Whamcloud - gitweb
LU-243 async lov_sync() operation
[fs/lustre-release.git] / lustre / obdfilter / filter.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/obdfilter/filter.c
37  *
38  * Author: Peter Braam <braam@clusterfs.com>
39  * Author: Andreas Dilger <adilger@clusterfs.com>
40  */
41
42 /*
43  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44  *            (which need to get journal_lock, may block if journal full).
45  *
46  * Invariant: Call filter_start_transno() before any journal ops to avoid the
47  *            same deadlock problem.  We can (and want) to get rid of the
48  *            transno sem in favour of the dir/inode i_mutex to avoid single
49  *            threaded operation on the OST.
50  */
51
52 #define DEBUG_SUBSYSTEM S_FILTER
53
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
56 #endif
57 #include <linux/module.h>
58 #include <linux/fs.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
65
66 #include <obd_cksum.h>
67 #include <obd_class.h>
68 #include <obd_lov.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
79
80 #include "filter_internal.h"
81
82 static struct lvfs_callback_ops filter_lvfs_ops;
83 cfs_mem_cache_t *ll_fmd_cachep;
84
85 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
86                              void *cb_data, int error)
87 {
88         struct obd_export *exp = cb_data;
89         LASSERT(exp->exp_obd == obd);
90         obd_transno_commit_cb(obd, transno, exp, error);
91         class_export_cb_put(exp);
92 }
93
94 int filter_version_get_check(struct obd_export *exp,
95                              struct obd_trans_info *oti, struct inode *inode)
96 {
97         __u64 curr_version;
98
99         if (inode == NULL || oti == NULL)
100                 RETURN(0);
101
102         curr_version = fsfilt_get_version(exp->exp_obd, inode);
103         if ((__s64)curr_version == -EOPNOTSUPP)
104                 RETURN(0);
105         /* VBR: version is checked always because costs nothing */
106         if (oti->oti_pre_version != 0 &&
107             oti->oti_pre_version != curr_version) {
108                 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
109                        oti->oti_pre_version, curr_version);
110                 cfs_spin_lock(&exp->exp_lock);
111                 exp->exp_vbr_failed = 1;
112                 cfs_spin_unlock(&exp->exp_lock);
113                 RETURN (-EOVERFLOW);
114         }
115         oti->oti_pre_version = curr_version;
116         RETURN(0);
117 }
118
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121                           struct obd_trans_info *oti, int rc, int force_sync)
122 {
123         struct obd_device_target *obt = &exp->exp_obd->u.obt;
124         struct tg_export_data *ted = &exp->exp_target_data;
125         struct lr_server_data *lsd = class_server_data(exp->exp_obd);
126         struct lsd_client_data *lcd;
127         __u64 last_rcvd;
128         loff_t off;
129         int err, log_pri = D_RPCTRACE;
130
131         /* Propagate error code. */
132         if (rc)
133                 RETURN(rc);
134
135         if (!exp->exp_obd->obd_replayable || oti == NULL)
136                 RETURN(rc);
137
138         cfs_mutex_down(&ted->ted_lcd_lock);
139         lcd = ted->ted_lcd;
140         /* if the export has already been disconnected, we have no last_rcvd slot,
141          * update server data with latest transno then */
142         if (lcd == NULL) {
143                 cfs_mutex_up(&ted->ted_lcd_lock);
144                 CWARN("commit transaction for disconnected client %s: rc %d\n",
145                       exp->exp_client_uuid.uuid, rc);
146                 err = filter_update_server_data(exp->exp_obd);
147                 RETURN(err);
148         }
149
150         /* we don't allocate new transnos for replayed requests */
151         cfs_spin_lock(&obt->obt_lut->lut_translock);
152         if (oti->oti_transno == 0) {
153                 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
154                 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
155         } else {
156                 last_rcvd = oti->oti_transno;
157                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
158                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
159         }
160         oti->oti_transno = last_rcvd;
161
162         LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
163         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
164         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
165         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
166         cfs_spin_unlock(&obt->obt_lut->lut_translock);
167
168         if (inode)
169                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
170
171         off = ted->ted_lr_off;
172         if (off <= 0) {
173                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
174                        ted->ted_lr_idx, ted->ted_lr_off);
175                 err = -EINVAL;
176         } else {
177                 class_export_cb_get(exp); /* released when the cb is called */
178                 if (!force_sync)
179                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
180                                                            last_rcvd,
181                                                            oti->oti_handle,
182                                                            filter_commit_cb,
183                                                            exp);
184
185                 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
186                                           lcd, sizeof(*lcd), &off,
187                                           force_sync | exp->exp_need_sync);
188                 if (force_sync)
189                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
190         }
191         if (err) {
192                 log_pri = D_ERROR;
193                 if (rc == 0)
194                         rc = err;
195         }
196
197         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
198                last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
199         cfs_mutex_up(&ted->ted_lcd_lock);
200         RETURN(rc);
201 }
202
203 void f_dput(struct dentry *dentry)
204 {
205         /* Can't go inside filter_ddelete because it can block */
206         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
207                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
208         LASSERT(atomic_read(&dentry->d_count) > 0);
209
210         dput(dentry);
211 }
212
213 static void init_brw_stats(struct brw_stats *brw_stats)
214 {
215         int i;
216         for (i = 0; i < BRW_LAST; i++)
217                 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
218 }
219
220 static int lprocfs_init_rw_stats(struct obd_device *obd,
221                                  struct lprocfs_stats **stats)
222 {
223         int num_stats;
224
225         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
226                                                         LPROC_FILTER_LAST - 1;
227         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
228         if (*stats == NULL)
229                 return -ENOMEM;
230
231         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
232         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
233                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
234         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
235                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
236
237         return(0);
238 }
239
240 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
241    plus the procfs overhead :( */
242 static int filter_export_stats_init(struct obd_device *obd,
243                                     struct obd_export *exp,
244                                     void *client_nid)
245 {
246         int rc, newnid = 0;
247         ENTRY;
248
249         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
250                 /* Self-export gets no proc entry */
251                 RETURN(0);
252
253         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
254         if (rc) {
255                 /* Mask error for already created
256                  * /proc entries */
257                 if (rc == -EALREADY)
258                         rc = 0;
259                 RETURN(rc);
260         }
261
262         if (newnid) {
263                 struct nid_stat *tmp = exp->exp_nid_stats;
264                 LASSERT(tmp != NULL);
265
266                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
267                 if (tmp->nid_brw_stats == NULL)
268                         GOTO(clean, rc = -ENOMEM);
269
270                 init_brw_stats(tmp->nid_brw_stats);
271                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
272                                         0644, &filter_per_nid_stats_fops,
273                                         exp->exp_nid_stats);
274                 if (rc)
275                         CWARN("Error adding the brw_stats file\n");
276
277                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
278                 if (rc)
279                         GOTO(clean, rc);
280
281                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
282                                             tmp->nid_stats);
283                 if (rc)
284                         GOTO(clean, rc);
285                 rc = lprocfs_nid_ldlm_stats_init(tmp);
286                 if (rc)
287                         GOTO(clean, rc);
288         }
289
290         RETURN(0);
291  clean:
292         return rc;
293 }
294
295 /* Add client data to the FILTER.  We use a bitmap to locate a free space
296  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
297  * Otherwise, we have just read the data from the last_rcvd file and
298  * we know its offset. */
299 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
300                              int cl_idx)
301 {
302         struct obd_device_target *obt = &obd->u.obt;
303         struct tg_export_data *ted = &exp->exp_target_data;
304         struct lr_server_data *lsd = class_server_data(obd);
305         unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
306         int new_client = (cl_idx == -1);
307
308         ENTRY;
309
310         LASSERT(bitmap != NULL);
311         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
312
313         /* Self-export */
314         if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
315                 RETURN(0);
316
317         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
318          * there's no need for extra complication here
319          */
320         if (new_client) {
321                 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
322         repeat:
323                 if (cl_idx >= LR_MAX_CLIENTS) {
324                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
325                                cl_idx);
326                         RETURN(-EOVERFLOW);
327                 }
328                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
329                         cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
330                                                         cl_idx);
331                         goto repeat;
332                 }
333         } else {
334                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
335                         CERROR("FILTER client %d: bit already set in bitmap!\n",
336                                cl_idx);
337                         LBUG();
338                 }
339         }
340
341         ted->ted_lr_idx = cl_idx;
342         ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
343                           cl_idx * le16_to_cpu(lsd->lsd_client_size);
344         cfs_init_mutex(&ted->ted_lcd_lock);
345         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
346
347         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
348                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
349
350         if (new_client) {
351                 struct lvfs_run_ctxt saved;
352                 loff_t off = ted->ted_lr_off;
353                 int rc;
354                 void *handle;
355
356                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
357                        ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
358
359                 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
360                         RETURN(-ENOSPC);
361
362                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
363                 /* Transaction needed to fix bug 1403 */
364                 handle = fsfilt_start(obd,
365                                       obt->obt_rcvd_filp->f_dentry->d_inode,
366                                       FSFILT_OP_SETATTR, NULL);
367                 if (IS_ERR(handle)) {
368                         rc = PTR_ERR(handle);
369                         CERROR("unable to start transaction: rc %d\n", rc);
370                 } else {
371                         ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
372                         exp->exp_last_request_time = cfs_time_current_sec();
373                         rc = fsfilt_add_journal_cb(obd, 0, handle,
374                                                    target_client_add_cb,
375                                                    class_export_cb_get(exp));
376                         if (rc == 0) {
377                                 cfs_spin_lock(&exp->exp_lock);
378                                 exp->exp_need_sync = 1;
379                                 cfs_spin_unlock(&exp->exp_lock);
380                         }
381                         rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
382                                                  ted->ted_lcd,
383                                                  sizeof(*ted->ted_lcd),
384                                                  &off, rc /* sync if no cb */);
385                         fsfilt_commit(obd,
386                                       obt->obt_rcvd_filp->f_dentry->d_inode,
387                                       handle, 0);
388                 }
389                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
390
391                 if (rc) {
392                         CERROR("error writing %s client idx %u: rc %d\n",
393                                LAST_RCVD, ted->ted_lr_idx, rc);
394                         RETURN(rc);
395                 }
396         }
397         RETURN(0);
398 }
399
400 static int filter_client_del(struct obd_export *exp)
401 {
402         struct tg_export_data *ted = &exp->exp_target_data;
403         struct obd_device_target *obt = &exp->exp_obd->u.obt;
404         struct lvfs_run_ctxt saved;
405         int rc;
406         loff_t off;
407         ENTRY;
408
409         if (ted->ted_lcd == NULL)
410                 RETURN(0);
411
412         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
413         if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
414                 GOTO(free, 0);
415
416         LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
417
418         off = ted->ted_lr_off;
419
420         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
421                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
422
423         /* Don't clear ted_lr_idx here as it is likely also unset.  At worst
424          * we leak a client slot that will be cleaned on the next recovery. */
425         if (off <= 0) {
426                 CERROR("%s: client idx %d has med_off %lld\n",
427                        exp->exp_obd->obd_name, ted->ted_lr_idx, off);
428                 GOTO(free, rc = -EINVAL);
429         }
430
431         /* Clear the bit _after_ zeroing out the client so we don't
432            race with filter_client_add and zero out new clients.*/
433         if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
434                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
435                        ted->ted_lr_idx);
436                 LBUG();
437         }
438
439         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
440         /* Make sure the server's last_transno is up to date.
