Whamcloud - gitweb
1c485bbae2c1d042acc5bf0785e9a13406a207b6
[fs/lustre-release.git] / lustre / obdfilter / filter.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/obdfilter/filter.c
37  *
38  * Author: Peter Braam <braam@clusterfs.com>
39  * Author: Andreas Dilger <adilger@clusterfs.com>
40  */
41
42 /*
43  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44  *            (which need to get journal_lock, may block if journal full).
45  *
46  * Invariant: Call filter_start_transno() before any journal ops to avoid the
47  *            same deadlock problem.  We can (and want) to get rid of the
48  *            transno sem in favour of the dir/inode i_mutex to avoid single
49  *            threaded operation on the OST.
50  */
51
52 #define DEBUG_SUBSYSTEM S_FILTER
53
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
56 #endif
57 #include <linux/module.h>
58 #include <linux/fs.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
65
66 #include <obd_cksum.h>
67 #include <obd_class.h>
68 #include <obd_lov.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
79
80 #include "filter_internal.h"
81
82 static struct lvfs_callback_ops filter_lvfs_ops;
83 cfs_mem_cache_t *ll_fmd_cachep;
84
85 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
86                              void *cb_data, int error)
87 {
88         struct obd_export *exp = cb_data;
89         LASSERT(exp->exp_obd == obd);
90         obd_transno_commit_cb(obd, transno, exp, error);
91         class_export_cb_put(exp);
92 }
93
94 int filter_version_get_check(struct obd_export *exp,
95                              struct obd_trans_info *oti, struct inode *inode)
96 {
97         __u64 curr_version;
98
99         if (inode == NULL || oti == NULL)
100                 RETURN(0);
101
102         curr_version = fsfilt_get_version(exp->exp_obd, inode);
103         if ((__s64)curr_version == -EOPNOTSUPP)
104                 RETURN(0);
105         /* VBR: version is checked always because costs nothing */
106         if (oti->oti_pre_version != 0 &&
107             oti->oti_pre_version != curr_version) {
108                 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
109                        oti->oti_pre_version, curr_version);
110                 cfs_spin_lock(&exp->exp_lock);
111                 exp->exp_vbr_failed = 1;
112                 cfs_spin_unlock(&exp->exp_lock);
113                 RETURN (-EOVERFLOW);
114         }
115         oti->oti_pre_version = curr_version;
116         RETURN(0);
117 }
118
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121                           struct obd_trans_info *oti, int rc, int force_sync)
122 {
123         struct obd_device_target *obt = &exp->exp_obd->u.obt;
124         struct tg_export_data *ted = &exp->exp_target_data;
125         struct lr_server_data *lsd = class_server_data(exp->exp_obd);
126         struct lsd_client_data *lcd;
127         __u64 last_rcvd;
128         loff_t off;
129         int err, log_pri = D_RPCTRACE;
130
131         /* Propagate error code. */
132         if (rc)
133                 RETURN(rc);
134
135         if (!exp->exp_obd->obd_replayable || oti == NULL)
136                 RETURN(rc);
137
138         cfs_mutex_down(&ted->ted_lcd_lock);
139         lcd = ted->ted_lcd;
140         /* if the export has already been disconnected, we have no last_rcvd slot,
141          * update server data with latest transno then */
142         if (lcd == NULL) {
143                 cfs_mutex_up(&ted->ted_lcd_lock);
144                 CWARN("commit transaction for disconnected client %s: rc %d\n",
145                       exp->exp_client_uuid.uuid, rc);
146                 err = filter_update_server_data(exp->exp_obd);
147                 RETURN(err);
148         }
149
150         /* we don't allocate new transnos for replayed requests */
151         cfs_spin_lock(&obt->obt_lut->lut_translock);
152         if (oti->oti_transno == 0) {
153                 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
154                 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
155         } else {
156                 last_rcvd = oti->oti_transno;
157                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
158                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
159         }
160         oti->oti_transno = last_rcvd;
161
162         LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
163         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
164         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
165         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
166         cfs_spin_unlock(&obt->obt_lut->lut_translock);
167
168         if (inode)
169                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
170
171         off = ted->ted_lr_off;
172         if (off <= 0) {
173                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
174                        ted->ted_lr_idx, ted->ted_lr_off);
175                 err = -EINVAL;
176         } else {
177                 class_export_cb_get(exp); /* released when the cb is called */
178                 if (!force_sync)
179                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
180                                                            last_rcvd,
181                                                            oti->oti_handle,
182                                                            filter_commit_cb,
183                                                            exp);
184
185                 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
186                                           lcd, sizeof(*lcd), &off,
187                                           force_sync | exp->exp_need_sync);
188                 if (force_sync)
189                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
190         }
191         if (err) {
192                 log_pri = D_ERROR;
193                 if (rc == 0)
194                         rc = err;
195         }
196
197         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
198                last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
199         cfs_mutex_up(&ted->ted_lcd_lock);
200         RETURN(rc);
201 }
202
203 void f_dput(struct dentry *dentry)
204 {
205         /* Can't go inside filter_ddelete because it can block */
206         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
207                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
208         LASSERT(atomic_read(&dentry->d_count) > 0);
209
210         dput(dentry);
211 }
212
213 static void init_brw_stats(struct brw_stats *brw_stats)
214 {
215         int i;
216         for (i = 0; i < BRW_LAST; i++)
217                 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
218 }
219
220 static int lprocfs_init_rw_stats(struct obd_device *obd,
221                                  struct lprocfs_stats **stats)
222 {
223         int num_stats;
224
225         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
226                                                         LPROC_FILTER_LAST - 1;
227         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
228         if (*stats == NULL)
229                 return -ENOMEM;
230
231         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
232         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
233                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
234         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
235                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
236
237         return(0);
238 }
239
240 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
241    plus the procfs overhead :( */
242 static int filter_export_stats_init(struct obd_device *obd,
243                                     struct obd_export *exp,
244                                     int reconnect,
245                                     void *client_nid)
246 {
247         int rc, newnid = 0;
248         ENTRY;
249
250         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
251                 /* Self-export gets no proc entry */
252                 RETURN(0);
253
254         rc = lprocfs_exp_setup(exp, client_nid, reconnect, &newnid);
255         if (rc) {
256                 /* Mask error for already created
257                  * /proc entries */
258                 if (rc == -EALREADY)
259                         rc = 0;
260                 RETURN(rc);
261         }
262
263         if (newnid) {
264                 struct nid_stat *tmp = exp->exp_nid_stats;
265                 LASSERT(tmp != NULL);
266
267                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
268                 if (tmp->nid_brw_stats == NULL)
269                         GOTO(clean, rc = -ENOMEM);
270
271                 init_brw_stats(tmp->nid_brw_stats);
272                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
273                                         0644, &filter_per_nid_stats_fops,
274                                         exp->exp_nid_stats);
275                 if (rc)
276                         CWARN("Error adding the brw_stats file\n");
277
278                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
279                 if (rc)
280                         GOTO(clean, rc);
281
282                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
283                                             tmp->nid_stats);
284                 if (rc)
285                         GOTO(clean, rc);
286                 rc = lprocfs_nid_ldlm_stats_init(tmp);
287                 if (rc)
288                         GOTO(clean, rc);
289         }
290
291         RETURN(0);
292  clean:
293         lprocfs_exp_cleanup(exp);
294         return rc;
295 }
296
297 /* Add client data to the FILTER.  We use a bitmap to locate a free space
298  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
299  * Otherwise, we have just read the data from the last_rcvd file and
300  * we know its offset. */
301 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
302                              int cl_idx)
303 {
304         struct obd_device_target *obt = &obd->u.obt;
305         struct tg_export_data *ted = &exp->exp_target_data;
306         struct lr_server_data *lsd = class_server_data(obd);
307         unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
308         int new_client = (cl_idx == -1);
309
310         ENTRY;
311
312         LASSERT(bitmap != NULL);
313         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
314
315         /* Self-export */
316         if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
317                 RETURN(0);
318
319         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
320          * there's no need for extra complication here
321          */
322         if (new_client) {
323                 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
324         repeat:
325                 if (cl_idx >= LR_MAX_CLIENTS) {
326                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
327                                cl_idx);
328                         RETURN(-EOVERFLOW);
329                 }
330                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
331                         cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
332                                                         cl_idx);
333                         goto repeat;
334                 }
335         } else {
336                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
337                         CERROR("FILTER client %d: bit already set in bitmap!\n",
338                                cl_idx);
339                         LBUG();
340                 }
341         }
342
343         ted->ted_lr_idx = cl_idx;
344         ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
345                           cl_idx * le16_to_cpu(lsd->lsd_client_size);
346         cfs_init_mutex(&ted->ted_lcd_lock);
347         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
348
349         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
350                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
351
352         if (new_client) {
353                 struct lvfs_run_ctxt saved;
354                 loff_t off = ted->ted_lr_off;
355                 int rc;
356                 void *handle;
357
358                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
359                        ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
360
361                 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
362                         RETURN(-ENOSPC);
363
364                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
365                 /* Transaction needed to fix bug 1403 */
366                 handle = fsfilt_start(obd,
367                                       obt->obt_rcvd_filp->f_dentry->d_inode,
368                                       FSFILT_OP_SETATTR, NULL);
369                 if (IS_ERR(handle)) {
370                         rc = PTR_ERR(handle);
371                         CERROR("unable to start transaction: rc %d\n", rc);
372                 } else {
373                         ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
374                         exp->exp_last_request_time = cfs_time_current_sec();
375                         rc = fsfilt_add_journal_cb(obd, 0, handle,
376                                                    target_client_add_cb,
377                                                    class_export_cb_get(exp));
378                         if (rc == 0) {
379                                 cfs_spin_lock(&exp->exp_lock);
380                                 exp->exp_need_sync = 1;
381                                 cfs_spin_unlock(&exp->exp_lock);
382                         }
383                         rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
384                                                  ted->ted_lcd,
385                                                  sizeof(*ted->ted_lcd),
386                                                  &off, rc /* sync if no cb */);
387                         fsfilt_commit(obd,
388                                       obt->obt_rcvd_filp->f_dentry->d_inode,
389                                       handle, 0);
390                 }
391                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
392
393                 if (rc) {
394                         CERROR("error writing %s client idx %u: rc %d\n",
395                                LAST_RCVD, ted->ted_lr_idx, rc);
396                         RETURN(rc);
397                 }
398         }
399         RETURN(0);
400 }
401
402 static int filter_client_del(struct obd_export *exp)
403 {
404         struct tg_export_data *ted = &exp->exp_target_data;
405         struct obd_device_target *obt = &exp->exp_obd->u.obt;
406         struct lvfs_run_ctxt saved;
407         int rc;
408         loff_t off;
409         ENTRY;
410
411         if (ted->ted_lcd == NULL)
412                 RETURN(0);
413
414         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
415         if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
416                 GOTO(free, 0);
417
418         LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
419
420         off = ted->ted_lr_off;
421
422         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
423                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
424
425         /* Don't clear ted_lr_idx here as it is likely also unset.  At worst
426          * we leak a client slot that will be cleaned on the next recovery. */
427         if (off <= 0) {
428                 CERROR("%s: client idx %d has med_off %lld\n",
429                        exp->exp_obd->obd_name, ted->ted_lr_idx, off);
430                 GOTO(free, rc = -EINVAL);
431         }
432
433         /* Clear the bit _after_ zeroing out the client so we don't
434            race with filter_client_add and zero out new clients.*/
435         if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
436                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
437                        ted->ted_lr_idx);
438                 LBUG();
439         }
440
441         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
442         /* Make sure the server's last_transno is up to date.
443          * This should be done before zeroing client slot so last_transno will
444          * be in server data or in client data in case of failure */
445         filter_update_server_data(exp->exp_obd);
446
447         cfs_mutex_down(&ted->ted_lcd_lock);
448         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
449         rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
450                                  ted->ted_lcd,
451                                  sizeof(*ted->ted_lcd), &off, 0);
452         cfs_mutex_up(&ted->ted_lcd_lock);
453         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
454
455         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
456                "zero out client %s at idx %u/%llu in %s, rc %d\n",
457                ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
458                LAST_RCVD, rc);
459         EXIT;
460 free:
461         return 0;
462 }
463
464 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
465 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
466                                          struct filter_mod_data *fmd)
467 {
468         LASSERT_SPIN_LOCKED(&fed->fed_lock);
469         if (--fmd->fmd_refcount == 0) {
470                 /* XXX when we have persistent reservations and the handle
471                  * is stored herein we need to drop it here. */
472                 fed->fed_mod_count--;
473                 cfs_list_del(&fmd->fmd_list);
474                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
475         }
476 }
477
478 /* drop fmd reference, free it if last ref */
479 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
480 {
481         struct filter_export_data *fed;
482
483         if (fmd == NULL)
484                 return;
485
486         fed = &exp->exp_filter_data;
487         cfs_spin_lock(&fed->fed_lock);
488         filter_fmd_put_nolock(fed, fmd); /* caller reference */
489         cfs_spin_unlock(&fed->fed_lock);
490 }
491
492 /* expire entries from the end of the list if there are too many
493  * or they are too old */
494 static void filter_fmd_expire_nolock(struct filter_obd *filter,
495                                      struct filter_export_data *fed,
496                                      struct filter_mod_data *keep)
497 {
498         struct filter_mod_data *fmd, *tmp;
499
500         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
501                 if (fmd == keep)
502                         break;
503
504                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
505                     fed->fed_mod_count < filter->fo_fmd_max_num)
506                         break;
507
508                 cfs_list_del_init(&fmd->fmd_list);
509                 filter_fmd_put_nolock(fed, fmd); /* list reference */
510         }
511 }
512
513 void filter_fmd_expire(struct obd_export *exp)
514 {
515         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
516         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
517                                  &exp->exp_filter_data, NULL);
518         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
519 }
520
521 /* find specified objid, group in export fmd list.
522  * caller must hold fed_lock and take fmd reference itself */
523 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
524                                                 struct filter_export_data *fed,
525                                                 obd_id objid, obd_seq group)
526 {
527         struct filter_mod_data *found = NULL, *fmd;
528
529         LASSERT_SPIN_LOCKED(&fed->fed_lock);
530
531         cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
532                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
533                         found = fmd;
534                         cfs_list_del(&fmd->fmd_list);
535                         cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
536                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
537                         break;
538                 }
539         }
540
541         filter_fmd_expire_nolock(filter, fed, found);
542
543         return found;
544 }
545
546 /* Find fmd based on objid and group, or return NULL if not found. */
547 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
548                                         obd_id objid, obd_seq group)
549 {
550         struct filter_mod_data *fmd;
551
552         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
553         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
554                                      &exp->exp_filter_data, objid, group);
555         if (fmd)
556                 fmd->fmd_refcount++;    /* caller reference */
557         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
558
559         return fmd;
560 }
561
562 /* Find fmd based on objid and group, or create a new one if none is found.
