Whamcloud - gitweb
b=22273 Don't remove lcd in case of client_add failure
[fs/lustre-release.git] / lustre / obdfilter / filter.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/obdfilter/filter.c
37  *
38  * Author: Peter Braam <braam@clusterfs.com>
39  * Author: Andreas Dilger <adilger@clusterfs.com>
40  */
41
42 /*
43  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44  *            (which need to get journal_lock, may block if journal full).
45  *
46  * Invariant: Call filter_start_transno() before any journal ops to avoid the
47  *            same deadlock problem.  We can (and want) to get rid of the
48  *            transno sem in favour of the dir/inode i_mutex to avoid single
49  *            threaded operation on the OST.
50  */
51
52 #define DEBUG_SUBSYSTEM S_FILTER
53
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
56 #endif
57 #include <linux/module.h>
58 #include <linux/fs.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
65
66 #include <obd_cksum.h>
67 #include <obd_class.h>
68 #include <obd_lov.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
79
80 #include "filter_internal.h"
81
82 static struct lvfs_callback_ops filter_lvfs_ops;
83 cfs_mem_cache_t *ll_fmd_cachep;
84
85 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
86                              void *cb_data, int error)
87 {
88         struct obd_export *exp = cb_data;
89         LASSERT(exp->exp_obd == obd);
90         obd_transno_commit_cb(obd, transno, exp, error);
91         class_export_cb_put(exp);
92 }
93
94 int filter_version_get_check(struct obd_export *exp,
95                              struct obd_trans_info *oti, struct inode *inode)
96 {
97         __u64 curr_version;
98
99         if (inode == NULL || oti == NULL)
100                 RETURN(0);
101
102         curr_version = fsfilt_get_version(exp->exp_obd, inode);
103         if ((__s64)curr_version == -EOPNOTSUPP)
104                 RETURN(0);
105         /* VBR: version is checked always because costs nothing */
106         if (oti->oti_pre_version != 0 &&
107             oti->oti_pre_version != curr_version) {
108                 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
109                        oti->oti_pre_version, curr_version);
110                 cfs_spin_lock(&exp->exp_lock);
111                 exp->exp_vbr_failed = 1;
112                 cfs_spin_unlock(&exp->exp_lock);
113                 RETURN (-EOVERFLOW);
114         }
115         oti->oti_pre_version = curr_version;
116         RETURN(0);
117 }
118
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121                           struct obd_trans_info *oti, int rc, int force_sync)
122 {
123         struct filter_obd *filter = &exp->exp_obd->u.filter;
124         struct filter_export_data *fed = &exp->exp_filter_data;
125         struct lr_server_data *fsd = class_server_data(exp->exp_obd);
126         struct lsd_client_data *lcd;
127         __u64 last_rcvd;
128         loff_t off;
129         int err, log_pri = D_RPCTRACE;
130
131         /* Propagate error code. */
132         if (rc)
133                 RETURN(rc);
134
135         if (!exp->exp_obd->obd_replayable || oti == NULL)
136                 RETURN(rc);
137
138         cfs_mutex_down(&fed->fed_lcd_lock);
139         lcd = fed->fed_lcd;
140         /* if the export has already been disconnected, we have no last_rcvd slot,
141          * update server data with latest transno then */
142         if (lcd == NULL) {
143                 cfs_mutex_up(&fed->fed_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(&filter->fo_translock);
152         if (oti->oti_transno == 0) {
153                 last_rcvd = le64_to_cpu(fsd->lsd_last_transno) + 1;
154                 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
155         } else {
156                 last_rcvd = oti->oti_transno;
157                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
158                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
159         }
160         oti->oti_transno = last_rcvd;
161
162         LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
163         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
164         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
165         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
166         cfs_spin_unlock(&filter->fo_translock);
167
168         if (inode)
169                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
170
171         off = fed->fed_lr_off;
172         if (off <= 0) {
173                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
174                        fed->fed_lr_idx, fed->fed_lr_off);
175                 err = -EINVAL;
176         } else {
177                 class_export_cb_get(exp); /* released when the cb is called */
178                 if (!force_sync)
179                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
180                                                            last_rcvd,
181                                                            oti->oti_handle,
182                                                            filter_commit_cb,
183                                                            exp);
184
185                 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
186                                           lcd, sizeof(*lcd), &off,
187                                           force_sync | exp->exp_need_sync);
188                 if (force_sync)
189                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
190         }
191         if (err) {
192                 log_pri = D_ERROR;
193                 if (rc == 0)
194                         rc = err;
195         }
196
197         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
198                last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
199         cfs_mutex_up(&fed->fed_lcd_lock);
200         RETURN(rc);
201 }
202
203 void f_dput(struct dentry *dentry)
204 {
205         /* Can't go inside filter_ddelete because it can block */
206         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
207                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
208         LASSERT(atomic_read(&dentry->d_count) > 0);
209
210         dput(dentry);
211 }
212
213 static void init_brw_stats(struct brw_stats *brw_stats)
214 {
215         int i;
216         for (i = 0; i < BRW_LAST; i++)
217                 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
218 }
219
220 static int lprocfs_init_rw_stats(struct obd_device *obd,
221                                  struct lprocfs_stats **stats)
222 {
223         int num_stats;
224
225         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
226                                                         LPROC_FILTER_LAST - 1;
227         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
228         if (*stats == NULL)
229                 return -ENOMEM;
230
231         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
232         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
233                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
234         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
235                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
236
237         return(0);
238 }
239
240 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
241    plus the procfs overhead :( */
242 static int filter_export_stats_init(struct obd_device *obd,
243                                     struct obd_export *exp,
244                                     void *client_nid)
245 {
246         int rc, newnid = 0;
247         ENTRY;
248
249         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
250                 /* Self-export gets no proc entry */
251                 RETURN(0);
252
253         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
254         if (rc) {
255                 /* Mask error for already created
256                  * /proc entries */
257                 if (rc == -EALREADY)
258                         rc = 0;
259                 RETURN(rc);
260         }
261
262         if (newnid) {
263                 struct nid_stat *tmp = exp->exp_nid_stats;
264                 LASSERT(tmp != NULL);
265
266                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
267                 if (tmp->nid_brw_stats == NULL)
268                         GOTO(clean, rc = -ENOMEM);
269
270                 init_brw_stats(tmp->nid_brw_stats);
271                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
272                                         0644, &filter_per_nid_stats_fops,
273                                         exp->exp_nid_stats);
274                 if (rc)
275                         CWARN("Error adding the brw_stats file\n");
276
277                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
278                 if (rc)
279                         GOTO(clean, rc);
280
281                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
282                                             tmp->nid_stats);
283                 if (rc)
284                         GOTO(clean, rc);
285                 /* Always add in ldlm_stats */
286                 tmp->nid_ldlm_stats = 
287                         lprocfs_alloc_stats(LDLM_LAST_OPC - LDLM_FIRST_OPC,
288                                             LPROCFS_STATS_FLAG_NOPERCPU);
289                 if (tmp->nid_ldlm_stats == NULL)
290                         GOTO(clean, rc = -ENOMEM);
291
292                 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
293                 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
294                                             tmp->nid_ldlm_stats);
295                 if (rc)
296                         GOTO(clean, rc);
297         }
298
299         RETURN(0);
300  clean:
301         lprocfs_exp_cleanup(exp);
302         return rc;
303 }
304
305 /* Add client data to the FILTER.  We use a bitmap to locate a free space
306  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
307  * Otherwise, we have just read the data from the last_rcvd file and
308  * we know its offset. */
309 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
310                              int cl_idx)
311 {
312         struct filter_obd *filter = &obd->u.filter;
313         struct filter_export_data *fed = &exp->exp_filter_data;
314         struct lr_server_data *fsd = class_server_data(obd);
315         unsigned long *bitmap = filter->fo_last_rcvd_slots;
316         int new_client = (cl_idx == -1);
317
318         ENTRY;
319
320         LASSERT(bitmap != NULL);
321         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
322
323         /* Self-export */
324         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
325                 RETURN(0);
326
327         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
328          * there's no need for extra complication here
329          */
330         if (new_client) {
331                 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
332         repeat:
333                 if (cl_idx >= LR_MAX_CLIENTS) {
334                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
335                                cl_idx);
336                         RETURN(-EOVERFLOW);
337                 }
338                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
339                         cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
340                                                         cl_idx);
341                         goto repeat;
342                 }
343         } else {
344                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
345                         CERROR("FILTER client %d: bit already set in bitmap!\n",
346                                cl_idx);
347                         LBUG();
348                 }
349         }
350
351         fed->fed_lr_idx = cl_idx;
352         fed->fed_lr_off = le32_to_cpu(fsd->lsd_client_start) +
353                           cl_idx * le16_to_cpu(fsd->lsd_client_size);
354         cfs_init_mutex(&fed->fed_lcd_lock);
355         LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
356
357         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
358                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
359
360         if (new_client) {
361                 struct lvfs_run_ctxt saved;
362                 loff_t off = fed->fed_lr_off;
363                 int rc;
364                 void *handle;
365
366                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
367                        fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
368
369                 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
370                         RETURN(-ENOSPC);
371
372                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
373                 /* Transaction needed to fix bug 1403 */
374                 handle = fsfilt_start(obd,
375                                       filter->fo_rcvd_filp->f_dentry->d_inode,
376                                       FSFILT_OP_SETATTR, NULL);
377                 if (IS_ERR(handle)) {
378                         rc = PTR_ERR(handle);
379                         CERROR("unable to start transaction: rc %d\n", rc);
380                 } else {
381                         fed->fed_lcd->lcd_last_epoch = fsd->lsd_start_epoch;
382                         exp->exp_last_request_time = cfs_time_current_sec();
383                         rc = fsfilt_add_journal_cb(obd, 0, handle,
384                                                    target_client_add_cb,
385                                                    class_export_cb_get(exp));
386                         if (rc == 0) {
387                                 cfs_spin_lock(&exp->exp_lock);
388                                 exp->exp_need_sync = 1;
389                                 cfs_spin_unlock(&exp->exp_lock);
390                         }
391                         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
392                                                  fed->fed_lcd,
393                                                  sizeof(*fed->fed_lcd),
394                                                  &off, rc /* sync if no cb */);
395                         fsfilt_commit(obd,
396                                       filter->fo_rcvd_filp->f_dentry->d_inode,
397                                       handle, 0);
398                 }
399                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
400
401                 if (rc) {
402                         CERROR("error writing %s client idx %u: rc %d\n",
403                                LAST_RCVD, fed->fed_lr_idx, rc);
404                         RETURN(rc);
405                 }
406         }
407         RETURN(0);
408 }
409
410 static int filter_client_del(struct obd_export *exp)
411 {
412         struct filter_export_data *fed = &exp->exp_filter_data;
413         struct filter_obd *filter = &exp->exp_obd->u.filter;
414         struct obd_device *obd = exp->exp_obd;
415         struct lvfs_run_ctxt saved;
416         struct lsd_client_data *lcd = fed->fed_lcd;
417         int rc;
418         loff_t off;
419         ENTRY;
420
421         if (lcd == NULL)
422                 RETURN(0);
423
424         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
425         if (strcmp(lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
426                 GOTO(free, 0);
427
428         LASSERT(filter->fo_last_rcvd_slots != NULL);
429
430         off = fed->fed_lr_off;
431
432         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
433                fed->fed_lr_idx, fed->fed_lr_off, lcd->lcd_uuid);
434
435         /* Don't clear fed_lr_idx here as it is likely also unset.  At worst
436          * we leak a client slot that will be cleaned on the next recovery. */
437         if (off <= 0) {
438                 CERROR("%s: client idx %d has med_off %lld\n",
439                        obd->obd_name, fed->fed_lr_idx, off);
440                 GOTO(free, rc = -EINVAL);
441         }
442
443         /* Clear the bit _after_ zeroing out the client so we don't
444            race with filter_client_add and zero out new clients.*/
445         if (!cfs_test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
446                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
447                        fed->fed_lr_idx);
448                 LBUG();
449         }
450
451         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
452         /* Make sure the server's last_transno is up to date.
