Whamcloud - gitweb
b=21871 Just cleanup export nid_stat while exp_refcount == 0.
[fs/lustre-release.git] / lustre / obdfilter / filter.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/obdfilter/filter.c
37  *
38  * Author: Peter Braam <braam@clusterfs.com>
39  * Author: Andreas Dilger <adilger@clusterfs.com>
40  */
41
42 /*
43  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44  *            (which need to get journal_lock, may block if journal full).
45  *
46  * Invariant: Call filter_start_transno() before any journal ops to avoid the
47  *            same deadlock problem.  We can (and want) to get rid of the
48  *            transno sem in favour of the dir/inode i_mutex to avoid single
49  *            threaded operation on the OST.
50  */
51
52 #define DEBUG_SUBSYSTEM S_FILTER
53
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
56 #endif
57 #include <linux/module.h>
58 #include <linux/fs.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
65
66 #include <obd_cksum.h>
67 #include <obd_class.h>
68 #include <obd_lov.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
79
80 #include "filter_internal.h"
81
82 static struct lvfs_callback_ops filter_lvfs_ops;
83 cfs_mem_cache_t *ll_fmd_cachep;
84
85 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
86                              void *cb_data, int error)
87 {
88         struct obd_export *exp = cb_data;
89         LASSERT(exp->exp_obd == obd);
90         obd_transno_commit_cb(obd, transno, exp, error);
91         class_export_cb_put(exp);
92 }
93
94 int filter_version_get_check(struct obd_export *exp,
95                              struct obd_trans_info *oti, struct inode *inode)
96 {
97         __u64 curr_version;
98
99         if (inode == NULL || oti == NULL)
100                 RETURN(0);
101
102         curr_version = fsfilt_get_version(exp->exp_obd, inode);
103         if ((__s64)curr_version == -EOPNOTSUPP)
104                 RETURN(0);
105         /* VBR: version is checked always because costs nothing */
106         if (oti->oti_pre_version != 0 &&
107             oti->oti_pre_version != curr_version) {
108                 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
109                        oti->oti_pre_version, curr_version);
110                 cfs_spin_lock(&exp->exp_lock);
111                 exp->exp_vbr_failed = 1;
112                 cfs_spin_unlock(&exp->exp_lock);
113                 RETURN (-EOVERFLOW);
114         }
115         oti->oti_pre_version = curr_version;
116         RETURN(0);
117 }
118
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121                           struct obd_trans_info *oti, int rc, int force_sync)
122 {
123         struct obd_device_target *obt = &exp->exp_obd->u.obt;
124         struct tg_export_data *ted = &exp->exp_target_data;
125         struct lr_server_data *lsd = class_server_data(exp->exp_obd);
126         struct lsd_client_data *lcd;
127         __u64 last_rcvd;
128         loff_t off;
129         int err, log_pri = D_RPCTRACE;
130
131         /* Propagate error code. */
132         if (rc)
133                 RETURN(rc);
134
135         if (!exp->exp_obd->obd_replayable || oti == NULL)
136                 RETURN(rc);
137
138         cfs_mutex_down(&ted->ted_lcd_lock);
139         lcd = ted->ted_lcd;
140         /* if the export has already been disconnected, we have no last_rcvd slot,
141          * update server data with latest transno then */
142         if (lcd == NULL) {
143                 cfs_mutex_up(&ted->ted_lcd_lock);
144                 CWARN("commit transaction for disconnected client %s: rc %d\n",
145                       exp->exp_client_uuid.uuid, rc);
146                 err = filter_update_server_data(exp->exp_obd);
147                 RETURN(err);
148         }
149
150         /* we don't allocate new transnos for replayed requests */
151         cfs_spin_lock(&obt->obt_lut->lut_translock);
152         if (oti->oti_transno == 0) {
153                 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
154                 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
155         } else {
156                 last_rcvd = oti->oti_transno;
157                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
158                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
159         }
160         oti->oti_transno = last_rcvd;
161
162         LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
163         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
164         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
165         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
166         cfs_spin_unlock(&obt->obt_lut->lut_translock);
167
168         if (inode)
169                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
170
171         off = ted->ted_lr_off;
172         if (off <= 0) {
173                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
174                        ted->ted_lr_idx, ted->ted_lr_off);
175                 err = -EINVAL;
176         } else {
177                 class_export_cb_get(exp); /* released when the cb is called */
178                 if (!force_sync)
179                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
180                                                            last_rcvd,
181                                                            oti->oti_handle,
182                                                            filter_commit_cb,
183                                                            exp);
184
185                 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
186                                           lcd, sizeof(*lcd), &off,
187                                           force_sync | exp->exp_need_sync);
188                 if (force_sync)
189                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
190         }
191         if (err) {
192                 log_pri = D_ERROR;
193                 if (rc == 0)
194                         rc = err;
195         }
196
197         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
198                last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
199         cfs_mutex_up(&ted->ted_lcd_lock);
200         RETURN(rc);
201 }
202
203 void f_dput(struct dentry *dentry)
204 {
205         /* Can't go inside filter_ddelete because it can block */
206         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
207                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
208         LASSERT(atomic_read(&dentry->d_count) > 0);
209
210         dput(dentry);
211 }
212
213 static void init_brw_stats(struct brw_stats *brw_stats)
214 {
215         int i;
216         for (i = 0; i < BRW_LAST; i++)
217                 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
218 }
219
220 static int lprocfs_init_rw_stats(struct obd_device *obd,
221                                  struct lprocfs_stats **stats)
222 {
223         int num_stats;
224
225         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
226                                                         LPROC_FILTER_LAST - 1;
227         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
228         if (*stats == NULL)
229                 return -ENOMEM;
230
231         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
232         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
233                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
234         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
235                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
236
237         return(0);
238 }
239
240 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
241    plus the procfs overhead :( */
242 static int filter_export_stats_init(struct obd_device *obd,
243                                     struct obd_export *exp,
244                                     int reconnect,
245                                     void *client_nid)
246 {
247         int rc, newnid = 0;
248         ENTRY;
249
250         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
251                 /* Self-export gets no proc entry */
252                 RETURN(0);
253
254         rc = lprocfs_exp_setup(exp, client_nid, reconnect, &newnid);
255         if (rc) {
256                 /* Mask error for already created
257                  * /proc entries */
258                 if (rc == -EALREADY)
259                         rc = 0;
260                 RETURN(rc);
261         }
262
263         if (newnid) {
264                 struct nid_stat *tmp = exp->exp_nid_stats;
265                 LASSERT(tmp != NULL);
266
267                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
268                 if (tmp->nid_brw_stats == NULL)
269                         GOTO(clean, rc = -ENOMEM);
270
271                 init_brw_stats(tmp->nid_brw_stats);
272                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
273                                         0644, &filter_per_nid_stats_fops,
274                                         exp->exp_nid_stats);
275                 if (rc)
276                         CWARN("Error adding the brw_stats file\n");
277
278                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
279                 if (rc)
280                         GOTO(clean, rc);
281
282                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
283                                             tmp->nid_stats);
284                 if (rc)
285                         GOTO(clean, rc);
286                 rc = lprocfs_nid_ldlm_stats_init(tmp);
287                 if (rc)
288                         GOTO(clean, rc);
289         }
290
291         RETURN(0);
292  clean:
293         return rc;
294 }
295
296 /* Add client data to the FILTER.  We use a bitmap to locate a free space
297  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
298  * Otherwise, we have just read the data from the last_rcvd file and
299  * we know its offset. */
300 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
301                              int cl_idx)
302 {
303         struct obd_device_target *obt = &obd->u.obt;
304         struct tg_export_data *ted = &exp->exp_target_data;
305         struct lr_server_data *lsd = class_server_data(obd);
306         unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
307         int new_client = (cl_idx == -1);
308
309         ENTRY;
310
311         LASSERT(bitmap != NULL);
312         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
313
314         /* Self-export */
315         if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
316                 RETURN(0);
317
318         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
319          * there's no need for extra complication here
320          */
321         if (new_client) {
322                 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
323         repeat:
324                 if (cl_idx >= LR_MAX_CLIENTS) {
325                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
326                                cl_idx);
327                         RETURN(-EOVERFLOW);
328                 }
329                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
330                         cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
331                                                         cl_idx);
332                         goto repeat;
333                 }
334         } else {
335                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
336                         CERROR("FILTER client %d: bit already set in bitmap!\n",
337                                cl_idx);
338                         LBUG();
339                 }
340         }
341
342         ted->ted_lr_idx = cl_idx;
343         ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
344                           cl_idx * le16_to_cpu(lsd->lsd_client_size);
345         cfs_init_mutex(&ted->ted_lcd_lock);
346         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
347
348         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
349                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
350
351         if (new_client) {
352                 struct lvfs_run_ctxt saved;
353                 loff_t off = ted->ted_lr_off;
354                 int rc;
355                 void *handle;
356
357                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
358                        ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
359
360                 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
361                         RETURN(-ENOSPC);
362
363                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
364                 /* Transaction needed to fix bug 1403 */
365                 handle = fsfilt_start(obd,
366                                       obt->obt_rcvd_filp->f_dentry->d_inode,
367                                       FSFILT_OP_SETATTR, NULL);
368                 if (IS_ERR(handle)) {
369                         rc = PTR_ERR(handle);
370                         CERROR("unable to start transaction: rc %d\n", rc);
371                 } else {
372                         ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
373                         exp->exp_last_request_time = cfs_time_current_sec();
374                         rc = fsfilt_add_journal_cb(obd, 0, handle,
375                                                    target_client_add_cb,
376                                                    class_export_cb_get(exp));
377                         if (rc == 0) {
378                                 cfs_spin_lock(&exp->exp_lock);
379                                 exp->exp_need_sync = 1;
380                                 cfs_spin_unlock(&exp->exp_lock);
381                         }
382                         rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
383                                                  ted->ted_lcd,
384                                                  sizeof(*ted->ted_lcd),
385                                                  &off, rc /* sync if no cb */);
386                         fsfilt_commit(obd,
387                                       obt->obt_rcvd_filp->f_dentry->d_inode,
388                                       handle, 0);
389                 }
390                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
391
392                 if (rc) {
393                         CERROR("error writing %s client idx %u: rc %d\n",
394                                LAST_RCVD, ted->ted_lr_idx, rc);
395                         RETURN(rc);
396                 }
397         }
398         RETURN(0);
399 }
400
401 static int filter_client_del(struct obd_export *exp)
402 {
403         struct tg_export_data *ted = &exp->exp_target_data;
404         struct obd_device_target *obt = &exp->exp_obd->u.obt;
405         struct lvfs_run_ctxt saved;
406         int rc;
407         loff_t off;
408         ENTRY;
409
410         if (ted->ted_lcd == NULL)
411                 RETURN(0);
412
413         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
414         if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
415                 GOTO(free, 0);
416
417         LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
418
419         off = ted->ted_lr_off;
420
421         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
422                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
423
424         /* Don't clear ted_lr_idx here as it is likely also unset.  At worst
425          * we leak a client slot that will be cleaned on the next recovery. */
426         if (off <= 0) {
427                 CERROR("%s: client idx %d has med_off %lld\n",
428                        exp->exp_obd->obd_name, ted->ted_lr_idx, off);
429                 GOTO(free, rc = -EINVAL);
430         }
431
432         /* Clear the bit _after_ zeroing out the client so we don't
433            race with filter_client_add and zero out new clients.*/
434         if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
435                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
436                        ted->ted_lr_idx);
437                 LBUG();
438         }
439
440         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
441         /* Make sure the server's last_transno is up to date.
