Whamcloud - gitweb
e213d3e741197ebadfde2a45b826ee65b8143838
[fs/lustre-release.git] / lustre / obdfilter / filter.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/obdfilter/filter.c
37  *
38  * Author: Peter Braam <braam@clusterfs.com>
39  * Author: Andreas Dilger <adilger@clusterfs.com>
40  */
41
42 /*
43  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44  *            (which need to get journal_lock, may block if journal full).
45  *
46  * Invariant: Call filter_start_transno() before any journal ops to avoid the
47  *            same deadlock problem.  We can (and want) to get rid of the
48  *            transno sem in favour of the dir/inode i_mutex to avoid single
49  *            threaded operation on the OST.
50  */
51
52 #define DEBUG_SUBSYSTEM S_FILTER
53
54 #include <linux/module.h>
55 #include <linux/fs.h>
56 #include <linux/dcache.h>
57 #include <linux/init.h>
58 #include <linux/version.h>
59 #include <linux/sched.h>
60 #include <linux/mount.h>
61 #include <linux/buffer_head.h>
62
63 #include <obd_cksum.h>
64 #include <obd_class.h>
65 #include <obd_lov.h>
66 #include <lustre_dlm.h>
67 #include <lustre_fsfilt.h>
68 #include <lprocfs_status.h>
69 #include <lustre_log.h>
70 #include <libcfs/list.h>
71 #include <lustre_disk.h>
72 #include <linux/slab.h>
73 #include <lustre_param.h>
74 #include <lustre/ll_fiemap.h>
75
76 #include "filter_internal.h"
77
78 static struct lvfs_callback_ops filter_lvfs_ops;
79 cfs_mem_cache_t *ll_fmd_cachep;
80
81 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
82                              void *cb_data, int error)
83 {
84         struct obd_export *exp = cb_data;
85         LASSERT(exp->exp_obd == obd);
86         obd_transno_commit_cb(obd, transno, exp, error);
87         class_export_cb_put(exp);
88 }
89
90 int filter_version_get_check(struct obd_export *exp,
91                              struct obd_trans_info *oti, struct inode *inode)
92 {
93         __u64 curr_version;
94
95         if (inode == NULL || oti == NULL)
96                 RETURN(0);
97
98         curr_version = fsfilt_get_version(exp->exp_obd, inode);
99         if ((__s64)curr_version == -EOPNOTSUPP)
100                 RETURN(0);
101         /* VBR: version is checked always because costs nothing */
102         if (oti->oti_pre_version != 0 &&
103             oti->oti_pre_version != curr_version) {
104                 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
105                        oti->oti_pre_version, curr_version);
106                 cfs_spin_lock(&exp->exp_lock);
107                 exp->exp_vbr_failed = 1;
108                 cfs_spin_unlock(&exp->exp_lock);
109                 RETURN (-EOVERFLOW);
110         }
111         oti->oti_pre_version = curr_version;
112         RETURN(0);
113 }
114
115 /* Assumes caller has already pushed us into the kernel context. */
116 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
117                           struct obd_trans_info *oti, int rc, int force_sync)
118 {
119         struct obd_device_target *obt = &exp->exp_obd->u.obt;
120         struct tg_export_data *ted = &exp->exp_target_data;
121         struct lr_server_data *lsd = class_server_data(exp->exp_obd);
122         struct lsd_client_data *lcd;
123         __u64 last_rcvd;
124         loff_t off;
125         int err, log_pri = D_RPCTRACE;
126
127         /* Propagate error code. */
128         if (rc)
129                 RETURN(rc);
130
131         if (!exp->exp_obd->obd_replayable || oti == NULL)
132                 RETURN(rc);
133
134         cfs_mutex_lock(&ted->ted_lcd_lock);
135         lcd = ted->ted_lcd;
136         /* if the export has already been disconnected, we have no last_rcvd slot,
137          * update server data with latest transno then */
138         if (lcd == NULL) {
139                 cfs_mutex_unlock(&ted->ted_lcd_lock);
140                 CWARN("commit transaction for disconnected client %s: rc %d\n",
141                       exp->exp_client_uuid.uuid, rc);
142                 err = filter_update_server_data(exp->exp_obd);
143                 RETURN(err);
144         }
145
146         /* we don't allocate new transnos for replayed requests */
147         cfs_spin_lock(&obt->obt_lut->lut_translock);
148         if (oti->oti_transno == 0) {
149                 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
150                 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
151                 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
152         } else {
153                 last_rcvd = oti->oti_transno;
154                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
155                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
156                 if (unlikely(last_rcvd < le64_to_cpu(lcd->lcd_last_transno))) {
157                         CERROR("Trying to overwrite bigger transno, on-disk: "
158                                LPU64", new: "LPU64"\n",
159                                le64_to_cpu(lcd->lcd_last_transno), last_rcvd);
160                         cfs_spin_lock(&exp->exp_lock);
161                         exp->exp_vbr_failed = 1;
162                         cfs_spin_unlock(&exp->exp_lock);
163                         cfs_spin_unlock(&obt->obt_lut->lut_translock);
164                         cfs_mutex_unlock(&ted->ted_lcd_lock);
165                         RETURN(-EOVERFLOW);
166                 }
167         }
168         oti->oti_transno = last_rcvd;
169
170         lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
171         lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
172         lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
173         cfs_spin_unlock(&obt->obt_lut->lut_translock);
174
175         if (inode)
176                 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
177
178         off = ted->ted_lr_off;
179         if (off <= 0) {
180                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
181                        ted->ted_lr_idx, ted->ted_lr_off);
182                 err = -EINVAL;
183         } else {
184                 class_export_cb_get(exp); /* released when the cb is called */
185                 if (!force_sync)
186                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
187                                                            last_rcvd,
188                                                            oti->oti_handle,
189                                                            filter_commit_cb,
190                                                            exp);
191
192                 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
193                                           lcd, sizeof(*lcd), &off,
194                                           force_sync | exp->exp_need_sync);
195                 if (force_sync)
196                         filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
197         }
198         if (err) {
199                 log_pri = D_ERROR;
200                 if (rc == 0)
201                         rc = err;
202         }
203
204         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
205                last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
206         cfs_mutex_unlock(&ted->ted_lcd_lock);
207         RETURN(rc);
208 }
209
210 void f_dput(struct dentry *dentry)
211 {
212         /* Can't go inside filter_ddelete because it can block */
213         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
214                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
215         LASSERT(atomic_read(&dentry->d_count) > 0);
216
217         dput(dentry);
218 }
219
220 static void init_brw_stats(struct brw_stats *brw_stats)
221 {
222         int i;
223         for (i = 0; i < BRW_LAST; i++)
224                 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
225 }
226
227 static int lprocfs_init_rw_stats(struct obd_device *obd,
228                                  struct lprocfs_stats **stats)
229 {
230         int num_stats;
231
232         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
233                                                         LPROC_FILTER_LAST - 1;
234         *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
235         if (*stats == NULL)
236                 return -ENOMEM;
237
238         lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
239         lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
240                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
241         lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
242                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
243
244         return(0);
245 }
246
247 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
248    plus the procfs overhead :( */
249 static int filter_export_stats_init(struct obd_device *obd,
250                                     struct obd_export *exp,
251                                     void *client_nid)
252 {
253         int rc, newnid = 0;
254         ENTRY;
255
256         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
257                 /* Self-export gets no proc entry */
258                 RETURN(0);
259
260         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
261         if (rc) {
262                 /* Mask error for already created
263                  * /proc entries */
264                 if (rc == -EALREADY)
265                         rc = 0;
266                 RETURN(rc);
267         }
268
269         if (newnid) {
270                 struct nid_stat *tmp = exp->exp_nid_stats;
271                 LASSERT(tmp != NULL);
272
273                 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
274                 if (tmp->nid_brw_stats == NULL)
275                         GOTO(clean, rc = -ENOMEM);
276
277                 init_brw_stats(tmp->nid_brw_stats);
278                 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
279                                         0644, &filter_per_nid_stats_fops,
280                                         exp->exp_nid_stats);
281                 if (rc)
282                         CWARN("Error adding the brw_stats file\n");
283
284                 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
285                 if (rc)
