Whamcloud - gitweb
8043556141c8272658b064d23b5a21de3073415a
[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                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
370                 /* Transaction needed to fix bug 1403 */
371                 handle = fsfilt_start(obd,
372                                       filter->fo_rcvd_filp->f_dentry->d_inode,
373                                       FSFILT_OP_SETATTR, NULL);
374                 if (IS_ERR(handle)) {
375                         rc = PTR_ERR(handle);
376                         CERROR("unable to start transaction: rc %d\n", rc);
377                 } else {
378                         fed->fed_lcd->lcd_last_epoch = fsd->lsd_start_epoch;
379                         exp->exp_last_request_time = cfs_time_current_sec();
380                         rc = fsfilt_add_journal_cb(obd, 0, handle,
381                                                    target_client_add_cb,
382                                                    class_export_cb_get(exp));
383                         if (rc == 0) {
384                                 cfs_spin_lock(&exp->exp_lock);
385                                 exp->exp_need_sync = 1;
386                                 cfs_spin_unlock(&exp->exp_lock);
387                         }
388                         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
389                                                  fed->fed_lcd,
390                                                  sizeof(*fed->fed_lcd),
391                                                  &off, rc /* sync if no cb */);
392                         fsfilt_commit(obd,
393                                       filter->fo_rcvd_filp->f_dentry->d_inode,
394                                       handle, 0);
395                 }
396                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
397
398                 if (rc) {
399                         CERROR("error writing %s client idx %u: rc %d\n",
400                                LAST_RCVD, fed->fed_lr_idx, rc);
401                         RETURN(rc);
402                 }
403         }
404         RETURN(0);
405 }
406
407 static int filter_client_del(struct obd_export *exp)
408 {
409         struct filter_export_data *fed = &exp->exp_filter_data;
410         struct filter_obd *filter = &exp->exp_obd->u.filter;
411         struct obd_device *obd = exp->exp_obd;
412         struct lvfs_run_ctxt saved;
413         struct lsd_client_data *lcd = fed->fed_lcd;
414         int rc;
415         loff_t off;
416         ENTRY;
417
418         if (lcd == NULL)
419                 RETURN(0);
420
421         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
422         if (strcmp(lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
423                 GOTO(free, 0);
424
425         LASSERT(filter->fo_last_rcvd_slots != NULL);
426
427         off = fed->fed_lr_off;
428
429         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
430                fed->fed_lr_idx, fed->fed_lr_off, lcd->lcd_uuid);
431
432         /* Don't clear fed_lr_idx here as it is likely also unset.  At worst
433          * we leak a client slot that will be cleaned on the next recovery. */
434         if (off <= 0) {
435                 CERROR("%s: client idx %d has med_off %lld\n",
436                        obd->obd_name, fed->fed_lr_idx, off);
437                 GOTO(free, rc = -EINVAL);
438         }
439
440         /* Clear the bit _after_ zeroing out the client so we don't
441            race with filter_client_add and zero out new clients.*/
442         if (!cfs_test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
443                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
444                        fed->fed_lr_idx);
445                 LBUG();
446         }
447
448         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
449         /* Make sure the server's last_transno is up to date.
450          * This should be done before zeroing client slot so last_transno will
451          * be in server data or in client data in case of failure */
452         filter_update_server_data(obd);
453
454         cfs_mutex_down(&fed->fed_lcd_lock);
455         memset(fed->fed_lcd->lcd_uuid, 0, sizeof fed->fed_lcd->lcd_uuid);
456         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
457                                  fed->fed_lcd,
458                                  sizeof(*fed->fed_lcd), &off, 0);
459         cfs_mutex_up(&fed->fed_lcd_lock);
460         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
461
462         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
463                "zero out client %s at idx %u/%llu in %s, rc %d\n",
464                lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
465                LAST_RCVD, rc);
466         EXIT;
467 free:
468         return 0;
469 }
470
471 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
472 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
473                                          struct filter_mod_data *fmd)
474 {
475         LASSERT_SPIN_LOCKED(&fed->fed_lock);
476         if (--fmd->fmd_refcount == 0) {
477                 /* XXX when we have persistent reservations and the handle
478                  * is stored herein we need to drop it here. */
479                 fed->fed_mod_count--;
480                 cfs_list_del(&fmd->fmd_list);
481                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
482         }
483 }
484
485 /* drop fmd reference, free it if last ref */
486 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
487 {
488         struct filter_export_data *fed;
489
490         if (fmd == NULL)
491                 return;
492
493         fed = &exp->exp_filter_data;
494         cfs_spin_lock(&fed->fed_lock);
495         filter_fmd_put_nolock(fed, fmd); /* caller reference */
496         cfs_spin_unlock(&fed->fed_lock);
497 }
498
499 /* expire entries from the end of the list if there are too many
500  * or they are too old */
501 static void filter_fmd_expire_nolock(struct filter_obd *filter,
502                                      struct filter_export_data *fed,
503                                      struct filter_mod_data *keep)
504 {
505         struct filter_mod_data *fmd, *tmp;
506
507         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
508                 if (fmd == keep)
509                         break;
510
511                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
512                     fed->fed_mod_count < filter->fo_fmd_max_num)
513                         break;
514
515                 cfs_list_del_init(&fmd->fmd_list);
516                 filter_fmd_put_nolock(fed, fmd); /* list reference */
517         }
518 }
519
520 void filter_fmd_expire(struct obd_export *exp)
521 {
522         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
523         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
524                                  &exp->exp_filter_data, NULL);
525         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
526 }
527
528 /* find specified objid, group in export fmd list.
529  * caller must hold fed_lock and take fmd reference itself */
530 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
531                                                 struct filter_export_data *fed,
532                                                 obd_id objid, obd_gr group)
533 {
534         struct filter_mod_data *found = NULL, *fmd;
535
536         LASSERT_SPIN_LOCKED(&fed->fed_lock);
537
538         cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
539                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
540                         found = fmd;
541                         cfs_list_del(&fmd->fmd_list);
542                         cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
543                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
544                         break;
545                 }
546         }
547
548         filter_fmd_expire_nolock(filter, fed, found);
549
550         return found;
551 }
552
553 /* Find fmd based on objid and group, or return NULL if not found. */
554 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
555                                         obd_id objid, obd_gr group)
556 {
557         struct filter_mod_data *fmd;
558
559         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
560         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
561                                      &exp->exp_filter_data, objid, group);
562         if (fmd)
563                 fmd->fmd_refcount++;    /* caller reference */
564         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
565
566         return fmd;
567 }
568
569 /* Find fmd based on objid and group, or create a new one if none is found.
570  * It is possible for this function to return NULL under memory pressure,
571  * or if objid = 0 is passed (which will only cause old entries to expire).
572  * Currently this is not fatal because any fmd state is transient and
573  * may also be freed when it gets sufficiently old. */
574 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
575                                        obd_id objid, obd_gr group)
576 {
577         struct filter_export_data *fed = &exp->exp_filter_data;
578         struct filter_mod_data *found = NULL, *fmd_new = NULL;
579
580         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
581
582         cfs_spin_lock(&fed->fed_lock);
583         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
584         if (fmd_new) {
585                 if (found == NULL) {
586                         cfs_list_add_tail(&fmd_new->fmd_list,
587                                           &fed->fed_mod_list);
588                         fmd_new->fmd_id = objid;
589                         fmd_new->fmd_gr = group;
590                         fmd_new->fmd_refcount++;   /* list reference */
591                         found = fmd_new;
592                         fed->fed_mod_count++;
593                 } else {
594                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
595                 }
596         }
597         if (found) {
598                 found->fmd_refcount++;          /* caller reference */
599                 found->fmd_expire = jiffies +
600                         exp->exp_obd->u.filter.fo_fmd_max_age;
601         }
602
603         cfs_spin_unlock(&fed->fed_lock);
604
605         return found;
606 }
607
608 #ifdef DO_FMD_DROP
609 /* drop fmd list reference so it will disappear when last reference is put.
610  * This isn't so critical because it would in fact only affect the one client
611  * that is doing the unlink and at worst we have an stale entry referencing
612  * an object that should never be used again. */
613 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
614 {
615         struct filter_mod_data *found = NULL;
616
617         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
618         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
619         if (found) {
620                 cfs_list_del_init(&found->fmd_list);
621                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
622         }
623         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
624 }
625 #else
626 #define filter_fmd_drop(exp, objid, group)
627 #endif
628
629 /* remove all entries from fmd list */
630 static void filter_fmd_cleanup(struct obd_export *exp)
631 {
632         struct filter_export_data *fed = &exp->exp_filter_data;
633         struct filter_mod_data *fmd = NULL, *tmp;
634
635         cfs_spin_lock(&fed->fed_lock);
636         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
637                 cfs_list_del_init(&fmd->fmd_list);
638                 filter_fmd_put_nolock(fed, fmd);
639         }
640         cfs_spin_unlock(&fed->fed_lock);
641 }
642
643 static int filter_init_export(struct obd_export *exp)
644 {
645         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
646         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
647
648         cfs_spin_lock(&exp->exp_lock);
649         exp->exp_connecting = 1;
650         cfs_spin_unlock(&exp->exp_lock);
651
652         return ldlm_init_export(exp);
653 }
654
655 static int filter_free_server_data(struct filter_obd *filter)
656 {
657         lut_fini(NULL, filter->fo_obt.obt_lut);
658         OBD_FREE_PTR(filter->fo_obt.obt_lut);
659         return 0;
660 }
661
662 /* assumes caller is already in kernel ctxt */
663 int filter_update_server_data(struct obd_device *obd)
664 {
665         struct file *filp = obd->u.obt.obt_rcvd_filp;
666         struct lr_server_data *fsd = class_server_data(obd);
667         loff_t off = 0;
668         int rc;
669         ENTRY;
670
671         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
672         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
673                le64_to_cpu(fsd->lsd_last_transno));
674         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
675                le64_to_cpu(fsd->lsd_mount_count));
676
677         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, 0);
678         if (rc)
679                 CERROR("error writing lr_server_data: rc = %d\n", rc);
680
681         RETURN(rc);
682 }
683
684 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
685                              int force_sync)
686 {
687         struct filter_obd *filter = &obd->u.filter;
688         __u64 tmp;
689         loff_t off = 0;
690         int rc;
691         ENTRY;
692
693         if (filter->fo_last_objid_files[group] == NULL) {
694                 CERROR("Object group "LPU64" not fully setup; not updating "
695                        "last_objid\n", group);
696                 RETURN(-EINVAL);
697         }
698
699         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
700                obd->obd_name, group, filter->fo_last_objids[group]);
701
702         tmp = cpu_to_le64(filter->fo_last_objids[group]);
703         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
704                                  &tmp, sizeof(tmp), &off, force_sync);
705         if (rc)
706                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
707                        group, rc);
708         RETURN(rc);
709 }
710 extern int ost_handle(struct ptlrpc_request *req);
711 /* assumes caller has already in kernel ctxt */
712 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
713 {
714         struct filter_obd *filter = &obd->u.filter;
715         struct lr_server_data *fsd;
716         struct lsd_client_data *lcd = NULL;
717         struct inode *inode = filp->f_dentry->d_inode;
718         unsigned long last_rcvd_size = i_size_read(inode);
719         struct lu_target *lut;
720         __u64 mount_count;
721         __u32 start_epoch;
722         int cl_idx;
723         loff_t off = 0;
724         int rc;
725
726         /* ensure padding in the struct is the correct size */
727         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
728                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
729         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
730                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
731
732         /* allocate and initialize lu_target */
733         OBD_ALLOC_PTR(lut);
734         if (lut == NULL)
735                 RETURN(-ENOMEM);
736         rc = lut_init(NULL, lut, obd, NULL);
737         if (rc)
738                 GOTO(err_fsd, rc);
739         fsd = class_server_data(obd);
740         if (last_rcvd_size == 0) {
741                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
742
743                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
744                 fsd->lsd_last_transno = 0;
745                 mount_count = fsd->lsd_mount_count = 0;
746                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
747                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
748                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
749                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
750                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
751                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
752                  * MDS first connects and assigns the OST index number. */
753                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
754                                                         OBD_INCOMPAT_OST);
755         } else {
756                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
757                 if (rc) {
758                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
759                                LAST_RCVD, rc);
760                         GOTO(err_fsd, rc);
761                 }
762                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
763                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
764                                            "using the wrong disk %s. Were the "
765                                            "/dev/ assignments rearranged?\n",
766                                            obd->obd_uuid.uuid, fsd->lsd_uuid);
767                         GOTO(err_fsd, rc = -EINVAL);
768                 }
769                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
770                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
771                 /* COMPAT_146 */
772                 /* Assume old last_rcvd format unless I_C_LR is set */
773                 if (!(fsd->lsd_feature_incompat &
774                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
775                         fsd->lsd_last_transno = fsd->lsd_compat14;
776                 /* end COMPAT_146 */
777                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
778                  * MDS first connects and assigns the OST index number. */
779                 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
780                                                          OBD_INCOMPAT_OST);
781         }
782
783         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
784                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
785                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
786                        ~FILTER_INCOMPAT_SUPP);
787                 GOTO(err_fsd, rc = -EINVAL);
788         }
789         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
790                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
791                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
792                        ~FILTER_ROCOMPAT_SUPP);
793                 /* Do something like remount filesystem read-only */
794                 GOTO(err_fsd, rc = -EINVAL);
795         }
796
797         start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
798
799         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
800                obd->obd_name, start_epoch);
801         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
802                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
803         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
804                obd->obd_name, mount_count + 1);
805         CDEBUG(D_INODE, "%s: server data size: %u\n",
806                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
807         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
808                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
809         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
810                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
811         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
812                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
813         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
814                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
815                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
816                 le16_to_cpu(fsd->lsd_client_size));
817
818         if (!obd->obd_replayable) {
819                 CWARN("%s: recovery support OFF\n", obd->obd_name);
820                 GOTO(out, rc = 0);
821         }
822
823         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
824              off < last_rcvd_size; cl_idx++) {
825                 __u64 last_rcvd;
826                 struct obd_export *exp;
827                 struct filter_export_data *fed;
828
829                 if (!lcd) {
830                         OBD_ALLOC_PTR(lcd);
831                         if (!lcd)
832                                 GOTO(err_client, rc = -ENOMEM);
833                 }
834
835                 /* Don't assume off is incremented properly by
836                  * fsfilt_read_record(), in case sizeof(*lcd)
837                  * isn't the same as fsd->lsd_client_size.  */
838                 off = le32_to_cpu(fsd->lsd_client_start) +
839                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
840                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
841                 if (rc) {
842                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
843                                LAST_RCVD, cl_idx, off, rc);
844                         break; /* read error shouldn't cause startup to fail */
845                 }
846
847                 if (lcd->lcd_uuid[0] == '\0') {
848                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
849                                cl_idx);
850                         continue;
851                 }
852
853                 check_lcd(obd->obd_name, cl_idx, lcd);
854
855                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
856
857                 /* These exports are cleaned up by filter_disconnect(), so they
858                  * need to be set up like real exports as filter_connect() does.