442          * This should be done before zeroing client slot so last_transno will
443          * be in server data or in client data in case of failure */
444         filter_update_server_data(exp->exp_obd);
445
446         cfs_mutex_down(&ted->ted_lcd_lock);
447         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
448         rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
449                                  ted->ted_lcd,
450                                  sizeof(*ted->ted_lcd), &off, 0);
451         cfs_mutex_up(&ted->ted_lcd_lock);
452         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
453
454         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
455                "zero out client %s at idx %u/%llu in %s, rc %d\n",
456                ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
457                LAST_RCVD, rc);
458         EXIT;
459 free:
460         return 0;
461 }
462
463 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
464 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
465                                          struct filter_mod_data *fmd)
466 {
467         LASSERT_SPIN_LOCKED(&fed->fed_lock);
468         if (--fmd->fmd_refcount == 0) {
469                 /* XXX when we have persistent reservations and the handle
470                  * is stored herein we need to drop it here. */
471                 fed->fed_mod_count--;
472                 cfs_list_del(&fmd->fmd_list);
473                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
474         }
475 }
476
477 /* drop fmd reference, free it if last ref */
478 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
479 {
480         struct filter_export_data *fed;
481
482         if (fmd == NULL)
483                 return;
484
485         fed = &exp->exp_filter_data;
486         cfs_spin_lock(&fed->fed_lock);
487         filter_fmd_put_nolock(fed, fmd); /* caller reference */
488         cfs_spin_unlock(&fed->fed_lock);
489 }
490
491 /* expire entries from the end of the list if there are too many
492  * or they are too old */
493 static void filter_fmd_expire_nolock(struct filter_obd *filter,
494                                      struct filter_export_data *fed,
495                                      struct filter_mod_data *keep)
496 {
497         struct filter_mod_data *fmd, *tmp;
498
499         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
500                 if (fmd == keep)
501                         break;
502
503                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
504                     fed->fed_mod_count < filter->fo_fmd_max_num)
505                         break;
506
507                 cfs_list_del_init(&fmd->fmd_list);
508                 filter_fmd_put_nolock(fed, fmd); /* list reference */
509         }
510 }
511
512 void filter_fmd_expire(struct obd_export *exp)
513 {
514         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
515         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
516                                  &exp->exp_filter_data, NULL);
517         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
518 }
519
520 /* find specified objid, group in export fmd list.
521  * caller must hold fed_lock and take fmd reference itself */
522 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
523                                                 struct filter_export_data *fed,
524                                                 obd_id objid, obd_seq group)
525 {
526         struct filter_mod_data *found = NULL, *fmd;
527
528         LASSERT_SPIN_LOCKED(&fed->fed_lock);
529
530         cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
531                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
532                         found = fmd;
533                         cfs_list_del(&fmd->fmd_list);
534                         cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
535                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
536                         break;
537                 }
538         }
539
540         filter_fmd_expire_nolock(filter, fed, found);
541
542         return found;
543 }
544
545 /* Find fmd based on objid and group, or return NULL if not found. */
546 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
547                                         obd_id objid, obd_seq group)
548 {
549         struct filter_mod_data *fmd;
550
551         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
552         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
553                                      &exp->exp_filter_data, objid, group);
554         if (fmd)
555                 fmd->fmd_refcount++;    /* caller reference */
556         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
557
558         return fmd;
559 }
560
561 /* Find fmd based on objid and group, or create a new one if none is found.
562  * It is possible for this function to return NULL under memory pressure,
563  * or if objid = 0 is passed (which will only cause old entries to expire).
564  * Currently this is not fatal because any fmd state is transient and
565  * may also be freed when it gets sufficiently old. */
566 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
567                                        obd_id objid, obd_seq group)
568 {
569         struct filter_export_data *fed = &exp->exp_filter_data;
570         struct filter_mod_data *found = NULL, *fmd_new = NULL;
571
572         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
573
574         cfs_spin_lock(&fed->fed_lock);
575         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
576         if (fmd_new) {
577                 if (found == NULL) {
578                         cfs_list_add_tail(&fmd_new->fmd_list,
579                                           &fed->fed_mod_list);
580                         fmd_new->fmd_id = objid;
581                         fmd_new->fmd_gr = group;
582                         fmd_new->fmd_refcount++;   /* list reference */
583                         found = fmd_new;
584                         fed->fed_mod_count++;
585                 } else {
586                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
587                 }
588         }
589         if (found) {
590                 found->fmd_refcount++;          /* caller reference */
591                 found->fmd_expire = jiffies +
592                         exp->exp_obd->u.filter.fo_fmd_max_age;
593         }
594
595         cfs_spin_unlock(&fed->fed_lock);
596
597         return found;
598 }
599
600 #ifdef DO_FMD_DROP
601 /* drop fmd list reference so it will disappear when last reference is put.
602  * This isn't so critical because it would in fact only affect the one client
603  * that is doing the unlink and at worst we have an stale entry referencing
604  * an object that should never be used again. */
605 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
606 {
607         struct filter_mod_data *found = NULL;
608
609         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
610         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
611         if (found) {
612                 cfs_list_del_init(&found->fmd_list);
613                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
614         }
615         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
616 }
617 #else
618 #define filter_fmd_drop(exp, objid, group)
619 #endif
620
621 /* remove all entries from fmd list */
622 static void filter_fmd_cleanup(struct obd_export *exp)
623 {
624         struct filter_export_data *fed = &exp->exp_filter_data;
625         struct filter_mod_data *fmd = NULL, *tmp;
626
627         cfs_spin_lock(&fed->fed_lock);
628         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
629                 cfs_list_del_init(&fmd->fmd_list);
630                 filter_fmd_put_nolock(fed, fmd);
631         }
632         cfs_spin_unlock(&fed->fed_lock);
633 }
634
635 static int filter_init_export(struct obd_export *exp)
636 {
637         int rc;
638         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
639         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
640
641         cfs_spin_lock(&exp->exp_lock);
642         exp->exp_connecting = 1;
643         cfs_spin_unlock(&exp->exp_lock);
644         rc = lut_client_alloc(exp);
645         if (rc == 0)
646                 rc = ldlm_init_export(exp);
647
648         return rc;
649 }
650
651 static int filter_free_server_data(struct obd_device_target *obt)
652 {
653         lut_fini(NULL, obt->obt_lut);
654         OBD_FREE_PTR(obt->obt_lut);
655         return 0;
656 }
657
658 /* assumes caller is already in kernel ctxt */
659 int filter_update_server_data(struct obd_device *obd)
660 {
661         struct file *filp = obd->u.obt.obt_rcvd_filp;
662         struct lr_server_data *lsd = class_server_data(obd);
663         loff_t off = 0;
664         int rc;
665         ENTRY;
666
667         CDEBUG(D_INODE, "server uuid      : %s\n", lsd->lsd_uuid);
668         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
669                le64_to_cpu(lsd->lsd_last_transno));
670         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
671                le64_to_cpu(lsd->lsd_mount_count));
672
673         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
674         if (rc)
675                 CERROR("error writing lr_server_data: rc = %d\n", rc);
676
677         RETURN(rc);
678 }
679
680 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
681                              int force_sync)
682 {
683         struct filter_obd *filter = &obd->u.filter;
684         __u64 tmp;
685         loff_t off = 0;
686         int rc;
687         ENTRY;
688
689         if (filter->fo_last_objid_files[group] == NULL) {
690                 CERROR("Object seq "LPU64" not fully setup; not updating "
691                        "last_objid\n", group);
692                 RETURN(-EINVAL);
693         }
694
695         CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
696                obd->obd_name, filter->fo_last_objids[group], group);
697
698         tmp = cpu_to_le64(filter->fo_last_objids[group]);
699         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
700                                  &tmp, sizeof(tmp), &off, force_sync);
701         if (rc)
702                 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
703                        group, rc);
704         RETURN(rc);
705 }
706 extern int ost_handle(struct ptlrpc_request *req);
707 /* assumes caller has already in kernel ctxt */
708 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
709 {
710         struct filter_obd *filter = &obd->u.filter;
711         struct lr_server_data *lsd;
712         struct lsd_client_data *lcd = NULL;
713         struct inode *inode = filp->f_dentry->d_inode;
714         unsigned long last_rcvd_size = i_size_read(inode);
715         struct lu_target *lut;
716         __u64 mount_count;
717         __u32 start_epoch;
718         int cl_idx;
719         loff_t off = 0;
720         int rc;
721
722         /* ensure padding in the struct is the correct size */
723         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
724                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
725         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
726                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
727
728         /* allocate and initialize lu_target */
729         OBD_ALLOC_PTR(lut);
730         if (lut == NULL)
731                 RETURN(-ENOMEM);
732         rc = lut_init(NULL, lut, obd, NULL);
733         if (rc)
734                 GOTO(err_lut, rc);
735         lsd = class_server_data(obd);
736         if (last_rcvd_size == 0) {
737                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
738
739                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
740                 lsd->lsd_last_transno = 0;
741                 mount_count = lsd->lsd_mount_count = 0;
742                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
743                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
744                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
745                 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
746                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
747                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
748                  * MDS first connects and assigns the OST index number. */
749                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
750                                                         OBD_INCOMPAT_OST);
751         } else {
752                 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
753                 if (rc) {
754                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
755                                LAST_RCVD, rc);
756                         GOTO(err_lut, rc);
757                 }
758                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
759                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
760                                            "using the wrong disk %s. Were the "
761                                            "/dev/ assignments rearranged?\n",
762                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
763                         GOTO(err_lut, rc = -EINVAL);
764                 }
765                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
766                 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
767                 /* COMPAT_146 */
768                 /* Assume old last_rcvd format unless I_C_LR is set */
769                 if (!(lsd->lsd_feature_incompat &
770                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
771                         lsd->lsd_last_transno = lsd->lsd_compat14;
772                 /* end COMPAT_146 */
773                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
774                  * MDS first connects and assigns the OST index number. */
775                 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
776                                                          OBD_INCOMPAT_OST);
777         }
778
779         if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
780                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
781                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
782                        ~FILTER_INCOMPAT_SUPP);
783                 GOTO(err_lut, rc = -EINVAL);
784         }
785         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
786                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
787                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
788                        ~FILTER_ROCOMPAT_SUPP);
789                 /* Do something like remount filesystem read-only */
790                 GOTO(err_lut, rc = -EINVAL);
791         }
792
793         start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
794
795         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
796                obd->obd_name, start_epoch);
797         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
798                obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
799         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
800                obd->obd_name, mount_count + 1);
801         CDEBUG(D_INODE, "%s: server data size: %u\n",
802                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
803         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
804                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
805         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
806                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
807         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
808                obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
809         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
810                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
811                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
812                 le16_to_cpu(lsd->lsd_client_size));
813
814         if (!obd->obd_replayable) {
815                 CWARN("%s: recovery support OFF\n", obd->obd_name);
816                 GOTO(out, rc = 0);
817         }
818
819         OBD_ALLOC_PTR(lcd);
820         if (!lcd)
821                 GOTO(err_client, rc = -ENOMEM);
822
823         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
824              off < last_rcvd_size; cl_idx++) {
825                 __u64 last_rcvd;
826                 struct obd_export *exp;
827                 struct filter_export_data *fed;
828
829                 /* Don't assume off is incremented properly by
830                  * fsfilt_read_record(), in case sizeof(*lcd)
831                  * isn't the same as lsd->lsd_client_size.  */
832                 off = le32_to_cpu(lsd->lsd_client_start) +
833                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
834                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
835                 if (rc) {
836                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
837                                LAST_RCVD, cl_idx, off, rc);
838                         break; /* read error shouldn't cause startup to fail */
839                 }
840
841                 if (lcd->lcd_uuid[0] == '\0') {
842                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
843                                cl_idx);
844                         continue;
845                 }
846
847                 check_lcd(obd->obd_name, cl_idx, lcd);
848
849                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
850
851                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
852                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
853                        last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
854
855                 /* These exports are cleaned up by filter_disconnect(), so they
856                  * need to be set up like real exports as filter_connect() does.
