Whamcloud - gitweb
Revert "b=21485 Keep in-memory client data until export destroy"
[fs/lustre-release.git] / lustre / obdfilter / filter.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/obdfilter/filter.c
37  *
38  * Author: Peter Braam <braam@clusterfs.com>
39  * Author: Andreas Dilger <adilger@clusterfs.com>
40  */
41
42 /*
43  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44  *            (which need to get journal_lock, may block if journal full).
45  *
46  * Invariant: Call filter_start_transno() before any journal ops to avoid the
47  *            same deadlock problem.  We can (and want) to get rid of the
48  *            transno sem in favour of the dir/inode i_mutex to avoid single
49  *            threaded operation on the OST.
50  */
51
52 #define DEBUG_SUBSYSTEM S_FILTER
53
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
56 #endif
57 #include <linux/module.h>
58 #include <linux/fs.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
65
66 #include <obd_cksum.h>
67 #include <obd_class.h>
68 #include <obd_lov.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
79
80 #include "filter_internal.h"
81
82 static struct lvfs_callback_ops filter_lvfs_ops;
83 cfs_mem_cache_t *ll_fmd_cachep;
84
85 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
86                              void *cb_data, int error)
87 {
88         struct obd_export *exp = cb_data;
89         LASSERT(exp->exp_obd == obd);
90         obd_transno_commit_cb(obd, transno, exp, error);
91         class_export_cb_put(exp);
92 }
93
94 int filter_version_get_check(struct obd_export *exp,
95                              struct obd_trans_info *oti, struct inode *inode)
96 {
97         __u64 curr_version;
98
99         if (inode == NULL || oti == NULL)
100                 RETURN(0);
101
102         curr_version = fsfilt_get_version(exp->exp_obd, inode);
103         if ((__s64)curr_version == -EOPNOTSUPP)
104                 RETURN(0);
105         /* VBR: version is checked always because costs nothing */
106         if (oti->oti_pre_version != 0 &&
107             oti->oti_pre_version != curr_version) {
108                 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
109                        oti->oti_pre_version, curr_version);
110                 cfs_spin_lock(&exp->exp_lock);
111                 exp->exp_vbr_failed = 1;
112                 cfs_spin_unlock(&exp->exp_lock);
113                 RETURN (-EOVERFLOW);
114         }
115         oti->oti_pre_version = curr_version;
116         RETURN(0);
117 }
118
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121                           struct obd_trans_info *oti, int rc, int force_sync)
122 {
123         struct filter_obd *filter = &exp->exp_obd->u.filter;
124         struct filter_export_data *fed = &exp->exp_filter_data;
125         struct lsd_client_data *lcd;
126         __u64 last_rcvd;
127         loff_t off;
128         int err, log_pri = D_RPCTRACE;
129
130         /* Propagate error code. */
131         if (rc)
132                 RETURN(rc);
133
134         if (!exp->exp_obd->obd_replayable || oti == NULL)
135                 RETURN(rc);
136
137         cfs_mutex_down(&fed->fed_lcd_lock);
138         lcd = fed->fed_lcd;
139         /* if the export has already been disconnected, we have no last_rcvd slot,
140          * update server data with latest transno then */
141         if (lcd == NULL) {
142                 cfs_mutex_up(&fed->fed_lcd_lock);
143                 CWARN("commit transaction for disconnected client %s: rc %d\n",
144                       exp->exp_client_uuid.uuid, rc);
145                 err = filter_update_server_data(exp->exp_obd,
146                                                 filter->fo_rcvd_filp,
147                                                 filter->fo_fsd);
148                 RETURN(err);
149         }
150
151         /* we don't allocate new transnos for replayed requests */
152         cfs_spin_lock(&filter->fo_translock);
153         if (oti->oti_transno == 0) {
154                 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
155                 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
156         } else {
157                 last_rcvd = oti->oti_transno;
158                 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
159                         filter->fo_fsd->lsd_last_transno =
160                                 cpu_to_le64(last_rcvd);
161         }
162         oti->oti_transno = last_rcvd;
163
164         LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
165         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
166         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
167         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
168         cfs_spin_unlock(&filter->fo_translock);
169
170         if (inode)
171                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
172
173         off = fed->fed_lr_off;
174         if (off <= 0) {
175                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
176                        fed->fed_lr_idx, fed->fed_lr_off);
177                 err = -EINVAL;
178         } else {
179                 class_export_cb_get(exp); /* released when the cb is called */
180                 if (!force_sync)
181                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
182                                                            last_rcvd,
183                                                            oti->oti_handle,
184                                                            filter_commit_cb,
185                                                            exp);
186
187                 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
188                                           lcd, sizeof(*lcd), &off,
189                                           force_sync | exp->exp_need_sync);
190                 if (force_sync)
191                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
192         }
193         if (err) {
194                 log_pri = D_ERROR;
195                 if (rc == 0)
196                         rc = err;
197         }
198
199         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
200                last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
201         cfs_mutex_up(&fed->fed_lcd_lock);
202         RETURN(rc);
203 }
204
205 void f_dput(struct dentry *dentry)
206 {
207         /* Can't go inside filter_ddelete because it can block */
208         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
209                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
210         LASSERT(atomic_read(&dentry->d_count) > 0);
211
212         dput(dentry);
213 }
214
215 static void init_brw_stats(struct brw_stats *brw_stats)
216 {
217         int i;
218         for (i = 0; i < BRW_LAST; i++)
219                 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
220 }
221
222 static int lprocfs_init_rw_stats(struct obd_device *obd,
223                                  struct lprocfs_stats **stats)
224 {
225         int num_stats;
226
227         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
228                                                         LPROC_FILTER_LAST - 1;
229         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
230         if (*stats == NULL)
231                 return -ENOMEM;
232
233         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
234         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
235                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
236         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
237                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
238
239         return(0);
240 }
241
242 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
243    plus the procfs overhead :( */
244 static int filter_export_stats_init(struct obd_device *obd,
245                                     struct obd_export *exp,
246                                     void *client_nid)
247 {
248         int rc, newnid = 0;
249         ENTRY;
250
251         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
252                 /* Self-export gets no proc entry */
253                 RETURN(0);
254
255         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
256         if (rc) {
257                 /* Mask error for already created
258                  * /proc entries */
259                 if (rc == -EALREADY)
260                         rc = 0;
261                 RETURN(rc);
262         }
263
264         if (newnid) {
265                 struct nid_stat *tmp = exp->exp_nid_stats;
266                 LASSERT(tmp != NULL);
267
268                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
269                 if (tmp->nid_brw_stats == NULL)
270                         GOTO(clean, rc = -ENOMEM);
271
272                 init_brw_stats(tmp->nid_brw_stats);
273                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
274                                         0644, &filter_per_nid_stats_fops,
275                                         exp->exp_nid_stats);
276                 if (rc)
277                         CWARN("Error adding the brw_stats file\n");
278
279                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
280                 if (rc)
281                         GOTO(clean, rc);
282
283                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
284                                             tmp->nid_stats);
285                 if (rc)
286                         GOTO(clean, rc);
287                 /* Always add in ldlm_stats */
288                 tmp->nid_ldlm_stats = 
289                         lprocfs_alloc_stats(LDLM_LAST_OPC - LDLM_FIRST_OPC,
290                                             LPROCFS_STATS_FLAG_NOPERCPU);
291                 if (tmp->nid_ldlm_stats == NULL)
292                         GOTO(clean, rc = -ENOMEM);
293
294                 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
295                 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
296                                             tmp->nid_ldlm_stats);
297                 if (rc)
298                         GOTO(clean, rc);
299         }
300
301         RETURN(0);
302  clean:
303         lprocfs_exp_cleanup(exp);
304         return rc;
305 }
306
307 /* Add client data to the FILTER.  We use a bitmap to locate a free space
308  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
309  * Otherwise, we have just read the data from the last_rcvd file and
310  * we know its offset. */
311 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
312                              int cl_idx)
313 {
314         struct filter_obd *filter = &obd->u.filter;
315         struct filter_export_data *fed = &exp->exp_filter_data;
316         unsigned long *bitmap = filter->fo_last_rcvd_slots;
317         int new_client = (cl_idx == -1);
318
319         ENTRY;
320
321         LASSERT(bitmap != NULL);
322         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
323
324         /* Self-export */
325         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
326                 RETURN(0);
327
328         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
329          * there's no need for extra complication here
330          */
331         if (new_client) {
332                 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
333         repeat:
334                 if (cl_idx >= LR_MAX_CLIENTS) {
335                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
336                                cl_idx);
337                         RETURN(-EOVERFLOW);
338                 }
339                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
340                         cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
341                                                         cl_idx);
342                         goto repeat;
343                 }
344         } else {
345                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
346                         CERROR("FILTER client %d: bit already set in bitmap!\n",
347                                cl_idx);
348                         LBUG();
349                 }
350         }
351
352         fed->fed_lr_idx = cl_idx;
353         fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
354                 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
355         cfs_init_mutex(&fed->fed_lcd_lock);
356         LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
357
358         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
359                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
360
361         if (new_client) {
362                 struct lvfs_run_ctxt saved;
363                 loff_t off = fed->fed_lr_off;
364                 int rc;
365                 void *handle;
366
367                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
368                        fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
369
370                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
371                 /* Transaction needed to fix bug 1403 */
372                 handle = fsfilt_start(obd,
373                                       filter->fo_rcvd_filp->f_dentry->d_inode,
374                                       FSFILT_OP_SETATTR, NULL);
375                 if (IS_ERR(handle)) {
376                         rc = PTR_ERR(handle);
377                         CERROR("unable to start transaction: rc %d\n", rc);
378                 } else {
379                         fed->fed_lcd->lcd_last_epoch =
380                                               filter->fo_fsd->lsd_start_epoch;
381                         exp->exp_last_request_time = cfs_time_current_sec();
382                         rc = fsfilt_add_journal_cb(obd, 0, handle,
383                                                    target_client_add_cb,
384                                                    class_export_cb_get(exp));
385                         if (rc == 0) {
386                                 cfs_spin_lock(&exp->exp_lock);
387                                 exp->exp_need_sync = 1;
388                                 cfs_spin_unlock(&exp->exp_lock);
389                         }
390                         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
391                                                  fed->fed_lcd,
392                                                  sizeof(*fed->fed_lcd),
393                                                  &off, rc /* sync if no cb */);
394                         fsfilt_commit(obd,
395                                       filter->fo_rcvd_filp->f_dentry->d_inode,
396                                       handle, 0);
397                 }
398                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
399
400                 if (rc) {
401                         CERROR("error writing %s client idx %u: rc %d\n",
402                                LAST_RCVD, fed->fed_lr_idx, rc);
403                         RETURN(rc);
404                 }
405         }
406         RETURN(0);
407 }
408
409 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
410
411 static int filter_client_free(struct obd_export *exp)
412 {
413         struct filter_export_data *fed = &exp->exp_filter_data;
414         struct filter_obd *filter = &exp->exp_obd->u.filter;
415         struct obd_device *obd = exp->exp_obd;
416         struct lvfs_run_ctxt saved;
417         struct lsd_client_data *lcd = fed->fed_lcd;
418         int rc;
419         loff_t off;
420         ENTRY;
421
422         if (lcd == NULL)
423                 RETURN(0);
424
425         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
426         if (strcmp(lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
427                 GOTO(free, 0);
428
429         LASSERT(filter->fo_last_rcvd_slots != NULL);
430
431         off = fed->fed_lr_off;
432
433         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
434                fed->fed_lr_idx, fed->fed_lr_off, lcd->lcd_uuid);
435
436         /* Don't clear fed_lr_idx here as it is likely also unset.  At worst
437          * we leak a client slot that will be cleaned on the next recovery. */
438         if (off <= 0) {
439                 CERROR("%s: client idx %d has med_off %lld\n",
440                        obd->obd_name, fed->fed_lr_idx, off);
441                 GOTO(free, rc = -EINVAL);
442         }
443
444         /* Clear the bit _after_ zeroing out the client so we don't
445            race with filter_client_add and zero out new clients.*/
446         if (!cfs_test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
447                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
448                        fed->fed_lr_idx);
449                 LBUG();
450         }
451
452         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
453         /* Make sure the server's last_transno is up to date.
