Whamcloud - gitweb
7fabcb42206043ecb8f0c642ebe4d60cd9e44079
[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_MDS1_N_BASE
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         struct obd_export *exp = cb_data;
91         LASSERT(exp->exp_obd == obd);
92         obd_transno_commit_cb(obd, transno, exp, error);
93         atomic_dec(&exp->exp_cb_count);
94         class_export_put(exp);
95 }
96
97 int filter_version_get_check(struct obd_export *exp,
98                              struct obd_trans_info *oti, struct inode *inode)
99 {
100         __u64 curr_version;
101
102         if (inode == NULL || oti == NULL)
103                 RETURN(0);
104
105         curr_version = fsfilt_get_version(exp->exp_obd, inode);
106         if ((__s64)curr_version == -EOPNOTSUPP)
107                 RETURN(0);
108         /* VBR: version is checked always because costs nothing */
109         if (oti->oti_pre_version != 0 &&
110             oti->oti_pre_version != curr_version) {
111                 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
112                        oti->oti_pre_version, curr_version);
113                 spin_lock(&exp->exp_lock);
114                 exp->exp_vbr_failed = 1;
115                 spin_unlock(&exp->exp_lock);
116                 RETURN (-EOVERFLOW);
117         }
118         oti->oti_pre_version = curr_version;
119         RETURN(0);
120 }
121
122 /* Assumes caller has already pushed us into the kernel context. */
123 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
124                           struct obd_trans_info *oti, int rc, int force_sync)
125 {
126         struct filter_obd *filter = &exp->exp_obd->u.filter;
127         struct filter_export_data *fed = &exp->exp_filter_data;
128         struct lsd_client_data *lcd = fed->fed_lcd;
129         __u64 last_rcvd;
130         loff_t off;
131         int err, log_pri = D_RPCTRACE;
132
133         /* Propagate error code. */
134         if (rc)
135                 RETURN(rc);
136
137         if (!exp->exp_obd->obd_replayable || oti == NULL)
138                 RETURN(rc);
139
140         /* we don't allocate new transnos for replayed requests */
141         spin_lock(&filter->fo_translock);
142         if (oti->oti_transno == 0) {
143                 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
144                 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
145         } else {
146                 last_rcvd = oti->oti_transno;
147                 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
148                         filter->fo_fsd->lsd_last_transno =
149                                 cpu_to_le64(last_rcvd);
150         }
151         oti->oti_transno = last_rcvd;
152         if (last_rcvd <= le64_to_cpu(lcd->lcd_last_transno)) {
153                 spin_unlock(&filter->fo_translock);
154                 LBUG();
155         }
156         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
157         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
158         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
159         spin_unlock(&filter->fo_translock);
160
161         if (inode)
162                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
163
164         off = fed->fed_lr_off;
165         if (off <= 0) {
166                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
167                        fed->fed_lr_idx, fed->fed_lr_off);
168                 err = -EINVAL;
169         } else {
170                 class_export_get(exp); /* released when the cb is called */
171                 atomic_inc(&exp->exp_cb_count);
172                 if (!force_sync)
173                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
174                                                            last_rcvd,
175                                                            oti->oti_handle,
176                                                            filter_commit_cb,
177                                                            exp);
178
179                 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
180                                           lcd, sizeof(*lcd), &off,
181                                           force_sync | exp->exp_need_sync);
182                 if (force_sync)
183                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
184         }
185         if (err) {
186                 log_pri = D_ERROR;
187                 if (rc == 0)
188                         rc = err;
189         }
190
191         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
192                last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
193
194         RETURN(rc);
195 }
196
197 void f_dput(struct dentry *dentry)
198 {
199         /* Can't go inside filter_ddelete because it can block */
200         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
201                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
202         LASSERT(atomic_read(&dentry->d_count) > 0);
203
204         dput(dentry);
205 }
206
207 static void init_brw_stats(struct brw_stats *brw_stats)
208 {
209         int i;
210         for (i = 0; i < BRW_LAST; i++)
211                 spin_lock_init(&brw_stats->hist[i].oh_lock);
212 }
213
214 static int lprocfs_init_rw_stats(struct obd_device *obd,
215                                  struct lprocfs_stats **stats)
216 {
217         int num_stats;
218
219         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
220                                                         LPROC_FILTER_LAST - 1;
221         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
222         if (*stats == NULL)
223                 return -ENOMEM;
224
225         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
226         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
227                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
228         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
229                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
230
231         return(0);
232 }
233
234 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
235    plus the procfs overhead :( */
236 static int filter_export_stats_init(struct obd_device *obd,
237                                     struct obd_export *exp,
238                                     void *client_nid)
239 {
240         int rc, newnid = 0;
241         ENTRY;
242
243         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
244                 /* Self-export gets no proc entry */
245                 RETURN(0);
246
247         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
248         if (rc) {
249                 /* Mask error for already created
250                  * /proc entries */
251                 if (rc == -EALREADY)
252                         rc = 0;
253                 RETURN(rc);
254         }
255
256         if (newnid) {
257                 struct nid_stat *tmp = exp->exp_nid_stats;
258                 LASSERT(tmp != NULL);
259
260                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
261                 if (tmp->nid_brw_stats == NULL)
262                         RETURN(-ENOMEM);
263
264                 init_brw_stats(tmp->nid_brw_stats);
265                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
266                                         0644, &filter_per_nid_stats_fops,
267                                         exp->exp_nid_stats);
268                 if (rc)
269                         CWARN("Error adding the brw_stats file\n");
270
271                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
272                 if (rc)
273                         RETURN(rc);
274
275                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
276                                             tmp->nid_stats);
277                 if (rc)
278                         RETURN(rc);
279                 /* Always add in ldlm_stats */
280                 tmp->nid_ldlm_stats = 
281                         lprocfs_alloc_stats(LDLM_LAST_OPC - LDLM_FIRST_OPC,
282                                             LPROCFS_STATS_FLAG_NOPERCPU);
283                 if (tmp->nid_ldlm_stats == NULL)
284                         return -ENOMEM;
285
286                 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
287                 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
288                                             tmp->nid_ldlm_stats);
289                 if (rc)
290                         RETURN(rc);
291         }
292
293         RETURN(0);
294 }
295
296 /* Add client data to the FILTER.  We use a bitmap to locate a free space
297  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
298  * Otherwise, we have just read the data from the last_rcvd file and
299  * we know its offset. */
300 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
301                              int cl_idx)
302 {
303         struct filter_obd *filter = &obd->u.filter;
304         struct filter_export_data *fed = &exp->exp_filter_data;
305         unsigned long *bitmap = filter->fo_last_rcvd_slots;
306         int new_client = (cl_idx == -1);
307
308         ENTRY;
309
310         LASSERT(bitmap != NULL);
311         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
312
313         /* Self-export */
314         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
315                 RETURN(0);
316
317         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
318          * there's no need for extra complication here
319          */
320         if (new_client) {
321                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
322         repeat:
323                 if (cl_idx >= LR_MAX_CLIENTS) {
324                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
325                                cl_idx);
326                         RETURN(-EOVERFLOW);
327                 }
328                 if (test_and_set_bit(cl_idx, bitmap)) {
329                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
330                                                     cl_idx);
331                         goto repeat;
332                 }
333         } else {
334                 if (test_and_set_bit(cl_idx, bitmap)) {
335                         CERROR("FILTER client %d: bit already set in bitmap!\n",
336                                cl_idx);
337                         LBUG();
338                 }
339         }
340
341         fed->fed_lr_idx = cl_idx;
342         fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
343                 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
344         LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
345
346         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
347                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
348
349         if (new_client) {
350                 struct lvfs_run_ctxt saved;
351                 loff_t off = fed->fed_lr_off;
352                 int rc;
353                 void *handle;
354
355                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
356                        fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
357
358                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
359                 /* Transaction needed to fix bug 1403 */
360                 handle = fsfilt_start(obd,
361                                       filter->fo_rcvd_filp->f_dentry->d_inode,
362                                       FSFILT_OP_SETATTR, NULL);
363                 if (IS_ERR(handle)) {
364                         rc = PTR_ERR(handle);
365                         CERROR("unable to start transaction: rc %d\n", rc);
366                 } else {
367                         fed->fed_lcd->lcd_last_epoch =
368                                               filter->fo_fsd->lsd_start_epoch;
369                         exp->exp_last_request_time = cfs_time_current_sec();
370                         rc = fsfilt_add_journal_cb(obd, 0, handle,
371                                                    target_client_add_cb, exp);
372                         if (rc == 0) {
373                                 spin_lock(&exp->exp_lock);
374                                 exp->exp_need_sync = 1;
375                                 spin_unlock(&exp->exp_lock);
376                         }
377                         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
378                                                  fed->fed_lcd,
379                                                  sizeof(*fed->fed_lcd),
380                                                  &off, rc /* sync if no cb */);
381                         fsfilt_commit(obd,
382                                       filter->fo_rcvd_filp->f_dentry->d_inode,
383                                       handle, 0);
384                 }
385                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
386
387                 if (rc) {
388                         CERROR("error writing %s client idx %u: rc %d\n",
389                                LAST_RCVD, fed->fed_lr_idx, rc);
390                         RETURN(rc);
391                 }
392         }
393         RETURN(0);
394 }
395
396 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
397
398 static int filter_client_free(struct obd_export *exp)
399 {
400         struct filter_export_data *fed = &exp->exp_filter_data;
401         struct filter_obd *filter = &exp->exp_obd->u.filter;
402         struct obd_device *obd = exp->exp_obd;
403         struct lvfs_run_ctxt saved;
404         int rc;
405         loff_t off;
406         ENTRY;
407
408         if (fed->fed_lcd == NULL)
409                 RETURN(0);
410
411         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
412         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
413                 GOTO(free, 0);
414
415         LASSERT(filter->fo_last_rcvd_slots != NULL);
416
417         off = fed->fed_lr_off;
418
419         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
420                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
421
422         /* Don't clear fed_lr_idx here as it is likely also unset.  At worst
423          * we leak a client slot that will be cleaned on the next recovery. */
424         if (off <= 0) {
425                 CERROR("%s: client idx %d has med_off %lld\n",
426                        obd->obd_name, fed->fed_lr_idx, off);
427                 GOTO(free, rc = -EINVAL);
428         }
429
430         /* Clear the bit _after_ zeroing out the client so we don't
431            race with filter_client_add and zero out new clients.*/
432         if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
433                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
434                        fed->fed_lr_idx);
435                 LBUG();
436         }
437
438         if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
439                 /* Don't force sync on disconnect if aborting recovery,
440                  * or it does num_clients * num_osts.  b=17194 */
441                 int need_sync = exp->exp_need_sync &&
442                                  !(exp->exp_flags&OBD_OPT_ABORT_RECOV);
443                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
444                 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
445                                          sizeof(zero_lcd), &off, 0);
446
447                 /* Make sure the server's last_transno is up to date. Do this
448                  * after the client is freed so we know all the client's
449                  * transactions have been committed. */
450                 if (rc == 0)
451                         filter_update_server_data(obd, filter->fo_rcvd_filp,
452                                                   filter->fo_fsd, need_sync);
453                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
454
455                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
456                        "zero out client %s at idx %u/%llu in %s %ssync rc %d\n",
457                        fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
458                        LAST_RCVD, need_sync ? "" : "a", rc);
459         }
460
461         if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
462                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
463                        fed->fed_lr_idx);
464                 LBUG();
465         }
466
467         EXIT;
468 free:
469         OBD_FREE_PTR(fed->fed_lcd);
470         fed->fed_lcd = NULL;
471
472         return 0;
473 }
474
475 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
476 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
477                                          struct filter_mod_data *fmd)
478 {
479         LASSERT_SPIN_LOCKED(&fed->fed_lock);
480         if (--fmd->fmd_refcount == 0) {
481                 /* XXX when we have persistent reservations and the handle
482                  * is stored herein we need to drop it here. */
483                 fed->fed_mod_count--;
484                 list_del(&fmd->fmd_list);
485                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
486         }
487 }
488
489 /* drop fmd reference, free it if last ref */
490 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
491 {
492         struct filter_export_data *fed;
493
494         if (fmd == NULL)
495                 return;
496
497         fed = &exp->exp_filter_data;
498         spin_lock(&fed->fed_lock);
499         filter_fmd_put_nolock(fed, fmd); /* caller reference */
500         spin_unlock(&fed->fed_lock);
501 }
502
503 /* expire entries from the end of the list if there are too many
504  * or they are too old */
505 static void filter_fmd_expire_nolock(struct filter_obd *filter,
506                                      struct filter_export_data *fed,
507                                      struct filter_mod_data *keep)
508 {
509         struct filter_mod_data *fmd, *tmp;
510
511         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
512                 if (fmd == keep)
513                         break;
514
515                 if (time_before(jiffies, fmd->fmd_expire) &&
516                     fed->fed_mod_count < filter->fo_fmd_max_num)
517                         break;
518
519                 list_del_init(&fmd->fmd_list);
520                 filter_fmd_put_nolock(fed, fmd); /* list reference */
521         }
522 }
523
524 void filter_fmd_expire(struct obd_export *exp)
525 {
526         spin_lock(&exp->exp_filter_data.fed_lock);
527         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
528                                  &exp->exp_filter_data, NULL);
529         spin_unlock(&exp->exp_filter_data.fed_lock);
530 }
531
532 /* find specified objid, group in export fmd list.
533  * caller must hold fed_lock and take fmd reference itself */
534 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
535                                                 struct filter_export_data *fed,
536                                                 obd_id objid, obd_gr group)
537 {
538         struct filter_mod_data *found = NULL, *fmd;
539
540         LASSERT_SPIN_LOCKED(&fed->fed_lock);
541
542         list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
543                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
544                         found = fmd;
545                         list_del(&fmd->fmd_list);
546                         list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
547                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
548                         break;
549                 }
550         }
551
552         filter_fmd_expire_nolock(filter, fed, found);
553
554         return found;
555 }
556
557 /* Find fmd based on objid and group, or return NULL if not found. */
558 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
559                                         obd_id objid, obd_gr group)
560 {
561         struct filter_mod_data *fmd;
562
563         spin_lock(&exp->exp_filter_data.fed_lock);
564         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
565                                      &exp->exp_filter_data, objid, group);
566         if (fmd)
567                 fmd->fmd_refcount++;    /* caller reference */
568         spin_unlock(&exp->exp_filter_data.fed_lock);
569
570         return fmd;
571 }
572
573 /* Find fmd based on objid and group, or create a new one if none is found.
