Whamcloud - gitweb
LU-694 ptlrpc: Job Stats
[fs/lustre-release.git] / lustre / obdfilter / filter.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
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_lock(&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_unlock(&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                 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
156         } else {
157                 last_rcvd = oti->oti_transno;
158                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
159                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
160                 if (unlikely(last_rcvd < le64_to_cpu(lcd->lcd_last_transno))) {
161                         CERROR("Trying to overwrite bigger transno, on-disk: "
162                                LPU64", new: "LPU64"\n",
163                                le64_to_cpu(lcd->lcd_last_transno), last_rcvd);
164                         cfs_spin_lock(&exp->exp_lock);
165                         exp->exp_vbr_failed = 1;
166                         cfs_spin_unlock(&exp->exp_lock);
167                         cfs_spin_unlock(&obt->obt_lut->lut_translock);
168                         cfs_mutex_unlock(&ted->ted_lcd_lock);
169                         RETURN(-EOVERFLOW);
170                 }
171         }
172         oti->oti_transno = last_rcvd;
173
174         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
175         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
176         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
177         cfs_spin_unlock(&obt->obt_lut->lut_translock);
178
179         if (inode)
180                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
181
182         off = ted->ted_lr_off;
183         if (off <= 0) {
184                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
185                        ted->ted_lr_idx, ted->ted_lr_off);
186                 err = -EINVAL;
187         } else {
188                 class_export_cb_get(exp); /* released when the cb is called */
189                 if (!force_sync)
190                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
191                                                            last_rcvd,
192                                                            oti->oti_handle,
193                                                            filter_commit_cb,
194                                                            exp);
195
196                 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
197                                           lcd, sizeof(*lcd), &off,
198                                           force_sync | exp->exp_need_sync);
199                 if (force_sync)
200                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
201         }
202         if (err) {
203                 log_pri = D_ERROR;
204                 if (rc == 0)
205                         rc = err;
206         }
207
208         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
209                last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
210         cfs_mutex_unlock(&ted->ted_lcd_lock);
211         RETURN(rc);
212 }
213
214 void f_dput(struct dentry *dentry)
215 {
216         /* Can't go inside filter_ddelete because it can block */
217         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
218                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
219         LASSERT(atomic_read(&dentry->d_count) > 0);
220
221         dput(dentry);
222 }
223
224 static void init_brw_stats(struct brw_stats *brw_stats)
225 {
226         int i;
227         for (i = 0; i < BRW_LAST; i++)
228                 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
229 }
230
231 static int lprocfs_init_rw_stats(struct obd_device *obd,
232                                  struct lprocfs_stats **stats)
233 {
234         int num_stats;
235
236         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
237                                                         LPROC_FILTER_LAST - 1;
238         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
239         if (*stats == NULL)
240                 return -ENOMEM;
241
242         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
243         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
244                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
245         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
246                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
247
248         return(0);
249 }
250
251 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
252    plus the procfs overhead :( */
253 static int filter_export_stats_init(struct obd_device *obd,
254                                     struct obd_export *exp,
255                                     void *client_nid)
256 {
257         int rc, newnid = 0;
258         ENTRY;
259
260         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
261                 /* Self-export gets no proc entry */
262                 RETURN(0);
263
264         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
265         if (rc) {
266                 /* Mask error for already created
267                  * /proc entries */
268                 if (rc == -EALREADY)
269                         rc = 0;
270                 RETURN(rc);
271         }
272
273         if (newnid) {
274                 struct nid_stat *tmp = exp->exp_nid_stats;
275                 LASSERT(tmp != NULL);
276
277                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
278                 if (tmp->nid_brw_stats == NULL)
279                         GOTO(clean, rc = -ENOMEM);
280
281                 init_brw_stats(tmp->nid_brw_stats);
282                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
283                                         0644, &filter_per_nid_stats_fops,
284                                         exp->exp_nid_stats);
285                 if (rc)
286                         CWARN("Error adding the brw_stats file\n");
287
288                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
289                 if (rc)
290                         GOTO(clean, rc);
291
292                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
293                                             tmp->nid_stats);
294                 if (rc)
295                         GOTO(clean, rc);
296                 rc = lprocfs_nid_ldlm_stats_init(tmp);
297                 if (rc)
298                         GOTO(clean, rc);
299         }
300
301         RETURN(0);
302  clean:
303         return rc;
304 }
305
306 /* Add client data to the FILTER.  We use a bitmap to locate a free space
307  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
308  * Otherwise, we have just read the data from the last_rcvd file and
309  * we know its offset. */
310 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
311                              int cl_idx)
312 {
313         struct obd_device_target *obt = &obd->u.obt;
314         struct tg_export_data *ted = &exp->exp_target_data;
315         struct lr_server_data *lsd = class_server_data(obd);
316         unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
317         int new_client = (cl_idx == -1);
318
319         ENTRY;
320
321         LASSERT(bitmap != NULL);
322         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
323
324         /* Self-export */
325         if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
326                 RETURN(0);
327
328         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
329          * there's no need for extra complication here
330          */
331         if (new_client) {
332                 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
333         repeat:
334                 if (cl_idx >= LR_MAX_CLIENTS) {
335                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
336                                cl_idx);
337                         RETURN(-EOVERFLOW);
338                 }
339                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
340                         cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
341                                                         cl_idx);
342                         goto repeat;
343                 }
344         } else {
345                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
346                         CERROR("FILTER client %d: bit already set in bitmap!\n",
347                                cl_idx);
348                         LBUG();
349                 }
350         }
351
352         ted->ted_lr_idx = cl_idx;
353         ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
354                           cl_idx * le16_to_cpu(lsd->lsd_client_size);
355         cfs_mutex_init(&ted->ted_lcd_lock);
356         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
357
358         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
359                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
360
361         if (new_client) {
362                 struct lvfs_run_ctxt saved;
363                 loff_t off = ted->ted_lr_off;
364                 int rc;
365                 void *handle;
366
367                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
368                        ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
369
370                 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
371                         RETURN(-ENOSPC);
372
373                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
374                 /* Transaction needed to fix bug 1403 */
375                 handle = fsfilt_start(obd,
376                                       obt->obt_rcvd_filp->f_dentry->d_inode,
377                                       FSFILT_OP_SETATTR, NULL);
378                 if (IS_ERR(handle)) {
379                         rc = PTR_ERR(handle);
380                         CERROR("unable to start transaction: rc %d\n", rc);
381                 } else {
382                         ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
383                         exp->exp_last_request_time = cfs_time_current_sec();
384                         rc = fsfilt_add_journal_cb(obd, 0, handle,
385                                                    target_client_add_cb,
386                                                    class_export_cb_get(exp));
387                         if (rc == 0) {
388                                 cfs_spin_lock(&exp->exp_lock);
389                                 exp->exp_need_sync = 1;
390                                 cfs_spin_unlock(&exp->exp_lock);
391                         }
392                         rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
393                                                  ted->ted_lcd,
394                                                  sizeof(*ted->ted_lcd),
395                                                  &off, rc /* sync if no cb */);
396                         fsfilt_commit(obd,
397                                       obt->obt_rcvd_filp->f_dentry->d_inode,
398                                       handle, 0);
399                 }
400                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
401
402                 if (rc) {
403                         CERROR("error writing %s client idx %u: rc %d\n",
404                                LAST_RCVD, ted->ted_lr_idx, rc);
405                         RETURN(rc);
406                 }
407         }
408         RETURN(0);
409 }
410
411 static int filter_client_del(struct obd_export *exp)
412 {
413         struct tg_export_data *ted = &exp->exp_target_data;
414         struct obd_device_target *obt = &exp->exp_obd->u.obt;
415         struct lvfs_run_ctxt saved;
416         int rc;
417         loff_t off;
418         ENTRY;
419
420         if (ted->ted_lcd == NULL)
421                 RETURN(0);
422
423         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
424         if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
425                 GOTO(free, 0);
426
427         LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
428
429         off = ted->ted_lr_off;
430
431         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
432                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
433
434         /* Don't clear ted_lr_idx here as it is likely also unset.  At worst
435          * we leak a client slot that will be cleaned on the next recovery. */
436         if (off <= 0) {
437                 CERROR("%s: client idx %d has med_off %lld\n",
438                        exp->exp_obd->obd_name, ted->ted_lr_idx, off);
439                 GOTO(free, rc = -EINVAL);
440         }
441
442         /* Clear the bit _after_ zeroing out the client so we don't
443            race with filter_client_add and zero out new clients.*/
444         if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
445                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
446                        ted->ted_lr_idx);
447                 LBUG();
448         }
449
450         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
451         /* Make sure the server's last_transno is up to date.
452          * This should be done before zeroing client slot so last_transno will
453          * be in server data or in client data in case of failure */
454         filter_update_server_data(exp->exp_obd);
455
456         cfs_mutex_lock(&ted->ted_lcd_lock);
457         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
458         rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
459                                  ted->ted_lcd,
460                                  sizeof(*ted->ted_lcd), &off, 0);
461         cfs_mutex_unlock(&ted->ted_lcd_lock);
462         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
463
464         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
465                "zero out client %s at idx %u/%llu in %s, rc %d\n",
466                ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
467                LAST_RCVD, rc);
468         EXIT;
469 free:
470         return 0;
471 }
472
473 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
474 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
475                                          struct filter_mod_data *fmd)
476 {
477         LASSERT_SPIN_LOCKED(&fed->fed_lock);
478         if (--fmd->fmd_refcount == 0) {
479                 /* XXX when we have persistent reservations and the handle
480                  * is stored herein we need to drop it here. */
481                 fed->fed_mod_count--;
482                 cfs_list_del(&fmd->fmd_list);
483                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
484         }
485 }
486
487 /* drop fmd reference, free it if last ref */
488 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
489 {
490         struct filter_export_data *fed;
491
492         if (fmd == NULL)
493                 return;
494
495         fed = &exp->exp_filter_data;
496         cfs_spin_lock(&fed->fed_lock);
497         filter_fmd_put_nolock(fed, fmd); /* caller reference */
498         cfs_spin_unlock(&fed->fed_lock);
499 }
500
501 /* expire entries from the end of the list if there are too many
502  * or they are too old */
503 static void filter_fmd_expire_nolock(struct filter_obd *filter,
504                                      struct filter_export_data *fed,
505                                      struct filter_mod_data *keep)
506 {
507         struct filter_mod_data *fmd, *tmp;
508
509         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
510                 if (fmd == keep)
511                         break;
512
513                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
514                     fed->fed_mod_count < filter->fo_fmd_max_num)
515                         break;
516
517                 cfs_list_del_init(&fmd->fmd_list);
518                 filter_fmd_put_nolock(fed, fmd); /* list reference */
519         }
520 }
521
522 void filter_fmd_expire(struct obd_export *exp)
523 {
524         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
525         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
526                                  &exp->exp_filter_data, NULL);
527         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
528 }
529
530 /* find specified objid, group in export fmd list.
531  * caller must hold fed_lock and take fmd reference itself */
532 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
533                                                 struct filter_export_data *fed,
534                                                 obd_id objid, obd_seq group)
535 {
536         struct filter_mod_data *found = NULL, *fmd;
537
538         LASSERT_SPIN_LOCKED(&fed->fed_lock);
539
540         cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
541                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
542                         found = fmd;
543                         cfs_list_del(&fmd->fmd_list);
544                         cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
545                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
546                         break;
547                 }
548         }
549
550         filter_fmd_expire_nolock(filter, fed, found);
551
552         return found;
553 }
554
555 /* Find fmd based on objid and group, or return NULL if not found. */
556 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
557                                         obd_id objid, obd_seq group)
558 {
559         struct filter_mod_data *fmd;
560
561         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
562         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
563                                      &exp->exp_filter_data, objid, group);
564         if (fmd)
565                 fmd->fmd_refcount++;    /* caller reference */
566         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
567
568         return fmd;
569 }
570
571 /* Find fmd based on objid and group, or create a new one if none is found.
572  * It is possible for this function to return NULL under memory pressure,
573  * or if objid = 0 is passed (which will only cause old entries to expire).