454          * This should be done before zeroing client slot so last_transno will
455          * be in server data or in client data in case of failure */
456         filter_update_server_data(obd, filter->fo_rcvd_filp, filter->fo_fsd);
457
458         cfs_mutex_down(&fed->fed_lcd_lock);
459         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
460                                  sizeof(zero_lcd), &off, 0);
461         fed->fed_lcd = NULL;
462         cfs_mutex_up(&fed->fed_lcd_lock);
463         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
464
465         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
466                "zero out client %s at idx %u/%llu in %s, rc %d\n",
467                lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
468                LAST_RCVD, rc);
469
470         if (!cfs_test_and_clear_bit(fed->fed_lr_idx,
471                                     filter->fo_last_rcvd_slots)) {
472                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
473                        fed->fed_lr_idx);
474                 LBUG();
475         }
476         OBD_FREE_PTR(lcd);
477         RETURN(0);
478 free:
479         cfs_mutex_down(&fed->fed_lcd_lock);
480         fed->fed_lcd = NULL;
481         cfs_mutex_up(&fed->fed_lcd_lock);
482         OBD_FREE_PTR(lcd);
483
484         return 0;
485 }
486
487 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
488 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
489                                          struct filter_mod_data *fmd)
490 {
491         LASSERT_SPIN_LOCKED(&fed->fed_lock);
492         if (--fmd->fmd_refcount == 0) {
493                 /* XXX when we have persistent reservations and the handle
494                  * is stored herein we need to drop it here. */
495                 fed->fed_mod_count--;
496                 cfs_list_del(&fmd->fmd_list);
497                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
498         }
499 }
500
501 /* drop fmd reference, free it if last ref */
502 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
503 {
504         struct filter_export_data *fed;
505
506         if (fmd == NULL)
507                 return;
508
509         fed = &exp->exp_filter_data;
510         cfs_spin_lock(&fed->fed_lock);
511         filter_fmd_put_nolock(fed, fmd); /* caller reference */
512         cfs_spin_unlock(&fed->fed_lock);
513 }
514
515 /* expire entries from the end of the list if there are too many
516  * or they are too old */
517 static void filter_fmd_expire_nolock(struct filter_obd *filter,
518                                      struct filter_export_data *fed,
519                                      struct filter_mod_data *keep)
520 {
521         struct filter_mod_data *fmd, *tmp;
522
523         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
524                 if (fmd == keep)
525                         break;
526
527                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
528                     fed->fed_mod_count < filter->fo_fmd_max_num)
529                         break;
530
531                 cfs_list_del_init(&fmd->fmd_list);
532                 filter_fmd_put_nolock(fed, fmd); /* list reference */
533         }
534 }
535
536 void filter_fmd_expire(struct obd_export *exp)
537 {
538         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
539         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
540                                  &exp->exp_filter_data, NULL);
541         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
542 }
543
544 /* find specified objid, group in export fmd list.
545  * caller must hold fed_lock and take fmd reference itself */
546 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
547                                                 struct filter_export_data *fed,
548                                                 obd_id objid, obd_gr group)
549 {
550         struct filter_mod_data *found = NULL, *fmd;
551
552         LASSERT_SPIN_LOCKED(&fed->fed_lock);
553
554         cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
555                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
556                         found = fmd;
557                         cfs_list_del(&fmd->fmd_list);
558                         cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
559                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
560                         break;
561                 }
562         }
563
564         filter_fmd_expire_nolock(filter, fed, found);
565
566         return found;
567 }
568
569 /* Find fmd based on objid and group, or return NULL if not found. */
570 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
571                                         obd_id objid, obd_gr group)
572 {
573         struct filter_mod_data *fmd;
574
575         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
576         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
577                                      &exp->exp_filter_data, objid, group);
578         if (fmd)
579                 fmd->fmd_refcount++;    /* caller reference */
580         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
581
582         return fmd;
583 }
584
585 /* Find fmd based on objid and group, or create a new one if none is found.
586  * It is possible for this function to return NULL under memory pressure,
587  * or if objid = 0 is passed (which will only cause old entries to expire).
588  * Currently this is not fatal because any fmd state is transient and
589  * may also be freed when it gets sufficiently old. */
590 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
591                                        obd_id objid, obd_gr group)
592 {
593         struct filter_export_data *fed = &exp->exp_filter_data;
594         struct filter_mod_data *found = NULL, *fmd_new = NULL;
595
596         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
597
598         cfs_spin_lock(&fed->fed_lock);
599         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
600         if (fmd_new) {
601                 if (found == NULL) {
602                         cfs_list_add_tail(&fmd_new->fmd_list,
603                                           &fed->fed_mod_list);
604                         fmd_new->fmd_id = objid;
605                         fmd_new->fmd_gr = group;
606                         fmd_new->fmd_refcount++;   /* list reference */
607                         found = fmd_new;
608                         fed->fed_mod_count++;
609                 } else {
610                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
611                 }
612         }
613         if (found) {
614                 found->fmd_refcount++;          /* caller reference */
615                 found->fmd_expire = jiffies +
616                         exp->exp_obd->u.filter.fo_fmd_max_age;
617         }
618
619         cfs_spin_unlock(&fed->fed_lock);
620
621         return found;
622 }
623
624 #ifdef DO_FMD_DROP
625 /* drop fmd list reference so it will disappear when last reference is put.
626  * This isn't so critical because it would in fact only affect the one client
627  * that is doing the unlink and at worst we have an stale entry referencing
628  * an object that should never be used again. */
629 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
630 {
631         struct filter_mod_data *found = NULL;
632
633         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
634         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
635         if (found) {
636                 cfs_list_del_init(&found->fmd_list);
637                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
638         }
639         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
640 }
641 #else
642 #define filter_fmd_drop(exp, objid, group)
643 #endif
644
645 /* remove all entries from fmd list */
646 static void filter_fmd_cleanup(struct obd_export *exp)
647 {
648         struct filter_export_data *fed = &exp->exp_filter_data;
649         struct filter_mod_data *fmd = NULL, *tmp;
650
651         cfs_spin_lock(&fed->fed_lock);
652         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
653                 cfs_list_del_init(&fmd->fmd_list);
654                 filter_fmd_put_nolock(fed, fmd);
655         }
656         cfs_spin_unlock(&fed->fed_lock);
657 }
658
659 static int filter_init_export(struct obd_export *exp)
660 {
661         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
662         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
663
664         cfs_spin_lock(&exp->exp_lock);
665         exp->exp_connecting = 1;
666         cfs_spin_unlock(&exp->exp_lock);
667
668         return ldlm_init_export(exp);
669 }
670
671 static int filter_free_server_data(struct filter_obd *filter)
672 {
673         OBD_FREE_PTR(filter->fo_fsd);
674         filter->fo_fsd = NULL;
675         OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
676         filter->fo_last_rcvd_slots = NULL;
677         return 0;
678 }
679
680 /* assumes caller is already in kernel ctxt */
681 int filter_update_server_data(struct obd_device *obd, struct file *filp,
682                               struct lr_server_data *fsd)
683 {
684         loff_t off = 0;
685         int rc;
686         ENTRY;
687
688         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
689         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
690                le64_to_cpu(fsd->lsd_last_transno));
691         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
692                le64_to_cpu(fsd->lsd_mount_count));
693
694         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, 0);
695         if (rc)
696                 CERROR("error writing lr_server_data: rc = %d\n", rc);
697
698         RETURN(rc);
699 }
700
701 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
702                              int force_sync)
703 {
704         struct filter_obd *filter = &obd->u.filter;
705         __u64 tmp;
706         loff_t off = 0;
707         int rc;
708         ENTRY;
709
710         if (filter->fo_last_objid_files[group] == NULL) {
711                 CERROR("Object group "LPU64" not fully setup; not updating "
712                        "last_objid\n", group);
713                 RETURN(-EINVAL);
714         }
715
716         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
717                obd->obd_name, group, filter->fo_last_objids[group]);
718
719         tmp = cpu_to_le64(filter->fo_last_objids[group]);
720         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
721                                  &tmp, sizeof(tmp), &off, force_sync);
722         if (rc)
723                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
724                        group, rc);
725         RETURN(rc);
726 }
727 extern int ost_handle(struct ptlrpc_request *req);
728 /* assumes caller has already in kernel ctxt */
729 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
730 {
731         struct filter_obd *filter = &obd->u.filter;
732         struct lr_server_data *fsd;
733         struct lsd_client_data *lcd = NULL;
734         struct inode *inode = filp->f_dentry->d_inode;
735         unsigned long last_rcvd_size = i_size_read(inode);
736         __u64 mount_count;
737         __u32 start_epoch;
738         int cl_idx;
739         loff_t off = 0;
740         int rc;
741
742         /* ensure padding in the struct is the correct size */
743         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
744                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
745         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
746                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
747
748         OBD_ALLOC(fsd, sizeof(*fsd));
749         if (!fsd)
750                 RETURN(-ENOMEM);
751         filter->fo_fsd = fsd;
752
753         OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
754         if (filter->fo_last_rcvd_slots == NULL) {
755                 OBD_FREE(fsd, sizeof(*fsd));
756                 RETURN(-ENOMEM);
757         }
758
759         if (last_rcvd_size == 0) {
760                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
761
762                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
763                 fsd->lsd_last_transno = 0;
764                 mount_count = fsd->lsd_mount_count = 0;
765                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
766                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
767                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
768                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
769                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
770                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
771                  * MDS first connects and assigns the OST index number. */
772                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
773                                                         OBD_INCOMPAT_OST);
774         } else {
775                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
776                 if (rc) {
777                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
778                                LAST_RCVD, rc);
779                         GOTO(err_fsd, rc);
780                 }
781                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
782                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
783                                            "using the wrong disk %s. Were the "
784                                            "/dev/ assignments rearranged?\n",
785                                            obd->obd_uuid.uuid, fsd->lsd_uuid);
786                         GOTO(err_fsd, rc = -EINVAL);
787                 }
788                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
789                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
790                 /* COMPAT_146 */
791                 /* Assume old last_rcvd format unless I_C_LR is set */
792                 if (!(fsd->lsd_feature_incompat &
793                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
794                         fsd->lsd_last_transno = fsd->lsd_compat14;
795                 /* end COMPAT_146 */
796                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
797                  * MDS first connects and assigns the OST index number. */
798                 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
799                                                          OBD_INCOMPAT_OST);
800         }
801
802         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
803                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
804                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
805                        ~FILTER_INCOMPAT_SUPP);
806                 GOTO(err_fsd, rc = -EINVAL);
807         }
808         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
809                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
810                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
811                        ~FILTER_ROCOMPAT_SUPP);
812                 /* Do something like remount filesystem read-only */
813                 GOTO(err_fsd, rc = -EINVAL);
814         }
815
816         start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
817
818         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
819                obd->obd_name, start_epoch);
820         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
821                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
822         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
823                obd->obd_name, mount_count + 1);
824         CDEBUG(D_INODE, "%s: server data size: %u\n",
825                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
826         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
827                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
828         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
829                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
830         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
831                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
832         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
833                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
834                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
835                 le16_to_cpu(fsd->lsd_client_size));
836
837         if (!obd->obd_replayable) {
838                 CWARN("%s: recovery support OFF\n", obd->obd_name);
839                 GOTO(out, rc = 0);
840         }
841
842         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
843              off < last_rcvd_size; cl_idx++) {
844                 __u64 last_rcvd;
845                 struct obd_export *exp;
846                 struct filter_export_data *fed;
847
848                 if (!lcd) {
849                         OBD_ALLOC_PTR(lcd);
850                         if (!lcd)
851                                 GOTO(err_client, rc = -ENOMEM);
852                 }
853
854                 /* Don't assume off is incremented properly by
855                  * fsfilt_read_record(), in case sizeof(*lcd)
856                  * isn't the same as fsd->lsd_client_size.  */
857                 off = le32_to_cpu(fsd->lsd_client_start) +
858                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
859                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
860                 if (rc) {
861                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
862                                LAST_RCVD, cl_idx, off, rc);
863                         break; /* read error shouldn't cause startup to fail */
864                 }
865
866                 if (lcd->lcd_uuid[0] == '\0') {
867                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
868                                cl_idx);
869                         continue;
870                 }
871
872                 check_lcd(obd->obd_name, cl_idx, lcd);
873
874                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
875
876                 /* These exports are cleaned up by filter_disconnect(), so they
877                  * need to be set up like real exports as filter_connect() does.