574  * It is possible for this function to return NULL under memory pressure,
575  * or if objid = 0 is passed (which will only cause old entries to expire).
576  * Currently this is not fatal because any fmd state is transient and
577  * may also be freed when it gets sufficiently old. */
578 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
579                                        obd_id objid, obd_gr group)
580 {
581         struct filter_export_data *fed = &exp->exp_filter_data;
582         struct filter_mod_data *found = NULL, *fmd_new = NULL;
583
584         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
585
586         spin_lock(&fed->fed_lock);
587         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
588         if (fmd_new) {
589                 if (found == NULL) {
590                         list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
591                         fmd_new->fmd_id = objid;
592                         fmd_new->fmd_gr = group;
593                         fmd_new->fmd_refcount++;   /* list reference */
594                         found = fmd_new;
595                         fed->fed_mod_count++;
596                 } else {
597                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
598                 }
599         }
600         if (found) {
601                 found->fmd_refcount++;          /* caller reference */
602                 found->fmd_expire = jiffies +
603                         exp->exp_obd->u.filter.fo_fmd_max_age;
604         }
605
606         spin_unlock(&fed->fed_lock);
607
608         return found;
609 }
610
611 #ifdef DO_FMD_DROP
612 /* drop fmd list reference so it will disappear when last reference is put.
613  * This isn't so critical because it would in fact only affect the one client
614  * that is doing the unlink and at worst we have an stale entry referencing
615  * an object that should never be used again. */
616 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
617 {
618         struct filter_mod_data *found = NULL;
619
620         spin_lock(&exp->exp_filter_data.fed_lock);
621         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
622         if (found) {
623                 list_del_init(&found->fmd_list);
624                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
625         }
626         spin_unlock(&exp->exp_filter_data.fed_lock);
627 }
628 #else
629 #define filter_fmd_drop(exp, objid, group)
630 #endif
631
632 /* remove all entries from fmd list */
633 static void filter_fmd_cleanup(struct obd_export *exp)
634 {
635         struct filter_export_data *fed = &exp->exp_filter_data;
636         struct filter_mod_data *fmd = NULL, *tmp;
637
638         spin_lock(&fed->fed_lock);
639         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
640                 list_del_init(&fmd->fmd_list);
641                 filter_fmd_put_nolock(fed, fmd);
642         }
643         spin_unlock(&fed->fed_lock);
644 }
645
646 static int filter_init_export(struct obd_export *exp)
647 {
648         spin_lock_init(&exp->exp_filter_data.fed_lock);
649         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
650
651         spin_lock(&exp->exp_lock);
652         exp->exp_connecting = 1;
653         spin_unlock(&exp->exp_lock);
654
655         return ldlm_init_export(exp);
656 }
657
658 static int filter_free_server_data(struct filter_obd *filter)
659 {
660         OBD_FREE_PTR(filter->fo_fsd);
661         filter->fo_fsd = NULL;
662         OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
663         filter->fo_last_rcvd_slots = NULL;
664         return 0;
665 }
666
667 /* assumes caller is already in kernel ctxt */
668 int filter_update_server_data(struct obd_device *obd, struct file *filp,
669                               struct lr_server_data *fsd, int force_sync)
670 {
671         loff_t off = 0;
672         int rc;
673         ENTRY;
674
675         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
676         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
677                le64_to_cpu(fsd->lsd_last_transno));
678         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
679                le64_to_cpu(fsd->lsd_mount_count));
680
681         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
682         if (rc)
683                 CERROR("error writing lr_server_data: rc = %d\n", rc);
684
685         RETURN(rc);
686 }
687
688 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
689                              int force_sync)
690 {
691         struct filter_obd *filter = &obd->u.filter;
692         __u64 tmp;
693         loff_t off = 0;
694         int rc;
695         ENTRY;
696
697         if (filter->fo_last_objid_files[group] == NULL) {
698                 CERROR("Object group "LPU64" not fully setup; not updating "
699                        "last_objid\n", group);
700                 RETURN(-EINVAL);
701         }
702
703         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
704                obd->obd_name, group, filter->fo_last_objids[group]);
705
706         tmp = cpu_to_le64(filter->fo_last_objids[group]);
707         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
708                                  &tmp, sizeof(tmp), &off, force_sync);
709         if (rc)
710                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
711                        group, rc);
712         RETURN(rc);
713 }
714 extern int ost_handle(struct ptlrpc_request *req);
715 /* assumes caller has already in kernel ctxt */
716 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
717 {
718         struct filter_obd *filter = &obd->u.filter;
719         struct lr_server_data *fsd;
720         struct lsd_client_data *lcd = NULL;
721         struct inode *inode = filp->f_dentry->d_inode;
722         unsigned long last_rcvd_size = i_size_read(inode);
723         __u64 mount_count;
724         __u32 start_epoch;
725         int cl_idx;
726         loff_t off = 0;
727         int rc;
728
729         /* ensure padding in the struct is the correct size */
730         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
731                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
732         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
733                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
734
735         OBD_ALLOC(fsd, sizeof(*fsd));
736         if (!fsd)
737                 RETURN(-ENOMEM);
738         filter->fo_fsd = fsd;
739
740         OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
741         if (filter->fo_last_rcvd_slots == NULL) {
742                 OBD_FREE(fsd, sizeof(*fsd));
743                 RETURN(-ENOMEM);
744         }
745
746         if (last_rcvd_size == 0) {
747                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
748
749                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
750                 fsd->lsd_last_transno = 0;
751                 mount_count = fsd->lsd_mount_count = 0;
752                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
753                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
754                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
755                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
756                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
757                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
758                  * MDS first connects and assigns the OST index number. */
759                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
760                                                         OBD_INCOMPAT_OST);
761         } else {
762                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
763                 if (rc) {
764                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
765                                LAST_RCVD, rc);
766                         GOTO(err_fsd, rc);
767                 }
768                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
769                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
770                                            "using the wrong disk %s. Were the "
771                                            "/dev/ assignments rearranged?\n",
772                                            obd->obd_uuid.uuid, fsd->lsd_uuid);
773                         GOTO(err_fsd, rc = -EINVAL);
774                 }
775                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
776                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
777                 /* COMPAT_146 */
778                 /* Assume old last_rcvd format unless I_C_LR is set */
779                 if (!(fsd->lsd_feature_incompat &
780                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
781                         fsd->lsd_last_transno = fsd->lsd_compat14;
782                 /* end COMPAT_146 */
783                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
784                  * MDS first connects and assigns the OST index number. */
785                 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
786                                                          OBD_INCOMPAT_OST);
787         }
788
789         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
790                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
791                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
792                        ~FILTER_INCOMPAT_SUPP);
793                 GOTO(err_fsd, rc = -EINVAL);
794         }
795         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
796                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
797                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
798                        ~FILTER_ROCOMPAT_SUPP);
799                 /* Do something like remount filesystem read-only */
800                 GOTO(err_fsd, rc = -EINVAL);
801         }
802
803         start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
804
805         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
806                obd->obd_name, start_epoch);
807         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
808                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
809         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
810                obd->obd_name, mount_count + 1);
811         CDEBUG(D_INODE, "%s: server data size: %u\n",
812                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
813         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
814                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
815         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
816                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
817         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
818                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
819         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
820                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
821                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
822                 le16_to_cpu(fsd->lsd_client_size));
823
824         if (!obd->obd_replayable) {
825                 CWARN("%s: recovery support OFF\n", obd->obd_name);
826                 GOTO(out, rc = 0);
827         }
828
829         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
830              off < last_rcvd_size; cl_idx++) {
831                 __u64 last_rcvd;
832                 struct obd_export *exp;
833                 struct filter_export_data *fed;
834
835                 if (!lcd) {
836                         OBD_ALLOC_PTR(lcd);
837                         if (!lcd)
838                                 GOTO(err_client, rc = -ENOMEM);
839                 }
840
841                 /* Don't assume off is incremented properly by
842                  * fsfilt_read_record(), in case sizeof(*lcd)
843                  * isn't the same as fsd->lsd_client_size.  */
844                 off = le32_to_cpu(fsd->lsd_client_start) +
845                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
846                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
847                 if (rc) {
848                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
849                                LAST_RCVD, cl_idx, off, rc);
850                         break; /* read error shouldn't cause startup to fail */
851                 }
852
853                 if (lcd->lcd_uuid[0] == '\0') {
854                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
855                                cl_idx);
856                         continue;
857                 }
858
859                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
860
861                 /* These exports are cleaned up by filter_disconnect(), so they
862                  * need to be set up like real exports as filter_connect() does.
863                  */
864                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
865
866                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
867                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
868                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
869                 if (IS_ERR(exp)) {
870                         if (PTR_ERR(exp) == -EALREADY) {
871                                 /* export already exists, zero out this one */
872                                 CERROR("Zeroing out duplicate export due to "
873                                        "bug 10479.\n");
874                                 lcd->lcd_uuid[0] = '\0';
875                         } else {
876                                 GOTO(err_client, rc = PTR_ERR(exp));
877                         }
878                 } else {
879                         fed = &exp->exp_filter_data;
880                         fed->fed_lcd = lcd;
881                         fed->fed_group = 0; /* will be assigned at connect */
882                         filter_export_stats_init(obd, exp, NULL);
883                         rc = filter_client_add(obd, exp, cl_idx);
884                         /* can't fail for existing client */
885                         LASSERTF(rc == 0, "rc = %d\n", rc);
886
887                         /* VBR: set export last committed */
888                         exp->exp_last_committed = last_rcvd;
889                         spin_lock(&exp->exp_lock);
890                         exp->exp_connecting = 0;
891                         exp->exp_in_recovery = 0;
892                         spin_unlock(&exp->exp_lock);
893                         spin_lock_bh(&obd->obd_processing_task_lock);
894                         obd->obd_max_recoverable_clients++;
895                         spin_unlock_bh(&obd->obd_processing_task_lock);
896                         lcd = NULL;
897                         class_export_put(exp);
898                 }
899
900                 /* Need to check last_rcvd even for duplicated exports. */
901                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
902                        cl_idx, last_rcvd);
903
904                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
905                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
906         }
907
908         if (lcd)
909                 OBD_FREE_PTR(lcd);
910
911         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
912 out:
913         filter->fo_mount_count = mount_count + 1;
914         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
915
916         /* save it, so mount count and last_transno is current */
917         rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
918         if (rc)
919                 GOTO(err_client, rc);
920
921         RETURN(0);
922
923 err_client:
924         class_disconnect_exports(obd);
925 err_fsd:
926         filter_free_server_data(filter);
927         RETURN(rc);
928 }
929
930 static int filter_cleanup_groups(struct obd_device *obd)
931 {
932         struct filter_obd *filter = &obd->u.filter;
933         struct file *filp;
934         struct dentry *dentry;
935         int i, j;
936         ENTRY;
937
938         if (filter->fo_dentry_O_groups != NULL) {
939                 for (i = 0; i < filter->fo_group_count; i++) {
940                         dentry = filter->fo_dentry_O_groups[i];
941                         if (dentry != NULL)
942                                 f_dput(dentry);
943                 }
944                 OBD_FREE(filter->fo_dentry_O_groups,
945                          filter->fo_group_count *
946                          sizeof(*filter->fo_dentry_O_groups));
947                 filter->fo_dentry_O_groups = NULL;
948         }
949         if (filter->fo_last_objid_files != NULL) {
950                 for (i = 0; i < filter->fo_group_count; i++) {
951                         filp = filter->fo_last_objid_files[i];
952                         if (filp != NULL)
953                                 filp_close(filp, 0);
954                 }
955                 OBD_FREE(filter->fo_last_objid_files,
956                          filter->fo_group_count *
957                          sizeof(*filter->fo_last_objid_files));
958                 filter->fo_last_objid_files = NULL;
959         }
960         if (filter->fo_dentry_O_sub != NULL) {
961                 for (i = 0; i < filter->fo_group_count; i++) {
962                         for (j = 0; j < filter->fo_subdir_count; j++) {
963                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
964                                 if (dentry != NULL)
965                                         f_dput(dentry);
966                         }
967                 }
968                 OBD_FREE(filter->fo_dentry_O_sub,
969                          filter->fo_group_count *
970                          sizeof(*filter->fo_dentry_O_sub));
971                 filter->fo_dentry_O_sub = NULL;
972         }
973         if (filter->fo_last_objids != NULL) {
974                 OBD_FREE(filter->fo_last_objids,
975                          filter->fo_group_count *
976                          sizeof(*filter->fo_last_objids));
977                 filter->fo_last_objids = NULL;
978         }
979         if (filter->fo_dentry_O != NULL) {
980                 f_dput(filter->fo_dentry_O);
981                 filter->fo_dentry_O = NULL;
982         }
983         RETURN(0);
984 }
985
986 static int filter_update_last_group(struct obd_device *obd, int group)
987 {
988         struct filter_obd *filter = &obd->u.filter;
989         struct file *filp = NULL;
990         int last_group = 0, rc;
991         loff_t off = 0;
992         ENTRY;
993
994         if (group <= filter->fo_committed_group)
995                 RETURN(0);
996
997         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
998         if (IS_ERR(filp)) {
999                 rc = PTR_ERR(filp);
1000                 filp = NULL;
1001                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1002                 GOTO(cleanup, rc);
1003         }
1004
1005         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1006         if (rc) {
1007                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1008                 GOTO(cleanup, rc);
1009         }
1010         LASSERTF(off == 0 || CHECK_MDS_GROUP(last_group),
1011                  "off = %llu and last_group = %d\n", off, last_group);
1012
1013         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1014                obd->obd_name, last_group, group);
1015
1016         off = 0;
1017         last_group = group;
1018         /* must be sync: bXXXX */
1019         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1020         if (rc) {
1021                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1022                 GOTO(cleanup, rc);
1023         }
1024
1025         filter->fo_committed_group = group;
1026 cleanup:
1027         if (filp)
1028                 filp_close(filp, 0);
1029         RETURN(rc);
1030 }
1031
1032 static int filter_read_group_internal(struct obd_device *obd, int group,
1033                                       int create)
1034 {
1035         struct filter_obd *filter = &obd->u.filter;
1036         __u64 *new_objids = NULL;
