Whamcloud - gitweb
LU-106 procfs: many proc entries are not accessed safely
[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  * Copyright (c) 2011 Whamcloud, Inc.
34  */
35 /*
36  * This file is part of Lustre, http://www.lustre.org/
37  * Lustre is a trademark of Sun Microsystems, Inc.
38  *
39  * lustre/obdfilter/filter.c
40  *
41  * Author: Peter Braam <braam@clusterfs.com>
42  * Author: Andreas Dilger <adilger@clusterfs.com>
43  */
44
45 /*
46  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
47  *            (which need to get journal_lock, may block if journal full).
48  *
49  * Invariant: Call filter_start_transno() before any journal ops to avoid the
50  *            same deadlock problem.  We can (and want) to get rid of the
51  *            transno sem in favour of the dir/inode i_mutex to avoid single
52  *            threaded operation on the OST.
53  */
54
55 #define DEBUG_SUBSYSTEM S_FILTER
56
57 #ifndef AUTOCONF_INCLUDED
58 #include <linux/config.h>
59 #endif
60 #include <linux/module.h>
61 #include <linux/fs.h>
62 #include <linux/dcache.h>
63 #include <linux/init.h>
64 #include <linux/version.h>
65 #include <linux/sched.h>
66 #include <linux/mount.h>
67 #include <linux/buffer_head.h>
68
69 #include <obd_cksum.h>
70 #include <obd_class.h>
71 #include <obd_lov.h>
72 #include <lustre_dlm.h>
73 #include <lustre_fsfilt.h>
74 #include <lprocfs_status.h>
75 #include <lustre_log.h>
76 #include <libcfs/list.h>
77 #include <lustre_disk.h>
78 #include <lustre_quota.h>
79 #include <linux/slab.h>
80 #include <lustre_param.h>
81 #include <lustre/ll_fiemap.h>
82
83 #include "filter_internal.h"
84
85 static struct lvfs_callback_ops filter_lvfs_ops;
86 cfs_mem_cache_t *ll_fmd_cachep;
87
88 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
89                              void *cb_data, int error)
90 {
91         struct obd_export *exp = cb_data;
92         LASSERT(exp->exp_obd == obd);
93         obd_transno_commit_cb(obd, transno, exp, error);
94         class_export_cb_put(exp);
95 }
96
97 int filter_version_get_check(struct obd_export *exp,
98                              struct obd_trans_info *oti, struct inode *inode)
99 {
100         __u64 curr_version;
101
102         if (inode == NULL || oti == NULL)
103                 RETURN(0);
104
105         curr_version = fsfilt_get_version(exp->exp_obd, inode);
106         if ((__s64)curr_version == -EOPNOTSUPP)
107                 RETURN(0);
108         /* VBR: version is checked always because costs nothing */
109         if (oti->oti_pre_version != 0 &&
110             oti->oti_pre_version != curr_version) {
111                 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
112                        oti->oti_pre_version, curr_version);
113                 cfs_spin_lock(&exp->exp_lock);
114                 exp->exp_vbr_failed = 1;
115                 cfs_spin_unlock(&exp->exp_lock);
116                 RETURN (-EOVERFLOW);
117         }
118         oti->oti_pre_version = curr_version;
119         RETURN(0);
120 }
121
122 /* Assumes caller has already pushed us into the kernel context. */
123 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
124                           struct obd_trans_info *oti, int rc, int force_sync)
125 {
126         struct obd_device_target *obt = &exp->exp_obd->u.obt;
127         struct tg_export_data *ted = &exp->exp_target_data;
128         struct lr_server_data *lsd = class_server_data(exp->exp_obd);
129         struct lsd_client_data *lcd;
130         __u64 last_rcvd;
131         loff_t off;
132         int err, log_pri = D_RPCTRACE;
133
134         /* Propagate error code. */
135         if (rc)
136                 RETURN(rc);
137
138         if (!exp->exp_obd->obd_replayable || oti == NULL)
139                 RETURN(rc);
140
141         cfs_mutex_down(&ted->ted_lcd_lock);
142         lcd = ted->ted_lcd;
143         /* if the export has already been disconnected, we have no last_rcvd slot,
144          * update server data with latest transno then */
145         if (lcd == NULL) {
146                 cfs_mutex_up(&ted->ted_lcd_lock);
147                 CWARN("commit transaction for disconnected client %s: rc %d\n",
148                       exp->exp_client_uuid.uuid, rc);
149                 err = filter_update_server_data(exp->exp_obd);
150                 RETURN(err);
151         }
152
153         /* we don't allocate new transnos for replayed requests */
154         cfs_spin_lock(&obt->obt_lut->lut_translock);
155         if (oti->oti_transno == 0) {
156                 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
157                 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
158                 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
159         } else {
160                 last_rcvd = oti->oti_transno;
161                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
162                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
163                 if (unlikely(last_rcvd < le64_to_cpu(lcd->lcd_last_transno))) {
164                         CERROR("Trying to overwrite bigger transno, on-disk: "
165                                LPU64", new: "LPU64"\n",
166                                le64_to_cpu(lcd->lcd_last_transno), last_rcvd);
167                         cfs_spin_lock(&exp->exp_lock);
168                         exp->exp_vbr_failed = 1;
169                         cfs_spin_unlock(&exp->exp_lock);
170                         cfs_spin_unlock(&obt->obt_lut->lut_translock);
171                         cfs_mutex_up(&ted->ted_lcd_lock);
172                         RETURN(-EOVERFLOW);
173                 }
174         }
175         oti->oti_transno = last_rcvd;
176
177         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
178         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
179         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
180         cfs_spin_unlock(&obt->obt_lut->lut_translock);
181
182         if (inode)
183                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
184
185         off = ted->ted_lr_off;
186         if (off <= 0) {
187                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
188                        ted->ted_lr_idx, ted->ted_lr_off);
189                 err = -EINVAL;
190         } else {
191                 class_export_cb_get(exp); /* released when the cb is called */
192                 if (!force_sync)
193                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
194                                                            last_rcvd,
195                                                            oti->oti_handle,
196                                                            filter_commit_cb,
197                                                            exp);
198
199                 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
200                                           lcd, sizeof(*lcd), &off,
201                                           force_sync | exp->exp_need_sync);
202                 if (force_sync)
203                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
204         }
205         if (err) {
206                 log_pri = D_ERROR;
207                 if (rc == 0)
208                         rc = err;
209         }
210
211         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
212                last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
213         cfs_mutex_up(&ted->ted_lcd_lock);
214         RETURN(rc);
215 }
216
217 void f_dput(struct dentry *dentry)
218 {
219         /* Can't go inside filter_ddelete because it can block */
220         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
221                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
222         LASSERT(atomic_read(&dentry->d_count) > 0);
223
224         dput(dentry);
225 }
226
227 static void init_brw_stats(struct brw_stats *brw_stats)
228 {
229         int i;
230         for (i = 0; i < BRW_LAST; i++)
231                 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
232 }
233
234 static int lprocfs_init_rw_stats(struct obd_device *obd,
235                                  struct lprocfs_stats **stats)
236 {
237         int num_stats;
238
239         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
240                                                         LPROC_FILTER_LAST - 1;
241         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
242         if (*stats == NULL)
243                 return -ENOMEM;
244
245         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
246         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
247                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
248         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
249                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
250
251         return(0);
252 }
253
254 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
255    plus the procfs overhead :( */
256 static int filter_export_stats_init(struct obd_device *obd,
257                                     struct obd_export *exp,
258                                     void *client_nid)
259 {
260         int rc, newnid = 0;
261         ENTRY;
262
263         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
264                 /* Self-export gets no proc entry */
265                 RETURN(0);
266
267         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
268         if (rc) {
269                 /* Mask error for already created
270                  * /proc entries */
271                 if (rc == -EALREADY)
272                         rc = 0;
273                 RETURN(rc);
274         }
275
276         if (newnid) {
277                 struct nid_stat *tmp = exp->exp_nid_stats;
278                 LASSERT(tmp != NULL);
279
280                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
281                 if (tmp->nid_brw_stats == NULL)
282                         GOTO(clean, rc = -ENOMEM);
283
284                 init_brw_stats(tmp->nid_brw_stats);
285                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
286                                         0644, &filter_per_nid_stats_fops,
287                                         exp->exp_nid_stats);
288                 if (rc)
289                         CWARN("Error adding the brw_stats file\n");
290
291                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
292                 if (rc)
293                         GOTO(clean, rc);
294
295                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
296                                             tmp->nid_stats);
297                 if (rc)
298                         GOTO(clean, rc);
299                 rc = lprocfs_nid_ldlm_stats_init(tmp);
300                 if (rc)
301                         GOTO(clean, rc);
302         }
303
304         RETURN(0);
305  clean:
306         return rc;
307 }
308
309 /* Add client data to the FILTER.  We use a bitmap to locate a free space
310  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
311  * Otherwise, we have just read the data from the last_rcvd file and
312  * we know its offset. */
313 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
314                              int cl_idx)
315 {
316         struct obd_device_target *obt = &obd->u.obt;
317         struct tg_export_data *ted = &exp->exp_target_data;
318         struct lr_server_data *lsd = class_server_data(obd);
319         unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
320         int new_client = (cl_idx == -1);
321
322         ENTRY;
323
324         LASSERT(bitmap != NULL);
325         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
326
327         /* Self-export */
328         if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
329                 RETURN(0);
330
331         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
332          * there's no need for extra complication here
333          */
334         if (new_client) {
335                 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
336         repeat:
337                 if (cl_idx >= LR_MAX_CLIENTS) {
338                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
339                                cl_idx);
340                         RETURN(-EOVERFLOW);
341                 }
342                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
343                         cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
344                                                         cl_idx);
345                         goto repeat;
346                 }
347         } else {
348                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
349                         CERROR("FILTER client %d: bit already set in bitmap!\n",
350                                cl_idx);
351                         LBUG();
352                 }
353         }
354
355         ted->ted_lr_idx = cl_idx;
356         ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
357                           cl_idx * le16_to_cpu(lsd->lsd_client_size);
358         cfs_init_mutex(&ted->ted_lcd_lock);
359         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
360
361         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
362                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
363
364         if (new_client) {
365                 struct lvfs_run_ctxt saved;
366                 loff_t off = ted->ted_lr_off;
367                 int rc;
368                 void *handle;
369
370                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
371                        ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
372
373                 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
374                         RETURN(-ENOSPC);
375
376                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
377                 /* Transaction needed to fix bug 1403 */
378                 handle = fsfilt_start(obd,
379                                       obt->obt_rcvd_filp->f_dentry->d_inode,
380                                       FSFILT_OP_SETATTR, NULL);
381                 if (IS_ERR(handle)) {
382                         rc = PTR_ERR(handle);
383                         CERROR("unable to start transaction: rc %d\n", rc);
384                 } else {
385                         ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
386                         exp->exp_last_request_time = cfs_time_current_sec();
387                         rc = fsfilt_add_journal_cb(obd, 0, handle,
388                                                    target_client_add_cb,
389                                                    class_export_cb_get(exp));
390                         if (rc == 0) {
391                                 cfs_spin_lock(&exp->exp_lock);
392                                 exp->exp_need_sync = 1;
393                                 cfs_spin_unlock(&exp->exp_lock);
394                         }
395                         rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
396                                                  ted->ted_lcd,
397                                                  sizeof(*ted->ted_lcd),
398                                                  &off, rc /* sync if no cb */);
399                         fsfilt_commit(obd,
400                                       obt->obt_rcvd_filp->f_dentry->d_inode,
401                                       handle, 0);
402                 }
403                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
404
405                 if (rc) {
406                         CERROR("error writing %s client idx %u: rc %d\n",
407                                LAST_RCVD, ted->ted_lr_idx, rc);
408                         RETURN(rc);
409                 }
410         }
411         RETURN(0);
412 }
413
414 static int filter_client_del(struct obd_export *exp)
415 {
416         struct tg_export_data *ted = &exp->exp_target_data;
417         struct obd_device_target *obt = &exp->exp_obd->u.obt;
418         struct lvfs_run_ctxt saved;
419         int rc;
420         loff_t off;
421         ENTRY;
422
423         if (ted->ted_lcd == NULL)
424                 RETURN(0);
425
426         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
427         if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
428                 GOTO(free, 0);
429
430         LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
431
432         off = ted->ted_lr_off;
433
434         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
435                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
436
437         /* Don't clear ted_lr_idx here as it is likely also unset.  At worst
438          * we leak a client slot that will be cleaned on the next recovery. */
439         if (off <= 0) {
440                 CERROR("%s: client idx %d has med_off %lld\n",
441                        exp->exp_obd->obd_name, ted->ted_lr_idx, off);
442                 GOTO(free, rc = -EINVAL);
443         }
444
445         /* Clear the bit _after_ zeroing out the client so we don't
446            race with filter_client_add and zero out new clients.*/
447         if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
448                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
449                        ted->ted_lr_idx);
450                 LBUG();
451         }
452
453         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
454         /* Make sure the server's last_transno is up to date.
