Whamcloud - gitweb
9dc6584449488afcb3676f812b2c3e9acd92e0c1
[fs/lustre-release.git] / lustre / obdfilter / filter.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
30  * Use is subject to license terms.
31  *
32  * Copyright (c) 2011, 2012, Whamcloud, Inc.
33  */
34 /*
35  * This file is part of Lustre, http://www.lustre.org/
36  * Lustre is a trademark of Sun Microsystems, Inc.
37  *
38  * lustre/obdfilter/filter.c
39  *
40  * Author: Peter Braam <braam@clusterfs.com>
41  * Author: Andreas Dilger <adilger@clusterfs.com>
42  */
43
44 /*
45  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
46  *            (which need to get journal_lock, may block if journal full).
47  *
48  * Invariant: Call filter_start_transno() before any journal ops to avoid the
49  *            same deadlock problem.  We can (and want) to get rid of the
50  *            transno sem in favour of the dir/inode i_mutex to avoid single
51  *            threaded operation on the OST.
52  */
53
54 #define DEBUG_SUBSYSTEM S_FILTER
55
56 #ifndef AUTOCONF_INCLUDED
57 #include <linux/config.h>
58 #endif
59 #include <linux/module.h>
60 #include <linux/fs.h>
61 #include <linux/dcache.h>
62 #include <linux/init.h>
63 #include <linux/version.h>
64 #include <linux/sched.h>
65 #include <linux/mount.h>
66 #include <linux/buffer_head.h>
67
68 #include <obd_cksum.h>
69 #include <obd_class.h>
70 #include <obd_lov.h>
71 #include <lustre_dlm.h>
72 #include <lustre_fsfilt.h>
73 #include <lprocfs_status.h>
74 #include <lustre_log.h>
75 #include <libcfs/list.h>
76 #include <lustre_disk.h>
77 #include <lustre_quota.h>
78 #include <linux/slab.h>
79 #include <lustre_param.h>
80 #include <lustre/ll_fiemap.h>
81
82 #include "filter_internal.h"
83
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         class_export_cb_put(exp);
94 }
95
96 int filter_version_get_check(struct obd_export *exp,
97                              struct obd_trans_info *oti, struct inode *inode)
98 {
99         __u64 curr_version;
100
101         if (inode == NULL || oti == NULL)
102                 RETURN(0);
103
104         curr_version = fsfilt_get_version(exp->exp_obd, inode);
105         if ((__s64)curr_version == -EOPNOTSUPP)
106                 RETURN(0);
107         /* VBR: version is checked always because costs nothing */
108         if (oti->oti_pre_version != 0 &&
109             oti->oti_pre_version != curr_version) {
110                 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
111                        oti->oti_pre_version, curr_version);
112                 cfs_spin_lock(&exp->exp_lock);
113                 exp->exp_vbr_failed = 1;
114                 cfs_spin_unlock(&exp->exp_lock);
115                 RETURN (-EOVERFLOW);
116         }
117         oti->oti_pre_version = curr_version;
118         RETURN(0);
119 }
120
121 /* Assumes caller has already pushed us into the kernel context. */
122 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
123                           struct obd_trans_info *oti, int rc, int force_sync)
124 {
125         struct obd_device_target *obt = &exp->exp_obd->u.obt;
126         struct tg_export_data *ted = &exp->exp_target_data;
127         struct lr_server_data *lsd = class_server_data(exp->exp_obd);
128         struct lsd_client_data *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         cfs_mutex_lock(&ted->ted_lcd_lock);
141         lcd = ted->ted_lcd;
142         /* if the export has already been disconnected, we have no last_rcvd slot,
143          * update server data with latest transno then */
144         if (lcd == NULL) {
145                 cfs_mutex_unlock(&ted->ted_lcd_lock);
146                 CWARN("commit transaction for disconnected client %s: rc %d\n",
147                       exp->exp_client_uuid.uuid, rc);
148                 err = filter_update_server_data(exp->exp_obd);
149                 RETURN(err);
150         }
151
152         /* we don't allocate new transnos for replayed requests */
153         cfs_spin_lock(&obt->obt_lut->lut_translock);
154         if (oti->oti_transno == 0) {
155                 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
156                 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
157                 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
158         } else {
159                 last_rcvd = oti->oti_transno;
160                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
161                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
162                 if (unlikely(last_rcvd < le64_to_cpu(lcd->lcd_last_transno))) {
163                         CERROR("Trying to overwrite bigger transno, on-disk: "
164                                LPU64", new: "LPU64"\n",
165                                le64_to_cpu(lcd->lcd_last_transno), last_rcvd);
166                         cfs_spin_lock(&exp->exp_lock);
167                         exp->exp_vbr_failed = 1;
168                         cfs_spin_unlock(&exp->exp_lock);
169                         cfs_spin_unlock(&obt->obt_lut->lut_translock);
170                         cfs_mutex_unlock(&ted->ted_lcd_lock);
171                         RETURN(-EOVERFLOW);
172                 }
173         }
174         oti->oti_transno = last_rcvd;
175
176         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
177         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
178         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
179         cfs_spin_unlock(&obt->obt_lut->lut_translock);
180
181         if (inode)
182                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
183
184         off = ted->ted_lr_off;
185         if (off <= 0) {
186                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
187                        ted->ted_lr_idx, ted->ted_lr_off);
188                 err = -EINVAL;
189         } else {
190                 class_export_cb_get(exp); /* released when the cb is called */
191                 if (!force_sync)
192                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
193                                                            last_rcvd,
194                                                            oti->oti_handle,
195                                                            filter_commit_cb,
196                                                            exp);
197
198                 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
199                                           lcd, sizeof(*lcd), &off,
200                                           force_sync | exp->exp_need_sync);
201                 if (force_sync)
202                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
203         }
204         if (err) {
205                 log_pri = D_ERROR;
206                 if (rc == 0)
207                         rc = err;
208         }
209
210         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
211                last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
212         cfs_mutex_unlock(&ted->ted_lcd_lock);
213         RETURN(rc);
214 }
215
216 void f_dput(struct dentry *dentry)
217 {
218         /* Can't go inside filter_ddelete because it can block */
219         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
220                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
221         LASSERT(atomic_read(&dentry->d_count) > 0);
222
223         dput(dentry);
224 }
225
226 static void init_brw_stats(struct brw_stats *brw_stats)
227 {
228         int i;
229         for (i = 0; i < BRW_LAST; i++)
230                 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
231 }
232
233 static int lprocfs_init_rw_stats(struct obd_device *obd,
234                                  struct lprocfs_stats **stats)
235 {
236         int num_stats;
237
238         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
239                                                         LPROC_FILTER_LAST - 1;
240         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
241         if (*stats == NULL)
242                 return -ENOMEM;
243
244         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
245         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
246                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
247         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
248                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
249
250         return(0);
251 }
252
253 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
254    plus the procfs overhead :( */
255 static int filter_export_stats_init(struct obd_device *obd,
256                                     struct obd_export *exp,
257                                     void *client_nid)
258 {
259         int rc, newnid = 0;
260         ENTRY;
261
262         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
263                 /* Self-export gets no proc entry */
264                 RETURN(0);
265
266         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
267         if (rc) {
268                 /* Mask error for already created
269                  * /proc entries */
270                 if (rc == -EALREADY)
271                         rc = 0;
272                 RETURN(rc);
273         }
274
275         if (newnid) {
276                 struct nid_stat *tmp = exp->exp_nid_stats;
277                 LASSERT(tmp != NULL);
278
279                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
280                 if (tmp->nid_brw_stats == NULL)
281                         GOTO(clean, rc = -ENOMEM);
282
283                 init_brw_stats(tmp->nid_brw_stats);
284                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
285                                         0644, &filter_per_nid_stats_fops,
286                                         exp->exp_nid_stats);
287                 if (rc)
288                         CWARN("Error adding the brw_stats file\n");
289
290                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
291                 if (rc)
292                         GOTO(clean, rc);
293
294                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
295                                             tmp->nid_stats);
296                 if (rc)
297                         GOTO(clean, rc);
298                 rc = lprocfs_nid_ldlm_stats_init(tmp);
299                 if (rc)
300                         GOTO(clean, rc);
301         }
302
303         RETURN(0);
304  clean:
305         return rc;
306 }
307
308 /* Add client data to the FILTER.  We use a bitmap to locate a free space
309  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
310  * Otherwise, we have just read the data from the last_rcvd file and
311  * we know its offset. */
312 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
313                              int cl_idx)
314 {
315         struct obd_device_target *obt = &obd->u.obt;
316         struct tg_export_data *ted = &exp->exp_target_data;
317         struct lr_server_data *lsd = class_server_data(obd);
318         unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
319         int new_client = (cl_idx == -1);
320
321         ENTRY;
322
323         LASSERT(bitmap != NULL);
324         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
325
326         /* Self-export */
327         if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
328                 RETURN(0);
329
330         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
331          * there's no need for extra complication here
332          */
333         if (new_client) {
334                 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
335         repeat:
336                 if (cl_idx >= LR_MAX_CLIENTS) {
337                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
338                                cl_idx);
339                         RETURN(-EOVERFLOW);
340                 }
341                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
342                         cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
343                                                         cl_idx);
344                         goto repeat;
345                 }
346         } else {
347                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
348                         CERROR("FILTER client %d: bit already set in bitmap!\n",
349                                cl_idx);
350                         LBUG();
351                 }
352         }
353
354         ted->ted_lr_idx = cl_idx;
355         ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
356                           cl_idx * le16_to_cpu(lsd->lsd_client_size);
357         cfs_mutex_init(&ted->ted_lcd_lock);
358         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
359
360         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
361                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
362
363         if (new_client) {
364                 struct lvfs_run_ctxt saved;
365                 loff_t off = ted->ted_lr_off;
366                 int rc;
367                 void *handle;
368
369                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
370                        ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
371
372                 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
373                         RETURN(-ENOSPC);
374
375                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
376                 /* Transaction needed to fix bug 1403 */
377                 handle = fsfilt_start(obd,
378                                       obt->obt_rcvd_filp->f_dentry->d_inode,
379                                       FSFILT_OP_SETATTR, NULL);
380                 if (IS_ERR(handle)) {
381                         rc = PTR_ERR(handle);
382                         CERROR("unable to start transaction: rc %d\n", rc);
383                 } else {
384                         ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
385                         exp->exp_last_request_time = cfs_time_current_sec();
386                         rc = fsfilt_add_journal_cb(obd, 0, handle,
387                                                    target_client_add_cb,
388                                                    class_export_cb_get(exp));
389                         if (rc == 0) {
390                                 cfs_spin_lock(&exp->exp_lock);
391                                 exp->exp_need_sync = 1;
392                                 cfs_spin_unlock(&exp->exp_lock);
393                         }
394                         rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
395                                                  ted->ted_lcd,
396                                                  sizeof(*ted->ted_lcd),
397                                                  &off, rc /* sync if no cb */);
398                         fsfilt_commit(obd,
399                                       obt->obt_rcvd_filp->f_dentry->d_inode,
400                                       handle, 0);
401                 }
402                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
403
404                 if (rc) {
405                         CERROR("error writing %s client idx %u: rc %d\n",
406                                LAST_RCVD, ted->ted_lr_idx, rc);
407                         RETURN(rc);
408                 }
409         }
410         RETURN(0);
411 }
412
413 static int filter_client_del(struct obd_export *exp)
414 {
415         struct tg_export_data *ted = &exp->exp_target_data;
416         struct obd_device_target *obt = &exp->exp_obd->u.obt;
417         struct lvfs_run_ctxt saved;
418         int rc;
419         loff_t off;
420         ENTRY;
421
422         if (ted->ted_lcd == NULL)
423                 RETURN(0);
424
425         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
426         if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
427                 GOTO(free, 0);
428
429         LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
430
431         off = ted->ted_lr_off;
432
433         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
434                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
435
436         /* Don't clear ted_lr_idx here as it is likely also unset.  At worst
437          * we leak a client slot that will be cleaned on the next recovery. */
438         if (off <= 0) {
439                 CERROR("%s: client idx %d has med_off %lld\n",
440                        exp->exp_obd->obd_name, ted->ted_lr_idx, off);
441                 GOTO(free, rc = -EINVAL);
442         }
443
444         /* Clear the bit _after_ zeroing out the client so we don't
445            race with filter_client_add and zero out new clients.*/
446         if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
447                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
448                        ted->ted_lr_idx);
449                 LBUG();
450         }
451
452         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
453         /* Make sure the server's last_transno is up to date.
454          * This should be done before zeroing client slot so last_transno will
455          * be in server data or in client data in case of failure */
456         filter_update_server_data(exp->exp_obd);
457
458         cfs_mutex_lock(&ted->ted_lcd_lock);
459         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
460         rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
461                                  ted->ted_lcd,
462                                  sizeof(*ted->ted_lcd), &off, 0);
463         cfs_mutex_unlock(&ted->ted_lcd_lock);
464         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
465
466         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
467                "zero out client %s at idx %u/%llu in %s, rc %d\n",
468                ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
469                LAST_RCVD, rc);
470         EXIT;
471 free:
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                 cfs_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         cfs_spin_lock(&fed->fed_lock);
499         filter_fmd_put_nolock(fed, fmd); /* caller reference */
500         cfs_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         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
512                 if (fmd == keep)
513                         break;
514
515                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
516                     fed->fed_mod_count < filter->fo_fmd_max_num)
517                         break;
518
519                 cfs_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         cfs_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         cfs_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_seq group)
537 {
538         struct filter_mod_data *found = NULL, *fmd;
539
540         LASSERT_SPIN_LOCKED(&fed->fed_lock);
541
542         cfs_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                         cfs_list_del(&fmd->fmd_list);
546                         cfs_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_seq group)
560 {
561         struct filter_mod_data *fmd;
562
563         cfs_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         cfs_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_seq 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         cfs_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                         cfs_list_add_tail(&fmd_new->fmd_list,
591                                           &fed->fed_mod_list);
592                         fmd_new->fmd_id = objid;
593                         fmd_new->fmd_gr = group;
594                         fmd_new->fmd_refcount++;   /* list reference */
595                         found = fmd_new;
596                         fed->fed_mod_count++;
597                 } else {
598                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
599                 }
600         }
601         if (found) {
602                 found->fmd_refcount++;          /* caller reference */
603                 found->fmd_expire = jiffies +
604                         exp->exp_obd->u.filter.fo_fmd_max_age;
605         }
606
607         cfs_spin_unlock(&fed->fed_lock);
608
609         return found;
610 }
611
612 #ifdef DO_FMD_DROP
613 /* drop fmd list reference so it will disappear when last reference is put.