453          * This should be done before zeroing client slot so last_transno will
454          * be in server data or in client data in case of failure */
455         filter_update_server_data(obd);
456
457         cfs_mutex_down(&fed->fed_lcd_lock);
458         memset(fed->fed_lcd->lcd_uuid, 0, sizeof fed->fed_lcd->lcd_uuid);
459         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
460                                  fed->fed_lcd,
461                                  sizeof(*fed->fed_lcd), &off, 0);
462         cfs_mutex_up(&fed->fed_lcd_lock);
463         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
464
465         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
466                "zero out client %s at idx %u/%llu in %s, rc %d\n",
467                lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
468                LAST_RCVD, rc);
469         EXIT;
470 free:
471         return 0;
472 }
473
474 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
475 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
476                                          struct filter_mod_data *fmd)
477 {
478         LASSERT_SPIN_LOCKED(&fed->fed_lock);
479         if (--fmd->fmd_refcount == 0) {
480                 /* XXX when we have persistent reservations and the handle
481                  * is stored herein we need to drop it here. */
482                 fed->fed_mod_count--;
483                 cfs_list_del(&fmd->fmd_list);
484                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
485         }
486 }
487
488 /* drop fmd reference, free it if last ref */
489 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
490 {
491         struct filter_export_data *fed;
492
493         if (fmd == NULL)
494                 return;
495
496         fed = &exp->exp_filter_data;
497         cfs_spin_lock(&fed->fed_lock);
498         filter_fmd_put_nolock(fed, fmd); /* caller reference */
499         cfs_spin_unlock(&fed->fed_lock);
500 }
501
502 /* expire entries from the end of the list if there are too many
503  * or they are too old */
504 static void filter_fmd_expire_nolock(struct filter_obd *filter,
505                                      struct filter_export_data *fed,
506                                      struct filter_mod_data *keep)
507 {
508         struct filter_mod_data *fmd, *tmp;
509
510         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
511                 if (fmd == keep)
512                         break;
513
514                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
515                     fed->fed_mod_count < filter->fo_fmd_max_num)
516                         break;
517
518                 cfs_list_del_init(&fmd->fmd_list);
519                 filter_fmd_put_nolock(fed, fmd); /* list reference */
520         }
521 }
522
523 void filter_fmd_expire(struct obd_export *exp)
524 {
525         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
526         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
527                                  &exp->exp_filter_data, NULL);
528         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
529 }
530
531 /* find specified objid, group in export fmd list.
532  * caller must hold fed_lock and take fmd reference itself */
533 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
534                                                 struct filter_export_data *fed,
535                                                 obd_id objid, obd_gr group)
536 {
537         struct filter_mod_data *found = NULL, *fmd;
538
539         LASSERT_SPIN_LOCKED(&fed->fed_lock);
540
541         cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
542                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
543                         found = fmd;
544                         cfs_list_del(&fmd->fmd_list);
545                         cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
546                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
547                         break;
548                 }
549         }
550
551         filter_fmd_expire_nolock(filter, fed, found);
552
553         return found;
554 }
555
556 /* Find fmd based on objid and group, or return NULL if not found. */
557 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
558                                         obd_id objid, obd_gr group)
559 {
560         struct filter_mod_data *fmd;
561
562         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
563         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
564                                      &exp->exp_filter_data, objid, group);
565         if (fmd)
566                 fmd->fmd_refcount++;    /* caller reference */
567         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
568
569         return fmd;
570 }
571
572 /* Find fmd based on objid and group, or create a new one if none is found.
573  * It is possible for this function to return NULL under memory pressure,
574  * or if objid = 0 is passed (which will only cause old entries to expire).
575  * Currently this is not fatal because any fmd state is transient and
576  * may also be freed when it gets sufficiently old. */
577 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
578                                        obd_id objid, obd_gr group)
579 {
580         struct filter_export_data *fed = &exp->exp_filter_data;
581         struct filter_mod_data *found = NULL, *fmd_new = NULL;
582
583         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
584
585         cfs_spin_lock(&fed->fed_lock);
586         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
587         if (fmd_new) {
588                 if (found == NULL) {
589                         cfs_list_add_tail(&fmd_new->fmd_list,
590                                           &fed->fed_mod_list);
591                         fmd_new->fmd_id = objid;
592                         fmd_new->fmd_gr = group;
593                         fmd_new->fmd_refcount++;   /* list reference */
594                         found = fmd_new;
595                         fed->fed_mod_count++;
596                 } else {
597                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
598                 }
599         }
600         if (found) {
601                 found->fmd_refcount++;          /* caller reference */
602                 found->fmd_expire = jiffies +
603                         exp->exp_obd->u.filter.fo_fmd_max_age;
604         }
605
606         cfs_spin_unlock(&fed->fed_lock);
607
608         return found;
609 }
610
611 #ifdef DO_FMD_DROP
612 /* drop fmd list reference so it will disappear when last reference is put.
613  * This isn't so critical because it would in fact only affect the one client
614  * that is doing the unlink and at worst we have an stale entry referencing
615  * an object that should never be used again. */
616 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
617 {
618         struct filter_mod_data *found = NULL;
619
620         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
621         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
622         if (found) {
623                 cfs_list_del_init(&found->fmd_list);
624                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
625         }
626         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
627 }
628 #else
629 #define filter_fmd_drop(exp, objid, group)
630 #endif
631
632 /* remove all entries from fmd list */
633 static void filter_fmd_cleanup(struct obd_export *exp)
634 {
635         struct filter_export_data *fed = &exp->exp_filter_data;
636         struct filter_mod_data *fmd = NULL, *tmp;
637
638         cfs_spin_lock(&fed->fed_lock);
639         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
640                 cfs_list_del_init(&fmd->fmd_list);
641                 filter_fmd_put_nolock(fed, fmd);
642         }
643         cfs_spin_unlock(&fed->fed_lock);
644 }
645
646 static int filter_init_export(struct obd_export *exp)
647 {
648         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
649         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
650
651         cfs_spin_lock(&exp->exp_lock);
652         exp->exp_connecting = 1;
653         cfs_spin_unlock(&exp->exp_lock);
654
655         return ldlm_init_export(exp);
656 }
657
658 static int filter_free_server_data(struct filter_obd *filter)
659 {
660         lut_fini(NULL, filter->fo_obt.obt_lut);
661         OBD_FREE_PTR(filter->fo_obt.obt_lut);
662         return 0;
663 }
664
665 /* assumes caller is already in kernel ctxt */
666 int filter_update_server_data(struct obd_device *obd)
667 {
668         struct file *filp = obd->u.obt.obt_rcvd_filp;
669         struct lr_server_data *fsd = class_server_data(obd);
670         loff_t off = 0;
671         int rc;
672         ENTRY;
673
674         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
675         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
676                le64_to_cpu(fsd->lsd_last_transno));
677         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
678                le64_to_cpu(fsd->lsd_mount_count));
679
680         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, 0);
681         if (rc)
682                 CERROR("error writing lr_server_data: rc = %d\n", rc);
683
684         RETURN(rc);
685 }
686
687 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
688                              int force_sync)
689 {
690         struct filter_obd *filter = &obd->u.filter;
691         __u64 tmp;
692         loff_t off = 0;
693         int rc;
694         ENTRY;
695
696         if (filter->fo_last_objid_files[group] == NULL) {
697                 CERROR("Object group "LPU64" not fully setup; not updating "
698                        "last_objid\n", group);
699                 RETURN(-EINVAL);
700         }
701
702         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
703                obd->obd_name, group, filter->fo_last_objids[group]);
704
705         tmp = cpu_to_le64(filter->fo_last_objids[group]);
706         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
707                                  &tmp, sizeof(tmp), &off, force_sync);
708         if (rc)
709                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
710                        group, rc);
711         RETURN(rc);
712 }
713 extern int ost_handle(struct ptlrpc_request *req);
714 /* assumes caller has already in kernel ctxt */
715 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
716 {
717         struct filter_obd *filter = &obd->u.filter;
718         struct lr_server_data *fsd;
719         struct lsd_client_data *lcd = NULL;
720         struct inode *inode = filp->f_dentry->d_inode;
721         unsigned long last_rcvd_size = i_size_read(inode);
722         struct lu_target *lut;
723         __u64 mount_count;
724         __u32 start_epoch;
725         int cl_idx;
726         loff_t off = 0;
727         int rc;
728
729         /* ensure padding in the struct is the correct size */
730         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
731                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
732         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
733                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
734
735         /* allocate and initialize lu_target */
736         OBD_ALLOC_PTR(lut);
737         if (lut == NULL)
738                 RETURN(-ENOMEM);
739         rc = lut_init(NULL, lut, obd, NULL);
740         if (rc)
741                 GOTO(err_fsd, rc);
742         fsd = class_server_data(obd);
743         if (last_rcvd_size == 0) {
744                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
745
746                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
747                 fsd->lsd_last_transno = 0;
748                 mount_count = fsd->lsd_mount_count = 0;
749                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
750                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
751                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
752                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
753                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
754                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
755                  * MDS first connects and assigns the OST index number. */
756                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
757                                                         OBD_INCOMPAT_OST);
758         } else {
759                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
760                 if (rc) {
761                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
762                                LAST_RCVD, rc);
763                         GOTO(err_fsd, rc);
764                 }
765                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
766                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
767                                            "using the wrong disk %s. Were the "
768                                            "/dev/ assignments rearranged?\n",
769                                            obd->obd_uuid.uuid, fsd->lsd_uuid);
770                         GOTO(err_fsd, rc = -EINVAL);
771                 }
772                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
773                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
774                 /* COMPAT_146 */
775                 /* Assume old last_rcvd format unless I_C_LR is set */
776                 if (!(fsd->lsd_feature_incompat &
777                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
778                         fsd->lsd_last_transno = fsd->lsd_compat14;
779                 /* end COMPAT_146 */
780                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
781                  * MDS first connects and assigns the OST index number. */
782                 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
783                                                          OBD_INCOMPAT_OST);
784         }
785
786         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
787                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
788                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
789                        ~FILTER_INCOMPAT_SUPP);
790                 GOTO(err_fsd, rc = -EINVAL);
791         }
792         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
793                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
794                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
795                        ~FILTER_ROCOMPAT_SUPP);
796                 /* Do something like remount filesystem read-only */
797                 GOTO(err_fsd, rc = -EINVAL);
798         }
799
800         start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
801
802         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
803                obd->obd_name, start_epoch);
804         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
805                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
806         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
807                obd->obd_name, mount_count + 1);
808         CDEBUG(D_INODE, "%s: server data size: %u\n",
809                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
810         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
811                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
812         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
813                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
814         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
815                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
816         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
817                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
818                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
819                 le16_to_cpu(fsd->lsd_client_size));
820
821         if (!obd->obd_replayable) {
822                 CWARN("%s: recovery support OFF\n", obd->obd_name);
823                 GOTO(out, rc = 0);
824         }
825
826         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
827              off < last_rcvd_size; cl_idx++) {
828                 __u64 last_rcvd;
829                 struct obd_export *exp;
830                 struct filter_export_data *fed;
831
832                 if (!lcd) {
833                         OBD_ALLOC_PTR(lcd);
834                         if (!lcd)
835                                 GOTO(err_client, rc = -ENOMEM);
836                 }
837
838                 /* Don't assume off is incremented properly by
839                  * fsfilt_read_record(), in case sizeof(*lcd)
840                  * isn't the same as fsd->lsd_client_size.  */
841                 off = le32_to_cpu(fsd->lsd_client_start) +
842                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
843                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
844                 if (rc) {
845                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
846                                LAST_RCVD, cl_idx, off, rc);
847                         break; /* read error shouldn't cause startup to fail */
848                 }
849
850                 if (lcd->lcd_uuid[0] == '\0') {
851                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
852                                cl_idx);
853                         continue;
854                 }
855
856                 check_lcd(obd->obd_name, cl_idx, lcd);
857
858                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
859
860                 /* These exports are cleaned up by filter_disconnect(), so they
861                  * need to be set up like real exports as filter_connect() does.