441          * This should be done before zeroing client slot so last_transno will
442          * be in server data or in client data in case of failure */
443         filter_update_server_data(exp->exp_obd);
444
445         cfs_mutex_down(&ted->ted_lcd_lock);
446         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
447         rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
448                                  ted->ted_lcd,
449                                  sizeof(*ted->ted_lcd), &off, 0);
450         cfs_mutex_up(&ted->ted_lcd_lock);
451         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
452
453         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
454                "zero out client %s at idx %u/%llu in %s, rc %d\n",
455                ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
456                LAST_RCVD, rc);
457         EXIT;
458 free:
459         return 0;
460 }
461
462 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
463 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
464                                          struct filter_mod_data *fmd)
465 {
466         LASSERT_SPIN_LOCKED(&fed->fed_lock);
467         if (--fmd->fmd_refcount == 0) {
468                 /* XXX when we have persistent reservations and the handle
469                  * is stored herein we need to drop it here. */
470                 fed->fed_mod_count--;
471                 cfs_list_del(&fmd->fmd_list);
472                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
473         }
474 }
475
476 /* drop fmd reference, free it if last ref */
477 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
478 {
479         struct filter_export_data *fed;
480
481         if (fmd == NULL)
482                 return;
483
484         fed = &exp->exp_filter_data;
485         cfs_spin_lock(&fed->fed_lock);
486         filter_fmd_put_nolock(fed, fmd); /* caller reference */
487         cfs_spin_unlock(&fed->fed_lock);
488 }
489
490 /* expire entries from the end of the list if there are too many
491  * or they are too old */
492 static void filter_fmd_expire_nolock(struct filter_obd *filter,
493                                      struct filter_export_data *fed,
494                                      struct filter_mod_data *keep)
495 {
496         struct filter_mod_data *fmd, *tmp;
497
498         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
499                 if (fmd == keep)
500                         break;
501
502                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
503                     fed->fed_mod_count < filter->fo_fmd_max_num)
504                         break;
505
506                 cfs_list_del_init(&fmd->fmd_list);
507                 filter_fmd_put_nolock(fed, fmd); /* list reference */
508         }
509 }
510
511 void filter_fmd_expire(struct obd_export *exp)
512 {
513         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
514         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
515                                  &exp->exp_filter_data, NULL);
516         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
517 }
518
519 /* find specified objid, group in export fmd list.
520  * caller must hold fed_lock and take fmd reference itself */
521 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
522                                                 struct filter_export_data *fed,
523                                                 obd_id objid, obd_seq group)
524 {
525         struct filter_mod_data *found = NULL, *fmd;
526
527         LASSERT_SPIN_LOCKED(&fed->fed_lock);
528
529         cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
530                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
531                         found = fmd;
532                         cfs_list_del(&fmd->fmd_list);
533                         cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
534                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
535                         break;
536                 }
537         }
538
539         filter_fmd_expire_nolock(filter, fed, found);
540
541         return found;
542 }
543
544 /* Find fmd based on objid and group, or return NULL if not found. */
545 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
546                                         obd_id objid, obd_seq group)
547 {
548         struct filter_mod_data *fmd;
549
550         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
551         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
552                                      &exp->exp_filter_data, objid, group);
553         if (fmd)
554                 fmd->fmd_refcount++;    /* caller reference */
555         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
556
557         return fmd;
558 }
559
560 /* Find fmd based on objid and group, or create a new one if none is found.
561  * It is possible for this function to return NULL under memory pressure,
562  * or if objid = 0 is passed (which will only cause old entries to expire).
563  * Currently this is not fatal because any fmd state is transient and
564  * may also be freed when it gets sufficiently old. */
565 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
566                                        obd_id objid, obd_seq group)
567 {
568         struct filter_export_data *fed = &exp->exp_filter_data;
569         struct filter_mod_data *found = NULL, *fmd_new = NULL;
570
571         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
572
573         cfs_spin_lock(&fed->fed_lock);
574         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
575         if (fmd_new) {
576                 if (found == NULL) {
577                         cfs_list_add_tail(&fmd_new->fmd_list,
578                                           &fed->fed_mod_list);
579                         fmd_new->fmd_id = objid;
580                         fmd_new->fmd_gr = group;
581                         fmd_new->fmd_refcount++;   /* list reference */
582                         found = fmd_new;
583                         fed->fed_mod_count++;
584                 } else {
585                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
586                 }
587         }
588         if (found) {
589                 found->fmd_refcount++;          /* caller reference */
590                 found->fmd_expire = jiffies +
591                         exp->exp_obd->u.filter.fo_fmd_max_age;
592         }
593
594         cfs_spin_unlock(&fed->fed_lock);
595
596         return found;
597 }
598
599 #ifdef DO_FMD_DROP
600 /* drop fmd list reference so it will disappear when last reference is put.
601  * This isn't so critical because it would in fact only affect the one client
602  * that is doing the unlink and at worst we have an stale entry referencing
603  * an object that should never be used again. */
604 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
605 {
606         struct filter_mod_data *found = NULL;
607
608         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
609         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
610         if (found) {
611                 cfs_list_del_init(&found->fmd_list);
612                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
613         }
614         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
615 }
616 #else
617 #define filter_fmd_drop(exp, objid, group)
618 #endif
619
620 /* remove all entries from fmd list */
621 static void filter_fmd_cleanup(struct obd_export *exp)
622 {
623         struct filter_export_data *fed = &exp->exp_filter_data;
624         struct filter_mod_data *fmd = NULL, *tmp;
625
626         cfs_spin_lock(&fed->fed_lock);
627         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
628                 cfs_list_del_init(&fmd->fmd_list);
629                 filter_fmd_put_nolock(fed, fmd);
630         }
631         cfs_spin_unlock(&fed->fed_lock);
632 }
633
634 static int filter_init_export(struct obd_export *exp)
635 {
636         int rc;
637         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
638         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
639
640         cfs_spin_lock(&exp->exp_lock);
641         exp->exp_connecting = 1;
642         cfs_spin_unlock(&exp->exp_lock);
643         rc = lut_client_alloc(exp);
644         if (rc == 0)
645                 rc = ldlm_init_export(exp);
646
647         return rc;
648 }
649
650 static int filter_free_server_data(struct obd_device_target *obt)
651 {
652         lut_fini(NULL, obt->obt_lut);
653         OBD_FREE_PTR(obt->obt_lut);
654         return 0;
655 }
656
657 /* assumes caller is already in kernel ctxt */
658 int filter_update_server_data(struct obd_device *obd)
659 {
660         struct file *filp = obd->u.obt.obt_rcvd_filp;
661         struct lr_server_data *lsd = class_server_data(obd);
662         loff_t off = 0;
663         int rc;
664         ENTRY;
665
666         CDEBUG(D_INODE, "server uuid      : %s\n", lsd->lsd_uuid);
667         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
668                le64_to_cpu(lsd->lsd_last_transno));
669         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
670                le64_to_cpu(lsd->lsd_mount_count));
671
672         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
673         if (rc)
674                 CERROR("error writing lr_server_data: rc = %d\n", rc);
675
676         RETURN(rc);
677 }
678
679 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
680                              int force_sync)
681 {
682         struct filter_obd *filter = &obd->u.filter;
683         __u64 tmp;
684         loff_t off = 0;
685         int rc;
686         ENTRY;
687
688         if (filter->fo_last_objid_files[group] == NULL) {
689                 CERROR("Object seq "LPU64" not fully setup; not updating "
690                        "last_objid\n", group);
691                 RETURN(-EINVAL);
692         }
693
694         CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
695                obd->obd_name, filter->fo_last_objids[group], group);
696
697         tmp = cpu_to_le64(filter->fo_last_objids[group]);
698         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
699                                  &tmp, sizeof(tmp), &off, force_sync);
700         if (rc)
701                 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
702                        group, rc);
703         RETURN(rc);
704 }
705 extern int ost_handle(struct ptlrpc_request *req);
706 /* assumes caller has already in kernel ctxt */
707 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
708 {
709         struct filter_obd *filter = &obd->u.filter;
710         struct lr_server_data *lsd;
711         struct lsd_client_data *lcd = NULL;
712         struct inode *inode = filp->f_dentry->d_inode;
713         unsigned long last_rcvd_size = i_size_read(inode);
714         struct lu_target *lut;
715         __u64 mount_count;
716         __u32 start_epoch;
717         int cl_idx;
718         loff_t off = 0;
719         int rc;
720
721         /* ensure padding in the struct is the correct size */
722         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
723                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
724         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
725                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
726
727         /* allocate and initialize lu_target */
728         OBD_ALLOC_PTR(lut);
729         if (lut == NULL)
730                 RETURN(-ENOMEM);
731         rc = lut_init(NULL, lut, obd, NULL);
732         if (rc)
733                 GOTO(err_lut, rc);
734         lsd = class_server_data(obd);
735         if (last_rcvd_size == 0) {
736                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
737
738                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
739                 lsd->lsd_last_transno = 0;
740                 mount_count = lsd->lsd_mount_count = 0;
741                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
742                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
743                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
744                 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
745                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
746                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
747                  * MDS first connects and assigns the OST index number. */
748                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
749                                                         OBD_INCOMPAT_OST);
750         } else {
751                 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
752                 if (rc) {
753                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
754                                LAST_RCVD, rc);
755                         GOTO(err_lut, rc);
756                 }
757                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
758                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
759                                            "using the wrong disk %s. Were the "
760                                            "/dev/ assignments rearranged?\n",
761                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
762                         GOTO(err_lut, rc = -EINVAL);
763                 }
764                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
765                 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
766                 /* COMPAT_146 */
767                 /* Assume old last_rcvd format unless I_C_LR is set */
768                 if (!(lsd->lsd_feature_incompat &
769                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
770                         lsd->lsd_last_transno = lsd->lsd_compat14;
771                 /* end COMPAT_146 */
772                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
773                  * MDS first connects and assigns the OST index number. */
774                 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
775                                                          OBD_INCOMPAT_OST);
776         }
777
778         if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
779                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
780                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
781                        ~FILTER_INCOMPAT_SUPP);
782                 GOTO(err_lut, rc = -EINVAL);
783         }
784         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
785                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
786                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
787                        ~FILTER_ROCOMPAT_SUPP);
788                 /* Do something like remount filesystem read-only */
789                 GOTO(err_lut, rc = -EINVAL);
790         }
791
792         start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
793
794         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
795                obd->obd_name, start_epoch);
796         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
797                obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
798         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
799                obd->obd_name, mount_count + 1);
800         CDEBUG(D_INODE, "%s: server data size: %u\n",
801                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
802         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
803                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
804         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
805                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
806         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
807                obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
808         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
809                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
810                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
811                 le16_to_cpu(lsd->lsd_client_size));
812
813         if (!obd->obd_replayable) {
814                 CWARN("%s: recovery support OFF\n", obd->obd_name);
815                 GOTO(out, rc = 0);
816         }
817
818         OBD_ALLOC_PTR(lcd);
819         if (!lcd)
820                 GOTO(err_client, rc = -ENOMEM);
821
822         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
823              off < last_rcvd_size; cl_idx++) {
824                 __u64 last_rcvd;
825                 struct obd_export *exp;
826                 struct filter_export_data *fed;
827
828                 /* Don't assume off is incremented properly by
829                  * fsfilt_read_record(), in case sizeof(*lcd)
830                  * isn't the same as lsd->lsd_client_size.  */
831                 off = le32_to_cpu(lsd->lsd_client_start) +
832                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
833                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
834                 if (rc) {
835                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
836                                LAST_RCVD, cl_idx, off, rc);
837                         break; /* read error shouldn't cause startup to fail */
838                 }
839
840                 if (lcd->lcd_uuid[0] == '\0') {
841                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
842                                cl_idx);
843                         continue;
844                 }
845
846                 check_lcd(obd->obd_name, cl_idx, lcd);
847
848                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
849
850                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
851                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
852                        last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
853
854                 /* These exports are cleaned up by filter_disconnect(), so they
855                  * need to be set up like real exports as filter_connect() does.