614  * This isn't so critical because it would in fact only affect the one client
615  * that is doing the unlink and at worst we have an stale entry referencing
616  * an object that should never be used again. */
617 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
618 {
619         struct filter_mod_data *found = NULL;
620
621         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
622         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
623         if (found) {
624                 cfs_list_del_init(&found->fmd_list);
625                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
626         }
627         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
628 }
629 #else
630 #define filter_fmd_drop(exp, objid, group)
631 #endif
632
633 /* remove all entries from fmd list */
634 static void filter_fmd_cleanup(struct obd_export *exp)
635 {
636         struct filter_export_data *fed = &exp->exp_filter_data;
637         struct filter_mod_data *fmd = NULL, *tmp;
638
639         cfs_spin_lock(&fed->fed_lock);
640         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
641                 cfs_list_del_init(&fmd->fmd_list);
642                 filter_fmd_put_nolock(fed, fmd);
643         }
644         cfs_spin_unlock(&fed->fed_lock);
645 }
646
647 static int filter_init_export(struct obd_export *exp)
648 {
649         int rc;
650         ENTRY;
651
652         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
653         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
654
655         cfs_spin_lock(&exp->exp_lock);
656         exp->exp_connecting = 1;
657         cfs_spin_unlock(&exp->exp_lock);
658
659         /* self-export doesn't need client data and ldlm initialization */
660         if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
661                                      &exp->exp_client_uuid)))
662                 RETURN(0);
663
664         rc = lut_client_alloc(exp);
665         if (rc == 0)
666                 rc = ldlm_init_export(exp);
667         if (rc)
668                 CERROR("%s: Can't initialize export: rc %d\n",
669                        exp->exp_obd->obd_name, rc);
670
671         RETURN(rc);
672 }
673
674 static int filter_free_server_data(struct obd_device_target *obt)
675 {
676         lut_fini(NULL, obt->obt_lut);
677         OBD_FREE_PTR(obt->obt_lut);
678         return 0;
679 }
680
681 /* assumes caller is already in kernel ctxt */
682 int filter_update_server_data(struct obd_device *obd)
683 {
684         struct file *filp = obd->u.obt.obt_rcvd_filp;
685         struct lr_server_data *lsd = class_server_data(obd);
686         loff_t off = 0;
687         int rc;
688         ENTRY;
689
690         CDEBUG(D_INODE, "server uuid      : %s\n", lsd->lsd_uuid);
691         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
692                le64_to_cpu(lsd->lsd_last_transno));
693         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
694                le64_to_cpu(lsd->lsd_mount_count));
695
696         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
697         if (rc)
698                 CERROR("error writing lr_server_data: rc = %d\n", rc);
699
700         RETURN(rc);
701 }
702
703 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
704                              int force_sync)
705 {
706         struct filter_obd *filter = &obd->u.filter;
707         __u64 tmp;
708         loff_t off = 0;
709         int rc;
710         ENTRY;
711
712         if (filter->fo_last_objid_files[group] == NULL) {
713                 CERROR("Object seq "LPU64" not fully setup; not updating "
714                        "last_objid\n", group);
715                 RETURN(-EINVAL);
716         }
717
718         CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
719                obd->obd_name, filter->fo_last_objids[group], group);
720
721         tmp = cpu_to_le64(filter->fo_last_objids[group]);
722         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
723                                  &tmp, sizeof(tmp), &off, force_sync);
724         if (rc)
725                 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
726                        group, rc);
727         RETURN(rc);
728 }
729 extern int ost_handle(struct ptlrpc_request *req);
730 /* assumes caller has already in kernel ctxt */
731 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
732 {
733         struct filter_obd *filter = &obd->u.filter;
734         struct lr_server_data *lsd;
735         struct lsd_client_data *lcd = NULL;
736         struct inode *inode = filp->f_dentry->d_inode;
737         unsigned long last_rcvd_size = i_size_read(inode);
738         struct lu_target *lut;
739         __u64 mount_count;
740         __u32 start_epoch;
741         int cl_idx;
742         loff_t off = 0;
743         int rc;
744
745         /* ensure padding in the struct is the correct size */
746         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
747                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
748         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
749                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
750
751         /* allocate and initialize lu_target */
752         OBD_ALLOC_PTR(lut);
753         if (lut == NULL)
754                 RETURN(-ENOMEM);
755         rc = lut_init(NULL, lut, obd, NULL);
756         if (rc)
757                 GOTO(err_lut, rc);
758         lsd = class_server_data(obd);
759         if (last_rcvd_size == 0) {
760                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
761
762                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
763                 lsd->lsd_last_transno = 0;
764                 mount_count = lsd->lsd_mount_count = 0;
765                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
766                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
767                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
768                 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
769                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
770                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
771                  * MDS first connects and assigns the OST index number. */
772                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
773                                                         OBD_INCOMPAT_OST);
774         } else {
775                 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
776                 if (rc) {
777                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
778                                LAST_RCVD, rc);
779                         GOTO(err_lut, rc);
780                 }
781                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
782                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
783                                            "using the wrong disk %s. Were the "
784                                            "/dev/ assignments rearranged?\n",
785                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
786                         GOTO(err_lut, rc = -EINVAL);
787                 }
788                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
789                 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
790                 /* COMPAT_146 */
791                 /* Assume old last_rcvd format unless I_C_LR is set */
792                 if (!(lsd->lsd_feature_incompat &
793                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
794                         lsd->lsd_last_transno = lsd->lsd_compat14;
795                 /* end COMPAT_146 */
796                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
797                  * MDS first connects and assigns the OST index number. */
798                 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
799                                                          OBD_INCOMPAT_OST);
800         }
801
802         if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
803                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
804                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
805                        ~FILTER_INCOMPAT_SUPP);
806                 GOTO(err_lut, rc = -EINVAL);
807         }
808         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
809                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
810                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
811                        ~FILTER_ROCOMPAT_SUPP);
812                 /* Do something like remount filesystem read-only */
813                 GOTO(err_lut, rc = -EINVAL);
814         }
815
816         start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
817
818         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
819                obd->obd_name, start_epoch);
820         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
821                obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
822         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
823                obd->obd_name, mount_count + 1);
824         CDEBUG(D_INODE, "%s: server data size: %u\n",
825                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
826         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
827                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
828         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
829                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
830         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
831                obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
832         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
833                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
834                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
835                 le16_to_cpu(lsd->lsd_client_size));
836
837         if (!obd->obd_replayable) {
838                 CWARN("%s: recovery support OFF\n", obd->obd_name);
839                 GOTO(out, rc = 0);
840         }
841
842         OBD_ALLOC_PTR(lcd);
843         if (!lcd)
844                 GOTO(err_client, rc = -ENOMEM);
845
846         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
847              off < last_rcvd_size; cl_idx++) {
848                 __u64 last_rcvd;
849                 struct obd_export *exp;
850                 struct filter_export_data *fed;
851
852                 /* Don't assume off is incremented properly by
853                  * fsfilt_read_record(), in case sizeof(*lcd)
854                  * isn't the same as lsd->lsd_client_size.  */
855                 off = le32_to_cpu(lsd->lsd_client_start) +
856                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
857                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
858                 if (rc) {
859                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
860                                LAST_RCVD, cl_idx, off, rc);
861                         break; /* read error shouldn't cause startup to fail */
862                 }
863
864                 if (lcd->lcd_uuid[0] == '\0') {
865                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
866                                cl_idx);
867                         continue;
868                 }
869
870                 check_lcd(obd->obd_name, cl_idx, lcd);
871
872                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
873
874                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
875                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
876                        last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
877
878                 /* These exports are cleaned up by filter_disconnect(), so they
879                  * need to be set up like real exports as filter_connect() does.
880                  */
881                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
882                 if (IS_ERR(exp)) {
883                         if (PTR_ERR(exp) == -EALREADY) {
884                                 /* export already exists, zero out this one */
885                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
886                                 continue;
887                         }
888                         OBD_FREE_PTR(lcd);
889                         GOTO(err_client, rc = PTR_ERR(exp));
890                 }
891
892                 fed = &exp->exp_filter_data;
893                 *fed->fed_ted.ted_lcd = *lcd;
894                 fed->fed_group = 0; /* will be assigned at connect */
895                 filter_export_stats_init(obd, exp, NULL);
896                 rc = filter_client_add(obd, exp, cl_idx);
897                 /* can't fail for existing client */
898                 LASSERTF(rc == 0, "rc = %d\n", rc);
899
900                 /* VBR: set export last committed */
901                 exp->exp_last_committed = last_rcvd;
902                 cfs_spin_lock(&exp->exp_lock);
903                 exp->exp_connecting = 0;
904                 exp->exp_in_recovery = 0;
905                 cfs_spin_unlock(&exp->exp_lock);
906                 obd->obd_max_recoverable_clients++;
907                 class_export_put(exp);
908
909                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
910                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
911         }
912         OBD_FREE_PTR(lcd);
913
914         obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
915 out:
916         obd->u.obt.obt_mount_count = mount_count + 1;
917         obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
918         lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
919
920         /* save it, so mount count and last_transno is current */
921         rc = filter_update_server_data(obd);
922         if (rc)
923                 GOTO(err_client, rc);
924
925         RETURN(0);
926
927 err_client:
928         class_disconnect_exports(obd);
929 err_lut:
930         filter_free_server_data(&obd->u.obt);
931         RETURN(rc);
932 }
933
934 static int filter_cleanup_groups(struct obd_device *obd)
935 {
936         struct filter_obd *filter = &obd->u.filter;
937         struct file *filp;
938         struct dentry *dentry;
939         int i, j;
940         ENTRY;
941
942         if (filter->fo_dentry_O_groups != NULL) {
943                 for (i = 0; i < filter->fo_group_count; i++) {
944                         dentry = filter->fo_dentry_O_groups[i];
945                         if (dentry != NULL)
946                                 f_dput(dentry);
947                 }
948                 OBD_FREE(filter->fo_dentry_O_groups,
949                          filter->fo_group_count *
950                          sizeof(*filter->fo_dentry_O_groups));
951                 filter->fo_dentry_O_groups = NULL;
952         }
953         if (filter->fo_last_objid_files != NULL) {
954                 for (i = 0; i < filter->fo_group_count; i++) {
955                         filp = filter->fo_last_objid_files[i];
956                         if (filp != NULL)
957                                 filp_close(filp, 0);
958                 }
959                 OBD_FREE(filter->fo_last_objid_files,
960                          filter->fo_group_count *
961                          sizeof(*filter->fo_last_objid_files));
962                 filter->fo_last_objid_files = NULL;
963         }
964         if (filter->fo_dentry_O_sub != NULL) {
965                 for (i = 0; i < filter->fo_group_count; i++) {
966                         for (j = 0; j < filter->fo_subdir_count; j++) {
967                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
968                                 if (dentry != NULL)
969                                         f_dput(dentry);
970                         }
971                 }
972                 OBD_FREE(filter->fo_dentry_O_sub,
973                          filter->fo_group_count *
974                          sizeof(*filter->fo_dentry_O_sub));
975                 filter->fo_dentry_O_sub = NULL;
976         }
977         if (filter->fo_last_objids != NULL) {
978                 OBD_FREE(filter->fo_last_objids,
979                          filter->fo_group_count *
980                          sizeof(*filter->fo_last_objids));
981                 filter->fo_last_objids = NULL;
982         }
983         if (filter->fo_dentry_O != NULL) {
984                 f_dput(filter->fo_dentry_O);
985                 filter->fo_dentry_O = NULL;
986         }
987         RETURN(0);
988 }
989
990 static int filter_update_last_group(struct obd_device *obd, int group)
991 {
992         struct filter_obd *filter = &obd->u.filter;
993         struct file *filp = NULL;
994         int last_group = 0, rc;
995         loff_t off = 0;
996         ENTRY;
997
998         if (group <= filter->fo_committed_group)
999                 RETURN(0);
1000
1001         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1002         if (IS_ERR(filp)) {
1003                 rc = PTR_ERR(filp);
1004                 filp = NULL;
1005                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1006                 GOTO(cleanup, rc);
1007         }
1008
1009         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1010         if (rc) {
1011                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1012                 GOTO(cleanup, rc);
1013         }
1014
1015         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1016                obd->obd_name, last_group, group);
1017
1018         off = 0;
1019         last_group = group;
1020         /* must be sync: bXXXX */
1021         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1022         if (rc) {
1023                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1024                 GOTO(cleanup, rc);
1025         }
1026
1027         filter->fo_committed_group = group;
1028 cleanup:
1029         if (filp)
1030                 filp_close(filp, 0);
1031         RETURN(rc);
1032 }
1033
1034 static int filter_read_group_internal(struct obd_device *obd, int group,