574  * Currently this is not fatal because any fmd state is transient and
575  * may also be freed when it gets sufficiently old. */
576 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
577                                        obd_id objid, obd_seq group)
578 {
579         struct filter_export_data *fed = &exp->exp_filter_data;
580         struct filter_mod_data *found = NULL, *fmd_new = NULL;
581
582         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
583
584         cfs_spin_lock(&fed->fed_lock);
585         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
586         if (fmd_new) {
587                 if (found == NULL) {
588                         cfs_list_add_tail(&fmd_new->fmd_list,
589                                           &fed->fed_mod_list);
590                         fmd_new->fmd_id = objid;
591                         fmd_new->fmd_gr = group;
592                         fmd_new->fmd_refcount++;   /* list reference */
593                         found = fmd_new;
594                         fed->fed_mod_count++;
595                 } else {
596                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
597                 }
598         }
599         if (found) {
600                 found->fmd_refcount++;          /* caller reference */
601                 found->fmd_expire = jiffies +
602                         exp->exp_obd->u.filter.fo_fmd_max_age;
603         }
604
605         cfs_spin_unlock(&fed->fed_lock);
606
607         return found;
608 }
609
610 #ifdef DO_FMD_DROP
611 /* drop fmd list reference so it will disappear when last reference is put.
612  * This isn't so critical because it would in fact only affect the one client
613  * that is doing the unlink and at worst we have an stale entry referencing
614  * an object that should never be used again. */
615 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
616 {
617         struct filter_mod_data *found = NULL;
618
619         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
620         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
621         if (found) {
622                 cfs_list_del_init(&found->fmd_list);
623                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
624         }
625         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
626 }
627 #else
628 #define filter_fmd_drop(exp, objid, group)
629 #endif
630
631 /* remove all entries from fmd list */
632 static void filter_fmd_cleanup(struct obd_export *exp)
633 {
634         struct filter_export_data *fed = &exp->exp_filter_data;
635         struct filter_mod_data *fmd = NULL, *tmp;
636
637         cfs_spin_lock(&fed->fed_lock);
638         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
639                 cfs_list_del_init(&fmd->fmd_list);
640                 filter_fmd_put_nolock(fed, fmd);
641         }
642         cfs_spin_unlock(&fed->fed_lock);
643 }
644
645 static int filter_init_export(struct obd_export *exp)
646 {
647         int rc;
648         ENTRY;
649
650         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
651         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
652
653         cfs_spin_lock(&exp->exp_lock);
654         exp->exp_connecting = 1;
655         cfs_spin_unlock(&exp->exp_lock);
656
657         /* self-export doesn't need client data and ldlm initialization */
658         if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
659                                      &exp->exp_client_uuid)))
660                 RETURN(0);
661
662         rc = lut_client_alloc(exp);
663         if (rc == 0)
664                 rc = ldlm_init_export(exp);
665         if (rc)
666                 CERROR("%s: Can't initialize export: rc %d\n",
667                        exp->exp_obd->obd_name, rc);
668
669         RETURN(rc);
670 }
671
672 static int filter_free_server_data(struct obd_device_target *obt)
673 {
674         lut_fini(NULL, obt->obt_lut);
675         OBD_FREE_PTR(obt->obt_lut);
676         return 0;
677 }
678
679 /* assumes caller is already in kernel ctxt */
680 int filter_update_server_data(struct obd_device *obd)
681 {
682         struct file *filp = obd->u.obt.obt_rcvd_filp;
683         struct lr_server_data *lsd = class_server_data(obd);
684         loff_t off = 0;
685         int rc;
686         ENTRY;
687
688         CDEBUG(D_INODE, "server uuid      : %s\n", lsd->lsd_uuid);
689         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
690                le64_to_cpu(lsd->lsd_last_transno));
691         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
692                le64_to_cpu(lsd->lsd_mount_count));
693
694         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
695         if (rc)
696                 CERROR("error writing lr_server_data: rc = %d\n", rc);
697
698         RETURN(rc);
699 }
700
701 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
702                              int force_sync)
703 {
704         struct filter_obd *filter = &obd->u.filter;
705         __u64 tmp;
706         loff_t off = 0;
707         int rc;
708         ENTRY;
709
710         if (filter->fo_last_objid_files[group] == NULL) {
711                 CERROR("Object seq "LPU64" not fully setup; not updating "
712                        "last_objid\n", group);
713                 RETURN(-EINVAL);
714         }
715
716         CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
717                obd->obd_name, filter->fo_last_objids[group], group);
718
719         tmp = cpu_to_le64(filter->fo_last_objids[group]);
720         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
721                                  &tmp, sizeof(tmp), &off, force_sync);
722         if (rc)
723                 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
724                        group, rc);
725         RETURN(rc);
726 }
727 extern int ost_handle(struct ptlrpc_request *req);
728 /* assumes caller has already in kernel ctxt */
729 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
730 {
731         struct filter_obd *filter = &obd->u.filter;
732         struct lr_server_data *lsd;
733         struct lsd_client_data *lcd = NULL;
734         struct inode *inode = filp->f_dentry->d_inode;
735         unsigned long last_rcvd_size = i_size_read(inode);
736         struct lu_target *lut;
737         __u64 mount_count;
738         __u32 start_epoch;
739         int cl_idx;
740         loff_t off = 0;
741         int rc;
742
743         /* ensure padding in the struct is the correct size */
744         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
745                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
746         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
747                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
748
749         /* allocate and initialize lu_target */
750         OBD_ALLOC_PTR(lut);
751         if (lut == NULL)
752                 RETURN(-ENOMEM);
753         rc = lut_init(NULL, lut, obd, NULL);
754         if (rc)
755                 GOTO(err_lut, rc);
756         lsd = class_server_data(obd);
757         if (last_rcvd_size == 0) {
758                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
759
760                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
761                 lsd->lsd_last_transno = 0;
762                 mount_count = lsd->lsd_mount_count = 0;
763                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
764                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
765                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
766                 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
767                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
768                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
769                  * MDS first connects and assigns the OST index number. */
770                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
771                                                         OBD_INCOMPAT_OST);
772         } else {
773                 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
774                 if (rc) {
775                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
776                                LAST_RCVD, rc);
777                         GOTO(err_lut, rc);
778                 }
779                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
780                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
781                                            "using the wrong disk %s. Were the "
782                                            "/dev/ assignments rearranged?\n",
783                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
784                         GOTO(err_lut, rc = -EINVAL);
785                 }
786                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
787                 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
788                 /* COMPAT_146 */
789                 /* Assume old last_rcvd format unless I_C_LR is set */
790                 if (!(lsd->lsd_feature_incompat &
791                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
792                         lsd->lsd_last_transno = lsd->lsd_compat14;
793                 /* end COMPAT_146 */
794                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
795                  * MDS first connects and assigns the OST index number. */
796                 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
797                                                          OBD_INCOMPAT_OST);
798         }
799
800         if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
801                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
802                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
803                        ~FILTER_INCOMPAT_SUPP);
804                 GOTO(err_lut, rc = -EINVAL);
805         }
806         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
807                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
808                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
809                        ~FILTER_ROCOMPAT_SUPP);
810                 /* Do something like remount filesystem read-only */
811                 GOTO(err_lut, rc = -EINVAL);
812         }
813
814         start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
815
816         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
817                obd->obd_name, start_epoch);
818         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
819                obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
820         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
821                obd->obd_name, mount_count + 1);
822         CDEBUG(D_INODE, "%s: server data size: %u\n",
823                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
824         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
825                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
826         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
827                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
828         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
829                obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
830         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
831                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
832                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
833                 le16_to_cpu(lsd->lsd_client_size));
834
835         if (!obd->obd_replayable) {
836                 CWARN("%s: recovery support OFF\n", obd->obd_name);
837                 GOTO(out, rc = 0);
838         }
839
840         OBD_ALLOC_PTR(lcd);
841         if (!lcd)
842                 GOTO(err_client, rc = -ENOMEM);
843
844         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
845              off < last_rcvd_size; cl_idx++) {
846                 __u64 last_rcvd;
847                 struct obd_export *exp;
848                 struct filter_export_data *fed;
849
850                 /* Don't assume off is incremented properly by
851                  * fsfilt_read_record(), in case sizeof(*lcd)
852                  * isn't the same as lsd->lsd_client_size.  */
853                 off = le32_to_cpu(lsd->lsd_client_start) +
854                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
855                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
856                 if (rc) {
857                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
858                                LAST_RCVD, cl_idx, off, rc);
859                         break; /* read error shouldn't cause startup to fail */
860                 }
861
862                 if (lcd->lcd_uuid[0] == '\0') {
863                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
864                                cl_idx);
865                         continue;
866                 }
867
868                 check_lcd(obd->obd_name, cl_idx, lcd);
869
870                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
871
872                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
873                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
874                        last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
875
876                 /* These exports are cleaned up by filter_disconnect(), so they
877                  * need to be set up like real exports as filter_connect() does.