878                  */
879                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
880
881                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
882                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
883                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
884                 if (IS_ERR(exp)) {
885                         if (PTR_ERR(exp) == -EALREADY) {
886                                 /* export already exists, zero out this one */
887                                 CERROR("Zeroing out duplicate export due to "
888                                        "bug 10479.\n");
889                                 lcd->lcd_uuid[0] = '\0';
890                         } else {
891                                 GOTO(err_client, rc = PTR_ERR(exp));
892                         }
893                 } else {
894                         fed = &exp->exp_filter_data;
895                         fed->fed_lcd = lcd;
896                         fed->fed_group = 0; /* will be assigned at connect */
897                         filter_export_stats_init(obd, exp, NULL);
898                         rc = filter_client_add(obd, exp, cl_idx);
899                         /* can't fail for existing client */
900                         LASSERTF(rc == 0, "rc = %d\n", rc);
901
902                         /* VBR: set export last committed */
903                         exp->exp_last_committed = last_rcvd;
904                         cfs_spin_lock(&exp->exp_lock);
905                         exp->exp_connecting = 0;
906                         exp->exp_in_recovery = 0;
907                         cfs_spin_unlock(&exp->exp_lock);
908                         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
909                         obd->obd_max_recoverable_clients++;
910                         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
911                         lcd = NULL;
912                         class_export_put(exp);
913                 }
914
915                 /* Need to check last_rcvd even for duplicated exports. */
916                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
917                        cl_idx, last_rcvd);
918
919                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
920                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
921         }
922
923         if (lcd)
924                 OBD_FREE_PTR(lcd);
925
926         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
927 out:
928         filter->fo_mount_count = mount_count + 1;
929         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
930
931         /* save it, so mount count and last_transno is current */
932         rc = filter_update_server_data(obd, filp, filter->fo_fsd);
933         if (rc)
934                 GOTO(err_client, rc);
935
936         RETURN(0);
937
938 err_client:
939         class_disconnect_exports(obd);
940 err_fsd:
941         filter_free_server_data(filter);
942         RETURN(rc);
943 }
944
945 static int filter_cleanup_groups(struct obd_device *obd)
946 {
947         struct filter_obd *filter = &obd->u.filter;
948         struct file *filp;
949         struct dentry *dentry;
950         int i, j;
951         ENTRY;
952
953         if (filter->fo_dentry_O_groups != NULL) {
954                 for (i = 0; i < filter->fo_group_count; i++) {
955                         dentry = filter->fo_dentry_O_groups[i];
956                         if (dentry != NULL)
957                                 f_dput(dentry);
958                 }
959                 OBD_FREE(filter->fo_dentry_O_groups,
960                          filter->fo_group_count *
961                          sizeof(*filter->fo_dentry_O_groups));
962                 filter->fo_dentry_O_groups = NULL;
963         }
964         if (filter->fo_last_objid_files != NULL) {
965                 for (i = 0; i < filter->fo_group_count; i++) {
966                         filp = filter->fo_last_objid_files[i];
967                         if (filp != NULL)
968                                 filp_close(filp, 0);
969                 }
970                 OBD_FREE(filter->fo_last_objid_files,
971                          filter->fo_group_count *
972                          sizeof(*filter->fo_last_objid_files));
973                 filter->fo_last_objid_files = NULL;
974         }
975         if (filter->fo_dentry_O_sub != NULL) {
976                 for (i = 0; i < filter->fo_group_count; i++) {
977                         for (j = 0; j < filter->fo_subdir_count; j++) {
978                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
979                                 if (dentry != NULL)
980                                         f_dput(dentry);
981                         }
982                 }
983                 OBD_FREE(filter->fo_dentry_O_sub,
984                          filter->fo_group_count *
985                          sizeof(*filter->fo_dentry_O_sub));
986                 filter->fo_dentry_O_sub = NULL;
987         }
988         if (filter->fo_last_objids != NULL) {
989                 OBD_FREE(filter->fo_last_objids,
990                          filter->fo_group_count *
991                          sizeof(*filter->fo_last_objids));
992                 filter->fo_last_objids = NULL;
993         }
994         if (filter->fo_dentry_O != NULL) {
995                 f_dput(filter->fo_dentry_O);
996                 filter->fo_dentry_O = NULL;
997         }
998         RETURN(0);
999 }
1000
1001 static int filter_update_last_group(struct obd_device *obd, int group)
1002 {
1003         struct filter_obd *filter = &obd->u.filter;
1004         struct file *filp = NULL;
1005         int last_group = 0, rc;
1006         loff_t off = 0;
1007         ENTRY;
1008
1009         if (group <= filter->fo_committed_group)
1010                 RETURN(0);
1011
1012         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1013         if (IS_ERR(filp)) {
1014                 rc = PTR_ERR(filp);
1015                 filp = NULL;
1016                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1017                 GOTO(cleanup, rc);
1018         }
1019
1020         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1021         if (rc) {
1022                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1023                 GOTO(cleanup, rc);
1024         }
1025
1026         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1027                obd->obd_name, last_group, group);
1028
1029         off = 0;
1030         last_group = group;
1031         /* must be sync: bXXXX */
1032         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1033         if (rc) {
1034                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1035                 GOTO(cleanup, rc);
1036         }
1037
1038         filter->fo_committed_group = group;
1039 cleanup:
1040         if (filp)
1041                 filp_close(filp, 0);
1042         RETURN(rc);
1043 }
1044
1045 static int filter_read_group_internal(struct obd_device *obd, int group,
1046                                       int create)
1047 {
1048         struct filter_obd *filter = &obd->u.filter;
1049         __u64 *new_objids = NULL;
1050         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1051         struct dentry **new_groups = NULL;
1052         struct file **new_files = NULL;
1053         struct dentry *dentry;
1054         struct file *filp;
1055         int old_count = filter->fo_group_count, rc, stage = 0, i;
1056         char name[25];
1057         __u64 last_objid;
1058         loff_t off = 0;
1059         int len = group + 1;
1060
1061         snprintf(name, 24, "%d", group);
1062         name[24] = '\0';
1063
1064         if (!create) {
1065                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1066                                            strlen(name));
1067                 if (IS_ERR(dentry)) {
1068                         CERROR("Cannot lookup expected object group %d: %ld\n",
1069                                group, PTR_ERR(dentry));
1070                         RETURN(PTR_ERR(dentry));
1071                 }
1072         } else {
1073                 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1074                                       name, 0700, 1);
1075                 if (IS_ERR(dentry)) {
1076                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1077                                PTR_ERR(dentry));
1078                         RETURN(PTR_ERR(dentry));
1079                 }
1080         }
1081         stage = 1;
1082
1083         snprintf(name, 24, "O/%d/LAST_ID", group);
1084         name[24] = '\0';
1085         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1086         if (IS_ERR(filp)) {
1087                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1088                 GOTO(cleanup, rc = PTR_ERR(filp));
1089         }
1090         stage = 2;
1091
1092         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1093         if (rc) {
1094                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1095                 GOTO(cleanup, rc);
1096         }
1097
1098         if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1099                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1100                 if (tmp_subdirs == NULL)
1101                         GOTO(cleanup, rc = -ENOMEM);
1102                 stage = 3;
1103
1104                 for (i = 0; i < filter->fo_subdir_count; i++) {
1105                         char dir[20];
1106                         snprintf(dir, sizeof(dir), "d%u", i);
1107
1108                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1109                                                               filter->fo_vfsmnt,
1110                                                               dir, 0700, 1);
1111                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1112                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1113                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1114                                        group, dir, rc);
1115                                 GOTO(cleanup, rc);
1116                         }
1117
1118                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1119                                tmp_subdirs->dentry[i]);
1120                 }
1121         }
1122
1123         /* 'group' is an index; we need an array of length 'group + 1' */
1124         if (group + 1 > old_count) {
1125                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1126                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1127                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1128                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1129                 stage = 4;
1130                 if (new_objids == NULL || new_subdirs == NULL ||
1131                     new_groups == NULL || new_files == NULL)
1132                         GOTO(cleanup, rc = -ENOMEM);
1133
1134                 if (old_count) {
1135                         memcpy(new_objids, filter->fo_last_objids,
1136                                old_count * sizeof(*new_objids));
1137                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1138                                old_count * sizeof(*new_subdirs));
1139                         memcpy(new_groups, filter->fo_dentry_O_groups,
1140                                old_count * sizeof(*new_groups));
1141                         memcpy(new_files, filter->fo_last_objid_files,
1142                                old_count * sizeof(*new_files));
1143
1144                         OBD_FREE(filter->fo_last_objids,
1145                                  old_count * sizeof(*new_objids));
1146                         OBD_FREE(filter->fo_dentry_O_sub,
1147                                  old_count * sizeof(*new_subdirs));
1148                         OBD_FREE(filter->fo_dentry_O_groups,
1149                                  old_count * sizeof(*new_groups));
1150                         OBD_FREE(filter->fo_last_objid_files,
1151                                  old_count * sizeof(*new_files));
1152                 }
1153                 filter->fo_last_objids = new_objids;
1154                 filter->fo_dentry_O_sub = new_subdirs;
1155                 filter->fo_dentry_O_groups = new_groups;
1156                 filter->fo_last_objid_files = new_files;
1157                 filter->fo_group_count = len;
1158         }
1159
1160         filter->fo_dentry_O_groups[group] = dentry;
1161         filter->fo_last_objid_files[group] = filp;
1162         if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1163                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1164                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1165         }
1166
1167         filter_update_last_group(obd, group);
1168
1169         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1170                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1171                 rc = filter_update_last_objid(obd, group, 1);
1172                 RETURN(rc);
1173         }
1174
1175         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1176         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1177                obd->obd_name, group, last_objid);
1178         RETURN(0);
1179  cleanup:
1180         switch (stage) {
1181         case 4:
1182                 if (new_objids != NULL)
1183                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1184                 if (new_subdirs != NULL)
1185                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1186                 if (new_groups != NULL)
1187                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1188                 if (new_files != NULL)
1189                         OBD_FREE(new_files, len * sizeof(*new_files));
1190         case 3:
1191                 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1192                         for (i = 0; i < filter->fo_subdir_count; i++) {
1193                                 if (tmp_subdirs->dentry[i] != NULL)
1194                                         dput(tmp_subdirs->dentry[i]);
1195                         }
1196                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1197                 }
1198         case 2:
1199                 filp_close(filp, 0);
1200         case 1:
1201                 dput(dentry);
1202         }
1203         RETURN(rc);
1204 }
1205
1206 static int filter_read_groups(struct obd_device *obd, int last_group,
1207                               int create)
1208 {
1209         struct filter_obd *filter = &obd->u.filter;
1210         int old_count, group, rc = 0;
1211
1212         cfs_down(&filter->fo_init_lock);
1213         old_count = filter->fo_group_count;
1214         for (group = old_count; group <= last_group; group++) {
1215                 rc = filter_read_group_internal(obd, group, create);
1216                 if (rc != 0)
1217                         break;
1218         }
1219         cfs_up(&filter->fo_init_lock);
1220         return rc;
1221 }
1222
1223 /* FIXME: object groups */
1224 static int filter_prep_groups(struct obd_device *obd)
1225 {
1226         struct filter_obd *filter = &obd->u.filter;
1227         struct dentry *O_dentry;
1228         struct file *filp;
1229         int    last_group, rc = 0, cleanup_phase = 0;
1230         loff_t off = 0;
1231         ENTRY;
1232
1233         O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1234                                 "O", 0700, 1);
