Whamcloud - gitweb
LU-85 build: update main bug report URL
[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 (c) 2001, 2010, Oracle and/or its affiliates. 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 obd_device_target *obt = &exp->exp_obd->u.obt;
124         struct tg_export_data *ted = &exp->exp_target_data;
125         struct lr_server_data *lsd = class_server_data(exp->exp_obd);
126         struct lsd_client_data *lcd;
127         __u64 last_rcvd;
128         loff_t off;
129         int err, log_pri = D_RPCTRACE;
130
131         /* Propagate error code. */
132         if (rc)
133                 RETURN(rc);
134
135         if (!exp->exp_obd->obd_replayable || oti == NULL)
136                 RETURN(rc);
137
138         cfs_mutex_down(&ted->ted_lcd_lock);
139         lcd = ted->ted_lcd;
140         /* if the export has already been disconnected, we have no last_rcvd slot,
141          * update server data with latest transno then */
142         if (lcd == NULL) {
143                 cfs_mutex_up(&ted->ted_lcd_lock);
144                 CWARN("commit transaction for disconnected client %s: rc %d\n",
145                       exp->exp_client_uuid.uuid, rc);
146                 err = filter_update_server_data(exp->exp_obd);
147                 RETURN(err);
148         }
149
150         /* we don't allocate new transnos for replayed requests */
151         cfs_spin_lock(&obt->obt_lut->lut_translock);
152         if (oti->oti_transno == 0) {
153                 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
154                 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
155                 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
156         } else {
157                 last_rcvd = oti->oti_transno;
158                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
159                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
160                 if (unlikely(last_rcvd < le64_to_cpu(lcd->lcd_last_transno))) {
161                         CERROR("Trying to overwrite bigger transno, on-disk: "
162                                LPU64", new: "LPU64"\n",
163                                le64_to_cpu(lcd->lcd_last_transno), last_rcvd);
164                         cfs_spin_lock(&exp->exp_lock);
165                         exp->exp_vbr_failed = 1;
166                         cfs_spin_unlock(&exp->exp_lock);
167                         cfs_spin_unlock(&obt->obt_lut->lut_translock);
168                         cfs_mutex_up(&ted->ted_lcd_lock);
169                         RETURN(-EOVERFLOW);
170                 }
171         }
172         oti->oti_transno = last_rcvd;
173
174         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
175         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
176         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
177         cfs_spin_unlock(&obt->obt_lut->lut_translock);
178
179         if (inode)
180                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
181
182         off = ted->ted_lr_off;
183         if (off <= 0) {
184                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
185                        ted->ted_lr_idx, ted->ted_lr_off);
186                 err = -EINVAL;
187         } else {
188                 class_export_cb_get(exp); /* released when the cb is called */
189                 if (!force_sync)
190                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
191                                                            last_rcvd,
192                                                            oti->oti_handle,
193                                                            filter_commit_cb,
194                                                            exp);
195
196                 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
197                                           lcd, sizeof(*lcd), &off,
198                                           force_sync | exp->exp_need_sync);
199                 if (force_sync)
200                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
201         }
202         if (err) {
203                 log_pri = D_ERROR;
204                 if (rc == 0)
205                         rc = err;
206         }
207
208         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
209                last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
210         cfs_mutex_up(&ted->ted_lcd_lock);
211         RETURN(rc);
212 }
213
214 void f_dput(struct dentry *dentry)
215 {
216         /* Can't go inside filter_ddelete because it can block */
217         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
218                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
219         LASSERT(atomic_read(&dentry->d_count) > 0);
220
221         dput(dentry);
222 }
223
224 static void init_brw_stats(struct brw_stats *brw_stats)
225 {
226         int i;
227         for (i = 0; i < BRW_LAST; i++)
228                 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
229 }
230
231 static int lprocfs_init_rw_stats(struct obd_device *obd,
232                                  struct lprocfs_stats **stats)
233 {
234         int num_stats;
235
236         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
237                                                         LPROC_FILTER_LAST - 1;
238         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
239         if (*stats == NULL)
240                 return -ENOMEM;
241
242         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
243         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
244                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
245         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
246                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
247
248         return(0);
249 }
250
251 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
252    plus the procfs overhead :( */
253 static int filter_export_stats_init(struct obd_device *obd,
254                                     struct obd_export *exp,
255                                     void *client_nid)
256 {
257         int rc, newnid = 0;
258         ENTRY;
259
260         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
261                 /* Self-export gets no proc entry */
262                 RETURN(0);
263
264         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
265         if (rc) {
266                 /* Mask error for already created
267                  * /proc entries */
268                 if (rc == -EALREADY)
269                         rc = 0;
270                 RETURN(rc);
271         }
272
273         if (newnid) {
274                 struct nid_stat *tmp = exp->exp_nid_stats;
275                 LASSERT(tmp != NULL);
276
277                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
278                 if (tmp->nid_brw_stats == NULL)
279                         GOTO(clean, rc = -ENOMEM);
280
281                 init_brw_stats(tmp->nid_brw_stats);
282                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
283                                         0644, &filter_per_nid_stats_fops,
284                                         exp->exp_nid_stats);
285                 if (rc)
286                         CWARN("Error adding the brw_stats file\n");
287
288                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
289                 if (rc)
290                         GOTO(clean, rc);
291
292                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
293                                             tmp->nid_stats);
294                 if (rc)
295                         GOTO(clean, rc);
296                 rc = lprocfs_nid_ldlm_stats_init(tmp);
297                 if (rc)
298                         GOTO(clean, rc);
299         }
300
301         RETURN(0);
302  clean:
303         return rc;
304 }
305
306 /* Add client data to the FILTER.  We use a bitmap to locate a free space
307  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
308  * Otherwise, we have just read the data from the last_rcvd file and
309  * we know its offset. */
310 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
311                              int cl_idx)
312 {
313         struct obd_device_target *obt = &obd->u.obt;
314         struct tg_export_data *ted = &exp->exp_target_data;
315         struct lr_server_data *lsd = class_server_data(obd);
316         unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
317         int new_client = (cl_idx == -1);
318
319         ENTRY;
320
321         LASSERT(bitmap != NULL);
322         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
323
324         /* Self-export */
325         if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
326                 RETURN(0);
327
328         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
329          * there's no need for extra complication here
330          */
331         if (new_client) {
332                 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
333         repeat:
334                 if (cl_idx >= LR_MAX_CLIENTS) {
335                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
336                                cl_idx);
337                         RETURN(-EOVERFLOW);
338                 }
339                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
340                         cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
341                                                         cl_idx);
342                         goto repeat;
343                 }
344         } else {
345                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
346                         CERROR("FILTER client %d: bit already set in bitmap!\n",
347                                cl_idx);
348                         LBUG();
349                 }
350         }
351
352         ted->ted_lr_idx = cl_idx;
353         ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
354                           cl_idx * le16_to_cpu(lsd->lsd_client_size);
355         cfs_init_mutex(&ted->ted_lcd_lock);
356         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
357
358         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
359                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
360
361         if (new_client) {
362                 struct lvfs_run_ctxt saved;
363                 loff_t off = ted->ted_lr_off;
364                 int rc;
365                 void *handle;
366
367                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
368                        ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
369
370                 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
371                         RETURN(-ENOSPC);
372
373                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
374                 /* Transaction needed to fix bug 1403 */
375                 handle = fsfilt_start(obd,
376                                       obt->obt_rcvd_filp->f_dentry->d_inode,
377                                       FSFILT_OP_SETATTR, NULL);
378                 if (IS_ERR(handle)) {
379                         rc = PTR_ERR(handle);
380                         CERROR("unable to start transaction: rc %d\n", rc);
381                 } else {
382                         ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
383                         exp->exp_last_request_time = cfs_time_current_sec();
384                         rc = fsfilt_add_journal_cb(obd, 0, handle,
385                                                    target_client_add_cb,
386                                                    class_export_cb_get(exp));
387                         if (rc == 0) {
388                                 cfs_spin_lock(&exp->exp_lock);
389                                 exp->exp_need_sync = 1;
390                                 cfs_spin_unlock(&exp->exp_lock);
391                         }
392                         rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
393                                                  ted->ted_lcd,
394                                                  sizeof(*ted->ted_lcd),
395                                                  &off, rc /* sync if no cb */);
396                         fsfilt_commit(obd,
397                                       obt->obt_rcvd_filp->f_dentry->d_inode,
398                                       handle, 0);
399                 }
400                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
401
402                 if (rc) {
403                         CERROR("error writing %s client idx %u: rc %d\n",
404                                LAST_RCVD, ted->ted_lr_idx, rc);
405                         RETURN(rc);
406                 }
407         }
408         RETURN(0);
409 }
410
411 static int filter_client_del(struct obd_export *exp)
412 {
413         struct tg_export_data *ted = &exp->exp_target_data;
414         struct obd_device_target *obt = &exp->exp_obd->u.obt;
415         struct lvfs_run_ctxt saved;
416         int rc;
417         loff_t off;
418         ENTRY;
419
420         if (ted->ted_lcd == NULL)
421                 RETURN(0);
422
423         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
424         if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
425                 GOTO(free, 0);
426
427         LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
428
429         off = ted->ted_lr_off;
430
431         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
432                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
433
434         /* Don't clear ted_lr_idx here as it is likely also unset.  At worst
435          * we leak a client slot that will be cleaned on the next recovery. */
436         if (off <= 0) {
437                 CERROR("%s: client idx %d has med_off %lld\n",
438                        exp->exp_obd->obd_name, ted->ted_lr_idx, off);
439                 GOTO(free, rc = -EINVAL);
440         }
441
442         /* Clear the bit _after_ zeroing out the client so we don't
443            race with filter_client_add and zero out new clients.*/
444         if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
445                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
446                        ted->ted_lr_idx);
447                 LBUG();
448         }
449
450         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
451         /* Make sure the server's last_transno is up to date.
452          * This should be done before zeroing client slot so last_transno will
453          * be in server data or in client data in case of failure */
454         filter_update_server_data(exp->exp_obd);
455
456         cfs_mutex_down(&ted->ted_lcd_lock);
457         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
458         rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
459                                  ted->ted_lcd,
460                                  sizeof(*ted->ted_lcd), &off, 0);
461         cfs_mutex_up(&ted->ted_lcd_lock);
462         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
463
464         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
465                "zero out client %s at idx %u/%llu in %s, rc %d\n",
466                ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
467                LAST_RCVD, rc);
468         EXIT;
469 free:
470         return 0;
471 }
472
473 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
474 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
475                                          struct filter_mod_data *fmd)
476 {
477         LASSERT_SPIN_LOCKED(&fed->fed_lock);
478         if (--fmd->fmd_refcount == 0) {
479                 /* XXX when we have persistent reservations and the handle
480                  * is stored herein we need to drop it here. */
481                 fed->fed_mod_count--;
482                 cfs_list_del(&fmd->fmd_list);
483                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
484         }
485 }
486
487 /* drop fmd reference, free it if last ref */
488 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
489 {
490         struct filter_export_data *fed;
491
492         if (fmd == NULL)
493                 return;
494
495         fed = &exp->exp_filter_data;
496         cfs_spin_lock(&fed->fed_lock);
497         filter_fmd_put_nolock(fed, fmd); /* caller reference */
498         cfs_spin_unlock(&fed->fed_lock);
499 }
500
501 /* expire entries from the end of the list if there are too many
502  * or they are too old */
503 static void filter_fmd_expire_nolock(struct filter_obd *filter,
504                                      struct filter_export_data *fed,
505                                      struct filter_mod_data *keep)
506 {
507         struct filter_mod_data *fmd, *tmp;
508
509         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
510                 if (fmd == keep)
511                         break;
512
513                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
514                     fed->fed_mod_count < filter->fo_fmd_max_num)
515                         break;
516
517                 cfs_list_del_init(&fmd->fmd_list);
518                 filter_fmd_put_nolock(fed, fmd); /* list reference */
519         }
520 }
521
522 void filter_fmd_expire(struct obd_export *exp)
523 {
524         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
525         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
526                                  &exp->exp_filter_data, NULL);
527         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
528 }
529
530 /* find specified objid, group in export fmd list.