862                  */
863                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
864
865                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
866                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
867                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
868                 if (IS_ERR(exp)) {
869                         if (PTR_ERR(exp) == -EALREADY) {
870                                 /* export already exists, zero out this one */
871                                 CERROR("Zeroing out duplicate export due to "
872                                        "bug 10479.\n");
873                                 lcd->lcd_uuid[0] = '\0';
874                         } else {
875                                 GOTO(err_client, rc = PTR_ERR(exp));
876                         }
877                 } else {
878                         fed = &exp->exp_filter_data;
879                         fed->fed_lcd = lcd;
880                         fed->fed_group = 0; /* will be assigned at connect */
881                         filter_export_stats_init(obd, exp, NULL);
882                         rc = filter_client_add(obd, exp, cl_idx);
883                         /* can't fail for existing client */
884                         LASSERTF(rc == 0, "rc = %d\n", rc);
885
886                         /* VBR: set export last committed */
887                         exp->exp_last_committed = last_rcvd;
888                         cfs_spin_lock(&exp->exp_lock);
889                         exp->exp_connecting = 0;
890                         exp->exp_in_recovery = 0;
891                         cfs_spin_unlock(&exp->exp_lock);
892                         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
893                         obd->obd_max_recoverable_clients++;
894                         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
895                         lcd = NULL;
896                         class_export_put(exp);
897                 }
898
899                 /* Need to check last_rcvd even for duplicated exports. */
900                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
901                        cl_idx, last_rcvd);
902
903                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
904                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
905         }
906
907         if (lcd)
908                 OBD_FREE_PTR(lcd);
909
910         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
911 out:
912         filter->fo_mount_count = mount_count + 1;
913         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
914
915         /* save it, so mount count and last_transno is current */
916         rc = filter_update_server_data(obd);
917         if (rc)
918                 GOTO(err_client, rc);
919
920         RETURN(0);
921
922 err_client:
923         class_disconnect_exports(obd);
924 err_fsd:
925         filter_free_server_data(filter);
926         RETURN(rc);
927 }
928
929 static int filter_cleanup_groups(struct obd_device *obd)
930 {
931         struct filter_obd *filter = &obd->u.filter;
932         struct file *filp;
933         struct dentry *dentry;
934         int i, j;
935         ENTRY;
936
937         if (filter->fo_dentry_O_groups != NULL) {
938                 for (i = 0; i < filter->fo_group_count; i++) {
939                         dentry = filter->fo_dentry_O_groups[i];
940                         if (dentry != NULL)
941                                 f_dput(dentry);
942                 }
943                 OBD_FREE(filter->fo_dentry_O_groups,
944                          filter->fo_group_count *
945                          sizeof(*filter->fo_dentry_O_groups));
946                 filter->fo_dentry_O_groups = NULL;
947         }
948         if (filter->fo_last_objid_files != NULL) {
949                 for (i = 0; i < filter->fo_group_count; i++) {
950                         filp = filter->fo_last_objid_files[i];
951                         if (filp != NULL)
952                                 filp_close(filp, 0);
953                 }
954                 OBD_FREE(filter->fo_last_objid_files,
955                          filter->fo_group_count *
956                          sizeof(*filter->fo_last_objid_files));
957                 filter->fo_last_objid_files = NULL;
958         }
959         if (filter->fo_dentry_O_sub != NULL) {
960                 for (i = 0; i < filter->fo_group_count; i++) {
961                         for (j = 0; j < filter->fo_subdir_count; j++) {
962                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
963                                 if (dentry != NULL)
964                                         f_dput(dentry);
965                         }
966                 }
967                 OBD_FREE(filter->fo_dentry_O_sub,
968                          filter->fo_group_count *
969                          sizeof(*filter->fo_dentry_O_sub));
970                 filter->fo_dentry_O_sub = NULL;
971         }
972         if (filter->fo_last_objids != NULL) {
973                 OBD_FREE(filter->fo_last_objids,
974                          filter->fo_group_count *
975                          sizeof(*filter->fo_last_objids));
976                 filter->fo_last_objids = NULL;
977         }
978         if (filter->fo_dentry_O != NULL) {
979                 f_dput(filter->fo_dentry_O);
980                 filter->fo_dentry_O = NULL;
981         }
982         RETURN(0);
983 }
984
985 static int filter_update_last_group(struct obd_device *obd, int group)
986 {
987         struct filter_obd *filter = &obd->u.filter;
988         struct file *filp = NULL;
989         int last_group = 0, rc;
990         loff_t off = 0;
991         ENTRY;
992
993         if (group <= filter->fo_committed_group)
994                 RETURN(0);
995
996         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
997         if (IS_ERR(filp)) {
998                 rc = PTR_ERR(filp);
999                 filp = NULL;
1000                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1001                 GOTO(cleanup, rc);
1002         }
1003
1004         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1005         if (rc) {
1006                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1007                 GOTO(cleanup, rc);
1008         }
1009
1010         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1011                obd->obd_name, last_group, group);
1012
1013         off = 0;
1014         last_group = group;
1015         /* must be sync: bXXXX */
1016         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1017         if (rc) {
1018                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1019                 GOTO(cleanup, rc);
1020         }
1021
1022         filter->fo_committed_group = group;
1023 cleanup:
1024         if (filp)
1025                 filp_close(filp, 0);
1026         RETURN(rc);
1027 }
1028
1029 static int filter_read_group_internal(struct obd_device *obd, int group,
1030                                       int create)
1031 {
1032         struct filter_obd *filter = &obd->u.filter;
1033         __u64 *new_objids = NULL;
1034         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1035         struct dentry **new_groups = NULL;
1036         struct file **new_files = NULL;
1037         struct dentry *dentry;
1038         struct file *filp;
1039         int old_count = filter->fo_group_count, rc, stage = 0, i;
1040         char name[25];
1041         __u64 last_objid;
1042         loff_t off = 0;
1043         int len = group + 1;
1044
1045         snprintf(name, 24, "%d", group);
1046         name[24] = '\0';
1047
1048         if (!create) {
1049                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1050                                            strlen(name));
1051                 if (IS_ERR(dentry)) {
1052                         CERROR("Cannot lookup expected object group %d: %ld\n",
1053                                group, PTR_ERR(dentry));
1054                         RETURN(PTR_ERR(dentry));
1055                 }
1056         } else {
1057                 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1058                                       name, 0700, 1);
1059                 if (IS_ERR(dentry)) {
1060                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1061                                PTR_ERR(dentry));
1062                         RETURN(PTR_ERR(dentry));
1063                 }
1064         }
1065         stage = 1;
1066
1067         snprintf(name, 24, "O/%d/LAST_ID", group);
1068         name[24] = '\0';
1069         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1070         if (IS_ERR(filp)) {
1071                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1072                 GOTO(cleanup, rc = PTR_ERR(filp));
1073         }
1074         stage = 2;
1075
1076         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1077         if (rc) {
1078                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1079                 GOTO(cleanup, rc);
1080         }
1081
1082         if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1083                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1084                 if (tmp_subdirs == NULL)
1085                         GOTO(cleanup, rc = -ENOMEM);
1086                 stage = 3;
1087
1088                 for (i = 0; i < filter->fo_subdir_count; i++) {
1089                         char dir[20];
1090                         snprintf(dir, sizeof(dir), "d%u", i);
1091
1092                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1093                                                               filter->fo_vfsmnt,
1094                                                               dir, 0700, 1);
1095                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1096                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1097                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1098                                        group, dir, rc);
1099                                 GOTO(cleanup, rc);
1100                         }
1101
1102                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1103                                tmp_subdirs->dentry[i]);
1104                 }
1105         }
1106
1107         /* 'group' is an index; we need an array of length 'group + 1' */
1108         if (group + 1 > old_count) {
1109                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1110                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1111                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1112                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1113                 stage = 4;
1114                 if (new_objids == NULL || new_subdirs == NULL ||
1115                     new_groups == NULL || new_files == NULL)
1116                         GOTO(cleanup, rc = -ENOMEM);
1117
1118                 if (old_count) {
1119                         memcpy(new_objids, filter->fo_last_objids,
1120                                old_count * sizeof(*new_objids));
1121                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1122                                old_count * sizeof(*new_subdirs));
1123                         memcpy(new_groups, filter->fo_dentry_O_groups,
1124                                old_count * sizeof(*new_groups));
1125                         memcpy(new_files, filter->fo_last_objid_files,
1126                                old_count * sizeof(*new_files));
1127
1128                         OBD_FREE(filter->fo_last_objids,
1129                                  old_count * sizeof(*new_objids));
1130                         OBD_FREE(filter->fo_dentry_O_sub,
1131                                  old_count * sizeof(*new_subdirs));
1132                         OBD_FREE(filter->fo_dentry_O_groups,
1133                                  old_count * sizeof(*new_groups));
1134                         OBD_FREE(filter->fo_last_objid_files,
1135                                  old_count * sizeof(*new_files));
1136                 }
1137                 filter->fo_last_objids = new_objids;
1138                 filter->fo_dentry_O_sub = new_subdirs;
1139                 filter->fo_dentry_O_groups = new_groups;
1140                 filter->fo_last_objid_files = new_files;
1141                 filter->fo_group_count = len;
1142         }
1143
1144         filter->fo_dentry_O_groups[group] = dentry;
1145         filter->fo_last_objid_files[group] = filp;
1146         if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1147                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1148                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1149         }
1150
1151         filter_update_last_group(obd, group);
1152
1153         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1154                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1155                 rc = filter_update_last_objid(obd, group, 1);
1156                 RETURN(rc);
1157         }
1158
1159         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1160         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1161                obd->obd_name, group, last_objid);
1162         RETURN(0);
1163  cleanup:
1164         switch (stage) {
1165         case 4:
1166                 if (new_objids != NULL)
1167                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1168                 if (new_subdirs != NULL)
1169                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1170                 if (new_groups != NULL)
1171                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1172                 if (new_files != NULL)
1173                         OBD_FREE(new_files, len * sizeof(*new_files));
1174         case 3:
1175                 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1176                         for (i = 0; i < filter->fo_subdir_count; i++) {
1177                                 if (tmp_subdirs->dentry[i] != NULL)
1178                                         dput(tmp_subdirs->dentry[i]);
1179                         }
1180                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1181                 }
1182         case 2:
1183                 filp_close(filp, 0);
1184         case 1:
1185                 dput(dentry);
1186         }
1187         RETURN(rc);
1188 }
1189
1190 static int filter_read_groups(struct obd_device *obd, int last_group,
1191                               int create)
1192 {
1193         struct filter_obd *filter = &obd->u.filter;
1194         int old_count, group, rc = 0;
1195
1196         cfs_down(&filter->fo_init_lock);
1197         old_count = filter->fo_group_count;
1198         for (group = old_count; group <= last_group; group++) {
1199                 rc = filter_read_group_internal(obd, group, create);
1200                 if (rc != 0)
1201                         break;
1202         }
1203         cfs_up(&filter->fo_init_lock);
1204         return rc;
1205 }
1206
1207 /* FIXME: object groups */
1208 static int filter_prep_groups(struct obd_device *obd)
1209 {
1210         struct filter_obd *filter = &obd->u.filter;
1211         struct dentry *O_dentry;
1212         struct file *filp;
1213         int    last_group, rc = 0, cleanup_phase = 0;
1214         loff_t off = 0;
1215         ENTRY;
1216
