Whamcloud - gitweb
1a9af708615d4a112d7359f0e0ec04c5b8afa276
[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 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
83 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS1_N_BASE
84 static struct lvfs_callback_ops filter_lvfs_ops;
85 cfs_mem_cache_t *ll_fmd_cachep;
86
87 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
88                              void *cb_data, int error)
89 {
90         obd_transno_commit_cb(obd, transno, error);
91 }
92
93 /* Assumes caller has already pushed us into the kernel context. */
94 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
95                           int rc, int force_sync)
96 {
97         struct filter_obd *filter = &exp->exp_obd->u.filter;
98         struct filter_export_data *fed = &exp->exp_filter_data;
99         struct lsd_client_data *lcd = fed->fed_lcd;
100         __u64 last_rcvd;
101         loff_t off;
102         int err, log_pri = D_RPCTRACE;
103
104         /* Propagate error code. */
105         if (rc)
106                 RETURN(rc);
107
108         if (!exp->exp_obd->obd_replayable || oti == NULL)
109                 RETURN(rc);
110
111         /* we don't allocate new transnos for replayed requests */
112         if (oti->oti_transno == 0) {
113                 spin_lock(&filter->fo_translock);
114                 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
115                 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
116                 spin_unlock(&filter->fo_translock);
117                 oti->oti_transno = last_rcvd;
118         } else {
119                 spin_lock(&filter->fo_translock);
120                 last_rcvd = oti->oti_transno;
121                 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
122                         filter->fo_fsd->lsd_last_transno =
123                                 cpu_to_le64(last_rcvd);
124                 spin_unlock(&filter->fo_translock);
125         }
126         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
127
128         /* could get xid from oti, if it's ever needed */
129         lcd->lcd_last_xid = 0;
130
131         off = fed->fed_lr_off;
132         if (off <= 0) {
133                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
134                        fed->fed_lr_idx, fed->fed_lr_off);
135                 err = -EINVAL;
136         } else {
137                 if (!force_sync)
138                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
139                                                            last_rcvd,
140                                                            oti->oti_handle,
141                                                            filter_commit_cb,
142                                                            NULL);
143
144                 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
145                                           lcd, sizeof(*lcd), &off,
146                                           force_sync | exp->exp_need_sync);
147                 if (force_sync)
148                         filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
149         }
150         if (err) {
151                 log_pri = D_ERROR;
152                 if (rc == 0)
153                         rc = err;
154         }
155
156         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
157                last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
158
159         RETURN(rc);
160 }
161
162 void f_dput(struct dentry *dentry)
163 {
164         /* Can't go inside filter_ddelete because it can block */
165         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
166                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
167         LASSERT(atomic_read(&dentry->d_count) > 0);
168
169         dput(dentry);
170 }
171
172 static void init_brw_stats(struct brw_stats *brw_stats)
173 {
174         int i;
175         for (i = 0; i < BRW_LAST; i++)
176                 spin_lock_init(&brw_stats->hist[i].oh_lock);
177 }
178
179 static int lprocfs_init_rw_stats(struct obd_device *obd,
180                                  struct lprocfs_stats **stats)
181 {
182         int num_stats;
183
184         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
185                                                         LPROC_FILTER_LAST - 1;
186         *stats = lprocfs_alloc_stats(num_stats, 0);
187         if (*stats == NULL)
188                 return -ENOMEM;
189
190         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
191         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
192                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
193         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
194                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
195
196         return(0);
197 }
198
199 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
200    plus the procfs overhead :( */
201 static int filter_export_stats_init(struct obd_device *obd,
202                                     struct obd_export *exp,
203                                     void *client_nid)
204 {
205         int rc, newnid = 0;
206         ENTRY;
207
208         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
209                 /* Self-export gets no proc entry */
210                 RETURN(0);
211
212         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
213         if (rc) {
214                 /* Mask error for already created
215                  * /proc entries */
216                 if (rc == -EALREADY)
217                         rc = 0;
218                 RETURN(rc);
219         }
220
221         if (newnid) {
222                 struct nid_stat *tmp = exp->exp_nid_stats;
223                 LASSERT(tmp != NULL);
224
225                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
226                 if (tmp->nid_brw_stats == NULL)
227                         RETURN(-ENOMEM);
228
229                 init_brw_stats(tmp->nid_brw_stats);
230                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
231                                         0644, &filter_per_nid_stats_fops,
232                                         exp->exp_nid_stats);
233                 if (rc)
234                         CWARN("Error adding the brw_stats file\n");
235
236                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
237                 if (rc)
238                         RETURN(rc);
239
240                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
241                                             tmp->nid_stats);
242                 if (rc)
243                         RETURN(rc);
244                 /* Always add in ldlm_stats */
245                 tmp->nid_ldlm_stats = lprocfs_alloc_stats(LDLM_LAST_OPC -
246                                                           LDLM_FIRST_OPC, 0);
247                 if (tmp->nid_ldlm_stats == NULL)
248                         return -ENOMEM;
249
250                 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
251                 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
252                                             tmp->nid_ldlm_stats);
253                 if (rc)
254                         RETURN(rc);
255         }
256
257         RETURN(0);
258 }
259
260 /* Add client data to the FILTER.  We use a bitmap to locate a free space
261  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
262  * Otherwise, we have just read the data from the last_rcvd file and
263  * we know its offset. */
264 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
265                              int cl_idx)
266 {
267         struct filter_obd *filter = &obd->u.filter;
268         struct filter_export_data *fed = &exp->exp_filter_data;
269         unsigned long *bitmap = filter->fo_last_rcvd_slots;
270         int new_client = (cl_idx == -1);
271
272         ENTRY;
273
274         LASSERT(bitmap != NULL);
275         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
276
277         /* Self-export */
278         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
279                 RETURN(0);
280
281         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
282          * there's no need for extra complication here
283          */
284         if (new_client) {
285                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
286         repeat:
287                 if (cl_idx >= LR_MAX_CLIENTS) {
288                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
289                                cl_idx);
290                         RETURN(-EOVERFLOW);
291                 }
292                 if (test_and_set_bit(cl_idx, bitmap)) {
293                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
294                                                     cl_idx);
295                         goto repeat;
296                 }
297         } else {
298                 if (test_and_set_bit(cl_idx, bitmap)) {
299                         CERROR("FILTER client %d: bit already set in bitmap!\n",
300                                cl_idx);
301                         LBUG();
302                 }
303         }
304
305         fed->fed_lr_idx = cl_idx;
306         fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
307                 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
308         LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
309
310         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
311                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
312
313         if (new_client) {
314                 struct lvfs_run_ctxt saved;
315                 loff_t off = fed->fed_lr_off;
316                 int rc;
317                 void *handle;
318
319                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
320                        fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
321
322                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
323                 /* Transaction needed to fix bug 1403 */
324                 handle = fsfilt_start(obd,
325                                       filter->fo_rcvd_filp->f_dentry->d_inode,
326                                       FSFILT_OP_SETATTR, NULL);
327                 if (IS_ERR(handle)) {
328                         rc = PTR_ERR(handle);
329                         CERROR("unable to start transaction: rc %d\n", rc);
330                 } else {
331                         rc = fsfilt_add_journal_cb(obd, 0, handle,
332                                                    target_client_add_cb, exp);
333                         if (rc == 0) {
334                                 spin_lock(&exp->exp_lock);
335                                 exp->exp_need_sync = 1;
336                                 spin_unlock(&exp->exp_lock);
337                         }
338                         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
339                                                  fed->fed_lcd,
340                                                  sizeof(*fed->fed_lcd),
341                                                  &off, rc /* sync if no cb */);
342                         fsfilt_commit(obd,
343                                       filter->fo_rcvd_filp->f_dentry->d_inode,
344                                       handle, 0);
345                 }
346                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
347
348                 if (rc) {
349                         CERROR("error writing %s client idx %u: rc %d\n",
350                                LAST_RCVD, fed->fed_lr_idx, rc);
351                         RETURN(rc);
352                 }
353         }
354         RETURN(0);
355 }
356
357 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
358
359 static int filter_client_free(struct obd_export *exp)
360 {
361         struct filter_export_data *fed = &exp->exp_filter_data;
362         struct filter_obd *filter = &exp->exp_obd->u.filter;
363         struct obd_device *obd = exp->exp_obd;
364         struct lvfs_run_ctxt saved;
365         int rc;
366         loff_t off;
367         ENTRY;
368
369         if (fed->fed_lcd == NULL)
370                 RETURN(0);
371
372         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
373         if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
374                 GOTO(free, 0);
375
376         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
377                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
378
379         LASSERT(filter->fo_last_rcvd_slots != NULL);
380
381         off = fed->fed_lr_off;
382
383         /* Don't clear fed_lr_idx here as it is likely also unset.  At worst
384          * we leak a client slot that will be cleaned on the next recovery. */
385         if (off <= 0) {
386                 CERROR("%s: client idx %d has med_off %lld\n",
387                        obd->obd_name, fed->fed_lr_idx, off);
388                 GOTO(free, rc = -EINVAL);
389         }
390
391         /* Clear the bit _after_ zeroing out the client so we don't
392            race with filter_client_add and zero out new clients.*/
393         if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
394                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
395                        fed->fed_lr_idx);
396                 LBUG();
397         }
398
399         if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
400                 /* Don't force sync on disconnect if aborting recovery,
401                  * or it does num_clients * num_osts.  b=17194 */
402                 int need_sync = (!exp->exp_libclient || exp->exp_need_sync) &&
403                                  !(exp->exp_flags&OBD_OPT_ABORT_RECOV);
404                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
405                 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
406                                          sizeof(zero_lcd), &off, 0);
407
408                 /* Make sure the server's last_transno is up to date. Do this
409                  * after the client is freed so we know all the client's
410                  * transactions have been committed. */
411                 if (rc == 0)
412                         filter_update_server_data(obd, filter->fo_rcvd_filp,
413                                                   filter->fo_fsd, need_sync);
414                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
415
416                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
417                        "zero out client %s at idx %u/%llu in %s %ssync rc %d\n",
418                        fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
419                        LAST_RCVD, need_sync ? "" : "a", rc);
420         }
421
422         if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
423                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
424                        fed->fed_lr_idx);
425                 LBUG();
426         }
427
428         EXIT;
429 free:
430         OBD_FREE_PTR(fed->fed_lcd);
431         fed->fed_lcd = NULL;
432
433         return 0;
434 }
435
436 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
437 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
438                                          struct filter_mod_data *fmd)
439 {
440         LASSERT_SPIN_LOCKED(&fed->fed_lock);
441         if (--fmd->fmd_refcount == 0) {
442                 /* XXX when we have persistent reservations and the handle
443                  * is stored herein we need to drop it here. */
444                 fed->fed_mod_count--;
445                 list_del(&fmd->fmd_list);
446                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
447         }
448 }
449
450 /* drop fmd reference, free it if last ref */
451 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
452 {
453         struct filter_export_data *fed;
454
455         if (fmd == NULL)
456                 return;
457
458         fed = &exp->exp_filter_data;
459         spin_lock(&fed->fed_lock);
460         filter_fmd_put_nolock(fed, fmd); /* caller reference */
461         spin_unlock(&fed->fed_lock);
462 }
463
464 /* expire entries from the end of the list if there are too many
465  * or they are too old */
466 static void filter_fmd_expire_nolock(struct filter_obd *filter,
467                                      struct filter_export_data *fed,
468                                      struct filter_mod_data *keep)
469 {
470         struct filter_mod_data *fmd, *tmp;
471
472         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
473                 if (fmd == keep)
474                         break;
475
476                 if (time_before(jiffies, fmd->fmd_expire) &&
477                     fed->fed_mod_count < filter->fo_fmd_max_num)
478                         break;
479
480                 list_del_init(&fmd->fmd_list);
481                 filter_fmd_put_nolock(fed, fmd); /* list reference */
482         }
483 }
484
485 void filter_fmd_expire(struct obd_export *exp)
486 {
487         spin_lock(&exp->exp_filter_data.fed_lock);
488         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
489                                  &exp->exp_filter_data, NULL);
490         spin_unlock(&exp->exp_filter_data.fed_lock);
491 }
492
493 /* find specified objid, group in export fmd list.
494  * caller must hold fed_lock and take fmd reference itself */
495 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
496                                                 struct filter_export_data *fed,
497                                                 obd_id objid, obd_gr group)
498 {
499         struct filter_mod_data *found = NULL, *fmd;
500
501         LASSERT_SPIN_LOCKED(&fed->fed_lock);
502
503         list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
504                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
505                         found = fmd;
506                         list_del(&fmd->fmd_list);
507                         list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
508                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
509                         break;
510                 }
511         }
512
513         filter_fmd_expire_nolock(filter, fed, found);
514
515         return found;
516 }
517
518 /* Find fmd based on objid and group, or return NULL if not found. */
519 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
520                                         obd_id objid, obd_gr group)
521 {
522         struct filter_mod_data *fmd;
523
524         spin_lock(&exp->exp_filter_data.fed_lock);
525         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
526                                      &exp->exp_filter_data, objid, group);
527         if (fmd)
528                 fmd->fmd_refcount++;    /* caller reference */
529         spin_unlock(&exp->exp_filter_data.fed_lock);
530
531         return fmd;
532 }
533
534 /* Find fmd based on objid and group, or create a new one if none is found.
535  * It is possible for this function to return NULL under memory pressure,
536  * or if objid = 0 is passed (which will only cause old entries to expire).
537  * Currently this is not fatal because any fmd state is transient and
538  * may also be freed when it gets sufficiently old. */
539 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
540                                        obd_id objid, obd_gr group)
541 {
542         struct filter_export_data *fed = &exp->exp_filter_data;
543         struct filter_mod_data *found = NULL, *fmd_new = NULL;
544
545         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
546
547         spin_lock(&fed->fed_lock);
548         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
549         if (fmd_new) {
550                 if (found == NULL) {
551                         list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
552                         fmd_new->fmd_id = objid;
553                         fmd_new->fmd_gr = group;
554                         fmd_new->fmd_refcount++;   /* list reference */
555                         found = fmd_new;
556                         fed->fed_mod_count++;
557                 } else {
558                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
559                 }
560         }
561         if (found) {
562                 found->fmd_refcount++;          /* caller reference */
563                 found->fmd_expire = jiffies +
564                         exp->exp_obd->u.filter.fo_fmd_max_age;
565         }
566
567         spin_unlock(&fed->fed_lock);
568
569         return found;
570 }
571
572 #ifdef DO_FMD_DROP
573 /* drop fmd list reference so it will disappear when last reference is put.