455          * This should be done before zeroing client slot so last_transno will
456          * be in server data or in client data in case of failure */
457         filter_update_server_data(exp->exp_obd);
458
459         cfs_mutex_down(&ted->ted_lcd_lock);
460         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
461         rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
462                                  ted->ted_lcd,
463                                  sizeof(*ted->ted_lcd), &off, 0);
464         cfs_mutex_up(&ted->ted_lcd_lock);
465         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
466
467         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
468                "zero out client %s at idx %u/%llu in %s, rc %d\n",
469                ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
470                LAST_RCVD, rc);
471         EXIT;
472 free:
473         return 0;
474 }
475
476 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
477 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
478                                          struct filter_mod_data *fmd)
479 {
480         LASSERT_SPIN_LOCKED(&fed->fed_lock);
481         if (--fmd->fmd_refcount == 0) {
482                 /* XXX when we have persistent reservations and the handle
483                  * is stored herein we need to drop it here. */
484                 fed->fed_mod_count--;
485                 cfs_list_del(&fmd->fmd_list);
486                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
487         }
488 }
489
490 /* drop fmd reference, free it if last ref */
491 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
492 {
493         struct filter_export_data *fed;
494
495         if (fmd == NULL)
496                 return;
497
498         fed = &exp->exp_filter_data;
499         cfs_spin_lock(&fed->fed_lock);
500         filter_fmd_put_nolock(fed, fmd); /* caller reference */
501         cfs_spin_unlock(&fed->fed_lock);
502 }
503
504 /* expire entries from the end of the list if there are too many
505  * or they are too old */
506 static void filter_fmd_expire_nolock(struct filter_obd *filter,
507                                      struct filter_export_data *fed,
508                                      struct filter_mod_data *keep)
509 {
510         struct filter_mod_data *fmd, *tmp;
511
512         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
513                 if (fmd == keep)
514                         break;
515
516                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
517                     fed->fed_mod_count < filter->fo_fmd_max_num)
518                         break;
519
520                 cfs_list_del_init(&fmd->fmd_list);
521                 filter_fmd_put_nolock(fed, fmd); /* list reference */
522         }
523 }
524
525 void filter_fmd_expire(struct obd_export *exp)
526 {
527         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
528         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
529                                  &exp->exp_filter_data, NULL);
530         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
531 }
532
533 /* find specified objid, group in export fmd list.
534  * caller must hold fed_lock and take fmd reference itself */
535 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
536                                                 struct filter_export_data *fed,
537                                                 obd_id objid, obd_seq group)
538 {
539         struct filter_mod_data *found = NULL, *fmd;
540
541         LASSERT_SPIN_LOCKED(&fed->fed_lock);
542
543         cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
544                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
545                         found = fmd;
546                         cfs_list_del(&fmd->fmd_list);
547                         cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
548                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
549                         break;
550                 }
551         }
552
553         filter_fmd_expire_nolock(filter, fed, found);
554
555         return found;
556 }
557
558 /* Find fmd based on objid and group, or return NULL if not found. */
559 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
560                                         obd_id objid, obd_seq group)
561 {
562         struct filter_mod_data *fmd;
563
564         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
565         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
566                                      &exp->exp_filter_data, objid, group);
567         if (fmd)
568                 fmd->fmd_refcount++;    /* caller reference */
569         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
570
571         return fmd;
572 }
573
574 /* Find fmd based on objid and group, or create a new one if none is found.
575  * It is possible for this function to return NULL under memory pressure,
576  * or if objid = 0 is passed (which will only cause old entries to expire).
577  * Currently this is not fatal because any fmd state is transient and
578  * may also be freed when it gets sufficiently old. */
579 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
580                                        obd_id objid, obd_seq group)
581 {
582         struct filter_export_data *fed = &exp->exp_filter_data;
583         struct filter_mod_data *found = NULL, *fmd_new = NULL;
584
585         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
586
587         cfs_spin_lock(&fed->fed_lock);
588         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
589         if (fmd_new) {
590                 if (found == NULL) {
591                         cfs_list_add_tail(&fmd_new->fmd_list,
592                                           &fed->fed_mod_list);
593                         fmd_new->fmd_id = objid;
594                         fmd_new->fmd_gr = group;
595                         fmd_new->fmd_refcount++;   /* list reference */
596                         found = fmd_new;
597                         fed->fed_mod_count++;
598                 } else {
599                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
600                 }
601         }
602         if (found) {
603                 found->fmd_refcount++;          /* caller reference */
604                 found->fmd_expire = jiffies +
605                         exp->exp_obd->u.filter.fo_fmd_max_age;
606         }
607
608         cfs_spin_unlock(&fed->fed_lock);
609
610         return found;
611 }
612
613 #ifdef DO_FMD_DROP
614 /* drop fmd list reference so it will disappear when last reference is put.
615  * This isn't so critical because it would in fact only affect the one client
616  * that is doing the unlink and at worst we have an stale entry referencing
617  * an object that should never be used again. */
618 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
619 {
620         struct filter_mod_data *found = NULL;
621
622         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
623         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
624         if (found) {
625                 cfs_list_del_init(&found->fmd_list);
626                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
627         }
628         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
629 }
630 #else
631 #define filter_fmd_drop(exp, objid, group)
632 #endif
633
634 /* remove all entries from fmd list */
635 static void filter_fmd_cleanup(struct obd_export *exp)
636 {
637         struct filter_export_data *fed = &exp->exp_filter_data;
638         struct filter_mod_data *fmd = NULL, *tmp;
639
640         cfs_spin_lock(&fed->fed_lock);
641         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
642                 cfs_list_del_init(&fmd->fmd_list);
643                 filter_fmd_put_nolock(fed, fmd);
644         }
645         cfs_spin_unlock(&fed->fed_lock);
646 }
647
648 static int filter_init_export(struct obd_export *exp)
649 {
650         int rc;
651         ENTRY;
652
653         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
654         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
655
656         cfs_spin_lock(&exp->exp_lock);
657         exp->exp_connecting = 1;
658         cfs_spin_unlock(&exp->exp_lock);
659
660         /* self-export doesn't need client data and ldlm initialization */
661         if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
662                                      &exp->exp_client_uuid)))
663                 RETURN(0);
664
665         rc = lut_client_alloc(exp);
666         if (rc == 0)
667                 rc = ldlm_init_export(exp);
668         if (rc)
669                 CERROR("%s: Can't initialize export: rc %d\n",
670                        exp->exp_obd->obd_name, rc);
671
672         RETURN(rc);
673 }
674
675 static int filter_free_server_data(struct obd_device_target *obt)
676 {
677         lut_fini(NULL, obt->obt_lut);
678         OBD_FREE_PTR(obt->obt_lut);
679         return 0;
680 }
681
682 /* assumes caller is already in kernel ctxt */
683 int filter_update_server_data(struct obd_device *obd)
684 {
685         struct file *filp = obd->u.obt.obt_rcvd_filp;
686         struct lr_server_data *lsd = class_server_data(obd);
687         loff_t off = 0;
688         int rc;
689         ENTRY;
690
691         CDEBUG(D_INODE, "server uuid      : %s\n", lsd->lsd_uuid);
692         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
693                le64_to_cpu(lsd->lsd_last_transno));
694         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
695                le64_to_cpu(lsd->lsd_mount_count));
696
697         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
698         if (rc)
699                 CERROR("error writing lr_server_data: rc = %d\n", rc);
700
701         RETURN(rc);
702 }
703
704 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
705                              int force_sync)
706 {
707         struct filter_obd *filter = &obd->u.filter;
708         __u64 tmp;
709         loff_t off = 0;
710         int rc;
711         ENTRY;
712
713         if (filter->fo_last_objid_files[group] == NULL) {
714                 CERROR("Object seq "LPU64" not fully setup; not updating "
715                        "last_objid\n", group);
716                 RETURN(-EINVAL);
717         }
718
719         CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
720                obd->obd_name, filter->fo_last_objids[group], group);
721
722         tmp = cpu_to_le64(filter->fo_last_objids[group]);
723         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
724                                  &tmp, sizeof(tmp), &off, force_sync);
725         if (rc)
726                 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
727                        group, rc);
728         RETURN(rc);
729 }
730 extern int ost_handle(struct ptlrpc_request *req);
731 /* assumes caller has already in kernel ctxt */
732 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
733 {
734         struct filter_obd *filter = &obd->u.filter;
735         struct lr_server_data *lsd;
736         struct lsd_client_data *lcd = NULL;
737         struct inode *inode = filp->f_dentry->d_inode;
738         unsigned long last_rcvd_size = i_size_read(inode);
739         struct lu_target *lut;
740         __u64 mount_count;
741         __u32 start_epoch;
742         int cl_idx;
743         loff_t off = 0;
744         int rc;
745
746         /* ensure padding in the struct is the correct size */
747         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
748                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
749         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
750                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
751
752         /* allocate and initialize lu_target */
753         OBD_ALLOC_PTR(lut);
754         if (lut == NULL)
755                 RETURN(-ENOMEM);
756         rc = lut_init(NULL, lut, obd, NULL);
757         if (rc)
758                 GOTO(err_lut, rc);
759         lsd = class_server_data(obd);
760         if (last_rcvd_size == 0) {
761                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
762
763                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
764                 lsd->lsd_last_transno = 0;
765                 mount_count = lsd->lsd_mount_count = 0;
766                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
767                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
768                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
769                 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
770                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
771                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
772                  * MDS first connects and assigns the OST index number. */
773                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
774                                                         OBD_INCOMPAT_OST);
775         } else {
776                 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
777                 if (rc) {
778                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
779                                LAST_RCVD, rc);
780                         GOTO(err_lut, rc);
781                 }
782                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
783                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
784                                            "using the wrong disk %s. Were the "
785                                            "/dev/ assignments rearranged?\n",
786                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
787                         GOTO(err_lut, rc = -EINVAL);
788                 }
789                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
790                 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
791                 /* COMPAT_146 */
792                 /* Assume old last_rcvd format unless I_C_LR is set */
793                 if (!(lsd->lsd_feature_incompat &
794                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
795                         lsd->lsd_last_transno = lsd->lsd_compat14;
796                 /* end COMPAT_146 */
797                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
798                  * MDS first connects and assigns the OST index number. */
799                 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
800                                                          OBD_INCOMPAT_OST);
801         }
802
803         if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
804                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
805                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
806                        ~FILTER_INCOMPAT_SUPP);
807                 GOTO(err_lut, rc = -EINVAL);
808         }
809         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
810                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
811                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
812                        ~FILTER_ROCOMPAT_SUPP);
813                 /* Do something like remount filesystem read-only */
814                 GOTO(err_lut, rc = -EINVAL);
815         }
816
817         start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
818
819         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
820                obd->obd_name, start_epoch);
821         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
822                obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
823         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
824                obd->obd_name, mount_count + 1);
825         CDEBUG(D_INODE, "%s: server data size: %u\n",
826                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
827         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
828                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
829         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
830                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
831         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
832                obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
833         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
834                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
835                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
836                 le16_to_cpu(lsd->lsd_client_size));
837
838         if (!obd->obd_replayable) {
839                 CWARN("%s: recovery support OFF\n", obd->obd_name);
840                 GOTO(out, rc = 0);
841         }
842
843         OBD_ALLOC_PTR(lcd);
844         if (!lcd)
845                 GOTO(err_client, rc = -ENOMEM);
846
847         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
848              off < last_rcvd_size; cl_idx++) {
849                 __u64 last_rcvd;
850                 struct obd_export *exp;
851                 struct filter_export_data *fed;
852
853                 /* Don't assume off is incremented properly by
854                  * fsfilt_read_record(), in case sizeof(*lcd)
855                  * isn't the same as lsd->lsd_client_size.  */
856                 off = le32_to_cpu(lsd->lsd_client_start) +
857                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
858                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
859                 if (rc) {
860                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
861                                LAST_RCVD, cl_idx, off, rc);
862                         break; /* read error shouldn't cause startup to fail */
863                 }
864
865                 if (lcd->lcd_uuid[0] == '\0') {
866                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
867                                cl_idx);
868                         continue;
869                 }
870
871                 check_lcd(obd->obd_name, cl_idx, lcd);
872
873                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
874
875                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
876                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
877                        last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
878
879                 /* These exports are cleaned up by filter_disconnect(), so they
880                  * need to be set up like real exports as filter_connect() does.