563  * It is possible for this function to return NULL under memory pressure,
564  * or if objid = 0 is passed (which will only cause old entries to expire).
565  * Currently this is not fatal because any fmd state is transient and
566  * may also be freed when it gets sufficiently old. */
567 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
568                                        obd_id objid, obd_seq group)
569 {
570         struct filter_export_data *fed = &exp->exp_filter_data;
571         struct filter_mod_data *found = NULL, *fmd_new = NULL;
572
573         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
574
575         cfs_spin_lock(&fed->fed_lock);
576         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
577         if (fmd_new) {
578                 if (found == NULL) {
579                         cfs_list_add_tail(&fmd_new->fmd_list,
580                                           &fed->fed_mod_list);
581                         fmd_new->fmd_id = objid;
582                         fmd_new->fmd_gr = group;
583                         fmd_new->fmd_refcount++;   /* list reference */
584                         found = fmd_new;
585                         fed->fed_mod_count++;
586                 } else {
587                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
588                 }
589         }
590         if (found) {
591                 found->fmd_refcount++;          /* caller reference */
592                 found->fmd_expire = jiffies +
593                         exp->exp_obd->u.filter.fo_fmd_max_age;
594         }
595
596         cfs_spin_unlock(&fed->fed_lock);
597
598         return found;
599 }
600
601 #ifdef DO_FMD_DROP
602 /* drop fmd list reference so it will disappear when last reference is put.
603  * This isn't so critical because it would in fact only affect the one client
604  * that is doing the unlink and at worst we have an stale entry referencing
605  * an object that should never be used again. */
606 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
607 {
608         struct filter_mod_data *found = NULL;
609
610         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
611         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
612         if (found) {
613                 cfs_list_del_init(&found->fmd_list);
614                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
615         }
616         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
617 }
618 #else
619 #define filter_fmd_drop(exp, objid, group)
620 #endif
621
622 /* remove all entries from fmd list */
623 static void filter_fmd_cleanup(struct obd_export *exp)
624 {
625         struct filter_export_data *fed = &exp->exp_filter_data;
626         struct filter_mod_data *fmd = NULL, *tmp;
627
628         cfs_spin_lock(&fed->fed_lock);
629         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
630                 cfs_list_del_init(&fmd->fmd_list);
631                 filter_fmd_put_nolock(fed, fmd);
632         }
633         cfs_spin_unlock(&fed->fed_lock);
634 }
635
636 static int filter_init_export(struct obd_export *exp)
637 {
638         int rc;
639         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
640         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
641
642         cfs_spin_lock(&exp->exp_lock);
643         exp->exp_connecting = 1;
644         cfs_spin_unlock(&exp->exp_lock);
645         rc = lut_client_alloc(exp);
646         if (rc == 0)
647                 rc = ldlm_init_export(exp);
648
649         return rc;
650 }
651
652 static int filter_free_server_data(struct obd_device_target *obt)
653 {
654         lut_fini(NULL, obt->obt_lut);
655         OBD_FREE_PTR(obt->obt_lut);
656         return 0;
657 }
658
659 /* assumes caller is already in kernel ctxt */
660 int filter_update_server_data(struct obd_device *obd)
661 {
662         struct file *filp = obd->u.obt.obt_rcvd_filp;
663         struct lr_server_data *lsd = class_server_data(obd);
664         loff_t off = 0;
665         int rc;
666         ENTRY;
667
668         CDEBUG(D_INODE, "server uuid      : %s\n", lsd->lsd_uuid);
669         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
670                le64_to_cpu(lsd->lsd_last_transno));
671         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
672                le64_to_cpu(lsd->lsd_mount_count));
673
674         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
675         if (rc)
676                 CERROR("error writing lr_server_data: rc = %d\n", rc);
677
678         RETURN(rc);
679 }
680
681 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
682                              int force_sync)
683 {
684         struct filter_obd *filter = &obd->u.filter;
685         __u64 tmp;
686         loff_t off = 0;
687         int rc;
688         ENTRY;
689
690         if (filter->fo_last_objid_files[group] == NULL) {
691                 CERROR("Object seq "LPU64" not fully setup; not updating "
692                        "last_objid\n", group);
693                 RETURN(-EINVAL);
694         }
695
696         CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
697                obd->obd_name, filter->fo_last_objids[group], group);
698
699         tmp = cpu_to_le64(filter->fo_last_objids[group]);
700         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
701                                  &tmp, sizeof(tmp), &off, force_sync);
702         if (rc)
703                 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
704                        group, rc);
705         RETURN(rc);
706 }
707 extern int ost_handle(struct ptlrpc_request *req);
708 /* assumes caller has already in kernel ctxt */
709 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
710 {
711         struct filter_obd *filter = &obd->u.filter;
712         struct lr_server_data *lsd;
713         struct lsd_client_data *lcd = NULL;
714         struct inode *inode = filp->f_dentry->d_inode;
715         unsigned long last_rcvd_size = i_size_read(inode);
716         struct lu_target *lut;
717         __u64 mount_count;
718         __u32 start_epoch;
719         int cl_idx;
720         loff_t off = 0;
721         int rc;
722
723         /* ensure padding in the struct is the correct size */
724         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
725                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
726         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
727                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
728
729         /* allocate and initialize lu_target */
730         OBD_ALLOC_PTR(lut);
731         if (lut == NULL)
732                 RETURN(-ENOMEM);
733         rc = lut_init(NULL, lut, obd, NULL);
734         if (rc)
735                 GOTO(err_lut, rc);
736         lsd = class_server_data(obd);
737         if (last_rcvd_size == 0) {
738                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
739
740                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
741                 lsd->lsd_last_transno = 0;
742                 mount_count = lsd->lsd_mount_count = 0;
743                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
744                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
745                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
746                 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
747                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
748                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
749                  * MDS first connects and assigns the OST index number. */
750                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
751                                                         OBD_INCOMPAT_OST);
752         } else {
753                 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
754                 if (rc) {
755                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
756                                LAST_RCVD, rc);
757                         GOTO(err_lut, rc);
758                 }
759                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
760                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
761                                            "using the wrong disk %s. Were the "
762                                            "/dev/ assignments rearranged?\n",
763                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
764                         GOTO(err_lut, rc = -EINVAL);
765                 }
766                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
767                 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
768                 /* COMPAT_146 */
769                 /* Assume old last_rcvd format unless I_C_LR is set */
770                 if (!(lsd->lsd_feature_incompat &
771                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
772                         lsd->lsd_last_transno = lsd->lsd_compat14;
773                 /* end COMPAT_146 */
774                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
775                  * MDS first connects and assigns the OST index number. */
776                 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
777                                                          OBD_INCOMPAT_OST);
778         }
779
780         if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
781                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
782                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
783                        ~FILTER_INCOMPAT_SUPP);
784                 GOTO(err_lut, rc = -EINVAL);
785         }
786         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
787                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
788                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
789                        ~FILTER_ROCOMPAT_SUPP);
790                 /* Do something like remount filesystem read-only */
791                 GOTO(err_lut, rc = -EINVAL);
792         }
793
794         start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
795
796         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
797                obd->obd_name, start_epoch);
798         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
799                obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
800         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
801                obd->obd_name, mount_count + 1);
802         CDEBUG(D_INODE, "%s: server data size: %u\n",
803                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
804         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
805                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
806         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
807                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
808         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
809                obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
810         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
811                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
812                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
813                 le16_to_cpu(lsd->lsd_client_size));
814
815         if (!obd->obd_replayable) {
816                 CWARN("%s: recovery support OFF\n", obd->obd_name);
817                 GOTO(out, rc = 0);
818         }
819
820         OBD_ALLOC_PTR(lcd);
821         if (!lcd)
822                 GOTO(err_client, rc = -ENOMEM);
823
824         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
825              off < last_rcvd_size; cl_idx++) {
826                 __u64 last_rcvd;
827                 struct obd_export *exp;
828                 struct filter_export_data *fed;
829
830                 /* Don't assume off is incremented properly by
831                  * fsfilt_read_record(), in case sizeof(*lcd)
832                  * isn't the same as lsd->lsd_client_size.  */
833                 off = le32_to_cpu(lsd->lsd_client_start) +
834                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
835                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
836                 if (rc) {
837                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
838                                LAST_RCVD, cl_idx, off, rc);
839                         break; /* read error shouldn't cause startup to fail */
840                 }
841
842                 if (lcd->lcd_uuid[0] == '\0') {
843                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
844                                cl_idx);
845                         continue;
846                 }
847
848                 check_lcd(obd->obd_name, cl_idx, lcd);
849
850                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
851
852                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
853                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
854                        last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
855
856                 /* These exports are cleaned up by filter_disconnect(), so they
857                  * need to be set up like real exports as filter_connect() does.
858                  */
859                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
860                 if (IS_ERR(exp)) {
861                         if (PTR_ERR(exp) == -EALREADY) {
862                                 /* export already exists, zero out this one */
863                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
864                                 continue;
865                         }
866                         OBD_FREE_PTR(lcd);
867                         GOTO(err_client, rc = PTR_ERR(exp));
868                 }
869
870                 fed = &exp->exp_filter_data;
871                 *fed->fed_ted.ted_lcd = *lcd;
872                 fed->fed_group = 0; /* will be assigned at connect */
873                 filter_export_stats_init(obd, exp, 0, NULL);
874                 rc = filter_client_add(obd, exp, cl_idx);
875                 /* can't fail for existing client */
876                 LASSERTF(rc == 0, "rc = %d\n", rc);
877
878                 /* VBR: set export last committed */
879                 exp->exp_last_committed = last_rcvd;
880                 cfs_spin_lock(&exp->exp_lock);
881                 exp->exp_connecting = 0;
882                 exp->exp_in_recovery = 0;
883                 cfs_spin_unlock(&exp->exp_lock);
884                 obd->obd_max_recoverable_clients++;
885                 class_export_put(exp);
886
887                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
888                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
889         }
890         OBD_FREE_PTR(lcd);
891
892         obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
893 out:
894         obd->u.obt.obt_mount_count = mount_count + 1;
895         lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
896
897         /* save it, so mount count and last_transno is current */
898         rc = filter_update_server_data(obd);
899         if (rc)
900                 GOTO(err_client, rc);
901
902         RETURN(0);
903
904 err_client:
905         class_disconnect_exports(obd);
906 err_lut:
907         filter_free_server_data(&obd->u.obt);
908         RETURN(rc);
909 }
910
911 static int filter_cleanup_groups(struct obd_device *obd)
912 {
913         struct filter_obd *filter = &obd->u.filter;
914         struct file *filp;
915         struct dentry *dentry;
916         int i, j;
917         ENTRY;
918
919         if (filter->fo_dentry_O_groups != NULL) {
920                 for (i = 0; i < filter->fo_group_count; i++) {
921                         dentry = filter->fo_dentry_O_groups[i];
922                         if (dentry != NULL)
923                                 f_dput(dentry);
924                 }
925                 OBD_FREE(filter->fo_dentry_O_groups,
926                          filter->fo_group_count *
927                          sizeof(*filter->fo_dentry_O_groups));
928                 filter->fo_dentry_O_groups = NULL;
929         }
930         if (filter->fo_last_objid_files != NULL) {
931                 for (i = 0; i < filter->fo_group_count; i++) {
932                         filp = filter->fo_last_objid_files[i];
933                         if (filp != NULL)
934                                 filp_close(filp, 0);
935                 }
936                 OBD_FREE(filter->fo_last_objid_files,
937                          filter->fo_group_count *
938                          sizeof(*filter->fo_last_objid_files));
939                 filter->fo_last_objid_files = NULL;
940         }
941         if (filter->fo_dentry_O_sub != NULL) {
942                 for (i = 0; i < filter->fo_group_count; i++) {
943                         for (j = 0; j < filter->fo_subdir_count; j++) {
944                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
945                                 if (dentry != NULL)
946                                         f_dput(dentry);
947                         }
948                 }
949                 OBD_FREE(filter->fo_dentry_O_sub,
950                          filter->fo_group_count *
951                          sizeof(*filter->fo_dentry_O_sub));
952                 filter->fo_dentry_O_sub = NULL;
953         }
954         if (filter->fo_last_objids != NULL) {
955                 OBD_FREE(filter->fo_last_objids,
956                          filter->fo_group_count *
957                          sizeof(*filter->fo_last_objids));
958                 filter->fo_last_objids = NULL;
959         }
960         if (filter->fo_dentry_O != NULL) {
961                 f_dput(filter->fo_dentry_O);
962                 filter->fo_dentry_O = NULL;
963         }
964         RETURN(0);
965 }
966
967 static int filter_update_last_group(struct obd_device *obd, int group)
968 {
969         struct filter_obd *filter = &obd->u.filter;
970         struct file *filp = NULL;
971         int last_group = 0, rc;
972         loff_t off = 0;
973         ENTRY;
974
975         if (group <= filter->fo_committed_group)
976                 RETURN(0);
977
978         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
979         if (IS_ERR(filp)) {
980                 rc = PTR_ERR(filp);
981                 filp = NULL;
982                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
983                 GOTO(cleanup, rc);
984         }
985
986         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
987         if (rc) {
988                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