856                  */
857                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
858                 if (IS_ERR(exp)) {
859                         if (PTR_ERR(exp) == -EALREADY) {
860                                 /* export already exists, zero out this one */
861                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
862                                 continue;
863                         }
864                         OBD_FREE_PTR(lcd);
865                         GOTO(err_client, rc = PTR_ERR(exp));
866                 }
867
868                 fed = &exp->exp_filter_data;
869                 *fed->fed_ted.ted_lcd = *lcd;
870                 fed->fed_group = 0; /* will be assigned at connect */
871                 filter_export_stats_init(obd, exp, NULL);
872                 rc = filter_client_add(obd, exp, cl_idx);
873                 /* can't fail for existing client */
874                 LASSERTF(rc == 0, "rc = %d\n", rc);
875
876                 /* VBR: set export last committed */
877                 exp->exp_last_committed = last_rcvd;
878                 cfs_spin_lock(&exp->exp_lock);
879                 exp->exp_connecting = 0;
880                 exp->exp_in_recovery = 0;
881                 cfs_spin_unlock(&exp->exp_lock);
882                 obd->obd_max_recoverable_clients++;
883                 class_export_put(exp);
884
885                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
886                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
887         }
888         OBD_FREE_PTR(lcd);
889
890         obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
891 out:
892         obd->u.obt.obt_mount_count = mount_count + 1;
893         lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
894
895         /* save it, so mount count and last_transno is current */
896         rc = filter_update_server_data(obd);
897         if (rc)
898                 GOTO(err_client, rc);
899
900         RETURN(0);
901
902 err_client:
903         class_disconnect_exports(obd);
904 err_lut:
905         filter_free_server_data(&obd->u.obt);
906         RETURN(rc);
907 }
908
909 static int filter_cleanup_groups(struct obd_device *obd)
910 {
911         struct filter_obd *filter = &obd->u.filter;
912         struct file *filp;
913         struct dentry *dentry;
914         int i, j;
915         ENTRY;
916
917         if (filter->fo_dentry_O_groups != NULL) {
918                 for (i = 0; i < filter->fo_group_count; i++) {
919                         dentry = filter->fo_dentry_O_groups[i];
920                         if (dentry != NULL)
921                                 f_dput(dentry);
922                 }
923                 OBD_FREE(filter->fo_dentry_O_groups,
924                          filter->fo_group_count *
925                          sizeof(*filter->fo_dentry_O_groups));
926                 filter->fo_dentry_O_groups = NULL;
927         }
928         if (filter->fo_last_objid_files != NULL) {
929                 for (i = 0; i < filter->fo_group_count; i++) {
930                         filp = filter->fo_last_objid_files[i];
931                         if (filp != NULL)
932                                 filp_close(filp, 0);
933                 }
934                 OBD_FREE(filter->fo_last_objid_files,
935                          filter->fo_group_count *
936                          sizeof(*filter->fo_last_objid_files));
937                 filter->fo_last_objid_files = NULL;
938         }
939         if (filter->fo_dentry_O_sub != NULL) {
940                 for (i = 0; i < filter->fo_group_count; i++) {
941                         for (j = 0; j < filter->fo_subdir_count; j++) {
942                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
943                                 if (dentry != NULL)
944                                         f_dput(dentry);
945                         }
946                 }
947                 OBD_FREE(filter->fo_dentry_O_sub,
948                          filter->fo_group_count *
949                          sizeof(*filter->fo_dentry_O_sub));
950                 filter->fo_dentry_O_sub = NULL;
951         }
952         if (filter->fo_last_objids != NULL) {
953                 OBD_FREE(filter->fo_last_objids,
954                          filter->fo_group_count *
955                          sizeof(*filter->fo_last_objids));
956                 filter->fo_last_objids = NULL;
957         }
958         if (filter->fo_dentry_O != NULL) {
959                 f_dput(filter->fo_dentry_O);
960                 filter->fo_dentry_O = NULL;
961         }
962         RETURN(0);
963 }
964
965 static int filter_update_last_group(struct obd_device *obd, int group)
966 {
967         struct filter_obd *filter = &obd->u.filter;
968         struct file *filp = NULL;
969         int last_group = 0, rc;
970         loff_t off = 0;
971         ENTRY;
972
973         if (group <= filter->fo_committed_group)
974                 RETURN(0);
975
976         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
977         if (IS_ERR(filp)) {
978                 rc = PTR_ERR(filp);
979                 filp = NULL;
980                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
981                 GOTO(cleanup, rc);
982         }
983
984         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
985         if (rc) {
986                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
987                 GOTO(cleanup, rc);
988         }
989
990         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
991                obd->obd_name, last_group, group);
992
993         off = 0;
994         last_group = group;
995         /* must be sync: bXXXX */
996         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
997         if (rc) {
998                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
999                 GOTO(cleanup, rc);
1000         }
1001
1002         filter->fo_committed_group = group;
1003 cleanup:
1004         if (filp)
1005                 filp_close(filp, 0);
1006         RETURN(rc);
1007 }
1008
1009 static int filter_read_group_internal(struct obd_device *obd, int group,
1010                                       int create)
1011 {
1012         struct filter_obd *filter = &obd->u.filter;
1013         __u64 *new_objids = NULL;
1014         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1015         struct dentry **new_groups = NULL;
1016         struct file **new_files = NULL;
1017         struct dentry *dentry;
1018         struct file *filp;
1019         int old_count = filter->fo_group_count, rc, stage = 0, i;
1020         char name[25];
1021         __u64 last_objid;
1022         loff_t off = 0;
1023         int len = group + 1;
1024
1025         snprintf(name, 24, "%d", group);
1026         name[24] = '\0';
1027
1028         if (!create) {
1029                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1030                                            strlen(name));
1031                 if (IS_ERR(dentry)) {
1032                         CERROR("Cannot lookup expected object group %d: %ld\n",
1033                                group, PTR_ERR(dentry));
1034                         RETURN(PTR_ERR(dentry));
1035                 }
1036         } else {
1037                 dentry = simple_mkdir(filter->fo_dentry_O,
1038                                       obd->u.obt.obt_vfsmnt, name, 0700, 1);
1039                 if (IS_ERR(dentry)) {
1040                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1041                                PTR_ERR(dentry));
1042                         RETURN(PTR_ERR(dentry));
1043                 }
1044         }
1045         stage = 1;
1046
1047         snprintf(name, 24, "O/%d/LAST_ID", group);
1048         name[24] = '\0';
1049         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1050         if (IS_ERR(filp)) {
1051                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1052                 GOTO(cleanup, rc = PTR_ERR(filp));
1053         }
1054         stage = 2;
1055
1056         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1057         if (rc) {
1058                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1059                 GOTO(cleanup, rc);
1060         }
1061
1062         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1063                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1064                 if (tmp_subdirs == NULL)
1065                         GOTO(cleanup, rc = -ENOMEM);
1066                 stage = 3;
1067
1068                 for (i = 0; i < filter->fo_subdir_count; i++) {
1069                         char dir[20];
1070                         snprintf(dir, sizeof(dir), "d%u", i);
1071
1072                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1073                                                               obd->u.obt.obt_vfsmnt,
1074                                                               dir, 0700, 1);
1075                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1076                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1077                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1078                                        group, dir, rc);
1079                                 GOTO(cleanup, rc);
1080                         }
1081
1082                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1083                                tmp_subdirs->dentry[i]);
1084                 }
1085         }
1086
1087         /* 'group' is an index; we need an array of length 'group + 1' */
1088         if (group + 1 > old_count) {
1089                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1090                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1091                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1092                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1093                 stage = 4;
1094                 if (new_objids == NULL || new_subdirs == NULL ||
1095                     new_groups == NULL || new_files == NULL)
1096                         GOTO(cleanup, rc = -ENOMEM);
1097
1098                 if (old_count) {
1099                         memcpy(new_objids, filter->fo_last_objids,
1100                                old_count * sizeof(*new_objids));
1101                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1102                                old_count * sizeof(*new_subdirs));
1103                         memcpy(new_groups, filter->fo_dentry_O_groups,
1104                                old_count * sizeof(*new_groups));
1105                         memcpy(new_files, filter->fo_last_objid_files,
1106                                old_count * sizeof(*new_files));
1107
1108                         OBD_FREE(filter->fo_last_objids,
1109                                  old_count * sizeof(*new_objids));
1110                         OBD_FREE(filter->fo_dentry_O_sub,
1111                                  old_count * sizeof(*new_subdirs));
1112                         OBD_FREE(filter->fo_dentry_O_groups,
1113                                  old_count * sizeof(*new_groups));
1114                         OBD_FREE(filter->fo_last_objid_files,
1115                                  old_count * sizeof(*new_files));
1116                 }
1117                 filter->fo_last_objids = new_objids;
1118                 filter->fo_dentry_O_sub = new_subdirs;
1119                 filter->fo_dentry_O_groups = new_groups;
1120                 filter->fo_last_objid_files = new_files;
1121                 filter->fo_group_count = len;
1122         }
1123
1124         filter->fo_dentry_O_groups[group] = dentry;
1125         filter->fo_last_objid_files[group] = filp;
1126         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1127                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1128                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1129         }
1130
1131         filter_update_last_group(obd, group);
1132
1133         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1134                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1135                 rc = filter_update_last_objid(obd, group, 1);
1136                 RETURN(rc);
1137         }
1138
1139         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1140         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1141                obd->obd_name, group, last_objid);
1142         RETURN(0);
1143  cleanup:
1144         switch (stage) {
1145         case 4:
1146                 if (new_objids != NULL)
1147                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1148                 if (new_subdirs != NULL)
1149                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1150                 if (new_groups != NULL)
1151                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1152                 if (new_files != NULL)
1153                         OBD_FREE(new_files, len * sizeof(*new_files));
1154         case 3:
1155                 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1156                         for (i = 0; i < filter->fo_subdir_count; i++) {
1157                                 if (tmp_subdirs->dentry[i] != NULL)
1158                                         dput(tmp_subdirs->dentry[i]);
1159                         }
1160                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1161                 }
1162         case 2:
1163                 filp_close(filp, 0);
1164         case 1:
1165                 dput(dentry);
1166         }
1167         RETURN(rc);
1168 }
1169
1170 static int filter_read_groups(struct obd_device *obd, int last_group,
1171                               int create)
1172 {
1173         struct filter_obd *filter = &obd->u.filter;
1174         int old_count, group, rc = 0;
1175
1176         cfs_down(&filter->fo_init_lock);
1177         old_count = filter->fo_group_count;
1178         for (group = old_count; group <= last_group; group++) {
1179                 rc = filter_read_group_internal(obd, group, create);
1180                 if (rc != 0)
1181                         break;
1182         }
1183         cfs_up(&filter->fo_init_lock);
1184         return rc;
1185 }
1186
1187 /* FIXME: object groups */
1188 static int filter_prep_groups(struct obd_device *obd)
1189 {
1190         struct filter_obd *filter = &obd->u.filter;
1191         struct dentry *O_dentry;
1192         struct file *filp;
1193         int    last_group, rc = 0, cleanup_phase = 0;
1194         loff_t off = 0;
1195         ENTRY;
1196
1197         O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1198                                 "O", 0700, 1);
1199         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1200         if (IS_ERR(O_dentry)) {
1201                 rc = PTR_ERR(O_dentry);
1202                 CERROR("cannot open/create O: rc = %d\n", rc);
1203                 GOTO(cleanup, rc);
1204         }
1205         filter->fo_dentry_O = O_dentry;
1206         cleanup_phase = 1; /* O_dentry */
1207
1208         /* we have to initialize all groups before first connections from
1209          * clients because they may send create/destroy for any group -bzzz */
1210         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1211         if (IS_ERR(filp)) {
1212                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1213                 GOTO(cleanup, rc = PTR_ERR(filp));
1214         }
1215         cleanup_phase = 2; /* filp */
1216
1217         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1218         if (rc) {
1219                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1220                 GOTO(cleanup, rc);
1221         }
1222
1223         if (off == 0)
1224                 last_group = FID_SEQ_OST_MDT0;
1225
1226         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1227               FID_SEQ_OST_MDT0, last_group);
1228         filter->fo_committed_group = last_group;
1229         rc = filter_read_groups(obd, last_group, 1);
1230         if (rc)
1231                 GOTO(cleanup, rc);
1232
1233         filp_close(filp, 0);