574  * This isn't so critical because it would in fact only affect the one client
575  * that is doing the unlink and at worst we have an stale entry referencing
576  * an object that should never be used again. */
577 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
578 {
579         struct filter_mod_data *found = NULL;
580
581         spin_lock(&exp->exp_filter_data.fed_lock);
582         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
583         if (found) {
584                 list_del_init(&found->fmd_list);
585                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
586         }
587         spin_unlock(&exp->exp_filter_data.fed_lock);
588 }
589 #else
590 #define filter_fmd_drop(exp, objid, group)
591 #endif
592
593 /* remove all entries from fmd list */
594 static void filter_fmd_cleanup(struct obd_export *exp)
595 {
596         struct filter_export_data *fed = &exp->exp_filter_data;
597         struct filter_mod_data *fmd = NULL, *tmp;
598
599         spin_lock(&fed->fed_lock);
600         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
601                 list_del_init(&fmd->fmd_list);
602                 filter_fmd_put_nolock(fed, fmd);
603         }
604         spin_unlock(&fed->fed_lock);
605 }
606
607 static int filter_init_export(struct obd_export *exp)
608 {
609         spin_lock_init(&exp->exp_filter_data.fed_lock);
610         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
611
612         spin_lock(&exp->exp_lock);
613         exp->exp_connecting = 1;
614         spin_unlock(&exp->exp_lock);
615
616         return ldlm_init_export(exp);
617 }
618
619 static int filter_free_server_data(struct filter_obd *filter)
620 {
621         OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
622         filter->fo_fsd = NULL;
623         OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
624         filter->fo_last_rcvd_slots = NULL;
625         return 0;
626 }
627
628 /* assumes caller is already in kernel ctxt */
629 int filter_update_server_data(struct obd_device *obd, struct file *filp,
630                               struct lr_server_data *fsd, int force_sync)
631 {
632         loff_t off = 0;
633         int rc;
634         ENTRY;
635
636         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
637         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
638                le64_to_cpu(fsd->lsd_last_transno));
639         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
640                le64_to_cpu(fsd->lsd_mount_count));
641
642         fsd->lsd_compat14 = fsd->lsd_last_transno;
643         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
644         if (rc)
645                 CERROR("error writing lr_server_data: rc = %d\n", rc);
646
647         RETURN(rc);
648 }
649
650 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
651                              int force_sync)
652 {
653         struct filter_obd *filter = &obd->u.filter;
654         __u64 tmp;
655         loff_t off = 0;
656         int rc;
657         ENTRY;
658
659         if (filter->fo_last_objid_files[group] == NULL) {
660                 CERROR("Object group "LPU64" not fully setup; not updating "
661                        "last_objid\n", group);
662                 RETURN(-EINVAL);
663         }
664
665         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
666                obd->obd_name, group, filter->fo_last_objids[group]);
667
668         tmp = cpu_to_le64(filter->fo_last_objids[group]);
669         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
670                                  &tmp, sizeof(tmp), &off, force_sync);
671         if (rc)
672                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
673                        group, rc);
674         RETURN(rc);
675 }
676 extern int ost_handle(struct ptlrpc_request *req);
677 /* assumes caller has already in kernel ctxt */
678 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
679 {
680         struct filter_obd *filter = &obd->u.filter;
681         struct lr_server_data *fsd;
682         struct lsd_client_data *lcd = NULL;
683         struct inode *inode = filp->f_dentry->d_inode;
684         unsigned long last_rcvd_size = i_size_read(inode);
685         __u64 mount_count;
686         int cl_idx;
687         loff_t off = 0;
688         int rc;
689
690         /* ensure padding in the struct is the correct size */
691         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
692                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
693         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
694                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
695
696         OBD_ALLOC(fsd, sizeof(*fsd));
697         if (!fsd)
698                 RETURN(-ENOMEM);
699         filter->fo_fsd = fsd;
700
701         OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
702         if (filter->fo_last_rcvd_slots == NULL) {
703                 OBD_FREE(fsd, sizeof(*fsd));
704                 RETURN(-ENOMEM);
705         }
706
707         if (last_rcvd_size == 0) {
708                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
709
710                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
711                 fsd->lsd_last_transno = 0;
712                 mount_count = fsd->lsd_mount_count = 0;
713                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
714                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
715                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
716                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
717                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
718                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
719         } else {
720                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
721                 if (rc) {
722                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
723                                LAST_RCVD, rc);
724                         GOTO(err_fsd, rc);
725                 }
726                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
727                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
728                                            "using the wrong disk %s. Were the "
729                                            "/dev/ assignments rearranged?\n",
730                                            obd->obd_uuid.uuid, fsd->lsd_uuid);
731                         GOTO(err_fsd, rc = -EINVAL);
732                 }
733                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
734                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
735                 /* COMPAT_146 */
736                 /* Assume old last_rcvd format unless I_C_LR is set */
737                 if (!(fsd->lsd_feature_incompat &
738                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
739                         fsd->lsd_last_transno = fsd->lsd_compat14;
740                 /* end COMPAT_146 */
741         }
742
743         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
744                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
745                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
746                        ~FILTER_INCOMPAT_SUPP);
747                 GOTO(err_fsd, rc = -EINVAL);
748         }
749         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
750                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
751                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
752                        ~FILTER_ROCOMPAT_SUPP);
753                 /* Do something like remount filesystem read-only */
754                 GOTO(err_fsd, rc = -EINVAL);
755         }
756
757         CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
758                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
759         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
760                obd->obd_name, mount_count + 1);
761         CDEBUG(D_INODE, "%s: server data size: %u\n",
762                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
763         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
764                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
765         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
766                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
767         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
768                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
769         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
770                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
771                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
772                 le16_to_cpu(fsd->lsd_client_size));
773
774         if (!obd->obd_replayable) {
775                 CWARN("%s: recovery support OFF\n", obd->obd_name);
776                 GOTO(out, rc = 0);
777         }
778
779         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
780              off < last_rcvd_size; cl_idx++) {
781                 __u64 last_rcvd;
782                 struct obd_export *exp;
783                 struct filter_export_data *fed;
784
785                 if (!lcd) {
786                         OBD_ALLOC_PTR(lcd);
787                         if (!lcd)
788                                 GOTO(err_client, rc = -ENOMEM);
789                 }
790
791                 /* Don't assume off is incremented properly by
792                  * fsfilt_read_record(), in case sizeof(*lcd)
793                  * isn't the same as fsd->lsd_client_size.  */
794                 off = le32_to_cpu(fsd->lsd_client_start) +
795                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
796                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
797                 if (rc) {
798                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
799                                LAST_RCVD, cl_idx, off, rc);
800                         break; /* read error shouldn't cause startup to fail */
801                 }
802
803                 if (lcd->lcd_uuid[0] == '\0') {
804                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
805                                cl_idx);
806                         continue;
807                 }
808
809                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
810
811                 /* These exports are cleaned up by filter_disconnect(), so they
812                  * need to be set up like real exports as filter_connect() does.
813                  */
814                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
815
816                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
817                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
818                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
819                 if (IS_ERR(exp)) {
820                         if (PTR_ERR(exp) == -EALREADY) {
821                                 /* export already exists, zero out this one */
822                                 CERROR("Zeroing out duplicate export due to "
823                                        "bug 10479.\n");
824                                 lcd->lcd_uuid[0] = '\0';
825                         } else {
826                                 GOTO(err_client, rc = PTR_ERR(exp));
827                         }
828                 } else {
829                         fed = &exp->exp_filter_data;
830                         fed->fed_lcd = lcd;
831                         fed->fed_group = 0; /* will be assigned at connect */
832                         filter_export_stats_init(obd, exp, NULL);
833                         rc = filter_client_add(obd, exp, cl_idx);
834                         /* can't fail for existing client */
835                         LASSERTF(rc == 0, "rc = %d\n", rc);
836
837                         lcd = NULL;
838                         spin_lock(&exp->exp_lock);
839                         exp->exp_connecting = 0;
840                         exp->exp_in_recovery = 0;
841                         spin_unlock(&exp->exp_lock);
842                         obd->obd_max_recoverable_clients++;
843                         class_export_put(exp);
844                 }
845
846                 /* Need to check last_rcvd even for duplicated exports. */
847                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
848                        cl_idx, last_rcvd);
849
850                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
851                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
852         }
853
854         if (lcd)
855                 OBD_FREE_PTR(lcd);
856
857         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
858
859         target_recovery_init(obd, ost_handle);
860
861 out:
862         filter->fo_mount_count = mount_count + 1;
863         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
864
865         /* save it, so mount count and last_transno is current */
866         rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
867         if (rc)
868                 GOTO(err_client, rc);
869
870         RETURN(0);
871
872 err_client:
873         target_recovery_fini(obd);
874 err_fsd:
875         filter_free_server_data(filter);
876         RETURN(rc);
877 }
878
879 static int filter_cleanup_groups(struct obd_device *obd)
880 {
881         struct filter_obd *filter = &obd->u.filter;
882         struct file *filp;
883         struct dentry *dentry;
884         int i, j;
885         ENTRY;
886
887         if (filter->fo_dentry_O_groups != NULL) {
888                 for (i = 0; i < filter->fo_group_count; i++) {
889                         dentry = filter->fo_dentry_O_groups[i];
890                         if (dentry != NULL)
891                                 f_dput(dentry);
892                 }
893                 OBD_FREE(filter->fo_dentry_O_groups,
894                          filter->fo_group_count *
895                          sizeof(*filter->fo_dentry_O_groups));
896                 filter->fo_dentry_O_groups = NULL;
897         }
898         if (filter->fo_last_objid_files != NULL) {
899                 for (i = 0; i < filter->fo_group_count; i++) {
900                         filp = filter->fo_last_objid_files[i];
901                         if (filp != NULL)
902                                 filp_close(filp, 0);
903                 }
904                 OBD_FREE(filter->fo_last_objid_files,
905                          filter->fo_group_count *
906                          sizeof(*filter->fo_last_objid_files));
907                 filter->fo_last_objid_files = NULL;
908         }
909         if (filter->fo_dentry_O_sub != NULL) {
910                 for (i = 0; i < filter->fo_group_count; i++) {
911                         for (j = 0; j < filter->fo_subdir_count; j++) {
912                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
913                                 if (dentry != NULL)
914                                         f_dput(dentry);
915                         }
916                 }
917                 OBD_FREE(filter->fo_dentry_O_sub,
918                          filter->fo_group_count *
919                          sizeof(*filter->fo_dentry_O_sub));
920                 filter->fo_dentry_O_sub = NULL;
921         }
922         if (filter->fo_last_objids != NULL) {
923                 OBD_FREE(filter->fo_last_objids,
924                          filter->fo_group_count *
925                          sizeof(*filter->fo_last_objids));
926                 filter->fo_last_objids = NULL;
927         }
928         if (filter->fo_dentry_O != NULL) {
929                 f_dput(filter->fo_dentry_O);
930                 filter->fo_dentry_O = NULL;
931         }
932         RETURN(0);
933 }
934
935 static int filter_update_last_group(struct obd_device *obd, int group)
936 {
937         struct filter_obd *filter = &obd->u.filter;
938         struct file *filp = NULL;
939         int last_group = 0, rc;
940         loff_t off = 0;
941         ENTRY;
942
943         if (group <= filter->fo_committed_group)
944                 RETURN(0);
945
946         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
947         if (IS_ERR(filp)) {
948                 rc = PTR_ERR(filp);
949                 filp = NULL;
950                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
951                 GOTO(cleanup, rc);
952         }
953
954         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
955         if (rc) {
956                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
957                 GOTO(cleanup, rc);
958         }
959         LASSERTF(off == 0 || CHECK_MDS_GROUP(last_group),
960                  "off = %llu and last_group = %d\n", off, last_group);
961
962         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
963                obd->obd_name, last_group, group);
964
965         off = 0;
966         last_group = group;
967         /* must be sync: bXXXX */
968         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
969         if (rc) {
970                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
971                 GOTO(cleanup, rc);
972         }
973
974         filter->fo_committed_group = group;
975 cleanup:
976         if (filp)
977                 filp_close(filp, 0);
978         RETURN(rc);
979 }
980
981 static int filter_read_group_internal(struct obd_device *obd, int group,
982                                       int create)
983 {
984         struct filter_obd *filter = &obd->u.filter;
985         __u64 *new_objids = NULL;
986         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
987         struct dentry **new_groups = NULL;
988         struct file **new_files = NULL;
989         struct dentry *dentry;
990         struct file *filp;
991         int old_count = filter->fo_group_count, rc, stage = 0, i;
992         char name[25];
993         __u64 last_objid;
994         loff_t off = 0;
995         int len = group + 1;
996
997         snprintf(name, 24, "%d", group);
998         name[24] = '\0';
999
1000         if (!create) {
1001                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1002                                            strlen(name));
1003                 if (IS_ERR(dentry)) {
1004                         CERROR("Cannot lookup expected object group %d: %ld\n",
1005                                group, PTR_ERR(dentry));
1006                         RETURN(PTR_ERR(dentry));
1007                 }
1008         } else {
1009                 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1010                                       name, 0700, 1);
1011                 if (IS_ERR(dentry)) {
1012                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1013                                PTR_ERR(dentry));