989                 GOTO(cleanup, rc);
990         }
991
992         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
993                obd->obd_name, last_group, group);
994
995         off = 0;
996         last_group = group;
997         /* must be sync: bXXXX */
998         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
999         if (rc) {
1000                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1001                 GOTO(cleanup, rc);
1002         }
1003
1004         filter->fo_committed_group = group;
1005 cleanup:
1006         if (filp)
1007                 filp_close(filp, 0);
1008         RETURN(rc);
1009 }
1010
1011 static int filter_read_group_internal(struct obd_device *obd, int group,
1012                                       int create)
1013 {
1014         struct filter_obd *filter = &obd->u.filter;
1015         __u64 *new_objids = NULL;
1016         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1017         struct dentry **new_groups = NULL;
1018         struct file **new_files = NULL;
1019         struct dentry *dentry;
1020         struct file *filp;
1021         int old_count = filter->fo_group_count, rc, stage = 0, i;
1022         char name[25];
1023         __u64 last_objid;
1024         loff_t off = 0;
1025         int len = group + 1;
1026
1027         snprintf(name, 24, "%d", group);
1028         name[24] = '\0';
1029
1030         if (!create) {
1031                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1032                                            strlen(name));
1033                 if (IS_ERR(dentry)) {
1034                         CERROR("Cannot lookup expected object group %d: %ld\n",
1035                                group, PTR_ERR(dentry));
1036                         RETURN(PTR_ERR(dentry));
1037                 }
1038         } else {
1039                 dentry = simple_mkdir(filter->fo_dentry_O,
1040                                       obd->u.obt.obt_vfsmnt, name, 0700, 1);
1041                 if (IS_ERR(dentry)) {
1042                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1043                                PTR_ERR(dentry));
1044                         RETURN(PTR_ERR(dentry));
1045                 }
1046         }
1047         stage = 1;
1048
1049         snprintf(name, 24, "O/%d/LAST_ID", group);
1050         name[24] = '\0';
1051         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1052         if (IS_ERR(filp)) {
1053                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1054                 GOTO(cleanup, rc = PTR_ERR(filp));
1055         }
1056         stage = 2;
1057
1058         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1059         if (rc) {
1060                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1061                 GOTO(cleanup, rc);
1062         }
1063
1064         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1065                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1066                 if (tmp_subdirs == NULL)
1067                         GOTO(cleanup, rc = -ENOMEM);
1068                 stage = 3;
1069
1070                 for (i = 0; i < filter->fo_subdir_count; i++) {
1071                         char dir[20];
1072                         snprintf(dir, sizeof(dir), "d%u", i);
1073
1074                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1075                                                               obd->u.obt.obt_vfsmnt,
1076                                                               dir, 0700, 1);
1077                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1078                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1079                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1080                                        group, dir, rc);
1081                                 GOTO(cleanup, rc);
1082                         }
1083
1084                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1085                                tmp_subdirs->dentry[i]);
1086                 }
1087         }
1088
1089         /* 'group' is an index; we need an array of length 'group + 1' */
1090         if (group + 1 > old_count) {
1091                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1092                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1093                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1094                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1095                 stage = 4;
1096                 if (new_objids == NULL || new_subdirs == NULL ||
1097                     new_groups == NULL || new_files == NULL)
1098                         GOTO(cleanup, rc = -ENOMEM);
1099
1100                 if (old_count) {
1101                         memcpy(new_objids, filter->fo_last_objids,
1102                                old_count * sizeof(*new_objids));
1103                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1104                                old_count * sizeof(*new_subdirs));
1105                         memcpy(new_groups, filter->fo_dentry_O_groups,
1106                                old_count * sizeof(*new_groups));
1107                         memcpy(new_files, filter->fo_last_objid_files,
1108                                old_count * sizeof(*new_files));
1109
1110                         OBD_FREE(filter->fo_last_objids,
1111                                  old_count * sizeof(*new_objids));
1112                         OBD_FREE(filter->fo_dentry_O_sub,
1113                                  old_count * sizeof(*new_subdirs));
1114                         OBD_FREE(filter->fo_dentry_O_groups,
1115                                  old_count * sizeof(*new_groups));
1116                         OBD_FREE(filter->fo_last_objid_files,
1117                                  old_count * sizeof(*new_files));
1118                 }
1119                 filter->fo_last_objids = new_objids;
1120                 filter->fo_dentry_O_sub = new_subdirs;
1121                 filter->fo_dentry_O_groups = new_groups;
1122                 filter->fo_last_objid_files = new_files;
1123                 filter->fo_group_count = len;
1124         }
1125
1126         filter->fo_dentry_O_groups[group] = dentry;
1127         filter->fo_last_objid_files[group] = filp;
1128         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1129                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1130                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1131         }
1132
1133         filter_update_last_group(obd, group);
1134
1135         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1136                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1137                 rc = filter_update_last_objid(obd, group, 1);
1138                 RETURN(rc);
1139         }
1140
1141         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1142         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1143                obd->obd_name, group, last_objid);
1144         RETURN(0);
1145  cleanup:
1146         switch (stage) {
1147         case 4:
1148                 if (new_objids != NULL)
1149                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1150                 if (new_subdirs != NULL)
1151                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1152                 if (new_groups != NULL)
1153                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1154                 if (new_files != NULL)
1155                         OBD_FREE(new_files, len * sizeof(*new_files));
1156         case 3:
1157                 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1158                         for (i = 0; i < filter->fo_subdir_count; i++) {
1159                                 if (tmp_subdirs->dentry[i] != NULL)
1160                                         dput(tmp_subdirs->dentry[i]);
1161                         }
1162                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1163                 }
1164         case 2:
1165                 filp_close(filp, 0);
1166         case 1:
1167                 dput(dentry);
1168         }
1169         RETURN(rc);
1170 }
1171
1172 static int filter_read_groups(struct obd_device *obd, int last_group,
1173                               int create)
1174 {
1175         struct filter_obd *filter = &obd->u.filter;
1176         int old_count, group, rc = 0;
1177
1178         cfs_down(&filter->fo_init_lock);
1179         old_count = filter->fo_group_count;
1180         for (group = old_count; group <= last_group; group++) {
1181                 rc = filter_read_group_internal(obd, group, create);
1182                 if (rc != 0)
1183                         break;
1184         }
1185         cfs_up(&filter->fo_init_lock);
1186         return rc;
1187 }
1188
1189 /* FIXME: object groups */
1190 static int filter_prep_groups(struct obd_device *obd)
1191 {
1192         struct filter_obd *filter = &obd->u.filter;
1193         struct dentry *O_dentry;
1194         struct file *filp;
1195         int    last_group, rc = 0, cleanup_phase = 0;
1196         loff_t off = 0;
1197         ENTRY;
1198
1199         O_dentry = simple_mkdir(current->fs->pwd, obd->u.obt.obt_vfsmnt,
1200                                 "O", 0700, 1);
1201         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1202         if (IS_ERR(O_dentry)) {
1203                 rc = PTR_ERR(O_dentry);
1204                 CERROR("cannot open/create O: rc = %d\n", rc);
1205                 GOTO(cleanup, rc);
1206         }
1207         filter->fo_dentry_O = O_dentry;
1208         cleanup_phase = 1; /* O_dentry */
1209
1210         /* we have to initialize all groups before first connections from
1211          * clients because they may send create/destroy for any group -bzzz */
1212         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1213         if (IS_ERR(filp)) {
1214                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1215                 GOTO(cleanup, rc = PTR_ERR(filp));
1216         }
1217         cleanup_phase = 2; /* filp */
1218
1219         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1220         if (rc) {
1221                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1222                 GOTO(cleanup, rc);
1223         }
1224
1225         if (off == 0)
1226                 last_group = FID_SEQ_OST_MDT0;
1227
1228         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1229               FID_SEQ_OST_MDT0, last_group);
1230         filter->fo_committed_group = last_group;
1231         rc = filter_read_groups(obd, last_group, 1);
1232         if (rc)
1233                 GOTO(cleanup, rc);
1234
1235         filp_close(filp, 0);
1236         RETURN(0);
1237
1238  cleanup:
1239         switch (cleanup_phase) {
1240         case 2:
1241                 filp_close(filp, 0);
1242         case 1:
1243                 filter_cleanup_groups(obd);
1244                 f_dput(filter->fo_dentry_O);
1245                 filter->fo_dentry_O = NULL;
1246         default:
1247                 break;
1248         }
1249         return rc;
1250
1251 }
1252
1253 /* setup the object store with correct subdirectories */
1254 static int filter_prep(struct obd_device *obd)
1255 {
1256         struct lvfs_run_ctxt saved;
1257         struct filter_obd *filter = &obd->u.filter;
1258         struct file *file;
1259         struct inode *inode;
1260         int rc = 0;
1261         ENTRY;
1262
1263         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1264         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1265         if (!file || IS_ERR(file)) {
1266                 rc = PTR_ERR(file);
1267                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1268                        LAST_RCVD, rc);
1269                 GOTO(out, rc);
1270         }
1271         obd->u.obt.obt_rcvd_filp = file;
1272         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1273                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1274                        file->f_dentry->d_inode->i_mode);
1275                 GOTO(err_filp, rc = -ENOENT);
1276         }
1277
1278         inode = file->f_dentry->d_parent->d_inode;
1279         /* We use i_op->unlink directly in filter_vfs_unlink() */
1280         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1281                 CERROR("%s: filesystem does not support create/unlink ops\n",
1282                        obd->obd_name);
1283                 GOTO(err_filp, rc = -EOPNOTSUPP);
1284         }
1285
1286         rc = filter_init_server_data(obd, file);
1287         if (rc) {
1288                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1289                 GOTO(err_filp, rc);
1290         }
1291         LASSERT(obd->u.obt.obt_lut);
1292         target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1293
1294         /* open and create health check io file*/
1295         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1296         if (IS_ERR(file)) {
1297                 rc = PTR_ERR(file);
1298                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1299                        HEALTH_CHECK, rc);
1300                 GOTO(err_server_data, rc);
1301         }
1302         filter->fo_obt.obt_health_check_filp = file;
1303         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1304                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1305                        file->f_dentry->d_inode->i_mode);
1306                 GOTO(err_health_check, rc = -ENOENT);
1307         }
1308         rc = lvfs_check_io_health(obd, file);
1309         if (rc)
1310                 GOTO(err_health_check, rc);
1311
1312         rc = filter_prep_groups(obd);
1313         if (rc)
1314                 GOTO(err_health_check, rc);
1315 out:
1316         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1317
1318         return(rc);
1319
1320 err_health_check:
1321         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1322                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1323         filter->fo_obt.obt_health_check_filp = NULL;
1324 err_server_data:
1325         target_recovery_fini(obd);
1326         filter_free_server_data(&obd->u.obt);
1327 err_filp:
1328         if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1329                 CERROR("can't close %s after error\n", LAST_RCVD);
1330         obd->u.obt.obt_rcvd_filp = NULL;
1331         goto out;
1332 }
1333
1334 /* cleanup the filter: write last used object id to status file */
1335 static void filter_post(struct obd_device *obd)
1336 {
1337         struct lvfs_run_ctxt saved;
1338         struct filter_obd *filter = &obd->u.filter;
1339         int rc, i;
1340
1341         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1342          * best to start a transaction with h_sync, because we removed this
1343          * from lastobjid */
1344
1345         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1346         rc = filter_update_server_data(obd);
1347         if (rc)
1348                 CERROR("error writing server data: rc = %d\n", rc);
1349
1350         for (i = 0; i < filter->fo_group_count; i++) {
1351                 rc = filter_update_last_objid(obd, i,
1352                                 (i == filter->fo_group_count - 1));
1353                 if (rc)
1354                         CERROR("error writing group %d lastobjid: rc = %d\n",
1355                                i, rc);
1356         }
1357
1358         rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1359         obd->u.obt.obt_rcvd_filp = NULL;
1360         if (rc)
1361                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1362
1363         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1364         filter->fo_obt.obt_health_check_filp = NULL;
1365         if (rc)
1366                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1367
1368         filter_cleanup_groups(obd);
1369         filter_free_server_data(&obd->u.obt);
1370         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1371
1372         filter_free_capa_keys(filter);
1373         cleanup_capa_hash(filter->fo_capa_hash);
1374 }
1375
1376 static void filter_set_last_id(struct filter_obd *filter,
1377                                obd_id id, obd_seq group)
1378 {
1379         LASSERT(group <= filter->fo_group_count);
1380
1381         cfs_spin_lock(&filter->fo_objidlock);
1382         filter->fo_last_objids[group] = id;
1383         cfs_spin_unlock(&filter->fo_objidlock);
1384 }
1385
1386 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1387 {
1388         obd_id id;
1389         LASSERT(group <= filter->fo_group_count);
1390         LASSERT(filter->fo_last_objids != NULL);
1391
1392         /* FIXME: object groups */
1393         cfs_spin_lock(&filter->fo_objidlock);
1394         id = filter->fo_last_objids[group];
1395         cfs_spin_unlock(&filter->fo_objidlock);
1396         return id;
1397 }
1398
1399 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1400 {
1401         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1402         return 0;
1403 }
1404
1405 /* We never dget the object parent, so DON'T dput it either */
1406 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1407 {
1408         struct filter_obd *filter = &obd->u.filter;
1409         struct filter_subdirs *subdirs;
1410         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1411
1412         if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1413                 return filter->fo_dentry_O_groups[group];
1414
1415         subdirs = &filter->fo_dentry_O_sub[group];
1416         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1417 }
1418
1419 /* We never dget the object parent, so DON'T dput it either */
1420 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1421                                   obd_id objid)
1422 {
1423         unsigned long now = jiffies;
1424         struct dentry *dparent = filter_parent(obd, group, objid);
1425         int rc;
1426
1427         if (IS_ERR(dparent))
1428                 return dparent;
1429         if (dparent == NULL)
1430                 return ERR_PTR(-ENOENT);
1431
1432         rc = filter_lock_dentry(obd, dparent);
1433         fsfilt_check_slow(obd, now, "parent lock");
1434         return rc ? ERR_PTR(rc) : dparent;
1435 }
1436
1437 /* We never dget the object parent, so DON'T dput it either */
1438 static void filter_parent_unlock(struct dentry *dparent)
1439 {
1440         UNLOCK_INODE_MUTEX(dparent->d_inode);
1441 }
1442
1443 /* How to get files, dentries, inodes from object id's.