857                  */
858                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
859                 if (IS_ERR(exp)) {
860                         if (PTR_ERR(exp) == -EALREADY) {
861                                 /* export already exists, zero out this one */
862                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
863                                 continue;
864                         }
865                         OBD_FREE_PTR(lcd);
866                         GOTO(err_client, rc = PTR_ERR(exp));
867                 }
868
869                 fed = &exp->exp_filter_data;
870                 *fed->fed_ted.ted_lcd = *lcd;
871                 fed->fed_group = 0; /* will be assigned at connect */
872                 filter_export_stats_init(obd, exp, 0, NULL);
873                 rc = filter_client_add(obd, exp, cl_idx);
874                 /* can't fail for existing client */
875                 LASSERTF(rc == 0, "rc = %d\n", rc);
876
877                 /* VBR: set export last committed */
878                 exp->exp_last_committed = last_rcvd;
879                 cfs_spin_lock(&exp->exp_lock);
880                 exp->exp_connecting = 0;
881                 exp->exp_in_recovery = 0;
882                 cfs_spin_unlock(&exp->exp_lock);
883                 obd->obd_max_recoverable_clients++;
884                 class_export_put(exp);
885
886                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
887                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
888         }
889         OBD_FREE_PTR(lcd);
890
891         obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
892 out:
893         obd->u.obt.obt_mount_count = mount_count + 1;
894         lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
895
896         /* save it, so mount count and last_transno is current */
897         rc = filter_update_server_data(obd);
898         if (rc)
899                 GOTO(err_client, rc);
900
901         RETURN(0);
902
903 err_client:
904         class_disconnect_exports(obd);
905 err_lut:
906         filter_free_server_data(&obd->u.obt);
907         RETURN(rc);
908 }
909
910 static int filter_cleanup_groups(struct obd_device *obd)
911 {
912         struct filter_obd *filter = &obd->u.filter;
913         struct file *filp;
914         struct dentry *dentry;
915         int i, j;
916         ENTRY;
917
918         if (filter->fo_dentry_O_groups != NULL) {
919                 for (i = 0; i < filter->fo_group_count; i++) {
920                         dentry = filter->fo_dentry_O_groups[i];
921                         if (dentry != NULL)
922                                 f_dput(dentry);
923                 }
924                 OBD_FREE(filter->fo_dentry_O_groups,
925                          filter->fo_group_count *
926                          sizeof(*filter->fo_dentry_O_groups));
927                 filter->fo_dentry_O_groups = NULL;
928         }
929         if (filter->fo_last_objid_files != NULL) {
930                 for (i = 0; i < filter->fo_group_count; i++) {
931                         filp = filter->fo_last_objid_files[i];
932                         if (filp != NULL)
933                                 filp_close(filp, 0);
934                 }
935                 OBD_FREE(filter->fo_last_objid_files,
936                          filter->fo_group_count *
937                          sizeof(*filter->fo_last_objid_files));
938                 filter->fo_last_objid_files = NULL;
939         }
940         if (filter->fo_dentry_O_sub != NULL) {
941                 for (i = 0; i < filter->fo_group_count; i++) {
942                         for (j = 0; j < filter->fo_subdir_count; j++) {
943                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
944                                 if (dentry != NULL)
945                                         f_dput(dentry);
946                         }
947                 }
948                 OBD_FREE(filter->fo_dentry_O_sub,
949                          filter->fo_group_count *
950                          sizeof(*filter->fo_dentry_O_sub));
951                 filter->fo_dentry_O_sub = NULL;
952         }
953         if (filter->fo_last_objids != NULL) {
954                 OBD_FREE(filter->fo_last_objids,
955                          filter->fo_group_count *
956                          sizeof(*filter->fo_last_objids));
957                 filter->fo_last_objids = NULL;
958         }
959         if (filter->fo_dentry_O != NULL) {
960                 f_dput(filter->fo_dentry_O);
961                 filter->fo_dentry_O = NULL;
962         }
963         RETURN(0);
964 }
965
966 static int filter_update_last_group(struct obd_device *obd, int group)
967 {
968         struct filter_obd *filter = &obd->u.filter;
969         struct file *filp = NULL;
970         int last_group = 0, rc;
971         loff_t off = 0;
972         ENTRY;
973
974         if (group <= filter->fo_committed_group)
975                 RETURN(0);
976
977         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
978         if (IS_ERR(filp)) {
979                 rc = PTR_ERR(filp);
980                 filp = NULL;
981                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
982                 GOTO(cleanup, rc);
983         }
984
985         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
986         if (rc) {
987                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
988                 GOTO(cleanup, rc);
989         }
990
991         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
992                obd->obd_name, last_group, group);
993
994         off = 0;
995         last_group = group;
996         /* must be sync: bXXXX */
997         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
998         if (rc) {
999                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1000                 GOTO(cleanup, rc);
1001         }
1002
1003         filter->fo_committed_group = group;
1004 cleanup:
1005         if (filp)
1006                 filp_close(filp, 0);
1007         RETURN(rc);
1008 }
1009
1010 static int filter_read_group_internal(struct obd_device *obd, int group,
1011                                       int create)
1012 {
1013         struct filter_obd *filter = &obd->u.filter;
1014         __u64 *new_objids = NULL;
1015         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1016         struct dentry **new_groups = NULL;
1017         struct file **new_files = NULL;
1018         struct dentry *dentry;
1019         struct file *filp;
1020         int old_count = filter->fo_group_count, rc, stage = 0, i;
1021         char name[25];
1022         __u64 last_objid;
1023         loff_t off = 0;
1024         int len = group + 1;
1025
1026         snprintf(name, 24, "%d", group);
1027         name[24] = '\0';
1028
1029         if (!create) {
1030                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1031                                            strlen(name));
1032                 if (IS_ERR(dentry)) {
1033                         CERROR("Cannot lookup expected object group %d: %ld\n",
1034                                group, PTR_ERR(dentry));
1035                         RETURN(PTR_ERR(dentry));
1036                 }
1037         } else {
1038                 dentry = simple_mkdir(filter->fo_dentry_O,
1039                                       obd->u.obt.obt_vfsmnt, name, 0700, 1);
1040                 if (IS_ERR(dentry)) {
1041                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1042                                PTR_ERR(dentry));
1043                         RETURN(PTR_ERR(dentry));
1044                 }
1045         }
1046         stage = 1;
1047
1048         snprintf(name, 24, "O/%d/LAST_ID", group);
1049         name[24] = '\0';
1050         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1051         if (IS_ERR(filp)) {
1052                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1053                 GOTO(cleanup, rc = PTR_ERR(filp));
1054         }
1055         stage = 2;
1056
1057         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1058         if (rc) {
1059                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1060                 GOTO(cleanup, rc);
1061         }
1062
1063         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1064                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1065                 if (tmp_subdirs == NULL)
1066                         GOTO(cleanup, rc = -ENOMEM);
1067                 stage = 3;
1068
1069                 for (i = 0; i < filter->fo_subdir_count; i++) {
1070                         char dir[20];
1071                         snprintf(dir, sizeof(dir), "d%u", i);
1072
1073                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1074                                                               obd->u.obt.obt_vfsmnt,
1075                                                               dir, 0700, 1);
1076                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1077                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1078                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1079                                        group, dir, rc);
1080                                 GOTO(cleanup, rc);
1081                         }
1082
1083                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1084                                tmp_subdirs->dentry[i]);
1085                 }
1086         }
1087
1088         /* 'group' is an index; we need an array of length 'group + 1' */
1089         if (group + 1 > old_count) {
1090                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1091                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1092                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1093                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1094                 stage = 4;
1095                 if (new_objids == NULL || new_subdirs == NULL ||
1096                     new_groups == NULL || new_files == NULL)
1097                         GOTO(cleanup, rc = -ENOMEM);
1098
1099                 if (old_count) {
1100                         memcpy(new_objids, filter->fo_last_objids,
1101                                old_count * sizeof(*new_objids));
1102                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1103                                old_count * sizeof(*new_subdirs));
1104                         memcpy(new_groups, filter->fo_dentry_O_groups,
1105                                old_count * sizeof(*new_groups));
1106                         memcpy(new_files, filter->fo_last_objid_files,
1107                                old_count * sizeof(*new_files));
1108
1109                         OBD_FREE(filter->fo_last_objids,
1110                                  old_count * sizeof(*new_objids));
1111                         OBD_FREE(filter->fo_dentry_O_sub,
1112                                  old_count * sizeof(*new_subdirs));
1113                         OBD_FREE(filter->fo_dentry_O_groups,
1114                                  old_count * sizeof(*new_groups));
1115                         OBD_FREE(filter->fo_last_objid_files,
1116                                  old_count * sizeof(*new_files));
1117                 }
1118                 filter->fo_last_objids = new_objids;
1119                 filter->fo_dentry_O_sub = new_subdirs;
1120                 filter->fo_dentry_O_groups = new_groups;
1121                 filter->fo_last_objid_files = new_files;
1122                 filter->fo_group_count = len;
1123         }
1124
1125         filter->fo_dentry_O_groups[group] = dentry;
1126         filter->fo_last_objid_files[group] = filp;
1127         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1128                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1129                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1130         }
1131
1132         filter_update_last_group(obd, group);
1133
1134         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1135                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1136                 rc = filter_update_last_objid(obd, group, 1);
1137                 RETURN(rc);
1138         }
1139
1140         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1141         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1142                obd->obd_name, group, last_objid);
1143         RETURN(0);
1144  cleanup:
1145         switch (stage) {
1146         case 4:
1147                 if (new_objids != NULL)
1148                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1149                 if (new_subdirs != NULL)
1150                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1151                 if (new_groups != NULL)
1152                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1153                 if (new_files != NULL)
1154                         OBD_FREE(new_files, len * sizeof(*new_files));
1155         case 3:
1156                 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1157                         for (i = 0; i < filter->fo_subdir_count; i++) {
1158                                 if (tmp_subdirs->dentry[i] != NULL)
1159                                         dput(tmp_subdirs->dentry[i]);
1160                         }
1161                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1162                 }
1163         case 2:
1164                 filp_close(filp, 0);
1165         case 1:
1166                 dput(dentry);
1167         }
1168         RETURN(rc);
1169 }
1170
1171 static int filter_read_groups(struct obd_device *obd, int last_group,
1172                               int create)
1173 {
1174         struct filter_obd *filter = &obd->u.filter;
1175         int old_count, group, rc = 0;
1176
1177         cfs_down(&filter->fo_init_lock);
1178         old_count = filter->fo_group_count;
1179         for (group = old_count; group <= last_group; group++) {
1180                 rc = filter_read_group_internal(obd, group, create);
1181                 if (rc != 0)
1182                         break;
1183         }
1184         cfs_up(&filter->fo_init_lock);
1185         return rc;
1186 }
1187
1188 /* FIXME: object groups */
1189 static int filter_prep_groups(struct obd_device *obd)
1190 {
1191         struct filter_obd *filter = &obd->u.filter;
1192         struct dentry *O_dentry;
1193         struct file *filp;
1194         int    last_group, rc = 0, cleanup_phase = 0;
1195         loff_t off = 0;
1196         ENTRY;
1197
1198         O_dentry = simple_mkdir(current->fs->pwd, obd->u.obt.obt_vfsmnt,
1199                                 "O", 0700, 1);
1200         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1201         if (IS_ERR(O_dentry)) {
1202                 rc = PTR_ERR(O_dentry);
1203                 CERROR("cannot open/create O: rc = %d\n", rc);
1204                 GOTO(cleanup, rc);
1205         }
1206         filter->fo_dentry_O = O_dentry;
1207         cleanup_phase = 1; /* O_dentry */
1208
1209         /* we have to initialize all groups before first connections from
1210          * clients because they may send create/destroy for any group -bzzz */
1211         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1212         if (IS_ERR(filp)) {
1213                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1214                 GOTO(cleanup, rc = PTR_ERR(filp));
1215         }
1216         cleanup_phase = 2; /* filp */
1217
1218         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1219         if (rc) {
1220                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1221                 GOTO(cleanup, rc);
1222         }
1223
1224         if (off == 0)
1225                 last_group = FID_SEQ_OST_MDT0;
1226
1227         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1228               FID_SEQ_OST_MDT0, last_group);
1229         filter->fo_committed_group = last_group;