1014                         RETURN(PTR_ERR(dentry));
1015                 }
1016         }
1017         stage = 1;
1018
1019         snprintf(name, 24, "O/%d/LAST_ID", group);
1020         name[24] = '\0';
1021         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1022         if (IS_ERR(filp)) {
1023                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1024                 GOTO(cleanup, rc = PTR_ERR(filp));
1025         }
1026         stage = 2;
1027
1028         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1029         if (rc) {
1030                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1031                 GOTO(cleanup, rc);
1032         }
1033
1034         if (filter->fo_subdir_count) {
1035                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1036                 if (tmp_subdirs == NULL)
1037                         GOTO(cleanup, rc = -ENOMEM);
1038                 stage = 3;
1039
1040                 for (i = 0; i < filter->fo_subdir_count; i++) {
1041                         char dir[20];
1042                         snprintf(dir, sizeof(dir), "d%u", i);
1043
1044                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1045                                                               filter->fo_vfsmnt,
1046                                                               dir, 0700, 1);
1047                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1048                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1049                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1050                                        group, dir, rc);
1051                                 GOTO(cleanup, rc);
1052                         }
1053
1054                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1055                                tmp_subdirs->dentry[i]);
1056                 }
1057         }
1058
1059         /* 'group' is an index; we need an array of length 'group + 1' */
1060         if (group + 1 > old_count) {
1061                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1062                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1063                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1064                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1065                 stage = 4;
1066                 if (new_objids == NULL || new_subdirs == NULL ||
1067                     new_groups == NULL || new_files == NULL)
1068                         GOTO(cleanup, rc = -ENOMEM);
1069
1070                 if (old_count) {
1071                         memcpy(new_objids, filter->fo_last_objids,
1072                                old_count * sizeof(*new_objids));
1073                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1074                                old_count * sizeof(*new_subdirs));
1075                         memcpy(new_groups, filter->fo_dentry_O_groups,
1076                                old_count * sizeof(*new_groups));
1077                         memcpy(new_files, filter->fo_last_objid_files,
1078                                old_count * sizeof(*new_files));
1079
1080                         OBD_FREE(filter->fo_last_objids,
1081                                  old_count * sizeof(*new_objids));
1082                         OBD_FREE(filter->fo_dentry_O_sub,
1083                                  old_count * sizeof(*new_subdirs));
1084                         OBD_FREE(filter->fo_dentry_O_groups,
1085                                  old_count * sizeof(*new_groups));
1086                         OBD_FREE(filter->fo_last_objid_files,
1087                                  old_count * sizeof(*new_files));
1088                 }
1089                 filter->fo_last_objids = new_objids;
1090                 filter->fo_dentry_O_sub = new_subdirs;
1091                 filter->fo_dentry_O_groups = new_groups;
1092                 filter->fo_last_objid_files = new_files;
1093                 filter->fo_group_count = len;
1094         }
1095
1096         filter->fo_dentry_O_groups[group] = dentry;
1097         filter->fo_last_objid_files[group] = filp;
1098         if (filter->fo_subdir_count) {
1099                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1100                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1101         }
1102
1103         filter_update_last_group(obd, group);
1104
1105         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1106                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1107                 rc = filter_update_last_objid(obd, group, 1);
1108                 RETURN(rc);
1109         }
1110
1111         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1112         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1113                obd->obd_name, group, last_objid);
1114         RETURN(0);
1115  cleanup:
1116         switch (stage) {
1117         case 4:
1118                 if (new_objids != NULL)
1119                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1120                 if (new_subdirs != NULL)
1121                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1122                 if (new_groups != NULL)
1123                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1124                 if (new_files != NULL)
1125                         OBD_FREE(new_files, len * sizeof(*new_files));
1126         case 3:
1127                 if (filter->fo_subdir_count) {
1128                         for (i = 0; i < filter->fo_subdir_count; i++) {
1129                                 if (tmp_subdirs->dentry[i] != NULL)
1130                                         dput(tmp_subdirs->dentry[i]);
1131                         }
1132                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1133                 }
1134         case 2:
1135                 filp_close(filp, 0);
1136         case 1:
1137                 dput(dentry);
1138         }
1139         RETURN(rc);
1140 }
1141
1142 static int filter_read_groups(struct obd_device *obd, int last_group,
1143                               int create)
1144 {
1145         struct filter_obd *filter = &obd->u.filter;
1146         int old_count, group, rc = 0;
1147
1148         down(&filter->fo_init_lock);
1149         old_count = filter->fo_group_count;
1150         for (group = old_count; group <= last_group; group++) {
1151
1152                 rc = filter_read_group_internal(obd, group, create);
1153                 if (rc != 0)
1154                         break;
1155         }
1156         up(&filter->fo_init_lock);
1157         return rc;
1158 }
1159
1160 /* FIXME: object groups */
1161 static int filter_prep_groups(struct obd_device *obd)
1162 {
1163         struct filter_obd *filter = &obd->u.filter;
1164         struct dentry *dentry, *O_dentry;
1165         struct file *filp;
1166         int    last_group, rc = 0, cleanup_phase = 0;
1167         loff_t off = 0;
1168         ENTRY;
1169
1170         O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1171                                 "O", 0700, 1);
1172         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1173         if (IS_ERR(O_dentry)) {
1174                 rc = PTR_ERR(O_dentry);
1175                 CERROR("cannot open/create O: rc = %d\n", rc);
1176                 GOTO(cleanup, rc);
1177         }
1178         filter->fo_dentry_O = O_dentry;
1179         cleanup_phase = 1; /* O_dentry */
1180
1181         /* Lookup "R" to tell if we're on an old OST FS and need to convert
1182          * from O/R/<dir>/<objid> to O/0/<dir>/<objid>.  This can be removed
1183          * some time post 1.0 when all old-style OSTs have converted along
1184          * with the init_objid hack. */
1185         dentry = ll_lookup_one_len("R", O_dentry, 1);
1186         if (IS_ERR(dentry))
1187                 GOTO(cleanup, rc = PTR_ERR(dentry));
1188         if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1189                 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1190                 ENTRY;
1191
1192                 CWARN("converting OST to new object layout\n");
1193                 if (IS_ERR(O0_dentry)) {
1194                         rc = PTR_ERR(O0_dentry);
1195                         CERROR("error looking up O/0: rc %d\n", rc);
1196                         GOTO(cleanup_R, rc);
1197                 }
1198
1199                 if (O0_dentry->d_inode) {
1200                         CERROR("Both O/R and O/0 exist. Fix manually.\n");
1201                         GOTO(cleanup_O0, rc = -EEXIST);
1202                 }
1203
1204                 LOCK_INODE_MUTEX(O_dentry->d_inode);
1205                 rc = ll_vfs_rename(O_dentry->d_inode, dentry, filter->fo_vfsmnt,
1206                                    O_dentry->d_inode, O0_dentry,
1207                                    filter->fo_vfsmnt);
1208                 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1209
1210                 if (rc) {
1211                         CERROR("error renaming O/R to O/0: rc %d\n", rc);
1212                         GOTO(cleanup_O0, rc);
1213                 }
1214                 filter->fo_fsd->lsd_feature_incompat |=
1215                         cpu_to_le32(OBD_INCOMPAT_GROUPS);
1216                 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1217                                                filter->fo_fsd, 1);
1218                 GOTO(cleanup_O0, rc);
1219
1220         cleanup_O0:
1221                 f_dput(O0_dentry);
1222         cleanup_R:
1223                 f_dput(dentry);
1224                 if (rc)
1225                         GOTO(cleanup, rc);
1226         } else {
1227                 f_dput(dentry);
1228         }
1229
1230         cleanup_phase = 2; /* groups */
1231
1232         /* we have to initialize all groups before first connections from
1233          * clients because they may send create/destroy for any group -bzzz */
1234         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1235         if (IS_ERR(filp)) {
1236                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1237                 GOTO(cleanup, rc = PTR_ERR(filp));
1238         }
1239         cleanup_phase = 3; /* filp */
1240
1241         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1242         if (rc) {
1243                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1244                 GOTO(cleanup, rc);
1245         }
1246         if (off == 0) {
1247                 last_group = FILTER_MIN_GROUPS;
1248         } else {
1249                 LASSERT_MDS_GROUP(last_group);
1250         }
1251
1252         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1253               FILTER_MIN_GROUPS, last_group);
1254         filter->fo_committed_group = last_group;
1255         rc = filter_read_groups(obd, last_group, 1);
1256         if (rc)
1257                 GOTO(cleanup, rc);
1258
1259         filp_close(filp, 0);
1260         RETURN(0);
1261
1262  cleanup:
1263         switch (cleanup_phase) {
1264         case 3:
1265                 filp_close(filp, 0);
1266         case 2:
1267                 filter_cleanup_groups(obd);
1268         case 1:
1269                 f_dput(filter->fo_dentry_O);
1270                 filter->fo_dentry_O = NULL;
1271         default:
1272                 break;
1273         }
1274         return rc;
1275
1276 }
1277
1278 /* setup the object store with correct subdirectories */
1279 static int filter_prep(struct obd_device *obd)
1280 {
1281         struct lvfs_run_ctxt saved;
1282         struct filter_obd *filter = &obd->u.filter;
1283         struct file *file;
1284         struct inode *inode;
1285         int rc = 0;
1286         ENTRY;
1287
1288         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1289         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1290         if (!file || IS_ERR(file)) {
1291                 rc = PTR_ERR(file);
1292                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1293                        LAST_RCVD, rc);
1294                 GOTO(out, rc);
1295         }
1296         filter->fo_rcvd_filp = file;
1297         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1298                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1299                        file->f_dentry->d_inode->i_mode);
1300                 GOTO(err_filp, rc = -ENOENT);
1301         }
1302
1303         inode = file->f_dentry->d_parent->d_inode;
1304         /* We use i_op->unlink directly in filter_vfs_unlink() */
1305         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1306                 CERROR("%s: filesystem does not support create/unlink ops\n",
1307                        obd->obd_name);
1308                 GOTO(err_filp, rc = -EOPNOTSUPP);
1309         }
1310
1311         rc = filter_init_server_data(obd, file);
1312         if (rc) {
1313                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1314                 GOTO(err_filp, rc);
1315         }
1316         /* open and create health check io file*/
1317         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1318         if (IS_ERR(file)) {
1319                 rc = PTR_ERR(file);
1320                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1321                        HEALTH_CHECK, rc);
1322                 GOTO(err_server_data, rc);
1323         }
1324         filter->fo_health_check_filp = file;
1325         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1326                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1327                        file->f_dentry->d_inode->i_mode);
1328                 GOTO(err_health_check, rc = -ENOENT);
1329         }
1330         rc = lvfs_check_io_health(obd, file);
1331         if (rc)
1332                 GOTO(err_health_check, rc);
1333
1334         rc = filter_prep_groups(obd);
1335         if (rc)
1336                 GOTO(err_health_check, rc);
1337 out:
1338         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1339
1340         return(rc);
1341
1342 err_health_check:
1343         if (filp_close(filter->fo_health_check_filp, 0))
1344                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1345         filter->fo_health_check_filp = NULL;
1346 err_server_data:
1347         target_recovery_fini(obd);
1348         filter_free_server_data(filter);
1349 err_filp:
1350         if (filp_close(filter->fo_rcvd_filp, 0))
1351                 CERROR("can't close %s after error\n", LAST_RCVD);
1352         filter->fo_rcvd_filp = NULL;
1353         goto out;
1354 }
1355
1356 /* cleanup the filter: write last used object id to status file */
1357 static void filter_post(struct obd_device *obd)
1358 {
1359         struct lvfs_run_ctxt saved;
1360         struct filter_obd *filter = &obd->u.filter;
1361         int rc, i;
1362
1363         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1364          * best to start a transaction with h_sync, because we removed this
1365          * from lastobjid */
1366
1367         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1368         rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1369                                        filter->fo_fsd, 0);
1370         if (rc)
1371                 CERROR("error writing server data: rc = %d\n", rc);
1372
1373         for (i = 0; i < filter->fo_group_count; i++) {
1374                 rc = filter_update_last_objid(obd, i,
1375                                 (i == filter->fo_group_count - 1));
1376                 if (rc)
1377                         CERROR("error writing group %d lastobjid: rc = %d\n",
1378                                i, rc);
1379         }
1380
1381         rc = filp_close(filter->fo_rcvd_filp, 0);
1382         filter->fo_rcvd_filp = NULL;
1383         if (rc)
1384                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1385
1386         rc = filp_close(filter->fo_health_check_filp, 0);
1387         filter->fo_health_check_filp = NULL;
1388         if (rc)
1389                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1390
1391         filter_cleanup_groups(obd);
1392         filter_free_server_data(filter);
1393         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1394
1395         filter_free_capa_keys(filter);
1396         cleanup_capa_hash(filter->fo_capa_hash);
1397 }
1398
1399 static void filter_set_last_id(struct filter_obd *filter,
1400                                obd_id id, obd_gr group)
1401 {
1402         LASSERT(filter->fo_fsd != NULL);
1403         LASSERT(group <= filter->fo_group_count);
1404
1405         spin_lock(&filter->fo_objidlock);
1406         filter->fo_last_objids[group] = id;
1407         spin_unlock(&filter->fo_objidlock);
1408 }
1409
1410 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1411 {
1412         obd_id id;
1413         LASSERT(filter->fo_fsd != NULL);
1414         LASSERT(group <= filter->fo_group_count);
1415
1416         /* FIXME: object groups */
1417         spin_lock(&filter->fo_objidlock);
1418         id = filter->fo_last_objids[group];
1419         spin_unlock(&filter->fo_objidlock);
1420         return id;
1421 }
1422
1423 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1424 {
1425         LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1426         return 0;
1427 }
1428
1429 /* We never dget the object parent, so DON'T dput it either */
1430 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1431 {
1432         struct filter_obd *filter = &obd->u.filter;
1433         struct filter_subdirs *subdirs;
1434         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1435
1436         if ((group > FILTER_GROUP_MDS0 && group < FILTER_GROUP_MDS1_N_BASE) ||
1437              filter->fo_subdir_count == 0)
1438                 return filter->fo_dentry_O_groups[group];
1439
1440         subdirs = &filter->fo_dentry_O_sub[group];
1441         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1442 }
1443
1444 /* We never dget the object parent, so DON'T dput it either */
1445 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1446                                   obd_id objid)
1447 {
1448         unsigned long now = jiffies;
1449         struct dentry *dparent = filter_parent(obd, group, objid);
1450         int rc;
1451
1452         if (IS_ERR(dparent))
1453                 return dparent;
1454         if (dparent == NULL)
1455                 return ERR_PTR(-ENOENT);
1456
1457         rc = filter_lock_dentry(obd, dparent);
1458         fsfilt_check_slow(obd, now, "parent lock");
1459         return rc ? ERR_PTR(rc) : dparent;
1460 }
1461
1462 /* We never dget the object parent, so DON'T dput it either */
1463 static void filter_parent_unlock(struct dentry *dparent)
1464 {
1465         UNLOCK_INODE_MUTEX(dparent->d_inode);
1466 }
1467
1468 /* How to get files, dentries, inodes from object id's.