878                  */
879                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
880                 if (IS_ERR(exp)) {
881                         if (PTR_ERR(exp) == -EALREADY) {
882                                 /* export already exists, zero out this one */
883                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
884                                 continue;
885                         }
886                         OBD_FREE_PTR(lcd);
887                         GOTO(err_client, rc = PTR_ERR(exp));
888                 }
889
890                 fed = &exp->exp_filter_data;
891                 *fed->fed_ted.ted_lcd = *lcd;
892                 fed->fed_group = 0; /* will be assigned at connect */
893                 filter_export_stats_init(obd, exp, NULL);
894                 rc = filter_client_add(obd, exp, cl_idx);
895                 /* can't fail for existing client */
896                 LASSERTF(rc == 0, "rc = %d\n", rc);
897
898                 /* VBR: set export last committed */
899                 exp->exp_last_committed = last_rcvd;
900                 cfs_spin_lock(&exp->exp_lock);
901                 exp->exp_connecting = 0;
902                 exp->exp_in_recovery = 0;
903                 cfs_spin_unlock(&exp->exp_lock);
904                 obd->obd_max_recoverable_clients++;
905                 class_export_put(exp);
906
907                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
908                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
909         }
910         OBD_FREE_PTR(lcd);
911
912         obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
913 out:
914         obd->u.obt.obt_mount_count = mount_count + 1;
915         obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
916         lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
917
918         /* save it, so mount count and last_transno is current */
919         rc = filter_update_server_data(obd);
920         if (rc)
921                 GOTO(err_client, rc);
922
923         RETURN(0);
924
925 err_client:
926         class_disconnect_exports(obd);
927 err_lut:
928         filter_free_server_data(&obd->u.obt);
929         RETURN(rc);
930 }
931
932 static int filter_cleanup_groups(struct obd_device *obd)
933 {
934         struct filter_obd *filter = &obd->u.filter;
935         struct file *filp;
936         struct dentry *dentry;
937         int i, j;
938         ENTRY;
939
940         if (filter->fo_dentry_O_groups != NULL) {
941                 for (i = 0; i < filter->fo_group_count; i++) {
942                         dentry = filter->fo_dentry_O_groups[i];
943                         if (dentry != NULL)
944                                 f_dput(dentry);
945                 }
946                 OBD_FREE(filter->fo_dentry_O_groups,
947                          filter->fo_group_count *
948                          sizeof(*filter->fo_dentry_O_groups));
949                 filter->fo_dentry_O_groups = NULL;
950         }
951         if (filter->fo_last_objid_files != NULL) {
952                 for (i = 0; i < filter->fo_group_count; i++) {
953                         filp = filter->fo_last_objid_files[i];
954                         if (filp != NULL)
955                                 filp_close(filp, 0);
956                 }
957                 OBD_FREE(filter->fo_last_objid_files,
958                          filter->fo_group_count *
959                          sizeof(*filter->fo_last_objid_files));
960                 filter->fo_last_objid_files = NULL;
961         }
962         if (filter->fo_dentry_O_sub != NULL) {
963                 for (i = 0; i < filter->fo_group_count; i++) {
964                         for (j = 0; j < filter->fo_subdir_count; j++) {
965                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
966                                 if (dentry != NULL)
967                                         f_dput(dentry);
968                         }
969                 }
970                 OBD_FREE(filter->fo_dentry_O_sub,
971                          filter->fo_group_count *
972                          sizeof(*filter->fo_dentry_O_sub));
973                 filter->fo_dentry_O_sub = NULL;
974         }
975         if (filter->fo_last_objids != NULL) {
976                 OBD_FREE(filter->fo_last_objids,
977                          filter->fo_group_count *
978                          sizeof(*filter->fo_last_objids));
979                 filter->fo_last_objids = NULL;
980         }
981         if (filter->fo_dentry_O != NULL) {
982                 f_dput(filter->fo_dentry_O);
983                 filter->fo_dentry_O = NULL;
984         }
985         RETURN(0);
986 }
987
988 static int filter_update_last_group(struct obd_device *obd, int group)
989 {
990         struct filter_obd *filter = &obd->u.filter;
991         struct file *filp = NULL;
992         int last_group = 0, rc;
993         loff_t off = 0;
994         ENTRY;
995
996         if (group <= filter->fo_committed_group)
997                 RETURN(0);
998
999         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1000         if (IS_ERR(filp)) {
1001                 rc = PTR_ERR(filp);
1002                 filp = NULL;
1003                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1004                 GOTO(cleanup, rc);
1005         }
1006
1007         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1008         if (rc) {
1009                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1010                 GOTO(cleanup, rc);
1011         }
1012
1013         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1014                obd->obd_name, last_group, group);
1015
1016         off = 0;
1017         last_group = group;
1018         /* must be sync: bXXXX */
1019         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1020         if (rc) {
1021                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1022                 GOTO(cleanup, rc);
1023         }
1024
1025         filter->fo_committed_group = group;
1026 cleanup:
1027         if (filp)
1028                 filp_close(filp, 0);
1029         RETURN(rc);
1030 }
1031
1032 static int filter_read_group_internal(struct obd_device *obd, int group,
1033                                       int create)
1034 {
1035         struct filter_obd *filter = &obd->u.filter;
1036         __u64 *new_objids = NULL;
1037         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1038         struct dentry **new_groups = NULL;
1039         struct file **new_files = NULL;
1040         struct dentry *dentry;
1041         struct file *filp;
1042         int old_count = filter->fo_group_count, rc, stage = 0, i;
1043         char name[25];
1044         __u64 last_objid;
1045         loff_t off = 0;
1046         int len = group + 1;
1047
1048         snprintf(name, 24, "%d", group);
1049         name[24] = '\0';
1050
1051         if (!create) {
1052                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1053                                            strlen(name));
1054                 if (IS_ERR(dentry)) {
1055                         CERROR("Cannot lookup expected object group %d: %ld\n",
1056                                group, PTR_ERR(dentry));
1057                         RETURN(PTR_ERR(dentry));
1058                 }
1059         } else {
1060                 dentry = simple_mkdir(filter->fo_dentry_O,
1061                                       obd->u.obt.obt_vfsmnt, name, 0700, 1);
1062                 if (IS_ERR(dentry)) {
1063                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1064                                PTR_ERR(dentry));
1065                         RETURN(PTR_ERR(dentry));
1066                 }
1067         }
1068         stage = 1;
1069
1070         snprintf(name, 24, "O/%d/LAST_ID", group);
1071         name[24] = '\0';
1072         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1073         if (IS_ERR(filp)) {
1074                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1075                 GOTO(cleanup, rc = PTR_ERR(filp));
1076         }
1077         stage = 2;
1078
1079         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1080         if (rc) {
1081                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1082                 GOTO(cleanup, rc);
1083         }
1084
1085         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1086                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1087                 if (tmp_subdirs == NULL)
1088                         GOTO(cleanup, rc = -ENOMEM);
1089                 stage = 3;
1090
1091                 for (i = 0; i < filter->fo_subdir_count; i++) {
1092                         char dir[20];
1093                         snprintf(dir, sizeof(dir), "d%u", i);
1094
1095                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1096                                                               obd->u.obt.obt_vfsmnt,
1097                                                               dir, 0700, 1);
1098                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1099                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1100                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1101                                        group, dir, rc);
1102                                 GOTO(cleanup, rc);
1103                         }
1104
1105                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1106                                tmp_subdirs->dentry[i]);
1107                 }
1108         }
1109
1110         /* 'group' is an index; we need an array of length 'group + 1' */
1111         if (group + 1 > old_count) {
1112                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1113                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1114                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1115                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1116                 stage = 4;
1117                 if (new_objids == NULL || new_subdirs == NULL ||
1118                     new_groups == NULL || new_files == NULL)
1119                         GOTO(cleanup, rc = -ENOMEM);
1120
1121                 if (old_count) {
1122                         memcpy(new_objids, filter->fo_last_objids,
1123                                old_count * sizeof(*new_objids));
1124                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1125                                old_count * sizeof(*new_subdirs));
1126                         memcpy(new_groups, filter->fo_dentry_O_groups,
1127                                old_count * sizeof(*new_groups));
1128                         memcpy(new_files, filter->fo_last_objid_files,
1129                                old_count * sizeof(*new_files));
1130
1131                         OBD_FREE(filter->fo_last_objids,
1132                                  old_count * sizeof(*new_objids));
1133                         OBD_FREE(filter->fo_dentry_O_sub,
1134                                  old_count * sizeof(*new_subdirs));
1135                         OBD_FREE(filter->fo_dentry_O_groups,
1136                                  old_count * sizeof(*new_groups));
1137                         OBD_FREE(filter->fo_last_objid_files,
1138                                  old_count * sizeof(*new_files));
1139                 }
1140                 filter->fo_last_objids = new_objids;
1141                 filter->fo_dentry_O_sub = new_subdirs;
1142                 filter->fo_dentry_O_groups = new_groups;
1143                 filter->fo_last_objid_files = new_files;
1144                 filter->fo_group_count = len;
1145         }
1146
1147         filter->fo_dentry_O_groups[group] = dentry;
1148         filter->fo_last_objid_files[group] = filp;
1149         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1150                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1151                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1152         }
1153
1154         filter_update_last_group(obd, group);
1155
1156         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1157                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1158                 rc = filter_update_last_objid(obd, group, 1);
1159                 RETURN(rc);
1160         }
1161
1162         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1163         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1164                obd->obd_name, group, last_objid);
1165         RETURN(0);
1166  cleanup:
1167         switch (stage) {
1168         case 4:
1169                 if (new_objids != NULL)
1170                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1171                 if (new_subdirs != NULL)
1172                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1173                 if (new_groups != NULL)
1174                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1175                 if (new_files != NULL)
1176                         OBD_FREE(new_files, len * sizeof(*new_files));
1177         case 3:
1178                 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1179                         for (i = 0; i < filter->fo_subdir_count; i++) {
1180                                 if (tmp_subdirs->dentry[i] != NULL)
1181                                         dput(tmp_subdirs->dentry[i]);
1182                         }
1183                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1184                 }
1185         case 2:
1186                 filp_close(filp, 0);
1187         case 1:
1188                 dput(dentry);
1189         }
1190         RETURN(rc);
1191 }
1192
1193 static int filter_read_groups(struct obd_device *obd, int last_group,
1194                               int create)
1195 {
1196         struct filter_obd *filter = &obd->u.filter;
1197         int old_count, group, rc = 0;
1198
1199         cfs_mutex_lock(&filter->fo_init_lock);
1200         old_count = filter->fo_group_count;
1201         for (group = old_count; group <= last_group; group++) {
1202                 rc = filter_read_group_internal(obd, group, create);
1203                 if (rc != 0)
1204                         break;
1205         }
1206         cfs_mutex_unlock(&filter->fo_init_lock);
1207         return rc;
1208 }
1209
1210 /* FIXME: object groups */
1211 static int filter_prep_groups(struct obd_device *obd)
1212 {
1213         struct filter_obd *filter = &obd->u.filter;
1214         struct dentry *O_dentry;
1215         struct file *filp;
1216         int    last_group, rc = 0, cleanup_phase = 0;
1217         loff_t off = 0;
1218         ENTRY;
1219
1220         O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1221                                 "O", 0700, 1);
1222         CDEBUG(D_INODE, "%s: got/created O: %p\n", obd->obd_name, O_dentry);
1223         if (IS_ERR(O_dentry)) {
1224                 rc = PTR_ERR(O_dentry);
1225                 CERROR("%s: cannot open/create O: rc = %d\n", obd->obd_name,rc);
1226                 GOTO(cleanup, rc);
1227         }
1228         filter->fo_dentry_O = O_dentry;
1229         cleanup_phase = 1; /* O_dentry */
1230
1231         /* we have to initialize all groups before first connections from
1232          * clients because they may send create/destroy for any group -bzzz */
1233         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1234         if (IS_ERR(filp)) {
1235                 CERROR("%s: cannot create LAST_GROUP: rc = %ld\n",
1236                        obd->obd_name, PTR_ERR(filp));
1237                 GOTO(cleanup, rc = PTR_ERR(filp));
1238         }
1239         cleanup_phase = 2; /* filp */
1240
1241         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1242         if (rc) {
1243                 CERROR("%s: error reading LAST_GROUP: rc %d\n",
1244                        obd->obd_name, rc);
1245                 GOTO(cleanup, rc);
1246         }
1247
1248         if (off == 0)
1249                 last_group = FID_SEQ_OST_MDT0;
1250
1251         CDEBUG(D_INODE, "%s: initialize group %u (max %u)\n", obd->obd_name,
1252                FID_SEQ_OST_MDT0, last_group);
1253         filter->fo_committed_group = last_group;
1254         rc = filter_read_groups(obd, last_group, 1);
1255         if (rc)
1256                 GOTO(cleanup, rc);
1257
1258         filp_close(filp, 0);
1259         RETURN(0);
1260
1261  cleanup:
1262         switch (cleanup_phase) {
1263         case 2:
1264                 filp_close(filp, 0);
1265         case 1:
1266                 filter_cleanup_groups(obd);
1267                 f_dput(filter->fo_dentry_O);
1268                 filter->fo_dentry_O = NULL;
1269         default:
1270                 break;
1271         }
1272         return rc;
1273
1274 }
1275
1276 /* setup the object store with correct subdirectories */
1277 static int filter_prep(struct obd_device *obd)
1278 {
1279         struct lvfs_run_ctxt saved;
1280         struct filter_obd *filter = &obd->u.filter;
1281         struct file *file;
1282         struct inode *inode;
1283         int rc = 0;
1284         ENTRY;
1285
1286         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1287         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1288         if (!file || IS_ERR(file)) {
1289                 rc = PTR_ERR(file);
1290                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1291                        LAST_RCVD, rc);
1292                 GOTO(out, rc);
1293         }