1235         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1236         if (IS_ERR(O_dentry)) {
1237                 rc = PTR_ERR(O_dentry);
1238                 CERROR("cannot open/create O: rc = %d\n", rc);
1239                 GOTO(cleanup, rc);
1240         }
1241         filter->fo_dentry_O = O_dentry;
1242         cleanup_phase = 1; /* O_dentry */
1243
1244         /* we have to initialize all groups before first connections from
1245          * clients because they may send create/destroy for any group -bzzz */
1246         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1247         if (IS_ERR(filp)) {
1248                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1249                 GOTO(cleanup, rc = PTR_ERR(filp));
1250         }
1251         cleanup_phase = 2; /* filp */
1252
1253         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1254         if (rc) {
1255                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1256                 GOTO(cleanup, rc);
1257         }
1258
1259         if (off == 0)
1260                 last_group = FILTER_GROUP_MDS0;
1261
1262         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1263               FILTER_GROUP_MDS0, last_group);
1264         filter->fo_committed_group = last_group;
1265         rc = filter_read_groups(obd, last_group, 1);
1266         if (rc)
1267                 GOTO(cleanup, rc);
1268
1269         filp_close(filp, 0);
1270         RETURN(0);
1271
1272  cleanup:
1273         switch (cleanup_phase) {
1274         case 2:
1275                 filp_close(filp, 0);
1276         case 1:
1277                 filter_cleanup_groups(obd);
1278                 f_dput(filter->fo_dentry_O);
1279                 filter->fo_dentry_O = NULL;
1280         default:
1281                 break;
1282         }
1283         return rc;
1284
1285 }
1286
1287 /* setup the object store with correct subdirectories */
1288 static int filter_prep(struct obd_device *obd)
1289 {
1290         struct lvfs_run_ctxt saved;
1291         struct filter_obd *filter = &obd->u.filter;
1292         struct file *file;
1293         struct inode *inode;
1294         int rc = 0;
1295         ENTRY;
1296
1297         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1298         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1299         if (!file || IS_ERR(file)) {
1300                 rc = PTR_ERR(file);
1301                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1302                        LAST_RCVD, rc);
1303                 GOTO(out, rc);
1304         }
1305         filter->fo_rcvd_filp = file;
1306         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1307                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1308                        file->f_dentry->d_inode->i_mode);
1309                 GOTO(err_filp, rc = -ENOENT);
1310         }
1311
1312         inode = file->f_dentry->d_parent->d_inode;
1313         /* We use i_op->unlink directly in filter_vfs_unlink() */
1314         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1315                 CERROR("%s: filesystem does not support create/unlink ops\n",
1316                        obd->obd_name);
1317                 GOTO(err_filp, rc = -EOPNOTSUPP);
1318         }
1319
1320         /** lu_target has very limited use in filter now */
1321         lut_init(NULL, &filter->fo_lut, obd, NULL);
1322
1323         rc = filter_init_server_data(obd, file);
1324         if (rc) {
1325                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1326                 GOTO(err_filp, rc);
1327         }
1328
1329         target_recovery_init(&filter->fo_lut, ost_handle);
1330
1331         /* open and create health check io file*/
1332         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1333         if (IS_ERR(file)) {
1334                 rc = PTR_ERR(file);
1335                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1336                        HEALTH_CHECK, rc);
1337                 GOTO(err_server_data, rc);
1338         }
1339         filter->fo_obt.obt_health_check_filp = file;
1340         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1341                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1342                        file->f_dentry->d_inode->i_mode);
1343                 GOTO(err_health_check, rc = -ENOENT);
1344         }
1345         rc = lvfs_check_io_health(obd, file);
1346         if (rc)
1347                 GOTO(err_health_check, rc);
1348
1349         rc = filter_prep_groups(obd);
1350         if (rc)
1351                 GOTO(err_health_check, rc);
1352 out:
1353         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1354
1355         return(rc);
1356
1357 err_health_check:
1358         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1359                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1360         filter->fo_obt.obt_health_check_filp = NULL;
1361 err_server_data:
1362         target_recovery_fini(obd);
1363         filter_free_server_data(filter);
1364 err_filp:
1365         if (filp_close(filter->fo_rcvd_filp, 0))
1366                 CERROR("can't close %s after error\n", LAST_RCVD);
1367         filter->fo_rcvd_filp = NULL;
1368         goto out;
1369 }
1370
1371 /* cleanup the filter: write last used object id to status file */
1372 static void filter_post(struct obd_device *obd)
1373 {
1374         struct lvfs_run_ctxt saved;
1375         struct filter_obd *filter = &obd->u.filter;
1376         int rc, i;
1377
1378         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1379          * best to start a transaction with h_sync, because we removed this
1380          * from lastobjid */
1381
1382         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1383         rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1384                                        filter->fo_fsd);
1385         if (rc)
1386                 CERROR("error writing server data: rc = %d\n", rc);
1387
1388         for (i = 0; i < filter->fo_group_count; i++) {
1389                 rc = filter_update_last_objid(obd, i,
1390                                 (i == filter->fo_group_count - 1));
1391                 if (rc)
1392                         CERROR("error writing group %d lastobjid: rc = %d\n",
1393                                i, rc);
1394         }
1395
1396         rc = filp_close(filter->fo_rcvd_filp, 0);
1397         filter->fo_rcvd_filp = NULL;
1398         if (rc)
1399                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1400
1401         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1402         filter->fo_obt.obt_health_check_filp = NULL;
1403         if (rc)
1404                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1405
1406         filter_cleanup_groups(obd);
1407         filter_free_server_data(filter);
1408         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1409
1410         filter_free_capa_keys(filter);
1411         cleanup_capa_hash(filter->fo_capa_hash);
1412 }
1413
1414 static void filter_set_last_id(struct filter_obd *filter,
1415                                obd_id id, obd_gr group)
1416 {
1417         LASSERT(filter->fo_fsd != NULL);
1418         LASSERT(group <= filter->fo_group_count);
1419
1420         cfs_spin_lock(&filter->fo_objidlock);
1421         filter->fo_last_objids[group] = id;
1422         cfs_spin_unlock(&filter->fo_objidlock);
1423 }
1424
1425 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1426 {
1427         obd_id id;
1428         LASSERT(filter->fo_fsd != NULL);
1429         LASSERT(group <= filter->fo_group_count);
1430         LASSERT(filter->fo_last_objids != NULL);
1431
1432         /* FIXME: object groups */
1433         cfs_spin_lock(&filter->fo_objidlock);
1434         id = filter->fo_last_objids[group];
1435         cfs_spin_unlock(&filter->fo_objidlock);
1436         return id;
1437 }
1438
1439 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1440 {
1441         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1442         return 0;
1443 }
1444
1445 /* We never dget the object parent, so DON'T dput it either */
1446 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1447 {
1448         struct filter_obd *filter = &obd->u.filter;
1449         struct filter_subdirs *subdirs;
1450         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1451
1452         if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1453                 return filter->fo_dentry_O_groups[group];
1454
1455         subdirs = &filter->fo_dentry_O_sub[group];
1456         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1457 }
1458
1459 /* We never dget the object parent, so DON'T dput it either */
1460 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1461                                   obd_id objid)
1462 {
1463         unsigned long now = jiffies;
1464         struct dentry *dparent = filter_parent(obd, group, objid);
1465         int rc;
1466
1467         if (IS_ERR(dparent))
1468                 return dparent;
1469         if (dparent == NULL)
1470                 return ERR_PTR(-ENOENT);
1471
1472         rc = filter_lock_dentry(obd, dparent);
1473         fsfilt_check_slow(obd, now, "parent lock");
1474         return rc ? ERR_PTR(rc) : dparent;
1475 }
1476
1477 /* We never dget the object parent, so DON'T dput it either */
1478 static void filter_parent_unlock(struct dentry *dparent)
1479 {
1480         UNLOCK_INODE_MUTEX(dparent->d_inode);
1481 }
1482
1483 /* How to get files, dentries, inodes from object id's.
1484  *
1485  * If dir_dentry is passed, the caller has already locked the parent
1486  * appropriately for this operation (normally a write lock).  If
1487  * dir_dentry is NULL, we do a read lock while we do the lookup to
1488  * avoid races with create/destroy and such changing the directory
1489  * internal to the filesystem code. */
1490 struct dentry *filter_fid2dentry(struct obd_device *obd,
1491                                  struct dentry *dir_dentry,
1492                                  obd_gr group, obd_id id)
1493 {
1494         struct dentry *dparent = dir_dentry;
1495         struct dentry *dchild;
1496         char name[32];
1497         int len;
1498         ENTRY;
1499
1500         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1501             obd->u.filter.fo_destroys_in_progress == 0) {
1502                 /* don't fail lookups for orphan recovery, it causes
1503                  * later LBUGs when objects still exist during precreate */
1504                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1505                 RETURN(ERR_PTR(-ENOENT));
1506         }
1507         if (id == 0) {
1508                 CERROR("fatal: invalid object id 0\n");
1509                 RETURN(ERR_PTR(-ESTALE));
1510         }
1511
1512         len = sprintf(name, LPU64, id);
1513         if (dir_dentry == NULL) {
1514                 dparent = filter_parent_lock(obd, group, id);
1515                 if (IS_ERR(dparent)) {
1516                         CERROR("%s: error getting object "LPU64":"LPU64
1517                                " parent: rc %ld\n", obd->obd_name,
1518                                id, group, PTR_ERR(dparent));
1519                         RETURN(dparent);
1520                 }
1521         }
1522         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1523                dparent->d_name.len, dparent->d_name.name, name);
1524         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1525         if (dir_dentry == NULL)
1526                 filter_parent_unlock(dparent);
1527         if (IS_ERR(dchild)) {
1528                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1529                        PTR_ERR(dchild));
1530                 RETURN(dchild);
1531         }
1532
1533         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1534                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1535                        obd->obd_name, id, dchild->d_inode->i_ino);
1536                 f_dput(dchild);
1537                 RETURN(ERR_PTR(-ENOENT));
1538         }
1539
1540         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1541                name, dchild, atomic_read(&dchild->d_count));
1542
1543         LASSERT(atomic_read(&dchild->d_count) > 0);
1544
1545         RETURN(dchild);
1546 }
1547
1548 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1549                                   obd_id group, struct lustre_handle *lockh)
1550 {
1551         int flags = LDLM_AST_DISCARD_DATA, rc;
1552         struct ldlm_res_id res_id;
1553         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1554         ENTRY;
1555
1556         osc_build_res_name(objid, group, &res_id);
1557         /* Tell the clients that the object is gone now and that they should
1558          * throw away any cached pages. */
1559         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1560                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1561                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1562                                     lockh);
1563         if (rc != ELDLM_OK)
1564                 lockh->cookie = 0;
1565         RETURN(rc);
1566 }
1567
1568 static void filter_fini_destroy(struct obd_device *obd,
1569                                 struct lustre_handle *lockh)
1570 {
1571         if (lustre_handle_is_used(lockh))
1572                 ldlm_lock_decref(lockh, LCK_PW);
1573 }
1574
1575 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1576  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1577  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1578  * reverse.  Caller must take i_sem before starting the transaction and we
1579  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1580 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1581                       struct vfsmount *mnt)
1582 {
1583         int rc;
1584         ENTRY;
1585
1586         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1587          * itself we already hold dir->i_mutex for child create/unlink ops */
1588         LASSERT(dentry->d_inode != NULL);
1589         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1590         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1591
1592
1593         /* may_delete() */
1594         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1595                 GOTO(out, rc = -ENOENT);
1596
1597         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1598         if (rc)
1599                 GOTO(out, rc);
1600
1601         if (IS_APPEND(dir))
1602                 GOTO(out, rc = -EPERM);
1603
1604         /* check_sticky() */
1605         if ((dentry->d_inode->i_uid != current->fsuid &&
1606              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1607             IS_IMMUTABLE(dentry->d_inode))
1608                 GOTO(out, rc = -EPERM);
1609
1610         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1611          *       that was done because of journal_start (which is already done
1612          *       here) or some other ordering issue. */
1613         DQUOT_INIT(dir);
1614
1615         rc = ll_security_inode_unlink(dir, dentry, mnt);
1616         if (rc)
1617                 GOTO(out, rc);
1618
1619         rc = dir->i_op->unlink(dir, dentry);
1620 out:
1621         /* need to drop i_mutex before we lose inode reference */
1622         UNLOCK_INODE_MUTEX(dentry->d_inode);
1623         if (rc == 0)
1624                 d_delete(dentry);
1625
1626         RETURN(rc);
1627 }
1628
1629 /* Caller must hold LCK_PW on parent and push us into kernel context.
1630  * Caller must hold child i_mutex, we drop it always.