1469  *
1470  * If dir_dentry is passed, the caller has already locked the parent
1471  * appropriately for this operation (normally a write lock).  If
1472  * dir_dentry is NULL, we do a read lock while we do the lookup to
1473  * avoid races with create/destroy and such changing the directory
1474  * internal to the filesystem code. */
1475 struct dentry *filter_fid2dentry(struct obd_device *obd,
1476                                  struct dentry *dir_dentry,
1477                                  obd_gr group, obd_id id)
1478 {
1479         struct dentry *dparent = dir_dentry;
1480         struct dentry *dchild;
1481         char name[32];
1482         int len;
1483         ENTRY;
1484
1485         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1486             obd->u.filter.fo_destroys_in_progress == 0) {
1487                 /* don't fail lookups for orphan recovery, it causes
1488                  * later LBUGs when objects still exist during precreate */
1489                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1490                 RETURN(ERR_PTR(-ENOENT));
1491         }
1492         if (id == 0) {
1493                 CERROR("fatal: invalid object id 0\n");
1494                 RETURN(ERR_PTR(-ESTALE));
1495         }
1496
1497         len = sprintf(name, LPU64, id);
1498         if (dir_dentry == NULL) {
1499                 dparent = filter_parent_lock(obd, group, id);
1500                 if (IS_ERR(dparent)) {
1501                         CERROR("%s: error getting object "LPU64":"LPU64
1502                                " parent: rc %ld\n", obd->obd_name,
1503                                id, group, PTR_ERR(dparent));
1504                         RETURN(dparent);
1505                 }
1506         }
1507         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1508                dparent->d_name.len, dparent->d_name.name, name);
1509         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1510         if (dir_dentry == NULL)
1511                 filter_parent_unlock(dparent);
1512         if (IS_ERR(dchild)) {
1513                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1514                        PTR_ERR(dchild));
1515                 RETURN(dchild);
1516         }
1517
1518         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1519                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1520                        obd->obd_name, id, dchild->d_inode->i_ino);
1521                 f_dput(dchild);
1522                 RETURN(ERR_PTR(-ENOENT));
1523         }
1524
1525         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1526                name, dchild, atomic_read(&dchild->d_count));
1527
1528         LASSERT(atomic_read(&dchild->d_count) > 0);
1529
1530         RETURN(dchild);
1531 }
1532
1533 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1534                                   obd_id group)
1535 {
1536         struct lustre_handle lockh;
1537         int flags = LDLM_AST_DISCARD_DATA, rc;
1538         struct ldlm_res_id res_id;
1539         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1540         ENTRY;
1541
1542         osc_build_res_name(objid, group, &res_id);
1543         /* Tell the clients that the object is gone now and that they should
1544          * throw away any cached pages. */
1545         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1546                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1547                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1548                                     NULL, &lockh);
1549
1550         /* We only care about the side-effects, just drop the lock. */
1551         if (rc == ELDLM_OK)
1552                 ldlm_lock_decref(&lockh, LCK_PW);
1553
1554         RETURN(rc);
1555 }
1556
1557 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1558  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1559  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1560  * reverse.  Caller must take i_sem before starting the transaction and we
1561  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1562 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1563                       struct vfsmount *mnt)
1564 {
1565         int rc;
1566         ENTRY;
1567
1568         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1569          * itself we already hold dir->i_mutex for child create/unlink ops */
1570         LASSERT(dentry->d_inode != NULL);
1571         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1572         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1573
1574
1575         /* may_delete() */
1576         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1577                 GOTO(out, rc = -ENOENT);
1578
1579         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1580         if (rc)
1581                 GOTO(out, rc);
1582
1583         if (IS_APPEND(dir))
1584                 GOTO(out, rc = -EPERM);
1585
1586         /* check_sticky() */
1587         if ((dentry->d_inode->i_uid != current->fsuid &&
1588              !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1589             IS_IMMUTABLE(dentry->d_inode))
1590                 GOTO(out, rc = -EPERM);
1591
1592         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1593          *       that was done because of journal_start (which is already done
1594          *       here) or some other ordering issue. */
1595         DQUOT_INIT(dir);
1596
1597         rc = ll_security_inode_unlink(dir, dentry, mnt);
1598         if (rc)
1599                 GOTO(out, rc);
1600
1601         rc = dir->i_op->unlink(dir, dentry);
1602 out:
1603         /* need to drop i_mutex before we lose inode reference */
1604         UNLOCK_INODE_MUTEX(dentry->d_inode);
1605         if (rc == 0)
1606                 d_delete(dentry);
1607
1608         RETURN(rc);
1609 }
1610
1611 /* Caller must hold LCK_PW on parent and push us into kernel context.
1612  * Caller must hold child i_mutex, we drop it always.
1613  * Caller is also required to ensure that dchild->d_inode exists. */
1614 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1615                                    obd_gr group, struct dentry *dparent,
1616                                    struct dentry *dchild)
1617 {
1618         struct inode *inode = dchild->d_inode;
1619         struct filter_obd *filter = &obd->u.filter;
1620         int rc;
1621
1622         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1623                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1624                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1625                        (unsigned long)inode->i_nlink,
1626                        atomic_read(&inode->i_count));
1627         }
1628
1629         rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1630         if (rc)
1631                 CERROR("error unlinking objid %.*s: rc %d\n",
1632                        dchild->d_name.len, dchild->d_name.name, rc);
1633         return(rc);
1634 }
1635
1636 struct filter_intent_args {
1637         struct ldlm_lock **victim;
1638         __u64 size;
1639         int *liblustre;
1640 };
1641
1642 static enum interval_iter filter_intent_cb(struct interval_node *n,
1643                                            void *args)
1644 {
1645         struct ldlm_interval *node = (struct ldlm_interval *)n;
1646         struct filter_intent_args *arg = (struct filter_intent_args*)args;
1647         __u64 size = arg->size;
1648         struct ldlm_lock **v = arg->victim;
1649         struct ldlm_lock *lck;
1650
1651         /* If the interval is lower than the current file size,
1652          * just break. */
1653         if (interval_high(n) <= size)
1654                 return INTERVAL_ITER_STOP;
1655
1656         list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1657                 /* Don't send glimpse ASTs to liblustre clients.
1658                  * They aren't listening for them, and they do
1659                  * entirely synchronous I/O anyways. */
1660                 if (lck->l_export == NULL ||
1661                     lck->l_export->exp_libclient == 1)
1662                         continue;
1663
1664                 if (*arg->liblustre)
1665                         *arg->liblustre = 0;
1666
1667                 if (*v == NULL) {
1668                         *v = LDLM_LOCK_GET(lck);
1669                 } else if ((*v)->l_policy_data.l_extent.start <
1670                            lck->l_policy_data.l_extent.start) {
1671                         LDLM_LOCK_RELEASE(*v);
1672                         *v = LDLM_LOCK_GET(lck);
1673                 }
1674
1675                 /* the same policy group - every lock has the
1676                  * same extent, so needn't do it any more */
1677                 break;
1678         }
1679
1680         return INTERVAL_ITER_CONT;
1681 }
1682
1683 static int filter_intent_policy(struct ldlm_namespace *ns,
1684                                 struct ldlm_lock **lockp, void *req_cookie,
1685                                 ldlm_mode_t mode, int flags, void *data)
1686 {
1687         CFS_LIST_HEAD(rpc_list);
1688         struct ptlrpc_request *req = req_cookie;
1689         struct ldlm_lock *lock = *lockp, *l = NULL;
1690         struct ldlm_resource *res = lock->l_resource;
1691         ldlm_processing_policy policy;
1692         struct ost_lvb *res_lvb, *reply_lvb;
1693         struct ldlm_reply *rep;
1694         ldlm_error_t err;
1695         int idx, rc, tmpflags = 0, only_liblustre = 1;
1696         struct ldlm_interval_tree *tree;
1697         struct filter_intent_args arg;
1698         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1699                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1700                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1701         ENTRY;
1702
1703         policy = ldlm_get_processing_policy(res);
1704         LASSERT(policy != NULL);
1705         LASSERT(req != NULL);
1706
1707         rc = lustre_pack_reply(req, 3, repsize, NULL);
1708         if (rc)
1709                 RETURN(req->rq_status = rc);
1710
1711         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1712         LASSERT(rep != NULL);
1713
1714         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1715                                    sizeof(*reply_lvb));
1716         LASSERT(reply_lvb != NULL);
1717
1718         //fixup_handle_for_resent_req(req, lock, &lockh);
1719
1720         /* Call the extent policy function to see if our request can be
1721          * granted, or is blocked.
1722          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1723          * lock, and should not be granted if the lock will be blocked.
1724          */
1725
1726         LASSERT(ns == res->lr_namespace);
1727         lock_res(res);
1728         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1729         check_res_locked(res);
1730
1731         /* FIXME: we should change the policy function slightly, to not make
1732          * this list at all, since we just turn around and free it */
1733         while (!list_empty(&rpc_list)) {
1734                 struct ldlm_lock *wlock =
1735                         list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1736                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1737                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1738                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1739                 list_del_init(&wlock->l_cp_ast);
1740                 LDLM_LOCK_RELEASE(wlock);
1741         }
1742
1743         /* The lock met with no resistance; we're finished. */
1744         if (rc == LDLM_ITER_CONTINUE) {
1745                 /* do not grant locks to the liblustre clients: they cannot
1746                  * handle ASTs robustly.  We need to do this while still
1747                  * holding ns_lock to avoid the lock remaining on the res_link
1748                  * list (and potentially being added to l_pending_list by an
1749                  * AST) when we are going to drop this lock ASAP. */
1750                 if (lock->l_export->exp_libclient ||
1751                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1752                         ldlm_resource_unlink_lock(lock);
1753                         err = ELDLM_LOCK_ABORTED;
1754                 } else {
1755                         err = ELDLM_LOCK_REPLACED;
1756                 }
1757                 unlock_res(res);
1758                 RETURN(err);
1759         }
1760
1761         /* Do not grant any lock, but instead send GL callbacks.  The extent
1762          * policy nicely created a list of all PW locks for us.  We will choose
1763          * the highest of those which are larger than the size in the LVB, if
1764          * any, and perform a glimpse callback. */
1765         res_lvb = res->lr_lvb_data;
1766         LASSERT(res_lvb != NULL);
1767         *reply_lvb = *res_lvb;
1768
1769         /*
1770          * ->ns_lock guarantees that no new locks are granted, and,
1771          * therefore, that res->lr_lvb_data cannot increase beyond the
1772          * end of already granted lock. As a result, it is safe to
1773          * check against "stale" reply_lvb->lvb_size value without
1774          * res->lr_lvb_sem.
1775          */
1776         arg.size = reply_lvb->lvb_size;
1777         arg.victim = &l;
1778         arg.liblustre = &only_liblustre;
1779         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1780                 tree = &res->lr_itree[idx];
1781                 if (tree->lit_mode == LCK_PR)
1782                         continue;
1783
1784                 interval_iterate_reverse(tree->lit_root,
1785                                          filter_intent_cb, &arg);
1786         }
1787         unlock_res(res);
1788
1789         /* There were no PW locks beyond the size in the LVB; finished. */
1790         if (l == NULL) {
1791                 if (only_liblustre) {
1792                         /* If we discovered a liblustre client with a PW lock,
1793                          * however, the LVB may be out of date!  The LVB is
1794                          * updated only on glimpse (which we don't do for
1795                          * liblustre clients) and cancel (which the client
1796                          * obviously has not yet done).  So if it has written
1797                          * data but kept the lock, the LVB is stale and needs
1798                          * to be updated from disk.
1799                          *
1800                          * Of course, this will all disappear when we switch to
1801                          * taking liblustre locks on the OST. */
1802                         ldlm_res_lvbo_update(res, NULL, 0, 1);
1803                 }
1804                 RETURN(ELDLM_LOCK_ABORTED);
1805         }
1806
1807         /*
1808          * This check is for lock taken in filter_prepare_destroy() that does
1809          * not have l_glimpse_ast set. So the logic is: if there is a lock
1810          * with no l_glimpse_ast set, this object is being destroyed already.
1811          *
1812          * Hence, if you are grabbing DLM locks on the server, always set
1813          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1814          */
1815         if (l->l_glimpse_ast == NULL) {
1816                 /* We are racing with unlink(); just return -ENOENT */
1817                 rep->lock_policy_res1 = -ENOENT;
1818                 goto out;
1819         }
1820
1821         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1822         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1823         /* Update the LVB from disk if the AST failed (this is a legal race) */
1824         /*
1825          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1826          * sending ast is not handled. This can result in lost client writes.
1827          */
1828         if (rc != 0)
1829                 ldlm_res_lvbo_update(res, NULL, 0, 1);
1830
1831         lock_res(res);
1832         *reply_lvb = *res_lvb;
1833         unlock_res(res);
1834
1835  out:
1836         LDLM_LOCK_RELEASE(l);
1837
1838         RETURN(ELDLM_LOCK_ABORTED);
1839 }
1840
1841 /*
1842  * per-obd_device iobuf pool.
1843  *
1844  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1845  * allocations in write-path has to be minimized (see bug 5137).
1846  *
1847  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1848  * OST threads (see ost_thread_{init,done}()).
1849  *
1850  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1851  * at the OST layer there are only (potentially) multiple obd_device of type
1852  * unknown at the time of OST thread creation.
1853  *
1854  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1855  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1856  * it's very own iobuf.
1857  *
1858  * Functions below
1859  *
1860  *     filter_kiobuf_pool_init()
1861  *
1862  *     filter_kiobuf_pool_done()
1863  *
1864  *     filter_iobuf_get()
1865  *
1866  * operate on this array. They are "generic" in a sense that they don't depend
1867  * on actual type of iobuf's (the latter depending on Linux kernel version).