1294         obd->u.obt.obt_rcvd_filp = file;
1295         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1296                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1297                        file->f_dentry->d_inode->i_mode);
1298                 GOTO(err_filp, rc = -ENOENT);
1299         }
1300
1301         inode = file->f_dentry->d_parent->d_inode;
1302         /* We use i_op->unlink directly in filter_vfs_unlink() */
1303         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1304                 CERROR("%s: filesystem does not support create/unlink ops\n",
1305                        obd->obd_name);
1306                 GOTO(err_filp, rc = -EOPNOTSUPP);
1307         }
1308
1309         rc = filter_init_server_data(obd, file);
1310         if (rc) {
1311                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1312                 GOTO(err_filp, rc);
1313         }
1314         LASSERT(obd->u.obt.obt_lut);
1315         target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1316
1317         /* open and create health check io file*/
1318         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1319         if (IS_ERR(file)) {
1320                 rc = PTR_ERR(file);
1321                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1322                        HEALTH_CHECK, rc);
1323                 GOTO(err_server_data, rc);
1324         }
1325         filter->fo_obt.obt_health_check_filp = file;
1326         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1327                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1328                        file->f_dentry->d_inode->i_mode);
1329                 GOTO(err_health_check, rc = -ENOENT);
1330         }
1331         rc = lvfs_check_io_health(obd, file);
1332         if (rc)
1333                 GOTO(err_health_check, rc);
1334
1335         rc = filter_prep_groups(obd);
1336         if (rc)
1337                 GOTO(err_health_check, rc);
1338 out:
1339         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1340
1341         return(rc);
1342
1343 err_health_check:
1344         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1345                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1346         filter->fo_obt.obt_health_check_filp = NULL;
1347 err_server_data:
1348         target_recovery_fini(obd);
1349         filter_free_server_data(&obd->u.obt);
1350 err_filp:
1351         if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1352                 CERROR("can't close %s after error\n", LAST_RCVD);
1353         obd->u.obt.obt_rcvd_filp = NULL;
1354         goto out;
1355 }
1356
1357 /* cleanup the filter: write last used object id to status file */
1358 static void filter_post(struct obd_device *obd)
1359 {
1360         struct lvfs_run_ctxt saved;
1361         struct filter_obd *filter = &obd->u.filter;
1362         int rc, i;
1363
1364         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1365          * best to start a transaction with h_sync, because we removed this
1366          * from lastobjid */
1367
1368         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1369         rc = filter_update_server_data(obd);
1370         if (rc)
1371                 CERROR("error writing server data: rc = %d\n", rc);
1372
1373         for (i = 0; i < filter->fo_group_count; i++) {
1374                 rc = filter_update_last_objid(obd, i,
1375                                 (i == filter->fo_group_count - 1));
1376                 if (rc)
1377                         CERROR("error writing group %d lastobjid: rc = %d\n",
1378                                i, rc);
1379         }
1380
1381         rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1382         obd->u.obt.obt_rcvd_filp = NULL;
1383         if (rc)
1384                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1385
1386         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1387         filter->fo_obt.obt_health_check_filp = NULL;
1388         if (rc)
1389                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1390
1391         filter_cleanup_groups(obd);
1392         filter_free_server_data(&obd->u.obt);
1393         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1394
1395         filter_free_capa_keys(filter);
1396         cleanup_capa_hash(filter->fo_capa_hash);
1397 }
1398
1399 static void filter_set_last_id(struct filter_obd *filter,
1400                                obd_id id, obd_seq group)
1401 {
1402         LASSERT(group <= filter->fo_group_count);
1403
1404         cfs_spin_lock(&filter->fo_objidlock);
1405         filter->fo_last_objids[group] = id;
1406         cfs_spin_unlock(&filter->fo_objidlock);
1407 }
1408
1409 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1410 {
1411         obd_id id;
1412         LASSERT(group <= filter->fo_group_count);
1413         LASSERT(filter->fo_last_objids != NULL);
1414
1415         /* FIXME: object groups */
1416         cfs_spin_lock(&filter->fo_objidlock);
1417         id = filter->fo_last_objids[group];
1418         cfs_spin_unlock(&filter->fo_objidlock);
1419         return id;
1420 }
1421
1422 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1423 {
1424         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1425         return 0;
1426 }
1427
1428 /* We never dget the object parent, so DON'T dput it either */
1429 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1430 {
1431         struct filter_obd *filter = &obd->u.filter;
1432         struct filter_subdirs *subdirs;
1433
1434         if (group >= filter->fo_group_count) /* FIXME: object groups */
1435                 return ERR_PTR(-EBADF);
1436
1437         if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1438                 return filter->fo_dentry_O_groups[group];
1439
1440         subdirs = &filter->fo_dentry_O_sub[group];
1441         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1442 }
1443
1444 /* We never dget the object parent, so DON'T dput it either */
1445 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1446                                   obd_id objid)
1447 {
1448         unsigned long now = jiffies;
1449         struct dentry *dparent = filter_parent(obd, group, objid);
1450         int rc;
1451
1452         if (IS_ERR(dparent))
1453                 return dparent;
1454         if (dparent == NULL)
1455                 return ERR_PTR(-ENOENT);
1456
1457         rc = filter_lock_dentry(obd, dparent);
1458         fsfilt_check_slow(obd, now, "parent lock");
1459         return rc ? ERR_PTR(rc) : dparent;
1460 }
1461
1462 /* We never dget the object parent, so DON'T dput it either */
1463 static void filter_parent_unlock(struct dentry *dparent)
1464 {
1465         UNLOCK_INODE_MUTEX(dparent->d_inode);
1466 }
1467
1468 /* How to get files, dentries, inodes from object id's.
1469  *
1470  * If dir_dentry is passed, the caller has already locked the parent
1471  * appropriately for this operation (normally a write lock).  If
1472  * dir_dentry is NULL, we do a read lock while we do the lookup to
1473  * avoid races with create/destroy and such changing the directory
1474  * internal to the filesystem code. */
1475 struct dentry *filter_fid2dentry(struct obd_device *obd,
1476                                  struct dentry *dir_dentry,
1477                                  obd_seq group, obd_id id)
1478 {
1479         struct dentry *dparent = dir_dentry;
1480         struct dentry *dchild;
1481         char name[32];
1482         int len;
1483         ENTRY;
1484
1485         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1486             obd->u.filter.fo_destroys_in_progress == 0) {
1487                 /* don't fail lookups for orphan recovery, it causes
1488                  * later LBUGs when objects still exist during precreate */
1489                 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1490                 RETURN(ERR_PTR(-ENOENT));
1491         }
1492         if (id == 0) {
1493                 CERROR("fatal: invalid object id 0\n");
1494                 RETURN(ERR_PTR(-ESTALE));
1495         }
1496
1497         len = sprintf(name, LPU64, id);
1498         if (dir_dentry == NULL) {
1499                 dparent = filter_parent_lock(obd, group, id);
1500                 if (IS_ERR(dparent)) {
1501                         CERROR("%s: error getting object "POSTID
1502                                " parent: rc %ld\n", obd->obd_name,
1503                                id, group, PTR_ERR(dparent));
1504                         RETURN(dparent);
1505                 }
1506         }
1507         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1508                dparent->d_name.len, dparent->d_name.name, name);
1509         /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1510         dchild = lookup_one_len(name, dparent, len);
1511         if (dir_dentry == NULL)
1512                 filter_parent_unlock(dparent);
1513         if (IS_ERR(dchild)) {
1514                 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1515                        obd->obd_name, id, group, PTR_ERR(dchild));
1516                 RETURN(dchild);
1517         }
1518
1519         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1520                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1521                        obd->obd_name, id, dchild->d_inode->i_ino);
1522                 f_dput(dchild);
1523                 RETURN(ERR_PTR(-ENOENT));
1524         }
1525
1526         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1527                name, dchild, atomic_read(&dchild->d_count));
1528
1529         LASSERT(atomic_read(&dchild->d_count) > 0);
1530
1531         RETURN(dchild);
1532 }
1533
1534 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1535                                   obd_id group, struct lustre_handle *lockh)
1536 {
1537         int flags = LDLM_AST_DISCARD_DATA, rc;
1538         struct ldlm_res_id res_id;
1539         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1540         ENTRY;
1541
1542         osc_build_res_name(objid, group, &res_id);
1543         /* Tell the clients that the object is gone now and that they should
1544          * throw away any cached pages. */
1545         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1546                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1547                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1548                                     lockh);
1549         if (rc != ELDLM_OK)
1550                 lockh->cookie = 0;
1551         RETURN(rc);
1552 }
1553
1554 static void filter_fini_destroy(struct obd_device *obd,
1555                                 struct lustre_handle *lockh)
1556 {
1557         if (lustre_handle_is_used(lockh))
1558                 ldlm_lock_decref(lockh, LCK_PW);
1559 }
1560
1561 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1562  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1563  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1564  * reverse.  Caller must take i_sem before starting the transaction and we
1565  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1566 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1567                       struct vfsmount *mnt)
1568 {
1569         int rc;
1570         ENTRY;
1571
1572         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1573          * itself we already hold dir->i_mutex for child create/unlink ops */
1574         LASSERT(dentry->d_inode != NULL);
1575         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1576         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1577
1578
1579         /* may_delete() */
1580         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1581                 GOTO(out, rc = -ENOENT);
1582
1583         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1584         if (rc)
1585                 GOTO(out, rc);
1586
1587         if (IS_APPEND(dir))
1588                 GOTO(out, rc = -EPERM);
1589
1590         /* check_sticky() */
1591         if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1592              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1593             IS_IMMUTABLE(dentry->d_inode))
1594                 GOTO(out, rc = -EPERM);
1595
1596         /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
1597         ll_vfs_dq_init(dir);
1598
1599         rc = ll_security_inode_unlink(dir, dentry, mnt);
1600         if (rc)
1601                 GOTO(out, rc);
1602
1603         rc = dir->i_op->unlink(dir, dentry);
1604 out:
1605         /* need to drop i_mutex before we lose inode reference */
1606         UNLOCK_INODE_MUTEX(dentry->d_inode);
1607         if (rc == 0)
1608                 d_delete(dentry);
1609
1610         RETURN(rc);
1611 }
1612
1613 /* Caller must hold LCK_PW on parent and push us into kernel context.
1614  * Caller must hold child i_mutex, we drop it always.
1615  * Caller is also required to ensure that dchild->d_inode exists. */
1616 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1617                                    obd_seq group, struct dentry *dparent,
1618                                    struct dentry *dchild)
1619 {
1620         struct inode *inode = dchild->d_inode;
1621         int rc;
1622
1623         /* There should be 2 references to the inode:
1624          *  1) taken by filter_prepare_destroy
1625          *  2) taken by filter_destroy */
1626         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1627                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1628                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1629                        (unsigned long)inode->i_nlink,
1630                        atomic_read(&inode->i_count));
1631         }
1632
1633         rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1634         if (rc)
1635                 CERROR("error unlinking objid %.*s: rc %d\n",
1636                        dchild->d_name.len, dchild->d_name.name, rc);
1637         return(rc);
1638 }
1639
1640 struct filter_intent_args {
1641         struct ldlm_lock **victim;
1642         __u64 size;
1643         int *liblustre;
1644 };
1645
1646 static enum interval_iter filter_intent_cb(struct interval_node *n,
1647                                            void *args)
1648 {
1649         struct ldlm_interval *node = (struct ldlm_interval *)n;
1650         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1651         __u64 size = arg->size;
1652         struct ldlm_lock **v = arg->victim;
1653         struct ldlm_lock *lck;
1654
1655         /* If the interval is lower than the current file size,
1656          * just break. */
1657         if (interval_high(n) <= size)
1658                 return INTERVAL_ITER_STOP;
1659
1660         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1661                 /* Don't send glimpse ASTs to liblustre clients.
1662                  * They aren't listening for them, and they do
1663                  * entirely synchronous I/O anyways. */
1664                 if (lck->l_export == NULL ||
1665                     lck->l_export->exp_libclient == 1)
1666                         continue;
1667
1668                 if (*arg->liblustre)
1669                         *arg->liblustre = 0;
1670
1671                 if (*v == NULL) {
1672                         *v = LDLM_LOCK_GET(lck);
1673                 } else if ((*v)->l_policy_data.l_extent.start <
1674                            lck->l_policy_data.l_extent.start) {
1675                         LDLM_LOCK_RELEASE(*v);
1676                         *v = LDLM_LOCK_GET(lck);
1677                 }
1678
1679                 /* the same policy group - every lock has the
1680                  * same extent, so needn't do it any more */
1681                 break;
1682         }
1683
1684         return INTERVAL_ITER_CONT;
1685 }
1686
1687 static int filter_intent_policy(struct ldlm_namespace *ns,
1688                                 struct ldlm_lock **lockp, void *req_cookie,
1689                                 ldlm_mode_t mode, int flags, void *data)
1690 {
1691         struct ptlrpc_request *req = req_cookie;
1692         struct ldlm_lock *lock = *lockp, *l = NULL;
1693         struct ldlm_resource *res = lock->l_resource;
1694         ldlm_processing_policy policy;
1695         struct ost_lvb *res_lvb, *reply_lvb;
1696         struct ldlm_reply *rep;
1697         ldlm_error_t err;
1698         int idx, rc, tmpflags = 0, only_liblustre = 1;
1699         struct ldlm_interval_tree *tree;
1700         struct filter_intent_args arg;
1701         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1702                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1703                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1704         ENTRY;
1705
1706         policy = ldlm_get_processing_policy(res);
1707         LASSERT(policy != NULL);
1708         LASSERT(req != NULL);
1709
1710         rc = lustre_pack_reply(req, 3, repsize, NULL);
1711         if (rc)
1712                 RETURN(req->rq_status = rc);
1713
1714         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1715         LASSERT(rep != NULL);
1716
1717         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1718                                    sizeof(*reply_lvb));
1719         LASSERT(reply_lvb != NULL);
1720
1721         //fixup_handle_for_resent_req(req, lock, &lockh);
1722
1723         /* Call the extent policy function to see if our request can be
1724          * granted, or is blocked.