286                         GOTO(clean, rc);
287
288                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
289                                             tmp->nid_stats);
290                 if (rc)
291                         GOTO(clean, rc);
292                 rc = lprocfs_nid_ldlm_stats_init(tmp);
293                 if (rc)
294                         GOTO(clean, rc);
295         }
296
297         RETURN(0);
298  clean:
299         return rc;
300 }
301
302 /* Add client data to the FILTER.  We use a bitmap to locate a free space
303  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
304  * Otherwise, we have just read the data from the last_rcvd file and
305  * we know its offset. */
306 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
307                              int cl_idx)
308 {
309         struct obd_device_target *obt = &obd->u.obt;
310         struct tg_export_data *ted = &exp->exp_target_data;
311         struct lr_server_data *lsd = class_server_data(obd);
312         unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
313         int new_client = (cl_idx == -1);
314
315         ENTRY;
316
317         LASSERT(bitmap != NULL);
318         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
319
320         /* Self-export */
321         if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
322                 RETURN(0);
323
324         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
325          * there's no need for extra complication here
326          */
327         if (new_client) {
328                 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
329         repeat:
330                 if (cl_idx >= LR_MAX_CLIENTS) {
331                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
332                                cl_idx);
333                         RETURN(-EOVERFLOW);
334                 }
335                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
336                         cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
337                                                         cl_idx);
338                         goto repeat;
339                 }
340         } else {
341                 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
342                         CERROR("FILTER client %d: bit already set in bitmap!\n",
343                                cl_idx);
344                         LBUG();
345                 }
346         }
347
348         ted->ted_lr_idx = cl_idx;
349         ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
350                           cl_idx * le16_to_cpu(lsd->lsd_client_size);
351         cfs_mutex_init(&ted->ted_lcd_lock);
352         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
353
354         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
355                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
356
357         if (new_client) {
358                 struct lvfs_run_ctxt saved;
359                 loff_t off = ted->ted_lr_off;
360                 int rc;
361                 void *handle;
362
363                 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
364                        ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
365
366                 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
367                         RETURN(-ENOSPC);
368
369                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
370                 /* Transaction needed to fix bug 1403 */
371                 handle = fsfilt_start(obd,
372                                       obt->obt_rcvd_filp->f_dentry->d_inode,
373                                       FSFILT_OP_SETATTR, NULL);
374                 if (IS_ERR(handle)) {
375                         rc = PTR_ERR(handle);
376                         CERROR("unable to start transaction: rc %d\n", rc);
377                 } else {
378                         ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
379                         exp->exp_last_request_time = cfs_time_current_sec();
380                         rc = fsfilt_add_journal_cb(obd, 0, handle,
381                                                    target_client_add_cb,
382                                                    class_export_cb_get(exp));
383                         if (rc == 0) {
384                                 cfs_spin_lock(&exp->exp_lock);
385                                 exp->exp_need_sync = 1;
386                                 cfs_spin_unlock(&exp->exp_lock);
387                         }
388                         rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
389                                                  ted->ted_lcd,
390                                                  sizeof(*ted->ted_lcd),
391                                                  &off, rc /* sync if no cb */);
392                         fsfilt_commit(obd,
393                                       obt->obt_rcvd_filp->f_dentry->d_inode,
394                                       handle, 0);
395                 }
396                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
397
398                 if (rc) {
399                         CERROR("error writing %s client idx %u: rc %d\n",
400                                LAST_RCVD, ted->ted_lr_idx, rc);
401                         RETURN(rc);
402                 }
403         }
404         RETURN(0);
405 }
406
407 static int filter_client_del(struct obd_export *exp)
408 {
409         struct tg_export_data *ted = &exp->exp_target_data;
410         struct obd_device_target *obt = &exp->exp_obd->u.obt;
411         struct lvfs_run_ctxt saved;
412         int rc;
413         loff_t off;
414         ENTRY;
415
416         if (ted->ted_lcd == NULL)
417                 RETURN(0);
418
419         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
420         if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
421                 GOTO(free, 0);
422
423         LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
424
425         off = ted->ted_lr_off;
426
427         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
428                ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
429
430         /* Don't clear ted_lr_idx here as it is likely also unset.  At worst
431          * we leak a client slot that will be cleaned on the next recovery. */
432         if (off <= 0) {
433                 CERROR("%s: client idx %d has med_off %lld\n",
434                        exp->exp_obd->obd_name, ted->ted_lr_idx, off);
435                 GOTO(free, rc = -EINVAL);
436         }
437
438         /* Clear the bit _after_ zeroing out the client so we don't
439            race with filter_client_add and zero out new clients.*/
440         if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
441                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
442                        ted->ted_lr_idx);
443                 LBUG();
444         }
445
446         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
447         /* Make sure the server's last_transno is up to date.
448          * This should be done before zeroing client slot so last_transno will
449          * be in server data or in client data in case of failure */
450         filter_update_server_data(exp->exp_obd);
451
452         cfs_mutex_lock(&ted->ted_lcd_lock);
453         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
454         rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
455                                  ted->ted_lcd,
456                                  sizeof(*ted->ted_lcd), &off, 0);
457         cfs_mutex_unlock(&ted->ted_lcd_lock);
458         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
459
460         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
461                "zero out client %s at idx %u/%llu in %s, rc %d\n",
462                ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
463                LAST_RCVD, rc);
464         EXIT;
465 free:
466         return 0;
467 }
468
469 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
470 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
471                                          struct filter_mod_data *fmd)
472 {
473         LASSERT_SPIN_LOCKED(&fed->fed_lock);
474         if (--fmd->fmd_refcount == 0) {
475                 /* XXX when we have persistent reservations and the handle
476                  * is stored herein we need to drop it here. */
477                 fed->fed_mod_count--;
478                 cfs_list_del(&fmd->fmd_list);
479                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
480         }
481 }
482
483 /* drop fmd reference, free it if last ref */
484 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
485 {
486         struct filter_export_data *fed;
487
488         if (fmd == NULL)
489                 return;
490
491         fed = &exp->exp_filter_data;
492         cfs_spin_lock(&fed->fed_lock);
493         filter_fmd_put_nolock(fed, fmd); /* caller reference */
494         cfs_spin_unlock(&fed->fed_lock);
495 }
496
497 /* expire entries from the end of the list if there are too many
498  * or they are too old */
499 static void filter_fmd_expire_nolock(struct filter_obd *filter,
500                                      struct filter_export_data *fed,
501                                      struct filter_mod_data *keep)
502 {
503         struct filter_mod_data *fmd, *tmp;
504
505         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
506                 if (fmd == keep)
507                         break;
508
509                 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
510                     fed->fed_mod_count < filter->fo_fmd_max_num)
511                         break;
512
513                 cfs_list_del_init(&fmd->fmd_list);
514                 filter_fmd_put_nolock(fed, fmd); /* list reference */
515         }
516 }
517
518 void filter_fmd_expire(struct obd_export *exp)
519 {
520         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
521         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
522                                  &exp->exp_filter_data, NULL);
523         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
524 }
525
526 /* find specified objid, group in export fmd list.
