Whamcloud - gitweb
0dd323bc36de79257fbd3edc9997028cd1f52fe0
[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         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
652         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
653
654         cfs_spin_lock(&exp->exp_lock);
655         exp->exp_connecting = 1;
656         cfs_spin_unlock(&exp->exp_lock);
657         rc = lut_client_alloc(exp);
658         if (rc == 0)
659                 rc = ldlm_init_export(exp);
660
661         return rc;
662 }
663
664 static int filter_free_server_data(struct obd_device_target *obt)
665 {
666         lut_fini(NULL, obt->obt_lut);
667         OBD_FREE_PTR(obt->obt_lut);
668         return 0;
669 }
670
671 /* assumes caller is already in kernel ctxt */
672 int filter_update_server_data(struct obd_device *obd)
673 {
674         struct file *filp = obd->u.obt.obt_rcvd_filp;
675         struct lr_server_data *lsd = class_server_data(obd);
676         loff_t off = 0;
677         int rc;
678         ENTRY;
679
680         CDEBUG(D_INODE, "server uuid      : %s\n", lsd->lsd_uuid);
681         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
682                le64_to_cpu(lsd->lsd_last_transno));
683         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
684                le64_to_cpu(lsd->lsd_mount_count));
685
686         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
687         if (rc)
688                 CERROR("error writing lr_server_data: rc = %d\n", rc);
689
690         RETURN(rc);
691 }
692
693 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
694                              int force_sync)
695 {
696         struct filter_obd *filter = &obd->u.filter;
697         __u64 tmp;
698         loff_t off = 0;
699         int rc;
700         ENTRY;
701
702         if (filter->fo_last_objid_files[group] == NULL) {
703                 CERROR("Object seq "LPU64" not fully setup; not updating "
704                        "last_objid\n", group);
705                 RETURN(-EINVAL);
706         }
707
708         CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
709                obd->obd_name, filter->fo_last_objids[group], group);
710
711         tmp = cpu_to_le64(filter->fo_last_objids[group]);
712         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
713                                  &tmp, sizeof(tmp), &off, force_sync);
714         if (rc)
715                 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
716                        group, rc);
717         RETURN(rc);
718 }
719 extern int ost_handle(struct ptlrpc_request *req);
720 /* assumes caller has already in kernel ctxt */
721 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
722 {
723         struct filter_obd *filter = &obd->u.filter;
724         struct lr_server_data *lsd;
725         struct lsd_client_data *lcd = NULL;
726         struct inode *inode = filp->f_dentry->d_inode;
727         unsigned long last_rcvd_size = i_size_read(inode);
728         struct lu_target *lut;
729         __u64 mount_count;
730         __u32 start_epoch;
731         int cl_idx;
732         loff_t off = 0;
733         int rc;
734
735         /* ensure padding in the struct is the correct size */
736         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
737                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
738         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
739                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
740
741         /* allocate and initialize lu_target */
742         OBD_ALLOC_PTR(lut);
743         if (lut == NULL)
744                 RETURN(-ENOMEM);
745         rc = lut_init(NULL, lut, obd, NULL);
746         if (rc)
747                 GOTO(err_lut, rc);
748         lsd = class_server_data(obd);
749         if (last_rcvd_size == 0) {
750                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
751
752                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
753                 lsd->lsd_last_transno = 0;
754                 mount_count = lsd->lsd_mount_count = 0;
755                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
756                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
757                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
758                 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
759                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
760                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
761                  * MDS first connects and assigns the OST index number. */
762                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
763                                                         OBD_INCOMPAT_OST);
764         } else {
765                 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
766                 if (rc) {
767                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
768                                LAST_RCVD, rc);
769                         GOTO(err_lut, rc);
770                 }
771                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
772                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
773                                            "using the wrong disk %s. Were the "
774                                            "/dev/ assignments rearranged?\n",
775                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
776                         GOTO(err_lut, rc = -EINVAL);
777                 }
778                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
779                 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
780                 /* COMPAT_146 */
781                 /* Assume old last_rcvd format unless I_C_LR is set */
782                 if (!(lsd->lsd_feature_incompat &
783                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
784                         lsd->lsd_last_transno = lsd->lsd_compat14;
785                 /* end COMPAT_146 */
786                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
787                  * MDS first connects and assigns the OST index number. */
788                 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
789                                                          OBD_INCOMPAT_OST);
790         }
791
792         if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
793                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
794                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
795                        ~FILTER_INCOMPAT_SUPP);
796                 GOTO(err_lut, rc = -EINVAL);
797         }
798         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
799                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
800                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
801                        ~FILTER_ROCOMPAT_SUPP);
802                 /* Do something like remount filesystem read-only */
803                 GOTO(err_lut, rc = -EINVAL);
804         }
805
806         start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
807
808         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
809                obd->obd_name, start_epoch);
810         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
811                obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
812         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
813                obd->obd_name, mount_count + 1);
814         CDEBUG(D_INODE, "%s: server data size: %u\n",
815                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
816         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
817                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
818         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
819                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
820         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
821                obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
822         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
823                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
824                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
825                 le16_to_cpu(lsd->lsd_client_size));
826
827         if (!obd->obd_replayable) {
828                 CWARN("%s: recovery support OFF\n", obd->obd_name);
829                 GOTO(out, rc = 0);
830         }
831
832         OBD_ALLOC_PTR(lcd);
833         if (!lcd)
834                 GOTO(err_client, rc = -ENOMEM);
835
836         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
837              off < last_rcvd_size; cl_idx++) {
838                 __u64 last_rcvd;
839                 struct obd_export *exp;
840                 struct filter_export_data *fed;
841
842                 /* Don't assume off is incremented properly by
843                  * fsfilt_read_record(), in case sizeof(*lcd)
844                  * isn't the same as lsd->lsd_client_size.  */
845                 off = le32_to_cpu(lsd->lsd_client_start) +
846                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
847                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
848                 if (rc) {
849                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
850                                LAST_RCVD, cl_idx, off, rc);
851                         break; /* read error shouldn't cause startup to fail */
852                 }
853
854                 if (lcd->lcd_uuid[0] == '\0') {
855                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
856                                cl_idx);
857                         continue;
858                 }
859
860                 check_lcd(obd->obd_name, cl_idx, lcd);
861
862                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
863
864                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
865                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
866                        last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
867
868                 /* These exports are cleaned up by filter_disconnect(), so they
869                  * need to be set up like real exports as filter_connect() does.
870                  */
871                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
872                 if (IS_ERR(exp)) {
873                         if (PTR_ERR(exp) == -EALREADY) {
874                                 /* export already exists, zero out this one */
875                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
876                                 continue;
877                         }
878                         OBD_FREE_PTR(lcd);
879                         GOTO(err_client, rc = PTR_ERR(exp));
880                 }
881
882                 fed = &exp->exp_filter_data;
883                 *fed->fed_ted.ted_lcd = *lcd;
884                 fed->fed_group = 0; /* will be assigned at connect */
885                 filter_export_stats_init(obd, exp, NULL);
886                 rc = filter_client_add(obd, exp, cl_idx);
887                 /* can't fail for existing client */
888                 LASSERTF(rc == 0, "rc = %d\n", rc);
889
890                 /* VBR: set export last committed */
891                 exp->exp_last_committed = last_rcvd;
892                 cfs_spin_lock(&exp->exp_lock);
893                 exp->exp_connecting = 0;
894                 exp->exp_in_recovery = 0;
895                 cfs_spin_unlock(&exp->exp_lock);
896                 obd->obd_max_recoverable_clients++;
897                 class_export_put(exp);
898
899                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
900                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
901         }
902         OBD_FREE_PTR(lcd);
903
904         obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
905 out:
906         obd->u.obt.obt_mount_count = mount_count + 1;
907         obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
908         lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
909
910         /* save it, so mount count and last_transno is current */
911         rc = filter_update_server_data(obd);
912         if (rc)
913                 GOTO(err_client, rc);
914
915         RETURN(0);
916
917 err_client:
918         class_disconnect_exports(obd);
919 err_lut:
920         filter_free_server_data(&obd->u.obt);
921         RETURN(rc);
922 }
923
924 static int filter_cleanup_groups(struct obd_device *obd)
925 {
926         struct filter_obd *filter = &obd->u.filter;
927         struct file *filp;
928         struct dentry *dentry;
929         int i, j;
930         ENTRY;
931
932         if (filter->fo_dentry_O_groups != NULL) {
933                 for (i = 0; i < filter->fo_group_count; i++) {
934                         dentry = filter->fo_dentry_O_groups[i];
935                         if (dentry != NULL)
936                                 f_dput(dentry);
937                 }
938                 OBD_FREE(filter->fo_dentry_O_groups,
939                          filter->fo_group_count *
940                          sizeof(*filter->fo_dentry_O_groups));
941                 filter->fo_dentry_O_groups = NULL;
942         }
943         if (filter->fo_last_objid_files != NULL) {
944                 for (i = 0; i < filter->fo_group_count; i++) {
945                         filp = filter->fo_last_objid_files[i];
946                         if (filp != NULL)
947                                 filp_close(filp, 0);
948                 }
949                 OBD_FREE(filter->fo_last_objid_files,
950                          filter->fo_group_count *
951                          sizeof(*filter->fo_last_objid_files));
952                 filter->fo_last_objid_files = NULL;
953         }
954         if (filter->fo_dentry_O_sub != NULL) {
955                 for (i = 0; i < filter->fo_group_count; i++) {
956                         for (j = 0; j < filter->fo_subdir_count; j++) {
957                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
958                                 if (dentry != NULL)
959                                         f_dput(dentry);
960                         }
961                 }
962                 OBD_FREE(filter->fo_dentry_O_sub,
963                          filter->fo_group_count *
964                          sizeof(*filter->fo_dentry_O_sub));
965                 filter->fo_dentry_O_sub = NULL;
966         }
967         if (filter->fo_last_objids != NULL) {
968                 OBD_FREE(filter->fo_last_objids,
969                          filter->fo_group_count *
970                          sizeof(*filter->fo_last_objids));
971                 filter->fo_last_objids = NULL;
972         }
973         if (filter->fo_dentry_O != NULL) {
974                 f_dput(filter->fo_dentry_O);
975                 filter->fo_dentry_O = NULL;
976         }
977         RETURN(0);
978 }
979
980 static int filter_update_last_group(struct obd_device *obd, int group)
981 {
982         struct filter_obd *filter = &obd->u.filter;
983         struct file *filp = NULL;
984         int last_group = 0, rc;
985         loff_t off = 0;
986         ENTRY;
987
988         if (group <= filter->fo_committed_group)
989                 RETURN(0);
990
991         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
992         if (IS_ERR(filp)) {
993                 rc = PTR_ERR(filp);
994                 filp = NULL;
995                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
996                 GOTO(cleanup, rc);
997         }
998
999         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1000         if (rc) {
1001                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1002                 GOTO(cleanup, rc);
1003         }
1004
1005         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1006                obd->obd_name, last_group, group);
1007
1008         off = 0;
1009         last_group = group;
1010         /* must be sync: bXXXX */
1011         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1012         if (rc) {
1013                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1014                 GOTO(cleanup, rc);
1015         }
1016
1017         filter->fo_committed_group = group;
1018 cleanup:
1019         if (filp)
1020                 filp_close(filp, 0);
1021         RETURN(rc);
1022 }
1023
1024 static int filter_read_group_internal(struct obd_device *obd, int group,
1025                                       int create)
1026 {
1027         struct filter_obd *filter = &obd->u.filter;
1028         __u64 *new_objids = NULL;
1029         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1030         struct dentry **new_groups = NULL;
1031         struct file **new_files = NULL;
1032         struct dentry *dentry;
1033         struct file *filp;
1034         int old_count = filter->fo_group_count, rc, stage = 0, i;
1035         char name[25];
1036         __u64 last_objid;
1037         loff_t off = 0;
1038         int len = group + 1;
1039
1040         snprintf(name, 24, "%d", group);
1041         name[24] = '\0';
1042
1043         if (!create) {
1044                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1045                                            strlen(name));