1035                                       int create)
1036 {
1037         struct filter_obd *filter = &obd->u.filter;
1038         __u64 *new_objids = NULL;
1039         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1040         struct dentry **new_groups = NULL;
1041         struct file **new_files = NULL;
1042         struct dentry *dentry;
1043         struct file *filp;
1044         int old_count = filter->fo_group_count, rc, stage = 0, i;
1045         char name[25];
1046         __u64 last_objid;
1047         loff_t off = 0;
1048         int len = group + 1;
1049
1050         snprintf(name, 24, "%d", group);
1051         name[24] = '\0';
1052
1053         if (!create) {
1054                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1055                                            strlen(name));
1056                 if (IS_ERR(dentry)) {
1057                         CERROR("Cannot lookup expected object group %d: %ld\n",
1058                                group, PTR_ERR(dentry));
1059                         RETURN(PTR_ERR(dentry));
1060                 }
1061         } else {
1062                 dentry = simple_mkdir(filter->fo_dentry_O,
1063                                       obd->u.obt.obt_vfsmnt, name, 0700, 1);
1064                 if (IS_ERR(dentry)) {
1065                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1066                                PTR_ERR(dentry));
1067                         RETURN(PTR_ERR(dentry));
1068                 }
1069         }
1070         stage = 1;
1071
1072         snprintf(name, 24, "O/%d/LAST_ID", group);
1073         name[24] = '\0';
1074         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1075         if (IS_ERR(filp)) {
1076                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1077                 GOTO(cleanup, rc = PTR_ERR(filp));
1078         }
1079         stage = 2;
1080
1081         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1082         if (rc) {
1083                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1084                 GOTO(cleanup, rc);
1085         }
1086
1087         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1088                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1089                 if (tmp_subdirs == NULL)
1090                         GOTO(cleanup, rc = -ENOMEM);
1091                 stage = 3;
1092
1093                 for (i = 0; i < filter->fo_subdir_count; i++) {
1094                         char dir[20];
1095                         snprintf(dir, sizeof(dir), "d%u", i);
1096
1097                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1098                                                               obd->u.obt.obt_vfsmnt,
1099                                                               dir, 0700, 1);
1100                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1101                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1102                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1103                                        group, dir, rc);
1104                                 GOTO(cleanup, rc);
1105                         }
1106
1107                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1108                                tmp_subdirs->dentry[i]);
1109                 }
1110         }
1111
1112         /* 'group' is an index; we need an array of length 'group + 1' */
1113         if (group + 1 > old_count) {
1114                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1115                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1116                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1117                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1118                 stage = 4;
1119                 if (new_objids == NULL || new_subdirs == NULL ||
1120                     new_groups == NULL || new_files == NULL)
1121                         GOTO(cleanup, rc = -ENOMEM);
1122
1123                 if (old_count) {
1124                         memcpy(new_objids, filter->fo_last_objids,
1125                                old_count * sizeof(*new_objids));
1126                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1127                                old_count * sizeof(*new_subdirs));
1128                         memcpy(new_groups, filter->fo_dentry_O_groups,
1129                                old_count * sizeof(*new_groups));
1130                         memcpy(new_files, filter->fo_last_objid_files,
1131                                old_count * sizeof(*new_files));
1132
1133                         OBD_FREE(filter->fo_last_objids,
1134                                  old_count * sizeof(*new_objids));
1135                         OBD_FREE(filter->fo_dentry_O_sub,
1136                                  old_count * sizeof(*new_subdirs));
1137                         OBD_FREE(filter->fo_dentry_O_groups,
1138                                  old_count * sizeof(*new_groups));
1139                         OBD_FREE(filter->fo_last_objid_files,
1140                                  old_count * sizeof(*new_files));
1141                 }
1142                 filter->fo_last_objids = new_objids;
1143                 filter->fo_dentry_O_sub = new_subdirs;
1144                 filter->fo_dentry_O_groups = new_groups;
1145                 filter->fo_last_objid_files = new_files;
1146                 filter->fo_group_count = len;
1147         }
1148
1149         filter->fo_dentry_O_groups[group] = dentry;
1150         filter->fo_last_objid_files[group] = filp;
1151         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1152                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1153                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1154         }
1155
1156         filter_update_last_group(obd, group);
1157
1158         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1159                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1160                 rc = filter_update_last_objid(obd, group, 1);
1161                 RETURN(rc);
1162         }
1163
1164         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1165         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1166                obd->obd_name, group, last_objid);
1167         RETURN(0);
1168  cleanup:
1169         switch (stage) {
1170         case 4:
1171                 if (new_objids != NULL)
1172                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1173                 if (new_subdirs != NULL)
1174                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1175                 if (new_groups != NULL)
1176                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1177                 if (new_files != NULL)
1178                         OBD_FREE(new_files, len * sizeof(*new_files));
1179         case 3:
1180                 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1181                         for (i = 0; i < filter->fo_subdir_count; i++) {
1182                                 if (tmp_subdirs->dentry[i] != NULL)
1183                                         dput(tmp_subdirs->dentry[i]);
1184                         }
1185                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1186                 }
1187         case 2:
1188                 filp_close(filp, 0);
1189         case 1:
1190                 dput(dentry);
1191         }
1192         RETURN(rc);
1193 }
1194
1195 static int filter_read_groups(struct obd_device *obd, int last_group,
1196                               int create)
1197 {
1198         struct filter_obd *filter = &obd->u.filter;
1199         int old_count, group, rc = 0;
1200
1201         cfs_mutex_lock(&filter->fo_init_lock);
1202         old_count = filter->fo_group_count;
1203         for (group = old_count; group <= last_group; group++) {
1204                 rc = filter_read_group_internal(obd, group, create);
1205                 if (rc != 0)
1206                         break;
1207         }
1208         cfs_mutex_unlock(&filter->fo_init_lock);
1209         return rc;
1210 }
1211
1212 /* FIXME: object groups */
1213 static int filter_prep_groups(struct obd_device *obd)
1214 {
1215         struct filter_obd *filter = &obd->u.filter;
1216         struct dentry *O_dentry;
1217         struct file *filp;
1218         int    last_group, rc = 0, cleanup_phase = 0;
1219         loff_t off = 0;
1220         ENTRY;
1221
1222         O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1223                                 "O", 0700, 1);
1224         CDEBUG(D_INODE, "%s: got/created O: %p\n", obd->obd_name, O_dentry);
1225         if (IS_ERR(O_dentry)) {
1226                 rc = PTR_ERR(O_dentry);
1227                 CERROR("%s: cannot open/create O: rc = %d\n", obd->obd_name,rc);
1228                 GOTO(cleanup, rc);
1229         }
1230         filter->fo_dentry_O = O_dentry;
1231         cleanup_phase = 1; /* O_dentry */
1232
1233         /* we have to initialize all groups before first connections from
1234          * clients because they may send create/destroy for any group -bzzz */
1235         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1236         if (IS_ERR(filp)) {
1237                 CERROR("%s: cannot create LAST_GROUP: rc = %ld\n",
1238                        obd->obd_name, PTR_ERR(filp));
1239                 GOTO(cleanup, rc = PTR_ERR(filp));
1240         }
1241         cleanup_phase = 2; /* filp */
1242
1243         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1244         if (rc) {
1245                 CERROR("%s: error reading LAST_GROUP: rc %d\n",
1246                        obd->obd_name, rc);
1247                 GOTO(cleanup, rc);
1248         }
1249
1250         if (off == 0)
1251                 last_group = FID_SEQ_OST_MDT0;
1252
1253         CDEBUG(D_INODE, "%s: initialize group %u (max %u)\n", obd->obd_name,
1254                FID_SEQ_OST_MDT0, last_group);
1255         filter->fo_committed_group = last_group;
1256         rc = filter_read_groups(obd, last_group, 1);
1257         if (rc)
1258                 GOTO(cleanup, rc);
1259
1260         filp_close(filp, 0);
1261         RETURN(0);
1262
1263  cleanup:
1264         switch (cleanup_phase) {
1265         case 2:
1266                 filp_close(filp, 0);
1267         case 1:
1268                 filter_cleanup_groups(obd);
1269                 f_dput(filter->fo_dentry_O);
1270                 filter->fo_dentry_O = NULL;
1271         default:
1272                 break;
1273         }
1274         return rc;
1275
1276 }
1277
1278 /* setup the object store with correct subdirectories */
1279 static int filter_prep(struct obd_device *obd)
1280 {
1281         struct lvfs_run_ctxt saved;
1282         struct filter_obd *filter = &obd->u.filter;
1283         struct file *file;
1284         struct inode *inode;
1285         int rc = 0;
1286         ENTRY;
1287
1288         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1289         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1290         if (!file || IS_ERR(file)) {
1291                 rc = PTR_ERR(file);
1292                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1293                        LAST_RCVD, rc);
1294                 GOTO(out, rc);
1295         }
1296         obd->u.obt.obt_rcvd_filp = file;
1297         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1298                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1299                        file->f_dentry->d_inode->i_mode);
1300                 GOTO(err_filp, rc = -ENOENT);
1301         }
1302
1303         inode = file->f_dentry->d_parent->d_inode;
1304         /* We use i_op->unlink directly in filter_vfs_unlink() */
1305         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1306                 CERROR("%s: filesystem does not support create/unlink ops\n",
1307                        obd->obd_name);
1308                 GOTO(err_filp, rc = -EOPNOTSUPP);
1309         }
1310
1311         rc = filter_init_server_data(obd, file);
1312         if (rc) {
1313                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1314                 GOTO(err_filp, rc);
1315         }
1316         LASSERT(obd->u.obt.obt_lut);
1317         target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1318
1319         /* open and create health check io file*/
1320         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1321         if (IS_ERR(file)) {
1322                 rc = PTR_ERR(file);
1323                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1324                        HEALTH_CHECK, rc);
1325                 GOTO(err_server_data, rc);
1326         }
1327         filter->fo_obt.obt_health_check_filp = file;
1328         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1329                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1330                        file->f_dentry->d_inode->i_mode);
1331                 GOTO(err_health_check, rc = -ENOENT);
1332         }
1333         rc = lvfs_check_io_health(obd, file);
1334         if (rc)
1335                 GOTO(err_health_check, rc);
1336
1337         rc = filter_prep_groups(obd);
1338         if (rc)
1339                 GOTO(err_health_check, rc);
1340 out:
1341         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1342
1343         return(rc);
1344
1345 err_health_check:
1346         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1347                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1348         filter->fo_obt.obt_health_check_filp = NULL;
1349 err_server_data:
1350         target_recovery_fini(obd);
1351         filter_free_server_data(&obd->u.obt);
1352 err_filp:
1353         if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1354                 CERROR("can't close %s after error\n", LAST_RCVD);
1355         obd->u.obt.obt_rcvd_filp = NULL;
1356         goto out;
1357 }
1358
1359 /* cleanup the filter: write last used object id to status file */
1360 static void filter_post(struct obd_device *obd)
1361 {
1362         struct lvfs_run_ctxt saved;
1363         struct filter_obd *filter = &obd->u.filter;
1364         int rc, i;
1365
1366         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1367          * best to start a transaction with h_sync, because we removed this
1368          * from lastobjid */
1369
1370         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1371         rc = filter_update_server_data(obd);
1372         if (rc)
1373                 CERROR("error writing server data: rc = %d\n", rc);
1374
1375         for (i = 0; i < filter->fo_group_count; i++) {
1376                 rc = filter_update_last_objid(obd, i,
1377                                 (i == filter->fo_group_count - 1));
1378                 if (rc)
1379                         CERROR("error writing group %d lastobjid: rc = %d\n",
1380                                i, rc);
1381         }
1382
1383         rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1384         obd->u.obt.obt_rcvd_filp = NULL;
1385         if (rc)
1386                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1387
1388         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1389         filter->fo_obt.obt_health_check_filp = NULL;
1390         if (rc)
1391                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1392
1393         filter_cleanup_groups(obd);
1394         filter_free_server_data(&obd->u.obt);
1395         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1396
1397         filter_free_capa_keys(filter);
1398         cleanup_capa_hash(filter->fo_capa_hash);
1399 }
1400
1401 static void filter_set_last_id(struct filter_obd *filter,
1402                                obd_id id, obd_seq group)
1403 {
1404         LASSERT(group <= filter->fo_group_count);
1405
1406         cfs_spin_lock(&filter->fo_objidlock);
1407         filter->fo_last_objids[group] = id;
1408         cfs_spin_unlock(&filter->fo_objidlock);
1409 }
1410
1411 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1412 {
1413         obd_id id;
1414         LASSERT(group <= filter->fo_group_count);
1415         LASSERT(filter->fo_last_objids != NULL);
1416
1417         /* FIXME: object groups */
1418         cfs_spin_lock(&filter->fo_objidlock);
1419         id = filter->fo_last_objids[group];
1420         cfs_spin_unlock(&filter->fo_objidlock);
1421         return id;
1422 }
1423
1424 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1425 {
1426         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1427         return 0;
1428 }
1429
1430 /* We never dget the object parent, so DON'T dput it either */
1431 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1432 {
1433         struct filter_obd *filter = &obd->u.filter;
1434         struct filter_subdirs *subdirs;
1435
1436         if (group >= filter->fo_group_count) /* FIXME: object groups */
1437                 return ERR_PTR(-EBADF);
1438
1439         if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1440                 return filter->fo_dentry_O_groups[group];
1441
1442         subdirs = &filter->fo_dentry_O_sub[group];
1443         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1444 }
1445
1446 /* We never dget the object parent, so DON'T dput it either */
1447 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1448                                   obd_id objid)
1449 {
1450         unsigned long now = jiffies;
1451         struct dentry *dparent = filter_parent(obd, group, objid);
1452         int rc;
1453
1454         if (IS_ERR(dparent))
1455                 return dparent;
1456         if (dparent == NULL)
1457                 return ERR_PTR(-ENOENT);
1458
1459         rc = filter_lock_dentry(obd, dparent);
1460         fsfilt_check_slow(obd, now, "parent lock");
1461         return rc ? ERR_PTR(rc) : dparent;
1462 }
1463
1464 /* We never dget the object parent, so DON'T dput it either */
1465 static void filter_parent_unlock(struct dentry *dparent)
1466 {
1467         UNLOCK_INODE_MUTEX(dparent->d_inode);
1468 }
1469
1470 /* How to get files, dentries, inodes from object id's.
