
1810                          * liblustre clients) and cancel (which the client
1811                          * obviously has not yet done).  So if it has written
1812                          * data but kept the lock, the LVB is stale and needs
1813                          * to be updated from disk.
1814                          *
1815                          * Of course, this will all disappear when we switch to
1816                          * taking liblustre locks on the OST. */
1817                         ldlm_res_lvbo_update(res, NULL, 1);
1818                 }
1819                 RETURN(ELDLM_LOCK_ABORTED);
1820         }
1821
1822         /*
1823          * This check is for lock taken in filter_prepare_destroy() that does
1824          * not have l_glimpse_ast set. So the logic is: if there is a lock
1825          * with no l_glimpse_ast set, this object is being destroyed already.
1826          *
1827          * Hence, if you are grabbing DLM locks on the server, always set
1828          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1829          */
1830         if (l->l_glimpse_ast == NULL) {
1831                 /* We are racing with unlink(); just return -ENOENT */
1832                 rep->lock_policy_res1 = -ENOENT;
1833                 goto out;
1834         }
1835
1836         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1837         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1838         /* Update the LVB from disk if the AST failed (this is a legal race) */
1839         /*
1840          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1841          * sending ast is not handled. This can result in lost client writes.
1842          */
1843         if (rc != 0)
1844                 ldlm_res_lvbo_update(res, NULL, 1);
1845
1846         lock_res(res);
1847         *reply_lvb = *res_lvb;
1848         unlock_res(res);
1849
1850  out:
1851         LDLM_LOCK_RELEASE(l);
1852
1853         RETURN(ELDLM_LOCK_ABORTED);
1854 }
1855
1856 /*
1857  * per-obd_device iobuf pool.
1858  *
1859  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1860  * allocations in write-path has to be minimized (see bug 5137).
1861  *
1862  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1863  * OST threads (see ost_thread_{init,done}()).
1864  *
1865  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1866  * at the OST layer there are only (potentially) multiple obd_device of type
1867  * unknown at the time of OST thread creation.
1868  *
1869  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1870  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1871  * it's very own iobuf.
1872  *
1873  * Functions below
1874  *
1875  *     filter_kiobuf_pool_init()
1876  *
1877  *     filter_kiobuf_pool_done()
1878  *
1879  *     filter_iobuf_get()
1880  *
1881  * operate on this array. They are "generic" in a sense that they don't depend
1882  * on actual type of iobuf's (the latter depending on Linux kernel version).
1883  */
1884
1885 /*
1886  * destroy pool created by filter_iobuf_pool_init
1887  */
1888 static void filter_iobuf_pool_done(struct filter_obd *filter)
1889 {
1890         struct filter_iobuf **pool;
1891         int i;
1892
1893         ENTRY;
1894
1895         pool = filter->fo_iobuf_pool;
1896         if (pool != NULL) {
1897                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1898                         if (pool[i] != NULL)
1899                                 filter_free_iobuf(pool[i]);
1900                 }
1901                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1902                 filter->fo_iobuf_pool = NULL;
1903         }
1904         EXIT;
1905 }
1906
1907 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1908 {
1909         struct filter_obd       *filter = &obd->u.filter;
1910         struct sptlrpc_rule_set  tmp_rset;
1911         int                      rc;
1912
1913         sptlrpc_rule_set_init(&tmp_rset);
1914         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1915         if (rc) {
1916                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1917                        obd->obd_name, rc);
1918                 return rc;
1919         }
1920
1921         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1922
1923         cfs_write_lock(&filter->fo_sptlrpc_lock);
1924         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1925         filter->fo_sptlrpc_rset = tmp_rset;
1926         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1927
1928         return 0;
1929 }
1930
1931 /*
1932  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1933  */
1934 static int filter_iobuf_pool_init(struct filter_obd *filter)
1935 {
1936         void **pool;
1937
1938         ENTRY;
1939
1940
1941         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1942                       GFP_KERNEL);
1943         if (filter->fo_iobuf_pool == NULL)
1944                 RETURN(-ENOMEM);
1945
1946         filter->fo_iobuf_count = OSS_THREADS_MAX;
1947
1948         RETURN(0);
1949 }
1950
1951 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1952  * many threads there will be so we allocate a large empty array and only
1953  * fill in those slots that are actually in use.
1954  * If we haven't allocated a pool entry for this thread before, do so now. */
1955 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1956 {
1957         int thread_id                    = (oti && oti->oti_thread) ?
1958                                            oti->oti_thread->t_id : -1;
1959         struct filter_iobuf  *pool       = NULL;
1960         struct filter_iobuf **pool_place = NULL;
1961
1962         if (thread_id >= 0) {
1963                 LASSERT(thread_id < filter->fo_iobuf_count);
1964                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1965         }
1966
1967         if (unlikely(pool == NULL)) {
1968                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1969                                           PTLRPC_MAX_BRW_PAGES);
1970                 if (pool_place != NULL)
1971                         *pool_place = pool;
1972         }
1973
1974         return pool;
1975 }
1976
1977 /* mount the file system (secretly).  lustre_cfg parameters are:
1978  * 1 = device
1979  * 2 = fstype
1980  * 3 = flags: failover=f, failout=n
1981  * 4 = mount options
1982  */
1983 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1984                         void *option)
1985 {
1986         struct filter_obd *filter = &obd->u.filter;
1987         struct vfsmount *mnt;
1988         struct lustre_mount_info *lmi;
1989         struct obd_uuid uuid;
1990         __u8 *uuid_ptr;
1991         char *str, *label;
1992         char ns_name[48];
1993         request_queue_t *q;
1994         int rc, i;
1995         ENTRY;
1996
1997         if (lcfg->lcfg_bufcount < 3 ||
1998             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1999             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
2000                 RETURN(-EINVAL);
2001
2002         lmi = server_get_mount(obd->obd_name);
2003         if (lmi) {
2004                 /* We already mounted in lustre_fill_super.