1631  * Caller is also required to ensure that dchild->d_inode exists. */
1632 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1633                                    obd_gr group, struct dentry *dparent,
1634                                    struct dentry *dchild)
1635 {
1636         struct inode *inode = dchild->d_inode;
1637         struct filter_obd *filter = &obd->u.filter;
1638         int rc;
1639
1640         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1641                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1642                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1643                        (unsigned long)inode->i_nlink,
1644                        atomic_read(&inode->i_count));
1645         }
1646
1647         rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1648         if (rc)
1649                 CERROR("error unlinking objid %.*s: rc %d\n",
1650                        dchild->d_name.len, dchild->d_name.name, rc);
1651         return(rc);
1652 }
1653
1654 struct filter_intent_args {
1655         struct ldlm_lock **victim;
1656         __u64 size;
1657         int *liblustre;
1658 };
1659
1660 static enum interval_iter filter_intent_cb(struct interval_node *n,
1661                                            void *args)
1662 {
1663         struct ldlm_interval *node = (struct ldlm_interval *)n;
1664         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1665         __u64 size = arg->size;
1666         struct ldlm_lock **v = arg->victim;
1667         struct ldlm_lock *lck;
1668
1669         /* If the interval is lower than the current file size,
1670          * just break. */
1671         if (interval_high(n) <= size)
1672                 return INTERVAL_ITER_STOP;
1673
1674         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1675                 /* Don't send glimpse ASTs to liblustre clients.
1676                  * They aren't listening for them, and they do
1677                  * entirely synchronous I/O anyways. */
1678                 if (lck->l_export == NULL ||
1679                     lck->l_export->exp_libclient == 1)
1680                         continue;
1681
1682                 if (*arg->liblustre)
1683                         *arg->liblustre = 0;
1684
1685                 if (*v == NULL) {
1686                         *v = LDLM_LOCK_GET(lck);
1687                 } else if ((*v)->l_policy_data.l_extent.start <
1688                            lck->l_policy_data.l_extent.start) {
1689                         LDLM_LOCK_RELEASE(*v);
1690                         *v = LDLM_LOCK_GET(lck);
1691                 }
1692
1693                 /* the same policy group - every lock has the
1694                  * same extent, so needn't do it any more */
1695                 break;
1696         }
1697
1698         return INTERVAL_ITER_CONT;
1699 }
1700
1701 static int filter_intent_policy(struct ldlm_namespace *ns,
1702                                 struct ldlm_lock **lockp, void *req_cookie,
1703                                 ldlm_mode_t mode, int flags, void *data)
1704 {
1705         CFS_LIST_HEAD(rpc_list);
1706         struct ptlrpc_request *req = req_cookie;
1707         struct ldlm_lock *lock = *lockp, *l = NULL;
1708         struct ldlm_resource *res = lock->l_resource;
1709         ldlm_processing_policy policy;
1710         struct ost_lvb *res_lvb, *reply_lvb;
1711         struct ldlm_reply *rep;
1712         ldlm_error_t err;
1713         int idx, rc, tmpflags = 0, only_liblustre = 1;
1714         struct ldlm_interval_tree *tree;
1715         struct filter_intent_args arg;
1716         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1717                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1718                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1719         ENTRY;
1720
1721         policy = ldlm_get_processing_policy(res);
1722         LASSERT(policy != NULL);
1723         LASSERT(req != NULL);
1724
1725         rc = lustre_pack_reply(req, 3, repsize, NULL);
1726         if (rc)
1727                 RETURN(req->rq_status = rc);
1728
1729         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1730         LASSERT(rep != NULL);
1731
1732         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1733                                    sizeof(*reply_lvb));
1734         LASSERT(reply_lvb != NULL);
1735
1736         //fixup_handle_for_resent_req(req, lock, &lockh);
1737
1738         /* Call the extent policy function to see if our request can be
1739          * granted, or is blocked.
1740          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1741          * lock, and should not be granted if the lock will be blocked.
1742          */
1743
1744         LASSERT(ns == res->lr_namespace);
1745         lock_res(res);
1746         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1747         check_res_locked(res);
1748
1749         /* FIXME: we should change the policy function slightly, to not make
1750          * this list at all, since we just turn around and free it */
1751         while (!cfs_list_empty(&rpc_list)) {
1752                 struct ldlm_lock *wlock =
1753                         cfs_list_entry(rpc_list.next, struct ldlm_lock,
1754                                        l_cp_ast);
1755                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1756                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1757                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1758                 cfs_list_del_init(&wlock->l_cp_ast);
1759                 LDLM_LOCK_RELEASE(wlock);
1760         }
1761
1762         /* The lock met with no resistance; we're finished. */
1763         if (rc == LDLM_ITER_CONTINUE) {
1764                 /* do not grant locks to the liblustre clients: they cannot
1765                  * handle ASTs robustly.  We need to do this while still
1766                  * holding ns_lock to avoid the lock remaining on the res_link
1767                  * list (and potentially being added to l_pending_list by an
1768                  * AST) when we are going to drop this lock ASAP. */
1769                 if (lock->l_export->exp_libclient ||
1770                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1771                         ldlm_resource_unlink_lock(lock);
1772                         err = ELDLM_LOCK_ABORTED;
1773                 } else {
1774                         err = ELDLM_LOCK_REPLACED;
1775                 }
1776                 unlock_res(res);
1777                 RETURN(err);
1778         }
1779
1780         /* Do not grant any lock, but instead send GL callbacks.  The extent
1781          * policy nicely created a list of all PW locks for us.  We will choose
1782          * the highest of those which are larger than the size in the LVB, if
1783          * any, and perform a glimpse callback. */
1784         res_lvb = res->lr_lvb_data;
1785         LASSERT(res_lvb != NULL);
1786         *reply_lvb = *res_lvb;
1787
1788         /*
1789          * ->ns_lock guarantees that no new locks are granted, and,
1790          * therefore, that res->lr_lvb_data cannot increase beyond the
1791          * end of already granted lock. As a result, it is safe to
1792          * check against "stale" reply_lvb->lvb_size value without
1793          * res->lr_lvb_sem.
1794          */
1795         arg.size = reply_lvb->lvb_size;
1796         arg.victim = &l;
1797         arg.liblustre = &only_liblustre;
1798         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1799                 tree = &res->lr_itree[idx];
1800                 if (tree->lit_mode == LCK_PR)
1801                         continue;
1802
1803                 interval_iterate_reverse(tree->lit_root,
1804                                          filter_intent_cb, &arg);
1805         }
1806         unlock_res(res);
1807
1808         /* There were no PW locks beyond the size in the LVB; finished. */
1809         if (l == NULL) {
1810                 if (only_liblustre) {
1811                         /* If we discovered a liblustre client with a PW lock,
1812                          * however, the LVB may be out of date!  The LVB is
1813                          * updated only on glimpse (which we don't do for
1814                          * liblustre clients) and cancel (which the client
1815                          * obviously has not yet done).  So if it has written
1816                          * data but kept the lock, the LVB is stale and needs
1817                          * to be updated from disk.
1818                          *
1819                          * Of course, this will all disappear when we switch to
1820                          * taking liblustre locks on the OST. */
1821                         ldlm_res_lvbo_update(res, NULL, 1);
1822                 }
1823                 RETURN(ELDLM_LOCK_ABORTED);
1824         }
1825
1826         /*
1827          * This check is for lock taken in filter_prepare_destroy() that does
1828          * not have l_glimpse_ast set. So the logic is: if there is a lock
1829          * with no l_glimpse_ast set, this object is being destroyed already.
1830          *
1831          * Hence, if you are grabbing DLM locks on the server, always set
1832          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1833          */
1834         if (l->l_glimpse_ast == NULL) {
1835                 /* We are racing with unlink(); just return -ENOENT */
1836                 rep->lock_policy_res1 = -ENOENT;
1837                 goto out;
1838         }
1839
1840         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1841         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1842         /* Update the LVB from disk if the AST failed (this is a legal race) */
1843         /*
1844          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1845          * sending ast is not handled. This can result in lost client writes.
1846          */
1847         if (rc != 0)
1848                 ldlm_res_lvbo_update(res, NULL, 1);
1849
1850         lock_res(res);
1851         *reply_lvb = *res_lvb;
1852         unlock_res(res);
1853
1854  out:
1855         LDLM_LOCK_RELEASE(l);
1856
1857         RETURN(ELDLM_LOCK_ABORTED);
1858 }
1859
1860 /*
1861  * per-obd_device iobuf pool.
1862  *
1863  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1864  * allocations in write-path has to be minimized (see bug 5137).
1865  *
1866  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1867  * OST threads (see ost_thread_{init,done}()).
1868  *
1869  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1870  * at the OST layer there are only (potentially) multiple obd_device of type
1871  * unknown at the time of OST thread creation.
1872  *
1873  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1874  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1875  * it's very own iobuf.
1876  *
1877  * Functions below
1878  *
1879  *     filter_kiobuf_pool_init()
1880  *
1881  *     filter_kiobuf_pool_done()
1882  *
1883  *     filter_iobuf_get()
1884  *
1885  * operate on this array. They are "generic" in a sense that they don't depend
1886  * on actual type of iobuf's (the latter depending on Linux kernel version).
1887  */
1888
1889 /*
1890  * destroy pool created by filter_iobuf_pool_init
1891  */
1892 static void filter_iobuf_pool_done(struct filter_obd *filter)
1893 {
1894         struct filter_iobuf **pool;
1895         int i;
1896
1897         ENTRY;
1898
1899         pool = filter->fo_iobuf_pool;
1900         if (pool != NULL) {
1901                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1902                         if (pool[i] != NULL)
1903                                 filter_free_iobuf(pool[i]);
1904                 }
1905                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1906                 filter->fo_iobuf_pool = NULL;
1907         }
1908         EXIT;
1909 }
1910
1911 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1912 {
1913         struct filter_obd       *filter = &obd->u.filter;
1914         struct sptlrpc_rule_set  tmp_rset;
1915         int                      rc;
1916
1917         sptlrpc_rule_set_init(&tmp_rset);
1918         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1919         if (rc) {
1920                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1921                        obd->obd_name, rc);
1922                 return rc;
1923         }
1924
1925         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1926
1927         cfs_write_lock(&filter->fo_sptlrpc_lock);
1928         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1929         filter->fo_sptlrpc_rset = tmp_rset;
1930         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1931
1932         return 0;
1933 }
1934
1935 /*
1936  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1937  */
1938 static int filter_iobuf_pool_init(struct filter_obd *filter)
1939 {
1940         void **pool;
1941
1942         ENTRY;
1943
1944
1945         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1946                       GFP_KERNEL);
1947         if (filter->fo_iobuf_pool == NULL)
1948                 RETURN(-ENOMEM);
1949
1950         filter->fo_iobuf_count = OSS_THREADS_MAX;
1951
1952         RETURN(0);
1953 }
1954
1955 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1956  * many threads there will be so we allocate a large empty array and only
1957  * fill in those slots that are actually in use.
1958  * If we haven't allocated a pool entry for this thread before, do so now. */
1959 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1960 {
1961         int thread_id                    = (oti && oti->oti_thread) ?
1962                                            oti->oti_thread->t_id : -1;
1963         struct filter_iobuf  *pool       = NULL;
1964         struct filter_iobuf **pool_place = NULL;
1965
1966         if (thread_id >= 0) {
1967                 LASSERT(thread_id < filter->fo_iobuf_count);
1968                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1969         }
1970
1971         if (unlikely(pool == NULL)) {
1972                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1973                                           PTLRPC_MAX_BRW_PAGES);
1974                 if (pool_place != NULL)
1975                         *pool_place = pool;
1976         }
1977
1978         return pool;
1979 }
1980
1981 /* mount the file system (secretly).  lustre_cfg parameters are:
1982  * 1 = device
1983  * 2 = fstype
1984  * 3 = flags: failover=f, failout=n
1985  * 4 = mount options
1986  */
1987 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1988                         void *option)
1989 {
1990         struct filter_obd *filter = &obd->u.filter;
1991         struct vfsmount *mnt;
1992         struct lustre_mount_info *lmi;
1993         struct obd_uuid uuid;
1994         __u8 *uuid_ptr;
1995         char *str, *label;
1996         char ns_name[48];
1997         request_queue_t *q;
1998         int rc, i;
1999         ENTRY;
2000
2001         if (lcfg->lcfg_bufcount < 3 ||
2002             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
2003             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
2004                 RETURN(-EINVAL);
2005
2006         lmi = server_get_mount(obd->obd_name);
2007         if (lmi) {
2008                 /* We already mounted in lustre_fill_super.