881                  */
882                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
883                 if (IS_ERR(exp)) {
884                         if (PTR_ERR(exp) == -EALREADY) {
885                                 /* export already exists, zero out this one */
886                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
887                                 continue;
888                         }
889                         OBD_FREE_PTR(lcd);
890                         GOTO(err_client, rc = PTR_ERR(exp));
891                 }
892
893                 fed = &exp->exp_filter_data;
894                 *fed->fed_ted.ted_lcd = *lcd;
895                 fed->fed_group = 0; /* will be assigned at connect */
896                 filter_export_stats_init(obd, exp, NULL);
897                 rc = filter_client_add(obd, exp, cl_idx);
898                 /* can't fail for existing client */
899                 LASSERTF(rc == 0, "rc = %d\n", rc);
900
901                 /* VBR: set export last committed */
902                 exp->exp_last_committed = last_rcvd;
903                 cfs_spin_lock(&exp->exp_lock);
904                 exp->exp_connecting = 0;
905                 exp->exp_in_recovery = 0;
906                 cfs_spin_unlock(&exp->exp_lock);
907                 obd->obd_max_recoverable_clients++;
908                 class_export_put(exp);
909
910                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
911                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
912         }
913         OBD_FREE_PTR(lcd);
914
915         obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
916 out:
917         obd->u.obt.obt_mount_count = mount_count + 1;
918         obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
919         lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
920
921         /* save it, so mount count and last_transno is current */
922         rc = filter_update_server_data(obd);
923         if (rc)
924                 GOTO(err_client, rc);
925
926         RETURN(0);
927
928 err_client:
929         class_disconnect_exports(obd);
930 err_lut:
931         filter_free_server_data(&obd->u.obt);
932         RETURN(rc);
933 }
934
935 static int filter_cleanup_groups(struct obd_device *obd)
936 {
937         struct filter_obd *filter = &obd->u.filter;
938         struct file *filp;
939         struct dentry *dentry;
940         int i, j;
941         ENTRY;
942
943         if (filter->fo_dentry_O_groups != NULL) {
944                 for (i = 0; i < filter->fo_group_count; i++) {
945                         dentry = filter->fo_dentry_O_groups[i];
946                         if (dentry != NULL)
947                                 f_dput(dentry);
948                 }
949                 OBD_FREE(filter->fo_dentry_O_groups,
950                          filter->fo_group_count *
951                          sizeof(*filter->fo_dentry_O_groups));
952                 filter->fo_dentry_O_groups = NULL;
953         }
954         if (filter->fo_last_objid_files != NULL) {
955                 for (i = 0; i < filter->fo_group_count; i++) {
956                         filp = filter->fo_last_objid_files[i];
957                         if (filp != NULL)
958                                 filp_close(filp, 0);
959                 }
960                 OBD_FREE(filter->fo_last_objid_files,
961                          filter->fo_group_count *
962                          sizeof(*filter->fo_last_objid_files));
963                 filter->fo_last_objid_files = NULL;
964         }
965         if (filter->fo_dentry_O_sub != NULL) {
966                 for (i = 0; i < filter->fo_group_count; i++) {
967                         for (j = 0; j < filter->fo_subdir_count; j++) {
968                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
969                                 if (dentry != NULL)
970                                         f_dput(dentry);
971                         }
972                 }
973                 OBD_FREE(filter->fo_dentry_O_sub,
974                          filter->fo_group_count *
975                          sizeof(*filter->fo_dentry_O_sub));
976                 filter->fo_dentry_O_sub = NULL;
977         }
978         if (filter->fo_last_objids != NULL) {
979                 OBD_FREE(filter->fo_last_objids,
980                          filter->fo_group_count *
981                          sizeof(*filter->fo_last_objids));
982                 filter->fo_last_objids = NULL;
983         }
984         if (filter->fo_dentry_O != NULL) {
985                 f_dput(filter->fo_dentry_O);
986                 filter->fo_dentry_O = NULL;
987         }
988         RETURN(0);
989 }
990
991 static int filter_update_last_group(struct obd_device *obd, int group)
992 {
993         struct filter_obd *filter = &obd->u.filter;
994         struct file *filp = NULL;
995         int last_group = 0, rc;
996         loff_t off = 0;
997         ENTRY;
998
999         if (group <= filter->fo_committed_group)
1000                 RETURN(0);
1001
1002         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1003         if (IS_ERR(filp)) {
1004                 rc = PTR_ERR(filp);
1005                 filp = NULL;
1006                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1007                 GOTO(cleanup, rc);
1008         }
1009
1010         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1011         if (rc) {
1012                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1013                 GOTO(cleanup, rc);
1014         }
1015
1016         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1017                obd->obd_name, last_group, group);
1018
1019         off = 0;
1020         last_group = group;
1021         /* must be sync: bXXXX */
1022         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1023         if (rc) {
1024                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1025                 GOTO(cleanup, rc);
1026         }
1027
1028         filter->fo_committed_group = group;
1029 cleanup:
1030         if (filp)
1031                 filp_close(filp, 0);
1032         RETURN(rc);
1033 }
1034
1035 static int filter_read_group_internal(struct obd_device *obd, int group,
1036                                       int create)
1037 {
1038         struct filter_obd *filter = &obd->u.filter;
1039         __u64 *new_objids = NULL;
1040         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1041         struct dentry **new_groups = NULL;
1042         struct file **new_files = NULL;
1043         struct dentry *dentry;
1044         struct file *filp;
1045         int old_count = filter->fo_group_count, rc, stage = 0, i;
1046         char name[25];
1047         __u64 last_objid;
1048         loff_t off = 0;
1049         int len = group + 1;
1050
1051         snprintf(name, 24, "%d", group);
1052         name[24] = '\0';
1053
1054         if (!create) {
1055                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1056                                            strlen(name));
1057                 if (IS_ERR(dentry)) {
1058                         CERROR("Cannot lookup expected object group %d: %ld\n",
1059                                group, PTR_ERR(dentry));
1060                         RETURN(PTR_ERR(dentry));
1061                 }
1062         } else {
1063                 dentry = simple_mkdir(filter->fo_dentry_O,
1064                                       obd->u.obt.obt_vfsmnt, name, 0700, 1);
1065                 if (IS_ERR(dentry)) {
1066                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1067                                PTR_ERR(dentry));
1068                         RETURN(PTR_ERR(dentry));
1069                 }
1070         }
1071         stage = 1;
1072
1073         snprintf(name, 24, "O/%d/LAST_ID", group);
1074         name[24] = '\0';
1075         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1076         if (IS_ERR(filp)) {
1077                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1078                 GOTO(cleanup, rc = PTR_ERR(filp));
1079         }
1080         stage = 2;
1081
1082         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1083         if (rc) {
1084                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1085                 GOTO(cleanup, rc);
1086         }
1087
1088         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1089                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1090                 if (tmp_subdirs == NULL)
1091                         GOTO(cleanup, rc = -ENOMEM);
1092                 stage = 3;
1093
1094                 for (i = 0; i < filter->fo_subdir_count; i++) {
1095                         char dir[20];
1096                         snprintf(dir, sizeof(dir), "d%u", i);
1097
1098                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1099                                                               obd->u.obt.obt_vfsmnt,
1100                                                               dir, 0700, 1);
1101                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1102                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1103                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1104                                        group, dir, rc);
1105                                 GOTO(cleanup, rc);
1106                         }
1107
1108                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1109                                tmp_subdirs->dentry[i]);
1110                 }
1111         }
1112
1113         /* 'group' is an index; we need an array of length 'group + 1' */
1114         if (group + 1 > old_count) {
1115                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1116                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1117                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1118                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1119                 stage = 4;
1120                 if (new_objids == NULL || new_subdirs == NULL ||
1121                     new_groups == NULL || new_files == NULL)
1122                         GOTO(cleanup, rc = -ENOMEM);
1123
1124                 if (old_count) {
1125                         memcpy(new_objids, filter->fo_last_objids,
1126                                old_count * sizeof(*new_objids));
1127                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1128                                old_count * sizeof(*new_subdirs));
1129                         memcpy(new_groups, filter->fo_dentry_O_groups,
1130                                old_count * sizeof(*new_groups));
1131                         memcpy(new_files, filter->fo_last_objid_files,
1132                                old_count * sizeof(*new_files));
1133
1134                         OBD_FREE(filter->fo_last_objids,
1135                                  old_count * sizeof(*new_objids));
1136                         OBD_FREE(filter->fo_dentry_O_sub,
1137                                  old_count * sizeof(*new_subdirs));
1138                         OBD_FREE(filter->fo_dentry_O_groups,
1139                                  old_count * sizeof(*new_groups));
1140                         OBD_FREE(filter->fo_last_objid_files,
1141                                  old_count * sizeof(*new_files));
1142                 }
1143                 filter->fo_last_objids = new_objids;
1144                 filter->fo_dentry_O_sub = new_subdirs;
1145                 filter->fo_dentry_O_groups = new_groups;
1146                 filter->fo_last_objid_files = new_files;
1147                 filter->fo_group_count = len;
1148         }
1149
1150         filter->fo_dentry_O_groups[group] = dentry;
1151         filter->fo_last_objid_files[group] = filp;
1152         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1153                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1154                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1155         }
1156
1157         filter_update_last_group(obd, group);
1158
1159         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1160                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1161                 rc = filter_update_last_objid(obd, group, 1);
1162                 RETURN(rc);
1163         }
1164
1165         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1166         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1167                obd->obd_name, group, last_objid);
1168         RETURN(0);
1169  cleanup:
1170         switch (stage) {
1171         case 4:
1172                 if (new_objids != NULL)
1173                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1174                 if (new_subdirs != NULL)
1175                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1176                 if (new_groups != NULL)
1177                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1178                 if (new_files != NULL)
1179                         OBD_FREE(new_files, len * sizeof(*new_files));
1180         case 3:
1181                 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1182                         for (i = 0; i < filter->fo_subdir_count; i++) {
1183                                 if (tmp_subdirs->dentry[i] != NULL)
1184                                         dput(tmp_subdirs->dentry[i]);
1185                         }
1186                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1187                 }
1188         case 2:
1189                 filp_close(filp, 0);
1190         case 1:
1191                 dput(dentry);
1192         }
1193         RETURN(rc);
1194 }
1195
1196 static int filter_read_groups(struct obd_device *obd, int last_group,
1197                               int create)
1198 {
1199         struct filter_obd *filter = &obd->u.filter;
1200         int old_count, group, rc = 0;
1201
1202         cfs_down(&filter->fo_init_lock);
1203         old_count = filter->fo_group_count;
1204         for (group = old_count; group <= last_group; group++) {
1205                 rc = filter_read_group_internal(obd, group, create);
1206                 if (rc != 0)
1207                         break;
1208         }
1209         cfs_up(&filter->fo_init_lock);
1210         return rc;
1211 }
1212
1213 /* FIXME: object groups */
1214 static int filter_prep_groups(struct obd_device *obd)
1215 {
1216         struct filter_obd *filter = &obd->u.filter;
1217         struct dentry *O_dentry;
1218         struct file *filp;
1219         int    last_group, rc = 0, cleanup_phase = 0;
1220         loff_t off = 0;
1221         ENTRY;
1222
1223         O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1224                                 "O", 0700, 1);
1225         CDEBUG(D_INODE, "%s: got/created O: %p\n", obd->obd_name, O_dentry);
1226         if (IS_ERR(O_dentry)) {
1227                 rc = PTR_ERR(O_dentry);
1228                 CERROR("%s: cannot open/create O: rc = %d\n", obd->obd_name,rc);
1229                 GOTO(cleanup, rc);
1230         }
1231         filter->fo_dentry_O = O_dentry;
1232         cleanup_phase = 1; /* O_dentry */
1233
1234         /* we have to initialize all groups before first connections from
