Whamcloud - gitweb
A couple of trivial spelling fixes.
[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 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
83 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS0
84 static struct lvfs_callback_ops filter_lvfs_ops;
85 cfs_mem_cache_t *ll_fmd_cachep;
86
87 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
88                              void *cb_data, int error)
89 {
90         obd_transno_commit_cb(obd, transno, error);
91 }
92
93 /* Assumes caller has already pushed us into the kernel context. */
94 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
95                           int rc, int force_sync)
96 {
97         struct filter_obd *filter = &exp->exp_obd->u.filter;
98         struct filter_export_data *fed = &exp->exp_filter_data;
99         struct lsd_client_data *lcd = fed->fed_lcd;
100         __u64 last_rcvd;
101         loff_t off;
102         int err, log_pri = D_RPCTRACE;
103
104         /* Propagate error code. */
105         if (rc)
106                 RETURN(rc);
107
108         if (!exp->exp_obd->obd_replayable || oti == NULL)
109                 RETURN(rc);
110
111         /* we don't allocate new transnos for replayed requests */
112         if (oti->oti_transno == 0) {
113                 spin_lock(&filter->fo_translock);
114                 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
115                 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
116                 spin_unlock(&filter->fo_translock);
117                 oti->oti_transno = last_rcvd;
118         } else {
119                 spin_lock(&filter->fo_translock);
120                 last_rcvd = oti->oti_transno;
121                 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
122                         filter->fo_fsd->lsd_last_transno =
123                                 cpu_to_le64(last_rcvd);
124                 spin_unlock(&filter->fo_translock);
125         }
126         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
127
128         /* could get xid from oti, if it's ever needed */
129         lcd->lcd_last_xid = 0;
130
131         off = fed->fed_lr_off;
132         if (off <= 0) {
133                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
134                        fed->fed_lr_idx, fed->fed_lr_off);
135                 err = -EINVAL;
136         } else {
137                 if (!force_sync)
138                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
139                                                            last_rcvd,
140                                                            oti->oti_handle,
141                                                            filter_commit_cb,
142                                                            NULL);
143
144                 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
145                                           lcd, sizeof(*lcd), &off,
146                                           force_sync | exp->exp_need_sync);
147                 if (force_sync)
148                         filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
149         }
150         if (err) {
151                 log_pri = D_ERROR;
152                 if (rc == 0)
153                         rc = err;
154         }
155
156         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
157                last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
158
159         RETURN(rc);
160 }
161
162 void f_dput(struct dentry *dentry)
163 {
164         /* Can't go inside filter_ddelete because it can block */
165         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
166                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
167         LASSERT(atomic_read(&dentry->d_count) > 0);
168
169         dput(dentry);
170 }
171
172 static void init_brw_stats(struct brw_stats *brw_stats)
173 {
174         int i;
175         for (i = 0; i < BRW_LAST; i++)
176                 spin_lock_init(&brw_stats->hist[i].oh_lock);
177 }
178
179 static int lprocfs_init_rw_stats(struct obd_device *obd,
180                                  struct lprocfs_stats **stats)
181 {
182         int num_stats;
183
184         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
185                                                         LPROC_FILTER_LAST - 1;
186         *stats = lprocfs_alloc_stats(num_stats, 0);
187         if (*stats == NULL)
188                 return -ENOMEM;
189
190         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
191         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
192                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
193         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
194                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
195
196         return(0);
197 }
198
199 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
200    plus the procfs overhead :( */
201 static int filter_export_stats_init(struct obd_device *obd,
202                                     struct obd_export *exp,
203                                     void *client_nid)
204 {
205         struct filter_export_data *fed = &exp->exp_filter_data;
206         int rc, newnid = 0;
207         ENTRY;
208
209         init_brw_stats(&fed->fed_brw_stats);
210
211         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
212                 /* Self-export gets no proc entry */
213                 RETURN(0);
214
215         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
216         if (rc)
217                 RETURN(rc);
218
219         if (newnid) {
220                 struct nid_stat *tmp = exp->exp_nid_stats;
221                 LASSERT(tmp != NULL);
222
223                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
224                 if (tmp->nid_brw_stats == NULL)
225                         RETURN(-ENOMEM);
226
227                 init_brw_stats(tmp->nid_brw_stats);
228                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
229                                         0644, &filter_per_nid_stats_fops,
230                                         exp->exp_nid_stats);
231                 if (rc)
232                         CWARN("Error adding the brw_stats file\n");
233
234                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
235                 if (rc)
236                         RETURN(rc);
237
238                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
239                                             tmp->nid_stats);
240                 if (rc)
241                         RETURN(rc);
242         }
243
244         RETURN(0);
245 }
246
247 /* Add client data to the FILTER.  We use a bitmap to locate a free space
248  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
249  * Otherwise, we have just read the data from the last_rcvd file and
250  * we know its offset. */
251 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
252                              int cl_idx)
253 {
254         struct filter_obd *filter = &obd->u.filter;
255         struct filter_export_data *fed = &exp->exp_filter_data;
256         unsigned long *bitmap = filter->fo_last_rcvd_slots;
257         int new_client = (cl_idx == -1);
258
259         ENTRY;
260
261         LASSERT(bitmap != NULL);
262         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
263
264         /* Self-export */
265         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
266                 RETURN(0);
267
268         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
269          * there's no need for extra complication here
270          */
271         if (new_client) {
272                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
273         repeat:
274                 if (cl_idx >= LR_MAX_CLIENTS) {
275                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
276                                cl_idx);
277                         RETURN(-EOVERFLOW);
278                 }
279                 if (test_and_set_bit(cl_idx, bitmap)) {
280                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
281                                                     cl_idx);
282                         goto repeat;
283                 }
284         } else {
285                 if (test_and_set_bit(cl_idx, bitmap)) {
286                         CERROR("FILTER client %d: bit already set in bitmap!\n",
287                                cl_idx);
288                         LBUG();
289                 }
290         }
291
292         fed->fed_lr_idx = cl_idx;
293         fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
294                 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
295         LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
296
297         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
298                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
299
300         if (new_client) {
301                 struct lvfs_run_ctxt saved;
302                 loff_t off = fed->fed_lr_off;
303                 int rc;
304                 void *handle;
305
306                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
307                        fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
308
309                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
310                 /* Transaction needed to fix bug 1403 */
311                 handle = fsfilt_start(obd,
312                                       filter->fo_rcvd_filp->f_dentry->d_inode,
313                                       FSFILT_OP_SETATTR, NULL);
314                 if (IS_ERR(handle)) {
315                         rc = PTR_ERR(handle);
316                         CERROR("unable to start transaction: rc %d\n", rc);
317                 } else {
318                         rc = fsfilt_add_journal_cb(obd, 0, handle,
319                                                    target_client_add_cb, exp);
320                         if (rc == 0) {
321                                 spin_lock(&exp->exp_lock);
322                                 exp->exp_need_sync = 1;
323                                 spin_unlock(&exp->exp_lock);
324                         }
325                         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
326                                                  fed->fed_lcd,
327                                                  sizeof(*fed->fed_lcd),
328                                                  &off, rc /* sync if no cb */);
329                         fsfilt_commit(obd,
330                                       filter->fo_rcvd_filp->f_dentry->d_inode,
331                                       handle, 0);
332                 }
333                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
334
335                 if (rc) {
336                         CERROR("error writing %s client idx %u: rc %d\n",
337                                LAST_RCVD, fed->fed_lr_idx, rc);
338                         RETURN(rc);
339                 }
340         }
341         RETURN(0);
342 }
343
344 static int filter_client_free(struct obd_export *exp)
345 {
346         struct filter_export_data *fed = &exp->exp_filter_data;
347         struct filter_obd *filter = &exp->exp_obd->u.filter;
348         struct obd_device *obd = exp->exp_obd;
349         struct lsd_client_data zero_lcd;
350         struct lvfs_run_ctxt saved;
351         int rc;
352         loff_t off;
353         ENTRY;
354
355         if (fed->fed_lcd == NULL)
356                 RETURN(0);
357
358         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
359         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
360                 GOTO(free, 0);
361
362         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
363                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
364
365         LASSERT(filter->fo_last_rcvd_slots != NULL);
366
367         off = fed->fed_lr_off;
368
369         /* Don't clear fed_lr_idx here as it is likely also unset.  At worst
370          * we leak a client slot that will be cleaned on the next recovery. */
371         if (off <= 0) {
372                 CERROR("%s: client idx %d has med_off %lld\n",
373                        obd->obd_name, fed->fed_lr_idx, off);
374                 GOTO(free, rc = -EINVAL);
375         }
376
377         /* Clear the bit _after_ zeroing out the client so we don't
378            race with filter_client_add and zero out new clients.*/
379         if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
380                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
381                        fed->fed_lr_idx);
382                 LBUG();
383         }
384
385         if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
386                 memset(&zero_lcd, 0, sizeof zero_lcd);
387                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
388                 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
389                                          sizeof(zero_lcd), &off,
390                                          (!exp->exp_libclient ||
391                                           exp->exp_need_sync));
392                 if (rc == 0)
393                         /* update server's transno */
394                         filter_update_server_data(obd, filter->fo_rcvd_filp,
395                                                   filter->fo_fsd,
396                                                   !exp->exp_libclient);
397                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
398
399                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
400                        "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
401                        fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
402                        LAST_RCVD, rc);
403         }
404
405         if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
406                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
407                        fed->fed_lr_idx);
408                 LBUG();
409         }
410
411         EXIT;
412 free:
413         OBD_FREE_PTR(fed->fed_lcd);
414         fed->fed_lcd = NULL;
415
416         return 0;
417 }
418
419 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
420 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
421                                          struct filter_mod_data *fmd)
422 {
423         LASSERT_SPIN_LOCKED(&fed->fed_lock);
424         if (--fmd->fmd_refcount == 0) {
425                 /* XXX when we have persistent reservations and the handle
426                  * is stored herein we need to drop it here. */
427                 fed->fed_mod_count--;
428                 list_del(&fmd->fmd_list);
429                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
430         }
431 }
432
433 /* drop fmd reference, free it if last ref */
434 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
435 {
436         struct filter_export_data *fed;
437
438         if (fmd == NULL)
439                 return;
440
441         fed = &exp->exp_filter_data;
442         spin_lock(&fed->fed_lock);
443         filter_fmd_put_nolock(fed, fmd); /* caller reference */
444         spin_unlock(&fed->fed_lock);
445 }
446
447 /* expire entries from the end of the list if there are too many
448  * or they are too old */
449 static void filter_fmd_expire_nolock(struct filter_obd *filter,
450                                      struct filter_export_data *fed,
451                                      struct filter_mod_data *keep)
452 {
453         struct filter_mod_data *fmd, *tmp;
454
455         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
456                 if (fmd == keep)
457                         break;
458
459                 if (time_before(jiffies, fmd->fmd_expire) &&
460                     fed->fed_mod_count < filter->fo_fmd_max_num)
461                         break;
462
463                 list_del_init(&fmd->fmd_list);
464                 filter_fmd_put_nolock(fed, fmd); /* list reference */
465         }
466 }
467
468 void filter_fmd_expire(struct obd_export *exp)
469 {
470         spin_lock(&exp->exp_filter_data.fed_lock);
471         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
472                                  &exp->exp_filter_data, NULL);
473         spin_unlock(&exp->exp_filter_data.fed_lock);
474 }
475
476 /* find specified objid, group in export fmd list.
477  * caller must hold fed_lock and take fmd reference itself */
478 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
479                                                 struct filter_export_data *fed,
480                                                 obd_id objid, obd_gr group)
481 {
482         struct filter_mod_data *found = NULL, *fmd;
483
484         LASSERT_SPIN_LOCKED(&fed->fed_lock);
485
486         list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
487                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
488                         found = fmd;
489                         list_del(&fmd->fmd_list);
490                         list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
491                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
492                         break;
493                 }
494         }
495
496         filter_fmd_expire_nolock(filter, fed, found);
497
498         return found;
499 }
500
501 /* Find fmd based on objid and group, or return NULL if not found. */
502 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
503                                         obd_id objid, obd_gr group)
504 {
505         struct filter_mod_data *fmd;
506
507         spin_lock(&exp->exp_filter_data.fed_lock);
508         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
509                                      &exp->exp_filter_data, objid, group);
510         if (fmd)
511                 fmd->fmd_refcount++;    /* caller reference */
512         spin_unlock(&exp->exp_filter_data.fed_lock);
513
514         return fmd;
515 }
516
517 /* Find fmd based on objid and group, or create a new one if none is found.