531  * caller must hold fed_lock and take fmd reference itself */
532 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
533                                                 struct filter_export_data *fed,
534                                                 obd_id objid, obd_seq group)
535 {
536         struct filter_mod_data *found = NULL, *fmd;
537
538         LASSERT_SPIN_LOCKED(&fed->fed_lock);
539
540         cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
541                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
542                         found = fmd;
543                         cfs_list_del(&fmd->fmd_list);
544                         cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
545                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
546                         break;
547                 }
548         }
549
550         filter_fmd_expire_nolock(filter, fed, found);
551
552         return found;
553 }
554
555 /* Find fmd based on objid and group, or return NULL if not found. */
556 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
557                                         obd_id objid, obd_seq group)
558 {
559         struct filter_mod_data *fmd;
560
561         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
562         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
563                                      &exp->exp_filter_data, objid, group);
564         if (fmd)
565                 fmd->fmd_refcount++;    /* caller reference */
566         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
567
568         return fmd;
569 }
570
571 /* Find fmd based on objid and group, or create a new one if none is found.
572  * It is possible for this function to return NULL under memory pressure,
573  * or if objid = 0 is passed (which will only cause old entries to expire).
574  * Currently this is not fatal because any fmd state is transient and
575  * may also be freed when it gets sufficiently old. */
576 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
577                                        obd_id objid, obd_seq group)
578 {
579         struct filter_export_data *fed = &exp->exp_filter_data;
580         struct filter_mod_data *found = NULL, *fmd_new = NULL;
581
582         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
583
584         cfs_spin_lock(&fed->fed_lock);
585         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
586         if (fmd_new) {
587                 if (found == NULL) {
588                         cfs_list_add_tail(&fmd_new->fmd_list,
589                                           &fed->fed_mod_list);
590                         fmd_new->fmd_id = objid;
591                         fmd_new->fmd_gr = group;
592                         fmd_new->fmd_refcount++;   /* list reference */
593                         found = fmd_new;
594                         fed->fed_mod_count++;
595                 } else {
596                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
597                 }
598         }
599         if (found) {
600                 found->fmd_refcount++;          /* caller reference */
601                 found->fmd_expire = jiffies +
602                         exp->exp_obd->u.filter.fo_fmd_max_age;
603         }
604
605         cfs_spin_unlock(&fed->fed_lock);
606
607         return found;
608 }
609
610 #ifdef DO_FMD_DROP
611 /* drop fmd list reference so it will disappear when last reference is put.
612  * This isn't so critical because it would in fact only affect the one client
613  * that is doing the unlink and at worst we have an stale entry referencing
614  * an object that should never be used again. */
615 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
616 {
617         struct filter_mod_data *found = NULL;
618
619         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
620         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
621         if (found) {
622                 cfs_list_del_init(&found->fmd_list);
623                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
624         }
625         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
626 }
627 #else
628 #define filter_fmd_drop(exp, objid, group)
629 #endif
630
631 /* remove all entries from fmd list */
632 static void filter_fmd_cleanup(struct obd_export *exp)
633 {
634         struct filter_export_data *fed = &exp->exp_filter_data;
635         struct filter_mod_data *fmd = NULL, *tmp;
636
637         cfs_spin_lock(&fed->fed_lock);
638         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
639                 cfs_list_del_init(&fmd->fmd_list);
640                 filter_fmd_put_nolock(fed, fmd);
641         }
642         cfs_spin_unlock(&fed->fed_lock);
643 }
644
645 static int filter_init_export(struct obd_export *exp)
646 {
647         int rc;
648         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
649         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
650
651         cfs_spin_lock(&exp->exp_lock);
652         exp->exp_connecting = 1;
653         cfs_spin_unlock(&exp->exp_lock);
654         rc = lut_client_alloc(exp);
655         if (rc == 0)
656                 rc = ldlm_init_export(exp);
657
658         return rc;
659 }
660
661 static int filter_free_server_data(struct obd_device_target *obt)
662 {
663         lut_fini(NULL, obt->obt_lut);
664         OBD_FREE_PTR(obt->obt_lut);
665         return 0;
666 }
667
668 /* assumes caller is already in kernel ctxt */
669 int filter_update_server_data(struct obd_device *obd)
670 {
671         struct file *filp = obd->u.obt.obt_rcvd_filp;
672         struct lr_server_data *lsd = class_server_data(obd);
673         loff_t off = 0;
674         int rc;
675         ENTRY;
676
677         CDEBUG(D_INODE, "server uuid      : %s\n", lsd->lsd_uuid);
678         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
679                le64_to_cpu(lsd->lsd_last_transno));
680         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
681                le64_to_cpu(lsd->lsd_mount_count));
682
683         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
684         if (rc)
685                 CERROR("error writing lr_server_data: rc = %d\n", rc);
686
687         RETURN(rc);
688 }
689
690 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
691                              int force_sync)
692 {
693         struct filter_obd *filter = &obd->u.filter;
694         __u64 tmp;
695         loff_t off = 0;
696         int rc;
697         ENTRY;
698
699         if (filter->fo_last_objid_files[group] == NULL) {
700                 CERROR("Object seq "LPU64" not fully setup; not updating "
701                        "last_objid\n", group);
702                 RETURN(-EINVAL);
703         }
704
705         CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
706                obd->obd_name, filter->fo_last_objids[group], group);
707
708         tmp = cpu_to_le64(filter->fo_last_objids[group]);
709         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
710                                  &tmp, sizeof(tmp), &off, force_sync);
711         if (rc)
712                 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
713                        group, rc);
714         RETURN(rc);
715 }
716 extern int ost_handle(struct ptlrpc_request *req);
717 /* assumes caller has already in kernel ctxt */
718 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
719 {
720         struct filter_obd *filter = &obd->u.filter;
721         struct lr_server_data *lsd;
722         struct lsd_client_data *lcd = NULL;
723         struct inode *inode = filp->f_dentry->d_inode;
724         unsigned long last_rcvd_size = i_size_read(inode);
725         struct lu_target *lut;
726         __u64 mount_count;
727         __u32 start_epoch;
728         int cl_idx;
729         loff_t off = 0;
730         int rc;
731
732         /* ensure padding in the struct is the correct size */
733         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
734                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
735         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
736                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
737
738         /* allocate and initialize lu_target */
739         OBD_ALLOC_PTR(lut);
740         if (lut == NULL)
741                 RETURN(-ENOMEM);
742         rc = lut_init(NULL, lut, obd, NULL);
743         if (rc)
744                 GOTO(err_lut, rc);
745         lsd = class_server_data(obd);
746         if (last_rcvd_size == 0) {
747                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
748
749                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
750                 lsd->lsd_last_transno = 0;
751                 mount_count = lsd->lsd_mount_count = 0;
752                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
753                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
754                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
755                 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
756                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
757                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
758                  * MDS first connects and assigns the OST index number. */
759                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
760                                                         OBD_INCOMPAT_OST);
761         } else {
762                 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
763                 if (rc) {
764                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
765                                LAST_RCVD, rc);
766                         GOTO(err_lut, rc);
767                 }
768                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
769                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
770                                            "using the wrong disk %s. Were the "
771                                            "/dev/ assignments rearranged?\n",
772                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
773                         GOTO(err_lut, rc = -EINVAL);
774                 }
775                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
776                 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
777                 /* COMPAT_146 */
778                 /* Assume old last_rcvd format unless I_C_LR is set */
779                 if (!(lsd->lsd_feature_incompat &
780                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
781                         lsd->lsd_last_transno = lsd->lsd_compat14;
782                 /* end COMPAT_146 */
783                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
784                  * MDS first connects and assigns the OST index number. */
785                 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
786                                                          OBD_INCOMPAT_OST);
787         }
788
789         if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
790                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
791                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
792                        ~FILTER_INCOMPAT_SUPP);
793                 GOTO(err_lut, rc = -EINVAL);
794         }
795         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
796                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
797                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
798                        ~FILTER_ROCOMPAT_SUPP);
799                 /* Do something like remount filesystem read-only */
800                 GOTO(err_lut, rc = -EINVAL);
801         }
802
803         start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
804
805         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
806                obd->obd_name, start_epoch);
807         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
808                obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
809         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
810                obd->obd_name, mount_count + 1);
811         CDEBUG(D_INODE, "%s: server data size: %u\n",
812                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
813         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
814                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
815         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
816                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
817         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
818                obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
819         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
820                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
821                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
822                 le16_to_cpu(lsd->lsd_client_size));
823
824         if (!obd->obd_replayable) {
825                 CWARN("%s: recovery support OFF\n", obd->obd_name);
826                 GOTO(out, rc = 0);
827         }
828
829         OBD_ALLOC_PTR(lcd);
830         if (!lcd)
831                 GOTO(err_client, rc = -ENOMEM);
832
833         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
834              off < last_rcvd_size; cl_idx++) {
835                 __u64 last_rcvd;
836                 struct obd_export *exp;
837                 struct filter_export_data *fed;
838
839                 /* Don't assume off is incremented properly by
840                  * fsfilt_read_record(), in case sizeof(*lcd)
841                  * isn't the same as lsd->lsd_client_size.  */
842                 off = le32_to_cpu(lsd->lsd_client_start) +
843                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
844                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
845                 if (rc) {
846                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
847                                LAST_RCVD, cl_idx, off, rc);
848                         break; /* read error shouldn't cause startup to fail */
849                 }
850
851                 if (lcd->lcd_uuid[0] == '\0') {
852                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
853                                cl_idx);
854                         continue;
855                 }
856
857                 check_lcd(obd->obd_name, cl_idx, lcd);
858
859                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
860
861                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
862                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
863                        last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
864
865                 /* These exports are cleaned up by filter_disconnect(), so they
866                  * need to be set up like real exports as filter_connect() does.