1046                 if (IS_ERR(dentry)) {
1047                         CERROR("Cannot lookup expected object group %d: %ld\n",
1048                                group, PTR_ERR(dentry));
1049                         RETURN(PTR_ERR(dentry));
1050                 }
1051         } else {
1052                 dentry = simple_mkdir(filter->fo_dentry_O,
1053                                       obd->u.obt.obt_vfsmnt, name, 0700, 1);
1054                 if (IS_ERR(dentry)) {
1055                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1056                                PTR_ERR(dentry));
1057                         RETURN(PTR_ERR(dentry));
1058                 }
1059         }
1060         stage = 1;
1061
1062         snprintf(name, 24, "O/%d/LAST_ID", group);
1063         name[24] = '\0';
1064         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1065         if (IS_ERR(filp)) {
1066                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1067                 GOTO(cleanup, rc = PTR_ERR(filp));
1068         }
1069         stage = 2;
1070
1071         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1072         if (rc) {
1073                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1074                 GOTO(cleanup, rc);
1075         }
1076
1077         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1078                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1079                 if (tmp_subdirs == NULL)
1080                         GOTO(cleanup, rc = -ENOMEM);
1081                 stage = 3;
1082
1083                 for (i = 0; i < filter->fo_subdir_count; i++) {
1084                         char dir[20];
1085                         snprintf(dir, sizeof(dir), "d%u", i);
1086
1087                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1088                                                               obd->u.obt.obt_vfsmnt,
1089                                                               dir, 0700, 1);
1090                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1091                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1092                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1093                                        group, dir, rc);
1094                                 GOTO(cleanup, rc);
1095                         }
1096
1097                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1098                                tmp_subdirs->dentry[i]);
1099                 }
1100         }
1101
1102         /* 'group' is an index; we need an array of length 'group + 1' */
1103         if (group + 1 > old_count) {
1104                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1105                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1106                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1107                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1108                 stage = 4;
1109                 if (new_objids == NULL || new_subdirs == NULL ||
1110                     new_groups == NULL || new_files == NULL)
1111                         GOTO(cleanup, rc = -ENOMEM);
1112
1113                 if (old_count) {
1114                         memcpy(new_objids, filter->fo_last_objids,
1115                                old_count * sizeof(*new_objids));
1116                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1117                                old_count * sizeof(*new_subdirs));
1118                         memcpy(new_groups, filter->fo_dentry_O_groups,
1119                                old_count * sizeof(*new_groups));
1120                         memcpy(new_files, filter->fo_last_objid_files,
1121                                old_count * sizeof(*new_files));
1122
1123                         OBD_FREE(filter->fo_last_objids,
1124                                  old_count * sizeof(*new_objids));
1125                         OBD_FREE(filter->fo_dentry_O_sub,
1126                                  old_count * sizeof(*new_subdirs));
1127                         OBD_FREE(filter->fo_dentry_O_groups,
1128                                  old_count * sizeof(*new_groups));
1129                         OBD_FREE(filter->fo_last_objid_files,
1130                                  old_count * sizeof(*new_files));
1131                 }
1132                 filter->fo_last_objids = new_objids;
1133                 filter->fo_dentry_O_sub = new_subdirs;
1134                 filter->fo_dentry_O_groups = new_groups;
1135                 filter->fo_last_objid_files = new_files;
1136                 filter->fo_group_count = len;
1137         }
1138
1139         filter->fo_dentry_O_groups[group] = dentry;
1140         filter->fo_last_objid_files[group] = filp;
1141         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1142                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1143                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1144         }
1145
1146         filter_update_last_group(obd, group);
1147
1148         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1149                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1150                 rc = filter_update_last_objid(obd, group, 1);
1151                 RETURN(rc);
1152         }
1153
1154         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1155         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1156                obd->obd_name, group, last_objid);
1157         RETURN(0);
1158  cleanup:
1159         switch (stage) {
1160         case 4:
1161                 if (new_objids != NULL)
1162                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1163                 if (new_subdirs != NULL)
1164                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1165                 if (new_groups != NULL)
1166                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1167                 if (new_files != NULL)
1168                         OBD_FREE(new_files, len * sizeof(*new_files));
1169         case 3:
1170                 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1171                         for (i = 0; i < filter->fo_subdir_count; i++) {
1172                                 if (tmp_subdirs->dentry[i] != NULL)
1173                                         dput(tmp_subdirs->dentry[i]);
1174                         }
1175                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1176                 }
1177         case 2:
1178                 filp_close(filp, 0);
1179         case 1:
1180                 dput(dentry);
1181         }
1182         RETURN(rc);
1183 }
1184
1185 static int filter_read_groups(struct obd_device *obd, int last_group,
1186                               int create)
1187 {
1188         struct filter_obd *filter = &obd->u.filter;
1189         int old_count, group, rc = 0;
1190
1191         cfs_down(&filter->fo_init_lock);
1192         old_count = filter->fo_group_count;
1193         for (group = old_count; group <= last_group; group++) {
1194                 rc = filter_read_group_internal(obd, group, create);
1195                 if (rc != 0)
1196                         break;
1197         }
1198         cfs_up(&filter->fo_init_lock);
1199         return rc;
1200 }
1201
1202 /* FIXME: object groups */
1203 static int filter_prep_groups(struct obd_device *obd)
1204 {
1205         struct filter_obd *filter = &obd->u.filter;
1206         struct dentry *O_dentry;
1207         struct file *filp;
1208         int    last_group, rc = 0, cleanup_phase = 0;
1209         loff_t off = 0;
1210         ENTRY;
1211
1212         O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1213                                 "O", 0700, 1);
1214         CDEBUG(D_INODE, "%s: got/created O: %p\n", obd->obd_name, O_dentry);
1215         if (IS_ERR(O_dentry)) {
1216                 rc = PTR_ERR(O_dentry);
1217                 CERROR("%s: cannot open/create O: rc = %d\n", obd->obd_name,rc);
1218                 GOTO(cleanup, rc);
1219         }
1220         filter->fo_dentry_O = O_dentry;
1221         cleanup_phase = 1; /* O_dentry */
1222
1223         /* we have to initialize all groups before first connections from
1224          * clients because they may send create/destroy for any group -bzzz */
1225         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1226         if (IS_ERR(filp)) {
1227                 CERROR("%s: cannot create LAST_GROUP: rc = %ld\n",
1228                        obd->obd_name, PTR_ERR(filp));
1229                 GOTO(cleanup, rc = PTR_ERR(filp));
1230         }
1231         cleanup_phase = 2; /* filp */
1232
1233         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1234         if (rc) {
1235                 CERROR("%s: error reading LAST_GROUP: rc %d\n",
1236                        obd->obd_name, rc);
1237                 GOTO(cleanup, rc);
1238         }
1239
1240         if (off == 0)
1241                 last_group = FID_SEQ_OST_MDT0;
1242
1243         CDEBUG(D_INODE, "%s: initialize group %u (max %u)\n", obd->obd_name,
1244                FID_SEQ_OST_MDT0, last_group);
1245         filter->fo_committed_group = last_group;
1246         rc = filter_read_groups(obd, last_group, 1);
1247         if (rc)
1248                 GOTO(cleanup, rc);
1249
1250         filp_close(filp, 0);
1251         RETURN(0);
1252
1253  cleanup:
1254         switch (cleanup_phase) {
1255         case 2:
1256                 filp_close(filp, 0);
1257         case 1:
1258                 filter_cleanup_groups(obd);
1259                 f_dput(filter->fo_dentry_O);
1260                 filter->fo_dentry_O = NULL;
1261         default:
1262                 break;
1263         }
1264         return rc;
1265
1266 }
1267
1268 /* setup the object store with correct subdirectories */
1269 static int filter_prep(struct obd_device *obd)
1270 {
1271         struct lvfs_run_ctxt saved;
1272         struct filter_obd *filter = &obd->u.filter;
1273         struct file *file;
1274         struct inode *inode;
1275         int rc = 0;
1276         ENTRY;
1277
1278         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1279         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1280         if (!file || IS_ERR(file)) {
1281                 rc = PTR_ERR(file);
1282                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1283                        LAST_RCVD, rc);
1284                 GOTO(out, rc);
1285         }
1286         obd->u.obt.obt_rcvd_filp = file;
1287         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1288                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1289                        file->f_dentry->d_inode->i_mode);
1290                 GOTO(err_filp, rc = -ENOENT);
1291         }
1292
1293         inode = file->f_dentry->d_parent->d_inode;
1294         /* We use i_op->unlink directly in filter_vfs_unlink() */
1295         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1296                 CERROR("%s: filesystem does not support create/unlink ops\n",
1297                        obd->obd_name);
1298                 GOTO(err_filp, rc = -EOPNOTSUPP);
1299         }
1300
1301         rc = filter_init_server_data(obd, file);
1302         if (rc) {
1303                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1304                 GOTO(err_filp, rc);
1305         }
1306         LASSERT(obd->u.obt.obt_lut);
1307         target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1308
1309         /* open and create health check io file*/
1310         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1311         if (IS_ERR(file)) {
1312                 rc = PTR_ERR(file);
1313                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1314                        HEALTH_CHECK, rc);
1315                 GOTO(err_server_data, rc);
1316         }
1317         filter->fo_obt.obt_health_check_filp = file;
1318         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1319                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1320                        file->f_dentry->d_inode->i_mode);
1321                 GOTO(err_health_check, rc = -ENOENT);
1322         }
1323         rc = lvfs_check_io_health(obd, file);
1324         if (rc)
1325                 GOTO(err_health_check, rc);
1326
1327         rc = filter_prep_groups(obd);
1328         if (rc)
1329                 GOTO(err_health_check, rc);
1330 out:
1331         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1332
1333         return(rc);
1334
1335 err_health_check:
1336         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1337                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1338         filter->fo_obt.obt_health_check_filp = NULL;
1339 err_server_data:
1340         target_recovery_fini(obd);
1341         filter_free_server_data(&obd->u.obt);
1342 err_filp:
1343         if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1344                 CERROR("can't close %s after error\n", LAST_RCVD);
1345         obd->u.obt.obt_rcvd_filp = NULL;
1346         goto out;
1347 }
1348
1349 /* cleanup the filter: write last used object id to status file */
1350 static void filter_post(struct obd_device *obd)
1351 {
1352         struct lvfs_run_ctxt saved;
1353         struct filter_obd *filter = &obd->u.filter;
1354         int rc, i;
1355
1356         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1357          * best to start a transaction with h_sync, because we removed this
1358          * from lastobjid */
1359
1360         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1361         rc = filter_update_server_data(obd);
1362         if (rc)
1363                 CERROR("error writing server data: rc = %d\n", rc);
1364
1365         for (i = 0; i < filter->fo_group_count; i++) {
1366                 rc = filter_update_last_objid(obd, i,
1367                                 (i == filter->fo_group_count - 1));
1368                 if (rc)
1369                         CERROR("error writing group %d lastobjid: rc = %d\n",
1370                                i, rc);
1371         }
1372
1373         rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1374         obd->u.obt.obt_rcvd_filp = NULL;
1375         if (rc)
1376                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1377
1378         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1379         filter->fo_obt.obt_health_check_filp = NULL;
1380         if (rc)
1381                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1382
1383         filter_cleanup_groups(obd);
1384         filter_free_server_data(&obd->u.obt);
1385         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1386
1387         filter_free_capa_keys(filter);
1388         cleanup_capa_hash(filter->fo_capa_hash);
1389 }
1390
1391 static void filter_set_last_id(struct filter_obd *filter,
1392                                obd_id id, obd_seq group)
1393 {
1394         LASSERT(group <= filter->fo_group_count);
1395
1396         cfs_spin_lock(&filter->fo_objidlock);
1397         filter->fo_last_objids[group] = id;
1398         cfs_spin_unlock(&filter->fo_objidlock);
1399 }
1400
1401 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1402 {
1403         obd_id id;
1404         LASSERT(group <= filter->fo_group_count);
1405         LASSERT(filter->fo_last_objids != NULL);
1406
1407         /* FIXME: object groups */
1408         cfs_spin_lock(&filter->fo_objidlock);
1409         id = filter->fo_last_objids[group];
1410         cfs_spin_unlock(&filter->fo_objidlock);
1411         return id;
1412 }
1413
1414 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1415 {
1416         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1417         return 0;
1418 }
1419
1420 /* We never dget the object parent, so DON'T dput it either */
1421 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1422 {
1423         struct filter_obd *filter = &obd->u.filter;
1424         struct filter_subdirs *subdirs;
1425
1426         if (group >= filter->fo_group_count) /* FIXME: object groups */
1427                 return ERR_PTR(-EBADF);
1428
1429         if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1430                 return filter->fo_dentry_O_groups[group];
1431
1432         subdirs = &filter->fo_dentry_O_sub[group];
1433         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1434 }
1435
1436 /* We never dget the object parent, so DON'T dput it either */
1437 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1438                                   obd_id objid)
1439 {
1440         unsigned long now = jiffies;
1441         struct dentry *dparent = filter_parent(obd, group, objid);
1442         int rc;
1443
1444         if (IS_ERR(dparent))
1445                 return dparent;
1446         if (dparent == NULL)
1447                 return ERR_PTR(-ENOENT);
1448
1449         rc = filter_lock_dentry(obd, dparent);
1450         fsfilt_check_slow(obd, now, "parent lock");
1451         return rc ? ERR_PTR(rc) : dparent;
1452 }
1453
1454 /* We never dget the object parent, so DON'T dput it either */
1455 static void filter_parent_unlock(struct dentry *dparent)
1456 {
1457         UNLOCK_INODE_MUTEX(dparent->d_inode);
1458 }
1459
1460 /* How to get files, dentries, inodes from object id's.
