Whamcloud - gitweb
LU-241 Support crc32c with hardware accelerated instruction as one of lustre checksums
[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  * Copyright (c) 2011 Whamcloud, Inc.
34  */
35 /*
36  * This file is part of Lustre, http://www.lustre.org/
37  * Lustre is a trademark of Sun Microsystems, Inc.
38  *
39  * lustre/obdfilter/filter.c
40  *
41  * Author: Peter Braam <braam@clusterfs.com>
42  * Author: Andreas Dilger <adilger@clusterfs.com>
43  */
44
45 /*
46  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
47  *            (which need to get journal_lock, may block if journal full).
48  *
49  * Invariant: Call filter_start_transno() before any journal ops to avoid the
50  *            same deadlock problem.  We can (and want) to get rid of the
51  *            transno sem in favour of the dir/inode i_mutex to avoid single
52  *            threaded operation on the OST.
53  */
54
55 #define DEBUG_SUBSYSTEM S_FILTER
56
57 #ifndef AUTOCONF_INCLUDED
58 #include <linux/config.h>
59 #endif
60 #include <linux/module.h>
61 #include <linux/fs.h>
62 #include <linux/dcache.h>
63 #include <linux/init.h>
64 #include <linux/version.h>
65 #include <linux/sched.h>
66 #include <linux/mount.h>
67 #include <linux/buffer_head.h>
68
69 #include <obd_cksum.h>
70 #include <obd_class.h>
71 #include <obd_lov.h>
72 #include <lustre_dlm.h>
73 #include <lustre_fsfilt.h>
74 #include <lprocfs_status.h>
75 #include <lustre_log.h>
76 #include <libcfs/list.h>
77 #include <lustre_disk.h>
78 #include <lustre_quota.h>
79 #include <linux/slab.h>
80 #include <lustre_param.h>
81 #include <lustre/ll_fiemap.h>
82
83 #include "filter_internal.h"
84
85 static struct lvfs_callback_ops filter_lvfs_ops;
86 cfs_mem_cache_t *ll_fmd_cachep;
87
88 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
89                              void *cb_data, int error)
90 {
91         struct obd_export *exp = cb_data;
92         LASSERT(exp->exp_obd == obd);
93         obd_transno_commit_cb(obd, transno, exp, error);
94         class_export_cb_put(exp);
95 }
96
97 int filter_version_get_check(struct obd_export *exp,
98                              struct obd_trans_info *oti, struct inode *inode)
99 {
100         __u64 curr_version;
101
102         if (inode == NULL || oti == NULL)
103                 RETURN(0);
104
105         curr_version = fsfilt_get_version(exp->exp_obd, inode);
106         if ((__s64)curr_version == -EOPNOTSUPP)
107                 RETURN(0);
108         /* VBR: version is checked always because costs nothing */
109         if (oti->oti_pre_version != 0 &&
110             oti->oti_pre_version != curr_version) {
111                 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
112                        oti->oti_pre_version, curr_version);
113                 cfs_spin_lock(&exp->exp_lock);
114                 exp->exp_vbr_failed = 1;
115                 cfs_spin_unlock(&exp->exp_lock);
116                 RETURN (-EOVERFLOW);
117         }
118         oti->oti_pre_version = curr_version;
119         RETURN(0);
120 }
121
122 /* Assumes caller has already pushed us into the kernel context. */
123 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
124                           struct obd_trans_info *oti, int rc, int force_sync)
125 {
126         struct obd_device_target *obt = &exp->exp_obd->u.obt;
127         struct tg_export_data *ted = &exp->exp_target_data;
128         struct lr_server_data *lsd = class_server_data(exp->exp_obd);
129         struct lsd_client_data *lcd;
130         __u64 last_rcvd;
131         loff_t off;
132         int err, log_pri = D_RPCTRACE;
133
134         /* Propagate error code. */
135         if (rc)
136                 RETURN(rc);
137
138         if (!exp->exp_obd->obd_replayable || oti == NULL)
139                 RETURN(rc);
140
141         cfs_mutex_down(&ted->ted_lcd_lock);
142         lcd = ted->ted_lcd;
143         /* if the export has already been disconnected, we have no last_rcvd slot,
144          * update server data with latest transno then */
145         if (lcd == NULL) {
146                 cfs_mutex_up(&ted->ted_lcd_lock);
147                 CWARN("commit transaction for disconnected client %s: rc %d\n",
148                       exp->exp_client_uuid.uuid, rc);
149                 err = filter_update_server_data(exp->exp_obd);
150                 RETURN(err);
151         }
152
153         /* we don't allocate new transnos for replayed requests */
154         cfs_spin_lock(&obt->obt_lut->lut_translock);
155         if (oti->oti_transno == 0) {
156                 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
157                 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
158                 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
159         } else {
160                 last_rcvd = oti->oti_transno;
161                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
162                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
163                 if (unlikely(last_rcvd < le64_to_cpu(lcd->lcd_last_transno))) {
164                         CERROR("Trying to overwrite bigger transno, on-disk: "
165                                LPU64", new: "LPU64"\n",
166                                le64_to_cpu(lcd->lcd_last_transno), last_rcvd);
167                         cfs_spin_lock(&exp->exp_lock);
168                         exp->exp_vbr_failed = 1;
169                         cfs_spin_unlock(&exp->exp_lock);
170                         cfs_spin_unlock(&obt->obt_lut->lut_translock);
171                         cfs_mutex_up(&ted->ted_lcd_lock);
172                         RETURN(-EOVERFLOW);
173                 }
174         }
175         oti->oti_transno = last_rcvd;
176
177         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
178         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
179         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
180         cfs_spin_unlock(&obt->obt_lut->lut_translock);
181
182         if (inode)
183                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
184
185         off = ted->ted_lr_off;
186         if (off <= 0) {
187                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
188                        ted->ted_lr_idx, ted->ted_lr_off);
189                 err = -EINVAL;
190         } else {
191                 class_export_cb_get(exp); /* released when the cb is called */
192                 if (!force_sync)
193                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
194                                                            last_rcvd,
195                                                            oti->oti_handle,
196                                                            filter_commit_cb,
197                                                            exp);
198
199                 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
200                                           lcd, sizeof(*lcd), &off,
201                                           force_sync | exp->exp_need_sync);
202                 if (force_sync)
203                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
204         }
205         if (err) {
206                 log_pri = D_ERROR;
207                 if (rc == 0)
208                         rc = err;
209         }
210
211         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
212                last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
213         cfs_mutex_up(&ted->ted_lcd_lock);
214         RETURN(rc);
215 }
216
217 void f_dput(struct dentry *dentry)
218 {
219         /* Can't go inside filter_ddelete because it can block */
220         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
221                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
222         LASSERT(atomic_read(&dentry->d_count) > 0);
223
224         dput(dentry);
225 }
226
227 static void init_brw_stats(struct brw_stats *brw_stats)
228 {
229         int i;
230         for (i = 0; i < BRW_LAST; i++)
231                 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
232 }
233
234 static int lprocfs_init_rw_stats(struct obd_device *obd,
235                                  struct lprocfs_stats **stats)
236 {
237         int num_stats;
238
239         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
240                                                         LPROC_FILTER_LAST - 1;
241         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
242         if (*stats == NULL)
243                 return -ENOMEM;
244
245         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
246         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
247                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
248         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
249                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
250
251         return(0);
252 }
253
254 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
255    plus the procfs overhead :( */
256 static int filter_export_stats_init(struct obd_device *obd,
257                                     struct obd_export *exp,
258                                     void *client_nid)
259 {
260         int rc, newnid = 0;
261         ENTRY;
262
263         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
264                 /* Self-export gets no proc entry */
265                 RETURN(0);
266
267         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
268         if (rc) {
269                 /* Mask error for already created
270                  * /proc entries */
271                 if (rc == -EALREADY)
272                         rc = 0;
273                 RETURN(rc);
274         }
275
276         if (newnid) {
277                 struct nid_stat *tmp = exp->exp_nid_stats;
278                 LASSERT(tmp != NULL);
279
280                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
281                 if (tmp->nid_brw_stats == NULL)
282                         GOTO(clean, rc = -ENOMEM);
283
284                 init_brw_stats(tmp->nid_brw_stats);
285                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
286                                         0644, &filter_per_nid_stats_fops,
287                                         exp->exp_nid_stats);
288                 if (rc)
289                         CWARN("Error adding the brw_stats file\n");
290
291                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
292                 if (rc)
293                         GOTO(clean, rc);
294
295                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
296                                             tmp->nid_stats);
297                 if (rc)
298                         GOTO(clean, rc);
299                 rc = lprocfs_nid_ldlm_stats_init(tmp);
300                 if (rc)
301                         GOTO(clean, rc);
302         }
303
304         RETURN(0);
305  clean:
306         return rc;
307 }
308
309 /* Add client data to the FILTER.  We use a bitmap to locate a free space
310  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
311  * Otherwise, we have just read the data from the last_rcvd file and
312  * we know its offset. */
313 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
314                              int cl_idx)
315 {
316         struct obd_device_target *obt = &obd->u.obt;
317         struct tg_export_data *ted = &exp->exp_target_data;
318         struct lr_server_data *lsd = class_server_data(obd);
319         unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
320         int new_client = (cl_idx == -1);
321
322         ENTRY;
323
324         LASSERT(bitmap != NULL);
325         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
326
327         /* Self-export */
328         if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
329                 RETURN(0);
330
331         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
332          * there's no need for extra complication here
333          */
334         if (new_client) {
335                 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
336         repeat:
337                 if (cl_idx >= LR_MAX_CLIENTS) {
338                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
339                                cl_idx);
340                         RETURN(-EOVERFLOW);
341                 }
342                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
343                         cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
344                                                         cl_idx);
345                         goto repeat;
346                 }
347         } else {
348                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
349                         CERROR("FILTER client %d: bit already set in bitmap!\n",
350                                cl_idx);
351                         LBUG();
352                 }
353         }
354
355         ted->ted_lr_idx = cl_idx;
356         ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
357                           cl_idx * le16_to_cpu(lsd->lsd_client_size);
358         cfs_init_mutex(&ted->ted_lcd_lock);
359         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
360
361         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
362                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
363
364         if (new_client) {
365                 struct lvfs_run_ctxt saved;
366                 loff_t off = ted->ted_lr_off;
367                 int rc;
368                 void *handle;
369
370                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
371                        ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
372
373                 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
374                         RETURN(-ENOSPC);
375
376                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
377                 /* Transaction needed to fix bug 1403 */
378                 handle = fsfilt_start(obd,
379                                       obt->obt_rcvd_filp->f_dentry->d_inode,
380                                       FSFILT_OP_SETATTR, NULL);
381                 if (IS_ERR(handle)) {
382                         rc = PTR_ERR(handle);
383                         CERROR("unable to start transaction: rc %d\n", rc);
384                 } else {
385                         ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
386                         exp->exp_last_request_time = cfs_time_current_sec();
387                         rc = fsfilt_add_journal_cb(obd, 0, handle,
388                                                    target_client_add_cb,
389                                                    class_export_cb_get(exp));
390                         if (rc == 0) {
391                                 cfs_spin_lock(&exp->exp_lock);
392                                 exp->exp_need_sync = 1;
393                                 cfs_spin_unlock(&exp->exp_lock);
394                         }
395                         rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
396                                                  ted->ted_lcd,
397                                                  sizeof(*ted->ted_lcd),
398                                                  &off, rc /* sync if no cb */);
399                         fsfilt_commit(obd,
400                                       obt->obt_rcvd_filp->f_dentry->d_inode,
401                                       handle, 0);
402                 }
403                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
404
405                 if (rc) {
406                         CERROR("error writing %s client idx %u: rc %d\n",
407                                LAST_RCVD, ted->ted_lr_idx, rc);
408                         RETURN(rc);
409                 }
410         }
411         RETURN(0);
412 }
413
414 static int filter_client_del(struct obd_export *exp)
415 {
416         struct tg_export_data *ted = &exp->exp_target_data;
417         struct obd_device_target *obt = &exp->exp_obd->u.obt;
418         struct lvfs_run_ctxt saved;
419         int rc;
420         loff_t off;
421         ENTRY;
422
423         if (ted->ted_lcd == NULL)
424                 RETURN(0);
425
426         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
427         if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
428                 GOTO(free, 0);
429
430         LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
431
432         off = ted->ted_lr_off;
433
434         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
435                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
436
437         /* Don't clear ted_lr_idx here as it is likely also unset.  At worst
438          * we leak a client slot that will be cleaned on the next recovery. */
439         if (off <= 0) {
440                 CERROR("%s: client idx %d has med_off %lld\n",
441                        exp->exp_obd->obd_name, ted->ted_lr_idx, off);
442                 GOTO(free, rc = -EINVAL);
443         }
444
445         /* Clear the bit _after_ zeroing out the client so we don't
446            race with filter_client_add and zero out new clients.*/
447         if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
448                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
449                        ted->ted_lr_idx);
450                 LBUG();
451         }
452
453         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
454         /* Make sure the server's last_transno is up to date.