867                  */
868                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
869                 if (IS_ERR(exp)) {
870                         if (PTR_ERR(exp) == -EALREADY) {
871                                 /* export already exists, zero out this one */
872                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
873                                 continue;
874                         }
875                         OBD_FREE_PTR(lcd);
876                         GOTO(err_client, rc = PTR_ERR(exp));
877                 }
878
879                 fed = &exp->exp_filter_data;
880                 *fed->fed_ted.ted_lcd = *lcd;
881                 fed->fed_group = 0; /* will be assigned at connect */
882                 filter_export_stats_init(obd, exp, NULL);
883                 rc = filter_client_add(obd, exp, cl_idx);
884                 /* can't fail for existing client */
885                 LASSERTF(rc == 0, "rc = %d\n", rc);
886
887                 /* VBR: set export last committed */
888                 exp->exp_last_committed = last_rcvd;
889                 cfs_spin_lock(&exp->exp_lock);
890                 exp->exp_connecting = 0;
891                 exp->exp_in_recovery = 0;
892                 cfs_spin_unlock(&exp->exp_lock);
893                 obd->obd_max_recoverable_clients++;
894                 class_export_put(exp);
895
896                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
897                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
898         }
899         OBD_FREE_PTR(lcd);
900
901         obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
902 out:
903         obd->u.obt.obt_mount_count = mount_count + 1;
904         lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
905
906         /* save it, so mount count and last_transno is current */
907         rc = filter_update_server_data(obd);
908         if (rc)
909                 GOTO(err_client, rc);
910
911         RETURN(0);
912
913 err_client:
914         class_disconnect_exports(obd);
915 err_lut:
916         filter_free_server_data(&obd->u.obt);
917         RETURN(rc);
918 }
919
920 static int filter_cleanup_groups(struct obd_device *obd)
921 {
922         struct filter_obd *filter = &obd->u.filter;
923         struct file *filp;
924         struct dentry *dentry;
925         int i, j;
926         ENTRY;
927
928         if (filter->fo_dentry_O_groups != NULL) {
929                 for (i = 0; i < filter->fo_group_count; i++) {
930                         dentry = filter->fo_dentry_O_groups[i];
931                         if (dentry != NULL)
932                                 f_dput(dentry);
933                 }
934                 OBD_FREE(filter->fo_dentry_O_groups,
935                          filter->fo_group_count *
936                          sizeof(*filter->fo_dentry_O_groups));
937                 filter->fo_dentry_O_groups = NULL;
938         }
939         if (filter->fo_last_objid_files != NULL) {
940                 for (i = 0; i < filter->fo_group_count; i++) {
941                         filp = filter->fo_last_objid_files[i];
942                         if (filp != NULL)
943                                 filp_close(filp, 0);
944                 }
945                 OBD_FREE(filter->fo_last_objid_files,
946                          filter->fo_group_count *
947                          sizeof(*filter->fo_last_objid_files));
948                 filter->fo_last_objid_files = NULL;
949         }
950         if (filter->fo_dentry_O_sub != NULL) {
951                 for (i = 0; i < filter->fo_group_count; i++) {
952                         for (j = 0; j < filter->fo_subdir_count; j++) {
953                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
954                                 if (dentry != NULL)
955                                         f_dput(dentry);
956                         }
957                 }
958                 OBD_FREE(filter->fo_dentry_O_sub,
959                          filter->fo_group_count *
960                          sizeof(*filter->fo_dentry_O_sub));
961                 filter->fo_dentry_O_sub = NULL;
962         }
963         if (filter->fo_last_objids != NULL) {
964                 OBD_FREE(filter->fo_last_objids,
965                          filter->fo_group_count *
966                          sizeof(*filter->fo_last_objids));
967                 filter->fo_last_objids = NULL;
968         }
969         if (filter->fo_dentry_O != NULL) {
970                 f_dput(filter->fo_dentry_O);
971                 filter->fo_dentry_O = NULL;
972         }
973         RETURN(0);
974 }
975
976 static int filter_update_last_group(struct obd_device *obd, int group)
977 {
978         struct filter_obd *filter = &obd->u.filter;
979         struct file *filp = NULL;
980         int last_group = 0, rc;
981         loff_t off = 0;
982         ENTRY;
983
984         if (group <= filter->fo_committed_group)
985                 RETURN(0);
986
987         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
988         if (IS_ERR(filp)) {
989                 rc = PTR_ERR(filp);
990                 filp = NULL;
991                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
992                 GOTO(cleanup, rc);
993         }
994
995         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
996         if (rc) {
997                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
998                 GOTO(cleanup, rc);
999         }
1000
1001         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1002                obd->obd_name, last_group, group);
1003
1004         off = 0;
1005         last_group = group;
1006         /* must be sync: bXXXX */
1007         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1008         if (rc) {
1009                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1010                 GOTO(cleanup, rc);
1011         }
1012
1013         filter->fo_committed_group = group;
1014 cleanup:
1015         if (filp)
1016                 filp_close(filp, 0);
1017         RETURN(rc);
1018 }
1019
1020 static int filter_read_group_internal(struct obd_device *obd, int group,
1021                                       int create)
1022 {
1023         struct filter_obd *filter = &obd->u.filter;
1024         __u64 *new_objids = NULL;
1025         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1026         struct dentry **new_groups = NULL;
1027         struct file **new_files = NULL;
1028         struct dentry *dentry;
1029         struct file *filp;
1030         int old_count = filter->fo_group_count, rc, stage = 0, i;
1031         char name[25];
1032         __u64 last_objid;
1033         loff_t off = 0;
1034         int len = group + 1;
1035
1036         snprintf(name, 24, "%d", group);
1037         name[24] = '\0';
1038
1039         if (!create) {
1040                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1041                                            strlen(name));
1042                 if (IS_ERR(dentry)) {
1043                         CERROR("Cannot lookup expected object group %d: %ld\n",
1044                                group, PTR_ERR(dentry));
1045                         RETURN(PTR_ERR(dentry));
1046                 }
1047         } else {
1048                 dentry = simple_mkdir(filter->fo_dentry_O,
1049                                       obd->u.obt.obt_vfsmnt, name, 0700, 1);
1050                 if (IS_ERR(dentry)) {
1051                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1052                                PTR_ERR(dentry));
1053                         RETURN(PTR_ERR(dentry));
1054                 }
1055         }
1056         stage = 1;
1057
1058         snprintf(name, 24, "O/%d/LAST_ID", group);
1059         name[24] = '\0';
1060         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1061         if (IS_ERR(filp)) {
1062                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1063                 GOTO(cleanup, rc = PTR_ERR(filp));
1064         }
1065         stage = 2;
1066
1067         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1068         if (rc) {
1069                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1070                 GOTO(cleanup, rc);
1071         }
1072
1073         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1074                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1075                 if (tmp_subdirs == NULL)
1076                         GOTO(cleanup, rc = -ENOMEM);
1077                 stage = 3;
1078
1079                 for (i = 0; i < filter->fo_subdir_count; i++) {
1080                         char dir[20];
1081                         snprintf(dir, sizeof(dir), "d%u", i);
1082
1083                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1084                                                               obd->u.obt.obt_vfsmnt,
1085                                                               dir, 0700, 1);
1086                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1087                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1088                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1089                                        group, dir, rc);
1090                                 GOTO(cleanup, rc);
1091                         }
1092
1093                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1094                                tmp_subdirs->dentry[i]);
1095                 }
1096         }
1097
1098         /* 'group' is an index; we need an array of length 'group + 1' */
1099         if (group + 1 > old_count) {
1100                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1101                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1102                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1103                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1104                 stage = 4;
1105                 if (new_objids == NULL || new_subdirs == NULL ||
1106                     new_groups == NULL || new_files == NULL)
1107                         GOTO(cleanup, rc = -ENOMEM);
1108
1109                 if (old_count) {
1110                         memcpy(new_objids, filter->fo_last_objids,
1111                                old_count * sizeof(*new_objids));
1112                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1113                                old_count * sizeof(*new_subdirs));
1114                         memcpy(new_groups, filter->fo_dentry_O_groups,
1115                                old_count * sizeof(*new_groups));
1116                         memcpy(new_files, filter->fo_last_objid_files,
1117                                old_count * sizeof(*new_files));
1118
1119                         OBD_FREE(filter->fo_last_objids,
1120                                  old_count * sizeof(*new_objids));
1121                         OBD_FREE(filter->fo_dentry_O_sub,
1122                                  old_count * sizeof(*new_subdirs));
1123                         OBD_FREE(filter->fo_dentry_O_groups,
1124                                  old_count * sizeof(*new_groups));
1125                         OBD_FREE(filter->fo_last_objid_files,
1126                                  old_count * sizeof(*new_files));
1127                 }
1128                 filter->fo_last_objids = new_objids;
1129                 filter->fo_dentry_O_sub = new_subdirs;
1130                 filter->fo_dentry_O_groups = new_groups;
1131                 filter->fo_last_objid_files = new_files;
1132                 filter->fo_group_count = len;
1133         }
1134
1135         filter->fo_dentry_O_groups[group] = dentry;
1136         filter->fo_last_objid_files[group] = filp;
1137         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1138                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1139                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1140         }
1141
1142         filter_update_last_group(obd, group);
1143
1144         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1145                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1146                 rc = filter_update_last_objid(obd, group, 1);
1147                 RETURN(rc);
1148         }
1149
1150         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1151         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1152                obd->obd_name, group, last_objid);
1153         RETURN(0);
1154  cleanup:
1155         switch (stage) {
1156         case 4:
1157                 if (new_objids != NULL)
1158                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1159                 if (new_subdirs != NULL)
1160                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1161                 if (new_groups != NULL)
1162                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1163                 if (new_files != NULL)
1164                         OBD_FREE(new_files, len * sizeof(*new_files));
1165         case 3:
1166                 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1167                         for (i = 0; i < filter->fo_subdir_count; i++) {
1168                                 if (tmp_subdirs->dentry[i] != NULL)
1169                                         dput(tmp_subdirs->dentry[i]);
1170                         }
1171                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1172                 }
1173         case 2:
1174                 filp_close(filp, 0);
1175         case 1:
1176                 dput(dentry);
1177         }
1178         RETURN(rc);
1179 }
1180
1181 static int filter_read_groups(struct obd_device *obd, int last_group,
1182                               int create)
1183 {
1184         struct filter_obd *filter = &obd->u.filter;
1185         int old_count, group, rc = 0;
1186
1187         cfs_down(&filter->fo_init_lock);
1188         old_count = filter->fo_group_count;
1189         for (group = old_count; group <= last_group; group++) {
1190                 rc = filter_read_group_internal(obd, group, create);
1191                 if (rc != 0)
1192                         break;
1193         }
1194         cfs_up(&filter->fo_init_lock);
1195         return rc;
1196 }
1197
1198 /* FIXME: object groups */
1199 static int filter_prep_groups(struct obd_device *obd)
1200 {
1201         struct filter_obd *filter = &obd->u.filter;
1202         struct dentry *O_dentry;
1203         struct file *filp;
1204         int    last_group, rc = 0, cleanup_phase = 0;
1205         loff_t off = 0;
1206         ENTRY;
1207
1208         O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1209                                 "O", 0700, 1);
1210         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1211         if (IS_ERR(O_dentry)) {
1212                 rc = PTR_ERR(O_dentry);
1213                 CERROR("cannot open/create O: rc = %d\n", rc);
1214                 GOTO(cleanup, rc);
1215         }
1216         filter->fo_dentry_O = O_dentry;
1217         cleanup_phase = 1; /* O_dentry */
1218
1219         /* we have to initialize all groups before first connections from
1220          * clients because they may send create/destroy for any group -bzzz */
1221         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1222         if (IS_ERR(filp)) {
1223                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1224                 GOTO(cleanup, rc = PTR_ERR(filp));
1225         }
1226         cleanup_phase = 2; /* filp */
1227
1228         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1229         if (rc) {
1230                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1231                 GOTO(cleanup, rc);
1232         }
1233
1234         if (off == 0)
1235                 last_group = FID_SEQ_OST_MDT0;
1236
1237         CWARN("%s: initialize groups [%u,%u]\n", obd->obd_name,
1238               FID_SEQ_OST_MDT0, last_group);
1239         filter->fo_committed_group = last_group;
1240         rc = filter_read_groups(obd, last_group, 1);
1241         if (rc)
1242                 GOTO(cleanup, rc);
1243
1244         filp_close(filp, 0);
1245         RETURN(0);
1246
1247  cleanup:
1248         switch (cleanup_phase) {
1249         case 2:
1250                 filp_close(filp, 0);
1251         case 1:
1252                 filter_cleanup_groups(obd);
1253                 f_dput(filter->fo_dentry_O);
1254                 filter->fo_dentry_O = NULL;
1255         default:
1256                 break;
1257         }
1258         return rc;
1259
1260 }
1261
1262 /* setup the object store with correct subdirectories */
1263 static int filter_prep(struct obd_device *obd)
1264 {
1265         struct lvfs_run_ctxt saved;
1266         struct filter_obd *filter = &obd->u.filter;
1267         struct file *file;
1268         struct inode *inode;
1269         int rc = 0;
1270         ENTRY;
1271
1272         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1273         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1274         if (!file || IS_ERR(file)) {
1275                 rc = PTR_ERR(file);
1276                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1277                        LAST_RCVD, rc);
1278                 GOTO(out, rc);
1279         }
1280         obd->u.obt.obt_rcvd_filp = file;
1281         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1282                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1283                        file->f_dentry->d_inode->i_mode);