1868  */
1869
1870 /*
1871  * destroy pool created by filter_iobuf_pool_init
1872  */
1873 static void filter_iobuf_pool_done(struct filter_obd *filter)
1874 {
1875         struct filter_iobuf **pool;
1876         int i;
1877
1878         ENTRY;
1879
1880         pool = filter->fo_iobuf_pool;
1881         if (pool != NULL) {
1882                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1883                         if (pool[i] != NULL)
1884                                 filter_free_iobuf(pool[i]);
1885                 }
1886                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1887                 filter->fo_iobuf_pool = NULL;
1888         }
1889         EXIT;
1890 }
1891
1892 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1893 {
1894         struct filter_obd       *filter = &obd->u.filter;
1895         struct sptlrpc_rule_set  tmp_rset;
1896         int                      rc;
1897
1898         sptlrpc_rule_set_init(&tmp_rset);
1899         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1900         if (rc) {
1901                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1902                        obd->obd_name, rc);
1903                 return rc;
1904         }
1905
1906         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1907
1908         write_lock(&filter->fo_sptlrpc_lock);
1909         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1910         filter->fo_sptlrpc_rset = tmp_rset;
1911         write_unlock(&filter->fo_sptlrpc_lock);
1912
1913         return 0;
1914 }
1915
1916 /*
1917  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1918  */
1919 static int filter_iobuf_pool_init(struct filter_obd *filter)
1920 {
1921         void **pool;
1922
1923         ENTRY;
1924
1925
1926         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1927                       GFP_KERNEL);
1928         if (filter->fo_iobuf_pool == NULL)
1929                 RETURN(-ENOMEM);
1930
1931         filter->fo_iobuf_count = OSS_THREADS_MAX;
1932
1933         RETURN(0);
1934 }
1935
1936 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1937  * many threads there will be so we allocate a large empty array and only
1938  * fill in those slots that are actually in use.
1939  * If we haven't allocated a pool entry for this thread before, do so now. */
1940 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1941 {
1942         int thread_id                    = (oti && oti->oti_thread) ?
1943                                            oti->oti_thread->t_id : -1;
1944         struct filter_iobuf  *pool       = NULL;
1945         struct filter_iobuf **pool_place = NULL;
1946
1947         if (thread_id >= 0) {
1948                 LASSERT(thread_id < filter->fo_iobuf_count);
1949                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1950         }
1951
1952         if (unlikely(pool == NULL)) {
1953                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1954                                           PTLRPC_MAX_BRW_PAGES);
1955                 if (pool_place != NULL)
1956                         *pool_place = pool;
1957         }
1958
1959         return pool;
1960 }
1961
1962 /* mount the file system (secretly).  lustre_cfg parameters are:
1963  * 1 = device
1964  * 2 = fstype
1965  * 3 = flags: failover=f, failout=n
1966  * 4 = mount options
1967  */
1968 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1969                         void *option)
1970 {
1971         struct filter_obd *filter = &obd->u.filter;
1972         struct vfsmount *mnt;
1973         struct lustre_mount_info *lmi;
1974         struct obd_uuid uuid;
1975         __u8 *uuid_ptr;
1976         char *str, *label;
1977         char ns_name[48];
1978         request_queue_t *q;
1979         int rc, i;
1980         ENTRY;
1981
1982         if (lcfg->lcfg_bufcount < 3 ||
1983             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1984             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1985                 RETURN(-EINVAL);
1986
1987         lmi = server_get_mount(obd->obd_name);
1988         if (lmi) {
1989                 /* We already mounted in lustre_fill_super.
1990                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1991                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1992                 mnt = lmi->lmi_mnt;
1993                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1994         } else {
1995                 /* old path - used by lctl */
1996                 CERROR("Using old MDS mount method\n");
1997                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1998                                     MS_NOATIME|MS_NODIRATIME,
1999                                     lustre_cfg_string(lcfg, 1), option);
2000                 if (IS_ERR(mnt)) {
2001                         rc = PTR_ERR(mnt);
2002                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2003                                            lustre_cfg_string(lcfg, 1), rc);
2004                         RETURN(rc);
2005                 }
2006
2007                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2008         }
2009         if (IS_ERR(obd->obd_fsops))
2010                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2011
2012         rc = filter_iobuf_pool_init(filter);
2013         if (rc != 0)
2014                 GOTO(err_ops, rc);
2015
2016         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2017                 CERROR("%s: Underlying device is marked as read-only. "
2018                        "Setup failed\n", obd->obd_name);
2019                 GOTO(err_ops, rc = -EROFS);
2020         }
2021
2022         /* failover is the default */
2023         obd->obd_replayable = 1;
2024
2025         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2026                 str = lustre_cfg_string(lcfg, 3);
2027                 if (strchr(str, 'n')) {
2028                         CWARN("%s: recovery disabled\n", obd->obd_name);
2029                         obd->obd_replayable = 0;
2030                 }
2031         }
2032
2033         filter->fo_vfsmnt = mnt;
2034         obd->u.obt.obt_sb = mnt->mnt_sb;
2035         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2036         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2037
2038         fsfilt_setup(obd, obd->u.obt.obt_sb);
2039
2040         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2041         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2042         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2043         obd->obd_lvfs_ctxt.fs = get_ds();
2044         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2045
2046         init_mutex(&filter->fo_init_lock);
2047         filter->fo_committed_group = 0;
2048
2049         rc = filter_prep(obd);
2050         if (rc)
2051                 GOTO(err_ops, rc);
2052
2053         filter->fo_destroys_in_progress = 0;
2054         for (i = 0; i < 32; i++)
2055                 sema_init(&filter->fo_create_locks[i], 1);
2056
2057         spin_lock_init(&filter->fo_translock);
2058         spin_lock_init(&filter->fo_objidlock);
2059         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2060         sema_init(&filter->fo_alloc_lock, 1);
2061         init_brw_stats(&filter->fo_filter_stats);
2062         filter->fo_read_cache = 1; /* enable read-only cache by default */
2063         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2064         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2065         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2066         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2067
2068         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2069         spin_lock_init(&filter->fo_llog_list_lock);
2070
2071         filter->fo_fl_oss_capa = 1;
2072         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2073         filter->fo_capa_hash = init_capa_hash();
2074         if (filter->fo_capa_hash == NULL)
2075                 GOTO(err_ops, rc = -ENOMEM);
2076
2077         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2078         obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2079                                                 LDLM_NAMESPACE_GREEDY);
2080         if (obd->obd_namespace == NULL)
2081                 GOTO(err_post, rc = -ENOMEM);
2082         obd->obd_namespace->ns_lvbp = obd;
2083         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2084         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2085
2086         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2087                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2088
2089         rc = obd_llog_init(obd, &obd->obd_olg, obd, 1, NULL, NULL);
2090         if (rc) {
2091                 CERROR("failed to setup llogging subsystems\n");
2092                 GOTO(err_post, rc);
2093         }
2094
2095         rwlock_init(&filter->fo_sptlrpc_lock);
2096         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2097         /* do this after llog being initialized */
2098         filter_adapt_sptlrpc_conf(obd, 1);
2099
2100         rc = lquota_setup(filter_quota_interface_ref, obd);
2101         if (rc)
2102                 GOTO(err_post, rc);
2103
2104         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2105         if (q->max_sectors < q->max_hw_sectors &&
2106             q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2107                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2108                               "for larger I/O requests: max_sectors = %u "
2109                               "could be up to max_hw_sectors=%u\n",
2110                               obd->obd_name, mnt->mnt_sb->s_id,
2111                               q->max_sectors, q->max_hw_sectors);
2112
2113         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2114         if (uuid_ptr != NULL) {
2115                 class_uuid_unparse(uuid_ptr, &uuid);
2116                 str = uuid.uuid;
2117         } else {
2118                 str = "no UUID";
2119         }
2120
2121         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2122
2123         if (obd->obd_recovering) {
2124                 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2125                               "recovery for at least %d:%.02d, or until %d "
2126                               "client%s reconnect. During this time new clients"
2127                               " will not be allowed to connect. "
2128                               "Recovery progress can be monitored by watching "
2129                               "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
2130                               obd->obd_name, lustre_cfg_string(lcfg, 1),
2131                               label ?: "", label ? "/" : "", str,
2132                               obd->obd_recovery_timeout / 60,
2133                               obd->obd_recovery_timeout % 60,
2134                               obd->obd_max_recoverable_clients,
2135                               (obd->obd_max_recoverable_clients == 1) ? "":"s",
2136                               obd->obd_name);
2137         } else {
2138                 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2139                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2140                               label ?: "", label ? "/" : "", str,
2141                               obd->obd_replayable ? "enabled" : "disabled");
2142         }
2143
2144         RETURN(0);
2145
2146 err_post:
2147         filter_post(obd);
2148 err_ops:
2149         fsfilt_put_ops(obd->obd_fsops);
2150         filter_iobuf_pool_done(filter);
2151 err_mntput:
2152         server_put_mount(obd->obd_name, mnt);
2153         obd->u.obt.obt_sb = 0;
2154         return rc;
2155 }
2156
2157 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2158 {
2159         struct lprocfs_static_vars lvars;
2160         unsigned long addr;
2161         struct page *page;
2162         int rc;
2163
2164         CLASSERT(offsetof(struct obd_device, u.obt) ==
2165                  offsetof(struct obd_device, u.filter.fo_obt));
2166
2167         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2168                 RETURN(-EINVAL);
2169
2170         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2171         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2172         if (!page)
2173                 RETURN(-ENOMEM);
2174         addr = (unsigned long)cfs_page_address(page);
2175         clear_page((void *)addr);
2176
2177         /* lprocfs must be setup before the filter so state can be safely added
2178          * to /proc incrementally as the filter is setup */
2179         lprocfs_filter_init_vars(&lvars);
2180         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2181             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2182                 /* Init obdfilter private stats here */
2183                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2184                                      LPROCFS_CNTR_AVGMINMAX,
2185                                      "read_bytes", "bytes");
2186                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2187                                      LPROCFS_CNTR_AVGMINMAX,
2188                                      "write_bytes", "bytes");
2189                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2190                                      LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2191                                      "get_page", "usec");
2192                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2193                                      LPROCFS_CNTR_AVGMINMAX,
2194                                      "get_page_failures", "num");
2195                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2196                                      LPROCFS_CNTR_AVGMINMAX,
2197                                      "cache_access", "pages");
2198                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2199                                      LPROCFS_CNTR_AVGMINMAX,
2200                                      "cache_hit", "pages");
2201                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2202                                      LPROCFS_CNTR_AVGMINMAX,
2203                                      "cache_miss", "pages");
2204
2205                 lproc_filter_attach_seqstat(obd);
2206                 obd->obd_proc_exports_entry = lprocfs_register("exports",
2207                                                         obd->obd_proc_entry,
2208                                                         NULL, NULL);
2209                 if (IS_ERR(obd->obd_proc_exports_entry)) {
2210                         rc = PTR_ERR(obd->obd_proc_exports_entry);
2211                         CERROR("error %d setting up lprocfs for %s\n",
2212                                rc, "exports");
2213                         obd->obd_proc_exports_entry = NULL;
2214                 }
2215         }
2216         if (obd->obd_proc_exports_entry)
2217                 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2218                                    lprocfs_nid_stats_clear_read,
2219                                    lprocfs_nid_stats_clear_write, obd, NULL);
2220
2221         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2222                LUSTRE_CFG_BUFLEN(lcfg, 4));
2223         rc = filter_common_setup(obd, lcfg, (void *)addr);
2224         OBD_PAGE_FREE(page);
2225
2226         if (rc) {
2227                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2228                 lprocfs_free_per_client_stats(obd);
2229                 lprocfs_free_obd_stats(obd);
2230                 lprocfs_obd_cleanup(obd);
2231         }
2232
2233         return rc;
2234 }
2235
2236 static struct llog_operations filter_mds_ost_repl_logops;
2237
2238 static struct llog_operations filter_size_orig_logops = {
2239         .lop_setup   = llog_obd_origin_setup,
2240         .lop_cleanup = llog_obd_origin_cleanup,
2241         .lop_add     = llog_obd_origin_add
2242 };
2243
2244 static int filter_olg_fini(struct obd_llog_group *olg)
2245 {
2246         struct llog_ctxt *ctxt;
2247         int rc = 0, rc2 = 0;
2248         ENTRY;
2249
2250         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2251         if (ctxt)
2252                 rc = llog_cleanup(ctxt);
2253
2254         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2255         if (ctxt) {
2256                 rc2 = llog_cleanup(ctxt);
2257                 if (!rc)
2258                         rc = rc2;
2259         }
2260
2261         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2262         if (ctxt) {
2263                 rc2 = llog_cleanup(ctxt);
2264                 if (!rc)
2265                         rc = rc2;
2266         }
2267
2268         RETURN(rc);
2269 }
2270
2271 static int
2272 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2273                 struct obd_device *tgt)
2274 {
2275         int rc;
2276         ENTRY;
2277
2278         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2279                         &filter_mds_ost_repl_logops);
2280         if (rc)
2281                 GOTO(cleanup, rc);
2282
2283         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2284                         &filter_size_orig_logops);
2285         if (rc)
2286                 GOTO(cleanup, rc);
2287         EXIT;
2288 cleanup:
2289         if (rc)
2290                 filter_olg_fini(olg);
2291         return rc;
2292 }
2293
2294 /**
2295  * Init the default olg, which is embeded in the obd_device, for filter.
2296  */
2297 static int
2298 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2299                         struct obd_device *tgt)
2300 {
2301         struct filter_obd *filter = &obd->u.filter;
2302         struct llog_ctxt *ctxt;
2303         int rc;
2304         ENTRY;
2305
2306         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2307         if (!filter->fo_lcm)
2308                 RETURN(-ENOMEM);
2309
2310         filter_mds_ost_repl_logops = llog_client_ops;
2311         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2312         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2313         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2314
2315         rc = filter_olg_init(obd, olg, tgt);
2316         if (rc)
2317                 GOTO(cleanup_lcm, rc);
2318
2319         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2320                         &llog_lvfs_ops);
2321         if (rc)
2322                 GOTO(cleanup_olg, rc);
2323
2324         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2325         if (!ctxt) {
2326                 CERROR("Can't get ctxt for %p:%x\n", olg,
2327                        LLOG_MDS_OST_REPL_CTXT);
2328                 GOTO(cleanup_olg, rc = -ENODEV);
2329         }
2330         ctxt->loc_lcm = filter->fo_lcm;
2331         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2332         llog_ctxt_put(ctxt);
2333
2334         RETURN(0);
2335 cleanup_olg:
2336         filter_olg_fini(olg);
2337 cleanup_lcm:
2338         llog_recov_thread_fini(filter->fo_lcm, 1);
2339         filter->fo_lcm = NULL;
2340         return rc;
2341 }
2342
2343 static int
2344 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2345                  struct obd_device *tgt, int count, struct llog_catid *catid,
2346                  struct obd_uuid *uuid)
2347 {
2348         struct filter_obd *filter = &obd->u.filter;
2349         struct llog_ctxt *ctxt;
2350         int rc;
2351         ENTRY;
2352
2353         LASSERT(olg != NULL);
2354         if (olg == &obd->obd_olg)
2355                 return filter_default_olg_init(obd, olg, tgt);
2356
2357         LASSERT(filter->fo_lcm != NULL);
2358         rc = filter_olg_init(obd, olg, tgt);
2359         if (rc)
2360                 RETURN(rc);
2361         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2362         if (!ctxt) {
2363                 CERROR("Can't get ctxt for %p:%x\n", olg,
2364                        LLOG_MDS_OST_REPL_CTXT);
2365                 filter_olg_fini(olg);
2366                 RETURN(-ENODEV);
2367         }
2368         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2369         ctxt->loc_lcm = filter->fo_lcm;
2370         llog_ctxt_put(ctxt);
2371         RETURN(rc);
2372 }
2373
2374 static int filter_llog_finish(struct obd_device *obd, int count)
2375 {
2376         struct filter_obd *filter = &obd->u.filter;
2377         struct llog_ctxt *ctxt;
2378         ENTRY;
2379
2380         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2381         if (ctxt) {
2382                 /*
2383                  * Make sure that no cached llcds left in recov_thread.