1235          * clients because they may send create/destroy for any group -bzzz */
1236         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1237         if (IS_ERR(filp)) {
1238                 CERROR("%s: cannot create LAST_GROUP: rc = %ld\n",
1239                        obd->obd_name, PTR_ERR(filp));
1240                 GOTO(cleanup, rc = PTR_ERR(filp));
1241         }
1242         cleanup_phase = 2; /* filp */
1243
1244         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1245         if (rc) {
1246                 CERROR("%s: error reading LAST_GROUP: rc %d\n",
1247                        obd->obd_name, rc);
1248                 GOTO(cleanup, rc);
1249         }
1250
1251         if (off == 0)
1252                 last_group = FID_SEQ_OST_MDT0;
1253
1254         CDEBUG(D_INODE, "%s: initialize group %u (max %u)\n", obd->obd_name,
1255                FID_SEQ_OST_MDT0, last_group);
1256         filter->fo_committed_group = last_group;
1257         rc = filter_read_groups(obd, last_group, 1);
1258         if (rc)
1259                 GOTO(cleanup, rc);
1260
1261         filp_close(filp, 0);
1262         RETURN(0);
1263
1264  cleanup:
1265         switch (cleanup_phase) {
1266         case 2:
1267                 filp_close(filp, 0);
1268         case 1:
1269                 filter_cleanup_groups(obd);
1270                 f_dput(filter->fo_dentry_O);
1271                 filter->fo_dentry_O = NULL;
1272         default:
1273                 break;
1274         }
1275         return rc;
1276
1277 }
1278
1279 /* setup the object store with correct subdirectories */
1280 static int filter_prep(struct obd_device *obd)
1281 {
1282         struct lvfs_run_ctxt saved;
1283         struct filter_obd *filter = &obd->u.filter;
1284         struct file *file;
1285         struct inode *inode;
1286         int rc = 0;
1287         ENTRY;
1288
1289         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1290         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1291         if (!file || IS_ERR(file)) {
1292                 rc = PTR_ERR(file);
1293                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1294                        LAST_RCVD, rc);
1295                 GOTO(out, rc);
1296         }
1297         obd->u.obt.obt_rcvd_filp = file;
1298         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1299                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1300                        file->f_dentry->d_inode->i_mode);
1301                 GOTO(err_filp, rc = -ENOENT);
1302         }
1303
1304         inode = file->f_dentry->d_parent->d_inode;
1305         /* We use i_op->unlink directly in filter_vfs_unlink() */
1306         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1307                 CERROR("%s: filesystem does not support create/unlink ops\n",
1308                        obd->obd_name);
1309                 GOTO(err_filp, rc = -EOPNOTSUPP);
1310         }
1311
1312         rc = filter_init_server_data(obd, file);
1313         if (rc) {
1314                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1315                 GOTO(err_filp, rc);
1316         }
1317         LASSERT(obd->u.obt.obt_lut);
1318         target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1319
1320         /* open and create health check io file*/
1321         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1322         if (IS_ERR(file)) {
1323                 rc = PTR_ERR(file);
1324                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1325                        HEALTH_CHECK, rc);
1326                 GOTO(err_server_data, rc);
1327         }
1328         filter->fo_obt.obt_health_check_filp = file;
1329         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1330                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1331                        file->f_dentry->d_inode->i_mode);
1332                 GOTO(err_health_check, rc = -ENOENT);
1333         }
1334         rc = lvfs_check_io_health(obd, file);
1335         if (rc)
1336                 GOTO(err_health_check, rc);
1337
1338         rc = filter_prep_groups(obd);
1339         if (rc)
1340                 GOTO(err_health_check, rc);
1341 out:
1342         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1343
1344         return(rc);
1345
1346 err_health_check:
1347         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1348                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1349         filter->fo_obt.obt_health_check_filp = NULL;
1350 err_server_data:
1351         target_recovery_fini(obd);
1352         filter_free_server_data(&obd->u.obt);
1353 err_filp:
1354         if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1355                 CERROR("can't close %s after error\n", LAST_RCVD);
1356         obd->u.obt.obt_rcvd_filp = NULL;
1357         goto out;
1358 }
1359
1360 /* cleanup the filter: write last used object id to status file */
1361 static void filter_post(struct obd_device *obd)
1362 {
1363         struct lvfs_run_ctxt saved;
1364         struct filter_obd *filter = &obd->u.filter;
1365         int rc, i;
1366
1367         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1368          * best to start a transaction with h_sync, because we removed this
1369          * from lastobjid */
1370
1371         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1372         rc = filter_update_server_data(obd);
1373         if (rc)
1374                 CERROR("error writing server data: rc = %d\n", rc);
1375
1376         for (i = 0; i < filter->fo_group_count; i++) {
1377                 rc = filter_update_last_objid(obd, i,
1378                                 (i == filter->fo_group_count - 1));
1379                 if (rc)
1380                         CERROR("error writing group %d lastobjid: rc = %d\n",
1381                                i, rc);
1382         }
1383
1384         rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1385         obd->u.obt.obt_rcvd_filp = NULL;
1386         if (rc)
1387                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1388
1389         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1390         filter->fo_obt.obt_health_check_filp = NULL;
1391         if (rc)
1392                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1393
1394         filter_cleanup_groups(obd);
1395         filter_free_server_data(&obd->u.obt);
1396         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1397
1398         filter_free_capa_keys(filter);
1399         cleanup_capa_hash(filter->fo_capa_hash);
1400 }
1401
1402 static void filter_set_last_id(struct filter_obd *filter,
1403                                obd_id id, obd_seq group)
1404 {
1405         LASSERT(group <= filter->fo_group_count);
1406
1407         cfs_spin_lock(&filter->fo_objidlock);
1408         filter->fo_last_objids[group] = id;
1409         cfs_spin_unlock(&filter->fo_objidlock);
1410 }
1411
1412 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1413 {
1414         obd_id id;
1415         LASSERT(group <= filter->fo_group_count);
1416         LASSERT(filter->fo_last_objids != NULL);
1417
1418         /* FIXME: object groups */
1419         cfs_spin_lock(&filter->fo_objidlock);
1420         id = filter->fo_last_objids[group];
1421         cfs_spin_unlock(&filter->fo_objidlock);
1422         return id;
1423 }
1424
1425 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1426 {
1427         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1428         return 0;
1429 }
1430
1431 /* We never dget the object parent, so DON'T dput it either */
1432 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1433 {
1434         struct filter_obd *filter = &obd->u.filter;
1435         struct filter_subdirs *subdirs;
1436
1437         if (group >= filter->fo_group_count) /* FIXME: object groups */
1438                 return ERR_PTR(-EBADF);
1439
1440         if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1441                 return filter->fo_dentry_O_groups[group];
1442
1443         subdirs = &filter->fo_dentry_O_sub[group];
1444         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1445 }
1446
1447 /* We never dget the object parent, so DON'T dput it either */
1448 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1449                                   obd_id objid)
1450 {
1451         unsigned long now = jiffies;
1452         struct dentry *dparent = filter_parent(obd, group, objid);
1453         int rc;
1454
1455         if (IS_ERR(dparent))
1456                 return dparent;
1457         if (dparent == NULL)
1458                 return ERR_PTR(-ENOENT);
1459
1460         rc = filter_lock_dentry(obd, dparent);
1461         fsfilt_check_slow(obd, now, "parent lock");
1462         return rc ? ERR_PTR(rc) : dparent;
1463 }
1464
1465 /* We never dget the object parent, so DON'T dput it either */
1466 static void filter_parent_unlock(struct dentry *dparent)
1467 {
1468         UNLOCK_INODE_MUTEX(dparent->d_inode);
1469 }
1470
1471 /* How to get files, dentries, inodes from object id's.
1472  *
1473  * If dir_dentry is passed, the caller has already locked the parent
1474  * appropriately for this operation (normally a write lock).  If
1475  * dir_dentry is NULL, we do a read lock while we do the lookup to
1476  * avoid races with create/destroy and such changing the directory
1477  * internal to the filesystem code. */
1478 struct dentry *filter_fid2dentry(struct obd_device *obd,
1479                                  struct dentry *dir_dentry,
1480                                  obd_seq group, obd_id id)
1481 {
1482         struct dentry *dparent = dir_dentry;
1483         struct dentry *dchild;
1484         char name[32];
1485         int len;
1486         ENTRY;
1487
1488         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1489             obd->u.filter.fo_destroys_in_progress == 0) {
1490                 /* don't fail lookups for orphan recovery, it causes
1491                  * later LBUGs when objects still exist during precreate */
1492                 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1493                 RETURN(ERR_PTR(-ENOENT));
1494         }
1495         if (id == 0) {
1496                 CERROR("fatal: invalid object id 0\n");
1497                 RETURN(ERR_PTR(-ESTALE));
1498         }
1499
1500         len = sprintf(name, LPU64, id);
1501         if (dir_dentry == NULL) {
1502                 dparent = filter_parent_lock(obd, group, id);
1503                 if (IS_ERR(dparent)) {
1504                         CERROR("%s: error getting object "POSTID
1505                                " parent: rc %ld\n", obd->obd_name,
1506                                id, group, PTR_ERR(dparent));
1507                         RETURN(dparent);
1508                 }
1509         }
1510         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1511                dparent->d_name.len, dparent->d_name.name, name);
1512         /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1513         dchild = lookup_one_len(name, dparent, len);
1514         if (dir_dentry == NULL)
1515                 filter_parent_unlock(dparent);
1516         if (IS_ERR(dchild)) {
1517                 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1518                        obd->obd_name, id, group, PTR_ERR(dchild));
1519                 RETURN(dchild);
1520         }
1521
1522         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1523                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1524                        obd->obd_name, id, dchild->d_inode->i_ino);
1525                 f_dput(dchild);
1526                 RETURN(ERR_PTR(-ENOENT));
1527         }
1528
1529         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1530                name, dchild, atomic_read(&dchild->d_count));
1531
1532         LASSERT(atomic_read(&dchild->d_count) > 0);
1533
1534         RETURN(dchild);
1535 }
1536
1537 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1538                                   obd_id group, struct lustre_handle *lockh)
1539 {
1540         int flags = LDLM_AST_DISCARD_DATA, rc;
1541         struct ldlm_res_id res_id;
1542         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1543         ENTRY;
1544
1545         osc_build_res_name(objid, group, &res_id);
1546         /* Tell the clients that the object is gone now and that they should
1547          * throw away any cached pages. */
1548         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1549                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1550                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1551                                     lockh);
1552         if (rc != ELDLM_OK)
1553                 lockh->cookie = 0;
1554         RETURN(rc);
1555 }
1556
1557 static void filter_fini_destroy(struct obd_device *obd,
1558                                 struct lustre_handle *lockh)
1559 {
1560         if (lustre_handle_is_used(lockh))
1561                 ldlm_lock_decref(lockh, LCK_PW);
1562 }
1563
1564 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1565  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1566  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1567  * reverse.  Caller must take i_sem before starting the transaction and we
1568  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1569 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1570                       struct vfsmount *mnt)
1571 {
1572         int rc;
1573         ENTRY;
1574
1575         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1576          * itself we already hold dir->i_mutex for child create/unlink ops */
1577         LASSERT(dentry->d_inode != NULL);
1578         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1579         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1580
1581
1582         /* may_delete() */
1583         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1584                 GOTO(out, rc = -ENOENT);
1585
1586         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1587         if (rc)
1588                 GOTO(out, rc);
1589
1590         if (IS_APPEND(dir))
1591                 GOTO(out, rc = -EPERM);
1592
1593         /* check_sticky() */
1594         if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1595              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1596             IS_IMMUTABLE(dentry->d_inode))
1597                 GOTO(out, rc = -EPERM);
1598
1599         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1600          *       that was done because of journal_start (which is already done
1601          *       here) or some other ordering issue. */
1602         ll_vfs_dq_init(dir);
1603
1604         rc = ll_security_inode_unlink(dir, dentry, mnt);
1605         if (rc)
1606                 GOTO(out, rc);
1607
1608         rc = dir->i_op->unlink(dir, dentry);
1609 out:
1610         /* need to drop i_mutex before we lose inode reference */
1611         UNLOCK_INODE_MUTEX(dentry->d_inode);
1612         if (rc == 0)
1613                 d_delete(dentry);
1614
1615         RETURN(rc);
1616 }
1617
1618 /* Caller must hold LCK_PW on parent and push us into kernel context.