1284                 GOTO(err_filp, rc = -ENOENT);
1285         }
1286
1287         inode = file->f_dentry->d_parent->d_inode;
1288         /* We use i_op->unlink directly in filter_vfs_unlink() */
1289         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1290                 CERROR("%s: filesystem does not support create/unlink ops\n",
1291                        obd->obd_name);
1292                 GOTO(err_filp, rc = -EOPNOTSUPP);
1293         }
1294
1295         rc = filter_init_server_data(obd, file);
1296         if (rc) {
1297                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1298                 GOTO(err_filp, rc);
1299         }
1300         LASSERT(obd->u.obt.obt_lut);
1301         target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1302
1303         /* open and create health check io file*/
1304         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1305         if (IS_ERR(file)) {
1306                 rc = PTR_ERR(file);
1307                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1308                        HEALTH_CHECK, rc);
1309                 GOTO(err_server_data, rc);
1310         }
1311         filter->fo_obt.obt_health_check_filp = file;
1312         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1313                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1314                        file->f_dentry->d_inode->i_mode);
1315                 GOTO(err_health_check, rc = -ENOENT);
1316         }
1317         rc = lvfs_check_io_health(obd, file);
1318         if (rc)
1319                 GOTO(err_health_check, rc);
1320
1321         rc = filter_prep_groups(obd);
1322         if (rc)
1323                 GOTO(err_health_check, rc);
1324 out:
1325         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1326
1327         return(rc);
1328
1329 err_health_check:
1330         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1331                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1332         filter->fo_obt.obt_health_check_filp = NULL;
1333 err_server_data:
1334         target_recovery_fini(obd);
1335         filter_free_server_data(&obd->u.obt);
1336 err_filp:
1337         if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1338                 CERROR("can't close %s after error\n", LAST_RCVD);
1339         obd->u.obt.obt_rcvd_filp = NULL;
1340         goto out;
1341 }
1342
1343 /* cleanup the filter: write last used object id to status file */
1344 static void filter_post(struct obd_device *obd)
1345 {
1346         struct lvfs_run_ctxt saved;
1347         struct filter_obd *filter = &obd->u.filter;
1348         int rc, i;
1349
1350         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1351          * best to start a transaction with h_sync, because we removed this
1352          * from lastobjid */
1353
1354         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1355         rc = filter_update_server_data(obd);
1356         if (rc)
1357                 CERROR("error writing server data: rc = %d\n", rc);
1358
1359         for (i = 0; i < filter->fo_group_count; i++) {
1360                 rc = filter_update_last_objid(obd, i,
1361                                 (i == filter->fo_group_count - 1));
1362                 if (rc)
1363                         CERROR("error writing group %d lastobjid: rc = %d\n",
1364                                i, rc);
1365         }
1366
1367         rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1368         obd->u.obt.obt_rcvd_filp = NULL;
1369         if (rc)
1370                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1371
1372         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1373         filter->fo_obt.obt_health_check_filp = NULL;
1374         if (rc)
1375                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1376
1377         filter_cleanup_groups(obd);
1378         filter_free_server_data(&obd->u.obt);
1379         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1380
1381         filter_free_capa_keys(filter);
1382         cleanup_capa_hash(filter->fo_capa_hash);
1383 }
1384
1385 static void filter_set_last_id(struct filter_obd *filter,
1386                                obd_id id, obd_seq group)
1387 {
1388         LASSERT(group <= filter->fo_group_count);
1389
1390         cfs_spin_lock(&filter->fo_objidlock);
1391         filter->fo_last_objids[group] = id;
1392         cfs_spin_unlock(&filter->fo_objidlock);
1393 }
1394
1395 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1396 {
1397         obd_id id;
1398         LASSERT(group <= filter->fo_group_count);
1399         LASSERT(filter->fo_last_objids != NULL);
1400
1401         /* FIXME: object groups */
1402         cfs_spin_lock(&filter->fo_objidlock);
1403         id = filter->fo_last_objids[group];
1404         cfs_spin_unlock(&filter->fo_objidlock);
1405         return id;
1406 }
1407
1408 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1409 {
1410         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1411         return 0;
1412 }
1413
1414 /* We never dget the object parent, so DON'T dput it either */
1415 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1416 {
1417         struct filter_obd *filter = &obd->u.filter;
1418         struct filter_subdirs *subdirs;
1419
1420         if (group >= filter->fo_group_count) /* FIXME: object groups */
1421                 return ERR_PTR(-EBADF);
1422
1423         if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1424                 return filter->fo_dentry_O_groups[group];
1425
1426         subdirs = &filter->fo_dentry_O_sub[group];
1427         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1428 }
1429
1430 /* We never dget the object parent, so DON'T dput it either */
1431 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1432                                   obd_id objid)
1433 {
1434         unsigned long now = jiffies;
1435         struct dentry *dparent = filter_parent(obd, group, objid);
1436         int rc;
1437
1438         if (IS_ERR(dparent))
1439                 return dparent;
1440         if (dparent == NULL)
1441                 return ERR_PTR(-ENOENT);
1442
1443         rc = filter_lock_dentry(obd, dparent);
1444         fsfilt_check_slow(obd, now, "parent lock");
1445         return rc ? ERR_PTR(rc) : dparent;
1446 }
1447
1448 /* We never dget the object parent, so DON'T dput it either */
1449 static void filter_parent_unlock(struct dentry *dparent)
1450 {
1451         UNLOCK_INODE_MUTEX(dparent->d_inode);
1452 }
1453
1454 /* How to get files, dentries, inodes from object id's.
1455  *
1456  * If dir_dentry is passed, the caller has already locked the parent
1457  * appropriately for this operation (normally a write lock).  If
1458  * dir_dentry is NULL, we do a read lock while we do the lookup to
1459  * avoid races with create/destroy and such changing the directory
1460  * internal to the filesystem code. */
1461 struct dentry *filter_fid2dentry(struct obd_device *obd,
1462                                  struct dentry *dir_dentry,
1463                                  obd_seq group, obd_id id)
1464 {
1465         struct dentry *dparent = dir_dentry;
1466         struct dentry *dchild;
1467         char name[32];
1468         int len;
1469         ENTRY;
1470
1471         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1472             obd->u.filter.fo_destroys_in_progress == 0) {
1473                 /* don't fail lookups for orphan recovery, it causes
1474                  * later LBUGs when objects still exist during precreate */
1475                 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1476                 RETURN(ERR_PTR(-ENOENT));
1477         }
1478         if (id == 0) {
1479                 CERROR("fatal: invalid object id 0\n");
1480                 RETURN(ERR_PTR(-ESTALE));
1481         }
1482
1483         len = sprintf(name, LPU64, id);
1484         if (dir_dentry == NULL) {
1485                 dparent = filter_parent_lock(obd, group, id);
1486                 if (IS_ERR(dparent)) {
1487                         CERROR("%s: error getting object "POSTID
1488                                " parent: rc %ld\n", obd->obd_name,
1489                                id, group, PTR_ERR(dparent));
1490                         RETURN(dparent);
1491                 }
1492         }
1493         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1494                dparent->d_name.len, dparent->d_name.name, name);
1495         /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1496         dchild = lookup_one_len(name, dparent, len);
1497         if (dir_dentry == NULL)
1498                 filter_parent_unlock(dparent);
1499         if (IS_ERR(dchild)) {
1500                 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1501                        obd->obd_name, id, group, PTR_ERR(dchild));
1502                 RETURN(dchild);
1503         }
1504
1505         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1506                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1507                        obd->obd_name, id, dchild->d_inode->i_ino);
1508                 f_dput(dchild);
1509                 RETURN(ERR_PTR(-ENOENT));
1510         }
1511
1512         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1513                name, dchild, atomic_read(&dchild->d_count));
1514
1515         LASSERT(atomic_read(&dchild->d_count) > 0);
1516
1517         RETURN(dchild);
1518 }
1519
1520 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1521                                   obd_id group, struct lustre_handle *lockh)
1522 {
1523         int flags = LDLM_AST_DISCARD_DATA, rc;
1524         struct ldlm_res_id res_id;
1525         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1526         ENTRY;
1527
1528         osc_build_res_name(objid, group, &res_id);
1529         /* Tell the clients that the object is gone now and that they should
1530          * throw away any cached pages. */
1531         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1532                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1533                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1534                                     lockh);
1535         if (rc != ELDLM_OK)
1536                 lockh->cookie = 0;
1537         RETURN(rc);
1538 }
1539
1540 static void filter_fini_destroy(struct obd_device *obd,
1541                                 struct lustre_handle *lockh)
1542 {
1543         if (lustre_handle_is_used(lockh))
1544                 ldlm_lock_decref(lockh, LCK_PW);
1545 }
1546
1547 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1548  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1549  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1550  * reverse.  Caller must take i_sem before starting the transaction and we
1551  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1552 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1553                       struct vfsmount *mnt)
1554 {
1555         int rc;
1556         ENTRY;
1557
1558         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1559          * itself we already hold dir->i_mutex for child create/unlink ops */
1560         LASSERT(dentry->d_inode != NULL);
1561         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1562         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1563
1564
1565         /* may_delete() */
1566         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1567                 GOTO(out, rc = -ENOENT);
1568
1569         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1570         if (rc)
1571                 GOTO(out, rc);
1572
1573         if (IS_APPEND(dir))
1574                 GOTO(out, rc = -EPERM);
1575
1576         /* check_sticky() */
1577         if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1578              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1579             IS_IMMUTABLE(dentry->d_inode))
1580                 GOTO(out, rc = -EPERM);
1581
1582         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1583          *       that was done because of journal_start (which is already done
1584          *       here) or some other ordering issue. */
1585         ll_vfs_dq_init(dir);
1586
1587         rc = ll_security_inode_unlink(dir, dentry, mnt);
1588         if (rc)
1589                 GOTO(out, rc);
1590
1591         rc = dir->i_op->unlink(dir, dentry);
1592 out:
1593         /* need to drop i_mutex before we lose inode reference */
1594         UNLOCK_INODE_MUTEX(dentry->d_inode);
1595         if (rc == 0)
1596                 d_delete(dentry);
1597
1598         RETURN(rc);
1599 }
1600
1601 /* Caller must hold LCK_PW on parent and push us into kernel context.
1602  * Caller must hold child i_mutex, we drop it always.
1603  * Caller is also required to ensure that dchild->d_inode exists. */
1604 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1605                                    obd_seq group, struct dentry *dparent,
1606                                    struct dentry *dchild)
1607 {
1608         struct inode *inode = dchild->d_inode;
1609         int rc;
1610
1611         /* There should be 2 references to the inode:
1612          *  1) taken by filter_prepare_destroy
1613          *  2) taken by filter_destroy */
1614         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1615                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1616                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1617                        (unsigned long)inode->i_nlink,
1618                        atomic_read(&inode->i_count));
1619         }
1620
1621         rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1622         if (rc)
1623                 CERROR("error unlinking objid %.*s: rc %d\n",
1624                        dchild->d_name.len, dchild->d_name.name, rc);
1625         return(rc);
1626 }
1627
1628 struct filter_intent_args {
1629         struct ldlm_lock **victim;
1630         __u64 size;
1631         int *liblustre;
1632 };
1633
1634 static enum interval_iter filter_intent_cb(struct interval_node *n,
1635                                            void *args)
1636 {
1637         struct ldlm_interval *node = (struct ldlm_interval *)n;
1638         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1639         __u64 size = arg->size;
1640         struct ldlm_lock **v = arg->victim;
1641         struct ldlm_lock *lck;
1642
1643         /* If the interval is lower than the current file size,
1644          * just break. */
1645         if (interval_high(n) <= size)
1646                 return INTERVAL_ITER_STOP;
1647
1648         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1649                 /* Don't send glimpse ASTs to liblustre clients.
1650                  * They aren't listening for them, and they do
1651                  * entirely synchronous I/O anyways. */
1652                 if (lck->l_export == NULL ||
1653                     lck->l_export->exp_libclient == 1)
1654                         continue;
1655
1656                 if (*arg->liblustre)
1657                         *arg->liblustre = 0;
1658
1659                 if (*v == NULL) {
1660                         *v = LDLM_LOCK_GET(lck);
1661                 } else if ((*v)->l_policy_data.l_extent.start <
1662                            lck->l_policy_data.l_extent.start) {
1663                         LDLM_LOCK_RELEASE(*v);
1664                         *v = LDLM_LOCK_GET(lck);
1665                 }
1666
1667                 /* the same policy group - every lock has the
1668                  * same extent, so needn't do it any more */
1669                 break;
1670         }
1671
1672         return INTERVAL_ITER_CONT;
1673 }
1674
1675 static int filter_intent_policy(struct ldlm_namespace *ns,
1676                                 struct ldlm_lock **lockp, void *req_cookie,
1677                                 ldlm_mode_t mode, int flags, void *data)
1678 {
1679         CFS_LIST_HEAD(rpc_list);
1680         struct ptlrpc_request *req = req_cookie;
1681         struct ldlm_lock *lock = *lockp, *l = NULL;
1682         struct ldlm_resource *res = lock->l_resource;
1683         ldlm_processing_policy policy;
1684         struct ost_lvb *res_lvb, *reply_lvb;
1685         struct ldlm_reply *rep;
1686         ldlm_error_t err;
1687         int idx, rc, tmpflags = 0, only_liblustre = 1;
1688         struct ldlm_interval_tree *tree;
1689         struct filter_intent_args arg;
1690         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1691                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1692                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1693         ENTRY;
1694
1695         policy = ldlm_get_processing_policy(res);
1696         LASSERT(policy != NULL);
1697         LASSERT(req != NULL);
1698
1699         rc = lustre_pack_reply(req, 3, repsize, NULL);
1700         if (rc)
1701                 RETURN(req->rq_status = rc);
1702
1703         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1704         LASSERT(rep != NULL);
1705
1706         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1707                                    sizeof(*reply_lvb));
1708         LASSERT(reply_lvb != NULL);
1709
1710         //fixup_handle_for_resent_req(req, lock, &lockh);
1711
1712         /* Call the extent policy function to see if our request can be
1713          * granted, or is blocked.