518  * It is possible for this function to return NULL under memory pressure,
519  * or if objid = 0 is passed (which will only cause old entries to expire).
520  * Currently this is not fatal because any fmd state is transient and
521  * may also be freed when it gets sufficiently old. */
522 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
523                                        obd_id objid, obd_gr group)
524 {
525         struct filter_export_data *fed = &exp->exp_filter_data;
526         struct filter_mod_data *found = NULL, *fmd_new = NULL;
527
528         OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO, sizeof(*fmd_new));
529
530         spin_lock(&fed->fed_lock);
531         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
532         if (fmd_new) {
533                 if (found == NULL) {
534                         list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
535                         fmd_new->fmd_id = objid;
536                         fmd_new->fmd_gr = group;
537                         fmd_new->fmd_refcount++;   /* list reference */
538                         found = fmd_new;
539                         fed->fed_mod_count++;
540                 } else {
541                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
542                 }
543         }
544         if (found) {
545                 found->fmd_refcount++;          /* caller reference */
546                 found->fmd_expire = jiffies +
547                         exp->exp_obd->u.filter.fo_fmd_max_age;
548         }
549
550         spin_unlock(&fed->fed_lock);
551
552         return found;
553 }
554
555 #ifdef DO_FMD_DROP
556 /* drop fmd list reference so it will disappear when last reference is put.
557  * This isn't so critical because it would in fact only affect the one client
558  * that is doing the unlink and at worst we have an stale entry referencing
559  * an object that should never be used again. */
560 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
561 {
562         struct filter_mod_data *found = NULL;
563
564         spin_lock(&exp->exp_filter_data.fed_lock);
565         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
566         if (found) {
567                 list_del_init(&found->fmd_list);
568                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
569         }
570         spin_unlock(&exp->exp_filter_data.fed_lock);
571 }
572 #else
573 #define filter_fmd_drop(exp, objid, group)
574 #endif
575
576 /* remove all entries from fmd list */
577 static void filter_fmd_cleanup(struct obd_export *exp)
578 {
579         struct filter_export_data *fed = &exp->exp_filter_data;
580         struct filter_mod_data *fmd = NULL, *tmp;
581
582         spin_lock(&fed->fed_lock);
583         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
584                 list_del_init(&fmd->fmd_list);
585                 filter_fmd_put_nolock(fed, fmd);
586         }
587         spin_unlock(&fed->fed_lock);
588 }
589
590 static int filter_init_export(struct obd_export *exp)
591 {
592         spin_lock_init(&exp->exp_filter_data.fed_lock);
593         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
594
595         spin_lock(&exp->exp_lock);
596         exp->exp_connecting = 1;
597         spin_unlock(&exp->exp_lock);
598
599         return ldlm_init_export(exp);
600 }
601
602 static int filter_free_server_data(struct filter_obd *filter)
603 {
604         OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
605         filter->fo_fsd = NULL;
606         OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
607         filter->fo_last_rcvd_slots = NULL;
608         return 0;
609 }
610
611 /* assumes caller is already in kernel ctxt */
612 int filter_update_server_data(struct obd_device *obd, struct file *filp,
613                               struct lr_server_data *fsd, int force_sync)
614 {
615         loff_t off = 0;
616         int rc;
617         ENTRY;
618
619         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
620         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
621                le64_to_cpu(fsd->lsd_last_transno));
622         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
623                le64_to_cpu(fsd->lsd_mount_count));
624
625         fsd->lsd_compat14 = fsd->lsd_last_transno;
626         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
627         if (rc)
628                 CERROR("error writing lr_server_data: rc = %d\n", rc);
629
630         RETURN(rc);
631 }
632
633 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
634                              int force_sync)
635 {
636         struct filter_obd *filter = &obd->u.filter;
637         __u64 tmp;
638         loff_t off = 0;
639         int rc;
640         ENTRY;
641
642         if (filter->fo_last_objid_files[group] == NULL) {
643                 CERROR("Object group "LPU64" not fully setup; not updating "
644                        "last_objid\n", group);
645                 RETURN(-EINVAL);
646         }
647
648         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
649                obd->obd_name, group, filter->fo_last_objids[group]);
650
651         tmp = cpu_to_le64(filter->fo_last_objids[group]);
652         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
653                                  &tmp, sizeof(tmp), &off, force_sync);
654         if (rc)
655                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
656                        group, rc);
657         RETURN(rc);
658 }
659 extern int ost_handle(struct ptlrpc_request *req);
660 /* assumes caller has already in kernel ctxt */
661 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
662 {
663         struct filter_obd *filter = &obd->u.filter;
664         struct lr_server_data *fsd;
665         struct lsd_client_data *lcd = NULL;
666         struct inode *inode = filp->f_dentry->d_inode;
667         unsigned long last_rcvd_size = i_size_read(inode);
668         __u64 mount_count;
669         int cl_idx;
670         loff_t off = 0;
671         int rc;
672
673         /* ensure padding in the struct is the correct size */
674         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
675                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
676         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
677                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
678
679         OBD_ALLOC(fsd, sizeof(*fsd));
680         if (!fsd)
681                 RETURN(-ENOMEM);
682         filter->fo_fsd = fsd;
683
684         OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
685         if (filter->fo_last_rcvd_slots == NULL) {
686                 OBD_FREE(fsd, sizeof(*fsd));
687                 RETURN(-ENOMEM);
688         }
689
690         if (last_rcvd_size == 0) {
691                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
692
693                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
694                 fsd->lsd_last_transno = 0;
695                 mount_count = fsd->lsd_mount_count = 0;
696                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
697                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
698                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
699                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
700                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
701                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
702         } else {
703                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
704                 if (rc) {
705                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
706                                LAST_RCVD, rc);
707                         GOTO(err_fsd, rc);
708                 }
709                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
710                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
711                                            "using the wrong disk %s. Were the "
712                                            "/dev/ assignments rearranged?\n",
713                                            obd->obd_uuid.uuid, fsd->lsd_uuid);
714                         GOTO(err_fsd, rc = -EINVAL);
715                 }
716                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
717                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
718                 /* COMPAT_146 */
719                 /* Assume old last_rcvd format unless I_C_LR is set */
720                 if (!(fsd->lsd_feature_incompat &
721                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
722                         fsd->lsd_last_transno = fsd->lsd_compat14;
723                 /* end COMPAT_146 */
724         }
725
726         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
727                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
728                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
729                        ~FILTER_INCOMPAT_SUPP);
730                 GOTO(err_fsd, rc = -EINVAL);
731         }
732         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
733                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
734                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
735                        ~FILTER_ROCOMPAT_SUPP);
736                 /* Do something like remount filesystem read-only */
737                 GOTO(err_fsd, rc = -EINVAL);
738         }
739
740         CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
741                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
742         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
743                obd->obd_name, mount_count + 1);
744         CDEBUG(D_INODE, "%s: server data size: %u\n",
745                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
746         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
747                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
748         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
749                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
750         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
751                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
752         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
753                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
754                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
755                 le16_to_cpu(fsd->lsd_client_size));
756
757         if (!obd->obd_replayable) {
758                 CWARN("%s: recovery support OFF\n", obd->obd_name);
759                 GOTO(out, rc = 0);
760         }
761
762         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
763              off < last_rcvd_size; cl_idx++) {
764                 __u64 last_rcvd;
765                 struct obd_export *exp;
766                 struct filter_export_data *fed;
767
768                 if (!lcd) {
769                         OBD_ALLOC_PTR(lcd);
770                         if (!lcd)
771                                 GOTO(err_client, rc = -ENOMEM);
772                 }
773
774                 /* Don't assume off is incremented properly by
775                  * fsfilt_read_record(), in case sizeof(*lcd)
776                  * isn't the same as fsd->lsd_client_size.  */
777                 off = le32_to_cpu(fsd->lsd_client_start) +
778                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
779                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
780                 if (rc) {
781                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
782                                LAST_RCVD, cl_idx, off, rc);
783                         break; /* read error shouldn't cause startup to fail */
784                 }
785
786                 if (lcd->lcd_uuid[0] == '\0') {
787                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
788                                cl_idx);
789                         continue;
790                 }
791
792                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
793
794                 /* These exports are cleaned up by filter_disconnect(), so they
795                  * need to be set up like real exports as filter_connect() does.
796                  */
797                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
798
799                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
800                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
801                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
802                 if (IS_ERR(exp)) {
803                         if (PTR_ERR(exp) == -EALREADY) {
804                                 /* export already exists, zero out this one */
805                                 CERROR("Zeroing out duplicate export due to "
806                                        "bug 10479.\n");
807                                 lcd->lcd_uuid[0] = '\0';
808                         } else {
809                                 GOTO(err_client, rc = PTR_ERR(exp));
810                         }
811                 } else {
812                         fed = &exp->exp_filter_data;
813                         fed->fed_lcd = lcd;
814                         fed->fed_group = 0; /* will be assigned at connect */
815                         filter_export_stats_init(obd, exp, NULL);
816                         rc = filter_client_add(obd, exp, cl_idx);
817                         /* can't fail for existing client */
818                         LASSERTF(rc == 0, "rc = %d\n", rc);
819
820                         lcd = NULL;
821                         spin_lock(&exp->exp_lock);
822                         exp->exp_connecting = 0;
823                         exp->exp_in_recovery = 0;
824                         spin_unlock(&exp->exp_lock);
825                         obd->obd_max_recoverable_clients++;
826                         class_export_put(exp);
827                 }
828
829                 /* Need to check last_rcvd even for duplicated exports. */
830                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
831                        cl_idx, last_rcvd);
832
833                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
834                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
835         }
836
837         if (lcd)
838                 OBD_FREE_PTR(lcd);
839
840         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
841
842         target_recovery_init(obd, ost_handle);
843
844 out:
845         filter->fo_mount_count = mount_count + 1;
846         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
847
848         /* save it, so mount count and last_transno is current */
849         rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
850         if (rc)
851                 GOTO(err_client, rc);
852
853         RETURN(0);
854
855 err_client:
856         target_recovery_fini(obd);
857 err_fsd:
858         filter_free_server_data(filter);
859         RETURN(rc);
860 }
861
862 static int filter_cleanup_groups(struct obd_device *obd)
863 {
864         struct filter_obd *filter = &obd->u.filter;
865         struct file *filp;
866         struct dentry *dentry;
867         int i, j;
868         ENTRY;
869
870         if (filter->fo_dentry_O_groups != NULL) {
871                 for (i = 0; i < filter->fo_group_count; i++) {
872                         dentry = filter->fo_dentry_O_groups[i];
873                         if (dentry != NULL)
874                                 f_dput(dentry);
875                 }
876                 OBD_FREE(filter->fo_dentry_O_groups,
877                          filter->fo_group_count *
878                          sizeof(*filter->fo_dentry_O_groups));
879                 filter->fo_dentry_O_groups = NULL;
880         }
881         if (filter->fo_last_objid_files != NULL) {
882                 for (i = 0; i < filter->fo_group_count; i++) {
883                         filp = filter->fo_last_objid_files[i];
884                         if (filp != NULL)
885                                 filp_close(filp, 0);
886                 }
887                 OBD_FREE(filter->fo_last_objid_files,
888                          filter->fo_group_count *
889                          sizeof(*filter->fo_last_objid_files));
890                 filter->fo_last_objid_files = NULL;
891         }
892         if (filter->fo_dentry_O_sub != NULL) {
893                 for (i = 0; i < filter->fo_group_count; i++) {
894                         for (j = 0; j < filter->fo_subdir_count; j++) {