1234         RETURN(0);
1235
1236  cleanup:
1237         switch (cleanup_phase) {
1238         case 2:
1239                 filp_close(filp, 0);
1240         case 1:
1241                 filter_cleanup_groups(obd);
1242                 f_dput(filter->fo_dentry_O);
1243                 filter->fo_dentry_O = NULL;
1244         default:
1245                 break;
1246         }
1247         return rc;
1248
1249 }
1250
1251 /* setup the object store with correct subdirectories */
1252 static int filter_prep(struct obd_device *obd)
1253 {
1254         struct lvfs_run_ctxt saved;
1255         struct filter_obd *filter = &obd->u.filter;
1256         struct file *file;
1257         struct inode *inode;
1258         int rc = 0;
1259         ENTRY;
1260
1261         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1262         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1263         if (!file || IS_ERR(file)) {
1264                 rc = PTR_ERR(file);
1265                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1266                        LAST_RCVD, rc);
1267                 GOTO(out, rc);
1268         }
1269         obd->u.obt.obt_rcvd_filp = file;
1270         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1271                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1272                        file->f_dentry->d_inode->i_mode);
1273                 GOTO(err_filp, rc = -ENOENT);
1274         }
1275
1276         inode = file->f_dentry->d_parent->d_inode;
1277         /* We use i_op->unlink directly in filter_vfs_unlink() */
1278         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1279                 CERROR("%s: filesystem does not support create/unlink ops\n",
1280                        obd->obd_name);
1281                 GOTO(err_filp, rc = -EOPNOTSUPP);
1282         }
1283
1284         rc = filter_init_server_data(obd, file);
1285         if (rc) {
1286                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1287                 GOTO(err_filp, rc);
1288         }
1289         LASSERT(obd->u.obt.obt_lut);
1290         target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1291
1292         /* open and create health check io file*/
1293         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1294         if (IS_ERR(file)) {
1295                 rc = PTR_ERR(file);
1296                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1297                        HEALTH_CHECK, rc);
1298                 GOTO(err_server_data, rc);
1299         }
1300         filter->fo_obt.obt_health_check_filp = file;
1301         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1302                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1303                        file->f_dentry->d_inode->i_mode);
1304                 GOTO(err_health_check, rc = -ENOENT);
1305         }
1306         rc = lvfs_check_io_health(obd, file);
1307         if (rc)
1308                 GOTO(err_health_check, rc);
1309
1310         rc = filter_prep_groups(obd);
1311         if (rc)
1312                 GOTO(err_health_check, rc);
1313 out:
1314         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1315
1316         return(rc);
1317
1318 err_health_check:
1319         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1320                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1321         filter->fo_obt.obt_health_check_filp = NULL;
1322 err_server_data:
1323         target_recovery_fini(obd);
1324         filter_free_server_data(&obd->u.obt);
1325 err_filp:
1326         if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1327                 CERROR("can't close %s after error\n", LAST_RCVD);
1328         obd->u.obt.obt_rcvd_filp = NULL;
1329         goto out;
1330 }
1331
1332 /* cleanup the filter: write last used object id to status file */
1333 static void filter_post(struct obd_device *obd)
1334 {
1335         struct lvfs_run_ctxt saved;
1336         struct filter_obd *filter = &obd->u.filter;
1337         int rc, i;
1338
1339         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1340          * best to start a transaction with h_sync, because we removed this
1341          * from lastobjid */
1342
1343         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1344         rc = filter_update_server_data(obd);
1345         if (rc)
1346                 CERROR("error writing server data: rc = %d\n", rc);
1347
1348         for (i = 0; i < filter->fo_group_count; i++) {
1349                 rc = filter_update_last_objid(obd, i,
1350                                 (i == filter->fo_group_count - 1));
1351                 if (rc)
1352                         CERROR("error writing group %d lastobjid: rc = %d\n",
1353                                i, rc);
1354         }
1355
1356         rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1357         obd->u.obt.obt_rcvd_filp = NULL;
1358         if (rc)
1359                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1360
1361         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1362         filter->fo_obt.obt_health_check_filp = NULL;
1363         if (rc)
1364                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1365
1366         filter_cleanup_groups(obd);
1367         filter_free_server_data(&obd->u.obt);
1368         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1369
1370         filter_free_capa_keys(filter);
1371         cleanup_capa_hash(filter->fo_capa_hash);
1372 }
1373
1374 static void filter_set_last_id(struct filter_obd *filter,
1375                                obd_id id, obd_seq group)
1376 {
1377         LASSERT(group <= filter->fo_group_count);
1378
1379         cfs_spin_lock(&filter->fo_objidlock);
1380         filter->fo_last_objids[group] = id;
1381         cfs_spin_unlock(&filter->fo_objidlock);
1382 }
1383
1384 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1385 {
1386         obd_id id;
1387         LASSERT(group <= filter->fo_group_count);
1388         LASSERT(filter->fo_last_objids != NULL);
1389
1390         /* FIXME: object groups */
1391         cfs_spin_lock(&filter->fo_objidlock);
1392         id = filter->fo_last_objids[group];
1393         cfs_spin_unlock(&filter->fo_objidlock);
1394         return id;
1395 }
1396
1397 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1398 {
1399         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1400         return 0;
1401 }
1402
1403 /* We never dget the object parent, so DON'T dput it either */
1404 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1405 {
1406         struct filter_obd *filter = &obd->u.filter;
1407         struct filter_subdirs *subdirs;
1408
1409         if (group >= filter->fo_group_count) /* FIXME: object groups */
1410                 return ERR_PTR(-EBADF);
1411
1412         if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1413                 return filter->fo_dentry_O_groups[group];
1414
1415         subdirs = &filter->fo_dentry_O_sub[group];
1416         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1417 }
1418
1419 /* We never dget the object parent, so DON'T dput it either */
1420 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1421                                   obd_id objid)
1422 {
1423         unsigned long now = jiffies;
1424         struct dentry *dparent = filter_parent(obd, group, objid);
1425         int rc;
1426
1427         if (IS_ERR(dparent))
1428                 return dparent;
1429         if (dparent == NULL)
1430                 return ERR_PTR(-ENOENT);
1431
1432         rc = filter_lock_dentry(obd, dparent);
1433         fsfilt_check_slow(obd, now, "parent lock");
1434         return rc ? ERR_PTR(rc) : dparent;
1435 }
1436
1437 /* We never dget the object parent, so DON'T dput it either */
1438 static void filter_parent_unlock(struct dentry *dparent)
1439 {
1440         UNLOCK_INODE_MUTEX(dparent->d_inode);
1441 }
1442
1443 /* How to get files, dentries, inodes from object id's.
1444  *
1445  * If dir_dentry is passed, the caller has already locked the parent
1446  * appropriately for this operation (normally a write lock).  If
1447  * dir_dentry is NULL, we do a read lock while we do the lookup to
1448  * avoid races with create/destroy and such changing the directory
1449  * internal to the filesystem code. */
1450 struct dentry *filter_fid2dentry(struct obd_device *obd,
1451                                  struct dentry *dir_dentry,
1452                                  obd_seq group, obd_id id)
1453 {
1454         struct dentry *dparent = dir_dentry;
1455         struct dentry *dchild;
1456         char name[32];
1457         int len;
1458         ENTRY;
1459
1460         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1461             obd->u.filter.fo_destroys_in_progress == 0) {
1462                 /* don't fail lookups for orphan recovery, it causes
1463                  * later LBUGs when objects still exist during precreate */
1464                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1465                 RETURN(ERR_PTR(-ENOENT));
1466         }
1467         if (id == 0) {
1468                 CERROR("fatal: invalid object id 0\n");
1469                 RETURN(ERR_PTR(-ESTALE));
1470         }
1471
1472         len = sprintf(name, LPU64, id);
1473         if (dir_dentry == NULL) {
1474                 dparent = filter_parent_lock(obd, group, id);
1475                 if (IS_ERR(dparent)) {
1476                         CERROR("%s: error getting object "POSTID
1477                                " parent: rc %ld\n", obd->obd_name,
1478                                id, group, PTR_ERR(dparent));
1479                         RETURN(dparent);
1480                 }
1481         }
1482         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1483                dparent->d_name.len, dparent->d_name.name, name);
1484         /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1485         dchild = lookup_one_len(name, dparent, len);
1486         if (dir_dentry == NULL)
1487                 filter_parent_unlock(dparent);
1488         if (IS_ERR(dchild)) {
1489                 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1490                        obd->obd_name, id, group, PTR_ERR(dchild));
1491                 RETURN(dchild);
1492         }
1493
1494         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1495                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1496                        obd->obd_name, id, dchild->d_inode->i_ino);
1497                 f_dput(dchild);
1498                 RETURN(ERR_PTR(-ENOENT));
1499         }
1500
1501         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1502                name, dchild, atomic_read(&dchild->d_count));
1503
1504         LASSERT(atomic_read(&dchild->d_count) > 0);
1505
1506         RETURN(dchild);
1507 }
1508
1509 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1510                                   obd_id group, struct lustre_handle *lockh)
1511 {
1512         int flags = LDLM_AST_DISCARD_DATA, rc;
1513         struct ldlm_res_id res_id;
1514         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1515         ENTRY;
1516
1517         osc_build_res_name(objid, group, &res_id);
1518         /* Tell the clients that the object is gone now and that they should
1519          * throw away any cached pages. */
1520         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1521                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1522                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1523                                     lockh);
1524         if (rc != ELDLM_OK)
1525                 lockh->cookie = 0;
1526         RETURN(rc);
1527 }
1528
1529 static void filter_fini_destroy(struct obd_device *obd,
1530                                 struct lustre_handle *lockh)
1531 {
1532         if (lustre_handle_is_used(lockh))
1533                 ldlm_lock_decref(lockh, LCK_PW);
1534 }
1535
1536 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1537  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1538  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1539  * reverse.  Caller must take i_sem before starting the transaction and we
1540  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1541 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1542                       struct vfsmount *mnt)
1543 {
1544         int rc;
1545         ENTRY;
1546
1547         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1548          * itself we already hold dir->i_mutex for child create/unlink ops */
1549         LASSERT(dentry->d_inode != NULL);
1550         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1551         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1552
1553
1554         /* may_delete() */
1555         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1556                 GOTO(out, rc = -ENOENT);
1557
1558         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1559         if (rc)
1560                 GOTO(out, rc);
1561
1562         if (IS_APPEND(dir))
1563                 GOTO(out, rc = -EPERM);
1564
1565         /* check_sticky() */
1566         if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1567              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1568             IS_IMMUTABLE(dentry->d_inode))
1569                 GOTO(out, rc = -EPERM);
1570
1571         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1572          *       that was done because of journal_start (which is already done
1573          *       here) or some other ordering issue. */
1574         ll_vfs_dq_init(dir);
1575
1576         rc = ll_security_inode_unlink(dir, dentry, mnt);
1577         if (rc)
1578                 GOTO(out, rc);
1579
1580         rc = dir->i_op->unlink(dir, dentry);
1581 out:
1582         /* need to drop i_mutex before we lose inode reference */
1583         UNLOCK_INODE_MUTEX(dentry->d_inode);
1584         if (rc == 0)
1585                 d_delete(dentry);
1586
1587         RETURN(rc);
1588 }
1589
1590 /* Caller must hold LCK_PW on parent and push us into kernel context.
1591  * Caller must hold child i_mutex, we drop it always.
1592  * Caller is also required to ensure that dchild->d_inode exists. */
1593 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1594                                    obd_seq group, struct dentry *dparent,
1595                                    struct dentry *dchild)
1596 {
1597         struct inode *inode = dchild->d_inode;
1598         int rc;
1599
1600         /* There should be 2 references to the inode:
1601          *  1) taken by filter_prepare_destroy
1602          *  2) taken by filter_destroy */
1603         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1604                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1605                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1606                        (unsigned long)inode->i_nlink,
1607                        atomic_read(&inode->i_count));
1608         }
1609
1610         rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1611         if (rc)
1612                 CERROR("error unlinking objid %.*s: rc %d\n",
1613                        dchild->d_name.len, dchild->d_name.name, rc);
1614         return(rc);
1615 }
1616
1617 struct filter_intent_args {
1618         struct ldlm_lock **victim;
1619         __u64 size;
1620         int *liblustre;
1621 };
1622
1623 static enum interval_iter filter_intent_cb(struct interval_node *n,
1624                                            void *args)
1625 {
1626         struct ldlm_interval *node = (struct ldlm_interval *)n;
1627         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1628         __u64 size = arg->size;
1629         struct ldlm_lock **v = arg->victim;
1630         struct ldlm_lock *lck;
1631
1632         /* If the interval is lower than the current file size,
1633          * just break. */
1634         if (interval_high(n) <= size)
1635                 return INTERVAL_ITER_STOP;
1636
1637         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1638                 /* Don't send glimpse ASTs to liblustre clients.