2384                  * We actually do sync in disconnect time, but disconnect
2385                  * may not come being marked rq_no_resend = 1.
2386                  */
2387                 llog_sync(ctxt, NULL);
2388
2389                 /*
2390                  * Balance class_import_get() in llog_receptor_accept().
2391                  * This is safe to do, as llog is already synchronized
2392                  * and its import may go.
2393                  */
2394                 mutex_down(&ctxt->loc_sem);
2395                 if (ctxt->loc_imp) {
2396                         class_import_put(ctxt->loc_imp);
2397                         ctxt->loc_imp = NULL;
2398                 }
2399                 mutex_up(&ctxt->loc_sem);
2400                 llog_ctxt_put(ctxt);
2401         }
2402
2403         if (filter->fo_lcm) {
2404                 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2405                 filter->fo_lcm = NULL;
2406         }
2407         RETURN(filter_olg_fini(&obd->obd_olg));
2408 }
2409
2410 /**
2411  * Find the group llog according to group index in the llog group list.
2412  */
2413 static struct obd_llog_group *
2414 filter_find_olg_internal(struct filter_obd *filter, int group)
2415 {
2416         struct obd_llog_group *olg;
2417
2418         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2419         list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2420                 if (olg->olg_group == group)
2421                         RETURN(olg);
2422         }
2423         RETURN(NULL);
2424 }
2425
2426 /**
2427  * Find the group llog according to group index on the filter
2428  */
2429 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2430 {
2431         struct obd_llog_group *olg = NULL;
2432         struct filter_obd *filter;
2433
2434         filter = &obd->u.filter;
2435
2436         if (group == FILTER_GROUP_LLOG)
2437                 RETURN(&obd->obd_olg);
2438
2439         spin_lock(&filter->fo_llog_list_lock);
2440         olg = filter_find_olg_internal(filter, group);
2441         spin_unlock(&filter->fo_llog_list_lock);
2442
2443         RETURN(olg);
2444 }
2445 /**
2446  * Find the llog_group of the filter according to the group. If it can not
2447  * find, create the llog_group, which only happens when mds is being synced
2448  * with OST.
2449  */
2450 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2451 {
2452         struct obd_llog_group *olg = NULL;
2453         struct filter_obd *filter;
2454         int rc;
2455
2456         filter = &obd->u.filter;
2457
2458         if (group == FILTER_GROUP_LLOG)
2459                 RETURN(&obd->obd_olg);
2460
2461         spin_lock(&filter->fo_llog_list_lock);
2462         olg = filter_find_olg_internal(filter, group);
2463         if (olg) {
2464                 if (olg->olg_initializing) {
2465                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2466                 } else {
2467                         GOTO(out_unlock, olg);
2468                 }
2469         }
2470         OBD_ALLOC_PTR(olg);
2471         if (olg == NULL)
2472                GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2473
2474         llog_group_init(olg, group);
2475         list_add(&olg->olg_list, &filter->fo_llog_list);
2476         olg->olg_initializing = 1;
2477         spin_unlock(&filter->fo_llog_list_lock);
2478
2479         rc = llog_cat_initialize(obd, olg, 1, NULL);
2480         if (rc) {
2481                spin_lock(&filter->fo_llog_list_lock);
2482                list_del(&olg->olg_list);
2483                spin_unlock(&filter->fo_llog_list_lock);
2484                OBD_FREE_PTR(olg);
2485                GOTO(out, olg = ERR_PTR(-ENOMEM));
2486         }
2487         spin_lock(&filter->fo_llog_list_lock);
2488         olg->olg_initializing = 0;
2489         spin_unlock(&filter->fo_llog_list_lock);
2490         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2491               obd->obd_name, group, olg);
2492 out:
2493         RETURN(olg);
2494
2495 out_unlock:
2496         spin_unlock(&filter->fo_llog_list_lock);
2497         GOTO(out, olg);
2498 }
2499
2500 static int filter_llog_connect(struct obd_export *exp,
2501                                struct llogd_conn_body *body)
2502 {
2503         struct obd_device *obd = exp->exp_obd;
2504         struct llog_ctxt *ctxt;
2505         struct obd_llog_group *olg;
2506         int rc;
2507         ENTRY;
2508
2509         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2510                obd->obd_name, body->lgdc_logid.lgl_oid,
2511                body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2512
2513         olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2514         if (!olg) {
2515                 CERROR(" %s: can not find olg of group %d\n",
2516                        obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2517                 RETURN(-ENOENT);
2518         }
2519         llog_group_set_export(olg, exp);
2520
2521         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2522         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2523                  body->lgdc_ctxt_idx);
2524
2525         CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2526               obd->obd_name, body->lgdc_logid.lgl_oid,
2527               body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2528
2529         rc = llog_connect(ctxt, &body->lgdc_logid,
2530                           &body->lgdc_gen, NULL);
2531         llog_ctxt_put(ctxt);
2532         if (rc != 0)
2533                 CERROR("failed to connect rc %d idx %d\n", rc,
2534                                 body->lgdc_ctxt_idx);
2535
2536         RETURN(rc);
2537 }
2538
2539 static int filter_llog_preclean(struct obd_device *obd)
2540 {
2541         struct obd_llog_group *olg, *tmp;
2542         struct filter_obd *filter;
2543         struct list_head  remove_list;
2544         int rc = 0;
2545         ENTRY;
2546
2547         rc = obd_llog_finish(obd, 0);
2548         if (rc)
2549                 CERROR("failed to cleanup llogging subsystem\n");
2550
2551         filter = &obd->u.filter;
2552         CFS_INIT_LIST_HEAD(&remove_list);
2553
2554         spin_lock(&filter->fo_llog_list_lock);
2555         while (!list_empty(&filter->fo_llog_list)) {
2556                 olg = list_entry(filter->fo_llog_list.next,
2557                                  struct obd_llog_group, olg_list);
2558                 list_del(&olg->olg_list);
2559                 list_add(&olg->olg_list, &remove_list);
2560         }
2561         spin_unlock(&filter->fo_llog_list_lock);
2562
2563         list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2564                 list_del_init(&olg->olg_list);
2565                 rc = filter_olg_fini(olg);
2566                 if (rc)
2567                         CERROR("failed to cleanup llogging subsystem for %u\n",
2568                                olg->olg_group);
2569                 OBD_FREE_PTR(olg);
2570         }
2571
2572         RETURN(rc);
2573 }
2574
2575 static int filter_precleanup(struct obd_device *obd,
2576                              enum obd_cleanup_stage stage)
2577 {
2578         int rc = 0;
2579         ENTRY;
2580
2581         switch(stage) {
2582         case OBD_CLEANUP_EARLY:
2583                 break;
2584         case OBD_CLEANUP_EXPORTS:
2585                 target_cleanup_recovery(obd);
2586                 rc = filter_llog_preclean(obd);
2587                 break;
2588         }
2589         RETURN(rc);
2590 }
2591
2592 static int filter_cleanup(struct obd_device *obd)
2593 {
2594         struct filter_obd *filter = &obd->u.filter;
2595         ENTRY;
2596
2597         if (obd->obd_fail)
2598                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2599                               "will be preserved.\n", obd->obd_name);
2600
2601         if (!list_empty(&obd->obd_exports)) {
2602                 CERROR("%s: still has clients!\n", obd->obd_name);
2603                 class_disconnect_exports(obd);
2604                 if (!list_empty(&obd->obd_exports)) {
2605                         CERROR("still has exports after forced cleanup?\n");
2606                         RETURN(-EBUSY);
2607                 }
2608         }
2609
2610         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2611         lprocfs_free_per_client_stats(obd);
2612         lprocfs_free_obd_stats(obd);
2613         lprocfs_obd_cleanup(obd);
2614         lquota_cleanup(filter_quota_interface_ref, obd);
2615
2616         /* Stop recovery before namespace cleanup. */
2617         target_stop_recovery_thread(obd);
2618         target_cleanup_recovery(obd);
2619
2620         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2621         obd->obd_namespace = NULL;
2622
2623         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2624
2625         if (obd->u.obt.obt_sb == NULL)
2626                 RETURN(0);
2627
2628         filter_post(obd);
2629
2630         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2631         shrink_dcache_sb(obd->u.obt.obt_sb);
2632
2633         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2634         obd->u.obt.obt_sb = NULL;
2635
2636         fsfilt_put_ops(obd->obd_fsops);
2637
2638         filter_iobuf_pool_done(filter);
2639
2640         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2641
2642         RETURN(0);
2643 }
2644
2645 static int filter_connect_internal(struct obd_export *exp,
2646                                    struct obd_connect_data *data)
2647 {
2648         struct filter_export_data *fed = &exp->exp_filter_data;
2649
2650         if (!data)
2651                 RETURN(0);
2652
2653         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2654                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2655                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2656                data->ocd_connect_flags, data->ocd_version,
2657                data->ocd_grant, data->ocd_index);
2658
2659         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2660                 CWARN("!!! This export (nid %s) used object group %d "
2661                        "earlier; now it's trying to use group %d!  This could "
2662                        "be a bug in the MDS. Please report to "
2663                        "http://bugzilla.lustre.org/\n",
2664                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2665                 RETURN(-EPROTO);
2666         }
2667         fed->fed_group = data->ocd_group;
2668
2669         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2670         exp->exp_connect_flags = data->ocd_connect_flags;
2671         data->ocd_version = LUSTRE_VERSION_CODE;
2672
2673         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2674         if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2675                   data->ocd_connect_flags & OBD_CONNECT_MDS))
2676                 RETURN(-EPROTO);
2677
2678         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2679                 obd_size left, want;
2680
2681                 spin_lock(&exp->exp_obd->obd_osfs_lock);
2682                 left = filter_grant_space_left(exp);
2683                 want = data->ocd_grant;
2684                 filter_grant(exp, fed->fed_grant, want, left);
2685                 data->ocd_grant = fed->fed_grant;
2686                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2687
2688                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2689                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2690                        exp->exp_client_uuid.uuid, exp,
2691                        data->ocd_grant, want, left);
2692         }
2693
2694         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2695                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2696                 struct lr_server_data *lsd = filter->fo_fsd;
2697                 int index = le32_to_cpu(lsd->lsd_ost_index);
2698
2699                 if (!(lsd->lsd_feature_compat &
2700                       cpu_to_le32(OBD_COMPAT_OST))) {
2701                         /* this will only happen on the first connect */
2702                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2703                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2704                         filter_update_server_data(exp->exp_obd,
2705                                                   filter->fo_rcvd_filp, lsd, 1);
2706                 } else if (index != data->ocd_index) {
2707                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2708                                            " %u doesn't match actual OST index"
2709                                            " %u in last_rcvd file, bad "
2710                                            "configuration?\n",
2711                                            obd_export_nid2str(exp), index,
2712                                            data->ocd_index);
2713                         RETURN(-EBADF);
2714                 }
2715         }
2716
2717         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2718                 data->ocd_brw_size = 65536;
2719         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2720                 data->ocd_brw_size = min(data->ocd_brw_size,
2721                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2722                 LASSERT(data->ocd_brw_size);
2723         }
2724
2725         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2726                 __u32 cksum_types = data->ocd_cksum_types;
2727
2728                 /* The client set in ocd_cksum_types the checksum types it
2729                  * supports. We have to mask off the algorithms that we don't
2730                  * support */
2731                 if (cksum_types & OBD_CKSUM_ALL)
2732                         data->ocd_cksum_types &= OBD_CKSUM_ALL;
2733                 else
2734                         data->ocd_cksum_types = OBD_CKSUM_CRC32;
2735
2736                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2737                                    "%x\n", exp->exp_obd->obd_name,
2738                                    obd_export_nid2str(exp), cksum_types,
2739                                    data->ocd_cksum_types);
2740         } else {
2741                 /* This client does not support OBD_CONNECT_CKSUM
2742                  * fall back to CRC32 */
2743                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2744                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2745                                    exp->exp_obd->obd_name,
2746                                    obd_export_nid2str(exp));
2747         }
2748
2749         /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2750          * FIXME: We don't strictly need the COMPAT flag for that,
2751          * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2752          * FIXME: We needed it for the index, as index 0 is valid. */
2753
2754         RETURN(0);
2755 }
2756
2757 static int filter_reconnect(const struct lu_env *env,
2758                             struct obd_export *exp, struct obd_device *obd,
2759                             struct obd_uuid *cluuid,
2760                             struct obd_connect_data *data,
2761                             void *localdata)
2762 {
2763         int rc;
2764         ENTRY;
2765
2766         if (exp == NULL || obd == NULL || cluuid == NULL)
2767                 RETURN(-EINVAL);
2768
2769         rc = filter_connect_internal(exp, data);
2770         if (rc == 0)
2771                 filter_export_stats_init(obd, exp, localdata);
2772
2773         RETURN(rc);
2774 }
2775
2776 /* nearly identical to mds_connect */
2777 static int filter_connect(const struct lu_env *env,
2778                           struct obd_export **exp, struct obd_device *obd,
2779                           struct obd_uuid *cluuid,