895                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
896                                 if (dentry != NULL)
897                                         f_dput(dentry);
898                         }
899                 }
900                 OBD_FREE(filter->fo_dentry_O_sub,
901                          filter->fo_group_count *
902                          sizeof(*filter->fo_dentry_O_sub));
903                 filter->fo_dentry_O_sub = NULL;
904         }
905         if (filter->fo_last_objids != NULL) {
906                 OBD_FREE(filter->fo_last_objids,
907                          filter->fo_group_count *
908                          sizeof(*filter->fo_last_objids));
909                 filter->fo_last_objids = NULL;
910         }
911         if (filter->fo_dentry_O != NULL) {
912                 f_dput(filter->fo_dentry_O);
913                 filter->fo_dentry_O = NULL;
914         }
915         RETURN(0);
916 }
917
918 static int filter_update_last_group(struct obd_device *obd, int group)
919 {
920         struct filter_obd *filter = &obd->u.filter;
921         struct file *filp = NULL;
922         int last_group = 0, rc;
923         loff_t off = 0;
924         ENTRY;
925
926         if (group <= filter->fo_committed_group)
927                 RETURN(0);
928
929         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
930         if (IS_ERR(filp)) {
931                 rc = PTR_ERR(filp);
932                 filp = NULL;
933                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
934                 GOTO(cleanup, rc);
935         }
936
937         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
938         if (rc) {
939                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
940                 GOTO(cleanup, rc);
941         }
942         LASSERT(off == 0 || last_group >= FILTER_MIN_GROUPS);
943         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
944                obd->obd_name, last_group, group);
945
946         off = 0;
947         last_group = group;
948         /* must be sync: bXXXX */
949         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
950         if (rc) {
951                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
952                 GOTO(cleanup, rc);
953         }
954
955         filter->fo_committed_group = group;
956 cleanup:
957         if (filp)
958                 filp_close(filp, 0);
959         RETURN(rc);
960 }
961
962 static int filter_read_group_internal(struct obd_device *obd, int group,
963                                       int create)
964 {
965         struct filter_obd *filter = &obd->u.filter;
966         __u64 *new_objids = NULL;
967         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
968         struct dentry **new_groups = NULL;
969         struct file **new_files = NULL;
970         struct dentry *dentry;
971         struct file *filp;
972         int old_count = filter->fo_group_count, rc, stage = 0, i;
973         char name[25];
974         __u64 last_objid;
975         loff_t off = 0;
976         int len = group + 1;
977
978         snprintf(name, 24, "%d", group);
979         name[24] = '\0';
980
981         if (!create) {
982                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
983                                            strlen(name));
984                 if (IS_ERR(dentry)) {
985                         CERROR("Cannot lookup expected object group %d: %ld\n",
986                                group, PTR_ERR(dentry));
987                         RETURN(PTR_ERR(dentry));
988                 }
989         } else {
990                 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
991                                       name, 0700, 1);
992                 if (IS_ERR(dentry)) {
993                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
994                                PTR_ERR(dentry));
995                         RETURN(PTR_ERR(dentry));
996                 }
997         }
998         stage = 1;
999
1000         snprintf(name, 24, "O/%d/LAST_ID", group);
1001         name[24] = '\0';
1002         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1003         if (IS_ERR(filp)) {
1004                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1005                 GOTO(cleanup, rc = PTR_ERR(filp));
1006         }
1007         stage = 2;
1008
1009         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1010         if (rc) {
1011                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1012                 GOTO(cleanup, rc);
1013         }
1014
1015         if (filter->fo_subdir_count) {
1016                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1017                 if (tmp_subdirs == NULL)
1018                         GOTO(cleanup, rc = -ENOMEM);
1019                 stage = 3;
1020
1021                 for (i = 0; i < filter->fo_subdir_count; i++) {
1022                         char dir[20];
1023                         snprintf(dir, sizeof(dir), "d%u", i);
1024
1025                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1026                                                               filter->fo_vfsmnt,
1027                                                               dir, 0700, 1);
1028                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1029                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1030                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1031                                        group, dir, rc);
1032                                 GOTO(cleanup, rc);
1033                         }
1034
1035                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1036                                tmp_subdirs->dentry[i]);
1037                 }
1038         }
1039
1040         /* 'group' is an index; we need an array of length 'group + 1' */
1041         if (group + 1 > old_count) {
1042                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1043                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1044                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1045                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1046                 stage = 4;
1047                 if (new_objids == NULL || new_subdirs == NULL ||
1048                     new_groups == NULL || new_files == NULL)
1049                         GOTO(cleanup, rc = -ENOMEM);
1050
1051                 if (old_count) {
1052                         memcpy(new_objids, filter->fo_last_objids,
1053                                old_count * sizeof(*new_objids));
1054                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1055                                old_count * sizeof(*new_subdirs));
1056                         memcpy(new_groups, filter->fo_dentry_O_groups,
1057                                old_count * sizeof(*new_groups));
1058                         memcpy(new_files, filter->fo_last_objid_files,
1059                                old_count * sizeof(*new_files));
1060
1061                         OBD_FREE(filter->fo_last_objids,
1062                                  old_count * sizeof(*new_objids));
1063                         OBD_FREE(filter->fo_dentry_O_sub,
1064                                  old_count * sizeof(*new_subdirs));
1065                         OBD_FREE(filter->fo_dentry_O_groups,
1066                                  old_count * sizeof(*new_groups));
1067                         OBD_FREE(filter->fo_last_objid_files,
1068                                  old_count * sizeof(*new_files));
1069                 }
1070                 filter->fo_last_objids = new_objids;
1071                 filter->fo_dentry_O_sub = new_subdirs;
1072                 filter->fo_dentry_O_groups = new_groups;
1073                 filter->fo_last_objid_files = new_files;
1074                 filter->fo_group_count = len;
1075         }
1076
1077         filter->fo_dentry_O_groups[group] = dentry;
1078         filter->fo_last_objid_files[group] = filp;
1079         if (filter->fo_subdir_count) {
1080                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1081                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1082         }
1083
1084         filter_update_last_group(obd, group);
1085
1086         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1087                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1088                 rc = filter_update_last_objid(obd, group, 1);
1089                 RETURN(rc);
1090         }
1091
1092         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1093         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1094                obd->obd_name, group, last_objid);
1095         RETURN(0);
1096  cleanup:
1097         switch (stage) {
1098         case 4:
1099                 if (new_objids != NULL)
1100                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1101                 if (new_subdirs != NULL)
1102                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1103                 if (new_groups != NULL)
1104                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1105                 if (new_files != NULL)
1106                         OBD_FREE(new_files, len * sizeof(*new_files));
1107         case 3:
1108                 if (filter->fo_subdir_count) {
1109                         for (i = 0; i < filter->fo_subdir_count; i++) {
1110                                 if (tmp_subdirs->dentry[i] != NULL)
1111                                         dput(tmp_subdirs->dentry[i]);
1112                         }
1113                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1114                 }
1115         case 2:
1116                 filp_close(filp, 0);
1117         case 1:
1118                 dput(dentry);
1119         }
1120         RETURN(rc);
1121 }
1122
1123 static int filter_read_groups(struct obd_device *obd, int last_group,
1124                               int create)
1125 {
1126         struct filter_obd *filter = &obd->u.filter;
1127         int old_count, group, rc = 0;
1128
1129         down(&filter->fo_init_lock);
1130         old_count = filter->fo_group_count;
1131         for (group = old_count; group <= last_group; group++) {
1132                 if (group == 0)
1133                         continue; /* no group zero */
1134
1135                 rc = filter_read_group_internal(obd, group, create);
1136                 if (rc != 0)
1137                         break;
1138         }
1139         up(&filter->fo_init_lock);
1140         return rc;
1141 }
1142
1143 /* FIXME: object groups */
1144 static int filter_prep_groups(struct obd_device *obd)
1145 {
1146         struct filter_obd *filter = &obd->u.filter;
1147         struct dentry *dentry, *O_dentry;
1148         struct file *filp;
1149         int    last_group, rc = 0, cleanup_phase = 0;
1150         loff_t off = 0;
1151         ENTRY;
1152
1153         O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1154                                 "O", 0700, 1);
1155         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1156         if (IS_ERR(O_dentry)) {
1157                 rc = PTR_ERR(O_dentry);
1158                 CERROR("cannot open/create O: rc = %d\n", rc);
1159                 GOTO(cleanup, rc);
1160         }
1161         filter->fo_dentry_O = O_dentry;
1162         cleanup_phase = 1; /* O_dentry */
1163
1164         /* Lookup "R" to tell if we're on an old OST FS and need to convert
1165          * from O/R/<dir>/<objid> to O/0/<dir>/<objid>.  This can be removed
1166          * some time post 1.0 when all old-style OSTs have converted along
1167          * with the init_objid hack. */
1168         dentry = ll_lookup_one_len("R", O_dentry, 1);
1169         if (IS_ERR(dentry))
1170                 GOTO(cleanup, rc = PTR_ERR(dentry));
1171         if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1172                 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1173                 ENTRY;
1174
1175                 CWARN("converting OST to new object layout\n");
1176                 if (IS_ERR(O0_dentry)) {
1177                         rc = PTR_ERR(O0_dentry);
1178                         CERROR("error looking up O/0: rc %d\n", rc);
1179                         GOTO(cleanup_R, rc);
1180                 }
1181
1182                 if (O0_dentry->d_inode) {
1183                         CERROR("Both O/R and O/0 exist. Fix manually.\n");
1184                         GOTO(cleanup_O0, rc = -EEXIST);
1185                 }
1186
1187                 LOCK_INODE_MUTEX(O_dentry->d_inode);
1188                 rc = ll_vfs_rename(O_dentry->d_inode, dentry, filter->fo_vfsmnt,
1189                                    O_dentry->d_inode, O0_dentry,
1190                                    filter->fo_vfsmnt);
1191                 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1192
1193                 if (rc) {
1194                         CERROR("error renaming O/R to O/0: rc %d\n", rc);
1195                         GOTO(cleanup_O0, rc);
1196                 }
1197                 filter->fo_fsd->lsd_feature_incompat |=
1198                         cpu_to_le32(OBD_INCOMPAT_GROUPS);
1199                 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1200                                                filter->fo_fsd, 1);
1201                 GOTO(cleanup_O0, rc);
1202
1203         cleanup_O0:
1204                 f_dput(O0_dentry);
1205         cleanup_R:
1206                 f_dput(dentry);
1207                 if (rc)
1208                         GOTO(cleanup, rc);
1209         } else {
1210                 f_dput(dentry);
1211         }
1212
1213         cleanup_phase = 2; /* groups */
1214
1215         /* we have to initialize all groups before first connections from
1216          * clients because they may send create/destroy for any group -bzzz */
1217         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1218         if (IS_ERR(filp)) {
1219                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1220                 GOTO(cleanup, rc = PTR_ERR(filp));
1221         }
1222         cleanup_phase = 3; /* filp */
1223
1224         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1225         if (rc) {
1226                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1227                 GOTO(cleanup, rc);
1228         }
1229         if (off == 0) {
1230                 last_group = FILTER_MIN_GROUPS;
1231         } else {
1232                 LASSERT(last_group >= FILTER_MIN_GROUPS);
1233         }
1234
1235         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1236               FILTER_MIN_GROUPS, last_group);
1237         filter->fo_committed_group = last_group;
1238         rc = filter_read_groups(obd, last_group, 1);
1239         if (rc)
1240                 GOTO(cleanup, rc);
1241
1242         filp_close(filp, 0);
1243         RETURN(0);
1244
1245  cleanup:
1246         switch (cleanup_phase) {
1247         case 3:
1248                 filp_close(filp, 0);
1249         case 2:
1250                 filter_cleanup_groups(obd);
1251         case 1:
1252                 f_dput(filter->fo_dentry_O);
1253                 filter->fo_dentry_O = NULL;
1254         default:
1255                 break;
1256         }
1257         return rc;
1258
1259 }
1260
1261 /* setup the object store with correct subdirectories */
1262 static int filter_prep(struct obd_device *obd)
1263 {
1264         struct lvfs_run_ctxt saved;
1265         struct filter_obd *filter = &obd->u.filter;
1266         struct file *file;
1267         struct inode *inode;
1268         int rc = 0;
1269         ENTRY;
1270
1271         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1272         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1273         if (!file || IS_ERR(file)) {
1274                 rc = PTR_ERR(file);
1275                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1276                        LAST_RCVD, rc);
1277                 GOTO(out, rc);
1278         }
1279         filter->fo_rcvd_filp = file;
1280         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1281                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1282                        file->f_dentry->d_inode->i_mode);
1283                 GOTO(err_filp, rc = -ENOENT);
1284         }
1285
1286         inode = file->f_dentry->d_parent->d_inode;
1287         /* We use i_op->unlink directly in filter_vfs_unlink() */
1288         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1289                 CERROR("%s: filesystem does not support create/unlink ops\n",