1230         rc = filter_read_groups(obd, last_group, 1);
1231         if (rc)
1232                 GOTO(cleanup, rc);
1233
1234         filp_close(filp, 0);
1235         RETURN(0);
1236
1237  cleanup:
1238         switch (cleanup_phase) {
1239         case 2:
1240                 filp_close(filp, 0);
1241         case 1:
1242                 filter_cleanup_groups(obd);
1243                 f_dput(filter->fo_dentry_O);
1244                 filter->fo_dentry_O = NULL;
1245         default:
1246                 break;
1247         }
1248         return rc;
1249
1250 }
1251
1252 /* setup the object store with correct subdirectories */
1253 static int filter_prep(struct obd_device *obd)
1254 {
1255         struct lvfs_run_ctxt saved;
1256         struct filter_obd *filter = &obd->u.filter;
1257         struct file *file;
1258         struct inode *inode;
1259         int rc = 0;
1260         ENTRY;
1261
1262         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1263         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1264         if (!file || IS_ERR(file)) {
1265                 rc = PTR_ERR(file);
1266                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1267                        LAST_RCVD, rc);
1268                 GOTO(out, rc);
1269         }
1270         obd->u.obt.obt_rcvd_filp = file;
1271         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1272                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1273                        file->f_dentry->d_inode->i_mode);
1274                 GOTO(err_filp, rc = -ENOENT);
1275         }
1276
1277         inode = file->f_dentry->d_parent->d_inode;
1278         /* We use i_op->unlink directly in filter_vfs_unlink() */
1279         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1280                 CERROR("%s: filesystem does not support create/unlink ops\n",
1281                        obd->obd_name);
1282                 GOTO(err_filp, rc = -EOPNOTSUPP);
1283         }
1284
1285         rc = filter_init_server_data(obd, file);
1286         if (rc) {
1287                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1288                 GOTO(err_filp, rc);
1289         }
1290         LASSERT(obd->u.obt.obt_lut);
1291         target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1292
1293         /* open and create health check io file*/
1294         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1295         if (IS_ERR(file)) {
1296                 rc = PTR_ERR(file);
1297                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1298                        HEALTH_CHECK, rc);
1299                 GOTO(err_server_data, rc);
1300         }
1301         filter->fo_obt.obt_health_check_filp = file;
1302         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1303                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1304                        file->f_dentry->d_inode->i_mode);
1305                 GOTO(err_health_check, rc = -ENOENT);
1306         }
1307         rc = lvfs_check_io_health(obd, file);
1308         if (rc)
1309                 GOTO(err_health_check, rc);
1310
1311         rc = filter_prep_groups(obd);
1312         if (rc)
1313                 GOTO(err_health_check, rc);
1314 out:
1315         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1316
1317         return(rc);
1318
1319 err_health_check:
1320         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1321                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1322         filter->fo_obt.obt_health_check_filp = NULL;
1323 err_server_data:
1324         target_recovery_fini(obd);
1325         filter_free_server_data(&obd->u.obt);
1326 err_filp:
1327         if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1328                 CERROR("can't close %s after error\n", LAST_RCVD);
1329         obd->u.obt.obt_rcvd_filp = NULL;
1330         goto out;
1331 }
1332
1333 /* cleanup the filter: write last used object id to status file */
1334 static void filter_post(struct obd_device *obd)
1335 {
1336         struct lvfs_run_ctxt saved;
1337         struct filter_obd *filter = &obd->u.filter;
1338         int rc, i;
1339
1340         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1341          * best to start a transaction with h_sync, because we removed this
1342          * from lastobjid */
1343
1344         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1345         rc = filter_update_server_data(obd);
1346         if (rc)
1347                 CERROR("error writing server data: rc = %d\n", rc);
1348
1349         for (i = 0; i < filter->fo_group_count; i++) {
1350                 rc = filter_update_last_objid(obd, i,
1351                                 (i == filter->fo_group_count - 1));
1352                 if (rc)
1353                         CERROR("error writing group %d lastobjid: rc = %d\n",
1354                                i, rc);
1355         }
1356
1357         rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1358         obd->u.obt.obt_rcvd_filp = NULL;
1359         if (rc)
1360                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1361
1362         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1363         filter->fo_obt.obt_health_check_filp = NULL;
1364         if (rc)
1365                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1366
1367         filter_cleanup_groups(obd);
1368         filter_free_server_data(&obd->u.obt);
1369         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1370
1371         filter_free_capa_keys(filter);
1372         cleanup_capa_hash(filter->fo_capa_hash);
1373 }
1374
1375 static void filter_set_last_id(struct filter_obd *filter,
1376                                obd_id id, obd_seq group)
1377 {
1378         LASSERT(group <= filter->fo_group_count);
1379
1380         cfs_spin_lock(&filter->fo_objidlock);
1381         filter->fo_last_objids[group] = id;
1382         cfs_spin_unlock(&filter->fo_objidlock);
1383 }
1384
1385 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1386 {
1387         obd_id id;
1388         LASSERT(group <= filter->fo_group_count);
1389         LASSERT(filter->fo_last_objids != NULL);
1390
1391         /* FIXME: object groups */
1392         cfs_spin_lock(&filter->fo_objidlock);
1393         id = filter->fo_last_objids[group];
1394         cfs_spin_unlock(&filter->fo_objidlock);
1395         return id;
1396 }
1397
1398 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1399 {
1400         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1401         return 0;
1402 }
1403
1404 /* We never dget the object parent, so DON'T dput it either */
1405 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1406 {
1407         struct filter_obd *filter = &obd->u.filter;
1408         struct filter_subdirs *subdirs;
1409         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1410
1411         if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1412                 return filter->fo_dentry_O_groups[group];
1413
1414         subdirs = &filter->fo_dentry_O_sub[group];
1415         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1416 }
1417
1418 /* We never dget the object parent, so DON'T dput it either */
1419 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1420                                   obd_id objid)
1421 {
1422         unsigned long now = jiffies;
1423         struct dentry *dparent = filter_parent(obd, group, objid);
1424         int rc;
1425
1426         if (IS_ERR(dparent))
1427                 return dparent;
1428         if (dparent == NULL)
1429                 return ERR_PTR(-ENOENT);
1430
1431         rc = filter_lock_dentry(obd, dparent);
1432         fsfilt_check_slow(obd, now, "parent lock");
1433         return rc ? ERR_PTR(rc) : dparent;
1434 }
1435
1436 /* We never dget the object parent, so DON'T dput it either */
1437 static void filter_parent_unlock(struct dentry *dparent)
1438 {
1439         UNLOCK_INODE_MUTEX(dparent->d_inode);
1440 }
1441
1442 /* How to get files, dentries, inodes from object id's.
1443  *
1444  * If dir_dentry is passed, the caller has already locked the parent
1445  * appropriately for this operation (normally a write lock).  If
1446  * dir_dentry is NULL, we do a read lock while we do the lookup to
1447  * avoid races with create/destroy and such changing the directory
1448  * internal to the filesystem code. */
1449 struct dentry *filter_fid2dentry(struct obd_device *obd,
1450                                  struct dentry *dir_dentry,
1451                                  obd_seq group, obd_id id)
1452 {
1453         struct dentry *dparent = dir_dentry;
1454         struct dentry *dchild;
1455         char name[32];
1456         int len;
1457         ENTRY;
1458
1459         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1460             obd->u.filter.fo_destroys_in_progress == 0) {
1461                 /* don't fail lookups for orphan recovery, it causes
1462                  * later LBUGs when objects still exist during precreate */
1463                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1464                 RETURN(ERR_PTR(-ENOENT));
1465         }
1466         if (id == 0) {
1467                 CERROR("fatal: invalid object id 0\n");
1468                 RETURN(ERR_PTR(-ESTALE));
1469         }
1470
1471         len = sprintf(name, LPU64, id);
1472         if (dir_dentry == NULL) {
1473                 dparent = filter_parent_lock(obd, group, id);
1474                 if (IS_ERR(dparent)) {
1475                         CERROR("%s: error getting object "POSTID
1476                                " parent: rc %ld\n", obd->obd_name,
1477                                id, group, PTR_ERR(dparent));
1478                         RETURN(dparent);
1479                 }
1480         }
1481         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1482                dparent->d_name.len, dparent->d_name.name, name);
1483         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1484         if (dir_dentry == NULL)
1485                 filter_parent_unlock(dparent);
1486         if (IS_ERR(dchild)) {
1487                 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1488                        obd->obd_name, id, group, PTR_ERR(dchild));
1489                 RETURN(dchild);
1490         }
1491
1492         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1493                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1494                        obd->obd_name, id, dchild->d_inode->i_ino);
1495                 f_dput(dchild);
1496                 RETURN(ERR_PTR(-ENOENT));
1497         }
1498
1499         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1500                name, dchild, atomic_read(&dchild->d_count));
1501
1502         LASSERT(atomic_read(&dchild->d_count) > 0);
1503
1504         RETURN(dchild);
1505 }
1506
1507 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1508                                   obd_id group, struct lustre_handle *lockh)
1509 {
1510         int flags = LDLM_AST_DISCARD_DATA, rc;
1511         struct ldlm_res_id res_id;
1512         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1513         ENTRY;
1514
1515         osc_build_res_name(objid, group, &res_id);
1516         /* Tell the clients that the object is gone now and that they should
1517          * throw away any cached pages. */
1518         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1519                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1520                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1521                                     lockh);
1522         if (rc != ELDLM_OK)
1523                 lockh->cookie = 0;
1524         RETURN(rc);
1525 }
1526
1527 static void filter_fini_destroy(struct obd_device *obd,
1528                                 struct lustre_handle *lockh)
1529 {
1530         if (lustre_handle_is_used(lockh))
1531                 ldlm_lock_decref(lockh, LCK_PW);
1532 }
1533
1534 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1535  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1536  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1537  * reverse.  Caller must take i_sem before starting the transaction and we
1538  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1539 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1540                       struct vfsmount *mnt)
1541 {
1542         int rc;
1543         ENTRY;
1544
1545         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1546          * itself we already hold dir->i_mutex for child create/unlink ops */
1547         LASSERT(dentry->d_inode != NULL);
1548         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1549         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1550
1551
1552         /* may_delete() */
1553         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1554                 GOTO(out, rc = -ENOENT);
1555
1556         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1557         if (rc)
1558                 GOTO(out, rc);
1559
1560         if (IS_APPEND(dir))
1561                 GOTO(out, rc = -EPERM);
1562
1563         /* check_sticky() */
1564         if ((dentry->d_inode->i_uid != current->fsuid &&
1565              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1566             IS_IMMUTABLE(dentry->d_inode))
1567                 GOTO(out, rc = -EPERM);
1568
1569         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1570          *       that was done because of journal_start (which is already done
1571          *       here) or some other ordering issue. */
1572         DQUOT_INIT(dir);
1573
1574         rc = ll_security_inode_unlink(dir, dentry, mnt);
1575         if (rc)
1576                 GOTO(out, rc);
1577
1578         rc = dir->i_op->unlink(dir, dentry);
1579 out:
1580         /* need to drop i_mutex before we lose inode reference */
1581         UNLOCK_INODE_MUTEX(dentry->d_inode);
1582         if (rc == 0)
1583                 d_delete(dentry);
1584
1585         RETURN(rc);
1586 }
1587
1588 /* Caller must hold LCK_PW on parent and push us into kernel context.
1589  * Caller must hold child i_mutex, we drop it always.
1590  * Caller is also required to ensure that dchild->d_inode exists. */
1591 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1592                                    obd_seq group, struct dentry *dparent,
1593                                    struct dentry *dchild)
1594 {
1595         struct inode *inode = dchild->d_inode;
1596         int rc;
1597
1598         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1599                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1600                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1601                        (unsigned long)inode->i_nlink,
1602                        atomic_read(&inode->i_count));
1603         }
1604
1605         rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1606         if (rc)
1607                 CERROR("error unlinking objid %.*s: rc %d\n",
1608                        dchild->d_name.len, dchild->d_name.name, rc);
1609         return(rc);
1610 }
1611
1612 struct filter_intent_args {
1613         struct ldlm_lock **victim;
1614         __u64 size;
1615         int *liblustre;
1616 };
1617
1618 static enum interval_iter filter_intent_cb(struct interval_node *n,
1619                                            void *args)
1620 {
1621         struct ldlm_interval *node = (struct ldlm_interval *)n;
1622         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1623         __u64 size = arg->size;
1624         struct ldlm_lock **v = arg->victim;
1625         struct ldlm_lock *lck;
1626
1627         /* If the interval is lower than the current file size,
1628          * just break. */
1629         if (interval_high(n) <= size)
1630                 return INTERVAL_ITER_STOP;
1631
1632         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1633                 /* Don't send glimpse ASTs to liblustre clients.