1714          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1715          * lock, and should not be granted if the lock will be blocked.
1716          */
1717
1718         LASSERT(ns == ldlm_res_to_ns(res));
1719         lock_res(res);
1720         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1721         check_res_locked(res);
1722
1723         /* FIXME: we should change the policy function slightly, to not make
1724          * this list at all, since we just turn around and free it */
1725         while (!cfs_list_empty(&rpc_list)) {
1726                 struct ldlm_lock *wlock =
1727                         cfs_list_entry(rpc_list.next, struct ldlm_lock,
1728                                        l_cp_ast);
1729                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1730                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1731                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1732                 cfs_list_del_init(&wlock->l_cp_ast);
1733                 LDLM_LOCK_RELEASE(wlock);
1734         }
1735
1736         /* The lock met with no resistance; we're finished. */
1737         if (rc == LDLM_ITER_CONTINUE) {
1738                 /* do not grant locks to the liblustre clients: they cannot
1739                  * handle ASTs robustly.  We need to do this while still
1740                  * holding lr_lock to avoid the lock remaining on the res_link
1741                  * list (and potentially being added to l_pending_list by an
1742                  * AST) when we are going to drop this lock ASAP. */
1743                 if (lock->l_export->exp_libclient ||
1744                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1745                         ldlm_resource_unlink_lock(lock);
1746                         err = ELDLM_LOCK_ABORTED;
1747                 } else {
1748                         err = ELDLM_LOCK_REPLACED;
1749                 }
1750                 unlock_res(res);
1751                 RETURN(err);
1752         }
1753
1754         /* Do not grant any lock, but instead send GL callbacks.  The extent
1755          * policy nicely created a list of all PW locks for us.  We will choose
1756          * the highest of those which are larger than the size in the LVB, if
1757          * any, and perform a glimpse callback. */
1758         res_lvb = res->lr_lvb_data;
1759         LASSERT(res_lvb != NULL);
1760         *reply_lvb = *res_lvb;
1761
1762         /*
1763          * lr_lock guarantees that no new locks are granted, and,
1764          * therefore, that res->lr_lvb_data cannot increase beyond the
1765          * end of already granted lock. As a result, it is safe to
1766          * check against "stale" reply_lvb->lvb_size value without
1767          * res->lr_lvb_sem.
1768          */
1769         arg.size = reply_lvb->lvb_size;
1770         arg.victim = &l;
1771         arg.liblustre = &only_liblustre;
1772         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1773                 tree = &res->lr_itree[idx];
1774                 if (tree->lit_mode == LCK_PR)
1775                         continue;
1776
1777                 interval_iterate_reverse(tree->lit_root,
1778                                          filter_intent_cb, &arg);
1779         }
1780         unlock_res(res);
1781
1782         /* There were no PW locks beyond the size in the LVB; finished. */
1783         if (l == NULL) {
1784                 if (only_liblustre) {
1785                         /* If we discovered a liblustre client with a PW lock,
1786                          * however, the LVB may be out of date!  The LVB is
1787                          * updated only on glimpse (which we don't do for
1788                          * liblustre clients) and cancel (which the client
1789                          * obviously has not yet done).  So if it has written
1790                          * data but kept the lock, the LVB is stale and needs
1791                          * to be updated from disk.
1792                          *
1793                          * Of course, this will all disappear when we switch to
1794                          * taking liblustre locks on the OST. */
1795                         ldlm_res_lvbo_update(res, NULL, 1);
1796                 }
1797                 RETURN(ELDLM_LOCK_ABORTED);
1798         }
1799
1800         /*
1801          * This check is for lock taken in filter_prepare_destroy() that does
1802          * not have l_glimpse_ast set. So the logic is: if there is a lock
1803          * with no l_glimpse_ast set, this object is being destroyed already.
1804          *
1805          * Hence, if you are grabbing DLM locks on the server, always set
1806          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1807          */
1808         if (l->l_glimpse_ast == NULL) {
1809                 /* We are racing with unlink(); just return -ENOENT */
1810                 rep->lock_policy_res1 = -ENOENT;
1811                 goto out;
1812         }
1813
1814         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1815         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1816
1817         lock_res(res);
1818         *reply_lvb = *res_lvb;
1819         unlock_res(res);
1820
1821  out:
1822         LDLM_LOCK_RELEASE(l);
1823
1824         RETURN(ELDLM_LOCK_ABORTED);
1825 }
1826
1827 /*
1828  * per-obd_device iobuf pool.
1829  *
1830  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1831  * allocations in write-path has to be minimized (see bug 5137).
1832  *
1833  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1834  * OST threads (see ost_thread_{init,done}()).
1835  *
1836  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1837  * at the OST layer there are only (potentially) multiple obd_device of type
1838  * unknown at the time of OST thread creation.
1839  *
1840  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1841  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1842  * it's very own iobuf.
1843  *
1844  * Functions below
1845  *
1846  *     filter_kiobuf_pool_init()
1847  *
1848  *     filter_kiobuf_pool_done()
1849  *
1850  *     filter_iobuf_get()
1851  *
1852  * operate on this array. They are "generic" in a sense that they don't depend
1853  * on actual type of iobuf's (the latter depending on Linux kernel version).
1854  */
1855
1856 /*
1857  * destroy pool created by filter_iobuf_pool_init
1858  */
1859 static void filter_iobuf_pool_done(struct filter_obd *filter)
1860 {
1861         struct filter_iobuf **pool;
1862         int i;
1863
1864         ENTRY;
1865
1866         pool = filter->fo_iobuf_pool;
1867         if (pool != NULL) {
1868                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1869                         if (pool[i] != NULL)
1870                                 filter_free_iobuf(pool[i]);
1871                 }
1872                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1873                 filter->fo_iobuf_pool = NULL;
1874         }
1875         EXIT;
1876 }
1877
1878 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1879 {
1880         struct filter_obd       *filter = &obd->u.filter;
1881         struct sptlrpc_rule_set  tmp_rset;
1882         int                      rc;
1883
1884         sptlrpc_rule_set_init(&tmp_rset);
1885         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1886         if (rc) {
1887                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1888                        obd->obd_name, rc);
1889                 return rc;
1890         }
1891
1892         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1893
1894         cfs_write_lock(&filter->fo_sptlrpc_lock);
1895         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1896         filter->fo_sptlrpc_rset = tmp_rset;
1897         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1898
1899         return 0;
1900 }
1901
1902 /*
1903  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1904  */
1905 static int filter_iobuf_pool_init(struct filter_obd *filter)
1906 {
1907         void **pool;
1908
1909         ENTRY;
1910
1911
1912         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1913                       GFP_KERNEL);
1914         if (filter->fo_iobuf_pool == NULL)
1915                 RETURN(-ENOMEM);
1916
1917         filter->fo_iobuf_count = OSS_THREADS_MAX;
1918
1919         RETURN(0);
1920 }
1921
1922 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1923  * many threads there will be so we allocate a large empty array and only
1924  * fill in those slots that are actually in use.
1925  * If we haven't allocated a pool entry for this thread before, do so now. */
1926 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1927 {
1928         int thread_id                    = (oti && oti->oti_thread) ?
1929                                            oti->oti_thread->t_id : -1;
1930         struct filter_iobuf  *pool       = NULL;
1931         struct filter_iobuf **pool_place = NULL;
1932
1933         if (thread_id >= 0) {
1934                 LASSERT(thread_id < filter->fo_iobuf_count);
1935                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1936         }
1937
1938         if (unlikely(pool == NULL)) {
1939                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1940                                           PTLRPC_MAX_BRW_PAGES);
1941                 if (pool_place != NULL)
1942                         *pool_place = pool;
1943         }
1944
1945         return pool;
1946 }
1947
1948 /* mount the file system (secretly).  lustre_cfg parameters are:
1949  * 1 = device
1950  * 2 = fstype
1951  * 3 = flags: failover=f, failout=n
1952  * 4 = mount options
1953  */
1954 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1955                         void *option)
1956 {
1957         struct filter_obd *filter = &obd->u.filter;
1958         struct vfsmount *mnt;
1959         struct lustre_mount_info *lmi;
1960         struct obd_uuid uuid;
1961         __u8 *uuid_ptr;
1962         char *str, *label;
1963         char ns_name[48];
1964         struct request_queue *q;
1965         int rc, i;
1966         ENTRY;
1967
1968         if (lcfg->lcfg_bufcount < 3 ||
1969             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1970             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1971                 RETURN(-EINVAL);
1972
1973         lmi = server_get_mount(obd->obd_name);
1974         if (lmi) {
1975                 /* We already mounted in lustre_fill_super.