2009                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2010                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2011                 mnt = lmi->lmi_mnt;
2012                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2013         } else {
2014                 /* old path - used by lctl */
2015                 CERROR("Using old MDS mount method\n");
2016                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2017                                     MS_NOATIME|MS_NODIRATIME,
2018                                     lustre_cfg_string(lcfg, 1), option);
2019                 if (IS_ERR(mnt)) {
2020                         rc = PTR_ERR(mnt);
2021                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2022                                            lustre_cfg_string(lcfg, 1), rc);
2023                         RETURN(rc);
2024                 }
2025
2026                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2027         }
2028         if (IS_ERR(obd->obd_fsops))
2029                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2030
2031         rc = filter_iobuf_pool_init(filter);
2032         if (rc != 0)
2033                 GOTO(err_ops, rc);
2034
2035         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2036                 CERROR("%s: Underlying device is marked as read-only. "
2037                        "Setup failed\n", obd->obd_name);
2038                 GOTO(err_ops, rc = -EROFS);
2039         }
2040
2041         /* failover is the default */
2042         obd->obd_replayable = 1;
2043
2044         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2045                 str = lustre_cfg_string(lcfg, 3);
2046                 if (strchr(str, 'n')) {
2047                         CWARN("%s: recovery disabled\n", obd->obd_name);
2048                         obd->obd_replayable = 0;
2049                 }
2050         }
2051
2052         filter->fo_vfsmnt = mnt;
2053         obd->u.obt.obt_sb = mnt->mnt_sb;
2054         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2055         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2056
2057         fsfilt_setup(obd, obd->u.obt.obt_sb);
2058
2059         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2060         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2061         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2062         obd->obd_lvfs_ctxt.fs = get_ds();
2063         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2064
2065         cfs_init_mutex(&filter->fo_init_lock);
2066         filter->fo_committed_group = 0;
2067         filter->fo_destroys_in_progress = 0;
2068         for (i = 0; i < 32; i++)
2069                 cfs_sema_init(&filter->fo_create_locks[i], 1);
2070
2071         cfs_spin_lock_init(&filter->fo_translock);
2072         cfs_spin_lock_init(&filter->fo_objidlock);
2073         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2074         cfs_sema_init(&filter->fo_alloc_lock, 1);
2075         init_brw_stats(&filter->fo_filter_stats);
2076         filter->fo_read_cache = 1; /* enable read-only cache by default */
2077         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2078         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2079         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2080         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2081
2082         rc = filter_prep(obd);
2083         if (rc)
2084                 GOTO(err_ops, rc);
2085
2086         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2087         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2088
2089         filter->fo_fl_oss_capa = 1;
2090
2091         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2092         filter->fo_capa_hash = init_capa_hash();
2093         if (filter->fo_capa_hash == NULL)
2094                 GOTO(err_post, rc = -ENOMEM);
2095
2096         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2097         obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2098                                                 LDLM_NAMESPACE_GREEDY);
2099         if (obd->obd_namespace == NULL)
2100                 GOTO(err_post, rc = -ENOMEM);
2101         obd->obd_namespace->ns_lvbp = obd;
2102         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2103         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2104
2105         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2106                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2107
2108         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2109         if (rc) {
2110                 CERROR("failed to setup llogging subsystems\n");
2111                 GOTO(err_post, rc);
2112         }
2113
2114         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2115         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2116         /* do this after llog being initialized */
2117         filter_adapt_sptlrpc_conf(obd, 1);
2118
2119         rc = lquota_setup(filter_quota_interface_ref, obd);
2120         if (rc)
2121                 GOTO(err_post, rc);
2122
2123         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2124         if (q->max_sectors < q->max_hw_sectors &&
2125             q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2126                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2127                               "for larger I/O requests: max_sectors = %u "
2128                               "could be up to max_hw_sectors=%u\n",
2129                               obd->obd_name, mnt->mnt_sb->s_id,
2130                               q->max_sectors, q->max_hw_sectors);
2131
2132         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2133         if (uuid_ptr != NULL) {
2134                 class_uuid_unparse(uuid_ptr, &uuid);
2135                 str = uuid.uuid;
2136         } else {
2137                 str = "no UUID";
2138         }
2139
2140         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2141         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2142                       obd->obd_name, label ?: str, lmi ? "on " : "",
2143                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2144                       obd->obd_replayable ? "enabled" : "disabled");
2145
2146         if (obd->obd_recovering)
2147                 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2148                               "or until %d client%s reconnect%s\n",
2149                               obd->obd_name,
2150                               obd->obd_recovery_timeout / 60,
2151                               obd->obd_recovery_timeout % 60,
2152                               obd->obd_max_recoverable_clients,
2153                               (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2154                               (obd->obd_max_recoverable_clients == 1) ? "s": "");
2155
2156
2157         RETURN(0);
2158
2159 err_post:
2160         filter_post(obd);
2161 err_ops:
2162         fsfilt_put_ops(obd->obd_fsops);
2163         filter_iobuf_pool_done(filter);
2164 err_mntput:
2165         server_put_mount(obd->obd_name, mnt);
2166         obd->u.obt.obt_sb = 0;
2167         return rc;
2168 }
2169
2170 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2171 {
2172         struct lprocfs_static_vars lvars;
2173         unsigned long addr;
2174         struct page *page;
2175         int rc;
2176
2177         CLASSERT(offsetof(struct obd_device, u.obt) ==
2178                  offsetof(struct obd_device, u.filter.fo_obt));
2179
2180         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2181                 RETURN(-EINVAL);
2182
2183         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2184         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2185         if (!page)
2186                 RETURN(-ENOMEM);
2187         addr = (unsigned long)cfs_page_address(page);
2188         clear_page((void *)addr);
2189
2190         /* lprocfs must be setup before the filter so state can be safely added
2191          * to /proc incrementally as the filter is setup */
2192         lprocfs_filter_init_vars(&lvars);
2193         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2194             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2195                 /* Init obdfilter private stats here */
2196                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2197                                      LPROCFS_CNTR_AVGMINMAX,
2198                                      "read_bytes", "bytes");
2199                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2200                                      LPROCFS_CNTR_AVGMINMAX,
2201                                      "write_bytes", "bytes");
2202                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2203                                      LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2204                                      "get_page", "usec");
2205                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2206                                      LPROCFS_CNTR_AVGMINMAX,
2207                                      "get_page_failures", "num");
2208                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2209                                      LPROCFS_CNTR_AVGMINMAX,
2210                                      "cache_access", "pages");
2211                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2212                                      LPROCFS_CNTR_AVGMINMAX,
2213                                      "cache_hit", "pages");
2214                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2215                                      LPROCFS_CNTR_AVGMINMAX,
2216                                      "cache_miss", "pages");
2217
2218                 lproc_filter_attach_seqstat(obd);
2219                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2220                                                         obd->obd_proc_entry,
2221                                                         NULL, NULL);
2222                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2223                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2224                         CERROR("error %d setting up lprocfs for %s\n",
2225                                rc, "exports");
2226                         obd->obd_proc_exports_entry = NULL;
2227                 }
2228         }
2229         if (obd->obd_proc_exports_entry)
2230                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2231                                    lprocfs_nid_stats_clear_read,
2232                                    lprocfs_nid_stats_clear_write, obd, NULL);
2233
2234         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2235                LUSTRE_CFG_BUFLEN(lcfg, 4));
2236         rc = filter_common_setup(obd, lcfg, (void *)addr);
2237         OBD_PAGE_FREE(page);
2238
2239         if (rc) {
2240                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2241                 lprocfs_free_per_client_stats(obd);
2242                 lprocfs_free_obd_stats(obd);
2243                 lprocfs_obd_cleanup(obd);
2244         }
2245
2246         return rc;
2247 }
2248
2249 static struct llog_operations filter_mds_ost_repl_logops;
2250
2251 static struct llog_operations filter_size_orig_logops = {
2252         .lop_setup   = llog_obd_origin_setup,
2253         .lop_cleanup = llog_obd_origin_cleanup,
2254         .lop_add     = llog_obd_origin_add
2255 };
2256
2257 static int filter_olg_fini(struct obd_llog_group *olg)
2258 {
2259         struct llog_ctxt *ctxt;
2260         int rc = 0, rc2 = 0;
2261         ENTRY;
2262
2263         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2264         if (ctxt)
2265                 rc = llog_cleanup(ctxt);
2266
2267         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2268         if (ctxt) {
2269                 rc2 = llog_cleanup(ctxt);
2270                 if (!rc)
2271                         rc = rc2;
2272         }
2273
2274         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2275         if (ctxt) {
2276                 rc2 = llog_cleanup(ctxt);
2277                 if (!rc)
2278                         rc = rc2;
2279         }
2280
2281         RETURN(rc);
2282 }
2283
2284 static int
2285 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2286                 struct obd_device *tgt)
2287 {
2288         int rc;
2289         ENTRY;
2290
2291         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2292                         &filter_mds_ost_repl_logops);
2293         if (rc)
2294                 GOTO(cleanup, rc);
2295
2296         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2297                         &filter_size_orig_logops);
2298         if (rc)
2299                 GOTO(cleanup, rc);
2300         EXIT;
2301 cleanup:
2302         if (rc)
2303                 filter_olg_fini(olg);
2304         return rc;
2305 }
2306
2307 /**
2308  * Init the default olg, which is embeded in the obd_device, for filter.
2309  */
2310 static int
2311 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2312                         struct obd_device *tgt)
2313 {
2314         struct filter_obd *filter = &obd->u.filter;
2315         struct llog_ctxt *ctxt;
2316         int rc;
2317         ENTRY;
2318
2319         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2320         if (!filter->fo_lcm)
2321                 RETURN(-ENOMEM);
2322
2323         filter_mds_ost_repl_logops = llog_client_ops;
2324         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2325         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2326         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2327
2328         rc = filter_olg_init(obd, olg, tgt);
2329         if (rc)
2330                 GOTO(cleanup_lcm, rc);
2331
2332         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2333                         &llog_lvfs_ops);
2334         if (rc)
2335                 GOTO(cleanup_olg, rc);
2336
2337         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2338         if (!ctxt) {
2339                 CERROR("Can't get ctxt for %p:%x\n", olg,
2340                        LLOG_MDS_OST_REPL_CTXT);
2341                 GOTO(cleanup_olg, rc = -ENODEV);
2342         }
2343         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2344         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2345         llog_ctxt_put(ctxt);
2346
2347         RETURN(0);
2348 cleanup_olg:
2349         filter_olg_fini(olg);
2350 cleanup_lcm:
2351         llog_recov_thread_fini(filter->fo_lcm, 1);
2352         filter->fo_lcm = NULL;
2353         return rc;
2354 }
2355
2356 static int
2357 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2358                  struct obd_device *tgt, int *index)
2359 {
2360         struct filter_obd *filter = &obd->u.filter;
2361         struct llog_ctxt *ctxt;
2362         int rc;
2363         ENTRY;
2364
2365         LASSERT(olg != NULL);
2366         if (olg == &obd->obd_olg)
2367                 return filter_default_olg_init(obd, olg, tgt);
2368
2369         LASSERT(filter->fo_lcm != NULL);
2370         rc = filter_olg_init(obd, olg, tgt);
2371         if (rc)
2372                 RETURN(rc);
2373         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2374         if (!ctxt) {
2375                 CERROR("Can't get ctxt for %p:%x\n", olg,
2376                        LLOG_MDS_OST_REPL_CTXT);
2377                 filter_olg_fini(olg);
2378                 RETURN(-ENODEV);
2379         }
2380         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2381         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2382         llog_ctxt_put(ctxt);
2383         RETURN(rc);
2384 }
2385
2386 static int filter_llog_finish(struct obd_device *obd, int count)
2387 {
2388         struct filter_obd *filter = &obd->u.filter;
2389         struct llog_ctxt *ctxt;
2390         ENTRY;
2391
2392         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2393         if (ctxt) {
2394                 /*
2395                  * Make sure that no cached llcds left in recov_thread.
2396                  * We actually do sync in disconnect time, but disconnect
2397                  * may not come being marked rq_no_resend = 1.
2398                  */
2399                 llog_sync(ctxt, NULL);
2400
2401                 /*
2402                  * Balance class_import_get() in llog_receptor_accept().
2403                  * This is safe to do, as llog is already synchronized
2404                  * and its import may go.
2405                  */
2406                 cfs_mutex_down(&ctxt->loc_sem);
2407                 if (ctxt->loc_imp) {
2408                         class_import_put(ctxt->loc_imp);
2409                         ctxt->loc_imp = NULL;
2410                 }
2411                 cfs_mutex_up(&ctxt->loc_sem);
2412                 llog_ctxt_put(ctxt);
2413         }
2414
2415         if (filter->fo_lcm) {
2416                 cfs_mutex_down(&ctxt->loc_sem);
2417                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2418                 filter->fo_lcm = NULL;
2419                 cfs_mutex_up(&ctxt->loc_sem);
2420         }
2421         RETURN(filter_olg_fini(&obd->obd_olg));
2422 }
2423
2424 /**
2425  * Find the group llog according to group index in the llog group list.