1471  *
1472  * If dir_dentry is passed, the caller has already locked the parent
1473  * appropriately for this operation (normally a write lock).  If
1474  * dir_dentry is NULL, we do a read lock while we do the lookup to
1475  * avoid races with create/destroy and such changing the directory
1476  * internal to the filesystem code. */
1477 struct dentry *filter_fid2dentry(struct obd_device *obd,
1478                                  struct dentry *dir_dentry,
1479                                  obd_seq group, obd_id id)
1480 {
1481         struct dentry *dparent = dir_dentry;
1482         struct dentry *dchild;
1483         char name[32];
1484         int len;
1485         ENTRY;
1486
1487         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1488             obd->u.filter.fo_destroys_in_progress == 0) {
1489                 /* don't fail lookups for orphan recovery, it causes
1490                  * later LBUGs when objects still exist during precreate */
1491                 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1492                 RETURN(ERR_PTR(-ENOENT));
1493         }
1494         if (id == 0) {
1495                 CERROR("fatal: invalid object id 0\n");
1496                 RETURN(ERR_PTR(-ESTALE));
1497         }
1498
1499         len = sprintf(name, LPU64, id);
1500         if (dir_dentry == NULL) {
1501                 dparent = filter_parent_lock(obd, group, id);
1502                 if (IS_ERR(dparent)) {
1503                         CERROR("%s: error getting object "POSTID
1504                                " parent: rc %ld\n", obd->obd_name,
1505                                id, group, PTR_ERR(dparent));
1506                         RETURN(dparent);
1507                 }
1508         }
1509         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1510                dparent->d_name.len, dparent->d_name.name, name);
1511         /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1512         dchild = lookup_one_len(name, dparent, len);
1513         if (dir_dentry == NULL)
1514                 filter_parent_unlock(dparent);
1515         if (IS_ERR(dchild)) {
1516                 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1517                        obd->obd_name, id, group, PTR_ERR(dchild));
1518                 RETURN(dchild);
1519         }
1520
1521         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1522                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1523                        obd->obd_name, id, dchild->d_inode->i_ino);
1524                 f_dput(dchild);
1525                 RETURN(ERR_PTR(-ENOENT));
1526         }
1527
1528         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1529                name, dchild, atomic_read(&dchild->d_count));
1530
1531         LASSERT(atomic_read(&dchild->d_count) > 0);
1532
1533         RETURN(dchild);
1534 }
1535
1536 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1537                                   obd_id group, struct lustre_handle *lockh)
1538 {
1539         int flags = LDLM_AST_DISCARD_DATA, rc;
1540         struct ldlm_res_id res_id;
1541         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1542         ENTRY;
1543
1544         osc_build_res_name(objid, group, &res_id);
1545         /* Tell the clients that the object is gone now and that they should
1546          * throw away any cached pages. */
1547         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1548                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1549                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1550                                     lockh);
1551         if (rc != ELDLM_OK)
1552                 lockh->cookie = 0;
1553         RETURN(rc);
1554 }
1555
1556 static void filter_fini_destroy(struct obd_device *obd,
1557                                 struct lustre_handle *lockh)
1558 {
1559         if (lustre_handle_is_used(lockh))
1560                 ldlm_lock_decref(lockh, LCK_PW);
1561 }
1562
1563 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1564  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1565  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1566  * reverse.  Caller must take i_sem before starting the transaction and we
1567  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1568 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1569                       struct vfsmount *mnt)
1570 {
1571         int rc;
1572         ENTRY;
1573
1574         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1575          * itself we already hold dir->i_mutex for child create/unlink ops */
1576         LASSERT(dentry->d_inode != NULL);
1577         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1578         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1579
1580
1581         /* may_delete() */
1582         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1583                 GOTO(out, rc = -ENOENT);
1584
1585         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1586         if (rc)
1587                 GOTO(out, rc);
1588
1589         if (IS_APPEND(dir))
1590                 GOTO(out, rc = -EPERM);
1591
1592         /* check_sticky() */
1593         if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1594              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1595             IS_IMMUTABLE(dentry->d_inode))
1596                 GOTO(out, rc = -EPERM);
1597
1598         /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
1599         ll_vfs_dq_init(dir);
1600
1601         rc = ll_security_inode_unlink(dir, dentry, mnt);
1602         if (rc)
1603                 GOTO(out, rc);
1604
1605         rc = dir->i_op->unlink(dir, dentry);
1606 out:
1607         /* need to drop i_mutex before we lose inode reference */
1608         UNLOCK_INODE_MUTEX(dentry->d_inode);
1609         if (rc == 0)
1610                 d_delete(dentry);
1611
1612         RETURN(rc);
1613 }
1614
1615 /* Caller must hold LCK_PW on parent and push us into kernel context.
1616  * Caller must hold child i_mutex, we drop it always.
1617  * Caller is also required to ensure that dchild->d_inode exists. */
1618 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1619                                    obd_seq group, struct dentry *dparent,
1620                                    struct dentry *dchild)
1621 {
1622         struct inode *inode = dchild->d_inode;
1623         int rc;
1624
1625         /* There should be 2 references to the inode:
1626          *  1) taken by filter_prepare_destroy
1627          *  2) taken by filter_destroy */
1628         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1629                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1630                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1631                        (unsigned long)inode->i_nlink,
1632                        atomic_read(&inode->i_count));
1633         }
1634
1635         rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1636         if (rc)
1637                 CERROR("error unlinking objid %.*s: rc %d\n",
1638                        dchild->d_name.len, dchild->d_name.name, rc);
1639         return(rc);
1640 }
1641
1642 struct filter_intent_args {
1643         struct ldlm_lock **victim;
1644         __u64 size;
1645         int *liblustre;
1646 };
1647
1648 static enum interval_iter filter_intent_cb(struct interval_node *n,
1649                                            void *args)
1650 {
1651         struct ldlm_interval *node = (struct ldlm_interval *)n;
1652         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1653         __u64 size = arg->size;
1654         struct ldlm_lock **v = arg->victim;
1655         struct ldlm_lock *lck;
1656
1657         /* If the interval is lower than the current file size,
1658          * just break. */
1659         if (interval_high(n) <= size)
1660                 return INTERVAL_ITER_STOP;
1661
1662         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1663                 /* Don't send glimpse ASTs to liblustre clients.
1664                  * They aren't listening for them, and they do
1665                  * entirely synchronous I/O anyways. */
1666                 if (lck->l_export == NULL ||
1667                     lck->l_export->exp_libclient == 1)
1668                         continue;
1669
1670                 if (*arg->liblustre)
1671                         *arg->liblustre = 0;
1672
1673                 if (*v == NULL) {
1674                         *v = LDLM_LOCK_GET(lck);
1675                 } else if ((*v)->l_policy_data.l_extent.start <
1676                            lck->l_policy_data.l_extent.start) {
1677                         LDLM_LOCK_RELEASE(*v);
1678                         *v = LDLM_LOCK_GET(lck);
1679                 }
1680
1681                 /* the same policy group - every lock has the
1682                  * same extent, so needn't do it any more */
1683                 break;
1684         }
1685
1686         return INTERVAL_ITER_CONT;
1687 }
1688
1689 static int filter_intent_policy(struct ldlm_namespace *ns,
1690                                 struct ldlm_lock **lockp, void *req_cookie,
1691                                 ldlm_mode_t mode, int flags, void *data)
1692 {
1693         struct ptlrpc_request *req = req_cookie;
1694         struct ldlm_lock *lock = *lockp, *l = NULL;
1695         struct ldlm_resource *res = lock->l_resource;
1696         ldlm_processing_policy policy;
1697         struct ost_lvb *res_lvb, *reply_lvb;
1698         struct ldlm_reply *rep;
1699         ldlm_error_t err;
1700         int idx, rc, tmpflags = 0, only_liblustre = 1;
1701         struct ldlm_interval_tree *tree;
1702         struct filter_intent_args arg;
1703         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1704                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1705                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1706         ENTRY;
1707
1708         policy = ldlm_get_processing_policy(res);
1709         LASSERT(policy != NULL);
1710         LASSERT(req != NULL);
1711
1712         rc = lustre_pack_reply(req, 3, repsize, NULL);
1713         if (rc)
1714                 RETURN(req->rq_status = rc);
1715
1716         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1717         LASSERT(rep != NULL);
1718
1719         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1720                                    sizeof(*reply_lvb));
1721         LASSERT(reply_lvb != NULL);
1722
1723         //fixup_handle_for_resent_req(req, lock, &lockh);
1724
1725         /* Call the extent policy function to see if our request can be
1726          * granted, or is blocked.
1727          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1728          * lock, and should not be granted if the lock will be blocked.
1729          */
1730
1731         if (flags & LDLM_FL_BLOCK_NOWAIT) {
1732                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_AGL_DELAY, 5);
1733
1734                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_AGL_NOLOCK))
1735                         RETURN(ELDLM_LOCK_ABORTED);
1736         }
1737
1738         LASSERT(ns == ldlm_res_to_ns(res));
1739         lock_res(res);
1740         rc = policy(lock, &tmpflags, 0, &err, NULL);
1741         check_res_locked(res);
1742
1743         /* The lock met with no resistance; we're finished. */
1744         if (rc == LDLM_ITER_CONTINUE) {
1745                 /* do not grant locks to the liblustre clients: they cannot
1746                  * handle ASTs robustly.  We need to do this while still
1747                  * holding lr_lock to avoid the lock remaining on the res_link
1748                  * list (and potentially being added to l_pending_list by an
1749                  * AST) when we are going to drop this lock ASAP. */
1750                 if (lock->l_export->exp_libclient ||
1751                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1752                         ldlm_resource_unlink_lock(lock);
1753                         err = ELDLM_LOCK_ABORTED;
1754                 } else {
1755                         err = ELDLM_LOCK_REPLACED;
1756                 }
1757                 unlock_res(res);
1758                 RETURN(err);
1759         } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1760                 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1761                  * callback for glimpse size. The real size user will trigger
1762                  * the glimpse callback when necessary. */
1763                 unlock_res(res);
1764                 RETURN(ELDLM_LOCK_ABORTED);
1765         }
1766
1767         /* Do not grant any lock, but instead send GL callbacks.  The extent
1768          * policy nicely created a list of all PW locks for us.  We will choose
1769          * the highest of those which are larger than the size in the LVB, if
1770          * any, and perform a glimpse callback. */
1771         res_lvb = res->lr_lvb_data;
1772         LASSERT(res_lvb != NULL);
1773         *reply_lvb = *res_lvb;
1774
1775         /*
1776          * lr_lock guarantees that no new locks are granted, and,
1777          * therefore, that res->lr_lvb_data cannot increase beyond the
1778          * end of already granted lock. As a result, it is safe to
1779          * check against "stale" reply_lvb->lvb_size value without
1780          * res->lr_lvb_mutex.
1781          */
1782         arg.size = reply_lvb->lvb_size;
1783         arg.victim = &l;
1784         arg.liblustre = &only_liblustre;
1785         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1786                 tree = &res->lr_itree[idx];
1787                 if (tree->lit_mode == LCK_PR)
1788                         continue;
1789
1790                 interval_iterate_reverse(tree->lit_root,
1791                                          filter_intent_cb, &arg);
1792         }
1793         unlock_res(res);
1794
1795         /* There were no PW locks beyond the size in the LVB; finished. */
1796         if (l == NULL) {
1797                 if (only_liblustre) {
1798                         /* If we discovered a liblustre client with a PW lock,
1799                          * however, the LVB may be out of date!  The LVB is
1800                          * updated only on glimpse (which we don't do for
1801                          * liblustre clients) and cancel (which the client
1802                          * obviously has not yet done).  So if it has written
1803                          * data but kept the lock, the LVB is stale and needs
1804                          * to be updated from disk.
1805                          *
1806                          * Of course, this will all disappear when we switch to
1807                          * taking liblustre locks on the OST. */
1808                         ldlm_res_lvbo_update(res, NULL, 1);
1809                 }
1810                 RETURN(ELDLM_LOCK_ABORTED);
1811         }
1812
1813         /*
1814          * This check is for lock taken in filter_prepare_destroy() that does
1815          * not have l_glimpse_ast set. So the logic is: if there is a lock
1816          * with no l_glimpse_ast set, this object is being destroyed already.
1817          *
1818          * Hence, if you are grabbing DLM locks on the server, always set
1819          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1820          */
1821         if (l->l_glimpse_ast == NULL) {
1822                 /* We are racing with unlink(); just return -ENOENT */
1823                 rep->lock_policy_res1 = -ENOENT;
1824                 goto out;
1825         }
1826
1827         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1828         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1829
1830         lock_res(res);
1831         *reply_lvb = *res_lvb;
1832         unlock_res(res);
1833
1834  out:
1835         LDLM_LOCK_RELEASE(l);
1836
1837         RETURN(ELDLM_LOCK_ABORTED);
1838 }
1839
1840 /*
1841  * per-obd_device iobuf pool.
1842  *
1843  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1844  * allocations in write-path has to be minimized (see bug 5137).
1845  *
1846  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1847  * OST threads (see ost_thread_{init,done}()).
1848  *
1849  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1850  * at the OST layer there are only (potentially) multiple obd_device of type
1851  * unknown at the time of OST thread creation.
1852  *
1853  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1854  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1855  * it's very own iobuf.
1856  *
1857  * Functions below
1858  *
1859  *     filter_kiobuf_pool_init()
1860  *
1861  *     filter_kiobuf_pool_done()
1862  *
1863  *     filter_iobuf_get()
1864  *
1865  * operate on this array. They are "generic" in a sense that they don't depend
1866  * on actual type of iobuf's (the latter depending on Linux kernel version).
1867  */
1868
1869 /*
1870  * destroy pool created by filter_iobuf_pool_init
1871  */
1872 static void filter_iobuf_pool_done(struct filter_obd *filter)
1873 {
1874         struct filter_iobuf **pool;
1875         int i;
1876
1877         ENTRY;
1878
1879         pool = filter->fo_iobuf_pool;
1880         if (pool != NULL) {
1881                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1882                         if (pool[i] != NULL)
1883                                 filter_free_iobuf(pool[i]);
1884                 }
1885                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1886                 filter->fo_iobuf_pool = NULL;
1887         }
1888         EXIT;
1889 }
1890
1891 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1892 {
1893         struct filter_obd       *filter = &obd->u.filter;
1894         struct sptlrpc_rule_set  tmp_rset;
1895         int                      rc;
1896
1897         sptlrpc_rule_set_init(&tmp_rset);
1898         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1899         if (rc) {
1900                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1901                        obd->obd_name, rc);
1902                 return rc;
1903         }
1904
1905         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1906
1907         cfs_write_lock(&filter->fo_sptlrpc_lock);
1908         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1909         filter->fo_sptlrpc_rset = tmp_rset;
1910         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1911
1912         return 0;
1913 }
1914
1915 /*
1916  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1917  */
1918 static int filter_iobuf_pool_init(struct filter_obd *filter)
1919 {
1920         void **pool;
1921
1922         ENTRY;
1923
1924
1925         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1926                       GFP_KERNEL);
1927         if (filter->fo_iobuf_pool == NULL)
1928                 RETURN(-ENOMEM);
1929
1930         filter->fo_iobuf_count = OSS_THREADS_MAX;
1931
1932         RETURN(0);
1933 }
1934
1935 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1936  * many threads there will be so we allocate a large empty array and only
1937  * fill in those slots that are actually in use.