2005                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2006                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2007                 mnt = lmi->lmi_mnt;
2008                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2009         } else {
2010                 /* old path - used by lctl */
2011                 CERROR("Using old MDS mount method\n");
2012                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2013                                     MS_NOATIME|MS_NODIRATIME,
2014                                     lustre_cfg_string(lcfg, 1), option);
2015                 if (IS_ERR(mnt)) {
2016                         rc = PTR_ERR(mnt);
2017                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2018                                            lustre_cfg_string(lcfg, 1), rc);
2019                         RETURN(rc);
2020                 }
2021
2022                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2023         }
2024         if (IS_ERR(obd->obd_fsops))
2025                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2026
2027         rc = filter_iobuf_pool_init(filter);
2028         if (rc != 0)
2029                 GOTO(err_ops, rc);
2030
2031         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2032                 CERROR("%s: Underlying device is marked as read-only. "
2033                        "Setup failed\n", obd->obd_name);
2034                 GOTO(err_ops, rc = -EROFS);
2035         }
2036
2037         /* failover is the default */
2038         obd->obd_replayable = 1;
2039
2040         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2041                 str = lustre_cfg_string(lcfg, 3);
2042                 if (strchr(str, 'n')) {
2043                         CWARN("%s: recovery disabled\n", obd->obd_name);
2044                         obd->obd_replayable = 0;
2045                 }
2046         }
2047
2048         filter->fo_vfsmnt = mnt;
2049         obd->u.obt.obt_sb = mnt->mnt_sb;
2050         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2051         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2052
2053         fsfilt_setup(obd, obd->u.obt.obt_sb);
2054
2055         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2056         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2057         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2058         obd->obd_lvfs_ctxt.fs = get_ds();
2059         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2060
2061         cfs_init_mutex(&filter->fo_init_lock);
2062         filter->fo_committed_group = 0;
2063         filter->fo_destroys_in_progress = 0;
2064         for (i = 0; i < 32; i++)
2065                 cfs_sema_init(&filter->fo_create_locks[i], 1);
2066
2067         cfs_spin_lock_init(&filter->fo_translock);
2068         cfs_spin_lock_init(&filter->fo_objidlock);
2069         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2070         cfs_sema_init(&filter->fo_alloc_lock, 1);
2071         init_brw_stats(&filter->fo_filter_stats);
2072         filter->fo_read_cache = 1; /* enable read-only cache by default */
2073         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2074         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2075         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2076         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2077
2078         rc = filter_prep(obd);
2079         if (rc)
2080                 GOTO(err_ops, rc);
2081
2082         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2083         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2084
2085         filter->fo_fl_oss_capa = 1;
2086
2087         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2088         filter->fo_capa_hash = init_capa_hash();
2089         if (filter->fo_capa_hash == NULL)
2090                 GOTO(err_post, rc = -ENOMEM);
2091
2092         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2093         obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2094                                                 LDLM_NAMESPACE_GREEDY);
2095         if (obd->obd_namespace == NULL)
2096                 GOTO(err_post, rc = -ENOMEM);
2097         obd->obd_namespace->ns_lvbp = obd;
2098         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2099         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2100
2101         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2102                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2103
2104         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2105         if (rc) {
2106                 CERROR("failed to setup llogging subsystems\n");
2107                 GOTO(err_post, rc);
2108         }
2109
2110         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2111         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2112         /* do this after llog being initialized */
2113         filter_adapt_sptlrpc_conf(obd, 1);
2114
2115         rc = lquota_setup(filter_quota_interface_ref, obd);
2116         if (rc)
2117                 GOTO(err_post, rc);
2118
2119         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2120         if (q->max_sectors < q->max_hw_sectors &&
2121             q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2122                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2123                               "for larger I/O requests: max_sectors = %u "
2124                               "could be up to max_hw_sectors=%u\n",
2125                               obd->obd_name, mnt->mnt_sb->s_id,
2126                               q->max_sectors, q->max_hw_sectors);
2127
2128         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2129         if (uuid_ptr != NULL) {
2130                 class_uuid_unparse(uuid_ptr, &uuid);
2131                 str = uuid.uuid;
2132         } else {
2133                 str = "no UUID";
2134         }
2135
2136         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2137
2138         if (obd->obd_recovering) {
2139                 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2140                               "recovery for at least %d:%.02d, or until %d "
2141                               "client%s reconnect%s.\n",
2142                               obd->obd_name, lustre_cfg_string(lcfg, 1),
2143                               label ?: "", label ? "/" : "", str,
2144                               obd->obd_recovery_timeout / 60,
2145                               obd->obd_recovery_timeout % 60,
2146                               obd->obd_max_recoverable_clients,