455          * This should be done before zeroing client slot so last_transno will
456          * be in server data or in client data in case of failure */
457         filter_update_server_data(exp->exp_obd);
458
459         cfs_mutex_down(&ted->ted_lcd_lock);
460         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
461         rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
462                                  ted->ted_lcd,
463                                  sizeof(*ted->ted_lcd), &off, 0);
464         cfs_mutex_up(&ted->ted_lcd_lock);
465         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
466
467         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
468                "zero out client %s at idx %u/%llu in %s, rc %d\n",
469                ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
470                LAST_RCVD, rc);
471         EXIT;
472 free:
473         return 0;
474 }
475
476 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
477 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
478                                          struct filter_mod_data *fmd)
479 {
480         LASSERT_SPIN_LOCKED(&fed->fed_lock);
481         if (--fmd->fmd_refcount == 0) {
482                 /* XXX when we have persistent reservations and the handle
483                  * is stored herein we need to drop it here. */
484                 fed->fed_mod_count--;
485                 cfs_list_del(&fmd->fmd_list);
486                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
487         }
488 }
489
490 /* drop fmd reference, free it if last ref */
491 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
492 {
493         struct filter_export_data *fed;
494
495         if (fmd == NULL)
496                 return;
497
498         fed = &exp->exp_filter_data;
499         cfs_spin_lock(&fed->fed_lock);
500         filter_fmd_put_nolock(fed, fmd); /* caller reference */
501         cfs_spin_unlock(&fed->fed_lock);
502 }
503
504 /* expire entries from the end of the list if there are too many
505  * or they are too old */
506 static void filter_fmd_expire_nolock(struct filter_obd *filter,
507                                      struct filter_export_data *fed,
508                                      struct filter_mod_data *keep)
509 {
510         struct filter_mod_data *fmd, *tmp;
511
512         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
513                 if (fmd == keep)
514                         break;
515
516                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
517                     fed->fed_mod_count < filter->fo_fmd_max_num)
518                         break;
519
520                 cfs_list_del_init(&fmd->fmd_list);
521                 filter_fmd_put_nolock(fed, fmd); /* list reference */
522         }
523 }
524
525 void filter_fmd_expire(struct obd_export *exp)
526 {
527         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
528         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
529                                  &exp->exp_filter_data, NULL);
530         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
531 }
532
533 /* find specified objid, group in export fmd list.
534  * caller must hold fed_lock and take fmd reference itself */
535 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
536                                                 struct filter_export_data *fed,
537                                                 obd_id objid, obd_seq group)
538 {
539         struct filter_mod_data *found = NULL, *fmd;
540
541         LASSERT_SPIN_LOCKED(&fed->fed_lock);
542
543         cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
544                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
545                         found = fmd;
546                         cfs_list_del(&fmd->fmd_list);
547                         cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
548                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
549                         break;
550                 }
551         }
552
553         filter_fmd_expire_nolock(filter, fed, found);
554
555         return found;
556 }
557
558 /* Find fmd based on objid and group, or return NULL if not found. */
559 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
560                                         obd_id objid, obd_seq group)
561 {
562         struct filter_mod_data *fmd;
563
564         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
565         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
566                                      &exp->exp_filter_data, objid, group);
567         if (fmd)
568                 fmd->fmd_refcount++;    /* caller reference */
569         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
570
571         return fmd;
572 }
573
574 /* Find fmd based on objid and group, or create a new one if none is found.
575  * It is possible for this function to return NULL under memory pressure,
576  * or if objid = 0 is passed (which will only cause old entries to expire).
577  * Currently this is not fatal because any fmd state is transient and
578  * may also be freed when it gets sufficiently old. */
579 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
580                                        obd_id objid, obd_seq group)
581 {
582         struct filter_export_data *fed = &exp->exp_filter_data;
583         struct filter_mod_data *found = NULL, *fmd_new = NULL;
584
585         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
586
587         cfs_spin_lock(&fed->fed_lock);
588         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
589         if (fmd_new) {
590                 if (found == NULL) {
591                         cfs_list_add_tail(&fmd_new->fmd_list,
592                                           &fed->fed_mod_list);
593                         fmd_new->fmd_id = objid;
594                         fmd_new->fmd_gr = group;
595                         fmd_new->fmd_refcount++;   /* list reference */
596                         found = fmd_new;
597                         fed->fed_mod_count++;
598                 } else {
599                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
600                 }
601         }
602         if (found) {
603                 found->fmd_refcount++;          /* caller reference */
604                 found->fmd_expire = jiffies +
605                         exp->exp_obd->u.filter.fo_fmd_max_age;
606         }
607
608         cfs_spin_unlock(&fed->fed_lock);
609
610         return found;
611 }
612
613 #ifdef DO_FMD_DROP
614 /* drop fmd list reference so it will disappear when last reference is put.
615  * This isn't so critical because it would in fact only affect the one client
616  * that is doing the unlink and at worst we have an stale entry referencing
617  * an object that should never be used again. */
618 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
619 {
620         struct filter_mod_data *found = NULL;
621
622         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
623         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
624         if (found) {
625                 cfs_list_del_init(&found->fmd_list);
626                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
627         }
628         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
629 }
630 #else
631 #define filter_fmd_drop(exp, objid, group)
632 #endif
633
634 /* remove all entries from fmd list */
635 static void filter_fmd_cleanup(struct obd_export *exp)
636 {
637         struct filter_export_data *fed = &exp->exp_filter_data;
638         struct filter_mod_data *fmd = NULL, *tmp;
639
640         cfs_spin_lock(&fed->fed_lock);
641         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
642                 cfs_list_del_init(&fmd->fmd_list);
643                 filter_fmd_put_nolock(fed, fmd);
644         }
645         cfs_spin_unlock(&fed->fed_lock);
646 }
647
648 static int filter_init_export(struct obd_export *exp)
649 {
650         int rc;
651         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
652         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
653
654         cfs_spin_lock(&exp->exp_lock);
655         exp->exp_connecting = 1;
656         cfs_spin_unlock(&exp->exp_lock);
657         rc = lut_client_alloc(exp);
658         if (rc == 0)
659                 rc = ldlm_init_export(exp);
660
661         return rc;
662 }
663
664 static int filter_free_server_data(struct obd_device_target *obt)
665 {
666         lut_fini(NULL, obt->obt_lut);
667         OBD_FREE_PTR(obt->obt_lut);
668         return 0;
669 }
670
671 /* assumes caller is already in kernel ctxt */
672 int filter_update_server_data(struct obd_device *obd)
673 {
674         struct file *filp = obd->u.obt.obt_rcvd_filp;
675         struct lr_server_data *lsd = class_server_data(obd);
676         loff_t off = 0;
677         int rc;
678         ENTRY;
679
680         CDEBUG(D_INODE, "server uuid      : %s\n", lsd->lsd_uuid);
681         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
682                le64_to_cpu(lsd->lsd_last_transno));
683         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
684                le64_to_cpu(lsd->lsd_mount_count));
685
686         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
687         if (rc)
688                 CERROR("error writing lr_server_data: rc = %d\n", rc);
689
690         RETURN(rc);
691 }
692
693 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
694                              int force_sync)
695 {
696         struct filter_obd *filter = &obd->u.filter;
697         __u64 tmp;
698         loff_t off = 0;
699         int rc;
700         ENTRY;
701
702         if (filter->fo_last_objid_files[group] == NULL) {
703                 CERROR("Object seq "LPU64" not fully setup; not updating "
704                        "last_objid\n", group);
705                 RETURN(-EINVAL);
706         }
707
708         CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
709                obd->obd_name, filter->fo_last_objids[group], group);
710
711         tmp = cpu_to_le64(filter->fo_last_objids[group]);
712         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
713                                  &tmp, sizeof(tmp), &off, force_sync);
714         if (rc)
715                 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
716                        group, rc);
717         RETURN(rc);
718 }
719 extern int ost_handle(struct ptlrpc_request *req);
720 /* assumes caller has already in kernel ctxt */
721 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
722 {
723         struct filter_obd *filter = &obd->u.filter;
724         struct lr_server_data *lsd;
725         struct lsd_client_data *lcd = NULL;
726         struct inode *inode = filp->f_dentry->d_inode;
727         unsigned long last_rcvd_size = i_size_read(inode);
728         struct lu_target *lut;
729         __u64 mount_count;
730         __u32 start_epoch;
731         int cl_idx;
732         loff_t off = 0;
733         int rc;
734
735         /* ensure padding in the struct is the correct size */
736         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
737                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
738         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
739                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
740
741         /* allocate and initialize lu_target */
742         OBD_ALLOC_PTR(lut);
743         if (lut == NULL)
744                 RETURN(-ENOMEM);
745         rc = lut_init(NULL, lut, obd, NULL);
746         if (rc)
747                 GOTO(err_lut, rc);
748         lsd = class_server_data(obd);
749         if (last_rcvd_size == 0) {
750                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
751
752                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
753                 lsd->lsd_last_transno = 0;
754                 mount_count = lsd->lsd_mount_count = 0;
755                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
756                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
757                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
758                 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
759                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
760                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
761                  * MDS first connects and assigns the OST index number. */
762                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
763                                                         OBD_INCOMPAT_OST);
764         } else {
765                 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
766                 if (rc) {
767                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
768                                LAST_RCVD, rc);
769                         GOTO(err_lut, rc);
770                 }
771                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
772                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
773                                            "using the wrong disk %s. Were the "
774                                            "/dev/ assignments rearranged?\n",
775                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
776                         GOTO(err_lut, rc = -EINVAL);
777                 }
778                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
779                 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
780                 /* COMPAT_146 */
781                 /* Assume old last_rcvd format unless I_C_LR is set */
782                 if (!(lsd->lsd_feature_incompat &
783                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
784                         lsd->lsd_last_transno = lsd->lsd_compat14;
785                 /* end COMPAT_146 */
786                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
787                  * MDS first connects and assigns the OST index number. */
788                 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
789                                                          OBD_INCOMPAT_OST);
790         }
791
792         if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
793                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
794                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
795                        ~FILTER_INCOMPAT_SUPP);
796                 GOTO(err_lut, rc = -EINVAL);
797         }
798         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
799                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
800                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
801                        ~FILTER_ROCOMPAT_SUPP);
802                 /* Do something like remount filesystem read-only */
803                 GOTO(err_lut, rc = -EINVAL);
804         }
805
806         start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
807
808         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
809                obd->obd_name, start_epoch);
810         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
811                obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
812         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
813                obd->obd_name, mount_count + 1);
814         CDEBUG(D_INODE, "%s: server data size: %u\n",
815                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
816         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
817                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
818         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
819                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
820         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
821                obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
822         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
823                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
824                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
825                 le16_to_cpu(lsd->lsd_client_size));
826
827         if (!obd->obd_replayable) {
828                 CWARN("%s: recovery support OFF\n", obd->obd_name);
829                 GOTO(out, rc = 0);
830         }
831
832         OBD_ALLOC_PTR(lcd);
833         if (!lcd)
834                 GOTO(err_client, rc = -ENOMEM);
835
836         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
837              off < last_rcvd_size; cl_idx++) {
838                 __u64 last_rcvd;
839                 struct obd_export *exp;
840                 struct filter_export_data *fed;
841
842                 /* Don't assume off is incremented properly by
843                  * fsfilt_read_record(), in case sizeof(*lcd)
844                  * isn't the same as lsd->lsd_client_size.  */
845                 off = le32_to_cpu(lsd->lsd_client_start) +
846                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
847                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
848                 if (rc) {
849                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
850                                LAST_RCVD, cl_idx, off, rc);
851                         break; /* read error shouldn't cause startup to fail */
852                 }
853
854                 if (lcd->lcd_uuid[0] == '\0') {
855                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
856                                cl_idx);
857                         continue;
858                 }
859
860                 check_lcd(obd->obd_name, cl_idx, lcd);
861
862                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
863
864                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
865                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
866                        last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
867
868                 /* These exports are cleaned up by filter_disconnect(), so they
869                  * need to be set up like real exports as filter_connect() does.