1217         O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1218                                 "O", 0700, 1);
1219         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1220         if (IS_ERR(O_dentry)) {
1221                 rc = PTR_ERR(O_dentry);
1222                 CERROR("cannot open/create O: rc = %d\n", rc);
1223                 GOTO(cleanup, rc);
1224         }
1225         filter->fo_dentry_O = O_dentry;
1226         cleanup_phase = 1; /* O_dentry */
1227
1228         /* we have to initialize all groups before first connections from
1229          * clients because they may send create/destroy for any group -bzzz */
1230         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1231         if (IS_ERR(filp)) {
1232                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1233                 GOTO(cleanup, rc = PTR_ERR(filp));
1234         }
1235         cleanup_phase = 2; /* filp */
1236
1237         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1238         if (rc) {
1239                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1240                 GOTO(cleanup, rc);
1241         }
1242
1243         if (off == 0)
1244                 last_group = FILTER_GROUP_MDS0;
1245
1246         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1247               FILTER_GROUP_MDS0, last_group);
1248         filter->fo_committed_group = last_group;
1249         rc = filter_read_groups(obd, last_group, 1);
1250         if (rc)
1251                 GOTO(cleanup, rc);
1252
1253         filp_close(filp, 0);
1254         RETURN(0);
1255
1256  cleanup:
1257         switch (cleanup_phase) {
1258         case 2:
1259                 filp_close(filp, 0);
1260         case 1:
1261                 filter_cleanup_groups(obd);
1262                 f_dput(filter->fo_dentry_O);
1263                 filter->fo_dentry_O = NULL;
1264         default:
1265                 break;
1266         }
1267         return rc;
1268
1269 }
1270
1271 /* setup the object store with correct subdirectories */
1272 static int filter_prep(struct obd_device *obd)
1273 {
1274         struct lvfs_run_ctxt saved;
1275         struct filter_obd *filter = &obd->u.filter;
1276         struct file *file;
1277         struct inode *inode;
1278         int rc = 0;
1279         ENTRY;
1280
1281         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1282         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1283         if (!file || IS_ERR(file)) {
1284                 rc = PTR_ERR(file);
1285                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1286                        LAST_RCVD, rc);
1287                 GOTO(out, rc);
1288         }
1289         filter->fo_rcvd_filp = file;
1290         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1291                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1292                        file->f_dentry->d_inode->i_mode);
1293                 GOTO(err_filp, rc = -ENOENT);
1294         }
1295
1296         inode = file->f_dentry->d_parent->d_inode;
1297         /* We use i_op->unlink directly in filter_vfs_unlink() */
1298         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1299                 CERROR("%s: filesystem does not support create/unlink ops\n",
1300                        obd->obd_name);
1301                 GOTO(err_filp, rc = -EOPNOTSUPP);
1302         }
1303
1304         rc = filter_init_server_data(obd, file);
1305         if (rc) {
1306                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1307                 GOTO(err_filp, rc);
1308         }
1309         LASSERT(obd->u.obt.obt_lut);
1310         target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1311
1312         /* open and create health check io file*/
1313         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1314         if (IS_ERR(file)) {
1315                 rc = PTR_ERR(file);
1316                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1317                        HEALTH_CHECK, rc);
1318                 GOTO(err_server_data, rc);
1319         }
1320         filter->fo_obt.obt_health_check_filp = file;
1321         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1322                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1323                        file->f_dentry->d_inode->i_mode);
1324                 GOTO(err_health_check, rc = -ENOENT);
1325         }
1326         rc = lvfs_check_io_health(obd, file);
1327         if (rc)
1328                 GOTO(err_health_check, rc);
1329
1330         rc = filter_prep_groups(obd);
1331         if (rc)
1332                 GOTO(err_health_check, rc);
1333 out:
1334         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1335
1336         return(rc);
1337
1338 err_health_check:
1339         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1340                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1341         filter->fo_obt.obt_health_check_filp = NULL;
1342 err_server_data:
1343         target_recovery_fini(obd);
1344         filter_free_server_data(filter);
1345 err_filp:
1346         if (filp_close(filter->fo_rcvd_filp, 0))
1347                 CERROR("can't close %s after error\n", LAST_RCVD);
1348         filter->fo_rcvd_filp = NULL;
1349         goto out;
1350 }
1351
1352 /* cleanup the filter: write last used object id to status file */
1353 static void filter_post(struct obd_device *obd)
1354 {
1355         struct lvfs_run_ctxt saved;
1356         struct filter_obd *filter = &obd->u.filter;
1357         int rc, i;
1358
1359         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1360          * best to start a transaction with h_sync, because we removed this
1361          * from lastobjid */
1362
1363         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1364         rc = filter_update_server_data(obd);
1365         if (rc)
1366                 CERROR("error writing server data: rc = %d\n", rc);
1367
1368         for (i = 0; i < filter->fo_group_count; i++) {
1369                 rc = filter_update_last_objid(obd, i,
1370                                 (i == filter->fo_group_count - 1));
1371                 if (rc)
1372                         CERROR("error writing group %d lastobjid: rc = %d\n",
1373                                i, rc);
1374         }
1375
1376         rc = filp_close(filter->fo_rcvd_filp, 0);
1377         filter->fo_rcvd_filp = NULL;
1378         if (rc)
1379                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1380
1381         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1382         filter->fo_obt.obt_health_check_filp = NULL;
1383         if (rc)
1384                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1385
1386         filter_cleanup_groups(obd);
1387         filter_free_server_data(filter);
1388         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1389
1390         filter_free_capa_keys(filter);
1391         cleanup_capa_hash(filter->fo_capa_hash);
1392 }
1393
1394 static void filter_set_last_id(struct filter_obd *filter,
1395                                obd_id id, obd_gr group)
1396 {
1397         LASSERT(group <= filter->fo_group_count);
1398
1399         cfs_spin_lock(&filter->fo_objidlock);
1400         filter->fo_last_objids[group] = id;
1401         cfs_spin_unlock(&filter->fo_objidlock);
1402 }
1403
1404 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1405 {
1406         obd_id id;
1407         LASSERT(group <= filter->fo_group_count);
1408         LASSERT(filter->fo_last_objids != NULL);
1409
1410         /* FIXME: object groups */
1411         cfs_spin_lock(&filter->fo_objidlock);
1412         id = filter->fo_last_objids[group];
1413         cfs_spin_unlock(&filter->fo_objidlock);
1414         return id;
1415 }
1416
1417 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1418 {
1419         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1420         return 0;
1421 }
1422
1423 /* We never dget the object parent, so DON'T dput it either */
1424 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1425 {
1426         struct filter_obd *filter = &obd->u.filter;
1427         struct filter_subdirs *subdirs;
1428         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1429
1430         if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1431                 return filter->fo_dentry_O_groups[group];
1432
1433         subdirs = &filter->fo_dentry_O_sub[group];
1434         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1435 }
1436
1437 /* We never dget the object parent, so DON'T dput it either */
1438 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1439                                   obd_id objid)
1440 {
1441         unsigned long now = jiffies;
1442         struct dentry *dparent = filter_parent(obd, group, objid);
1443         int rc;
1444
1445         if (IS_ERR(dparent))
1446                 return dparent;
1447         if (dparent == NULL)
1448                 return ERR_PTR(-ENOENT);
1449
1450         rc = filter_lock_dentry(obd, dparent);
1451         fsfilt_check_slow(obd, now, "parent lock");
1452         return rc ? ERR_PTR(rc) : dparent;
1453 }
1454
1455 /* We never dget the object parent, so DON'T dput it either */
1456 static void filter_parent_unlock(struct dentry *dparent)
1457 {
1458         UNLOCK_INODE_MUTEX(dparent->d_inode);
1459 }
1460
1461 /* How to get files, dentries, inodes from object id's.
1462  *
1463  * If dir_dentry is passed, the caller has already locked the parent
1464  * appropriately for this operation (normally a write lock).  If
1465  * dir_dentry is NULL, we do a read lock while we do the lookup to
1466  * avoid races with create/destroy and such changing the directory
1467  * internal to the filesystem code. */
1468 struct dentry *filter_fid2dentry(struct obd_device *obd,
1469                                  struct dentry *dir_dentry,
1470                                  obd_gr group, obd_id id)
1471 {
1472         struct dentry *dparent = dir_dentry;
1473         struct dentry *dchild;
1474         char name[32];
1475         int len;
1476         ENTRY;
1477
1478         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1479             obd->u.filter.fo_destroys_in_progress == 0) {
1480                 /* don't fail lookups for orphan recovery, it causes
1481                  * later LBUGs when objects still exist during precreate */
1482                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1483                 RETURN(ERR_PTR(-ENOENT));
1484         }
1485         if (id == 0) {
1486                 CERROR("fatal: invalid object id 0\n");
1487                 RETURN(ERR_PTR(-ESTALE));
1488         }
1489
1490         len = sprintf(name, LPU64, id);
1491         if (dir_dentry == NULL) {
1492                 dparent = filter_parent_lock(obd, group, id);
1493                 if (IS_ERR(dparent)) {
1494                         CERROR("%s: error getting object "LPU64":"LPU64
1495                                " parent: rc %ld\n", obd->obd_name,
1496                                id, group, PTR_ERR(dparent));
1497                         RETURN(dparent);
1498                 }
1499         }
1500         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1501                dparent->d_name.len, dparent->d_name.name, name);
1502         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1503         if (dir_dentry == NULL)
1504                 filter_parent_unlock(dparent);
1505         if (IS_ERR(dchild)) {
1506                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1507                        PTR_ERR(dchild));
1508                 RETURN(dchild);
1509         }
1510
1511         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1512                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1513                        obd->obd_name, id, dchild->d_inode->i_ino);
1514                 f_dput(dchild);
1515                 RETURN(ERR_PTR(-ENOENT));
1516         }
1517
1518         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1519                name, dchild, atomic_read(&dchild->d_count));
1520
1521         LASSERT(atomic_read(&dchild->d_count) > 0);
1522
1523         RETURN(dchild);
1524 }
1525
1526 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1527                                   obd_id group, struct lustre_handle *lockh)
1528 {
1529         int flags = LDLM_AST_DISCARD_DATA, rc;
1530         struct ldlm_res_id res_id;
1531         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1532         ENTRY;
1533
1534         osc_build_res_name(objid, group, &res_id);
1535         /* Tell the clients that the object is gone now and that they should
1536          * throw away any cached pages. */
1537         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1538                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1539                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1540                                     lockh);
1541         if (rc != ELDLM_OK)
1542                 lockh->cookie = 0;
1543         RETURN(rc);
1544 }
1545
1546 static void filter_fini_destroy(struct obd_device *obd,
1547                                 struct lustre_handle *lockh)
1548 {
1549         if (lustre_handle_is_used(lockh))
1550                 ldlm_lock_decref(lockh, LCK_PW);
1551 }
1552
1553 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1554  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1555  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1556  * reverse.  Caller must take i_sem before starting the transaction and we
1557  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1558 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1559                       struct vfsmount *mnt)
1560 {
1561         int rc;
1562         ENTRY;
1563
1564         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1565          * itself we already hold dir->i_mutex for child create/unlink ops */
1566         LASSERT(dentry->d_inode != NULL);
1567         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1568         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1569
1570
1571         /* may_delete() */
1572         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1573                 GOTO(out, rc = -ENOENT);
1574
1575         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1576         if (rc)
1577                 GOTO(out, rc);
1578
1579         if (IS_APPEND(dir))
1580                 GOTO(out, rc = -EPERM);
1581
1582         /* check_sticky() */
1583         if ((dentry->d_inode->i_uid != current->fsuid &&
1584              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1585             IS_IMMUTABLE(dentry->d_inode))
1586                 GOTO(out, rc = -EPERM);
1587
1588         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1589          *       that was done because of journal_start (which is already done
1590          *       here) or some other ordering issue. */
1591         DQUOT_INIT(dir);
1592
1593         rc = ll_security_inode_unlink(dir, dentry, mnt);
1594         if (rc)
1595                 GOTO(out, rc);
1596
1597         rc = dir->i_op->unlink(dir, dentry);
1598 out:
1599         /* need to drop i_mutex before we lose inode reference */
1600         UNLOCK_INODE_MUTEX(dentry->d_inode);
1601         if (rc == 0)
1602                 d_delete(dentry);
1603
1604         RETURN(rc);
1605 }
1606
1607 /* Caller must hold LCK_PW on parent and push us into kernel context.