1290                        obd->obd_name);
1291                 GOTO(err_filp, rc = -EOPNOTSUPP);
1292         }
1293
1294         rc = filter_init_server_data(obd, file);
1295         if (rc) {
1296                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1297                 GOTO(err_filp, rc);
1298         }
1299         /* open and create health check io file*/
1300         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1301         if (IS_ERR(file)) {
1302                 rc = PTR_ERR(file);
1303                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1304                        HEALTH_CHECK, rc);
1305                 GOTO(err_server_data, rc);
1306         }
1307         filter->fo_health_check_filp = file;
1308         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1309                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1310                        file->f_dentry->d_inode->i_mode);
1311                 GOTO(err_health_check, rc = -ENOENT);
1312         }
1313         rc = lvfs_check_io_health(obd, file);
1314         if (rc)
1315                 GOTO(err_health_check, rc);
1316
1317         rc = filter_prep_groups(obd);
1318         if (rc)
1319                 GOTO(err_health_check, rc);
1320 out:
1321         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1322
1323         return(rc);
1324
1325 err_health_check:
1326         if (filp_close(filter->fo_health_check_filp, 0))
1327                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1328         filter->fo_health_check_filp = NULL;
1329 err_server_data:
1330         target_recovery_fini(obd);
1331         filter_free_server_data(filter);
1332 err_filp:
1333         if (filp_close(filter->fo_rcvd_filp, 0))
1334                 CERROR("can't close %s after error\n", LAST_RCVD);
1335         filter->fo_rcvd_filp = NULL;
1336         goto out;
1337 }
1338
1339 /* cleanup the filter: write last used object id to status file */
1340 static void filter_post(struct obd_device *obd)
1341 {
1342         struct lvfs_run_ctxt saved;
1343         struct filter_obd *filter = &obd->u.filter;
1344         int rc, i;
1345
1346         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1347          * best to start a transaction with h_sync, because we removed this
1348          * from lastobjid */
1349
1350         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1351         rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1352                                        filter->fo_fsd, 0);
1353         if (rc)
1354                 CERROR("error writing server data: rc = %d\n", rc);
1355
1356         for (i = 1; i < filter->fo_group_count; i++) {
1357                 rc = filter_update_last_objid(obd, i,
1358                                 (i == filter->fo_group_count - 1));
1359                 if (rc)
1360                         CERROR("error writing group %d lastobjid: rc = %d\n",
1361                                i, rc);
1362         }
1363
1364         rc = filp_close(filter->fo_rcvd_filp, 0);
1365         filter->fo_rcvd_filp = NULL;
1366         if (rc)
1367                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1368
1369         rc = filp_close(filter->fo_health_check_filp, 0);
1370         filter->fo_health_check_filp = NULL;
1371         if (rc)
1372                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1373
1374         filter_cleanup_groups(obd);
1375         filter_free_server_data(filter);
1376         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1377
1378         filter_free_capa_keys(filter);
1379         cleanup_capa_hash(filter->fo_capa_hash);
1380 }
1381
1382 static void filter_set_last_id(struct filter_obd *filter,
1383                                obd_id id, obd_gr group)
1384 {
1385         LASSERT(filter->fo_fsd != NULL);
1386         LASSERT(group <= filter->fo_group_count);
1387
1388         spin_lock(&filter->fo_objidlock);
1389         filter->fo_last_objids[group] = id;
1390         spin_unlock(&filter->fo_objidlock);
1391 }
1392
1393 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1394 {
1395         obd_id id;
1396         LASSERT(filter->fo_fsd != NULL);
1397         LASSERT(group <= filter->fo_group_count);
1398
1399         /* FIXME: object groups */
1400         spin_lock(&filter->fo_objidlock);
1401         id = filter->fo_last_objids[group];
1402         spin_unlock(&filter->fo_objidlock);
1403
1404         return id;
1405 }
1406
1407 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1408 {
1409         LOCK_INODE_MUTEX(dparent->d_inode);
1410         return 0;
1411 }
1412
1413 /* We never dget the object parent, so DON'T dput it either */
1414 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1415 {
1416         struct filter_obd *filter = &obd->u.filter;
1417         struct filter_subdirs *subdirs;
1418         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1419
1420         if ((group > 0 && group < FILTER_GROUP_MDS0) || 
1421              filter->fo_subdir_count == 0)
1422                 return filter->fo_dentry_O_groups[group];
1423
1424         subdirs = &filter->fo_dentry_O_sub[group];
1425         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1426 }
1427
1428 /* We never dget the object parent, so DON'T dput it either */
1429 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1430                                   obd_id objid)
1431 {
1432         unsigned long now = jiffies;
1433         struct dentry *dparent = filter_parent(obd, group, objid);
1434         int rc;
1435
1436         if (IS_ERR(dparent))
1437                 return dparent;
1438         if (dparent == NULL)
1439                 return ERR_PTR(-ENOENT);
1440
1441         rc = filter_lock_dentry(obd, dparent);
1442         fsfilt_check_slow(obd, now, "parent lock");
1443         return rc ? ERR_PTR(rc) : dparent;
1444 }
1445
1446 /* We never dget the object parent, so DON'T dput it either */
1447 static void filter_parent_unlock(struct dentry *dparent)
1448 {
1449         UNLOCK_INODE_MUTEX(dparent->d_inode);
1450 }
1451
1452 /* How to get files, dentries, inodes from object id's.
1453  *
1454  * If dir_dentry is passed, the caller has already locked the parent
1455  * appropriately for this operation (normally a write lock).  If
1456  * dir_dentry is NULL, we do a read lock while we do the lookup to
1457  * avoid races with create/destroy and such changing the directory
1458  * internal to the filesystem code. */
1459 struct dentry *filter_fid2dentry(struct obd_device *obd,
1460                                  struct dentry *dir_dentry,
1461                                  obd_gr group, obd_id id)
1462 {
1463         struct dentry *dparent = dir_dentry;
1464         struct dentry *dchild;
1465         char name[32];
1466         int len;
1467         ENTRY;
1468
1469         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1470             obd->u.filter.fo_destroys_in_progress == 0) {
1471                 /* don't fail lookups for orphan recovery, it causes
1472                  * later LBUGs when objects still exist during precreate */
1473                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1474                 RETURN(ERR_PTR(-ENOENT));
1475         }
1476         if (id == 0) {
1477                 CERROR("fatal: invalid object id 0\n");
1478                 RETURN(ERR_PTR(-ESTALE));
1479         }
1480
1481         len = sprintf(name, LPU64, id);
1482         if (dir_dentry == NULL) {
1483                 dparent = filter_parent_lock(obd, group, id);
1484                 if (IS_ERR(dparent)) {
1485                         CERROR("%s: error getting object "LPU64":"LPU64
1486                                " parent: rc %ld\n", obd->obd_name,
1487                                id, group, PTR_ERR(dparent));
1488                         RETURN(dparent);
1489                 }
1490         }
1491         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1492                dparent->d_name.len, dparent->d_name.name, name);
1493         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1494         if (dir_dentry == NULL)
1495                 filter_parent_unlock(dparent);
1496         if (IS_ERR(dchild)) {
1497                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1498                        PTR_ERR(dchild));
1499                 RETURN(dchild);
1500         }
1501
1502         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1503                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1504                        obd->obd_name, id, dchild->d_inode->i_ino);
1505                 f_dput(dchild);
1506                 RETURN(ERR_PTR(-ENOENT));
1507         }
1508
1509         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1510                name, dchild, atomic_read(&dchild->d_count));
1511
1512         LASSERT(atomic_read(&dchild->d_count) > 0);
1513
1514         RETURN(dchild);
1515 }
1516
1517 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1518                                   obd_id group)
1519 {
1520         struct lustre_handle lockh;
1521         int flags = LDLM_AST_DISCARD_DATA, rc;
1522         struct ldlm_res_id res_id;
1523         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1524         ENTRY;
1525
1526         osc_build_res_name(objid, group, &res_id);
1527         /* Tell the clients that the object is gone now and that they should
1528          * throw away any cached pages. */
1529         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1530                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1531                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1532                                     &lockh);
1533
1534         /* We only care about the side-effects, just drop the lock. */
1535         if (rc == ELDLM_OK)
1536                 ldlm_lock_decref(&lockh, LCK_PW);
1537
1538         RETURN(rc);
1539 }
1540
1541 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1542  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1543  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1544  * reverse.  Caller must take i_sem before starting the transaction and we
1545  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1546 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1547                       struct vfsmount *mnt)
1548 {
1549         int rc;
1550         ENTRY;
1551
1552         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1553          * itself we already hold dir->i_mutex for child create/unlink ops */
1554         LASSERT(dentry->d_inode != NULL);
1555         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1556         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1557
1558
1559         /* may_delete() */
1560         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1561                 GOTO(out, rc = -ENOENT);
1562
1563         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1564         if (rc)
1565                 GOTO(out, rc);
1566
1567         if (IS_APPEND(dir))
1568                 GOTO(out, rc = -EPERM);
1569
1570         /* check_sticky() */
1571         if ((dentry->d_inode->i_uid != current->fsuid &&
1572              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1573             IS_IMMUTABLE(dentry->d_inode))
1574                 GOTO(out, rc = -EPERM);
1575
1576         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1577          *       that was done because of journal_start (which is already done
1578          *       here) or some other ordering issue. */
1579         DQUOT_INIT(dir);
1580
1581         rc = ll_security_inode_unlink(dir, dentry, mnt);
1582         if (rc)
1583                 GOTO(out, rc);
1584
1585         rc = dir->i_op->unlink(dir, dentry);
1586 out:
1587         /* need to drop i_mutex before we lose inode reference */
1588         UNLOCK_INODE_MUTEX(dentry->d_inode);
1589         if (rc == 0)
1590                 d_delete(dentry);
1591
1592         RETURN(rc);
1593 }
1594
1595 /* Caller must hold LCK_PW on parent and push us into kernel context.
1596  * Caller must hold child i_mutex, we drop it always.
1597  * Caller is also required to ensure that dchild->d_inode exists. */
1598 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1599                                    obd_gr group, struct dentry *dparent,
1600                                    struct dentry *dchild)
1601 {
1602         struct inode *inode = dchild->d_inode;
1603         struct filter_obd *filter = &obd->u.filter;
1604         int rc;
1605
1606         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1607                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1608                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1609                        (unsigned long)inode->i_nlink,
1610                        atomic_read(&inode->i_count));
1611         }
1612
1613         rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1614         if (rc)
1615                 CERROR("error unlinking objid %.*s: rc %d\n",
1616                        dchild->d_name.len, dchild->d_name.name, rc);
1617         return(rc);
1618 }
1619
1620 struct filter_intent_args {
1621         struct ldlm_lock **victim;
1622         __u64 size;
1623         int *liblustre;
1624 };
1625
1626 static enum interval_iter filter_intent_cb(struct interval_node *n,
1627                                            void *args)
1628 {
1629         struct ldlm_interval *node = (struct ldlm_interval *)n;
1630         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1631         __u64 size = arg->size;
1632         struct ldlm_lock **v = arg->victim;
1633         struct ldlm_lock *lck;
1634
1635         /* If the interval is lower than the current file size,
1636          * just break. */
1637         if (interval_high(n) <= size)
1638                 return INTERVAL_ITER_STOP;
1639
1640         list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1641                 /* Don't send glimpse ASTs to liblustre clients.
1642                  * They aren't listening for them, and they do
1643                  * entirely synchronous I/O anyways. */
1644                 if (lck->l_export == NULL ||
1645                     lck->l_export->exp_libclient == 1)
1646                         continue;
1647
1648                 if (*arg->liblustre)
1649                         *arg->liblustre = 0;
1650
1651                 if (*v == NULL) {
1652                         *v = LDLM_LOCK_GET(lck);
1653                 } else if ((*v)->l_policy_data.l_extent.start <
1654                            lck->l_policy_data.l_extent.start) {
1655                         LDLM_LOCK_PUT(*v);
1656                         *v = LDLM_LOCK_GET(lck);
1657                 }
1658
1659                 /* the same policy group - every lock has the
1660                  * same extent, so needn't do it any more */
1661                 break;
1662         }
1663
1664         return INTERVAL_ITER_CONT;
1665 }
1666
1667 static int filter_intent_policy(struct ldlm_namespace *ns,
1668                                 struct ldlm_lock **lockp, void *req_cookie,
1669                                 ldlm_mode_t mode, int flags, void *data)
1670 {
1671         CFS_LIST_HEAD(rpc_list);
1672         struct ptlrpc_request *req = req_cookie;
1673         struct ldlm_lock *lock = *lockp, *l = NULL;
1674         struct ldlm_resource *res = lock->l_resource;
1675         ldlm_processing_policy policy;
1676         struct ost_lvb *res_lvb, *reply_lvb;
1677         struct ldlm_reply *rep;
1678         ldlm_error_t err;
1679         int idx, rc, tmpflags = 0, only_liblustre = 1;
1680         struct ldlm_interval_tree *tree;
1681         struct filter_intent_args arg;
1682         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1683                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1684                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1685         ENTRY;
1686
1687         policy = ldlm_get_processing_policy(res);
1688         LASSERT(policy != NULL);
1689         LASSERT(req != NULL);
1690
1691         rc = lustre_pack_reply(req, 3, repsize, NULL);
1692         if (rc)
1693                 RETURN(req->rq_status = rc);
1694
1695         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1696         LASSERT(rep != NULL);
1697
1698         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1699                                    sizeof(*reply_lvb));
1700         LASSERT(reply_lvb != NULL);
1701
1702         //fixup_handle_for_resent_req(req, lock, &lockh);
1703
1704         /* Call the extent policy function to see if our request can be
1705          * granted, or is blocked.