527  * caller must hold fed_lock and take fmd reference itself */
528 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
529                                                 struct filter_export_data *fed,
530                                                 obd_id objid, obd_seq group)
531 {
532         struct filter_mod_data *found = NULL, *fmd;
533
534         LASSERT_SPIN_LOCKED(&fed->fed_lock);
535
536         cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
537                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
538                         found = fmd;
539                         cfs_list_del(&fmd->fmd_list);
540                         cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
541                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
542                         break;
543                 }
544         }
545
546         filter_fmd_expire_nolock(filter, fed, found);
547
548         return found;
549 }
550
551 /* Find fmd based on objid and group, or return NULL if not found. */
552 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
553                                         obd_id objid, obd_seq group)
554 {
555         struct filter_mod_data *fmd;
556
557         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
558         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
559                                      &exp->exp_filter_data, objid, group);
560         if (fmd)
561                 fmd->fmd_refcount++;    /* caller reference */
562         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
563
564         return fmd;
565 }
566
567 /* Find fmd based on objid and group, or create a new one if none is found.
568  * It is possible for this function to return NULL under memory pressure,
569  * or if objid = 0 is passed (which will only cause old entries to expire).
570  * Currently this is not fatal because any fmd state is transient and
571  * may also be freed when it gets sufficiently old. */
572 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
573                                        obd_id objid, obd_seq group)
574 {
575         struct filter_export_data *fed = &exp->exp_filter_data;
576         struct filter_mod_data *found = NULL, *fmd_new = NULL;
577
578         OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
579
580         cfs_spin_lock(&fed->fed_lock);
581         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
582         if (fmd_new) {
583                 if (found == NULL) {
584                         cfs_list_add_tail(&fmd_new->fmd_list,
585                                           &fed->fed_mod_list);
586                         fmd_new->fmd_id = objid;
587                         fmd_new->fmd_gr = group;
588                         fmd_new->fmd_refcount++;   /* list reference */
589                         found = fmd_new;
590                         fed->fed_mod_count++;
591                 } else {
592                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
593                 }
594         }
595         if (found) {
596                 found->fmd_refcount++;          /* caller reference */
597                 found->fmd_expire = jiffies +
598                         exp->exp_obd->u.filter.fo_fmd_max_age;
599         }
600
601         cfs_spin_unlock(&fed->fed_lock);
602
603         return found;
604 }
605
606 #ifdef DO_FMD_DROP
607 /* drop fmd list reference so it will disappear when last reference is put.
608  * This isn't so critical because it would in fact only affect the one client
609  * that is doing the unlink and at worst we have an stale entry referencing
610  * an object that should never be used again. */
611 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
612 {
613         struct filter_mod_data *found = NULL;
614
615         cfs_spin_lock(&exp->exp_filter_data.fed_lock);
616         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
617         if (found) {
618                 cfs_list_del_init(&found->fmd_list);
619                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
620         }
621         cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
622 }
623 #else
624 #define filter_fmd_drop(exp, objid, group)
625 #endif
626
627 /* remove all entries from fmd list */
628 static void filter_fmd_cleanup(struct obd_export *exp)
629 {
630         struct filter_export_data *fed = &exp->exp_filter_data;
631         struct filter_mod_data *fmd = NULL, *tmp;
632
633         cfs_spin_lock(&fed->fed_lock);
634         cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
635                 cfs_list_del_init(&fmd->fmd_list);
636                 filter_fmd_put_nolock(fed, fmd);
637         }
638         cfs_spin_unlock(&fed->fed_lock);
639 }
640
641 static int filter_init_export(struct obd_export *exp)
642 {
643         int rc;
644         ENTRY;
645
646         cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
647         CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
648
649         cfs_spin_lock(&exp->exp_lock);
650         exp->exp_connecting = 1;
651         cfs_spin_unlock(&exp->exp_lock);
652
653         /* self-export doesn't need client data and ldlm initialization */
654         if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
655                                      &exp->exp_client_uuid)))
656                 RETURN(0);
657
658         rc = lut_client_alloc(exp);
659         if (rc == 0)
660                 rc = ldlm_init_export(exp);
661         if (rc)
662                 CERROR("%s: Can't initialize export: rc %d\n",
663                        exp->exp_obd->obd_name, rc);
664
665         RETURN(rc);
666 }
667
668 static int filter_free_server_data(struct obd_device_target *obt)
669 {
670         lut_fini(NULL, obt->obt_lut);
671         OBD_FREE_PTR(obt->obt_lut);
672         return 0;
673 }
674
675 /* assumes caller is already in kernel ctxt */
676 int filter_update_server_data(struct obd_device *obd)
677 {
678         struct file *filp = obd->u.obt.obt_rcvd_filp;
679         struct lr_server_data *lsd = class_server_data(obd);
680         loff_t off = 0;
681         int rc;
682         ENTRY;
683
684         CDEBUG(D_INODE, "server uuid      : %s\n", lsd->lsd_uuid);
685         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
686                le64_to_cpu(lsd->lsd_last_transno));
687         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
688                le64_to_cpu(lsd->lsd_mount_count));
689
690         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
691         if (rc)
692                 CERROR("error writing lr_server_data: rc = %d\n", rc);
693
694         RETURN(rc);
695 }
696
697 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
698                              int force_sync)
699 {
700         struct filter_obd *filter = &obd->u.filter;
701         __u64 tmp;
702         loff_t off = 0;
703         int rc;
704         ENTRY;
705
706         if (filter->fo_last_objid_files[group] == NULL) {
707                 CERROR("Object seq "LPU64" not fully setup; not updating "
708                        "last_objid\n", group);
709                 RETURN(-EINVAL);
710         }
711
712         CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
713                obd->obd_name, filter->fo_last_objids[group], group);
714
715         tmp = cpu_to_le64(filter->fo_last_objids[group]);
716         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
717                                  &tmp, sizeof(tmp), &off, force_sync);
718         if (rc)
719                 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
720                        group, rc);
721         RETURN(rc);
722 }
723 extern int ost_handle(struct ptlrpc_request *req);
724 /* assumes caller has already in kernel ctxt */
725 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
726 {
727         struct filter_obd *filter = &obd->u.filter;
728         struct lr_server_data *lsd;
729         struct lsd_client_data *lcd = NULL;
730         struct inode *inode = filp->f_dentry->d_inode;
731         unsigned long last_rcvd_size = i_size_read(inode);
732         struct lu_target *lut;
733         __u64 mount_count;
734         __u32 start_epoch;
735         int cl_idx;
736         loff_t off = 0;
737         int rc;
738
739         /* ensure padding in the struct is the correct size */
740         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
741                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
742         CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
743                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
744
745         /* allocate and initialize lu_target */
746         OBD_ALLOC_PTR(lut);
747         if (lut == NULL)
748                 RETURN(-ENOMEM);
749         rc = lut_init(NULL, lut, obd, NULL);
750         if (rc)
751                 GOTO(err_lut, rc);
752         lsd = class_server_data(obd);
753         if (last_rcvd_size == 0) {
754                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
755
756                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
757                 lsd->lsd_last_transno = 0;
758                 mount_count = lsd->lsd_mount_count = 0;
759                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
760                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
761                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
762                 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
763                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
764                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
765                  * MDS first connects and assigns the OST index number. */
766                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
767                                                         OBD_INCOMPAT_OST);
768         } else {
769                 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