870                  */
871                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
872                 if (IS_ERR(exp)) {
873                         if (PTR_ERR(exp) == -EALREADY) {
874                                 /* export already exists, zero out this one */
875                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
876                                 continue;
877                         }
878                         OBD_FREE_PTR(lcd);
879                         GOTO(err_client, rc = PTR_ERR(exp));
880                 }
881
882                 fed = &exp->exp_filter_data;
883                 *fed->fed_ted.ted_lcd = *lcd;
884                 fed->fed_group = 0; /* will be assigned at connect */
885                 filter_export_stats_init(obd, exp, NULL);
886                 rc = filter_client_add(obd, exp, cl_idx);
887                 /* can't fail for existing client */
888                 LASSERTF(rc == 0, "rc = %d\n", rc);
889
890                 /* VBR: set export last committed */
891                 exp->exp_last_committed = last_rcvd;
892                 cfs_spin_lock(&exp->exp_lock);
893                 exp->exp_connecting = 0;
894                 exp->exp_in_recovery = 0;
895                 cfs_spin_unlock(&exp->exp_lock);
896                 obd->obd_max_recoverable_clients++;
897                 class_export_put(exp);
898
899                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
900                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
901         }
902         OBD_FREE_PTR(lcd);
903
904         obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
905 out:
906         obd->u.obt.obt_mount_count = mount_count + 1;
907         lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
908
909         /* save it, so mount count and last_transno is current */
910         rc = filter_update_server_data(obd);
911         if (rc)
912                 GOTO(err_client, rc);
913
914         RETURN(0);
915
916 err_client:
917         class_disconnect_exports(obd);
918 err_lut:
919         filter_free_server_data(&obd->u.obt);
920         RETURN(rc);
921 }
922
923 static int filter_cleanup_groups(struct obd_device *obd)
924 {
925         struct filter_obd *filter = &obd->u.filter;
926         struct file *filp;
927         struct dentry *dentry;
928         int i, j;
929         ENTRY;
930
931         if (filter->fo_dentry_O_groups != NULL) {
932                 for (i = 0; i < filter->fo_group_count; i++) {
933                         dentry = filter->fo_dentry_O_groups[i];
934                         if (dentry != NULL)
935                                 f_dput(dentry);
936                 }
937                 OBD_FREE(filter->fo_dentry_O_groups,
938                          filter->fo_group_count *
939                          sizeof(*filter->fo_dentry_O_groups));
940                 filter->fo_dentry_O_groups = NULL;
941         }
942         if (filter->fo_last_objid_files != NULL) {
943                 for (i = 0; i < filter->fo_group_count; i++) {
944                         filp = filter->fo_last_objid_files[i];
945                         if (filp != NULL)
946                                 filp_close(filp, 0);
947                 }
948                 OBD_FREE(filter->fo_last_objid_files,
949                          filter->fo_group_count *
950                          sizeof(*filter->fo_last_objid_files));
951                 filter->fo_last_objid_files = NULL;
952         }
953         if (filter->fo_dentry_O_sub != NULL) {
954                 for (i = 0; i < filter->fo_group_count; i++) {
955                         for (j = 0; j < filter->fo_subdir_count; j++) {
956                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
957                                 if (dentry != NULL)
958                                         f_dput(dentry);
959                         }
960                 }
961                 OBD_FREE(filter->fo_dentry_O_sub,
962                          filter->fo_group_count *
963                          sizeof(*filter->fo_dentry_O_sub));
964                 filter->fo_dentry_O_sub = NULL;
965         }
966         if (filter->fo_last_objids != NULL) {
967                 OBD_FREE(filter->fo_last_objids,
968                          filter->fo_group_count *
969                          sizeof(*filter->fo_last_objids));
970                 filter->fo_last_objids = NULL;
971         }
972         if (filter->fo_dentry_O != NULL) {
973                 f_dput(filter->fo_dentry_O);
974                 filter->fo_dentry_O = NULL;
975         }
976         RETURN(0);
977 }
978
979 static int filter_update_last_group(struct obd_device *obd, int group)
980 {
981         struct filter_obd *filter = &obd->u.filter;
982         struct file *filp = NULL;
983         int last_group = 0, rc;
984         loff_t off = 0;
985         ENTRY;
986
987         if (group <= filter->fo_committed_group)
988                 RETURN(0);
989
990         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
991         if (IS_ERR(filp)) {
992                 rc = PTR_ERR(filp);
993                 filp = NULL;
994                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
995                 GOTO(cleanup, rc);
996         }
997
998         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
999         if (rc) {
1000                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1001                 GOTO(cleanup, rc);
1002         }
1003
1004         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1005                obd->obd_name, last_group, group);
1006
1007         off = 0;
1008         last_group = group;
1009         /* must be sync: bXXXX */
1010         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1011         if (rc) {
1012                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1013                 GOTO(cleanup, rc);
1014         }
1015
1016         filter->fo_committed_group = group;
1017 cleanup:
1018         if (filp)
1019                 filp_close(filp, 0);
1020         RETURN(rc);
1021 }
1022
1023 static int filter_read_group_internal(struct obd_device *obd, int group,
1024                                       int create)
1025 {
1026         struct filter_obd *filter = &obd->u.filter;
1027         __u64 *new_objids = NULL;
1028         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1029         struct dentry **new_groups = NULL;
1030         struct file **new_files = NULL;
1031         struct dentry *dentry;
1032         struct file *filp;
1033         int old_count = filter->fo_group_count, rc, stage = 0, i;
1034         char name[25];
1035         __u64 last_objid;
1036         loff_t off = 0;
1037         int len = group + 1;
1038
1039         snprintf(name, 24, "%d", group);
1040         name[24] = '\0';
1041
1042         if (!create) {
1043                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1044                                            strlen(name));
1045                 if (IS_ERR(dentry)) {
1046                         CERROR("Cannot lookup expected object group %d: %ld\n",
1047                                group, PTR_ERR(dentry));
1048                         RETURN(PTR_ERR(dentry));
1049                 }
1050         } else {
1051                 dentry = simple_mkdir(filter->fo_dentry_O,
1052                                       obd->u.obt.obt_vfsmnt, name, 0700, 1);
1053                 if (IS_ERR(dentry)) {
1054                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1055                                PTR_ERR(dentry));
1056                         RETURN(PTR_ERR(dentry));
1057                 }
1058         }
1059         stage = 1;
1060
1061         snprintf(name, 24, "O/%d/LAST_ID", group);
1062         name[24] = '\0';
1063         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1064         if (IS_ERR(filp)) {
1065                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1066                 GOTO(cleanup, rc = PTR_ERR(filp));
1067         }
1068         stage = 2;
1069
1070         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1071         if (rc) {
1072                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1073                 GOTO(cleanup, rc);
1074         }
1075
1076         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1077                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1078                 if (tmp_subdirs == NULL)
1079                         GOTO(cleanup, rc = -ENOMEM);
1080                 stage = 3;
1081
1082                 for (i = 0; i < filter->fo_subdir_count; i++) {
1083                         char dir[20];
1084                         snprintf(dir, sizeof(dir), "d%u", i);
1085
1086                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1087                                                               obd->u.obt.obt_vfsmnt,
1088                                                               dir, 0700, 1);
1089                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1090                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1091                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1092                                        group, dir, rc);
1093                                 GOTO(cleanup, rc);
1094                         }
1095
1096                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1097                                tmp_subdirs->dentry[i]);
1098                 }
1099         }
1100
1101         /* 'group' is an index; we need an array of length 'group + 1' */
1102         if (group + 1 > old_count) {
1103                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1104                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1105                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1106                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1107                 stage = 4;
1108                 if (new_objids == NULL || new_subdirs == NULL ||
1109                     new_groups == NULL || new_files == NULL)
1110                         GOTO(cleanup, rc = -ENOMEM);
1111
1112                 if (old_count) {
1113                         memcpy(new_objids, filter->fo_last_objids,
1114                                old_count * sizeof(*new_objids));
1115                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1116                                old_count * sizeof(*new_subdirs));
1117                         memcpy(new_groups, filter->fo_dentry_O_groups,
1118                                old_count * sizeof(*new_groups));
1119                         memcpy(new_files, filter->fo_last_objid_files,
1120                                old_count * sizeof(*new_files));
1121
1122                         OBD_FREE(filter->fo_last_objids,
1123                                  old_count * sizeof(*new_objids));
1124                         OBD_FREE(filter->fo_dentry_O_sub,
1125                                  old_count * sizeof(*new_subdirs));
1126                         OBD_FREE(filter->fo_dentry_O_groups,
1127                                  old_count * sizeof(*new_groups));
1128                         OBD_FREE(filter->fo_last_objid_files,
1129                                  old_count * sizeof(*new_files));
1130                 }
1131                 filter->fo_last_objids = new_objids;
1132                 filter->fo_dentry_O_sub = new_subdirs;
1133                 filter->fo_dentry_O_groups = new_groups;
1134                 filter->fo_last_objid_files = new_files;
1135                 filter->fo_group_count = len;
1136         }
1137
1138         filter->fo_dentry_O_groups[group] = dentry;
1139         filter->fo_last_objid_files[group] = filp;
1140         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1141                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1142                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1143         }
1144
1145         filter_update_last_group(obd, group);
1146
1147         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1148                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1149                 rc = filter_update_last_objid(obd, group, 1);
1150                 RETURN(rc);
1151         }
1152
1153         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1154         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1155                obd->obd_name, group, last_objid);
1156         RETURN(0);
1157  cleanup:
1158         switch (stage) {
1159         case 4:
1160                 if (new_objids != NULL)
1161                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1162                 if (new_subdirs != NULL)
1163                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1164                 if (new_groups != NULL)
1165                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1166                 if (new_files != NULL)
1167                         OBD_FREE(new_files, len * sizeof(*new_files));
1168         case 3:
1169                 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1170                         for (i = 0; i < filter->fo_subdir_count; i++) {
1171                                 if (tmp_subdirs->dentry[i] != NULL)
1172                                         dput(tmp_subdirs->dentry[i]);
1173                         }
1174                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1175                 }
1176         case 2:
1177                 filp_close(filp, 0);
1178         case 1:
1179                 dput(dentry);
1180         }
1181         RETURN(rc);
1182 }
1183
1184 static int filter_read_groups(struct obd_device *obd, int last_group,
1185                               int create)
1186 {
1187         struct filter_obd *filter = &obd->u.filter;
1188         int old_count, group, rc = 0;
1189
1190         cfs_down(&filter->fo_init_lock);
1191         old_count = filter->fo_group_count;
1192         for (group = old_count; group <= last_group; group++) {
1193                 rc = filter_read_group_internal(obd, group, create);
1194                 if (rc != 0)
1195                         break;
1196         }
1197         cfs_up(&filter->fo_init_lock);
1198         return rc;
1199 }
1200
1201 /* FIXME: object groups */
1202 static int filter_prep_groups(struct obd_device *obd)
1203 {
1204         struct filter_obd *filter = &obd->u.filter;
1205         struct dentry *O_dentry;
1206         struct file *filp;
1207         int    last_group, rc = 0, cleanup_phase = 0;
1208         loff_t off = 0;
1209         ENTRY;
1210
1211         O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1212                                 "O", 0700, 1);
1213         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1214         if (IS_ERR(O_dentry)) {
1215                 rc = PTR_ERR(O_dentry);
1216                 CERROR("cannot open/create O: rc = %d\n", rc);
1217                 GOTO(cleanup, rc);
1218         }
1219         filter->fo_dentry_O = O_dentry;
1220         cleanup_phase = 1; /* O_dentry */
1221
1222         /* we have to initialize all groups before first connections from
1223          * clients because they may send create/destroy for any group -bzzz */
1224         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1225         if (IS_ERR(filp)) {
1226                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1227                 GOTO(cleanup, rc = PTR_ERR(filp));
1228         }
1229         cleanup_phase = 2; /* filp */
1230
1231         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1232         if (rc) {
1233                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1234                 GOTO(cleanup, rc);
1235         }
1236
1237         if (off == 0)
1238                 last_group = FID_SEQ_OST_MDT0;
1239
1240         CWARN("%s: initialize groups [%u,%u]\n", obd->obd_name,
1241               FID_SEQ_OST_MDT0, last_group);
1242         filter->fo_committed_group = last_group;
1243         rc = filter_read_groups(obd, last_group, 1);
1244         if (rc)
1245                 GOTO(cleanup, rc);
1246
1247         filp_close(filp, 0);