1938  * If we haven't allocated a pool entry for this thread before, do so now. */
1939 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1940 {
1941         int thread_id                    = (oti && oti->oti_thread) ?
1942                                            oti->oti_thread->t_id : -1;
1943         struct filter_iobuf  *pool       = NULL;
1944         struct filter_iobuf **pool_place = NULL;
1945
1946         if (thread_id >= 0) {
1947                 LASSERT(thread_id < filter->fo_iobuf_count);
1948                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1949         }
1950
1951         if (unlikely(pool == NULL)) {
1952                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1953                                           PTLRPC_MAX_BRW_PAGES);
1954                 if (pool_place != NULL)
1955                         *pool_place = pool;
1956         }
1957
1958         return pool;
1959 }
1960
1961 /* mount the file system (secretly).  lustre_cfg parameters are:
1962  * 1 = device
1963  * 2 = fstype
1964  * 3 = flags: failover=f, failout=n
1965  * 4 = mount options
1966  */
1967 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1968                         void *option)
1969 {
1970         struct filter_obd *filter = &obd->u.filter;
1971         struct vfsmount *mnt;
1972         struct file_system_type *type;
1973         struct lustre_mount_info *lmi;
1974         struct obd_uuid uuid;
1975         __u8 *uuid_ptr;
1976         char *str, *label;
1977         char ns_name[48];
1978         struct request_queue *q;
1979         int rc, i;
1980         ENTRY;
1981
1982         if (lcfg->lcfg_bufcount < 3 ||
1983             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1984             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1985                 RETURN(-EINVAL);
1986
1987         lmi = server_get_mount(obd->obd_name);
1988         if (lmi) {
1989                 /* We already mounted in lustre_fill_super.
1990                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1991                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1992                 mnt = lmi->lmi_mnt;
1993                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1994         } else {
1995                 /* old path - used by lctl */
1996                 CERROR("Using old MDS mount method\n");
1997                 type = get_fs_type(lustre_cfg_string(lcfg, 2));
1998                 if (!type) {
1999                         CERROR("get_fs_type failed\n");
2000                         RETURN(-ENODEV);
2001                 }
2002                 mnt = vfs_kern_mount(type, MS_NOATIME|MS_NODIRATIME,
2003                                      lustre_cfg_string(lcfg, 1), option);
2004                 cfs_module_put(type->owner);
2005                 if (IS_ERR(mnt)) {
2006                         rc = PTR_ERR(mnt);
2007                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2008                                            lustre_cfg_string(lcfg, 1), rc);
2009                         RETURN(rc);
2010                 }
2011
2012                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2013         }
2014         if (IS_ERR(obd->obd_fsops))
2015                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2016
2017         rc = filter_iobuf_pool_init(filter);
2018         if (rc != 0)
2019                 GOTO(err_ops, rc);
2020
2021         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2022                 CERROR("%s: Underlying device is marked as read-only. "
2023                        "Setup failed\n", obd->obd_name);
2024                 GOTO(err_ops, rc = -EROFS);
2025         }
2026
2027         /* failover is the default */
2028         obd->obd_replayable = 1;
2029
2030         /* disable connection until configuration finishes */
2031         obd->obd_no_conn = 1;
2032
2033         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2034                 str = lustre_cfg_string(lcfg, 3);
2035                 if (strchr(str, 'n')) {
2036                         CWARN("%s: recovery disabled\n", obd->obd_name);
2037                         obd->obd_replayable = 0;
2038                 }
2039         }
2040
2041         obd->u.obt.obt_magic = OBT_MAGIC;
2042         obd->u.obt.obt_vfsmnt = mnt;
2043         obd->u.obt.obt_sb = mnt->mnt_sb;
2044         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2045         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2046
2047         rc = fsfilt_setup(obd, obd->u.obt.obt_sb);
2048         if (rc)
2049                 GOTO(err_ops, rc);
2050
2051         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2052         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2053         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2054         obd->obd_lvfs_ctxt.fs = get_ds();
2055         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2056
2057         cfs_mutex_init(&filter->fo_init_lock);
2058         filter->fo_committed_group = 0;
2059         filter->fo_destroys_in_progress = 0;
2060         for (i = 0; i < 32; i++)
2061                 cfs_mutex_init(&filter->fo_create_locks[i]);
2062
2063         cfs_spin_lock_init(&filter->fo_objidlock);
2064         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2065         cfs_mutex_init(&filter->fo_alloc_lock);
2066         init_brw_stats(&filter->fo_filter_stats);
2067         cfs_spin_lock_init(&filter->fo_flags_lock);
2068         filter->fo_read_cache = 1; /* enable read-only cache by default */
2069         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2070         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2071         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2072         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2073         filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2074         filter_slc_set(filter); /* initialize sync on lock cancel */
2075
2076         rc = filter_prep(obd);
2077         if (rc)
2078                 GOTO(err_ops, rc);
2079
2080         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2081         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2082
2083         filter->fo_fl_oss_capa = 1;
2084
2085         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2086         filter->fo_capa_hash = init_capa_hash();
2087         if (filter->fo_capa_hash == NULL)
2088                 GOTO(err_post, rc = -ENOMEM);
2089
2090         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2091         obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2092                                                 LDLM_NAMESPACE_SERVER,
2093                                                 LDLM_NAMESPACE_GREEDY,
2094                                                 LDLM_NS_TYPE_OST);
2095         if (obd->obd_namespace == NULL)
2096                 GOTO(err_post, rc = -ENOMEM);
2097         obd->obd_namespace->ns_lvbp = obd;
2098         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2099         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2100
2101         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2102                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2103
2104         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2105         if (rc) {
2106                 CERROR("failed to setup llogging subsystems\n");
2107                 GOTO(err_post, rc);
2108         }
2109
2110         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2111         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2112         /* do this after llog being initialized */
2113         filter_adapt_sptlrpc_conf(obd, 1);
2114
2115         rc = lquota_setup(filter_quota_interface_ref, obd);
2116         if (rc)
2117                 GOTO(err_post, rc);
2118
2119         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2120         if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2121             queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2122                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2123                               "for larger I/O requests: max_sectors = %u "
2124                               "could be up to max_hw_sectors=%u\n",
2125                               obd->obd_name, mnt->mnt_sb->s_id,
2126                               queue_max_sectors(q), queue_max_hw_sectors(q));
2127
2128         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2129         if (uuid_ptr != NULL) {
2130                 class_uuid_unparse(uuid_ptr, &uuid);
2131                 str = uuid.uuid;
2132         } else {
2133                 str = "no UUID";
2134         }
2135
2136         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2137         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2138                       obd->obd_name, label ?: str, lmi ? "on " : "",
2139                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2140                       obd->obd_replayable ? "enabled" : "disabled");
2141
2142         RETURN(0);
2143
2144 err_post:
2145         filter_post(obd);
2146 err_ops:
2147         fsfilt_put_ops(obd->obd_fsops);
2148         filter_iobuf_pool_done(filter);
2149 err_mntput:
2150         server_put_mount(obd->obd_name, mnt);
2151         obd->u.obt.obt_sb = 0;
2152         return rc;
2153 }
2154
2155 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2156 {
2157         struct lprocfs_static_vars lvars;
2158         cfs_proc_dir_entry_t *entry;
2159         unsigned long addr;
2160         struct page *page;
2161         int rc;
2162         ENTRY;
2163
2164         CLASSERT(offsetof(struct obd_device, u.obt) ==
2165                  offsetof(struct obd_device, u.filter.fo_obt));
2166
2167         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2168                 RETURN(-EINVAL);
2169
2170         /* lprocfs must be setup before the filter so state can be safely added
2171          * to /proc incrementally as the filter is setup */
2172         lprocfs_filter_init_vars(&lvars);
2173         rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2174         if (rc) {
2175                 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2176                        obd->obd_name, rc);
2177                 RETURN(rc);
2178         }
2179
2180         rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2181         if (rc) {
2182                 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2183                        obd->obd_name, rc);
2184                 GOTO(obd_cleanup, rc);
2185         }
2186
2187         /* Init obdfilter private stats here */
2188         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2189                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2190         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2191                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2192         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2193                              LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2194                              "get_page", "usec");
2195         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2196                              LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2197         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2198                              LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2199         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2200                              LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2201         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2202                              LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2203
2204         rc = lproc_filter_attach_seqstat(obd);
2205         if (rc) {
2206                 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2207                 GOTO(free_obd_stats, rc);
2208         }
2209
2210         entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2211         if (IS_ERR(entry)) {
2212                 rc = PTR_ERR(entry);
2213                 CERROR("%s: error %d setting up lprocfs for %s\n",
2214                        obd->obd_name, rc, "exports");
2215                 GOTO(free_obd_stats, rc);
2216         }
2217         obd->obd_proc_exports_entry = entry;
2218
2219         entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2220                                    lprocfs_nid_stats_clear_read,
2221                                    lprocfs_nid_stats_clear_write, obd, NULL);
2222         if (IS_ERR(entry)) {
2223                 rc = PTR_ERR(entry);
2224                 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2225                        obd->obd_name, rc);
2226                 GOTO(free_obd_stats, rc);
2227         }
2228
2229         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2230         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2231         if (!page)
2232                 GOTO(remove_entry_clear, rc = -ENOMEM);
2233         addr = (unsigned long)cfs_page_address(page);
2234         clear_page((void *)addr);
2235         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2236                LUSTRE_CFG_BUFLEN(lcfg, 4));
2237         rc = filter_common_setup(obd, lcfg, (void *)addr);
2238         OBD_PAGE_FREE(page);
2239         if (rc) {
2240                 CERROR("%s: filter_common_setup failed: %d.\n",
2241                        obd->obd_name, rc);
2242                 GOTO(remove_entry_clear, rc);
2243         }
2244
2245         RETURN(0);
2246
2247 remove_entry_clear:
2248         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2249 free_obd_stats:
2250         lprocfs_free_obd_stats(obd);
2251 obd_cleanup:
2252         lprocfs_obd_cleanup(obd);
2253         return rc;
2254 }
2255
2256 static struct llog_operations filter_mds_ost_repl_logops;
2257
2258 static struct llog_operations filter_size_orig_logops = {
2259         .lop_setup   = llog_obd_origin_setup,
2260         .lop_cleanup = llog_obd_origin_cleanup,
2261         .lop_add     = llog_obd_origin_add
2262 };
2263
2264 static int filter_olg_fini(struct obd_llog_group *olg)
2265 {
2266         struct llog_ctxt *ctxt;
2267         int rc = 0, rc2 = 0;
2268         ENTRY;
2269
2270         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2271         if (ctxt)
2272                 rc = llog_cleanup(ctxt);
2273
2274         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2275         if (ctxt) {
2276                 rc2 = llog_cleanup(ctxt);
2277                 if (!rc)
2278                         rc = rc2;
2279         }
2280
2281         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2282         if (ctxt) {
2283                 rc2 = llog_cleanup(ctxt);
2284                 if (!rc)
2285                         rc = rc2;
2286         }
2287
2288         RETURN(rc);
2289 }
2290
2291 static int
2292 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2293                 struct obd_device *tgt)
2294 {
2295         int rc;
2296         ENTRY;
2297
2298         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2299                         &filter_mds_ost_repl_logops);
2300         if (rc)
2301                 GOTO(cleanup, rc);
2302
2303         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2304                         &filter_size_orig_logops);
2305         if (rc)
2306                 GOTO(cleanup, rc);
2307         EXIT;
2308 cleanup:
2309         if (rc)
2310                 filter_olg_fini(olg);
2311         return rc;
2312 }
2313
2314 /**
2315  * Init the default olg, which is embeded in the obd_device, for filter.
2316  */
2317 static int
2318 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2319                         struct obd_device *tgt)
2320 {
2321         struct filter_obd *filter = &obd->u.filter;
2322         struct llog_ctxt *ctxt;
2323         int rc;
2324         ENTRY;
2325
2326         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2327         if (!filter->fo_lcm)
2328                 RETURN(-ENOMEM);
2329
2330         filter_mds_ost_repl_logops = llog_client_ops;
2331         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2332         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2333         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2334
2335         rc = filter_olg_init(obd, olg, tgt);
2336         if (rc)
2337                 GOTO(cleanup_lcm, rc);
2338
2339         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2340                         &llog_lvfs_ops);
2341         if (rc)
2342                 GOTO(cleanup_olg, rc);
2343
2344         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2345         if (!ctxt) {
2346                 CERROR("Can't get ctxt for %p:%x\n", olg,
2347                        LLOG_MDS_OST_REPL_CTXT);
2348                 GOTO(cleanup_olg, rc = -ENODEV);
2349         }
2350         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2351         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2352         llog_ctxt_put(ctxt);
2353
2354         RETURN(0);
2355 cleanup_olg:
2356         filter_olg_fini(olg);
2357 cleanup_lcm:
2358         llog_recov_thread_fini(filter->fo_lcm, 1);
2359         filter->fo_lcm = NULL;
2360         return rc;
2361 }
2362
2363 static int
2364 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2365                  struct obd_device *tgt, int *index)
2366 {
2367         struct filter_obd *filter = &obd->u.filter;
2368         struct llog_ctxt *ctxt;
2369         int rc;
2370         ENTRY;
2371
2372         LASSERT(olg != NULL);
2373         if (olg == &obd->obd_olg)
2374                 return filter_default_olg_init(obd, olg, tgt);
2375
2376         LASSERT(filter->fo_lcm != NULL);
2377         rc = filter_olg_init(obd, olg, tgt);
2378         if (rc)
2379                 RETURN(rc);
2380         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2381         if (!ctxt) {
2382                 CERROR("Can't get ctxt for %p:%x\n", olg,
2383                        LLOG_MDS_OST_REPL_CTXT);
2384                 filter_olg_fini(olg);
2385                 RETURN(-ENODEV);
2386         }
2387         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2388         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2389         llog_ctxt_put(ctxt);
2390         RETURN(rc);
2391 }
2392
2393 static int filter_llog_finish(struct obd_device *obd, int count)
2394 {
2395         struct filter_obd *filter = &obd->u.filter;
2396         struct llog_ctxt *ctxt;
2397         ENTRY;
2398
2399         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2400         if (ctxt) {
2401                 /*
2402                  * Make sure that no cached llcds left in recov_thread.
2403                  * We actually do sync in disconnect time, but disconnect
2404                  * may not come being marked rq_no_resend = 1.