1725          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1726          * lock, and should not be granted if the lock will be blocked.
1727          */
1728
1729         if (flags & LDLM_FL_BLOCK_NOWAIT) {
1730                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_AGL_DELAY, 5);
1731
1732                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_AGL_NOLOCK))
1733                         RETURN(ELDLM_LOCK_ABORTED);
1734         }
1735
1736         LASSERT(ns == ldlm_res_to_ns(res));
1737         lock_res(res);
1738         rc = policy(lock, &tmpflags, 0, &err, NULL);
1739         check_res_locked(res);
1740
1741         /* The lock met with no resistance; we're finished. */
1742         if (rc == LDLM_ITER_CONTINUE) {
1743                 /* do not grant locks to the liblustre clients: they cannot
1744                  * handle ASTs robustly.  We need to do this while still
1745                  * holding lr_lock to avoid the lock remaining on the res_link
1746                  * list (and potentially being added to l_pending_list by an
1747                  * AST) when we are going to drop this lock ASAP. */
1748                 if (lock->l_export->exp_libclient ||
1749                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1750                         ldlm_resource_unlink_lock(lock);
1751                         err = ELDLM_LOCK_ABORTED;
1752                 } else {
1753                         err = ELDLM_LOCK_REPLACED;
1754                 }
1755                 unlock_res(res);
1756                 RETURN(err);
1757         } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1758                 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1759                  * callback for glimpse size. The real size user will trigger
1760                  * the glimpse callback when necessary. */
1761                 unlock_res(res);
1762                 RETURN(ELDLM_LOCK_ABORTED);
1763         }
1764
1765         /* Do not grant any lock, but instead send GL callbacks.  The extent
1766          * policy nicely created a list of all PW locks for us.  We will choose
1767          * the highest of those which are larger than the size in the LVB, if
1768          * any, and perform a glimpse callback. */
1769         res_lvb = res->lr_lvb_data;
1770         LASSERT(res_lvb != NULL);
1771         *reply_lvb = *res_lvb;
1772
1773         /*
1774          * lr_lock guarantees that no new locks are granted, and,
1775          * therefore, that res->lr_lvb_data cannot increase beyond the
1776          * end of already granted lock. As a result, it is safe to
1777          * check against "stale" reply_lvb->lvb_size value without
1778          * res->lr_lvb_mutex.
1779          */
1780         arg.size = reply_lvb->lvb_size;
1781         arg.victim = &l;
1782         arg.liblustre = &only_liblustre;
1783         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1784                 tree = &res->lr_itree[idx];
1785                 if (tree->lit_mode == LCK_PR)
1786                         continue;
1787
1788                 interval_iterate_reverse(tree->lit_root,
1789                                          filter_intent_cb, &arg);
1790         }
1791         unlock_res(res);
1792
1793         /* There were no PW locks beyond the size in the LVB; finished. */
1794         if (l == NULL) {
1795                 if (only_liblustre) {
1796                         /* If we discovered a liblustre client with a PW lock,
1797                          * however, the LVB may be out of date!  The LVB is
1798                          * updated only on glimpse (which we don't do for
1799                          * liblustre clients) and cancel (which the client
1800                          * obviously has not yet done).  So if it has written
1801                          * data but kept the lock, the LVB is stale and needs
1802                          * to be updated from disk.
1803                          *
1804                          * Of course, this will all disappear when we switch to
1805                          * taking liblustre locks on the OST. */
1806                         ldlm_res_lvbo_update(res, NULL, 1);
1807                 }
1808                 RETURN(ELDLM_LOCK_ABORTED);
1809         }
1810
1811         /*
1812          * This check is for lock taken in filter_prepare_destroy() that does
1813          * not have l_glimpse_ast set. So the logic is: if there is a lock
1814          * with no l_glimpse_ast set, this object is being destroyed already.
1815          *
1816          * Hence, if you are grabbing DLM locks on the server, always set
1817          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1818          */
1819         if (l->l_glimpse_ast == NULL) {
1820                 /* We are racing with unlink(); just return -ENOENT */
1821                 rep->lock_policy_res1 = -ENOENT;
1822                 goto out;
1823         }
1824
1825         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1826         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1827
1828         lock_res(res);
1829         *reply_lvb = *res_lvb;
1830         unlock_res(res);
1831
1832  out:
1833         LDLM_LOCK_RELEASE(l);
1834
1835         RETURN(ELDLM_LOCK_ABORTED);
1836 }
1837
1838 /*
1839  * per-obd_device iobuf pool.
1840  *
1841  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1842  * allocations in write-path has to be minimized (see bug 5137).
1843  *
1844  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1845  * OST threads (see ost_thread_{init,done}()).
1846  *
1847  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1848  * at the OST layer there are only (potentially) multiple obd_device of type
1849  * unknown at the time of OST thread creation.
1850  *
1851  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1852  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1853  * it's very own iobuf.
1854  *
1855  * Functions below
1856  *
1857  *     filter_kiobuf_pool_init()
1858  *
1859  *     filter_kiobuf_pool_done()
1860  *
1861  *     filter_iobuf_get()
1862  *
1863  * operate on this array. They are "generic" in a sense that they don't depend
1864  * on actual type of iobuf's (the latter depending on Linux kernel version).
1865  */
1866
1867 /*
1868  * destroy pool created by filter_iobuf_pool_init
1869  */
1870 static void filter_iobuf_pool_done(struct filter_obd *filter)
1871 {
1872         struct filter_iobuf **pool;
1873         int i;
1874
1875         ENTRY;
1876
1877         pool = filter->fo_iobuf_pool;
1878         if (pool != NULL) {
1879                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1880                         if (pool[i] != NULL)
1881                                 filter_free_iobuf(pool[i]);
1882                 }
1883                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1884                 filter->fo_iobuf_pool = NULL;
1885         }
1886         EXIT;
1887 }
1888
1889 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1890 {
1891         struct filter_obd       *filter = &obd->u.filter;
1892         struct sptlrpc_rule_set  tmp_rset;
1893         int                      rc;
1894
1895         sptlrpc_rule_set_init(&tmp_rset);
1896         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1897         if (rc) {
1898                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1899                        obd->obd_name, rc);
1900                 return rc;
1901         }
1902
1903         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1904
1905         cfs_write_lock(&filter->fo_sptlrpc_lock);
1906         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1907         filter->fo_sptlrpc_rset = tmp_rset;
1908         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1909
1910         return 0;
1911 }
1912
1913 /*
1914  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1915  */
1916 static int filter_iobuf_pool_init(struct filter_obd *filter)
1917 {
1918         void **pool;
1919
1920         ENTRY;
1921
1922
1923         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1924                       GFP_KERNEL);
1925         if (filter->fo_iobuf_pool == NULL)
1926                 RETURN(-ENOMEM);
1927
1928         filter->fo_iobuf_count = OSS_THREADS_MAX;
1929
1930         RETURN(0);
1931 }
1932
1933 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1934  * many threads there will be so we allocate a large empty array and only
1935  * fill in those slots that are actually in use.
1936  * If we haven't allocated a pool entry for this thread before, do so now. */
1937 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1938 {
1939         int thread_id                    = (oti && oti->oti_thread) ?
1940                                            oti->oti_thread->t_id : -1;
1941         struct filter_iobuf  *pool       = NULL;
1942         struct filter_iobuf **pool_place = NULL;
1943
1944         if (thread_id >= 0) {
1945                 LASSERT(thread_id < filter->fo_iobuf_count);
1946                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1947         }
1948
1949         if (unlikely(pool == NULL)) {
1950                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1951                                           PTLRPC_MAX_BRW_PAGES);
1952                 if (pool_place != NULL)
1953                         *pool_place = pool;
1954         }
1955
1956         return pool;
1957 }
1958
1959 /* mount the file system (secretly).  lustre_cfg parameters are:
1960  * 1 = device
1961  * 2 = fstype
1962  * 3 = flags: failover=f, failout=n
1963  * 4 = mount options
1964  */
1965 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1966                         void *option)
1967 {
1968         struct filter_obd *filter = &obd->u.filter;
1969         struct vfsmount *mnt;
1970         struct file_system_type *type;
1971         struct lustre_mount_info *lmi;
1972         struct obd_uuid uuid;
1973         __u8 *uuid_ptr;
1974         char *str, *label;
1975         char ns_name[48];
1976         struct request_queue *q;
1977         int rc, i;
1978         ENTRY;
1979
1980         if (lcfg->lcfg_bufcount < 3 ||
1981             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1982             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1983                 RETURN(-EINVAL);
1984
1985         lmi = server_get_mount(obd->obd_name);
1986         if (lmi) {
1987                 /* We already mounted in lustre_fill_super.