1248         RETURN(0);
1249
1250  cleanup:
1251         switch (cleanup_phase) {
1252         case 2:
1253                 filp_close(filp, 0);
1254         case 1:
1255                 filter_cleanup_groups(obd);
1256                 f_dput(filter->fo_dentry_O);
1257                 filter->fo_dentry_O = NULL;
1258         default:
1259                 break;
1260         }
1261         return rc;
1262
1263 }
1264
1265 /* setup the object store with correct subdirectories */
1266 static int filter_prep(struct obd_device *obd)
1267 {
1268         struct lvfs_run_ctxt saved;
1269         struct filter_obd *filter = &obd->u.filter;
1270         struct file *file;
1271         struct inode *inode;
1272         int rc = 0;
1273         ENTRY;
1274
1275         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1276         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1277         if (!file || IS_ERR(file)) {
1278                 rc = PTR_ERR(file);
1279                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1280                        LAST_RCVD, rc);
1281                 GOTO(out, rc);
1282         }
1283         obd->u.obt.obt_rcvd_filp = file;
1284         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1285                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1286                        file->f_dentry->d_inode->i_mode);
1287                 GOTO(err_filp, rc = -ENOENT);
1288         }
1289
1290         inode = file->f_dentry->d_parent->d_inode;
1291         /* We use i_op->unlink directly in filter_vfs_unlink() */
1292         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1293                 CERROR("%s: filesystem does not support create/unlink ops\n",
1294                        obd->obd_name);
1295                 GOTO(err_filp, rc = -EOPNOTSUPP);
1296         }
1297
1298         rc = filter_init_server_data(obd, file);
1299         if (rc) {
1300                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1301                 GOTO(err_filp, rc);
1302         }
1303         LASSERT(obd->u.obt.obt_lut);
1304         target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1305
1306         /* open and create health check io file*/
1307         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1308         if (IS_ERR(file)) {
1309                 rc = PTR_ERR(file);
1310                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1311                        HEALTH_CHECK, rc);
1312                 GOTO(err_server_data, rc);
1313         }
1314         filter->fo_obt.obt_health_check_filp = file;
1315         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1316                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1317                        file->f_dentry->d_inode->i_mode);
1318                 GOTO(err_health_check, rc = -ENOENT);
1319         }
1320         rc = lvfs_check_io_health(obd, file);
1321         if (rc)
1322                 GOTO(err_health_check, rc);
1323
1324         rc = filter_prep_groups(obd);
1325         if (rc)
1326                 GOTO(err_health_check, rc);
1327 out:
1328         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1329
1330         return(rc);
1331
1332 err_health_check:
1333         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1334                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1335         filter->fo_obt.obt_health_check_filp = NULL;
1336 err_server_data:
1337         target_recovery_fini(obd);
1338         filter_free_server_data(&obd->u.obt);
1339 err_filp:
1340         if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1341                 CERROR("can't close %s after error\n", LAST_RCVD);
1342         obd->u.obt.obt_rcvd_filp = NULL;
1343         goto out;
1344 }
1345
1346 /* cleanup the filter: write last used object id to status file */
1347 static void filter_post(struct obd_device *obd)
1348 {
1349         struct lvfs_run_ctxt saved;
1350         struct filter_obd *filter = &obd->u.filter;
1351         int rc, i;
1352
1353         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1354          * best to start a transaction with h_sync, because we removed this
1355          * from lastobjid */
1356
1357         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1358         rc = filter_update_server_data(obd);
1359         if (rc)
1360                 CERROR("error writing server data: rc = %d\n", rc);
1361
1362         for (i = 0; i < filter->fo_group_count; i++) {
1363                 rc = filter_update_last_objid(obd, i,
1364                                 (i == filter->fo_group_count - 1));
1365                 if (rc)
1366                         CERROR("error writing group %d lastobjid: rc = %d\n",
1367                                i, rc);
1368         }
1369
1370         rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1371         obd->u.obt.obt_rcvd_filp = NULL;
1372         if (rc)
1373                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1374
1375         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1376         filter->fo_obt.obt_health_check_filp = NULL;
1377         if (rc)
1378                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1379
1380         filter_cleanup_groups(obd);
1381         filter_free_server_data(&obd->u.obt);
1382         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1383
1384         filter_free_capa_keys(filter);
1385         cleanup_capa_hash(filter->fo_capa_hash);
1386 }
1387
1388 static void filter_set_last_id(struct filter_obd *filter,
1389                                obd_id id, obd_seq group)
1390 {
1391         LASSERT(group <= filter->fo_group_count);
1392
1393         cfs_spin_lock(&filter->fo_objidlock);
1394         filter->fo_last_objids[group] = id;
1395         cfs_spin_unlock(&filter->fo_objidlock);
1396 }
1397
1398 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1399 {
1400         obd_id id;
1401         LASSERT(group <= filter->fo_group_count);
1402         LASSERT(filter->fo_last_objids != NULL);
1403
1404         /* FIXME: object groups */
1405         cfs_spin_lock(&filter->fo_objidlock);
1406         id = filter->fo_last_objids[group];
1407         cfs_spin_unlock(&filter->fo_objidlock);
1408         return id;
1409 }
1410
1411 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1412 {
1413         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1414         return 0;
1415 }
1416
1417 /* We never dget the object parent, so DON'T dput it either */
1418 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1419 {
1420         struct filter_obd *filter = &obd->u.filter;
1421         struct filter_subdirs *subdirs;
1422
1423         if (group >= filter->fo_group_count) /* FIXME: object groups */
1424                 return ERR_PTR(-EBADF);
1425
1426         if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1427                 return filter->fo_dentry_O_groups[group];
1428
1429         subdirs = &filter->fo_dentry_O_sub[group];
1430         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1431 }
1432
1433 /* We never dget the object parent, so DON'T dput it either */
1434 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1435                                   obd_id objid)
1436 {
1437         unsigned long now = jiffies;
1438         struct dentry *dparent = filter_parent(obd, group, objid);
1439         int rc;
1440
1441         if (IS_ERR(dparent))
1442                 return dparent;
1443         if (dparent == NULL)
1444                 return ERR_PTR(-ENOENT);
1445
1446         rc = filter_lock_dentry(obd, dparent);
1447         fsfilt_check_slow(obd, now, "parent lock");
1448         return rc ? ERR_PTR(rc) : dparent;
1449 }
1450
1451 /* We never dget the object parent, so DON'T dput it either */
1452 static void filter_parent_unlock(struct dentry *dparent)
1453 {
1454         UNLOCK_INODE_MUTEX(dparent->d_inode);
1455 }
1456
1457 /* How to get files, dentries, inodes from object id's.
1458  *
1459  * If dir_dentry is passed, the caller has already locked the parent
1460  * appropriately for this operation (normally a write lock).  If
1461  * dir_dentry is NULL, we do a read lock while we do the lookup to
1462  * avoid races with create/destroy and such changing the directory
1463  * internal to the filesystem code. */
1464 struct dentry *filter_fid2dentry(struct obd_device *obd,
1465                                  struct dentry *dir_dentry,
1466                                  obd_seq group, obd_id id)
1467 {
1468         struct dentry *dparent = dir_dentry;
1469         struct dentry *dchild;
1470         char name[32];
1471         int len;
1472         ENTRY;
1473
1474         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1475             obd->u.filter.fo_destroys_in_progress == 0) {
1476                 /* don't fail lookups for orphan recovery, it causes
1477                  * later LBUGs when objects still exist during precreate */
1478                 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1479                 RETURN(ERR_PTR(-ENOENT));
1480         }
1481         if (id == 0) {
1482                 CERROR("fatal: invalid object id 0\n");
1483                 RETURN(ERR_PTR(-ESTALE));
1484         }
1485
1486         len = sprintf(name, LPU64, id);
1487         if (dir_dentry == NULL) {
1488                 dparent = filter_parent_lock(obd, group, id);
1489                 if (IS_ERR(dparent)) {
1490                         CERROR("%s: error getting object "POSTID
1491                                " parent: rc %ld\n", obd->obd_name,
1492                                id, group, PTR_ERR(dparent));
1493                         RETURN(dparent);
1494                 }
1495         }
1496         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1497                dparent->d_name.len, dparent->d_name.name, name);
1498         /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1499         dchild = lookup_one_len(name, dparent, len);
1500         if (dir_dentry == NULL)
1501                 filter_parent_unlock(dparent);
1502         if (IS_ERR(dchild)) {
1503                 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1504                        obd->obd_name, id, group, PTR_ERR(dchild));
1505                 RETURN(dchild);
1506         }
1507
1508         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1509                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1510                        obd->obd_name, id, dchild->d_inode->i_ino);
1511                 f_dput(dchild);
1512                 RETURN(ERR_PTR(-ENOENT));
1513         }
1514
1515         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1516                name, dchild, atomic_read(&dchild->d_count));
1517
1518         LASSERT(atomic_read(&dchild->d_count) > 0);
1519
1520         RETURN(dchild);
1521 }
1522
1523 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1524                                   obd_id group, struct lustre_handle *lockh)
1525 {
1526         int flags = LDLM_AST_DISCARD_DATA, rc;
1527         struct ldlm_res_id res_id;
1528         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1529         ENTRY;
1530
1531         osc_build_res_name(objid, group, &res_id);
1532         /* Tell the clients that the object is gone now and that they should
1533          * throw away any cached pages. */
1534         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1535                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1536                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1537                                     lockh);
1538         if (rc != ELDLM_OK)
1539                 lockh->cookie = 0;
1540         RETURN(rc);
1541 }
1542
1543 static void filter_fini_destroy(struct obd_device *obd,
1544                                 struct lustre_handle *lockh)
1545 {
1546         if (lustre_handle_is_used(lockh))
1547                 ldlm_lock_decref(lockh, LCK_PW);
1548 }
1549
1550 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1551  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1552  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1553  * reverse.  Caller must take i_sem before starting the transaction and we
1554  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1555 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1556                       struct vfsmount *mnt)
1557 {
1558         int rc;
1559         ENTRY;
1560
1561         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1562          * itself we already hold dir->i_mutex for child create/unlink ops */
1563         LASSERT(dentry->d_inode != NULL);
1564         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1565         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1566
1567
1568         /* may_delete() */
1569         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1570                 GOTO(out, rc = -ENOENT);
1571
1572         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1573         if (rc)
1574                 GOTO(out, rc);
1575
1576         if (IS_APPEND(dir))
1577                 GOTO(out, rc = -EPERM);
1578
1579         /* check_sticky() */
1580         if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1581              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1582             IS_IMMUTABLE(dentry->d_inode))
1583                 GOTO(out, rc = -EPERM);
1584
1585         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1586          *       that was done because of journal_start (which is already done
1587          *       here) or some other ordering issue. */
1588         ll_vfs_dq_init(dir);
1589
1590         rc = ll_security_inode_unlink(dir, dentry, mnt);
1591         if (rc)
1592                 GOTO(out, rc);
1593
1594         rc = dir->i_op->unlink(dir, dentry);
1595 out:
1596         /* need to drop i_mutex before we lose inode reference */
1597         UNLOCK_INODE_MUTEX(dentry->d_inode);
1598         if (rc == 0)
1599                 d_delete(dentry);
1600
1601         RETURN(rc);
1602 }
1603
1604 /* Caller must hold LCK_PW on parent and push us into kernel context.
1605  * Caller must hold child i_mutex, we drop it always.
1606  * Caller is also required to ensure that dchild->d_inode exists. */
1607 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1608                                    obd_seq group, struct dentry *dparent,
1609                                    struct dentry *dchild)
1610 {
1611         struct inode *inode = dchild->d_inode;
1612         int rc;
1613
1614         /* There should be 2 references to the inode:
1615          *  1) taken by filter_prepare_destroy
1616          *  2) taken by filter_destroy */
1617         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1618                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1619                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1620                        (unsigned long)inode->i_nlink,
1621                        atomic_read(&inode->i_count));
1622         }
1623
1624         rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1625         if (rc)
1626                 CERROR("error unlinking objid %.*s: rc %d\n",
1627                        dchild->d_name.len, dchild->d_name.name, rc);
1628         return(rc);
1629 }
1630
1631 struct filter_intent_args {
1632         struct ldlm_lock **victim;
1633         __u64 size;
1634         int *liblustre;
1635 };
1636
1637 static enum interval_iter filter_intent_cb(struct interval_node *n,
1638                                            void *args)
1639 {
1640         struct ldlm_interval *node = (struct ldlm_interval *)n;
1641         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1642         __u64 size = arg->size;
1643         struct ldlm_lock **v = arg->victim;
1644         struct ldlm_lock *lck;
1645
1646         /* If the interval is lower than the current file size,
1647          * just break. */
1648         if (interval_high(n) <= size)
1649                 return INTERVAL_ITER_STOP;
1650
1651         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1652                 /* Don't send glimpse ASTs to liblustre clients.