859                  */
860                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
861
862                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
863                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
864                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
865                 if (IS_ERR(exp)) {
866                         if (PTR_ERR(exp) == -EALREADY) {
867                                 /* export already exists, zero out this one */
868                                 CERROR("Zeroing out duplicate export due to "
869                                        "bug 10479.\n");
870                                 lcd->lcd_uuid[0] = '\0';
871                         } else {
872                                 GOTO(err_client, rc = PTR_ERR(exp));
873                         }
874                 } else {
875                         fed = &exp->exp_filter_data;
876                         fed->fed_lcd = lcd;
877                         fed->fed_group = 0; /* will be assigned at connect */
878                         filter_export_stats_init(obd, exp, NULL);
879                         rc = filter_client_add(obd, exp, cl_idx);
880                         /* can't fail for existing client */
881                         LASSERTF(rc == 0, "rc = %d\n", rc);
882
883                         /* VBR: set export last committed */
884                         exp->exp_last_committed = last_rcvd;
885                         cfs_spin_lock(&exp->exp_lock);
886                         exp->exp_connecting = 0;
887                         exp->exp_in_recovery = 0;
888                         cfs_spin_unlock(&exp->exp_lock);
889                         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
890                         obd->obd_max_recoverable_clients++;
891                         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
892                         lcd = NULL;
893                         class_export_put(exp);
894                 }
895
896                 /* Need to check last_rcvd even for duplicated exports. */
897                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
898                        cl_idx, last_rcvd);
899
900                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
901                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
902         }
903
904         if (lcd)
905                 OBD_FREE_PTR(lcd);
906
907         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
908 out:
909         filter->fo_mount_count = mount_count + 1;
910         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
911
912         /* save it, so mount count and last_transno is current */
913         rc = filter_update_server_data(obd);
914         if (rc)
915                 GOTO(err_client, rc);
916
917         RETURN(0);
918
919 err_client:
920         class_disconnect_exports(obd);
921 err_fsd:
922         filter_free_server_data(filter);
923         RETURN(rc);
924 }
925
926 static int filter_cleanup_groups(struct obd_device *obd)
927 {
928         struct filter_obd *filter = &obd->u.filter;
929         struct file *filp;
930         struct dentry *dentry;
931         int i, j;
932         ENTRY;
933
934         if (filter->fo_dentry_O_groups != NULL) {
935                 for (i = 0; i < filter->fo_group_count; i++) {
936                         dentry = filter->fo_dentry_O_groups[i];
937                         if (dentry != NULL)
938                                 f_dput(dentry);
939                 }
940                 OBD_FREE(filter->fo_dentry_O_groups,
941                          filter->fo_group_count *
942                          sizeof(*filter->fo_dentry_O_groups));
943                 filter->fo_dentry_O_groups = NULL;
944         }
945         if (filter->fo_last_objid_files != NULL) {
946                 for (i = 0; i < filter->fo_group_count; i++) {
947                         filp = filter->fo_last_objid_files[i];
948                         if (filp != NULL)
949                                 filp_close(filp, 0);
950                 }
951                 OBD_FREE(filter->fo_last_objid_files,
952                          filter->fo_group_count *
953                          sizeof(*filter->fo_last_objid_files));
954                 filter->fo_last_objid_files = NULL;
955         }
956         if (filter->fo_dentry_O_sub != NULL) {
957                 for (i = 0; i < filter->fo_group_count; i++) {
958                         for (j = 0; j < filter->fo_subdir_count; j++) {
959                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
960                                 if (dentry != NULL)
961                                         f_dput(dentry);
962                         }
963                 }
964                 OBD_FREE(filter->fo_dentry_O_sub,
965                          filter->fo_group_count *
966                          sizeof(*filter->fo_dentry_O_sub));
967                 filter->fo_dentry_O_sub = NULL;
968         }
969         if (filter->fo_last_objids != NULL) {
970                 OBD_FREE(filter->fo_last_objids,
971                          filter->fo_group_count *
972                          sizeof(*filter->fo_last_objids));
973                 filter->fo_last_objids = NULL;
974         }
975         if (filter->fo_dentry_O != NULL) {
976                 f_dput(filter->fo_dentry_O);
977                 filter->fo_dentry_O = NULL;
978         }
979         RETURN(0);
980 }
981
982 static int filter_update_last_group(struct obd_device *obd, int group)
983 {
984         struct filter_obd *filter = &obd->u.filter;
985         struct file *filp = NULL;
986         int last_group = 0, rc;
987         loff_t off = 0;
988         ENTRY;
989
990         if (group <= filter->fo_committed_group)
991                 RETURN(0);
992
993         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
994         if (IS_ERR(filp)) {
995                 rc = PTR_ERR(filp);
996                 filp = NULL;
997                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
998                 GOTO(cleanup, rc);
999         }
1000
1001         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1002         if (rc) {
1003                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1004                 GOTO(cleanup, rc);
1005         }
1006
1007         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1008                obd->obd_name, last_group, group);
1009
1010         off = 0;
1011         last_group = group;
1012         /* must be sync: bXXXX */
1013         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1014         if (rc) {
1015                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1016                 GOTO(cleanup, rc);
1017         }
1018
1019         filter->fo_committed_group = group;
1020 cleanup:
1021         if (filp)
1022                 filp_close(filp, 0);
1023         RETURN(rc);
1024 }
1025
1026 static int filter_read_group_internal(struct obd_device *obd, int group,
1027                                       int create)
1028 {
1029         struct filter_obd *filter = &obd->u.filter;
1030         __u64 *new_objids = NULL;
1031         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1032         struct dentry **new_groups = NULL;
1033         struct file **new_files = NULL;
1034         struct dentry *dentry;
1035         struct file *filp;
1036         int old_count = filter->fo_group_count, rc, stage = 0, i;
1037         char name[25];
1038         __u64 last_objid;
1039         loff_t off = 0;
1040         int len = group + 1;
1041
1042         snprintf(name, 24, "%d", group);
1043         name[24] = '\0';
1044
1045         if (!create) {
1046                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1047                                            strlen(name));
1048                 if (IS_ERR(dentry)) {
1049                         CERROR("Cannot lookup expected object group %d: %ld\n",
1050                                group, PTR_ERR(dentry));
1051                         RETURN(PTR_ERR(dentry));
1052                 }
1053         } else {
1054                 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1055                                       name, 0700, 1);
1056                 if (IS_ERR(dentry)) {
1057                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1058                                PTR_ERR(dentry));
1059                         RETURN(PTR_ERR(dentry));
1060                 }
1061         }
1062         stage = 1;
1063
1064         snprintf(name, 24, "O/%d/LAST_ID", group);
1065         name[24] = '\0';
1066         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1067         if (IS_ERR(filp)) {
1068                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1069                 GOTO(cleanup, rc = PTR_ERR(filp));
1070         }
1071         stage = 2;
1072
1073         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1074         if (rc) {
1075                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1076                 GOTO(cleanup, rc);
1077         }
1078
1079         if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1080                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1081                 if (tmp_subdirs == NULL)
1082                         GOTO(cleanup, rc = -ENOMEM);
1083                 stage = 3;
1084
1085                 for (i = 0; i < filter->fo_subdir_count; i++) {
1086                         char dir[20];
1087                         snprintf(dir, sizeof(dir), "d%u", i);
1088
1089                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1090                                                               filter->fo_vfsmnt,
1091                                                               dir, 0700, 1);
1092                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1093                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1094                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1095                                        group, dir, rc);
1096                                 GOTO(cleanup, rc);
1097                         }
1098
1099                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1100                                tmp_subdirs->dentry[i]);
1101                 }
1102         }
1103
1104         /* 'group' is an index; we need an array of length 'group + 1' */
1105         if (group + 1 > old_count) {
1106                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1107                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1108                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1109                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1110                 stage = 4;
1111                 if (new_objids == NULL || new_subdirs == NULL ||
1112                     new_groups == NULL || new_files == NULL)
1113                         GOTO(cleanup, rc = -ENOMEM);
1114
1115                 if (old_count) {
1116                         memcpy(new_objids, filter->fo_last_objids,
1117                                old_count * sizeof(*new_objids));
1118                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1119                                old_count * sizeof(*new_subdirs));
1120                         memcpy(new_groups, filter->fo_dentry_O_groups,
1121                                old_count * sizeof(*new_groups));
1122                         memcpy(new_files, filter->fo_last_objid_files,
1123                                old_count * sizeof(*new_files));
1124
1125                         OBD_FREE(filter->fo_last_objids,
1126                                  old_count * sizeof(*new_objids));
1127                         OBD_FREE(filter->fo_dentry_O_sub,
1128                                  old_count * sizeof(*new_subdirs));
1129                         OBD_FREE(filter->fo_dentry_O_groups,
1130                                  old_count * sizeof(*new_groups));
1131                         OBD_FREE(filter->fo_last_objid_files,
1132                                  old_count * sizeof(*new_files));
1133                 }
1134                 filter->fo_last_objids = new_objids;
1135                 filter->fo_dentry_O_sub = new_subdirs;
1136                 filter->fo_dentry_O_groups = new_groups;
1137                 filter->fo_last_objid_files = new_files;
1138                 filter->fo_group_count = len;
1139         }
1140
1141         filter->fo_dentry_O_groups[group] = dentry;
1142         filter->fo_last_objid_files[group] = filp;
1143         if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1144                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1145                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1146         }
1147
1148         filter_update_last_group(obd, group);
1149
1150         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1151                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1152                 rc = filter_update_last_objid(obd, group, 1);
1153                 RETURN(rc);
1154         }
1155
1156         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1157         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1158                obd->obd_name, group, last_objid);
1159         RETURN(0);
1160  cleanup:
1161         switch (stage) {
1162         case 4:
1163                 if (new_objids != NULL)
1164                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1165                 if (new_subdirs != NULL)
1166                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1167                 if (new_groups != NULL)
1168                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1169                 if (new_files != NULL)
1170                         OBD_FREE(new_files, len * sizeof(*new_files));
1171         case 3:
1172                 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1173                         for (i = 0; i < filter->fo_subdir_count; i++) {
1174                                 if (tmp_subdirs->dentry[i] != NULL)
1175                                         dput(tmp_subdirs->dentry[i]);
1176                         }
1177                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1178                 }
1179         case 2:
1180                 filp_close(filp, 0);
1181         case 1:
1182                 dput(dentry);
1183         }
1184         RETURN(rc);
1185 }
1186
1187 static int filter_read_groups(struct obd_device *obd, int last_group,
1188                               int create)
1189 {
1190         struct filter_obd *filter = &obd->u.filter;
1191         int old_count, group, rc = 0;
1192
1193         cfs_down(&filter->fo_init_lock);
1194         old_count = filter->fo_group_count;
1195         for (group = old_count; group <= last_group; group++) {
1196                 rc = filter_read_group_internal(obd, group, create);
1197                 if (rc != 0)
1198                         break;
1199         }
1200         cfs_up(&filter->fo_init_lock);
1201         return rc;
1202 }
1203
1204 /* FIXME: object groups */
1205 static int filter_prep_groups(struct obd_device *obd)
1206 {
1207         struct filter_obd *filter = &obd->u.filter;
1208         struct dentry *O_dentry;
1209         struct file *filp;
1210         int    last_group, rc = 0, cleanup_phase = 0;
1211         loff_t off = 0;
1212         ENTRY;
1213
1214         O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1215                                 "O", 0700, 1);
1216         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1217         if (IS_ERR(O_dentry)) {
1218                 rc = PTR_ERR(O_dentry);
1219                 CERROR("cannot open/create O: rc = %d\n", rc);
1220                 GOTO(cleanup, rc);
1221         }
1222         filter->fo_dentry_O = O_dentry;
1223         cleanup_phase = 1; /* O_dentry */
1224
1225         /* we have to initialize all groups before first connections from
1226          * clients because they may send create/destroy for any group -bzzz */
1227         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1228         if (IS_ERR(filp)) {
1229                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1230                 GOTO(cleanup, rc = PTR_ERR(filp));
1231         }
1232         cleanup_phase = 2; /* filp */
1233
1234         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1235         if (rc) {