1706          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1707          * lock, and should not be granted if the lock will be blocked.
1708          */
1709
1710         LASSERT(ns == res->lr_namespace);
1711         lock_res(res);
1712         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1713         check_res_locked(res);
1714
1715         /* FIXME: we should change the policy function slightly, to not make
1716          * this list at all, since we just turn around and free it */
1717         while (!list_empty(&rpc_list)) {
1718                 struct ldlm_lock *wlock =
1719                         list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1720                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1721                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1722                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1723                 list_del_init(&wlock->l_cp_ast);
1724                 LDLM_LOCK_PUT(wlock);
1725         }
1726
1727         /* The lock met with no resistance; we're finished. */
1728         if (rc == LDLM_ITER_CONTINUE) {
1729                 /* do not grant locks to the liblustre clients: they cannot
1730                  * handle ASTs robustly.  We need to do this while still
1731                  * holding ns_lock to avoid the lock remaining on the res_link
1732                  * list (and potentially being added to l_pending_list by an
1733                  * AST) when we are going to drop this lock ASAP. */
1734                 if (lock->l_export->exp_libclient ||
1735                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1736                         ldlm_resource_unlink_lock(lock);
1737                         err = ELDLM_LOCK_ABORTED;
1738                 } else {
1739                         err = ELDLM_LOCK_REPLACED;
1740                 }
1741                 unlock_res(res);
1742                 RETURN(err);
1743         }
1744
1745         /* Do not grant any lock, but instead send GL callbacks.  The extent
1746          * policy nicely created a list of all PW locks for us.  We will choose
1747          * the highest of those which are larger than the size in the LVB, if
1748          * any, and perform a glimpse callback. */
1749         res_lvb = res->lr_lvb_data;
1750         LASSERT(res_lvb != NULL);
1751         *reply_lvb = *res_lvb;
1752
1753         /*
1754          * ->ns_lock guarantees that no new locks are granted, and,
1755          * therefore, that res->lr_lvb_data cannot increase beyond the
1756          * end of already granted lock. As a result, it is safe to
1757          * check against "stale" reply_lvb->lvb_size value without
1758          * res->lr_lvb_sem.
1759          */
1760         arg.size = reply_lvb->lvb_size;
1761         arg.victim = &l;
1762         arg.liblustre = &only_liblustre;
1763         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1764                 tree = &res->lr_itree[idx];
1765                 if (tree->lit_mode == LCK_PR)
1766                         continue;
1767
1768                 interval_iterate_reverse(tree->lit_root,
1769                                          filter_intent_cb, &arg);
1770         }
1771         unlock_res(res);
1772
1773         /* There were no PW locks beyond the size in the LVB; finished. */
1774         if (l == NULL) {
1775                 if (only_liblustre) {
1776                         /* If we discovered a liblustre client with a PW lock,
1777                          * however, the LVB may be out of date!  The LVB is
1778                          * updated only on glimpse (which we don't do for
1779                          * liblustre clients) and cancel (which the client
1780                          * obviously has not yet done).  So if it has written
1781                          * data but kept the lock, the LVB is stale and needs
1782                          * to be updated from disk.
1783                          *
1784                          * Of course, this will all disappear when we switch to
1785                          * taking liblustre locks on the OST. */
1786                         ldlm_res_lvbo_update(res, NULL, 0, 1);
1787                 }
1788                 RETURN(ELDLM_LOCK_ABORTED);
1789         }
1790
1791         /*
1792          * This check is for lock taken in filter_prepare_destroy() that does
1793          * not have l_glimpse_ast set. So the logic is: if there is a lock
1794          * with no l_glimpse_ast set, this object is being destroyed already.
1795          *
1796          * Hence, if you are grabbing DLM locks on the server, always set
1797          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1798          */
1799         if (l->l_glimpse_ast == NULL) {
1800                 /* We are racing with unlink(); just return -ENOENT */
1801                 rep->lock_policy_res1 = -ENOENT;
1802                 goto out;
1803         }
1804
1805         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1806         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1807         /* Update the LVB from disk if the AST failed (this is a legal race) */
1808         /*
1809          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1810          * sending ast is not handled. This can result in lost client writes.
1811          */
1812         if (rc != 0)
1813                 ldlm_res_lvbo_update(res, NULL, 0, 1);
1814
1815         lock_res(res);
1816         *reply_lvb = *res_lvb;
1817         unlock_res(res);
1818
1819  out:
1820         LDLM_LOCK_PUT(l);
1821
1822         RETURN(ELDLM_LOCK_ABORTED);
1823 }
1824
1825 /*
1826  * per-obd_device iobuf pool.
1827  *
1828  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1829  * allocations in write-path has to be minimized (see bug 5137).
1830  *
1831  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1832  * OST threads (see ost_thread_{init,done}()).
1833  *
1834  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1835  * at the OST layer there are only (potentially) multiple obd_device of type
1836  * unknown at the time of OST thread creation.
1837  *
1838  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1839  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1840  * it's very own iobuf.
1841  *
1842  * Functions below
1843  *
1844  *     filter_kiobuf_pool_init()
1845  *
1846  *     filter_kiobuf_pool_done()
1847  *
1848  *     filter_iobuf_get()
1849  *
1850  * operate on this array. They are "generic" in a sense that they don't depend
1851  * on actual type of iobuf's (the latter depending on Linux kernel version).
1852  */
1853
1854 /*
1855  * destroy pool created by filter_iobuf_pool_init
1856  */
1857 static void filter_iobuf_pool_done(struct filter_obd *filter)
1858 {
1859         struct filter_iobuf **pool;
1860         int i;
1861
1862         ENTRY;
1863
1864         pool = filter->fo_iobuf_pool;
1865         if (pool != NULL) {
1866                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1867                         if (pool[i] != NULL)
1868                                 filter_free_iobuf(pool[i]);
1869                 }
1870                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1871                 filter->fo_iobuf_pool = NULL;
1872         }
1873         EXIT;
1874 }
1875
1876 /*
1877  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1878  */
1879 static int filter_iobuf_pool_init(struct filter_obd *filter)
1880 {
1881         void **pool;
1882
1883         ENTRY;
1884
1885
1886         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1887                       GFP_KERNEL);
1888         if (filter->fo_iobuf_pool == NULL)
1889                 RETURN(-ENOMEM);
1890
1891         filter->fo_iobuf_count = OSS_THREADS_MAX;
1892
1893         RETURN(0);
1894 }
1895
1896 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1897  * many threads there will be so we allocate a large empty array and only
1898  * fill in those slots that are actually in use.
1899  * If we haven't allocated a pool entry for this thread before, do so now. */
1900 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1901 {
1902         int thread_id                    = oti ? oti->oti_thread_id : -1;
1903         struct filter_iobuf  *pool       = NULL;
1904         struct filter_iobuf **pool_place = NULL;
1905
1906         if (thread_id >= 0) {
1907                 LASSERT(thread_id < filter->fo_iobuf_count);
1908                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1909         }
1910
1911         if (unlikely(pool == NULL)) {
1912                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1913                                           PTLRPC_MAX_BRW_PAGES);
1914                 if (pool_place != NULL)
1915                         *pool_place = pool;
1916         }
1917
1918         return pool;
1919 }
1920
1921 /* mount the file system (secretly).  lustre_cfg parameters are:
1922  * 1 = device
1923  * 2 = fstype
1924  * 3 = flags: failover=f, failout=n
1925  * 4 = mount options
1926  */
1927 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1928                         void *option)
1929 {
1930         struct filter_obd *filter = &obd->u.filter;
1931         struct vfsmount *mnt;
1932         struct lustre_mount_info *lmi;
1933         struct obd_uuid uuid;
1934         __u8 *uuid_ptr;
1935         char *str, *label;
1936         char ns_name[48];
1937         request_queue_t *q;
1938         int rc, i;
1939         ENTRY;
1940
1941         if (lcfg->lcfg_bufcount < 3 ||
1942             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1943             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1944                 RETURN(-EINVAL);
1945
1946         lmi = server_get_mount(obd->obd_name);
1947         if (lmi) {
1948                 /* We already mounted in lustre_fill_super.