1608  * Caller must hold child i_mutex, we drop it always.
1609  * Caller is also required to ensure that dchild->d_inode exists. */
1610 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1611                                    obd_gr group, struct dentry *dparent,
1612                                    struct dentry *dchild)
1613 {
1614         struct inode *inode = dchild->d_inode;
1615         struct filter_obd *filter = &obd->u.filter;
1616         int rc;
1617
1618         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1619                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1620                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1621                        (unsigned long)inode->i_nlink,
1622                        atomic_read(&inode->i_count));
1623         }
1624
1625         rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1626         if (rc)
1627                 CERROR("error unlinking objid %.*s: rc %d\n",
1628                        dchild->d_name.len, dchild->d_name.name, rc);
1629         return(rc);
1630 }
1631
1632 struct filter_intent_args {
1633         struct ldlm_lock **victim;
1634         __u64 size;
1635         int *liblustre;
1636 };
1637
1638 static enum interval_iter filter_intent_cb(struct interval_node *n,
1639                                            void *args)
1640 {
1641         struct ldlm_interval *node = (struct ldlm_interval *)n;
1642         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1643         __u64 size = arg->size;
1644         struct ldlm_lock **v = arg->victim;
1645         struct ldlm_lock *lck;
1646
1647         /* If the interval is lower than the current file size,
1648          * just break. */
1649         if (interval_high(n) <= size)
1650                 return INTERVAL_ITER_STOP;
1651
1652         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1653                 /* Don't send glimpse ASTs to liblustre clients.
1654                  * They aren't listening for them, and they do
1655                  * entirely synchronous I/O anyways. */
1656                 if (lck->l_export == NULL ||
1657                     lck->l_export->exp_libclient == 1)
1658                         continue;
1659
1660                 if (*arg->liblustre)
1661                         *arg->liblustre = 0;
1662
1663                 if (*v == NULL) {
1664                         *v = LDLM_LOCK_GET(lck);
1665                 } else if ((*v)->l_policy_data.l_extent.start <
1666                            lck->l_policy_data.l_extent.start) {
1667                         LDLM_LOCK_RELEASE(*v);
1668                         *v = LDLM_LOCK_GET(lck);
1669                 }
1670
1671                 /* the same policy group - every lock has the
1672                  * same extent, so needn't do it any more */
1673                 break;
1674         }
1675
1676         return INTERVAL_ITER_CONT;
1677 }
1678
1679 static int filter_intent_policy(struct ldlm_namespace *ns,
1680                                 struct ldlm_lock **lockp, void *req_cookie,
1681                                 ldlm_mode_t mode, int flags, void *data)
1682 {
1683         CFS_LIST_HEAD(rpc_list);
1684         struct ptlrpc_request *req = req_cookie;
1685         struct ldlm_lock *lock = *lockp, *l = NULL;
1686         struct ldlm_resource *res = lock->l_resource;
1687         ldlm_processing_policy policy;
1688         struct ost_lvb *res_lvb, *reply_lvb;
1689         struct ldlm_reply *rep;
1690         ldlm_error_t err;
1691         int idx, rc, tmpflags = 0, only_liblustre = 1;
1692         struct ldlm_interval_tree *tree;
1693         struct filter_intent_args arg;
1694         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1695                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1696                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1697         ENTRY;
1698
1699         policy = ldlm_get_processing_policy(res);
1700         LASSERT(policy != NULL);
1701         LASSERT(req != NULL);
1702
1703         rc = lustre_pack_reply(req, 3, repsize, NULL);
1704         if (rc)
1705                 RETURN(req->rq_status = rc);
1706
1707         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1708         LASSERT(rep != NULL);
1709
1710         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1711                                    sizeof(*reply_lvb));
1712         LASSERT(reply_lvb != NULL);
1713
1714         //fixup_handle_for_resent_req(req, lock, &lockh);
1715
1716         /* Call the extent policy function to see if our request can be
1717          * granted, or is blocked.
1718          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1719          * lock, and should not be granted if the lock will be blocked.
1720          */
1721
1722         LASSERT(ns == res->lr_namespace);
1723         lock_res(res);
1724         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1725         check_res_locked(res);
1726
1727         /* FIXME: we should change the policy function slightly, to not make
1728          * this list at all, since we just turn around and free it */
1729         while (!cfs_list_empty(&rpc_list)) {
1730                 struct ldlm_lock *wlock =
1731                         cfs_list_entry(rpc_list.next, struct ldlm_lock,
1732                                        l_cp_ast);
1733                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1734                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1735                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1736                 cfs_list_del_init(&wlock->l_cp_ast);
1737                 LDLM_LOCK_RELEASE(wlock);
1738         }
1739
1740         /* The lock met with no resistance; we're finished. */
1741         if (rc == LDLM_ITER_CONTINUE) {
1742                 /* do not grant locks to the liblustre clients: they cannot
1743                  * handle ASTs robustly.  We need to do this while still
1744                  * holding ns_lock to avoid the lock remaining on the res_link
1745                  * list (and potentially being added to l_pending_list by an
1746                  * AST) when we are going to drop this lock ASAP. */
1747                 if (lock->l_export->exp_libclient ||
1748                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1749                         ldlm_resource_unlink_lock(lock);
1750                         err = ELDLM_LOCK_ABORTED;
1751                 } else {
1752                         err = ELDLM_LOCK_REPLACED;
1753                 }
1754                 unlock_res(res);
1755                 RETURN(err);
1756         }
1757
1758         /* Do not grant any lock, but instead send GL callbacks.  The extent
1759          * policy nicely created a list of all PW locks for us.  We will choose
1760          * the highest of those which are larger than the size in the LVB, if
1761          * any, and perform a glimpse callback. */
1762         res_lvb = res->lr_lvb_data;
1763         LASSERT(res_lvb != NULL);
1764         *reply_lvb = *res_lvb;
1765
1766         /*
1767          * ->ns_lock guarantees that no new locks are granted, and,
1768          * therefore, that res->lr_lvb_data cannot increase beyond the
1769          * end of already granted lock. As a result, it is safe to
1770          * check against "stale" reply_lvb->lvb_size value without
1771          * res->lr_lvb_sem.
1772          */
1773         arg.size = reply_lvb->lvb_size;
1774         arg.victim = &l;
1775         arg.liblustre = &only_liblustre;
1776         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1777                 tree = &res->lr_itree[idx];
1778                 if (tree->lit_mode == LCK_PR)
1779                         continue;
1780
1781                 interval_iterate_reverse(tree->lit_root,
1782                                          filter_intent_cb, &arg);
1783         }
1784         unlock_res(res);
1785
1786         /* There were no PW locks beyond the size in the LVB; finished. */
1787         if (l == NULL) {
1788                 if (only_liblustre) {
1789                         /* If we discovered a liblustre client with a PW lock,
1790                          * however, the LVB may be out of date!  The LVB is
1791                          * updated only on glimpse (which we don't do for
1792                          * liblustre clients) and cancel (which the client
1793                          * obviously has not yet done).  So if it has written
1794                          * data but kept the lock, the LVB is stale and needs
1795                          * to be updated from disk.
1796                          *
1797                          * Of course, this will all disappear when we switch to
1798                          * taking liblustre locks on the OST. */
1799                         ldlm_res_lvbo_update(res, NULL, 1);
1800                 }
1801                 RETURN(ELDLM_LOCK_ABORTED);
1802         }
1803
1804         /*
1805          * This check is for lock taken in filter_prepare_destroy() that does
1806          * not have l_glimpse_ast set. So the logic is: if there is a lock
1807          * with no l_glimpse_ast set, this object is being destroyed already.
1808          *
1809          * Hence, if you are grabbing DLM locks on the server, always set
1810          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1811          */
1812         if (l->l_glimpse_ast == NULL) {
1813                 /* We are racing with unlink(); just return -ENOENT */
1814                 rep->lock_policy_res1 = -ENOENT;
1815                 goto out;
1816         }
1817
1818         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1819         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1820         /* Update the LVB from disk if the AST failed (this is a legal race) */
1821         /*
1822          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1823          * sending ast is not handled. This can result in lost client writes.
1824          */
1825         if (rc != 0)
1826                 ldlm_res_lvbo_update(res, NULL, 1);
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 lustre_mount_info *lmi;
1971         struct obd_uuid uuid;
1972         __u8 *uuid_ptr;
1973         char *str, *label;
1974         char ns_name[48];
1975         request_queue_t *q;
1976         int rc, i;
1977         ENTRY;
1978
1979         if (lcfg->lcfg_bufcount < 3 ||
1980             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1981             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1982                 RETURN(-EINVAL);
1983
1984         lmi = server_get_mount(obd->obd_name);
1985         if (lmi) {
1986                 /* We already mounted in lustre_fill_super.