1236                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1237                 GOTO(cleanup, rc);
1238         }
1239
1240         if (off == 0)
1241                 last_group = FILTER_GROUP_MDS0;
1242
1243         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1244               FILTER_GROUP_MDS0, last_group);
1245         filter->fo_committed_group = last_group;
1246         rc = filter_read_groups(obd, last_group, 1);
1247         if (rc)
1248                 GOTO(cleanup, rc);
1249
1250         filp_close(filp, 0);
1251         RETURN(0);
1252
1253  cleanup:
1254         switch (cleanup_phase) {
1255         case 2:
1256                 filp_close(filp, 0);
1257         case 1:
1258                 filter_cleanup_groups(obd);
1259                 f_dput(filter->fo_dentry_O);
1260                 filter->fo_dentry_O = NULL;
1261         default:
1262                 break;
1263         }
1264         return rc;
1265
1266 }
1267
1268 /* setup the object store with correct subdirectories */
1269 static int filter_prep(struct obd_device *obd)
1270 {
1271         struct lvfs_run_ctxt saved;
1272         struct filter_obd *filter = &obd->u.filter;
1273         struct file *file;
1274         struct inode *inode;
1275         int rc = 0;
1276         ENTRY;
1277
1278         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1279         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1280         if (!file || IS_ERR(file)) {
1281                 rc = PTR_ERR(file);
1282                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1283                        LAST_RCVD, rc);
1284                 GOTO(out, rc);
1285         }
1286         filter->fo_rcvd_filp = file;
1287         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1288                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1289                        file->f_dentry->d_inode->i_mode);
1290                 GOTO(err_filp, rc = -ENOENT);
1291         }
1292
1293         inode = file->f_dentry->d_parent->d_inode;
1294         /* We use i_op->unlink directly in filter_vfs_unlink() */
1295         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1296                 CERROR("%s: filesystem does not support create/unlink ops\n",
1297                        obd->obd_name);
1298                 GOTO(err_filp, rc = -EOPNOTSUPP);
1299         }
1300
1301         rc = filter_init_server_data(obd, file);
1302         if (rc) {
1303                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1304                 GOTO(err_filp, rc);
1305         }
1306         LASSERT(obd->u.obt.obt_lut);
1307         target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1308
1309         /* open and create health check io file*/
1310         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1311         if (IS_ERR(file)) {
1312                 rc = PTR_ERR(file);
1313                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1314                        HEALTH_CHECK, rc);
1315                 GOTO(err_server_data, rc);
1316         }
1317         filter->fo_obt.obt_health_check_filp = file;
1318         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1319                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1320                        file->f_dentry->d_inode->i_mode);
1321                 GOTO(err_health_check, rc = -ENOENT);
1322         }
1323         rc = lvfs_check_io_health(obd, file);
1324         if (rc)
1325                 GOTO(err_health_check, rc);
1326
1327         rc = filter_prep_groups(obd);
1328         if (rc)
1329                 GOTO(err_health_check, rc);
1330 out:
1331         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1332
1333         return(rc);
1334
1335 err_health_check:
1336         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1337                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1338         filter->fo_obt.obt_health_check_filp = NULL;
1339 err_server_data:
1340         target_recovery_fini(obd);
1341         filter_free_server_data(filter);
1342 err_filp:
1343         if (filp_close(filter->fo_rcvd_filp, 0))
1344                 CERROR("can't close %s after error\n", LAST_RCVD);
1345         filter->fo_rcvd_filp = NULL;
1346         goto out;
1347 }
1348
1349 /* cleanup the filter: write last used object id to status file */
1350 static void filter_post(struct obd_device *obd)
1351 {
1352         struct lvfs_run_ctxt saved;
1353         struct filter_obd *filter = &obd->u.filter;
1354         int rc, i;
1355
1356         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1357          * best to start a transaction with h_sync, because we removed this
1358          * from lastobjid */
1359
1360         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1361         rc = filter_update_server_data(obd);
1362         if (rc)
1363                 CERROR("error writing server data: rc = %d\n", rc);
1364
1365         for (i = 0; i < filter->fo_group_count; i++) {
1366                 rc = filter_update_last_objid(obd, i,
1367                                 (i == filter->fo_group_count - 1));
1368                 if (rc)
1369                         CERROR("error writing group %d lastobjid: rc = %d\n",
1370                                i, rc);
1371         }
1372
1373         rc = filp_close(filter->fo_rcvd_filp, 0);
1374         filter->fo_rcvd_filp = NULL;
1375         if (rc)
1376                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1377
1378         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1379         filter->fo_obt.obt_health_check_filp = NULL;
1380         if (rc)
1381                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1382
1383         filter_cleanup_groups(obd);
1384         filter_free_server_data(filter);
1385         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1386
1387         filter_free_capa_keys(filter);
1388         cleanup_capa_hash(filter->fo_capa_hash);
1389 }
1390
1391 static void filter_set_last_id(struct filter_obd *filter,
1392                                obd_id id, obd_gr group)
1393 {
1394         LASSERT(group <= filter->fo_group_count);
1395
1396         cfs_spin_lock(&filter->fo_objidlock);
1397         filter->fo_last_objids[group] = id;
1398         cfs_spin_unlock(&filter->fo_objidlock);
1399 }
1400
1401 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1402 {
1403         obd_id id;
1404         LASSERT(group <= filter->fo_group_count);
1405         LASSERT(filter->fo_last_objids != NULL);
1406
1407         /* FIXME: object groups */
1408         cfs_spin_lock(&filter->fo_objidlock);
1409         id = filter->fo_last_objids[group];
1410         cfs_spin_unlock(&filter->fo_objidlock);
1411         return id;
1412 }
1413
1414 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1415 {
1416         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1417         return 0;
1418 }
1419
1420 /* We never dget the object parent, so DON'T dput it either */
1421 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1422 {
1423         struct filter_obd *filter = &obd->u.filter;
1424         struct filter_subdirs *subdirs;
1425         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1426
1427         if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1428                 return filter->fo_dentry_O_groups[group];
1429
1430         subdirs = &filter->fo_dentry_O_sub[group];
1431         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1432 }
1433
1434 /* We never dget the object parent, so DON'T dput it either */
1435 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1436                                   obd_id objid)
1437 {
1438         unsigned long now = jiffies;
1439         struct dentry *dparent = filter_parent(obd, group, objid);
1440         int rc;
1441
1442         if (IS_ERR(dparent))
1443                 return dparent;
1444         if (dparent == NULL)
1445                 return ERR_PTR(-ENOENT);
1446
1447         rc = filter_lock_dentry(obd, dparent);
1448         fsfilt_check_slow(obd, now, "parent lock");
1449         return rc ? ERR_PTR(rc) : dparent;
1450 }
1451
1452 /* We never dget the object parent, so DON'T dput it either */
1453 static void filter_parent_unlock(struct dentry *dparent)
1454 {
1455         UNLOCK_INODE_MUTEX(dparent->d_inode);
1456 }
1457
1458 /* How to get files, dentries, inodes from object id's.
1459  *
1460  * If dir_dentry is passed, the caller has already locked the parent
1461  * appropriately for this operation (normally a write lock).  If
1462  * dir_dentry is NULL, we do a read lock while we do the lookup to
1463  * avoid races with create/destroy and such changing the directory
1464  * internal to the filesystem code. */
1465 struct dentry *filter_fid2dentry(struct obd_device *obd,
1466                                  struct dentry *dir_dentry,
1467                                  obd_gr group, obd_id id)
1468 {
1469         struct dentry *dparent = dir_dentry;
1470         struct dentry *dchild;
1471         char name[32];
1472         int len;
1473         ENTRY;
1474
1475         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1476             obd->u.filter.fo_destroys_in_progress == 0) {
1477                 /* don't fail lookups for orphan recovery, it causes
1478                  * later LBUGs when objects still exist during precreate */
1479                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1480                 RETURN(ERR_PTR(-ENOENT));
1481         }
1482         if (id == 0) {
1483                 CERROR("fatal: invalid object id 0\n");
1484                 RETURN(ERR_PTR(-ESTALE));
1485         }
1486
1487         len = sprintf(name, LPU64, id);
1488         if (dir_dentry == NULL) {
1489                 dparent = filter_parent_lock(obd, group, id);
1490                 if (IS_ERR(dparent)) {
1491                         CERROR("%s: error getting object "LPU64":"LPU64
1492                                " parent: rc %ld\n", obd->obd_name,
1493                                id, group, PTR_ERR(dparent));
1494                         RETURN(dparent);
1495                 }
1496         }
1497         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1498                dparent->d_name.len, dparent->d_name.name, name);
1499         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1500         if (dir_dentry == NULL)
1501                 filter_parent_unlock(dparent);
1502         if (IS_ERR(dchild)) {
1503                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1504                        PTR_ERR(dchild));
1505                 RETURN(dchild);
1506         }
1507
1508         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1509                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1510                        obd->obd_name, id, dchild->d_inode->i_ino);
1511                 f_dput(dchild);
1512                 RETURN(ERR_PTR(-ENOENT));
1513         }
1514
1515         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1516                name, dchild, atomic_read(&dchild->d_count));
1517
1518         LASSERT(atomic_read(&dchild->d_count) > 0);
1519
1520         RETURN(dchild);
1521 }
1522
1523 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1524                                   obd_id group, struct lustre_handle *lockh)
1525 {
1526         int flags = LDLM_AST_DISCARD_DATA, rc;
1527         struct ldlm_res_id res_id;
1528         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1529         ENTRY;
1530
1531         osc_build_res_name(objid, group, &res_id);
1532         /* Tell the clients that the object is gone now and that they should
1533          * throw away any cached pages. */
1534         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1535                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1536                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1537                                     lockh);
1538         if (rc != ELDLM_OK)
1539                 lockh->cookie = 0;
1540         RETURN(rc);
1541 }
1542
1543 static void filter_fini_destroy(struct obd_device *obd,
1544                                 struct lustre_handle *lockh)
1545 {
1546         if (lustre_handle_is_used(lockh))
1547                 ldlm_lock_decref(lockh, LCK_PW);
1548 }
1549
1550 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1551  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1552  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1553  * reverse.  Caller must take i_sem before starting the transaction and we
1554  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1555 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1556                       struct vfsmount *mnt)
1557 {
1558         int rc;
1559         ENTRY;
1560
1561         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1562          * itself we already hold dir->i_mutex for child create/unlink ops */
1563         LASSERT(dentry->d_inode != NULL);
1564         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1565         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1566
1567
1568         /* may_delete() */
1569         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1570                 GOTO(out, rc = -ENOENT);
1571
1572         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1573         if (rc)
1574                 GOTO(out, rc);
1575
1576         if (IS_APPEND(dir))
1577                 GOTO(out, rc = -EPERM);
1578
1579         /* check_sticky() */
1580         if ((dentry->d_inode->i_uid != current->fsuid &&
1581              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1582             IS_IMMUTABLE(dentry->d_inode))
1583                 GOTO(out, rc = -EPERM);
1584
1585         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1586          *       that was done because of journal_start (which is already done
1587          *       here) or some other ordering issue. */
1588         DQUOT_INIT(dir);
1589
1590         rc = ll_security_inode_unlink(dir, dentry, mnt);
1591         if (rc)
1592                 GOTO(out, rc);
1593
1594         rc = dir->i_op->unlink(dir, dentry);
1595 out:
1596         /* need to drop i_mutex before we lose inode reference */
1597         UNLOCK_INODE_MUTEX(dentry->d_inode);
1598         if (rc == 0)
1599                 d_delete(dentry);
1600
1601         RETURN(rc);
1602 }
1603
1604 /* Caller must hold LCK_PW on parent and push us into kernel context.
1605  * Caller must hold child i_mutex, we drop it always.
1606  * Caller is also required to ensure that dchild->d_inode exists. */
1607 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1608                                    obd_gr group, struct dentry *dparent,
1609                                    struct dentry *dchild)
1610 {
1611         struct inode *inode = dchild->d_inode;
1612         struct filter_obd *filter = &obd->u.filter;
1613         int rc;
1614
1615         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1616                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1617                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1618                        (unsigned long)inode->i_nlink,
1619                        atomic_read(&inode->i_count));
1620         }
1621
1622         rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1623         if (rc)
1624                 CERROR("error unlinking objid %.*s: rc %d\n",
1625                        dchild->d_name.len, dchild->d_name.name, rc);
1626         return(rc);
1627 }
1628
1629 struct filter_intent_args {
1630         struct ldlm_lock **victim;
1631         __u64 size;
1632         int *liblustre;
1633 };
1634
1635 static enum interval_iter filter_intent_cb(struct interval_node *n,
1636                                            void *args)
1637 {
1638         struct ldlm_interval *node = (struct ldlm_interval *)n;
1639         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1640         __u64 size = arg->size;
1641         struct ldlm_lock **v = arg->victim;
1642         struct ldlm_lock *lck;
1643
1644         /* If the interval is lower than the current file size,
1645          * just break. */
1646         if (interval_high(n) <= size)
1647                 return INTERVAL_ITER_STOP;
1648
1649         cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1650                 /* Don't send glimpse ASTs to liblustre clients.