1619  * Caller must hold child i_mutex, we drop it always.
1620  * Caller is also required to ensure that dchild->d_inode exists. */
1621 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1622                                    obd_seq group, struct dentry *dparent,
1623                                    struct dentry *dchild)
1624 {
1625         struct inode *inode = dchild->d_inode;
1626         int rc;
1627
1628         /* There should be 2 references to the inode:
1629          *  1) taken by filter_prepare_destroy
1630          *  2) taken by filter_destroy */
1631         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1632                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1633                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1634                        (unsigned long)inode->i_nlink,
1635                        atomic_read(&inode->i_count));
1636         }
1637
1638         rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1639         if (rc)
1640                 CERROR("error unlinking objid %.*s: rc %d\n",
1641                        dchild->d_name.len, dchild->d_name.name, rc);
1642         return(rc);
1643 }
1644
1645 struct filter_intent_args {
1646         struct ldlm_lock **victim;
1647         __u64 size;
1648         int *liblustre;
1649 };
1650
1651 static enum interval_iter filter_intent_cb(struct interval_node *n,
1652                                            void *args)
1653 {
1654         struct ldlm_interval *node = (struct ldlm_interval *)n;
1655         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1656         __u64 size = arg->size;
1657         struct ldlm_lock **v = arg->victim;
1658         struct ldlm_lock *lck;
1659
1660         /* If the interval is lower than the current file size,
1661          * just break. */
1662         if (interval_high(n) <= size)
1663                 return INTERVAL_ITER_STOP;
1664
1665         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1666                 /* Don't send glimpse ASTs to liblustre clients.
1667                  * They aren't listening for them, and they do
1668                  * entirely synchronous I/O anyways. */
1669                 if (lck->l_export == NULL ||
1670                     lck->l_export->exp_libclient == 1)
1671                         continue;
1672
1673                 if (*arg->liblustre)
1674                         *arg->liblustre = 0;
1675
1676                 if (*v == NULL) {
1677                         *v = LDLM_LOCK_GET(lck);
1678                 } else if ((*v)->l_policy_data.l_extent.start <
1679                            lck->l_policy_data.l_extent.start) {
1680                         LDLM_LOCK_RELEASE(*v);
1681                         *v = LDLM_LOCK_GET(lck);
1682                 }
1683
1684                 /* the same policy group - every lock has the
1685                  * same extent, so needn't do it any more */
1686                 break;
1687         }
1688
1689         return INTERVAL_ITER_CONT;
1690 }
1691
1692 static int filter_intent_policy(struct ldlm_namespace *ns,
1693                                 struct ldlm_lock **lockp, void *req_cookie,
1694                                 ldlm_mode_t mode, int flags, void *data)
1695 {
1696         CFS_LIST_HEAD(rpc_list);
1697         struct ptlrpc_request *req = req_cookie;
1698         struct ldlm_lock *lock = *lockp, *l = NULL;
1699         struct ldlm_resource *res = lock->l_resource;
1700         ldlm_processing_policy policy;
1701         struct ost_lvb *res_lvb, *reply_lvb;
1702         struct ldlm_reply *rep;
1703         ldlm_error_t err;
1704         int idx, rc, tmpflags = 0, only_liblustre = 1;
1705         struct ldlm_interval_tree *tree;
1706         struct filter_intent_args arg;
1707         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1708                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1709                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1710         ENTRY;
1711
1712         policy = ldlm_get_processing_policy(res);
1713         LASSERT(policy != NULL);
1714         LASSERT(req != NULL);
1715
1716         rc = lustre_pack_reply(req, 3, repsize, NULL);
1717         if (rc)
1718                 RETURN(req->rq_status = rc);
1719
1720         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1721         LASSERT(rep != NULL);
1722
1723         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1724                                    sizeof(*reply_lvb));
1725         LASSERT(reply_lvb != NULL);
1726
1727         //fixup_handle_for_resent_req(req, lock, &lockh);
1728
1729         /* Call the extent policy function to see if our request can be
1730          * granted, or is blocked.
1731          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1732          * lock, and should not be granted if the lock will be blocked.
1733          */
1734
1735         LASSERT(ns == ldlm_res_to_ns(res));
1736         lock_res(res);
1737         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1738         check_res_locked(res);
1739
1740         /* FIXME: we should change the policy function slightly, to not make
1741          * this list at all, since we just turn around and free it */
1742         while (!cfs_list_empty(&rpc_list)) {
1743                 struct ldlm_lock *wlock =
1744                         cfs_list_entry(rpc_list.next, struct ldlm_lock,
1745                                        l_cp_ast);
1746                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1747                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1748                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1749                 cfs_list_del_init(&wlock->l_cp_ast);
1750                 LDLM_LOCK_RELEASE(wlock);
1751         }
1752
1753         /* The lock met with no resistance; we're finished. */
1754         if (rc == LDLM_ITER_CONTINUE) {
1755                 /* do not grant locks to the liblustre clients: they cannot
1756                  * handle ASTs robustly.  We need to do this while still
1757                  * holding lr_lock to avoid the lock remaining on the res_link
1758                  * list (and potentially being added to l_pending_list by an
1759                  * AST) when we are going to drop this lock ASAP. */
1760                 if (lock->l_export->exp_libclient ||
1761                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1762                         ldlm_resource_unlink_lock(lock);
1763                         err = ELDLM_LOCK_ABORTED;
1764                 } else {
1765                         err = ELDLM_LOCK_REPLACED;
1766                 }
1767                 unlock_res(res);
1768                 RETURN(err);
1769         }
1770
1771         /* Do not grant any lock, but instead send GL callbacks.  The extent
1772          * policy nicely created a list of all PW locks for us.  We will choose
1773          * the highest of those which are larger than the size in the LVB, if
1774          * any, and perform a glimpse callback. */
1775         res_lvb = res->lr_lvb_data;
1776         LASSERT(res_lvb != NULL);
1777         *reply_lvb = *res_lvb;
1778
1779         /*
1780          * lr_lock guarantees that no new locks are granted, and,
1781          * therefore, that res->lr_lvb_data cannot increase beyond the
1782          * end of already granted lock. As a result, it is safe to
1783          * check against "stale" reply_lvb->lvb_size value without
1784          * res->lr_lvb_sem.
1785          */
1786         arg.size = reply_lvb->lvb_size;
1787         arg.victim = &l;
1788         arg.liblustre = &only_liblustre;
1789         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1790                 tree = &res->lr_itree[idx];
1791                 if (tree->lit_mode == LCK_PR)
1792                         continue;
1793
1794                 interval_iterate_reverse(tree->lit_root,
1795                                          filter_intent_cb, &arg);
1796         }
1797         unlock_res(res);
1798
1799         /* There were no PW locks beyond the size in the LVB; finished. */
1800         if (l == NULL) {
1801                 if (only_liblustre) {
1802                         /* If we discovered a liblustre client with a PW lock,
1803                          * however, the LVB may be out of date!  The LVB is
1804                          * updated only on glimpse (which we don't do for
1805                          * liblustre clients) and cancel (which the client
1806                          * obviously has not yet done).  So if it has written
1807                          * data but kept the lock, the LVB is stale and needs
1808                          * to be updated from disk.
1809                          *
1810                          * Of course, this will all disappear when we switch to
1811                          * taking liblustre locks on the OST. */
1812                         ldlm_res_lvbo_update(res, NULL, 1);
1813                 }
1814                 RETURN(ELDLM_LOCK_ABORTED);
1815         }
1816
1817         /*
1818          * This check is for lock taken in filter_prepare_destroy() that does
1819          * not have l_glimpse_ast set. So the logic is: if there is a lock
1820          * with no l_glimpse_ast set, this object is being destroyed already.
1821          *
1822          * Hence, if you are grabbing DLM locks on the server, always set
1823          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1824          */
1825         if (l->l_glimpse_ast == NULL) {
1826                 /* We are racing with unlink(); just return -ENOENT */
1827                 rep->lock_policy_res1 = -ENOENT;
1828                 goto out;
1829         }
1830
1831         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1832         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1833
1834         lock_res(res);
1835         *reply_lvb = *res_lvb;
1836         unlock_res(res);
1837
1838  out:
1839         LDLM_LOCK_RELEASE(l);
1840
1841         RETURN(ELDLM_LOCK_ABORTED);
1842 }
1843
1844 /*
1845  * per-obd_device iobuf pool.
1846  *
1847  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1848  * allocations in write-path has to be minimized (see bug 5137).
1849  *
1850  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1851  * OST threads (see ost_thread_{init,done}()).
1852  *
1853  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1854  * at the OST layer there are only (potentially) multiple obd_device of type
1855  * unknown at the time of OST thread creation.
1856  *
1857  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1858  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1859  * it's very own iobuf.
1860  *
1861  * Functions below
1862  *
1863  *     filter_kiobuf_pool_init()
1864  *
1865  *     filter_kiobuf_pool_done()
1866  *
1867  *     filter_iobuf_get()
1868  *
1869  * operate on this array. They are "generic" in a sense that they don't depend
1870  * on actual type of iobuf's (the latter depending on Linux kernel version).
1871  */
1872
1873 /*
1874  * destroy pool created by filter_iobuf_pool_init
1875  */
1876 static void filter_iobuf_pool_done(struct filter_obd *filter)
1877 {
1878         struct filter_iobuf **pool;
1879         int i;
1880
1881         ENTRY;
1882
1883         pool = filter->fo_iobuf_pool;
1884         if (pool != NULL) {
1885                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1886                         if (pool[i] != NULL)
1887                                 filter_free_iobuf(pool[i]);
1888                 }
1889                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1890                 filter->fo_iobuf_pool = NULL;
1891         }
1892         EXIT;
1893 }
1894
1895 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1896 {
1897         struct filter_obd       *filter = &obd->u.filter;
1898         struct sptlrpc_rule_set  tmp_rset;
1899         int                      rc;
1900
1901         sptlrpc_rule_set_init(&tmp_rset);
1902         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1903         if (rc) {
1904                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1905                        obd->obd_name, rc);
1906                 return rc;
1907         }
1908
1909         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1910
1911         cfs_write_lock(&filter->fo_sptlrpc_lock);
1912         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1913         filter->fo_sptlrpc_rset = tmp_rset;
1914         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1915
1916         return 0;
1917 }
1918
1919 /*
1920  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1921  */
1922 static int filter_iobuf_pool_init(struct filter_obd *filter)
1923 {
1924         void **pool;
1925
1926         ENTRY;
1927
1928
1929         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1930                       GFP_KERNEL);
1931         if (filter->fo_iobuf_pool == NULL)
1932                 RETURN(-ENOMEM);
1933
1934         filter->fo_iobuf_count = OSS_THREADS_MAX;
1935
1936         RETURN(0);
1937 }
1938
1939 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1940  * many threads there will be so we allocate a large empty array and only
1941  * fill in those slots that are actually in use.