1639                  * They aren't listening for them, and they do
1640                  * entirely synchronous I/O anyways. */
1641                 if (lck->l_export == NULL ||
1642                     lck->l_export->exp_libclient == 1)
1643                         continue;
1644
1645                 if (*arg->liblustre)
1646                         *arg->liblustre = 0;
1647
1648                 if (*v == NULL) {
1649                         *v = LDLM_LOCK_GET(lck);
1650                 } else if ((*v)->l_policy_data.l_extent.start <
1651                            lck->l_policy_data.l_extent.start) {
1652                         LDLM_LOCK_RELEASE(*v);
1653                         *v = LDLM_LOCK_GET(lck);
1654                 }
1655
1656                 /* the same policy group - every lock has the
1657                  * same extent, so needn't do it any more */
1658                 break;
1659         }
1660
1661         return INTERVAL_ITER_CONT;
1662 }
1663
1664 static int filter_intent_policy(struct ldlm_namespace *ns,
1665                                 struct ldlm_lock **lockp, void *req_cookie,
1666                                 ldlm_mode_t mode, int flags, void *data)
1667 {
1668         CFS_LIST_HEAD(rpc_list);
1669         struct ptlrpc_request *req = req_cookie;
1670         struct ldlm_lock *lock = *lockp, *l = NULL;
1671         struct ldlm_resource *res = lock->l_resource;
1672         ldlm_processing_policy policy;
1673         struct ost_lvb *res_lvb, *reply_lvb;
1674         struct ldlm_reply *rep;
1675         ldlm_error_t err;
1676         int idx, rc, tmpflags = 0, only_liblustre = 1;
1677         struct ldlm_interval_tree *tree;
1678         struct filter_intent_args arg;
1679         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1680                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1681                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1682         ENTRY;
1683
1684         policy = ldlm_get_processing_policy(res);
1685         LASSERT(policy != NULL);
1686         LASSERT(req != NULL);
1687
1688         rc = lustre_pack_reply(req, 3, repsize, NULL);
1689         if (rc)
1690                 RETURN(req->rq_status = rc);
1691
1692         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1693         LASSERT(rep != NULL);
1694
1695         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1696                                    sizeof(*reply_lvb));
1697         LASSERT(reply_lvb != NULL);
1698
1699         //fixup_handle_for_resent_req(req, lock, &lockh);
1700
1701         /* Call the extent policy function to see if our request can be
1702          * granted, or is blocked.
1703          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1704          * lock, and should not be granted if the lock will be blocked.
1705          */
1706
1707         LASSERT(ns == ldlm_res_to_ns(res));
1708         lock_res(res);
1709         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1710         check_res_locked(res);
1711
1712         /* FIXME: we should change the policy function slightly, to not make
1713          * this list at all, since we just turn around and free it */
1714         while (!cfs_list_empty(&rpc_list)) {
1715                 struct ldlm_lock *wlock =
1716                         cfs_list_entry(rpc_list.next, struct ldlm_lock,
1717                                        l_cp_ast);
1718                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1719                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1720                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1721                 cfs_list_del_init(&wlock->l_cp_ast);
1722                 LDLM_LOCK_RELEASE(wlock);
1723         }
1724
1725         /* The lock met with no resistance; we're finished. */
1726         if (rc == LDLM_ITER_CONTINUE) {
1727                 /* do not grant locks to the liblustre clients: they cannot
1728                  * handle ASTs robustly.  We need to do this while still
1729                  * holding lr_lock to avoid the lock remaining on the res_link
1730                  * list (and potentially being added to l_pending_list by an
1731                  * AST) when we are going to drop this lock ASAP. */
1732                 if (lock->l_export->exp_libclient ||
1733                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1734                         ldlm_resource_unlink_lock(lock);
1735                         err = ELDLM_LOCK_ABORTED;
1736                 } else {
1737                         err = ELDLM_LOCK_REPLACED;
1738                 }
1739                 unlock_res(res);
1740                 RETURN(err);
1741         }
1742
1743         /* Do not grant any lock, but instead send GL callbacks.  The extent
1744          * policy nicely created a list of all PW locks for us.  We will choose
1745          * the highest of those which are larger than the size in the LVB, if
1746          * any, and perform a glimpse callback. */
1747         res_lvb = res->lr_lvb_data;
1748         LASSERT(res_lvb != NULL);
1749         *reply_lvb = *res_lvb;
1750
1751         /*
1752          * lr_lock guarantees that no new locks are granted, and,
1753          * therefore, that res->lr_lvb_data cannot increase beyond the
1754          * end of already granted lock. As a result, it is safe to
1755          * check against "stale" reply_lvb->lvb_size value without
1756          * res->lr_lvb_sem.
1757          */
1758         arg.size = reply_lvb->lvb_size;
1759         arg.victim = &l;
1760         arg.liblustre = &only_liblustre;
1761         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1762                 tree = &res->lr_itree[idx];
1763                 if (tree->lit_mode == LCK_PR)
1764                         continue;
1765
1766                 interval_iterate_reverse(tree->lit_root,
1767                                          filter_intent_cb, &arg);
1768         }
1769         unlock_res(res);
1770
1771         /* There were no PW locks beyond the size in the LVB; finished. */
1772         if (l == NULL) {
1773                 if (only_liblustre) {
1774                         /* If we discovered a liblustre client with a PW lock,
1775                          * however, the LVB may be out of date!  The LVB is
1776                          * updated only on glimpse (which we don't do for
1777                          * liblustre clients) and cancel (which the client
1778                          * obviously has not yet done).  So if it has written
1779                          * data but kept the lock, the LVB is stale and needs
1780                          * to be updated from disk.
1781                          *
1782                          * Of course, this will all disappear when we switch to
1783                          * taking liblustre locks on the OST. */
1784                         ldlm_res_lvbo_update(res, NULL, 1);
1785                 }
1786                 RETURN(ELDLM_LOCK_ABORTED);
1787         }
1788
1789         /*
1790          * This check is for lock taken in filter_prepare_destroy() that does
1791          * not have l_glimpse_ast set. So the logic is: if there is a lock
1792          * with no l_glimpse_ast set, this object is being destroyed already.
1793          *
1794          * Hence, if you are grabbing DLM locks on the server, always set
1795          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1796          */
1797         if (l->l_glimpse_ast == NULL) {
1798                 /* We are racing with unlink(); just return -ENOENT */
1799                 rep->lock_policy_res1 = -ENOENT;
1800                 goto out;
1801         }
1802
1803         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1804         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1805
1806         lock_res(res);
1807         *reply_lvb = *res_lvb;
1808         unlock_res(res);
1809
1810  out:
1811         LDLM_LOCK_RELEASE(l);
1812
1813         RETURN(ELDLM_LOCK_ABORTED);
1814 }
1815
1816 /*
1817  * per-obd_device iobuf pool.
1818  *
1819  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1820  * allocations in write-path has to be minimized (see bug 5137).
1821  *
1822  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1823  * OST threads (see ost_thread_{init,done}()).
1824  *
1825  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1826  * at the OST layer there are only (potentially) multiple obd_device of type
1827  * unknown at the time of OST thread creation.
1828  *
1829  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1830  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1831  * it's very own iobuf.
1832  *
1833  * Functions below
1834  *
1835  *     filter_kiobuf_pool_init()
1836  *
1837  *     filter_kiobuf_pool_done()
1838  *
1839  *     filter_iobuf_get()
1840  *
1841  * operate on this array. They are "generic" in a sense that they don't depend
1842  * on actual type of iobuf's (the latter depending on Linux kernel version).
1843  */
1844
1845 /*
1846  * destroy pool created by filter_iobuf_pool_init
1847  */
1848 static void filter_iobuf_pool_done(struct filter_obd *filter)
1849 {
1850         struct filter_iobuf **pool;
1851         int i;
1852
1853         ENTRY;
1854
1855         pool = filter->fo_iobuf_pool;
1856         if (pool != NULL) {
1857                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1858                         if (pool[i] != NULL)
1859                                 filter_free_iobuf(pool[i]);
1860                 }
1861                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1862                 filter->fo_iobuf_pool = NULL;
1863         }
1864         EXIT;
1865 }
1866
1867 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1868 {
1869         struct filter_obd       *filter = &obd->u.filter;
1870         struct sptlrpc_rule_set  tmp_rset;
1871         int                      rc;
1872
1873         sptlrpc_rule_set_init(&tmp_rset);
1874         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1875         if (rc) {
1876                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1877                        obd->obd_name, rc);
1878                 return rc;
1879         }
1880
1881         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1882
1883         cfs_write_lock(&filter->fo_sptlrpc_lock);
1884         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1885         filter->fo_sptlrpc_rset = tmp_rset;
1886         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1887
1888         return 0;
1889 }
1890
1891 /*
1892  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1893  */
1894 static int filter_iobuf_pool_init(struct filter_obd *filter)
1895 {
1896         void **pool;
1897
1898         ENTRY;
1899
1900
1901         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1902                       GFP_KERNEL);
1903         if (filter->fo_iobuf_pool == NULL)
1904                 RETURN(-ENOMEM);
1905
1906         filter->fo_iobuf_count = OSS_THREADS_MAX;
1907
1908         RETURN(0);
1909 }
1910
1911 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1912  * many threads there will be so we allocate a large empty array and only
1913  * fill in those slots that are actually in use.
1914  * If we haven't allocated a pool entry for this thread before, do so now. */
1915 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1916 {
1917         int thread_id                    = (oti && oti->oti_thread) ?
1918                                            oti->oti_thread->t_id : -1;
1919         struct filter_iobuf  *pool       = NULL;
1920         struct filter_iobuf **pool_place = NULL;
1921
1922         if (thread_id >= 0) {
1923                 LASSERT(thread_id < filter->fo_iobuf_count);
1924                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1925         }
1926
1927         if (unlikely(pool == NULL)) {
1928                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1929                                           PTLRPC_MAX_BRW_PAGES);
1930                 if (pool_place != NULL)
1931                         *pool_place = pool;
1932         }
1933
1934         return pool;
1935 }
1936
1937 /* mount the file system (secretly).  lustre_cfg parameters are:
1938  * 1 = device
1939  * 2 = fstype
1940  * 3 = flags: failover=f, failout=n
1941  * 4 = mount options
1942  */
1943 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1944                         void *option)
1945 {
1946         struct filter_obd *filter = &obd->u.filter;
1947         struct vfsmount *mnt;
1948         struct lustre_mount_info *lmi;
1949         struct obd_uuid uuid;
1950         __u8 *uuid_ptr;
1951         char *str, *label;
1952         char ns_name[48];
1953         struct request_queue *q;
1954         int rc, i;
1955         ENTRY;
1956
1957         if (lcfg->lcfg_bufcount < 3 ||
1958             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1959             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1960                 RETURN(-EINVAL);
1961
1962         lmi = server_get_mount(obd->obd_name);
1963         if (lmi) {
1964                 /* We already mounted in lustre_fill_super.