2780                           struct obd_connect_data *data, void *localdata)
2781 {
2782         struct lvfs_run_ctxt saved;
2783         struct lustre_handle conn = { 0 };
2784         struct obd_export *lexp;
2785         struct filter_export_data *fed;
2786         struct lsd_client_data *lcd = NULL;
2787         __u32 group;
2788         int rc;
2789         ENTRY;
2790
2791         if (exp == NULL || obd == NULL || cluuid == NULL)
2792                 RETURN(-EINVAL);
2793
2794         rc = class_connect(&conn, obd, cluuid);
2795         if (rc)
2796                 RETURN(rc);
2797         lexp = class_conn2export(&conn);
2798         LASSERT(lexp != NULL);
2799
2800         fed = &lexp->exp_filter_data;
2801
2802         rc = filter_connect_internal(lexp, data);
2803         if (rc)
2804                 GOTO(cleanup, rc);
2805
2806         filter_export_stats_init(obd, lexp, localdata);
2807         if (obd->obd_replayable) {
2808                 OBD_ALLOC(lcd, sizeof(*lcd));
2809                 if (!lcd) {
2810                         CERROR("filter: out of memory for client data\n");
2811                         GOTO(cleanup, rc = -ENOMEM);
2812                 }
2813
2814                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2815                 fed->fed_lcd = lcd;
2816                 rc = filter_client_add(obd, lexp, -1);
2817                 if (rc)
2818                         GOTO(cleanup, rc);
2819         }
2820
2821         group = data->ocd_group;
2822
2823         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2824               obd->obd_name, lexp->exp_handle.h_cookie, group);
2825
2826         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2827         rc = filter_read_groups(obd, group, 1);
2828         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2829         if (rc != 0) {
2830                 CERROR("can't read group %u\n", group);
2831                 GOTO(cleanup, rc);
2832         }
2833
2834         GOTO(cleanup, rc);
2835
2836 cleanup:
2837         if (rc) {
2838                 if (lcd) {
2839                         OBD_FREE_PTR(lcd);
2840                         fed->fed_lcd = NULL;
2841                 }
2842                 class_disconnect(lexp);
2843                 *exp = NULL;
2844         } else {
2845                 *exp = lexp;
2846         }
2847
2848         RETURN(rc);
2849 }
2850
2851 /* Do extra sanity checks for grant accounting.  We do this at connect,
2852  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2853  * always get rid of it or turn it off when we know accounting is good. */
2854 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2855 {
2856         struct filter_export_data *fed;
2857         struct obd_export *exp;
2858         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2859         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2860         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2861
2862         if (list_empty(&obd->obd_exports))
2863                 return;
2864
2865         /* We don't want to do this for large machines that do lots of
2866            mounts or unmounts.  It burns... */
2867         if (obd->obd_num_exports > 100)
2868                 return;
2869
2870         spin_lock(&obd->obd_osfs_lock);
2871         spin_lock(&obd->obd_dev_lock);
2872         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2873                 int error = 0;
2874                 fed = &exp->exp_filter_data;
2875                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2876                     fed->fed_dirty < 0)
2877                         error = 1;
2878                 if (maxsize > 0) { /* we may not have done a statfs yet */
2879                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2880                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2881                                  exp->exp_client_uuid.uuid, exp,
2882                                  fed->fed_grant, fed->fed_pending, maxsize);
2883                         LASSERTF(fed->fed_dirty <= maxsize,
2884                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2885                                  exp->exp_client_uuid.uuid, exp,
2886                                  fed->fed_dirty, maxsize);
2887                 }
2888                 if (error)
2889                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2890                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2891                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2892                 else
2893                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2894                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2895                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2896                 tot_granted += fed->fed_grant + fed->fed_pending;
2897                 tot_pending += fed->fed_pending;
2898                 tot_dirty += fed->fed_dirty;
2899         }
2900         fo_tot_granted = obd->u.filter.fo_tot_granted;
2901         fo_tot_pending = obd->u.filter.fo_tot_pending;
2902         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2903         spin_unlock(&obd->obd_dev_lock);
2904         spin_unlock(&obd->obd_osfs_lock);
2905
2906         /* Do these assertions outside the spinlocks so we don't kill system */
2907         if (tot_granted != fo_tot_granted)
2908                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2909                        func, tot_granted, fo_tot_granted);
2910         if (tot_pending != fo_tot_pending)
2911                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2912                        func, tot_pending, fo_tot_pending);
2913         if (tot_dirty != fo_tot_dirty)
2914                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2915                        func, tot_dirty, fo_tot_dirty);
2916         if (tot_pending > tot_granted)
2917                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2918                        func, tot_pending, tot_granted);
2919         if (tot_granted > maxsize)
2920                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2921                        func, tot_granted, maxsize);
2922         if (tot_dirty > maxsize)
2923                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2924                        func, tot_dirty, maxsize);
2925 }
2926
2927 /* Remove this client from the grant accounting totals.  We also remove
2928  * the export from the obd device under the osfs and dev locks to ensure
2929  * that the filter_grant_sanity_check() calculations are always valid.
2930  * The client should do something similar when it invalidates its import. */
2931 static void filter_grant_discard(struct obd_export *exp)
2932 {
2933         struct obd_device *obd = exp->exp_obd;
2934         struct filter_obd *filter = &obd->u.filter;
2935         struct filter_export_data *fed = &exp->exp_filter_data;
2936
2937         spin_lock(&obd->obd_osfs_lock);
2938         spin_lock(&obd->obd_dev_lock);
2939         list_del_init(&exp->exp_obd_chain);
2940         spin_unlock(&obd->obd_dev_lock);
2941
2942         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2943                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2944                  obd->obd_name, filter->fo_tot_granted,
2945                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2946         filter->fo_tot_granted -= fed->fed_grant;
2947         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2948                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2949                  obd->obd_name, filter->fo_tot_pending,
2950                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2951         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2952         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2953                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2954                  obd->obd_name, filter->fo_tot_dirty,
2955                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2956         filter->fo_tot_dirty -= fed->fed_dirty;
2957         fed->fed_dirty = 0;
2958         fed->fed_grant = 0;
2959
2960         spin_unlock(&obd->obd_osfs_lock);
2961 }
2962
2963 static int filter_destroy_export(struct obd_export *exp)
2964 {
2965         ENTRY;
2966
2967         if (exp->exp_filter_data.fed_pending)
2968                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2969                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2970                        exp, exp->exp_filter_data.fed_pending);
2971
2972         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2973
2974         target_destroy_export(exp);
2975         ldlm_destroy_export(exp);
2976
2977         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2978                 RETURN(0);
2979
2980
2981         if (exp->exp_obd->obd_replayable)
2982                 filter_client_free(exp);
2983         else
2984                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2985
2986         filter_grant_discard(exp);
2987         filter_fmd_cleanup(exp);
2988
2989         if (!(exp->exp_flags & OBD_OPT_FORCE))
2990                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2991
2992         RETURN(0);
2993 }
2994
2995 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2996 {
2997         struct obd_llog_group *olg_min, *olg;
2998         struct filter_obd *filter;
2999         int worked = -1, group;
3000         struct llog_ctxt *ctxt;
3001         ENTRY;
3002
3003         filter = &obd->u.filter;
3004
3005         /* we can't sync log holding spinlock. also, we do not want to get
3006          * into livelock. so we do following: loop over MDS's exports in
3007          * group order and skip already synced llogs -bzzz */
3008         do {
3009                 /* look for group with min. number, but > worked */
3010                 olg_min = NULL;
3011                 group = 1 << 30;
3012                 spin_lock(&filter->fo_llog_list_lock);
3013                 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3014                         if (olg->olg_group <= worked) {
3015                                 /* this group is already synced */
3016                                 continue;
3017                         }
3018                         if (group < olg->olg_group) {
3019                                 /* we have group with smaller number to sync */
3020                                 continue;
3021                         }
3022                         /* store current minimal group */
3023                         olg_min = olg;
3024                         group = olg->olg_group;
3025                 }
3026                 spin_unlock(&filter->fo_llog_list_lock);
3027
3028                 if (olg_min == NULL)
3029                         break;
3030
3031                 worked = olg_min->olg_group;
3032                 if (olg_min->olg_exp &&
3033                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3034                         int err;
3035                         ctxt = llog_group_get_ctxt(olg_min,
3036                                                    LLOG_MDS_OST_REPL_CTXT);
3037                         if (ctxt) {
3038                                 err = llog_sync(ctxt, olg_min->olg_exp);
3039                                 llog_ctxt_put(ctxt);
3040                                 if (err) {
3041                                         CERROR("error flushing logs to MDS: "
3042                                                "rc %d\n", err);
3043                                 }
3044                         }
3045                 }
3046         } while (olg_min != NULL);
3047 }
3048
3049 /* Also incredibly similar to mds_disconnect */
3050 static int filter_disconnect(struct obd_export *exp)
3051 {
3052         struct obd_device *obd = exp->exp_obd;
3053         int rc;
3054         ENTRY;
3055
3056         LASSERT(exp);
3057         class_export_get(exp);
3058
3059         if (!(exp->exp_flags & OBD_OPT_FORCE))
3060                 filter_grant_sanity_check(obd, __FUNCTION__);
3061         filter_grant_discard(exp);
3062
3063         /* Flush any remaining cancel messages out to the target */
3064         filter_sync_llogs(obd, exp);
3065
3066         lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3067
3068         /* Disconnect early so that clients can't keep using export */
3069         rc = class_disconnect(exp);
3070         if (exp->exp_obd->obd_namespace != NULL)
3071                 ldlm_cancel_locks_for_export(exp);
3072
3073         fsfilt_sync(obd, obd->u.obt.obt_sb);
3074
3075         lprocfs_exp_cleanup(exp);
3076
3077         class_export_put(exp);
3078         RETURN(rc);
3079 }
3080
3081 /* reverse import is changed, sync all cancels */
3082 static void filter_revimp_update(struct obd_export *exp)
3083 {
3084         ENTRY;
3085
3086         LASSERT(exp);
3087         class_export_get(exp);
3088
3089         /* flush any remaining cancel messages out to the target */
3090         filter_sync_llogs(exp->exp_obd, exp);
3091         class_export_put(exp);
3092         EXIT;
3093 }
3094
3095 static int filter_ping(struct obd_export *exp)
3096 {
3097         filter_fmd_expire(exp);
3098
3099         return 0;
3100 }
3101
3102 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3103                                   const char *what, int quiet)
3104 {
3105         struct dentry *dchild = NULL;
3106         obd_gr group = 0;
3107
3108         if (oa->o_valid & OBD_MD_FLGROUP)
3109                 group = oa->o_gr;
3110
3111         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3112
3113         if (IS_ERR(dchild)) {
3114                 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3115                        what, group, oa->o_id);
3116                 RETURN(dchild);
3117         }
3118
3119         if (dchild->d_inode == NULL) {
3120                 if (!quiet)
3121                         CERROR("%s: %s on non-existent object: "LPU64"\n",
3122                                obd->obd_name, what, oa->o_id);
3123                 f_dput(dchild);
3124                 RETURN(ERR_PTR(-ENOENT));
3125         }
3126
3127         return dchild;
3128 }
3129
3130 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3131 {
3132         struct dentry *dentry = NULL;
3133         struct obd_device *obd;
3134         int rc = 0;
3135         ENTRY;
3136
3137         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3138                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3139         if (rc)
3140                 RETURN(rc);
3141
3142         obd = class_exp2obd(exp);
3143         if (obd == NULL) {
3144                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3145                 RETURN(-EINVAL);
3146         }
3147
3148         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3149         if (IS_ERR(dentry))
3150                 RETURN(PTR_ERR(dentry));
3151
3152         /* Limit the valid bits in the return data to what we actually use */
3153         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3154         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3155
3156         f_dput(dentry);
3157         RETURN(rc);
3158 }
3159
3160 /* this should be enabled/disabled in condition to enabled/disabled large
3161  * inodes (fast EAs) in backing store FS. */
3162 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3163                         void *handle, struct obdo *oa)
3164 {
3165         struct obd_device *obd = exp->exp_obd;
3166         int rc = 0;
3167         ENTRY;
3168
3169         if (oa->o_valid & OBD_MD_FLFID) {
3170                 struct filter_fid ff;
3171
3172                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3173                         oa->o_gr = 0;
3174                 /* packing fid and converting it to LE for storing into EA.