1653                  * They aren't listening for them, and they do
1654                  * entirely synchronous I/O anyways. */
1655                 if (lck->l_export == NULL ||
1656                     lck->l_export->exp_libclient == 1)
1657                         continue;
1658
1659                 if (*arg->liblustre)
1660                         *arg->liblustre = 0;
1661
1662                 if (*v == NULL) {
1663                         *v = LDLM_LOCK_GET(lck);
1664                 } else if ((*v)->l_policy_data.l_extent.start <
1665                            lck->l_policy_data.l_extent.start) {
1666                         LDLM_LOCK_RELEASE(*v);
1667                         *v = LDLM_LOCK_GET(lck);
1668                 }
1669
1670                 /* the same policy group - every lock has the
1671                  * same extent, so needn't do it any more */
1672                 break;
1673         }
1674
1675         return INTERVAL_ITER_CONT;
1676 }
1677
1678 static int filter_intent_policy(struct ldlm_namespace *ns,
1679                                 struct ldlm_lock **lockp, void *req_cookie,
1680                                 ldlm_mode_t mode, int flags, void *data)
1681 {
1682         CFS_LIST_HEAD(rpc_list);
1683         struct ptlrpc_request *req = req_cookie;
1684         struct ldlm_lock *lock = *lockp, *l = NULL;
1685         struct ldlm_resource *res = lock->l_resource;
1686         ldlm_processing_policy policy;
1687         struct ost_lvb *res_lvb, *reply_lvb;
1688         struct ldlm_reply *rep;
1689         ldlm_error_t err;
1690         int idx, rc, tmpflags = 0, only_liblustre = 1;
1691         struct ldlm_interval_tree *tree;
1692         struct filter_intent_args arg;
1693         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1694                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1695                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1696         ENTRY;
1697
1698         policy = ldlm_get_processing_policy(res);
1699         LASSERT(policy != NULL);
1700         LASSERT(req != NULL);
1701
1702         rc = lustre_pack_reply(req, 3, repsize, NULL);
1703         if (rc)
1704                 RETURN(req->rq_status = rc);
1705
1706         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1707         LASSERT(rep != NULL);
1708
1709         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1710                                    sizeof(*reply_lvb));
1711         LASSERT(reply_lvb != NULL);
1712
1713         //fixup_handle_for_resent_req(req, lock, &lockh);
1714
1715         /* Call the extent policy function to see if our request can be
1716          * granted, or is blocked.
1717          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1718          * lock, and should not be granted if the lock will be blocked.
1719          */
1720
1721         LASSERT(ns == ldlm_res_to_ns(res));
1722         lock_res(res);
1723         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1724         check_res_locked(res);
1725
1726         /* FIXME: we should change the policy function slightly, to not make
1727          * this list at all, since we just turn around and free it */
1728         while (!cfs_list_empty(&rpc_list)) {
1729                 struct ldlm_lock *wlock =
1730                         cfs_list_entry(rpc_list.next, struct ldlm_lock,
1731                                        l_cp_ast);
1732                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1733                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1734                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1735                 cfs_list_del_init(&wlock->l_cp_ast);
1736                 LDLM_LOCK_RELEASE(wlock);
1737         }
1738
1739         /* The lock met with no resistance; we're finished. */
1740         if (rc == LDLM_ITER_CONTINUE) {
1741                 /* do not grant locks to the liblustre clients: they cannot
1742                  * handle ASTs robustly.  We need to do this while still
1743                  * holding lr_lock to avoid the lock remaining on the res_link
1744                  * list (and potentially being added to l_pending_list by an
1745                  * AST) when we are going to drop this lock ASAP. */
1746                 if (lock->l_export->exp_libclient ||
1747                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1748                         ldlm_resource_unlink_lock(lock);
1749                         err = ELDLM_LOCK_ABORTED;
1750                 } else {
1751                         err = ELDLM_LOCK_REPLACED;
1752                 }
1753                 unlock_res(res);
1754                 RETURN(err);
1755         }
1756
1757         /* Do not grant any lock, but instead send GL callbacks.  The extent
1758          * policy nicely created a list of all PW locks for us.  We will choose
1759          * the highest of those which are larger than the size in the LVB, if
1760          * any, and perform a glimpse callback. */
1761         res_lvb = res->lr_lvb_data;
1762         LASSERT(res_lvb != NULL);
1763         *reply_lvb = *res_lvb;
1764
1765         /*
1766          * lr_lock guarantees that no new locks are granted, and,
1767          * therefore, that res->lr_lvb_data cannot increase beyond the
1768          * end of already granted lock. As a result, it is safe to
1769          * check against "stale" reply_lvb->lvb_size value without
1770          * res->lr_lvb_sem.
1771          */
1772         arg.size = reply_lvb->lvb_size;
1773         arg.victim = &l;
1774         arg.liblustre = &only_liblustre;
1775         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1776                 tree = &res->lr_itree[idx];
1777                 if (tree->lit_mode == LCK_PR)
1778                         continue;
1779
1780                 interval_iterate_reverse(tree->lit_root,
1781                                          filter_intent_cb, &arg);
1782         }
1783         unlock_res(res);
1784
1785         /* There were no PW locks beyond the size in the LVB; finished. */
1786         if (l == NULL) {
1787                 if (only_liblustre) {
1788                         /* If we discovered a liblustre client with a PW lock,
1789                          * however, the LVB may be out of date!  The LVB is
1790                          * updated only on glimpse (which we don't do for
1791                          * liblustre clients) and cancel (which the client
1792                          * obviously has not yet done).  So if it has written
1793                          * data but kept the lock, the LVB is stale and needs
1794                          * to be updated from disk.
1795                          *
1796                          * Of course, this will all disappear when we switch to
1797                          * taking liblustre locks on the OST. */
1798                         ldlm_res_lvbo_update(res, NULL, 1);
1799                 }
1800                 RETURN(ELDLM_LOCK_ABORTED);
1801         }
1802
1803         /*
1804          * This check is for lock taken in filter_prepare_destroy() that does
1805          * not have l_glimpse_ast set. So the logic is: if there is a lock
1806          * with no l_glimpse_ast set, this object is being destroyed already.
1807          *
1808          * Hence, if you are grabbing DLM locks on the server, always set
1809          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1810          */
1811         if (l->l_glimpse_ast == NULL) {
1812                 /* We are racing with unlink(); just return -ENOENT */
1813                 rep->lock_policy_res1 = -ENOENT;
1814                 goto out;
1815         }
1816
1817         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1818         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1819
1820         lock_res(res);
1821         *reply_lvb = *res_lvb;
1822         unlock_res(res);
1823
1824  out:
1825         LDLM_LOCK_RELEASE(l);
1826
1827         RETURN(ELDLM_LOCK_ABORTED);
1828 }
1829
1830 /*
1831  * per-obd_device iobuf pool.
1832  *
1833  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1834  * allocations in write-path has to be minimized (see bug 5137).
1835  *
1836  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1837  * OST threads (see ost_thread_{init,done}()).
1838  *
1839  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1840  * at the OST layer there are only (potentially) multiple obd_device of type
1841  * unknown at the time of OST thread creation.
1842  *
1843  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1844  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1845  * it's very own iobuf.
1846  *
1847  * Functions below
1848  *
1849  *     filter_kiobuf_pool_init()
1850  *
1851  *     filter_kiobuf_pool_done()
1852  *
1853  *     filter_iobuf_get()
1854  *
1855  * operate on this array. They are "generic" in a sense that they don't depend
1856  * on actual type of iobuf's (the latter depending on Linux kernel version).
1857  */
1858
1859 /*
1860  * destroy pool created by filter_iobuf_pool_init
1861  */
1862 static void filter_iobuf_pool_done(struct filter_obd *filter)
1863 {
1864         struct filter_iobuf **pool;
1865         int i;
1866
1867         ENTRY;
1868
1869         pool = filter->fo_iobuf_pool;
1870         if (pool != NULL) {
1871                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1872                         if (pool[i] != NULL)
1873                                 filter_free_iobuf(pool[i]);
1874                 }
1875                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1876                 filter->fo_iobuf_pool = NULL;
1877         }
1878         EXIT;
1879 }
1880
1881 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1882 {
1883         struct filter_obd       *filter = &obd->u.filter;
1884         struct sptlrpc_rule_set  tmp_rset;
1885         int                      rc;
1886
1887         sptlrpc_rule_set_init(&tmp_rset);
1888         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1889         if (rc) {
1890                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1891                        obd->obd_name, rc);
1892                 return rc;
1893         }
1894
1895         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1896
1897         cfs_write_lock(&filter->fo_sptlrpc_lock);
1898         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1899         filter->fo_sptlrpc_rset = tmp_rset;
1900         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1901
1902         return 0;
1903 }
1904
1905 /*
1906  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1907  */
1908 static int filter_iobuf_pool_init(struct filter_obd *filter)
1909 {
1910         void **pool;
1911
1912         ENTRY;
1913
1914
1915         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1916                       GFP_KERNEL);
1917         if (filter->fo_iobuf_pool == NULL)
1918                 RETURN(-ENOMEM);
1919
1920         filter->fo_iobuf_count = OSS_THREADS_MAX;
1921
1922         RETURN(0);
1923 }
1924
1925 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1926  * many threads there will be so we allocate a large empty array and only
1927  * fill in those slots that are actually in use.
1928  * If we haven't allocated a pool entry for this thread before, do so now. */
1929 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1930 {
1931         int thread_id                    = (oti && oti->oti_thread) ?
1932                                            oti->oti_thread->t_id : -1;
1933         struct filter_iobuf  *pool       = NULL;
1934         struct filter_iobuf **pool_place = NULL;
1935
1936         if (thread_id >= 0) {
1937                 LASSERT(thread_id < filter->fo_iobuf_count);
1938                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1939         }
1940
1941         if (unlikely(pool == NULL)) {
1942                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1943                                           PTLRPC_MAX_BRW_PAGES);
1944                 if (pool_place != NULL)
1945                         *pool_place = pool;
1946         }
1947
1948         return pool;
1949 }
1950
1951 /* mount the file system (secretly).  lustre_cfg parameters are:
1952  * 1 = device
1953  * 2 = fstype
1954  * 3 = flags: failover=f, failout=n
1955  * 4 = mount options
1956  */
1957 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1958                         void *option)
1959 {
1960         struct filter_obd *filter = &obd->u.filter;
1961         struct vfsmount *mnt;
1962         struct lustre_mount_info *lmi;
1963         struct obd_uuid uuid;
1964         __u8 *uuid_ptr;
1965         char *str, *label;
1966         char ns_name[48];
1967         struct request_queue *q;
1968         int rc, i;
1969         ENTRY;
1970
1971         if (lcfg->lcfg_bufcount < 3 ||
1972             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1973             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1974                 RETURN(-EINVAL);
1975
1976         lmi = server_get_mount(obd->obd_name);
1977         if (lmi) {
1978                 /* We already mounted in lustre_fill_super.