2147                               (obd->obd_max_recoverable_clients == 1) ? "":"s",
2148                               (obd->obd_max_recoverable_clients == 1) ? "s":"");
2149         } else {
2150                 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2151                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2152                               label ?: "", label ? "/" : "", str,
2153                               obd->obd_replayable ? "enabled" : "disabled");
2154         }
2155
2156         RETURN(0);
2157
2158 err_post:
2159         filter_post(obd);
2160 err_ops:
2161         fsfilt_put_ops(obd->obd_fsops);
2162         filter_iobuf_pool_done(filter);
2163 err_mntput:
2164         server_put_mount(obd->obd_name, mnt);
2165         obd->u.obt.obt_sb = 0;
2166         return rc;
2167 }
2168
2169 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2170 {
2171         struct lprocfs_static_vars lvars;
2172         unsigned long addr;
2173         struct page *page;
2174         int rc;
2175
2176         CLASSERT(offsetof(struct obd_device, u.obt) ==
2177                  offsetof(struct obd_device, u.filter.fo_obt));
2178
2179         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2180                 RETURN(-EINVAL);
2181
2182         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2183         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2184         if (!page)
2185                 RETURN(-ENOMEM);
2186         addr = (unsigned long)cfs_page_address(page);
2187         clear_page((void *)addr);
2188
2189         /* lprocfs must be setup before the filter so state can be safely added
2190          * to /proc incrementally as the filter is setup */
2191         lprocfs_filter_init_vars(&lvars);
2192         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2193             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2194                 /* Init obdfilter private stats here */
2195                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2196                                      LPROCFS_CNTR_AVGMINMAX,
2197                                      "read_bytes", "bytes");
2198                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2199                                      LPROCFS_CNTR_AVGMINMAX,
2200                                      "write_bytes", "bytes");
2201                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2202                                      LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2203                                      "get_page", "usec");
2204                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2205                                      LPROCFS_CNTR_AVGMINMAX,
2206                                      "get_page_failures", "num");
2207                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2208                                      LPROCFS_CNTR_AVGMINMAX,
2209                                      "cache_access", "pages");
2210                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2211                                      LPROCFS_CNTR_AVGMINMAX,
2212                                      "cache_hit", "pages");
2213                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2214                                      LPROCFS_CNTR_AVGMINMAX,
2215                                      "cache_miss", "pages");
2216
2217                 lproc_filter_attach_seqstat(obd);
2218                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2219                                                         obd->obd_proc_entry,
2220                                                         NULL, NULL);
2221                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2222                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2223                         CERROR("error %d setting up lprocfs for %s\n",
2224                                rc, "exports");
2225                         obd->obd_proc_exports_entry = NULL;
2226                 }
2227         }
2228         if (obd->obd_proc_exports_entry)
2229                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2230                                    lprocfs_nid_stats_clear_read,
2231                                    lprocfs_nid_stats_clear_write, obd, NULL);
2232
2233         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2234                LUSTRE_CFG_BUFLEN(lcfg, 4));
2235         rc = filter_common_setup(obd, lcfg, (void *)addr);
2236         OBD_PAGE_FREE(page);
2237
2238         if (rc) {
2239                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2240                 lprocfs_free_per_client_stats(obd);
2241                 lprocfs_free_obd_stats(obd);
2242                 lprocfs_obd_cleanup(obd);
2243         }
2244
2245         return rc;
2246 }
2247
2248 static struct llog_operations filter_mds_ost_repl_logops;
2249
2250 static struct llog_operations filter_size_orig_logops = {
2251         .lop_setup   = llog_obd_origin_setup,
2252         .lop_cleanup = llog_obd_origin_cleanup,
2253         .lop_add     = llog_obd_origin_add
2254 };
2255
2256 static int filter_olg_fini(struct obd_llog_group *olg)
2257 {
2258         struct llog_ctxt *ctxt;
2259         int rc = 0, rc2 = 0;
2260         ENTRY;
2261
2262         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2263         if (ctxt)
2264                 rc = llog_cleanup(ctxt);
2265
2266         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2267         if (ctxt) {
2268                 rc2 = llog_cleanup(ctxt);
2269                 if (!rc)
2270                         rc = rc2;
2271         }
2272
2273         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2274         if (ctxt) {
2275                 rc2 = llog_cleanup(ctxt);
2276                 if (!rc)
2277                         rc = rc2;
2278         }
2279
2280         RETURN(rc);
2281 }
2282
2283 static int
2284 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2285                 struct obd_device *tgt)
2286 {
2287         int rc;
2288         ENTRY;
2289
2290         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2291                         &filter_mds_ost_repl_logops);
2292         if (rc)
2293                 GOTO(cleanup, rc);
2294
2295         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2296                         &filter_size_orig_logops);
2297         if (rc)
2298                 GOTO(cleanup, rc);
2299         EXIT;
2300 cleanup:
2301         if (rc)
2302                 filter_olg_fini(olg);
2303         return rc;
2304 }
2305
2306 /**
2307  * Init the default olg, which is embeded in the obd_device, for filter.