1976                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1977                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1978                 mnt = lmi->lmi_mnt;
1979                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1980
1981                 /* gets recovery timeouts from mount data */
1982                 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1983                         obd->obd_recovery_timeout =
1984                                 lsi->lsi_lmd->lmd_recovery_time_soft;
1985                 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1986                         obd->obd_recovery_time_hard =
1987                                 lsi->lsi_lmd->lmd_recovery_time_hard;
1988         } else {
1989                 /* old path - used by lctl */
1990                 CERROR("Using old MDS mount method\n");
1991                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1992                                     MS_NOATIME|MS_NODIRATIME,
1993                                     lustre_cfg_string(lcfg, 1), option);
1994                 if (IS_ERR(mnt)) {
1995                         rc = PTR_ERR(mnt);
1996                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1997                                            lustre_cfg_string(lcfg, 1), rc);
1998                         RETURN(rc);
1999                 }
2000
2001                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2002         }
2003         if (IS_ERR(obd->obd_fsops))
2004                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2005
2006         rc = filter_iobuf_pool_init(filter);
2007         if (rc != 0)
2008                 GOTO(err_ops, rc);
2009
2010         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2011                 CERROR("%s: Underlying device is marked as read-only. "
2012                        "Setup failed\n", obd->obd_name);
2013                 GOTO(err_ops, rc = -EROFS);
2014         }
2015
2016         /* failover is the default */
2017         obd->obd_replayable = 1;
2018
2019         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2020                 str = lustre_cfg_string(lcfg, 3);
2021                 if (strchr(str, 'n')) {
2022                         CWARN("%s: recovery disabled\n", obd->obd_name);
2023                         obd->obd_replayable = 0;
2024                 }
2025         }
2026
2027         obd->u.obt.obt_vfsmnt = mnt;
2028         obd->u.obt.obt_sb = mnt->mnt_sb;
2029         obd->u.obt.obt_magic = OBT_MAGIC;
2030         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2031         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2032
2033         fsfilt_setup(obd, obd->u.obt.obt_sb);
2034
2035         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2036         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2037         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2038         obd->obd_lvfs_ctxt.fs = get_ds();
2039         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2040
2041         cfs_init_mutex(&filter->fo_init_lock);
2042         filter->fo_committed_group = 0;
2043         filter->fo_destroys_in_progress = 0;
2044         for (i = 0; i < 32; i++)
2045                 cfs_sema_init(&filter->fo_create_locks[i], 1);
2046
2047         cfs_spin_lock_init(&filter->fo_objidlock);
2048         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2049         cfs_sema_init(&filter->fo_alloc_lock, 1);
2050         init_brw_stats(&filter->fo_filter_stats);
2051         cfs_spin_lock_init(&filter->fo_flags_lock);
2052         filter->fo_read_cache = 1; /* enable read-only cache by default */
2053         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2054         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2055         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2056         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2057         filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2058         filter_slc_set(filter); /* initialize sync on lock cancel */
2059
2060         rc = filter_prep(obd);
2061         if (rc)
2062                 GOTO(err_ops, rc);
2063
2064         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2065         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2066
2067         filter->fo_fl_oss_capa = 1;
2068
2069         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2070         filter->fo_capa_hash = init_capa_hash();
2071         if (filter->fo_capa_hash == NULL)
2072                 GOTO(err_post, rc = -ENOMEM);
2073
2074         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2075         obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2076                                                 LDLM_NAMESPACE_SERVER,
2077                                                 LDLM_NAMESPACE_GREEDY,
2078                                                 LDLM_NS_TYPE_OST);
2079         if (obd->obd_namespace == NULL)
2080                 GOTO(err_post, rc = -ENOMEM);
2081         obd->obd_namespace->ns_lvbp = obd;
2082         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2083         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2084
2085         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2086                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2087
2088         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2089         if (rc) {
2090                 CERROR("failed to setup llogging subsystems\n");
2091                 GOTO(err_post, rc);
2092         }
2093
2094         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2095         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2096         /* do this after llog being initialized */
2097         filter_adapt_sptlrpc_conf(obd, 1);
2098
2099         rc = lquota_setup(filter_quota_interface_ref, obd);
2100         if (rc)
2101                 GOTO(err_post, rc);
2102
2103         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2104         if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2105             queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2106                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2107                               "for larger I/O requests: max_sectors = %u "
2108                               "could be up to max_hw_sectors=%u\n",
2109                               obd->obd_name, mnt->mnt_sb->s_id,
2110                               queue_max_sectors(q), queue_max_hw_sectors(q));
2111
2112         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2113         if (uuid_ptr != NULL) {
2114                 class_uuid_unparse(uuid_ptr, &uuid);
2115                 str = uuid.uuid;
2116         } else {
2117                 str = "no UUID";
2118         }
2119
2120         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2121         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2122                       obd->obd_name, label ?: str, lmi ? "on " : "",
2123                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2124                       obd->obd_replayable ? "enabled" : "disabled");
2125
2126         if (obd->obd_recovering)
2127                 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2128                               "or until %d client%s reconnect%s\n",
2129                               obd->obd_name,
2130                               obd->obd_recovery_timeout / 60,
2131                               obd->obd_recovery_timeout % 60,
2132                               obd->obd_max_recoverable_clients,
2133                               (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2134                               (obd->obd_max_recoverable_clients == 1) ? "s": "");
2135
2136
2137         RETURN(0);
2138
2139 err_post:
2140         filter_post(obd);
2141 err_ops:
2142         fsfilt_put_ops(obd->obd_fsops);
2143         filter_iobuf_pool_done(filter);
2144 err_mntput:
2145         server_put_mount(obd->obd_name, mnt);
2146         obd->u.obt.obt_sb = 0;
2147         return rc;
2148 }
2149
2150 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2151 {
2152         struct lprocfs_static_vars lvars;
2153         cfs_proc_dir_entry_t *entry;
2154         unsigned long addr;
2155         struct page *page;
2156         int rc;
2157         ENTRY;
2158
2159         CLASSERT(offsetof(struct obd_device, u.obt) ==
2160                  offsetof(struct obd_device, u.filter.fo_obt));
2161
2162         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2163                 RETURN(-EINVAL);
2164
2165         /* lprocfs must be setup before the filter so state can be safely added
2166          * to /proc incrementally as the filter is setup */
2167         lprocfs_filter_init_vars(&lvars);
2168         rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2169         if (rc) {
2170                 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2171                        obd->obd_name, rc);
2172                 RETURN(rc);
2173         }
2174
2175         rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2176         if (rc) {
2177                 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2178                        obd->obd_name, rc);
2179                 GOTO(obd_cleanup, rc);
2180         }
2181
2182         /* Init obdfilter private stats here */
2183         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2184                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2185         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2186                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2187         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2188                              LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2189                              "get_page", "usec");
2190         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2191                              LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2192         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2193                              LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2194         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2195                              LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2196         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2197                              LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2198
2199         rc = lproc_filter_attach_seqstat(obd);
2200         if (rc) {
2201                 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2202                 GOTO(free_obd_stats, rc);
2203         }
2204
2205         entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2206         if (IS_ERR(entry)) {
2207                 rc = PTR_ERR(entry);
2208                 CERROR("%s: error %d setting up lprocfs for %s\n",
2209                        obd->obd_name, rc, "exports");
2210                 GOTO(free_obd_stats, rc);
2211         }
2212         obd->obd_proc_exports_entry = entry;
2213
2214         entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2215                                    lprocfs_nid_stats_clear_read,
2216                                    lprocfs_nid_stats_clear_write, obd, NULL);
2217         if (IS_ERR(entry)) {
2218                 rc = PTR_ERR(entry);
2219                 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2220                        obd->obd_name, rc);
2221                 GOTO(free_obd_stats, rc);
2222         }
2223
2224         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2225         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2226         if (!page)
2227                 GOTO(remove_entry_clear, rc = -ENOMEM);
2228         addr = (unsigned long)cfs_page_address(page);
2229         clear_page((void *)addr);
2230         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2231                LUSTRE_CFG_BUFLEN(lcfg, 4));
2232         rc = filter_common_setup(obd, lcfg, (void *)addr);
2233         OBD_PAGE_FREE(page);
2234         if (rc) {
2235                 CERROR("%s: filter_common_setup failed: %d.\n",
2236                        obd->obd_name, rc);
2237                 GOTO(remove_entry_clear, rc);
2238         }
2239
2240         RETURN(0);
2241
2242 remove_entry_clear:
2243         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2244 free_obd_stats:
2245         lprocfs_free_obd_stats(obd);
2246 obd_cleanup:
2247         lprocfs_obd_cleanup(obd);
2248         return rc;
2249 }
2250
2251 static struct llog_operations filter_mds_ost_repl_logops;
2252
2253 static struct llog_operations filter_size_orig_logops = {
2254         .lop_setup   = llog_obd_origin_setup,
2255         .lop_cleanup = llog_obd_origin_cleanup,
2256         .lop_add     = llog_obd_origin_add
2257 };
2258
2259 static int filter_olg_fini(struct obd_llog_group *olg)
2260 {
2261         struct llog_ctxt *ctxt;
2262         int rc = 0, rc2 = 0;
2263         ENTRY;
2264
2265         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2266         if (ctxt)
2267                 rc = llog_cleanup(ctxt);
2268
2269         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2270         if (ctxt) {
2271                 rc2 = llog_cleanup(ctxt);
2272                 if (!rc)
2273                         rc = rc2;
2274         }
2275
2276         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2277         if (ctxt) {
2278                 rc2 = llog_cleanup(ctxt);
2279                 if (!rc)
2280                         rc = rc2;
2281         }
2282
2283         RETURN(rc);
2284 }
2285
2286 static int
2287 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2288                 struct obd_device *tgt)
2289 {
2290         int rc;
2291         ENTRY;
2292
2293         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2294                         &filter_mds_ost_repl_logops);
2295         if (rc)
2296                 GOTO(cleanup, rc);
2297
2298         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2299                         &filter_size_orig_logops);
2300         if (rc)
2301                 GOTO(cleanup, rc);
2302         EXIT;
2303 cleanup:
2304         if (rc)
2305                 filter_olg_fini(olg);
2306         return rc;
2307 }
2308
2309 /**
2310  * Init the default olg, which is embeded in the obd_device, for filter.
2311  */
2312 static int
2313 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2314                         struct obd_device *tgt)
2315 {
2316         struct filter_obd *filter = &obd->u.filter;
2317         struct llog_ctxt *ctxt;
2318         int rc;
2319         ENTRY;
2320
2321         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2322         if (!filter->fo_lcm)
2323                 RETURN(-ENOMEM);
2324
2325         filter_mds_ost_repl_logops = llog_client_ops;
2326         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2327         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2328         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2329
2330         rc = filter_olg_init(obd, olg, tgt);
2331         if (rc)
2332                 GOTO(cleanup_lcm, rc);
2333
2334         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2335                         &llog_lvfs_ops);
2336         if (rc)
2337                 GOTO(cleanup_olg, rc);
2338
2339         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2340         if (!ctxt) {
2341                 CERROR("Can't get ctxt for %p:%x\n", olg,
2342                        LLOG_MDS_OST_REPL_CTXT);
2343                 GOTO(cleanup_olg, rc = -ENODEV);
2344         }
2345         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2346         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2347         llog_ctxt_put(ctxt);
2348
2349         RETURN(0);
2350 cleanup_olg:
2351         filter_olg_fini(olg);
2352 cleanup_lcm:
2353         llog_recov_thread_fini(filter->fo_lcm, 1);
2354         filter->fo_lcm = NULL;
2355         return rc;
2356 }
2357
2358 static int
2359 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2360                  struct obd_device *tgt, int *index)
2361 {
2362         struct filter_obd *filter = &obd->u.filter;
2363         struct llog_ctxt *ctxt;
2364         int rc;
2365         ENTRY;
2366
2367         LASSERT(olg != NULL);
2368         if (olg == &obd->obd_olg)
2369                 return filter_default_olg_init(obd, olg, tgt);
2370
2371         LASSERT(filter->fo_lcm != NULL);
2372         rc = filter_olg_init(obd, olg, tgt);
2373         if (rc)
2374                 RETURN(rc);
2375         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2376         if (!ctxt) {
2377                 CERROR("Can't get ctxt for %p:%x\n", olg,
2378                        LLOG_MDS_OST_REPL_CTXT);
2379                 filter_olg_fini(olg);
2380                 RETURN(-ENODEV);
2381         }
2382         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2383         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2384         llog_ctxt_put(ctxt);
2385         RETURN(rc);
2386 }
2387
2388 static int filter_llog_finish(struct obd_device *obd, int count)
2389 {
2390         struct filter_obd *filter = &obd->u.filter;
2391         struct llog_ctxt *ctxt;
2392         ENTRY;
2393
2394         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2395         if (ctxt) {
2396                 /*
2397                  * Make sure that no cached llcds left in recov_thread.
2398                  * We actually do sync in disconnect time, but disconnect
2399                  * may not come being marked rq_no_resend = 1.
2400                  */
2401                 llog_sync(ctxt, NULL);
2402
2403                 /*
2404                  * Balance class_import_get() in llog_receptor_accept().
2405                  * This is safe to do, as llog is already synchronized
2406                  * and its import may go.
2407                  */
2408                 cfs_mutex_down(&ctxt->loc_sem);
2409                 if (ctxt->loc_imp) {
2410                         class_import_put(ctxt->loc_imp);
2411                         ctxt->loc_imp = NULL;
2412                 }
2413                 cfs_mutex_up(&ctxt->loc_sem);
2414                 llog_ctxt_put(ctxt);
2415         }
2416
2417         if (filter->fo_lcm) {
2418                 cfs_mutex_down(&ctxt->loc_sem);
2419                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2420                 filter->fo_lcm = NULL;
2421                 cfs_mutex_up(&ctxt->loc_sem);
2422         }
2423         RETURN(filter_olg_fini(&obd->obd_olg));
2424 }
2425
2426 /**
2427  * Find the group llog according to group index in the llog group list.
2428  */
2429 static struct obd_llog_group *
2430 filter_find_olg_internal(struct filter_obd *filter, int group)
2431 {
2432         struct obd_llog_group *olg;
2433
2434         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2435         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2436                 if (olg->olg_seq == group)
2437                         RETURN(olg);
2438         }
2439         RETURN(NULL);
2440 }
2441
2442 /**
2443  * Find the group llog according to group index on the filter
2444  */
2445 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2446 {
2447         struct obd_llog_group *olg = NULL;
2448         struct filter_obd *filter;
2449
2450         filter = &obd->u.filter;
2451
2452         if (group == FID_SEQ_LLOG)
2453                 RETURN(&obd->obd_olg);
2454
2455         cfs_spin_lock(&filter->fo_llog_list_lock);
2456         olg = filter_find_olg_internal(filter, group);
2457         cfs_spin_unlock(&filter->fo_llog_list_lock);
2458
2459         RETURN(olg);
2460 }
2461 /**
2462  * Find the llog_group of the filter according to the group. If it can not
2463  * find, create the llog_group, which only happens when mds is being synced
2464  * with OST.