1461  *
1462  * If dir_dentry is passed, the caller has already locked the parent
1463  * appropriately for this operation (normally a write lock).  If
1464  * dir_dentry is NULL, we do a read lock while we do the lookup to
1465  * avoid races with create/destroy and such changing the directory
1466  * internal to the filesystem code. */
1467 struct dentry *filter_fid2dentry(struct obd_device *obd,
1468                                  struct dentry *dir_dentry,
1469                                  obd_seq group, obd_id id)
1470 {
1471         struct dentry *dparent = dir_dentry;
1472         struct dentry *dchild;
1473         char name[32];
1474         int len;
1475         ENTRY;
1476
1477         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1478             obd->u.filter.fo_destroys_in_progress == 0) {
1479                 /* don't fail lookups for orphan recovery, it causes
1480                  * later LBUGs when objects still exist during precreate */
1481                 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1482                 RETURN(ERR_PTR(-ENOENT));
1483         }
1484         if (id == 0) {
1485                 CERROR("fatal: invalid object id 0\n");
1486                 RETURN(ERR_PTR(-ESTALE));
1487         }
1488
1489         len = sprintf(name, LPU64, id);
1490         if (dir_dentry == NULL) {
1491                 dparent = filter_parent_lock(obd, group, id);
1492                 if (IS_ERR(dparent)) {
1493                         CERROR("%s: error getting object "POSTID
1494                                " parent: rc %ld\n", obd->obd_name,
1495                                id, group, PTR_ERR(dparent));
1496                         RETURN(dparent);
1497                 }
1498         }
1499         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1500                dparent->d_name.len, dparent->d_name.name, name);
1501         /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1502         dchild = lookup_one_len(name, dparent, len);
1503         if (dir_dentry == NULL)
1504                 filter_parent_unlock(dparent);
1505         if (IS_ERR(dchild)) {
1506                 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1507                        obd->obd_name, id, group, PTR_ERR(dchild));
1508                 RETURN(dchild);
1509         }
1510
1511         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1512                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1513                        obd->obd_name, id, dchild->d_inode->i_ino);
1514                 f_dput(dchild);
1515                 RETURN(ERR_PTR(-ENOENT));
1516         }
1517
1518         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1519                name, dchild, atomic_read(&dchild->d_count));
1520
1521         LASSERT(atomic_read(&dchild->d_count) > 0);
1522
1523         RETURN(dchild);
1524 }
1525
1526 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1527                                   obd_id group, struct lustre_handle *lockh)
1528 {
1529         int flags = LDLM_AST_DISCARD_DATA, rc;
1530         struct ldlm_res_id res_id;
1531         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1532         ENTRY;
1533
1534         osc_build_res_name(objid, group, &res_id);
1535         /* Tell the clients that the object is gone now and that they should
1536          * throw away any cached pages. */
1537         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1538                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1539                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1540                                     lockh);
1541         if (rc != ELDLM_OK)
1542                 lockh->cookie = 0;
1543         RETURN(rc);
1544 }
1545
1546 static void filter_fini_destroy(struct obd_device *obd,
1547                                 struct lustre_handle *lockh)
1548 {
1549         if (lustre_handle_is_used(lockh))
1550                 ldlm_lock_decref(lockh, LCK_PW);
1551 }
1552
1553 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1554  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1555  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1556  * reverse.  Caller must take i_sem before starting the transaction and we
1557  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1558 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1559                       struct vfsmount *mnt)
1560 {
1561         int rc;
1562         ENTRY;
1563
1564         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1565          * itself we already hold dir->i_mutex for child create/unlink ops */
1566         LASSERT(dentry->d_inode != NULL);
1567         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1568         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1569
1570
1571         /* may_delete() */
1572         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1573                 GOTO(out, rc = -ENOENT);
1574
1575         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1576         if (rc)
1577                 GOTO(out, rc);
1578
1579         if (IS_APPEND(dir))
1580                 GOTO(out, rc = -EPERM);
1581
1582         /* check_sticky() */
1583         if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1584              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1585             IS_IMMUTABLE(dentry->d_inode))
1586                 GOTO(out, rc = -EPERM);
1587
1588         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1589          *       that was done because of journal_start (which is already done
1590          *       here) or some other ordering issue. */
1591         ll_vfs_dq_init(dir);
1592
1593         rc = ll_security_inode_unlink(dir, dentry, mnt);
1594         if (rc)
1595                 GOTO(out, rc);
1596
1597         rc = dir->i_op->unlink(dir, dentry);
1598 out:
1599         /* need to drop i_mutex before we lose inode reference */
1600         UNLOCK_INODE_MUTEX(dentry->d_inode);
1601         if (rc == 0)
1602                 d_delete(dentry);
1603
1604         RETURN(rc);
1605 }
1606
1607 /* Caller must hold LCK_PW on parent and push us into kernel context.
1608  * Caller must hold child i_mutex, we drop it always.
1609  * Caller is also required to ensure that dchild->d_inode exists. */
1610 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1611                                    obd_seq group, struct dentry *dparent,
1612                                    struct dentry *dchild)
1613 {
1614         struct inode *inode = dchild->d_inode;
1615         int rc;
1616
1617         /* There should be 2 references to the inode:
1618          *  1) taken by filter_prepare_destroy
1619          *  2) taken by filter_destroy */
1620         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1621                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1622                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1623                        (unsigned long)inode->i_nlink,
1624                        atomic_read(&inode->i_count));
1625         }
1626
1627         rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1628         if (rc)
1629                 CERROR("error unlinking objid %.*s: rc %d\n",
1630                        dchild->d_name.len, dchild->d_name.name, rc);
1631         return(rc);
1632 }
1633
1634 struct filter_intent_args {
1635         struct ldlm_lock **victim;
1636         __u64 size;
1637         int *liblustre;
1638 };
1639
1640 static enum interval_iter filter_intent_cb(struct interval_node *n,
1641                                            void *args)
1642 {
1643         struct ldlm_interval *node = (struct ldlm_interval *)n;
1644         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1645         __u64 size = arg->size;
1646         struct ldlm_lock **v = arg->victim;
1647         struct ldlm_lock *lck;
1648
1649         /* If the interval is lower than the current file size,
1650          * just break. */
1651         if (interval_high(n) <= size)
1652                 return INTERVAL_ITER_STOP;
1653
1654         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1655                 /* Don't send glimpse ASTs to liblustre clients.
1656                  * They aren't listening for them, and they do
1657                  * entirely synchronous I/O anyways. */
1658                 if (lck->l_export == NULL ||
1659                     lck->l_export->exp_libclient == 1)
1660                         continue;
1661
1662                 if (*arg->liblustre)
1663                         *arg->liblustre = 0;
1664
1665                 if (*v == NULL) {
1666                         *v = LDLM_LOCK_GET(lck);
1667                 } else if ((*v)->l_policy_data.l_extent.start <
1668                            lck->l_policy_data.l_extent.start) {
1669                         LDLM_LOCK_RELEASE(*v);
1670                         *v = LDLM_LOCK_GET(lck);
1671                 }
1672
1673                 /* the same policy group - every lock has the
1674                  * same extent, so needn't do it any more */
1675                 break;
1676         }
1677
1678         return INTERVAL_ITER_CONT;
1679 }
1680
1681 static int filter_intent_policy(struct ldlm_namespace *ns,
1682                                 struct ldlm_lock **lockp, void *req_cookie,
1683                                 ldlm_mode_t mode, int flags, void *data)
1684 {
1685         CFS_LIST_HEAD(rpc_list);
1686         struct ptlrpc_request *req = req_cookie;
1687         struct ldlm_lock *lock = *lockp, *l = NULL;
1688         struct ldlm_resource *res = lock->l_resource;
1689         ldlm_processing_policy policy;
1690         struct ost_lvb *res_lvb, *reply_lvb;
1691         struct ldlm_reply *rep;
1692         ldlm_error_t err;
1693         int idx, rc, tmpflags = 0, only_liblustre = 1;
1694         struct ldlm_interval_tree *tree;
1695         struct filter_intent_args arg;
1696         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1697                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1698                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1699         ENTRY;
1700
1701         policy = ldlm_get_processing_policy(res);
1702         LASSERT(policy != NULL);
1703         LASSERT(req != NULL);
1704
1705         rc = lustre_pack_reply(req, 3, repsize, NULL);
1706         if (rc)
1707                 RETURN(req->rq_status = rc);
1708
1709         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1710         LASSERT(rep != NULL);
1711
1712         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1713                                    sizeof(*reply_lvb));
1714         LASSERT(reply_lvb != NULL);
1715
1716         //fixup_handle_for_resent_req(req, lock, &lockh);
1717
1718         /* Call the extent policy function to see if our request can be
1719          * granted, or is blocked.
1720          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1721          * lock, and should not be granted if the lock will be blocked.
1722          */
1723
1724         LASSERT(ns == ldlm_res_to_ns(res));
1725         lock_res(res);
1726         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1727         check_res_locked(res);
1728
1729         /* FIXME: we should change the policy function slightly, to not make
1730          * this list at all, since we just turn around and free it */
1731         while (!cfs_list_empty(&rpc_list)) {
1732                 struct ldlm_lock *wlock =
1733                         cfs_list_entry(rpc_list.next, struct ldlm_lock,
1734                                        l_cp_ast);
1735                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1736                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1737                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1738                 cfs_list_del_init(&wlock->l_cp_ast);
1739                 LDLM_LOCK_RELEASE(wlock);
1740         }
1741
1742         /* The lock met with no resistance; we're finished. */
1743         if (rc == LDLM_ITER_CONTINUE) {
1744                 /* do not grant locks to the liblustre clients: they cannot
1745                  * handle ASTs robustly.  We need to do this while still
1746                  * holding lr_lock to avoid the lock remaining on the res_link
1747                  * list (and potentially being added to l_pending_list by an
1748                  * AST) when we are going to drop this lock ASAP. */
1749                 if (lock->l_export->exp_libclient ||
1750                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1751                         ldlm_resource_unlink_lock(lock);
1752                         err = ELDLM_LOCK_ABORTED;
1753                 } else {
1754                         err = ELDLM_LOCK_REPLACED;
1755                 }
1756                 unlock_res(res);
1757                 RETURN(err);
1758         }
1759
1760         /* Do not grant any lock, but instead send GL callbacks.  The extent
1761          * policy nicely created a list of all PW locks for us.  We will choose
1762          * the highest of those which are larger than the size in the LVB, if
1763          * any, and perform a glimpse callback. */
1764         res_lvb = res->lr_lvb_data;
1765         LASSERT(res_lvb != NULL);
1766         *reply_lvb = *res_lvb;
1767
1768         /*
1769          * lr_lock guarantees that no new locks are granted, and,
1770          * therefore, that res->lr_lvb_data cannot increase beyond the
1771          * end of already granted lock. As a result, it is safe to
1772          * check against "stale" reply_lvb->lvb_size value without
1773          * res->lr_lvb_sem.