1651                  * They aren't listening for them, and they do
1652                  * entirely synchronous I/O anyways. */
1653                 if (lck->l_export == NULL ||
1654                     lck->l_export->exp_libclient == 1)
1655                         continue;
1656
1657                 if (*arg->liblustre)
1658                         *arg->liblustre = 0;
1659
1660                 if (*v == NULL) {
1661                         *v = LDLM_LOCK_GET(lck);
1662                 } else if ((*v)->l_policy_data.l_extent.start <
1663                            lck->l_policy_data.l_extent.start) {
1664                         LDLM_LOCK_RELEASE(*v);
1665                         *v = LDLM_LOCK_GET(lck);
1666                 }
1667
1668                 /* the same policy group - every lock has the
1669                  * same extent, so needn't do it any more */
1670                 break;
1671         }
1672
1673         return INTERVAL_ITER_CONT;
1674 }
1675
1676 static int filter_intent_policy(struct ldlm_namespace *ns,
1677                                 struct ldlm_lock **lockp, void *req_cookie,
1678                                 ldlm_mode_t mode, int flags, void *data)
1679 {
1680         CFS_LIST_HEAD(rpc_list);
1681         struct ptlrpc_request *req = req_cookie;
1682         struct ldlm_lock *lock = *lockp, *l = NULL;
1683         struct ldlm_resource *res = lock->l_resource;
1684         ldlm_processing_policy policy;
1685         struct ost_lvb *res_lvb, *reply_lvb;
1686         struct ldlm_reply *rep;
1687         ldlm_error_t err;
1688         int idx, rc, tmpflags = 0, only_liblustre = 1;
1689         struct ldlm_interval_tree *tree;
1690         struct filter_intent_args arg;
1691         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1692                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1693                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1694         ENTRY;
1695
1696         policy = ldlm_get_processing_policy(res);
1697         LASSERT(policy != NULL);
1698         LASSERT(req != NULL);
1699
1700         rc = lustre_pack_reply(req, 3, repsize, NULL);
1701         if (rc)
1702                 RETURN(req->rq_status = rc);
1703
1704         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1705         LASSERT(rep != NULL);
1706
1707         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1708                                    sizeof(*reply_lvb));
1709         LASSERT(reply_lvb != NULL);
1710
1711         //fixup_handle_for_resent_req(req, lock, &lockh);
1712
1713         /* Call the extent policy function to see if our request can be
1714          * granted, or is blocked.
1715          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1716          * lock, and should not be granted if the lock will be blocked.
1717          */
1718
1719         LASSERT(ns == res->lr_namespace);
1720         lock_res(res);
1721         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1722         check_res_locked(res);
1723
1724         /* FIXME: we should change the policy function slightly, to not make
1725          * this list at all, since we just turn around and free it */
1726         while (!cfs_list_empty(&rpc_list)) {
1727                 struct ldlm_lock *wlock =
1728                         cfs_list_entry(rpc_list.next, struct ldlm_lock,
1729                                        l_cp_ast);
1730                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1731                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1732                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1733                 cfs_list_del_init(&wlock->l_cp_ast);
1734                 LDLM_LOCK_RELEASE(wlock);
1735         }
1736
1737         /* The lock met with no resistance; we're finished. */
1738         if (rc == LDLM_ITER_CONTINUE) {
1739                 /* do not grant locks to the liblustre clients: they cannot
1740                  * handle ASTs robustly.  We need to do this while still
1741                  * holding ns_lock to avoid the lock remaining on the res_link
1742                  * list (and potentially being added to l_pending_list by an
1743                  * AST) when we are going to drop this lock ASAP. */
1744                 if (lock->l_export->exp_libclient ||
1745                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1746                         ldlm_resource_unlink_lock(lock);
1747                         err = ELDLM_LOCK_ABORTED;
1748                 } else {
1749                         err = ELDLM_LOCK_REPLACED;
1750                 }
1751                 unlock_res(res);
1752                 RETURN(err);
1753         }
1754
1755         /* Do not grant any lock, but instead send GL callbacks.  The extent
1756          * policy nicely created a list of all PW locks for us.  We will choose
1757          * the highest of those which are larger than the size in the LVB, if
1758          * any, and perform a glimpse callback. */
1759         res_lvb = res->lr_lvb_data;
1760         LASSERT(res_lvb != NULL);
1761         *reply_lvb = *res_lvb;
1762
1763         /*
1764          * ->ns_lock guarantees that no new locks are granted, and,
1765          * therefore, that res->lr_lvb_data cannot increase beyond the
1766          * end of already granted lock. As a result, it is safe to
1767          * check against "stale" reply_lvb->lvb_size value without
1768          * res->lr_lvb_sem.
1769          */
1770         arg.size = reply_lvb->lvb_size;
1771         arg.victim = &l;
1772         arg.liblustre = &only_liblustre;
1773         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1774                 tree = &res->lr_itree[idx];
1775                 if (tree->lit_mode == LCK_PR)
1776                         continue;
1777
1778                 interval_iterate_reverse(tree->lit_root,
1779                                          filter_intent_cb, &arg);
1780         }
1781         unlock_res(res);
1782
1783         /* There were no PW locks beyond the size in the LVB; finished. */
1784         if (l == NULL) {
1785                 if (only_liblustre) {
1786                         /* If we discovered a liblustre client with a PW lock,
1787                          * however, the LVB may be out of date!  The LVB is
1788                          * updated only on glimpse (which we don't do for
1789                          * liblustre clients) and cancel (which the client
1790                          * obviously has not yet done).  So if it has written
1791                          * data but kept the lock, the LVB is stale and needs
1792                          * to be updated from disk.
1793                          *
1794                          * Of course, this will all disappear when we switch to
1795                          * taking liblustre locks on the OST. */
1796                         ldlm_res_lvbo_update(res, NULL, 1);
1797                 }
1798                 RETURN(ELDLM_LOCK_ABORTED);
1799         }
1800
1801         /*
1802          * This check is for lock taken in filter_prepare_destroy() that does
1803          * not have l_glimpse_ast set. So the logic is: if there is a lock
1804          * with no l_glimpse_ast set, this object is being destroyed already.
1805          *
1806          * Hence, if you are grabbing DLM locks on the server, always set
1807          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1808          */
1809         if (l->l_glimpse_ast == NULL) {
1810                 /* We are racing with unlink(); just return -ENOENT */
1811                 rep->lock_policy_res1 = -ENOENT;
1812                 goto out;
1813         }
1814
1815         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1816         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1817         /* Update the LVB from disk if the AST failed (this is a legal race) */
1818         /*
1819          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1820          * sending ast is not handled. This can result in lost client writes.
1821          */
1822         if (rc != 0)
1823                 ldlm_res_lvbo_update(res, NULL, 1);
1824
1825         lock_res(res);
1826         *reply_lvb = *res_lvb;
1827         unlock_res(res);
1828
1829  out:
1830         LDLM_LOCK_RELEASE(l);
1831
1832         RETURN(ELDLM_LOCK_ABORTED);
1833 }
1834
1835 /*
1836  * per-obd_device iobuf pool.
1837  *
1838  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1839  * allocations in write-path has to be minimized (see bug 5137).
1840  *
1841  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1842  * OST threads (see ost_thread_{init,done}()).
1843  *
1844  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1845  * at the OST layer there are only (potentially) multiple obd_device of type
1846  * unknown at the time of OST thread creation.
1847  *
1848  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1849  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1850  * it's very own iobuf.
1851  *
1852  * Functions below
1853  *
1854  *     filter_kiobuf_pool_init()
1855  *
1856  *     filter_kiobuf_pool_done()
1857  *
1858  *     filter_iobuf_get()
1859  *
1860  * operate on this array. They are "generic" in a sense that they don't depend
1861  * on actual type of iobuf's (the latter depending on Linux kernel version).
1862  */
1863
1864 /*
1865  * destroy pool created by filter_iobuf_pool_init
1866  */
1867 static void filter_iobuf_pool_done(struct filter_obd *filter)
1868 {
1869         struct filter_iobuf **pool;
1870         int i;
1871
1872         ENTRY;
1873
1874         pool = filter->fo_iobuf_pool;
1875         if (pool != NULL) {
1876                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1877                         if (pool[i] != NULL)
1878                                 filter_free_iobuf(pool[i]);
1879                 }
1880                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1881                 filter->fo_iobuf_pool = NULL;
1882         }
1883         EXIT;
1884 }
1885
1886 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1887 {
1888         struct filter_obd       *filter = &obd->u.filter;
1889         struct sptlrpc_rule_set  tmp_rset;
1890         int                      rc;
1891
1892         sptlrpc_rule_set_init(&tmp_rset);
1893         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1894         if (rc) {
1895                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1896                        obd->obd_name, rc);
1897                 return rc;
1898         }
1899
1900         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1901
1902         cfs_write_lock(&filter->fo_sptlrpc_lock);
1903         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1904         filter->fo_sptlrpc_rset = tmp_rset;
1905         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1906
1907         return 0;
1908 }
1909
1910 /*
1911  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1912  */
1913 static int filter_iobuf_pool_init(struct filter_obd *filter)
1914 {
1915         void **pool;
1916
1917         ENTRY;
1918
1919
1920         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1921                       GFP_KERNEL);
1922         if (filter->fo_iobuf_pool == NULL)
1923                 RETURN(-ENOMEM);
1924
1925         filter->fo_iobuf_count = OSS_THREADS_MAX;
1926
1927         RETURN(0);
1928 }
1929
1930 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1931  * many threads there will be so we allocate a large empty array and only
1932  * fill in those slots that are actually in use.
1933  * If we haven't allocated a pool entry for this thread before, do so now. */
1934 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1935 {
1936         int thread_id                    = (oti && oti->oti_thread) ?
1937                                            oti->oti_thread->t_id : -1;
1938         struct filter_iobuf  *pool       = NULL;
1939         struct filter_iobuf **pool_place = NULL;
1940
1941         if (thread_id >= 0) {
1942                 LASSERT(thread_id < filter->fo_iobuf_count);
1943                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1944         }
1945
1946         if (unlikely(pool == NULL)) {
1947                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1948                                           PTLRPC_MAX_BRW_PAGES);
1949                 if (pool_place != NULL)
1950                         *pool_place = pool;
1951         }
1952
1953         return pool;
1954 }
1955
1956 /* mount the file system (secretly).  lustre_cfg parameters are:
1957  * 1 = device
1958  * 2 = fstype
1959  * 3 = flags: failover=f, failout=n
1960  * 4 = mount options
1961  */
1962 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1963                         void *option)
1964 {
1965         struct filter_obd *filter = &obd->u.filter;
1966         struct vfsmount *mnt;
1967         struct lustre_mount_info *lmi;
1968         struct obd_uuid uuid;
1969         __u8 *uuid_ptr;
1970         char *str, *label;
1971         char ns_name[48];
1972         request_queue_t *q;
1973         int rc, i;
1974         ENTRY;
1975
1976         if (lcfg->lcfg_bufcount < 3 ||
1977             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1978             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1979                 RETURN(-EINVAL);
1980
1981         lmi = server_get_mount(obd->obd_name);
1982         if (lmi) {
1983                 /* We already mounted in lustre_fill_super.