2308  */
2309 static int
2310 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2311                         struct obd_device *tgt)
2312 {
2313         struct filter_obd *filter = &obd->u.filter;
2314         struct llog_ctxt *ctxt;
2315         int rc;
2316         ENTRY;
2317
2318         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2319         if (!filter->fo_lcm)
2320                 RETURN(-ENOMEM);
2321
2322         filter_mds_ost_repl_logops = llog_client_ops;
2323         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2324         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2325         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2326
2327         rc = filter_olg_init(obd, olg, tgt);
2328         if (rc)
2329                 GOTO(cleanup_lcm, rc);
2330
2331         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2332                         &llog_lvfs_ops);
2333         if (rc)
2334                 GOTO(cleanup_olg, rc);
2335
2336         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2337         if (!ctxt) {
2338                 CERROR("Can't get ctxt for %p:%x\n", olg,
2339                        LLOG_MDS_OST_REPL_CTXT);
2340                 GOTO(cleanup_olg, rc = -ENODEV);
2341         }
2342         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2343         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2344         llog_ctxt_put(ctxt);
2345
2346         RETURN(0);
2347 cleanup_olg:
2348         filter_olg_fini(olg);
2349 cleanup_lcm:
2350         llog_recov_thread_fini(filter->fo_lcm, 1);
2351         filter->fo_lcm = NULL;
2352         return rc;
2353 }
2354
2355 static int
2356 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2357                  struct obd_device *tgt, int *index)
2358 {
2359         struct filter_obd *filter = &obd->u.filter;
2360         struct llog_ctxt *ctxt;
2361         int rc;
2362         ENTRY;
2363
2364         LASSERT(olg != NULL);
2365         if (olg == &obd->obd_olg)
2366                 return filter_default_olg_init(obd, olg, tgt);
2367
2368         LASSERT(filter->fo_lcm != NULL);
2369         rc = filter_olg_init(obd, olg, tgt);
2370         if (rc)
2371                 RETURN(rc);
2372         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2373         if (!ctxt) {
2374                 CERROR("Can't get ctxt for %p:%x\n", olg,
2375                        LLOG_MDS_OST_REPL_CTXT);
2376                 filter_olg_fini(olg);
2377                 RETURN(-ENODEV);
2378         }
2379         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2380         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2381         llog_ctxt_put(ctxt);
2382         RETURN(rc);
2383 }
2384
2385 static int filter_llog_finish(struct obd_device *obd, int count)
2386 {
2387         struct filter_obd *filter = &obd->u.filter;
2388         struct llog_ctxt *ctxt;
2389         ENTRY;
2390
2391         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2392         if (ctxt) {
2393                 /*
2394                  * Make sure that no cached llcds left in recov_thread.
2395                  * We actually do sync in disconnect time, but disconnect
2396                  * may not come being marked rq_no_resend = 1.
2397                  */
2398                 llog_sync(ctxt, NULL);
2399
2400                 /*
2401                  * Balance class_import_get() in llog_receptor_accept().
2402                  * This is safe to do, as llog is already synchronized
2403                  * and its import may go.
2404                  */
2405                 cfs_mutex_down(&ctxt->loc_sem);
2406                 if (ctxt->loc_imp) {
2407                         class_import_put(ctxt->loc_imp);
2408                         ctxt->loc_imp = NULL;
2409                 }
2410                 cfs_mutex_up(&ctxt->loc_sem);
2411                 llog_ctxt_put(ctxt);
2412         }
2413
2414         if (filter->fo_lcm) {
2415                 cfs_mutex_down(&ctxt->loc_sem);
2416                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2417                 filter->fo_lcm = NULL;
2418                 cfs_mutex_up(&ctxt->loc_sem);
2419         }
2420         RETURN(filter_olg_fini(&obd->obd_olg));
2421 }
2422
2423 /**
2424  * Find the group llog according to group index in the llog group list.
2425  */
2426 static struct obd_llog_group *
2427 filter_find_olg_internal(struct filter_obd *filter, int group)
2428 {
2429         struct obd_llog_group *olg;
2430
2431         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2432         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2433                 if (olg->olg_group == group)
2434                         RETURN(olg);
2435         }
2436         RETURN(NULL);
2437 }
2438
2439 /**
2440  * Find the group llog according to group index on the filter
2441  */
2442 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2443 {
2444         struct obd_llog_group *olg = NULL;
2445         struct filter_obd *filter;
2446
2447         filter = &obd->u.filter;
2448
2449         if (group == FILTER_GROUP_LLOG)
2450                 RETURN(&obd->obd_olg);
2451
2452         cfs_spin_lock(&filter->fo_llog_list_lock);
2453         olg = filter_find_olg_internal(filter, group);
2454         cfs_spin_unlock(&filter->fo_llog_list_lock);
2455
2456         RETURN(olg);
2457 }
2458 /**
2459  * Find the llog_group of the filter according to the group. If it can not
2460  * find, create the llog_group, which only happens when mds is being synced
2461  * with OST.