1988                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1989                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1990                 mnt = lmi->lmi_mnt;
1991                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1992         } else {
1993                 /* old path - used by lctl */
1994                 CERROR("Using old MDS mount method\n");
1995                 type = get_fs_type(lustre_cfg_string(lcfg, 2));
1996                 if (!type) {
1997                         CERROR("get_fs_type failed\n");
1998                         RETURN(-ENODEV);
1999                 }
2000                 mnt = vfs_kern_mount(type, MS_NOATIME|MS_NODIRATIME,
2001                                      lustre_cfg_string(lcfg, 1), option);
2002                 cfs_module_put(type->owner);
2003                 if (IS_ERR(mnt)) {
2004                         rc = PTR_ERR(mnt);
2005                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2006                                            lustre_cfg_string(lcfg, 1), rc);
2007                         RETURN(rc);
2008                 }
2009
2010                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2011         }
2012         if (IS_ERR(obd->obd_fsops))
2013                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2014
2015         rc = filter_iobuf_pool_init(filter);
2016         if (rc != 0)
2017                 GOTO(err_ops, rc);
2018
2019         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2020                 CERROR("%s: Underlying device is marked as read-only. "
2021                        "Setup failed\n", obd->obd_name);
2022                 GOTO(err_ops, rc = -EROFS);
2023         }
2024
2025         /* failover is the default */
2026         obd->obd_replayable = 1;
2027
2028         /* disable connection until configuration finishes */
2029         obd->obd_no_conn = 1;
2030
2031         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2032                 str = lustre_cfg_string(lcfg, 3);
2033                 if (strchr(str, 'n')) {
2034                         CWARN("%s: recovery disabled\n", obd->obd_name);
2035                         obd->obd_replayable = 0;
2036                 }
2037         }
2038
2039         obd->u.obt.obt_magic = OBT_MAGIC;
2040         obd->u.obt.obt_vfsmnt = mnt;
2041         obd->u.obt.obt_sb = mnt->mnt_sb;
2042         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2043         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2044
2045         rc = fsfilt_setup(obd, obd->u.obt.obt_sb);
2046         if (rc)
2047                 GOTO(err_ops, rc);
2048
2049         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2050         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2051         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2052         obd->obd_lvfs_ctxt.fs = get_ds();
2053         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2054
2055         cfs_mutex_init(&filter->fo_init_lock);
2056         filter->fo_committed_group = 0;
2057         filter->fo_destroys_in_progress = 0;
2058         for (i = 0; i < 32; i++)
2059                 cfs_mutex_init(&filter->fo_create_locks[i]);
2060
2061         cfs_spin_lock_init(&filter->fo_objidlock);
2062         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2063         cfs_mutex_init(&filter->fo_alloc_lock);
2064         init_brw_stats(&filter->fo_filter_stats);
2065         cfs_spin_lock_init(&filter->fo_flags_lock);
2066         filter->fo_read_cache = 1; /* enable read-only cache by default */
2067         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2068         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2069         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2070         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2071         filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2072         filter_slc_set(filter); /* initialize sync on lock cancel */
2073
2074         rc = filter_prep(obd);
2075         if (rc)
2076                 GOTO(err_ops, rc);
2077
2078         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2079         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2080
2081         filter->fo_fl_oss_capa = 1;
2082
2083         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2084         filter->fo_capa_hash = init_capa_hash();
2085         if (filter->fo_capa_hash == NULL)
2086                 GOTO(err_post, rc = -ENOMEM);
2087
2088         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2089         obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2090                                                 LDLM_NAMESPACE_SERVER,
2091                                                 LDLM_NAMESPACE_GREEDY,
2092                                                 LDLM_NS_TYPE_OST);
2093         if (obd->obd_namespace == NULL)
2094                 GOTO(err_post, rc = -ENOMEM);
2095         obd->obd_namespace->ns_lvbp = obd;
2096         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2097         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2098
2099         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2100                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2101
2102         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2103         if (rc) {
2104                 CERROR("failed to setup llogging subsystems\n");
2105                 GOTO(err_post, rc);
2106         }
2107
2108         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2109         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2110         /* do this after llog being initialized */
2111         filter_adapt_sptlrpc_conf(obd, 1);
2112
2113         rc = lquota_setup(filter_quota_interface_ref, obd);
2114         if (rc)
2115                 GOTO(err_post, rc);
2116
2117         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2118         if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2119             queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2120                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2121                               "for larger I/O requests: max_sectors = %u "
2122                               "could be up to max_hw_sectors=%u\n",
2123                               obd->obd_name, mnt->mnt_sb->s_id,
2124                               queue_max_sectors(q), queue_max_hw_sectors(q));
2125
2126         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2127         if (uuid_ptr != NULL) {
2128                 class_uuid_unparse(uuid_ptr, &uuid);
2129                 str = uuid.uuid;
2130         } else {
2131                 str = "no UUID";
2132         }
2133
2134         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2135         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2136                       obd->obd_name, label ?: str, lmi ? "on " : "",
2137                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2138                       obd->obd_replayable ? "enabled" : "disabled");
2139
2140         RETURN(0);
2141
2142 err_post:
2143         filter_post(obd);
2144 err_ops:
2145         fsfilt_put_ops(obd->obd_fsops);
2146         filter_iobuf_pool_done(filter);
2147 err_mntput:
2148         server_put_mount(obd->obd_name, mnt);
2149         obd->u.obt.obt_sb = 0;
2150         return rc;
2151 }
2152
2153 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2154 {
2155         struct lprocfs_static_vars lvars;
2156         cfs_proc_dir_entry_t *entry;
2157         unsigned long addr;
2158         struct page *page;
2159         int rc;
2160         ENTRY;
2161
2162         CLASSERT(offsetof(struct obd_device, u.obt) ==
2163                  offsetof(struct obd_device, u.filter.fo_obt));
2164
2165         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2166                 RETURN(-EINVAL);
2167
2168         /* lprocfs must be setup before the filter so state can be safely added
2169          * to /proc incrementally as the filter is setup */
2170         lprocfs_filter_init_vars(&lvars);
2171         rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2172         if (rc) {
2173                 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2174                        obd->obd_name, rc);
2175                 RETURN(rc);
2176         }
2177
2178         rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2179         if (rc) {
2180                 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2181                        obd->obd_name, rc);
2182                 GOTO(obd_cleanup, rc);
2183         }
2184
2185         /* Init obdfilter private stats here */
2186         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2187                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2188         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2189                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2190         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2191                              LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2192                              "get_page", "usec");
2193         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2194                              LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2195         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2196                              LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2197         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2198                              LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2199         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2200                              LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2201
2202         rc = lproc_filter_attach_seqstat(obd);
2203         if (rc) {
2204                 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2205                 GOTO(free_obd_stats, rc);
2206         }
2207
2208         entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2209         if (IS_ERR(entry)) {
2210                 rc = PTR_ERR(entry);
2211                 CERROR("%s: error %d setting up lprocfs for %s\n",
2212                        obd->obd_name, rc, "exports");
2213                 GOTO(free_obd_stats, rc);
2214         }
2215         obd->obd_proc_exports_entry = entry;
2216
2217         entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2218                                    lprocfs_nid_stats_clear_read,
2219                                    lprocfs_nid_stats_clear_write, obd, NULL);
2220         if (IS_ERR(entry)) {
2221                 rc = PTR_ERR(entry);
2222                 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2223                        obd->obd_name, rc);
2224                 GOTO(free_obd_stats, rc);
2225         }
2226
2227         rc = lprocfs_job_stats_init(obd, LPROC_FILTER_STATS_LAST,
2228                                     filter_stats_counter_init);
2229         if (rc)
2230                 GOTO(remove_entry_clear, rc);
2231
2232         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2233         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2234         if (!page)
2235                 GOTO(job_stats_fini, rc = -ENOMEM);
2236         addr = (unsigned long)cfs_page_address(page);
2237         clear_page((void *)addr);
2238         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2239                LUSTRE_CFG_BUFLEN(lcfg, 4));
2240         rc = filter_common_setup(obd, lcfg, (void *)addr);
2241         OBD_PAGE_FREE(page);
2242         if (rc) {
2243                 CERROR("%s: filter_common_setup failed: %d.\n",
2244                        obd->obd_name, rc);
2245                 GOTO(job_stats_fini, rc);
2246         }
2247
2248         RETURN(0);
2249
2250 job_stats_fini:
2251         lprocfs_job_stats_fini(obd);
2252 remove_entry_clear:
2253         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2254 free_obd_stats:
2255         lprocfs_free_obd_stats(obd);
2256 obd_cleanup:
2257         lprocfs_obd_cleanup(obd);
2258         return rc;
2259 }
2260
2261 static struct llog_operations filter_mds_ost_repl_logops;
2262
2263 static struct llog_operations filter_size_orig_logops = {
2264         .lop_setup   = llog_obd_origin_setup,
2265         .lop_cleanup = llog_obd_origin_cleanup,
2266         .lop_add     = llog_obd_origin_add
2267 };
2268
2269 static int filter_olg_fini(struct obd_llog_group *olg)
2270 {
2271         struct llog_ctxt *ctxt;
2272         int rc = 0, rc2 = 0;
2273         ENTRY;
2274
2275         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2276         if (ctxt)
2277                 rc = llog_cleanup(ctxt);
2278
2279         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2280         if (ctxt) {
2281                 rc2 = llog_cleanup(ctxt);
2282                 if (!rc)
2283                         rc = rc2;
2284         }
2285
2286         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2287         if (ctxt) {
2288                 rc2 = llog_cleanup(ctxt);
2289                 if (!rc)
2290                         rc = rc2;
2291         }
2292
2293         RETURN(rc);
2294 }
2295
2296 static int
2297 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2298                 struct obd_device *tgt)
2299 {
2300         int rc;
2301         ENTRY;
2302
2303         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2304                         &filter_mds_ost_repl_logops);
2305         if (rc)
2306                 GOTO(cleanup, rc);
2307
2308         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2309                         &filter_size_orig_logops);
2310         if (rc)
2311                 GOTO(cleanup, rc);
2312         EXIT;
2313 cleanup:
2314         if (rc)
2315                 filter_olg_fini(olg);
2316         return rc;
2317 }
2318
2319 /**
2320  * Init the default olg, which is embeded in the obd_device, for filter.
2321  */
2322 static int
2323 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2324                         struct obd_device *tgt)
2325 {
2326         struct filter_obd *filter = &obd->u.filter;
2327         struct llog_ctxt *ctxt;
2328         int rc;
2329         ENTRY;
2330
2331         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2332         if (!filter->fo_lcm)
2333                 RETURN(-ENOMEM);
2334
2335         filter_mds_ost_repl_logops = llog_client_ops;
2336         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2337         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2338         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2339
2340         rc = filter_olg_init(obd, olg, tgt);
2341         if (rc)
2342                 GOTO(cleanup_lcm, rc);
2343
2344         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2345                         &llog_lvfs_ops);
2346         if (rc)
2347                 GOTO(cleanup_olg, rc);
2348
2349         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2350         if (!ctxt) {
2351                 CERROR("Can't get ctxt for %p:%x\n", olg,
2352                        LLOG_MDS_OST_REPL_CTXT);
2353                 GOTO(cleanup_olg, rc = -ENODEV);
2354         }
2355         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2356         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2357         llog_ctxt_put(ctxt);
2358
2359         RETURN(0);
2360 cleanup_olg:
2361         filter_olg_fini(olg);
2362 cleanup_lcm:
2363         llog_recov_thread_fini(filter->fo_lcm, 1);
2364         filter->fo_lcm = NULL;
2365         return rc;
2366 }
2367
2368 static int
2369 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2370                  struct obd_device *tgt, int *index)
2371 {
2372         struct filter_obd *filter = &obd->u.filter;
2373         struct llog_ctxt *ctxt;
2374         int rc;
2375         ENTRY;
2376
2377         LASSERT(olg != NULL);
2378         if (olg == &obd->obd_olg)
2379                 return filter_default_olg_init(obd, olg, tgt);
2380
2381         LASSERT(filter->fo_lcm != NULL);
2382         rc = filter_olg_init(obd, olg, tgt);
2383         if (rc)
2384                 RETURN(rc);
2385         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2386         if (!ctxt) {
2387                 CERROR("Can't get ctxt for %p:%x\n", olg,
2388                        LLOG_MDS_OST_REPL_CTXT);
2389                 filter_olg_fini(olg);
2390                 RETURN(-ENODEV);
2391         }
2392         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2393         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2394         llog_ctxt_put(ctxt);
2395         RETURN(rc);
2396 }
2397
2398 static int filter_llog_finish(struct obd_device *obd, int count)
2399 {
2400         struct filter_obd *filter = &obd->u.filter;
2401         struct llog_ctxt *ctxt;
2402         ENTRY;
2403
2404         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2405         if (ctxt) {
2406                 /*
2407                  * Make sure that no cached llcds left in recov_thread.
2408                  * We actually do sync in disconnect time, but disconnect
2409                  * may not come being marked rq_no_resend = 1.
2410                  */
2411                 llog_sync(ctxt, NULL);
2412
2413                 /*
2414                  * Balance class_import_get() in llog_receptor_accept().
2415                  * This is safe to do, as llog is already synchronized
2416                  * and its import may go.
2417                  */
2418                 cfs_mutex_lock(&ctxt->loc_mutex);
2419                 if (ctxt->loc_imp) {
2420                         class_import_put(ctxt->loc_imp);
2421                         ctxt->loc_imp = NULL;
2422                 }
2423                 cfs_mutex_unlock(&ctxt->loc_mutex);
2424                 llog_ctxt_put(ctxt);
2425         }
2426
2427         if (filter->fo_lcm) {
2428                 cfs_mutex_lock(&ctxt->loc_mutex);
2429                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2430                 filter->fo_lcm = NULL;
2431                 cfs_mutex_unlock(&ctxt->loc_mutex);
2432         }
2433         RETURN(filter_olg_fini(&obd->obd_olg));
2434 }
2435
2436 /**
2437  * Find the group llog according to group index in the llog group list.
2438  */
2439 static struct obd_llog_group *
2440 filter_find_olg_internal(struct filter_obd *filter, int group)
2441 {
2442         struct obd_llog_group *olg;
2443
2444         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2445         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2446                 if (olg->olg_seq == group)
2447                         RETURN(olg);
2448         }
2449         RETURN(NULL);
2450 }
2451
2452 /**
2453  * Find the group llog according to group index on the filter
2454  */
2455 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2456 {
2457         struct obd_llog_group *olg = NULL;
2458         struct filter_obd *filter;
2459
2460         filter = &obd->u.filter;
2461
2462         if (group == FID_SEQ_LLOG)
2463                 RETURN(&obd->obd_olg);
2464
2465         cfs_spin_lock(&filter->fo_llog_list_lock);
2466         olg = filter_find_olg_internal(filter, group);
2467         cfs_spin_unlock(&filter->fo_llog_list_lock);
2468
2469         RETURN(olg);
2470 }
2471 /**
2472  * Find the llog_group of the filter according to the group. If it can not
2473  * find, create the llog_group, which only happens when mds is being synced
2474  * with OST.