1949                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1950                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1951                 mnt = lmi->lmi_mnt;
1952                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1953         } else {
1954                 /* old path - used by lctl */
1955                 CERROR("Using old MDS mount method\n");
1956                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1957                                     MS_NOATIME|MS_NODIRATIME,
1958                                     lustre_cfg_string(lcfg, 1), option);
1959                 if (IS_ERR(mnt)) {
1960                         rc = PTR_ERR(mnt);
1961                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1962                                            lustre_cfg_string(lcfg, 1), rc);
1963                         RETURN(rc);
1964                 }
1965
1966                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1967         }
1968         if (IS_ERR(obd->obd_fsops))
1969                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1970
1971         rc = filter_iobuf_pool_init(filter);
1972         if (rc != 0)
1973                 GOTO(err_ops, rc);
1974
1975         LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1976
1977         /* failover is the default */
1978         obd->obd_replayable = 1;
1979
1980         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1981                 str = lustre_cfg_string(lcfg, 3);
1982                 if (strchr(str, 'n')) {
1983                         CWARN("%s: recovery disabled\n", obd->obd_name);
1984                         obd->obd_replayable = 0;
1985                 }
1986         }
1987
1988         filter->fo_vfsmnt = mnt;
1989         obd->u.obt.obt_sb = mnt->mnt_sb;
1990         filter->fo_fstype = mnt->mnt_sb->s_type->name;
1991         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1992
1993         fsfilt_setup(obd, obd->u.obt.obt_sb);
1994
1995         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1996         obd->obd_lvfs_ctxt.pwdmnt = mnt;
1997         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1998         obd->obd_lvfs_ctxt.fs = get_ds();
1999         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2000
2001         init_mutex(&filter->fo_init_lock);
2002         filter->fo_committed_group = 0;
2003
2004         rc = filter_prep(obd);
2005         if (rc)
2006                 GOTO(err_ops, rc);
2007
2008         filter->fo_destroys_in_progress = 0;
2009         for (i = 0; i < 32; i++)
2010                 sema_init(&filter->fo_create_locks[i], 1);
2011
2012         spin_lock_init(&filter->fo_translock);
2013         spin_lock_init(&filter->fo_objidlock);
2014         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2015         sema_init(&filter->fo_alloc_lock, 1);
2016         init_brw_stats(&filter->fo_filter_stats);
2017         filter->fo_read_cache = 1; /* enable read-only cache by default */
2018         filter->fo_writethrough_cache = 0; /* disable writethrough cache */
2019         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2020         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2021         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2022
2023         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2024         spin_lock_init(&filter->fo_llog_list_lock);
2025
2026         filter->fo_sptlrpc_lock = RW_LOCK_UNLOCKED;
2027         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2028
2029         filter->fo_fl_oss_capa = 0;
2030         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2031         filter->fo_capa_hash = init_capa_hash();
2032         if (filter->fo_capa_hash == NULL)
2033                 GOTO(err_ops, rc = -ENOMEM);
2034
2035         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2036         obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2037                                                 LDLM_NAMESPACE_GREEDY);
2038         if (obd->obd_namespace == NULL)
2039                 GOTO(err_post, rc = -ENOMEM);
2040         obd->obd_namespace->ns_lvbp = obd;
2041         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2042         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2043
2044         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2045                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2046
2047         rc = obd_llog_init(obd, &obd->obd_olg, obd, 1, NULL, NULL);
2048         if (rc) {
2049                 CERROR("failed to setup llogging subsystems\n");
2050                 GOTO(err_post, rc);
2051         }
2052
2053         rc = lquota_setup(filter_quota_interface_ref, obd);
2054         if (rc)
2055                 GOTO(err_post, rc);
2056
2057         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2058         if (q->max_sectors < q->max_hw_sectors &&
2059             q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2060                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2061                               "for larger I/O requests: max_sectors = %u "
2062                               "could be up to max_hw_sectors=%u\n",
2063                               obd->obd_name, mnt->mnt_sb->s_id,
2064                               q->max_sectors, q->max_hw_sectors);
2065
2066         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2067         if (uuid_ptr != NULL) {
2068                 class_uuid_unparse(uuid_ptr, &uuid);
2069                 str = uuid.uuid;
2070         } else {
2071                 str = "no UUID";
2072         }
2073
2074         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2075
2076         if (obd->obd_recovering) {
2077                 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2078                               "recovery for at least %d:%.02d, or until %d "
2079                               "client%s reconnect. During this time new clients"
2080                               " will not be allowed to connect. "
2081                               "Recovery progress can be monitored by watching "
2082                               "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
2083                               obd->obd_name, lustre_cfg_string(lcfg, 1),
2084                               label ?: "", label ? "/" : "", str,
2085                               obd->obd_recovery_timeout / 60,
2086                               obd->obd_recovery_timeout % 60,
2087                               obd->obd_max_recoverable_clients,
2088                               (obd->obd_max_recoverable_clients == 1) ? "":"s",
2089                               obd->obd_name);
2090         } else {
2091                 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2092                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2093                               label ?: "", label ? "/" : "", str,
2094                               obd->obd_replayable ? "enabled" : "disabled");
2095         }
2096
2097         RETURN(0);
2098
2099 err_post:
2100         filter_post(obd);
2101 err_ops:
2102         fsfilt_put_ops(obd->obd_fsops);
2103         filter_iobuf_pool_done(filter);
2104 err_mntput:
2105         server_put_mount(obd->obd_name, mnt);
2106         obd->u.obt.obt_sb = 0;
2107         return rc;
2108 }
2109
2110 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2111 {
2112         struct lprocfs_static_vars lvars;
2113         unsigned long addr;
2114         struct page *page;
2115         int rc;
2116
2117         CLASSERT(offsetof(struct obd_device, u.obt) ==
2118                  offsetof(struct obd_device, u.filter.fo_obt));
2119
2120         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2121                 RETURN(-EINVAL);
2122
2123         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2124         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2125         if (!page)
2126                 RETURN(-ENOMEM);
2127         addr = (unsigned long)cfs_page_address(page);
2128         clear_page((void *)addr);
2129
2130         /* lprocfs must be setup before the filter so state can be safely added
2131          * to /proc incrementally as the filter is setup */
2132         lprocfs_filter_init_vars(&lvars);
2133         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2134             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2135                 /* Init obdfilter private stats here */
2136                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2137                                      LPROCFS_CNTR_AVGMINMAX,
2138                                      "read_bytes", "bytes");
2139                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2140                                      LPROCFS_CNTR_AVGMINMAX,
2141                                      "write_bytes", "bytes");
2142                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2143                                      LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2144                                      "get_page", "usec");
2145                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2146                                      LPROCFS_CNTR_AVGMINMAX,
2147                                      "get_page_failures", "num");
2148                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2149                                      LPROCFS_CNTR_AVGMINMAX,
2150                                      "cache_access", "pages");
2151                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2152                                      LPROCFS_CNTR_AVGMINMAX,
2153                                      "cache_hit", "pages");
2154                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2155                                      LPROCFS_CNTR_AVGMINMAX,
2156                                      "cache_miss", "pages");
2157
2158                 lproc_filter_attach_seqstat(obd);
2159                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2160                                                         obd->obd_proc_entry,
2161                                                         NULL, NULL);
2162                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2163                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2164                         CERROR("error %d setting up lprocfs for %s\n",
2165                                rc, "exports");
2166                         obd->obd_proc_exports_entry = NULL;
2167                 }
2168         }
2169         if (obd->obd_proc_exports_entry)
2170                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2171                                    lprocfs_nid_stats_clear_read,
2172                                    lprocfs_nid_stats_clear_write, obd, NULL);
2173
2174         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2175                LUSTRE_CFG_BUFLEN(lcfg, 4));
2176         rc = filter_common_setup(obd, lcfg, (void *)addr);
2177         OBD_PAGE_FREE(page);
2178
2179         if (rc) {
2180                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2181                 lprocfs_free_per_client_stats(obd);
2182                 lprocfs_free_obd_stats(obd);
2183                 lprocfs_obd_cleanup(obd);
2184         }
2185
2186         return rc;
2187 }
2188
2189 static struct llog_operations filter_mds_ost_repl_logops;
2190
2191 static struct llog_operations filter_size_orig_logops = {
2192         .lop_setup   = llog_obd_origin_setup,
2193         .lop_cleanup = llog_obd_origin_cleanup,
2194         .lop_add     = llog_obd_origin_add
2195 };
2196
2197 static int filter_olg_fini(struct obd_llog_group *olg)
2198 {
2199         struct llog_ctxt *ctxt;
2200         int rc = 0, rc2 = 0;
2201         ENTRY;
2202
2203         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2204         if (ctxt)
2205                 rc = llog_cleanup(ctxt);
2206
2207         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2208         if (ctxt)
2209                 rc2 = llog_cleanup(ctxt);
2210         if (!rc)
2211                 rc = rc2;
2212
2213         RETURN(rc);
2214 }
2215
2216 static int
2217 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2218                 struct obd_device *tgt)
2219 {
2220         int rc;
2221         ENTRY;
2222
2223         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2224                         &filter_mds_ost_repl_logops);
2225         if (rc)
2226                 GOTO(cleanup, rc);
2227
2228         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2229                         &filter_size_orig_logops);
2230         if (rc)
2231                 GOTO(cleanup, rc);
2232         EXIT;
2233 cleanup:
2234         if (rc)
2235                 filter_olg_fini(olg);
2236         return rc;
2237 }
2238
2239 /**
2240  * Init the default olg, which is embeded in the obd_device, for filter.
2241  */
2242 static int
2243 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2244                         struct obd_device *tgt)
2245 {
2246         struct filter_obd *filter = &obd->u.filter;
2247         struct llog_ctxt *ctxt;
2248         int rc;
2249         ENTRY;
2250
2251         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2252         if (!filter->fo_lcm)
2253                 RETURN(-ENOMEM);
2254
2255         filter_mds_ost_repl_logops = llog_client_ops;
2256         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2257         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2258         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2259
2260         rc = filter_olg_init(obd, olg, tgt);
2261         if (rc)
2262                 GOTO(cleanup_lcm, rc);
2263
2264         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2265         if (!ctxt) {
2266                 CERROR("Can't get ctxt for %p:%x\n", olg,
2267                        LLOG_MDS_OST_REPL_CTXT);
2268                 GOTO(cleanup_olg, rc = -ENODEV);
2269         }
2270         ctxt->loc_lcm = filter->fo_lcm;
2271         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2272         llog_ctxt_put(ctxt);
2273
2274         RETURN(0);
2275 cleanup_olg:
2276         filter_olg_fini(olg);
2277 cleanup_lcm:
2278         llog_recov_thread_fini(filter->fo_lcm, 1);
2279         filter->fo_lcm = NULL;
2280         return rc;
2281 }
2282
2283 static int
2284 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2285                  struct obd_device *tgt, int count, struct llog_catid *catid,
2286                  struct obd_uuid *uuid)
2287 {
2288         struct filter_obd *filter = &obd->u.filter;
2289         struct llog_ctxt *ctxt;
2290         int rc;
2291         ENTRY;
2292
2293         LASSERT(olg != NULL);
2294         if (olg == &obd->obd_olg)
2295                 return filter_default_olg_init(obd, olg, tgt);
2296
2297         LASSERT(filter->fo_lcm != NULL);
2298         rc = filter_olg_init(obd, olg, tgt);
2299         if (rc)
2300                 RETURN(rc);
2301         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2302         if (!ctxt) {
2303                 CERROR("Can't get ctxt for %p:%x\n", olg,
2304                        LLOG_MDS_OST_REPL_CTXT);
2305                 filter_olg_fini(olg);
2306                 RETURN(-ENODEV);
2307         }
2308         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2309         ctxt->loc_lcm = filter->fo_lcm;
2310         llog_ctxt_put(ctxt);
2311         RETURN(rc);
2312 }
2313
2314 static int filter_llog_finish(struct obd_device *obd, int count)
2315 {
2316         struct filter_obd *filter = &obd->u.filter;
2317         struct llog_ctxt *ctxt;
2318         ENTRY;
2319
2320         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2321         LASSERT(ctxt != NULL);
2322         mutex_down(&ctxt->loc_sem);
2323         if (ctxt->loc_imp) {
2324                 /*
2325                  * Balance class_import_get() in llog_receptor_accept(). This
2326                  * is safe to do here, as llog is already synchronized and its
2327                  * import may go.
2328                  */
2329                 class_import_put(ctxt->loc_imp);
2330                 ctxt->loc_imp = NULL;
2331         }
2332         mutex_up(&ctxt->loc_sem);
2333         llog_ctxt_put(ctxt);
2334
2335         if (filter->fo_lcm) {
2336                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2337                 filter->fo_lcm = NULL;
2338         }
2339         RETURN(filter_olg_fini(&obd->obd_olg));
2340 }
2341
2342 /**
2343  * Find the group llog according to group index in the llog group list.
2344  */
2345 static struct obd_llog_group *
2346 filter_find_olg_internal(struct filter_obd *filter, int group)
2347 {
2348         struct obd_llog_group *olg;
2349
2350         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2351         list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2352                 if (olg->olg_group == group)
2353                         RETURN(olg);
2354         }
2355         RETURN(NULL);
2356 }
2357
2358 /**
2359  * Find the group llog according to group index on the filter
2360  */
2361 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2362 {
2363         struct obd_llog_group *olg = NULL;
2364         struct filter_obd *filter;
2365
2366         filter = &obd->u.filter;
2367
2368         if (group == FILTER_GROUP_LLOG)
2369                 RETURN(&obd->obd_olg);
2370
2371         spin_lock(&filter->fo_llog_list_lock);
2372         olg = filter_find_olg_internal(filter, group);
2373         spin_unlock(&filter->fo_llog_list_lock);
2374
2375         RETURN(olg);
2376 }
2377 /**
2378  * Find the llog_group of the filter according to the group. If it can not
2379  * find, create the llog_group, which only happens when mds is being synced
2380  * with OST.