1984                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1985                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1986                 mnt = lmi->lmi_mnt;
1987                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1988         } else {
1989                 /* old path - used by lctl */
1990                 CERROR("Using old MDS mount method\n");
1991                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1992                                     MS_NOATIME|MS_NODIRATIME,
1993                                     lustre_cfg_string(lcfg, 1), option);
1994                 if (IS_ERR(mnt)) {
1995                         rc = PTR_ERR(mnt);
1996                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1997                                            lustre_cfg_string(lcfg, 1), rc);
1998                         RETURN(rc);
1999                 }
2000
2001                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2002         }
2003         if (IS_ERR(obd->obd_fsops))
2004                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2005
2006         rc = filter_iobuf_pool_init(filter);
2007         if (rc != 0)
2008                 GOTO(err_ops, rc);
2009
2010         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2011                 CERROR("%s: Underlying device is marked as read-only. "
2012                        "Setup failed\n", obd->obd_name);
2013                 GOTO(err_ops, rc = -EROFS);
2014         }
2015
2016         /* failover is the default */
2017         obd->obd_replayable = 1;
2018
2019         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2020                 str = lustre_cfg_string(lcfg, 3);
2021                 if (strchr(str, 'n')) {
2022                         CWARN("%s: recovery disabled\n", obd->obd_name);
2023                         obd->obd_replayable = 0;
2024                 }
2025         }
2026
2027         filter->fo_vfsmnt = mnt;
2028         obd->u.obt.obt_sb = mnt->mnt_sb;
2029         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2030         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2031
2032         fsfilt_setup(obd, obd->u.obt.obt_sb);
2033
2034         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2035         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2036         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2037         obd->obd_lvfs_ctxt.fs = get_ds();
2038         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2039
2040         cfs_init_mutex(&filter->fo_init_lock);
2041         filter->fo_committed_group = 0;
2042         filter->fo_destroys_in_progress = 0;
2043         for (i = 0; i < 32; i++)
2044                 cfs_sema_init(&filter->fo_create_locks[i], 1);
2045
2046         cfs_spin_lock_init(&filter->fo_objidlock);
2047         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2048         cfs_sema_init(&filter->fo_alloc_lock, 1);
2049         init_brw_stats(&filter->fo_filter_stats);
2050         filter->fo_read_cache = 1; /* enable read-only cache by default */
2051         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2052         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2053         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2054         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2055
2056         rc = filter_prep(obd);
2057         if (rc)
2058                 GOTO(err_ops, rc);
2059
2060         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2061         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2062
2063         filter->fo_fl_oss_capa = 1;
2064
2065         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2066         filter->fo_capa_hash = init_capa_hash();
2067         if (filter->fo_capa_hash == NULL)
2068                 GOTO(err_post, rc = -ENOMEM);
2069
2070         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2071         obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2072                                                 LDLM_NAMESPACE_GREEDY);
2073         if (obd->obd_namespace == NULL)
2074                 GOTO(err_post, rc = -ENOMEM);
2075         obd->obd_namespace->ns_lvbp = obd;
2076         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2077         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2078
2079         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2080                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2081
2082         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2083         if (rc) {
2084                 CERROR("failed to setup llogging subsystems\n");
2085                 GOTO(err_post, rc);
2086         }
2087
2088         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2089         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2090         /* do this after llog being initialized */
2091         filter_adapt_sptlrpc_conf(obd, 1);
2092
2093         rc = lquota_setup(filter_quota_interface_ref, obd);
2094         if (rc)
2095                 GOTO(err_post, rc);
2096
2097         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2098         if (q->max_sectors < q->max_hw_sectors &&
2099             q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2100                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2101                               "for larger I/O requests: max_sectors = %u "
2102                               "could be up to max_hw_sectors=%u\n",
2103                               obd->obd_name, mnt->mnt_sb->s_id,
2104                               q->max_sectors, q->max_hw_sectors);
2105
2106         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2107         if (uuid_ptr != NULL) {
2108                 class_uuid_unparse(uuid_ptr, &uuid);
2109                 str = uuid.uuid;
2110         } else {
2111                 str = "no UUID";
2112         }
2113
2114         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2115         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2116                       obd->obd_name, label ?: str, lmi ? "on " : "",
2117                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2118                       obd->obd_replayable ? "enabled" : "disabled");
2119
2120         if (obd->obd_recovering)
2121                 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2122                               "or until %d client%s reconnect%s\n",
2123                               obd->obd_name,
2124                               obd->obd_recovery_timeout / 60,
2125                               obd->obd_recovery_timeout % 60,
2126                               obd->obd_max_recoverable_clients,
2127                               (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2128                               (obd->obd_max_recoverable_clients == 1) ? "s": "");
2129
2130
2131         RETURN(0);
2132
2133 err_post:
2134         filter_post(obd);
2135 err_ops:
2136         fsfilt_put_ops(obd->obd_fsops);
2137         filter_iobuf_pool_done(filter);
2138 err_mntput:
2139         server_put_mount(obd->obd_name, mnt);
2140         obd->u.obt.obt_sb = 0;
2141         return rc;
2142 }
2143
2144 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2145 {
2146         struct lprocfs_static_vars lvars;
2147         unsigned long addr;
2148         struct page *page;
2149         int rc;
2150
2151         CLASSERT(offsetof(struct obd_device, u.obt) ==
2152                  offsetof(struct obd_device, u.filter.fo_obt));
2153
2154         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2155                 RETURN(-EINVAL);
2156
2157         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2158         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2159         if (!page)
2160                 RETURN(-ENOMEM);
2161         addr = (unsigned long)cfs_page_address(page);
2162         clear_page((void *)addr);
2163
2164         /* lprocfs must be setup before the filter so state can be safely added
2165          * to /proc incrementally as the filter is setup */
2166         lprocfs_filter_init_vars(&lvars);
2167         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2168             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2169                 /* Init obdfilter private stats here */
2170                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2171                                      LPROCFS_CNTR_AVGMINMAX,
2172                                      "read_bytes", "bytes");
2173                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2174                                      LPROCFS_CNTR_AVGMINMAX,
2175                                      "write_bytes", "bytes");
2176                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2177                                      LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2178                                      "get_page", "usec");
2179                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2180                                      LPROCFS_CNTR_AVGMINMAX,
2181                                      "get_page_failures", "num");
2182                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2183                                      LPROCFS_CNTR_AVGMINMAX,
2184                                      "cache_access", "pages");
2185                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2186                                      LPROCFS_CNTR_AVGMINMAX,
2187                                      "cache_hit", "pages");
2188                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2189                                      LPROCFS_CNTR_AVGMINMAX,
2190                                      "cache_miss", "pages");
2191
2192                 lproc_filter_attach_seqstat(obd);
2193                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2194                                                         obd->obd_proc_entry,
2195                                                         NULL, NULL);
2196                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2197                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2198                         CERROR("error %d setting up lprocfs for %s\n",
2199                                rc, "exports");
2200                         obd->obd_proc_exports_entry = NULL;
2201                 }
2202         }
2203         if (obd->obd_proc_exports_entry)
2204                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2205                                    lprocfs_nid_stats_clear_read,
2206                                    lprocfs_nid_stats_clear_write, obd, NULL);
2207
2208         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2209                LUSTRE_CFG_BUFLEN(lcfg, 4));
2210         rc = filter_common_setup(obd, lcfg, (void *)addr);
2211         OBD_PAGE_FREE(page);
2212
2213         if (rc) {
2214                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2215                 lprocfs_free_per_client_stats(obd);
2216                 lprocfs_free_obd_stats(obd);
2217                 lprocfs_obd_cleanup(obd);
2218         }
2219
2220         return rc;
2221 }
2222
2223 static struct llog_operations filter_mds_ost_repl_logops;
2224
2225 static struct llog_operations filter_size_orig_logops = {
2226         .lop_setup   = llog_obd_origin_setup,
2227         .lop_cleanup = llog_obd_origin_cleanup,
2228         .lop_add     = llog_obd_origin_add
2229 };
2230
2231 static int filter_olg_fini(struct obd_llog_group *olg)
2232 {
2233         struct llog_ctxt *ctxt;
2234         int rc = 0, rc2 = 0;
2235         ENTRY;
2236
2237         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2238         if (ctxt)
2239                 rc = llog_cleanup(ctxt);
2240
2241         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2242         if (ctxt) {
2243                 rc2 = llog_cleanup(ctxt);
2244                 if (!rc)
2245                         rc = rc2;
2246         }
2247
2248         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2249         if (ctxt) {
2250                 rc2 = llog_cleanup(ctxt);
2251                 if (!rc)
2252                         rc = rc2;
2253         }
2254
2255         RETURN(rc);
2256 }
2257
2258 static int
2259 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2260                 struct obd_device *tgt)
2261 {
2262         int rc;
2263         ENTRY;
2264
2265         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2266                         &filter_mds_ost_repl_logops);
2267         if (rc)
2268                 GOTO(cleanup, rc);
2269
2270         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2271                         &filter_size_orig_logops);
2272         if (rc)
2273                 GOTO(cleanup, rc);
2274         EXIT;
2275 cleanup:
2276         if (rc)
2277                 filter_olg_fini(olg);
2278         return rc;
2279 }
2280
2281 /**
2282  * Init the default olg, which is embeded in the obd_device, for filter.
2283  */
2284 static int
2285 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2286                         struct obd_device *tgt)
2287 {
2288         struct filter_obd *filter = &obd->u.filter;
2289         struct llog_ctxt *ctxt;
2290         int rc;
2291         ENTRY;
2292
2293         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2294         if (!filter->fo_lcm)
2295                 RETURN(-ENOMEM);
2296
2297         filter_mds_ost_repl_logops = llog_client_ops;
2298         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2299         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2300         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2301
2302         rc = filter_olg_init(obd, olg, tgt);
2303         if (rc)
2304                 GOTO(cleanup_lcm, rc);
2305
2306         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2307                         &llog_lvfs_ops);
2308         if (rc)
2309                 GOTO(cleanup_olg, rc);
2310
2311         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2312         if (!ctxt) {
2313                 CERROR("Can't get ctxt for %p:%x\n", olg,
2314                        LLOG_MDS_OST_REPL_CTXT);
2315                 GOTO(cleanup_olg, rc = -ENODEV);
2316         }
2317         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2318         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2319         llog_ctxt_put(ctxt);
2320
2321         RETURN(0);
2322 cleanup_olg:
2323         filter_olg_fini(olg);
2324 cleanup_lcm:
2325         llog_recov_thread_fini(filter->fo_lcm, 1);
2326         filter->fo_lcm = NULL;
2327         return rc;
2328 }
2329
2330 static int
2331 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2332                  struct obd_device *tgt, int *index)
2333 {
2334         struct filter_obd *filter = &obd->u.filter;
2335         struct llog_ctxt *ctxt;
2336         int rc;
2337         ENTRY;
2338
2339         LASSERT(olg != NULL);
2340         if (olg == &obd->obd_olg)
2341                 return filter_default_olg_init(obd, olg, tgt);
2342
2343         LASSERT(filter->fo_lcm != NULL);
2344         rc = filter_olg_init(obd, olg, tgt);
2345         if (rc)
2346                 RETURN(rc);
2347         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2348         if (!ctxt) {
2349                 CERROR("Can't get ctxt for %p:%x\n", olg,
2350                        LLOG_MDS_OST_REPL_CTXT);
2351                 filter_olg_fini(olg);
2352                 RETURN(-ENODEV);
2353         }
2354         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2355         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2356         llog_ctxt_put(ctxt);
2357         RETURN(rc);
2358 }
2359
2360 static int filter_llog_finish(struct obd_device *obd, int count)
2361 {
2362         struct filter_obd *filter = &obd->u.filter;
2363         struct llog_ctxt *ctxt;
2364         ENTRY;
2365
2366         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2367         if (ctxt) {
2368                 /*
2369                  * Make sure that no cached llcds left in recov_thread.
2370                  * We actually do sync in disconnect time, but disconnect
2371                  * may not come being marked rq_no_resend = 1.
2372                  */
2373                 llog_sync(ctxt, NULL);
2374
2375                 /*
2376                  * Balance class_import_get() in llog_receptor_accept().
2377                  * This is safe to do, as llog is already synchronized
2378                  * and its import may go.
2379                  */
2380                 cfs_mutex_down(&ctxt->loc_sem);
2381                 if (ctxt->loc_imp) {
2382                         class_import_put(ctxt->loc_imp);
2383                         ctxt->loc_imp = NULL;
2384                 }
2385                 cfs_mutex_up(&ctxt->loc_sem);
2386                 llog_ctxt_put(ctxt);
2387         }
2388
2389         if (filter->fo_lcm) {
2390                 cfs_mutex_down(&ctxt->loc_sem);
2391                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2392                 filter->fo_lcm = NULL;
2393                 cfs_mutex_up(&ctxt->loc_sem);
2394         }
2395         RETURN(filter_olg_fini(&obd->obd_olg));
2396 }
2397
2398 /**
2399  * Find the group llog according to group index in the llog group list.
2400  */
2401 static struct obd_llog_group *
2402 filter_find_olg_internal(struct filter_obd *filter, int group)
2403 {
2404         struct obd_llog_group *olg;
2405
2406         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2407         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2408                 if (olg->olg_group == group)
2409                         RETURN(olg);
2410         }
2411         RETURN(NULL);
2412 }
2413
2414 /**
2415  * Find the group llog according to group index on the filter
2416  */
2417 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2418 {
2419         struct obd_llog_group *olg = NULL;
2420         struct filter_obd *filter;
2421
2422         filter = &obd->u.filter;
2423
2424         if (group == FILTER_GROUP_LLOG)
2425                 RETURN(&obd->obd_olg);
2426
2427         cfs_spin_lock(&filter->fo_llog_list_lock);
2428         olg = filter_find_olg_internal(filter, group);
2429         cfs_spin_unlock(&filter->fo_llog_list_lock);
2430
2431         RETURN(olg);
2432 }
2433 /**
2434  * Find the llog_group of the filter according to the group. If it can not
2435  * find, create the llog_group, which only happens when mds is being synced
2436  * with OST.