770                 if (rc) {
771                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
772                                LAST_RCVD, rc);
773                         GOTO(err_lut, rc);
774                 }
775                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
776                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
777                                            "using the wrong disk %s. Were the "
778                                            "/dev/ assignments rearranged?\n",
779                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
780                         GOTO(err_lut, rc = -EINVAL);
781                 }
782                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
783                 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
784                 /* COMPAT_146 */
785                 /* Assume old last_rcvd format unless I_C_LR is set */
786                 if (!(lsd->lsd_feature_incompat &
787                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
788                         lsd->lsd_last_transno = lsd->lsd_compat14;
789                 /* end COMPAT_146 */
790                 /* OBD_COMPAT_OST is set in filter_connect_internal when the
791                  * MDS first connects and assigns the OST index number. */
792                 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
793                                                          OBD_INCOMPAT_OST);
794         }
795
796         if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
797                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
798                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
799                        ~FILTER_INCOMPAT_SUPP);
800                 GOTO(err_lut, rc = -EINVAL);
801         }
802         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
803                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
804                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
805                        ~FILTER_ROCOMPAT_SUPP);
806                 /* Do something like remount filesystem read-only */
807                 GOTO(err_lut, rc = -EINVAL);
808         }
809
810         start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
811
812         CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
813                obd->obd_name, start_epoch);
814         CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
815                obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
816         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
817                obd->obd_name, mount_count + 1);
818         CDEBUG(D_INODE, "%s: server data size: %u\n",
819                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
820         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
821                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
822         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
823                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
824         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
825                obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
826         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
827                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
828                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
829                 le16_to_cpu(lsd->lsd_client_size));
830
831         if (!obd->obd_replayable) {
832                 CWARN("%s: recovery support OFF\n", obd->obd_name);
833                 GOTO(out, rc = 0);
834         }
835
836         OBD_ALLOC_PTR(lcd);
837         if (!lcd)
838                 GOTO(err_client, rc = -ENOMEM);
839
840         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
841              off < last_rcvd_size; cl_idx++) {
842                 __u64 last_rcvd;
843                 struct obd_export *exp;
844                 struct filter_export_data *fed;
845
846                 /* Don't assume off is incremented properly by
847                  * fsfilt_read_record(), in case sizeof(*lcd)
848                  * isn't the same as lsd->lsd_client_size.  */
849                 off = le32_to_cpu(lsd->lsd_client_start) +
850                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
851                 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
852                 if (rc) {
853                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
854                                LAST_RCVD, cl_idx, off, rc);
855                         break; /* read error shouldn't cause startup to fail */
856                 }
857
858                 if (lcd->lcd_uuid[0] == '\0') {
859                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
860                                cl_idx);
861                         continue;
862                 }
863
864                 check_lcd(obd->obd_name, cl_idx, lcd);
865
866                 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
867
868                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
869                        " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
870                        last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
871
872                 /* These exports are cleaned up by filter_disconnect(), so they
873                  * need to be set up like real exports as filter_connect() does.
874                  */
875                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
876                 if (IS_ERR(exp)) {
877                         if (PTR_ERR(exp) == -EALREADY) {
878                                 /* export already exists, zero out this one */
879                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
880                                 continue;
881                         }
882                         OBD_FREE_PTR(lcd);
883                         GOTO(err_client, rc = PTR_ERR(exp));
884                 }
885
886                 fed = &exp->exp_filter_data;
887                 *fed->fed_ted.ted_lcd = *lcd;
888                 fed->fed_group = 0; /* will be assigned at connect */
889                 filter_export_stats_init(obd, exp, NULL);
890                 rc = filter_client_add(obd, exp, cl_idx);
891                 /* can't fail for existing client */
892                 LASSERTF(rc == 0, "rc = %d\n", rc);
893
894                 /* VBR: set export last committed */
895                 exp->exp_last_committed = last_rcvd;
896                 cfs_spin_lock(&exp->exp_lock);
897                 exp->exp_connecting = 0;
898                 exp->exp_in_recovery = 0;
899                 cfs_spin_unlock(&exp->exp_lock);
900                 obd->obd_max_recoverable_clients++;
901                 class_export_put(exp);
902
903                 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
904                         lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
905         }
906         OBD_FREE_PTR(lcd);
907
908         obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
909 out:
910         obd->u.obt.obt_mount_count = mount_count + 1;
911         obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
912         lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
913
914         /* save it, so mount count and last_transno is current */
915         rc = filter_update_server_data(obd);
916         if (rc)
917                 GOTO(err_client, rc);
918
919         RETURN(0);
920
921 err_client:
922         class_disconnect_exports(obd);
923 err_lut:
924         filter_free_server_data(&obd->u.obt);
925         RETURN(rc);
926 }
927
928 static int filter_cleanup_groups(struct obd_device *obd)
929 {
930         struct filter_obd *filter = &obd->u.filter;
931         struct file *filp;
932         struct dentry *dentry;
933         int i, j;
934         ENTRY;
935
936         if (filter->fo_dentry_O_groups != NULL) {
937                 for (i = 0; i < filter->fo_group_count; i++) {
938                         dentry = filter->fo_dentry_O_groups[i];
939                         if (dentry != NULL)
940                                 f_dput(dentry);
941                 }
942                 OBD_FREE(filter->fo_dentry_O_groups,
943                          filter->fo_group_count *
944                          sizeof(*filter->fo_dentry_O_groups));
945                 filter->fo_dentry_O_groups = NULL;
946         }
947         if (filter->fo_last_objid_files != NULL) {
948                 for (i = 0; i < filter->fo_group_count; i++) {
949                         filp = filter->fo_last_objid_files[i];
950                         if (filp != NULL)
951                                 filp_close(filp, 0);
952                 }
953                 OBD_FREE(filter->fo_last_objid_files,
954                          filter->fo_group_count *
955                          sizeof(*filter->fo_last_objid_files));
956                 filter->fo_last_objid_files = NULL;
957         }
958         if (filter->fo_dentry_O_sub != NULL) {
959                 for (i = 0; i < filter->fo_group_count; i++) {
960                         for (j = 0; j < filter->fo_subdir_count; j++) {
961                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
962                                 if (dentry != NULL)
963                                         f_dput(dentry);
964                         }
965                 }
966                 OBD_FREE(filter->fo_dentry_O_sub,
967                          filter->fo_group_count *
968                          sizeof(*filter->fo_dentry_O_sub));
969                 filter->fo_dentry_O_sub = NULL;
970         }
971         if (filter->fo_last_objids != NULL) {
972                 OBD_FREE(filter->fo_last_objids,
973                          filter->fo_group_count *
974                          sizeof(*filter->fo_last_objids));
975                 filter->fo_last_objids = NULL;