1444  *
1445  * If dir_dentry is passed, the caller has already locked the parent
1446  * appropriately for this operation (normally a write lock).  If
1447  * dir_dentry is NULL, we do a read lock while we do the lookup to
1448  * avoid races with create/destroy and such changing the directory
1449  * internal to the filesystem code. */
1450 struct dentry *filter_fid2dentry(struct obd_device *obd,
1451                                  struct dentry *dir_dentry,
1452                                  obd_seq group, obd_id id)
1453 {
1454         struct dentry *dparent = dir_dentry;
1455         struct dentry *dchild;
1456         char name[32];
1457         int len;
1458         ENTRY;
1459
1460         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1461             obd->u.filter.fo_destroys_in_progress == 0) {
1462                 /* don't fail lookups for orphan recovery, it causes
1463                  * later LBUGs when objects still exist during precreate */
1464                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1465                 RETURN(ERR_PTR(-ENOENT));
1466         }
1467         if (id == 0) {
1468                 CERROR("fatal: invalid object id 0\n");
1469                 RETURN(ERR_PTR(-ESTALE));
1470         }
1471
1472         len = sprintf(name, LPU64, id);
1473         if (dir_dentry == NULL) {
1474                 dparent = filter_parent_lock(obd, group, id);
1475                 if (IS_ERR(dparent)) {
1476                         CERROR("%s: error getting object "POSTID
1477                                " parent: rc %ld\n", obd->obd_name,
1478                                id, group, PTR_ERR(dparent));
1479                         RETURN(dparent);
1480                 }
1481         }
1482         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1483                dparent->d_name.len, dparent->d_name.name, name);
1484         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1485         if (dir_dentry == NULL)
1486                 filter_parent_unlock(dparent);
1487         if (IS_ERR(dchild)) {
1488                 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1489                        obd->obd_name, id, group, PTR_ERR(dchild));
1490                 RETURN(dchild);
1491         }
1492
1493         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1494                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1495                        obd->obd_name, id, dchild->d_inode->i_ino);
1496                 f_dput(dchild);
1497                 RETURN(ERR_PTR(-ENOENT));
1498         }
1499
1500         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1501                name, dchild, atomic_read(&dchild->d_count));
1502
1503         LASSERT(atomic_read(&dchild->d_count) > 0);
1504
1505         RETURN(dchild);
1506 }
1507
1508 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1509                                   obd_id group, struct lustre_handle *lockh)
1510 {
1511         int flags = LDLM_AST_DISCARD_DATA, rc;
1512         struct ldlm_res_id res_id;
1513         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1514         ENTRY;
1515
1516         osc_build_res_name(objid, group, &res_id);
1517         /* Tell the clients that the object is gone now and that they should
1518          * throw away any cached pages. */
1519         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1520                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1521                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1522                                     lockh);
1523         if (rc != ELDLM_OK)
1524                 lockh->cookie = 0;
1525         RETURN(rc);
1526 }
1527
1528 static void filter_fini_destroy(struct obd_device *obd,
1529                                 struct lustre_handle *lockh)
1530 {
1531         if (lustre_handle_is_used(lockh))
1532                 ldlm_lock_decref(lockh, LCK_PW);
1533 }
1534
1535 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1536  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1537  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1538  * reverse.  Caller must take i_sem before starting the transaction and we
1539  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1540 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1541                       struct vfsmount *mnt)
1542 {
1543         int rc;
1544         ENTRY;
1545
1546         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1547          * itself we already hold dir->i_mutex for child create/unlink ops */
1548         LASSERT(dentry->d_inode != NULL);
1549         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1550         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1551
1552
1553         /* may_delete() */
1554         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1555                 GOTO(out, rc = -ENOENT);
1556
1557         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1558         if (rc)
1559                 GOTO(out, rc);
1560
1561         if (IS_APPEND(dir))
1562                 GOTO(out, rc = -EPERM);
1563
1564         /* check_sticky() */
1565         if ((dentry->d_inode->i_uid != current->fsuid &&
1566              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1567             IS_IMMUTABLE(dentry->d_inode))
1568                 GOTO(out, rc = -EPERM);
1569
1570         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1571          *       that was done because of journal_start (which is already done
1572          *       here) or some other ordering issue. */
1573         DQUOT_INIT(dir);
1574
1575         rc = ll_security_inode_unlink(dir, dentry, mnt);
1576         if (rc)
1577                 GOTO(out, rc);
1578
1579         rc = dir->i_op->unlink(dir, dentry);
1580 out:
1581         /* need to drop i_mutex before we lose inode reference */
1582         UNLOCK_INODE_MUTEX(dentry->d_inode);
1583         if (rc == 0)
1584                 d_delete(dentry);
1585
1586         RETURN(rc);
1587 }
1588
1589 /* Caller must hold LCK_PW on parent and push us into kernel context.
1590  * Caller must hold child i_mutex, we drop it always.
1591  * Caller is also required to ensure that dchild->d_inode exists. */
1592 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1593                                    obd_seq group, struct dentry *dparent,
1594                                    struct dentry *dchild)
1595 {
1596         struct inode *inode = dchild->d_inode;
1597         int rc;
1598
1599         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1600                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1601                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1602                        (unsigned long)inode->i_nlink,
1603                        atomic_read(&inode->i_count));
1604         }
1605
1606         rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1607         if (rc)
1608                 CERROR("error unlinking objid %.*s: rc %d\n",
1609                        dchild->d_name.len, dchild->d_name.name, rc);
1610         return(rc);
1611 }
1612
1613 struct filter_intent_args {
1614         struct ldlm_lock **victim;
1615         __u64 size;
1616         int *liblustre;
1617 };
1618
1619 static enum interval_iter filter_intent_cb(struct interval_node *n,
1620                                            void *args)
1621 {
1622         struct ldlm_interval *node = (struct ldlm_interval *)n;
1623         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1624         __u64 size = arg->size;
1625         struct ldlm_lock **v = arg->victim;
1626         struct ldlm_lock *lck;
1627
1628         /* If the interval is lower than the current file size,
1629          * just break. */
1630         if (interval_high(n) <= size)
1631                 return INTERVAL_ITER_STOP;
1632
1633         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1634                 /* Don't send glimpse ASTs to liblustre clients.
1635                  * They aren't listening for them, and they do
1636                  * entirely synchronous I/O anyways. */
1637                 if (lck->l_export == NULL ||
1638                     lck->l_export->exp_libclient == 1)
1639                         continue;
1640
1641                 if (*arg->liblustre)
1642                         *arg->liblustre = 0;
1643
1644                 if (*v == NULL) {
1645                         *v = LDLM_LOCK_GET(lck);
1646                 } else if ((*v)->l_policy_data.l_extent.start <
1647                            lck->l_policy_data.l_extent.start) {
1648                         LDLM_LOCK_RELEASE(*v);
1649                         *v = LDLM_LOCK_GET(lck);
1650                 }
1651
1652                 /* the same policy group - every lock has the
1653                  * same extent, so needn't do it any more */
1654                 break;
1655         }
1656
1657         return INTERVAL_ITER_CONT;
1658 }
1659
1660 static int filter_intent_policy(struct ldlm_namespace *ns,
1661                                 struct ldlm_lock **lockp, void *req_cookie,
1662                                 ldlm_mode_t mode, int flags, void *data)
1663 {
1664         CFS_LIST_HEAD(rpc_list);
1665         struct ptlrpc_request *req = req_cookie;
1666         struct ldlm_lock *lock = *lockp, *l = NULL;
1667         struct ldlm_resource *res = lock->l_resource;
1668         ldlm_processing_policy policy;
1669         struct ost_lvb *res_lvb, *reply_lvb;
1670         struct ldlm_reply *rep;
1671         ldlm_error_t err;
1672         int idx, rc, tmpflags = 0, only_liblustre = 1;
1673         struct ldlm_interval_tree *tree;
1674         struct filter_intent_args arg;
1675         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1676                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1677                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1678         ENTRY;
1679
1680         policy = ldlm_get_processing_policy(res);
1681         LASSERT(policy != NULL);
1682         LASSERT(req != NULL);
1683
1684         rc = lustre_pack_reply(req, 3, repsize, NULL);
1685         if (rc)
1686                 RETURN(req->rq_status = rc);
1687
1688         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1689         LASSERT(rep != NULL);
1690
1691         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1692                                    sizeof(*reply_lvb));
1693         LASSERT(reply_lvb != NULL);
1694
1695         //fixup_handle_for_resent_req(req, lock, &lockh);
1696
1697         /* Call the extent policy function to see if our request can be
1698          * granted, or is blocked.
1699          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1700          * lock, and should not be granted if the lock will be blocked.
1701          */
1702
1703         LASSERT(ns == res->lr_namespace);
1704         lock_res(res);
1705         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1706         check_res_locked(res);
1707
1708         /* FIXME: we should change the policy function slightly, to not make
1709          * this list at all, since we just turn around and free it */
1710         while (!cfs_list_empty(&rpc_list)) {
1711                 struct ldlm_lock *wlock =
1712                         cfs_list_entry(rpc_list.next, struct ldlm_lock,
1713                                        l_cp_ast);
1714                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1715                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1716                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1717                 cfs_list_del_init(&wlock->l_cp_ast);
1718                 LDLM_LOCK_RELEASE(wlock);
1719         }
1720
1721         /* The lock met with no resistance; we're finished. */
1722         if (rc == LDLM_ITER_CONTINUE) {
1723                 /* do not grant locks to the liblustre clients: they cannot
1724                  * handle ASTs robustly.  We need to do this while still
1725                  * holding ns_lock to avoid the lock remaining on the res_link
1726                  * list (and potentially being added to l_pending_list by an
1727                  * AST) when we are going to drop this lock ASAP. */
1728                 if (lock->l_export->exp_libclient ||
1729                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1730                         ldlm_resource_unlink_lock(lock);
1731                         err = ELDLM_LOCK_ABORTED;
1732                 } else {
1733                         err = ELDLM_LOCK_REPLACED;
1734                 }
1735                 unlock_res(res);
1736                 RETURN(err);
1737         }
1738
1739         /* Do not grant any lock, but instead send GL callbacks.  The extent
1740          * policy nicely created a list of all PW locks for us.  We will choose
1741          * the highest of those which are larger than the size in the LVB, if
1742          * any, and perform a glimpse callback. */
1743         res_lvb = res->lr_lvb_data;
1744         LASSERT(res_lvb != NULL);
1745         *reply_lvb = *res_lvb;
1746
1747         /*
1748          * ->ns_lock guarantees that no new locks are granted, and,
1749          * therefore, that res->lr_lvb_data cannot increase beyond the
1750          * end of already granted lock. As a result, it is safe to
1751          * check against "stale" reply_lvb->lvb_size value without
1752          * res->lr_lvb_sem.
1753          */
1754         arg.size = reply_lvb->lvb_size;
1755         arg.victim = &l;
1756         arg.liblustre = &only_liblustre;
1757         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1758                 tree = &res->lr_itree[idx];
1759                 if (tree->lit_mode == LCK_PR)
1760                         continue;
1761
1762                 interval_iterate_reverse(tree->lit_root,
1763                                          filter_intent_cb, &arg);
1764         }
1765         unlock_res(res);
1766
1767         /* There were no PW locks beyond the size in the LVB; finished. */
1768         if (l == NULL) {
1769                 if (only_liblustre) {
1770                         /* If we discovered a liblustre client with a PW lock,
1771                          * however, the LVB may be out of date!  The LVB is
1772                          * updated only on glimpse (which we don't do for
1773                          * liblustre clients) and cancel (which the client
1774                          * obviously has not yet done).  So if it has written
1775                          * data but kept the lock, the LVB is stale and needs
1776                          * to be updated from disk.
1777                          *
1778                          * Of course, this will all disappear when we switch to
1779                          * taking liblustre locks on the OST. */
1780                         ldlm_res_lvbo_update(res, NULL, 1);
1781                 }
1782                 RETURN(ELDLM_LOCK_ABORTED);
1783         }
1784
1785         /*
1786          * This check is for lock taken in filter_prepare_destroy() that does
1787          * not have l_glimpse_ast set. So the logic is: if there is a lock
1788          * with no l_glimpse_ast set, this object is being destroyed already.
1789          *
1790          * Hence, if you are grabbing DLM locks on the server, always set
1791          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1792          */
1793         if (l->l_glimpse_ast == NULL) {
1794                 /* We are racing with unlink(); just return -ENOENT */
1795                 rep->lock_policy_res1 = -ENOENT;
1796                 goto out;
1797         }
1798
1799         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1800         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1801         /* Update the LVB from disk if the AST failed (this is a legal race) */
1802         /*
1803          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1804          * sending ast is not handled. This can result in lost client writes.
1805          */
1806         if (rc != 0)
1807                 ldlm_res_lvbo_update(res, NULL, 1);
1808
1809         lock_res(res);
1810         *reply_lvb = *res_lvb;
1811         unlock_res(res);
1812
1813  out:
1814         LDLM_LOCK_RELEASE(l);
1815
1816         RETURN(ELDLM_LOCK_ABORTED);
1817 }
1818
1819 /*
1820  * per-obd_device iobuf pool.
1821  *
1822  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1823  * allocations in write-path has to be minimized (see bug 5137).
1824  *
1825  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1826  * OST threads (see ost_thread_{init,done}()).
1827  *
1828  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1829  * at the OST layer there are only (potentially) multiple obd_device of type
1830  * unknown at the time of OST thread creation.
1831  *
1832  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1833  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1834  * it's very own iobuf.
1835  *
1836  * Functions below
1837  *
1838  *     filter_kiobuf_pool_init()
1839  *
1840  *     filter_kiobuf_pool_done()
1841  *
1842  *     filter_iobuf_get()
1843  *
1844  * operate on this array. They are "generic" in a sense that they don't depend
1845  * on actual type of iobuf's (the latter depending on Linux kernel version).
1846  */
1847
1848 /*
1849  * destroy pool created by filter_iobuf_pool_init
1850  */
1851 static void filter_iobuf_pool_done(struct filter_obd *filter)
1852 {
1853         struct filter_iobuf **pool;
1854         int i;
1855
1856         ENTRY;
1857
1858         pool = filter->fo_iobuf_pool;
1859         if (pool != NULL) {
1860                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1861                         if (pool[i] != NULL)
1862                                 filter_free_iobuf(pool[i]);
1863                 }
1864                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1865                 filter->fo_iobuf_pool = NULL;
1866         }
1867         EXIT;
1868 }
1869
1870 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1871 {
1872         struct filter_obd       *filter = &obd->u.filter;
1873         struct sptlrpc_rule_set  tmp_rset;
1874         int                      rc;
1875
1876         sptlrpc_rule_set_init(&tmp_rset);
1877         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1878         if (rc) {
1879                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1880                        obd->obd_name, rc);
1881                 return rc;
1882         }
1883
1884         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1885
1886         cfs_write_lock(&filter->fo_sptlrpc_lock);
1887         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1888         filter->fo_sptlrpc_rset = tmp_rset;
1889         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1890
1891         return 0;
1892 }
1893
1894 /*
1895  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1896  */
1897 static int filter_iobuf_pool_init(struct filter_obd *filter)
1898 {
1899         void **pool;
1900
1901         ENTRY;
1902
1903
1904         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1905                       GFP_KERNEL);
1906         if (filter->fo_iobuf_pool == NULL)
1907                 RETURN(-ENOMEM);
1908
1909         filter->fo_iobuf_count = OSS_THREADS_MAX;
1910
1911         RETURN(0);
1912 }
1913
1914 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1915  * many threads there will be so we allocate a large empty array and only
1916  * fill in those slots that are actually in use.