1774          */
1775         arg.size = reply_lvb->lvb_size;
1776         arg.victim = &l;
1777         arg.liblustre = &only_liblustre;
1778         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1779                 tree = &res->lr_itree[idx];
1780                 if (tree->lit_mode == LCK_PR)
1781                         continue;
1782
1783                 interval_iterate_reverse(tree->lit_root,
1784                                          filter_intent_cb, &arg);
1785         }
1786         unlock_res(res);
1787
1788         /* There were no PW locks beyond the size in the LVB; finished. */
1789         if (l == NULL) {
1790                 if (only_liblustre) {
1791                         /* If we discovered a liblustre client with a PW lock,
1792                          * however, the LVB may be out of date!  The LVB is
1793                          * updated only on glimpse (which we don't do for
1794                          * liblustre clients) and cancel (which the client
1795                          * obviously has not yet done).  So if it has written
1796                          * data but kept the lock, the LVB is stale and needs
1797                          * to be updated from disk.
1798                          *
1799                          * Of course, this will all disappear when we switch to
1800                          * taking liblustre locks on the OST. */
1801                         ldlm_res_lvbo_update(res, NULL, 1);
1802                 }
1803                 RETURN(ELDLM_LOCK_ABORTED);
1804         }
1805
1806         /*
1807          * This check is for lock taken in filter_prepare_destroy() that does
1808          * not have l_glimpse_ast set. So the logic is: if there is a lock
1809          * with no l_glimpse_ast set, this object is being destroyed already.
1810          *
1811          * Hence, if you are grabbing DLM locks on the server, always set
1812          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1813          */
1814         if (l->l_glimpse_ast == NULL) {
1815                 /* We are racing with unlink(); just return -ENOENT */
1816                 rep->lock_policy_res1 = -ENOENT;
1817                 goto out;
1818         }
1819
1820         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1821         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1822
1823         lock_res(res);
1824         *reply_lvb = *res_lvb;
1825         unlock_res(res);
1826
1827  out:
1828         LDLM_LOCK_RELEASE(l);
1829
1830         RETURN(ELDLM_LOCK_ABORTED);
1831 }
1832
1833 /*
1834  * per-obd_device iobuf pool.
1835  *
1836  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1837  * allocations in write-path has to be minimized (see bug 5137).
1838  *
1839  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1840  * OST threads (see ost_thread_{init,done}()).
1841  *
1842  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1843  * at the OST layer there are only (potentially) multiple obd_device of type
1844  * unknown at the time of OST thread creation.
1845  *
1846  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1847  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1848  * it's very own iobuf.
1849  *
1850  * Functions below
1851  *
1852  *     filter_kiobuf_pool_init()
1853  *
1854  *     filter_kiobuf_pool_done()
1855  *
1856  *     filter_iobuf_get()
1857  *
1858  * operate on this array. They are "generic" in a sense that they don't depend
1859  * on actual type of iobuf's (the latter depending on Linux kernel version).
1860  */
1861
1862 /*
1863  * destroy pool created by filter_iobuf_pool_init
1864  */
1865 static void filter_iobuf_pool_done(struct filter_obd *filter)
1866 {
1867         struct filter_iobuf **pool;
1868         int i;
1869
1870         ENTRY;
1871
1872         pool = filter->fo_iobuf_pool;
1873         if (pool != NULL) {
1874                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1875                         if (pool[i] != NULL)
1876                                 filter_free_iobuf(pool[i]);
1877                 }
1878                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1879                 filter->fo_iobuf_pool = NULL;
1880         }
1881         EXIT;
1882 }
1883
1884 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1885 {
1886         struct filter_obd       *filter = &obd->u.filter;
1887         struct sptlrpc_rule_set  tmp_rset;
1888         int                      rc;
1889
1890         sptlrpc_rule_set_init(&tmp_rset);
1891         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1892         if (rc) {
1893                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1894                        obd->obd_name, rc);
1895                 return rc;
1896         }
1897
1898         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1899
1900         cfs_write_lock(&filter->fo_sptlrpc_lock);
1901         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1902         filter->fo_sptlrpc_rset = tmp_rset;
1903         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1904
1905         return 0;
1906 }
1907
1908 /*
1909  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1910  */
1911 static int filter_iobuf_pool_init(struct filter_obd *filter)
1912 {
1913         void **pool;
1914
1915         ENTRY;
1916
1917
1918         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1919                       GFP_KERNEL);
1920         if (filter->fo_iobuf_pool == NULL)
1921                 RETURN(-ENOMEM);
1922
1923         filter->fo_iobuf_count = OSS_THREADS_MAX;
1924
1925         RETURN(0);
1926 }
1927
1928 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1929  * many threads there will be so we allocate a large empty array and only
1930  * fill in those slots that are actually in use.
1931  * If we haven't allocated a pool entry for this thread before, do so now. */
1932 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1933 {
1934         int thread_id                    = (oti && oti->oti_thread) ?
1935                                            oti->oti_thread->t_id : -1;
1936         struct filter_iobuf  *pool       = NULL;
1937         struct filter_iobuf **pool_place = NULL;
1938
1939         if (thread_id >= 0) {
1940                 LASSERT(thread_id < filter->fo_iobuf_count);
1941                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1942         }
1943
1944         if (unlikely(pool == NULL)) {
1945                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1946                                           PTLRPC_MAX_BRW_PAGES);
1947                 if (pool_place != NULL)
1948                         *pool_place = pool;
1949         }
1950
1951         return pool;
1952 }
1953
1954 /* mount the file system (secretly).  lustre_cfg parameters are:
1955  * 1 = device
1956  * 2 = fstype
1957  * 3 = flags: failover=f, failout=n
1958  * 4 = mount options
1959  */
1960 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1961                         void *option)
1962 {
1963         struct filter_obd *filter = &obd->u.filter;
1964         struct vfsmount *mnt;
1965         struct lustre_mount_info *lmi;
1966         struct obd_uuid uuid;
1967         __u8 *uuid_ptr;
1968         char *str, *label;
1969         char ns_name[48];
1970         struct request_queue *q;
1971         int rc, i;
1972         ENTRY;
1973
1974         if (lcfg->lcfg_bufcount < 3 ||
1975             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1976             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1977                 RETURN(-EINVAL);
1978
1979         lmi = server_get_mount(obd->obd_name);
1980         if (lmi) {
1981                 /* We already mounted in lustre_fill_super.
1982                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1983                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1984                 mnt = lmi->lmi_mnt;
1985                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1986         } else {
1987                 /* old path - used by lctl */
1988                 CERROR("Using old MDS mount method\n");
1989                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1990                                     MS_NOATIME|MS_NODIRATIME,
1991                                     lustre_cfg_string(lcfg, 1), option);
1992                 if (IS_ERR(mnt)) {
1993                         rc = PTR_ERR(mnt);
1994                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1995                                            lustre_cfg_string(lcfg, 1), rc);
1996                         RETURN(rc);
1997                 }
1998
1999                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2000         }
2001         if (IS_ERR(obd->obd_fsops))
2002                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2003
2004         rc = filter_iobuf_pool_init(filter);
2005         if (rc != 0)
2006                 GOTO(err_ops, rc);
2007
2008         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2009                 CERROR("%s: Underlying device is marked as read-only. "
2010                        "Setup failed\n", obd->obd_name);
2011                 GOTO(err_ops, rc = -EROFS);
2012         }
2013
2014         /* failover is the default */
2015         obd->obd_replayable = 1;
2016
2017         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2018                 str = lustre_cfg_string(lcfg, 3);
2019                 if (strchr(str, 'n')) {
2020                         CWARN("%s: recovery disabled\n", obd->obd_name);
2021                         obd->obd_replayable = 0;
2022                 }
2023         }
2024
2025         obd->u.obt.obt_magic = OBT_MAGIC;
2026         obd->u.obt.obt_vfsmnt = mnt;
2027         obd->u.obt.obt_sb = mnt->mnt_sb;
2028         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2029         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2030
2031         fsfilt_setup(obd, obd->u.obt.obt_sb);
2032
2033         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2034         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2035         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2036         obd->obd_lvfs_ctxt.fs = get_ds();
2037         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2038
2039         cfs_init_mutex(&filter->fo_init_lock);
2040         filter->fo_committed_group = 0;
2041         filter->fo_destroys_in_progress = 0;
2042         for (i = 0; i < 32; i++)
2043                 cfs_sema_init(&filter->fo_create_locks[i], 1);
2044
2045         cfs_spin_lock_init(&filter->fo_objidlock);
2046         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2047         cfs_sema_init(&filter->fo_alloc_lock, 1);
2048         init_brw_stats(&filter->fo_filter_stats);
2049         cfs_spin_lock_init(&filter->fo_flags_lock);
2050         filter->fo_read_cache = 1; /* enable read-only cache by default */
2051         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2052         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2053         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2054         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2055         filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2056         filter_slc_set(filter); /* initialize sync on lock cancel */
2057
2058         rc = filter_prep(obd);
2059         if (rc)
2060                 GOTO(err_ops, rc);
2061
2062         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2063         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2064
2065         filter->fo_fl_oss_capa = 1;
2066
2067         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2068         filter->fo_capa_hash = init_capa_hash();
2069         if (filter->fo_capa_hash == NULL)
2070                 GOTO(err_post, rc = -ENOMEM);
2071
2072         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2073         obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2074                                                 LDLM_NAMESPACE_SERVER,
2075                                                 LDLM_NAMESPACE_GREEDY,
2076                                                 LDLM_NS_TYPE_OST);
2077         if (obd->obd_namespace == NULL)
2078                 GOTO(err_post, rc = -ENOMEM);
2079         obd->obd_namespace->ns_lvbp = obd;
2080         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2081         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2082
2083         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2084                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2085
2086         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2087         if (rc) {
2088                 CERROR("failed to setup llogging subsystems\n");
2089                 GOTO(err_post, rc);
2090         }
2091
2092         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2093         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2094         /* do this after llog being initialized */
2095         filter_adapt_sptlrpc_conf(obd, 1);
2096
2097         rc = lquota_setup(filter_quota_interface_ref, obd);
2098         if (rc)
2099                 GOTO(err_post, rc);
2100
2101         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2102         if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2103             queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2104                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2105                               "for larger I/O requests: max_sectors = %u "
2106                               "could be up to max_hw_sectors=%u\n",