2405                  */
2406                 llog_sync(ctxt, NULL);
2407
2408                 /*
2409                  * Balance class_import_get() in llog_receptor_accept().
2410                  * This is safe to do, as llog is already synchronized
2411                  * and its import may go.
2412                  */
2413                 cfs_mutex_lock(&ctxt->loc_mutex);
2414                 if (ctxt->loc_imp) {
2415                         class_import_put(ctxt->loc_imp);
2416                         ctxt->loc_imp = NULL;
2417                 }
2418                 cfs_mutex_unlock(&ctxt->loc_mutex);
2419                 llog_ctxt_put(ctxt);
2420         }
2421
2422         if (filter->fo_lcm) {
2423                 cfs_mutex_lock(&ctxt->loc_mutex);
2424                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2425                 filter->fo_lcm = NULL;
2426                 cfs_mutex_unlock(&ctxt->loc_mutex);
2427         }
2428         RETURN(filter_olg_fini(&obd->obd_olg));
2429 }
2430
2431 /**
2432  * Find the group llog according to group index in the llog group list.
2433  */
2434 static struct obd_llog_group *
2435 filter_find_olg_internal(struct filter_obd *filter, int group)
2436 {
2437         struct obd_llog_group *olg;
2438
2439         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2440         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2441                 if (olg->olg_seq == group)
2442                         RETURN(olg);
2443         }
2444         RETURN(NULL);
2445 }
2446
2447 /**
2448  * Find the group llog according to group index on the filter
2449  */
2450 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2451 {
2452         struct obd_llog_group *olg = NULL;
2453         struct filter_obd *filter;
2454
2455         filter = &obd->u.filter;
2456
2457         if (group == FID_SEQ_LLOG)
2458                 RETURN(&obd->obd_olg);
2459
2460         cfs_spin_lock(&filter->fo_llog_list_lock);
2461         olg = filter_find_olg_internal(filter, group);
2462         cfs_spin_unlock(&filter->fo_llog_list_lock);
2463
2464         RETURN(olg);
2465 }
2466 /**
2467  * Find the llog_group of the filter according to the group. If it can not
2468  * find, create the llog_group, which only happens when mds is being synced
2469  * with OST.
2470  */
2471 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2472 {
2473         struct obd_llog_group *olg = NULL, *olg_new = NULL;
2474         struct filter_obd *filter;
2475         int rc;
2476
2477         filter = &obd->u.filter;
2478
2479         if (group == FID_SEQ_LLOG)
2480                 RETURN(&obd->obd_olg);
2481
2482         OBD_ALLOC_PTR(olg_new);
2483         if (olg_new == NULL)
2484                RETURN(ERR_PTR(-ENOMEM));
2485
2486         cfs_spin_lock(&filter->fo_llog_list_lock);
2487         olg = filter_find_olg_internal(filter, group);
2488         if (olg) {
2489                 if (olg->olg_initializing) {
2490                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2491                 } else {
2492                         GOTO(out_unlock, olg);
2493                 }
2494         } else {
2495                 /* set as the newly allocated one */
2496                 olg = olg_new;
2497                 olg_new = NULL;
2498         }
2499
2500         llog_group_init(olg, group);
2501         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2502         olg->olg_initializing = 1;
2503         cfs_spin_unlock(&filter->fo_llog_list_lock);
2504
2505         rc = obd_llog_init(obd, olg, obd, NULL);
2506         if (rc) {
2507                cfs_spin_lock(&filter->fo_llog_list_lock);
2508                cfs_list_del(&olg->olg_list);
2509                cfs_spin_unlock(&filter->fo_llog_list_lock);
2510                OBD_FREE_PTR(olg);
2511                GOTO(out, olg = ERR_PTR(-ENOMEM));
2512         }
2513         cfs_spin_lock(&filter->fo_llog_list_lock);
2514         olg->olg_initializing = 0;
2515         cfs_spin_unlock(&filter->fo_llog_list_lock);
2516         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2517               obd->obd_name, group, olg);
2518 out:
2519         RETURN(olg);
2520
2521 out_unlock:
2522         cfs_spin_unlock(&filter->fo_llog_list_lock);
2523         if (olg_new)
2524                OBD_FREE_PTR(olg_new);
2525         goto out;
2526 }
2527
2528 static int filter_llog_connect(struct obd_export *exp,
2529                                struct llogd_conn_body *body)
2530 {
2531         struct obd_device *obd = exp->exp_obd;
2532         struct llog_ctxt *ctxt;
2533         struct obd_llog_group *olg;
2534         int rc;
2535         ENTRY;
2536
2537         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2538                obd->obd_name, body->lgdc_logid.lgl_oid,
2539                body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2540
2541         olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2542         if (!olg) {
2543                 CERROR(" %s: can not find olg of group %d\n",
2544                        obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2545                 RETURN(-ENOENT);
2546         }
2547         llog_group_set_export(olg, exp);
2548
2549         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2550         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2551                  body->lgdc_ctxt_idx);
2552
2553         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2554               obd->obd_name, body->lgdc_logid.lgl_oid,
2555               body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2556
2557         cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2558         obd->u.filter.fo_mds_ost_sync = 1;
2559         cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2560         rc = llog_connect(ctxt, &body->lgdc_logid,
2561                           &body->lgdc_gen, NULL);
2562         llog_ctxt_put(ctxt);
2563         if (rc != 0)
2564                 CERROR("failed to connect rc %d idx %d\n", rc,
2565                                 body->lgdc_ctxt_idx);
2566
2567         RETURN(rc);
2568 }
2569
2570 static int filter_llog_preclean(struct obd_device *obd)
2571 {
2572         struct obd_llog_group *olg, *tmp;
2573         struct filter_obd *filter;
2574         cfs_list_t  remove_list;
2575         int rc = 0;
2576         ENTRY;
2577
2578         rc = obd_llog_finish(obd, 0);
2579         if (rc)
2580                 CERROR("failed to cleanup llogging subsystem\n");
2581
2582         filter = &obd->u.filter;
2583         CFS_INIT_LIST_HEAD(&remove_list);
2584
2585         cfs_spin_lock(&filter->fo_llog_list_lock);
2586         while (!cfs_list_empty(&filter->fo_llog_list)) {
2587                 olg = cfs_list_entry(filter->fo_llog_list.next,
2588                                      struct obd_llog_group, olg_list);
2589                 cfs_list_del(&olg->olg_list);
2590                 cfs_list_add(&olg->olg_list, &remove_list);
2591         }
2592         cfs_spin_unlock(&filter->fo_llog_list_lock);
2593
2594         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2595                 cfs_list_del_init(&olg->olg_list);
2596                 rc = filter_olg_fini(olg);
2597                 if (rc)
2598                         CERROR("failed to cleanup llogging subsystem for %u\n",
2599                                olg->olg_seq);
2600                 OBD_FREE_PTR(olg);
2601         }
2602
2603         RETURN(rc);
2604 }
2605
2606 static int filter_precleanup(struct obd_device *obd,
2607                              enum obd_cleanup_stage stage)
2608 {
2609         int rc = 0;
2610         ENTRY;
2611
2612         switch(stage) {
2613         case OBD_CLEANUP_EARLY:
2614                 break;
2615         case OBD_CLEANUP_EXPORTS:
2616                 /* Stop recovery before namespace cleanup. */
2617                 target_recovery_fini(obd);
2618
2619                 obd_exports_barrier(obd);
2620                 obd_zombie_barrier();
2621
2622                 rc = filter_llog_preclean(obd);
2623                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2624                 lprocfs_free_per_client_stats(obd);
2625                 lprocfs_obd_cleanup(obd);
2626                 lprocfs_free_obd_stats(obd);
2627                 lquota_cleanup(filter_quota_interface_ref, obd);
2628                 break;
2629         }
2630         RETURN(rc);
2631 }
2632
2633 static int filter_cleanup(struct obd_device *obd)
2634 {
2635         struct filter_obd *filter = &obd->u.filter;
2636         ENTRY;
2637
2638         if (obd->obd_fail)
2639                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2640                               "will be preserved.\n", obd->obd_name);
2641
2642         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2643         obd->obd_namespace = NULL;
2644
2645         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2646
2647         if (obd->u.obt.obt_sb == NULL)
2648                 RETURN(0);
2649
2650         filter_post(obd);
2651
2652         ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2653         shrink_dcache_sb(obd->u.obt.obt_sb);
2654
2655         server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2656         obd->u.obt.obt_sb = NULL;
2657
2658         fsfilt_put_ops(obd->obd_fsops);
2659
2660         filter_iobuf_pool_done(filter);
2661
2662         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2663
2664         RETURN(0);
2665 }
2666
2667 static int filter_connect_internal(struct obd_export *exp,
2668                                    struct obd_connect_data *data,
2669                                    int reconnect)
2670 {
2671         struct filter_export_data *fed = &exp->exp_filter_data;
2672
2673         if (!data)
2674                 RETURN(0);
2675
2676         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2677                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2678                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2679                data->ocd_connect_flags, data->ocd_version,
2680                data->ocd_grant, data->ocd_index);
2681
2682         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2683                 CWARN("!!! This export (nid %s) used object group %d "
2684                        "earlier; now it's trying to use group %d!  This could "
2685                        "be a bug in the MDS. Please report to "
2686                        "http://bugs.whamcloud.com/\n",
2687                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2688                 RETURN(-EPROTO);
2689         }
2690         fed->fed_group = data->ocd_group;
2691
2692         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2693         exp->exp_connect_flags = data->ocd_connect_flags;
2694         data->ocd_version = LUSTRE_VERSION_CODE;
2695
2696         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2697         if (data->ocd_connect_flags & OBD_CONNECT_MDS)
2698                 CWARN("%s: Received MDS connection for group %u\n",
2699                       exp->exp_obd->obd_name, data->ocd_group);
2700         else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
2701                 RETURN(-EPROTO);
2702
2703         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2704                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2705                 obd_size left, want;
2706
2707                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2708                 left = filter_grant_space_left(exp);
2709                 want = data->ocd_grant;
2710                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2711                 data->ocd_grant = fed->fed_grant;
2712                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2713
2714                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2715                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2716                        exp->exp_client_uuid.uuid, exp,
2717                        data->ocd_grant, want, left);
2718
2719                 filter->fo_tot_granted_clients ++;
2720         }
2721
2722         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2723                 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2724                 int index = le32_to_cpu(lsd->lsd_ost_index);
2725
2726                 if (!(lsd->lsd_feature_compat &
2727                       cpu_to_le32(OBD_COMPAT_OST))) {
2728                         /* this will only happen on the first connect */
2729                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2730                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2731                         /* sync is not needed here as filter_client_add will
2732                          * set exp_need_sync flag */
2733                         filter_update_server_data(exp->exp_obd);
2734                 } else if (index != data->ocd_index) {
2735                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2736                                            " %u doesn't match actual OST index"
2737                                            " %u in last_rcvd file, bad "
2738                                            "configuration?\n",
2739                                            obd_export_nid2str(exp), index,
2740                                            data->ocd_index);
2741                         RETURN(-EBADF);
2742                 }
2743                 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2744                  * FIXME: We don't strictly need the COMPAT flag for that,
2745                  * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2746                  * FIXME: We needed it for the index, as index 0 is valid. */
2747         }
2748
2749         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2750                 data->ocd_brw_size = 65536;
2751         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2752                 data->ocd_brw_size = min(data->ocd_brw_size,
2753                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2754                 if (data->ocd_brw_size == 0) {
2755                         CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2756                                " ocd_version: %x ocd_grant: %d ocd_index: %u "
2757                                "ocd_brw_size is unexpectedly zero, "
2758                                "network data corruption?"