1634                  * They aren't listening for them, and they do
1635                  * entirely synchronous I/O anyways. */
1636                 if (lck->l_export == NULL ||
1637                     lck->l_export->exp_libclient == 1)
1638                         continue;
1639
1640                 if (*arg->liblustre)
1641                         *arg->liblustre = 0;
1642
1643                 if (*v == NULL) {
1644                         *v = LDLM_LOCK_GET(lck);
1645                 } else if ((*v)->l_policy_data.l_extent.start <
1646                            lck->l_policy_data.l_extent.start) {
1647                         LDLM_LOCK_RELEASE(*v);
1648                         *v = LDLM_LOCK_GET(lck);
1649                 }
1650
1651                 /* the same policy group - every lock has the
1652                  * same extent, so needn't do it any more */
1653                 break;
1654         }
1655
1656         return INTERVAL_ITER_CONT;
1657 }
1658
1659 static int filter_intent_policy(struct ldlm_namespace *ns,
1660                                 struct ldlm_lock **lockp, void *req_cookie,
1661                                 ldlm_mode_t mode, int flags, void *data)
1662 {
1663         CFS_LIST_HEAD(rpc_list);
1664         struct ptlrpc_request *req = req_cookie;
1665         struct ldlm_lock *lock = *lockp, *l = NULL;
1666         struct ldlm_resource *res = lock->l_resource;
1667         ldlm_processing_policy policy;
1668         struct ost_lvb *res_lvb, *reply_lvb;
1669         struct ldlm_reply *rep;
1670         ldlm_error_t err;
1671         int idx, rc, tmpflags = 0, only_liblustre = 1;
1672         struct ldlm_interval_tree *tree;
1673         struct filter_intent_args arg;
1674         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1675                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1676                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1677         ENTRY;
1678
1679         policy = ldlm_get_processing_policy(res);
1680         LASSERT(policy != NULL);
1681         LASSERT(req != NULL);
1682
1683         rc = lustre_pack_reply(req, 3, repsize, NULL);
1684         if (rc)
1685                 RETURN(req->rq_status = rc);
1686
1687         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1688         LASSERT(rep != NULL);
1689
1690         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1691                                    sizeof(*reply_lvb));
1692         LASSERT(reply_lvb != NULL);
1693
1694         //fixup_handle_for_resent_req(req, lock, &lockh);
1695
1696         /* Call the extent policy function to see if our request can be
1697          * granted, or is blocked.
1698          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1699          * lock, and should not be granted if the lock will be blocked.
1700          */
1701
1702         LASSERT(ns == res->lr_namespace);
1703         lock_res(res);
1704         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1705         check_res_locked(res);
1706
1707         /* FIXME: we should change the policy function slightly, to not make
1708          * this list at all, since we just turn around and free it */
1709         while (!cfs_list_empty(&rpc_list)) {
1710                 struct ldlm_lock *wlock =
1711                         cfs_list_entry(rpc_list.next, struct ldlm_lock,
1712                                        l_cp_ast);
1713                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1714                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1715                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1716                 cfs_list_del_init(&wlock->l_cp_ast);
1717                 LDLM_LOCK_RELEASE(wlock);
1718         }
1719
1720         /* The lock met with no resistance; we're finished. */
1721         if (rc == LDLM_ITER_CONTINUE) {
1722                 /* do not grant locks to the liblustre clients: they cannot
1723                  * handle ASTs robustly.  We need to do this while still
1724                  * holding ns_lock to avoid the lock remaining on the res_link
1725                  * list (and potentially being added to l_pending_list by an
1726                  * AST) when we are going to drop this lock ASAP. */
1727                 if (lock->l_export->exp_libclient ||
1728                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1729                         ldlm_resource_unlink_lock(lock);
1730                         err = ELDLM_LOCK_ABORTED;
1731                 } else {
1732                         err = ELDLM_LOCK_REPLACED;
1733                 }
1734                 unlock_res(res);
1735                 RETURN(err);
1736         }
1737
1738         /* Do not grant any lock, but instead send GL callbacks.  The extent
1739          * policy nicely created a list of all PW locks for us.  We will choose
1740          * the highest of those which are larger than the size in the LVB, if
1741          * any, and perform a glimpse callback. */
1742         res_lvb = res->lr_lvb_data;
1743         LASSERT(res_lvb != NULL);
1744         *reply_lvb = *res_lvb;
1745
1746         /*
1747          * ->ns_lock guarantees that no new locks are granted, and,
1748          * therefore, that res->lr_lvb_data cannot increase beyond the
1749          * end of already granted lock. As a result, it is safe to
1750          * check against "stale" reply_lvb->lvb_size value without
1751          * res->lr_lvb_sem.
1752          */
1753         arg.size = reply_lvb->lvb_size;
1754         arg.victim = &l;
1755         arg.liblustre = &only_liblustre;
1756         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1757                 tree = &res->lr_itree[idx];
1758                 if (tree->lit_mode == LCK_PR)
1759                         continue;
1760
1761                 interval_iterate_reverse(tree->lit_root,
1762                                          filter_intent_cb, &arg);
1763         }
1764         unlock_res(res);
1765
1766         /* There were no PW locks beyond the size in the LVB; finished. */
1767         if (l == NULL) {
1768                 if (only_liblustre) {
1769                         /* If we discovered a liblustre client with a PW lock,
1770                          * however, the LVB may be out of date!  The LVB is
1771                          * updated only on glimpse (which we don't do for
1772                          * liblustre clients) and cancel (which the client
1773                          * obviously has not yet done).  So if it has written
1774                          * data but kept the lock, the LVB is stale and needs
1775                          * to be updated from disk.
1776                          *
1777                          * Of course, this will all disappear when we switch to
1778                          * taking liblustre locks on the OST. */
1779                         ldlm_res_lvbo_update(res, NULL, 1);
1780                 }
1781                 RETURN(ELDLM_LOCK_ABORTED);
1782         }
1783
1784         /*
1785          * This check is for lock taken in filter_prepare_destroy() that does
1786          * not have l_glimpse_ast set. So the logic is: if there is a lock
1787          * with no l_glimpse_ast set, this object is being destroyed already.
1788          *
1789          * Hence, if you are grabbing DLM locks on the server, always set
1790          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1791          */
1792         if (l->l_glimpse_ast == NULL) {
1793                 /* We are racing with unlink(); just return -ENOENT */
1794                 rep->lock_policy_res1 = -ENOENT;
1795                 goto out;
1796         }
1797
1798         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1799         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1800         /* Update the LVB from disk if the AST failed (this is a legal race) */
1801         /*
1802          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1803          * sending ast is not handled. This can result in lost client writes.
1804          */
1805         if (rc != 0)
1806                 ldlm_res_lvbo_update(res, NULL, 1);
1807
1808         lock_res(res);
1809         *reply_lvb = *res_lvb;
1810         unlock_res(res);
1811
1812  out:
1813         LDLM_LOCK_RELEASE(l);
1814
1815         RETURN(ELDLM_LOCK_ABORTED);
1816 }
1817
1818 /*
1819  * per-obd_device iobuf pool.
1820  *
1821  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1822  * allocations in write-path has to be minimized (see bug 5137).
1823  *
1824  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1825  * OST threads (see ost_thread_{init,done}()).
1826  *
1827  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1828  * at the OST layer there are only (potentially) multiple obd_device of type
1829  * unknown at the time of OST thread creation.
1830  *
1831  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1832  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1833  * it's very own iobuf.
1834  *
1835  * Functions below
1836  *
1837  *     filter_kiobuf_pool_init()
1838  *
1839  *     filter_kiobuf_pool_done()
1840  *
1841  *     filter_iobuf_get()
1842  *
1843  * operate on this array. They are "generic" in a sense that they don't depend
1844  * on actual type of iobuf's (the latter depending on Linux kernel version).
1845  */
1846
1847 /*
1848  * destroy pool created by filter_iobuf_pool_init
1849  */
1850 static void filter_iobuf_pool_done(struct filter_obd *filter)
1851 {
1852         struct filter_iobuf **pool;
1853         int i;
1854
1855         ENTRY;
1856
1857         pool = filter->fo_iobuf_pool;
1858         if (pool != NULL) {
1859                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1860                         if (pool[i] != NULL)
1861                                 filter_free_iobuf(pool[i]);
1862                 }
1863                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1864                 filter->fo_iobuf_pool = NULL;
1865         }
1866         EXIT;
1867 }
1868
1869 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1870 {
1871         struct filter_obd       *filter = &obd->u.filter;
1872         struct sptlrpc_rule_set  tmp_rset;
1873         int                      rc;
1874
1875         sptlrpc_rule_set_init(&tmp_rset);
1876         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1877         if (rc) {
1878                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1879                        obd->obd_name, rc);
1880                 return rc;
1881         }
1882
1883         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1884
1885         cfs_write_lock(&filter->fo_sptlrpc_lock);
1886         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1887         filter->fo_sptlrpc_rset = tmp_rset;
1888         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1889
1890         return 0;
1891 }
1892
1893 /*
1894  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1895  */
1896 static int filter_iobuf_pool_init(struct filter_obd *filter)
1897 {
1898         void **pool;
1899
1900         ENTRY;
1901
1902
1903         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1904                       GFP_KERNEL);
1905         if (filter->fo_iobuf_pool == NULL)
1906                 RETURN(-ENOMEM);
1907
1908         filter->fo_iobuf_count = OSS_THREADS_MAX;
1909
1910         RETURN(0);
1911 }
1912
1913 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1914  * many threads there will be so we allocate a large empty array and only
1915  * fill in those slots that are actually in use.
1916  * If we haven't allocated a pool entry for this thread before, do so now. */
1917 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1918 {
1919         int thread_id                    = (oti && oti->oti_thread) ?
1920                                            oti->oti_thread->t_id : -1;
1921         struct filter_iobuf  *pool       = NULL;
1922         struct filter_iobuf **pool_place = NULL;
1923
1924         if (thread_id >= 0) {
1925                 LASSERT(thread_id < filter->fo_iobuf_count);
1926                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1927         }
1928
1929         if (unlikely(pool == NULL)) {
1930                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1931                                           PTLRPC_MAX_BRW_PAGES);
1932                 if (pool_place != NULL)
1933                         *pool_place = pool;
1934         }
1935
1936         return pool;
1937 }
1938
1939 /* mount the file system (secretly).  lustre_cfg parameters are:
1940  * 1 = device
1941  * 2 = fstype
1942  * 3 = flags: failover=f, failout=n
1943  * 4 = mount options
1944  */
1945 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1946                         void *option)
1947 {
1948         struct filter_obd *filter = &obd->u.filter;
1949         struct vfsmount *mnt;
1950         struct lustre_mount_info *lmi;
1951         struct obd_uuid uuid;
1952         __u8 *uuid_ptr;
1953         char *str, *label;
1954         char ns_name[48];
1955         request_queue_t *q;
1956         int rc, i;
1957         ENTRY;
1958
1959         if (lcfg->lcfg_bufcount < 3 ||
1960             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1961             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1962                 RETURN(-EINVAL);
1963
1964         lmi = server_get_mount(obd->obd_name);
1965         if (lmi) {
1966                 /* We already mounted in lustre_fill_super.