1942  * If we haven't allocated a pool entry for this thread before, do so now. */
1943 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1944 {
1945         int thread_id                    = (oti && oti->oti_thread) ?
1946                                            oti->oti_thread->t_id : -1;
1947         struct filter_iobuf  *pool       = NULL;
1948         struct filter_iobuf **pool_place = NULL;
1949
1950         if (thread_id >= 0) {
1951                 LASSERT(thread_id < filter->fo_iobuf_count);
1952                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1953         }
1954
1955         if (unlikely(pool == NULL)) {
1956                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1957                                           PTLRPC_MAX_BRW_PAGES);
1958                 if (pool_place != NULL)
1959                         *pool_place = pool;
1960         }
1961
1962         return pool;
1963 }
1964
1965 /* mount the file system (secretly).  lustre_cfg parameters are:
1966  * 1 = device
1967  * 2 = fstype
1968  * 3 = flags: failover=f, failout=n
1969  * 4 = mount options
1970  */
1971 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1972                         void *option)
1973 {
1974         struct filter_obd *filter = &obd->u.filter;
1975         struct vfsmount *mnt;
1976         struct lustre_mount_info *lmi;
1977         struct obd_uuid uuid;
1978         __u8 *uuid_ptr;
1979         char *str, *label;
1980         char ns_name[48];
1981         struct request_queue *q;
1982         int rc, i;
1983         ENTRY;
1984
1985         if (lcfg->lcfg_bufcount < 3 ||
1986             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1987             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1988                 RETURN(-EINVAL);
1989
1990         lmi = server_get_mount(obd->obd_name);
1991         if (lmi) {
1992                 /* We already mounted in lustre_fill_super.
1993                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1994                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1995                 mnt = lmi->lmi_mnt;
1996                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1997         } else {
1998                 /* old path - used by lctl */
1999                 CERROR("Using old MDS mount method\n");
2000                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2001                                     MS_NOATIME|MS_NODIRATIME,
2002                                     lustre_cfg_string(lcfg, 1), option);
2003                 if (IS_ERR(mnt)) {
2004                         rc = PTR_ERR(mnt);
2005                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2006                                            lustre_cfg_string(lcfg, 1), rc);
2007                         RETURN(rc);
2008                 }
2009
2010                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2011         }
2012         if (IS_ERR(obd->obd_fsops))
2013                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2014
2015         rc = filter_iobuf_pool_init(filter);
2016         if (rc != 0)
2017                 GOTO(err_ops, rc);
2018
2019         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2020                 CERROR("%s: Underlying device is marked as read-only. "
2021                        "Setup failed\n", obd->obd_name);
2022                 GOTO(err_ops, rc = -EROFS);
2023         }
2024
2025         /* failover is the default */
2026         obd->obd_replayable = 1;
2027
2028         /* disable connection until configuration finishes */
2029         obd->obd_no_conn = 1;
2030
2031         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2032                 str = lustre_cfg_string(lcfg, 3);
2033                 if (strchr(str, 'n')) {
2034                         CWARN("%s: recovery disabled\n", obd->obd_name);
2035                         obd->obd_replayable = 0;
2036                 }
2037         }
2038
2039         obd->u.obt.obt_magic = OBT_MAGIC;
2040         obd->u.obt.obt_vfsmnt = mnt;
2041         obd->u.obt.obt_sb = mnt->mnt_sb;
2042         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2043         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2044
2045         fsfilt_setup(obd, obd->u.obt.obt_sb);
2046
2047         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2048         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2049         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2050         obd->obd_lvfs_ctxt.fs = get_ds();
2051         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2052
2053         cfs_init_mutex(&filter->fo_init_lock);
2054         filter->fo_committed_group = 0;
2055         filter->fo_destroys_in_progress = 0;
2056         for (i = 0; i < 32; i++)
2057                 cfs_sema_init(&filter->fo_create_locks[i], 1);
2058
2059         cfs_spin_lock_init(&filter->fo_objidlock);
2060         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2061         cfs_sema_init(&filter->fo_alloc_lock, 1);
2062         init_brw_stats(&filter->fo_filter_stats);
2063         cfs_spin_lock_init(&filter->fo_flags_lock);
2064         filter->fo_read_cache = 1; /* enable read-only cache by default */
2065         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2066         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2067         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2068         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2069         filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2070         filter_slc_set(filter); /* initialize sync on lock cancel */
2071
2072         rc = filter_prep(obd);
2073         if (rc)
2074                 GOTO(err_ops, rc);
2075
2076         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2077         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2078
2079         filter->fo_fl_oss_capa = 1;
2080
2081         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2082         filter->fo_capa_hash = init_capa_hash();
2083         if (filter->fo_capa_hash == NULL)
2084                 GOTO(err_post, rc = -ENOMEM);
2085
2086         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2087         obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2088                                                 LDLM_NAMESPACE_SERVER,
2089                                                 LDLM_NAMESPACE_GREEDY,
2090                                                 LDLM_NS_TYPE_OST);
2091         if (obd->obd_namespace == NULL)
2092                 GOTO(err_post, rc = -ENOMEM);
2093         obd->obd_namespace->ns_lvbp = obd;
2094         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2095         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2096
2097         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2098                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2099
2100         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2101         if (rc) {
2102                 CERROR("failed to setup llogging subsystems\n");
2103                 GOTO(err_post, rc);
2104         }
2105
2106         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2107         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2108         /* do this after llog being initialized */
2109         filter_adapt_sptlrpc_conf(obd, 1);
2110
2111         rc = lquota_setup(filter_quota_interface_ref, obd);
2112         if (rc)
2113                 GOTO(err_post, rc);
2114
2115         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2116         if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2117             queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2118                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2119                               "for larger I/O requests: max_sectors = %u "
2120                               "could be up to max_hw_sectors=%u\n",
2121                               obd->obd_name, mnt->mnt_sb->s_id,
2122                               queue_max_sectors(q), queue_max_hw_sectors(q));
2123
2124         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2125         if (uuid_ptr != NULL) {
2126                 class_uuid_unparse(uuid_ptr, &uuid);
2127                 str = uuid.uuid;
2128         } else {
2129                 str = "no UUID";
2130         }
2131
2132         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2133         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2134                       obd->obd_name, label ?: str, lmi ? "on " : "",
2135                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2136                       obd->obd_replayable ? "enabled" : "disabled");
2137
2138         RETURN(0);
2139
2140 err_post:
2141         filter_post(obd);
2142 err_ops:
2143         fsfilt_put_ops(obd->obd_fsops);
2144         filter_iobuf_pool_done(filter);
2145 err_mntput:
2146         server_put_mount(obd->obd_name, mnt);
2147         obd->u.obt.obt_sb = 0;
2148         return rc;
2149 }
2150
2151 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2152 {
2153         struct lprocfs_static_vars lvars;
2154         cfs_proc_dir_entry_t *entry;
2155         unsigned long addr;
2156         struct page *page;
2157         int rc;
2158         ENTRY;
2159
2160         CLASSERT(offsetof(struct obd_device, u.obt) ==
2161                  offsetof(struct obd_device, u.filter.fo_obt));
2162
2163         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2164                 RETURN(-EINVAL);
2165
2166         /* lprocfs must be setup before the filter so state can be safely added
2167          * to /proc incrementally as the filter is setup */
2168         lprocfs_filter_init_vars(&lvars);
2169         rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2170         if (rc) {
2171                 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2172                        obd->obd_name, rc);
2173                 RETURN(rc);
2174         }
2175
2176         rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2177         if (rc) {
2178                 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2179                        obd->obd_name, rc);
2180                 GOTO(obd_cleanup, rc);
2181         }
2182
2183         /* Init obdfilter private stats here */
2184         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2185                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2186         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2187                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2188         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2189                              LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2190                              "get_page", "usec");
2191         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2192                              LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2193         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2194                              LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2195         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2196                              LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2197         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2198                              LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2199
2200         rc = lproc_filter_attach_seqstat(obd);
2201         if (rc) {
2202                 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2203                 GOTO(free_obd_stats, rc);
2204         }
2205
2206         entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2207         if (IS_ERR(entry)) {
2208                 rc = PTR_ERR(entry);
2209                 CERROR("%s: error %d setting up lprocfs for %s\n",
2210                        obd->obd_name, rc, "exports");
2211                 GOTO(free_obd_stats, rc);
2212         }
2213         obd->obd_proc_exports_entry = entry;
2214
2215         entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2216                                    lprocfs_nid_stats_clear_read,
2217                                    lprocfs_nid_stats_clear_write, obd, NULL);
2218         if (IS_ERR(entry)) {
2219                 rc = PTR_ERR(entry);
2220                 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2221                        obd->obd_name, rc);
2222                 GOTO(free_obd_stats, rc);
2223         }
2224
2225         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2226         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2227         if (!page)
2228                 GOTO(remove_entry_clear, rc = -ENOMEM);
2229         addr = (unsigned long)cfs_page_address(page);
2230         clear_page((void *)addr);
2231         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2232                LUSTRE_CFG_BUFLEN(lcfg, 4));
2233         rc = filter_common_setup(obd, lcfg, (void *)addr);
2234         OBD_PAGE_FREE(page);
2235         if (rc) {
2236                 CERROR("%s: filter_common_setup failed: %d.\n",
2237                        obd->obd_name, rc);
2238                 GOTO(remove_entry_clear, rc);
2239         }
2240
2241         RETURN(0);
2242
2243 remove_entry_clear:
2244         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2245 free_obd_stats:
2246         lprocfs_free_obd_stats(obd);
2247 obd_cleanup:
2248         lprocfs_obd_cleanup(obd);
2249         return rc;
2250 }
2251
2252 static struct llog_operations filter_mds_ost_repl_logops;
2253
2254 static struct llog_operations filter_size_orig_logops = {
2255         .lop_setup   = llog_obd_origin_setup,
2256         .lop_cleanup = llog_obd_origin_cleanup,
2257         .lop_add     = llog_obd_origin_add
2258 };
2259
2260 static int filter_olg_fini(struct obd_llog_group *olg)
2261 {
2262         struct llog_ctxt *ctxt;
2263         int rc = 0, rc2 = 0;
2264         ENTRY;
2265
2266         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2267         if (ctxt)
2268                 rc = llog_cleanup(ctxt);
2269
2270         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2271         if (ctxt) {
2272                 rc2 = llog_cleanup(ctxt);
2273                 if (!rc)
2274                         rc = rc2;
2275         }
2276
2277         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2278         if (ctxt) {
2279                 rc2 = llog_cleanup(ctxt);
2280                 if (!rc)
2281                         rc = rc2;
2282         }
2283
2284         RETURN(rc);
2285 }
2286
2287 static int
2288 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2289                 struct obd_device *tgt)
2290 {
2291         int rc;
2292         ENTRY;
2293
2294         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2295                         &filter_mds_ost_repl_logops);
2296         if (rc)
2297                 GOTO(cleanup, rc);
2298
2299         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2300                         &filter_size_orig_logops);
2301         if (rc)
2302                 GOTO(cleanup, rc);
2303         EXIT;
2304 cleanup:
2305         if (rc)
2306                 filter_olg_fini(olg);
2307         return rc;
2308 }
2309
2310 /**
2311  * Init the default olg, which is embeded in the obd_device, for filter.