2759                                "Refusing connection of this client\n",
2760                                 exp->exp_obd->obd_name,
2761                                 exp->exp_client_uuid.uuid,
2762                                 exp, data->ocd_connect_flags, data->ocd_version,
2763                                 data->ocd_grant, data->ocd_index);
2764                         RETURN(-EPROTO);
2765                 }
2766         }
2767
2768         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2769                 __u32 cksum_types = data->ocd_cksum_types;
2770
2771                 /* The client set in ocd_cksum_types the checksum types it
2772                  * supports. We have to mask off the algorithms that we don't
2773                  * support */
2774                 data->ocd_cksum_types &= cksum_types_supported();
2775
2776                 /* 1.6.4- only support CRC32 and didn't set ocd_cksum_types */
2777                 if (unlikely(data->ocd_cksum_types == 0))
2778                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2779
2780                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2781                                    "%x\n", exp->exp_obd->obd_name,
2782                                    obd_export_nid2str(exp), cksum_types,
2783                                    data->ocd_cksum_types);
2784         } else {
2785                 /* This client does not support OBD_CONNECT_CKSUM
2786                  * fall back to CRC32 */
2787                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2788                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2789                                    exp->exp_obd->obd_name,
2790                                    obd_export_nid2str(exp));
2791         }
2792
2793         if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2794                 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2795
2796         RETURN(0);
2797 }
2798
2799 static int filter_reconnect(const struct lu_env *env,
2800                             struct obd_export *exp, struct obd_device *obd,
2801                             struct obd_uuid *cluuid,
2802                             struct obd_connect_data *data,
2803                             void *localdata)
2804 {
2805         int rc;
2806         ENTRY;
2807
2808         if (exp == NULL || obd == NULL || cluuid == NULL)
2809                 RETURN(-EINVAL);
2810
2811         rc = filter_connect_internal(exp, data, 1);
2812         if (rc == 0)
2813                 filter_export_stats_init(obd, exp, localdata);
2814
2815         RETURN(rc);
2816 }
2817
2818 static int filter_connect(const struct lu_env *env,
2819                           struct obd_export **exp, struct obd_device *obd,
2820                           struct obd_uuid *cluuid,
2821                           struct obd_connect_data *data, void *localdata)
2822 {
2823         struct lvfs_run_ctxt saved;
2824         struct lustre_handle conn = { 0 };
2825         struct obd_export *lexp;
2826         int rc;
2827         ENTRY;
2828
2829         if (exp == NULL || obd == NULL || cluuid == NULL)
2830                 RETURN(-EINVAL);
2831
2832         rc = class_connect(&conn, obd, cluuid);
2833         if (rc)
2834                 RETURN(rc);
2835         lexp = class_conn2export(&conn);
2836         LASSERT(lexp != NULL);
2837
2838         rc = filter_connect_internal(lexp, data, 0);
2839         if (rc)
2840                 GOTO(cleanup, rc);
2841
2842         filter_export_stats_init(obd, lexp, localdata);
2843         if (obd->obd_replayable) {
2844                 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2845                 LASSERT(lcd);
2846                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2847                 rc = filter_client_add(obd, lexp, -1);
2848                 if (rc)
2849                         GOTO(cleanup, rc);
2850         }
2851
2852         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2853         rc = filter_read_groups(obd, data->ocd_group, 1);
2854         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2855         if (rc != 0) {
2856                 CERROR("can't read group %u\n", data->ocd_group);
2857                 GOTO(cleanup, rc);
2858         }
2859
2860         GOTO(cleanup, rc);
2861
2862 cleanup:
2863         if (rc) {
2864                 class_disconnect(lexp);
2865                 *exp = NULL;
2866         } else {
2867                 *exp = lexp;
2868         }
2869
2870         RETURN(rc);
2871 }
2872
2873 /* Do extra sanity checks for grant accounting.  We do this at connect,
2874  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2875  * always get rid of it or turn it off when we know accounting is good. */
2876 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2877 {
2878         struct filter_export_data *fed;
2879         struct obd_export *exp;
2880         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2881         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2882         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2883
2884         if (cfs_list_empty(&obd->obd_exports))
2885                 return;
2886
2887         /* We don't want to do this for large machines that do lots of
2888            mounts or unmounts.  It burns... */
2889         if (obd->obd_num_exports > 100)
2890                 return;
2891
2892         cfs_spin_lock(&obd->obd_osfs_lock);
2893         cfs_spin_lock(&obd->obd_dev_lock);
2894         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2895                 int error = 0;
2896                 fed = &exp->exp_filter_data;
2897                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2898                     fed->fed_dirty < 0)
2899                         error = 1;
2900                 if (maxsize > 0) { /* we may not have done a statfs yet */
2901                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2902                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2903                                  exp->exp_client_uuid.uuid, exp,
2904                                  fed->fed_grant, fed->fed_pending, maxsize);
2905                         LASSERTF(fed->fed_dirty <= maxsize,
2906                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2907                                  exp->exp_client_uuid.uuid, exp,
2908                                  fed->fed_dirty, maxsize);
2909                 }
2910                 if (error)
2911                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2912                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2913                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2914                 else
2915                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2916                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2917                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2918                 tot_granted += fed->fed_grant + fed->fed_pending;
2919                 tot_pending += fed->fed_pending;
2920                 tot_dirty += fed->fed_dirty;
2921         }
2922         fo_tot_granted = obd->u.filter.fo_tot_granted;
2923         fo_tot_pending = obd->u.filter.fo_tot_pending;
2924         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2925         cfs_spin_unlock(&obd->obd_dev_lock);
2926         cfs_spin_unlock(&obd->obd_osfs_lock);
2927
2928         /* Do these assertions outside the spinlocks so we don't kill system */
2929         if (tot_granted != fo_tot_granted)
2930                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2931                        func, tot_granted, fo_tot_granted);
2932         if (tot_pending != fo_tot_pending)
2933                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2934                        func, tot_pending, fo_tot_pending);
2935         if (tot_dirty != fo_tot_dirty)
2936                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2937                        func, tot_dirty, fo_tot_dirty);
2938         if (tot_pending > tot_granted)
2939                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2940                        func, tot_pending, tot_granted);
2941         if (tot_granted > maxsize)
2942                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2943                        func, tot_granted, maxsize);
2944         if (tot_dirty > maxsize)
2945                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2946                        func, tot_dirty, maxsize);
2947 }
2948
2949 /* Remove this client from the grant accounting totals.  We also remove
2950  * the export from the obd device under the osfs and dev locks to ensure
2951  * that the filter_grant_sanity_check() calculations are always valid.
2952  * The client should do something similar when it invalidates its import. */
2953 static void filter_grant_discard(struct obd_export *exp)
2954 {
2955         struct obd_device *obd = exp->exp_obd;
2956         struct filter_obd *filter = &obd->u.filter;
2957         struct filter_export_data *fed = &exp->exp_filter_data;
2958
2959         cfs_spin_lock(&obd->obd_osfs_lock);
2960         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2961                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2962                  obd->obd_name, filter->fo_tot_granted,
2963                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2964         filter->fo_tot_granted -= fed->fed_grant;
2965         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2966                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2967                  obd->obd_name, filter->fo_tot_pending,
2968                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2969         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2970         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2971                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2972                  obd->obd_name, filter->fo_tot_dirty,
2973                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2974         filter->fo_tot_dirty -= fed->fed_dirty;
2975         fed->fed_dirty = 0;
2976         fed->fed_grant = 0;
2977
2978         cfs_spin_unlock(&obd->obd_osfs_lock);
2979 }
2980
2981 static int filter_destroy_export(struct obd_export *exp)
2982 {
2983         struct filter_export_data *fed = &exp->exp_filter_data;
2984         ENTRY;
2985
2986         if (fed->fed_pending)
2987                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2988                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2989                        exp, fed->fed_pending);
2990
2991         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2992
2993         target_destroy_export(exp);
2994
2995         if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
2996                                      &exp->exp_client_uuid)))
2997                RETURN(0);
2998
2999         ldlm_destroy_export(exp);
3000         lut_client_free(exp);
3001
3002         if (!exp->exp_obd->obd_replayable)
3003                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3004
3005         filter_grant_discard(exp);
3006         filter_fmd_cleanup(exp);
3007
3008         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3009                 struct filter_obd *filter = &exp->exp_obd->u.filter;
3010                 if (filter->fo_tot_granted_clients > 0)
3011                         filter->fo_tot_granted_clients --;
3012         }
3013
3014         if (!(exp->exp_flags & OBD_OPT_FORCE))
3015                 filter_grant_sanity_check(exp->exp_obd, __func__);
3016
3017         RETURN(0);
3018 }
3019
3020 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3021 {
3022         struct obd_llog_group *olg_min, *olg;
3023         struct filter_obd *filter;
3024         int worked = -1, group;
3025         struct llog_ctxt *ctxt;
3026         ENTRY;
3027
3028         filter = &obd->u.filter;
3029
3030         /* we can't sync log holding spinlock. also, we do not want to get
3031          * into livelock. so we do following: loop over MDS's exports in
3032          * group order and skip already synced llogs -bzzz */
3033         do {
3034                 /* look for group with min. number, but > worked */
3035                 olg_min = NULL;
3036                 group = 1 << 30;
3037                 cfs_spin_lock(&filter->fo_llog_list_lock);
3038                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3039                         if (olg->olg_seq <= worked) {
3040                                 /* this group is already synced */
3041                                 continue;
3042                         }
3043                         if (group < olg->olg_seq) {
3044                                 /* we have group with smaller number to sync */
3045                                 continue;
3046                         }
3047                         /* store current minimal group */
3048                         olg_min = olg;
3049                         group = olg->olg_seq;
3050                 }
3051                 cfs_spin_unlock(&filter->fo_llog_list_lock);
3052
3053                 if (olg_min == NULL)
3054                         break;
3055
3056                 worked = olg_min->olg_seq;
3057                 if (olg_min->olg_exp &&
3058                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3059                         int err;
3060                         ctxt = llog_group_get_ctxt(olg_min,
3061                                                    LLOG_MDS_OST_REPL_CTXT);
3062                         if (ctxt) {
3063                                 err = llog_sync(ctxt, olg_min->olg_exp);
3064                                 llog_ctxt_put(ctxt);
3065                                 if (err) {
3066                                         CERROR("error flushing logs to MDS: "
3067                                                "rc %d\n", err);
3068                                 }
3069                         }
3070                 }
3071         } while (olg_min != NULL);
3072 }
3073
3074 /* Also incredibly similar to mds_disconnect */
3075 static int filter_disconnect(struct obd_export *exp)
3076 {
3077         struct obd_device *obd = exp->exp_obd;
3078         int rc;
3079         ENTRY;
3080
3081         LASSERT(exp);
3082         class_export_get(exp);
3083
3084         if (!(exp->exp_flags & OBD_OPT_FORCE))
3085                 filter_grant_sanity_check(obd, __func__);
3086         filter_grant_discard(exp);
3087
3088         /* Flush any remaining cancel messages out to the target */
3089         filter_sync_llogs(obd, exp);
3090
3091         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3092
3093         rc = server_disconnect_export(exp);
3094
3095         /* Do not erase record for recoverable client. */
3096         if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3097                 filter_client_del(exp);
3098         else
3099                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3100
3101         class_export_put(exp);
3102         RETURN(rc);
3103 }
3104
3105 /* reverse import is changed, sync all cancels */
3106 static void filter_revimp_update(struct obd_export *exp)
3107 {
3108         ENTRY;
3109
3110         LASSERT(exp);
3111         class_export_get(exp);
3112
3113         /* flush any remaining cancel messages out to the target */
3114         filter_sync_llogs(exp->exp_obd, exp);
3115         class_export_put(exp);
3116         EXIT;
3117 }
3118
3119 static int filter_ping(struct obd_export *exp)
3120 {
3121         filter_fmd_expire(exp);
3122         return 0;
3123 }
3124
3125 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3126                                   const char *what, int quiet)
3127 {
3128         struct dentry *dchild = NULL;
3129
3130         dchild = filter_fid2dentry(obd, NULL,  ostid->oi_seq, ostid->oi_id);
3131
3132         if (IS_ERR(dchild)) {
3133                 CERROR("%s error looking up object: "POSTID"\n",
3134                        what, ostid->oi_id, ostid->oi_seq);
3135                 RETURN(dchild);
3136         }
3137
3138         if (dchild->d_inode == NULL) {
3139                 if (!quiet)
3140                         CERROR("%s: %s on non-existent object: "POSTID" \n",
3141                                obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3142                 f_dput(dchild);
3143                 RETURN(ERR_PTR(-ENOENT));
3144         }
3145
3146 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0)
3147         /* Try to correct for a bug in 2.1.0 (LU-221) that caused negative
3148          * timestamps to appear to be in the far future, due old timestamp
3149          * being stored on disk as an unsigned value.  This fixes up any
3150          * bad values stored on disk before returning them to the client,
3151          * and ensures any timestamp updates are correct.  LU-1042 */
3152         if (unlikely(LTIME_S(dchild->d_inode->i_atime) == LU221_BAD_TIME))
3153                 LTIME_S(dchild->d_inode->i_atime) = 0;
3154         if (unlikely(LTIME_S(dchild->d_inode->i_mtime) == LU221_BAD_TIME))
3155                 LTIME_S(dchild->d_inode->i_mtime) = 0;
3156         if (unlikely(LTIME_S(dchild->d_inode->i_ctime) == LU221_BAD_TIME))
3157                 LTIME_S(dchild->d_inode->i_ctime) = 0;
3158 #else
3159 #warning "remove old LU-221/LU-1042 workaround code"
3160 #endif
3161
3162         return dchild;
3163 }
3164
3165 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3166 {
3167         struct dentry *dentry = NULL;
3168         struct obd_device *obd;
3169         __u64 curr_version;
3170         int rc = 0;
3171         ENTRY;
3172
3173         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3174                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3175         if (rc)
3176                 RETURN(rc);
3177
3178         obd = class_exp2obd(exp);
3179         if (obd == NULL) {
3180                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3181                 RETURN(-EINVAL);
3182         }
3183
3184         dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3185         if (IS_ERR(dentry))
3186                 RETURN(PTR_ERR(dentry));
3187
3188         /* Limit the valid bits in the return data to what we actually use */
3189         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3190         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3191
3192         /* Store inode version in reply */
3193         curr_version = fsfilt_get_version(exp->exp_obd, dentry->d_inode);
3194         if ((__s64)curr_version != -EOPNOTSUPP) {
3195                 oinfo->oi_oa->o_valid |= OBD_MD_FLDATAVERSION;
3196                 oinfo->oi_oa->o_data_version = curr_version;
3197         }
3198
3199         f_dput(dentry);
3200         RETURN(rc);
3201 }
3202
3203 /* this should be enabled/disabled in condition to enabled/disabled large
3204  * inodes (fast EAs) in backing store FS. */
3205 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3206                         void *handle, struct obdo *oa)
3207 {
3208         struct obd_device *obd = exp->exp_obd;
3209         int rc = 0;
3210         ENTRY;
3211
3212         if (oa->o_valid & OBD_MD_FLFID) {
3213                 struct filter_fid ff;
3214
3215                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3216                         oa->o_seq = 0;
3217                 /* packing fid and converting it to LE for storing into EA.