1967                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1968                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1969                 mnt = lmi->lmi_mnt;
1970                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1971
1972                 /* gets recovery timeouts from mount data */
1973                 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1974                         obd->obd_recovery_timeout =
1975                                 lsi->lsi_lmd->lmd_recovery_time_soft;
1976                 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1977                         obd->obd_recovery_time_hard =
1978                                 lsi->lsi_lmd->lmd_recovery_time_hard;
1979         } else {
1980                 /* old path - used by lctl */
1981                 CERROR("Using old MDS mount method\n");
1982                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1983                                     MS_NOATIME|MS_NODIRATIME,
1984                                     lustre_cfg_string(lcfg, 1), option);
1985                 if (IS_ERR(mnt)) {
1986                         rc = PTR_ERR(mnt);
1987                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1988                                            lustre_cfg_string(lcfg, 1), rc);
1989                         RETURN(rc);
1990                 }
1991
1992                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1993         }
1994         if (IS_ERR(obd->obd_fsops))
1995                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1996
1997         rc = filter_iobuf_pool_init(filter);
1998         if (rc != 0)
1999                 GOTO(err_ops, rc);
2000
2001         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2002                 CERROR("%s: Underlying device is marked as read-only. "
2003                        "Setup failed\n", obd->obd_name);
2004                 GOTO(err_ops, rc = -EROFS);
2005         }
2006
2007         /* failover is the default */
2008         obd->obd_replayable = 1;
2009
2010         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2011                 str = lustre_cfg_string(lcfg, 3);
2012                 if (strchr(str, 'n')) {
2013                         CWARN("%s: recovery disabled\n", obd->obd_name);
2014                         obd->obd_replayable = 0;
2015                 }
2016         }
2017
2018         obd->u.obt.obt_vfsmnt = mnt;
2019         obd->u.obt.obt_sb = mnt->mnt_sb;
2020         obd->u.obt.obt_magic = OBT_MAGIC;
2021         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2022         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2023
2024         fsfilt_setup(obd, obd->u.obt.obt_sb);
2025
2026         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2027         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2028         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2029         obd->obd_lvfs_ctxt.fs = get_ds();
2030         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2031
2032         cfs_init_mutex(&filter->fo_init_lock);
2033         filter->fo_committed_group = 0;
2034         filter->fo_destroys_in_progress = 0;
2035         for (i = 0; i < 32; i++)
2036                 cfs_sema_init(&filter->fo_create_locks[i], 1);
2037
2038         cfs_spin_lock_init(&filter->fo_objidlock);
2039         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2040         cfs_sema_init(&filter->fo_alloc_lock, 1);
2041         init_brw_stats(&filter->fo_filter_stats);
2042         cfs_spin_lock_init(&filter->fo_flags_lock);
2043         filter->fo_read_cache = 1; /* enable read-only cache by default */
2044         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2045         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2046         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2047         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2048         filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2049         filter_slc_set(filter); /* initialize sync on lock cancel */
2050
2051         rc = filter_prep(obd);
2052         if (rc)
2053                 GOTO(err_ops, rc);
2054
2055         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2056         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2057
2058         filter->fo_fl_oss_capa = 1;
2059
2060         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2061         filter->fo_capa_hash = init_capa_hash();
2062         if (filter->fo_capa_hash == NULL)
2063                 GOTO(err_post, rc = -ENOMEM);
2064
2065         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2066         obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2067                                                 LDLM_NAMESPACE_GREEDY);
2068         if (obd->obd_namespace == NULL)
2069                 GOTO(err_post, rc = -ENOMEM);
2070         obd->obd_namespace->ns_lvbp = obd;
2071         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2072         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2073
2074         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2075                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2076
2077         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2078         if (rc) {
2079                 CERROR("failed to setup llogging subsystems\n");
2080                 GOTO(err_post, rc);
2081         }
2082
2083         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2084         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2085         /* do this after llog being initialized */
2086         filter_adapt_sptlrpc_conf(obd, 1);
2087
2088         rc = lquota_setup(filter_quota_interface_ref, obd);
2089         if (rc)
2090                 GOTO(err_post, rc);
2091
2092         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2093         if (q->max_sectors < q->max_hw_sectors &&
2094             q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2095                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2096                               "for larger I/O requests: max_sectors = %u "
2097                               "could be up to max_hw_sectors=%u\n",
2098                               obd->obd_name, mnt->mnt_sb->s_id,
2099                               q->max_sectors, q->max_hw_sectors);
2100
2101         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2102         if (uuid_ptr != NULL) {
2103                 class_uuid_unparse(uuid_ptr, &uuid);
2104                 str = uuid.uuid;
2105         } else {
2106                 str = "no UUID";
2107         }
2108
2109         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2110         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2111                       obd->obd_name, label ?: str, lmi ? "on " : "",
2112                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2113                       obd->obd_replayable ? "enabled" : "disabled");
2114
2115         if (obd->obd_recovering)
2116                 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2117                               "or until %d client%s reconnect%s\n",
2118                               obd->obd_name,
2119                               obd->obd_recovery_timeout / 60,
2120                               obd->obd_recovery_timeout % 60,
2121                               obd->obd_max_recoverable_clients,
2122                               (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2123                               (obd->obd_max_recoverable_clients == 1) ? "s": "");
2124
2125
2126         RETURN(0);
2127
2128 err_post:
2129         filter_post(obd);
2130 err_ops:
2131         fsfilt_put_ops(obd->obd_fsops);
2132         filter_iobuf_pool_done(filter);
2133 err_mntput:
2134         server_put_mount(obd->obd_name, mnt);
2135         obd->u.obt.obt_sb = 0;
2136         return rc;
2137 }
2138
2139 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2140 {
2141         struct lprocfs_static_vars lvars;
2142         unsigned long addr;
2143         struct page *page;
2144         int rc;
2145
2146         CLASSERT(offsetof(struct obd_device, u.obt) ==
2147                  offsetof(struct obd_device, u.filter.fo_obt));
2148
2149         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2150                 RETURN(-EINVAL);
2151
2152         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2153         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2154         if (!page)
2155                 RETURN(-ENOMEM);
2156         addr = (unsigned long)cfs_page_address(page);
2157         clear_page((void *)addr);
2158
2159         /* lprocfs must be setup before the filter so state can be safely added
2160          * to /proc incrementally as the filter is setup */
2161         lprocfs_filter_init_vars(&lvars);
2162         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2163             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2164                 /* Init obdfilter private stats here */
2165                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2166                                      LPROCFS_CNTR_AVGMINMAX,
2167                                      "read_bytes", "bytes");
2168                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2169                                      LPROCFS_CNTR_AVGMINMAX,
2170                                      "write_bytes", "bytes");
2171                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2172                                      LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2173                                      "get_page", "usec");
2174                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2175                                      LPROCFS_CNTR_AVGMINMAX,
2176                                      "get_page_failures", "num");
2177                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2178                                      LPROCFS_CNTR_AVGMINMAX,
2179                                      "cache_access", "pages");
2180                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2181                                      LPROCFS_CNTR_AVGMINMAX,
2182                                      "cache_hit", "pages");
2183                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2184                                      LPROCFS_CNTR_AVGMINMAX,
2185                                      "cache_miss", "pages");
2186
2187                 lproc_filter_attach_seqstat(obd);
2188                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2189                                                         obd->obd_proc_entry,
2190                                                         NULL, NULL);
2191                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2192                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2193                         CERROR("error %d setting up lprocfs for %s\n",
2194                                rc, "exports");
2195                         obd->obd_proc_exports_entry = NULL;
2196                 }
2197         }
2198         if (obd->obd_proc_exports_entry)
2199                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2200                                    lprocfs_nid_stats_clear_read,
2201                                    lprocfs_nid_stats_clear_write, obd, NULL);
2202
2203         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2204                LUSTRE_CFG_BUFLEN(lcfg, 4));
2205         rc = filter_common_setup(obd, lcfg, (void *)addr);
2206         OBD_PAGE_FREE(page);
2207
2208         if (rc) {
2209                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2210                 lprocfs_free_per_client_stats(obd);
2211                 lprocfs_free_obd_stats(obd);
2212                 lprocfs_obd_cleanup(obd);
2213         }
2214
2215         return rc;
2216 }
2217
2218 static struct llog_operations filter_mds_ost_repl_logops;
2219
2220 static struct llog_operations filter_size_orig_logops = {
2221         .lop_setup   = llog_obd_origin_setup,
2222         .lop_cleanup = llog_obd_origin_cleanup,
2223         .lop_add     = llog_obd_origin_add
2224 };
2225
2226 static int filter_olg_fini(struct obd_llog_group *olg)
2227 {
2228         struct llog_ctxt *ctxt;
2229         int rc = 0, rc2 = 0;
2230         ENTRY;
2231
2232         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2233         if (ctxt)
2234                 rc = llog_cleanup(ctxt);
2235
2236         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2237         if (ctxt) {
2238                 rc2 = llog_cleanup(ctxt);
2239                 if (!rc)
2240                         rc = rc2;
2241         }
2242
2243         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2244         if (ctxt) {
2245                 rc2 = llog_cleanup(ctxt);
2246                 if (!rc)
2247                         rc = rc2;
2248         }
2249
2250         RETURN(rc);
2251 }
2252
2253 static int
2254 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2255                 struct obd_device *tgt)
2256 {
2257         int rc;
2258         ENTRY;
2259
2260         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2261                         &filter_mds_ost_repl_logops);
2262         if (rc)
2263                 GOTO(cleanup, rc);
2264
2265         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2266                         &filter_size_orig_logops);
2267         if (rc)
2268                 GOTO(cleanup, rc);
2269         EXIT;
2270 cleanup:
2271         if (rc)
2272                 filter_olg_fini(olg);
2273         return rc;
2274 }
2275
2276 /**
2277  * Init the default olg, which is embeded in the obd_device, for filter.
2278  */
2279 static int
2280 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2281                         struct obd_device *tgt)
2282 {
2283         struct filter_obd *filter = &obd->u.filter;
2284         struct llog_ctxt *ctxt;
2285         int rc;
2286         ENTRY;
2287
2288         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2289         if (!filter->fo_lcm)
2290                 RETURN(-ENOMEM);
2291
2292         filter_mds_ost_repl_logops = llog_client_ops;
2293         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2294         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2295         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2296
2297         rc = filter_olg_init(obd, olg, tgt);
2298         if (rc)
2299                 GOTO(cleanup_lcm, rc);
2300
2301         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2302                         &llog_lvfs_ops);
2303         if (rc)
2304                 GOTO(cleanup_olg, rc);
2305
2306         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2307         if (!ctxt) {
2308                 CERROR("Can't get ctxt for %p:%x\n", olg,
2309                        LLOG_MDS_OST_REPL_CTXT);
2310                 GOTO(cleanup_olg, rc = -ENODEV);
2311         }
2312         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2313         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2314         llog_ctxt_put(ctxt);
2315
2316         RETURN(0);
2317 cleanup_olg:
2318         filter_olg_fini(olg);
2319 cleanup_lcm:
2320         llog_recov_thread_fini(filter->fo_lcm, 1);
2321         filter->fo_lcm = NULL;
2322         return rc;
2323 }
2324
2325 static int
2326 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2327                  struct obd_device *tgt, int *index)
2328 {
2329         struct filter_obd *filter = &obd->u.filter;
2330         struct llog_ctxt *ctxt;
2331         int rc;
2332         ENTRY;
2333
2334         LASSERT(olg != NULL);
2335         if (olg == &obd->obd_olg)
2336                 return filter_default_olg_init(obd, olg, tgt);
2337
2338         LASSERT(filter->fo_lcm != NULL);
2339         rc = filter_olg_init(obd, olg, tgt);
2340         if (rc)
2341                 RETURN(rc);
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                 filter_olg_fini(olg);
2347                 RETURN(-ENODEV);
2348         }
2349         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2350         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2351         llog_ctxt_put(ctxt);
2352         RETURN(rc);
2353 }
2354
2355 static int filter_llog_finish(struct obd_device *obd, int count)
2356 {
2357         struct filter_obd *filter = &obd->u.filter;
2358         struct llog_ctxt *ctxt;
2359         ENTRY;
2360
2361         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2362         if (ctxt) {
2363                 /*
2364                  * Make sure that no cached llcds left in recov_thread.
2365                  * We actually do sync in disconnect time, but disconnect
2366                  * may not come being marked rq_no_resend = 1.
2367                  */
2368                 llog_sync(ctxt, NULL);
2369
2370                 /*
2371                  * Balance class_import_get() in llog_receptor_accept().
2372                  * This is safe to do, as llog is already synchronized
2373                  * and its import may go.
2374                  */
2375                 cfs_mutex_down(&ctxt->loc_sem);
2376                 if (ctxt->loc_imp) {
2377                         class_import_put(ctxt->loc_imp);
2378                         ctxt->loc_imp = NULL;
2379                 }
2380                 cfs_mutex_up(&ctxt->loc_sem);
2381                 llog_ctxt_put(ctxt);
2382         }
2383
2384         if (filter->fo_lcm) {
2385                 cfs_mutex_down(&ctxt->loc_sem);
2386                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2387                 filter->fo_lcm = NULL;
2388                 cfs_mutex_up(&ctxt->loc_sem);
2389         }
2390         RETURN(filter_olg_fini(&obd->obd_olg));
2391 }
2392
2393 /**
2394  * Find the group llog according to group index in the llog group list.
2395  */
2396 static struct obd_llog_group *
2397 filter_find_olg_internal(struct filter_obd *filter, int group)
2398 {
2399         struct obd_llog_group *olg;
2400
2401         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2402         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2403                 if (olg->olg_seq == group)
2404                         RETURN(olg);
2405         }
2406         RETURN(NULL);
2407 }
2408
2409 /**
2410  * Find the group llog according to group index on the filter
2411  */
2412 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2413 {
2414         struct obd_llog_group *olg = NULL;
2415         struct filter_obd *filter;
2416
2417         filter = &obd->u.filter;
2418
2419         if (group == FID_SEQ_LLOG)
2420                 RETURN(&obd->obd_olg);
2421
2422         cfs_spin_lock(&filter->fo_llog_list_lock);
2423         olg = filter_find_olg_internal(filter, group);
2424         cfs_spin_unlock(&filter->fo_llog_list_lock);
2425
2426         RETURN(olg);
2427 }
2428 /**
2429  * Find the llog_group of the filter according to the group. If it can not
2430  * find, create the llog_group, which only happens when mds is being synced
2431  * with OST.