2107                               obd->obd_name, mnt->mnt_sb->s_id,
2108                               queue_max_sectors(q), queue_max_hw_sectors(q));
2109
2110         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2111         if (uuid_ptr != NULL) {
2112                 class_uuid_unparse(uuid_ptr, &uuid);
2113                 str = uuid.uuid;
2114         } else {
2115                 str = "no UUID";
2116         }
2117
2118         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2119         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2120                       obd->obd_name, label ?: str, lmi ? "on " : "",
2121                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2122                       obd->obd_replayable ? "enabled" : "disabled");
2123
2124         if (obd->obd_recovering)
2125                 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2126                               "or until %d client%s reconnect%s\n",
2127                               obd->obd_name,
2128                               obd->obd_recovery_timeout / 60,
2129                               obd->obd_recovery_timeout % 60,
2130                               obd->obd_max_recoverable_clients,
2131                               (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2132                               (obd->obd_max_recoverable_clients == 1) ? "s": "");
2133
2134
2135         RETURN(0);
2136
2137 err_post:
2138         filter_post(obd);
2139 err_ops:
2140         fsfilt_put_ops(obd->obd_fsops);
2141         filter_iobuf_pool_done(filter);
2142 err_mntput:
2143         server_put_mount(obd->obd_name, mnt);
2144         obd->u.obt.obt_sb = 0;
2145         return rc;
2146 }
2147
2148 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2149 {
2150         struct lprocfs_static_vars lvars;
2151         cfs_proc_dir_entry_t *entry;
2152         unsigned long addr;
2153         struct page *page;
2154         int rc;
2155         ENTRY;
2156
2157         CLASSERT(offsetof(struct obd_device, u.obt) ==
2158                  offsetof(struct obd_device, u.filter.fo_obt));
2159
2160         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2161                 RETURN(-EINVAL);
2162
2163         /* lprocfs must be setup before the filter so state can be safely added
2164          * to /proc incrementally as the filter is setup */
2165         lprocfs_filter_init_vars(&lvars);
2166         rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2167         if (rc) {
2168                 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2169                        obd->obd_name, rc);
2170                 RETURN(rc);
2171         }
2172
2173         rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2174         if (rc) {
2175                 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2176                        obd->obd_name, rc);
2177                 GOTO(obd_cleanup, rc);
2178         }
2179
2180         /* Init obdfilter private stats here */
2181         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2182                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2183         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2184                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2185         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2186                              LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2187                              "get_page", "usec");
2188         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2189                              LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2190         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2191                              LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2192         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2193                              LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2194         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2195                              LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2196
2197         rc = lproc_filter_attach_seqstat(obd);
2198         if (rc) {
2199                 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2200                 GOTO(free_obd_stats, rc);
2201         }
2202
2203         entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2204         if (IS_ERR(entry)) {
2205                 rc = PTR_ERR(entry);
2206                 CERROR("%s: error %d setting up lprocfs for %s\n",
2207                        obd->obd_name, rc, "exports");
2208                 GOTO(free_obd_stats, rc);
2209         }
2210         obd->obd_proc_exports_entry = entry;
2211
2212         entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2213                                    lprocfs_nid_stats_clear_read,
2214                                    lprocfs_nid_stats_clear_write, obd, NULL);
2215         if (IS_ERR(entry)) {
2216                 rc = PTR_ERR(entry);
2217                 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2218                        obd->obd_name, rc);
2219                 GOTO(free_obd_stats, rc);
2220         }
2221
2222         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2223         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2224         if (!page)
2225                 GOTO(remove_entry_clear, rc = -ENOMEM);
2226         addr = (unsigned long)cfs_page_address(page);
2227         clear_page((void *)addr);
2228         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2229                LUSTRE_CFG_BUFLEN(lcfg, 4));
2230         rc = filter_common_setup(obd, lcfg, (void *)addr);
2231         OBD_PAGE_FREE(page);
2232         if (rc) {
2233                 CERROR("%s: filter_common_setup failed: %d.\n",
2234                        obd->obd_name, rc);
2235                 GOTO(remove_entry_clear, rc);
2236         }
2237
2238         RETURN(0);
2239
2240 remove_entry_clear:
2241         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2242 free_obd_stats:
2243         lprocfs_free_obd_stats(obd);
2244 obd_cleanup:
2245         lprocfs_obd_cleanup(obd);
2246         return rc;
2247 }
2248
2249 static struct llog_operations filter_mds_ost_repl_logops;
2250
2251 static struct llog_operations filter_size_orig_logops = {
2252         .lop_setup   = llog_obd_origin_setup,
2253         .lop_cleanup = llog_obd_origin_cleanup,
2254         .lop_add     = llog_obd_origin_add
2255 };
2256
2257 static int filter_olg_fini(struct obd_llog_group *olg)
2258 {
2259         struct llog_ctxt *ctxt;
2260         int rc = 0, rc2 = 0;
2261         ENTRY;
2262
2263         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2264         if (ctxt)
2265                 rc = llog_cleanup(ctxt);
2266
2267         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2268         if (ctxt) {
2269                 rc2 = llog_cleanup(ctxt);
2270                 if (!rc)
2271                         rc = rc2;
2272         }
2273
2274         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2275         if (ctxt) {
2276                 rc2 = llog_cleanup(ctxt);
2277                 if (!rc)
2278                         rc = rc2;
2279         }
2280
2281         RETURN(rc);
2282 }
2283
2284 static int
2285 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2286                 struct obd_device *tgt)
2287 {
2288         int rc;
2289         ENTRY;
2290
2291         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2292                         &filter_mds_ost_repl_logops);
2293         if (rc)
2294                 GOTO(cleanup, rc);
2295
2296         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2297                         &filter_size_orig_logops);
2298         if (rc)
2299                 GOTO(cleanup, rc);
2300         EXIT;
2301 cleanup:
2302         if (rc)
2303                 filter_olg_fini(olg);
2304         return rc;
2305 }
2306
2307 /**
2308  * Init the default olg, which is embeded in the obd_device, for filter.
2309  */
2310 static int
2311 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2312                         struct obd_device *tgt)
2313 {
2314         struct filter_obd *filter = &obd->u.filter;
2315         struct llog_ctxt *ctxt;
2316         int rc;
2317         ENTRY;
2318
2319         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2320         if (!filter->fo_lcm)
2321                 RETURN(-ENOMEM);
2322
2323         filter_mds_ost_repl_logops = llog_client_ops;
2324         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2325         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2326         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2327
2328         rc = filter_olg_init(obd, olg, tgt);
2329         if (rc)
2330                 GOTO(cleanup_lcm, rc);
2331
2332         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2333                         &llog_lvfs_ops);
2334         if (rc)
2335                 GOTO(cleanup_olg, rc);
2336
2337         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2338         if (!ctxt) {
2339                 CERROR("Can't get ctxt for %p:%x\n", olg,
2340                        LLOG_MDS_OST_REPL_CTXT);
2341                 GOTO(cleanup_olg, rc = -ENODEV);
2342         }
2343         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2344         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2345         llog_ctxt_put(ctxt);
2346
2347         RETURN(0);
2348 cleanup_olg:
2349         filter_olg_fini(olg);
2350 cleanup_lcm:
2351         llog_recov_thread_fini(filter->fo_lcm, 1);
2352         filter->fo_lcm = NULL;
2353         return rc;
2354 }
2355
2356 static int
2357 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2358                  struct obd_device *tgt, int *index)
2359 {
2360         struct filter_obd *filter = &obd->u.filter;
2361         struct llog_ctxt *ctxt;
2362         int rc;
2363         ENTRY;
2364
2365         LASSERT(olg != NULL);
2366         if (olg == &obd->obd_olg)
2367                 return filter_default_olg_init(obd, olg, tgt);
2368
2369         LASSERT(filter->fo_lcm != NULL);
2370         rc = filter_olg_init(obd, olg, tgt);
2371         if (rc)
2372                 RETURN(rc);
2373         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2374         if (!ctxt) {
2375                 CERROR("Can't get ctxt for %p:%x\n", olg,
2376                        LLOG_MDS_OST_REPL_CTXT);
2377                 filter_olg_fini(olg);
2378                 RETURN(-ENODEV);
2379         }
2380         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2381         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2382         llog_ctxt_put(ctxt);
2383         RETURN(rc);
2384 }
2385
2386 static int filter_llog_finish(struct obd_device *obd, int count)
2387 {
2388         struct filter_obd *filter = &obd->u.filter;
2389         struct llog_ctxt *ctxt;
2390         ENTRY;
2391
2392         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2393         if (ctxt) {
2394                 /*
2395                  * Make sure that no cached llcds left in recov_thread.
2396                  * We actually do sync in disconnect time, but disconnect
2397                  * may not come being marked rq_no_resend = 1.
2398                  */
2399                 llog_sync(ctxt, NULL);
2400
2401                 /*
2402                  * Balance class_import_get() in llog_receptor_accept().
2403                  * This is safe to do, as llog is already synchronized
2404                  * and its import may go.
2405                  */
2406                 cfs_mutex_down(&ctxt->loc_sem);
2407                 if (ctxt->loc_imp) {
2408                         class_import_put(ctxt->loc_imp);
2409                         ctxt->loc_imp = NULL;
2410                 }
2411                 cfs_mutex_up(&ctxt->loc_sem);
2412                 llog_ctxt_put(ctxt);
2413         }
2414
2415         if (filter->fo_lcm) {
2416                 cfs_mutex_down(&ctxt->loc_sem);
2417                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2418                 filter->fo_lcm = NULL;
2419                 cfs_mutex_up(&ctxt->loc_sem);
2420         }
2421         RETURN(filter_olg_fini(&obd->obd_olg));
2422 }
2423
2424 /**
2425  * Find the group llog according to group index in the llog group list.
2426  */
2427 static struct obd_llog_group *
2428 filter_find_olg_internal(struct filter_obd *filter, int group)
2429 {
2430         struct obd_llog_group *olg;
2431
2432         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2433         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2434                 if (olg->olg_seq == group)
2435                         RETURN(olg);
2436         }
2437         RETURN(NULL);
2438 }
2439
2440 /**
2441  * Find the group llog according to group index on the filter
2442  */
2443 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2444 {
2445         struct obd_llog_group *olg = NULL;
2446         struct filter_obd *filter;
2447
2448         filter = &obd->u.filter;
2449
2450         if (group == FID_SEQ_LLOG)
2451                 RETURN(&obd->obd_olg);
2452
2453         cfs_spin_lock(&filter->fo_llog_list_lock);
2454         olg = filter_find_olg_internal(filter, group);
2455         cfs_spin_unlock(&filter->fo_llog_list_lock);
2456
2457         RETURN(olg);
2458 }
2459 /**
2460  * Find the llog_group of the filter according to the group. If it can not
2461  * find, create the llog_group, which only happens when mds is being synced
2462  * with OST.