2381  */
2382 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2383 {
2384         struct obd_llog_group *olg = NULL;
2385         struct filter_obd *filter;
2386         int rc;
2387
2388         filter = &obd->u.filter;
2389
2390         if (group == FILTER_GROUP_LLOG)
2391                 RETURN(&obd->obd_olg);
2392
2393         spin_lock(&filter->fo_llog_list_lock);
2394         olg = filter_find_olg_internal(filter, group);
2395         if (olg) {
2396                 if (olg->olg_initializing) {
2397                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2398                 } else {
2399                         GOTO(out_unlock, olg);
2400                 }
2401         }
2402         OBD_ALLOC_PTR(olg);
2403         if (olg == NULL)
2404                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2405
2406         llog_group_init(olg, group);
2407         list_add(&olg->olg_list, &filter->fo_llog_list);
2408         olg->olg_initializing = 1;
2409         spin_unlock(&filter->fo_llog_list_lock);
2410
2411         rc = llog_cat_initialize(obd, olg, 1, NULL);
2412         if (rc) {
2413                spin_lock(&filter->fo_llog_list_lock);
2414                list_del(&olg->olg_list);
2415                spin_unlock(&filter->fo_llog_list_lock);
2416                OBD_FREE_PTR(olg);
2417                GOTO(out, olg = ERR_PTR(-ENOMEM));
2418         }
2419         spin_lock(&filter->fo_llog_list_lock);
2420         olg->olg_initializing = 0;
2421         spin_unlock(&filter->fo_llog_list_lock);
2422         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2423               obd->obd_name, group, olg);
2424 out:
2425         RETURN(olg);
2426
2427 out_unlock:
2428         spin_unlock(&filter->fo_llog_list_lock);
2429         GOTO(out, olg);
2430 }
2431
2432 static int filter_llog_connect(struct obd_export *exp,
2433                                struct llogd_conn_body *body)
2434 {
2435         struct obd_device *obd = exp->exp_obd;
2436         struct llog_ctxt *ctxt;
2437         struct obd_llog_group *olg;
2438         int rc;
2439         ENTRY;
2440
2441         CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2442                 (unsigned) body->lgdc_logid.lgl_ogr,
2443                 (unsigned) body->lgdc_logid.lgl_oid,
2444                 (unsigned) body->lgdc_logid.lgl_ogen);
2445
2446         olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2447         if (!olg) {
2448                 CERROR(" %s: can not find olg of group %d\n",
2449                        obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2450                 RETURN(-ENOENT);
2451         }
2452         llog_group_set_export(olg, exp);
2453
2454         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2455         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2456                  body->lgdc_ctxt_idx);
2457         rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2458                           &body->lgdc_gen, NULL);
2459         llog_ctxt_put(ctxt);
2460         if (rc != 0)
2461                 CERROR("failed to connect rc %d idx %d\n", rc,
2462                                 body->lgdc_ctxt_idx);
2463
2464         RETURN(rc);
2465 }
2466
2467 static int filter_llog_preclean(struct obd_device *obd)
2468 {
2469         struct obd_llog_group *olg, *tmp;
2470         struct filter_obd *filter;
2471         struct list_head  remove_list;
2472         int rc = 0;
2473         ENTRY;
2474
2475         rc = obd_llog_finish(obd, 0);
2476         if (rc)
2477                 CERROR("failed to cleanup llogging subsystem\n");
2478
2479         filter = &obd->u.filter;
2480         CFS_INIT_LIST_HEAD(&remove_list);
2481
2482         spin_lock(&filter->fo_llog_list_lock);
2483         while (!list_empty(&filter->fo_llog_list)) {
2484                 olg = list_entry(filter->fo_llog_list.next,
2485                                  struct obd_llog_group, olg_list);
2486                 list_del(&olg->olg_list);
2487                 list_add(&olg->olg_list, &remove_list);
2488         }
2489         spin_unlock(&filter->fo_llog_list_lock);
2490
2491         list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2492                 list_del_init(&olg->olg_list);
2493                 rc = filter_olg_fini(olg);
2494                 if (rc)
2495                         CERROR("failed to cleanup llogging subsystem for %u\n",
2496                                olg->olg_group);
2497                 OBD_FREE_PTR(olg);
2498         }
2499
2500         RETURN(rc);
2501 }
2502
2503 static int filter_precleanup(struct obd_device *obd,
2504                              enum obd_cleanup_stage stage)
2505 {
2506         int rc = 0;
2507         ENTRY;
2508
2509         switch(stage) {
2510         case OBD_CLEANUP_EARLY:
2511                 break;
2512         case OBD_CLEANUP_EXPORTS:
2513                 target_cleanup_recovery(obd);
2514                 rc = filter_llog_preclean(obd);
2515                 break;
2516         }
2517         RETURN(rc);
2518 }
2519
2520 static int filter_cleanup(struct obd_device *obd)
2521 {
2522         struct filter_obd *filter = &obd->u.filter;
2523         ENTRY;
2524
2525         if (obd->obd_fail)
2526                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2527                               "will be preserved.\n", obd->obd_name);
2528
2529         if (!list_empty(&obd->obd_exports)) {
2530                 CERROR("%s: still has clients!\n", obd->obd_name);
2531                 class_disconnect_exports(obd);
2532                 if (!list_empty(&obd->obd_exports)) {
2533                         CERROR("still has exports after forced cleanup?\n");
2534                         RETURN(-EBUSY);
2535                 }
2536         }
2537
2538         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2539         lprocfs_free_per_client_stats(obd);
2540         lprocfs_free_obd_stats(obd);
2541         lprocfs_obd_cleanup(obd);
2542         lquota_cleanup(filter_quota_interface_ref, obd);
2543
2544         /* Stop recovery before namespace cleanup. */
2545         target_stop_recovery_thread(obd);
2546         target_cleanup_recovery(obd);
2547
2548         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2549         obd->obd_namespace = NULL;
2550
2551         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2552
2553         if (obd->u.obt.obt_sb == NULL)
2554                 RETURN(0);
2555
2556         filter_post(obd);
2557
2558         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2559         shrink_dcache_sb(obd->u.obt.obt_sb);
2560
2561         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2562         obd->u.obt.obt_sb = NULL;
2563
2564         fsfilt_put_ops(obd->obd_fsops);
2565
2566         filter_iobuf_pool_done(filter);
2567
2568         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2569
2570         RETURN(0);
2571 }
2572
2573 static int filter_connect_internal(struct obd_export *exp,
2574                                    struct obd_connect_data *data)
2575 {
2576         struct filter_export_data *fed = &exp->exp_filter_data;
2577
2578         if (!data)
2579                 RETURN(0);
2580
2581         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2582                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2583                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2584                data->ocd_connect_flags, data->ocd_version,
2585                data->ocd_grant, data->ocd_index);
2586
2587         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2588                 CWARN("!!! This export (nid %s) used object group %d "
2589                        "earlier; now it's trying to use group %d!  This could "
2590                        "be a bug in the MDS. Please report to "
2591                        "http://bugzilla.lustre.org/\n",
2592                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2593                 RETURN(-EPROTO);
2594         }
2595         fed->fed_group = data->ocd_group;
2596
2597         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2598         exp->exp_connect_flags = data->ocd_connect_flags;
2599         data->ocd_version = LUSTRE_VERSION_CODE;
2600
2601         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2602                 obd_size left, want;
2603
2604                 spin_lock(&exp->exp_obd->obd_osfs_lock);
2605                 left = filter_grant_space_left(exp);
2606                 want = data->ocd_grant;
2607                 filter_grant(exp, fed->fed_grant, want, left);
2608                 data->ocd_grant = fed->fed_grant;
2609                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2610
2611                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2612                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2613                        exp->exp_client_uuid.uuid, exp,
2614                        data->ocd_grant, want, left);
2615         }
2616
2617         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2618                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2619                 struct lr_server_data *lsd = filter->fo_fsd;
2620                 int index = le32_to_cpu(lsd->lsd_ost_index);
2621
2622                 if (!(lsd->lsd_feature_compat &
2623                       cpu_to_le32(OBD_COMPAT_OST))) {
2624                         /* this will only happen on the first connect */
2625                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2626                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2627                         filter_update_server_data(exp->exp_obd,
2628                                                   filter->fo_rcvd_filp, lsd, 1);
2629                 } else if (index != data->ocd_index) {
2630                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2631                                            " %u doesn't match actual OST index"
2632                                            " %u in last_rcvd file, bad "
2633                                            "configuration?\n",
2634                                            obd_export_nid2str(exp), index,
2635                                            data->ocd_index);
2636                         RETURN(-EBADF);
2637                 }
2638         }
2639
2640         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2641                 data->ocd_brw_size = 65536;
2642         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2643                 data->ocd_brw_size = min(data->ocd_brw_size,
2644                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2645                 LASSERT(data->ocd_brw_size);
2646         }
2647
2648         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2649                 __u32 cksum_types = data->ocd_cksum_types;
2650
2651                 /* The client set in ocd_cksum_types the checksum types it
2652                  * supports. We have to mask off the algorithms that we don't
2653                  * support */
2654                 if (cksum_types & OBD_CKSUM_ALL)
2655                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2656                 else
2657                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2658
2659                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2660                                    "%x\n", exp->exp_obd->obd_name,
2661                                    obd_export_nid2str(exp), cksum_types,
2662                                    data->ocd_cksum_types);
2663         } else {
2664                 /* This client does not support OBD_CONNECT_CKSUM
2665                  * fall back to CRC32 */
2666                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2667                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2668                                    exp->exp_obd->obd_name,
2669                                    obd_export_nid2str(exp));
2670         }
2671
2672         /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2673          * FIXME: We don't strictly need the COMPAT flag for that,
2674          * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2675          * FIXME: We needed it for the index, as index 0 is valid. */
2676
2677         RETURN(0);
2678 }
2679
2680 static int filter_reconnect(const struct lu_env *env,
2681                             struct obd_export *exp, struct obd_device *obd,
2682                             struct obd_uuid *cluuid,
2683                             struct obd_connect_data *data)
2684 {
2685         int rc;
2686         ENTRY;
2687
2688         if (exp == NULL || obd == NULL || cluuid == NULL)
2689                 RETURN(-EINVAL);
2690
2691         rc = filter_connect_internal(exp, data);
2692
2693         RETURN(rc);
2694 }
2695
2696 /* nearly identical to mds_connect */
2697 static int filter_connect(const struct lu_env *env,
2698                           struct lustre_handle *conn, struct obd_device *obd,
2699                           struct obd_uuid *cluuid,
2700                           struct obd_connect_data *data, void *localdata)
2701 {
2702         struct lvfs_run_ctxt saved;
2703         struct obd_export *exp;
2704         struct filter_export_data *fed;
2705         struct lsd_client_data *lcd = NULL;
2706         __u32 group;
2707         int rc;
2708         ENTRY;
2709
2710         if (conn == NULL || obd == NULL || cluuid == NULL)
2711                 RETURN(-EINVAL);
2712
2713         rc = class_connect(conn, obd, cluuid);
2714         if (rc)
2715                 RETURN(rc);
2716         exp = class_conn2export(conn);
2717         LASSERT(exp != NULL);
2718
2719         fed = &exp->exp_filter_data;
2720
2721         rc = filter_connect_internal(exp, data);
2722         if (rc)
2723                 GOTO(cleanup, rc);
2724
2725         filter_export_stats_init(obd, exp, localdata);
2726         if (obd->obd_replayable) {
2727                 OBD_ALLOC(lcd, sizeof(*lcd));
2728                 if (!lcd) {
2729                         CERROR("filter: out of memory for client data\n");
2730                         GOTO(cleanup, rc = -ENOMEM);
2731                 }
2732
2733                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2734                 fed->fed_lcd = lcd;
2735                 rc = filter_client_add(obd, exp, -1);
2736                 if (rc)
2737                         GOTO(cleanup, rc);
2738         }
2739
2740         group = data->ocd_group;
2741         if (group == 0)
2742                 GOTO(cleanup, rc);
2743
2744         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2745               obd->obd_name, exp->exp_handle.h_cookie, group);
2746
2747         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2748         rc = filter_read_groups(obd, group, 1);
2749         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2750         if (rc != 0) {
2751                 CERROR("can't read group %u\n", group);
2752                 GOTO(cleanup, rc);
2753         }
2754
2755         GOTO(cleanup, rc);
2756
2757 cleanup:
2758         if (rc) {
2759                 if (lcd) {
2760                         OBD_FREE_PTR(lcd);
2761                         fed->fed_lcd = NULL;
2762                 }
2763                 class_disconnect(exp);
2764         } else {
2765                 class_export_put(exp);
2766         }
2767
2768         RETURN(rc);
2769 }
2770
2771 /* Do extra sanity checks for grant accounting.  We do this at connect,
2772  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2773  * always get rid of it or turn it off when we know accounting is good. */
2774 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2775 {
2776         struct filter_export_data *fed;
2777         struct obd_export *exp;
2778         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2779         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2780         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2781
2782         if (list_empty(&obd->obd_exports))
2783                 return;
2784
2785         /* We don't want to do this for large machines that do lots of
2786            mounts or unmounts.  It burns... */
2787         if (obd->obd_num_exports > 100)
2788                 return;
2789
2790         spin_lock(&obd->obd_osfs_lock);
2791         spin_lock(&obd->obd_dev_lock);
2792         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2793                 int error = 0;
2794                 fed = &exp->exp_filter_data;
2795                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2796                     fed->fed_dirty < 0)
2797                         error = 1;
2798                 if (maxsize > 0) { /* we may not have done a statfs yet */
2799                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2800                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2801                                  exp->exp_client_uuid.uuid, exp,
2802                                  fed->fed_grant, fed->fed_pending, maxsize);
2803                         LASSERTF(fed->fed_dirty <= maxsize,
2804                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2805                                  exp->exp_client_uuid.uuid, exp,
2806                                  fed->fed_dirty, maxsize);
2807                 }
2808                 if (error)
2809                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2810                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2811                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2812                 else
2813                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2814                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2815                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2816                 tot_granted += fed->fed_grant + fed->fed_pending;
2817                 tot_pending += fed->fed_pending;
2818                 tot_dirty += fed->fed_dirty;
2819         }
2820         fo_tot_granted = obd->u.filter.fo_tot_granted;
2821         fo_tot_pending = obd->u.filter.fo_tot_pending;
2822         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2823         spin_unlock(&obd->obd_dev_lock);
2824         spin_unlock(&obd->obd_osfs_lock);
2825
2826         /* Do these assertions outside the spinlocks so we don't kill system */
2827         if (tot_granted != fo_tot_granted)
2828                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2829                        func, tot_granted, fo_tot_granted);
2830         if (tot_pending != fo_tot_pending)
2831                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2832                        func, tot_pending, fo_tot_pending);
2833         if (tot_dirty != fo_tot_dirty)
2834                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2835                        func, tot_dirty, fo_tot_dirty);
2836         if (tot_pending > tot_granted)
2837                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2838                        func, tot_pending, tot_granted);
2839         if (tot_granted > maxsize)
2840                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2841                        func, tot_granted, maxsize);
2842         if (tot_dirty > maxsize)
2843                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2844                        func, tot_dirty, maxsize);
2845 }
2846
2847 /* Remove this client from the grant accounting totals.  We also remove
2848  * the export from the obd device under the osfs and dev locks to ensure
2849  * that the filter_grant_sanity_check() calculations are always valid.