2426  */
2427 static struct obd_llog_group *
2428 filter_find_olg_internal(struct filter_obd *filter, int group)
2429 {
2430         struct obd_llog_group *olg;
2431
2432         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2433         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2434                 if (olg->olg_group == group)
2435                         RETURN(olg);
2436         }
2437         RETURN(NULL);
2438 }
2439
2440 /**
2441  * Find the group llog according to group index on the filter
2442  */
2443 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2444 {
2445         struct obd_llog_group *olg = NULL;
2446         struct filter_obd *filter;
2447
2448         filter = &obd->u.filter;
2449
2450         if (group == FILTER_GROUP_LLOG)
2451                 RETURN(&obd->obd_olg);
2452
2453         cfs_spin_lock(&filter->fo_llog_list_lock);
2454         olg = filter_find_olg_internal(filter, group);
2455         cfs_spin_unlock(&filter->fo_llog_list_lock);
2456
2457         RETURN(olg);
2458 }
2459 /**
2460  * Find the llog_group of the filter according to the group. If it can not
2461  * find, create the llog_group, which only happens when mds is being synced
2462  * with OST.
2463  */
2464 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2465 {
2466         struct obd_llog_group *olg = NULL;
2467         struct filter_obd *filter;
2468         int rc;
2469
2470         filter = &obd->u.filter;
2471
2472         if (group == FILTER_GROUP_LLOG)
2473                 RETURN(&obd->obd_olg);
2474
2475         cfs_spin_lock(&filter->fo_llog_list_lock);
2476         olg = filter_find_olg_internal(filter, group);
2477         if (olg) {
2478                 if (olg->olg_initializing) {
2479                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2480                 } else {
2481                         GOTO(out_unlock, olg);
2482                 }
2483         }
2484         OBD_ALLOC_PTR(olg);
2485         if (olg == NULL)
2486                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2487
2488         llog_group_init(olg, group);
2489         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2490         olg->olg_initializing = 1;
2491         cfs_spin_unlock(&filter->fo_llog_list_lock);
2492
2493         rc = obd_llog_init(obd, olg, obd, NULL);
2494         if (rc) {
2495                cfs_spin_lock(&filter->fo_llog_list_lock);
2496                cfs_list_del(&olg->olg_list);
2497                cfs_spin_unlock(&filter->fo_llog_list_lock);
2498                OBD_FREE_PTR(olg);
2499                GOTO(out, olg = ERR_PTR(-ENOMEM));
2500         }
2501         cfs_spin_lock(&filter->fo_llog_list_lock);
2502         olg->olg_initializing = 0;
2503         cfs_spin_unlock(&filter->fo_llog_list_lock);
2504         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2505               obd->obd_name, group, olg);
2506 out:
2507         RETURN(olg);
2508
2509 out_unlock:
2510         cfs_spin_unlock(&filter->fo_llog_list_lock);
2511         GOTO(out, olg);
2512 }
2513
2514 static int filter_llog_connect(struct obd_export *exp,
2515                                struct llogd_conn_body *body)
2516 {
2517         struct obd_device *obd = exp->exp_obd;
2518         struct llog_ctxt *ctxt;
2519         struct obd_llog_group *olg;
2520         int rc;
2521         ENTRY;
2522
2523         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2524                obd->obd_name, body->lgdc_logid.lgl_oid,
2525                body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2526
2527         olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2528         if (!olg) {
2529                 CERROR(" %s: can not find olg of group %d\n",
2530                        obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2531                 RETURN(-ENOENT);
2532         }
2533         llog_group_set_export(olg, exp);
2534
2535         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2536         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2537                  body->lgdc_ctxt_idx);
2538
2539         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2540               obd->obd_name, body->lgdc_logid.lgl_oid,
2541               body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2542
2543         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2544         obd->u.filter.fo_mds_ost_sync = 1;
2545         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2546         rc = llog_connect(ctxt, &body->lgdc_logid,
2547                           &body->lgdc_gen, NULL);
2548         llog_ctxt_put(ctxt);
2549         if (rc != 0)
2550                 CERROR("failed to connect rc %d idx %d\n", rc,
2551                                 body->lgdc_ctxt_idx);
2552
2553         RETURN(rc);
2554 }
2555
2556 static int filter_llog_preclean(struct obd_device *obd)
2557 {
2558         struct obd_llog_group *olg, *tmp;
2559         struct filter_obd *filter;
2560         cfs_list_t  remove_list;
2561         int rc = 0;
2562         ENTRY;
2563
2564         rc = obd_llog_finish(obd, 0);
2565         if (rc)
2566                 CERROR("failed to cleanup llogging subsystem\n");
2567
2568         filter = &obd->u.filter;
2569         CFS_INIT_LIST_HEAD(&remove_list);
2570
2571         cfs_spin_lock(&filter->fo_llog_list_lock);
2572         while (!cfs_list_empty(&filter->fo_llog_list)) {
2573                 olg = cfs_list_entry(filter->fo_llog_list.next,
2574                                      struct obd_llog_group, olg_list);
2575                 cfs_list_del(&olg->olg_list);
2576                 cfs_list_add(&olg->olg_list, &remove_list);
2577         }
2578         cfs_spin_unlock(&filter->fo_llog_list_lock);
2579
2580         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2581                 cfs_list_del_init(&olg->olg_list);
2582                 rc = filter_olg_fini(olg);
2583                 if (rc)
2584                         CERROR("failed to cleanup llogging subsystem for %u\n",
2585                                olg->olg_group);
2586                 OBD_FREE_PTR(olg);
2587         }
2588
2589         RETURN(rc);
2590 }
2591
2592 static int filter_precleanup(struct obd_device *obd,
2593                              enum obd_cleanup_stage stage)
2594 {
2595         int rc = 0;
2596         ENTRY;
2597
2598         switch(stage) {
2599         case OBD_CLEANUP_EARLY:
2600                 break;
2601         case OBD_CLEANUP_EXPORTS:
2602                 /* Stop recovery before namespace cleanup. */
2603                 target_recovery_fini(obd);
2604                 rc = filter_llog_preclean(obd);
2605                 break;
2606         }
2607         RETURN(rc);
2608 }
2609
2610 static int filter_cleanup(struct obd_device *obd)
2611 {
2612         struct filter_obd *filter = &obd->u.filter;
2613         ENTRY;
2614
2615         if (obd->obd_fail)
2616                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2617                               "will be preserved.\n", obd->obd_name);
2618
2619         obd_exports_barrier(obd);
2620         obd_zombie_barrier();
2621
2622         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2623         lprocfs_free_per_client_stats(obd);
2624         lprocfs_free_obd_stats(obd);
2625         lprocfs_obd_cleanup(obd);
2626         lquota_cleanup(filter_quota_interface_ref, obd);
2627
2628         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2629         obd->obd_namespace = NULL;
2630
2631         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2632
2633         if (obd->u.obt.obt_sb == NULL)
2634                 RETURN(0);
2635
2636         filter_post(obd);
2637
2638         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2639         shrink_dcache_sb(obd->u.obt.obt_sb);
2640
2641         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2642         obd->u.obt.obt_sb = NULL;
2643
2644         fsfilt_put_ops(obd->obd_fsops);
2645
2646         filter_iobuf_pool_done(filter);
2647
2648         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2649
2650         RETURN(0);
2651 }
2652
2653 static int filter_connect_internal(struct obd_export *exp,
2654                                    struct obd_connect_data *data,
2655                                    int reconnect)
2656 {
2657         struct filter_export_data *fed = &exp->exp_filter_data;
2658
2659         if (!data)
2660                 RETURN(0);
2661
2662         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2663                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2664                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2665                data->ocd_connect_flags, data->ocd_version,
2666                data->ocd_grant, data->ocd_index);
2667
2668         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2669                 CWARN("!!! This export (nid %s) used object group %d "
2670                        "earlier; now it's trying to use group %d!  This could "
2671                        "be a bug in the MDS. Please report to "
2672                        "http://bugzilla.lustre.org/\n",
2673                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2674                 RETURN(-EPROTO);
2675         }
2676         fed->fed_group = data->ocd_group;
2677
2678         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2679         exp->exp_connect_flags = data->ocd_connect_flags;
2680         data->ocd_version = LUSTRE_VERSION_CODE;
2681
2682         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2683         if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2684                   data->ocd_connect_flags & OBD_CONNECT_MDS))
2685                 RETURN(-EPROTO);
2686
2687         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2688                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2689                 obd_size left, want;
2690
2691                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2692                 left = filter_grant_space_left(exp);
2693                 want = data->ocd_grant;
2694                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2695                 data->ocd_grant = fed->fed_grant;
2696                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2697
2698                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2699                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2700                        exp->exp_client_uuid.uuid, exp,
2701                        data->ocd_grant, want, left);
2702
2703                 filter->fo_tot_granted_clients ++;
2704         }
2705
2706         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2707                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2708                 struct lr_server_data *lsd = filter->fo_fsd;
2709                 int index = le32_to_cpu(lsd->lsd_ost_index);
2710
2711                 if (!(lsd->lsd_feature_compat &
2712                       cpu_to_le32(OBD_COMPAT_OST))) {
2713                         /* this will only happen on the first connect */
2714                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2715                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2716                         /* sync is not needed here as filter_client_add will
2717                          * set exp_need_sync flag */
2718                         filter_update_server_data(exp->exp_obd,
2719                                                   filter->fo_rcvd_filp, lsd);
2720                 } else if (index != data->ocd_index) {
2721                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2722                                            " %u doesn't match actual OST index"
2723                                            " %u in last_rcvd file, bad "
2724                                            "configuration?\n",
2725                                            obd_export_nid2str(exp), index,
2726                                            data->ocd_index);
2727                         RETURN(-EBADF);
2728                 }
2729                 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2730                  * FIXME: We don't strictly need the COMPAT flag for that,
2731                  * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2732                  * FIXME: We needed it for the index, as index 0 is valid. */
2733         }
2734
2735         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2736                 data->ocd_brw_size = 65536;
2737         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2738                 data->ocd_brw_size = min(data->ocd_brw_size,
2739                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2740                 LASSERT(data->ocd_brw_size);
2741         }
2742
2743         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2744                 __u32 cksum_types = data->ocd_cksum_types;
2745
2746                 /* The client set in ocd_cksum_types the checksum types it
2747                  * supports. We have to mask off the algorithms that we don't
2748                  * support */
2749                 if (cksum_types & OBD_CKSUM_ALL)
2750                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2751                 else
2752                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2753
2754                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2755                                    "%x\n", exp->exp_obd->obd_name,
2756                                    obd_export_nid2str(exp), cksum_types,
2757                                    data->ocd_cksum_types);
2758         } else {
2759                 /* This client does not support OBD_CONNECT_CKSUM
2760                  * fall back to CRC32 */
2761                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2762                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2763                                    exp->exp_obd->obd_name,
2764                                    obd_export_nid2str(exp));
2765         }
2766
2767         RETURN(0);
2768 }
2769
2770 static int filter_reconnect(const struct lu_env *env,
2771                             struct obd_export *exp, struct obd_device *obd,
2772                             struct obd_uuid *cluuid,
2773                             struct obd_connect_data *data,
2774                             void *localdata)
2775 {
2776         int rc;
2777         ENTRY;
2778
2779         if (exp == NULL || obd == NULL || cluuid == NULL)
2780                 RETURN(-EINVAL);
2781
2782         rc = filter_connect_internal(exp, data, 1);
2783         if (rc == 0)
2784                 filter_export_stats_init(obd, exp, localdata);
2785
2786         RETURN(rc);
2787 }
2788
2789 /* nearly identical to mds_connect */
2790 static int filter_connect(const struct lu_env *env,
2791                           struct obd_export **exp, struct obd_device *obd,
2792                           struct obd_uuid *cluuid,
2793                           struct obd_connect_data *data, void *localdata)
2794 {
2795         struct lvfs_run_ctxt saved;
2796         struct lustre_handle conn = { 0 };
2797         struct obd_export *lexp;
2798         struct filter_export_data *fed;
2799         struct lsd_client_data *lcd = NULL;
2800         __u32 group;
2801         int rc;
2802         ENTRY;
2803
2804         if (exp == NULL || obd == NULL || cluuid == NULL)
2805                 RETURN(-EINVAL);
2806
2807         rc = class_connect(&conn, obd, cluuid);
2808         if (rc)
2809                 RETURN(rc);
2810         lexp = class_conn2export(&conn);
2811         LASSERT(lexp != NULL);
2812
2813         fed = &lexp->exp_filter_data;
2814
2815         rc = filter_connect_internal(lexp, data, 0);
2816         if (rc)
2817                 GOTO(cleanup, rc);