1037         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1038         struct dentry **new_groups = NULL;
1039         struct file **new_files = NULL;
1040         struct dentry *dentry;
1041         struct file *filp;
1042         int old_count = filter->fo_group_count, rc, stage = 0, i;
1043         char name[25];
1044         __u64 last_objid;
1045         loff_t off = 0;
1046         int len = group + 1;
1047
1048         snprintf(name, 24, "%d", group);
1049         name[24] = '\0';
1050
1051         if (!create) {
1052                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1053                                            strlen(name));
1054                 if (IS_ERR(dentry)) {
1055                         CERROR("Cannot lookup expected object group %d: %ld\n",
1056                                group, PTR_ERR(dentry));
1057                         RETURN(PTR_ERR(dentry));
1058                 }
1059         } else {
1060                 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1061                                       name, 0700, 1);
1062                 if (IS_ERR(dentry)) {
1063                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1064                                PTR_ERR(dentry));
1065                         RETURN(PTR_ERR(dentry));
1066                 }
1067         }
1068         stage = 1;
1069
1070         snprintf(name, 24, "O/%d/LAST_ID", group);
1071         name[24] = '\0';
1072         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1073         if (IS_ERR(filp)) {
1074                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1075                 GOTO(cleanup, rc = PTR_ERR(filp));
1076         }
1077         stage = 2;
1078
1079         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1080         if (rc) {
1081                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1082                 GOTO(cleanup, rc);
1083         }
1084
1085         if (filter->fo_subdir_count) {
1086                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1087                 if (tmp_subdirs == NULL)
1088                         GOTO(cleanup, rc = -ENOMEM);
1089                 stage = 3;
1090
1091                 for (i = 0; i < filter->fo_subdir_count; i++) {
1092                         char dir[20];
1093                         snprintf(dir, sizeof(dir), "d%u", i);
1094
1095                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1096                                                               filter->fo_vfsmnt,
1097                                                               dir, 0700, 1);
1098                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1099                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1100                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1101                                        group, dir, rc);
1102                                 GOTO(cleanup, rc);
1103                         }
1104
1105                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1106                                tmp_subdirs->dentry[i]);
1107                 }
1108         }
1109
1110         /* 'group' is an index; we need an array of length 'group + 1' */
1111         if (group + 1 > old_count) {
1112                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1113                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1114                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1115                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1116                 stage = 4;
1117                 if (new_objids == NULL || new_subdirs == NULL ||
1118                     new_groups == NULL || new_files == NULL)
1119                         GOTO(cleanup, rc = -ENOMEM);
1120
1121                 if (old_count) {
1122                         memcpy(new_objids, filter->fo_last_objids,
1123                                old_count * sizeof(*new_objids));
1124                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1125                                old_count * sizeof(*new_subdirs));
1126                         memcpy(new_groups, filter->fo_dentry_O_groups,
1127                                old_count * sizeof(*new_groups));
1128                         memcpy(new_files, filter->fo_last_objid_files,
1129                                old_count * sizeof(*new_files));
1130
1131                         OBD_FREE(filter->fo_last_objids,
1132                                  old_count * sizeof(*new_objids));
1133                         OBD_FREE(filter->fo_dentry_O_sub,
1134                                  old_count * sizeof(*new_subdirs));
1135                         OBD_FREE(filter->fo_dentry_O_groups,
1136                                  old_count * sizeof(*new_groups));
1137                         OBD_FREE(filter->fo_last_objid_files,
1138                                  old_count * sizeof(*new_files));
1139                 }
1140                 filter->fo_last_objids = new_objids;
1141                 filter->fo_dentry_O_sub = new_subdirs;
1142                 filter->fo_dentry_O_groups = new_groups;
1143                 filter->fo_last_objid_files = new_files;
1144                 filter->fo_group_count = len;
1145         }
1146
1147         filter->fo_dentry_O_groups[group] = dentry;
1148         filter->fo_last_objid_files[group] = filp;
1149         if (filter->fo_subdir_count) {
1150                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1151                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1152         }
1153
1154         filter_update_last_group(obd, group);
1155
1156         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1157                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1158                 rc = filter_update_last_objid(obd, group, 1);
1159                 RETURN(rc);
1160         }
1161
1162         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1163         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1164                obd->obd_name, group, last_objid);
1165         RETURN(0);
1166  cleanup:
1167         switch (stage) {
1168         case 4:
1169                 if (new_objids != NULL)
1170                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1171                 if (new_subdirs != NULL)
1172                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1173                 if (new_groups != NULL)
1174                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1175                 if (new_files != NULL)
1176                         OBD_FREE(new_files, len * sizeof(*new_files));
1177         case 3:
1178                 if (filter->fo_subdir_count) {
1179                         for (i = 0; i < filter->fo_subdir_count; i++) {
1180                                 if (tmp_subdirs->dentry[i] != NULL)
1181                                         dput(tmp_subdirs->dentry[i]);
1182                         }
1183                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1184                 }
1185         case 2:
1186                 filp_close(filp, 0);
1187         case 1:
1188                 dput(dentry);
1189         }
1190         RETURN(rc);
1191 }
1192
1193 static int filter_read_groups(struct obd_device *obd, int last_group,
1194                               int create)
1195 {
1196         struct filter_obd *filter = &obd->u.filter;
1197         int old_count, group, rc = 0;
1198
1199         down(&filter->fo_init_lock);
1200         old_count = filter->fo_group_count;
1201         for (group = old_count; group <= last_group; group++) {
1202
1203                 rc = filter_read_group_internal(obd, group, create);
1204                 if (rc != 0)
1205                         break;
1206         }
1207         up(&filter->fo_init_lock);
1208         return rc;
1209 }
1210
1211 /* FIXME: object groups */
1212 static int filter_prep_groups(struct obd_device *obd)
1213 {
1214         struct filter_obd *filter = &obd->u.filter;
1215         struct dentry *dentry, *O_dentry;
1216         struct file *filp;
1217         int    last_group, rc = 0, cleanup_phase = 0;
1218         loff_t off = 0;
1219         ENTRY;
1220
1221         O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1222                                 "O", 0700, 1);
1223         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1224         if (IS_ERR(O_dentry)) {
1225                 rc = PTR_ERR(O_dentry);
1226                 CERROR("cannot open/create O: rc = %d\n", rc);
1227                 GOTO(cleanup, rc);
1228         }
1229         filter->fo_dentry_O = O_dentry;
1230         cleanup_phase = 1; /* O_dentry */
1231
1232         /* Lookup "R" to tell if we're on an old OST FS and need to convert
1233          * from O/R/<dir>/<objid> to O/0/<dir>/<objid>.  This can be removed
1234          * some time post 1.0 when all old-style OSTs have converted along
1235          * with the init_objid hack. */
1236         dentry = ll_lookup_one_len("R", O_dentry, 1);
1237         if (IS_ERR(dentry))
1238                 GOTO(cleanup, rc = PTR_ERR(dentry));
1239         if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1240                 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1241                 ENTRY;
1242
1243                 CWARN("converting OST to new object layout\n");
1244                 if (IS_ERR(O0_dentry)) {
1245                         rc = PTR_ERR(O0_dentry);
1246                         CERROR("error looking up O/0: rc %d\n", rc);
1247                         GOTO(cleanup_R, rc);
1248                 }
1249
1250                 if (O0_dentry->d_inode) {
1251                         CERROR("Both O/R and O/0 exist. Fix manually.\n");
1252                         GOTO(cleanup_O0, rc = -EEXIST);
1253                 }
1254
1255                 LOCK_INODE_MUTEX(O_dentry->d_inode);
1256                 rc = ll_vfs_rename(O_dentry->d_inode, dentry, filter->fo_vfsmnt,
1257                                    O_dentry->d_inode, O0_dentry,
1258                                    filter->fo_vfsmnt);
1259                 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1260
1261                 if (rc) {
1262                         CERROR("error renaming O/R to O/0: rc %d\n", rc);
1263                         GOTO(cleanup_O0, rc);
1264                 }
1265                 filter->fo_fsd->lsd_feature_incompat |=
1266                         cpu_to_le32(OBD_INCOMPAT_GROUPS);
1267                 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1268                                                filter->fo_fsd, 1);
1269                 GOTO(cleanup_O0, rc);
1270
1271         cleanup_O0:
1272                 f_dput(O0_dentry);
1273         cleanup_R:
1274                 f_dput(dentry);
1275                 if (rc)
1276                         GOTO(cleanup, rc);
1277         } else {
1278                 f_dput(dentry);
1279         }
1280
1281         cleanup_phase = 2; /* groups */
1282
1283         /* we have to initialize all groups before first connections from
1284          * clients because they may send create/destroy for any group -bzzz */
1285         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1286         if (IS_ERR(filp)) {
1287                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1288                 GOTO(cleanup, rc = PTR_ERR(filp));
1289         }
1290         cleanup_phase = 3; /* filp */
1291
1292         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1293         if (rc) {
1294                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1295                 GOTO(cleanup, rc);
1296         }
1297         if (off == 0) {
1298                 last_group = FILTER_MIN_GROUPS;
1299         } else {
1300                 LASSERT_MDS_GROUP(last_group);
1301         }
1302
1303         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1304               FILTER_MIN_GROUPS, last_group);
1305         filter->fo_committed_group = last_group;
1306         rc = filter_read_groups(obd, last_group, 1);
1307         if (rc)
1308                 GOTO(cleanup, rc);
1309
1310         filp_close(filp, 0);
1311         RETURN(0);
1312
1313  cleanup:
1314         switch (cleanup_phase) {
1315         case 3:
1316                 filp_close(filp, 0);
1317         case 2:
1318                 filter_cleanup_groups(obd);
1319         case 1:
1320                 f_dput(filter->fo_dentry_O);
1321                 filter->fo_dentry_O = NULL;
1322         default:
1323                 break;
1324         }
1325         return rc;
1326
1327 }
1328
1329 /* setup the object store with correct subdirectories */
1330 static int filter_prep(struct obd_device *obd)
1331 {
1332         struct lvfs_run_ctxt saved;
1333         struct filter_obd *filter = &obd->u.filter;
1334         struct file *file;
1335         struct inode *inode;
1336         int rc = 0;
1337         ENTRY;
1338
1339         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1340         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1341         if (!file || IS_ERR(file)) {
1342                 rc = PTR_ERR(file);
1343                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1344                        LAST_RCVD, rc);
1345                 GOTO(out, rc);
1346         }
1347         filter->fo_rcvd_filp = file;
1348         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1349                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1350                        file->f_dentry->d_inode->i_mode);
1351                 GOTO(err_filp, rc = -ENOENT);
1352         }
1353
1354         inode = file->f_dentry->d_parent->d_inode;
1355         /* We use i_op->unlink directly in filter_vfs_unlink() */
1356         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1357                 CERROR("%s: filesystem does not support create/unlink ops\n",
1358                        obd->obd_name);
1359                 GOTO(err_filp, rc = -EOPNOTSUPP);
1360         }
1361
1362         /** lu_target has very limited use in filter now */
1363         lut_init(NULL, &filter->fo_lut, obd, NULL);
1364
1365         rc = filter_init_server_data(obd, file);
1366         if (rc) {
1367                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1368                 GOTO(err_filp, rc);
1369         }
1370
1371         target_recovery_init(&filter->fo_lut, ost_handle);
1372
1373         /* open and create health check io file*/
1374         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1375         if (IS_ERR(file)) {
1376                 rc = PTR_ERR(file);
1377                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1378                        HEALTH_CHECK, rc);
1379                 GOTO(err_server_data, rc);
1380         }
1381         filter->fo_health_check_filp = file;
1382         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1383                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1384                        file->f_dentry->d_inode->i_mode);
1385                 GOTO(err_health_check, rc = -ENOENT);
1386         }
1387         rc = lvfs_check_io_health(obd, file);
1388         if (rc)
1389                 GOTO(err_health_check, rc);
1390
1391         rc = filter_prep_groups(obd);
1392         if (rc)
1393                 GOTO(err_health_check, rc);
1394 out:
1395         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1396
1397         return(rc);
1398
1399 err_health_check:
1400         if (filp_close(filter->fo_health_check_filp, 0))
1401                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1402         filter->fo_health_check_filp = NULL;
1403 err_server_data:
1404         target_recovery_fini(obd);
1405         filter_free_server_data(filter);
1406 err_filp:
1407         if (filp_close(filter->fo_rcvd_filp, 0))
1408                 CERROR("can't close %s after error\n", LAST_RCVD);
1409         filter->fo_rcvd_filp = NULL;
1410         goto out;
1411 }
1412
1413 /* cleanup the filter: write last used object id to status file */
1414 static void filter_post(struct obd_device *obd)
1415 {
1416         struct lvfs_run_ctxt saved;
1417         struct filter_obd *filter = &obd->u.filter;
1418         int rc, i;
1419
1420         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1421          * best to start a transaction with h_sync, because we removed this
1422          * from lastobjid */
1423
1424         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1425         rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1426                                        filter->fo_fsd, 0);
1427         if (rc)
1428                 CERROR("error writing server data: rc = %d\n", rc);
1429
1430         for (i = 0; i < filter->fo_group_count; i++) {
1431                 rc = filter_update_last_objid(obd, i,
1432                                 (i == filter->fo_group_count - 1));
1433                 if (rc)
1434                         CERROR("error writing group %d lastobjid: rc = %d\n",
1435                                i, rc);
1436         }
1437
1438         rc = filp_close(filter->fo_rcvd_filp, 0);
1439         filter->fo_rcvd_filp = NULL;
1440         if (rc)
1441                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1442
1443         rc = filp_close(filter->fo_health_check_filp, 0);
1444         filter->fo_health_check_filp = NULL;
1445         if (rc)
1446                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1447
1448         filter_cleanup_groups(obd);
1449         filter_free_server_data(filter);
1450         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1451
1452         filter_free_capa_keys(filter);
1453         cleanup_capa_hash(filter->fo_capa_hash);
1454 }
1455
1456 static void filter_set_last_id(struct filter_obd *filter,
1457                                obd_id id, obd_gr group)
1458 {
1459         LASSERT(filter->fo_fsd != NULL);
1460         LASSERT(group <= filter->fo_group_count);
1461
1462         spin_lock(&filter->fo_objidlock);
1463         filter->fo_last_objids[group] = id;
1464         spin_unlock(&filter->fo_objidlock);
1465 }
1466
1467 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1468 {
1469         obd_id id;
1470         LASSERT(filter->fo_fsd != NULL);
1471         LASSERT(group <= filter->fo_group_count);
1472
1473         /* FIXME: object groups */
1474         spin_lock(&filter->fo_objidlock);
1475         id = filter->fo_last_objids[group];
1476         spin_unlock(&filter->fo_objidlock);
1477         return id;
1478 }
1479
1480 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1481 {
1482         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1483         return 0;
1484 }
1485
1486 /* We never dget the object parent, so DON'T dput it either */
1487 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1488 {
1489         struct filter_obd *filter = &obd->u.filter;
1490         struct filter_subdirs *subdirs;
1491         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1492
1493         if ((group > FILTER_GROUP_MDS0 && group < FILTER_GROUP_MDS1_N_BASE) ||
1494              filter->fo_subdir_count == 0)
1495                 return filter->fo_dentry_O_groups[group];
1496
1497         subdirs = &filter->fo_dentry_O_sub[group];
1498         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1499 }
1500
1501 /* We never dget the object parent, so DON'T dput it either */
1502 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1503                                   obd_id objid)
1504 {
1505         unsigned long now = jiffies;
1506         struct dentry *dparent = filter_parent(obd, group, objid);
1507         int rc;
1508
1509         if (IS_ERR(dparent))
1510                 return dparent;
1511         if (dparent == NULL)
1512                 return ERR_PTR(-ENOENT);
1513
1514         rc = filter_lock_dentry(obd, dparent);
1515         fsfilt_check_slow(obd, now, "parent lock");
1516         return rc ? ERR_PTR(rc) : dparent;
1517 }
1518
1519 /* We never dget the object parent, so DON'T dput it either */
1520 static void filter_parent_unlock(struct dentry *dparent)
1521 {
1522         UNLOCK_INODE_MUTEX(dparent->d_inode);
1523 }
1524
1525 /* How to get files, dentries, inodes from object id's.