1917  * If we haven't allocated a pool entry for this thread before, do so now. */
1918 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1919 {
1920         int thread_id                    = (oti && oti->oti_thread) ?
1921                                            oti->oti_thread->t_id : -1;
1922         struct filter_iobuf  *pool       = NULL;
1923         struct filter_iobuf **pool_place = NULL;
1924
1925         if (thread_id >= 0) {
1926                 LASSERT(thread_id < filter->fo_iobuf_count);
1927                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1928         }
1929
1930         if (unlikely(pool == NULL)) {
1931                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1932                                           PTLRPC_MAX_BRW_PAGES);
1933                 if (pool_place != NULL)
1934                         *pool_place = pool;
1935         }
1936
1937         return pool;
1938 }
1939
1940 /* mount the file system (secretly).  lustre_cfg parameters are:
1941  * 1 = device
1942  * 2 = fstype
1943  * 3 = flags: failover=f, failout=n
1944  * 4 = mount options
1945  */
1946 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1947                         void *option)
1948 {
1949         struct filter_obd *filter = &obd->u.filter;
1950         struct vfsmount *mnt;
1951         struct lustre_mount_info *lmi;
1952         struct obd_uuid uuid;
1953         __u8 *uuid_ptr;
1954         char *str, *label;
1955         char ns_name[48];
1956         request_queue_t *q;
1957         int rc, i;
1958         ENTRY;
1959
1960         if (lcfg->lcfg_bufcount < 3 ||
1961             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1962             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1963                 RETURN(-EINVAL);
1964
1965         lmi = server_get_mount(obd->obd_name);
1966         if (lmi) {
1967                 /* We already mounted in lustre_fill_super.
1968                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1969                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1970                 mnt = lmi->lmi_mnt;
1971                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1972
1973                 /* gets recovery timeouts from mount data */
1974                 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1975                         obd->obd_recovery_timeout =
1976                                 lsi->lsi_lmd->lmd_recovery_time_soft;
1977                 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1978                         obd->obd_recovery_time_hard =
1979                                 lsi->lsi_lmd->lmd_recovery_time_hard;
1980         } else {
1981                 /* old path - used by lctl */
1982                 CERROR("Using old MDS mount method\n");
1983                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1984                                     MS_NOATIME|MS_NODIRATIME,
1985                                     lustre_cfg_string(lcfg, 1), option);
1986                 if (IS_ERR(mnt)) {
1987                         rc = PTR_ERR(mnt);
1988                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1989                                            lustre_cfg_string(lcfg, 1), rc);
1990                         RETURN(rc);
1991                 }
1992
1993                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1994         }
1995         if (IS_ERR(obd->obd_fsops))
1996                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1997
1998         rc = filter_iobuf_pool_init(filter);
1999         if (rc != 0)
2000                 GOTO(err_ops, rc);
2001
2002         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2003                 CERROR("%s: Underlying device is marked as read-only. "
2004                        "Setup failed\n", obd->obd_name);
2005                 GOTO(err_ops, rc = -EROFS);
2006         }
2007
2008         /* failover is the default */
2009         obd->obd_replayable = 1;
2010
2011         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2012                 str = lustre_cfg_string(lcfg, 3);
2013                 if (strchr(str, 'n')) {
2014                         CWARN("%s: recovery disabled\n", obd->obd_name);
2015                         obd->obd_replayable = 0;
2016                 }
2017         }
2018
2019         obd->u.obt.obt_vfsmnt = mnt;
2020         obd->u.obt.obt_sb = mnt->mnt_sb;
2021         obd->u.obt.obt_magic = OBT_MAGIC;
2022         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2023         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2024
2025         fsfilt_setup(obd, obd->u.obt.obt_sb);
2026
2027         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2028         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2029         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2030         obd->obd_lvfs_ctxt.fs = get_ds();
2031         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2032
2033         cfs_init_mutex(&filter->fo_init_lock);
2034         filter->fo_committed_group = 0;
2035         filter->fo_destroys_in_progress = 0;
2036         for (i = 0; i < 32; i++)
2037                 cfs_sema_init(&filter->fo_create_locks[i], 1);
2038
2039         cfs_spin_lock_init(&filter->fo_objidlock);
2040         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2041         cfs_sema_init(&filter->fo_alloc_lock, 1);
2042         init_brw_stats(&filter->fo_filter_stats);
2043         cfs_spin_lock_init(&filter->fo_flags_lock);
2044         filter->fo_read_cache = 1; /* enable read-only cache by default */
2045         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2046         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2047         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2048         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2049         filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2050         filter_slc_set(filter); /* initialize sync on lock cancel */
2051
2052         rc = filter_prep(obd);
2053         if (rc)
2054                 GOTO(err_ops, rc);
2055
2056         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2057         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2058
2059         filter->fo_fl_oss_capa = 1;
2060
2061         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2062         filter->fo_capa_hash = init_capa_hash();
2063         if (filter->fo_capa_hash == NULL)
2064                 GOTO(err_post, rc = -ENOMEM);
2065
2066         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2067         obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2068                                                 LDLM_NAMESPACE_GREEDY);
2069         if (obd->obd_namespace == NULL)
2070                 GOTO(err_post, rc = -ENOMEM);
2071         obd->obd_namespace->ns_lvbp = obd;
2072         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2073         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2074
2075         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2076                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2077
2078         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2079         if (rc) {
2080                 CERROR("failed to setup llogging subsystems\n");
2081                 GOTO(err_post, rc);
2082         }
2083
2084         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2085         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2086         /* do this after llog being initialized */
2087         filter_adapt_sptlrpc_conf(obd, 1);
2088
2089         rc = lquota_setup(filter_quota_interface_ref, obd);
2090         if (rc)
2091                 GOTO(err_post, rc);
2092
2093         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2094         if (q->max_sectors < q->max_hw_sectors &&
2095             q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2096                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2097                               "for larger I/O requests: max_sectors = %u "
2098                               "could be up to max_hw_sectors=%u\n",
2099                               obd->obd_name, mnt->mnt_sb->s_id,
2100                               q->max_sectors, q->max_hw_sectors);
2101
2102         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2103         if (uuid_ptr != NULL) {
2104                 class_uuid_unparse(uuid_ptr, &uuid);
2105                 str = uuid.uuid;
2106         } else {
2107                 str = "no UUID";
2108         }
2109
2110         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2111         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2112                       obd->obd_name, label ?: str, lmi ? "on " : "",
2113                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2114                       obd->obd_replayable ? "enabled" : "disabled");
2115
2116         if (obd->obd_recovering)
2117                 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2118                               "or until %d client%s reconnect%s\n",
2119                               obd->obd_name,
2120                               obd->obd_recovery_timeout / 60,
2121                               obd->obd_recovery_timeout % 60,
2122                               obd->obd_max_recoverable_clients,
2123                               (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2124                               (obd->obd_max_recoverable_clients == 1) ? "s": "");
2125
2126
2127         RETURN(0);
2128
2129 err_post:
2130         filter_post(obd);
2131 err_ops:
2132         fsfilt_put_ops(obd->obd_fsops);
2133         filter_iobuf_pool_done(filter);
2134 err_mntput:
2135         server_put_mount(obd->obd_name, mnt);
2136         obd->u.obt.obt_sb = 0;
2137         return rc;
2138 }
2139
2140 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2141 {
2142         struct lprocfs_static_vars lvars;
2143         unsigned long addr;
2144         struct page *page;
2145         int rc;
2146
2147         CLASSERT(offsetof(struct obd_device, u.obt) ==
2148                  offsetof(struct obd_device, u.filter.fo_obt));
2149
2150         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2151                 RETURN(-EINVAL);
2152
2153         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2154         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2155         if (!page)
2156                 RETURN(-ENOMEM);
2157         addr = (unsigned long)cfs_page_address(page);
2158         clear_page((void *)addr);
2159
2160         /* lprocfs must be setup before the filter so state can be safely added
2161          * to /proc incrementally as the filter is setup */
2162         lprocfs_filter_init_vars(&lvars);
2163         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2164             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2165                 /* Init obdfilter private stats here */
2166                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2167                                      LPROCFS_CNTR_AVGMINMAX,
2168                                      "read_bytes", "bytes");
2169                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2170                                      LPROCFS_CNTR_AVGMINMAX,
2171                                      "write_bytes", "bytes");
2172                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2173                                      LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2174                                      "get_page", "usec");
2175                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2176                                      LPROCFS_CNTR_AVGMINMAX,
2177                                      "get_page_failures", "num");
2178                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2179                                      LPROCFS_CNTR_AVGMINMAX,
2180                                      "cache_access", "pages");
2181                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2182                                      LPROCFS_CNTR_AVGMINMAX,
2183                                      "cache_hit", "pages");
2184                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2185                                      LPROCFS_CNTR_AVGMINMAX,
2186                                      "cache_miss", "pages");
2187
2188                 lproc_filter_attach_seqstat(obd);
2189                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2190                                                         obd->obd_proc_entry,
2191                                                         NULL, NULL);
2192                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2193                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2194                         CERROR("error %d setting up lprocfs for %s\n",
2195                                rc, "exports");
2196                         obd->obd_proc_exports_entry = NULL;
2197                 }
2198         }
2199         if (obd->obd_proc_exports_entry)
2200                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2201                                    lprocfs_nid_stats_clear_read,
2202                                    lprocfs_nid_stats_clear_write, obd, NULL);
2203
2204         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2205                LUSTRE_CFG_BUFLEN(lcfg, 4));
2206         rc = filter_common_setup(obd, lcfg, (void *)addr);
2207         OBD_PAGE_FREE(page);
2208
2209         if (rc) {
2210                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2211                 lprocfs_free_per_client_stats(obd);
2212                 lprocfs_free_obd_stats(obd);
2213                 lprocfs_obd_cleanup(obd);
2214         }
2215
2216         return rc;
2217 }
2218
2219 static struct llog_operations filter_mds_ost_repl_logops;
2220
2221 static struct llog_operations filter_size_orig_logops = {
2222         .lop_setup   = llog_obd_origin_setup,
2223         .lop_cleanup = llog_obd_origin_cleanup,
2224         .lop_add     = llog_obd_origin_add
2225 };
2226
2227 static int filter_olg_fini(struct obd_llog_group *olg)
2228 {
2229         struct llog_ctxt *ctxt;
2230         int rc = 0, rc2 = 0;
2231         ENTRY;
2232
2233         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2234         if (ctxt)
2235                 rc = llog_cleanup(ctxt);
2236
2237         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2238         if (ctxt) {
2239                 rc2 = llog_cleanup(ctxt);
2240                 if (!rc)
2241                         rc = rc2;
2242         }
2243
2244         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2245         if (ctxt) {
2246                 rc2 = llog_cleanup(ctxt);
2247                 if (!rc)
2248                         rc = rc2;
2249         }
2250
2251         RETURN(rc);
2252 }
2253
2254 static int
2255 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2256                 struct obd_device *tgt)
2257 {
2258         int rc;
2259         ENTRY;
2260
2261         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2262                         &filter_mds_ost_repl_logops);
2263         if (rc)
2264                 GOTO(cleanup, rc);
2265
2266         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2267                         &filter_size_orig_logops);
2268         if (rc)
2269                 GOTO(cleanup, rc);
2270         EXIT;
2271 cleanup:
2272         if (rc)
2273                 filter_olg_fini(olg);
2274         return rc;
2275 }
2276
2277 /**
2278  * Init the default olg, which is embeded in the obd_device, for filter.
2279  */
2280 static int
2281 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2282                         struct obd_device *tgt)
2283 {
2284         struct filter_obd *filter = &obd->u.filter;
2285         struct llog_ctxt *ctxt;
2286         int rc;
2287         ENTRY;
2288
2289         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2290         if (!filter->fo_lcm)
2291                 RETURN(-ENOMEM);
2292
2293         filter_mds_ost_repl_logops = llog_client_ops;
2294         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2295         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2296         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2297
2298         rc = filter_olg_init(obd, olg, tgt);
2299         if (rc)
2300                 GOTO(cleanup_lcm, rc);
2301
2302         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2303                         &llog_lvfs_ops);
2304         if (rc)
2305                 GOTO(cleanup_olg, rc);
2306
2307         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2308         if (!ctxt) {
2309                 CERROR("Can't get ctxt for %p:%x\n", olg,
2310                        LLOG_MDS_OST_REPL_CTXT);
2311                 GOTO(cleanup_olg, rc = -ENODEV);
2312         }
2313         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2314         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2315         llog_ctxt_put(ctxt);
2316
2317         RETURN(0);
2318 cleanup_olg:
2319         filter_olg_fini(olg);
2320 cleanup_lcm:
2321         llog_recov_thread_fini(filter->fo_lcm, 1);
2322         filter->fo_lcm = NULL;
2323         return rc;
2324 }
2325
2326 static int
2327 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2328                  struct obd_device *tgt, int *index)
2329 {
2330         struct filter_obd *filter = &obd->u.filter;
2331         struct llog_ctxt *ctxt;
2332         int rc;
2333         ENTRY;
2334
2335         LASSERT(olg != NULL);
2336         if (olg == &obd->obd_olg)
2337                 return filter_default_olg_init(obd, olg, tgt);
2338
2339         LASSERT(filter->fo_lcm != NULL);
2340         rc = filter_olg_init(obd, olg, tgt);
2341         if (rc)
2342                 RETURN(rc);
2343         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2344         if (!ctxt) {
2345                 CERROR("Can't get ctxt for %p:%x\n", olg,
2346                        LLOG_MDS_OST_REPL_CTXT);
2347                 filter_olg_fini(olg);
2348                 RETURN(-ENODEV);
2349         }
2350         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2351         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2352         llog_ctxt_put(ctxt);
2353         RETURN(rc);
2354 }
2355
2356 static int filter_llog_finish(struct obd_device *obd, int count)
2357 {
2358         struct filter_obd *filter = &obd->u.filter;
2359         struct llog_ctxt *ctxt;
2360         ENTRY;
2361
2362         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2363         if (ctxt) {
2364                 /*
2365                  * Make sure that no cached llcds left in recov_thread.