2437  */
2438 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2439 {
2440         struct obd_llog_group *olg = NULL;
2441         struct filter_obd *filter;
2442         int rc;
2443
2444         filter = &obd->u.filter;
2445
2446         if (group == FILTER_GROUP_LLOG)
2447                 RETURN(&obd->obd_olg);
2448
2449         cfs_spin_lock(&filter->fo_llog_list_lock);
2450         olg = filter_find_olg_internal(filter, group);
2451         if (olg) {
2452                 if (olg->olg_initializing) {
2453                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2454                 } else {
2455                         GOTO(out_unlock, olg);
2456                 }
2457         }
2458         OBD_ALLOC_PTR(olg);
2459         if (olg == NULL)
2460                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2461
2462         llog_group_init(olg, group);
2463         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2464         olg->olg_initializing = 1;
2465         cfs_spin_unlock(&filter->fo_llog_list_lock);
2466
2467         rc = obd_llog_init(obd, olg, obd, NULL);
2468         if (rc) {
2469                cfs_spin_lock(&filter->fo_llog_list_lock);
2470                cfs_list_del(&olg->olg_list);
2471                cfs_spin_unlock(&filter->fo_llog_list_lock);
2472                OBD_FREE_PTR(olg);
2473                GOTO(out, olg = ERR_PTR(-ENOMEM));
2474         }
2475         cfs_spin_lock(&filter->fo_llog_list_lock);
2476         olg->olg_initializing = 0;
2477         cfs_spin_unlock(&filter->fo_llog_list_lock);
2478         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2479               obd->obd_name, group, olg);
2480 out:
2481         RETURN(olg);
2482
2483 out_unlock:
2484         cfs_spin_unlock(&filter->fo_llog_list_lock);
2485         GOTO(out, olg);
2486 }
2487
2488 static int filter_llog_connect(struct obd_export *exp,
2489                                struct llogd_conn_body *body)
2490 {
2491         struct obd_device *obd = exp->exp_obd;
2492         struct llog_ctxt *ctxt;
2493         struct obd_llog_group *olg;
2494         int rc;
2495         ENTRY;
2496
2497         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2498                obd->obd_name, body->lgdc_logid.lgl_oid,
2499                body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2500
2501         olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2502         if (!olg) {
2503                 CERROR(" %s: can not find olg of group %d\n",
2504                        obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2505                 RETURN(-ENOENT);
2506         }
2507         llog_group_set_export(olg, exp);
2508
2509         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2510         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2511                  body->lgdc_ctxt_idx);
2512
2513         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2514               obd->obd_name, body->lgdc_logid.lgl_oid,
2515               body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2516
2517         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2518         obd->u.filter.fo_mds_ost_sync = 1;
2519         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2520         rc = llog_connect(ctxt, &body->lgdc_logid,
2521                           &body->lgdc_gen, NULL);
2522         llog_ctxt_put(ctxt);
2523         if (rc != 0)
2524                 CERROR("failed to connect rc %d idx %d\n", rc,
2525                                 body->lgdc_ctxt_idx);
2526
2527         RETURN(rc);
2528 }
2529
2530 static int filter_llog_preclean(struct obd_device *obd)
2531 {
2532         struct obd_llog_group *olg, *tmp;
2533         struct filter_obd *filter;
2534         cfs_list_t  remove_list;
2535         int rc = 0;
2536         ENTRY;
2537
2538         rc = obd_llog_finish(obd, 0);
2539         if (rc)
2540                 CERROR("failed to cleanup llogging subsystem\n");
2541
2542         filter = &obd->u.filter;
2543         CFS_INIT_LIST_HEAD(&remove_list);
2544
2545         cfs_spin_lock(&filter->fo_llog_list_lock);
2546         while (!cfs_list_empty(&filter->fo_llog_list)) {
2547                 olg = cfs_list_entry(filter->fo_llog_list.next,
2548                                      struct obd_llog_group, olg_list);
2549                 cfs_list_del(&olg->olg_list);
2550                 cfs_list_add(&olg->olg_list, &remove_list);
2551         }
2552         cfs_spin_unlock(&filter->fo_llog_list_lock);
2553
2554         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2555                 cfs_list_del_init(&olg->olg_list);
2556                 rc = filter_olg_fini(olg);
2557                 if (rc)
2558                         CERROR("failed to cleanup llogging subsystem for %u\n",
2559                                olg->olg_group);
2560                 OBD_FREE_PTR(olg);
2561         }
2562
2563         RETURN(rc);
2564 }
2565
2566 static int filter_precleanup(struct obd_device *obd,
2567                              enum obd_cleanup_stage stage)
2568 {
2569         int rc = 0;
2570         ENTRY;
2571
2572         switch(stage) {
2573         case OBD_CLEANUP_EARLY:
2574                 break;
2575         case OBD_CLEANUP_EXPORTS:
2576                 /* Stop recovery before namespace cleanup. */
2577                 target_recovery_fini(obd);
2578                 rc = filter_llog_preclean(obd);
2579                 break;
2580         }
2581         RETURN(rc);
2582 }
2583
2584 static int filter_cleanup(struct obd_device *obd)
2585 {
2586         struct filter_obd *filter = &obd->u.filter;
2587         ENTRY;
2588
2589         if (obd->obd_fail)
2590                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2591                               "will be preserved.\n", obd->obd_name);
2592
2593         obd_exports_barrier(obd);
2594         obd_zombie_barrier();
2595
2596         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2597         lprocfs_free_per_client_stats(obd);
2598         lprocfs_free_obd_stats(obd);
2599         lprocfs_obd_cleanup(obd);
2600         lquota_cleanup(filter_quota_interface_ref, obd);
2601
2602         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2603         obd->obd_namespace = NULL;
2604
2605         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2606
2607         if (obd->u.obt.obt_sb == NULL)
2608                 RETURN(0);
2609
2610         filter_post(obd);
2611
2612         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2613         shrink_dcache_sb(obd->u.obt.obt_sb);
2614
2615         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2616         obd->u.obt.obt_sb = NULL;
2617
2618         fsfilt_put_ops(obd->obd_fsops);
2619
2620         filter_iobuf_pool_done(filter);
2621
2622         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2623
2624         RETURN(0);
2625 }
2626
2627 static int filter_connect_internal(struct obd_export *exp,
2628                                    struct obd_connect_data *data,
2629                                    int reconnect)
2630 {
2631         struct filter_export_data *fed = &exp->exp_filter_data;
2632
2633         if (!data)
2634                 RETURN(0);
2635
2636         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2637                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2638                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2639                data->ocd_connect_flags, data->ocd_version,
2640                data->ocd_grant, data->ocd_index);
2641
2642         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2643                 CWARN("!!! This export (nid %s) used object group %d "
2644                        "earlier; now it's trying to use group %d!  This could "
2645                        "be a bug in the MDS. Please report to "
2646                        "http://bugzilla.lustre.org/\n",
2647                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2648                 RETURN(-EPROTO);
2649         }
2650         fed->fed_group = data->ocd_group;
2651
2652         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2653         exp->exp_connect_flags = data->ocd_connect_flags;
2654         data->ocd_version = LUSTRE_VERSION_CODE;
2655
2656         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2657         if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2658                   data->ocd_connect_flags & OBD_CONNECT_MDS))
2659                 RETURN(-EPROTO);
2660
2661         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2662                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2663                 obd_size left, want;
2664
2665                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2666                 left = filter_grant_space_left(exp);
2667                 want = data->ocd_grant;
2668                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2669                 data->ocd_grant = fed->fed_grant;
2670                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2671
2672                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2673                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2674                        exp->exp_client_uuid.uuid, exp,
2675                        data->ocd_grant, want, left);
2676
2677                 filter->fo_tot_granted_clients ++;
2678         }
2679
2680         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2681                 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2682                 int index = le32_to_cpu(lsd->lsd_ost_index);
2683
2684                 if (!(lsd->lsd_feature_compat &
2685                       cpu_to_le32(OBD_COMPAT_OST))) {
2686                         /* this will only happen on the first connect */
2687                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2688                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2689                         /* sync is not needed here as filter_client_add will
2690                          * set exp_need_sync flag */
2691                         filter_update_server_data(exp->exp_obd);
2692                 } else if (index != data->ocd_index) {
2693                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2694                                            " %u doesn't match actual OST index"
2695                                            " %u in last_rcvd file, bad "
2696                                            "configuration?\n",
2697                                            obd_export_nid2str(exp), index,
2698                                            data->ocd_index);
2699                         RETURN(-EBADF);
2700                 }
2701                 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2702                  * FIXME: We don't strictly need the COMPAT flag for that,
2703                  * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2704                  * FIXME: We needed it for the index, as index 0 is valid. */
2705         }
2706
2707         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2708                 data->ocd_brw_size = 65536;
2709         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2710                 data->ocd_brw_size = min(data->ocd_brw_size,
2711                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2712                 LASSERT(data->ocd_brw_size);
2713         }
2714
2715         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2716                 __u32 cksum_types = data->ocd_cksum_types;
2717
2718                 /* The client set in ocd_cksum_types the checksum types it
2719                  * supports. We have to mask off the algorithms that we don't
2720                  * support */
2721                 if (cksum_types & OBD_CKSUM_ALL)
2722                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2723                 else
2724                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2725
2726                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2727                                    "%x\n", exp->exp_obd->obd_name,
2728                                    obd_export_nid2str(exp), cksum_types,
2729                                    data->ocd_cksum_types);
2730         } else {
2731                 /* This client does not support OBD_CONNECT_CKSUM
2732                  * fall back to CRC32 */
2733                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2734                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2735                                    exp->exp_obd->obd_name,
2736                                    obd_export_nid2str(exp));
2737         }
2738
2739         RETURN(0);
2740 }
2741
2742 static int filter_reconnect(const struct lu_env *env,
2743                             struct obd_export *exp, struct obd_device *obd,
2744                             struct obd_uuid *cluuid,
2745                             struct obd_connect_data *data,
2746                             void *localdata)
2747 {
2748         int rc;
2749         ENTRY;
2750
2751         if (exp == NULL || obd == NULL || cluuid == NULL)
2752                 RETURN(-EINVAL);
2753
2754         rc = filter_connect_internal(exp, data, 1);
2755         if (rc == 0)
2756                 filter_export_stats_init(obd, exp, localdata);
2757
2758         RETURN(rc);
2759 }
2760
2761 /* nearly identical to mds_connect */
2762 static int filter_connect(const struct lu_env *env,
2763                           struct obd_export **exp, struct obd_device *obd,
2764                           struct obd_uuid *cluuid,
2765                           struct obd_connect_data *data, void *localdata)
2766 {
2767         struct lvfs_run_ctxt saved;
2768         struct lustre_handle conn = { 0 };
2769         struct obd_export *lexp;
2770         struct filter_export_data *fed;
2771         struct lsd_client_data *lcd = NULL;
2772         __u32 group;
2773         int rc;
2774         ENTRY;
2775
2776         if (exp == NULL || obd == NULL || cluuid == NULL)
2777                 RETURN(-EINVAL);
2778
2779         rc = class_connect(&conn, obd, cluuid);
2780         if (rc)
2781                 RETURN(rc);
2782         lexp = class_conn2export(&conn);
2783         LASSERT(lexp != NULL);
2784
2785         fed = &lexp->exp_filter_data;
2786
2787         rc = filter_connect_internal(lexp, data, 0);
2788         if (rc)
2789                 GOTO(cleanup, rc);
2790
2791         filter_export_stats_init(obd, lexp, localdata);
2792         if (obd->obd_replayable) {
2793                 OBD_ALLOC(lcd, sizeof(*lcd));
2794                 if (!lcd) {
2795                         CERROR("filter: out of memory for client data\n");
2796                         GOTO(cleanup, rc = -ENOMEM);
2797                 }
2798
2799                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2800                 fed->fed_lcd = lcd;
2801                 rc = filter_client_add(obd, lexp, -1);
2802                 if (rc)
2803                         GOTO(cleanup, rc);
2804         }
2805
2806         group = data->ocd_group;
2807
2808         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2809               obd->obd_name, lexp->exp_handle.h_cookie, group);
2810
2811         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2812         rc = filter_read_groups(obd, group, 1);
2813         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2814         if (rc != 0) {
2815                 CERROR("can't read group %u\n", group);
2816                 GOTO(cleanup, rc);
2817         }
2818
2819         GOTO(cleanup, rc);
2820
2821 cleanup:
2822         if (rc) {
2823                 if (lcd) {
2824                         OBD_FREE_PTR(lcd);
2825                         fed->fed_lcd = NULL;
2826                 }
2827                 class_disconnect(lexp);
2828                 lprocfs_exp_cleanup(lexp);
2829                 *exp = NULL;
2830         } else {
2831                 *exp = lexp;
2832         }
2833
2834         RETURN(rc);
2835 }
2836
2837 /* Do extra sanity checks for grant accounting.  We do this at connect,
2838  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2839  * always get rid of it or turn it off when we know accounting is good. */
2840 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2841 {
2842         struct filter_export_data *fed;
2843         struct obd_export *exp;
2844         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2845         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2846         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2847
2848         if (cfs_list_empty(&obd->obd_exports))
2849                 return;
2850
2851         /* We don't want to do this for large machines that do lots of
2852            mounts or unmounts.  It burns... */
2853         if (obd->obd_num_exports > 100)
2854                 return;
2855
2856         cfs_spin_lock(&obd->obd_osfs_lock);
2857         cfs_spin_lock(&obd->obd_dev_lock);
2858         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2859                 int error = 0;
2860                 fed = &exp->exp_filter_data;
2861                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2862                     fed->fed_dirty < 0)
2863                         error = 1;
2864                 if (maxsize > 0) { /* we may not have done a statfs yet */
2865                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2866                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2867                                  exp->exp_client_uuid.uuid, exp,
2868                                  fed->fed_grant, fed->fed_pending, maxsize);
2869                         LASSERTF(fed->fed_dirty <= maxsize,
2870                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2871                                  exp->exp_client_uuid.uuid, exp,
2872                                  fed->fed_dirty, maxsize);
2873                 }
2874                 if (error)
2875                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2876                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2877                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2878                 else
2879                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2880                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2881                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2882                 tot_granted += fed->fed_grant + fed->fed_pending;
2883                 tot_pending += fed->fed_pending;
2884                 tot_dirty += fed->fed_dirty;
2885         }
2886         fo_tot_granted = obd->u.filter.fo_tot_granted;
2887         fo_tot_pending = obd->u.filter.fo_tot_pending;
2888         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2889         cfs_spin_unlock(&obd->obd_dev_lock);
2890         cfs_spin_unlock(&obd->obd_osfs_lock);
2891
2892         /* Do these assertions outside the spinlocks so we don't kill system */
2893         if (tot_granted != fo_tot_granted)
2894                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2895                        func, tot_granted, fo_tot_granted);
2896         if (tot_pending != fo_tot_pending)
2897                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2898                        func, tot_pending, fo_tot_pending);
2899         if (tot_dirty != fo_tot_dirty)
2900                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2901                        func, tot_dirty, fo_tot_dirty);
2902         if (tot_pending > tot_granted)
2903                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2904                        func, tot_pending, tot_granted);
2905         if (tot_granted > maxsize)
2906                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2907                        func, tot_granted, maxsize);
2908         if (tot_dirty > maxsize)
2909                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2910                        func, tot_dirty, maxsize);
2911 }
2912
2913 /* Remove this client from the grant accounting totals.  We also remove
2914  * the export from the obd device under the osfs and dev locks to ensure
2915  * that the filter_grant_sanity_check() calculations are always valid.