2475  */
2476 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2477 {
2478         struct obd_llog_group *olg = NULL, *olg_new = NULL;
2479         struct filter_obd *filter;
2480         int rc;
2481
2482         filter = &obd->u.filter;
2483
2484         if (group == FID_SEQ_LLOG)
2485                 RETURN(&obd->obd_olg);
2486
2487         OBD_ALLOC_PTR(olg_new);
2488         if (olg_new == NULL)
2489                RETURN(ERR_PTR(-ENOMEM));
2490
2491         cfs_spin_lock(&filter->fo_llog_list_lock);
2492         olg = filter_find_olg_internal(filter, group);
2493         if (olg) {
2494                 if (olg->olg_initializing) {
2495                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2496                 } else {
2497                         GOTO(out_unlock, olg);
2498                 }
2499         } else {
2500                 /* set as the newly allocated one */
2501                 olg = olg_new;
2502                 olg_new = NULL;
2503         }
2504
2505         llog_group_init(olg, group);
2506         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2507         olg->olg_initializing = 1;
2508         cfs_spin_unlock(&filter->fo_llog_list_lock);
2509
2510         rc = obd_llog_init(obd, olg, obd, NULL);
2511         if (rc) {
2512                cfs_spin_lock(&filter->fo_llog_list_lock);
2513                cfs_list_del(&olg->olg_list);
2514                cfs_spin_unlock(&filter->fo_llog_list_lock);
2515                OBD_FREE_PTR(olg);
2516                GOTO(out, olg = ERR_PTR(-ENOMEM));
2517         }
2518         cfs_spin_lock(&filter->fo_llog_list_lock);
2519         olg->olg_initializing = 0;
2520         cfs_spin_unlock(&filter->fo_llog_list_lock);
2521         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2522               obd->obd_name, group, olg);
2523 out:
2524         RETURN(olg);
2525
2526 out_unlock:
2527         cfs_spin_unlock(&filter->fo_llog_list_lock);
2528         if (olg_new)
2529                OBD_FREE_PTR(olg_new);
2530         goto out;
2531 }
2532
2533 static int filter_llog_connect(struct obd_export *exp,
2534                                struct llogd_conn_body *body)
2535 {
2536         struct obd_device *obd = exp->exp_obd;
2537         struct llog_ctxt *ctxt;
2538         struct obd_llog_group *olg;
2539         int rc;
2540         ENTRY;
2541
2542         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2543                obd->obd_name, body->lgdc_logid.lgl_oid,
2544                body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2545
2546         olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2547         if (!olg) {
2548                 CERROR(" %s: can not find olg of group %d\n",
2549                        obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2550                 RETURN(-ENOENT);
2551         }
2552         llog_group_set_export(olg, exp);
2553
2554         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2555         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2556                  body->lgdc_ctxt_idx);
2557
2558         CDEBUG(D_HA, "%s: Recovery from log "LPX64"/"LPX64":%x\n",
2559                obd->obd_name, body->lgdc_logid.lgl_oid,
2560                body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2561
2562         cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2563         obd->u.filter.fo_mds_ost_sync = 1;
2564         cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2565         rc = llog_connect(ctxt, &body->lgdc_logid,
2566                           &body->lgdc_gen, NULL);
2567         llog_ctxt_put(ctxt);
2568         if (rc != 0)
2569                 CERROR("failed to connect rc %d idx %d\n", rc,
2570                                 body->lgdc_ctxt_idx);
2571
2572         RETURN(rc);
2573 }
2574
2575 static int filter_llog_preclean(struct obd_device *obd)
2576 {
2577         struct obd_llog_group *olg, *tmp;
2578         struct filter_obd *filter;
2579         cfs_list_t  remove_list;
2580         int rc = 0;
2581         ENTRY;
2582
2583         rc = obd_llog_finish(obd, 0);
2584         if (rc)
2585                 CERROR("failed to cleanup llogging subsystem\n");
2586
2587         filter = &obd->u.filter;
2588         CFS_INIT_LIST_HEAD(&remove_list);
2589
2590         cfs_spin_lock(&filter->fo_llog_list_lock);
2591         while (!cfs_list_empty(&filter->fo_llog_list)) {
2592                 olg = cfs_list_entry(filter->fo_llog_list.next,
2593                                      struct obd_llog_group, olg_list);
2594                 cfs_list_del(&olg->olg_list);
2595                 cfs_list_add(&olg->olg_list, &remove_list);
2596         }
2597         cfs_spin_unlock(&filter->fo_llog_list_lock);
2598
2599         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2600                 cfs_list_del_init(&olg->olg_list);
2601                 rc = filter_olg_fini(olg);
2602                 if (rc)
2603                         CERROR("failed to cleanup llogging subsystem for %u\n",
2604                                olg->olg_seq);
2605                 OBD_FREE_PTR(olg);
2606         }
2607
2608         RETURN(rc);
2609 }
2610
2611 static int filter_precleanup(struct obd_device *obd,
2612                              enum obd_cleanup_stage stage)
2613 {
2614         int rc = 0;
2615         ENTRY;
2616
2617         switch(stage) {
2618         case OBD_CLEANUP_EARLY:
2619                 break;
2620         case OBD_CLEANUP_EXPORTS:
2621                 /* Stop recovery before namespace cleanup. */
2622                 target_recovery_fini(obd);
2623
2624                 obd_exports_barrier(obd);
2625                 obd_zombie_barrier();
2626
2627                 rc = filter_llog_preclean(obd);
2628                 lprocfs_job_stats_fini(obd);
2629                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2630                 lprocfs_free_per_client_stats(obd);
2631                 lprocfs_obd_cleanup(obd);
2632                 lprocfs_free_obd_stats(obd);
2633                 lquota_cleanup(filter_quota_interface_ref, obd);
2634                 break;
2635         }
2636         RETURN(rc);
2637 }
2638
2639 static int filter_cleanup(struct obd_device *obd)
2640 {
2641         struct filter_obd *filter = &obd->u.filter;
2642         ENTRY;
2643
2644         if (obd->obd_fail)
2645                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2646                               "will be preserved.\n", obd->obd_name);
2647
2648         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2649         obd->obd_namespace = NULL;
2650
2651         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2652
2653         if (obd->u.obt.obt_sb == NULL)
2654                 RETURN(0);
2655
2656         filter_post(obd);
2657
2658         ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2659         shrink_dcache_sb(obd->u.obt.obt_sb);
2660
2661         server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2662         obd->u.obt.obt_sb = NULL;
2663
2664         fsfilt_put_ops(obd->obd_fsops);
2665
2666         filter_iobuf_pool_done(filter);
2667
2668         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2669
2670         RETURN(0);
2671 }
2672
2673 static int filter_connect_internal(struct obd_export *exp,
2674                                    struct obd_connect_data *data,
2675                                    int reconnect)
2676 {
2677         struct filter_export_data *fed = &exp->exp_filter_data;
2678
2679         if (!data)
2680                 RETURN(0);
2681
2682         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2683                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2684                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2685                data->ocd_connect_flags, data->ocd_version,
2686                data->ocd_grant, data->ocd_index);
2687
2688         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2689                 CWARN("!!! This export (nid %s) used object group %d "
2690                        "earlier; now it's trying to use group %d!  This could "
2691                        "be a bug in the MDS. Please report to "
2692                        "http://bugs.whamcloud.com/\n",
2693                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2694                 RETURN(-EPROTO);
2695         }
2696         fed->fed_group = data->ocd_group;
2697
2698         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2699         exp->exp_connect_flags = data->ocd_connect_flags;
2700         data->ocd_version = LUSTRE_VERSION_CODE;
2701
2702         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2703         if (data->ocd_connect_flags & OBD_CONNECT_MDS)
2704                 CDEBUG(D_HA, "%s: Received MDS connection for group %u\n",
2705                        exp->exp_obd->obd_name, data->ocd_group);
2706         else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
2707                 RETURN(-EPROTO);
2708
2709         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2710                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2711                 obd_size left, want;
2712
2713                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2714                 left = filter_grant_space_left(exp);
2715                 want = data->ocd_grant;
2716                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2717                 data->ocd_grant = fed->fed_grant;
2718                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2719
2720                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2721                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2722                        exp->exp_client_uuid.uuid, exp,
2723                        data->ocd_grant, want, left);
2724
2725                 filter->fo_tot_granted_clients ++;
2726         }
2727
2728         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2729                 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2730                 int index = le32_to_cpu(lsd->lsd_ost_index);
2731
2732                 if (!(lsd->lsd_feature_compat &
2733                       cpu_to_le32(OBD_COMPAT_OST))) {
2734                         /* this will only happen on the first connect */
2735                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2736                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2737                         /* sync is not needed here as filter_client_add will
2738                          * set exp_need_sync flag */
2739                         filter_update_server_data(exp->exp_obd);
2740                 } else if (index != data->ocd_index) {
2741                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2742                                            " %u doesn't match actual OST index"
2743                                            " %u in last_rcvd file, bad "
2744                                            "configuration?\n",
2745                                            obd_export_nid2str(exp), index,
2746                                            data->ocd_index);
2747                         RETURN(-EBADF);
2748                 }
2749                 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2750                  * FIXME: We don't strictly need the COMPAT flag for that,
2751                  * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2752                  * FIXME: We needed it for the index, as index 0 is valid. */
2753         }
2754
2755         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2756                 data->ocd_brw_size = 65536;
2757         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2758                 data->ocd_brw_size = min(data->ocd_brw_size,
2759                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2760                 if (data->ocd_brw_size == 0) {
2761                         CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2762                                " ocd_version: %x ocd_grant: %d ocd_index: %u "
2763                                "ocd_brw_size is unexpectedly zero, "
2764                                "network data corruption?"