1526  *
1527  * If dir_dentry is passed, the caller has already locked the parent
1528  * appropriately for this operation (normally a write lock).  If
1529  * dir_dentry is NULL, we do a read lock while we do the lookup to
1530  * avoid races with create/destroy and such changing the directory
1531  * internal to the filesystem code. */
1532 struct dentry *filter_fid2dentry(struct obd_device *obd,
1533                                  struct dentry *dir_dentry,
1534                                  obd_gr group, obd_id id)
1535 {
1536         struct dentry *dparent = dir_dentry;
1537         struct dentry *dchild;
1538         char name[32];
1539         int len;
1540         ENTRY;
1541
1542         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1543             obd->u.filter.fo_destroys_in_progress == 0) {
1544                 /* don't fail lookups for orphan recovery, it causes
1545                  * later LBUGs when objects still exist during precreate */
1546                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1547                 RETURN(ERR_PTR(-ENOENT));
1548         }
1549         if (id == 0) {
1550                 CERROR("fatal: invalid object id 0\n");
1551                 RETURN(ERR_PTR(-ESTALE));
1552         }
1553
1554         len = sprintf(name, LPU64, id);
1555         if (dir_dentry == NULL) {
1556                 dparent = filter_parent_lock(obd, group, id);
1557                 if (IS_ERR(dparent)) {
1558                         CERROR("%s: error getting object "LPU64":"LPU64
1559                                " parent: rc %ld\n", obd->obd_name,
1560                                id, group, PTR_ERR(dparent));
1561                         RETURN(dparent);
1562                 }
1563         }
1564         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1565                dparent->d_name.len, dparent->d_name.name, name);
1566         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1567         if (dir_dentry == NULL)
1568                 filter_parent_unlock(dparent);
1569         if (IS_ERR(dchild)) {
1570                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1571                        PTR_ERR(dchild));
1572                 RETURN(dchild);
1573         }
1574
1575         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1576                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1577                        obd->obd_name, id, dchild->d_inode->i_ino);
1578                 f_dput(dchild);
1579                 RETURN(ERR_PTR(-ENOENT));
1580         }
1581
1582         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1583                name, dchild, atomic_read(&dchild->d_count));
1584
1585         LASSERT(atomic_read(&dchild->d_count) > 0);
1586
1587         RETURN(dchild);
1588 }
1589
1590 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1591                                   obd_id group, struct lustre_handle *lockh)
1592 {
1593         int flags = LDLM_AST_DISCARD_DATA, rc;
1594         struct ldlm_res_id res_id;
1595         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1596         ENTRY;
1597
1598         osc_build_res_name(objid, group, &res_id);
1599         /* Tell the clients that the object is gone now and that they should
1600          * throw away any cached pages. */
1601         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1602                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1603                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1604                                     NULL, lockh);
1605         if (rc != ELDLM_OK)
1606                 lockh->cookie = 0;
1607         RETURN(rc);
1608 }
1609
1610 static void filter_fini_destroy(struct obd_device *obd,
1611                                 struct lustre_handle *lockh)
1612 {
1613         if (lockh->cookie)
1614                 ldlm_lock_decref(lockh, LCK_PW);
1615 }
1616
1617 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1618  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1619  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1620  * reverse.  Caller must take i_sem before starting the transaction and we
1621  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1622 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1623                       struct vfsmount *mnt)
1624 {
1625         int rc;
1626         ENTRY;
1627
1628         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1629          * itself we already hold dir->i_mutex for child create/unlink ops */
1630         LASSERT(dentry->d_inode != NULL);
1631         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1632         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1633
1634
1635         /* may_delete() */
1636         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1637                 GOTO(out, rc = -ENOENT);
1638
1639         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1640         if (rc)
1641                 GOTO(out, rc);
1642
1643         if (IS_APPEND(dir))
1644                 GOTO(out, rc = -EPERM);
1645
1646         /* check_sticky() */
1647         if ((dentry->d_inode->i_uid != current->fsuid &&
1648              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1649             IS_IMMUTABLE(dentry->d_inode))
1650                 GOTO(out, rc = -EPERM);
1651
1652         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1653          *       that was done because of journal_start (which is already done
1654          *       here) or some other ordering issue. */
1655         DQUOT_INIT(dir);
1656
1657         rc = ll_security_inode_unlink(dir, dentry, mnt);
1658         if (rc)
1659                 GOTO(out, rc);
1660
1661         rc = dir->i_op->unlink(dir, dentry);
1662 out:
1663         /* need to drop i_mutex before we lose inode reference */
1664         UNLOCK_INODE_MUTEX(dentry->d_inode);
1665         if (rc == 0)
1666                 d_delete(dentry);
1667
1668         RETURN(rc);
1669 }
1670
1671 /* Caller must hold LCK_PW on parent and push us into kernel context.
1672  * Caller must hold child i_mutex, we drop it always.
1673  * Caller is also required to ensure that dchild->d_inode exists. */
1674 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1675                                    obd_gr group, struct dentry *dparent,
1676                                    struct dentry *dchild)
1677 {
1678         struct inode *inode = dchild->d_inode;
1679         struct filter_obd *filter = &obd->u.filter;
1680         int rc;
1681
1682         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1683                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1684                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1685                        (unsigned long)inode->i_nlink,
1686                        atomic_read(&inode->i_count));
1687         }
1688
1689         rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1690         if (rc)
1691                 CERROR("error unlinking objid %.*s: rc %d\n",
1692                        dchild->d_name.len, dchild->d_name.name, rc);
1693         return(rc);
1694 }
1695
1696 struct filter_intent_args {
1697         struct ldlm_lock **victim;
1698         __u64 size;
1699         int *liblustre;
1700 };
1701
1702 static enum interval_iter filter_intent_cb(struct interval_node *n,
1703                                            void *args)
1704 {
1705         struct ldlm_interval *node = (struct ldlm_interval *)n;
1706         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1707         __u64 size = arg->size;
1708         struct ldlm_lock **v = arg->victim;
1709         struct ldlm_lock *lck;
1710
1711         /* If the interval is lower than the current file size,
1712          * just break. */
1713         if (interval_high(n) <= size)
1714                 return INTERVAL_ITER_STOP;
1715
1716         list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1717                 /* Don't send glimpse ASTs to liblustre clients.
1718                  * They aren't listening for them, and they do
1719                  * entirely synchronous I/O anyways. */
1720                 if (lck->l_export == NULL ||
1721                     lck->l_export->exp_libclient == 1)
1722                         continue;
1723
1724                 if (*arg->liblustre)
1725                         *arg->liblustre = 0;
1726
1727                 if (*v == NULL) {
1728                         *v = LDLM_LOCK_GET(lck);
1729                 } else if ((*v)->l_policy_data.l_extent.start <
1730                            lck->l_policy_data.l_extent.start) {
1731                         LDLM_LOCK_RELEASE(*v);
1732                         *v = LDLM_LOCK_GET(lck);
1733                 }
1734
1735                 /* the same policy group - every lock has the
1736                  * same extent, so needn't do it any more */
1737                 break;
1738         }
1739
1740         return INTERVAL_ITER_CONT;
1741 }
1742
1743 static int filter_intent_policy(struct ldlm_namespace *ns,
1744                                 struct ldlm_lock **lockp, void *req_cookie,
1745                                 ldlm_mode_t mode, int flags, void *data)
1746 {
1747         CFS_LIST_HEAD(rpc_list);
1748         struct ptlrpc_request *req = req_cookie;
1749         struct ldlm_lock *lock = *lockp, *l = NULL;
1750         struct ldlm_resource *res = lock->l_resource;
1751         ldlm_processing_policy policy;
1752         struct ost_lvb *res_lvb, *reply_lvb;
1753         struct ldlm_reply *rep;
1754         ldlm_error_t err;
1755         int idx, rc, tmpflags = 0, only_liblustre = 1;
1756         struct ldlm_interval_tree *tree;
1757         struct filter_intent_args arg;
1758         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1759                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1760                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1761         ENTRY;
1762
1763         policy = ldlm_get_processing_policy(res);
1764         LASSERT(policy != NULL);
1765         LASSERT(req != NULL);
1766
1767         rc = lustre_pack_reply(req, 3, repsize, NULL);
1768         if (rc)
1769                 RETURN(req->rq_status = rc);
1770
1771         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1772         LASSERT(rep != NULL);
1773
1774         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1775                                    sizeof(*reply_lvb));
1776         LASSERT(reply_lvb != NULL);
1777
1778         //fixup_handle_for_resent_req(req, lock, &lockh);
1779
1780         /* Call the extent policy function to see if our request can be
1781          * granted, or is blocked.
1782          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1783          * lock, and should not be granted if the lock will be blocked.
1784          */
1785
1786         LASSERT(ns == res->lr_namespace);
1787         lock_res(res);
1788         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1789         check_res_locked(res);
1790
1791         /* FIXME: we should change the policy function slightly, to not make
1792          * this list at all, since we just turn around and free it */
1793         while (!list_empty(&rpc_list)) {
1794                 struct ldlm_lock *wlock =
1795                         list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1796                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1797                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1798                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1799                 list_del_init(&wlock->l_cp_ast);
1800                 LDLM_LOCK_RELEASE(wlock);
1801         }
1802
1803         /* The lock met with no resistance; we're finished. */
1804         if (rc == LDLM_ITER_CONTINUE) {
1805                 /* do not grant locks to the liblustre clients: they cannot
1806                  * handle ASTs robustly.  We need to do this while still
1807                  * holding ns_lock to avoid the lock remaining on the res_link
1808                  * list (and potentially being added to l_pending_list by an
1809                  * AST) when we are going to drop this lock ASAP. */
1810                 if (lock->l_export->exp_libclient ||
1811                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1812                         ldlm_resource_unlink_lock(lock);
1813                         err = ELDLM_LOCK_ABORTED;
1814                 } else {
1815                         err = ELDLM_LOCK_REPLACED;
1816                 }
1817                 unlock_res(res);
1818                 RETURN(err);
1819         }
1820
1821         /* Do not grant any lock, but instead send GL callbacks.  The extent
1822          * policy nicely created a list of all PW locks for us.  We will choose
1823          * the highest of those which are larger than the size in the LVB, if
1824          * any, and perform a glimpse callback. */
1825         res_lvb = res->lr_lvb_data;
1826         LASSERT(res_lvb != NULL);
1827         *reply_lvb = *res_lvb;
1828
1829         /*
1830          * ->ns_lock guarantees that no new locks are granted, and,
1831          * therefore, that res->lr_lvb_data cannot increase beyond the
1832          * end of already granted lock. As a result, it is safe to
1833          * check against "stale" reply_lvb->lvb_size value without
1834          * res->lr_lvb_sem.
1835          */
1836         arg.size = reply_lvb->lvb_size;
1837         arg.victim = &l;
1838         arg.liblustre = &only_liblustre;
1839         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1840                 tree = &res->lr_itree[idx];
1841                 if (tree->lit_mode == LCK_PR)
1842                         continue;
1843
1844                 interval_iterate_reverse(tree->lit_root,
1845                                          filter_intent_cb, &arg);
1846         }
1847         unlock_res(res);
1848
1849         /* There were no PW locks beyond the size in the LVB; finished. */
1850         if (l == NULL) {
1851                 if (only_liblustre) {
1852                         /* If we discovered a liblustre client with a PW lock,
1853                          * however, the LVB may be out of date!  The LVB is
1854                          * updated only on glimpse (which we don't do for
1855                          * liblustre clients) and cancel (which the client
1856                          * obviously has not yet done).  So if it has written
1857                          * data but kept the lock, the LVB is stale and needs
1858                          * to be updated from disk.
1859                          *
1860                          * Of course, this will all disappear when we switch to
1861                          * taking liblustre locks on the OST. */
1862                         ldlm_res_lvbo_update(res, NULL, 0, 1);
1863                 }
1864                 RETURN(ELDLM_LOCK_ABORTED);
1865         }
1866
1867         /*
1868          * This check is for lock taken in filter_prepare_destroy() that does
1869          * not have l_glimpse_ast set. So the logic is: if there is a lock
1870          * with no l_glimpse_ast set, this object is being destroyed already.
1871          *
1872          * Hence, if you are grabbing DLM locks on the server, always set
1873          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1874          */
1875         if (l->l_glimpse_ast == NULL) {
1876                 /* We are racing with unlink(); just return -ENOENT */
1877                 rep->lock_policy_res1 = -ENOENT;
1878                 goto out;
1879         }
1880
1881         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1882         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1883         /* Update the LVB from disk if the AST failed (this is a legal race) */
1884         /*
1885          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1886          * sending ast is not handled. This can result in lost client writes.