2462  */
2463 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2464 {
2465         struct obd_llog_group *olg = NULL;
2466         struct filter_obd *filter;
2467         int rc;
2468
2469         filter = &obd->u.filter;
2470
2471         if (group == FILTER_GROUP_LLOG)
2472                 RETURN(&obd->obd_olg);
2473
2474         cfs_spin_lock(&filter->fo_llog_list_lock);
2475         olg = filter_find_olg_internal(filter, group);
2476         if (olg) {
2477                 if (olg->olg_initializing) {
2478                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2479                 } else {
2480                         GOTO(out_unlock, olg);
2481                 }
2482         }
2483         OBD_ALLOC_PTR(olg);
2484         if (olg == NULL)
2485                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2486
2487         llog_group_init(olg, group);
2488         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2489         olg->olg_initializing = 1;
2490         cfs_spin_unlock(&filter->fo_llog_list_lock);
2491
2492         rc = obd_llog_init(obd, olg, obd, NULL);
2493         if (rc) {
2494                cfs_spin_lock(&filter->fo_llog_list_lock);
2495                cfs_list_del(&olg->olg_list);
2496                cfs_spin_unlock(&filter->fo_llog_list_lock);
2497                OBD_FREE_PTR(olg);
2498                GOTO(out, olg = ERR_PTR(-ENOMEM));
2499         }
2500         cfs_spin_lock(&filter->fo_llog_list_lock);
2501         olg->olg_initializing = 0;
2502         cfs_spin_unlock(&filter->fo_llog_list_lock);
2503         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2504               obd->obd_name, group, olg);
2505 out:
2506         RETURN(olg);
2507
2508 out_unlock:
2509         cfs_spin_unlock(&filter->fo_llog_list_lock);
2510         GOTO(out, olg);
2511 }
2512
2513 static int filter_llog_connect(struct obd_export *exp,
2514                                struct llogd_conn_body *body)
2515 {
2516         struct obd_device *obd = exp->exp_obd;
2517         struct llog_ctxt *ctxt;
2518         struct obd_llog_group *olg;
2519         int rc;
2520         ENTRY;
2521
2522         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2523                obd->obd_name, body->lgdc_logid.lgl_oid,
2524                body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2525
2526         olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2527         if (!olg) {
2528                 CERROR(" %s: can not find olg of group %d\n",
2529                        obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2530                 RETURN(-ENOENT);
2531         }
2532         llog_group_set_export(olg, exp);
2533
2534         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2535         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2536                  body->lgdc_ctxt_idx);
2537
2538         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2539               obd->obd_name, body->lgdc_logid.lgl_oid,
2540               body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2541
2542         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2543         obd->u.filter.fo_mds_ost_sync = 1;
2544         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2545         rc = llog_connect(ctxt, &body->lgdc_logid,
2546                           &body->lgdc_gen, NULL);
2547         llog_ctxt_put(ctxt);
2548         if (rc != 0)
2549                 CERROR("failed to connect rc %d idx %d\n", rc,
2550                                 body->lgdc_ctxt_idx);
2551
2552         RETURN(rc);
2553 }
2554
2555 static int filter_llog_preclean(struct obd_device *obd)
2556 {
2557         struct obd_llog_group *olg, *tmp;
2558         struct filter_obd *filter;
2559         cfs_list_t  remove_list;
2560         int rc = 0;
2561         ENTRY;
2562
2563         rc = obd_llog_finish(obd, 0);
2564         if (rc)
2565                 CERROR("failed to cleanup llogging subsystem\n");
2566
2567         filter = &obd->u.filter;
2568         CFS_INIT_LIST_HEAD(&remove_list);
2569
2570         cfs_spin_lock(&filter->fo_llog_list_lock);
2571         while (!cfs_list_empty(&filter->fo_llog_list)) {
2572                 olg = cfs_list_entry(filter->fo_llog_list.next,
2573                                      struct obd_llog_group, olg_list);
2574                 cfs_list_del(&olg->olg_list);
2575                 cfs_list_add(&olg->olg_list, &remove_list);
2576         }
2577         cfs_spin_unlock(&filter->fo_llog_list_lock);
2578
2579         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2580                 cfs_list_del_init(&olg->olg_list);
2581                 rc = filter_olg_fini(olg);
2582                 if (rc)
2583                         CERROR("failed to cleanup llogging subsystem for %u\n",
2584                                olg->olg_group);
2585                 OBD_FREE_PTR(olg);
2586         }
2587
2588         RETURN(rc);
2589 }
2590
2591 static int filter_precleanup(struct obd_device *obd,
2592                              enum obd_cleanup_stage stage)
2593 {
2594         int rc = 0;
2595         ENTRY;
2596
2597         switch(stage) {
2598         case OBD_CLEANUP_EARLY:
2599                 break;
2600         case OBD_CLEANUP_EXPORTS:
2601                 /* Stop recovery before namespace cleanup. */
2602                 target_recovery_fini(obd);
2603                 rc = filter_llog_preclean(obd);
2604                 break;
2605         }
2606         RETURN(rc);
2607 }
2608
2609 static int filter_cleanup(struct obd_device *obd)
2610 {
2611         struct filter_obd *filter = &obd->u.filter;
2612         ENTRY;
2613
2614         if (obd->obd_fail)
2615                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2616                               "will be preserved.\n", obd->obd_name);
2617
2618         obd_exports_barrier(obd);
2619         obd_zombie_barrier();
2620
2621         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2622         lprocfs_free_per_client_stats(obd);
2623         lprocfs_free_obd_stats(obd);
2624         lprocfs_obd_cleanup(obd);
2625         lquota_cleanup(filter_quota_interface_ref, obd);
2626
2627         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2628         obd->obd_namespace = NULL;
2629
2630         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2631
2632         if (obd->u.obt.obt_sb == NULL)
2633                 RETURN(0);
2634
2635         filter_post(obd);
2636
2637         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2638         shrink_dcache_sb(obd->u.obt.obt_sb);
2639
2640         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2641         obd->u.obt.obt_sb = NULL;
2642
2643         fsfilt_put_ops(obd->obd_fsops);
2644
2645         filter_iobuf_pool_done(filter);
2646
2647         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2648
2649         RETURN(0);
2650 }
2651
2652 static int filter_connect_internal(struct obd_export *exp,
2653                                    struct obd_connect_data *data,
2654                                    int reconnect)
2655 {
2656         struct filter_export_data *fed = &exp->exp_filter_data;
2657
2658         if (!data)
2659                 RETURN(0);
2660
2661         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2662                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2663                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2664                data->ocd_connect_flags, data->ocd_version,
2665                data->ocd_grant, data->ocd_index);
2666