2366                  * We actually do sync in disconnect time, but disconnect
2367                  * may not come being marked rq_no_resend = 1.
2368                  */
2369                 llog_sync(ctxt, NULL);
2370
2371                 /*
2372                  * Balance class_import_get() in llog_receptor_accept().
2373                  * This is safe to do, as llog is already synchronized
2374                  * and its import may go.
2375                  */
2376                 cfs_mutex_down(&ctxt->loc_sem);
2377                 if (ctxt->loc_imp) {
2378                         class_import_put(ctxt->loc_imp);
2379                         ctxt->loc_imp = NULL;
2380                 }
2381                 cfs_mutex_up(&ctxt->loc_sem);
2382                 llog_ctxt_put(ctxt);
2383         }
2384
2385         if (filter->fo_lcm) {
2386                 cfs_mutex_down(&ctxt->loc_sem);
2387                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2388                 filter->fo_lcm = NULL;
2389                 cfs_mutex_up(&ctxt->loc_sem);
2390         }
2391         RETURN(filter_olg_fini(&obd->obd_olg));
2392 }
2393
2394 /**
2395  * Find the group llog according to group index in the llog group list.
2396  */
2397 static struct obd_llog_group *
2398 filter_find_olg_internal(struct filter_obd *filter, int group)
2399 {
2400         struct obd_llog_group *olg;
2401
2402         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2403         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2404                 if (olg->olg_seq == group)
2405                         RETURN(olg);
2406         }
2407         RETURN(NULL);
2408 }
2409
2410 /**
2411  * Find the group llog according to group index on the filter
2412  */
2413 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2414 {
2415         struct obd_llog_group *olg = NULL;
2416         struct filter_obd *filter;
2417
2418         filter = &obd->u.filter;
2419
2420         if (group == FID_SEQ_LLOG)
2421                 RETURN(&obd->obd_olg);
2422
2423         cfs_spin_lock(&filter->fo_llog_list_lock);
2424         olg = filter_find_olg_internal(filter, group);
2425         cfs_spin_unlock(&filter->fo_llog_list_lock);
2426
2427         RETURN(olg);
2428 }
2429 /**
2430  * Find the llog_group of the filter according to the group. If it can not
2431  * find, create the llog_group, which only happens when mds is being synced
2432  * with OST.
2433  */
2434 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2435 {
2436         struct obd_llog_group *olg = NULL;
2437         struct filter_obd *filter;
2438         int rc;
2439
2440         filter = &obd->u.filter;
2441
2442         if (group == FID_SEQ_LLOG)
2443                 RETURN(&obd->obd_olg);
2444
2445         cfs_spin_lock(&filter->fo_llog_list_lock);
2446         olg = filter_find_olg_internal(filter, group);
2447         if (olg) {
2448                 if (olg->olg_initializing) {
2449                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2450                 } else {
2451                         GOTO(out_unlock, olg);
2452                 }
2453         }
2454         OBD_ALLOC_PTR(olg);
2455         if (olg == NULL)
2456                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2457
2458         llog_group_init(olg, group);
2459         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2460         olg->olg_initializing = 1;
2461         cfs_spin_unlock(&filter->fo_llog_list_lock);
2462
2463         rc = obd_llog_init(obd, olg, obd, NULL);
2464         if (rc) {
2465                cfs_spin_lock(&filter->fo_llog_list_lock);
2466                cfs_list_del(&olg->olg_list);
2467                cfs_spin_unlock(&filter->fo_llog_list_lock);
2468                OBD_FREE_PTR(olg);
2469                GOTO(out, olg = ERR_PTR(-ENOMEM));
2470         }
2471         cfs_spin_lock(&filter->fo_llog_list_lock);
2472         olg->olg_initializing = 0;
2473         cfs_spin_unlock(&filter->fo_llog_list_lock);
2474         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2475               obd->obd_name, group, olg);
2476 out:
2477         RETURN(olg);
2478
2479 out_unlock:
2480         cfs_spin_unlock(&filter->fo_llog_list_lock);
2481         GOTO(out, olg);
2482 }
2483
2484 static int filter_llog_connect(struct obd_export *exp,
2485                                struct llogd_conn_body *body)
2486 {
2487         struct obd_device *obd = exp->exp_obd;
2488         struct llog_ctxt *ctxt;
2489         struct obd_llog_group *olg;
2490         int rc;
2491         ENTRY;
2492
2493         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2494                obd->obd_name, body->lgdc_logid.lgl_oid,
2495                body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2496
2497         olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2498         if (!olg) {
2499                 CERROR(" %s: can not find olg of group %d\n",
2500                        obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2501                 RETURN(-ENOENT);
2502         }
2503         llog_group_set_export(olg, exp);
2504
2505         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2506         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2507                  body->lgdc_ctxt_idx);
2508
2509         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2510               obd->obd_name, body->lgdc_logid.lgl_oid,
2511               body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2512
2513         cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2514         obd->u.filter.fo_mds_ost_sync = 1;
2515         cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2516         rc = llog_connect(ctxt, &body->lgdc_logid,
2517                           &body->lgdc_gen, NULL);
2518         llog_ctxt_put(ctxt);
2519         if (rc != 0)
2520                 CERROR("failed to connect rc %d idx %d\n", rc,
2521                                 body->lgdc_ctxt_idx);
2522
2523         RETURN(rc);
2524 }
2525
2526 static int filter_llog_preclean(struct obd_device *obd)
2527 {
2528         struct obd_llog_group *olg, *tmp;
2529         struct filter_obd *filter;
2530         cfs_list_t  remove_list;
2531         int rc = 0;
2532         ENTRY;
2533
2534         rc = obd_llog_finish(obd, 0);
2535         if (rc)
2536                 CERROR("failed to cleanup llogging subsystem\n");
2537
2538         filter = &obd->u.filter;
2539         CFS_INIT_LIST_HEAD(&remove_list);
2540
2541         cfs_spin_lock(&filter->fo_llog_list_lock);
2542         while (!cfs_list_empty(&filter->fo_llog_list)) {
2543                 olg = cfs_list_entry(filter->fo_llog_list.next,
2544                                      struct obd_llog_group, olg_list);
2545                 cfs_list_del(&olg->olg_list);
2546                 cfs_list_add(&olg->olg_list, &remove_list);
2547         }
2548         cfs_spin_unlock(&filter->fo_llog_list_lock);
2549
2550         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2551                 cfs_list_del_init(&olg->olg_list);
2552                 rc = filter_olg_fini(olg);
2553                 if (rc)
2554                         CERROR("failed to cleanup llogging subsystem for %u\n",
2555                                olg->olg_seq);
2556                 OBD_FREE_PTR(olg);
2557         }
2558
2559         RETURN(rc);
2560 }
2561
2562 static int filter_precleanup(struct obd_device *obd,
2563                              enum obd_cleanup_stage stage)
2564 {
2565         int rc = 0;
2566         ENTRY;
2567
2568         switch(stage) {
2569         case OBD_CLEANUP_EARLY:
2570                 break;
2571         case OBD_CLEANUP_EXPORTS:
2572                 /* Stop recovery before namespace cleanup. */
2573                 target_recovery_fini(obd);
2574                 rc = filter_llog_preclean(obd);
2575                 break;
2576         }
2577         RETURN(rc);
2578 }
2579
2580 static int filter_cleanup(struct obd_device *obd)
2581 {
2582         struct filter_obd *filter = &obd->u.filter;
2583         ENTRY;
2584
2585         if (obd->obd_fail)
2586                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2587                               "will be preserved.\n", obd->obd_name);
2588
2589         obd_exports_barrier(obd);
2590         obd_zombie_barrier();
2591
2592         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2593         lprocfs_free_per_client_stats(obd);
2594         lprocfs_free_obd_stats(obd);
2595         lprocfs_obd_cleanup(obd);
2596         lquota_cleanup(filter_quota_interface_ref, obd);
2597
2598         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2599         obd->obd_namespace = NULL;
2600
2601         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2602
2603         if (obd->u.obt.obt_sb == NULL)
2604                 RETURN(0);
2605
2606         filter_post(obd);
2607
2608         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2609         shrink_dcache_sb(obd->u.obt.obt_sb);
2610
2611         server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2612         obd->u.obt.obt_sb = NULL;
2613
2614         fsfilt_put_ops(obd->obd_fsops);
2615
2616         filter_iobuf_pool_done(filter);
2617
2618         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2619
2620         RETURN(0);
2621 }
2622
2623 static int filter_connect_internal(struct obd_export *exp,
2624                                    struct obd_connect_data *data,
2625                                    int reconnect)
2626 {
2627         struct filter_export_data *fed = &exp->exp_filter_data;
2628
2629         if (!data)
2630                 RETURN(0);
2631
2632         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2633                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2634                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2635                data->ocd_connect_flags, data->ocd_version,
2636                data->ocd_grant, data->ocd_index);
2637
2638         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2639                 CWARN("!!! This export (nid %s) used object group %d "
2640                        "earlier; now it's trying to use group %d!  This could "
2641                        "be a bug in the MDS. Please report to "
2642                        "http://bugzilla.lustre.org/\n",
2643                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2644                 RETURN(-EPROTO);
2645         }
2646         fed->fed_group = data->ocd_group;
2647
2648         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2649         exp->exp_connect_flags = data->ocd_connect_flags;
2650         data->ocd_version = LUSTRE_VERSION_CODE;
2651
2652         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2653         if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2654                   data->ocd_connect_flags & OBD_CONNECT_MDS))
2655                 RETURN(-EPROTO);
2656
2657         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2658                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2659                 obd_size left, want;
2660
2661                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2662                 left = filter_grant_space_left(exp);
2663                 want = data->ocd_grant;
2664                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2665                 data->ocd_grant = fed->fed_grant;
2666                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2667
2668                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2669                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2670                        exp->exp_client_uuid.uuid, exp,
2671                        data->ocd_grant, want, left);
2672
2673                 filter->fo_tot_granted_clients ++;
2674         }
2675
2676         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2677                 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2678                 int index = le32_to_cpu(lsd->lsd_ost_index);
2679
2680                 if (!(lsd->lsd_feature_compat &
2681                       cpu_to_le32(OBD_COMPAT_OST))) {
2682                         /* this will only happen on the first connect */
2683                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2684                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2685                         /* sync is not needed here as filter_client_add will
2686                          * set exp_need_sync flag */
2687                         filter_update_server_data(exp->exp_obd);
2688                 } else if (index != data->ocd_index) {
2689                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2690                                            " %u doesn't match actual OST index"
2691                                            " %u in last_rcvd file, bad "
2692                                            "configuration?\n",
2693                                            obd_export_nid2str(exp), index,
2694                                            data->ocd_index);
2695                         RETURN(-EBADF);
2696                 }
2697                 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2698                  * FIXME: We don't strictly need the COMPAT flag for that,
2699                  * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2700                  * FIXME: We needed it for the index, as index 0 is valid. */
2701         }
2702
2703         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2704                 data->ocd_brw_size = 65536;
2705         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2706                 data->ocd_brw_size = min(data->ocd_brw_size,
2707                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2708                 LASSERT(data->ocd_brw_size);
2709         }
2710
2711         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2712                 __u32 cksum_types = data->ocd_cksum_types;
2713
2714                 /* The client set in ocd_cksum_types the checksum types it
2715                  * supports. We have to mask off the algorithms that we don't
2716                  * support */
2717                 if (cksum_types & OBD_CKSUM_ALL)
2718                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2719                 else
2720                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2721
2722                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2723                                    "%x\n", exp->exp_obd->obd_name,
2724                                    obd_export_nid2str(exp), cksum_types,
2725                                    data->ocd_cksum_types);
2726         } else {
2727                 /* This client does not support OBD_CONNECT_CKSUM
2728                  * fall back to CRC32 */
2729                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2730                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2731                                    exp->exp_obd->obd_name,
2732                                    obd_export_nid2str(exp));
2733         }
2734
2735         RETURN(0);
2736 }
2737
2738 static int filter_reconnect(const struct lu_env *env,
2739                             struct obd_export *exp, struct obd_device *obd,
2740                             struct obd_uuid *cluuid,
2741                             struct obd_connect_data *data,
2742                             void *localdata)
2743 {
2744         int rc;
2745         ENTRY;
2746
2747         if (exp == NULL || obd == NULL || cluuid == NULL)
2748                 RETURN(-EINVAL);
2749
2750         rc = filter_connect_internal(exp, data, 1);
2751         if (rc == 0)
2752                 filter_export_stats_init(obd, exp, 1, localdata);
2753
2754         RETURN(rc);
2755 }
2756
2757 /* nearly identical to mds_connect */
2758 static int filter_connect(const struct lu_env *env,
2759                           struct obd_export **exp, struct obd_device *obd,
2760                           struct obd_uuid *cluuid,
2761                           struct obd_connect_data *data, void *localdata)
2762 {
2763         struct lvfs_run_ctxt saved;
2764         struct lustre_handle conn = { 0 };
2765         struct obd_export *lexp;
2766         __u32 group;
2767         int rc;
2768         ENTRY;
2769
2770         if (exp == NULL || obd == NULL || cluuid == NULL)
2771                 RETURN(-EINVAL);
2772
2773         rc = class_connect(&conn, obd, cluuid);
2774         if (rc)
2775                 RETURN(rc);
2776         lexp = class_conn2export(&conn);
2777         LASSERT(lexp != NULL);
2778
2779         rc = filter_connect_internal(lexp, data, 0);
2780         if (rc)
2781                 GOTO(cleanup, rc);
2782
2783         filter_export_stats_init(obd, lexp, 0, localdata);
2784         if (obd->obd_replayable) {
2785                 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2786                 LASSERT(lcd);
2787                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2788                 rc = filter_client_add(obd, lexp, -1);
2789                 if (rc)
2790                         GOTO(cleanup, rc);
2791         }
2792
2793         group = data->ocd_group;
2794
2795         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2796               obd->obd_name, lexp->exp_handle.h_cookie, group);
2797
2798         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2799         rc = filter_read_groups(obd, group, 1);
2800         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2801         if (rc != 0) {
2802                 CERROR("can't read group %u\n", group);
2803                 GOTO(cleanup, rc);
2804         }
2805
2806         GOTO(cleanup, rc);
2807
2808 cleanup:
2809         if (rc) {
2810                 class_disconnect(lexp);
2811                 lprocfs_exp_cleanup(lexp);
2812                 *exp = NULL;
2813         } else {
2814                 *exp = lexp;
2815         }
2816
2817         RETURN(rc);
2818 }
2819
2820 /* Do extra sanity checks for grant accounting.  We do this at connect,
2821  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2822  * always get rid of it or turn it off when we know accounting is good. */
2823 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2824 {
2825         struct filter_export_data *fed;
2826         struct obd_export *exp;
2827         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2828         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2829         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2830
2831         if (cfs_list_empty(&obd->obd_exports))
2832                 return;
2833
2834         /* We don't want to do this for large machines that do lots of
2835            mounts or unmounts.  It burns... */
2836         if (obd->obd_num_exports > 100)
2837                 return;
2838
2839         cfs_spin_lock(&obd->obd_osfs_lock);
2840         cfs_spin_lock(&obd->obd_dev_lock);
2841         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2842                 int error = 0;
2843                 fed = &exp->exp_filter_data;
2844                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2845                     fed->fed_dirty < 0)
2846                         error = 1;
2847                 if (maxsize > 0) { /* we may not have done a statfs yet */
2848                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2849                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2850                                  exp->exp_client_uuid.uuid, exp,
2851                                  fed->fed_grant, fed->fed_pending, maxsize);
2852                         LASSERTF(fed->fed_dirty <= maxsize,
2853                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2854                                  exp->exp_client_uuid.uuid, exp,
2855                                  fed->fed_dirty, maxsize);
2856                 }
2857                 if (error)
2858                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2859                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2860                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2861                 else
2862                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2863                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2864                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2865                 tot_granted += fed->fed_grant + fed->fed_pending;
2866                 tot_pending += fed->fed_pending;
2867                 tot_dirty += fed->fed_dirty;
2868         }
2869         fo_tot_granted = obd->u.filter.fo_tot_granted;
2870         fo_tot_pending = obd->u.filter.fo_tot_pending;
2871         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2872         cfs_spin_unlock(&obd->obd_dev_lock);
2873         cfs_spin_unlock(&obd->obd_osfs_lock);
2874
2875         /* Do these assertions outside the spinlocks so we don't kill system */
2876         if (tot_granted != fo_tot_granted)
2877                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2878                        func, tot_granted, fo_tot_granted);
2879         if (tot_pending != fo_tot_pending)
2880                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2881                        func, tot_pending, fo_tot_pending);
2882         if (tot_dirty != fo_tot_dirty)
2883                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2884                        func, tot_dirty, fo_tot_dirty);
2885         if (tot_pending > tot_granted)
2886                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2887                        func, tot_pending, tot_granted);
2888         if (tot_granted > maxsize)
2889                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2890                        func, tot_granted, maxsize);
2891         if (tot_dirty > maxsize)
2892                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2893                        func, tot_dirty, maxsize);
2894 }
2895
2896 /* Remove this client from the grant accounting totals.  We also remove
2897  * the export from the obd device under the osfs and dev locks to ensure
2898  * that the filter_grant_sanity_check() calculations are always valid.