1979                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1980                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1981                 mnt = lmi->lmi_mnt;
1982                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1983
1984                 /* gets recovery timeouts from mount data */
1985                 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1986                         obd->obd_recovery_timeout =
1987                                 lsi->lsi_lmd->lmd_recovery_time_soft;
1988                 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1989                         obd->obd_recovery_time_hard =
1990                                 lsi->lsi_lmd->lmd_recovery_time_hard;
1991         } else {
1992                 /* old path - used by lctl */
1993                 CERROR("Using old MDS mount method\n");
1994                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1995                                     MS_NOATIME|MS_NODIRATIME,
1996                                     lustre_cfg_string(lcfg, 1), option);
1997                 if (IS_ERR(mnt)) {
1998                         rc = PTR_ERR(mnt);
1999                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2000                                            lustre_cfg_string(lcfg, 1), rc);
2001                         RETURN(rc);
2002                 }
2003
2004                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2005         }
2006         if (IS_ERR(obd->obd_fsops))
2007                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2008
2009         rc = filter_iobuf_pool_init(filter);
2010         if (rc != 0)
2011                 GOTO(err_ops, rc);
2012
2013         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2014                 CERROR("%s: Underlying device is marked as read-only. "
2015                        "Setup failed\n", obd->obd_name);
2016                 GOTO(err_ops, rc = -EROFS);
2017         }
2018
2019         /* failover is the default */
2020         obd->obd_replayable = 1;
2021
2022         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2023                 str = lustre_cfg_string(lcfg, 3);
2024                 if (strchr(str, 'n')) {
2025                         CWARN("%s: recovery disabled\n", obd->obd_name);
2026                         obd->obd_replayable = 0;
2027                 }
2028         }
2029
2030         obd->u.obt.obt_vfsmnt = mnt;
2031         obd->u.obt.obt_sb = mnt->mnt_sb;
2032         obd->u.obt.obt_magic = OBT_MAGIC;
2033         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2034         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2035
2036         fsfilt_setup(obd, obd->u.obt.obt_sb);
2037
2038         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2039         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2040         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2041         obd->obd_lvfs_ctxt.fs = get_ds();
2042         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2043
2044         cfs_init_mutex(&filter->fo_init_lock);
2045         filter->fo_committed_group = 0;
2046         filter->fo_destroys_in_progress = 0;
2047         for (i = 0; i < 32; i++)
2048                 cfs_sema_init(&filter->fo_create_locks[i], 1);
2049
2050         cfs_spin_lock_init(&filter->fo_objidlock);
2051         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2052         cfs_sema_init(&filter->fo_alloc_lock, 1);
2053         init_brw_stats(&filter->fo_filter_stats);
2054         cfs_spin_lock_init(&filter->fo_flags_lock);
2055         filter->fo_read_cache = 1; /* enable read-only cache by default */
2056         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2057         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2058         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2059         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2060         filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2061         filter_slc_set(filter); /* initialize sync on lock cancel */
2062
2063         rc = filter_prep(obd);
2064         if (rc)
2065                 GOTO(err_ops, rc);
2066
2067         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2068         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2069
2070         filter->fo_fl_oss_capa = 1;
2071
2072         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2073         filter->fo_capa_hash = init_capa_hash();
2074         if (filter->fo_capa_hash == NULL)
2075                 GOTO(err_post, rc = -ENOMEM);
2076
2077         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2078         obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2079                                                 LDLM_NAMESPACE_SERVER,
2080                                                 LDLM_NAMESPACE_GREEDY,
2081                                                 LDLM_NS_TYPE_OST);
2082         if (obd->obd_namespace == NULL)
2083                 GOTO(err_post, rc = -ENOMEM);
2084         obd->obd_namespace->ns_lvbp = obd;
2085         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2086         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2087
2088         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2089                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2090
2091         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2092         if (rc) {
2093                 CERROR("failed to setup llogging subsystems\n");
2094                 GOTO(err_post, rc);
2095         }
2096
2097         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2098         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2099         /* do this after llog being initialized */
2100         filter_adapt_sptlrpc_conf(obd, 1);
2101
2102         rc = lquota_setup(filter_quota_interface_ref, obd);
2103         if (rc)
2104                 GOTO(err_post, rc);
2105
2106         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2107         if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2108             queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2109                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2110                               "for larger I/O requests: max_sectors = %u "
2111                               "could be up to max_hw_sectors=%u\n",
2112                               obd->obd_name, mnt->mnt_sb->s_id,
2113                               queue_max_sectors(q), queue_max_hw_sectors(q));
2114
2115         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2116         if (uuid_ptr != NULL) {
2117                 class_uuid_unparse(uuid_ptr, &uuid);
2118                 str = uuid.uuid;
2119         } else {
2120                 str = "no UUID";
2121         }
2122
2123         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2124         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2125                       obd->obd_name, label ?: str, lmi ? "on " : "",
2126                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2127                       obd->obd_replayable ? "enabled" : "disabled");
2128
2129         if (obd->obd_recovering)
2130                 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2131                               "or until %d client%s reconnect%s\n",
2132                               obd->obd_name,
2133                               obd->obd_recovery_timeout / 60,
2134                               obd->obd_recovery_timeout % 60,
2135                               obd->obd_max_recoverable_clients,
2136                               (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2137                               (obd->obd_max_recoverable_clients == 1) ? "s": "");
2138
2139
2140         RETURN(0);
2141
2142 err_post:
2143         filter_post(obd);
2144 err_ops:
2145         fsfilt_put_ops(obd->obd_fsops);
2146         filter_iobuf_pool_done(filter);
2147 err_mntput:
2148         server_put_mount(obd->obd_name, mnt);
2149         obd->u.obt.obt_sb = 0;
2150         return rc;
2151 }
2152
2153 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2154 {
2155         struct lprocfs_static_vars lvars;
2156         cfs_proc_dir_entry_t *entry;
2157         unsigned long addr;
2158         struct page *page;
2159         int rc;
2160         ENTRY;
2161
2162         CLASSERT(offsetof(struct obd_device, u.obt) ==
2163                  offsetof(struct obd_device, u.filter.fo_obt));
2164
2165         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2166                 RETURN(-EINVAL);
2167
2168         /* lprocfs must be setup before the filter so state can be safely added
2169          * to /proc incrementally as the filter is setup */
2170         lprocfs_filter_init_vars(&lvars);
2171         rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2172         if (rc) {
2173                 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2174                        obd->obd_name, rc);
2175                 RETURN(rc);
2176         }
2177
2178         rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2179         if (rc) {
2180                 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2181                        obd->obd_name, rc);
2182                 GOTO(obd_cleanup, rc);
2183         }
2184
2185         /* Init obdfilter private stats here */
2186         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2187                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2188         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2189                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2190         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2191                              LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2192                              "get_page", "usec");
2193         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2194                              LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2195         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2196                              LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2197         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2198                              LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2199         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2200                              LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2201
2202         rc = lproc_filter_attach_seqstat(obd);
2203         if (rc) {
2204                 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2205                 GOTO(free_obd_stats, rc);
2206         }
2207
2208         entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2209         if (IS_ERR(entry)) {
2210                 rc = PTR_ERR(entry);
2211                 CERROR("%s: error %d setting up lprocfs for %s\n",
2212                        obd->obd_name, rc, "exports");
2213                 GOTO(free_obd_stats, rc);
2214         }
2215         obd->obd_proc_exports_entry = entry;
2216
2217         entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2218                                    lprocfs_nid_stats_clear_read,
2219                                    lprocfs_nid_stats_clear_write, obd, NULL);
2220         if (IS_ERR(entry)) {
2221                 rc = PTR_ERR(entry);
2222                 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2223                        obd->obd_name, rc);
2224                 GOTO(free_obd_stats, rc);
2225         }
2226
2227         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2228         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2229         if (!page)
2230                 GOTO(remove_entry_clear, rc = -ENOMEM);
2231         addr = (unsigned long)cfs_page_address(page);
2232         clear_page((void *)addr);
2233         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2234                LUSTRE_CFG_BUFLEN(lcfg, 4));
2235         rc = filter_common_setup(obd, lcfg, (void *)addr);
2236         OBD_PAGE_FREE(page);
2237         if (rc) {
2238                 CERROR("%s: filter_common_setup failed: %d.\n",
2239                        obd->obd_name, rc);
2240                 GOTO(remove_entry_clear, rc);
2241         }
2242
2243         RETURN(0);
2244
2245 remove_entry_clear:
2246         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2247 free_obd_stats:
2248         lprocfs_free_obd_stats(obd);
2249 obd_cleanup:
2250         lprocfs_obd_cleanup(obd);
2251         return rc;
2252 }
2253
2254 static struct llog_operations filter_mds_ost_repl_logops;
2255
2256 static struct llog_operations filter_size_orig_logops = {
2257         .lop_setup   = llog_obd_origin_setup,
2258         .lop_cleanup = llog_obd_origin_cleanup,
2259         .lop_add     = llog_obd_origin_add
2260 };
2261
2262 static int filter_olg_fini(struct obd_llog_group *olg)
2263 {
2264         struct llog_ctxt *ctxt;
2265         int rc = 0, rc2 = 0;
2266         ENTRY;
2267
2268         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2269         if (ctxt)
2270                 rc = llog_cleanup(ctxt);
2271
2272         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2273         if (ctxt) {
2274                 rc2 = llog_cleanup(ctxt);
2275                 if (!rc)
2276                         rc = rc2;
2277         }
2278
2279         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2280         if (ctxt) {
2281                 rc2 = llog_cleanup(ctxt);
2282                 if (!rc)
2283                         rc = rc2;
2284         }
2285
2286         RETURN(rc);
2287 }
2288
2289 static int
2290 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2291                 struct obd_device *tgt)
2292 {
2293         int rc;
2294         ENTRY;
2295
2296         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2297                         &filter_mds_ost_repl_logops);
2298         if (rc)
2299                 GOTO(cleanup, rc);
2300
2301         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2302                         &filter_size_orig_logops);
2303         if (rc)
2304                 GOTO(cleanup, rc);
2305         EXIT;
2306 cleanup:
2307         if (rc)
2308                 filter_olg_fini(olg);
2309         return rc;
2310 }
2311
2312 /**
2313  * Init the default olg, which is embeded in the obd_device, for filter.
2314  */
2315 static int
2316 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2317                         struct obd_device *tgt)
2318 {
2319         struct filter_obd *filter = &obd->u.filter;
2320         struct llog_ctxt *ctxt;
2321         int rc;
2322         ENTRY;
2323
2324         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2325         if (!filter->fo_lcm)
2326                 RETURN(-ENOMEM);
2327
2328         filter_mds_ost_repl_logops = llog_client_ops;
2329         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2330         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2331         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2332
2333         rc = filter_olg_init(obd, olg, tgt);
2334         if (rc)
2335                 GOTO(cleanup_lcm, rc);
2336
2337         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2338                         &llog_lvfs_ops);
2339         if (rc)
2340                 GOTO(cleanup_olg, rc);
2341
2342         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2343         if (!ctxt) {
2344                 CERROR("Can't get ctxt for %p:%x\n", olg,
2345                        LLOG_MDS_OST_REPL_CTXT);
2346                 GOTO(cleanup_olg, rc = -ENODEV);
2347         }
2348         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2349         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2350         llog_ctxt_put(ctxt);
2351
2352         RETURN(0);
2353 cleanup_olg:
2354         filter_olg_fini(olg);
2355 cleanup_lcm:
2356         llog_recov_thread_fini(filter->fo_lcm, 1);
2357         filter->fo_lcm = NULL;
2358         return rc;
2359 }
2360
2361 static int
2362 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2363                  struct obd_device *tgt, int *index)
2364 {
2365         struct filter_obd *filter = &obd->u.filter;
2366         struct llog_ctxt *ctxt;
2367         int rc;
2368         ENTRY;
2369
2370         LASSERT(olg != NULL);
2371         if (olg == &obd->obd_olg)
2372                 return filter_default_olg_init(obd, olg, tgt);
2373
2374         LASSERT(filter->fo_lcm != NULL);
2375         rc = filter_olg_init(obd, olg, tgt);
2376         if (rc)
2377                 RETURN(rc);
2378         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2379         if (!ctxt) {
2380                 CERROR("Can't get ctxt for %p:%x\n", olg,
2381                        LLOG_MDS_OST_REPL_CTXT);
2382                 filter_olg_fini(olg);
2383                 RETURN(-ENODEV);
2384         }
2385         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2386         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2387         llog_ctxt_put(ctxt);
2388         RETURN(rc);
2389 }
2390
2391 static int filter_llog_finish(struct obd_device *obd, int count)
2392 {
2393         struct filter_obd *filter = &obd->u.filter;
2394         struct llog_ctxt *ctxt;
2395         ENTRY;
2396
2397         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2398         if (ctxt) {
2399                 /*
2400                  * Make sure that no cached llcds left in recov_thread.
2401                  * We actually do sync in disconnect time, but disconnect
2402                  * may not come being marked rq_no_resend = 1.
2403                  */
2404                 llog_sync(ctxt, NULL);
2405
2406                 /*
2407                  * Balance class_import_get() in llog_receptor_accept().
2408                  * This is safe to do, as llog is already synchronized
2409                  * and its import may go.
2410                  */
2411                 cfs_mutex_down(&ctxt->loc_sem);
2412                 if (ctxt->loc_imp) {
2413                         class_import_put(ctxt->loc_imp);
2414                         ctxt->loc_imp = NULL;
2415                 }
2416                 cfs_mutex_up(&ctxt->loc_sem);
2417                 llog_ctxt_put(ctxt);
2418         }
2419
2420         if (filter->fo_lcm) {
2421                 cfs_mutex_down(&ctxt->loc_sem);
2422                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2423                 filter->fo_lcm = NULL;
2424                 cfs_mutex_up(&ctxt->loc_sem);
2425         }
2426         RETURN(filter_olg_fini(&obd->obd_olg));
2427 }
2428
2429 /**
2430  * Find the group llog according to group index in the llog group list.