2667         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2668                 CWARN("!!! This export (nid %s) used object group %d "
2669                        "earlier; now it's trying to use group %d!  This could "
2670                        "be a bug in the MDS. Please report to "
2671                        "http://bugzilla.lustre.org/\n",
2672                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2673                 RETURN(-EPROTO);
2674         }
2675         fed->fed_group = data->ocd_group;
2676
2677         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2678         exp->exp_connect_flags = data->ocd_connect_flags;
2679         data->ocd_version = LUSTRE_VERSION_CODE;
2680
2681         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2682         if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2683                   data->ocd_connect_flags & OBD_CONNECT_MDS))
2684                 RETURN(-EPROTO);
2685
2686         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2687                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2688                 obd_size left, want;
2689
2690                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2691                 left = filter_grant_space_left(exp);
2692                 want = data->ocd_grant;
2693                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2694                 data->ocd_grant = fed->fed_grant;
2695                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2696
2697                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2698                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2699                        exp->exp_client_uuid.uuid, exp,
2700                        data->ocd_grant, want, left);
2701
2702                 filter->fo_tot_granted_clients ++;
2703         }
2704
2705         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2706                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2707                 struct lr_server_data *lsd = filter->fo_fsd;
2708                 int index = le32_to_cpu(lsd->lsd_ost_index);
2709
2710                 if (!(lsd->lsd_feature_compat &
2711                       cpu_to_le32(OBD_COMPAT_OST))) {
2712                         /* this will only happen on the first connect */
2713                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2714                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2715                         /* sync is not needed here as filter_client_add will
2716                          * set exp_need_sync flag */
2717                         filter_update_server_data(exp->exp_obd,
2718                                                   filter->fo_rcvd_filp, lsd);
2719                 } else if (index != data->ocd_index) {
2720                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2721                                            " %u doesn't match actual OST index"
2722                                            " %u in last_rcvd file, bad "
2723                                            "configuration?\n",
2724                                            obd_export_nid2str(exp), index,
2725                                            data->ocd_index);
2726                         RETURN(-EBADF);
2727                 }
2728                 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2729                  * FIXME: We don't strictly need the COMPAT flag for that,
2730                  * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2731                  * FIXME: We needed it for the index, as index 0 is valid. */
2732         }
2733
2734         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2735                 data->ocd_brw_size = 65536;
2736         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2737                 data->ocd_brw_size = min(data->ocd_brw_size,
2738                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2739                 LASSERT(data->ocd_brw_size);
2740         }
2741
2742         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2743                 __u32 cksum_types = data->ocd_cksum_types;
2744
2745                 /* The client set in ocd_cksum_types the checksum types it
2746                  * supports. We have to mask off the algorithms that we don't
2747                  * support */
2748                 if (cksum_types & OBD_CKSUM_ALL)
2749                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2750                 else
2751                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2752
2753                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2754                                    "%x\n", exp->exp_obd->obd_name,
2755                                    obd_export_nid2str(exp), cksum_types,
2756                                    data->ocd_cksum_types);
2757         } else {
2758                 /* This client does not support OBD_CONNECT_CKSUM
2759                  * fall back to CRC32 */
2760                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2761                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2762                                    exp->exp_obd->obd_name,
2763                                    obd_export_nid2str(exp));
2764         }
2765
2766         RETURN(0);
2767 }
2768
2769 static int filter_reconnect(const struct lu_env *env,
2770                             struct obd_export *exp, struct obd_device *obd,
2771                             struct obd_uuid *cluuid,
2772                             struct obd_connect_data *data,
2773                             void *localdata)
2774 {
2775         int rc;
2776         ENTRY;
2777
2778         if (exp == NULL || obd == NULL || cluuid == NULL)
2779                 RETURN(-EINVAL);
2780
2781         rc = filter_connect_internal(exp, data, 1);
2782         if (rc == 0)
2783                 filter_export_stats_init(obd, exp, localdata);
2784
2785         RETURN(rc);
2786 }
2787
2788 /* nearly identical to mds_connect */
2789 static int filter_connect(const struct lu_env *env,
2790                           struct obd_export **exp, struct obd_device *obd,
2791                           struct obd_uuid *cluuid,
2792                           struct obd_connect_data *data, void *localdata)
2793 {
2794         struct lvfs_run_ctxt saved;
2795         struct lustre_handle conn = { 0 };
2796         struct obd_export *lexp;
2797         struct filter_export_data *fed;
2798         struct lsd_client_data *lcd = NULL;
2799         __u32 group;
2800         int rc;
2801         ENTRY;
2802
2803         if (exp == NULL || obd == NULL || cluuid == NULL)
2804                 RETURN(-EINVAL);
2805
2806         rc = class_connect(&conn, obd, cluuid);
2807         if (rc)
2808                 RETURN(rc);
2809         lexp = class_conn2export(&conn);
2810         LASSERT(lexp != NULL);
2811
2812         fed = &lexp->exp_filter_data;
2813
2814         rc = filter_connect_internal(lexp, data, 0);
2815         if (rc)
2816                 GOTO(cleanup, rc);
2817
2818         filter_export_stats_init(obd, lexp, localdata);
2819         if (obd->obd_replayable) {
2820                 OBD_ALLOC(lcd, sizeof(*lcd));
2821                 if (!lcd) {
2822                         CERROR("filter: out of memory for client data\n");
2823                         GOTO(cleanup, rc = -ENOMEM);
2824                 }
2825
2826                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2827                 fed->fed_lcd = lcd;
2828                 rc = filter_client_add(obd, lexp, -1);
2829                 if (rc)
2830                         GOTO(cleanup, rc);