1887          */
1888         if (rc != 0)
1889                 ldlm_res_lvbo_update(res, NULL, 0, 1);
1890
1891         lock_res(res);
1892         *reply_lvb = *res_lvb;
1893         unlock_res(res);
1894
1895  out:
1896         LDLM_LOCK_RELEASE(l);
1897
1898         RETURN(ELDLM_LOCK_ABORTED);
1899 }
1900
1901 /*
1902  * per-obd_device iobuf pool.
1903  *
1904  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1905  * allocations in write-path has to be minimized (see bug 5137).
1906  *
1907  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1908  * OST threads (see ost_thread_{init,done}()).
1909  *
1910  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1911  * at the OST layer there are only (potentially) multiple obd_device of type
1912  * unknown at the time of OST thread creation.
1913  *
1914  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1915  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1916  * it's very own iobuf.
1917  *
1918  * Functions below
1919  *
1920  *     filter_kiobuf_pool_init()
1921  *
1922  *     filter_kiobuf_pool_done()
1923  *
1924  *     filter_iobuf_get()
1925  *
1926  * operate on this array. They are "generic" in a sense that they don't depend
1927  * on actual type of iobuf's (the latter depending on Linux kernel version).
1928  */
1929
1930 /*
1931  * destroy pool created by filter_iobuf_pool_init
1932  */
1933 static void filter_iobuf_pool_done(struct filter_obd *filter)
1934 {
1935         struct filter_iobuf **pool;
1936         int i;
1937
1938         ENTRY;
1939
1940         pool = filter->fo_iobuf_pool;
1941         if (pool != NULL) {
1942                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1943                         if (pool[i] != NULL)
1944                                 filter_free_iobuf(pool[i]);
1945                 }
1946                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1947                 filter->fo_iobuf_pool = NULL;
1948         }
1949         EXIT;
1950 }
1951
1952 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1953 {
1954         struct filter_obd       *filter = &obd->u.filter;
1955         struct sptlrpc_rule_set  tmp_rset;
1956         int                      rc;
1957
1958         sptlrpc_rule_set_init(&tmp_rset);
1959         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1960         if (rc) {
1961                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1962                        obd->obd_name, rc);
1963                 return rc;
1964         }
1965
1966         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1967
1968         write_lock(&filter->fo_sptlrpc_lock);
1969         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1970         filter->fo_sptlrpc_rset = tmp_rset;
1971         write_unlock(&filter->fo_sptlrpc_lock);
1972
1973         return 0;
1974 }
1975
1976 /*
1977  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1978  */
1979 static int filter_iobuf_pool_init(struct filter_obd *filter)
1980 {
1981         void **pool;
1982
1983         ENTRY;
1984
1985
1986         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1987                       GFP_KERNEL);
1988         if (filter->fo_iobuf_pool == NULL)
1989                 RETURN(-ENOMEM);
1990
1991         filter->fo_iobuf_count = OSS_THREADS_MAX;
1992
1993         RETURN(0);
1994 }
1995
1996 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1997  * many threads there will be so we allocate a large empty array and only
1998  * fill in those slots that are actually in use.
1999  * If we haven't allocated a pool entry for this thread before, do so now. */
2000 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
2001 {
2002         int thread_id                    = (oti && oti->oti_thread) ?
2003                                            oti->oti_thread->t_id : -1;
2004         struct filter_iobuf  *pool       = NULL;
2005         struct filter_iobuf **pool_place = NULL;
2006
2007         if (thread_id >= 0) {
2008                 LASSERT(thread_id < filter->fo_iobuf_count);
2009                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
2010         }
2011
2012         if (unlikely(pool == NULL)) {
2013                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
2014                                           PTLRPC_MAX_BRW_PAGES);
2015                 if (pool_place != NULL)
2016                         *pool_place = pool;
2017         }
2018
2019         return pool;
2020 }
2021
2022 /* mount the file system (secretly).  lustre_cfg parameters are:
2023  * 1 = device
2024  * 2 = fstype
2025  * 3 = flags: failover=f, failout=n
2026  * 4 = mount options
2027  */
2028 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
2029                         void *option)
2030 {
2031         struct filter_obd *filter = &obd->u.filter;
2032         struct vfsmount *mnt;
2033         struct lustre_mount_info *lmi;
2034         struct obd_uuid uuid;
2035         __u8 *uuid_ptr;
2036         char *str, *label;
2037         char ns_name[48];
2038         request_queue_t *q;
2039         int rc, i;
2040         ENTRY;
2041
2042         if (lcfg->lcfg_bufcount < 3 ||
2043             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
2044             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
2045                 RETURN(-EINVAL);
2046
2047         lmi = server_get_mount(obd->obd_name);
2048         if (lmi) {
2049                 /* We already mounted in lustre_fill_super.
2050                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2051                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2052                 mnt = lmi->lmi_mnt;
2053                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2054         } else {
2055                 /* old path - used by lctl */
2056                 CERROR("Using old MDS mount method\n");
2057                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2058                                     MS_NOATIME|MS_NODIRATIME,
2059                                     lustre_cfg_string(lcfg, 1), option);
2060                 if (IS_ERR(mnt)) {
2061                         rc = PTR_ERR(mnt);
2062                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2063                                            lustre_cfg_string(lcfg, 1), rc);
2064                         RETURN(rc);
2065                 }
2066
2067                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2068         }
2069         if (IS_ERR(obd->obd_fsops))
2070                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2071
2072         rc = filter_iobuf_pool_init(filter);
2073         if (rc != 0)
2074                 GOTO(err_ops, rc);
2075
2076         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2077                 CERROR("%s: Underlying device is marked as read-only. "
2078                        "Setup failed\n", obd->obd_name);
2079                 GOTO(err_ops, rc = -EROFS);
2080         }
2081
2082         /* failover is the default */
2083         obd->obd_replayable = 1;
2084
2085         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2086                 str = lustre_cfg_string(lcfg, 3);
2087                 if (strchr(str, 'n')) {
2088                         CWARN("%s: recovery disabled\n", obd->obd_name);
2089                         obd->obd_replayable = 0;
2090                 }
2091         }
2092
2093         filter->fo_vfsmnt = mnt;
2094         obd->u.obt.obt_sb = mnt->mnt_sb;
2095         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2096         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2097
2098         fsfilt_setup(obd, obd->u.obt.obt_sb);
2099
2100         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2101         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2102         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2103         obd->obd_lvfs_ctxt.fs = get_ds();
2104         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2105
2106         init_mutex(&filter->fo_init_lock);
2107         filter->fo_committed_group = 0;
2108         filter->fo_destroys_in_progress = 0;
2109         for (i = 0; i < 32; i++)
2110                 sema_init(&filter->fo_create_locks[i], 1);
2111
2112         spin_lock_init(&filter->fo_translock);
2113         spin_lock_init(&filter->fo_objidlock);
2114         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2115         sema_init(&filter->fo_alloc_lock, 1);
2116         init_brw_stats(&filter->fo_filter_stats);
2117         filter->fo_read_cache = 1; /* enable read-only cache by default */
2118         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2119         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2120         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2121         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2122
2123         rc = filter_prep(obd);
2124         if (rc)
2125                 GOTO(err_ops, rc);
2126
2127         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2128         spin_lock_init(&filter->fo_llog_list_lock);
2129
2130         filter->fo_fl_oss_capa = 1;
2131
2132         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2133         filter->fo_capa_hash = init_capa_hash();
2134         if (filter->fo_capa_hash == NULL)
2135                 GOTO(err_post, rc = -ENOMEM);
2136
2137         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2138         obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2139                                                 LDLM_NAMESPACE_GREEDY);
2140         if (obd->obd_namespace == NULL)
2141                 GOTO(err_post, rc = -ENOMEM);
2142         obd->obd_namespace->ns_lvbp = obd;
2143         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2144         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2145
2146         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2147                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2148
2149         rc = obd_llog_init(obd, &obd->obd_olg, obd, 1, NULL, NULL);
2150         if (rc) {
2151                 CERROR("failed to setup llogging subsystems\n");
2152                 GOTO(err_post, rc);
2153         }
2154
2155         rwlock_init(&filter->fo_sptlrpc_lock);
2156         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2157         /* do this after llog being initialized */
2158         filter_adapt_sptlrpc_conf(obd, 1);
2159
2160         rc = lquota_setup(filter_quota_interface_ref, obd);
2161         if (rc)
2162                 GOTO(err_post, rc);
2163
2164         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2165         if (q->max_sectors < q->max_hw_sectors &&
2166             q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2167                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2168                               "for larger I/O requests: max_sectors = %u "
2169                               "could be up to max_hw_sectors=%u\n",
2170                               obd->obd_name, mnt->mnt_sb->s_id,
2171                               q->max_sectors, q->max_hw_sectors);
2172
2173         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2174         if (uuid_ptr != NULL) {
2175                 class_uuid_unparse(uuid_ptr, &uuid);
2176                 str = uuid.uuid;
2177         } else {
2178                 str = "no UUID";
2179         }
2180
2181         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2182
2183         if (obd->obd_recovering) {
2184                 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2185                               "recovery for at least %d:%.02d, or until %d "
2186                               "client%s reconnect%s.\n",
2187                               obd->obd_name, lustre_cfg_string(lcfg, 1),
2188                               label ?: "", label ? "/" : "", str,
2189                               obd->obd_recovery_timeout / 60,
2190                               obd->obd_recovery_timeout % 60,
2191                               obd->obd_max_recoverable_clients,
2192                               (obd->obd_max_recoverable_clients == 1) ? "":"s",
2193                               (obd->obd_max_recoverable_clients == 1) ? "s":"");
2194         } else {
2195                 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2196                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2197                               label ?: "", label ? "/" : "", str,
2198                               obd->obd_replayable ? "enabled" : "disabled");
2199         }
2200
2201         RETURN(0);
2202
2203 err_post:
2204         filter_post(obd);
2205 err_ops:
2206         fsfilt_put_ops(obd->obd_fsops);
2207         filter_iobuf_pool_done(filter);
2208 err_mntput:
2209         server_put_mount(obd->obd_name, mnt);
2210         obd->u.obt.obt_sb = 0;
2211         return rc;
2212 }
2213
2214 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2215 {
2216         struct lprocfs_static_vars lvars;
2217         unsigned long addr;
2218         struct page *page;
2219         int rc;
2220
2221         CLASSERT(offsetof(struct obd_device, u.obt) ==
2222                  offsetof(struct obd_device, u.filter.fo_obt));
2223
2224         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2225                 RETURN(-EINVAL);
2226
2227         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2228         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2229         if (!page)
2230                 RETURN(-ENOMEM);
2231         addr = (unsigned long)cfs_page_address(page);
2232         clear_page((void *)addr);
2233
2234         /* lprocfs must be setup before the filter so state can be safely added
2235          * to /proc incrementally as the filter is setup */
2236         lprocfs_filter_init_vars(&lvars);
2237         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2238             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2239                 /* Init obdfilter private stats here */
2240                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2241                                      LPROCFS_CNTR_AVGMINMAX,
2242                                      "read_bytes", "bytes");
2243                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2244                                      LPROCFS_CNTR_AVGMINMAX,
2245                                      "write_bytes", "bytes");
2246                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2247                                      LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2248                                      "get_page", "usec");
2249                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2250                                      LPROCFS_CNTR_AVGMINMAX,
2251                                      "get_page_failures", "num");
2252                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2253                                      LPROCFS_CNTR_AVGMINMAX,
2254                                      "cache_access", "pages");
2255                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2256                                      LPROCFS_CNTR_AVGMINMAX,
2257                                      "cache_hit", "pages");
2258                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2259                                      LPROCFS_CNTR_AVGMINMAX,
2260                                      "cache_miss", "pages");
2261
2262                 lproc_filter_attach_seqstat(obd);
2263                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2264                                                         obd->obd_proc_entry,
2265                                                         NULL, NULL);
2266                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2267                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2268                         CERROR("error %d setting up lprocfs for %s\n",
2269                                rc, "exports");
2270                         obd->obd_proc_exports_entry = NULL;
2271                 }
2272         }
2273         if (obd->obd_proc_exports_entry)
2274                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2275                                    lprocfs_nid_stats_clear_read,
2276                                    lprocfs_nid_stats_clear_write, obd, NULL);
2277
2278         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2279                LUSTRE_CFG_BUFLEN(lcfg, 4));
2280         rc = filter_common_setup(obd, lcfg, (void *)addr);
2281         OBD_PAGE_FREE(page);
2282
2283         if (rc) {
2284                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2285                 lprocfs_free_per_client_stats(obd);
2286                 lprocfs_free_obd_stats(obd);
2287                 lprocfs_obd_cleanup(obd);
2288         }
2289
2290         return rc;
2291 }
2292
2293 static struct llog_operations filter_mds_ost_repl_logops;
2294
2295 static struct llog_operations filter_size_orig_logops = {
2296         .lop_setup   = llog_obd_origin_setup,
2297         .lop_cleanup = llog_obd_origin_cleanup,
2298         .lop_add     = llog_obd_origin_add
2299 };
2300
2301 static int filter_olg_fini(struct obd_llog_group *olg)
2302 {
2303         struct llog_ctxt *ctxt;
2304         int rc = 0, rc2 = 0;
2305         ENTRY;
2306
2307         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2308         if (ctxt)
2309                 rc = llog_cleanup(ctxt);
2310
2311         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2312         if (ctxt) {
2313                 rc2 = llog_cleanup(ctxt);
2314                 if (!rc)
2315                         rc = rc2;
2316         }
2317
2318         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2319         if (ctxt) {
2320                 rc2 = llog_cleanup(ctxt);
2321                 if (!rc)
2322                         rc = rc2;
2323         }
2324
2325         RETURN(rc);
2326 }
2327
2328 static int
2329 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2330                 struct obd_device *tgt)
2331 {
2332         int rc;
2333         ENTRY;
2334
2335         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2336                         &filter_mds_ost_repl_logops);
2337         if (rc)
2338                 GOTO(cleanup, rc);
2339
2340         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2341                         &filter_size_orig_logops);
2342         if (rc)
2343                 GOTO(cleanup, rc);
2344         EXIT;
2345 cleanup:
2346         if (rc)
2347                 filter_olg_fini(olg);
2348         return rc;
2349 }
2350
2351 /**
2352  * Init the default olg, which is embeded in the obd_device, for filter.