2431  */
2432 static struct obd_llog_group *
2433 filter_find_olg_internal(struct filter_obd *filter, int group)
2434 {
2435         struct obd_llog_group *olg;
2436
2437         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2438         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2439                 if (olg->olg_seq == group)
2440                         RETURN(olg);
2441         }
2442         RETURN(NULL);
2443 }
2444
2445 /**
2446  * Find the group llog according to group index on the filter
2447  */
2448 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2449 {
2450         struct obd_llog_group *olg = NULL;
2451         struct filter_obd *filter;
2452
2453         filter = &obd->u.filter;
2454
2455         if (group == FID_SEQ_LLOG)
2456                 RETURN(&obd->obd_olg);
2457
2458         cfs_spin_lock(&filter->fo_llog_list_lock);
2459         olg = filter_find_olg_internal(filter, group);
2460         cfs_spin_unlock(&filter->fo_llog_list_lock);
2461
2462         RETURN(olg);
2463 }
2464 /**
2465  * Find the llog_group of the filter according to the group. If it can not
2466  * find, create the llog_group, which only happens when mds is being synced
2467  * with OST.
2468  */
2469 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2470 {
2471         struct obd_llog_group *olg = NULL, *olg_new = NULL;
2472         struct filter_obd *filter;
2473         int rc;
2474
2475         filter = &obd->u.filter;
2476
2477         if (group == FID_SEQ_LLOG)
2478                 RETURN(&obd->obd_olg);
2479
2480         OBD_ALLOC_PTR(olg_new);
2481         if (olg_new == NULL)
2482                RETURN(ERR_PTR(-ENOMEM));
2483
2484         cfs_spin_lock(&filter->fo_llog_list_lock);
2485         olg = filter_find_olg_internal(filter, group);
2486         if (olg) {
2487                 if (olg->olg_initializing) {
2488                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2489                 } else {
2490                         GOTO(out_unlock, olg);
2491                 }
2492         } else {
2493                 /* set as the newly allocated one */
2494                 olg = olg_new;
2495                 olg_new = NULL;
2496         }
2497
2498         llog_group_init(olg, group);
2499         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2500         olg->olg_initializing = 1;
2501         cfs_spin_unlock(&filter->fo_llog_list_lock);
2502
2503         rc = obd_llog_init(obd, olg, obd, NULL);
2504         if (rc) {
2505                cfs_spin_lock(&filter->fo_llog_list_lock);
2506                cfs_list_del(&olg->olg_list);
2507                cfs_spin_unlock(&filter->fo_llog_list_lock);
2508                OBD_FREE_PTR(olg);
2509                GOTO(out, olg = ERR_PTR(-ENOMEM));
2510         }
2511         cfs_spin_lock(&filter->fo_llog_list_lock);
2512         olg->olg_initializing = 0;
2513         cfs_spin_unlock(&filter->fo_llog_list_lock);
2514         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2515               obd->obd_name, group, olg);
2516 out:
2517         RETURN(olg);
2518
2519 out_unlock:
2520         cfs_spin_unlock(&filter->fo_llog_list_lock);
2521         if (olg_new)
2522                OBD_FREE_PTR(olg_new);
2523         goto out;
2524 }
2525
2526 static int filter_llog_connect(struct obd_export *exp,
2527                                struct llogd_conn_body *body)
2528 {
2529         struct obd_device *obd = exp->exp_obd;
2530         struct llog_ctxt *ctxt;
2531         struct obd_llog_group *olg;
2532         int rc;
2533         ENTRY;
2534
2535         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2536                obd->obd_name, body->lgdc_logid.lgl_oid,
2537                body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2538
2539         olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2540         if (!olg) {
2541                 CERROR(" %s: can not find olg of group %d\n",
2542                        obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2543                 RETURN(-ENOENT);
2544         }
2545         llog_group_set_export(olg, exp);
2546
2547         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2548         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2549                  body->lgdc_ctxt_idx);
2550
2551         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2552               obd->obd_name, body->lgdc_logid.lgl_oid,
2553               body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2554
2555         cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2556         obd->u.filter.fo_mds_ost_sync = 1;
2557         cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2558         rc = llog_connect(ctxt, &body->lgdc_logid,
2559                           &body->lgdc_gen, NULL);
2560         llog_ctxt_put(ctxt);
2561         if (rc != 0)
2562                 CERROR("failed to connect rc %d idx %d\n", rc,
2563                                 body->lgdc_ctxt_idx);
2564
2565         RETURN(rc);
2566 }
2567
2568 static int filter_llog_preclean(struct obd_device *obd)
2569 {
2570         struct obd_llog_group *olg, *tmp;
2571         struct filter_obd *filter;
2572         cfs_list_t  remove_list;
2573         int rc = 0;
2574         ENTRY;
2575
2576         rc = obd_llog_finish(obd, 0);
2577         if (rc)
2578                 CERROR("failed to cleanup llogging subsystem\n");
2579
2580         filter = &obd->u.filter;
2581         CFS_INIT_LIST_HEAD(&remove_list);
2582
2583         cfs_spin_lock(&filter->fo_llog_list_lock);
2584         while (!cfs_list_empty(&filter->fo_llog_list)) {
2585                 olg = cfs_list_entry(filter->fo_llog_list.next,
2586                                      struct obd_llog_group, olg_list);
2587                 cfs_list_del(&olg->olg_list);
2588                 cfs_list_add(&olg->olg_list, &remove_list);
2589         }
2590         cfs_spin_unlock(&filter->fo_llog_list_lock);
2591
2592         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2593                 cfs_list_del_init(&olg->olg_list);
2594                 rc = filter_olg_fini(olg);
2595                 if (rc)
2596                         CERROR("failed to cleanup llogging subsystem for %u\n",
2597                                olg->olg_seq);
2598                 OBD_FREE_PTR(olg);
2599         }
2600
2601         RETURN(rc);
2602 }
2603
2604 static int filter_precleanup(struct obd_device *obd,
2605                              enum obd_cleanup_stage stage)
2606 {
2607         int rc = 0;
2608         ENTRY;
2609
2610         switch(stage) {
2611         case OBD_CLEANUP_EARLY:
2612                 break;
2613         case OBD_CLEANUP_EXPORTS:
2614                 /* Stop recovery before namespace cleanup. */
2615                 target_recovery_fini(obd);
2616                 rc = filter_llog_preclean(obd);
2617                 break;
2618         }
2619         RETURN(rc);
2620 }
2621
2622 static int filter_cleanup(struct obd_device *obd)
2623 {
2624         struct filter_obd *filter = &obd->u.filter;
2625         ENTRY;
2626
2627         if (obd->obd_fail)
2628                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2629                               "will be preserved.\n", obd->obd_name);
2630
2631         obd_exports_barrier(obd);
2632         obd_zombie_barrier();
2633
2634         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2635         lprocfs_free_per_client_stats(obd);
2636         lprocfs_free_obd_stats(obd);
2637         lprocfs_obd_cleanup(obd);
2638         lquota_cleanup(filter_quota_interface_ref, obd);
2639
2640         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2641         obd->obd_namespace = NULL;
2642
2643         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2644
2645         if (obd->u.obt.obt_sb == NULL)
2646                 RETURN(0);
2647
2648         filter_post(obd);
2649
2650         ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2651         shrink_dcache_sb(obd->u.obt.obt_sb);
2652
2653         server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2654         obd->u.obt.obt_sb = NULL;
2655
2656         fsfilt_put_ops(obd->obd_fsops);
2657
2658         filter_iobuf_pool_done(filter);
2659
2660         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2661
2662         RETURN(0);
2663 }
2664
2665 static int filter_connect_internal(struct obd_export *exp,
2666                                    struct obd_connect_data *data,
2667                                    int reconnect)
2668 {
2669         struct filter_export_data *fed = &exp->exp_filter_data;
2670
2671         if (!data)
2672                 RETURN(0);
2673
2674         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2675                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2676                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2677                data->ocd_connect_flags, data->ocd_version,
2678                data->ocd_grant, data->ocd_index);
2679
2680         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2681                 CWARN("!!! This export (nid %s) used object group %d "
2682                        "earlier; now it's trying to use group %d!  This could "
2683                        "be a bug in the MDS. Please report to "
2684                        "http://bugs.whamcloud.com/\n",
2685                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2686                 RETURN(-EPROTO);
2687         }
2688         fed->fed_group = data->ocd_group;
2689
2690         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2691         exp->exp_connect_flags = data->ocd_connect_flags;
2692         data->ocd_version = LUSTRE_VERSION_CODE;
2693
2694         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2695         if (data->ocd_connect_flags & OBD_CONNECT_MDS)
2696                 CWARN("%s: Received MDS connection for group %u\n",
2697                       exp->exp_obd->obd_name, data->ocd_group);
2698         else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
2699                 RETURN(-EPROTO);
2700
2701         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2702                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2703                 obd_size left, want;
2704
2705                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2706                 left = filter_grant_space_left(exp);
2707                 want = data->ocd_grant;
2708                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2709                 data->ocd_grant = fed->fed_grant;
2710                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2711
2712                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2713                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2714                        exp->exp_client_uuid.uuid, exp,
2715                        data->ocd_grant, want, left);
2716
2717                 filter->fo_tot_granted_clients ++;
2718         }
2719
2720         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2721                 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2722                 int index = le32_to_cpu(lsd->lsd_ost_index);
2723
2724                 if (!(lsd->lsd_feature_compat &
2725                       cpu_to_le32(OBD_COMPAT_OST))) {
2726                         /* this will only happen on the first connect */
2727                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2728                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2729                         /* sync is not needed here as filter_client_add will
2730                          * set exp_need_sync flag */
2731                         filter_update_server_data(exp->exp_obd);
2732                 } else if (index != data->ocd_index) {
2733                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2734                                            " %u doesn't match actual OST index"
2735                                            " %u in last_rcvd file, bad "
2736                                            "configuration?\n",
2737                                            obd_export_nid2str(exp), index,
2738                                            data->ocd_index);
2739                         RETURN(-EBADF);
2740                 }
2741                 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2742                  * FIXME: We don't strictly need the COMPAT flag for that,
2743                  * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2744                  * FIXME: We needed it for the index, as index 0 is valid. */
2745         }
2746
2747         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2748                 data->ocd_brw_size = 65536;
2749         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2750                 data->ocd_brw_size = min(data->ocd_brw_size,
2751                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2752                 if (data->ocd_brw_size == 0) {
2753                         CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2754                                " ocd_version: %x ocd_grant: %d ocd_index: %u "
2755                                "ocd_brw_size is unexpectedly zero, "
2756                                "network data corruption?"