2831         }
2832
2833         group = data->ocd_group;
2834
2835         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2836               obd->obd_name, lexp->exp_handle.h_cookie, group);
2837
2838         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2839         rc = filter_read_groups(obd, group, 1);
2840         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2841         if (rc != 0) {
2842                 CERROR("can't read group %u\n", group);
2843                 GOTO(cleanup, rc);
2844         }
2845
2846         GOTO(cleanup, rc);
2847
2848 cleanup:
2849         if (rc) {
2850                 if (lcd) {
2851                         OBD_FREE_PTR(lcd);
2852                         fed->fed_lcd = NULL;
2853                 }
2854                 class_disconnect(lexp);
2855                 *exp = NULL;
2856         } else {
2857                 *exp = lexp;
2858         }
2859
2860         RETURN(rc);
2861 }
2862
2863 /* Do extra sanity checks for grant accounting.  We do this at connect,
2864  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2865  * always get rid of it or turn it off when we know accounting is good. */
2866 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2867 {
2868         struct filter_export_data *fed;
2869         struct obd_export *exp;
2870         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2871         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2872         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2873
2874         if (cfs_list_empty(&obd->obd_exports))
2875                 return;
2876
2877         /* We don't want to do this for large machines that do lots of
2878            mounts or unmounts.  It burns... */
2879         if (obd->obd_num_exports > 100)
2880                 return;
2881
2882         cfs_spin_lock(&obd->obd_osfs_lock);
2883         cfs_spin_lock(&obd->obd_dev_lock);
2884         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2885                 int error = 0;
2886                 fed = &exp->exp_filter_data;
2887                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2888                     fed->fed_dirty < 0)
2889                         error = 1;
2890                 if (maxsize > 0) { /* we may not have done a statfs yet */
2891                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2892                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2893                                  exp->exp_client_uuid.uuid, exp,
2894                                  fed->fed_grant, fed->fed_pending, maxsize);
2895                         LASSERTF(fed->fed_dirty <= maxsize,
2896                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2897                                  exp->exp_client_uuid.uuid, exp,
2898                                  fed->fed_dirty, maxsize);
2899                 }
2900                 if (error)
2901                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2902                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2903                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2904                 else
2905                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2906                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2907                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2908                 tot_granted += fed->fed_grant + fed->fed_pending;
2909                 tot_pending += fed->fed_pending;
2910                 tot_dirty += fed->fed_dirty;
2911         }
2912         fo_tot_granted = obd->u.filter.fo_tot_granted;
2913         fo_tot_pending = obd->u.filter.fo_tot_pending;
2914         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2915         cfs_spin_unlock(&obd->obd_dev_lock);
2916         cfs_spin_unlock(&obd->obd_osfs_lock);
2917
2918         /* Do these assertions outside the spinlocks so we don't kill system */
2919         if (tot_granted != fo_tot_granted)
2920                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2921                        func, tot_granted, fo_tot_granted);
2922         if (tot_pending != fo_tot_pending)
2923                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2924                        func, tot_pending, fo_tot_pending);
2925         if (tot_dirty != fo_tot_dirty)
2926                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2927                        func, tot_dirty, fo_tot_dirty);
2928         if (tot_pending > tot_granted)
2929                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2930                        func, tot_pending, tot_granted);
2931         if (tot_granted > maxsize)
2932                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2933                        func, tot_granted, maxsize);
2934         if (tot_dirty > maxsize)
2935                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2936                        func, tot_dirty, maxsize);
2937 }
2938
2939 /* Remove this client from the grant accounting totals.  We also remove
2940  * the export from the obd device under the osfs and dev locks to ensure
2941  * that the filter_grant_sanity_check() calculations are always valid.
2942  * The client should do something similar when it invalidates its import. */
2943 static void filter_grant_discard(struct obd_export *exp)
2944 {
2945         struct obd_device *obd = exp->exp_obd;
2946         struct filter_obd *filter = &obd->u.filter;
2947         struct filter_export_data *fed = &exp->exp_filter_data;
2948
2949         cfs_spin_lock(&obd->obd_osfs_lock);
2950         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2951                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2952                  obd->obd_name, filter->fo_tot_granted,
2953                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2954         filter->fo_tot_granted -= fed->fed_grant;
2955         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2956                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2957                  obd->obd_name, filter->fo_tot_pending,
2958                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2959         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2960         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2961                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2962                  obd->obd_name, filter->fo_tot_dirty,
2963                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2964         filter->fo_tot_dirty -= fed->fed_dirty;
2965         fed->fed_dirty = 0;
2966         fed->fed_grant = 0;
2967
2968         cfs_spin_unlock(&obd->obd_osfs_lock);
2969 }
2970
2971 static int filter_destroy_export(struct obd_export *exp)
2972 {
2973         ENTRY;
2974
2975         if (exp->exp_filter_data.fed_pending)
2976                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2977                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2978                        exp, exp->exp_filter_data.fed_pending);
2979
2980         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2981
2982         target_destroy_export(exp);
2983         ldlm_destroy_export(exp);
2984
2985         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2986                 RETURN(0);
2987
2988
2989         if (!exp->exp_obd->obd_replayable)
2990                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2991
2992         filter_grant_discard(exp);
2993         filter_fmd_cleanup(exp);
2994