2463  */
2464 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2465 {
2466         struct obd_llog_group *olg = NULL, *olg_new = NULL;
2467         struct filter_obd *filter;
2468         int rc;
2469
2470         filter = &obd->u.filter;
2471
2472         if (group == FID_SEQ_LLOG)
2473                 RETURN(&obd->obd_olg);
2474
2475         OBD_ALLOC_PTR(olg_new);
2476         if (olg_new == NULL)
2477                RETURN(ERR_PTR(-ENOMEM));
2478
2479         cfs_spin_lock(&filter->fo_llog_list_lock);
2480         olg = filter_find_olg_internal(filter, group);
2481         if (olg) {
2482                 if (olg->olg_initializing) {
2483                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2484                 } else {
2485                         GOTO(out_unlock, olg);
2486                 }
2487         } else {
2488                 /* set as the newly allocated one */
2489                 olg = olg_new;
2490                 olg_new = NULL;
2491         }
2492
2493         llog_group_init(olg, group);
2494         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2495         olg->olg_initializing = 1;
2496         cfs_spin_unlock(&filter->fo_llog_list_lock);
2497
2498         rc = obd_llog_init(obd, olg, obd, NULL);
2499         if (rc) {
2500                cfs_spin_lock(&filter->fo_llog_list_lock);
2501                cfs_list_del(&olg->olg_list);
2502                cfs_spin_unlock(&filter->fo_llog_list_lock);
2503                OBD_FREE_PTR(olg);
2504                GOTO(out, olg = ERR_PTR(-ENOMEM));
2505         }
2506         cfs_spin_lock(&filter->fo_llog_list_lock);
2507         olg->olg_initializing = 0;
2508         cfs_spin_unlock(&filter->fo_llog_list_lock);
2509         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2510               obd->obd_name, group, olg);
2511 out:
2512         RETURN(olg);
2513
2514 out_unlock:
2515         cfs_spin_unlock(&filter->fo_llog_list_lock);
2516         if (olg_new)
2517                OBD_FREE_PTR(olg_new);
2518         goto out;
2519 }
2520
2521 static int filter_llog_connect(struct obd_export *exp,
2522                                struct llogd_conn_body *body)
2523 {
2524         struct obd_device *obd = exp->exp_obd;
2525         struct llog_ctxt *ctxt;
2526         struct obd_llog_group *olg;
2527         int rc;
2528         ENTRY;
2529
2530         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2531                obd->obd_name, body->lgdc_logid.lgl_oid,
2532                body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2533
2534         olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2535         if (!olg) {
2536                 CERROR(" %s: can not find olg of group %d\n",
2537                        obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2538                 RETURN(-ENOENT);
2539         }
2540         llog_group_set_export(olg, exp);
2541
2542         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2543         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2544                  body->lgdc_ctxt_idx);
2545
2546         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2547               obd->obd_name, body->lgdc_logid.lgl_oid,
2548               body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2549
2550         cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2551         obd->u.filter.fo_mds_ost_sync = 1;
2552         cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2553         rc = llog_connect(ctxt, &body->lgdc_logid,
2554                           &body->lgdc_gen, NULL);
2555         llog_ctxt_put(ctxt);
2556         if (rc != 0)
2557                 CERROR("failed to connect rc %d idx %d\n", rc,
2558                                 body->lgdc_ctxt_idx);
2559
2560         RETURN(rc);
2561 }
2562
2563 static int filter_llog_preclean(struct obd_device *obd)
2564 {
2565         struct obd_llog_group *olg, *tmp;
2566         struct filter_obd *filter;
2567         cfs_list_t  remove_list;
2568         int rc = 0;
2569         ENTRY;
2570
2571         rc = obd_llog_finish(obd, 0);
2572         if (rc)
2573                 CERROR("failed to cleanup llogging subsystem\n");
2574
2575         filter = &obd->u.filter;
2576         CFS_INIT_LIST_HEAD(&remove_list);
2577
2578         cfs_spin_lock(&filter->fo_llog_list_lock);
2579         while (!cfs_list_empty(&filter->fo_llog_list)) {
2580                 olg = cfs_list_entry(filter->fo_llog_list.next,
2581                                      struct obd_llog_group, olg_list);
2582                 cfs_list_del(&olg->olg_list);
2583                 cfs_list_add(&olg->olg_list, &remove_list);
2584         }
2585         cfs_spin_unlock(&filter->fo_llog_list_lock);
2586
2587         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2588                 cfs_list_del_init(&olg->olg_list);
2589                 rc = filter_olg_fini(olg);
2590                 if (rc)
2591                         CERROR("failed to cleanup llogging subsystem for %u\n",
2592                                olg->olg_seq);
2593                 OBD_FREE_PTR(olg);
2594         }
2595
2596         RETURN(rc);
2597 }
2598
2599 static int filter_precleanup(struct obd_device *obd,
2600                              enum obd_cleanup_stage stage)
2601 {
2602         int rc = 0;
2603         ENTRY;
2604
2605         switch(stage) {
2606         case OBD_CLEANUP_EARLY:
2607                 break;
2608         case OBD_CLEANUP_EXPORTS:
2609                 /* Stop recovery before namespace cleanup. */
2610                 target_recovery_fini(obd);
2611                 rc = filter_llog_preclean(obd);
2612                 break;
2613         }
2614         RETURN(rc);
2615 }
2616
2617 static int filter_cleanup(struct obd_device *obd)
2618 {
2619         struct filter_obd *filter = &obd->u.filter;
2620         ENTRY;
2621
2622         if (obd->obd_fail)
2623                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2624                               "will be preserved.\n", obd->obd_name);
2625
2626         obd_exports_barrier(obd);
2627         obd_zombie_barrier();
2628
2629         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2630         lprocfs_free_per_client_stats(obd);
2631         lprocfs_free_obd_stats(obd);
2632         lprocfs_obd_cleanup(obd);
2633         lquota_cleanup(filter_quota_interface_ref, obd);
2634
2635         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2636         obd->obd_namespace = NULL;
2637
2638         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2639
2640         if (obd->u.obt.obt_sb == NULL)
2641                 RETURN(0);
2642
2643         filter_post(obd);
2644
2645         ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2646         shrink_dcache_sb(obd->u.obt.obt_sb);
2647
2648         server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2649         obd->u.obt.obt_sb = NULL;
2650
2651         fsfilt_put_ops(obd->obd_fsops);
2652
2653         filter_iobuf_pool_done(filter);
2654
2655         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2656
2657         RETURN(0);
2658 }
2659
2660 static int filter_connect_internal(struct obd_export *exp,
2661                                    struct obd_connect_data *data,
2662                                    int reconnect)
2663 {
2664         struct filter_export_data *fed = &exp->exp_filter_data;
2665
2666         if (!data)
2667                 RETURN(0);
2668
2669         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2670                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2671                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2672                data->ocd_connect_flags, data->ocd_version,
2673                data->ocd_grant, data->ocd_index);
2674
2675         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2676                 CWARN("!!! This export (nid %s) used object group %d "
2677                        "earlier; now it's trying to use group %d!  This could "
2678                        "be a bug in the MDS. Please report to "
2679                        "http://bugs.whamcloud.com/\n",
2680                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2681                 RETURN(-EPROTO);
2682         }
2683         fed->fed_group = data->ocd_group;
2684
2685         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2686         exp->exp_connect_flags = data->ocd_connect_flags;
2687         data->ocd_version = LUSTRE_VERSION_CODE;
2688
2689         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2690         if (data->ocd_connect_flags & OBD_CONNECT_MDS)
2691                 CWARN("%s: Received MDS connection for group %u\n",
2692                       exp->exp_obd->obd_name, data->ocd_group);
2693         else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
2694                 RETURN(-EPROTO);
2695
2696         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2697                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2698                 obd_size left, want;
2699
2700                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2701                 left = filter_grant_space_left(exp);
2702                 want = data->ocd_grant;
2703                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2704                 data->ocd_grant = fed->fed_grant;
2705                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2706
2707                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2708                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2709                        exp->exp_client_uuid.uuid, exp,
2710                        data->ocd_grant, want, left);
2711
2712                 filter->fo_tot_granted_clients ++;
2713         }
2714
2715         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2716                 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2717                 int index = le32_to_cpu(lsd->lsd_ost_index);
2718
2719                 if (!(lsd->lsd_feature_compat &
2720                       cpu_to_le32(OBD_COMPAT_OST))) {
2721                         /* this will only happen on the first connect */
2722                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2723                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2724                         /* sync is not needed here as filter_client_add will
2725                          * set exp_need_sync flag */
2726                         filter_update_server_data(exp->exp_obd);
2727                 } else if (index != data->ocd_index) {
2728                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2729                                            " %u doesn't match actual OST index"
2730                                            " %u in last_rcvd file, bad "
2731                                            "configuration?\n",
2732                                            obd_export_nid2str(exp), index,
2733                                            data->ocd_index);
2734                         RETURN(-EBADF);
2735                 }
2736                 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2737                  * FIXME: We don't strictly need the COMPAT flag for that,
2738                  * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2739                  * FIXME: We needed it for the index, as index 0 is valid. */
2740         }
2741
2742         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2743                 data->ocd_brw_size = 65536;
2744         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2745                 data->ocd_brw_size = min(data->ocd_brw_size,
2746                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2747                 if (data->ocd_brw_size == 0) {
2748                         CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2749                                " ocd_version: %x ocd_grant: %d ocd_index: %u "
2750                                "ocd_brw_size is unexpectedly zero, "
2751                                "network data corruption?"
2752                                "Refusing connection of this client\n",
2753                                 exp->exp_obd->obd_name,
2754                                 exp->exp_client_uuid.uuid,
2755                                 exp, data->ocd_connect_flags, data->ocd_version,
2756                                 data->ocd_grant, data->ocd_index);
2757                         RETURN(-EPROTO);
2758                 }
2759         }
2760
2761         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2762                 __u32 cksum_types = data->ocd_cksum_types;
2763
2764                 /* The client set in ocd_cksum_types the checksum types it
2765                  * supports. We have to mask off the algorithms that we don't
2766                  * support */
2767                 data->ocd_cksum_types &= cksum_types_supported();
2768
2769                 /* 1.6.4- only support CRC32 and didn't set ocd_cksum_types */
2770                 if (unlikely(data->ocd_cksum_types == 0))
2771                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2772
2773                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2774                                    "%x\n", exp->exp_obd->obd_name,
2775                                    obd_export_nid2str(exp), cksum_types,
2776                                    data->ocd_cksum_types);
2777         } else {
2778                 /* This client does not support OBD_CONNECT_CKSUM
2779                  * fall back to CRC32 */
2780                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2781                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2782                                    exp->exp_obd->obd_name,
2783                                    obd_export_nid2str(exp));
2784         }
2785
2786         if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2787                 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2788
2789         RETURN(0);
2790 }
2791
2792 static int filter_reconnect(const struct lu_env *env,
2793                             struct obd_export *exp, struct obd_device *obd,
2794                             struct obd_uuid *cluuid,
2795                             struct obd_connect_data *data,
2796                             void *localdata)
2797 {
2798         int rc;
2799         ENTRY;
2800
2801         if (exp == NULL || obd == NULL || cluuid == NULL)
2802                 RETURN(-EINVAL);
2803
2804         rc = filter_connect_internal(exp, data, 1);
2805         if (rc == 0)
2806                 filter_export_stats_init(obd, exp, localdata);
2807
2808         RETURN(rc);
2809 }
2810
2811 static int filter_connect(const struct lu_env *env,
2812                           struct obd_export **exp, struct obd_device *obd,
2813                           struct obd_uuid *cluuid,
2814                           struct obd_connect_data *data, void *localdata)
2815 {
2816         struct lvfs_run_ctxt saved;
2817         struct lustre_handle conn = { 0 };
2818         struct obd_export *lexp;
2819         int rc;
2820         ENTRY;
2821
2822         if (exp == NULL || obd == NULL || cluuid == NULL)
2823                 RETURN(-EINVAL);
2824
2825         rc = class_connect(&conn, obd, cluuid);
2826         if (rc)
2827                 RETURN(rc);
2828         lexp = class_conn2export(&conn);
2829         LASSERT(lexp != NULL);
2830
2831         rc = filter_connect_internal(lexp, data, 0);
2832         if (rc)
2833                 GOTO(cleanup, rc);
2834
2835         filter_export_stats_init(obd, lexp, localdata);
2836         if (obd->obd_replayable) {
2837                 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2838                 LASSERT(lcd);
2839                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2840                 rc = filter_client_add(obd, lexp, -1);
2841                 if (rc)
2842                         GOTO(cleanup, rc);
2843         }
2844
2845         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2846         rc = filter_read_groups(obd, data->ocd_group, 1);
2847         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2848         if (rc != 0) {
2849                 CERROR("can't read group %u\n", data->ocd_group);
2850                 GOTO(cleanup, rc);
2851         }
2852
2853         GOTO(cleanup, rc);
2854
2855 cleanup:
2856         if (rc) {
2857                 class_disconnect(lexp);
2858                 *exp = NULL;
2859         } else {
2860                 *exp = lexp;
2861         }
2862
2863         RETURN(rc);
2864 }
2865
2866 /* Do extra sanity checks for grant accounting.  We do this at connect,
2867  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2868  * always get rid of it or turn it off when we know accounting is good. */
2869 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2870 {
2871         struct filter_export_data *fed;
2872         struct obd_export *exp;
2873         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2874         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2875         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2876
2877         if (cfs_list_empty(&obd->obd_exports))
2878                 return;
2879
2880         /* We don't want to do this for large machines that do lots of
2881            mounts or unmounts.  It burns... */
2882         if (obd->obd_num_exports > 100)
2883                 return;
2884
2885         cfs_spin_lock(&obd->obd_osfs_lock);
2886         cfs_spin_lock(&obd->obd_dev_lock);
2887         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2888                 int error = 0;
2889                 fed = &exp->exp_filter_data;
2890                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2891                     fed->fed_dirty < 0)
2892                         error = 1;
2893                 if (maxsize > 0) { /* we may not have done a statfs yet */
2894                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2895                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2896                                  exp->exp_client_uuid.uuid, exp,
2897                                  fed->fed_grant, fed->fed_pending, maxsize);
2898                         LASSERTF(fed->fed_dirty <= maxsize,
2899                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2900                                  exp->exp_client_uuid.uuid, exp,
2901                                  fed->fed_dirty, maxsize);
2902                 }
2903                 if (error)
2904                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2905                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2906                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2907                 else
2908                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2909                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2910                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2911                 tot_granted += fed->fed_grant + fed->fed_pending;
2912                 tot_pending += fed->fed_pending;
2913                 tot_dirty += fed->fed_dirty;
2914         }
2915         fo_tot_granted = obd->u.filter.fo_tot_granted;
2916         fo_tot_pending = obd->u.filter.fo_tot_pending;
2917         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2918         cfs_spin_unlock(&obd->obd_dev_lock);
2919         cfs_spin_unlock(&obd->obd_osfs_lock);
2920
2921         /* Do these assertions outside the spinlocks so we don't kill system */
2922         if (tot_granted != fo_tot_granted)
2923                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2924                        func, tot_granted, fo_tot_granted);
2925         if (tot_pending != fo_tot_pending)
2926                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2927                        func, tot_pending, fo_tot_pending);
2928         if (tot_dirty != fo_tot_dirty)
2929                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2930                        func, tot_dirty, fo_tot_dirty);
2931         if (tot_pending > tot_granted)
2932                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2933                        func, tot_pending, tot_granted);
2934         if (tot_granted > maxsize)
2935                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2936                        func, tot_granted, maxsize);
2937         if (tot_dirty > maxsize)
2938                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2939                        func, tot_dirty, maxsize);
2940 }
2941
2942 /* Remove this client from the grant accounting totals.  We also remove
2943  * the export from the obd device under the osfs and dev locks to ensure
2944  * that the filter_grant_sanity_check() calculations are always valid.