1965                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1966                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1967                 mnt = lmi->lmi_mnt;
1968                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1969
1970                 /* gets recovery timeouts from mount data */
1971                 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1972                         obd->obd_recovery_timeout =
1973                                 lsi->lsi_lmd->lmd_recovery_time_soft;
1974                 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1975                         obd->obd_recovery_time_hard =
1976                                 lsi->lsi_lmd->lmd_recovery_time_hard;
1977         } else {
1978                 /* old path - used by lctl */
1979                 CERROR("Using old MDS mount method\n");
1980                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1981                                     MS_NOATIME|MS_NODIRATIME,
1982                                     lustre_cfg_string(lcfg, 1), option);
1983                 if (IS_ERR(mnt)) {
1984                         rc = PTR_ERR(mnt);
1985                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1986                                            lustre_cfg_string(lcfg, 1), rc);
1987                         RETURN(rc);
1988                 }
1989
1990                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1991         }
1992         if (IS_ERR(obd->obd_fsops))
1993                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1994
1995         rc = filter_iobuf_pool_init(filter);
1996         if (rc != 0)
1997                 GOTO(err_ops, rc);
1998
1999         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2000                 CERROR("%s: Underlying device is marked as read-only. "
2001                        "Setup failed\n", obd->obd_name);
2002                 GOTO(err_ops, rc = -EROFS);
2003         }
2004
2005         /* failover is the default */
2006         obd->obd_replayable = 1;
2007
2008         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2009                 str = lustre_cfg_string(lcfg, 3);
2010                 if (strchr(str, 'n')) {
2011                         CWARN("%s: recovery disabled\n", obd->obd_name);
2012                         obd->obd_replayable = 0;
2013                 }
2014         }
2015
2016         obd->u.obt.obt_vfsmnt = mnt;
2017         obd->u.obt.obt_sb = mnt->mnt_sb;
2018         obd->u.obt.obt_magic = OBT_MAGIC;
2019         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2020         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2021
2022         fsfilt_setup(obd, obd->u.obt.obt_sb);
2023
2024         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2025         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2026         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2027         obd->obd_lvfs_ctxt.fs = get_ds();
2028         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2029
2030         cfs_init_mutex(&filter->fo_init_lock);
2031         filter->fo_committed_group = 0;
2032         filter->fo_destroys_in_progress = 0;
2033         for (i = 0; i < 32; i++)
2034                 cfs_sema_init(&filter->fo_create_locks[i], 1);
2035
2036         cfs_spin_lock_init(&filter->fo_objidlock);
2037         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2038         cfs_sema_init(&filter->fo_alloc_lock, 1);
2039         init_brw_stats(&filter->fo_filter_stats);
2040         cfs_spin_lock_init(&filter->fo_flags_lock);
2041         filter->fo_read_cache = 1; /* enable read-only cache by default */
2042         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2043         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2044         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2045         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2046         filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2047         filter_slc_set(filter); /* initialize sync on lock cancel */
2048
2049         rc = filter_prep(obd);
2050         if (rc)
2051                 GOTO(err_ops, rc);
2052
2053         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2054         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2055
2056         filter->fo_fl_oss_capa = 1;
2057
2058         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2059         filter->fo_capa_hash = init_capa_hash();
2060         if (filter->fo_capa_hash == NULL)
2061                 GOTO(err_post, rc = -ENOMEM);
2062
2063         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2064         obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2065                                                 LDLM_NAMESPACE_SERVER,
2066                                                 LDLM_NAMESPACE_GREEDY,
2067                                                 LDLM_NS_TYPE_OST);
2068         if (obd->obd_namespace == NULL)
2069                 GOTO(err_post, rc = -ENOMEM);
2070         obd->obd_namespace->ns_lvbp = obd;
2071         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2072         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2073
2074         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2075                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2076
2077         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2078         if (rc) {
2079                 CERROR("failed to setup llogging subsystems\n");
2080                 GOTO(err_post, rc);
2081         }
2082
2083         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2084         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2085         /* do this after llog being initialized */
2086         filter_adapt_sptlrpc_conf(obd, 1);
2087
2088         rc = lquota_setup(filter_quota_interface_ref, obd);
2089         if (rc)
2090                 GOTO(err_post, rc);
2091
2092         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2093         if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2094             queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2095                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2096                               "for larger I/O requests: max_sectors = %u "
2097                               "could be up to max_hw_sectors=%u\n",
2098                               obd->obd_name, mnt->mnt_sb->s_id,
2099                               queue_max_sectors(q), queue_max_hw_sectors(q));
2100
2101         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2102         if (uuid_ptr != NULL) {
2103                 class_uuid_unparse(uuid_ptr, &uuid);
2104                 str = uuid.uuid;
2105         } else {
2106                 str = "no UUID";
2107         }
2108
2109         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2110         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2111                       obd->obd_name, label ?: str, lmi ? "on " : "",
2112                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2113                       obd->obd_replayable ? "enabled" : "disabled");
2114
2115         if (obd->obd_recovering)
2116                 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2117                               "or until %d client%s reconnect%s\n",
2118                               obd->obd_name,
2119                               obd->obd_recovery_timeout / 60,
2120                               obd->obd_recovery_timeout % 60,
2121                               obd->obd_max_recoverable_clients,
2122                               (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2123                               (obd->obd_max_recoverable_clients == 1) ? "s": "");
2124
2125
2126         RETURN(0);
2127
2128 err_post:
2129         filter_post(obd);
2130 err_ops:
2131         fsfilt_put_ops(obd->obd_fsops);
2132         filter_iobuf_pool_done(filter);
2133 err_mntput:
2134         server_put_mount(obd->obd_name, mnt);
2135         obd->u.obt.obt_sb = 0;
2136         return rc;
2137 }
2138
2139 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2140 {
2141         struct lprocfs_static_vars lvars;
2142         unsigned long addr;
2143         struct page *page;
2144         int rc;
2145
2146         CLASSERT(offsetof(struct obd_device, u.obt) ==
2147                  offsetof(struct obd_device, u.filter.fo_obt));
2148
2149         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2150                 RETURN(-EINVAL);
2151
2152         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2153         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2154         if (!page)
2155                 RETURN(-ENOMEM);
2156         addr = (unsigned long)cfs_page_address(page);
2157         clear_page((void *)addr);
2158
2159         /* lprocfs must be setup before the filter so state can be safely added
2160          * to /proc incrementally as the filter is setup */
2161         lprocfs_filter_init_vars(&lvars);
2162         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2163             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2164                 /* Init obdfilter private stats here */
2165                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2166                                      LPROCFS_CNTR_AVGMINMAX,
2167                                      "read_bytes", "bytes");
2168                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2169                                      LPROCFS_CNTR_AVGMINMAX,
2170                                      "write_bytes", "bytes");
2171                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2172                                      LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2173                                      "get_page", "usec");
2174                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2175                                      LPROCFS_CNTR_AVGMINMAX,
2176                                      "get_page_failures", "num");
2177                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2178                                      LPROCFS_CNTR_AVGMINMAX,
2179                                      "cache_access", "pages");
2180                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2181                                      LPROCFS_CNTR_AVGMINMAX,
2182                                      "cache_hit", "pages");
2183                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2184                                      LPROCFS_CNTR_AVGMINMAX,
2185                                      "cache_miss", "pages");
2186
2187                 lproc_filter_attach_seqstat(obd);
2188                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2189                                                         obd->obd_proc_entry,
2190                                                         NULL, NULL);
2191                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2192                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2193                         CERROR("error %d setting up lprocfs for %s\n",
2194                                rc, "exports");
2195                         obd->obd_proc_exports_entry = NULL;
2196                 }
2197         }
2198         if (obd->obd_proc_exports_entry)
2199                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2200                                    lprocfs_nid_stats_clear_read,
2201                                    lprocfs_nid_stats_clear_write, obd, NULL);
2202
2203         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2204                LUSTRE_CFG_BUFLEN(lcfg, 4));
2205         rc = filter_common_setup(obd, lcfg, (void *)addr);
2206         OBD_PAGE_FREE(page);
2207
2208         if (rc) {
2209                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2210                 lprocfs_free_per_client_stats(obd);
2211                 lprocfs_free_obd_stats(obd);
2212                 lprocfs_obd_cleanup(obd);
2213         }
2214
2215         return rc;
2216 }
2217
2218 static struct llog_operations filter_mds_ost_repl_logops;
2219
2220 static struct llog_operations filter_size_orig_logops = {
2221         .lop_setup   = llog_obd_origin_setup,
2222         .lop_cleanup = llog_obd_origin_cleanup,
2223         .lop_add     = llog_obd_origin_add
2224 };
2225
2226 static int filter_olg_fini(struct obd_llog_group *olg)
2227 {
2228         struct llog_ctxt *ctxt;
2229         int rc = 0, rc2 = 0;
2230         ENTRY;
2231
2232         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2233         if (ctxt)
2234                 rc = llog_cleanup(ctxt);
2235
2236         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2237         if (ctxt) {
2238                 rc2 = llog_cleanup(ctxt);
2239                 if (!rc)
2240                         rc = rc2;
2241         }
2242
2243         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2244         if (ctxt) {
2245                 rc2 = llog_cleanup(ctxt);
2246                 if (!rc)
2247                         rc = rc2;
2248         }
2249
2250         RETURN(rc);
2251 }
2252
2253 static int
2254 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2255                 struct obd_device *tgt)
2256 {
2257         int rc;
2258         ENTRY;
2259
2260         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2261                         &filter_mds_ost_repl_logops);
2262         if (rc)
2263                 GOTO(cleanup, rc);
2264
2265         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2266                         &filter_size_orig_logops);
2267         if (rc)
2268                 GOTO(cleanup, rc);
2269         EXIT;
2270 cleanup:
2271         if (rc)
2272                 filter_olg_fini(olg);
2273         return rc;
2274 }
2275
2276 /**
2277  * Init the default olg, which is embeded in the obd_device, for filter.
2278  */
2279 static int
2280 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2281                         struct obd_device *tgt)
2282 {
2283         struct filter_obd *filter = &obd->u.filter;
2284         struct llog_ctxt *ctxt;
2285         int rc;
2286         ENTRY;
2287
2288         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2289         if (!filter->fo_lcm)
2290                 RETURN(-ENOMEM);
2291
2292         filter_mds_ost_repl_logops = llog_client_ops;
2293         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2294         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2295         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2296
2297         rc = filter_olg_init(obd, olg, tgt);
2298         if (rc)
2299                 GOTO(cleanup_lcm, rc);
2300
2301         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2302                         &llog_lvfs_ops);
2303         if (rc)
2304                 GOTO(cleanup_olg, rc);
2305
2306         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2307         if (!ctxt) {
2308                 CERROR("Can't get ctxt for %p:%x\n", olg,
2309                        LLOG_MDS_OST_REPL_CTXT);
2310                 GOTO(cleanup_olg, rc = -ENODEV);
2311         }
2312         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2313         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2314         llog_ctxt_put(ctxt);
2315
2316         RETURN(0);
2317 cleanup_olg:
2318         filter_olg_fini(olg);
2319 cleanup_lcm:
2320         llog_recov_thread_fini(filter->fo_lcm, 1);
2321         filter->fo_lcm = NULL;
2322         return rc;
2323 }
2324
2325 static int
2326 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2327                  struct obd_device *tgt, int *index)
2328 {
2329         struct filter_obd *filter = &obd->u.filter;
2330         struct llog_ctxt *ctxt;
2331         int rc;
2332         ENTRY;
2333
2334         LASSERT(olg != NULL);
2335         if (olg == &obd->obd_olg)
2336                 return filter_default_olg_init(obd, olg, tgt);
2337
2338         LASSERT(filter->fo_lcm != NULL);
2339         rc = filter_olg_init(obd, olg, tgt);
2340         if (rc)
2341                 RETURN(rc);
2342         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2343         if (!ctxt) {
2344                 CERROR("Can't get ctxt for %p:%x\n", olg,
2345                        LLOG_MDS_OST_REPL_CTXT);
2346                 filter_olg_fini(olg);
2347                 RETURN(-ENODEV);
2348         }
2349         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2350         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2351         llog_ctxt_put(ctxt);
2352         RETURN(rc);
2353 }
2354
2355 static int filter_llog_finish(struct obd_device *obd, int count)
2356 {
2357         struct filter_obd *filter = &obd->u.filter;
2358         struct llog_ctxt *ctxt;
2359         ENTRY;
2360
2361         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2362         if (ctxt) {
2363                 /*
2364                  * Make sure that no cached llcds left in recov_thread.
2365                  * We actually do sync in disconnect time, but disconnect
2366                  * may not come being marked rq_no_resend = 1.
2367                  */
2368                 llog_sync(ctxt, NULL);
2369
2370                 /*
2371                  * Balance class_import_get() in llog_receptor_accept().
2372                  * This is safe to do, as llog is already synchronized
2373                  * and its import may go.
2374                  */
2375                 cfs_mutex_down(&ctxt->loc_sem);
2376                 if (ctxt->loc_imp) {
2377                         class_import_put(ctxt->loc_imp);
2378                         ctxt->loc_imp = NULL;
2379                 }
2380                 cfs_mutex_up(&ctxt->loc_sem);
2381                 llog_ctxt_put(ctxt);
2382         }
2383
2384         if (filter->fo_lcm) {
2385                 cfs_mutex_down(&ctxt->loc_sem);
2386                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2387                 filter->fo_lcm = NULL;
2388                 cfs_mutex_up(&ctxt->loc_sem);
2389         }
2390         RETURN(filter_olg_fini(&obd->obd_olg));
2391 }
2392
2393 /**
2394  * Find the group llog according to group index in the llog group list.
2395  */
2396 static struct obd_llog_group *
2397 filter_find_olg_internal(struct filter_obd *filter, int group)
2398 {
2399         struct obd_llog_group *olg;
2400
2401         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2402         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2403                 if (olg->olg_seq == group)
2404                         RETURN(olg);
2405         }
2406         RETURN(NULL);
2407 }
2408
2409 /**
2410  * Find the group llog according to group index on the filter
2411  */
2412 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2413 {
2414         struct obd_llog_group *olg = NULL;
2415         struct filter_obd *filter;
2416
2417         filter = &obd->u.filter;
2418
2419         if (group == FID_SEQ_LLOG)
2420                 RETURN(&obd->obd_olg);
2421
2422         cfs_spin_lock(&filter->fo_llog_list_lock);
2423         olg = filter_find_olg_internal(filter, group);
2424         cfs_spin_unlock(&filter->fo_llog_list_lock);
2425
2426         RETURN(olg);
2427 }
2428 /**
2429  * Find the llog_group of the filter according to the group. If it can not
2430  * find, create the llog_group, which only happens when mds is being synced
2431  * with OST.