2850  * The client should do something similar when it invalidates its import. */
2851 static void filter_grant_discard(struct obd_export *exp)
2852 {
2853         struct obd_device *obd = exp->exp_obd;
2854         struct filter_obd *filter = &obd->u.filter;
2855         struct filter_export_data *fed = &exp->exp_filter_data;
2856
2857         spin_lock(&obd->obd_osfs_lock);
2858         spin_lock(&obd->obd_dev_lock);
2859         list_del_init(&exp->exp_obd_chain);
2860         spin_unlock(&obd->obd_dev_lock);
2861
2862         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2863                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2864                  obd->obd_name, filter->fo_tot_granted,
2865                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2866         filter->fo_tot_granted -= fed->fed_grant;
2867         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2868                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2869                  obd->obd_name, filter->fo_tot_pending,
2870                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2871         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2872         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2873                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2874                  obd->obd_name, filter->fo_tot_dirty,
2875                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2876         filter->fo_tot_dirty -= fed->fed_dirty;
2877         fed->fed_dirty = 0;
2878         fed->fed_grant = 0;
2879
2880         spin_unlock(&obd->obd_osfs_lock);
2881 }
2882
2883 static int filter_destroy_export(struct obd_export *exp)
2884 {
2885         ENTRY;
2886
2887         if (exp->exp_filter_data.fed_pending)
2888                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2889                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2890                        exp, exp->exp_filter_data.fed_pending);
2891
2892         /* Not ported yet the b1_6 quota functionality
2893          * lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2894          */
2895
2896         target_destroy_export(exp);
2897         ldlm_destroy_export(exp);
2898
2899         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2900                 RETURN(0);
2901
2902
2903         if (exp->exp_obd->obd_replayable)
2904                 filter_client_free(exp);
2905         else
2906                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2907
2908         filter_grant_discard(exp);
2909         filter_fmd_cleanup(exp);
2910
2911         if (!(exp->exp_flags & OBD_OPT_FORCE))
2912                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2913
2914         RETURN(0);
2915 }
2916
2917 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2918 {
2919         struct obd_llog_group *olg_min, *olg;
2920         struct filter_obd *filter;
2921         int worked = 0, group;
2922         struct llog_ctxt *ctxt;
2923         ENTRY;
2924
2925         filter = &obd->u.filter;
2926
2927         /* we can't sync log holding spinlock. also, we do not want to get
2928          * into livelock. so we do following: loop over MDS's exports in
2929          * group order and skip already synced llogs -bzzz */
2930         do {
2931                 /* look for group with min. number, but > worked */
2932                 olg_min = NULL;
2933                 group = 1 << 30;
2934                 spin_lock(&filter->fo_llog_list_lock);
2935                 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2936                         if (olg->olg_group <= worked) {
2937                                 /* this group is already synced */
2938                                 continue;
2939                         }
2940                         if (group < olg->olg_group) {
2941                                 /* we have group with smaller number to sync */
2942                                 continue;
2943                         }
2944                         /* store current minimal group */
2945                         olg_min = olg;
2946                         group = olg->olg_group;
2947                 }
2948                 spin_unlock(&filter->fo_llog_list_lock);
2949
2950                 if (olg_min == NULL)
2951                         break;
2952
2953                 worked = olg_min->olg_group;
2954                 if (olg_min->olg_exp &&
2955                     (dexp == olg_min->olg_exp || dexp == NULL)) {
2956                         int err;
2957                         ctxt = llog_group_get_ctxt(olg_min,
2958                                                 LLOG_MDS_OST_REPL_CTXT);
2959                         LASSERT(ctxt != NULL);
2960                         err = llog_sync(ctxt, olg_min->olg_exp);
2961                         llog_ctxt_put(ctxt);
2962                         if (err)
2963                                 CERROR("error flushing logs to MDS: rc %d\n",
2964                                        err);
2965                 }
2966         } while (olg_min != NULL);
2967 }
2968
2969 /* also incredibly similar to mds_disconnect */
2970 static int filter_disconnect(struct obd_export *exp)
2971 {
2972         struct obd_device *obd = exp->exp_obd;
2973         int rc;
2974         ENTRY;
2975
2976         LASSERT(exp);
2977         class_export_get(exp);
2978
2979         if (!(exp->exp_flags & OBD_OPT_FORCE))
2980                 filter_grant_sanity_check(obd, __FUNCTION__);
2981         filter_grant_discard(exp);
2982
2983         /* Disconnect early so that clients can't keep using export */
2984         rc = class_disconnect(exp);
2985         if (exp->exp_obd->obd_namespace != NULL)
2986                 ldlm_cancel_locks_for_export(exp);
2987
2988         fsfilt_sync(obd, obd->u.obt.obt_sb);
2989
2990         lprocfs_exp_cleanup(exp);
2991
2992         /* flush any remaining cancel messages out to the target */
2993         filter_sync_llogs(obd, exp);
2994         class_export_put(exp);
2995         RETURN(rc);
2996 }
2997
2998 /* reverse import is changed, sync all cancels */
2999 static void filter_revimp_update(struct obd_export *exp)
3000 {
3001         ENTRY;
3002
3003         LASSERT(exp);
3004         class_export_get(exp);
3005
3006         /* flush any remaining cancel messages out to the target */
3007         filter_sync_llogs(exp->exp_obd, exp);
3008         class_export_put(exp);
3009         EXIT;
3010 }
3011
3012 static int filter_ping(struct obd_export *exp)
3013 {
3014         filter_fmd_expire(exp);
3015
3016         return 0;
3017 }
3018
3019 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3020                                   const char *what, int quiet)
3021 {
3022         struct dentry *dchild = NULL;
3023         obd_gr group = 0;
3024
3025         if (oa->o_valid & OBD_MD_FLGROUP)
3026                 group = oa->o_gr;
3027
3028         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3029
3030         if (IS_ERR(dchild)) {
3031                 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3032                        what, group, oa->o_id);
3033                 RETURN(dchild);
3034         }
3035
3036         if (dchild->d_inode == NULL) {
3037                 if (!quiet)
3038                         CERROR("%s: %s on non-existent object: "LPU64"\n",
3039                                obd->obd_name, what, oa->o_id);
3040                 f_dput(dchild);
3041                 RETURN(ERR_PTR(-ENOENT));
3042         }
3043
3044         return dchild;
3045 }
3046
3047 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3048 {
3049         struct dentry *dentry = NULL;
3050         struct obd_device *obd;
3051         int rc = 0;
3052         ENTRY;
3053
3054         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3055                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3056         if (rc)
3057                 RETURN(rc);
3058
3059         obd = class_exp2obd(exp);
3060         if (obd == NULL) {
3061                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3062                 RETURN(-EINVAL);
3063         }
3064
3065         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3066         if (IS_ERR(dentry))
3067                 RETURN(PTR_ERR(dentry));
3068
3069         /* Limit the valid bits in the return data to what we actually use */
3070         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3071         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3072
3073         f_dput(dentry);
3074         RETURN(rc);
3075 }
3076
3077 /* this should be enabled/disabled in condition to enabled/disabled large
3078  * inodes (fast EAs) in backing store FS. */
3079 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3080                         void *handle, struct obdo *oa)
3081 {
3082         struct obd_device *obd = exp->exp_obd;
3083         int rc = 0;
3084         ENTRY;
3085
3086         if (oa->o_valid & OBD_MD_FLFID) {
3087                 struct filter_fid ff;
3088
3089                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3090                         oa->o_gr = 0;
3091                 /* packing fid and converting it to LE for storing into EA.
3092                  * Here ->o_stripe_idx should be filled by LOV and rest of
3093                  * fields - by client. */
3094                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3095                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3096                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3097                 ff.ff_objid = cpu_to_le64(oa->o_id);
3098                 ff.ff_group = cpu_to_le64(oa->o_gr);
3099
3100                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3101                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3102                        oa->o_generation, oa->o_id, oa->o_gr);
3103
3104                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3105                 if (rc)
3106                         CERROR("store fid in object failed! rc: %d\n", rc);
3107         } else {
3108                 CDEBUG(D_HA, "OSS object without fid info!\n");
3109         }
3110
3111         RETURN(rc);
3112 }
3113
3114 /* this is called from filter_truncate() until we have filter_punch() */
3115 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3116                             struct obdo *oa, struct obd_trans_info *oti)
3117 {
3118         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3119         struct llog_cookie *fcc = NULL;
3120         struct filter_obd *filter;
3121         int rc, err, locked = 0, sync = 0;
3122         unsigned int ia_valid;
3123         struct inode *inode;
3124         struct iattr iattr;
3125         void *handle;
3126         ENTRY;
3127
3128         LASSERT(dentry != NULL);
3129         LASSERT(!IS_ERR(dentry));
3130
3131         inode = dentry->d_inode;
3132         LASSERT(inode != NULL);
3133
3134         filter = &exp->exp_obd->u.filter;
3135         iattr_from_obdo(&iattr, oa, oa->o_valid);
3136         ia_valid = iattr.ia_valid;
3137
3138         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3139                 OBD_ALLOC(fcc, sizeof(*fcc));
3140                 if (fcc != NULL)
3141                         *fcc = oa->o_lcookie;
3142         }
3143
3144         if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
3145                 DQUOT_INIT(inode);
3146                 LOCK_INODE_MUTEX(inode);
3147                 locked = 1;
3148         }
3149
3150         /* If the inode still has SUID+SGID bits set (see filter_precreate())
3151          * then we will accept the UID+GID sent by the client during write for
3152          * initializing the ownership of this inode.  We only allow this to
3153          * happen once so clear these bits in setattr. In 2.6 kernels it is
3154          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3155          * the flags that are actually being set. */
3156         if (ia_valid & (ATTR_UID | ATTR_GID)) {
3157                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3158                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3159
3160                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3161                         if (!(ia_valid & ATTR_MODE)) {
3162                                 iattr.ia_mode = inode->i_mode;
3163                                 iattr.ia_valid |= ATTR_MODE;
3164                         }
3165                         iattr.ia_mode &= ~S_ISUID;
3166                 }
3167                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3168                         if (!(iattr.ia_valid & ATTR_MODE)) {
3169                                 iattr.ia_mode = inode->i_mode;
3170                                 iattr.ia_valid |= ATTR_MODE;
3171                         }
3172                         iattr.ia_mode &= ~S_ISGID;
3173                 }
3174
3175                 orig_ids[USRQUOTA] = inode->i_uid;
3176                 orig_ids[GRPQUOTA] = inode->i_gid;
3177                 handle = fsfilt_start_log(exp->exp_obd, inode,
3178                                           FSFILT_OP_SETATTR, oti, 1);
3179                 if (IS_ERR(handle))
3180                         GOTO(out_unlock, rc = PTR_ERR(handle));
3181
3182                 /* update inode EA only once when inode is suid bit marked. As
3183                  * on 2.6.x UID and GID may be set separately, we check here
3184                  * only one of them to avoid double setting. */
3185                 if (inode->i_mode & S_ISUID)
3186                         filter_update_fidea(exp, inode, handle, oa);
3187         } else {
3188                 handle = fsfilt_start(exp->exp_obd, inode,
3189                                       FSFILT_OP_SETATTR, oti);
3190                 if (IS_ERR(handle))
3191                         GOTO(out_unlock, rc = PTR_ERR(handle));
3192         }
3193         if (oa->o_valid & OBD_MD_FLFLAGS) {
3194                 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3195                                       EXT3_IOC_SETFLAGS, (long)&oa->o_flags);