976         }
977         if (filter->fo_dentry_O != NULL) {
978                 f_dput(filter->fo_dentry_O);
979                 filter->fo_dentry_O = NULL;
980         }
981         RETURN(0);
982 }
983
984 static int filter_update_last_group(struct obd_device *obd, int group)
985 {
986         struct filter_obd *filter = &obd->u.filter;
987         struct file *filp = NULL;
988         int last_group = 0, rc;
989         loff_t off = 0;
990         ENTRY;
991
992         if (group <= filter->fo_committed_group)
993                 RETURN(0);
994
995         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
996         if (IS_ERR(filp)) {
997                 rc = PTR_ERR(filp);
998                 filp = NULL;
999                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1000                 GOTO(cleanup, rc);
1001         }
1002
1003         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1004         if (rc) {
1005                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1006                 GOTO(cleanup, rc);
1007         }
1008
1009         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1010                obd->obd_name, last_group, group);
1011
1012         off = 0;
1013         last_group = group;
1014         /* must be sync: bXXXX */
1015         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1016         if (rc) {
1017                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1018                 GOTO(cleanup, rc);
1019         }
1020
1021         filter->fo_committed_group = group;
1022 cleanup:
1023         if (filp)
1024                 filp_close(filp, 0);
1025         RETURN(rc);
1026 }
1027
1028 static int filter_read_group_internal(struct obd_device *obd, int group,
1029                                       int create)
1030 {
1031         struct filter_obd *filter = &obd->u.filter;
1032         __u64 *new_objids = NULL;
1033         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1034         struct dentry **new_groups = NULL;
1035         struct file **new_files = NULL;
1036         struct dentry *dentry;
1037         struct file *filp;
1038         int old_count = filter->fo_group_count, rc, stage = 0, i;
1039         char name[25];
1040         __u64 last_objid;
1041         loff_t off = 0;
1042         int len = group + 1;
1043
1044         snprintf(name, 24, "%d", group);
1045         name[24] = '\0';
1046
1047         if (!create) {
1048                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1049                                            strlen(name));
1050                 if (IS_ERR(dentry)) {
1051                         CERROR("Cannot lookup expected object group %d: %ld\n",
1052                                group, PTR_ERR(dentry));
1053                         RETURN(PTR_ERR(dentry));
1054                 }
1055         } else {
1056                 dentry = simple_mkdir(filter->fo_dentry_O,
1057                                       obd->u.obt.obt_vfsmnt, name, 0700, 1);
1058                 if (IS_ERR(dentry)) {
1059                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1060                                PTR_ERR(dentry));
1061                         RETURN(PTR_ERR(dentry));
1062                 }
1063         }
1064         stage = 1;
1065
1066         snprintf(name, 24, "O/%d/LAST_ID", group);
1067         name[24] = '\0';
1068         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1069         if (IS_ERR(filp)) {
1070                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1071                 GOTO(cleanup, rc = PTR_ERR(filp));
1072         }
1073         stage = 2;
1074
1075         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1076         if (rc) {
1077                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1078                 GOTO(cleanup, rc);
1079         }
1080
1081         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1082                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1083                 if (tmp_subdirs == NULL)
1084                         GOTO(cleanup, rc = -ENOMEM);
1085                 stage = 3;
1086
1087                 for (i = 0; i < filter->fo_subdir_count; i++) {
1088                         char dir[20];
1089                         snprintf(dir, sizeof(dir), "d%u", i);
1090
1091                         tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1092                                                               obd->u.obt.obt_vfsmnt,
1093                                                               dir, 0700, 1);
1094                         if (IS_ERR(tmp_subdirs->dentry[i])) {
1095                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1096                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1097                                        group, dir, rc);
1098                                 GOTO(cleanup, rc);
1099                         }
1100
1101                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1102                                tmp_subdirs->dentry[i]);
1103                 }
1104         }
1105
1106         /* 'group' is an index; we need an array of length 'group + 1' */
1107         if (group + 1 > old_count) {
1108                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1109                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1110                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1111                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1112                 stage = 4;
1113                 if (new_objids == NULL || new_subdirs == NULL ||
1114                     new_groups == NULL || new_files == NULL)
1115                         GOTO(cleanup, rc = -ENOMEM);
1116
1117                 if (old_count) {
1118                         memcpy(new_objids, filter->fo_last_objids,
1119                                old_count * sizeof(*new_objids));
1120                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1121                                old_count * sizeof(*new_subdirs));
1122                         memcpy(new_groups, filter->fo_dentry_O_groups,
1123                                old_count * sizeof(*new_groups));
1124                         memcpy(new_files, filter->fo_last_objid_files,
1125                                old_count * sizeof(*new_files));
1126
1127                         OBD_FREE(filter->fo_last_objids,
1128                                  old_count * sizeof(*new_objids));
1129                         OBD_FREE(filter->fo_dentry_O_sub,
1130                                  old_count * sizeof(*new_subdirs));
1131                         OBD_FREE(filter->fo_dentry_O_groups,
1132                                  old_count * sizeof(*new_groups));
1133                         OBD_FREE(filter->fo_last_objid_files,
1134                                  old_count * sizeof(*new_files));
1135                 }
1136                 filter->fo_last_objids = new_objids;
1137                 filter->fo_dentry_O_sub = new_subdirs;
1138                 filter->fo_dentry_O_groups = new_groups;
1139                 filter->fo_last_objid_files = new_files;
1140                 filter->fo_group_count = len;
1141         }
1142
1143         filter->fo_dentry_O_groups[group] = dentry;
1144         filter->fo_last_objid_files[group] = filp;
1145         if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1146                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1147                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1148         }
1149
1150         filter_update_last_group(obd, group);
1151
1152         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1153                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1154                 rc = filter_update_last_objid(obd, group, 1);
1155                 RETURN(rc);
1156         }
1157
1158         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1159         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1160                obd->obd_name, group, last_objid);
1161         RETURN(0);
1162  cleanup:
1163         switch (stage) {
1164         case 4:
1165                 if (new_objids != NULL)
1166                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1167                 if (new_subdirs != NULL)
1168                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1169                 if (new_groups != NULL)
1170                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1171                 if (new_files != NULL)
1172                         OBD_FREE(new_files, len * sizeof(*new_files));
1173         case 3:
1174                 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1175                         for (i = 0; i < filter->fo_subdir_count; i++) {
1176                                 if (tmp_subdirs->dentry[i] != NULL)
1177                                         dput(tmp_subdirs->dentry[i]);
1178                         }
1179                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1180                 }
1181         case 2:
1182                 filp_close(filp, 0);
1183         case 1:
1184                 dput(dentry);
1185         }
1186         RETURN(rc);
1187 }
1188
1189 static int filter_read_groups(struct obd_device *obd, int last_group,
1190                               int create)
1191 {
1192         struct filter_obd *filter = &obd->u.filter;
1193         int old_count, group, rc = 0;
1194
1195         cfs_mutex_lock(&filter->fo_init_lock);
1196         old_count = filter->fo_group_count;
1197         for (group = old_count; group <= last_group; group++) {
1198                 rc = filter_read_group_internal(obd, group, create);
1199                 if (rc != 0)
1200                         break;
1201         }
1202         cfs_mutex_unlock(&filter->fo_init_lock);
1203         return rc;
1204 }
1205
1206 /* FIXME: object groups */
1207 static int filter_prep_groups(struct obd_device *obd)