3218                  * Here ->o_stripe_idx should be filled by LOV and rest of
3219                  * fields - by client. */
3220                 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3221                 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3222                 /* XXX: we are ignoring o_parent_ver here, since this should
3223                  *      be the same for all objects in this fileset. */
3224                 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3225                 ff.ff_objid = cpu_to_le64(oa->o_id);
3226                 ff.ff_seq = cpu_to_le64(oa->o_seq);
3227
3228                 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3229                        LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3230                        oa->o_seq);
3231
3232                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3233                 if (rc)
3234                         CERROR("store fid in object failed! rc: %d\n", rc);
3235         } else {
3236                 CDEBUG(D_HA, "OSS object without fid info!\n");
3237         }
3238
3239         RETURN(rc);
3240 }
3241
3242 /* this is called from filter_truncate() until we have filter_punch() */
3243 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3244                             struct obdo *oa, struct obd_trans_info *oti)
3245 {
3246         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3247         struct llog_cookie *fcc = NULL;
3248         struct filter_obd *filter;
3249         int rc, err, sync = 0;
3250         loff_t old_size = 0;
3251         unsigned int ia_valid;
3252         struct inode *inode;
3253         struct page *page = NULL;
3254         struct iattr iattr;
3255         void *handle;
3256         ENTRY;
3257
3258         LASSERT(dentry != NULL);
3259         LASSERT(!IS_ERR(dentry));
3260
3261         inode = dentry->d_inode;
3262         LASSERT(inode != NULL);
3263
3264         filter = &exp->exp_obd->u.filter;
3265         iattr_from_obdo(&iattr, oa, oa->o_valid);
3266         ia_valid = iattr.ia_valid;
3267
3268         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3269                 OBD_ALLOC(fcc, sizeof(*fcc));
3270                 if (fcc != NULL)
3271                         *fcc = oa->o_lcookie;
3272         }
3273         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3274                 unsigned long now = jiffies;
3275                 /* Filter truncates and writes are serialized by
3276                  * i_alloc_sem, see the comment in
3277                  * filter_preprw_write.*/
3278                 if (ia_valid & ATTR_SIZE)
3279                         down_write(&inode->i_alloc_sem);
3280                 LOCK_INODE_MUTEX(inode);
3281                 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
3282                 old_size = i_size_read(inode);
3283         }
3284
3285         /* VBR: version recovery check */
3286         rc = filter_version_get_check(exp, oti, inode);
3287         if (rc)
3288                 GOTO(out_unlock, rc);
3289
3290         /* Let's pin the last page so that ldiskfs_truncate
3291          * should not start GFP_FS allocation. */
3292         if (ia_valid & ATTR_SIZE) {
3293                 page = grab_cache_page(inode->i_mapping,
3294                                        iattr.ia_size >> PAGE_CACHE_SHIFT);
3295                 if (page == NULL)
3296                         GOTO(out_unlock, rc = -ENOMEM);
3297
3298                 unlock_page(page);
3299         }
3300
3301         /* If the inode still has SUID+SGID bits set (see filter_precreate())
3302          * then we will accept the UID+GID sent by the client during write for
3303          * initializing the ownership of this inode.  We only allow this to
3304          * happen once so clear these bits in setattr. In 2.6 kernels it is
3305          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3306          * the flags that are actually being set. */
3307         if (ia_valid & (ATTR_UID | ATTR_GID)) {
3308                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3309                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3310
3311                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3312                         if (!(ia_valid & ATTR_MODE)) {
3313                                 iattr.ia_mode = inode->i_mode;
3314                                 iattr.ia_valid |= ATTR_MODE;
3315                         }
3316                         iattr.ia_mode &= ~S_ISUID;
3317                 }
3318                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3319                         if (!(iattr.ia_valid & ATTR_MODE)) {
3320                                 iattr.ia_mode = inode->i_mode;
3321                                 iattr.ia_valid |= ATTR_MODE;
3322                         }
3323                         iattr.ia_mode &= ~S_ISGID;
3324                 }
3325
3326                 orig_ids[USRQUOTA] = inode->i_uid;
3327                 orig_ids[GRPQUOTA] = inode->i_gid;
3328                 handle = fsfilt_start_log(exp->exp_obd, inode,
3329                                           FSFILT_OP_SETATTR, oti, 1);
3330                 if (IS_ERR(handle))
3331                         GOTO(out_unlock, rc = PTR_ERR(handle));
3332
3333                 /* update inode EA only once when inode is suid bit marked. As
3334                  * on 2.6.x UID and GID may be set separately, we check here
3335                  * only one of them to avoid double setting. */
3336                 if (inode->i_mode & S_ISUID)
3337                         filter_update_fidea(exp, inode, handle, oa);
3338         } else {
3339                 handle = fsfilt_start(exp->exp_obd, inode,
3340                                       FSFILT_OP_SETATTR, oti);
3341                 if (IS_ERR(handle))
3342                         GOTO(out_unlock, rc = PTR_ERR(handle));
3343         }
3344
3345         /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
3346         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3347                 ll_vfs_dq_init(inode);
3348
3349         if (oa->o_valid & OBD_MD_FLFLAGS) {
3350                 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3351                                       FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3352         } else {
3353                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3354                 if (fcc != NULL)
3355                         /* set cancel cookie callback function */
3356                         sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3357                                                      filter_cancel_cookies_cb,
3358                                                      fcc);
3359         }
3360
3361         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3362                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3363
3364        /* The truncate might have used up our transaction credits.  Make sure
3365         * we have two left for the last_rcvd and VBR inode version updates. */
3366         err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3367
3368         /* Update inode version only if data has changed => size has changed */
3369         rc = filter_finish_transno(exp, ia_valid & ATTR_SIZE ? inode : NULL,
3370                                    oti, rc, sync);
3371
3372         if (sync) {
3373                 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3374                 fcc = NULL;
3375         }
3376
3377         err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3378         if (err) {
3379                 CERROR("error on commit, err = %d\n", err);
3380                 if (!rc)
3381                         rc = err;
3382         } else {
3383                 fcc = NULL;
3384         }
3385
3386         /* For a partial-page truncate flush the page to disk immediately
3387          * to avoid data corruption during direct disk write. b=17397 */
3388         if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3389             old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3390                 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3391                                                iattr.ia_size + 1);
3392                 if (!rc)
3393                         rc = err;
3394         }
3395
3396         EXIT;
3397
3398 out_unlock:
3399         if (page)
3400                 page_cache_release(page);
3401
3402         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3403                 UNLOCK_INODE_MUTEX(inode);
3404         if (ia_valid & ATTR_SIZE)
3405                 up_write(&inode->i_alloc_sem);
3406         if (fcc)
3407                 OBD_FREE(fcc, sizeof(*fcc));
3408
3409         /* trigger quota release */
3410         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3411                 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3412                 int rc2 = lquota_adjust(filter_quota_interface_ref,
3413                                         exp->exp_obd, cur_ids,
3414                                         orig_ids, rc, FSFILT_OP_SETATTR);
3415                 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3416                        "filter adjust qunit. (rc:%d)\n", rc2);
3417         }
3418         return rc;
3419 }
3420
3421 /* this is called from filter_truncate() until we have filter_punch() */
3422 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3423                    struct obd_trans_info *oti)
3424 {
3425         struct obdo *oa = oinfo->oi_oa;
3426         struct lustre_capa *capa = oinfo_capa(oinfo);
3427         struct ldlm_res_id res_id;
3428         struct filter_mod_data *fmd;
3429         struct lvfs_run_ctxt saved;
3430         struct filter_obd *filter;
3431         struct ldlm_resource *res;
3432         struct dentry *dentry;
3433         __u64 opc = CAPA_OPC_META_WRITE;
3434         int rc;
3435         ENTRY;
3436
3437         if (oinfo->oi_flags & OBD_FL_PUNCH)
3438                 opc |= CAPA_OPC_OSS_TRUNC;
3439
3440         rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3441         if (rc)
3442                 RETURN(rc);
3443
3444         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3445                 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3446                 if (rc)
3447                         RETURN(rc);
3448         }
3449
3450         osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3451         /* This would be very bad - accidentally truncating a file when
3452          * changing the time or similar - bug 12203. */
3453         if (oa->o_valid & OBD_MD_FLSIZE &&
3454             oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3455                 static char mdsinum[48];
3456
3457                 if (oa->o_valid & OBD_MD_FLFID)
3458                         snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3459                                  oa->o_parent_seq, oa->o_parent_oid,
3460                                  oa->o_parent_ver);
3461                 else
3462                         mdsinum[0] = '\0';
3463
3464                 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3465                        "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3466                        oa->o_id, oa->o_seq, mdsinum);
3467                 RETURN(-EPERM);
3468         }
3469
3470         dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3471         if (IS_ERR(dentry))
3472                 RETURN(PTR_ERR(dentry));
3473
3474         filter = &exp->exp_obd->u.filter;
3475         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3476
3477         /*
3478          * We need to be atomic against a concurrent write
3479          * (which takes the semaphore for reading). fmd_mactime_xid
3480          * checks will have no effect if a write request with lower
3481          * xid starts just before a setattr and finishes later than
3482          * the setattr (see bug 21489, comment 27).
3483          */
3484         if (oa->o_valid &
3485             (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3486                 unsigned long now = jiffies;
3487                 down_write(&dentry->d_inode->i_alloc_sem);
3488                 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem");
3489                 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3490                 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3491                         fmd->fmd_mactime_xid = oti->oti_xid;
3492                 filter_fmd_put(exp, fmd);
3493                 up_write(&dentry->d_inode->i_alloc_sem);
3494         }
3495
3496         /* setting objects attributes (including owner/group) */
3497         rc = filter_setattr_internal(exp, dentry, oa, oti);
3498         if (rc)
3499                 GOTO(out_unlock, rc);
3500
3501         res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3502                                 &res_id, LDLM_EXTENT, 0);
3503
3504         if (res != NULL) {
3505                 LDLM_RESOURCE_ADDREF(res);
3506                 rc = ldlm_res_lvbo_update(res, NULL, 0);
3507                 LDLM_RESOURCE_DELREF(res);
3508                 ldlm_resource_putref(res);
3509         }
3510
3511         oa->o_valid = OBD_MD_FLID;
3512
3513         /* Quota release need uid/gid info */
3514         obdo_from_inode(oa, dentry->d_inode,
3515                         FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3516
3517         EXIT;
3518 out_unlock:
3519         f_dput(dentry);
3520         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3521         return rc;
3522 }
3523
3524 /* XXX identical to osc_unpackmd */
3525 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3526                            struct lov_mds_md *lmm, int lmm_bytes)
3527 {
3528         int lsm_size;
3529         ENTRY;
3530
3531         if (lmm != NULL) {
3532                 if (lmm_bytes < sizeof (*lmm)) {
3533                         CERROR("lov_mds_md too small: %d, need %d\n",
3534                                lmm_bytes, (int)sizeof(*lmm));
3535                         RETURN(-EINVAL);
3536                 }
3537                 /* XXX LOV_MAGIC etc check? */
3538
3539                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3540                         CERROR("lov_mds_md: zero lmm_object_id\n");
3541                         RETURN(-EINVAL);
3542                 }
3543         }
3544
3545         lsm_size = lov_stripe_md_size(1);
3546         if (lsmp == NULL)
3547                 RETURN(lsm_size);
3548
3549         if (*lsmp != NULL && lmm == NULL) {
3550                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3551                 OBD_FREE(*lsmp, lsm_size);
3552                 *lsmp = NULL;
3553                 RETURN(0);
3554         }
3555
3556         if (*lsmp == NULL) {
3557                 OBD_ALLOC(*lsmp, lsm_size);
3558                 if (*lsmp == NULL)
3559                         RETURN(-ENOMEM);
3560
3561                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3562                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3563                         OBD_FREE(*lsmp, lsm_size);
3564                         RETURN(-ENOMEM);
3565                 }
3566                 loi_init((*lsmp)->lsm_oinfo[0]);
3567         }
3568
3569         if (lmm != NULL) {
3570                 /* XXX zero *lsmp? */
3571                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3572                 LASSERT((*lsmp)->lsm_object_id);
3573         }
3574
3575         (*lsmp)->lsm_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
3576
3577         RETURN(lsm_size);
3578 }
3579
3580 /* caller must hold fo_create_locks[oa->o_seq] */
3581 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3582                                      struct filter_obd *filter)
3583 {
3584         struct obdo doa = { 0 }; /* XXX obdo on stack */
3585         obd_id last, id;
3586         int rc = 0;
3587         int skip_orphan;
3588         ENTRY;
3589
3590         LASSERT_MUTEX_LOCKED(&filter->fo_create_locks[oa->o_seq]);
3591
3592         memset(&doa, 0, sizeof(doa));
3593
3594         doa.o_valid |= OBD_MD_FLGROUP;
3595         doa.o_seq = oa->o_seq;
3596         doa.o_mode = S_IFREG;
3597
3598         if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3599                 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3600                        exp->exp_obd->obd_name, doa.o_seq);
3601                 RETURN(0);
3602         }
3603
3604         last = filter_last_id(filter, doa.o_seq);
3605
3606         skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3607
3608         CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3609                exp->exp_obd->obd_name, oa->o_id + 1, last,
3610                skip_orphan ? ", orphan objids won't be reused any more." : ".");
3611
3612         for (id = last; id > oa->o_id; id--) {
3613                 doa.o_id = id;
3614                 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3615                 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3616                         CEMERG("error destroying precreate objid "LPU64": %d\n",
3617                                id, rc);
3618
3619                 /* update last_id on disk periodically so that if we restart
3620                  * we don't need to re-scan all of the just-deleted objects. */
3621                 if ((id & 511) == 0 && !skip_orphan) {
3622                         filter_set_last_id(filter, id - 1, doa.o_seq);
3623                         filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3624                 }
3625         }
3626
3627         CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3628                exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3629
3630         if (!skip_orphan) {
3631                 filter_set_last_id(filter, id, doa.o_seq);
3632                 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3633         } else {
3634                 /*
3635                  * We have destroyed orphan objects, but don't want to reuse
3636                  * them. Therefore we don't reset last_id to the last created
3637                  * objects. Instead, we report back to the MDS the object id
3638                  * of the last orphan, so that the MDS can restart allocating
3639                  * objects from this id + 1 and thus skip the whole orphan
3640                  * object id range
3641                  */
3642                 oa->o_id = last;
3643                 rc = 0;
3644         }
3645         cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3646
3647         RETURN(rc);
3648 }
3649
3650 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3651                             obd_seq group, int *num);
3652 /* returns a negative error or a nonnegative number of files to create */
3653 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3654                                    obd_seq group, struct obd_trans_info *oti)
3655 {
3656         struct obd_device *obd = exp->exp_obd;
3657         struct filter_obd *filter = &obd->u.filter;
3658         int diff, rc;
3659         ENTRY;
3660
3661         /* delete orphans request */
3662         if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3663                 obd_id last = filter_last_id(filter, group);
3664
3665                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3666                         CERROR("%s: dropping old orphan cleanup request\n",
3667                                obd->obd_name);
3668                         RETURN(0);
3669                 }
3670                 /* This causes inflight precreates to abort and drop lock */
3671                 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3672                 cfs_mutex_lock(&filter->fo_create_locks[group]);
3673                 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3674                         CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3675                                exp->exp_obd->obd_name, group);
3676                         cfs_mutex_unlock(&filter->fo_create_locks[group]);
3677                         RETURN(0);
3678                 }
3679                 diff = oa->o_id - last;
3680                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3681                        last, diff);
3682
3683                 if (-diff > OST_MAX_PRECREATE) {
3684                         CERROR("%s: ignoring bogus orphan destroy request: "
3685                                "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3686                                oa->o_id, last);
3687                         /* FIXME: should reset precreate_next_id on MDS */
3688                         GOTO(out, rc = -EINVAL);
3689                 }
3690                 if (diff < 0) {