3175                  * Here ->o_stripe_idx should be filled by LOV and rest of
3176                  * fields - by client. */
3177                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3178                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3179                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3180                 ff.ff_objid = cpu_to_le64(oa->o_id);
3181                 ff.ff_group = cpu_to_le64(oa->o_gr);
3182
3183                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3184                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3185                        oa->o_generation, oa->o_id, oa->o_gr);
3186
3187                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3188                 if (rc)
3189                         CERROR("store fid in object failed! rc: %d\n", rc);
3190         } else {
3191                 CDEBUG(D_HA, "OSS object without fid info!\n");
3192         }
3193
3194         RETURN(rc);
3195 }
3196
3197 /* this is called from filter_truncate() until we have filter_punch() */
3198 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3199                             struct obdo *oa, struct obd_trans_info *oti)
3200 {
3201         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3202         struct llog_cookie *fcc = NULL;
3203         struct filter_obd *filter;
3204         int rc, err, locked = 0, sync = 0;
3205         loff_t old_size = 0;
3206         unsigned int ia_valid;
3207         struct inode *inode;
3208         struct iattr iattr;
3209         void *handle;
3210         ENTRY;
3211
3212         LASSERT(dentry != NULL);
3213         LASSERT(!IS_ERR(dentry));
3214
3215         inode = dentry->d_inode;
3216         LASSERT(inode != NULL);
3217
3218         filter = &exp->exp_obd->u.filter;
3219         iattr_from_obdo(&iattr, oa, oa->o_valid);
3220         ia_valid = iattr.ia_valid;
3221
3222         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3223                 OBD_ALLOC(fcc, sizeof(*fcc));
3224                 if (fcc != NULL)
3225                         *fcc = oa->o_lcookie;
3226         }
3227
3228         if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
3229                 DQUOT_INIT(inode);
3230                 LOCK_INODE_MUTEX(inode);
3231                 old_size = i_size_read(inode);
3232                 locked = 1;
3233         }
3234
3235         /* If the inode still has SUID+SGID bits set (see filter_precreate())
3236          * then we will accept the UID+GID sent by the client during write for
3237          * initializing the ownership of this inode.  We only allow this to
3238          * happen once so clear these bits in setattr. In 2.6 kernels it is
3239          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3240          * the flags that are actually being set. */
3241         if (ia_valid & (ATTR_UID | ATTR_GID)) {
3242                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3243                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3244
3245                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3246                         if (!(ia_valid & ATTR_MODE)) {
3247                                 iattr.ia_mode = inode->i_mode;
3248                                 iattr.ia_valid |= ATTR_MODE;
3249                         }
3250                         iattr.ia_mode &= ~S_ISUID;
3251                 }
3252                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3253                         if (!(iattr.ia_valid & ATTR_MODE)) {
3254                                 iattr.ia_mode = inode->i_mode;
3255                                 iattr.ia_valid |= ATTR_MODE;
3256                         }
3257                         iattr.ia_mode &= ~S_ISGID;
3258                 }
3259
3260                 orig_ids[USRQUOTA] = inode->i_uid;
3261                 orig_ids[GRPQUOTA] = inode->i_gid;
3262                 handle = fsfilt_start_log(exp->exp_obd, inode,
3263                                           FSFILT_OP_SETATTR, oti, 1);
3264                 if (IS_ERR(handle))
3265                         GOTO(out_unlock, rc = PTR_ERR(handle));
3266
3267                 /* update inode EA only once when inode is suid bit marked. As
3268                  * on 2.6.x UID and GID may be set separately, we check here
3269                  * only one of them to avoid double setting. */
3270                 if (inode->i_mode & S_ISUID)
3271                         filter_update_fidea(exp, inode, handle, oa);
3272         } else {
3273                 handle = fsfilt_start(exp->exp_obd, inode,
3274                                       FSFILT_OP_SETATTR, oti);
3275                 if (IS_ERR(handle))
3276                         GOTO(out_unlock, rc = PTR_ERR(handle));
3277         }
3278         if (oa->o_valid & OBD_MD_FLFLAGS) {
3279                 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3280                                       EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3281         } else {
3282                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3283                 if (fcc != NULL)
3284                         /* set cancel cookie callback function */
3285                         sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3286                                                      filter_cancel_cookies_cb,
3287                                                      fcc);
3288         }
3289
3290         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3291                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3292
3293         /* The truncate might have used up our transaction credits.  Make
3294          * sure we have one left for the last_rcvd update. */
3295         err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3296
3297         rc = filter_finish_transno(exp, oti, rc, sync);
3298         if (sync) {
3299                 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3300                 fcc = NULL;
3301         }
3302
3303         err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3304         if (err) {
3305                 CERROR("error on commit, err = %d\n", err);
3306                 if (!rc)
3307                         rc = err;
3308         } else {
3309                 fcc = NULL;
3310         }
3311
3312         /* For a partial-page truncate flush the page to disk immediately
3313          * to avoid data corruption during direct disk write. b=17397 */
3314         if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3315             old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3316                 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3317                                                iattr.ia_size + 1);
3318                 if (!rc)
3319                         rc = err;
3320         }
3321
3322         if (locked) {
3323                 UNLOCK_INODE_MUTEX(inode);
3324                 locked = 0;
3325         }
3326
3327         EXIT;
3328 out_unlock:
3329         if (locked)
3330                 UNLOCK_INODE_MUTEX(inode);
3331
3332         if (fcc)
3333                 OBD_FREE(fcc, sizeof(*fcc));
3334
3335         /* trigger quota release */
3336         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3337                 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3338                 int rc2 = lquota_adjust(filter_quota_interface_ref,
3339                                         exp->exp_obd, cur_ids,
3340                                         orig_ids, rc, FSFILT_OP_SETATTR);
3341                 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3342                        "filter adjust qunit. (rc:%d)\n", rc2);
3343         }
3344         return rc;
3345 }
3346
3347 /* this is called from filter_truncate() until we have filter_punch() */
3348 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3349                    struct obd_trans_info *oti)
3350 {
3351         struct obdo *oa = oinfo->oi_oa;
3352         struct lustre_capa *capa = oinfo_capa(oinfo);
3353         struct ldlm_res_id res_id;
3354         struct filter_mod_data *fmd;
3355         struct lvfs_run_ctxt saved;
3356         struct filter_obd *filter;
3357         struct ldlm_resource *res;
3358         struct dentry *dentry;
3359         __u64 opc = CAPA_OPC_META_WRITE;
3360         int rc;
3361         ENTRY;
3362
3363         if (oa->o_valid & OBD_FL_TRUNC)
3364                 opc |= CAPA_OPC_OSS_TRUNC;
3365         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa, opc);
3366         if (rc)
3367                 RETURN(rc);
3368
3369         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3370                 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
3371                 if (rc)
3372                         RETURN(rc);
3373         }
3374
3375         osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3376         /* This would be very bad - accidentally truncating a file when
3377          * changing the time or similar - bug 12203. */
3378         if (oa->o_valid & OBD_MD_FLSIZE &&
3379             oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3380                 static char mdsinum[48];
3381
3382                 if (oa->o_valid & OBD_MD_FLFID)
3383                         snprintf(mdsinum, sizeof(mdsinum) - 1,
3384                                  " of inode "LPU64"/%u", oa->o_fid,
3385                                  oa->o_generation);
3386                 else
3387                         mdsinum[0] = '\0';
3388
3389                 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3390                        " %s\n",
3391                        exp->exp_obd->obd_name, obd_export_nid2str(exp),
3392                        oa->o_id, mdsinum);
3393                 RETURN(-EPERM);
3394         }
3395
3396         dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3397         if (IS_ERR(dentry))
3398                 RETURN(PTR_ERR(dentry));
3399
3400         filter = &exp->exp_obd->u.filter;
3401         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3402         lock_kernel();
3403
3404         if (oa->o_valid &
3405             (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3406                 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3407                 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3408                         fmd->fmd_mactime_xid = oti->oti_xid;
3409                 filter_fmd_put(exp, fmd);
3410         }
3411
3412         /* setting objects attributes (including owner/group) */
3413         rc = filter_setattr_internal(exp, dentry, oa, oti);
3414         if (rc)
3415                 GOTO(out_unlock, rc);
3416
3417         res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3418                                 &res_id, LDLM_EXTENT, 0);
3419
3420         if (res != NULL) {
3421                 LDLM_RESOURCE_ADDREF(res);
3422                 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3423                 LDLM_RESOURCE_DELREF(res);
3424                 ldlm_resource_putref(res);
3425         }
3426
3427         oa->o_valid = OBD_MD_FLID;
3428
3429         /* Quota release need uid/gid info */
3430         obdo_from_inode(oa, dentry->d_inode,
3431                         FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3432
3433         EXIT;
3434 out_unlock:
3435         unlock_kernel();
3436         f_dput(dentry);
3437         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3438         return rc;
3439 }
3440
3441 /* XXX identical to osc_unpackmd */
3442 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3443                            struct lov_mds_md *lmm, int lmm_bytes)
3444 {
3445         int lsm_size;
3446         ENTRY;
3447
3448         if (lmm != NULL) {
3449                 if (lmm_bytes < sizeof (*lmm)) {
3450                         CERROR("lov_mds_md too small: %d, need %d\n",
3451                                lmm_bytes, (int)sizeof(*lmm));
3452                         RETURN(-EINVAL);
3453                 }
3454                 /* XXX LOV_MAGIC etc check? */
3455
3456                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3457                         CERROR("lov_mds_md: zero lmm_object_id\n");
3458                         RETURN(-EINVAL);
3459                 }
3460         }
3461
3462         lsm_size = lov_stripe_md_size(1);
3463         if (lsmp == NULL)
3464                 RETURN(lsm_size);
3465
3466         if (*lsmp != NULL && lmm == NULL) {
3467                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3468                 OBD_FREE(*lsmp, lsm_size);
3469                 *lsmp = NULL;
3470                 RETURN(0);
3471         }
3472
3473         if (*lsmp == NULL) {
3474                 OBD_ALLOC(*lsmp, lsm_size);
3475                 if (*lsmp == NULL)
3476                         RETURN(-ENOMEM);
3477
3478                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3479                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3480                         OBD_FREE(*lsmp, lsm_size);
3481                         RETURN(-ENOMEM);
3482                 }
3483                 loi_init((*lsmp)->lsm_oinfo[0]);
3484         }
3485
3486         if (lmm != NULL) {
3487                 /* XXX zero *lsmp? */
3488                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3489                 LASSERT((*lsmp)->lsm_object_id);
3490         }
3491
3492         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3493
3494         RETURN(lsm_size);
3495 }
3496
3497 /* caller must hold fo_create_locks[oa->o_gr] */
3498 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3499                                      struct filter_obd *filter)
3500 {
3501         struct obdo doa; /* XXX obdo on stack */
3502         obd_id last, id;
3503         int rc = 0;
3504         int skip_orphan;
3505         ENTRY;
3506
3507         LASSERT(oa);
3508         LASSERT_MDS_GROUP(oa->o_gr);
3509         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3510         LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3511
3512         memset(&doa, 0, sizeof(doa));
3513
3514         doa.o_valid |= OBD_MD_FLGROUP;
3515         doa.o_gr = oa->o_gr;
3516         doa.o_mode = S_IFREG;
3517
3518         if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3519                 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3520                        exp->exp_obd->obd_name, doa.o_gr);
3521                 RETURN(0);
3522         }
3523
3524         last = filter_last_id(filter, doa.o_gr);
3525
3526         skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3527
3528         CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3529                exp->exp_obd->obd_name, oa->o_id + 1, last,
3530                skip_orphan ? ", orphan objids won't be reused any more." : ".");
3531
3532         for (id = last; id > oa->o_id; id--) {
3533                 doa.o_id = id;
3534                 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3535                 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3536                         CEMERG("error destroying precreate objid "LPU64": %d\n",
3537                                id, rc);
3538
3539                 /* update last_id on disk periodically so that if we restart
3540                  * we don't need to re-scan all of the just-deleted objects. */
3541                 if ((id & 511) == 0 && !skip_orphan) {
3542                         filter_set_last_id(filter, id - 1, doa.o_gr);
3543                         filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3544                 }
3545         }
3546
3547         CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3548                exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3549
3550         if (!skip_orphan) {
3551                 filter_set_last_id(filter, id, doa.o_gr);
3552                 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3553         } else {
3554                 /* don't reuse orphan object, return last used objid */
3555                 oa->o_id = last;
3556                 rc = 0;
3557         }
3558         clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3559
3560         RETURN(rc);
3561 }
3562
3563 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3564                             obd_gr group, int *num);
3565 /* returns a negative error or a nonnegative number of files to create */
3566 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3567                                    obd_gr group, struct obd_trans_info *oti)
3568 {
3569         struct obd_device *obd = exp->exp_obd;
3570         struct filter_obd *filter = &obd->u.filter;
3571         int diff, rc;
3572         ENTRY;
3573
3574         /* delete orphans request */
3575         if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3576                 obd_id last = filter_last_id(filter, group);
3577
3578                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3579                         CERROR("%s: dropping old orphan cleanup request\n",
3580                                obd->obd_name);
3581                         RETURN(0);
3582                 }
3583                 /* This causes inflight precreates to abort and drop lock */
3584                 set_bit(group, &filter->fo_destroys_in_progress);
3585                 down(&filter->fo_create_locks[group]);
3586                 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3587                         CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3588                                exp->exp_obd->obd_name, group);
3589                         up(&filter->fo_create_locks[group]);
3590                         RETURN(0);
3591                 }
3592                 diff = oa->o_id - last;
3593                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3594                        last, diff);
3595
3596                 if (-diff > OST_MAX_PRECREATE) {
3597                         CERROR("%s: ignoring bogus orphan destroy request: "
3598                                "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3599                                oa->o_id, last);
3600                         /* FIXME: should reset precreate_next_id on MDS */
3601                         GOTO(out, rc = -EINVAL);
3602                 }
3603                 if (diff < 0) {
3604                         rc = filter_destroy_precreated(exp, oa, filter);
3605                         if (rc)
3606                                 CERROR("%s: unable to write lastobjid, but "
3607                                        "orphans were deleted\n", obd->obd_name);
3608                         GOTO(out, rc);
3609                 } else {
3610                         /* XXX: Used by MDS for the first time! */
3611                         clear_bit(group, &filter->fo_destroys_in_progress);
3612                 }
3613         } else {
3614                 down(&filter->fo_create_locks[group]);
3615                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3616                         CERROR("%s: dropping old precreate request\n",
3617                                obd->obd_name);
3618                         GOTO(out, rc = 0);
3619                 }
3620                 /* only precreate if group == 0 and o_id is specfied */
3621                 if (group == FILTER_GROUP_LLOG || oa->o_id == 0)
3622                         diff = 1;
3623                 else
3624                         diff = oa->o_id - filter_last_id(filter, group);
3625                 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3626                        filter_last_id(filter, group), diff);
3627
3628                 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3629                          oa->o_id, filter_last_id(filter, group), diff);
3630         }
3631
3632         if (diff > 0) {
3633                 oa->o_id = filter_last_id(&obd->u.filter, group);
3634                 rc = filter_precreate(obd, oa, group, &diff);
3635                 oa->o_id = filter_last_id(&obd->u.filter, group);
3636                 oa->o_gr = group;
3637                 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3638                 GOTO(out, rc);
3639         }
3640         /* else diff == 0 */
3641         GOTO(out, rc = 0);
3642 out:
3643         up(&filter->fo_create_locks[group]);
3644         return rc;
3645 }
3646
3647 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3648                          __u64 max_age, __u32 flags)
3649 {
3650         struct filter_obd *filter = &obd->u.filter;
3651         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3652         int rc;
3653         ENTRY;
3654
3655         /* at least try to account for cached pages.  its still racey and
3656          * might be under-reporting if clients haven't announced their
3657          * caches with brw recently */
3658         spin_lock(&obd->obd_osfs_lock);
3659         rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3660         memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3661         spin_unlock(&obd->obd_osfs_lock);
3662
3663         CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3664                " pending "LPU64" free "LPU64" avail "LPU64"\n",
3665                filter->fo_tot_dirty, filter->fo_tot_granted,
3666                filter->fo_tot_pending,
3667                osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3668
3669         filter_grant_sanity_check(obd, __FUNCTION__);
3670
3671         osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3672                                ((filter->fo_tot_dirty + filter->fo_tot_pending +
3673                                  osfs->os_bsize - 1) >> blockbits));
3674
3675         /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3676          * stop creating files on MDS if OST is not good shape to create
3677          * objects.*/
3678         osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3679                 EROFS : 0;
3680         RETURN(rc);