2945  * The client should do something similar when it invalidates its import. */
2946 static void filter_grant_discard(struct obd_export *exp)
2947 {
2948         struct obd_device *obd = exp->exp_obd;
2949         struct filter_obd *filter = &obd->u.filter;
2950         struct filter_export_data *fed = &exp->exp_filter_data;
2951
2952         cfs_spin_lock(&obd->obd_osfs_lock);
2953         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2954                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2955                  obd->obd_name, filter->fo_tot_granted,
2956                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2957         filter->fo_tot_granted -= fed->fed_grant;
2958         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2959                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2960                  obd->obd_name, filter->fo_tot_pending,
2961                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2962         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2963         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2964                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2965                  obd->obd_name, filter->fo_tot_dirty,
2966                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2967         filter->fo_tot_dirty -= fed->fed_dirty;
2968         fed->fed_dirty = 0;
2969         fed->fed_grant = 0;
2970
2971         cfs_spin_unlock(&obd->obd_osfs_lock);
2972 }
2973
2974 static int filter_destroy_export(struct obd_export *exp)
2975 {
2976         struct filter_export_data *fed = &exp->exp_filter_data;
2977         ENTRY;
2978
2979         if (fed->fed_pending)
2980                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2981                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2982                        exp, fed->fed_pending);
2983
2984         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2985
2986         target_destroy_export(exp);
2987         ldlm_destroy_export(exp);
2988         lut_client_free(exp);
2989
2990         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2991                 RETURN(0);
2992
2993         if (!exp->exp_obd->obd_replayable)
2994                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2995
2996         filter_grant_discard(exp);
2997         filter_fmd_cleanup(exp);
2998
2999         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3000                 struct filter_obd *filter = &exp->exp_obd->u.filter;
3001                 if (filter->fo_tot_granted_clients > 0)
3002                         filter->fo_tot_granted_clients --;
3003         }
3004
3005         if (!(exp->exp_flags & OBD_OPT_FORCE))
3006                 filter_grant_sanity_check(exp->exp_obd, __func__);
3007
3008         RETURN(0);
3009 }
3010
3011 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3012 {
3013         struct obd_llog_group *olg_min, *olg;
3014         struct filter_obd *filter;
3015         int worked = -1, group;
3016         struct llog_ctxt *ctxt;
3017         ENTRY;
3018
3019         filter = &obd->u.filter;
3020
3021         /* we can't sync log holding spinlock. also, we do not want to get
3022          * into livelock. so we do following: loop over MDS's exports in
3023          * group order and skip already synced llogs -bzzz */
3024         do {
3025                 /* look for group with min. number, but > worked */
3026                 olg_min = NULL;
3027                 group = 1 << 30;
3028                 cfs_spin_lock(&filter->fo_llog_list_lock);
3029                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3030                         if (olg->olg_seq <= worked) {
3031                                 /* this group is already synced */
3032                                 continue;
3033                         }
3034                         if (group < olg->olg_seq) {
3035                                 /* we have group with smaller number to sync */
3036                                 continue;
3037                         }
3038                         /* store current minimal group */
3039                         olg_min = olg;
3040                         group = olg->olg_seq;
3041                 }
3042                 cfs_spin_unlock(&filter->fo_llog_list_lock);
3043
3044                 if (olg_min == NULL)
3045                         break;
3046
3047                 worked = olg_min->olg_seq;
3048                 if (olg_min->olg_exp &&
3049                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3050                         int err;
3051                         ctxt = llog_group_get_ctxt(olg_min,
3052                                                    LLOG_MDS_OST_REPL_CTXT);
3053                         if (ctxt) {
3054                                 err = llog_sync(ctxt, olg_min->olg_exp);
3055                                 llog_ctxt_put(ctxt);
3056                                 if (err) {
3057                                         CERROR("error flushing logs to MDS: "
3058                                                "rc %d\n", err);
3059                                 }
3060                         }
3061                 }
3062         } while (olg_min != NULL);
3063 }
3064
3065 /* Also incredibly similar to mds_disconnect */
3066 static int filter_disconnect(struct obd_export *exp)
3067 {
3068         struct obd_device *obd = exp->exp_obd;
3069         int rc;
3070         ENTRY;
3071
3072         LASSERT(exp);
3073         class_export_get(exp);
3074
3075         if (!(exp->exp_flags & OBD_OPT_FORCE))
3076                 filter_grant_sanity_check(obd, __func__);
3077         filter_grant_discard(exp);
3078
3079         /* Flush any remaining cancel messages out to the target */
3080         filter_sync_llogs(obd, exp);
3081
3082         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3083
3084         rc = server_disconnect_export(exp);
3085
3086         /* Do not erase record for recoverable client. */
3087         if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3088                 filter_client_del(exp);
3089         else
3090                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3091
3092         class_export_put(exp);
3093         RETURN(rc);
3094 }
3095
3096 /* reverse import is changed, sync all cancels */
3097 static void filter_revimp_update(struct obd_export *exp)
3098 {
3099         ENTRY;
3100
3101         LASSERT(exp);
3102         class_export_get(exp);
3103
3104         /* flush any remaining cancel messages out to the target */
3105         filter_sync_llogs(exp->exp_obd, exp);
3106         class_export_put(exp);
3107         EXIT;
3108 }
3109
3110 static int filter_ping(struct obd_export *exp)
3111 {
3112         filter_fmd_expire(exp);
3113         return 0;
3114 }
3115
3116 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3117                                   const char *what, int quiet)
3118 {
3119         struct dentry *dchild = NULL;
3120
3121         dchild = filter_fid2dentry(obd, NULL,  ostid->oi_seq, ostid->oi_id);
3122
3123         if (IS_ERR(dchild)) {
3124                 CERROR("%s error looking up object: "POSTID"\n",
3125                        what, ostid->oi_id, ostid->oi_seq);
3126                 RETURN(dchild);
3127         }
3128
3129         if (dchild->d_inode == NULL) {
3130                 if (!quiet)
3131                         CERROR("%s: %s on non-existent object: "POSTID" \n",
3132                                obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3133                 f_dput(dchild);
3134                 RETURN(ERR_PTR(-ENOENT));
3135         }
3136
3137         return dchild;
3138 }
3139
3140 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3141 {
3142         struct dentry *dentry = NULL;
3143         struct obd_device *obd;
3144         int rc = 0;
3145         ENTRY;
3146
3147         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3148                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3149         if (rc)
3150                 RETURN(rc);
3151
3152         obd = class_exp2obd(exp);
3153         if (obd == NULL) {
3154                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3155                 RETURN(-EINVAL);
3156         }
3157
3158         dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3159         if (IS_ERR(dentry))
3160                 RETURN(PTR_ERR(dentry));
3161
3162         /* Limit the valid bits in the return data to what we actually use */
3163         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3164         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3165
3166         f_dput(dentry);
3167         RETURN(rc);
3168 }
3169
3170 /* this should be enabled/disabled in condition to enabled/disabled large
3171  * inodes (fast EAs) in backing store FS. */
3172 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3173                         void *handle, struct obdo *oa)
3174 {
3175         struct obd_device *obd = exp->exp_obd;
3176         int rc = 0;
3177         ENTRY;
3178
3179         if (oa->o_valid & OBD_MD_FLFID) {
3180                 struct filter_fid ff;
3181
3182                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3183                         oa->o_seq = 0;
3184                 /* packing fid and converting it to LE for storing into EA.
3185                  * Here ->o_stripe_idx should be filled by LOV and rest of
3186                  * fields - by client. */
3187                 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3188                 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3189                 /* XXX: we are ignoring o_parent_ver here, since this should
3190                  *      be the same for all objects in this fileset. */
3191                 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3192                 ff.ff_objid = cpu_to_le64(oa->o_id);
3193                 ff.ff_seq = cpu_to_le64(oa->o_seq);
3194
3195                 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3196                        LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3197                        oa->o_seq);
3198
3199                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3200                 if (rc)
3201                         CERROR("store fid in object failed! rc: %d\n", rc);
3202         } else {
3203                 CDEBUG(D_HA, "OSS object without fid info!\n");
3204         }
3205
3206         RETURN(rc);
3207 }
3208
3209 /* this is called from filter_truncate() until we have filter_punch() */
3210 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3211                             struct obdo *oa, struct obd_trans_info *oti)
3212 {
3213         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3214         struct llog_cookie *fcc = NULL;
3215         struct filter_obd *filter;
3216         int rc, err, sync = 0;
3217         loff_t old_size = 0;
3218         unsigned int ia_valid;
3219         struct inode *inode;
3220         struct page *page = NULL;
3221         struct iattr iattr;
3222         void *handle;
3223         ENTRY;
3224
3225         LASSERT(dentry != NULL);
3226         LASSERT(!IS_ERR(dentry));
3227
3228         inode = dentry->d_inode;
3229         LASSERT(inode != NULL);
3230
3231         filter = &exp->exp_obd->u.filter;
3232         iattr_from_obdo(&iattr, oa, oa->o_valid);
3233         ia_valid = iattr.ia_valid;
3234
3235         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3236                 OBD_ALLOC(fcc, sizeof(*fcc));
3237                 if (fcc != NULL)
3238                         *fcc = oa->o_lcookie;
3239         }
3240         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3241                 unsigned long now = jiffies;
3242                 ll_vfs_dq_init(inode);
3243                 /* Filter truncates and writes are serialized by
3244                  * i_alloc_sem, see the comment in
3245                  * filter_preprw_write.*/
3246                 if (ia_valid & ATTR_SIZE)
3247                         down_write(&inode->i_alloc_sem);
3248                 LOCK_INODE_MUTEX(inode);
3249                 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
3250                 old_size = i_size_read(inode);
3251         }
3252
3253         /* VBR: version recovery check */
3254         rc = filter_version_get_check(exp, oti, inode);
3255         if (rc)
3256                 GOTO(out_unlock, rc);
3257
3258         /* Let's pin the last page so that ldiskfs_truncate
3259          * should not start GFP_FS allocation. */
3260         if (ia_valid & ATTR_SIZE) {
3261                 page = grab_cache_page(inode->i_mapping,
3262                                        iattr.ia_size >> PAGE_CACHE_SHIFT);
3263                 if (page == NULL)
3264                         GOTO(out_unlock, rc = -ENOMEM);
3265
3266                 unlock_page(page);
3267         }
3268
3269         /* If the inode still has SUID+SGID bits set (see filter_precreate())
3270          * then we will accept the UID+GID sent by the client during write for
3271          * initializing the ownership of this inode.  We only allow this to
3272          * happen once so clear these bits in setattr. In 2.6 kernels it is
3273          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3274          * the flags that are actually being set. */
3275         if (ia_valid & (ATTR_UID | ATTR_GID)) {
3276                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3277                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3278
3279                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3280                         if (!(ia_valid & ATTR_MODE)) {
3281                                 iattr.ia_mode = inode->i_mode;
3282                                 iattr.ia_valid |= ATTR_MODE;
3283                         }
3284                         iattr.ia_mode &= ~S_ISUID;
3285                 }
3286                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3287                         if (!(iattr.ia_valid & ATTR_MODE)) {
3288                                 iattr.ia_mode = inode->i_mode;
3289                                 iattr.ia_valid |= ATTR_MODE;
3290                         }
3291                         iattr.ia_mode &= ~S_ISGID;
3292                 }
3293
3294                 orig_ids[USRQUOTA] = inode->i_uid;
3295                 orig_ids[GRPQUOTA] = inode->i_gid;
3296                 handle = fsfilt_start_log(exp->exp_obd, inode,
3297                                           FSFILT_OP_SETATTR, oti, 1);
3298                 if (IS_ERR(handle))
3299                         GOTO(out_unlock, rc = PTR_ERR(handle));
3300
3301                 /* update inode EA only once when inode is suid bit marked. As
3302                  * on 2.6.x UID and GID may be set separately, we check here
3303                  * only one of them to avoid double setting. */
3304                 if (inode->i_mode & S_ISUID)
3305                         filter_update_fidea(exp, inode, handle, oa);
3306         } else {
3307                 handle = fsfilt_start(exp->exp_obd, inode,
3308                                       FSFILT_OP_SETATTR, oti);
3309                 if (IS_ERR(handle))
3310                         GOTO(out_unlock, rc = PTR_ERR(handle));
3311         }
3312         if (oa->o_valid & OBD_MD_FLFLAGS) {
3313                 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3314                                       FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3315         } else {
3316                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3317                 if (fcc != NULL)
3318                         /* set cancel cookie callback function */