1987                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1988                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1989                 mnt = lmi->lmi_mnt;
1990                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1991         } else {
1992                 /* old path - used by lctl */
1993                 CERROR("Using old MDS mount method\n");
1994                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1995                                     MS_NOATIME|MS_NODIRATIME,
1996                                     lustre_cfg_string(lcfg, 1), option);
1997                 if (IS_ERR(mnt)) {
1998                         rc = PTR_ERR(mnt);
1999                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2000                                            lustre_cfg_string(lcfg, 1), rc);
2001                         RETURN(rc);
2002                 }
2003
2004                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2005         }
2006         if (IS_ERR(obd->obd_fsops))
2007                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2008
2009         rc = filter_iobuf_pool_init(filter);
2010         if (rc != 0)
2011                 GOTO(err_ops, rc);
2012
2013         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2014                 CERROR("%s: Underlying device is marked as read-only. "
2015                        "Setup failed\n", obd->obd_name);
2016                 GOTO(err_ops, rc = -EROFS);
2017         }
2018
2019         /* failover is the default */
2020         obd->obd_replayable = 1;
2021
2022         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2023                 str = lustre_cfg_string(lcfg, 3);
2024                 if (strchr(str, 'n')) {
2025                         CWARN("%s: recovery disabled\n", obd->obd_name);
2026                         obd->obd_replayable = 0;
2027                 }
2028         }
2029
2030         filter->fo_vfsmnt = mnt;
2031         obd->u.obt.obt_sb = mnt->mnt_sb;
2032         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2033         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2034
2035         fsfilt_setup(obd, obd->u.obt.obt_sb);
2036
2037         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2038         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2039         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2040         obd->obd_lvfs_ctxt.fs = get_ds();
2041         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2042
2043         cfs_init_mutex(&filter->fo_init_lock);
2044         filter->fo_committed_group = 0;
2045         filter->fo_destroys_in_progress = 0;
2046         for (i = 0; i < 32; i++)
2047                 cfs_sema_init(&filter->fo_create_locks[i], 1);
2048
2049         cfs_spin_lock_init(&filter->fo_objidlock);
2050         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2051         cfs_sema_init(&filter->fo_alloc_lock, 1);
2052         init_brw_stats(&filter->fo_filter_stats);
2053         filter->fo_read_cache = 1; /* enable read-only cache by default */
2054         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2055         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2056         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2057         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2058
2059         rc = filter_prep(obd);
2060         if (rc)
2061                 GOTO(err_ops, rc);
2062
2063         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2064         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2065
2066         filter->fo_fl_oss_capa = 1;
2067
2068         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2069         filter->fo_capa_hash = init_capa_hash();
2070         if (filter->fo_capa_hash == NULL)
2071                 GOTO(err_post, rc = -ENOMEM);
2072
2073         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2074         obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2075                                                 LDLM_NAMESPACE_GREEDY);
2076         if (obd->obd_namespace == NULL)
2077                 GOTO(err_post, rc = -ENOMEM);
2078         obd->obd_namespace->ns_lvbp = obd;
2079         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2080         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2081
2082         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2083                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2084
2085         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2086         if (rc) {
2087                 CERROR("failed to setup llogging subsystems\n");
2088                 GOTO(err_post, rc);
2089         }
2090
2091         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2092         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2093         /* do this after llog being initialized */
2094         filter_adapt_sptlrpc_conf(obd, 1);
2095
2096         rc = lquota_setup(filter_quota_interface_ref, obd);
2097         if (rc)
2098                 GOTO(err_post, rc);
2099
2100         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2101         if (q->max_sectors < q->max_hw_sectors &&
2102             q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2103                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2104                               "for larger I/O requests: max_sectors = %u "
2105                               "could be up to max_hw_sectors=%u\n",
2106                               obd->obd_name, mnt->mnt_sb->s_id,
2107                               q->max_sectors, q->max_hw_sectors);
2108
2109         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2110         if (uuid_ptr != NULL) {
2111                 class_uuid_unparse(uuid_ptr, &uuid);
2112                 str = uuid.uuid;
2113         } else {
2114                 str = "no UUID";
2115         }
2116
2117         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2118         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2119                       obd->obd_name, label ?: str, lmi ? "on " : "",
2120                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2121                       obd->obd_replayable ? "enabled" : "disabled");
2122
2123         if (obd->obd_recovering)
2124                 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2125                               "or until %d client%s reconnect%s\n",
2126                               obd->obd_name,
2127                               obd->obd_recovery_timeout / 60,
2128                               obd->obd_recovery_timeout % 60,
2129                               obd->obd_max_recoverable_clients,
2130                               (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2131                               (obd->obd_max_recoverable_clients == 1) ? "s": "");
2132
2133
2134         RETURN(0);
2135
2136 err_post:
2137         filter_post(obd);
2138 err_ops:
2139         fsfilt_put_ops(obd->obd_fsops);
2140         filter_iobuf_pool_done(filter);
2141 err_mntput:
2142         server_put_mount(obd->obd_name, mnt);
2143         obd->u.obt.obt_sb = 0;
2144         return rc;
2145 }
2146
2147 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2148 {
2149         struct lprocfs_static_vars lvars;
2150         unsigned long addr;
2151         struct page *page;
2152         int rc;
2153
2154         CLASSERT(offsetof(struct obd_device, u.obt) ==
2155                  offsetof(struct obd_device, u.filter.fo_obt));
2156
2157         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2158                 RETURN(-EINVAL);
2159
2160         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2161         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2162         if (!page)
2163                 RETURN(-ENOMEM);
2164         addr = (unsigned long)cfs_page_address(page);
2165         clear_page((void *)addr);
2166
2167         /* lprocfs must be setup before the filter so state can be safely added
2168          * to /proc incrementally as the filter is setup */
2169         lprocfs_filter_init_vars(&lvars);
2170         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2171             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2172                 /* Init obdfilter private stats here */
2173                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2174                                      LPROCFS_CNTR_AVGMINMAX,
2175                                      "read_bytes", "bytes");
2176                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2177                                      LPROCFS_CNTR_AVGMINMAX,
2178                                      "write_bytes", "bytes");
2179                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2180                                      LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2181                                      "get_page", "usec");
2182                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2183                                      LPROCFS_CNTR_AVGMINMAX,
2184                                      "get_page_failures", "num");
2185                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2186                                      LPROCFS_CNTR_AVGMINMAX,
2187                                      "cache_access", "pages");
2188                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2189                                      LPROCFS_CNTR_AVGMINMAX,
2190                                      "cache_hit", "pages");
2191                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2192                                      LPROCFS_CNTR_AVGMINMAX,
2193                                      "cache_miss", "pages");
2194
2195                 lproc_filter_attach_seqstat(obd);
2196                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2197                                                         obd->obd_proc_entry,
2198                                                         NULL, NULL);
2199                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2200                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2201                         CERROR("error %d setting up lprocfs for %s\n",
2202                                rc, "exports");
2203                         obd->obd_proc_exports_entry = NULL;
2204                 }
2205         }
2206         if (obd->obd_proc_exports_entry)
2207                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2208                                    lprocfs_nid_stats_clear_read,
2209                                    lprocfs_nid_stats_clear_write, obd, NULL);
2210
2211         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2212                LUSTRE_CFG_BUFLEN(lcfg, 4));
2213         rc = filter_common_setup(obd, lcfg, (void *)addr);
2214         OBD_PAGE_FREE(page);
2215
2216         if (rc) {
2217                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2218                 lprocfs_free_per_client_stats(obd);
2219                 lprocfs_free_obd_stats(obd);
2220                 lprocfs_obd_cleanup(obd);
2221         }
2222
2223         return rc;
2224 }
2225
2226 static struct llog_operations filter_mds_ost_repl_logops;
2227
2228 static struct llog_operations filter_size_orig_logops = {
2229         .lop_setup   = llog_obd_origin_setup,
2230         .lop_cleanup = llog_obd_origin_cleanup,
2231         .lop_add     = llog_obd_origin_add
2232 };
2233
2234 static int filter_olg_fini(struct obd_llog_group *olg)
2235 {
2236         struct llog_ctxt *ctxt;
2237         int rc = 0, rc2 = 0;
2238         ENTRY;
2239
2240         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2241         if (ctxt)
2242                 rc = llog_cleanup(ctxt);
2243
2244         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2245         if (ctxt) {
2246                 rc2 = llog_cleanup(ctxt);
2247                 if (!rc)
2248                         rc = rc2;
2249         }
2250
2251         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2252         if (ctxt) {
2253                 rc2 = llog_cleanup(ctxt);
2254                 if (!rc)
2255                         rc = rc2;
2256         }
2257
2258         RETURN(rc);
2259 }
2260
2261 static int
2262 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2263                 struct obd_device *tgt)
2264 {
2265         int rc;
2266         ENTRY;
2267
2268         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2269                         &filter_mds_ost_repl_logops);
2270         if (rc)
2271                 GOTO(cleanup, rc);
2272
2273         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2274                         &filter_size_orig_logops);
2275         if (rc)
2276                 GOTO(cleanup, rc);
2277         EXIT;
2278 cleanup:
2279         if (rc)
2280                 filter_olg_fini(olg);
2281         return rc;
2282 }
2283
2284 /**
2285  * Init the default olg, which is embeded in the obd_device, for filter.
2286  */
2287 static int
2288 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2289                         struct obd_device *tgt)
2290 {
2291         struct filter_obd *filter = &obd->u.filter;
2292         struct llog_ctxt *ctxt;
2293         int rc;
2294         ENTRY;
2295
2296         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2297         if (!filter->fo_lcm)
2298                 RETURN(-ENOMEM);
2299
2300         filter_mds_ost_repl_logops = llog_client_ops;
2301         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2302         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2303         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2304
2305         rc = filter_olg_init(obd, olg, tgt);
2306         if (rc)
2307                 GOTO(cleanup_lcm, rc);
2308
2309         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2310                         &llog_lvfs_ops);
2311         if (rc)
2312                 GOTO(cleanup_olg, rc);
2313
2314         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2315         if (!ctxt) {
2316                 CERROR("Can't get ctxt for %p:%x\n", olg,
2317                        LLOG_MDS_OST_REPL_CTXT);
2318                 GOTO(cleanup_olg, rc = -ENODEV);
2319         }
2320         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2321         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2322         llog_ctxt_put(ctxt);
2323
2324         RETURN(0);
2325 cleanup_olg:
2326         filter_olg_fini(olg);
2327 cleanup_lcm:
2328         llog_recov_thread_fini(filter->fo_lcm, 1);
2329         filter->fo_lcm = NULL;
2330         return rc;
2331 }
2332
2333 static int
2334 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2335                  struct obd_device *tgt, int *index)
2336 {
2337         struct filter_obd *filter = &obd->u.filter;
2338         struct llog_ctxt *ctxt;
2339         int rc;
2340         ENTRY;
2341
2342         LASSERT(olg != NULL);
2343         if (olg == &obd->obd_olg)
2344                 return filter_default_olg_init(obd, olg, tgt);
2345
2346         LASSERT(filter->fo_lcm != NULL);
2347         rc = filter_olg_init(obd, olg, tgt);
2348         if (rc)
2349                 RETURN(rc);
2350         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2351         if (!ctxt) {
2352                 CERROR("Can't get ctxt for %p:%x\n", olg,
2353                        LLOG_MDS_OST_REPL_CTXT);
2354                 filter_olg_fini(olg);
2355                 RETURN(-ENODEV);
2356         }
2357         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2358         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2359         llog_ctxt_put(ctxt);
2360         RETURN(rc);
2361 }
2362
2363 static int filter_llog_finish(struct obd_device *obd, int count)
2364 {
2365         struct filter_obd *filter = &obd->u.filter;
2366         struct llog_ctxt *ctxt;
2367         ENTRY;
2368
2369         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2370         if (ctxt) {
2371                 /*
2372                  * Make sure that no cached llcds left in recov_thread.
2373                  * We actually do sync in disconnect time, but disconnect
2374                  * may not come being marked rq_no_resend = 1.
2375                  */
2376                 llog_sync(ctxt, NULL);
2377
2378                 /*
2379                  * Balance class_import_get() in llog_receptor_accept().
2380                  * This is safe to do, as llog is already synchronized
2381                  * and its import may go.
2382                  */
2383                 cfs_mutex_down(&ctxt->loc_sem);
2384                 if (ctxt->loc_imp) {
2385                         class_import_put(ctxt->loc_imp);
2386                         ctxt->loc_imp = NULL;
2387                 }
2388                 cfs_mutex_up(&ctxt->loc_sem);
2389                 llog_ctxt_put(ctxt);
2390         }
2391
2392         if (filter->fo_lcm) {
2393                 cfs_mutex_down(&ctxt->loc_sem);
2394                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2395                 filter->fo_lcm = NULL;
2396                 cfs_mutex_up(&ctxt->loc_sem);
2397         }
2398         RETURN(filter_olg_fini(&obd->obd_olg));
2399 }
2400
2401 /**
2402  * Find the group llog according to group index in the llog group list.
2403  */
2404 static struct obd_llog_group *
2405 filter_find_olg_internal(struct filter_obd *filter, int group)
2406 {
2407         struct obd_llog_group *olg;
2408
2409         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2410         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2411                 if (olg->olg_group == group)
2412                         RETURN(olg);
2413         }
2414         RETURN(NULL);
2415 }
2416
2417 /**
2418  * Find the group llog according to group index on the filter
2419  */
2420 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2421 {
2422         struct obd_llog_group *olg = NULL;
2423         struct filter_obd *filter;
2424
2425         filter = &obd->u.filter;
2426
2427         if (group == FILTER_GROUP_LLOG)
2428                 RETURN(&obd->obd_olg);
2429
2430         cfs_spin_lock(&filter->fo_llog_list_lock);
2431         olg = filter_find_olg_internal(filter, group);
2432         cfs_spin_unlock(&filter->fo_llog_list_lock);
2433
2434         RETURN(olg);
2435 }
2436 /**
2437  * Find the llog_group of the filter according to the group. If it can not
2438  * find, create the llog_group, which only happens when mds is being synced
2439  * with OST.