2353  */
2354 static int
2355 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2356                         struct obd_device *tgt)
2357 {
2358         struct filter_obd *filter = &obd->u.filter;
2359         struct llog_ctxt *ctxt;
2360         int rc;
2361         ENTRY;
2362
2363         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2364         if (!filter->fo_lcm)
2365                 RETURN(-ENOMEM);
2366
2367         filter_mds_ost_repl_logops = llog_client_ops;
2368         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2369         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2370         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2371
2372         rc = filter_olg_init(obd, olg, tgt);
2373         if (rc)
2374                 GOTO(cleanup_lcm, rc);
2375
2376         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2377                         &llog_lvfs_ops);
2378         if (rc)
2379                 GOTO(cleanup_olg, rc);
2380
2381         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2382         if (!ctxt) {
2383                 CERROR("Can't get ctxt for %p:%x\n", olg,
2384                        LLOG_MDS_OST_REPL_CTXT);
2385                 GOTO(cleanup_olg, rc = -ENODEV);
2386         }
2387         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2388         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2389         llog_ctxt_put(ctxt);
2390
2391         RETURN(0);
2392 cleanup_olg:
2393         filter_olg_fini(olg);
2394 cleanup_lcm:
2395         llog_recov_thread_fini(filter->fo_lcm, 1);
2396         filter->fo_lcm = NULL;
2397         return rc;
2398 }
2399
2400 static int
2401 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2402                  struct obd_device *tgt, int count, struct llog_catid *catid,
2403                  struct obd_uuid *uuid)
2404 {
2405         struct filter_obd *filter = &obd->u.filter;
2406         struct llog_ctxt *ctxt;
2407         int rc;
2408         ENTRY;
2409
2410         LASSERT(olg != NULL);
2411         if (olg == &obd->obd_olg)
2412                 return filter_default_olg_init(obd, olg, tgt);
2413
2414         LASSERT(filter->fo_lcm != NULL);
2415         rc = filter_olg_init(obd, olg, tgt);
2416         if (rc)
2417                 RETURN(rc);
2418         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2419         if (!ctxt) {
2420                 CERROR("Can't get ctxt for %p:%x\n", olg,
2421                        LLOG_MDS_OST_REPL_CTXT);
2422                 filter_olg_fini(olg);
2423                 RETURN(-ENODEV);
2424         }
2425         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2426         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2427         llog_ctxt_put(ctxt);
2428         RETURN(rc);
2429 }
2430
2431 static int filter_llog_finish(struct obd_device *obd, int count)
2432 {
2433         struct filter_obd *filter = &obd->u.filter;
2434         struct llog_ctxt *ctxt;
2435         ENTRY;
2436
2437         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2438         if (ctxt) {
2439                 /*
2440                  * Make sure that no cached llcds left in recov_thread.
2441                  * We actually do sync in disconnect time, but disconnect
2442                  * may not come being marked rq_no_resend = 1.
2443                  */
2444                 llog_sync(ctxt, NULL);
2445
2446                 /*
2447                  * Balance class_import_get() in llog_receptor_accept().
2448                  * This is safe to do, as llog is already synchronized
2449                  * and its import may go.
2450                  */
2451                 mutex_down(&ctxt->loc_sem);
2452                 if (ctxt->loc_imp) {
2453                         class_import_put(ctxt->loc_imp);
2454                         ctxt->loc_imp = NULL;
2455                 }
2456                 mutex_up(&ctxt->loc_sem);
2457                 llog_ctxt_put(ctxt);
2458         }
2459
2460         if (filter->fo_lcm) {
2461                 mutex_down(&ctxt->loc_sem);
2462                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2463                 filter->fo_lcm = NULL;
2464                 mutex_up(&ctxt->loc_sem);
2465         }
2466         RETURN(filter_olg_fini(&obd->obd_olg));
2467 }
2468
2469 /**
2470  * Find the group llog according to group index in the llog group list.
2471  */
2472 static struct obd_llog_group *
2473 filter_find_olg_internal(struct filter_obd *filter, int group)
2474 {
2475         struct obd_llog_group *olg;
2476
2477         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2478         list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2479                 if (olg->olg_group == group)
2480                         RETURN(olg);
2481         }
2482         RETURN(NULL);
2483 }
2484
2485 /**
2486  * Find the group llog according to group index on the filter
2487  */
2488 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2489 {
2490         struct obd_llog_group *olg = NULL;
2491         struct filter_obd *filter;
2492
2493         filter = &obd->u.filter;
2494
2495         if (group == FILTER_GROUP_LLOG)
2496                 RETURN(&obd->obd_olg);
2497
2498         spin_lock(&filter->fo_llog_list_lock);
2499         olg = filter_find_olg_internal(filter, group);
2500         spin_unlock(&filter->fo_llog_list_lock);
2501
2502         RETURN(olg);
2503 }
2504 /**
2505  * Find the llog_group of the filter according to the group. If it can not
2506  * find, create the llog_group, which only happens when mds is being synced
2507  * with OST.
2508  */
2509 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2510 {
2511         struct obd_llog_group *olg = NULL;
2512         struct filter_obd *filter;
2513         int rc;
2514
2515         filter = &obd->u.filter;
2516
2517         if (group == FILTER_GROUP_LLOG)
2518                 RETURN(&obd->obd_olg);
2519
2520         spin_lock(&filter->fo_llog_list_lock);
2521         olg = filter_find_olg_internal(filter, group);
2522         if (olg) {
2523                 if (olg->olg_initializing) {
2524                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2525                 } else {
2526                         GOTO(out_unlock, olg);
2527                 }
2528         }
2529         OBD_ALLOC_PTR(olg);
2530         if (olg == NULL)
2531                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2532
2533         llog_group_init(olg, group);
2534         list_add(&olg->olg_list, &filter->fo_llog_list);
2535         olg->olg_initializing = 1;
2536         spin_unlock(&filter->fo_llog_list_lock);
2537
2538         rc = llog_cat_initialize(obd, olg, 1, NULL);
2539         if (rc) {
2540                spin_lock(&filter->fo_llog_list_lock);
2541                list_del(&olg->olg_list);
2542                spin_unlock(&filter->fo_llog_list_lock);
2543                OBD_FREE_PTR(olg);
2544                GOTO(out, olg = ERR_PTR(-ENOMEM));
2545         }
2546         spin_lock(&filter->fo_llog_list_lock);
2547         olg->olg_initializing = 0;
2548         spin_unlock(&filter->fo_llog_list_lock);
2549         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2550               obd->obd_name, group, olg);
2551 out:
2552         RETURN(olg);
2553
2554 out_unlock:
2555         spin_unlock(&filter->fo_llog_list_lock);
2556         GOTO(out, olg);
2557 }
2558
2559 static int filter_llog_connect(struct obd_export *exp,
2560                                struct llogd_conn_body *body)
2561 {
2562         struct obd_device *obd = exp->exp_obd;
2563         struct llog_ctxt *ctxt;
2564         struct obd_llog_group *olg;
2565         int rc;
2566         ENTRY;
2567
2568         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2569                obd->obd_name, body->lgdc_logid.lgl_oid,
2570                body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2571
2572         olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2573         if (!olg) {
2574                 CERROR(" %s: can not find olg of group %d\n",
2575                        obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2576                 RETURN(-ENOENT);
2577         }
2578         llog_group_set_export(olg, exp);
2579
2580         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2581         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2582                  body->lgdc_ctxt_idx);
2583
2584         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2585               obd->obd_name, body->lgdc_logid.lgl_oid,
2586               body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2587
2588         spin_lock_bh(&obd->obd_processing_task_lock);
2589         obd->u.filter.fo_mds_ost_sync = 1;
2590         spin_unlock_bh(&obd->obd_processing_task_lock);
2591         rc = llog_connect(ctxt, &body->lgdc_logid,
2592                           &body->lgdc_gen, NULL);
2593         llog_ctxt_put(ctxt);
2594         if (rc != 0)
2595                 CERROR("failed to connect rc %d idx %d\n", rc,
2596                                 body->lgdc_ctxt_idx);
2597
2598         RETURN(rc);
2599 }
2600
2601 static int filter_llog_preclean(struct obd_device *obd)
2602 {
2603         struct obd_llog_group *olg, *tmp;
2604         struct filter_obd *filter;
2605         struct list_head  remove_list;
2606         int rc = 0;
2607         ENTRY;
2608
2609         rc = obd_llog_finish(obd, 0);
2610         if (rc)
2611                 CERROR("failed to cleanup llogging subsystem\n");
2612
2613         filter = &obd->u.filter;
2614         CFS_INIT_LIST_HEAD(&remove_list);
2615
2616         spin_lock(&filter->fo_llog_list_lock);
2617         while (!list_empty(&filter->fo_llog_list)) {
2618                 olg = list_entry(filter->fo_llog_list.next,
2619                                  struct obd_llog_group, olg_list);
2620                 list_del(&olg->olg_list);
2621                 list_add(&olg->olg_list, &remove_list);
2622         }
2623         spin_unlock(&filter->fo_llog_list_lock);
2624
2625         list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2626                 list_del_init(&olg->olg_list);
2627                 rc = filter_olg_fini(olg);
2628                 if (rc)
2629                         CERROR("failed to cleanup llogging subsystem for %u\n",
2630                                olg->olg_group);
2631                 OBD_FREE_PTR(olg);
2632         }
2633
2634         RETURN(rc);
2635 }
2636
2637 static int filter_precleanup(struct obd_device *obd,
2638                              enum obd_cleanup_stage stage)
2639 {
2640         int rc = 0;
2641         ENTRY;
2642
2643         switch(stage) {
2644         case OBD_CLEANUP_EARLY:
2645                 break;
2646         case OBD_CLEANUP_EXPORTS:
2647                 /* Stop recovery before namespace cleanup. */
2648                 target_recovery_fini(obd);
2649                 rc = filter_llog_preclean(obd);
2650                 break;
2651         }
2652         RETURN(rc);
2653 }
2654
2655 static int filter_cleanup(struct obd_device *obd)
2656 {
2657         struct filter_obd *filter = &obd->u.filter;
2658         ENTRY;
2659
2660         if (obd->obd_fail)
2661                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2662                               "will be preserved.\n", obd->obd_name);
2663
2664         obd_exports_barrier(obd);
2665         obd_zombie_barrier();
2666
2667         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2668         lprocfs_free_per_client_stats(obd);
2669         lprocfs_free_obd_stats(obd);
2670         lprocfs_obd_cleanup(obd);
2671         lquota_cleanup(filter_quota_interface_ref, obd);
2672
2673         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2674         obd->obd_namespace = NULL;
2675
2676         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2677
2678         if (obd->u.obt.obt_sb == NULL)
2679                 RETURN(0);
2680
2681         filter_post(obd);
2682
2683         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2684         shrink_dcache_sb(obd->u.obt.obt_sb);
2685
2686         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2687         obd->u.obt.obt_sb = NULL;
2688
2689         fsfilt_put_ops(obd->obd_fsops);
2690
2691         filter_iobuf_pool_done(filter);
2692
2693         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2694
2695         RETURN(0);
2696 }
2697
2698 static int filter_connect_internal(struct obd_export *exp,
2699                                    struct obd_connect_data *data)
2700 {
2701         struct filter_export_data *fed = &exp->exp_filter_data;
2702
2703         if (!data)
2704                 RETURN(0);
2705
2706         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2707                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2708                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2709                data->ocd_connect_flags, data->ocd_version,
2710                data->ocd_grant, data->ocd_index);
2711
2712         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2713                 CWARN("!!! This export (nid %s) used object group %d "
2714                        "earlier; now it's trying to use group %d!  This could "
2715                        "be a bug in the MDS. Please report to "
2716                        "http://bugzilla.lustre.org/\n",
2717                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2718                 RETURN(-EPROTO);
2719         }
2720         fed->fed_group = data->ocd_group;
2721
2722         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2723         exp->exp_connect_flags = data->ocd_connect_flags;
2724         data->ocd_version = LUSTRE_VERSION_CODE;
2725
2726         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2727         if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2728                   data->ocd_connect_flags & OBD_CONNECT_MDS))
2729                 RETURN(-EPROTO);
2730
2731         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2732                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2733                 obd_size left, want;
2734
2735                 spin_lock(&exp->exp_obd->obd_osfs_lock);
2736                 left = filter_grant_space_left(exp);
2737                 want = data->ocd_grant;
2738                 filter_grant(exp, fed->fed_grant, want, left);
2739                 data->ocd_grant = fed->fed_grant;
2740                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2741
2742                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2743                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2744                        exp->exp_client_uuid.uuid, exp,
2745                        data->ocd_grant, want, left);
2746                 
2747                 filter->fo_tot_granted_clients ++;
2748         }
2749
2750         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2751                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2752                 struct lr_server_data *lsd = filter->fo_fsd;
2753                 int index = le32_to_cpu(lsd->lsd_ost_index);
2754
2755                 if (!(lsd->lsd_feature_compat &
2756                       cpu_to_le32(OBD_COMPAT_OST))) {
2757                         /* this will only happen on the first connect */
2758                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2759                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2760                         filter_update_server_data(exp->exp_obd,
2761                                                   filter->fo_rcvd_filp, lsd, 1);
2762                 } else if (index != data->ocd_index) {
2763                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2764                                            " %u doesn't match actual OST index"
2765                                            " %u in last_rcvd file, bad "
2766                                            "configuration?\n",
2767                                            obd_export_nid2str(exp), index,
2768                                            data->ocd_index);
2769                         RETURN(-EBADF);
2770                 }
2771                 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2772                  * FIXME: We don't strictly need the COMPAT flag for that,
2773                  * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2774                  * FIXME: We needed it for the index, as index 0 is valid. */
2775         }
2776
2777         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2778                 data->ocd_brw_size = 65536;
2779         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2780                 data->ocd_brw_size = min(data->ocd_brw_size,
2781                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2782                 LASSERT(data->ocd_brw_size);
2783         }
2784
2785         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2786                 __u32 cksum_types = data->ocd_cksum_types;
2787
2788                 /* The client set in ocd_cksum_types the checksum types it
2789                  * supports. We have to mask off the algorithms that we don't
2790                  * support */
2791                 if (cksum_types & OBD_CKSUM_ALL)
2792                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2793                 else
2794                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2795
2796                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2797                                    "%x\n", exp->exp_obd->obd_name,
2798                                    obd_export_nid2str(exp), cksum_types,