3319                         sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3320                                                      filter_cancel_cookies_cb,
3321                                                      fcc);
3322         }
3323
3324         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3325                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3326
3327        /* The truncate might have used up our transaction credits.  Make sure
3328         * we have two left for the last_rcvd and VBR inode version updates. */
3329         err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3330
3331         /* Update inode version only if data has changed => size has changed */
3332         rc = filter_finish_transno(exp, ia_valid & ATTR_SIZE ? inode : NULL,
3333                                    oti, rc, sync);
3334
3335         if (sync) {
3336                 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3337                 fcc = NULL;
3338         }
3339
3340         err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3341         if (err) {
3342                 CERROR("error on commit, err = %d\n", err);
3343                 if (!rc)
3344                         rc = err;
3345         } else {
3346                 fcc = NULL;
3347         }
3348
3349         /* For a partial-page truncate flush the page to disk immediately
3350          * to avoid data corruption during direct disk write. b=17397 */
3351         if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3352             old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3353                 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3354                                                iattr.ia_size + 1);
3355                 if (!rc)
3356                         rc = err;
3357         }
3358
3359         EXIT;
3360
3361 out_unlock:
3362         if (page)
3363                 page_cache_release(page);
3364
3365         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3366                 UNLOCK_INODE_MUTEX(inode);
3367         if (ia_valid & ATTR_SIZE)
3368                 up_write(&inode->i_alloc_sem);
3369         if (fcc)
3370                 OBD_FREE(fcc, sizeof(*fcc));
3371
3372         /* trigger quota release */
3373         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3374                 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3375                 int rc2 = lquota_adjust(filter_quota_interface_ref,
3376                                         exp->exp_obd, cur_ids,
3377                                         orig_ids, rc, FSFILT_OP_SETATTR);
3378                 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3379                        "filter adjust qunit. (rc:%d)\n", rc2);
3380         }
3381         return rc;
3382 }
3383
3384 /* this is called from filter_truncate() until we have filter_punch() */
3385 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3386                    struct obd_trans_info *oti)
3387 {
3388         struct obdo *oa = oinfo->oi_oa;
3389         struct lustre_capa *capa = oinfo_capa(oinfo);
3390         struct ldlm_res_id res_id;
3391         struct filter_mod_data *fmd;
3392         struct lvfs_run_ctxt saved;
3393         struct filter_obd *filter;
3394         struct ldlm_resource *res;
3395         struct dentry *dentry;
3396         __u64 opc = CAPA_OPC_META_WRITE;
3397         int rc;
3398         ENTRY;
3399
3400         if (oa->o_valid & OBD_FL_TRUNC)
3401                 opc |= CAPA_OPC_OSS_TRUNC;
3402
3403         rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3404         if (rc)
3405                 RETURN(rc);
3406
3407         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3408                 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3409                 if (rc)
3410                         RETURN(rc);
3411         }
3412
3413         osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3414         /* This would be very bad - accidentally truncating a file when
3415          * changing the time or similar - bug 12203. */
3416         if (oa->o_valid & OBD_MD_FLSIZE &&
3417             oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3418                 static char mdsinum[48];
3419
3420                 if (oa->o_valid & OBD_MD_FLFID)
3421                         snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3422                                  oa->o_parent_seq, oa->o_parent_oid,
3423                                  oa->o_parent_ver);
3424                 else
3425                         mdsinum[0] = '\0';
3426
3427                 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3428                        "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3429                        oa->o_id, oa->o_seq, mdsinum);
3430                 RETURN(-EPERM);
3431         }
3432
3433         dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3434         if (IS_ERR(dentry))
3435                 RETURN(PTR_ERR(dentry));
3436
3437         filter = &exp->exp_obd->u.filter;
3438         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3439
3440         /*
3441          * We need to be atomic against a concurrent write
3442          * (which takes the semaphore for reading). fmd_mactime_xid
3443          * checks will have no effect if a write request with lower
3444          * xid starts just before a setattr and finishes later than
3445          * the setattr (see bug 21489, comment 27).
3446          */
3447         if (oa->o_valid &
3448             (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3449                 unsigned long now = jiffies;
3450                 down_write(&dentry->d_inode->i_alloc_sem);
3451                 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem");
3452                 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3453                 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3454                         fmd->fmd_mactime_xid = oti->oti_xid;
3455                 filter_fmd_put(exp, fmd);
3456                 up_write(&dentry->d_inode->i_alloc_sem);
3457         }
3458
3459         /* setting objects attributes (including owner/group) */
3460         rc = filter_setattr_internal(exp, dentry, oa, oti);
3461         if (rc)
3462                 GOTO(out_unlock, rc);
3463
3464         res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3465                                 &res_id, LDLM_EXTENT, 0);
3466
3467         if (res != NULL) {
3468                 LDLM_RESOURCE_ADDREF(res);
3469                 rc = ldlm_res_lvbo_update(res, NULL, 0);
3470                 LDLM_RESOURCE_DELREF(res);
3471                 ldlm_resource_putref(res);
3472         }
3473
3474         oa->o_valid = OBD_MD_FLID;
3475
3476         /* Quota release need uid/gid info */
3477         obdo_from_inode(oa, dentry->d_inode, NULL,
3478                         FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3479
3480         EXIT;
3481 out_unlock:
3482         f_dput(dentry);
3483         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3484         return rc;
3485 }
3486
3487 /* XXX identical to osc_unpackmd */
3488 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3489                            struct lov_mds_md *lmm, int lmm_bytes)
3490 {
3491         int lsm_size;
3492         ENTRY;
3493
3494         if (lmm != NULL) {
3495                 if (lmm_bytes < sizeof (*lmm)) {
3496                         CERROR("lov_mds_md too small: %d, need %d\n",
3497                                lmm_bytes, (int)sizeof(*lmm));
3498                         RETURN(-EINVAL);
3499                 }
3500                 /* XXX LOV_MAGIC etc check? */
3501
3502                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3503                         CERROR("lov_mds_md: zero lmm_object_id\n");
3504                         RETURN(-EINVAL);
3505                 }
3506         }
3507
3508         lsm_size = lov_stripe_md_size(1);
3509         if (lsmp == NULL)
3510                 RETURN(lsm_size);
3511
3512         if (*lsmp != NULL && lmm == NULL) {
3513                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3514                 OBD_FREE(*lsmp, lsm_size);
3515                 *lsmp = NULL;
3516                 RETURN(0);
3517         }
3518
3519         if (*lsmp == NULL) {
3520                 OBD_ALLOC(*lsmp, lsm_size);
3521                 if (*lsmp == NULL)
3522                         RETURN(-ENOMEM);
3523
3524                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3525                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3526                         OBD_FREE(*lsmp, lsm_size);
3527                         RETURN(-ENOMEM);
3528                 }
3529                 loi_init((*lsmp)->lsm_oinfo[0]);
3530         }
3531
3532         if (lmm != NULL) {
3533                 /* XXX zero *lsmp? */
3534                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3535                 LASSERT((*lsmp)->lsm_object_id);
3536         }
3537
3538         (*lsmp)->lsm_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
3539
3540         RETURN(lsm_size);
3541 }
3542
3543 /* caller must hold fo_create_locks[oa->o_seq] */
3544 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3545                                      struct filter_obd *filter)
3546 {
3547         struct obdo doa = { 0 }; /* XXX obdo on stack */
3548         obd_id last, id;
3549         int rc = 0;
3550         int skip_orphan;
3551         ENTRY;
3552
3553         LASSERT(down_trylock(&filter->fo_create_locks[oa->o_seq]) != 0);
3554
3555         memset(&doa, 0, sizeof(doa));
3556
3557         doa.o_valid |= OBD_MD_FLGROUP;
3558         doa.o_seq = oa->o_seq;
3559         doa.o_mode = S_IFREG;
3560
3561         if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3562                 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3563                        exp->exp_obd->obd_name, doa.o_seq);
3564                 RETURN(0);
3565         }
3566
3567         last = filter_last_id(filter, doa.o_seq);
3568
3569         skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3570
3571         CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3572                exp->exp_obd->obd_name, oa->o_id + 1, last,
3573                skip_orphan ? ", orphan objids won't be reused any more." : ".");
3574
3575         for (id = last; id > oa->o_id; id--) {
3576                 doa.o_id = id;
3577                 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3578                 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3579                         CEMERG("error destroying precreate objid "LPU64": %d\n",
3580                                id, rc);
3581
3582                 /* update last_id on disk periodically so that if we restart
3583                  * we don't need to re-scan all of the just-deleted objects. */
3584                 if ((id & 511) == 0 && !skip_orphan) {
3585                         filter_set_last_id(filter, id - 1, doa.o_seq);
3586                         filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3587                 }
3588         }
3589
3590         CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3591                exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3592
3593         if (!skip_orphan) {
3594                 filter_set_last_id(filter, id, doa.o_seq);
3595                 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3596         } else {
3597                 /*
3598                  * We have destroyed orphan objects, but don't want to reuse
3599                  * them. Therefore we don't reset last_id to the last created
3600                  * objects. Instead, we report back to the MDS the object id
3601                  * of the last orphan, so that the MDS can restart allocating
3602                  * objects from this id + 1 and thus skip the whole orphan
3603                  * object id range
3604                  */
3605                 oa->o_id = last;
3606                 rc = 0;
3607         }
3608         cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3609
3610         RETURN(rc);
3611 }
3612
3613 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3614                             obd_seq group, int *num);
3615 /* returns a negative error or a nonnegative number of files to create */
3616 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3617                                    obd_seq group, struct obd_trans_info *oti)
3618 {
3619         struct obd_device *obd = exp->exp_obd;
3620         struct filter_obd *filter = &obd->u.filter;
3621         int diff, rc;
3622         ENTRY;
3623
3624         /* delete orphans request */
3625         if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3626                 obd_id last = filter_last_id(filter, group);
3627
3628                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3629                         CERROR("%s: dropping old orphan cleanup request\n",
3630                                obd->obd_name);
3631                         RETURN(0);
3632                 }
3633                 /* This causes inflight precreates to abort and drop lock */
3634                 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3635                 cfs_down(&filter->fo_create_locks[group]);
3636                 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3637                         CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3638                                exp->exp_obd->obd_name, group);
3639                         cfs_up(&filter->fo_create_locks[group]);
3640                         RETURN(0);
3641                 }
3642                 diff = oa->o_id - last;
3643                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3644                        last, diff);
3645
3646                 if (-diff > OST_MAX_PRECREATE) {
3647                         CERROR("%s: ignoring bogus orphan destroy request: "
3648                                "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3649                                oa->o_id, last);
3650                         /* FIXME: should reset precreate_next_id on MDS */
3651                         GOTO(out, rc = -EINVAL);
3652                 }
3653                 if (diff < 0) {
3654                         rc = filter_destroy_precreated(exp, oa, filter);
3655                         if (rc)
3656                                 CERROR("%s: unable to write lastobjid, but "
3657                                        "orphans were deleted\n", obd->obd_name);
3658                         GOTO(out, rc);
3659                 } else {
3660                         /* XXX: Used by MDS for the first time! */
3661                         cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3662                 }
3663         } else {
3664                 cfs_down(&filter->fo_create_locks[group]);
3665                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3666                         CERROR("%s: dropping old precreate request\n",
3667                                obd->obd_name);
3668                         GOTO(out, rc = 0);
3669                 }
3670                 /* only precreate if group == 0 and o_id is specfied */
3671                 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3672                         diff = 1;
3673                 else
3674                         diff = oa->o_id - filter_last_id(filter, group);
3675                 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3676                        filter_last_id(filter, group), diff);
3677
3678                 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3679                          oa->o_id, filter_last_id(filter, group), diff);
3680         }
3681
3682