2916  * The client should do something similar when it invalidates its import. */
2917 static void filter_grant_discard(struct obd_export *exp)
2918 {
2919         struct obd_device *obd = exp->exp_obd;
2920         struct filter_obd *filter = &obd->u.filter;
2921         struct filter_export_data *fed = &exp->exp_filter_data;
2922
2923         cfs_spin_lock(&obd->obd_osfs_lock);
2924         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2925                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2926                  obd->obd_name, filter->fo_tot_granted,
2927                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2928         filter->fo_tot_granted -= fed->fed_grant;
2929         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2930                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2931                  obd->obd_name, filter->fo_tot_pending,
2932                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2933         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2934         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2935                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2936                  obd->obd_name, filter->fo_tot_dirty,
2937                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2938         filter->fo_tot_dirty -= fed->fed_dirty;
2939         fed->fed_dirty = 0;
2940         fed->fed_grant = 0;
2941
2942         cfs_spin_unlock(&obd->obd_osfs_lock);
2943 }
2944
2945 static int filter_destroy_export(struct obd_export *exp)
2946 {
2947         struct filter_export_data *fed = &exp->exp_filter_data;
2948         ENTRY;
2949
2950         if (fed->fed_pending)
2951                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2952                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2953                        exp, fed->fed_pending);
2954
2955         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2956
2957         target_destroy_export(exp);
2958         ldlm_destroy_export(exp);
2959
2960         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2961                 RETURN(0);
2962
2963         lut_client_free(exp);
2964
2965         if (!exp->exp_obd->obd_replayable)
2966                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2967
2968         filter_grant_discard(exp);
2969         filter_fmd_cleanup(exp);
2970
2971         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2972                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2973                 if (filter->fo_tot_granted_clients > 0)
2974                         filter->fo_tot_granted_clients --;
2975         }
2976
2977         if (!(exp->exp_flags & OBD_OPT_FORCE))
2978                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2979
2980         RETURN(0);
2981 }
2982
2983 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2984 {
2985         struct obd_llog_group *olg_min, *olg;
2986         struct filter_obd *filter;
2987         int worked = -1, group;
2988         struct llog_ctxt *ctxt;
2989         ENTRY;
2990
2991         filter = &obd->u.filter;
2992
2993         /* we can't sync log holding spinlock. also, we do not want to get
2994          * into livelock. so we do following: loop over MDS's exports in
2995          * group order and skip already synced llogs -bzzz */
2996         do {
2997                 /* look for group with min. number, but > worked */
2998                 olg_min = NULL;
2999                 group = 1 << 30;
3000                 cfs_spin_lock(&filter->fo_llog_list_lock);
3001                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3002                         if (olg->olg_group <= worked) {
3003                                 /* this group is already synced */
3004                                 continue;
3005                         }
3006                         if (group < olg->olg_group) {
3007                                 /* we have group with smaller number to sync */
3008                                 continue;
3009                         }
3010                         /* store current minimal group */
3011                         olg_min = olg;
3012                         group = olg->olg_group;
3013                 }
3014                 cfs_spin_unlock(&filter->fo_llog_list_lock);
3015
3016                 if (olg_min == NULL)
3017                         break;
3018
3019                 worked = olg_min->olg_group;
3020                 if (olg_min->olg_exp &&
3021                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3022                         int err;
3023                         ctxt = llog_group_get_ctxt(olg_min,
3024                                                    LLOG_MDS_OST_REPL_CTXT);
3025                         if (ctxt) {
3026                                 err = llog_sync(ctxt, olg_min->olg_exp);
3027                                 llog_ctxt_put(ctxt);
3028                                 if (err) {
3029                                         CERROR("error flushing logs to MDS: "
3030                                                "rc %d\n", err);
3031                                 }
3032                         }
3033                 }
3034         } while (olg_min != NULL);
3035 }
3036
3037 /* Also incredibly similar to mds_disconnect */
3038 static int filter_disconnect(struct obd_export *exp)
3039 {
3040         struct obd_device *obd = exp->exp_obd;
3041         int rc;
3042         ENTRY;
3043
3044         LASSERT(exp);
3045         class_export_get(exp);
3046
3047         if (!(exp->exp_flags & OBD_OPT_FORCE))
3048                 filter_grant_sanity_check(obd, __FUNCTION__);
3049         filter_grant_discard(exp);
3050
3051         /* Flush any remaining cancel messages out to the target */
3052         filter_sync_llogs(obd, exp);
3053
3054         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3055
3056         rc = server_disconnect_export(exp);
3057
3058         if (exp->exp_obd->obd_replayable)
3059                 filter_client_del(exp);
3060         else
3061                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3062
3063         class_export_put(exp);
3064         RETURN(rc);
3065 }
3066
3067 /* reverse import is changed, sync all cancels */
3068 static void filter_revimp_update(struct obd_export *exp)
3069 {
3070         ENTRY;
3071
3072         LASSERT(exp);
3073         class_export_get(exp);
3074
3075         /* flush any remaining cancel messages out to the target */
3076         filter_sync_llogs(exp->exp_obd, exp);
3077         class_export_put(exp);
3078         EXIT;
3079 }
3080
3081 static int filter_ping(struct obd_export *exp)
3082 {
3083         filter_fmd_expire(exp);
3084         return 0;
3085 }
3086
3087 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3088                                   const char *what, int quiet)
3089 {
3090         struct dentry *dchild = NULL;
3091         obd_gr group = 0;
3092
3093         if (oa->o_valid & OBD_MD_FLGROUP)
3094                 group = oa->o_gr;
3095
3096         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3097
3098         if (IS_ERR(dchild)) {
3099                 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3100                        what, group, oa->o_id);
3101                 RETURN(dchild);
3102         }
3103
3104         if (dchild->d_inode == NULL) {
3105                 if (!quiet)
3106                         CERROR("%s: %s on non-existent object: "LPU64"\n",
3107                                obd->obd_name, what, oa->o_id);
3108                 f_dput(dchild);
3109                 RETURN(ERR_PTR(-ENOENT));
3110         }
3111
3112         return dchild;
3113 }
3114
3115 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3116 {
3117         struct dentry *dentry = NULL;
3118         struct obd_device *obd;
3119         int rc = 0;
3120         ENTRY;
3121
3122         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3123         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3124                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3125         if (rc)
3126                 RETURN(rc);
3127
3128         obd = class_exp2obd(exp);
3129         if (obd == NULL) {
3130                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3131                 RETURN(-EINVAL);
3132         }
3133
3134         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3135         if (IS_ERR(dentry))
3136                 RETURN(PTR_ERR(dentry));
3137
3138         /* Limit the valid bits in the return data to what we actually use */
3139         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3140         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3141
3142         f_dput(dentry);
3143         RETURN(rc);
3144 }
3145
3146 /* this should be enabled/disabled in condition to enabled/disabled large
3147  * inodes (fast EAs) in backing store FS. */
3148 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3149                         void *handle, struct obdo *oa)
3150 {
3151         struct obd_device *obd = exp->exp_obd;
3152         int rc = 0;
3153         ENTRY;
3154
3155         if (oa->o_valid & OBD_MD_FLFID) {
3156                 struct filter_fid ff;
3157
3158                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3159                         oa->o_gr = 0;
3160                 /* packing fid and converting it to LE for storing into EA.
3161                  * Here ->o_stripe_idx should be filled by LOV and rest of
3162                  * fields - by client. */
3163                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3164                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3165                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3166                 ff.ff_objid = cpu_to_le64(oa->o_id);
3167                 ff.ff_group = cpu_to_le64(oa->o_gr);
3168
3169                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3170                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3171                        oa->o_generation, oa->o_id, oa->o_gr);
3172
3173                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3174                 if (rc)
3175                         CERROR("store fid in object failed! rc: %d\n", rc);
3176         } else {
3177                 CDEBUG(D_HA, "OSS object without fid info!\n");
3178         }
3179
3180         RETURN(rc);
3181 }
3182
3183 /* this is called from filter_truncate() until we have filter_punch() */
3184 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3185                             struct obdo *oa, struct obd_trans_info *oti)
3186 {
3187         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3188         struct llog_cookie *fcc = NULL;
3189         struct filter_obd *filter;
3190         int rc, err, sync = 0;
3191         loff_t old_size = 0;
3192         unsigned int ia_valid;
3193         struct inode *inode;
3194         struct page *page = NULL;
3195         struct iattr iattr;
3196         void *handle;
3197         ENTRY;
3198
3199         LASSERT(dentry != NULL);
3200         LASSERT(!IS_ERR(dentry));
3201
3202         inode = dentry->d_inode;
3203         LASSERT(inode != NULL);
3204
3205         filter = &exp->exp_obd->u.filter;
3206         iattr_from_obdo(&iattr, oa, oa->o_valid);
3207         ia_valid = iattr.ia_valid;
3208
3209         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3210                 OBD_ALLOC(fcc, sizeof(*fcc));
3211                 if (fcc != NULL)
3212                         *fcc = oa->o_lcookie;
3213         }
3214         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3215                 DQUOT_INIT(inode);
3216                 /* Filter truncates and writes are serialized by
3217                  * i_alloc_sem, see the comment in
3218                  * filter_preprw_write.*/
3219                 if (ia_valid & ATTR_SIZE)
3220                         down_write(&inode->i_alloc_sem);
3221                 LOCK_INODE_MUTEX(inode);
3222                 old_size = i_size_read(inode);
3223         }
3224
3225         /* VBR: version recovery check */
3226         rc = filter_version_get_check(exp, oti, inode);
3227         if (rc)
3228                 GOTO(out_unlock, rc);
3229
3230         /* Let's pin the last page so that ldiskfs_truncate
3231          * should not start GFP_FS allocation. */
3232         if (ia_valid & ATTR_SIZE) {
3233                 page = grab_cache_page(inode->i_mapping,
3234                                        iattr.ia_size >> PAGE_CACHE_SHIFT);
3235                 if (page == NULL)
3236                         GOTO(out_unlock, rc = -ENOMEM);
3237
3238                 unlock_page(page);
3239         }
3240
3241         /* If the inode still has SUID+SGID bits set (see filter_precreate())
3242          * then we will accept the UID+GID sent by the client during write for
3243          * initializing the ownership of this inode.  We only allow this to
3244          * happen once so clear these bits in setattr. In 2.6 kernels it is
3245          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3246          * the flags that are actually being set. */
3247         if (ia_valid & (ATTR_UID | ATTR_GID)) {
3248                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3249                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3250
3251                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3252                         if (!(ia_valid & ATTR_MODE)) {
3253                                 iattr.ia_mode = inode->i_mode;
3254                                 iattr.ia_valid |= ATTR_MODE;
3255                         }
3256                         iattr.ia_mode &= ~S_ISUID;
3257                 }
3258                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3259                         if (!(iattr.ia_valid & ATTR_MODE)) {
3260                                 iattr.ia_mode = inode->i_mode;
3261                                 iattr.ia_valid |= ATTR_MODE;
3262                         }
3263                         iattr.ia_mode &= ~S_ISGID;
3264                 }
3265
3266                 orig_ids[USRQUOTA] = inode->i_uid;
3267                 orig_ids[GRPQUOTA] = inode->i_gid;
3268                 handle = fsfilt_start_log(exp->exp_obd, inode,
3269                                           FSFILT_OP_SETATTR, oti, 1);
3270                 if (IS_ERR(handle))
3271                         GOTO(out_unlock, rc = PTR_ERR(handle));
3272
3273                 /* update inode EA only once when inode is suid bit marked. As
3274                  * on 2.6.x UID and GID may be set separately, we check here
3275                  * only one of them to avoid double setting. */
3276                 if (inode->i_mode & S_ISUID)
3277                         filter_update_fidea(exp, inode, handle, oa);
3278         } else {
3279                 handle = fsfilt_start(exp->exp_obd, inode,
3280                                       FSFILT_OP_SETATTR, oti);
3281                 if (IS_ERR(handle))
3282                         GOTO(out_unlock, rc = PTR_ERR(handle));
3283         }
3284         if (oa->o_valid & OBD_MD_FLFLAGS) {
3285                 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3286                                       FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3287         } else {
3288                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3289                 if (fcc != NULL)
3290                         /* set cancel cookie callback function */
3291                         sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3292                                                      filter_cancel_cookies_cb,
3293                                                      fcc);
3294         }
3295
3296         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3297                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3298
3299         /* The truncate might have used up our transaction credits.  Make
3300          * sure we have one left for the last_rcvd update. */
3301         err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3302