2757                                "Refusing connection of this client\n",
2758                                 exp->exp_obd->obd_name,
2759                                 exp->exp_client_uuid.uuid,
2760                                 exp, data->ocd_connect_flags, data->ocd_version,
2761                                 data->ocd_grant, data->ocd_index);
2762                         RETURN(-EPROTO);
2763                 }
2764         }
2765
2766         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2767                 __u32 cksum_types = data->ocd_cksum_types;
2768
2769                 /* The client set in ocd_cksum_types the checksum types it
2770                  * supports. We have to mask off the algorithms that we don't
2771                  * support */
2772                 data->ocd_cksum_types &= cksum_types_supported();
2773
2774                 /* 1.6.4- only support CRC32 and didn't set ocd_cksum_types */
2775                 if (unlikely(data->ocd_cksum_types == 0))
2776                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2777
2778                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2779                                    "%x\n", exp->exp_obd->obd_name,
2780                                    obd_export_nid2str(exp), cksum_types,
2781                                    data->ocd_cksum_types);
2782         } else {
2783                 /* This client does not support OBD_CONNECT_CKSUM
2784                  * fall back to CRC32 */
2785                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2786                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2787                                    exp->exp_obd->obd_name,
2788                                    obd_export_nid2str(exp));
2789         }
2790
2791         if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2792                 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2793
2794         RETURN(0);
2795 }
2796
2797 static int filter_reconnect(const struct lu_env *env,
2798                             struct obd_export *exp, struct obd_device *obd,
2799                             struct obd_uuid *cluuid,
2800                             struct obd_connect_data *data,
2801                             void *localdata)
2802 {
2803         int rc;
2804         ENTRY;
2805
2806         if (exp == NULL || obd == NULL || cluuid == NULL)
2807                 RETURN(-EINVAL);
2808
2809         rc = filter_connect_internal(exp, data, 1);
2810         if (rc == 0)
2811                 filter_export_stats_init(obd, exp, localdata);
2812
2813         RETURN(rc);
2814 }
2815
2816 static int filter_connect(const struct lu_env *env,
2817                           struct obd_export **exp, struct obd_device *obd,
2818                           struct obd_uuid *cluuid,
2819                           struct obd_connect_data *data, void *localdata)
2820 {
2821         struct lvfs_run_ctxt saved;
2822         struct lustre_handle conn = { 0 };
2823         struct obd_export *lexp;
2824         int rc;
2825         ENTRY;
2826
2827         if (exp == NULL || obd == NULL || cluuid == NULL)
2828                 RETURN(-EINVAL);
2829
2830         rc = class_connect(&conn, obd, cluuid);
2831         if (rc)
2832                 RETURN(rc);
2833         lexp = class_conn2export(&conn);
2834         LASSERT(lexp != NULL);
2835
2836         rc = filter_connect_internal(lexp, data, 0);
2837         if (rc)
2838                 GOTO(cleanup, rc);
2839
2840         filter_export_stats_init(obd, lexp, localdata);
2841         if (obd->obd_replayable) {
2842                 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2843                 LASSERT(lcd);
2844                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2845                 rc = filter_client_add(obd, lexp, -1);
2846                 if (rc)
2847                         GOTO(cleanup, rc);
2848         }
2849
2850         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2851         rc = filter_read_groups(obd, data->ocd_group, 1);
2852         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2853         if (rc != 0) {
2854                 CERROR("can't read group %u\n", data->ocd_group);
2855                 GOTO(cleanup, rc);
2856         }
2857
2858         GOTO(cleanup, rc);
2859
2860 cleanup:
2861         if (rc) {
2862                 class_disconnect(lexp);
2863                 *exp = NULL;
2864         } else {
2865                 *exp = lexp;
2866         }
2867
2868         RETURN(rc);
2869 }
2870
2871 /* Do extra sanity checks for grant accounting.  We do this at connect,
2872  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2873  * always get rid of it or turn it off when we know accounting is good. */
2874 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2875 {
2876         struct filter_export_data *fed;
2877         struct obd_export *exp;
2878         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2879         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2880         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2881
2882         if (cfs_list_empty(&obd->obd_exports))
2883                 return;
2884
2885         /* We don't want to do this for large machines that do lots of
2886            mounts or unmounts.  It burns... */
2887         if (obd->obd_num_exports > 100)
2888                 return;
2889
2890         cfs_spin_lock(&obd->obd_osfs_lock);
2891         cfs_spin_lock(&obd->obd_dev_lock);
2892         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2893                 int error = 0;
2894                 fed = &exp->exp_filter_data;
2895                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2896                     fed->fed_dirty < 0)
2897                         error = 1;
2898                 if (maxsize > 0) { /* we may not have done a statfs yet */
2899                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2900                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2901                                  exp->exp_client_uuid.uuid, exp,
2902                                  fed->fed_grant, fed->fed_pending, maxsize);
2903                         LASSERTF(fed->fed_dirty <= maxsize,
2904                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2905                                  exp->exp_client_uuid.uuid, exp,
2906                                  fed->fed_dirty, maxsize);
2907                 }
2908                 if (error)
2909                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2910                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2911                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2912                 else
2913                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2914                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2915                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2916                 tot_granted += fed->fed_grant + fed->fed_pending;
2917                 tot_pending += fed->fed_pending;
2918                 tot_dirty += fed->fed_dirty;
2919         }
2920         fo_tot_granted = obd->u.filter.fo_tot_granted;
2921         fo_tot_pending = obd->u.filter.fo_tot_pending;
2922         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2923         cfs_spin_unlock(&obd->obd_dev_lock);
2924         cfs_spin_unlock(&obd->obd_osfs_lock);
2925
2926         /* Do these assertions outside the spinlocks so we don't kill system */
2927         if (tot_granted != fo_tot_granted)
2928                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2929                        func, tot_granted, fo_tot_granted);
2930         if (tot_pending != fo_tot_pending)
2931                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2932                        func, tot_pending, fo_tot_pending);
2933         if (tot_dirty != fo_tot_dirty)
2934                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2935                        func, tot_dirty, fo_tot_dirty);
2936         if (tot_pending > tot_granted)
2937                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2938                        func, tot_pending, tot_granted);
2939         if (tot_granted > maxsize)
2940                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2941                        func, tot_granted, maxsize);
2942         if (tot_dirty > maxsize)
2943                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2944                        func, tot_dirty, maxsize);
2945 }
2946
2947 /* Remove this client from the grant accounting totals.  We also remove
2948  * the export from the obd device under the osfs and dev locks to ensure
2949  * that the filter_grant_sanity_check() calculations are always valid.
2950  * The client should do something similar when it invalidates its import. */
2951 static void filter_grant_discard(struct obd_export *exp)
2952 {
2953         struct obd_device *obd = exp->exp_obd;
2954         struct filter_obd *filter = &obd->u.filter;
2955         struct filter_export_data *fed = &exp->exp_filter_data;
2956
2957         cfs_spin_lock(&obd->obd_osfs_lock);
2958         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2959                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2960                  obd->obd_name, filter->fo_tot_granted,
2961                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2962         filter->fo_tot_granted -= fed->fed_grant;
2963         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2964                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2965                  obd->obd_name, filter->fo_tot_pending,
2966                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2967         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2968         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2969                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2970                  obd->obd_name, filter->fo_tot_dirty,
2971                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2972         filter->fo_tot_dirty -= fed->fed_dirty;
2973         fed->fed_dirty = 0;
2974         fed->fed_grant = 0;
2975
2976         cfs_spin_unlock(&obd->obd_osfs_lock);
2977 }
2978
2979 static int filter_destroy_export(struct obd_export *exp)
2980 {
2981         struct filter_export_data *fed = &exp->exp_filter_data;
2982         ENTRY;
2983
2984         if (fed->fed_pending)
2985                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2986                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2987                        exp, fed->fed_pending);
2988
2989         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2990
2991         target_destroy_export(exp);
2992         ldlm_destroy_export(exp);
2993         lut_client_free(exp);
2994
2995         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2996                 RETURN(0);
2997
2998         if (!exp->exp_obd->obd_replayable)
2999                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3000
3001         filter_grant_discard(exp);
3002         filter_fmd_cleanup(exp);
3003
3004         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3005                 struct filter_obd *filter = &exp->exp_obd->u.filter;
3006                 if (filter->fo_tot_granted_clients > 0)
3007                         filter->fo_tot_granted_clients --;
3008         }
3009
3010         if (!(exp->exp_flags & OBD_OPT_FORCE))
3011                 filter_grant_sanity_check(exp->exp_obd, __func__);
3012
3013         RETURN(0);
3014 }
3015
3016 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3017 {
3018         struct obd_llog_group *olg_min, *olg;
3019         struct filter_obd *filter;
3020         int worked = -1, group;
3021         struct llog_ctxt *ctxt;
3022         ENTRY;
3023
3024         filter = &obd->u.filter;
3025
3026         /* we can't sync log holding spinlock. also, we do not want to get
3027          * into livelock. so we do following: loop over MDS's exports in
3028          * group order and skip already synced llogs -bzzz */
3029         do {
3030                 /* look for group with min. number, but > worked */
3031                 olg_min = NULL;
3032                 group = 1 << 30;
3033                 cfs_spin_lock(&filter->fo_llog_list_lock);
3034                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3035                         if (olg->olg_seq <= worked) {
3036                                 /* this group is already synced */
3037                                 continue;
3038                         }
3039                         if (group < olg->olg_seq) {
3040                                 /* we have group with smaller number to sync */
3041                                 continue;
3042                         }
3043                         /* store current minimal group */
3044                         olg_min = olg;
3045                         group = olg->olg_seq;
3046                 }
3047                 cfs_spin_unlock(&filter->fo_llog_list_lock);
3048
3049                 if (olg_min == NULL)
3050                         break;
3051
3052                 worked = olg_min->olg_seq;
3053                 if (olg_min->olg_exp &&
3054                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3055                         int err;
3056                         ctxt = llog_group_get_ctxt(olg_min,
3057                                                    LLOG_MDS_OST_REPL_CTXT);
3058                         if (ctxt) {
3059                                 err = llog_sync(ctxt, olg_min->olg_exp);
3060                                 llog_ctxt_put(ctxt);
3061                                 if (err) {
3062                                         CERROR("error flushing logs to MDS: "
3063                                                "rc %d\n", err);
3064                                 }
3065                         }
3066                 }
3067         } while (olg_min != NULL);
3068 }
3069
3070 /* Also incredibly similar to mds_disconnect */
3071 static int filter_disconnect(struct obd_export *exp)
3072 {
3073         struct obd_device *obd = exp->exp_obd;
3074         int rc;
3075         ENTRY;
3076
3077         LASSERT(exp);
3078         class_export_get(exp);
3079
3080         if (!(exp->exp_flags & OBD_OPT_FORCE))
3081                 filter_grant_sanity_check(obd, __func__);
3082         filter_grant_discard(exp);
3083
3084         /* Flush any remaining cancel messages out to the target */
3085         filter_sync_llogs(obd, exp);
3086
3087         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3088
3089         rc = server_disconnect_export(exp);
3090
3091         /* Do not erase record for recoverable client. */
3092         if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3093                 filter_client_del(exp);
3094         else
3095                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3096
3097         class_export_put(exp);
3098         RETURN(rc);
3099 }
3100
3101 /* reverse import is changed, sync all cancels */
3102 static void filter_revimp_update(struct obd_export *exp)
3103 {
3104         ENTRY;
3105
3106         LASSERT(exp);
3107         class_export_get(exp);
3108
3109         /* flush any remaining cancel messages out to the target */
3110         filter_sync_llogs(exp->exp_obd, exp);
3111         class_export_put(exp);
3112         EXIT;
3113 }
3114
3115 static int filter_ping(struct obd_export *exp)
3116 {
3117         filter_fmd_expire(exp);
3118         return 0;
3119 }
3120
3121 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3122                                   const char *what, int quiet)
3123 {
3124         struct dentry *dchild = NULL;
3125
3126         dchild = filter_fid2dentry(obd, NULL,  ostid->oi_seq, ostid->oi_id);
3127
3128         if (IS_ERR(dchild)) {
3129                 CERROR("%s error looking up object: "POSTID"\n",
3130                        what, ostid->oi_id, ostid->oi_seq);
3131                 RETURN(dchild);
3132         }
3133
3134         if (dchild->d_inode == NULL) {
3135                 if (!quiet)
3136                         CERROR("%s: %s on non-existent object: "POSTID" \n",
3137                                obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3138                 f_dput(dchild);
3139                 RETURN(ERR_PTR(-ENOENT));
3140         }
3141
3142         return dchild;
3143 }
3144
3145 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3146 {
3147         struct dentry *dentry = NULL;
3148         struct obd_device *obd;
3149         int rc = 0;
3150         ENTRY;
3151
3152         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3153                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3154         if (rc)
3155                 RETURN(rc);
3156
3157         obd = class_exp2obd(exp);
3158         if (obd == NULL) {
3159                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3160                 RETURN(-EINVAL);
3161         }
3162
3163         dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3164         if (IS_ERR(dentry))
3165                 RETURN(PTR_ERR(dentry));
3166
3167         /* Limit the valid bits in the return data to what we actually use */
3168         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3169         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3170
3171         f_dput(dentry);
3172         RETURN(rc);
3173 }
3174
3175 /* this should be enabled/disabled in condition to enabled/disabled large
3176  * inodes (fast EAs) in backing store FS. */
3177 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3178                         void *handle, struct obdo *oa)
3179 {
3180         struct obd_device *obd = exp->exp_obd;
3181         int rc = 0;
3182         ENTRY;
3183
3184         if (oa->o_valid & OBD_MD_FLFID) {
3185                 struct filter_fid ff;
3186
3187                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3188                         oa->o_seq = 0;
3189                 /* packing fid and converting it to LE for storing into EA.
3190                  * Here ->o_stripe_idx should be filled by LOV and rest of
3191                  * fields - by client. */
3192                 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3193                 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3194                 /* XXX: we are ignoring o_parent_ver here, since this should
3195                  *      be the same for all objects in this fileset. */
3196                 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3197                 ff.ff_objid = cpu_to_le64(oa->o_id);
3198                 ff.ff_seq = cpu_to_le64(oa->o_seq);
3199
3200                 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3201                        LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3202