1208 {
1209         struct filter_obd *filter = &obd->u.filter;
1210         struct dentry *O_dentry;
1211         struct file *filp;
1212         int    last_group, rc = 0, cleanup_phase = 0;
1213         loff_t off = 0;
1214         ENTRY;
1215
1216         O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1217                                 "O", 0700, 1);
1218         CDEBUG(D_INODE, "%s: got/created O: %p\n", obd->obd_name, O_dentry);
1219         if (IS_ERR(O_dentry)) {
1220                 rc = PTR_ERR(O_dentry);
1221                 CERROR("%s: cannot open/create O: rc = %d\n", obd->obd_name,rc);
1222                 GOTO(cleanup, rc);
1223         }
1224         filter->fo_dentry_O = O_dentry;
1225         cleanup_phase = 1; /* O_dentry */
1226
1227         /* we have to initialize all groups before first connections from
1228          * clients because they may send create/destroy for any group -bzzz */
1229         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1230         if (IS_ERR(filp)) {
1231                 CERROR("%s: cannot create LAST_GROUP: rc = %ld\n",
1232                        obd->obd_name, PTR_ERR(filp));
1233                 GOTO(cleanup, rc = PTR_ERR(filp));
1234         }
1235         cleanup_phase = 2; /* filp */
1236
1237         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1238         if (rc) {
1239                 CERROR("%s: error reading LAST_GROUP: rc %d\n",
1240                        obd->obd_name, rc);
1241                 GOTO(cleanup, rc);
1242         }
1243
1244         if (off == 0)
1245                 last_group = FID_SEQ_OST_MDT0;
1246
1247         CDEBUG(D_INODE, "%s: initialize group %u (max %u)\n", obd->obd_name,
1248                FID_SEQ_OST_MDT0, last_group);
1249         filter->fo_committed_group = last_group;
1250         rc = filter_read_groups(obd, last_group, 1);
1251         if (rc)
1252                 GOTO(cleanup, rc);
1253
1254         filp_close(filp, 0);
1255         RETURN(0);
1256
1257  cleanup:
1258         switch (cleanup_phase) {
1259         case 2:
1260                 filp_close(filp, 0);
1261         case 1:
1262                 filter_cleanup_groups(obd);
1263                 f_dput(filter->fo_dentry_O);
1264                 filter->fo_dentry_O = NULL;
1265         default:
1266                 break;
1267         }
1268         return rc;
1269
1270 }
1271
1272 /* setup the object store with correct subdirectories */
1273 static int filter_prep(struct obd_device *obd)
1274 {
1275         struct lvfs_run_ctxt saved;
1276         struct filter_obd *filter = &obd->u.filter;
1277         struct file *file;
1278         struct inode *inode;
1279         int rc = 0;
1280         ENTRY;
1281
1282         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1283         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1284         if (!file || IS_ERR(file)) {
1285                 rc = PTR_ERR(file);
1286                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1287                        LAST_RCVD, rc);
1288                 GOTO(out, rc);
1289         }
1290         obd->u.obt.obt_rcvd_filp = file;
1291         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1292                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1293                        file->f_dentry->d_inode->i_mode);
1294                 GOTO(err_filp, rc = -ENOENT);
1295         }
1296
1297         inode = file->f_dentry->d_parent->d_inode;
1298         /* We use i_op->unlink directly in filter_vfs_unlink() */
1299         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1300                 CERROR("%s: filesystem does not support create/unlink ops\n",
1301                        obd->obd_name);
1302                 GOTO(err_filp, rc = -EOPNOTSUPP);
1303         }
1304
1305         rc = filter_init_server_data(obd, file);
1306         if (rc) {
1307                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1308                 GOTO(err_filp, rc);
1309         }
1310         LASSERT(obd->u.obt.obt_lut);
1311         target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1312
1313         /* open and create health check io file*/
1314         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1315         if (IS_ERR(file)) {
1316                 rc = PTR_ERR(file);
1317                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1318                        HEALTH_CHECK, rc);
1319                 GOTO(err_server_data, rc);
1320         }
1321         filter->fo_obt.obt_health_check_filp = file;
1322         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1323                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1324                        file->f_dentry->d_inode->i_mode);
1325                 GOTO(err_health_check, rc = -ENOENT);
1326         }
1327         rc = lvfs_check_io_health(obd, file);
1328         if (rc)
1329                 GOTO(err_health_check, rc);
1330
1331         rc = filter_prep_groups(obd);
1332         if (rc)
1333                 GOTO(err_health_check, rc);
1334 out:
1335         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1336
1337         return(rc);
1338
1339 err_health_check:
1340         if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1341                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1342         filter->fo_obt.obt_health_check_filp = NULL;
1343 err_server_data:
1344         target_recovery_fini(obd);
1345         filter_free_server_data(&obd->u.obt);
1346 err_filp:
1347         if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1348                 CERROR("can't close %s after error\n", LAST_RCVD);
1349         obd->u.obt.obt_rcvd_filp = NULL;
1350         goto out;
1351 }
1352
1353 /* cleanup the filter: write last used object id to status file */
1354 static void filter_post(struct obd_device *obd)
1355 {
1356         struct lvfs_run_ctxt saved;
1357         struct filter_obd *filter = &obd->u.filter;
1358         int rc, i;
1359
1360         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1361          * best to start a transaction with h_sync, because we removed this
1362          * from lastobjid */
1363
1364         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1365         rc = filter_update_server_data(obd);
1366         if (rc)
1367                 CERROR("error writing server data: rc = %d\n", rc);
1368
1369         for (i = 0; i < filter->fo_group_count; i++) {
1370                 rc = filter_update_last_objid(obd, i,
1371                                 (i == filter->fo_group_count - 1));
1372                 if (rc)
1373                         CERROR("error writing group %d lastobjid: rc = %d\n",
1374                                i, rc);
1375         }
1376
1377         rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1378         obd->u.obt.obt_rcvd_filp = NULL;
1379         if (rc)
1380                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1381
1382         rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1383         filter->fo_obt.obt_health_check_filp = NULL;
1384         if (rc)
1385                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1386
1387         filter_cleanup_groups(obd);
1388         filter_free_server_data(&obd->u.obt);
1389         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1390
1391         filter_free_capa_keys(filter);
1392         cleanup_capa_hash(filter->fo_capa_hash);
1393 }
1394
1395 static void filter_set_last_id(struct filter_obd *filter,
1396                                obd_id id, obd_seq group)
1397 {
1398         LASSERT(group <= filter->fo_group_count);
1399
1400         cfs_spin_lock(&filter->fo_objidlock);
1401         filter->fo_last_objids[group] = id;
1402         cfs_spin_unlock(&filter->fo_objidlock);
1403 }
1404
1405 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1406 {
1407         obd_id id;
1408         LASSERT(group <= filter->fo_group_count);
1409         LASSERT(filter->fo_last_objids != NULL);
1410
1411         /* FIXME: object groups */
1412         cfs_spin_lock(&filter->fo_objidlock);
1413         id = filter->fo_last_objids[group];
1414         cfs_spin_unlock(&filter->fo_objidlock);
1415         return id;
1416 }
1417
1418 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1419 {
1420         mutex_lock_nested(&dparent->d_inode->i_mutex, I_MUTEX_PARENT);
1421         return 0;
1422 }
1423
1424 /* We never dget the object parent, so DON'T dput it either */
1425 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1426 {
1427         struct filter_obd *filter = &obd->u.filter;
1428         struct filter_subdirs *subdirs;
1429
1430         if (group >= filter->fo_group_count) /* FIXME: object groups */
1431                 return ERR_PTR(-EBADF);
1432
1433         if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1434                 return filter->fo_dentry_O_groups[group];
1435
1436         subdirs = &filter->fo_dentry_O_sub[group];
1437         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1438 }
1439
1440 /* We never dget the object parent, so DON'T dput it either */
1441 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1442                                   obd_id objid)
1443 {
1444         unsigned long now = jiffies;
1445         struct dentry *dparent = filter_parent(obd, group, objid);
1446         int rc;
1447
1448         if (IS_ERR(dparent))
1449                 return dparent;
1450         if (dparent == NULL)
1451                 return ERR_PTR(-ENOENT);
1452
1453         rc = filter_lock_dentry(obd, dparent);
1454         fsfilt_check_slow(obd, now, "parent lock");
1455         return rc ? ERR_PTR(rc) : dparent;
1456 }
1457
1458 /* We never dget the object parent, so DON'T dput it either */
1459 static void filter_parent_unlock(struct dentry *dparent)
1460 {
1461         mutex_unlock(&dparent->d_inode->i_mutex);
1462 }
1463
1464 /* How to get files, dentries, inodes from object id's.