2432  */
2433 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2434 {
2435         struct obd_llog_group *olg = NULL;
2436         struct filter_obd *filter;
2437         int rc;
2438
2439         filter = &obd->u.filter;
2440
2441         if (group == FID_SEQ_LLOG)
2442                 RETURN(&obd->obd_olg);
2443
2444         cfs_spin_lock(&filter->fo_llog_list_lock);
2445         olg = filter_find_olg_internal(filter, group);
2446         if (olg) {
2447                 if (olg->olg_initializing) {
2448                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2449                 } else {
2450                         GOTO(out_unlock, olg);
2451                 }
2452         }
2453         OBD_ALLOC_PTR(olg);
2454         if (olg == NULL)
2455                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2456
2457         llog_group_init(olg, group);
2458         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2459         olg->olg_initializing = 1;
2460         cfs_spin_unlock(&filter->fo_llog_list_lock);
2461
2462         rc = obd_llog_init(obd, olg, obd, NULL);
2463         if (rc) {
2464                cfs_spin_lock(&filter->fo_llog_list_lock);
2465                cfs_list_del(&olg->olg_list);
2466                cfs_spin_unlock(&filter->fo_llog_list_lock);
2467                OBD_FREE_PTR(olg);
2468                GOTO(out, olg = ERR_PTR(-ENOMEM));
2469         }
2470         cfs_spin_lock(&filter->fo_llog_list_lock);
2471         olg->olg_initializing = 0;
2472         cfs_spin_unlock(&filter->fo_llog_list_lock);
2473         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2474               obd->obd_name, group, olg);
2475 out:
2476         RETURN(olg);
2477
2478 out_unlock:
2479         cfs_spin_unlock(&filter->fo_llog_list_lock);
2480         GOTO(out, olg);
2481 }
2482
2483 static int filter_llog_connect(struct obd_export *exp,
2484                                struct llogd_conn_body *body)
2485 {
2486         struct obd_device *obd = exp->exp_obd;
2487         struct llog_ctxt *ctxt;
2488         struct obd_llog_group *olg;
2489         int rc;
2490         ENTRY;
2491
2492         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2493                obd->obd_name, body->lgdc_logid.lgl_oid,
2494                body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2495
2496         olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2497         if (!olg) {
2498                 CERROR(" %s: can not find olg of group %d\n",
2499                        obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2500                 RETURN(-ENOENT);
2501         }
2502         llog_group_set_export(olg, exp);
2503
2504         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2505         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2506                  body->lgdc_ctxt_idx);
2507
2508         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2509               obd->obd_name, body->lgdc_logid.lgl_oid,
2510               body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2511
2512         cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2513         obd->u.filter.fo_mds_ost_sync = 1;
2514         cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2515         rc = llog_connect(ctxt, &body->lgdc_logid,
2516                           &body->lgdc_gen, NULL);
2517         llog_ctxt_put(ctxt);
2518         if (rc != 0)
2519                 CERROR("failed to connect rc %d idx %d\n", rc,
2520                                 body->lgdc_ctxt_idx);
2521
2522         RETURN(rc);
2523 }
2524
2525 static int filter_llog_preclean(struct obd_device *obd)
2526 {
2527         struct obd_llog_group *olg, *tmp;
2528         struct filter_obd *filter;
2529         cfs_list_t  remove_list;
2530         int rc = 0;
2531         ENTRY;
2532
2533         rc = obd_llog_finish(obd, 0);
2534         if (rc)
2535                 CERROR("failed to cleanup llogging subsystem\n");
2536
2537         filter = &obd->u.filter;
2538         CFS_INIT_LIST_HEAD(&remove_list);
2539
2540         cfs_spin_lock(&filter->fo_llog_list_lock);
2541         while (!cfs_list_empty(&filter->fo_llog_list)) {
2542                 olg = cfs_list_entry(filter->fo_llog_list.next,
2543                                      struct obd_llog_group, olg_list);
2544                 cfs_list_del(&olg->olg_list);
2545                 cfs_list_add(&olg->olg_list, &remove_list);
2546         }
2547         cfs_spin_unlock(&filter->fo_llog_list_lock);
2548
2549         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2550                 cfs_list_del_init(&olg->olg_list);
2551                 rc = filter_olg_fini(olg);
2552                 if (rc)
2553                         CERROR("failed to cleanup llogging subsystem for %u\n",
2554                                olg->olg_seq);
2555                 OBD_FREE_PTR(olg);
2556         }
2557
2558         RETURN(rc);
2559 }
2560
2561 static int filter_precleanup(struct obd_device *obd,
2562                              enum obd_cleanup_stage stage)
2563 {
2564         int rc = 0;
2565         ENTRY;
2566
2567         switch(stage) {
2568         case OBD_CLEANUP_EARLY:
2569                 break;
2570         case OBD_CLEANUP_EXPORTS:
2571                 /* Stop recovery before namespace cleanup. */
2572                 target_recovery_fini(obd);
2573                 rc = filter_llog_preclean(obd);
2574                 break;
2575         }
2576         RETURN(rc);
2577 }
2578
2579 static int filter_cleanup(struct obd_device *obd)
2580 {
2581         struct filter_obd *filter = &obd->u.filter;
2582         ENTRY;
2583
2584         if (obd->obd_fail)
2585                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2586                               "will be preserved.\n", obd->obd_name);
2587
2588         obd_exports_barrier(obd);
2589         obd_zombie_barrier();
2590
2591         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2592         lprocfs_free_per_client_stats(obd);
2593         lprocfs_free_obd_stats(obd);
2594         lprocfs_obd_cleanup(obd);
2595         lquota_cleanup(filter_quota_interface_ref, obd);
2596
2597         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2598         obd->obd_namespace = NULL;
2599
2600         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2601
2602         if (obd->u.obt.obt_sb == NULL)
2603                 RETURN(0);
2604
2605         filter_post(obd);
2606
2607         ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2608         shrink_dcache_sb(obd->u.obt.obt_sb);
2609
2610         server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2611         obd->u.obt.obt_sb = NULL;
2612
2613         fsfilt_put_ops(obd->obd_fsops);
2614
2615         filter_iobuf_pool_done(filter);
2616
2617         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2618
2619         RETURN(0);
2620 }
2621
2622 static int filter_connect_internal(struct obd_export *exp,
2623                                    struct obd_connect_data *data,
2624                                    int reconnect)
2625 {
2626         struct filter_export_data *fed = &exp->exp_filter_data;
2627
2628         if (!data)
2629                 RETURN(0);
2630
2631         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2632                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2633                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2634                data->ocd_connect_flags, data->ocd_version,
2635                data->ocd_grant, data->ocd_index);
2636
2637         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2638                 CWARN("!!! This export (nid %s) used object group %d "
2639                        "earlier; now it's trying to use group %d!  This could "
2640                        "be a bug in the MDS. Please report to "
2641                        "http://bugzilla.lustre.org/\n",
2642                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2643                 RETURN(-EPROTO);
2644         }
2645         fed->fed_group = data->ocd_group;
2646
2647         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2648         exp->exp_connect_flags = data->ocd_connect_flags;
2649         data->ocd_version = LUSTRE_VERSION_CODE;
2650
2651         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2652         if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2653                   data->ocd_connect_flags & OBD_CONNECT_MDS))
2654                 RETURN(-EPROTO);
2655
2656         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2657                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2658                 obd_size left, want;
2659
2660                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2661                 left = filter_grant_space_left(exp);
2662                 want = data->ocd_grant;
2663                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2664                 data->ocd_grant = fed->fed_grant;
2665                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2666
2667                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2668                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2669                        exp->exp_client_uuid.uuid, exp,
2670                        data->ocd_grant, want, left);
2671
2672                 filter->fo_tot_granted_clients ++;
2673         }
2674
2675         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2676                 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2677                 int index = le32_to_cpu(lsd->lsd_ost_index);
2678
2679                 if (!(lsd->lsd_feature_compat &
2680                       cpu_to_le32(OBD_COMPAT_OST))) {
2681                         /* this will only happen on the first connect */
2682                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2683                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2684                         /* sync is not needed here as filter_client_add will
2685                          * set exp_need_sync flag */
2686                         filter_update_server_data(exp->exp_obd);
2687                 } else if (index != data->ocd_index) {
2688                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2689                                            " %u doesn't match actual OST index"
2690                                            " %u in last_rcvd file, bad "
2691                                            "configuration?\n",
2692                                            obd_export_nid2str(exp), index,
2693                                            data->ocd_index);
2694                         RETURN(-EBADF);
2695                 }
2696                 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2697                  * FIXME: We don't strictly need the COMPAT flag for that,
2698                  * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2699                  * FIXME: We needed it for the index, as index 0 is valid. */
2700         }
2701
2702         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2703                 data->ocd_brw_size = 65536;
2704         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2705                 data->ocd_brw_size = min(data->ocd_brw_size,
2706                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2707                 if (data->ocd_brw_size == 0) {
2708                         CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2709                                " ocd_version: %x ocd_grant: %d ocd_index: %u "
2710                                "ocd_brw_size is unexpectedly zero, "
2711                                "network data corruption?"