2799                                    data->ocd_cksum_types);
2800         } else {
2801                 /* This client does not support OBD_CONNECT_CKSUM
2802                  * fall back to CRC32 */
2803                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2804                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2805                                    exp->exp_obd->obd_name,
2806                                    obd_export_nid2str(exp));
2807         }
2808
2809         RETURN(0);
2810 }
2811
2812 static int filter_reconnect(const struct lu_env *env,
2813                             struct obd_export *exp, struct obd_device *obd,
2814                             struct obd_uuid *cluuid,
2815                             struct obd_connect_data *data,
2816                             void *localdata)
2817 {
2818         int rc;
2819         ENTRY;
2820
2821         if (exp == NULL || obd == NULL || cluuid == NULL)
2822                 RETURN(-EINVAL);
2823
2824         rc = filter_connect_internal(exp, data);
2825         if (rc == 0)
2826                 filter_export_stats_init(obd, exp, localdata);
2827
2828         RETURN(rc);
2829 }
2830
2831 /* nearly identical to mds_connect */
2832 static int filter_connect(const struct lu_env *env,
2833                           struct obd_export **exp, struct obd_device *obd,
2834                           struct obd_uuid *cluuid,
2835                           struct obd_connect_data *data, void *localdata)
2836 {
2837         struct lvfs_run_ctxt saved;
2838         struct lustre_handle conn = { 0 };
2839         struct obd_export *lexp;
2840         struct filter_export_data *fed;
2841         struct lsd_client_data *lcd = NULL;
2842         __u32 group;
2843         int rc;
2844         ENTRY;
2845
2846         if (exp == NULL || obd == NULL || cluuid == NULL)
2847                 RETURN(-EINVAL);
2848
2849         rc = class_connect(&conn, obd, cluuid);
2850         if (rc)
2851                 RETURN(rc);
2852         lexp = class_conn2export(&conn);
2853         LASSERT(lexp != NULL);
2854
2855         fed = &lexp->exp_filter_data;
2856
2857         rc = filter_connect_internal(lexp, data);
2858         if (rc)
2859                 GOTO(cleanup, rc);
2860
2861         filter_export_stats_init(obd, lexp, localdata);
2862         if (obd->obd_replayable) {
2863                 OBD_ALLOC(lcd, sizeof(*lcd));
2864                 if (!lcd) {
2865                         CERROR("filter: out of memory for client data\n");
2866                         GOTO(cleanup, rc = -ENOMEM);
2867                 }
2868
2869                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2870                 fed->fed_lcd = lcd;
2871                 rc = filter_client_add(obd, lexp, -1);
2872                 if (rc)
2873                         GOTO(cleanup, rc);
2874         }
2875
2876         group = data->ocd_group;
2877
2878         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2879               obd->obd_name, lexp->exp_handle.h_cookie, group);
2880
2881         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2882         rc = filter_read_groups(obd, group, 1);
2883         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2884         if (rc != 0) {
2885                 CERROR("can't read group %u\n", group);
2886                 GOTO(cleanup, rc);
2887         }
2888
2889         GOTO(cleanup, rc);
2890
2891 cleanup:
2892         if (rc) {
2893                 if (lcd) {
2894                         OBD_FREE_PTR(lcd);
2895                         fed->fed_lcd = NULL;
2896                 }
2897                 class_disconnect(lexp);
2898                 *exp = NULL;
2899         } else {
2900                 *exp = lexp;
2901         }
2902
2903         RETURN(rc);
2904 }
2905
2906 /* Do extra sanity checks for grant accounting.  We do this at connect,
2907  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2908  * always get rid of it or turn it off when we know accounting is good. */
2909 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2910 {
2911         struct filter_export_data *fed;
2912         struct obd_export *exp;
2913         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2914         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2915         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2916
2917         if (list_empty(&obd->obd_exports))
2918                 return;
2919
2920         /* We don't want to do this for large machines that do lots of
2921            mounts or unmounts.  It burns... */
2922         if (obd->obd_num_exports > 100)
2923                 return;
2924
2925         spin_lock(&obd->obd_osfs_lock);
2926         spin_lock(&obd->obd_dev_lock);
2927         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2928                 int error = 0;
2929                 fed = &exp->exp_filter_data;
2930                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2931                     fed->fed_dirty < 0)
2932                         error = 1;
2933                 if (maxsize > 0) { /* we may not have done a statfs yet */
2934                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2935                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2936                                  exp->exp_client_uuid.uuid, exp,
2937                                  fed->fed_grant, fed->fed_pending, maxsize);
2938                         LASSERTF(fed->fed_dirty <= maxsize,
2939                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2940                                  exp->exp_client_uuid.uuid, exp,
2941                                  fed->fed_dirty, maxsize);
2942                 }
2943                 if (error)
2944                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2945                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2946                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2947                 else
2948                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2949                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2950                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2951                 tot_granted += fed->fed_grant + fed->fed_pending;
2952                 tot_pending += fed->fed_pending;
2953                 tot_dirty += fed->fed_dirty;
2954         }
2955         fo_tot_granted = obd->u.filter.fo_tot_granted;
2956         fo_tot_pending = obd->u.filter.fo_tot_pending;
2957         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2958         spin_unlock(&obd->obd_dev_lock);
2959         spin_unlock(&obd->obd_osfs_lock);
2960
2961         /* Do these assertions outside the spinlocks so we don't kill system */
2962         if (tot_granted != fo_tot_granted)
2963                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2964                        func, tot_granted, fo_tot_granted);
2965         if (tot_pending != fo_tot_pending)
2966                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2967                        func, tot_pending, fo_tot_pending);
2968         if (tot_dirty != fo_tot_dirty)
2969                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2970                        func, tot_dirty, fo_tot_dirty);
2971         if (tot_pending > tot_granted)
2972                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2973                        func, tot_pending, tot_granted);
2974         if (tot_granted > maxsize)
2975                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2976                        func, tot_granted, maxsize);
2977         if (tot_dirty > maxsize)
2978                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2979                        func, tot_dirty, maxsize);
2980 }
2981
2982 /* Remove this client from the grant accounting totals.  We also remove
2983  * the export from the obd device under the osfs and dev locks to ensure
2984  * that the filter_grant_sanity_check() calculations are always valid.
2985  * The client should do something similar when it invalidates its import. */
2986 static void filter_grant_discard(struct obd_export *exp)
2987 {
2988         struct obd_device *obd = exp->exp_obd;
2989         struct filter_obd *filter = &obd->u.filter;
2990         struct filter_export_data *fed = &exp->exp_filter_data;
2991
2992         spin_lock(&obd->obd_osfs_lock);
2993         spin_lock(&obd->obd_dev_lock);
2994         list_del_init(&exp->exp_obd_chain);
2995         spin_unlock(&obd->obd_dev_lock);
2996
2997         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2998                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2999                  obd->obd_name, filter->fo_tot_granted,
3000                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
3001         filter->fo_tot_granted -= fed->fed_grant;
3002         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
3003                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
3004                  obd->obd_name, filter->fo_tot_pending,
3005                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
3006         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
3007         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
3008                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
3009                  obd->obd_name, filter->fo_tot_dirty,
3010                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
3011         filter->fo_tot_dirty -= fed->fed_dirty;
3012         fed->fed_dirty = 0;
3013         fed->fed_grant = 0;
3014
3015         spin_unlock(&obd->obd_osfs_lock);
3016 }
3017
3018 static int filter_destroy_export(struct obd_export *exp)
3019 {
3020         ENTRY;
3021
3022         if (exp->exp_filter_data.fed_pending)
3023                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
3024                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
3025                        exp, exp->exp_filter_data.fed_pending);
3026
3027         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3028
3029         target_destroy_export(exp);
3030         ldlm_destroy_export(exp);
3031
3032         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
3033                 RETURN(0);
3034
3035
3036         if (exp->exp_obd->obd_replayable)
3037                 filter_client_free(exp);
3038         else
3039                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3040
3041         filter_grant_discard(exp);
3042         filter_fmd_cleanup(exp);
3043
3044         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3045                 struct filter_obd *filter = &exp->exp_obd->u.filter;
3046                 if (filter->fo_tot_granted_clients > 0)
3047                         filter->fo_tot_granted_clients --;
3048         }
3049
3050         if (!(exp->exp_flags & OBD_OPT_FORCE))
3051                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3052
3053         RETURN(0);
3054 }
3055
3056 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3057 {
3058         struct obd_llog_group *olg_min, *olg;
3059         struct filter_obd *filter;
3060         int worked = -1, group;
3061         struct llog_ctxt *ctxt;
3062         ENTRY;
3063
3064         filter = &obd->u.filter;
3065
3066         /* we can't sync log holding spinlock. also, we do not want to get
3067          * into livelock. so we do following: loop over MDS's exports in
3068          * group order and skip already synced llogs -bzzz */
3069         do {
3070                 /* look for group with min. number, but > worked */
3071                 olg_min = NULL;
3072                 group = 1 << 30;
3073                 spin_lock(&filter->fo_llog_list_lock);
3074                 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3075                         if (olg->olg_group <= worked) {
3076                                 /* this group is already synced */
3077                                 continue;
3078                         }
3079                         if (group < olg->olg_group) {
3080                                 /* we have group with smaller number to sync */
3081                                 continue;
3082                         }
3083                         /* store current minimal group */
3084                         olg_min = olg;
3085                         group = olg->olg_group;
3086                 }
3087                 spin_unlock(&filter->fo_llog_list_lock);
3088
3089                 if (olg_min == NULL)
3090                         break;
3091
3092                 worked = olg_min->olg_group;
3093                 if (olg_min->olg_exp &&
3094                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3095                         int err;
3096                         ctxt = llog_group_get_ctxt(olg_min,
3097                                                    LLOG_MDS_OST_REPL_CTXT);
3098                         if (ctxt) {
3099                                 err = llog_sync(ctxt, olg_min->olg_exp);
3100                                 llog_ctxt_put(ctxt);
3101                                 if (err) {
3102                                         CERROR("error flushing logs to MDS: "
3103                                                "rc %d\n", err);
3104                                 }
3105                         }
3106                 }
3107         } while (olg_min != NULL);
3108 }
3109
3110 /* Also incredibly similar to mds_disconnect */
3111 static int filter_disconnect(struct obd_export *exp)
3112 {
3113         struct obd_device *obd = exp->exp_obd;
3114         int rc;
3115         ENTRY;
3116
3117         LASSERT(exp);
3118         class_export_get(exp);
3119
3120         if (!(exp->exp_flags & OBD_OPT_FORCE))
3121                 filter_grant_sanity_check(obd, __FUNCTION__);
3122         filter_grant_discard(exp);
3123
3124         /* Flush any remaining cancel messages out to the target */
3125         filter_sync_llogs(obd, exp);
3126
3127         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3128
3129         /* Disconnect early so that clients can't keep using export */
3130         rc = class_disconnect(exp);
3131         if (exp->exp_obd->obd_namespace != NULL)
3132                 ldlm_cancel_locks_for_export(exp);
3133
3134         fsfilt_sync(obd, obd->u.obt.obt_sb);
3135
3136         lprocfs_exp_cleanup(exp);
3137
3138         class_export_put(exp);
3139         RETURN(rc);
3140 }
3141
3142 /* reverse import is changed, sync all cancels */
3143 static void filter_revimp_update(struct obd_export *exp)
3144 {
3145         ENTRY;
3146
3147         LASSERT(exp);
3148         class_export_get(exp);
3149
3150         /* flush any remaining cancel messages out to the target */
3151         filter_sync_llogs(exp->exp_obd, exp);
3152         class_export_put(exp);
3153         EXIT;
3154 }
3155
3156 static int filter_ping(struct obd_export *exp)
3157 {
3158         filter_fmd_expire(exp);
3159         return 0;
3160 }
3161
3162 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3163                                   const char *what, int quiet)
3164 {
3165         struct dentry *dchild = NULL;
3166         obd_gr group = 0;
3167
3168         if (oa->o_valid & OBD_MD_FLGROUP)
3169                 group = oa->o_gr;
3170
3171         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3172
3173         if (IS_ERR(dchild)) {
3174                 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3175                        what, group, oa->o_id);
3176                 RETURN(dchild);
3177         }
3178
3179         if (dchild->d_inode == NULL) {
3180                 if (!quiet)
3181                         CERROR("%s: %s on non-existent object: "LPU64"\n",
3182                                obd->obd_name, what, oa->o_id);
3183                 f_dput(dchild);
3184                 RETURN(ERR_PTR(-ENOENT));
3185         }
3186
3187         return dchild;
3188 }
3189
3190 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3191 {
3192         struct dentry *dentry = NULL;
3193         struct obd_device *obd;
3194         int rc = 0;
3195         ENTRY;
3196
3197         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3198                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3199         if (rc)
3200                 RETURN(rc);
3201
3202         obd = class_exp2obd(exp);
3203         if (obd == NULL) {
3204                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3205                 RETURN(-EINVAL);
3206         }
3207
3208         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3209         if (IS_ERR(dentry))
3210                 RETURN(PTR_ERR(dentry));
3211
3212         /* Limit the valid bits in the return data to what we actually use */
3213         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3214         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3215
3216         f_dput(dentry);
3217         RETURN(rc);
3218 }
3219
3220 /* this should be enabled/disabled in condition to enabled/disabled large
3221  * inodes (fast EAs) in backing store FS. */
3222 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3223                         void *handle, struct obdo *oa)
3224 {
3225         struct obd_device *obd = exp->exp_obd;
3226         int rc = 0;
3227         ENTRY;
3228
3229         if (oa->o_valid & OBD_MD_FLFID) {
3230                 struct filter_fid ff;
3231
3232                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3233                         oa->o_gr = 0;
3234                 /* packing fid and converting it to LE for storing into EA.