2765                                "Refusing connection of this client\n",
2766                                 exp->exp_obd->obd_name,
2767                                 exp->exp_client_uuid.uuid,
2768                                 exp, data->ocd_connect_flags, data->ocd_version,
2769                                 data->ocd_grant, data->ocd_index);
2770                         RETURN(-EPROTO);
2771                 }
2772         }
2773
2774         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2775                 __u32 cksum_types = data->ocd_cksum_types;
2776
2777                 /* The client set in ocd_cksum_types the checksum types it
2778                  * supports. We have to mask off the algorithms that we don't
2779                  * support */
2780                 data->ocd_cksum_types &= cksum_types_supported();
2781
2782                 /* 1.6.4- only support CRC32 and didn't set ocd_cksum_types */
2783                 if (unlikely(data->ocd_cksum_types == 0))
2784                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2785
2786                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2787                                    "%x\n", exp->exp_obd->obd_name,
2788                                    obd_export_nid2str(exp), cksum_types,
2789                                    data->ocd_cksum_types);
2790         } else {
2791                 /* This client does not support OBD_CONNECT_CKSUM
2792                  * fall back to CRC32 */
2793                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2794                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2795                                    exp->exp_obd->obd_name,
2796                                    obd_export_nid2str(exp));
2797         }
2798
2799         if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2800                 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2801
2802         RETURN(0);
2803 }
2804
2805 static int filter_reconnect(const struct lu_env *env,
2806                             struct obd_export *exp, struct obd_device *obd,
2807                             struct obd_uuid *cluuid,
2808                             struct obd_connect_data *data,
2809                             void *localdata)
2810 {
2811         int rc;
2812         ENTRY;
2813
2814         if (exp == NULL || obd == NULL || cluuid == NULL)
2815                 RETURN(-EINVAL);
2816
2817         rc = filter_connect_internal(exp, data, 1);
2818         if (rc == 0)
2819                 filter_export_stats_init(obd, exp, localdata);
2820
2821         RETURN(rc);
2822 }
2823
2824 static int filter_connect(const struct lu_env *env,
2825                           struct obd_export **exp, struct obd_device *obd,
2826                           struct obd_uuid *cluuid,
2827                           struct obd_connect_data *data, void *localdata)
2828 {
2829         struct lvfs_run_ctxt saved;
2830         struct lustre_handle conn = { 0 };
2831         struct obd_export *lexp;
2832         int rc;
2833         ENTRY;
2834
2835         if (exp == NULL || obd == NULL || cluuid == NULL)
2836                 RETURN(-EINVAL);
2837
2838         rc = class_connect(&conn, obd, cluuid);
2839         if (rc)
2840                 RETURN(rc);
2841         lexp = class_conn2export(&conn);
2842         LASSERT(lexp != NULL);
2843
2844         rc = filter_connect_internal(lexp, data, 0);
2845         if (rc)
2846                 GOTO(cleanup, rc);
2847
2848         filter_export_stats_init(obd, lexp, localdata);
2849         if (obd->obd_replayable) {
2850                 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2851                 LASSERT(lcd);
2852                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2853                 rc = filter_client_add(obd, lexp, -1);
2854                 if (rc)
2855                         GOTO(cleanup, rc);
2856         }
2857
2858         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2859         rc = filter_read_groups(obd, data->ocd_group, 1);
2860         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2861         if (rc != 0) {
2862                 CERROR("can't read group %u\n", data->ocd_group);
2863                 GOTO(cleanup, rc);
2864         }
2865
2866         GOTO(cleanup, rc);
2867
2868 cleanup:
2869         if (rc) {
2870                 class_disconnect(lexp);
2871                 *exp = NULL;
2872         } else {
2873                 *exp = lexp;
2874         }
2875
2876         RETURN(rc);
2877 }
2878
2879 /* Do extra sanity checks for grant accounting.  We do this at connect,
2880  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2881  * always get rid of it or turn it off when we know accounting is good. */
2882 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2883 {
2884         struct filter_export_data *fed;
2885         struct obd_export *exp;
2886         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2887         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2888         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2889
2890         if (cfs_list_empty(&obd->obd_exports))
2891                 return;
2892
2893         /* We don't want to do this for large machines that do lots of
2894            mounts or unmounts.  It burns... */
2895         if (obd->obd_num_exports > 100)
2896                 return;
2897
2898         cfs_spin_lock(&obd->obd_osfs_lock);
2899         cfs_spin_lock(&obd->obd_dev_lock);
2900         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2901                 int error = 0;
2902                 fed = &exp->exp_filter_data;
2903                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2904                     fed->fed_dirty < 0)
2905                         error = 1;
2906                 if (maxsize > 0) { /* we may not have done a statfs yet */
2907                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2908                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2909                                  exp->exp_client_uuid.uuid, exp,
2910                                  fed->fed_grant, fed->fed_pending, maxsize);
2911                         LASSERTF(fed->fed_dirty <= maxsize,
2912                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2913                                  exp->exp_client_uuid.uuid, exp,
2914                                  fed->fed_dirty, maxsize);
2915                 }
2916                 if (error)
2917                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2918                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2919                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2920                 else
2921                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2922                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2923                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2924                 tot_granted += fed->fed_grant + fed->fed_pending;
2925                 tot_pending += fed->fed_pending;
2926                 tot_dirty += fed->fed_dirty;
2927         }
2928         fo_tot_granted = obd->u.filter.fo_tot_granted;
2929         fo_tot_pending = obd->u.filter.fo_tot_pending;
2930         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2931         cfs_spin_unlock(&obd->obd_dev_lock);
2932         cfs_spin_unlock(&obd->obd_osfs_lock);
2933
2934         /* Do these assertions outside the spinlocks so we don't kill system */
2935         if (tot_granted != fo_tot_granted)
2936                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2937                        func, tot_granted, fo_tot_granted);
2938         if (tot_pending != fo_tot_pending)
2939                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2940                        func, tot_pending, fo_tot_pending);
2941         if (tot_dirty != fo_tot_dirty)
2942                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2943                        func, tot_dirty, fo_tot_dirty);
2944         if (tot_pending > tot_granted)
2945                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2946                        func, tot_pending, tot_granted);
2947         if (tot_granted > maxsize)
2948                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2949                        func, tot_granted, maxsize);
2950         if (tot_dirty > maxsize)
2951                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2952                        func, tot_dirty, maxsize);
2953 }
2954
2955 /* Remove this client from the grant accounting totals.  We also remove
2956  * the export from the obd device under the osfs and dev locks to ensure
2957  * that the filter_grant_sanity_check() calculations are always valid.
2958  * The client should do something similar when it invalidates its import. */
2959 static void filter_grant_discard(struct obd_export *exp)
2960 {
2961         struct obd_device *obd = exp->exp_obd;
2962         struct filter_obd *filter = &obd->u.filter;
2963         struct filter_export_data *fed = &exp->exp_filter_data;
2964
2965         cfs_spin_lock(&obd->obd_osfs_lock);
2966         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2967                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2968                  obd->obd_name, filter->fo_tot_granted,
2969                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2970         filter->fo_tot_granted -= fed->fed_grant;
2971         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2972                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2973                  obd->obd_name, filter->fo_tot_pending,
2974                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2975         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2976         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2977                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2978                  obd->obd_name, filter->fo_tot_dirty,
2979                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2980         filter->fo_tot_dirty -= fed->fed_dirty;
2981         fed->fed_dirty = 0;
2982         fed->fed_grant = 0;
2983
2984         cfs_spin_unlock(&obd->obd_osfs_lock);
2985 }
2986
2987 static int filter_destroy_export(struct obd_export *exp)
2988 {
2989         struct filter_export_data *fed = &exp->exp_filter_data;
2990         ENTRY;
2991
2992         if (fed->fed_pending)
2993                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2994                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2995                        exp, fed->fed_pending);
2996
2997         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2998
2999         target_destroy_export(exp);
3000
3001         if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
3002                                      &exp->exp_client_uuid)))
3003                RETURN(0);
3004
3005         ldlm_destroy_export(exp);
3006         lut_client_free(exp);
3007
3008         if (!exp->exp_obd->obd_replayable)
3009                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3010
3011         filter_grant_discard(exp);
3012         filter_fmd_cleanup(exp);
3013
3014         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3015                 struct filter_obd *filter = &exp->exp_obd->u.filter;
3016                 if (filter->fo_tot_granted_clients > 0)
3017                         filter->fo_tot_granted_clients --;
3018         }
3019
3020         if (!(exp->exp_flags & OBD_OPT_FORCE))
3021                 filter_grant_sanity_check(exp->exp_obd, __func__);
3022
3023         RETURN(0);
3024 }
3025
3026 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3027 {
3028         struct obd_llog_group *olg_min, *olg;
3029         struct filter_obd *filter;
3030         int worked = -1, group;
3031         struct llog_ctxt *ctxt;
3032         ENTRY;
3033
3034         filter = &obd->u.filter;
3035
3036         /* we can't sync log holding spinlock. also, we do not want to get
3037          * into livelock. so we do following: loop over MDS's exports in
3038          * group order and skip already synced llogs -bzzz */
3039         do {
3040                 /* look for group with min. number, but > worked */
3041                 olg_min = NULL;
3042                 group = 1 << 30;
3043                 cfs_spin_lock(&filter->fo_llog_list_lock);
3044                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3045                         if (olg->olg_seq <= worked) {
3046                                 /* this group is already synced */
3047                                 continue;
3048                         }
3049                         if (group < olg->olg_seq) {
3050                                 /* we have group with smaller number to sync */
3051                                 continue;
3052                         }
3053                         /* store current minimal group */
3054                         olg_min = olg;
3055                         group = olg->olg_seq;
3056                 }
3057                 cfs_spin_unlock(&filter->fo_llog_list_lock);
3058
3059                 if (olg_min == NULL)
3060                         break;
3061
3062                 worked = olg_min->olg_seq;
3063                 if (olg_min->olg_exp &&
3064                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3065                         int err;
3066                         ctxt = llog_group_get_ctxt(olg_min,
3067                                                    LLOG_MDS_OST_REPL_CTXT);
3068                         if (ctxt) {
3069                                 err = llog_sync(ctxt, olg_min->olg_exp);
3070                                 llog_ctxt_put(ctxt);
3071                                 if (err) {
3072                                         CERROR("error flushing logs to MDS: "
3073                                                "rc %d\n", err);
3074                                 }
3075                         }
3076                 }
3077         } while (olg_min != NULL);
3078 }
3079
3080 /* Also incredibly similar to mds_disconnect */
3081 static int filter_disconnect(struct obd_export *exp)
3082 {
3083         struct obd_device *obd = exp->exp_obd;
3084         int rc;
3085         ENTRY;
3086
3087         LASSERT(exp);
3088         class_export_get(exp);
3089
3090         if (!(exp->exp_flags & OBD_OPT_FORCE))
3091                 filter_grant_sanity_check(obd, __func__);
3092         filter_grant_discard(exp);
3093
3094         /* Flush any remaining cancel messages out to the target */
3095         filter_sync_llogs(obd, exp);
3096
3097         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3098
3099         rc = server_disconnect_export(exp);
3100
3101         /* Do not erase record for recoverable client. */
3102         if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3103                 filter_client_del(exp);
3104         else
3105                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3106
3107         class_export_put(exp);
3108         RETURN(rc);
3109 }
3110
3111 /* reverse import is changed, sync all cancels */
3112 static void filter_revimp_update(struct obd_export *exp)
3113 {
3114         ENTRY;
3115
3116         LASSERT(exp);
3117         class_export_get(exp);
3118
3119         /* flush any remaining cancel messages out to the target */
3120         filter_sync_llogs(exp->exp_obd, exp);
3121         class_export_put(exp);
3122         EXIT;
3123 }
3124
3125 static int filter_ping(const struct lu_env *env, struct obd_export *exp)
3126 {
3127         filter_fmd_expire(exp);
3128         return 0;
3129 }
3130
3131 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3132                                   const char *what, int quiet)
3133 {
3134         struct dentry *dchild = NULL;
3135
3136         dchild = filter_fid2dentry(obd, NULL,  ostid->oi_seq, ostid->oi_id);
3137
3138         if (IS_ERR(dchild)) {
3139                 CERROR("%s error looking up object: "POSTID"\n",
3140                        what, ostid->oi_id, ostid->oi_seq);
3141                 RETURN(dchild);
3142         }
3143
3144         if (dchild->d_inode == NULL) {
3145                 if (!quiet)
3146                         CERROR("%s: %s on non-existent object: "POSTID" \n",
3147                                obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3148                 f_dput(dchild);
3149                 RETURN(ERR_PTR(-ENOENT));
3150         }
3151
3152 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0)
3153         /* Try to correct for a bug in 2.1.0 (LU-221) that caused negative
3154          * timestamps to appear to be in the far future, due old timestamp
3155          * being stored on disk as an unsigned value.  This fixes up any
3156          * bad values stored on disk before returning them to the client,
3157          * and ensures any timestamp updates are correct.  LU-1042 */
3158         if (unlikely(LTIME_S(dchild->d_inode->i_atime) == LU221_BAD_TIME))
3159                 LTIME_S(dchild->d_inode->i_atime) = 0;
3160         if (unlikely(LTIME_S(dchild->d_inode->i_mtime) == LU221_BAD_TIME))
3161                 LTIME_S(dchild->d_inode->i_mtime) = 0;
3162         if (unlikely(LTIME_S(dchild->d_inode->i_ctime) == LU221_BAD_TIME))
3163                 LTIME_S(dchild->d_inode->i_ctime) = 0;
3164 #else
3165 #warning "remove old LU-221/LU-1042 workaround code"
3166 #endif
3167
3168         return dchild;
3169 }
3170
3171 static int filter_getattr(const struct lu_env *env, struct obd_export *exp,
3172                           struct obd_info *oinfo)
3173 {
3174         struct dentry *dentry = NULL;
3175         struct obd_device *obd;
3176         __u64 curr_version;
3177         int rc = 0;
3178         ENTRY;
3179
3180         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3181                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3182         if (rc)
3183                 RETURN(rc);
3184
3185         obd = class_exp2obd(exp);
3186         if (obd == NULL) {
3187                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3188                 RETURN(-EINVAL);
3189         }
3190
3191         dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3192         if (IS_ERR(dentry))
3193                 RETURN(PTR_ERR(dentry));
3194
3195         /* Limit the valid bits in the return data to what we actually use */
3196         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3197         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3198
3199         /* Store inode version in reply */
3200         curr_version = fsfilt_get_version(exp->exp_obd, dentry->d_inode);
3201         if ((__s64)curr_version != -EOPNOTSUPP) {
3202                 oinfo->oi_oa->o_valid |= OBD_MD_FLDATAVERSION;
3203                 oinfo->oi_oa->o_data_version = curr_version;
3204         }
3205
3206       &nbs