2312  */
2313 static int
2314 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2315                         struct obd_device *tgt)
2316 {
2317         struct filter_obd *filter = &obd->u.filter;
2318         struct llog_ctxt *ctxt;
2319         int rc;
2320         ENTRY;
2321
2322         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2323         if (!filter->fo_lcm)
2324                 RETURN(-ENOMEM);
2325
2326         filter_mds_ost_repl_logops = llog_client_ops;
2327         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2328         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2329         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2330
2331         rc = filter_olg_init(obd, olg, tgt);
2332         if (rc)
2333                 GOTO(cleanup_lcm, rc);
2334
2335         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2336                         &llog_lvfs_ops);
2337         if (rc)
2338                 GOTO(cleanup_olg, rc);
2339
2340         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2341         if (!ctxt) {
2342                 CERROR("Can't get ctxt for %p:%x\n", olg,
2343                        LLOG_MDS_OST_REPL_CTXT);
2344                 GOTO(cleanup_olg, rc = -ENODEV);
2345         }
2346         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2347         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2348         llog_ctxt_put(ctxt);
2349
2350         RETURN(0);
2351 cleanup_olg:
2352         filter_olg_fini(olg);
2353 cleanup_lcm:
2354         llog_recov_thread_fini(filter->fo_lcm, 1);
2355         filter->fo_lcm = NULL;
2356         return rc;
2357 }
2358
2359 static int
2360 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2361                  struct obd_device *tgt, int *index)
2362 {
2363         struct filter_obd *filter = &obd->u.filter;
2364         struct llog_ctxt *ctxt;
2365         int rc;
2366         ENTRY;
2367
2368         LASSERT(olg != NULL);
2369         if (olg == &obd->obd_olg)
2370                 return filter_default_olg_init(obd, olg, tgt);
2371
2372         LASSERT(filter->fo_lcm != NULL);
2373         rc = filter_olg_init(obd, olg, tgt);
2374         if (rc)
2375                 RETURN(rc);
2376         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2377         if (!ctxt) {
2378                 CERROR("Can't get ctxt for %p:%x\n", olg,
2379                        LLOG_MDS_OST_REPL_CTXT);
2380                 filter_olg_fini(olg);
2381                 RETURN(-ENODEV);
2382         }
2383         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2384         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2385         llog_ctxt_put(ctxt);
2386         RETURN(rc);
2387 }
2388
2389 static int filter_llog_finish(struct obd_device *obd, int count)
2390 {
2391         struct filter_obd *filter = &obd->u.filter;
2392         struct llog_ctxt *ctxt;
2393         ENTRY;
2394
2395         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2396         if (ctxt) {
2397                 /*
2398                  * Make sure that no cached llcds left in recov_thread.
2399                  * We actually do sync in disconnect time, but disconnect
2400                  * may not come being marked rq_no_resend = 1.
2401                  */
2402                 llog_sync(ctxt, NULL);
2403
2404                 /*
2405                  * Balance class_import_get() in llog_receptor_accept().
2406                  * This is safe to do, as llog is already synchronized
2407                  * and its import may go.
2408                  */
2409                 cfs_mutex_down(&ctxt->loc_sem);
2410                 if (ctxt->loc_imp) {
2411                         class_import_put(ctxt->loc_imp);
2412                         ctxt->loc_imp = NULL;
2413                 }
2414                 cfs_mutex_up(&ctxt->loc_sem);
2415                 llog_ctxt_put(ctxt);
2416         }
2417
2418         if (filter->fo_lcm) {
2419                 cfs_mutex_down(&ctxt->loc_sem);
2420                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2421                 filter->fo_lcm = NULL;
2422                 cfs_mutex_up(&ctxt->loc_sem);
2423         }
2424         RETURN(filter_olg_fini(&obd->obd_olg));
2425 }
2426
2427 /**
2428  * Find the group llog according to group index in the llog group list.
2429  */
2430 static struct obd_llog_group *
2431 filter_find_olg_internal(struct filter_obd *filter, int group)
2432 {
2433         struct obd_llog_group *olg;
2434
2435         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2436         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2437                 if (olg->olg_seq == group)
2438                         RETURN(olg);
2439         }
2440         RETURN(NULL);
2441 }
2442
2443 /**
2444  * Find the group llog according to group index on the filter
2445  */
2446 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2447 {
2448         struct obd_llog_group *olg = NULL;
2449         struct filter_obd *filter;
2450
2451         filter = &obd->u.filter;
2452
2453         if (group == FID_SEQ_LLOG)
2454                 RETURN(&obd->obd_olg);
2455
2456         cfs_spin_lock(&filter->fo_llog_list_lock);
2457         olg = filter_find_olg_internal(filter, group);
2458         cfs_spin_unlock(&filter->fo_llog_list_lock);
2459
2460         RETURN(olg);
2461 }
2462 /**
2463  * Find the llog_group of the filter according to the group. If it can not
2464  * find, create the llog_group, which only happens when mds is being synced
2465  * with OST.
2466  */
2467 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2468 {
2469         struct obd_llog_group *olg = NULL, *olg_new = NULL;
2470         struct filter_obd *filter;
2471         int rc;
2472
2473         filter = &obd->u.filter;
2474
2475         if (group == FID_SEQ_LLOG)
2476                 RETURN(&obd->obd_olg);
2477
2478         OBD_ALLOC_PTR(olg_new);
2479         if (olg_new == NULL)
2480                RETURN(ERR_PTR(-ENOMEM));
2481
2482         cfs_spin_lock(&filter->fo_llog_list_lock);
2483         olg = filter_find_olg_internal(filter, group);
2484         if (olg) {
2485                 if (olg->olg_initializing) {
2486                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2487                 } else {
2488                         GOTO(out_unlock, olg);
2489                 }
2490         } else {
2491                 /* set as the newly allocated one */
2492                 olg = olg_new;
2493                 olg_new = NULL;
2494         }
2495
2496         llog_group_init(olg, group);
2497         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2498         olg->olg_initializing = 1;
2499         cfs_spin_unlock(&filter->fo_llog_list_lock);
2500
2501         rc = obd_llog_init(obd, olg, obd, NULL);
2502         if (rc) {
2503                cfs_spin_lock(&filter->fo_llog_list_lock);
2504                cfs_list_del(&olg->olg_list);
2505                cfs_spin_unlock(&filter->fo_llog_list_lock);
2506                OBD_FREE_PTR(olg);
2507                GOTO(out, olg = ERR_PTR(-ENOMEM));
2508         }
2509         cfs_spin_lock(&filter->fo_llog_list_lock);
2510         olg->olg_initializing = 0;
2511         cfs_spin_unlock(&filter->fo_llog_list_lock);
2512         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2513               obd->obd_name, group, olg);
2514 out:
2515         RETURN(olg);
2516
2517 out_unlock:
2518         cfs_spin_unlock(&filter->fo_llog_list_lock);
2519         if (olg_new)
2520                OBD_FREE_PTR(olg_new);
2521         goto out;
2522 }
2523
2524 static int filter_llog_connect(struct obd_export *exp,
2525                                struct llogd_conn_body *body)
2526 {
2527         struct obd_device *obd = exp->exp_obd;
2528         struct llog_ctxt *ctxt;
2529         struct obd_llog_group *olg;
2530         int rc;
2531         ENTRY;
2532
2533         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2534                obd->obd_name, body->lgdc_logid.lgl_oid,
2535                body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2536
2537         olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2538         if (!olg) {
2539                 CERROR(" %s: can not find olg of group %d\n",
2540                        obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2541                 RETURN(-ENOENT);
2542         }
2543         llog_group_set_export(olg, exp);
2544
2545         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2546         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2547                  body->lgdc_ctxt_idx);
2548
2549         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2550               obd->obd_name, body->lgdc_logid.lgl_oid,
2551               body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2552
2553         cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2554         obd->u.filter.fo_mds_ost_sync = 1;
2555         cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2556         rc = llog_connect(ctxt, &body->lgdc_logid,
2557                           &body->lgdc_gen, NULL);
2558         llog_ctxt_put(ctxt);
2559         if (rc != 0)
2560                 CERROR("failed to connect rc %d idx %d\n", rc,
2561                                 body->lgdc_ctxt_idx);
2562
2563         RETURN(rc);
2564 }
2565
2566 static int filter_llog_preclean(struct obd_device *obd)
2567 {
2568         struct obd_llog_group *olg, *tmp;
2569         struct filter_obd *filter;
2570         cfs_list_t  remove_list;
2571         int rc = 0;
2572         ENTRY;
2573
2574         rc = obd_llog_finish(obd, 0);
2575         if (rc)
2576                 CERROR("failed to cleanup llogging subsystem\n");
2577
2578         filter = &obd->u.filter;
2579         CFS_INIT_LIST_HEAD(&remove_list);
2580
2581         cfs_spin_lock(&filter->fo_llog_list_lock);
2582         while (!cfs_list_empty(&filter->fo_llog_list)) {
2583                 olg = cfs_list_entry(filter->fo_llog_list.next,
2584                                      struct obd_llog_group, olg_list);
2585                 cfs_list_del(&olg->olg_list);
2586                 cfs_list_add(&olg->olg_list, &remove_list);
2587         }
2588         cfs_spin_unlock(&filter->fo_llog_list_lock);
2589
2590         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2591                 cfs_list_del_init(&olg->olg_list);
2592                 rc = filter_olg_fini(olg);
2593                 if (rc)
2594                         CERROR("failed to cleanup llogging subsystem for %u\n",
2595                                olg->olg_seq);
2596                 OBD_FREE_PTR(olg);
2597         }
2598
2599         RETURN(rc);
2600 }
2601
2602 static int filter_precleanup(struct obd_device *obd,
2603                              enum obd_cleanup_stage stage)
2604 {
2605         int rc = 0;
2606         ENTRY;
2607
2608         switch(stage) {
2609         case OBD_CLEANUP_EARLY:
2610                 break;
2611         case OBD_CLEANUP_EXPORTS:
2612                 /* Stop recovery before namespace cleanup. */
2613                 target_recovery_fini(obd);
2614
2615                 obd_exports_barrier(obd);
2616                 obd_zombie_barrier();
2617
2618                 rc = filter_llog_preclean(obd);
2619                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2620                 lprocfs_free_per_client_stats(obd);
2621                 lprocfs_obd_cleanup(obd);
2622                 lprocfs_free_obd_stats(obd);
2623                 lquota_cleanup(filter_quota_interface_ref, obd);
2624                 break;
2625         }
2626         RETURN(rc);
2627 }
2628
2629 static int filter_cleanup(struct obd_device *obd)
2630 {
2631         struct filter_obd *filter = &obd->u.filter;
2632         ENTRY;
2633
2634         if (obd->obd_fail)
2635                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2636                               "will be preserved.\n", obd->obd_name);
2637
2638         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2639         obd->obd_namespace = NULL;
2640
2641         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2642
2643         if (obd->u.obt.obt_sb == NULL)
2644                 RETURN(0);
2645
2646         filter_post(obd);
2647
2648         ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2649         shrink_dcache_sb(obd->u.obt.obt_sb);
2650
2651         server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2652         obd->u.obt.obt_sb = NULL;
2653
2654         fsfilt_put_ops(obd->obd_fsops);
2655
2656         filter_iobuf_pool_done(filter);
2657
2658         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2659
2660         RETURN(0);
2661 }
2662
2663 static int filter_connect_internal(struct obd_export *exp,
2664                                    struct obd_connect_data *data,
2665                                    int reconnect)
2666 {
2667         struct filter_export_data *fed = &exp->exp_filter_data;
2668
2669         if (!data)
2670                 RETURN(0);
2671
2672         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2673                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2674                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2675                data->ocd_connect_flags, data->ocd_version,
2676                data->ocd_grant, data->ocd_index);
2677
2678         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2679                 CWARN("!!! This export (nid %s) used object group %d "
2680                        "earlier; now it's trying to use group %d!  This could "
2681                        "be a bug in the MDS. Please report to "
2682                        "http://bugs.whamcloud.com/\n",
2683                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2684                 RETURN(-EPROTO);
2685         }
2686         fed->fed_group = data->ocd_group;
2687
2688         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2689         exp->exp_connect_flags = data->ocd_connect_flags;
2690         data->ocd_version = LUSTRE_VERSION_CODE;
2691
2692         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2693         if (data->ocd_connect_flags & OBD_CONNECT_MDS)
2694                 CWARN("%s: Received MDS connection for group %u\n",
2695                       exp->exp_obd->obd_name, data->ocd_group);
2696         else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
2697                 RETURN(-EPROTO);
2698
2699         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2700                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2701                 obd_size left, want;
2702
2703                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2704                 left = filter_grant_space_left(exp);
2705                 want = data->ocd_grant;
2706                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2707                 data->ocd_grant = fed->fed_grant;
2708                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2709
2710                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2711                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2712                        exp->exp_client_uuid.uuid, exp,
2713                        data->ocd_grant, want, left);
2714
2715                 filter->fo_tot_granted_clients ++;
2716         }
2717
2718         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2719                 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2720                 int index = le32_to_cpu(lsd->lsd_ost_index);
2721
2722                 if (!(lsd->lsd_feature_compat &
2723                       cpu_to_le32(OBD_COMPAT_OST))) {
2724                         /* this will only happen on the first connect */
2725                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2726                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2727                         /* sync is not needed here as filter_client_add will
2728                          * set exp_need_sync flag */
2729                         filter_update_server_data(exp->exp_obd);
2730                 } else if (index != data->ocd_index) {
2731                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2732                                            " %u doesn't match actual OST index"
2733                                            " %u in last_rcvd file, bad "
2734                                            "configuration?\n",
2735                                            obd_export_nid2str(exp), index,
2736                                            data->ocd_index);
2737                         RETURN(-EBADF);
2738                 }
2739                 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2740                  * FIXME: We don't strictly need the COMPAT flag for that,
2741                  * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2742                  * FIXME: We needed it for the index, as index 0 is valid. */
2743         }
2744
2745         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2746                 data->ocd_brw_size = 65536;
2747         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2748                 data->ocd_brw_size = min(data->ocd_brw_size,
2749                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2750                 if (data->ocd_brw_size == 0) {
2751                         CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2752                                " ocd_version: %x ocd_grant: %d ocd_index: %u "
2753                                "ocd_brw_size is unexpectedly zero, "
2754                                "network data corruption?"