2432  */
2433 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2434 {
2435         struct obd_llog_group *olg = NULL;
2436         struct filter_obd *filter;
2437         int rc;
2438
2439         filter = &obd->u.filter;
2440
2441         if (group == FID_SEQ_LLOG)
2442                 RETURN(&obd->obd_olg);
2443
2444         cfs_spin_lock(&filter->fo_llog_list_lock);
2445         olg = filter_find_olg_internal(filter, group);
2446         if (olg) {
2447                 if (olg->olg_initializing) {
2448                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2449                 } else {
2450                         GOTO(out_unlock, olg);
2451                 }
2452         }
2453         OBD_ALLOC_PTR(olg);
2454         if (olg == NULL)
2455                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2456
2457         llog_group_init(olg, group);
2458         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2459         olg->olg_initializing = 1;
2460         cfs_spin_unlock(&filter->fo_llog_list_lock);
2461
2462         rc = obd_llog_init(obd, olg, obd, NULL);
2463         if (rc) {
2464                cfs_spin_lock(&filter->fo_llog_list_lock);
2465                cfs_list_del(&olg->olg_list);
2466                cfs_spin_unlock(&filter->fo_llog_list_lock);
2467                OBD_FREE_PTR(olg);
2468                GOTO(out, olg = ERR_PTR(-ENOMEM));
2469         }
2470         cfs_spin_lock(&filter->fo_llog_list_lock);
2471         olg->olg_initializing = 0;
2472         cfs_spin_unlock(&filter->fo_llog_list_lock);
2473         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2474               obd->obd_name, group, olg);
2475 out:
2476         RETURN(olg);
2477
2478 out_unlock:
2479         cfs_spin_unlock(&filter->fo_llog_list_lock);
2480         GOTO(out, olg);
2481 }
2482
2483 static int filter_llog_connect(struct obd_export *exp,
2484                                struct llogd_conn_body *body)
2485 {
2486         struct obd_device *obd = exp->exp_obd;
2487         struct llog_ctxt *ctxt;
2488         struct obd_llog_group *olg;
2489         int rc;
2490         ENTRY;
2491
2492         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2493                obd->obd_name, body->lgdc_logid.lgl_oid,
2494                body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2495
2496         olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2497         if (!olg) {
2498                 CERROR(" %s: can not find olg of group %d\n",
2499                        obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2500                 RETURN(-ENOENT);
2501         }
2502         llog_group_set_export(olg, exp);
2503
2504         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2505         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2506                  body->lgdc_ctxt_idx);
2507
2508         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2509               obd->obd_name, body->lgdc_logid.lgl_oid,
2510               body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2511
2512         cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2513         obd->u.filter.fo_mds_ost_sync = 1;
2514         cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2515         rc = llog_connect(ctxt, &body->lgdc_logid,
2516                           &body->lgdc_gen, NULL);
2517         llog_ctxt_put(ctxt);
2518         if (rc != 0)
2519                 CERROR("failed to connect rc %d idx %d\n", rc,
2520                                 body->lgdc_ctxt_idx);
2521
2522         RETURN(rc);
2523 }
2524
2525 static int filter_llog_preclean(struct obd_device *obd)
2526 {
2527         struct obd_llog_group *olg, *tmp;
2528         struct filter_obd *filter;
2529         cfs_list_t  remove_list;
2530         int rc = 0;
2531         ENTRY;
2532
2533         rc = obd_llog_finish(obd, 0);
2534         if (rc)
2535                 CERROR("failed to cleanup llogging subsystem\n");
2536
2537         filter = &obd->u.filter;
2538         CFS_INIT_LIST_HEAD(&remove_list);
2539
2540         cfs_spin_lock(&filter->fo_llog_list_lock);
2541         while (!cfs_list_empty(&filter->fo_llog_list)) {
2542                 olg = cfs_list_entry(filter->fo_llog_list.next,
2543                                      struct obd_llog_group, olg_list);
2544                 cfs_list_del(&olg->olg_list);
2545                 cfs_list_add(&olg->olg_list, &remove_list);
2546         }
2547         cfs_spin_unlock(&filter->fo_llog_list_lock);
2548
2549         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2550                 cfs_list_del_init(&olg->olg_list);
2551                 rc = filter_olg_fini(olg);
2552                 if (rc)
2553                         CERROR("failed to cleanup llogging subsystem for %u\n",
2554                                olg->olg_seq);
2555                 OBD_FREE_PTR(olg);
2556         }
2557
2558         RETURN(rc);
2559 }
2560
2561 static int filter_precleanup(struct obd_device *obd,
2562                              enum obd_cleanup_stage stage)
2563 {
2564         int rc = 0;
2565         ENTRY;
2566
2567         switch(stage) {
2568         case OBD_CLEANUP_EARLY:
2569                 break;
2570         case OBD_CLEANUP_EXPORTS:
2571                 /* Stop recovery before namespace cleanup. */
2572                 target_recovery_fini(obd);
2573                 rc = filter_llog_preclean(obd);
2574                 break;
2575         }
2576         RETURN(rc);
2577 }
2578
2579 static int filter_cleanup(struct obd_device *obd)
2580 {
2581         struct filter_obd *filter = &obd->u.filter;
2582         ENTRY;
2583
2584         if (obd->obd_fail)
2585                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2586                               "will be preserved.\n", obd->obd_name);
2587
2588         obd_exports_barrier(obd);
2589         obd_zombie_barrier();
2590
2591         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2592         lprocfs_free_per_client_stats(obd);
2593         lprocfs_free_obd_stats(obd);
2594         lprocfs_obd_cleanup(obd);
2595         lquota_cleanup(filter_quota_interface_ref, obd);
2596
2597         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2598         obd->obd_namespace = NULL;
2599
2600         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2601
2602         if (obd->u.obt.obt_sb == NULL)
2603                 RETURN(0);
2604
2605         filter_post(obd);
2606
2607         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2608         shrink_dcache_sb(obd->u.obt.obt_sb);
2609
2610         server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2611         obd->u.obt.obt_sb = NULL;
2612
2613         fsfilt_put_ops(obd->obd_fsops);
2614
2615         filter_iobuf_pool_done(filter);
2616
2617         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2618
2619         RETURN(0);
2620 }
2621
2622 static int filter_connect_internal(struct obd_export *exp,
2623                                    struct obd_connect_data *data,
2624                                    int reconnect)
2625 {
2626         struct filter_export_data *fed = &exp->exp_filter_data;
2627
2628         if (!data)
2629                 RETURN(0);
2630
2631         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2632                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2633                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2634                data->ocd_connect_flags, data->ocd_version,
2635                data->ocd_grant, data->ocd_index);
2636
2637         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2638                 CWARN("!!! This export (nid %s) used object group %d "
2639                        "earlier; now it's trying to use group %d!  This could "
2640                        "be a bug in the MDS. Please report to "
2641                        "http://bugzilla.lustre.org/\n",
2642                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2643                 RETURN(-EPROTO);
2644         }
2645         fed->fed_group = data->ocd_group;
2646
2647         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2648         exp->exp_connect_flags = data->ocd_connect_flags;
2649         data->ocd_version = LUSTRE_VERSION_CODE;
2650
2651         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2652         if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2653                   data->ocd_connect_flags & OBD_CONNECT_MDS))
2654                 RETURN(-EPROTO);
2655
2656         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2657                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2658                 obd_size left, want;
2659
2660                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2661                 left = filter_grant_space_left(exp);
2662                 want = data->ocd_grant;
2663                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2664                 data->ocd_grant = fed->fed_grant;
2665                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2666
2667                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2668                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2669                        exp->exp_client_uuid.uuid, exp,
2670                        data->ocd_grant, want, left);
2671
2672                 filter->fo_tot_granted_clients ++;
2673         }
2674
2675         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2676                 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2677                 int index = le32_to_cpu(lsd->lsd_ost_index);
2678
2679                 if (!(lsd->lsd_feature_compat &
2680                       cpu_to_le32(OBD_COMPAT_OST))) {
2681                         /* this will only happen on the first connect */
2682                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2683                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2684                         /* sync is not needed here as filter_client_add will
2685                          * set exp_need_sync flag */
2686                         filter_update_server_data(exp->exp_obd);
2687                 } else if (index != data->ocd_index) {
2688                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2689                                            " %u doesn't match actual OST index"
2690                                            " %u in last_rcvd file, bad "
2691                                            "configuration?\n",
2692                                            obd_export_nid2str(exp), index,
2693                                            data->ocd_index);
2694                         RETURN(-EBADF);
2695                 }
2696                 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2697                  * FIXME: We don't strictly need the COMPAT flag for that,
2698                  * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2699                  * FIXME: We needed it for the index, as index 0 is valid. */
2700         }
2701
2702         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2703                 data->ocd_brw_size = 65536;
2704         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2705                 data->ocd_brw_size = min(data->ocd_brw_size,
2706                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2707                 LASSERT(data->ocd_brw_size);
2708         }
2709
2710         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2711                 __u32 cksum_types = data->ocd_cksum_types;
2712
2713                 /* The client set in ocd_cksum_types the checksum types it
2714                  * supports. We have to mask off the algorithms that we don't
2715                  * support */
2716                 if (cksum_types & OBD_CKSUM_ALL)
2717                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2718                 else
2719                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2720
2721                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2722                                    "%x\n", exp->exp_obd->obd_name,
2723                                    obd_export_nid2str(exp), cksum_types,
2724                                    data->ocd_cksum_types);
2725         } else {
2726                 /* This client does not support OBD_CONNECT_CKSUM
2727                  * fall back to CRC32 */
2728                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2729                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2730                                    exp->exp_obd->obd_name,
2731                                    obd_export_nid2str(exp));
2732         }
2733
2734         RETURN(0);
2735 }
2736
2737 static int filter_reconnect(const struct lu_env *env,
2738                             struct obd_export *exp, struct obd_device *obd,
2739                             struct obd_uuid *cluuid,
2740                             struct obd_connect_data *data,
2741                             void *localdata)
2742 {
2743         int rc;
2744         ENTRY;
2745
2746         if (exp == NULL || obd == NULL || cluuid == NULL)
2747                 RETURN(-EINVAL);
2748
2749         rc = filter_connect_internal(exp, data, 1);
2750         if (rc == 0)
2751                 filter_export_stats_init(obd, exp, 1, localdata);
2752
2753         RETURN(rc);
2754 }
2755
2756 /* nearly identical to mds_connect */
2757 static int filter_connect(const struct lu_env *env,
2758                           struct obd_export **exp, struct obd_device *obd,
2759                           struct obd_uuid *cluuid,
2760                           struct obd_connect_data *data, void *localdata)
2761 {
2762         struct lvfs_run_ctxt saved;
2763         struct lustre_handle conn = { 0 };
2764         struct obd_export *lexp;
2765         __u32 group;
2766         int rc;
2767         ENTRY;
2768
2769         if (exp == NULL || obd == NULL || cluuid == NULL)
2770                 RETURN(-EINVAL);
2771
2772         rc = class_connect(&conn, obd, cluuid);
2773         if (rc)
2774                 RETURN(rc);
2775         lexp = class_conn2export(&conn);
2776         LASSERT(lexp != NULL);
2777
2778         rc = filter_connect_internal(lexp, data, 0);
2779         if (rc)
2780                 GOTO(cleanup, rc);
2781
2782         filter_export_stats_init(obd, lexp, 0, localdata);
2783         if (obd->obd_replayable) {
2784                 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2785                 LASSERT(lcd);
2786                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2787                 rc = filter_client_add(obd, lexp, -1);
2788                 if (rc)
2789                         GOTO(cleanup, rc);
2790         }
2791
2792         group = data->ocd_group;
2793
2794         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2795               obd->obd_name, lexp->exp_handle.h_cookie, group);
2796
2797         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2798         rc = filter_read_groups(obd, group, 1);
2799         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2800         if (rc != 0) {
2801                 CERROR("can't read group %u\n", group);
2802                 GOTO(cleanup, rc);
2803         }
2804
2805         GOTO(cleanup, rc);
2806
2807 cleanup:
2808         if (rc) {
2809                 class_disconnect(lexp);
2810                 *exp = NULL;
2811         } else {
2812                 *exp = lexp;
2813         }
2814
2815         RETURN(rc);
2816 }
2817
2818 /* Do extra sanity checks for grant accounting.  We do this at connect,
2819  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2820  * always get rid of it or turn it off when we know accounting is good. */
2821 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2822 {
2823         struct filter_export_data *fed;
2824         struct obd_export *exp;
2825         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2826         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2827         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2828
2829         if (cfs_list_empty(&obd->obd_exports))
2830                 return;
2831
2832         /* We don't want to do this for large machines that do lots of
2833            mounts or unmounts.  It burns... */
2834         if (obd->obd_num_exports > 100)
2835                 return;
2836
2837         cfs_spin_lock(&obd->obd_osfs_lock);
2838         cfs_spin_lock(&obd->obd_dev_lock);
2839         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2840                 int error = 0;
2841                 fed = &exp->exp_filter_data;
2842                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2843                     fed->fed_dirty < 0)
2844                         error = 1;
2845                 if (maxsize > 0) { /* we may not have done a statfs yet */
2846                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2847                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2848                                  exp->exp_client_uuid.uuid, exp,
2849                                  fed->fed_grant, fed->fed_pending, maxsize);
2850                         LASSERTF(fed->fed_dirty <= maxsize,
2851                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2852                                  exp->exp_client_uuid.uuid, exp,
2853                                  fed->fed_dirty, maxsize);
2854                 }
2855                 if (error)
2856                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2857                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2858                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2859                 else
2860                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2861                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2862                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2863                 tot_granted += fed->fed_grant + fed->fed_pending;
2864                 tot_pending += fed->fed_pending;
2865                 tot_dirty += fed->fed_dirty;
2866         }
2867         fo_tot_granted = obd->u.filter.fo_tot_granted;
2868         fo_tot_pending = obd->u.filter.fo_tot_pending;
2869         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2870         cfs_spin_unlock(&obd->obd_dev_lock);
2871         cfs_spin_unlock(&obd->obd_osfs_lock);
2872
2873         /* Do these assertions outside the spinlocks so we don't kill system */
2874         if (tot_granted != fo_tot_granted)
2875                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2876                        func, tot_granted, fo_tot_granted);
2877         if (tot_pending != fo_tot_pending)
2878                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2879                        func, tot_pending, fo_tot_pending);
2880         if (tot_dirty != fo_tot_dirty)
2881                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2882                        func, tot_dirty, fo_tot_dirty);
2883         if (tot_pending > tot_granted)
2884                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2885                        func, tot_pending, tot_granted);
2886         if (tot_granted > maxsize)
2887                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2888                        func, tot_granted, maxsize);
2889         if (tot_dirty > maxsize)
2890                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2891                        func, tot_dirty, maxsize);
2892 }
2893
2894 /* Remove this client from the grant accounting totals.  We also remove
2895  * the export from the obd device under the osfs and dev locks to ensure
2896  * that the filter_grant_sanity_check() calculations are always valid.