1465  *
1466  * If dir_dentry is passed, the caller has already locked the parent
1467  * appropriately for this operation (normally a write lock).  If
1468  * dir_dentry is NULL, we do a read lock while we do the lookup to
1469  * avoid races with create/destroy and such changing the directory
1470  * internal to the filesystem code. */
1471 struct dentry *filter_fid2dentry(struct obd_device *obd,
1472                                  struct dentry *dir_dentry,
1473                                  obd_seq group, obd_id id)
1474 {
1475         struct dentry *dparent = dir_dentry;
1476         struct dentry *dchild;
1477         char name[32];
1478         int len;
1479         ENTRY;
1480
1481         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1482             obd->u.filter.fo_destroys_in_progress == 0) {
1483                 /* don't fail lookups for orphan recovery, it causes
1484                  * later LBUGs when objects still exist during precreate */
1485                 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1486                 RETURN(ERR_PTR(-ENOENT));
1487         }
1488         if (id == 0) {
1489                 CERROR("fatal: invalid object id 0\n");
1490                 RETURN(ERR_PTR(-ESTALE));
1491         }
1492
1493         len = sprintf(name, LPU64, id);
1494         if (dir_dentry == NULL) {
1495                 dparent = filter_parent_lock(obd, group, id);
1496                 if (IS_ERR(dparent)) {
1497                         CERROR("%s: error getting object "POSTID
1498                                " parent: rc %ld\n", obd->obd_name,
1499                                id, group, PTR_ERR(dparent));
1500                         RETURN(dparent);
1501                 }
1502         }
1503         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1504                dparent->d_name.len, dparent->d_name.name, name);
1505         /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1506         dchild = lookup_one_len(name, dparent, len);
1507         if (dir_dentry == NULL)
1508                 filter_parent_unlock(dparent);
1509         if (IS_ERR(dchild)) {
1510                 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1511                        obd->obd_name, id, group, PTR_ERR(dchild));
1512                 RETURN(dchild);
1513         }
1514
1515         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1516                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1517                        obd->obd_name, id, dchild->d_inode->i_ino);
1518                 f_dput(dchild);
1519                 RETURN(ERR_PTR(-ENOENT));
1520         }
1521
1522         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1523                name, dchild, atomic_read(&dchild->d_count));
1524
1525         LASSERT(atomic_read(&dchild->d_count) > 0);
1526
1527         RETURN(dchild);
1528 }
1529
1530 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1531                                   obd_id group, struct lustre_handle *lockh)
1532 {
1533         int flags = LDLM_AST_DISCARD_DATA, rc;
1534         struct ldlm_res_id res_id;
1535         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1536         ENTRY;
1537
1538         osc_build_res_name(objid, group, &res_id);
1539         /* Tell the clients that the object is gone now and that they should
1540          * throw away any cached pages. */
1541         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1542                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1543                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1544                                     lockh);
1545         if (rc != ELDLM_OK)
1546                 lockh->cookie = 0;
1547         RETURN(rc);
1548 }
1549
1550 static void filter_fini_destroy(struct obd_device *obd,
1551                                 struct lustre_handle *lockh)
1552 {
1553         if (lustre_handle_is_used(lockh))
1554                 ldlm_lock_decref(lockh, LCK_PW);
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(mutex_trylock(&dir->i_mutex) == 0);
1572         LASSERT(mutex_trylock(&dentry->d_inode->i_mutex) == 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 != cfs_curproc_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         /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
1593         ll_vfs_dq_init(dir);
1594
1595         rc = ll_security_inode_unlink(dir, dentry, mnt);
1596         if (rc)
1597                 GOTO(out, rc);
1598
1599         rc = dir->i_op->unlink(dir, dentry);
1600 out:
1601         /* need to drop i_mutex before we lose inode reference */
1602         mutex_unlock(&dentry->d_inode->i_mutex);
1603         if (rc == 0)
1604                 d_delete(dentry);
1605
1606         RETURN(rc);
1607 }
1608
1609 /* Caller must hold LCK_PW on parent and push us into kernel context.
1610  * Caller must hold child i_mutex, we drop it always.
1611  * Caller is also required to ensure that dchild->d_inode exists. */
1612 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1613                                    obd_seq group, struct dentry *dparent,
1614                                    struct dentry *dchild)
1615 {
1616         struct inode *inode = dchild->d_inode;
1617         int rc;
1618
1619         /* There should be 2 references to the inode:
1620          *  1) taken by filter_prepare_destroy
1621          *  2) taken by filter_destroy */
1622         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
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, obd->u.obt.obt_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         cfs_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         struct ptlrpc_request *req = req_cookie;
1688         struct ldlm_lock *lock = *lockp, *l = NULL;
1689         struct ldlm_resource *res = lock->l_resource;
1690         ldlm_processing_policy policy;
1691         struct ost_lvb *res_lvb, *reply_lvb;
1692         struct ldlm_reply *rep;
1693         ldlm_error_t err;
1694         int idx, rc, tmpflags = 0, only_liblustre = 1;
1695         struct ldlm_interval_tree *tree;
1696         struct filter_intent_args arg;
1697         __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1698                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1699                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1700         struct ldlm_glimpse_work        gl_work;
1701         CFS_LIST_HEAD(gl_list);
1702         ENTRY;
1703
1704         policy = ldlm_get_processing_policy(res);
1705         LASSERT(policy != NULL);
1706         LASSERT(req != NULL);
1707
1708         rc = lustre_pack_reply(req, 3, repsize, NULL);
1709         if (rc)
1710                 RETURN(req->rq_status = rc);
1711
1712         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1713         LASSERT(rep != NULL);
1714
1715         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1716                                    sizeof(*reply_lvb));
1717         LASSERT(reply_lvb != NULL);
1718
1719         //fixup_handle_for_resent_req(req, lock, &lockh);
1720
1721         /* Call the extent policy function to see if our request can be
1722          * granted, or is blocked.