2440  */
2441 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2442 {
2443         struct obd_llog_group *olg = NULL;
2444         struct filter_obd *filter;
2445         int rc;
2446
2447         filter = &obd->u.filter;
2448
2449         if (group == FILTER_GROUP_LLOG)
2450                 RETURN(&obd->obd_olg);
2451
2452         cfs_spin_lock(&filter->fo_llog_list_lock);
2453         olg = filter_find_olg_internal(filter, group);
2454         if (olg) {
2455                 if (olg->olg_initializing) {
2456                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2457                 } else {
2458                         GOTO(out_unlock, olg);
2459                 }
2460         }
2461         OBD_ALLOC_PTR(olg);
2462         if (olg == NULL)
2463                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2464
2465         llog_group_init(olg, group);
2466         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2467         olg->olg_initializing = 1;
2468         cfs_spin_unlock(&filter->fo_llog_list_lock);
2469
2470         rc = obd_llog_init(obd, olg, obd, NULL);
2471         if (rc) {
2472                cfs_spin_lock(&filter->fo_llog_list_lock);
2473                cfs_list_del(&olg->olg_list);
2474                cfs_spin_unlock(&filter->fo_llog_list_lock);
2475                OBD_FREE_PTR(olg);
2476                GOTO(out, olg = ERR_PTR(-ENOMEM));
2477         }
2478         cfs_spin_lock(&filter->fo_llog_list_lock);
2479         olg->olg_initializing = 0;
2480         cfs_spin_unlock(&filter->fo_llog_list_lock);
2481         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2482               obd->obd_name, group, olg);
2483 out:
2484         RETURN(olg);
2485
2486 out_unlock:
2487         cfs_spin_unlock(&filter->fo_llog_list_lock);
2488         GOTO(out, olg);
2489 }
2490
2491 static int filter_llog_connect(struct obd_export *exp,
2492                                struct llogd_conn_body *body)
2493 {
2494         struct obd_device *obd = exp->exp_obd;
2495         struct llog_ctxt *ctxt;
2496         struct obd_llog_group *olg;
2497         int rc;
2498         ENTRY;
2499
2500         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2501                obd->obd_name, body->lgdc_logid.lgl_oid,
2502                body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2503
2504         olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2505         if (!olg) {
2506                 CERROR(" %s: can not find olg of group %d\n",
2507                        obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2508                 RETURN(-ENOENT);
2509         }
2510         llog_group_set_export(olg, exp);
2511
2512         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2513         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2514                  body->lgdc_ctxt_idx);
2515
2516         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2517               obd->obd_name, body->lgdc_logid.lgl_oid,
2518               body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2519
2520         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2521         obd->u.filter.fo_mds_ost_sync = 1;
2522         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2523         rc = llog_connect(ctxt, &body->lgdc_logid,
2524                           &body->lgdc_gen, NULL);
2525         llog_ctxt_put(ctxt);
2526         if (rc != 0)
2527                 CERROR("failed to connect rc %d idx %d\n", rc,
2528                                 body->lgdc_ctxt_idx);
2529
2530         RETURN(rc);
2531 }
2532
2533 static int filter_llog_preclean(struct obd_device *obd)
2534 {
2535         struct obd_llog_group *olg, *tmp;
2536         struct filter_obd *filter;
2537         cfs_list_t  remove_list;
2538         int rc = 0;
2539         ENTRY;
2540
2541         rc = obd_llog_finish(obd, 0);
2542         if (rc)
2543                 CERROR("failed to cleanup llogging subsystem\n");
2544
2545         filter = &obd->u.filter;
2546         CFS_INIT_LIST_HEAD(&remove_list);
2547
2548         cfs_spin_lock(&filter->fo_llog_list_lock);
2549         while (!cfs_list_empty(&filter->fo_llog_list)) {
2550                 olg = cfs_list_entry(filter->fo_llog_list.next,
2551                                      struct obd_llog_group, olg_list);
2552                 cfs_list_del(&olg->olg_list);
2553                 cfs_list_add(&olg->olg_list, &remove_list);
2554         }
2555         cfs_spin_unlock(&filter->fo_llog_list_lock);
2556
2557         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2558                 cfs_list_del_init(&olg->olg_list);
2559                 rc = filter_olg_fini(olg);
2560                 if (rc)
2561                         CERROR("failed to cleanup llogging subsystem for %u\n",
2562                                olg->olg_group);
2563                 OBD_FREE_PTR(olg);
2564         }
2565
2566         RETURN(rc);
2567 }
2568
2569 static int filter_precleanup(struct obd_device *obd,
2570                              enum obd_cleanup_stage stage)
2571 {
2572         int rc = 0;
2573         ENTRY;
2574
2575         switch(stage) {
2576         case OBD_CLEANUP_EARLY:
2577                 break;
2578         case OBD_CLEANUP_EXPORTS:
2579                 /* Stop recovery before namespace cleanup. */
2580                 target_recovery_fini(obd);
2581                 rc = filter_llog_preclean(obd);
2582                 break;
2583         }
2584         RETURN(rc);
2585 }
2586
2587 static int filter_cleanup(struct obd_device *obd)
2588 {
2589         struct filter_obd *filter = &obd->u.filter;
2590         ENTRY;
2591
2592         if (obd->obd_fail)
2593                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2594                               "will be preserved.\n", obd->obd_name);
2595
2596         obd_exports_barrier(obd);
2597         obd_zombie_barrier();
2598
2599         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2600         lprocfs_free_per_client_stats(obd);
2601         lprocfs_free_obd_stats(obd);
2602         lprocfs_obd_cleanup(obd);
2603         lquota_cleanup(filter_quota_interface_ref, obd);
2604
2605         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2606         obd->obd_namespace = NULL;
2607
2608         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2609
2610         if (obd->u.obt.obt_sb == NULL)
2611                 RETURN(0);
2612
2613         filter_post(obd);
2614
2615         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2616         shrink_dcache_sb(obd->u.obt.obt_sb);
2617
2618         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2619         obd->u.obt.obt_sb = NULL;
2620
2621         fsfilt_put_ops(obd->obd_fsops);
2622
2623         filter_iobuf_pool_done(filter);
2624
2625         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2626
2627         RETURN(0);
2628 }
2629
2630 static int filter_connect_internal(struct obd_export *exp,
2631                                    struct obd_connect_data *data,
2632                                    int reconnect)
2633 {
2634         struct filter_export_data *fed = &exp->exp_filter_data;
2635
2636         if (!data)
2637                 RETURN(0);
2638
2639         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2640                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2641                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2642                data->ocd_connect_flags, data->ocd_version,
2643                data->ocd_grant, data->ocd_index);
2644
2645         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2646                 CWARN("!!! This export (nid %s) used object group %d "
2647                        "earlier; now it's trying to use group %d!  This could "
2648                        "be a bug in the MDS. Please report to "
2649                        "http://bugzilla.lustre.org/\n",
2650                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2651                 RETURN(-EPROTO);
2652         }
2653         fed->fed_group = data->ocd_group;
2654
2655         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2656         exp->exp_connect_flags = data->ocd_connect_flags;
2657         data->ocd_version = LUSTRE_VERSION_CODE;
2658
2659         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2660         if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2661                   data->ocd_connect_flags & OBD_CONNECT_MDS))
2662                 RETURN(-EPROTO);
2663
2664         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2665                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2666                 obd_size left, want;
2667
2668                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2669                 left = filter_grant_space_left(exp);
2670                 want = data->ocd_grant;
2671                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2672                 data->ocd_grant = fed->fed_grant;
2673                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2674
2675                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2676                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2677                        exp->exp_client_uuid.uuid, exp,
2678                        data->ocd_grant, want, left);
2679
2680                 filter->fo_tot_granted_clients ++;
2681         }
2682
2683         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2684                 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2685                 int index = le32_to_cpu(lsd->lsd_ost_index);
2686
2687                 if (!(lsd->lsd_feature_compat &
2688                       cpu_to_le32(OBD_COMPAT_OST))) {
2689                         /* this will only happen on the first connect */
2690                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2691                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2692                         /* sync is not needed here as filter_client_add will
2693                          * set exp_need_sync flag */
2694                         filter_update_server_data(exp->exp_obd);
2695                 } else if (index != data->ocd_index) {
2696                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2697                                            " %u doesn't match actual OST index"
2698                                            " %u in last_rcvd file, bad "
2699                                            "configuration?\n",
2700                                            obd_export_nid2str(exp), index,
2701                                            data->ocd_index);
2702                         RETURN(-EBADF);
2703                 }
2704                 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2705                  * FIXME: We don't strictly need the COMPAT flag for that,
2706                  * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2707                  * FIXME: We needed it for the index, as index 0 is valid. */
2708         }
2709
2710         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2711                 data->ocd_brw_size = 65536;
2712         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2713                 data->ocd_brw_size = min(data->ocd_brw_size,
2714                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2715                 LASSERT(data->ocd_brw_size);
2716         }
2717
2718         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2719                 __u32 cksum_types = data->ocd_cksum_types;
2720
2721                 /* The client set in ocd_cksum_types the checksum types it
2722                  * supports. We have to mask off the algorithms that we don't
2723                  * support */
2724                 if (cksum_types & OBD_CKSUM_ALL)
2725                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2726                 else
2727                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2728
2729                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2730                                    "%x\n", exp->exp_obd->obd_name,
2731                                    obd_export_nid2str(exp), cksum_types,
2732                                    data->ocd_cksum_types);
2733         } else {
2734                 /* This client does not support OBD_CONNECT_CKSUM
2735                  * fall back to CRC32 */
2736                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2737                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2738                                    exp->exp_obd->obd_name,
2739                                    obd_export_nid2str(exp));
2740         }
2741
2742         RETURN(0);
2743 }
2744
2745 static int filter_reconnect(const struct lu_env *env,
2746                             struct obd_export *exp, struct obd_device *obd,
2747                             struct obd_uuid *cluuid,
2748                             struct obd_connect_data *data,
2749                             void *localdata)
2750 {
2751         int rc;
2752         ENTRY;
2753
2754         if (exp == NULL || obd == NULL || cluuid == NULL)
2755                 RETURN(-EINVAL);
2756
2757         rc = filter_connect_internal(exp, data, 1);
2758         if (rc == 0)
2759                 filter_export_stats_init(obd, exp, localdata);
2760
2761         RETURN(rc);
2762 }
2763
2764 /* nearly identical to mds_connect */
2765 static int filter_connect(const struct lu_env *env,
2766                           struct obd_export **exp, struct obd_device *obd,
2767                           struct obd_uuid *cluuid,
2768                           struct obd_connect_data *data, void *localdata)
2769 {
2770         struct lvfs_run_ctxt saved;
2771         struct lustre_handle conn = { 0 };
2772         struct obd_export *lexp;
2773         struct filter_export_data *fed;
2774         struct lsd_client_data *lcd = NULL;
2775         __u32 group;
2776         int rc;
2777         ENTRY;
2778
2779         if (exp == NULL || obd == NULL || cluuid == NULL)
2780                 RETURN(-EINVAL);
2781
2782         rc = class_connect(&conn, obd, cluuid);
2783         if (rc)
2784                 RETURN(rc);
2785         lexp = class_conn2export(&conn);
2786         LASSERT(lexp != NULL);
2787
2788         fed = &lexp->exp_filter_data;
2789
2790         rc = filter_connect_internal(lexp, data, 0);
2791         if (rc)
2792                 GOTO(cleanup, rc);
2793
2794         filter_export_stats_init(obd, lexp, localdata);
2795         if (obd->obd_replayable) {
2796                 OBD_ALLOC(lcd, sizeof(*lcd));
2797                 if (!lcd) {
2798                         CERROR("filter: out of memory for client data\n");
2799                         GOTO(cleanup, rc = -ENOMEM);
2800                 }
2801
2802                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2803                 fed->fed_lcd = lcd;
2804                 rc = filter_client_add(obd, lexp, -1);
2805                 if (rc)
2806                         GOTO(cleanup, rc);
2807         }
2808
2809         group = data->ocd_group;
2810
2811         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2812               obd->obd_name, lexp->exp_handle.h_cookie, group);
2813
2814         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2815         rc = filter_read_groups(obd, group, 1);
2816         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2817         if (rc != 0) {
2818                 CERROR("can't read group %u\n", group);
2819                 GOTO(cleanup, rc);
2820         }
2821
2822         GOTO(cleanup, rc);
2823
2824 cleanup:
2825         if (rc) {
2826                 class_disconnect(lexp);
2827                 lprocfs_exp_cleanup(lexp);
2828                 *exp = NULL;
2829         } else {
2830                 *exp = lexp;
2831         }
2832
2833         RETURN(rc);
2834 }
2835
2836 /* Do extra sanity checks for grant accounting.  We do this at connect,
2837  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2838  * always get rid of it or turn it off when we know accounting is good. */
2839 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2840 {
2841         struct filter_export_data *fed;
2842         struct obd_export *exp;
2843         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2844         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2845         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2846
2847         if (cfs_list_empty(&obd->obd_exports))
2848                 return;
2849
2850         /* We don't want to do this for large machines that do lots of
2851            mounts or unmounts.  It burns... */
2852         if (obd->obd_num_exports > 100)
2853                 return;
2854
2855         cfs_spin_lock(&obd->obd_osfs_lock);
2856         cfs_spin_lock(&obd->obd_dev_lock);
2857         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2858                 int error = 0;
2859                 fed = &exp->exp_filter_data;
2860                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2861                     fed->fed_dirty < 0)
2862                         error = 1;
2863                 if (maxsize > 0) { /* we may not have done a statfs yet */
2864                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2865                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2866                                  exp->exp_client_uuid.uuid, exp,
2867                                  fed->fed_grant, fed->fed_pending, maxsize);
2868                         LASSERTF(fed->fed_dirty <= maxsize,
2869                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2870                                  exp->exp_client_uuid.uuid, exp,
2871                                  fed->fed_dirty, maxsize);
2872                 }
2873                 if (error)
2874                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2875                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2876                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2877                 else
2878                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2879                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2880                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2881                 tot_granted += fed->fed_grant + fed->fed_pending;
2882                 tot_pending += fed->fed_pending;
2883                 tot_dirty += fed->fed_dirty;
2884         }
2885         fo_tot_granted = obd->u.filter.fo_tot_granted;
2886         fo_tot_pending = obd->u.filter.fo_tot_pending;
2887         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2888         cfs_spin_unlock(&obd->obd_dev_lock);
2889         cfs_spin_unlock(&obd->obd_osfs_lock);
2890
2891         /* Do these assertions outside the spinlocks so we don't kill system */
2892         if (tot_granted != fo_tot_granted)
2893                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2894                        func, tot_granted, fo_tot_granted);
2895         if (tot_pending != fo_tot_pending)
2896                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2897                        func, tot_pending, fo_tot_pending);
2898         if (tot_dirty != fo_tot_dirty)
2899                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2900                        func, tot_dirty, fo_tot_dirty);
2901         if (tot_pending > tot_granted)
2902                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2903                        func, tot_pending, tot_granted);
2904         if (tot_granted > maxsize)
2905                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2906                        func, tot_granted, maxsize);
2907         if (tot_dirty > maxsize)
2908                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2909                        func, tot_dirty, maxsize);
2910 }
2911
2912 /* Remove this client from the grant accounting totals.  We also remove
2913  * the export from the obd device under the osfs and dev locks to ensure
2914  * that the filter_grant_sanity_check() calculations are always valid.