2897  * The client should do something similar when it invalidates its import. */
2898 static void filter_grant_discard(struct obd_export *exp)
2899 {
2900         struct obd_device *obd = exp->exp_obd;
2901         struct filter_obd *filter = &obd->u.filter;
2902         struct filter_export_data *fed = &exp->exp_filter_data;
2903
2904         cfs_spin_lock(&obd->obd_osfs_lock);
2905         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2906                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2907                  obd->obd_name, filter->fo_tot_granted,
2908                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2909         filter->fo_tot_granted -= fed->fed_grant;
2910         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2911                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2912                  obd->obd_name, filter->fo_tot_pending,
2913                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2914         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2915         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2916                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2917                  obd->obd_name, filter->fo_tot_dirty,
2918                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2919         filter->fo_tot_dirty -= fed->fed_dirty;
2920         fed->fed_dirty = 0;
2921         fed->fed_grant = 0;
2922
2923         cfs_spin_unlock(&obd->obd_osfs_lock);
2924 }
2925
2926 static int filter_destroy_export(struct obd_export *exp)
2927 {
2928         struct filter_export_data *fed = &exp->exp_filter_data;
2929         ENTRY;
2930
2931         if (fed->fed_pending)
2932                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2933                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2934                        exp, fed->fed_pending);
2935
2936         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2937
2938         target_destroy_export(exp);
2939         ldlm_destroy_export(exp);
2940         lut_client_free(exp);
2941
2942         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2943                 RETURN(0);
2944
2945         if (!exp->exp_obd->obd_replayable)
2946                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2947
2948         filter_grant_discard(exp);
2949         filter_fmd_cleanup(exp);
2950
2951         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2952                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2953                 if (filter->fo_tot_granted_clients > 0)
2954                         filter->fo_tot_granted_clients --;
2955         }
2956
2957         if (!(exp->exp_flags & OBD_OPT_FORCE))
2958                 filter_grant_sanity_check(exp->exp_obd, __func__);
2959
2960         RETURN(0);
2961 }
2962
2963 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2964 {
2965         struct obd_llog_group *olg_min, *olg;
2966         struct filter_obd *filter;
2967         int worked = -1, group;
2968         struct llog_ctxt *ctxt;
2969         ENTRY;
2970
2971         filter = &obd->u.filter;
2972
2973         /* we can't sync log holding spinlock. also, we do not want to get
2974          * into livelock. so we do following: loop over MDS's exports in
2975          * group order and skip already synced llogs -bzzz */
2976         do {
2977                 /* look for group with min. number, but > worked */
2978                 olg_min = NULL;
2979                 group = 1 << 30;
2980                 cfs_spin_lock(&filter->fo_llog_list_lock);
2981                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2982                         if (olg->olg_seq <= worked) {
2983                                 /* this group is already synced */
2984                                 continue;
2985                         }
2986                         if (group < olg->olg_seq) {
2987                                 /* we have group with smaller number to sync */
2988                                 continue;
2989                         }
2990                         /* store current minimal group */
2991                         olg_min = olg;
2992                         group = olg->olg_seq;
2993                 }
2994                 cfs_spin_unlock(&filter->fo_llog_list_lock);
2995
2996                 if (olg_min == NULL)
2997                         break;
2998
2999                 worked = olg_min->olg_seq;
3000                 if (olg_min->olg_exp &&
3001                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3002                         int err;
3003                         ctxt = llog_group_get_ctxt(olg_min,
3004                                                    LLOG_MDS_OST_REPL_CTXT);
3005                         if (ctxt) {
3006                                 err = llog_sync(ctxt, olg_min->olg_exp);
3007                                 llog_ctxt_put(ctxt);
3008                                 if (err) {
3009                                         CERROR("error flushing logs to MDS: "
3010                                                "rc %d\n", err);
3011                                 }
3012                         }
3013                 }
3014         } while (olg_min != NULL);
3015 }
3016
3017 /* Also incredibly similar to mds_disconnect */
3018 static int filter_disconnect(struct obd_export *exp)
3019 {
3020         struct obd_device *obd = exp->exp_obd;
3021         int rc;
3022         ENTRY;
3023
3024         LASSERT(exp);
3025         class_export_get(exp);
3026
3027         if (!(exp->exp_flags & OBD_OPT_FORCE))
3028                 filter_grant_sanity_check(obd, __func__);
3029         filter_grant_discard(exp);
3030
3031         /* Flush any remaining cancel messages out to the target */
3032         filter_sync_llogs(obd, exp);
3033
3034         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3035
3036         rc = server_disconnect_export(exp);
3037
3038         /* Do not erase record for recoverable client. */
3039         if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3040                 filter_client_del(exp);
3041         else
3042                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3043
3044         class_export_put(exp);
3045         RETURN(rc);
3046 }
3047
3048 /* reverse import is changed, sync all cancels */
3049 static void filter_revimp_update(struct obd_export *exp)
3050 {
3051         ENTRY;
3052
3053         LASSERT(exp);
3054         class_export_get(exp);
3055
3056         /* flush any remaining cancel messages out to the target */
3057         filter_sync_llogs(exp->exp_obd, exp);
3058         class_export_put(exp);
3059         EXIT;
3060 }
3061
3062 static int filter_ping(struct obd_export *exp)
3063 {
3064         filter_fmd_expire(exp);
3065         return 0;
3066 }
3067
3068 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3069                                   const char *what, int quiet)
3070 {
3071         struct dentry *dchild = NULL;
3072
3073         dchild = filter_fid2dentry(obd, NULL,  ostid->oi_seq, ostid->oi_id);
3074
3075         if (IS_ERR(dchild)) {
3076                 CERROR("%s error looking up object: "POSTID"\n",
3077                        what, ostid->oi_id, ostid->oi_seq);
3078                 RETURN(dchild);
3079         }
3080
3081         if (dchild->d_inode == NULL) {
3082                 if (!quiet)
3083                         CERROR("%s: %s on non-existent object: "POSTID" \n",
3084                                obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3085                 f_dput(dchild);
3086                 RETURN(ERR_PTR(-ENOENT));
3087         }
3088
3089         return dchild;
3090 }
3091
3092 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3093 {
3094         struct dentry *dentry = NULL;
3095         struct obd_device *obd;
3096         int rc = 0;
3097         ENTRY;
3098
3099         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3100                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3101         if (rc)
3102                 RETURN(rc);
3103
3104         obd = class_exp2obd(exp);
3105         if (obd == NULL) {
3106                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3107                 RETURN(-EINVAL);
3108         }
3109
3110         dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3111         if (IS_ERR(dentry))
3112                 RETURN(PTR_ERR(dentry));
3113
3114         /* Limit the valid bits in the return data to what we actually use */
3115         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3116         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3117
3118         f_dput(dentry);
3119         RETURN(rc);
3120 }
3121
3122 /* this should be enabled/disabled in condition to enabled/disabled large
3123  * inodes (fast EAs) in backing store FS. */
3124 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3125                         void *handle, struct obdo *oa)
3126 {
3127         struct obd_device *obd = exp->exp_obd;
3128         int rc = 0;
3129         ENTRY;
3130
3131         if (oa->o_valid & OBD_MD_FLFID) {
3132                 struct filter_fid ff;
3133
3134                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3135                         oa->o_seq = 0;
3136                 /* packing fid and converting it to LE for storing into EA.
3137                  * Here ->o_stripe_idx should be filled by LOV and rest of
3138                  * fields - by client. */
3139                 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3140                 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3141                 /* XXX: we are ignoring o_parent_ver here, since this should
3142                  *      be the same for all objects in this fileset. */
3143                 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3144                 ff.ff_objid = cpu_to_le64(oa->o_id);
3145                 ff.ff_seq = cpu_to_le64(oa->o_seq);
3146
3147                 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3148                        LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3149                        oa->o_seq);
3150
3151                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3152                 if (rc)
3153                         CERROR("store fid in object failed! rc: %d\n", rc);
3154         } else {
3155                 CDEBUG(D_HA, "OSS object without fid info!\n");
3156         }
3157
3158         RETURN(rc);
3159 }
3160
3161 /* this is called from filter_truncate() until we have filter_punch() */
3162 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3163                             struct obdo *oa, struct obd_trans_info *oti)
3164 {
3165         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3166         struct llog_cookie *fcc = NULL;
3167         struct filter_obd *filter;
3168         int rc, err, sync = 0;
3169         loff_t old_size = 0;
3170         unsigned int ia_valid;
3171         struct inode *inode;
3172         struct page *page = NULL;
3173         struct iattr iattr;
3174         void *handle;
3175         ENTRY;
3176
3177         LASSERT(dentry != NULL);
3178         LASSERT(!IS_ERR(dentry));
3179
3180         inode = dentry->d_inode;
3181         LASSERT(inode != NULL);
3182
3183         filter = &exp->exp_obd->u.filter;
3184         iattr_from_obdo(&iattr, oa, oa->o_valid);
3185         ia_valid = iattr.ia_valid;
3186
3187         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3188                 OBD_ALLOC(fcc, sizeof(*fcc));
3189                 if (fcc != NULL)
3190                         *fcc = oa->o_lcookie;
3191         }
3192         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3193                 DQUOT_INIT(inode);
3194                 /* Filter truncates and writes are serialized by
3195                  * i_alloc_sem, see the comment in
3196                  * filter_preprw_write.*/
3197                 if (ia_valid & ATTR_SIZE)
3198                         down_write(&inode->i_alloc_sem);
3199                 LOCK_INODE_MUTEX(inode);
3200                 old_size = i_size_read(inode);
3201         }
3202
3203         /* VBR: ve