2755                                "Refusing connection of this client\n",
2756                                 exp->exp_obd->obd_name,
2757                                 exp->exp_client_uuid.uuid,
2758                                 exp, data->ocd_connect_flags, data->ocd_version,
2759                                 data->ocd_grant, data->ocd_index);
2760                         RETURN(-EPROTO);
2761                 }
2762         }
2763
2764         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2765                 __u32 cksum_types = data->ocd_cksum_types;
2766
2767                 /* The client set in ocd_cksum_types the checksum types it
2768                  * supports. We have to mask off the algorithms that we don't
2769                  * support */
2770                 data->ocd_cksum_types &= cksum_types_supported();
2771
2772                 /* 1.6.4- only support CRC32 and didn't set ocd_cksum_types */
2773                 if (unlikely(data->ocd_cksum_types == 0))
2774                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2775
2776                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2777                                    "%x\n", exp->exp_obd->obd_name,
2778                                    obd_export_nid2str(exp), cksum_types,
2779                                    data->ocd_cksum_types);
2780         } else {
2781                 /* This client does not support OBD_CONNECT_CKSUM
2782                  * fall back to CRC32 */
2783                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2784                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2785                                    exp->exp_obd->obd_name,
2786                                    obd_export_nid2str(exp));
2787         }
2788
2789         if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2790                 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2791
2792         RETURN(0);
2793 }
2794
2795 static int filter_reconnect(const struct lu_env *env,
2796                             struct obd_export *exp, struct obd_device *obd,
2797                             struct obd_uuid *cluuid,
2798                             struct obd_connect_data *data,
2799                             void *localdata)
2800 {
2801         int rc;
2802         ENTRY;
2803
2804         if (exp == NULL || obd == NULL || cluuid == NULL)
2805                 RETURN(-EINVAL);
2806
2807         rc = filter_connect_internal(exp, data, 1);
2808         if (rc == 0)
2809                 filter_export_stats_init(obd, exp, localdata);
2810
2811         RETURN(rc);
2812 }
2813
2814 static int filter_connect(const struct lu_env *env,
2815                           struct obd_export **exp, struct obd_device *obd,
2816                           struct obd_uuid *cluuid,
2817                           struct obd_connect_data *data, void *localdata)
2818 {
2819         struct lvfs_run_ctxt saved;
2820         struct lustre_handle conn = { 0 };
2821         struct obd_export *lexp;
2822         int rc;
2823         ENTRY;
2824
2825         if (exp == NULL || obd == NULL || cluuid == NULL)
2826                 RETURN(-EINVAL);
2827
2828         rc = class_connect(&conn, obd, cluuid);
2829         if (rc)
2830                 RETURN(rc);
2831         lexp = class_conn2export(&conn);
2832         LASSERT(lexp != NULL);
2833
2834         rc = filter_connect_internal(lexp, data, 0);
2835         if (rc)
2836                 GOTO(cleanup, rc);
2837
2838         filter_export_stats_init(obd, lexp, localdata);
2839         if (obd->obd_replayable) {
2840                 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2841                 LASSERT(lcd);
2842                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2843                 rc = filter_client_add(obd, lexp, -1);
2844                 if (rc)
2845                         GOTO(cleanup, rc);
2846         }
2847
2848         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2849         rc = filter_read_groups(obd, data->ocd_group, 1);
2850         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2851         if (rc != 0) {
2852                 CERROR("can't read group %u\n", data->ocd_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
2991         if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
2992                                      &exp->exp_client_uuid)))
2993                RETURN(0);
2994
2995         ldlm_destroy_export(exp);
2996         lut_client_free(exp);
2997
2998         if (!exp->exp_obd->obd_replayable)
2999                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3000
3001         filter_grant_discard(exp);
3002         filter_fmd_cleanup(exp);
3003
3004         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3005                 struct filter_obd *filter = &exp->exp_obd->u.filter;
3006                 if (filter->fo_tot_granted_clients > 0)
3007                         filter->fo_tot_granted_clients --;
3008         }
3009
3010         if (!(exp->exp_flags & OBD_OPT_FORCE))
3011                 filter_grant_sanity_check(exp->exp_obd, __func__);
3012
3013         RETURN(0);
3014 }
3015
3016 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3017 {
3018         struct obd_llog_group *olg_min, *olg;
3019         struct filter_obd *filter;
3020         int worked = -1, group;
3021         struct llog_ctxt *ctxt;
3022         ENTRY;
3023
3024         filter = &obd->u.filter;
3025
3026         /* we can't sync log holding spinlock. also, we do not want to get
3027          * into livelock. so we do following: loop over MDS's exports in
3028          * group order and skip already synced llogs -bzzz */
3029         do {
3030                 /* look for group with min. number, but > worked */
3031                 olg_min = NULL;
3032                 group = 1 << 30;
3033                 cfs_spin_lock(&filter->fo_llog_list_lock);
3034                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3035                         if (olg->olg_seq <= worked) {
3036                                 /* this group is already synced */
3037                                 continue;
3038                         }
3039                         if (group < olg->olg_seq) {
3040                                 /* we have group with smaller number to sync */
3041                                 continue;
3042                         }
3043                         /* store current minimal group */
3044                         olg_min = olg;
3045                         group = olg->olg_seq;
3046                 }
3047                 cfs_spin_unlock(&filter->fo_llog_list_lock);
3048
3049                 if (olg_min == NULL)
3050                         break;
3051
3052                 worked = olg_min->olg_seq;
3053                 if (olg_min->olg_exp &&
3054                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3055                         int err;
3056                         ctxt = llog_group_get_ctxt(olg_min,
3057                                                    LLOG_MDS_OST_REPL_CTXT);
3058                         if (ctxt) {
3059                                 err = llog_sync(ctxt, olg_min->olg_exp);
3060                                 llog_ctxt_put(ctxt);
3061                                 if (err) {
3062                                         CERROR("error flushing logs to MDS: "
3063                                                "rc %d\n", err);
3064                                 }
3065                         }
3066                 }
3067         } while (olg_min != NULL);
3068 }
3069
3070 /* Also incredibly similar to mds_disconnect */
3071 static int filter_disconnect(struct obd_export *exp)
3072 {
3073         struct obd_device *obd = exp->exp_obd;
3074         int rc;
3075         ENTRY;
3076
3077         LASSERT(exp);
3078         class_export_get(exp);
3079
3080         if (!(exp->exp_flags & OBD_OPT_FORCE))
3081                 filter_grant_sanity_check(obd, __func__);
3082         filter_grant_discard(exp);
3083
3084         /* Flush any remaining cancel messages out to the target */
3085         filter_sync_llogs(obd, exp);
3086
3087         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3088
3089         rc = server_disconnect_export(exp);
3090
3091         /* Do not erase record for recoverable client. */
3092         if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3093                 filter_client_del(exp);
3094         else
3095                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3096
3097         class_export_put(exp);
3098         RETURN(rc);
3099 }
3100
3101 /* reverse import is changed, sync all cancels */
3102 static void filter_revimp_update(struct obd_export *exp)
3103 {
3104         ENTRY;
3105
3106         LASSERT(exp);
3107         class_export_get(exp);
3108
3109         /* flush any remaining cancel messages out to the target */
3110         filter_sync_llogs(exp->exp_obd, exp);
3111         class_export_put(exp);
3112         EXIT;
3113 }
3114
3115 static int filter_ping(struct obd_export *exp)
3116 {
3117         filter_fmd_expire(exp);
3118         return 0;
3119 }
3120
3121 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3122                                   const char *what, int quiet)
3123 {
3124         struct dentry *dchild = NULL;
3125
3126         dchild = filter_fid2dentry(obd, NULL,  ostid->oi_seq, ostid->oi_id);
3127
3128         if (IS_ERR(dchild)) {
3129                 CERROR("%s error looking up object: "POSTID"\n",
3130                        what, ostid->oi_id, ostid->oi_seq);
3131                 RETURN(dchild);
3132         }
3133
3134         if (dchild->d_inode == NULL) {
3135                 if (!quiet)
3136                         CERROR("%s: %s on non-existent object: "POSTID" \n",
3137                                obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3138                 f_dput(dchild);
3139                 RETURN(ERR_PTR(-ENOENT));
3140         }
3141
3142         return dchild;
3143 }
3144
3145 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3146 {
3147         struct dentry *dentry = NULL;
3148         struct obd_device *obd;
3149         int rc = 0;
3150         ENTRY;
3151
3152         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3153                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3154         if (rc)
3155                 RETURN(rc);
3156
3157         obd = class_exp2obd(exp);
3158         if (obd == NULL) {
3159                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3160                 RETURN(-EINVAL);
3161         }
3162
3163         dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3164         if (IS_ERR(dentry))
3165                 RETURN(PTR_ERR(dentry));
3166
3167         /* Limit the valid bits in the return data to what we actually use */
3168         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3169         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3170
3171         f_dput(dentry);
3172         RETURN(rc);
3173 }
3174
3175 /* this should be enabled/disabled in condition to enabled/disabled large
3176  * inodes (fast EAs) in backing store FS. */
3177 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3178                         void *handle, struct obdo *oa)
3179 {
3180         struct obd_device *obd = exp->exp_obd;
3181         int rc = 0;
3182         ENTRY;
3183
3184         if (oa->o_valid & OBD_MD_FLFID) {
3185                 struct filter_fid ff;
3186
3187                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3188                         oa->o_seq = 0;
3189                 /* packing fid and converting it to LE for storing into EA.
3190                  * Here ->o_stripe_idx should be filled by LOV and rest of
3191                  * fields - by client. */
3192                 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3193                 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3194                 /* XXX: we are ignoring o_parent_ver here, since this should
3195                  *      be the same for all objects in this fileset. */
3196                 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3197                 ff.ff_objid = cpu_to_le64(oa->o_id);
3198                 ff.ff_seq = cpu_to_le64(oa->o_seq);
3199
3200                 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3201                        LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3202                        oa->o_seq);
3203
3204                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3205        &nbs