2465  */
2466 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2467 {
2468         struct obd_llog_group *olg = NULL, *olg_new = NULL;
2469         struct filter_obd *filter;
2470         int rc;
2471
2472         filter = &obd->u.filter;
2473
2474         if (group == FID_SEQ_LLOG)
2475                 RETURN(&obd->obd_olg);
2476
2477         OBD_ALLOC_PTR(olg_new);
2478         if (olg_new == NULL)
2479                RETURN(ERR_PTR(-ENOMEM));
2480
2481         cfs_spin_lock(&filter->fo_llog_list_lock);
2482         olg = filter_find_olg_internal(filter, group);
2483         if (olg) {
2484                 if (olg->olg_initializing) {
2485                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2486                 } else {
2487                         GOTO(out_unlock, olg);
2488                 }
2489         } else {
2490                 /* set as the newly allocated one */
2491                 olg = olg_new;
2492                 olg_new = NULL;
2493         }
2494
2495         llog_group_init(olg, group);
2496         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2497         olg->olg_initializing = 1;
2498         cfs_spin_unlock(&filter->fo_llog_list_lock);
2499
2500         rc = obd_llog_init(obd, olg, obd, NULL);
2501         if (rc) {
2502                cfs_spin_lock(&filter->fo_llog_list_lock);
2503                cfs_list_del(&olg->olg_list);
2504                cfs_spin_unlock(&filter->fo_llog_list_lock);
2505                OBD_FREE_PTR(olg);
2506                GOTO(out, olg = ERR_PTR(-ENOMEM));
2507         }
2508         cfs_spin_lock(&filter->fo_llog_list_lock);
2509         olg->olg_initializing = 0;
2510         cfs_spin_unlock(&filter->fo_llog_list_lock);
2511         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2512               obd->obd_name, group, olg);
2513 out:
2514         RETURN(olg);
2515
2516 out_unlock:
2517         cfs_spin_unlock(&filter->fo_llog_list_lock);
2518         if (olg_new)
2519                OBD_FREE_PTR(olg_new);
2520         goto out;
2521 }
2522
2523 static int filter_llog_connect(struct obd_export *exp,
2524                                struct llogd_conn_body *body)
2525 {
2526         struct obd_device *obd = exp->exp_obd;
2527         struct llog_ctxt *ctxt;
2528         struct obd_llog_group *olg;
2529         int rc;
2530         ENTRY;
2531
2532         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2533                obd->obd_name, body->lgdc_logid.lgl_oid,
2534                body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2535
2536         olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2537         if (!olg) {
2538                 CERROR(" %s: can not find olg of group %d\n",
2539                        obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2540                 RETURN(-ENOENT);
2541         }
2542         llog_group_set_export(olg, exp);
2543
2544         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2545         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2546                  body->lgdc_ctxt_idx);
2547
2548         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2549               obd->obd_name, body->lgdc_logid.lgl_oid,
2550               body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2551
2552         cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2553         obd->u.filter.fo_mds_ost_sync = 1;
2554         cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2555         rc = llog_connect(ctxt, &body->lgdc_logid,
2556                           &body->lgdc_gen, NULL);
2557         llog_ctxt_put(ctxt);
2558         if (rc != 0)
2559                 CERROR("failed to connect rc %d idx %d\n", rc,
2560                                 body->lgdc_ctxt_idx);
2561
2562         RETURN(rc);
2563 }
2564
2565 static int filter_llog_preclean(struct obd_device *obd)
2566 {
2567         struct obd_llog_group *olg, *tmp;
2568         struct filter_obd *filter;
2569         cfs_list_t  remove_list;
2570         int rc = 0;
2571         ENTRY;
2572
2573         rc = obd_llog_finish(obd, 0);
2574         if (rc)
2575                 CERROR("failed to cleanup llogging subsystem\n");
2576
2577         filter = &obd->u.filter;
2578         CFS_INIT_LIST_HEAD(&remove_list);
2579
2580         cfs_spin_lock(&filter->fo_llog_list_lock);
2581         while (!cfs_list_empty(&filter->fo_llog_list)) {
2582                 olg = cfs_list_entry(filter->fo_llog_list.next,
2583                                      struct obd_llog_group, olg_list);
2584                 cfs_list_del(&olg->olg_list);
2585                 cfs_list_add(&olg->olg_list, &remove_list);
2586         }
2587         cfs_spin_unlock(&filter->fo_llog_list_lock);
2588
2589         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2590                 cfs_list_del_init(&olg->olg_list);
2591                 rc = filter_olg_fini(olg);
2592                 if (rc)
2593                         CERROR("failed to cleanup llogging subsystem for %u\n",
2594                                olg->olg_seq);
2595                 OBD_FREE_PTR(olg);
2596         }
2597
2598         RETURN(rc);
2599 }
2600
2601 static int filter_precleanup(struct obd_device *obd,
2602                              enum obd_cleanup_stage stage)
2603 {
2604         int rc = 0;
2605         ENTRY;
2606
2607         switch(stage) {
2608         case OBD_CLEANUP_EARLY:
2609                 break;
2610         case OBD_CLEANUP_EXPORTS:
2611                 /* Stop recovery before namespace cleanup. */
2612                 target_recovery_fini(obd);
2613                 rc = filter_llog_preclean(obd);
2614                 break;
2615         }
2616         RETURN(rc);
2617 }
2618
2619 static int filter_cleanup(struct obd_device *obd)
2620 {
2621         struct filter_obd *filter = &obd->u.filter;
2622         ENTRY;
2623
2624         if (obd->obd_fail)
2625                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2626                               "will be preserved.\n", obd->obd_name);
2627
2628         obd_exports_barrier(obd);
2629         obd_zombie_barrier();
2630
2631         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2632         lprocfs_free_per_client_stats(obd);
2633         lprocfs_free_obd_stats(obd);
2634         lprocfs_obd_cleanup(obd);
2635         lquota_cleanup(filter_quota_interface_ref, obd);
2636
2637         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2638         obd->obd_namespace = NULL;
2639
2640         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2641
2642         if (obd->u.obt.obt_sb == NULL)
2643                 RETURN(0);
2644
2645         filter_post(obd);
2646
2647         ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2648         shrink_dcache_sb(obd->u.obt.obt_sb);
2649
2650         server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2651         obd->u.obt.obt_sb = NULL;
2652
2653         fsfilt_put_ops(obd->obd_fsops);
2654
2655         filter_iobuf_pool_done(filter);
2656
2657         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2658
2659         RETURN(0);
2660 }
2661
2662 static int filter_connect_internal(struct obd_export *exp,
2663                                    struct obd_connect_data *data,
2664                                    int reconnect)
2665 {
2666         struct filter_export_data *fed = &exp->exp_filter_data;
2667
2668         if (!data)
2669                 RETURN(0);
2670
2671         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2672                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2673                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2674                data->ocd_connect_flags, data->ocd_version,
2675                data->ocd_grant, data->ocd_index);
2676
2677         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2678                 CWARN("!!! This export (nid %s) used object group %d "
2679                        "earlier; now it's trying to use group %d!  This could "
2680                        "be a bug in the MDS. Please report to "
2681                        "http://bugs.whamcloud.com/\n",
2682                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2683                 RETURN(-EPROTO);
2684         }
2685         fed->fed_group = data->ocd_group;
2686
2687         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2688         exp->exp_connect_flags = data->ocd_connect_flags;
2689         data->ocd_version = LUSTRE_VERSION_CODE;
2690
2691         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2692         if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2693                   data->ocd_connect_flags & OBD_CONNECT_MDS))
2694                 RETURN(-EPROTO);
2695
2696         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2697                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2698                 obd_size left, want;
2699
2700                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2701                 left = filter_grant_space_left(exp);
2702                 want = data->ocd_grant;
2703                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2704                 data->ocd_grant = fed->fed_grant;
2705                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2706
2707                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2708                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2709                        exp->exp_client_uuid.uuid, exp,
2710                        data->ocd_grant, want, left);
2711
2712                 filter->fo_tot_granted_clients ++;
2713         }
2714
2715         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2716                 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2717                 int index = le32_to_cpu(lsd->lsd_ost_index);
2718
2719                 if (!(lsd->lsd_feature_compat &
2720                       cpu_to_le32(OBD_COMPAT_OST))) {
2721                         /* this will only happen on the first connect */
2722                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2723                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2724                         /* sync is not needed here as filter_client_add will
2725                          * set exp_need_sync flag */
2726                         filter_update_server_data(exp->exp_obd);
2727                 } else if (index != data->ocd_index) {
2728                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2729                                            " %u doesn't match actual OST index"
2730                                            " %u in last_rcvd file, bad "
2731                                            "configuration?\n",
2732                                            obd_export_nid2str(exp), index,
2733                                            data->ocd_index);
2734                         RETURN(-EBADF);
2735                 }
2736                 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2737                  * FIXME: We don't strictly need the COMPAT flag for that,
2738                  * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2739                  * FIXME: We needed it for the index, as index 0 is valid. */
2740         }
2741
2742         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2743                 data->ocd_brw_size = 65536;
2744         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2745                 data->ocd_brw_size = min(data->ocd_brw_size,
2746                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2747                 if (data->ocd_brw_size == 0) {
2748                         CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2749                                " ocd_version: %x ocd_grant: %d ocd_index: %u "
2750                                "ocd_brw_size is unexpectedly zero, "
2751                                "network data corruption?"