2995         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2996                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2997                 if (filter->fo_tot_granted_clients > 0)
2998                         filter->fo_tot_granted_clients --;
2999         }
3000
3001         if (!(exp->exp_flags & OBD_OPT_FORCE))
3002                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3003
3004         RETURN(0);
3005 }
3006
3007 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3008 {
3009         struct obd_llog_group *olg_min, *olg;
3010         struct filter_obd *filter;
3011         int worked = -1, group;
3012         struct llog_ctxt *ctxt;
3013         ENTRY;
3014
3015         filter = &obd->u.filter;
3016
3017         /* we can't sync log holding spinlock. also, we do not want to get
3018          * into livelock. so we do following: loop over MDS's exports in
3019          * group order and skip already synced llogs -bzzz */
3020         do {
3021                 /* look for group with min. number, but > worked */
3022                 olg_min = NULL;
3023                 group = 1 << 30;
3024                 cfs_spin_lock(&filter->fo_llog_list_lock);
3025                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3026                         if (olg->olg_group <= worked) {
3027                                 /* this group is already synced */
3028                                 continue;
3029                         }
3030                         if (group < olg->olg_group) {
3031                                 /* we have group with smaller number to sync */
3032                                 continue;
3033                         }
3034                         /* store current minimal group */
3035                         olg_min = olg;
3036                         group = olg->olg_group;
3037                 }
3038                 cfs_spin_unlock(&filter->fo_llog_list_lock);
3039
3040                 if (olg_min == NULL)
3041                         break;
3042
3043                 worked = olg_min->olg_group;
3044                 if (olg_min->olg_exp &&
3045                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3046                         int err;
3047                         ctxt = llog_group_get_ctxt(olg_min,
3048                                                    LLOG_MDS_OST_REPL_CTXT);
3049                         if (ctxt) {
3050                                 err = llog_sync(ctxt, olg_min->olg_exp);
3051                                 llog_ctxt_put(ctxt);
3052                                 if (err) {
3053                                         CERROR("error flushing logs to MDS: "
3054                                                "rc %d\n", err);
3055                                 }
3056                         }
3057                 }
3058         } while (olg_min != NULL);
3059 }
3060
3061 /* Also incredibly similar to mds_disconnect */
3062 static int filter_disconnect(struct obd_export *exp)
3063 {
3064         struct obd_device *obd = exp->exp_obd;
3065         int rc;
3066         ENTRY;
3067
3068         LASSERT(exp);
3069         class_export_get(exp);
3070
3071         if (!(exp->exp_flags & OBD_OPT_FORCE))
3072                 filter_grant_sanity_check(obd, __FUNCTION__);
3073         filter_grant_discard(exp);
3074
3075         /* Flush any remaining cancel messages out to the target */
3076         filter_sync_llogs(obd, exp);
3077
3078         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3079
3080         rc = server_disconnect_export(exp);
3081
3082         if (exp->exp_obd->obd_replayable)
3083                 filter_client_free(exp);
3084         else
3085                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3086
3087         class_export_put(exp);
3088         RETURN(rc);
3089 }
3090
3091 /* reverse import is changed, sync all cancels */
3092 static void filter_revimp_update(struct obd_export *exp)
3093 {
3094         ENTRY;
3095
3096         LASSERT(exp);
3097         class_export_get(exp);
3098
3099         /* flush any remaining cancel messages out to the target */
3100         filter_sync_llogs(exp->exp_obd, exp);
3101         class_export_put(exp);
3102         EXIT;
3103 }
3104
3105 static int filter_ping(struct obd_export *exp)
3106 {
3107         filter_fmd_expire(exp);
3108         return 0;
3109 }
3110
3111 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3112                                   const char *what, int quiet)
3113 {
3114         struct dentry *dchild = NULL;
3115         obd_gr group = 0;
3116
3117         if (oa->o_valid & OBD_MD_FLGROUP)
3118                 group = oa->o_gr;
3119
3120         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3121
3122         if (IS_ERR(dchild)) {
3123                 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3124                        what, group, oa->o_id);
3125                 RETURN(dchild);
3126         }
3127
3128         if (dchild->d_inode == NULL) {
3129                 if (!quiet)
3130                         CERROR("%s: %s on non-existent object: "LPU64"\n",
3131                                obd->obd_name, what, oa->o_id);
3132                 f_dput(dchild);
3133                 RETURN(ERR_PTR(-ENOENT));
3134         }
3135
3136         return dchild;
3137 }
3138
3139 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3140 {
3141         struct dentry *dentry = NULL;
3142         struct obd_device *obd;
3143         int rc = 0;
3144         ENTRY;
3145
3146         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3147         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3148                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3149         if (rc)
3150                 RETURN(rc);
3151
3152         obd = class_exp2obd(exp);
3153         if (obd == NULL) {
3154                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3155                 RETURN(-EINVAL);
3156         }
3157
3158         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3159         if (IS_ERR(dentry))
3160                 RETURN(PTR_ERR(dentry));
3161
3162         /* Limit the valid bits in the return data to what we actually use */
3163         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3164         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3165
3166         f_dput(dentry);
3167         RETURN(rc);
3168 }
3169
3170 /* this should be enabled/disabled in condition to enabled/disabled large
3171  * inodes (fast EAs) in backing store FS. */
3172 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3173                         void *handle, struct obdo *oa)
3174 {
3175         struct obd_device *obd = exp->exp_obd;
3176         int rc = 0;
3177         ENTRY;
3178
3179         if (oa->o_valid & OBD_MD_FLFID) {
3180                 struct filter_fid ff;
3181
3182                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3183                         oa->o_gr = 0;
3184                 /* packing fid and converting it to LE for storing into EA.
3185                  * Here ->o_stripe_idx should be filled by LOV and rest of
3186                  * fields - by client. */
3187                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3188                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3189                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3190                 ff.ff_objid = cpu_to_le64(oa->o_id);
3191                 ff.ff_group = cpu_to_le64(oa->o_gr);
3192
3193                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3194                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,