2915  * The client should do something similar when it invalidates its import. */
2916 static void filter_grant_discard(struct obd_export *exp)
2917 {
2918         struct obd_device *obd = exp->exp_obd;
2919         struct filter_obd *filter = &obd->u.filter;
2920         struct filter_export_data *fed = &exp->exp_filter_data;
2921
2922         cfs_spin_lock(&obd->obd_osfs_lock);
2923         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2924                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2925                  obd->obd_name, filter->fo_tot_granted,
2926                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2927         filter->fo_tot_granted -= fed->fed_grant;
2928         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2929                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2930                  obd->obd_name, filter->fo_tot_pending,
2931                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2932         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2933         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2934                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2935                  obd->obd_name, filter->fo_tot_dirty,
2936                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2937         filter->fo_tot_dirty -= fed->fed_dirty;
2938         fed->fed_dirty = 0;
2939         fed->fed_grant = 0;
2940
2941         cfs_spin_unlock(&obd->obd_osfs_lock);
2942 }
2943
2944 static int filter_destroy_export(struct obd_export *exp)
2945 {
2946         struct filter_export_data *fed = &exp->exp_filter_data;
2947         ENTRY;
2948
2949         if (fed->fed_pending)
2950                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2951                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2952                        exp, fed->fed_pending);
2953
2954         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2955
2956         target_destroy_export(exp);
2957         ldlm_destroy_export(exp);
2958
2959         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2960                 RETURN(0);
2961
2962         lut_client_free(exp);
2963
2964         if (!exp->exp_obd->obd_replayable)
2965                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2966
2967         filter_grant_discard(exp);
2968         filter_fmd_cleanup(exp);
2969
2970         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2971                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2972                 if (filter->fo_tot_granted_clients > 0)
2973                         filter->fo_tot_granted_clients --;
2974         }
2975
2976         if (!(exp->exp_flags & OBD_OPT_FORCE))
2977                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2978
2979         RETURN(0);
2980 }
2981
2982 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2983 {
2984         struct obd_llog_group *olg_min, *olg;
2985         struct filter_obd *filter;
2986         int worked = -1, group;
2987         struct llog_ctxt *ctxt;
2988         ENTRY;
2989
2990         filter = &obd->u.filter;
2991
2992         /* we can't sync log holding spinlock. also, we do not want to get
2993          * into livelock. so we do following: loop over MDS's exports in
2994          * group order and skip already synced llogs -bzzz */
2995         do {
2996                 /* look for group with min. number, but > worked */
2997                 olg_min = NULL;
2998                 group = 1 << 30;
2999                 cfs_spin_lock(&filter->fo_llog_list_lock);
3000                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3001                         if (olg->olg_group <= worked) {
3002                                 /* this group is already synced */
3003                                 continue;
3004                         }
3005                         if (group < olg->olg_group) {
3006                                 /* we have group with smaller number to sync */
3007                                 continue;
3008                         }
3009                         /* store current minimal group */
3010                         olg_min = olg;
3011                         group = olg->olg_group;
3012                 }
3013                 cfs_spin_unlock(&filter->fo_llog_list_lock);
3014
3015                 if (olg_min == NULL)
3016                         break;
3017
3018                 worked = olg_min->olg_group;
3019                 if (olg_min->olg_exp &&
3020                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3021                         int err;
3022                         ctxt = llog_group_get_ctxt(olg_min,
3023                                                    LLOG_MDS_OST_REPL_CTXT);
3024                         if (ctxt) {
3025                                 err = llog_sync(ctxt, olg_min->olg_exp);
3026                                 llog_ctxt_put(ctxt);
3027                                 if (err) {
3028                                         CERROR("error flushing logs to MDS: "
3029                                                "rc %d\n", err);
3030                                 }
3031                         }
3032                 }
3033         } while (olg_min != NULL);
3034 }
3035
3036 /* Also incredibly similar to mds_disconnect */
3037 static int filter_disconnect(struct obd_export *exp)
3038 {
3039         struct obd_device *obd = exp->exp_obd;
3040         int rc;
3041         ENTRY;
3042
3043         LASSERT(exp);
3044         class_export_get(exp);
3045
3046         if (!(exp->exp_flags & OBD_OPT_FORCE))
3047                 filter_grant_sanity_check(obd, __FUNCTION__);
3048         filter_grant_discard(exp);
3049
3050         /* Flush any remaining cancel messages out to the target */
3051         filter_sync_llogs(obd, exp);
3052
3053         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3054
3055         rc = server_disconnect_export(exp);
3056
3057         if (exp->exp_obd->obd_replayable)
3058                 filter_client_del(exp);
3059         else
3060                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3061
3062         class_export_put(exp);
3063         RETURN(rc);
3064 }
3065
3066 /* reverse import is changed, sync all cancels */
3067 static void filter_revimp_update(struct obd_export *exp)
3068 {
3069         ENTRY;
3070
3071         LASSERT(exp);
3072         class_export_get(exp);
3073
3074         /* flush any remaining cancel messages out to the target */
3075         filter_sync_llogs(exp->exp_obd, exp);
3076         class_export_put(exp);
3077         EXIT;
3078 }
3079
3080 static int filter_ping(struct obd_export *exp)
3081 {
3082         filter_fmd_expire(exp);
3083         return 0;
3084 }
3085
3086 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3087                                   const char *what, int quiet)
3088 {
3089         struct dentry *dchild = NULL;
3090         obd_gr group = 0;
3091
3092         if (oa->o_valid & OBD_MD_FLGROUP)
3093                 group = oa->o_gr;
3094
3095         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3096
3097         if (IS_ERR(dchild)) {
3098                 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3099                        what, group, oa->o_id);
3100                 RETURN(dchild);
3101         }
3102
3103         if (dchild->d_inode == NULL) {
3104                 if (!quiet)
3105                         CERROR("%s: %s on non-existent object: "LPU64"\n",
3106                                obd->obd_name, what, oa->o_id);
3107                 f_dput(dchild);
3108                 RETURN(ERR_PTR(-ENOENT));
3109         }
3110
3111         return dchild;
3112 }
3113
3114 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3115 {
3116         struct dentry *dentry = NULL;
3117         struct obd_device *obd;
3118         int rc = 0;
3119         ENTRY;
3120
3121         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3122         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3123                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3124         if (rc)
3125                 RETURN(rc);
3126
3127         obd = class_exp2obd(exp);
3128         if (obd == NULL) {
3129                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3130                 RETURN(-EINVAL);
3131         }
3132
3133         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3134         if (IS_ERR(dentry))
3135                 RETURN(PTR_ERR(dentry));
3136
3137         /* Limit the valid bits in the return data to what we actually use */
3138         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3139         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3140
3141         f_dput(dentry);
3142         RETURN(rc);
3143 }
3144
3145 /* this should be enabled/disabled in condition to enabled/disabled large
3146  * inodes (fast EAs) in backing store FS. */
3147 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3148                         void *handle, struct obdo *oa)
3149 {
3150         struct obd_device *obd = exp->exp_obd;
3151         int rc = 0;
3152         ENTRY;
3153
3154         if (oa->o_valid & OBD_MD_FLFID) {
3155                 struct filter_fid ff;
3156
3157                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3158                         oa->o_gr = 0;
3159                 /* packing fid and converting it to LE for storing into EA.
3160                  * Here ->o_stripe_idx should be filled by LOV and rest of
3161                  * fields - by client. */
3162                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3163                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3164                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3165                 ff.ff_objid = cpu_to_le64(oa->o_id);
3166                 ff.ff_group = cpu_to_le64(oa->o_gr);
3167
3168                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3169                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3170                        oa->o_generation, oa->o_id, oa->o_gr);
3171
3172                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3173                 if (rc)
3174                         CERROR("store fid in object failed! rc: %d\n", rc);
3175         } else {
3176                 CDEBUG(D_HA, "OSS object without fid info!\n");
3177         }
3178
3179         RETURN(rc);
3180 }
3181
3182 /* this is called from filter_truncate() until we have filter_punch() */
3183 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3184                             struct obdo *oa, struct obd_trans_info *oti)
3185 {
3186         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3187         struct llog_cookie *fcc = NULL;
3188         struct filter_obd *filter;
3189         int rc, err, sync = 0;
3190         loff_t old_size = 0;
3191         unsigned int ia_valid;
3192         struct inode *inode;
3193         struct page *page = NULL;
3194         struct iattr iattr;
3195         void *handle;
3196         ENTRY;
3197
3198         LASSERT(dentry != NULL);
3199         LASSERT(!IS_ERR(dentry));
3200