2712                                "Refusing connection of this client\n",
2713                                 exp->exp_obd->obd_name,
2714                                 exp->exp_client_uuid.uuid,
2715                                 exp, data->ocd_connect_flags, data->ocd_version,
2716                                 data->ocd_grant, data->ocd_index);
2717                         RETURN(-EPROTO);
2718                 }
2719         }
2720
2721         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2722                 __u32 cksum_types = data->ocd_cksum_types;
2723
2724                 /* The client set in ocd_cksum_types the checksum types it
2725                  * supports. We have to mask off the algorithms that we don't
2726                  * support */
2727                 if (cksum_types & OBD_CKSUM_ALL)
2728                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2729                 else
2730                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2731
2732                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2733                                    "%x\n", exp->exp_obd->obd_name,
2734                                    obd_export_nid2str(exp), cksum_types,
2735                                    data->ocd_cksum_types);
2736         } else {
2737                 /* This client does not support OBD_CONNECT_CKSUM
2738                  * fall back to CRC32 */
2739                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2740                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2741                                    exp->exp_obd->obd_name,
2742                                    obd_export_nid2str(exp));
2743         }
2744
2745         RETURN(0);
2746 }
2747
2748 static int filter_reconnect(const struct lu_env *env,
2749                             struct obd_export *exp, struct obd_device *obd,
2750                             struct obd_uuid *cluuid,
2751                             struct obd_connect_data *data,
2752                             void *localdata)
2753 {
2754         int rc;
2755         ENTRY;
2756
2757         if (exp == NULL || obd == NULL || cluuid == NULL)
2758                 RETURN(-EINVAL);
2759
2760         rc = filter_connect_internal(exp, data, 1);
2761         if (rc == 0)
2762                 filter_export_stats_init(obd, exp, localdata);
2763
2764         RETURN(rc);
2765 }
2766
2767 /* nearly identical to mds_connect */
2768 static int filter_connect(const struct lu_env *env,
2769                           struct obd_export **exp, struct obd_device *obd,
2770                           struct obd_uuid *cluuid,
2771                           struct obd_connect_data *data, void *localdata)
2772 {
2773         struct lvfs_run_ctxt saved;
2774         struct lustre_handle conn = { 0 };
2775         struct obd_export *lexp;
2776         __u32 group;
2777         int rc;
2778         ENTRY;
2779
2780         if (exp == NULL || obd == NULL || cluuid == NULL)
2781                 RETURN(-EINVAL);
2782
2783         rc = class_connect(&conn, obd, cluuid);
2784         if (rc)
2785                 RETURN(rc);
2786         lexp = class_conn2export(&conn);
2787         LASSERT(lexp != NULL);
2788
2789         rc = filter_connect_internal(lexp, data, 0);
2790         if (rc)
2791                 GOTO(cleanup, rc);
2792
2793         filter_export_stats_init(obd, lexp, localdata);
2794         if (obd->obd_replayable) {
2795                 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2796                 LASSERT(lcd);
2797                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2798                 rc = filter_client_add(obd, lexp, -1);
2799                 if (rc)
2800                         GOTO(cleanup, rc);
2801         }
2802
2803         group = data->ocd_group;
2804
2805         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2806               obd->obd_name, lexp->exp_handle.h_cookie, group);
2807
2808         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2809         rc = filter_read_groups(obd, group, 1);
2810         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2811         if (rc != 0) {
2812                 CERROR("can't read group %u\n", group);
2813                 GOTO(cleanup, rc);
2814         }
2815
2816         GOTO(cleanup, rc);
2817
2818 cleanup:
2819         if (rc) {
2820                 class_disconnect(lexp);
2821                 *exp = NULL;
2822         } else {
2823                 *exp = lexp;
2824         }
2825
2826         RETURN(rc);
2827 }
2828
2829 /* Do extra sanity checks for grant accounting.  We do this at connect,
2830  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2831  * always get rid of it or turn it off when we know accounting is good. */
2832 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2833 {
2834         struct filter_export_data *fed;
2835         struct obd_export *exp;
2836         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2837         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2838         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2839
2840         if (cfs_list_empty(&obd->obd_exports))
2841                 return;
2842
2843         /* We don't want to do this for large machines that do lots of
2844            mounts or unmounts.  It burns... */
2845         if (obd->obd_num_exports > 100)
2846                 return;
2847
2848         cfs_spin_lock(&obd->obd_osfs_lock);
2849         cfs_spin_lock(&obd->obd_dev_lock);
2850         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2851                 int error = 0;
2852                 fed = &exp->exp_filter_data;
2853                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2854                     fed->fed_dirty < 0)
2855                         error = 1;
2856                 if (maxsize > 0) { /* we may not have done a statfs yet */
2857                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2858                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2859                                  exp->exp_client_uuid.uuid, exp,
2860                                  fed->fed_grant, fed->fed_pending, maxsize);
2861                         LASSERTF(fed->fed_dirty <= maxsize,
2862                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2863                                  exp->exp_client_uuid.uuid, exp,
2864                                  fed->fed_dirty, maxsize);
2865                 }
2866                 if (error)
2867                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2868                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2869                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2870                 else
2871                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2872                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2873                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2874                 tot_granted += fed->fed_grant + fed->fed_pending;
2875                 tot_pending += fed->fed_pending;
2876                 tot_dirty += fed->fed_dirty;
2877         }
2878         fo_tot_granted = obd->u.filter.fo_tot_granted;
2879         fo_tot_pending = obd->u.filter.fo_tot_pending;
2880         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2881         cfs_spin_unlock(&obd->obd_dev_lock);
2882         cfs_spin_unlock(&obd->obd_osfs_lock);
2883
2884         /* Do these assertions outside the spinlocks so we don't kill system */
2885         if (tot_granted != fo_tot_granted)
2886                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2887                        func, tot_granted, fo_tot_granted);
2888         if (tot_pending != fo_tot_pending)
2889                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2890                        func, tot_pending, fo_tot_pending);
2891         if (tot_dirty != fo_tot_dirty)
2892                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2893                        func, tot_dirty, fo_tot_dirty);
2894         if (tot_pending > tot_granted)
2895                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2896                        func, tot_pending, tot_granted);
2897         if (tot_granted > maxsize)
2898                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2899                        func, tot_granted, maxsize);
2900         if (tot_dirty > maxsize)
2901                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2902                        func, tot_dirty, maxsize);
2903 }
2904
2905 /* Remove this client from the grant accounting totals.  We also remove
2906  * the export from the obd device under the osfs and dev locks to ensure
2907  * that the filter_grant_sanity_check() calculations are always valid.
2908  * The client should do something similar when it invalidates its import. */
2909 static void filter_grant_discard(struct obd_export *exp)
2910 {
2911         struct obd_device *obd = exp->exp_obd;
2912         struct filter_obd *filter = &obd->u.filter;
2913         struct filter_export_data *fed = &exp->exp_filter_data;
2914
2915         cfs_spin_lock(&obd->obd_osfs_lock);
2916         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2917                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2918                  obd->obd_name, filter->fo_tot_granted,
2919                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2920         filter->fo_tot_granted -= fed->fed_grant;
2921         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2922                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2923                  obd->obd_name, filter->fo_tot_pending,
2924                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2925         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2926         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2927                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2928                  obd->obd_name, filter->fo_tot_dirty,
2929                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2930         filter->fo_tot_dirty -= fed->fed_dirty;
2931         fed->fed_dirty = 0;
2932         fed->fed_grant = 0;
2933
2934         cfs_spin_unlock(&obd->obd_osfs_lock);
2935 }
2936
2937 static int filter_destroy_export(struct obd_export *exp)
2938 {
2939         struct filter_export_data *fed = &exp->exp_filter_data;
2940         ENTRY;
2941
2942         if (fed->fed_pending)
2943                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2944                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2945                        exp, fed->fed_pending);
2946
2947         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2948
2949         target_destroy_export(exp);
2950         ldlm_destroy_export(exp);
2951         lut_client_free(exp);
2952
2953         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2954                 RETURN(0);
2955
2956         if (!exp->exp_obd->obd_replayable)
2957                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2958
2959         filter_grant_discard(exp);
2960         filter_fmd_cleanup(exp);
2961
2962         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2963                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2964                 if (filter->fo_tot_granted_clients > 0)
2965                         filter->fo_tot_granted_clients --;
2966         }
2967
2968         if (!(exp->exp_flags & OBD_OPT_FORCE))
2969                 filter_grant_sanity_check(exp->exp_obd, __func__);
2970
2971         RETURN(0);
2972 }
2973
2974 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2975 {
2976         struct obd_llog_group *olg_min, *olg;
2977         struct filter_obd *filter;
2978         int worked = -1, group;
2979         struct llog_ctxt *ctxt;
2980         ENTRY;
2981
2982         filter = &obd->u.filter;
2983
2984         /* we can't sync log holding spinlock. also, we do not want to get
2985          * into livelock. so we do following: loop over MDS's exports in
2986          * group order and skip already synced llogs -bzzz */
2987         do {
2988                 /* look for group with min. number, but > worked */
2989                 olg_min = NULL;
2990                 group = 1 << 30;
2991                 cfs_spin_lock(&filter->fo_llog_list_lock);
2992                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2993                         if (olg->olg_seq <= worked) {
2994                                 /* this group is already synced */
2995                                 continue;
2996                         }
2997                         if (group < olg->olg_seq) {
2998                                 /* we have group with smaller number to sync */
2999                                 continue;
3000                         }
3001                         /* store current minimal group */
3002                         olg_min = olg;
3003                         group = olg->olg_seq;
3004                 }
3005                 cfs_spin_unlock(&filter->fo_llog_list_lock);
3006
3007                 if (olg_min == NULL)
3008                         break;
3009
3010                 worked = olg_min->olg_seq;
3011                 if (olg_min->olg_exp &&
3012                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3013                         int err;
3014                         ctxt = llog_group_get_ctxt(olg_min,
3015                                                    LLOG_MDS_OST_REPL_CTXT);
3016                         if (ctxt) {
3017                                 err = llog_sync(ctxt, olg_min->olg_exp);
3018                                 llog_ctxt_put(ctxt);
3019                                 if (err) {
3020                                         CERROR("error flushing logs to MDS: "
3021                                                "rc %d\n", err);
3022                                 }
3023                         }
3024                 }
3025         } while (olg_min != NULL);
3026 }
3027
3028 /* Also incredibly similar to mds_disconnect */
3029 static int filter_disconnect(struct obd_export *exp)
3030 {
3031         struct obd_device *obd = exp->exp_obd;
3032         int rc;
3033         ENTRY;
3034
3035         LASSERT(exp);
3036         class_export_get(exp);
3037
3038         if (!(exp->exp_flags & OBD_OPT_FORCE))
3039                 filter_grant_sanity_check(obd, __func__);
3040         filter_grant_discard(exp);
3041
3042         /* Flush any remaining cancel messages out to the target */
3043         filter_sync_llogs(obd, exp);
3044
3045         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3046
3047         rc = server_disconnect_export(exp);
3048
3049         /* Do not erase record for recoverable client. */
3050         if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3051                 filter_client_del(exp);
3052         else
3053                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3054
3055         class_export_put(exp);
3056         RETURN(rc);
3057 }
3058
3059 /* reverse import is changed, sync all cancels */
3060 static void filter_revimp_update(struct obd_export *exp)
3061 {
3062         ENTRY;
3063
3064         LASSERT(exp);
3065         class_export_get(exp);
3066
3067         /* flush any remaining cancel messages out to the target */
3068         filter_sync_llogs(exp->exp_obd, exp);
3069         class_export_put(exp);
3070         EXIT;
3071 }
3072
3073 static int filter_ping(struct obd_export *exp)
3074 {
3075         filter_fmd_expire(exp);
3076         return 0;
3077 }
3078
3079 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3080                                   const char *what, int quiet)
3081 {
3082         struct dentry *dchild = NULL;
3083
3084         dchild = filter_fid2dentry(obd, NULL,  ostid->oi_seq, ostid->oi_id);
3085
3086         if (IS_ERR(dchild)) {
3087                 CERROR("%s error looking up object: "POSTID"\n",
3088                        what, ostid->oi_id, ostid->oi_seq);
3089                 RETURN(dchild);
3090         }
3091
3092         if (dchild->d_inode == NULL) {
3093                 if (!quiet)
3094                         CERROR("%s: %s on non-existent object: "POSTID" \n",
3095                                obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3096                 f_dput(dchild);
3097                 RETURN(ERR_PTR(-ENOENT));
3098         }
3099
3100         return dchild;
3101 }
3102
3103 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3104 {
3105         struct dentry *dentry = NULL;
3106         struct obd_device *obd;
3107         int rc = 0;
3108         ENTRY;
3109
3110         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3111                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3112         if (rc)
3113                 RETURN(rc);
3114
3115         obd = class_exp2obd(exp);
3116         if (obd == NULL) {
3117                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3118                 RETURN(-EINVAL);
3119         }
3120
3121         dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3122         if (IS_ERR(dentry))
3123                 RETURN(PTR_ERR(dentry));
3124
3125         /* Limit the valid bits in the return data to what we actually use */
3126         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3127         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3128
3129         f_dput(dentry);
3130         RETURN(rc);
3131 }
3132
3133 /* this should be enabled/disabled in condition to enabled/disabled large
3134  * inodes (fast EAs) in backing store FS. */
3135 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3136                         void *handle, struct obdo *oa)
3137 {
3138         struct obd_device *obd = exp->exp_obd;
3139         int rc = 0;
3140         ENTRY;
3141
3142         if (oa->o_valid & OBD_MD_FLFID) {
3143                 struct filter_fid ff;
3144
3145                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3146                         oa->o_seq = 0;
3147                 /* packing fid and converting it to LE for storing into EA.
3148                  * Here ->o_stripe_idx should be filled by LOV and rest of
3149                  * fields - by client. */
3150                 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3151                 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3152                 /* XXX: we are ignoring o_parent_ver here, since this should
3153                  *      be the same for all objects in this fileset. */
3154                 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3155                 ff.ff_objid = cpu_to_le64(oa->o_id);
3156                 ff.ff_seq = cpu_to_le64(oa->o_seq);
3157
3158                 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3159                        LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3160                        oa->o_seq);
3161
3162                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3163                 if (rc)
3164                         CERROR("store fid in object failed! rc: %d\n", rc);
3165         } else {
3166                 CDEBUG(D_HA, "OSS object without fid info!\n");
3167         }
3168
3169         RETURN(rc);
3170 }
3171
3172 /* this is called from filter_truncate() until we have filter_punch() */
3173 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3174                             struct obdo *oa, struct obd_trans_info *oti)
3175 {
3176         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3177         struct llog_cookie *fcc = NULL;
3178         struct filter_obd *filter;
3179         int rc, err, sync = 0;
3180         loff_t old_size = 0;
3181         unsigned int ia_valid;
3182         struct inode *inode;
3183         struct page *page = NULL;
3184         struct iattr iattr;
3185         void *handle;
3186         ENTRY;
3187
3188         LASSERT(dentry != NULL);
3189         LASSERT(!IS_ERR(dentry));
3190
3191         inode = dentry->d_inode;
3192         LASSERT(inode != NULL);
3193
3194         filter = &exp->exp_obd->u.filter;
3195         iattr_from_obdo(&iattr, oa, oa->o_valid);
3196         ia_valid = iattr.ia_valid;
3197
3198         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3199                 OBD_ALLOC(fcc, sizeof(*fcc));