2752                                "Refusing connection of this client\n",
2753                                 exp->exp_obd->obd_name,
2754                                 exp->exp_client_uuid.uuid,
2755                                 exp, data->ocd_connect_flags, data->ocd_version,
2756                                 data->ocd_grant, data->ocd_index);
2757                         RETURN(-EPROTO);
2758                 }
2759         }
2760
2761         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2762                 __u32 cksum_types = data->ocd_cksum_types;
2763
2764                 /* The client set in ocd_cksum_types the checksum types it
2765                  * supports. We have to mask off the algorithms that we don't
2766                  * support */
2767                 if (cksum_types & OBD_CKSUM_ALL)
2768                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2769                 else
2770                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2771
2772                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2773                                    "%x\n", exp->exp_obd->obd_name,
2774                                    obd_export_nid2str(exp), cksum_types,
2775                                    data->ocd_cksum_types);
2776         } else {
2777                 /* This client does not support OBD_CONNECT_CKSUM
2778                  * fall back to CRC32 */
2779                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2780                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2781                                    exp->exp_obd->obd_name,
2782                                    obd_export_nid2str(exp));
2783         }
2784
2785         RETURN(0);
2786 }
2787
2788 static int filter_reconnect(const struct lu_env *env,
2789                             struct obd_export *exp, struct obd_device *obd,
2790                             struct obd_uuid *cluuid,
2791                             struct obd_connect_data *data,
2792                             void *localdata)
2793 {
2794         int rc;
2795         ENTRY;
2796
2797         if (exp == NULL || obd == NULL || cluuid == NULL)
2798                 RETURN(-EINVAL);
2799
2800         rc = filter_connect_internal(exp, data, 1);
2801         if (rc == 0)
2802                 filter_export_stats_init(obd, exp, localdata);
2803
2804         RETURN(rc);
2805 }
2806
2807 /* nearly identical to mds_connect */
2808 static int filter_connect(const struct lu_env *env,
2809                           struct obd_export **exp, struct obd_device *obd,
2810                           struct obd_uuid *cluuid,
2811                           struct obd_connect_data *data, void *localdata)
2812 {
2813         struct lvfs_run_ctxt saved;
2814         struct lustre_handle conn = { 0 };
2815         struct obd_export *lexp;
2816         __u32 group;
2817         int rc;
2818         ENTRY;
2819
2820         if (exp == NULL || obd == NULL || cluuid == NULL)
2821                 RETURN(-EINVAL);
2822
2823         rc = class_connect(&conn, obd, cluuid);
2824         if (rc)
2825                 RETURN(rc);
2826         lexp = class_conn2export(&conn);
2827         LASSERT(lexp != NULL);
2828
2829         rc = filter_connect_internal(lexp, data, 0);
2830         if (rc)
2831                 GOTO(cleanup, rc);
2832
2833         filter_export_stats_init(obd, lexp, localdata);
2834         if (obd->obd_replayable) {
2835                 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2836                 LASSERT(lcd);
2837                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2838                 rc = filter_client_add(obd, lexp, -1);
2839                 if (rc)
2840                         GOTO(cleanup, rc);
2841         }
2842
2843         group = data->ocd_group;
2844
2845         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2846               obd->obd_name, lexp->exp_handle.h_cookie, group);
2847
2848         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2849         rc = filter_read_groups(obd, group, 1);
2850         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2851         if (rc != 0) {
2852                 CERROR("can't read group %u\n", group);
2853                 GOTO(cleanup, rc);
2854         }
2855
2856         GOTO(cleanup, rc);
2857
2858 cleanup:
2859         if (rc) {
2860                 class_disconnect(lexp);
2861                 *exp = NULL;
2862         } else {
2863                 *exp = lexp;
2864         }
2865
2866         RETURN(rc);
2867 }
2868
2869 /* Do extra sanity checks for grant accounting.  We do this at connect,
2870  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2871  * always get rid of it or turn it off when we know accounting is good. */
2872 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2873 {
2874         struct filter_export_data *fed;
2875         struct obd_export *exp;
2876         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2877         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2878         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2879
2880         if (cfs_list_empty(&obd->obd_exports))
2881                 return;
2882
2883         /* We don't want to do this for large machines that do lots of
2884            mounts or unmounts.  It burns... */
2885         if (obd->obd_num_exports > 100)
2886                 return;
2887
2888         cfs_spin_lock(&obd->obd_osfs_lock);
2889         cfs_spin_lock(&obd->obd_dev_lock);
2890         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2891                 int error = 0;
2892                 fed = &exp->exp_filter_data;
2893                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2894                     fed->fed_dirty < 0)
2895                         error = 1;
2896                 if (maxsize > 0) { /* we may not have done a statfs yet */
2897                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2898                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2899                                  exp->exp_client_uuid.uuid, exp,
2900                                  fed->fed_grant, fed->fed_pending, maxsize);
2901                         LASSERTF(fed->fed_dirty <= maxsize,
2902                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2903                                  exp->exp_client_uuid.uuid, exp,
2904                                  fed->fed_dirty, maxsize);
2905                 }
2906                 if (error)
2907                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2908                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2909                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2910                 else
2911                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2912                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2913                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2914                 tot_granted += fed->fed_grant + fed->fed_pending;
2915                 tot_pending += fed->fed_pending;
2916                 tot_dirty += fed->fed_dirty;
2917         }
2918         fo_tot_granted = obd->u.filter.fo_tot_granted;
2919         fo_tot_pending = obd->u.filter.fo_tot_pending;
2920         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2921         cfs_spin_unlock(&obd->obd_dev_lock);
2922         cfs_spin_unlock(&obd->obd_osfs_lock);
2923
2924         /* Do these assertions outside the spinlocks so we don't kill system */
2925         if (tot_granted != fo_tot_granted)
2926                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2927                        func, tot_granted, fo_tot_granted);
2928         if (tot_pending != fo_tot_pending)
2929                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2930                        func, tot_pending, fo_tot_pending);
2931         if (tot_dirty != fo_tot_dirty)
2932                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2933                        func, tot_dirty, fo_tot_dirty);
2934         if (tot_pending > tot_granted)
2935                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2936                        func, tot_pending, tot_granted);
2937         if (tot_granted > maxsize)
2938                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2939                        func, tot_granted, maxsize);
2940         if (tot_dirty > maxsize)
2941                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2942                        func, tot_dirty, maxsize);
2943 }
2944
2945 /* Remove this client from the grant accounting totals.  We also remove
2946  * the export from the obd device under the osfs and dev locks to ensure
2947  * that the filter_grant_sanity_check() calculations are always valid.
2948  * The client should do something similar when it invalidates its import. */
2949 static void filter_grant_discard(struct obd_export *exp)
2950 {
2951         struct obd_device *obd = exp->exp_obd;
2952         struct filter_obd *filter = &obd->u.filter;
2953         struct filter_export_data *fed = &exp->exp_filter_data;
2954
2955         cfs_spin_lock(&obd->obd_osfs_lock);
2956         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2957                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2958                  obd->obd_name, filter->fo_tot_granted,
2959                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2960         filter->fo_tot_granted -= fed->fed_grant;
2961         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2962                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2963                  obd->obd_name, filter->fo_tot_pending,
2964                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2965         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2966         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2967                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2968                  obd->obd_name, filter->fo_tot_dirty,
2969                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2970         filter->fo_tot_dirty -= fed->fed_dirty;
2971         fed->fed_dirty = 0;
2972         fed->fed_grant = 0;
2973
2974         cfs_spin_unlock(&obd->obd_osfs_lock);
2975 }
2976
2977 static int filter_destroy_export(struct obd_export *exp)
2978 {
2979         struct filter_export_data *fed = &exp->exp_filter_data;
2980         ENTRY;
2981
2982         if (fed->fed_pending)
2983                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2984                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2985                        exp, fed->fed_pending);
2986
2987         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2988
2989         target_destroy_export(exp);
2990         ldlm_destroy_export(exp);
2991         lut_client_free(exp);
2992
2993         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2994                 RETURN(0);
2995
2996         if (!exp->exp_obd->obd_replayable)
2997                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2998
2999         filter_grant_discard(exp);
3000         filter_fmd_cleanup(exp);
3001
3002         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3003                 struct filter_obd *filter = &exp->exp_obd->u.filter;
3004                 if (filter->fo_tot_granted_clients > 0)
3005                         filter->fo_tot_granted_clients --;
3006         }
3007
3008         if (!(exp->exp_flags & OBD_OPT_FORCE))
3009                 filter_grant_sanity_check(exp->exp_obd, __func__);
3010
3011         RETURN(0);
3012 }
3013
3014 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3015 {
3016         struct obd_llog_group *olg_min, *olg;
3017         struct filter_obd *filter;
3018         int worked = -1, group;
3019         struct llog_ctxt *ctxt;
3020         ENTRY;
3021
3022         filter = &obd->u.filter;
3023
3024         /* we can't sync log holding spinlock. also, we do not want to get
3025          * into livelock. so we do following: loop over MDS's exports in
3026          * group order and skip already synced llogs -bzzz */
3027         do {
3028                 /* look for group with min. number, but > worked */
3029                 olg_min = NULL;
3030                 group = 1 << 30;
3031                 cfs_spin_lock(&filter->fo_llog_list_lock);
3032                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3033                         if (olg->olg_seq <= worked) {
3034                                 /* this group is already synced */
3035                                 continue;
3036                         }
3037                         if (group < olg->olg_seq) {
3038                                 /* we have group with smaller number to sync */
3039                                 continue;
3040                         }
3041                         /* store current minimal group */
3042                         olg_min = olg;
3043                         group = olg->olg_seq;
3044                 }
3045                 cfs_spin_unlock(&filter->fo_llog_list_lock);
3046
3047                 if (olg_min == NULL)
3048                         break;
3049
3050                 worked = olg_min->olg_seq;
3051                 if (olg_min->olg_exp &&
3052                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3053                         int err;
3054                         ctxt = llog_group_get_ctxt(olg_min,
3055                                                    LLOG_MDS_OST_REPL_CTXT);
3056                         if (ctxt) {
3057                                 err = llog_sync(ctxt, olg_min->olg_exp);
3058                                 llog_ctxt_put(ctxt);
3059                                 if (err) {
3060                                         CERROR("error flushing logs to MDS: "
3061                                                "rc %d\n", err);
3062                                 }
3063                         }
3064                 }
3065         } while (olg_min != NULL);
3066 }
3067
3068 /* Also incredibly similar to mds_disconnect */
3069 static int filter_disconnect(struct obd_export *exp)
3070 {
3071         struct obd_device *obd = exp->exp_obd;
3072         int rc;
3073         ENTRY;
3074
3075         LASSERT(exp);
3076         class_export_get(exp);
3077
3078         if (!(exp->exp_flags & OBD_OPT_FORCE))
3079                 filter_grant_sanity_check(obd, __func__);
3080         filter_grant_discard(exp);
3081
3082         /* Flush any remaining cancel messages out to the target */
3083         filter_sync_llogs(obd, exp);
3084
3085         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3086
3087         rc = server_disconnect_export(exp);
3088
3089         /* Do not erase record for recoverable client. */
3090         if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3091                 filter_client_del(exp);
3092         else
3093                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3094
3095         class_export_put(exp);
3096         RETURN(rc);
3097 }
3098
3099 /* reverse import is changed, sync all cancels */
3100 static void filter_revimp_update(struct obd_export *exp)
3101 {
3102         ENTRY;
3103
3104         LASSERT(exp);
3105         class_export_get(exp);
3106
3107         /* flush any remaining cancel messages out to the target */
3108         filter_sync_llogs(exp->exp_obd, exp);
3109         class_export_put(exp);
3110         EXIT;
3111 }
3112
3113 static int filter_ping(struct obd_export *exp)
3114 {
3115         filter_fmd_expire(exp);
3116         return 0;
3117 }
3118
3119 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3120                                   const char *what, int quiet)
3121 {
3122         struct dentry *dchild = NULL;
3123
3124         dchild = filter_fid2dentry(obd, NULL,  ostid->oi_seq, ostid->oi_id);
3125
3126         if (IS_ERR(dchild)) {
3127                 CERROR("%s error looking up object: "POSTID"\n",
3128                        what, ostid->oi_id, ostid->oi_seq);
3129                 RETURN(dchild);
3130         }
3131
3132         if (dchild->d_inode == NULL) {
3133                 if (!quiet)
3134                         CERROR("%s: %s on non-existent object: "POSTID" \n",
3135                                obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3136                 f_dput(dchild);
3137                 RETURN(ERR_PTR(-ENOENT));
3138         }
3139
3140         return dchild;
3141 }
3142
3143 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3144 {
3145         struct dentry *dentry = NULL;
3146         struct obd_device *obd;
3147         int rc = 0;
3148         ENTRY;
3149
3150         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3151                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3152         if (rc)
3153                 RETURN(rc);
3154
3155         obd = class_exp2obd(exp);
3156         if (obd == NULL) {
3157                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3158                 RETURN(-EINVAL);
3159         }
3160
3161         dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3162         if (IS_ERR(dentry))
3163                 RETURN(PTR_ERR(dentry));
3164
3165         /* Limit the valid bits in the return data to what we actually use */
3166         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3167         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3168
3169         f_dput(dentry);
3170         RETURN(rc);
3171 }
3172
3173 /* this should be enabled/disabled in condition to enabled/disabled large
3174  * inodes (fast EAs) in backing store FS. */
3175 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3176                         void *handle, struct obdo *oa)
3177 {
3178         struct obd_device *obd = exp->exp_obd;
3179         int rc = 0;
3180         ENTRY;
3181
3182         if (oa->o_valid & OBD_MD_FLFID) {
3183                 struct filter_fid ff;
3184
3185                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3186                         oa->o_seq = 0;
3187                 /* packing fid and converting it to LE for storing into EA.
3188                  * Here ->o_stripe_idx should be filled by LOV and rest of
3189                  * fields - by client. */
3190                 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3191                 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3192                 /* XXX: we are ignoring o_parent_ver here, since this should
3193                  *      be the same for all objects in this fileset. */
3194                 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3195                 ff.ff_objid = cpu_to_le64(oa->o_id);
3196                 ff.ff_seq = cpu_to_le64(oa->o_seq);
3197
3198                 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3199                        LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3200                        oa->o_seq);
3201
3202                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");