3235                  * Here ->o_stripe_idx should be filled by LOV and rest of
3236                  * fields - by client. */
3237                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3238                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3239                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3240                 ff.ff_objid = cpu_to_le64(oa->o_id);
3241                 ff.ff_group = cpu_to_le64(oa->o_gr);
3242
3243                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3244                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3245                        oa->o_generation, oa->o_id, oa->o_gr);
3246
3247                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3248                 if (rc)
3249                         CERROR("store fid in object failed! rc: %d\n", rc);
3250         } else {
3251                 CDEBUG(D_HA, "OSS object without fid info!\n");
3252         }
3253
3254         RETURN(rc);
3255 }
3256
3257 /* this is called from filter_truncate() until we have filter_punch() */
3258 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3259                             struct obdo *oa, struct obd_trans_info *oti)
3260 {
3261         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3262         struct llog_cookie *fcc = NULL;
3263         struct filter_obd *filter;
3264         int rc, err, sync = 0;
3265         loff_t old_size = 0;
3266         unsigned int ia_valid;
3267         struct inode *inode;
3268         struct iattr iattr;
3269         void *handle;
3270         ENTRY;
3271
3272         LASSERT(dentry != NULL);
3273         LASSERT(!IS_ERR(dentry));
3274
3275         inode = dentry->d_inode;
3276         LASSERT(inode != NULL);
3277
3278         filter = &exp->exp_obd->u.filter;
3279         iattr_from_obdo(&iattr, oa, oa->o_valid);
3280         ia_valid = iattr.ia_valid;
3281
3282         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3283                 OBD_ALLOC(fcc, sizeof(*fcc));
3284                 if (fcc != NULL)
3285                         *fcc = oa->o_lcookie;
3286         }
3287         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3288                 DQUOT_INIT(inode);
3289                 /* Filter truncates and writes are serialized by
3290                  * i_alloc_sem, see the comment in
3291                  * filter_preprw_write.*/
3292                 if (ia_valid & ATTR_SIZE)
3293                         down_write(&inode->i_alloc_sem);
3294                 LOCK_INODE_MUTEX(inode);
3295                 old_size = i_size_read(inode);
3296         }
3297
3298         /* VBR: version recovery check */
3299         rc = filter_version_get_check(exp, oti, inode);
3300         if (rc)
3301                 GOTO(out_unlock, rc);
3302
3303         /* If the inode still has SUID+SGID bits set (see filter_precreate())
3304          * then we will accept the UID+GID sent by the client during write for
3305          * initializing the ownership of this inode.  We only allow this to
3306          * happen once so clear these bits in setattr. In 2.6 kernels it is
3307          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3308          * the flags that are actually being set. */
3309         if (ia_valid & (ATTR_UID | ATTR_GID)) {
3310                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3311                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3312
3313                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3314                         if (!(ia_valid & ATTR_MODE)) {
3315                                 iattr.ia_mode = inode->i_mode;
3316                                 iattr.ia_valid |= ATTR_MODE;
3317                         }
3318                         iattr.ia_mode &= ~S_ISUID;
3319                 }
3320                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3321                         if (!(iattr.ia_valid & ATTR_MODE)) {
3322                                 iattr.ia_mode = inode->i_mode;
3323                                 iattr.ia_valid |= ATTR_MODE;
3324                         }
3325                         iattr.ia_mode &= ~S_ISGID;
3326                 }
3327
3328                 orig_ids[USRQUOTA] = inode->i_uid;
3329                 orig_ids[GRPQUOTA] = inode->i_gid;
3330                 handle = fsfilt_start_log(exp->exp_obd, inode,
3331                                           FSFILT_OP_SETATTR, oti, 1);
3332                 if (IS_ERR(handle))
3333                         GOTO(out_unlock, rc = PTR_ERR(handle));
3334
3335                 /* update inode EA only once when inode is suid bit marked. As
3336                  * on 2.6.x UID and GID may be set separately, we check here
3337                  * only one of them to avoid double setting. */
3338                 if (inode->i_mode & S_ISUID)
3339                         filter_update_fidea(exp, inode, handle, oa);
3340         } else {
3341                 handle = fsfilt_start(exp->exp_obd, inode,
3342                                       FSFILT_OP_SETATTR, oti);
3343                 if (IS_ERR(handle))
3344                         GOTO(out_unlock, rc = PTR_ERR(handle));
3345         }
3346         if (oa->o_valid & OBD_MD_FLFLAGS) {
3347                 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3348                                       EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3349         } else {
3350                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3351                 if (fcc != NULL)
3352                         /* set cancel cookie callback function */
3353                         sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3354                                                      filter_cancel_cookies_cb,
3355                                                      fcc);
3356         }
3357
3358         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3359                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3360
3361         /* The truncate might have used up our transaction credits.  Make
3362          * sure we have one left for the last_rcvd update. */
3363         err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3364
3365         rc = filter_finish_transno(exp, inode, oti, rc, sync);
3366         if (sync) {
3367                 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3368                 fcc = NULL;
3369         }
3370
3371         err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3372         if (err) {
3373                 CERROR("error on commit, err = %d\n", err);
3374                 if (!rc)
3375                         rc = err;
3376         } else {
3377                 fcc = NULL;
3378         }
3379
3380         /* For a partial-page truncate flush the page to disk immediately
3381          * to avoid data corruption during direct disk write. b=17397 */
3382         if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3383             old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3384                 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3385                                                iattr.ia_size + 1);
3386                 if (!rc)
3387                         rc = err;
3388         }
3389
3390         EXIT;
3391 out_unlock:
3392         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3393                 UNLOCK_INODE_MUTEX(inode);
3394         if (ia_valid & ATTR_SIZE)
3395                 up_write(&inode->i_alloc_sem);
3396         if (fcc)
3397                 OBD_FREE(fcc, sizeof(*fcc));
3398
3399         /* trigger quota release */
3400         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3401                 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3402                 int rc2 = lquota_adjust(filter_quota_interface_ref,
3403                                         exp->exp_obd, cur_ids,
3404                                         orig_ids, rc, FSFILT_OP_SETATTR);
3405                 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3406                        "filter adjust qunit. (rc:%d)\n", rc2);
3407         }
3408         return rc;
3409 }
3410
3411 /* this is called from filter_truncate() until we have filter_punch() */
3412 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3413                    struct obd_trans_info *oti)
3414 {
3415         struct obdo *oa = oinfo->oi_oa;
3416         struct lustre_capa *capa = oinfo_capa(oinfo);
3417         struct ldlm_res_id res_id;
3418         struct filter_mod_data *fmd;
3419         struct lvfs_run_ctxt saved;
3420         struct filter_obd *filter;
3421         struct ldlm_resource *res;
3422         struct dentry *dentry;
3423         __u64 opc = CAPA_OPC_META_WRITE;
3424         int rc;
3425         ENTRY;
3426
3427         if (oa->o_valid & OBD_FL_TRUNC)
3428                 opc |= CAPA_OPC_OSS_TRUNC;
3429         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa, opc);
3430         if (rc)
3431                 RETURN(rc);
3432
3433         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3434                 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
3435                 if (rc)
3436                         RETURN(rc);
3437         }
3438
3439         osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3440         /* This would be very bad - accidentally truncating a file when
3441          * changing the time or similar - bug 12203. */
3442         if (oa->o_valid & OBD_MD_FLSIZE &&
3443             oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3444                 static char mdsinum[48];
3445
3446                 if (oa->o_valid & OBD_MD_FLFID)
3447                         snprintf(mdsinum, sizeof(mdsinum) - 1,
3448                                  " of inode "LPU64"/%u", oa->o_fid,
3449                                  oa->o_generation);
3450                 else
3451                         mdsinum[0] = '\0';
3452
3453                 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3454                        " %s\n",
3455                        exp->exp_obd->obd_name, obd_export_nid2str(exp),
3456                        oa->o_id, mdsinum);
3457                 RETURN(-EPERM);
3458         }
3459
3460         dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3461         if (IS_ERR(dentry))
3462                 RETURN(PTR_ERR(dentry));
3463
3464         filter = &exp->exp_obd->u.filter;
3465         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3466         lock_kernel();
3467
3468         if (oa->o_valid &
3469             (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3470                 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3471                 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3472                         fmd->fmd_mactime_xid = oti->oti_xid;
3473                 filter_fmd_put(exp, fmd);
3474         }
3475
3476         /* setting objects attributes (including owner/group) */
3477         rc = filter_setattr_internal(exp, dentry, oa, oti);
3478         if (rc)
3479                 GOTO(out_unlock, rc);
3480
3481         res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3482                                 &res_id, LDLM_EXTENT, 0);
3483
3484         if (res != NULL) {
3485                 LDLM_RESOURCE_ADDREF(res);
3486                 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3487                 LDLM_RESOURCE_DELREF(res);
3488                 ldlm_resource_putref(res);
3489         }
3490
3491         oa->o_valid = OBD_MD_FLID;
3492
3493         /* Quota release need uid/gid info */
3494         obdo_from_inode(oa, dentry->d_inode,
3495                         FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3496
3497         EXIT;
3498 out_unlock:
3499         unlock_kernel();
3500         f_dput(dentry);
3501         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3502         return rc;
3503 }
3504
3505 /* XXX identical to osc_unpackmd */
3506 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3507                            struct lov_mds_md *lmm, int lmm_bytes)
3508 {
3509         int lsm_size;
3510         ENTRY;
3511
3512         if (lmm != NULL) {
3513                 if (lmm_bytes < sizeof (*lmm)) {
3514                         CERROR("lov_mds_md too small: %d, need %d\n",
3515                                lmm_bytes, (int)sizeof(*lmm));
3516                         RETURN(-EINVAL);
3517                 }
3518                 /* XXX LOV_MAGIC etc check? */
3519
3520                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3521                         CERROR("lov_mds_md: zero lmm_object_id\n");
3522                         RETURN(-EINVAL);
3523                 }
3524         }
3525
3526         lsm_size = lov_stripe_md_size(1);
3527         if (lsmp == NULL)
3528                 RETURN(lsm_size);
3529
3530         if (*lsmp != NULL && lmm == NULL) {
3531                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3532                 OBD_FREE(*lsmp, lsm_size);
3533                 *lsmp = NULL;
3534                 RETURN(0);
3535         }
3536
3537         if (*lsmp == NULL) {
3538                 OBD_ALLOC(*lsmp, lsm_size);
3539                 if (*lsmp == NULL)
3540                         RETURN(-ENOMEM);
3541
3542                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3543                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3544                         OBD_FREE(*lsmp, lsm_size);
3545                         RETURN(-ENOMEM);
3546                 }
3547                 loi_init((*lsmp)->lsm_oinfo[0]);
3548         }
3549
3550         if (lmm != NULL) {
3551                 /* XXX zero *lsmp? */
3552                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3553                 LASSERT((*lsmp)->lsm_object_id);
3554         }
3555
3556         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3557
3558         RETURN(lsm_size);
3559 }
3560
3561 /* caller must hold fo_create_locks[oa->o_gr] */
3562 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3563                                      struct filter_obd *filter)
3564 {
3565         struct obdo doa = { 0 }; /* XXX obdo on stack */
3566         obd_id last, id;
3567         int rc = 0;
3568         int skip_orphan;
3569         ENTRY;
3570
3571         LASSERT(oa);
3572         LASSERT_MDS_GROUP(oa->o_gr);
3573         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3574         LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3575
3576         memset(&doa, 0, sizeof(doa));
3577
3578         doa.o_valid |= OBD_MD_FLGROUP;
3579         doa.o_gr = oa->o_gr;
3580         doa.o_mode = S_IFREG;
3581
3582         if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3583                 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3584                        exp->exp_obd->obd_name, doa.o_gr);
3585                 RETURN(0);
3586         }
3587
3588         last = filter_last_id(filter, doa.o_gr);
3589
3590         skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3591
3592         CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3593                exp->exp_obd->obd_name, oa->o_id + 1, last,
3594                skip_orphan ? ", orphan objids won't be reused any more." : ".");
3595
3596         for (id = last; id > oa->o_id; id--) {
3597                 doa.o_id = id;
3598                 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3599                 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3600                         CEMERG("error destroying precreate objid "LPU64": %d\n",
3601                                id, rc);
3602
3603                 /* update last_id on disk periodically so that if we restart
3604                  * we don't need to re-scan all of the just-deleted objects. */
3605                 if ((id & 511) == 0 && !skip_orphan) {
3606                         filter_set_last_id(filter, id - 1, doa.o_gr);
3607                         filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3608                 }
3609         }
3610
3611         CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3612                exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3613
3614         if (!skip_orphan) {
3615                 filter_set_last_id(filter, id, doa.o_gr);
3616                 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3617         } else {
3618                 /* don't reuse orphan object, return last used objid */
3619                 oa->o_id = last;
3620                 rc = 0;
3621         }
3622         clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3623
3624         RETURN(rc);
3625 }
3626
3627 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3628                             obd_gr group, int *num);
3629 /* returns a negative error or a nonnegative number of files to create */
3630 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3631                                    obd_gr group, struct obd_trans_info *oti)
3632 {
3633         struct obd_device *obd = exp->exp_obd;
3634         struct filter_obd *filter = &obd->u.filter;
3635         int diff, rc;
3636         ENTRY;
3637
3638         /* delete orphans request */
3639         if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3640                 obd_id last = filter_last_id(filter, group);
3641
3642                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3643                         CERROR("%s: dropping old orphan cleanup request\n",
3644                                obd->obd_name);
3645                         RETURN(0);
3646                 }
3647                 /* This causes inflight precreates to abort and drop lock */
3648                 set_bit(group, &filter->fo_destroys_in_progress);
3649                 down(&filter->fo_create_locks[group]);
3650                 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3651                         CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3652                                exp->exp_obd->obd_name, group);
3653                         up(&filter->fo_create_locks[group]);
3654                         RETURN(0);
3655                 }
3656                 diff = oa->o_id - last;
3657                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3658                        last, diff);
3659
3660                 if (-diff > OST_MAX_PRECREATE) {
3661                         CERROR("%s: ignoring bogus orphan destroy request: "
3662                                "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3663                                oa->o_id, last);
3664                         /* FIXME: should reset precreate_next_id on MDS */
3665                         GOTO(out, rc = -EINVAL);
3666                 }
3667                 if (diff < 0) {
3668                         rc = filter_destroy_precreated(exp, oa, filter);
3669                         if (rc)
3670                                 CERROR("%s: unable to write lastobjid, but "
3671                                        "orphans were deleted\n", obd->obd_name);
3672                         GOTO(out, rc);
3673                 } else {
3674                         /* XXX: Used by MDS for the first time! */
3675                         clear_bit(group, &filter->fo_destroys_in_progress);
3676                 }
3677         } else {