2899  * The client should do something similar when it invalidates its import. */
2900 static void filter_grant_discard(struct obd_export *exp)
2901 {
2902         struct obd_device *obd = exp->exp_obd;
2903         struct filter_obd *filter = &obd->u.filter;
2904         struct filter_export_data *fed = &exp->exp_filter_data;
2905
2906         cfs_spin_lock(&obd->obd_osfs_lock);
2907         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2908                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2909                  obd->obd_name, filter->fo_tot_granted,
2910                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2911         filter->fo_tot_granted -= fed->fed_grant;
2912         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2913                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2914                  obd->obd_name, filter->fo_tot_pending,
2915                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2916         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2917         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2918                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2919                  obd->obd_name, filter->fo_tot_dirty,
2920                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2921         filter->fo_tot_dirty -= fed->fed_dirty;
2922         fed->fed_dirty = 0;
2923         fed->fed_grant = 0;
2924
2925         cfs_spin_unlock(&obd->obd_osfs_lock);
2926 }
2927
2928 static int filter_destroy_export(struct obd_export *exp)
2929 {
2930         struct filter_export_data *fed = &exp->exp_filter_data;
2931         ENTRY;
2932
2933         if (fed->fed_pending)
2934                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2935                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2936                        exp, fed->fed_pending);
2937
2938         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2939
2940         target_destroy_export(exp);
2941         ldlm_destroy_export(exp);
2942         lut_client_free(exp);
2943
2944         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2945                 RETURN(0);
2946
2947         if (!exp->exp_obd->obd_replayable)
2948                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2949
2950         filter_grant_discard(exp);
2951         filter_fmd_cleanup(exp);
2952
2953         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2954                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2955                 if (filter->fo_tot_granted_clients > 0)
2956                         filter->fo_tot_granted_clients --;
2957         }
2958
2959         if (!(exp->exp_flags & OBD_OPT_FORCE))
2960                 filter_grant_sanity_check(exp->exp_obd, __func__);
2961
2962         RETURN(0);
2963 }
2964
2965 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2966 {
2967         struct obd_llog_group *olg_min, *olg;
2968         struct filter_obd *filter;
2969         int worked = -1, group;
2970         struct llog_ctxt *ctxt;
2971         ENTRY;
2972
2973         filter = &obd->u.filter;
2974
2975         /* we can't sync log holding spinlock. also, we do not want to get
2976          * into livelock. so we do following: loop over MDS's exports in
2977          * group order and skip already synced llogs -bzzz */
2978         do {
2979                 /* look for group with min. number, but > worked */
2980                 olg_min = NULL;
2981                 group = 1 << 30;
2982                 cfs_spin_lock(&filter->fo_llog_list_lock);
2983                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2984                         if (olg->olg_seq <= worked) {
2985                                 /* this group is already synced */
2986                                 continue;
2987                         }
2988                         if (group < olg->olg_seq) {
2989                                 /* we have group with smaller number to sync */
2990                                 continue;
2991                         }
2992                         /* store current minimal group */
2993                         olg_min = olg;
2994                         group = olg->olg_seq;
2995                 }
2996                 cfs_spin_unlock(&filter->fo_llog_list_lock);
2997
2998                 if (olg_min == NULL)
2999                         break;
3000
3001                 worked = olg_min->olg_seq;
3002                 if (olg_min->olg_exp &&
3003                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3004                         int err;
3005                         ctxt = llog_group_get_ctxt(olg_min,
3006                                                    LLOG_MDS_OST_REPL_CTXT);
3007                         if (ctxt) {
3008                                 err = llog_sync(ctxt, olg_min->olg_exp);
3009                                 llog_ctxt_put(ctxt);
3010                                 if (err) {
3011                                         CERROR("error flushing logs to MDS: "
3012                                                "rc %d\n", err);
3013                                 }
3014                         }
3015                 }
3016         } while (olg_min != NULL);
3017 }
3018
3019 /* Also incredibly similar to mds_disconnect */
3020 static int filter_disconnect(struct obd_export *exp)
3021 {
3022         struct obd_device *obd = exp->exp_obd;
3023         int rc;
3024         ENTRY;
3025
3026         LASSERT(exp);
3027         class_export_get(exp);
3028
3029         if (!(exp->exp_flags & OBD_OPT_FORCE))
3030                 filter_grant_sanity_check(obd, __func__);
3031         filter_grant_discard(exp);
3032
3033         /* Flush any remaining cancel messages out to the target */
3034         filter_sync_llogs(obd, exp);
3035
3036         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3037
3038         rc = server_disconnect_export(exp);
3039
3040         /* Do not erase record for recoverable client. */
3041         if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3042                 filter_client_del(exp);
3043         else
3044                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3045
3046         class_export_put(exp);
3047         RETURN(rc);
3048 }
3049
3050 /* reverse import is changed, sync all cancels */
3051 static void filter_revimp_update(struct obd_export *exp)
3052 {
3053         ENTRY;
3054
3055         LASSERT(exp);
3056         class_export_get(exp);
3057
3058         /* flush any remaining cancel messages out to the target */
3059         filter_sync_llogs(exp->exp_obd, exp);
3060         class_export_put(exp);
3061         EXIT;
3062 }
3063
3064 static int filter_ping(struct obd_export *exp)
3065 {
3066         filter_fmd_expire(exp);
3067         return 0;
3068 }
3069
3070 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3071                                   const char *what, int quiet)
3072 {
3073         struct dentry *dchild = NULL;
3074
3075         dchild = filter_fid2dentry(obd, NULL,  ostid->oi_seq, ostid->oi_id);
3076
3077         if (IS_ERR(dchild)) {
3078                 CERROR("%s error looking up object: "POSTID"\n",
3079                        what, ostid->oi_id, ostid->oi_seq);
3080                 RETURN(dchild);
3081         }
3082
3083         if (dchild->d_inode == NULL) {
3084                 if (!quiet)
3085                         CERROR("%s: %s on non-existent object: "POSTID" \n",
3086                                obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3087                 f_dput(dchild);
3088                 RETURN(ERR_PTR(-ENOENT));
3089         }
3090
3091         return dchild;
3092 }
3093
3094 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3095 {
3096         struct dentry *dentry = NULL;
3097         struct obd_device *obd;
3098         int rc = 0;
3099         ENTRY;
3100
3101         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3102                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3103         if (rc)
3104                 RETURN(rc);
3105
3106         obd = class_exp2obd(exp);
3107         if (obd == NULL) {
3108                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3109                 RETURN(-EINVAL);
3110         }
3111
3112         dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3113         if (IS_ERR(dentry))
3114                 RETURN(PTR_ERR(dentry));
3115
3116         /* Limit the valid bits in the return data to what we actually use */
3117         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3118         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3119
3120         f_dput(dentry);
3121         RETURN(rc);
3122 }
3123
3124 /* this should be enabled/disabled in condition to enabled/disabled large
3125  * inodes (fast EAs) in backing store FS. */
3126 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3127                         void *handle, struct obdo *oa)
3128 {
3129         struct obd_device *obd = exp->exp_obd;
3130         int rc = 0;
3131         ENTRY;
3132
3133         if (oa->o_valid & OBD_MD_FLFID) {
3134                 struct filter_fid ff;
3135
3136                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3137                         oa->o_seq = 0;
3138                 /* packing fid and converting it to LE for storing into EA.