3303         rc = filter_finish_transno(exp, inode, oti, rc, sync);
3304         if (sync) {
3305                 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3306                 fcc = NULL;
3307         }
3308
3309         err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3310         if (err) {
3311                 CERROR("error on commit, err = %d\n", err);
3312                 if (!rc)
3313                         rc = err;
3314         } else {
3315                 fcc = NULL;
3316         }
3317
3318         /* For a partial-page truncate flush the page to disk immediately
3319          * to avoid data corruption during direct disk write. b=17397 */
3320         if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3321             old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3322                 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3323                                                iattr.ia_size + 1);
3324                 if (!rc)
3325                         rc = err;
3326         }
3327
3328         EXIT;
3329
3330 out_unlock:
3331         if (page)
3332                 page_cache_release(page);
3333
3334         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3335                 UNLOCK_INODE_MUTEX(inode);
3336         if (ia_valid & ATTR_SIZE)
3337                 up_write(&inode->i_alloc_sem);
3338         if (fcc)
3339                 OBD_FREE(fcc, sizeof(*fcc));
3340
3341         /* trigger quota release */
3342         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3343                 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3344                 int rc2 = lquota_adjust(filter_quota_interface_ref,
3345                                         exp->exp_obd, cur_ids,
3346                                         orig_ids, rc, FSFILT_OP_SETATTR);
3347                 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3348                        "filter adjust qunit. (rc:%d)\n", rc2);
3349         }
3350         return rc;
3351 }
3352
3353 /* this is called from filter_truncate() until we have filter_punch() */
3354 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3355                    struct obd_trans_info *oti)
3356 {
3357         struct obdo *oa = oinfo->oi_oa;
3358         struct lustre_capa *capa = oinfo_capa(oinfo);
3359         struct ldlm_res_id res_id;
3360         struct filter_mod_data *fmd;
3361         struct lvfs_run_ctxt saved;
3362         struct filter_obd *filter;
3363         struct ldlm_resource *res;
3364         struct dentry *dentry;
3365         __u64 opc = CAPA_OPC_META_WRITE;
3366         int rc;
3367         ENTRY;
3368
3369         if (oa->o_valid & OBD_FL_TRUNC)
3370                 opc |= CAPA_OPC_OSS_TRUNC;
3371
3372         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3373         rc = filter_auth_capa(exp, NULL, oa->o_gr, capa, opc);
3374         if (rc)
3375                 RETURN(rc);
3376
3377         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3378                 rc = filter_capa_fixoa(exp, oa, oa->o_gr, capa);
3379                 if (rc)
3380                         RETURN(rc);
3381         }
3382
3383         osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3384         /* This would be very bad - accidentally truncating a file when
3385          * changing the time or similar - bug 12203. */
3386         if (oa->o_valid & OBD_MD_FLSIZE &&
3387             oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3388                 static char mdsinum[48];
3389
3390                 if (oa->o_valid & OBD_MD_FLFID)
3391                         snprintf(mdsinum, sizeof(mdsinum) - 1,
3392                                  " of inode "LPU64"/%u", oa->o_fid,
3393                                  oa->o_generation);
3394                 else
3395                         mdsinum[0] = '\0';
3396
3397                 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3398                        " %s\n",
3399                        exp->exp_obd->obd_name, obd_export_nid2str(exp),
3400                        oa->o_id, mdsinum);
3401                 RETURN(-EPERM);
3402         }
3403
3404         dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3405         if (IS_ERR(dentry))
3406                 RETURN(PTR_ERR(dentry));
3407
3408         filter = &exp->exp_obd->u.filter;
3409         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3410
3411         /*
3412          * We need to be atomic against a concurrent write
3413          * (which takes the semaphore for reading). fmd_mactime_xid
3414          * checks will have no effect if a write request with lower
3415          * xid starts just before a setattr and finishes later than
3416          * the setattr (see bug 21489, comment 27).
3417          */
3418         if (oa->o_valid &
3419             (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3420                 down_write(&dentry->d_inode->i_alloc_sem);
3421                 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3422                 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3423                         fmd->fmd_mactime_xid = oti->oti_xid;
3424                 filter_fmd_put(exp, fmd);
3425                 up_write(&dentry->d_inode->i_alloc_sem);
3426         }
3427
3428         /* setting objects attributes (including owner/group) */
3429         rc = filter_setattr_internal(exp, dentry, oa, oti);
3430         if (rc)
3431                 GOTO(out_unlock, rc);
3432
3433         res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3434                                 &res_id, LDLM_EXTENT, 0);
3435
3436         if (res != NULL) {
3437                 LDLM_RESOURCE_ADDREF(res);
3438                 rc = ldlm_res_lvbo_update(res, NULL, 0);
3439                 LDLM_RESOURCE_DELREF(res);
3440                 ldlm_resource_putref(res);
3441         }
3442
3443         oa->o_valid = OBD_MD_FLID;
3444
3445         /* Quota release need uid/gid info */
3446         obdo_from_inode(oa, dentry->d_inode,
3447                         FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3448
3449         EXIT;
3450 out_unlock:
3451         f_dput(dentry);
3452         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3453         return rc;
3454 }
3455
3456 /* XXX identical to osc_unpackmd */
3457 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3458                            struct lov_mds_md *lmm, int lmm_bytes)
3459 {
3460         int lsm_size;
3461         ENTRY;
3462
3463         if (lmm != NULL) {
3464                 if (lmm_bytes < sizeof (*lmm)) {
3465                         CERROR("lov_mds_md too small: %d, need %d\n",
3466                                lmm_bytes, (int)sizeof(*lmm));
3467                         RETURN(-EINVAL);
3468                 }
3469                 /* XXX LOV_MAGIC etc check? */
3470
3471                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3472                         CERROR("lov_mds_md: zero lmm_object_id\n");
3473                         RETURN(-EINVAL);
3474                 }
3475         }
3476
3477         lsm_size = lov_stripe_md_size(1);
3478         if (lsmp == NULL)
3479                 RETURN(lsm_size);
3480
3481         if (*lsmp != NULL && lmm == NULL) {
3482                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3483                 OBD_FREE(*lsmp, lsm_size);
3484                 *lsmp = NULL;
3485                 RETURN(0);
3486         }
3487
3488         if (*lsmp == NULL) {
3489                 OBD_ALLOC(*lsmp, lsm_size);
3490                 if (*lsmp == NULL)
3491                         RETURN(-ENOMEM);
3492
3493                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3494                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3495                         OBD_FREE(*lsmp, lsm_size);
3496                         RETURN(-ENOMEM);
3497                 }
3498                 loi_init((*lsmp)->lsm_oinfo[0]);
3499         }
3500
3501         if (lmm != NULL) {
3502                 /* XXX zero *lsmp? */
3503                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3504                 LASSERT((*lsmp)->lsm_object_id);
3505         }
3506
3507         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3508
3509         RETURN(lsm_size);
3510 }
3511
3512 /* caller must hold fo_create_locks[oa->o_gr] */
3513 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3514                                      struct filter_obd *filter)
3515 {
3516         struct obdo doa = { 0 }; /* XXX obdo on stack */
3517         obd_id last, id;
3518         int rc = 0;
3519         int skip_orphan;
3520         ENTRY;
3521
3522         LASSERT(oa);
3523         LASSERT_MDS_GROUP(oa->o_gr);
3524         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3525         LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3526
3527         memset(&doa, 0, sizeof(doa));
3528
3529         doa.o_valid |= OBD_MD_FLGROUP;
3530         doa.o_gr = oa->o_gr;
3531         doa.o_mode = S_IFREG;
3532
3533         if (!cfs_test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3534                 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3535                        exp->exp_obd->obd_name, doa.o_gr);
3536                 RETURN(0);
3537         }
3538
3539         last = filter_last_id(filter, doa.o_gr);
3540
3541         skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3542
3543         CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3544                exp->exp_obd->obd_name, oa->o_id + 1, last,
3545                skip_orphan ? ", orphan objids won't be reused any more." : ".");
3546
3547         for (id = last; id > oa->o_id; id--) {
3548                 doa.o_id = id;
3549                 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3550                 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3551                         CEMERG("error destroying precreate objid "LPU64": %d\n",
3552                                id, rc);
3553
3554                 /* update last_id on disk periodically so that if we restart
3555                  * we don't need to re-scan all of the just-deleted objects. */
3556                 if ((id & 511) == 0 && !skip_orphan) {
3557                         filter_set_last_id(filter, id - 1, doa.o_gr);
3558                         filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3559                 }
3560         }
3561
3562         CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3563                exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3564
3565         if (!skip_orphan) {
3566                 filter_set_last_id(filter, id, doa.o_gr);
3567                 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3568         } else {
3569                 /* don't reuse orphan object, return last used objid */
3570                 oa->o_id = last;
3571                 rc = 0;
3572         }
3573         cfs_clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3574
3575         RETURN(rc);
3576 }
3577
3578 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3579                             obd_gr group, int *num);
3580 /* returns a negative error or a nonnegative number of files to create */
3581 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3582                                    obd_gr group, struct obd_trans_info *oti)
3583 {
3584         struct obd_device *obd = exp->exp_obd;
3585         struct filter_obd *filter = &obd->u.filter;
3586         int diff, rc;
3587         ENTRY;
3588
3589         /* delete orphans request */
3590         if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3591                 obd_id last = filter_last_id(filter, group);
3592
3593                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3594                         CERROR("%s: dropping old orphan cleanup request\n",
3595                                obd->obd_name);
3596                         RETURN(0);
3597                 }
3598                 /* This causes inflight precreates to abort and drop lock */
3599                 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3600                 cfs_down(&filter->fo_create_locks[group]);
3601                 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3602                         CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3603                                exp->exp_obd->obd_name, group);
3604                         cfs_up(&filter->fo_create_locks[group]);
3605                         RETURN(0);
3606                 }
3607                 diff = oa->o_id - last;
3608                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3609                        last, diff);
3610
3611                 if (-diff > OST_MAX_PRECREATE) {
3612                         CERROR("%s: ignoring bogus orphan destroy request: "
3613                                "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3614                                oa->o_id, last);
3615                         /* FIXME: should reset precreate_next_id on MDS */
3616                         GOTO(out, rc = -EINVAL);
3617                 }
3618                 if (diff < 0) {
3619                         rc = filter_destroy_precreated(exp, oa, filter);
3620                         if (rc)
3621                                 CERROR("%s: unable to write lastobjid, but "
3622                                        "orphans were deleted\n", obd->obd_name);
3623                         GOTO(out, rc);
3624                 } else {
3625                         /* XXX: Used by MDS for the first time! */
3626                         cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3627                 }
3628         } else {
3629                 cfs_down(&filter->fo_create_locks[group]);
3630                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3631                         CERROR("%s: dropping old precreate request\n",
3632                                obd->obd_name);
3633                         GOTO(out, rc = 0);
3634                 }
3635                 /* only precreate if group == 0 and o_id is specfied */
3636                 if (!filter_group_is_mds(group) || oa->o_id == 0)
3637                         diff = 1;
3638                 else
3639                         diff = oa->o_id - filter_last_id(filter, group);
3640                 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3641                        filter_last_id(filter, group), diff);
3642
3643                 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3644                          oa->o_id, filter_last_id(filter, group), diff);
3645         }
3646
3647         if (diff > 0) {
3648                 oa->o_id = filter_last_id(&obd->u.filter, group);
3649                 rc = filter_precreate(obd, oa, group, &diff);
3650                 oa->o_id = filter_last_id(&obd->u.filter, group);
3651                 oa->o_gr = group;
3652                 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3653                 GOTO(out, rc);
3654         }
3655         /* else diff == 0 */
3656         GOTO(out, rc = 0);
3657 out:
3658         cfs_up(&filter->fo_create_locks[group]);
3659         return rc;
3660 }
3661
3662 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3663                          __u64 max_age, __u32 flags)
3664 {
3665         struct filter_obd *filter = &obd->u.filter;
3666         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3667         int rc;
3668         ENTRY;
3669
3670         /* at least try to account for cached pages.  its still racey and
3671          * might be under-reporting if clients haven't announced their
3672          * caches with brw recently */
3673         cfs_spin_lock(&obd->obd_osfs_lock);
3674         rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3675         memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3676         cfs_spin_unlock(&obd->obd_osfs_lock);
3677
3678         CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3679                " pending "LPU64" free "LPU64" avail "LPU64"\n",
3680                filter->fo_tot_dirty, filter->fo_tot_granted,