2818
2819         filter_export_stats_init(obd, lexp, localdata);
2820         if (obd->obd_replayable) {
2821                 OBD_ALLOC(lcd, sizeof(*lcd));
2822                 if (!lcd) {
2823                         CERROR("filter: out of memory for client data\n");
2824                         GOTO(cleanup, rc = -ENOMEM);
2825                 }
2826
2827                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2828                 fed->fed_lcd = lcd;
2829                 rc = filter_client_add(obd, lexp, -1);
2830                 if (rc)
2831                         GOTO(cleanup, rc);
2832         }
2833
2834         group = data->ocd_group;
2835
2836         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2837               obd->obd_name, lexp->exp_handle.h_cookie, group);
2838
2839         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2840         rc = filter_read_groups(obd, group, 1);
2841         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2842         if (rc != 0) {
2843                 CERROR("can't read group %u\n", group);
2844                 GOTO(cleanup, rc);
2845         }
2846
2847         GOTO(cleanup, rc);
2848
2849 cleanup:
2850         if (rc) {
2851                 if (lcd) {
2852                         OBD_FREE_PTR(lcd);
2853                         fed->fed_lcd = NULL;
2854                 }
2855                 class_disconnect(lexp);
2856                 lprocfs_exp_cleanup(lexp);
2857                 *exp = NULL;
2858         } else {
2859                 *exp = lexp;
2860         }
2861
2862         RETURN(rc);
2863 }
2864
2865 /* Do extra sanity checks for grant accounting.  We do this at connect,
2866  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2867  * always get rid of it or turn it off when we know accounting is good. */
2868 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2869 {
2870         struct filter_export_data *fed;
2871         struct obd_export *exp;
2872         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2873         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2874         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2875
2876         if (cfs_list_empty(&obd->obd_exports))
2877                 return;
2878
2879         /* We don't want to do this for large machines that do lots of
2880            mounts or unmounts.  It burns... */
2881         if (obd->obd_num_exports > 100)
2882                 return;
2883
2884         cfs_spin_lock(&obd->obd_osfs_lock);
2885         cfs_spin_lock(&obd->obd_dev_lock);
2886         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2887                 int error = 0;
2888                 fed = &exp->exp_filter_data;
2889                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2890                     fed->fed_dirty < 0)
2891                         error = 1;
2892                 if (maxsize > 0) { /* we may not have done a statfs yet */
2893                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2894                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2895                                  exp->exp_client_uuid.uuid, exp,
2896                                  fed->fed_grant, fed->fed_pending, maxsize);
2897                         LASSERTF(fed->fed_dirty <= maxsize,
2898                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2899                                  exp->exp_client_uuid.uuid, exp,
2900                                  fed->fed_dirty, maxsize);
2901                 }
2902                 if (error)
2903                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2904                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2905                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2906                 else
2907                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2908                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2909                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2910                 tot_granted += fed->fed_grant + fed->fed_pending;
2911                 tot_pending += fed->fed_pending;
2912                 tot_dirty += fed->fed_dirty;
2913         }
2914         fo_tot_granted = obd->u.filter.fo_tot_granted;
2915         fo_tot_pending = obd->u.filter.fo_tot_pending;
2916         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2917         cfs_spin_unlock(&obd->obd_dev_lock);
2918         cfs_spin_unlock(&obd->obd_osfs_lock);
2919
2920         /* Do these assertions outside the spinlocks so we don't kill system */
2921         if (tot_granted != fo_tot_granted)
2922                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2923                        func, tot_granted, fo_tot_granted);
2924         if (tot_pending != fo_tot_pending)
2925                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2926                        func, tot_pending, fo_tot_pending);
2927         if (tot_dirty != fo_tot_dirty)
2928                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2929                        func, tot_dirty, fo_tot_dirty);
2930         if (tot_pending > tot_granted)
2931                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2932                        func, tot_pending, tot_granted);
2933         if (tot_granted > maxsize)
2934                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2935                        func, tot_granted, maxsize);
2936         if (tot_dirty > maxsize)
2937                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2938                        func, tot_dirty, maxsize);
2939 }
2940
2941 /* Remove this client from the grant accounting totals.  We also remove
2942  * the export from the obd device under the osfs and dev locks to ensure
2943  * that the filter_grant_sanity_check() calculations are always valid.
2944  * The client should do something similar when it invalidates its import. */
2945 static void filter_grant_discard(struct obd_export *exp)
2946 {
2947         struct obd_device *obd = exp->exp_obd;
2948         struct filter_obd *filter = &obd->u.filter;
2949         struct filter_export_data *fed = &exp->exp_filter_data;
2950
2951         cfs_spin_lock(&obd->obd_osfs_lock);
2952         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2953                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2954                  obd->obd_name, filter->fo_tot_granted,
2955                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2956         filter->fo_tot_granted -= fed->fed_grant;
2957         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2958                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2959                  obd->obd_name, filter->fo_tot_pending,
2960                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2961         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2962         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2963                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2964                  obd->obd_name, filter->fo_tot_dirty,
2965                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2966         filter->fo_tot_dirty -= fed->fed_dirty;
2967         fed->fed_dirty = 0;
2968         fed->fed_grant = 0;
2969
2970         cfs_spin_unlock(&obd->obd_osfs_lock);
2971 }
2972
2973 static int filter_destroy_export(struct obd_export *exp)
2974 {
2975         ENTRY;
2976
2977         if (exp->exp_filter_data.fed_pending)
2978                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2979                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2980                        exp, exp->exp_filter_data.fed_pending);
2981
2982         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2983
2984         target_destroy_export(exp);
2985         ldlm_destroy_export(exp);
2986
2987         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2988                 RETURN(0);
2989
2990
2991         if (!exp->exp_obd->obd_replayable)
2992                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2993
2994         filter_grant_discard(exp);
2995         filter_fmd_cleanup(exp);
2996
2997         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2998                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2999                 if (filter->fo_tot_granted_clients > 0)
3000                         filter->fo_tot_granted_clients --;
3001         }
3002
3003         if (!(exp->exp_flags & OBD_OPT_FORCE))
3004                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3005
3006         RETURN(0);
3007 }
3008
3009 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3010 {
3011         struct obd_llog_group *olg_min, *olg;
3012         struct filter_obd *filter;
3013         int worked = -1, group;
3014         struct llog_ctxt *ctxt;
3015         ENTRY;
3016
3017         filter = &obd->u.filter;
3018
3019         /* we can't sync log holding spinlock. also, we do not want to get
3020          * into livelock. so we do following: loop over MDS's exports in
3021          * group order and skip already synced llogs -bzzz */
3022         do {
3023                 /* look for group with min. number, but > worked */
3024                 olg_min = NULL;
3025                 group = 1 << 30;
3026                 cfs_spin_lock(&filter->fo_llog_list_lock);
3027                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3028                         if (olg->olg_group <= worked) {
3029                                 /* this group is already synced */
3030                                 continue;
3031                         }
3032                         if (group < olg->olg_group) {
3033                                 /* we have group with smaller number to sync */
3034                                 continue;
3035                         }
3036                         /* store current minimal group */
3037                         olg_min = olg;
3038                         group = olg->olg_group;
3039                 }
3040                 cfs_spin_unlock(&filter->fo_llog_list_lock);
3041
3042                 if (olg_min == NULL)
3043                         break;
3044
3045                 worked = olg_min->olg_group;
3046                 if (olg_min->olg_exp &&
3047                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3048                         int err;
3049                         ctxt = llog_group_get_ctxt(olg_min,
3050                                                    LLOG_MDS_OST_REPL_CTXT);
3051                         if (ctxt) {
3052                                 err = llog_sync(ctxt, olg_min->olg_exp);
3053                                 llog_ctxt_put(ctxt);
3054                                 if (err) {
3055                                         CERROR("error flushing logs to MDS: "
3056                                                "rc %d\n", err);
3057                                 }
3058                         }
3059                 }
3060         } while (olg_min != NULL);
3061 }
3062
3063 /* Also incredibly similar to mds_disconnect */
3064 static int filter_disconnect(struct obd_export *exp)
3065 {
3066         struct obd_device *obd = exp->exp_obd;
3067         int rc;
3068         ENTRY;
3069
3070         LASSERT(exp);
3071         class_export_get(exp);
3072
3073         if (!(exp->exp_flags & OBD_OPT_FORCE))
3074                 filter_grant_sanity_check(obd, __FUNCTION__);
3075         filter_grant_discard(exp);
3076
3077         /* Flush any remaining cancel messages out to the target */
3078         filter_sync_llogs(obd, exp);
3079
3080         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3081
3082         rc = server_disconnect_export(exp);
3083
3084         if (exp->exp_obd->obd_replayable)
3085                 filter_client_free(exp);
3086         else
3087                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3088
3089         class_export_put(exp);
3090         RETURN(rc);
3091 }
3092
3093 /* reverse import is changed, sync all cancels */
3094 static void filter_revimp_update(struct obd_export *exp)
3095 {
3096         ENTRY;
3097
3098         LASSERT(exp);
3099         class_export_get(exp);
3100
3101         /* flush any remaining cancel messages out to the target */
3102         filter_sync_llogs(exp->exp_obd, exp);
3103         class_export_put(exp);
3104         EXIT;
3105 }
3106
3107 static int filter_ping(struct obd_export *exp)
3108 {
3109         filter_fmd_expire(exp);
3110         return 0;
3111 }
3112
3113 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3114                                   const char *what, int quiet)
3115 {
3116         struct dentry *dchild = NULL;
3117         obd_gr group = 0;
3118
3119         if (oa->o_valid & OBD_MD_FLGROUP)
3120                 group = oa->o_gr;
3121
3122         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3123
3124         if (IS_ERR(dchild)) {
3125                 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3126                        what, group, oa->o_id);
3127                 RETURN(dchild);
3128         }
3129
3130         if (dchild->d_inode == NULL) {
3131                 if (!quiet)
3132                         CERROR("%s: %s on non-existent object: "LPU64"\n",
3133                                obd->obd_name, what, oa->o_id);
3134                 f_dput(dchild);
3135                 RETURN(ERR_PTR(-ENOENT));
3136         }
3137
3138         return dchild;
3139 }
3140
3141 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3142 {
3143         struct dentry *dentry = NULL;
3144         struct obd_device *obd;
3145         int rc = 0;
3146         ENTRY;
3147
3148         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3149         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3150                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3151         if (rc)
3152                 RETURN(rc);
3153
3154         obd = class_exp2obd(exp);
3155         if (obd == NULL) {
3156                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3157                 RETURN(-EINVAL);
3158         }
3159
3160         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3161         if (IS_ERR(dentry))
3162                 RETURN(PTR_ERR(dentry));
3163
3164         /* Limit the valid bits in the return data to what we actually use */
3165         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3166         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3167
3168         f_dput(dentry);
3169         RETURN(rc);
3170 }
3171
3172 /* this should be enabled/disabled in condition to enabled/disabled large
3173  * inodes (fast EAs) in backing store FS. */
3174 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3175                         void *handle, struct obdo *oa)
3176 {
3177         struct obd_device *obd = exp->exp_obd;
3178         int rc = 0;
3179         ENTRY;
3180
3181         if (oa->o_valid & OBD_MD_FLFID) {
3182                 struct filter_fid ff;
3183
3184                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3185                         oa->o_gr = 0;
3186                 /* packing fid and converting it to LE for storing into EA.
3187                  * Here ->o_stripe_idx should be filled by LOV and rest of
3188                  * fields - by client. */
3189                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3190                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3191                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3192                 ff.ff_objid = cpu_to_le64(oa->o_id);
3193                 ff.ff_group = cpu_to_le64(oa->o_gr);
3194
3195                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3196                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,