3139                  * Here ->o_stripe_idx should be filled by LOV and rest of
3140                  * fields - by client. */
3141                 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3142                 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3143                 /* XXX: we are ignoring o_parent_ver here, since this should
3144                  *      be the same for all objects in this fileset. */
3145                 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3146                 ff.ff_objid = cpu_to_le64(oa->o_id);
3147                 ff.ff_seq = cpu_to_le64(oa->o_seq);
3148
3149                 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3150                        LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3151                        oa->o_seq);
3152
3153                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3154                 if (rc)
3155                         CERROR("store fid in object failed! rc: %d\n", rc);
3156         } else {
3157                 CDEBUG(D_HA, "OSS object without fid info!\n");
3158         }
3159
3160         RETURN(rc);
3161 }
3162
3163 /* this is called from filter_truncate() until we have filter_punch() */
3164 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3165                             struct obdo *oa, struct obd_trans_info *oti)
3166 {
3167         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3168         struct llog_cookie *fcc = NULL;
3169         struct filter_obd *filter;
3170         int rc, err, sync = 0;
3171         loff_t old_size = 0;
3172         unsigned int ia_valid;
3173         struct inode *inode;
3174         struct page *page = NULL;
3175         struct iattr iattr;
3176         void *handle;
3177         ENTRY;
3178
3179         LASSERT(dentry != NULL);
3180         LASSERT(!IS_ERR(dentry));
3181
3182         inode = dentry->d_inode;
3183         LASSERT(inode != NULL);
3184
3185         filter = &exp->exp_obd->u.filter;
3186         iattr_from_obdo(&iattr, oa, oa->o_valid);
3187         ia_valid = iattr.ia_valid;
3188
3189         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3190                 OBD_ALLOC(fcc, sizeof(*fcc));
3191                 if (fcc != NULL)
3192                         *fcc = oa->o_lcookie;
3193         }
3194         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3195                 DQUOT_INIT(inode);
3196                 /* Filter truncates and writes are serialized by
3197                  * i_alloc_sem, see the comment in
3198                  * filter_preprw_write.*/
3199                 if (ia_valid & ATTR_SIZE)
3200                         down_write(&inode->i_alloc_sem);
3201                 LOCK_INODE_MUTEX(inode);
3202                 old_size = i_size_read(inode);
3203         }
3204
3205         /* VBR: version recovery check */
3206         rc = filter_version_get_check(exp, oti, inode);
3207         if (rc)
3208                 GOTO(out_unlock, rc);
3209
3210         /* Let's pin the last page so that ldiskfs_truncate
3211          * should not start GFP_FS allocation. */
3212         if (ia_valid & ATTR_SIZE) {
3213                 page = grab_cache_page(inode->i_mapping,
3214                                        iattr.ia_size >> PAGE_CACHE_SHIFT);
3215                 if (page == NULL)
3216                         GOTO(out_unlock, rc = -ENOMEM);
3217
3218                 unlock_page(page);
3219         }
3220
3221         /* If the inode still has SUID+SGID bits set (see filter_precreate())
3222          * then we will accept the UID+GID sent by the client during write for
3223          * initializing the ownership of this inode.  We only allow this to
3224          * happen once so clear these bits in setattr. In 2.6 kernels it is
3225          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3226          * the flags that are actually being set. */
3227         if (ia_valid & (ATTR_UID | ATTR_GID)) {
3228                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3229                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3230
3231                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3232                         if (!(ia_valid & ATTR_MODE)) {
3233                                 iattr.ia_mode = inode->i_mode;
3234                                 iattr.ia_valid |= ATTR_MODE;
3235                         }
3236                         iattr.ia_mode &= ~S_ISUID;
3237                 }
3238                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3239                         if (!(iattr.ia_valid & ATTR_MODE)) {
3240                                 iattr.ia_mode = inode->i_mode;
3241                                 iattr.ia_valid |= ATTR_MODE;
3242                         }
3243                         iattr.ia_mode &= ~S_ISGID;
3244                 }
3245
3246                 orig_ids[USRQUOTA] = inode->i_uid;
3247                 orig_ids[GRPQUOTA] = inode->i_gid;
3248                 handle = fsfilt_start_log(exp->exp_obd, inode,
3249                                           FSFILT_OP_SETATTR, oti, 1);
3250                 if (IS_ERR(handle))
3251                         GOTO(out_unlock, rc = PTR_ERR(handle));
3252
3253                 /* update inode EA only once when inode is suid bit marked. As
3254                  * on 2.6.x UID and GID may be set separately, we check here
3255                  * only one of them to avoid double setting. */
3256                 if (inode->i_mode & S_ISUID)
3257                         filter_update_fidea(exp, inode, handle, oa);
3258         } else {
3259                 handle = fsfilt_start(exp->exp_obd, inode,
3260                                       FSFILT_OP_SETATTR, oti);
3261                 if (IS_ERR(handle))
3262                         GOTO(out_unlock, rc = PTR_ERR(handle));
3263         }
3264         if (oa->o_valid & OBD_MD_FLFLAGS) {
3265                 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3266                                       FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3267         } else {
3268                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3269                 if (fcc != NULL)
3270                         /* set cancel cookie callback function */
3271                         sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3272                                                      filter_cancel_cookies_cb,
3273                                                      fcc);
3274         }
3275
3276         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3277                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3278
3279        /* The truncate might have used up our transaction credits.  Make sure
3280         * we have two left for the last_rcvd and VBR inode version updates. */
3281         err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3282
3283         rc = filter_finish_transno(exp, inode, oti, rc, sync);
3284         if (sync) {
3285                 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3286                 fcc = NULL;
3287         }
3288
3289         err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3290         if (err) {
3291                 CERROR("error on commit, err = %d\n", err);
3292                 if (!rc)
3293                         rc = err;
3294         } else {
3295                 fcc = NULL;
3296         }
3297
3298         /* For a partial-page truncate flush the page to disk immediately
3299          * to avoid data corruption during direct disk write. b=17397 */
3300         if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3301             old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3302                 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3303                                                iattr.ia_size + 1);
3304                 if (!rc)
3305                         rc = err;
3306         }
3307
3308         EXIT;
3309
3310 out_unlock:
3311         if (page)
3312                 page_cache_release(page);
3313
3314         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3315                 UNLOCK_INODE_MUTEX(inode);
3316         if (ia_valid & ATTR_SIZE)
3317                 up_write(&inode->i_alloc_sem);
3318         if (fcc)
3319                 OBD_FREE(fcc, sizeof(*fcc));
3320
3321         /* trigger quota release */
3322         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3323                 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3324                 int rc2 = lquota_adjust(filter_quota_interface_ref,
3325                                         exp->exp_obd, cur_ids,
3326                                         orig_ids, rc, FSFILT_OP_SETATTR);
3327                 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3328                        "filter adjust qunit. (rc:%d)\n", rc2);
3329         }
3330         return rc;
3331 }
3332
3333 /* this is called from filter_truncate() until we have filter_punch() */
3334 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3335                    struct obd_trans_info *oti)
3336 {
3337         struct obdo *oa = oinfo->oi_oa;
3338         struct lustre_capa *capa = oinfo_capa(oinfo);
3339         struct ldlm_res_id res_id;
3340         struct filter_mod_data *fmd;
3341         struct lvfs_run_ctxt saved;
3342         struct filter_obd *filter;
3343         struct ldlm_resource *res;
3344         struct dentry *dentry;
3345         __u64 opc = CAPA_OPC_META_WRITE;
3346         int rc;
3347         ENTRY;
3348
3349         if (oa->o_valid & OBD_FL_TRUNC)
3350                 opc |= CAPA_OPC_OSS_TRUNC;
3351
3352         rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3353         if (rc)
3354                 RETURN(rc);
3355
3356         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3357                 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3358                 if (rc)
3359                         RETURN(rc);
3360         }
3361
3362         osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3363         /* This would be very bad - accidentally truncating a file when
3364          * changing the time or similar - bug 12203. */
3365         if (oa->o_valid & OBD_MD_FLSIZE &&
3366             oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3367                 static char mdsinum[48];
3368
3369                 if (oa->o_valid & OBD_MD_FLFID)
3370                         snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3371                                  oa->o_parent_seq, oa->o_parent_oid,
3372                                  oa->o_parent_ver);
3373                 else
3374                         mdsinum[0] = '\0';
3375
3376                 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3377                        "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3378                        oa->o_id, oa->o_seq, mdsinum);
3379                 RETURN(-EPERM);
3380         }
3381
3382         dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3383         if (IS_ERR(dentry))
3384                 RETURN(PTR_ERR(dentry));
3385
3386         filter = &exp->exp_obd->u.filter;
3387         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3388
3389         /*
3390          * We need to be atomic against a concurrent write
3391          * (which takes the semaphore for reading). fmd_mactime_xid
3392          * checks will have no effect if a write request with lower
3393          * xid starts just before a setattr and finishes later than
3394          * the setattr (see bug 21489, comment 27).
3395          */
3396         if (oa->o_valid &
3397             (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3398                 down_write(&dentry->d_inode->i_alloc_sem);
3399                 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3400                 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3401                         fmd->fmd_mactime_xid = oti->oti_xid;
3402                 filter_fmd_put(exp, fmd);
3403                 up_write(&dentry->d_inode->i_alloc_sem);
3404         }
3405
3406         /* setting objects attributes (including owner/group) */
3407         rc = filter_setattr_internal(exp, dentry, oa, oti);
3408         if (rc)
3409                 GOTO(out_unlock, rc);
3410
3411         res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3412                                 &res_id, LDLM_EXTENT, 0);
3413
3414         if (res != NULL) {
3415                 LDLM_RESOURCE_ADDREF(res);
3416                 rc = ldlm_res_lvbo_update(res, NULL, 0);
3417                 LDLM_RESOURCE_DELREF(res);
3418                 ldlm_resource_putref(res);
3419         }
3420
3421         oa->o_valid = OBD_MD_FLID;
3422
3423         /* Quota release need uid/gid info */
3424         obdo_from_inode(oa, dentry->d_inode, NULL,
3425                         FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3426
3427         EXIT;
3428 out_unlock:
3429         f_dput(dentry);
3430         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3431         return rc;
3432 }
3433
3434 /* XXX identical to osc_unpackmd */
3435 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3436                            struct lov_mds_md *lmm, int lmm_bytes)
3437 {
3438         int lsm_size;
3439         ENTRY;
3440
3441         if (lmm != NULL) {
3442                 if (lmm_bytes < sizeof (*lmm)) {
3443                         CERROR("lov_mds_md too small: %d, need %d\n",
3444                                lmm_bytes, (int)sizeof(*lmm));
3445                         RETURN(-EINVAL);
3446                 }
3447                 /* XXX LOV_MAGIC etc check? */
3448
3449                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3450                         CERROR("lov_mds_md: zero lmm_object_id\n");
3451                         RETURN(-EINVAL);
3452                 }
3453         }
3454
3455         lsm_size = lov_stripe_md_size(1);
3456         if (lsmp == NULL)
3457                 RETURN(lsm_size);
3458
3459         if (*lsmp != NULL && lmm == NULL) {
3460                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3461                 OBD_FREE(*lsmp, lsm_size);
3462                 *lsmp = NULL;
3463                 RETURN(0);
3464         }
3465
3466         if (*lsmp == NULL) {
3467                 OBD_ALLOC(*lsmp, lsm_size);
3468                 if (*lsmp == NULL)
3469                         RETURN(-ENOMEM);
3470
3471                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3472                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3473                         OBD_FREE(*lsmp, lsm_size);
3474                         RETURN(-ENOMEM);
3475                 }
3476                 loi_init((*lsmp)->lsm_oinfo[0]);
3477         }
3478
3479         if (lmm != NULL) {
3480                 /* XXX zero *lsmp? */
3481                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3482                 LASSERT((*lsmp)->lsm_object_id);
3483         }
3484
3485         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3486
3487         RETURN(lsm_size);
3488 }
3489
3490 /* caller must hold fo_create_locks[oa->o_seq] */
3491 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3492                                      struct filter_obd *filter)
3493 {
3494         struct obdo doa = { 0 }; /* XXX obdo on stack */
3495         obd_id last, id;
3496         int rc = 0;
3497         int skip_orphan;
3498         ENTRY;
3499
3500         LASSERT(down_trylock(&filter->fo_create_locks[oa->o_seq]) != 0);
3501
3502         memset(&doa, 0, sizeof(doa));
3503
3504         doa.o_valid |= OBD_MD_FLGROUP;
3505         doa.o_seq = oa->o_seq;
3506         doa.o_mode = S_IFREG;
3507
3508         if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3509                 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3510                        exp->exp_obd->obd_name, doa.o_seq);
3511                 RETURN(0);
3512         }
3513
3514         last = filter_last_id(filter, doa.o_seq);
3515
3516         skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3517
3518         CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3519                exp->exp_obd->obd_name, oa->o_id + 1, last,
3520                skip_orphan ? ", orphan objids won't be reused any more." : ".");
3521
3522         for (id = last; id > oa->o_id; id--) {
3523                 doa.o_id = id;
3524                 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3525                 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3526                         CEMERG("error destroying precreate objid "LPU64": %d\n",
3527                                id, rc);
3528
3529                 /* update last_id on disk periodically so that if we restart
3530                  * we don't need to re-scan all of the just-deleted objects. */
3531                 if ((id & 511) == 0 && !skip_orphan) {
3532                         filter_set_last_id(filter, id - 1, doa.o_seq);
3533                         filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3534                 }
3535         }
3536
3537         CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3538                exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3539
3540         if (!skip_orphan) {
3541                 filter_set_last_id(filter, id, doa.o_seq);
3542                 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3543         } else {
3544                 /*
3545                  * We have destroyed orphan objects, but don't want to reuse
3546                  * them. Therefore we don't reset last_id to the last created
3547                  * objects. Instead, we report back to the MDS the object id
3548                  * of the last orphan, so that the MDS can restart allocating
3549                  * objects from this id + 1 and thus skip the whole orphan
3550                  * object id range
3551                  */
3552                 oa->o_id = last;
3553                 rc = 0;
3554         }
3555         cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3556
3557         RETURN(rc);
3558 }
3559
3560 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3561                             obd_seq group, int *num);
3562 /* returns a negative error or a nonnegative number of files to create */
3563 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3564                                    obd_seq group, struct obd_trans_info *oti)
3565 {
3566         struct obd_device *obd = exp->exp_obd;
3567         struct filter_obd *filter = &obd->u.filter;
3568         int diff, rc;
3569         ENTRY;
3570
3571         /* delete orphans request */
3572         if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3573                 obd_id last = filter_last_id(filter, group);
3574
3575                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3576                         CERROR("%s: dropping old orphan cleanup request\n",
3577                                obd->obd_name);
3578                         RETURN(0);
3579                 }
3580                 /* This causes inflight precreates to abort and drop lock */
3581                 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3582                 cfs_down(&filter->fo_create_locks[group]);
3583                 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3584                         CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3585                                exp->exp_obd->obd_name, group);
3586                         cfs_up(&filter->fo_create_locks[group]);
3587                         RETURN(0);
3588                 }
3589                 diff = oa->o_id - last;
3590                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3591                        last, diff);
3592
3593                 if (-diff > OST_MAX_PRECREATE) {
3594                         CERROR("%s: ignoring bogus orphan destroy request: "
3595                                "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3596                                oa->o_id, last);
3597                         /* FIXME: should reset precreate_next_id on MDS */
3598                         GOTO(out, rc = -EINVAL);
3599                 }
3600                 if (diff < 0) {
3601                         rc = filter_destroy_precreated(exp, oa, filter);
3602                         if (rc)
3603                                 CERROR("%s: unable to write lastobjid, but "
3604                                        "orphans were deleted\n", obd->obd_name);
3605                         GOTO(out, rc);
3606                 } else {
3607                         /* XXX: Used by MDS for the first time! */
3608                         cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3609                 }
3610         } else {
3611                 cfs_down(&filter->fo_create_locks[group]);
3612                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3613                         CERROR("%s: dropping old precreate request\n",
3614                                obd->obd_name);
3615                         GOTO(out, rc = 0);
3616                 }
3617                 /* only precreate if group == 0 and o_id is specfied */
3618                 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3619                         diff = 1;
3620                 else
3621                         diff = oa->o_id - filter_last_id(filter, group);
3622                 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3623                        filter_last_id(filter, group), diff);
3624
3625                 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3626                          oa->o_id, filter_last_id(filter, group), diff);
3627         }
3628
3629         if (diff > 0) {
3630                 oa->o_id = filter_last_id(&obd->u.filter, group);
3631                 rc = filter_precreate(obd, oa, group, &diff);
3632                 oa->o_id = filter_last_id(&obd->u.filter, group);
3633                 oa->o_seq = group;
3634                 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3635                 GOTO(out, rc);
3636         }
3637         /* else diff == 0 */
3638         GOTO(out, rc = 0);
3639 out:
3640         cfs_up(&filter->fo_create_locks[group]);
3641         return rc;
3642 }
3643
3644 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3645                          __u64 max_age, __u32 flags)
3646 {
3647         struct filter_obd *filter = &obd->u.filter;
3648         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3649         int rc;
3650         ENTRY;
3651
3652         /* at least try to account for cached pages.  its still racey and
3653          * might be under-reporting if clients haven't announced their
3654          * caches with brw recently */
3655         cfs_spin_lock(&obd->obd_osfs_lock);
3656         rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3657         memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3658         cfs_spin_unlock(&obd->obd_osfs_lock);
3659
3660         CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3661                " pending "LPU64" free "LPU64" avail "LPU64"\n",
3662                filter->fo_tot_dirty, filter->fo_tot_granted,
3663                filter->fo_tot_pending,
3664                osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3665
3666         filter_grant_sanity_check(obd, __func__);
3667
3668         osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3669                                ((filter->fo_tot_dirty + filter->fo_tot_pending +
3670                                  osfs->os_bsize - 1) >> blockbits));
3671
3672         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3673                 struct lr_server_data *lsd = class_server_data(obd);
3674                 int index = le32_to_cpu(lsd->lsd_ost_index);
3675
3676                 if (obd_fail_val == -1 ||
3677                     index == obd_fail_val)
3678                         osfs->os_bfree = osfs->os_bavail = 2;
3679                 else if (obd_fail_loc & OBD_FAIL_ONCE)
3680               &n