1723          * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1724          * lock, and should not be granted if the lock will be blocked.
1725          */
1726
1727         if (flags & LDLM_FL_BLOCK_NOWAIT) {
1728                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_AGL_DELAY, 5);
1729
1730                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_AGL_NOLOCK))
1731                         RETURN(ELDLM_LOCK_ABORTED);
1732         }
1733
1734         LASSERT(ns == ldlm_res_to_ns(res));
1735         lock_res(res);
1736         rc = policy(lock, &tmpflags, 0, &err, NULL);
1737         check_res_locked(res);
1738
1739         /* The lock met with no resistance; we're finished. */
1740         if (rc == LDLM_ITER_CONTINUE) {
1741                 /* do not grant locks to the liblustre clients: they cannot
1742                  * handle ASTs robustly.  We need to do this while still
1743                  * holding lr_lock to avoid the lock remaining on the res_link
1744                  * list (and potentially being added to l_pending_list by an
1745                  * AST) when we are going to drop this lock ASAP. */
1746                 if (lock->l_export->exp_libclient ||
1747                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1748                         ldlm_resource_unlink_lock(lock);
1749                         err = ELDLM_LOCK_ABORTED;
1750                 } else {
1751                         err = ELDLM_LOCK_REPLACED;
1752                 }
1753                 unlock_res(res);
1754                 RETURN(err);
1755         } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1756                 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1757                  * callback for glimpse size. The real size user will trigger
1758                  * the glimpse callback when necessary. */
1759                 unlock_res(res);
1760                 RETURN(ELDLM_LOCK_ABORTED);
1761         }
1762
1763         /* Do not grant any lock, but instead send GL callbacks.  The extent
1764          * policy nicely created a list of all PW locks for us.  We will choose
1765          * the highest of those which are larger than the size in the LVB, if
1766          * any, and perform a glimpse callback. */
1767         res_lvb = res->lr_lvb_data;
1768         LASSERT(res_lvb != NULL);
1769         *reply_lvb = *res_lvb;
1770
1771         /*
1772          * lr_lock guarantees that no new locks are granted, and,
1773          * therefore, that res->lr_lvb_data cannot increase beyond the
1774          * end of already granted lock. As a result, it is safe to
1775          * check against "stale" reply_lvb->lvb_size value without
1776          * res->lr_lvb_mutex.
1777          */
1778         arg.size = reply_lvb->lvb_size;
1779         arg.victim = &l;
1780         arg.liblustre = &only_liblustre;
1781         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1782                 tree = &res->lr_itree[idx];
1783                 if (tree->lit_mode == LCK_PR)
1784                         continue;
1785
1786                 interval_iterate_reverse(tree->lit_root,
1787                                          filter_intent_cb, &arg);
1788         }
1789         unlock_res(res);
1790
1791         /* There were no PW locks beyond the size in the LVB; finished. */
1792         if (l == NULL) {
1793                 if (only_liblustre) {
1794                         /* If we discovered a liblustre client with a PW lock,
1795                          * however, the LVB may be out of date!  The LVB is
1796                          * updated only on glimpse (which we don't do for
1797                          * liblustre clients) and cancel (which the client
1798                          * obviously has not yet done).  So if it has written
1799                          * data but kept the lock, the LVB is stale and needs
1800                          * to be updated from disk.
1801                          *
1802                          * Of course, this will all disappear when we switch to
1803                          * taking liblustre locks on the OST. */
1804                         ldlm_res_lvbo_update(res, NULL, 1);
1805                 }
1806                 RETURN(ELDLM_LOCK_ABORTED);
1807         }
1808
1809         /*
1810          * This check is for lock taken in filter_prepare_destroy() that does
1811          * not have l_glimpse_ast set. So the logic is: if there is a lock
1812          * with no l_glimpse_ast set, this object is being destroyed already.
1813          *
1814          * Hence, if you are grabbing DLM locks on the server, always set
1815          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1816          */
1817         if (l->l_glimpse_ast == NULL) {
1818                 /* We are racing with unlink(); just return -ENOENT */
1819                 rep->lock_policy_res1 = -ENOENT;
1820                 goto out;
1821         }
1822
1823         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1824
1825         /* Populate the gl_work structure.
1826          * Grab additional reference on the lock which will be released in
1827          * ldlm_work_gl_ast_lock() */
1828         gl_work.gl_lock = LDLM_LOCK_GET(l);
1829         /* The glimpse callback is sent to one single extent lock. As a result,
1830          * the gl_work list is just composed of one element */
1831         cfs_list_add_tail(&gl_work.gl_list, &gl_list);
1832         /* There is actually no need for a glimpse descriptor when glimpsing
1833          * extent locks */
1834         gl_work.gl_desc = NULL;
1835         /* the ldlm_glimpse_work structure is allocated on the stack */
1836         gl_work.gl_flags = LDLM_GL_WORK_NOFREE;
1837
1838         rc = ldlm_glimpse_locks(res, &gl_list); /* this will update the LVB */
1839
1840         if (!cfs_list_empty(&gl_list))
1841                 LDLM_LOCK_RELEASE(l);
1842
1843         lock_res(res);
1844         *reply_lvb = *res_lvb;
1845         unlock_res(res);
1846
1847  out:
1848         LDLM_LOCK_RELEASE(l);
1849
1850         RETURN(ELDLM_LOCK_ABORTED);
1851 }
1852
1853 /*
1854  * per-obd_device iobuf pool.
1855  *
1856  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1857  * allocations in write-path has to be minimized (see bug 5137).
1858  *
1859  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1860  * OST threads (see ost_thread_{init,done}()).
1861  *
1862  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1863  * at the OST layer there are only (potentially) multiple obd_device of type
1864  * unknown at the time of OST thread creation.
1865  *
1866  * We create a cfs_hash for struct filter_obd (->fo_iobuf_hash field) on
1867  * initializing, each OST thread will create it's own iobuf on the first
1868  * access and insert it into ->fo_iobuf_hash with thread ID as key,
1869  * so the iobuf can be found again by thread ID.
1870  *
1871  * Functions below
1872  *
1873  *     filter_iobuf_pool_init()
1874  *
1875  *     filter_iobuf_pool_done()
1876  *
1877  *     filter_iobuf_get()
1878  *
1879  * operate on this array. They are "generic" in a sense that they don't depend
1880  * on actual type of iobuf's (the latter depending on Linux kernel version).
1881  */
1882
1883 /*
1884  * destroy pool created by filter_iobuf_pool_init
1885  */
1886 static void filter_iobuf_pool_done(struct filter_obd *filter)
1887 {
1888         ENTRY;
1889
1890         if (filter->fo_iobuf_hash != NULL) {
1891                 cfs_hash_putref(filter->fo_iobuf_hash);
1892                 filter->fo_iobuf_hash = NULL;
1893         }
1894         EXIT;
1895 }
1896
1897 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1898 {
1899         struct filter_obd       *filter = &obd->u.filter;
1900         struct sptlrpc_rule_set  tmp_rset;
1901         int                      rc;
1902
1903         sptlrpc_rule_set_init(&tmp_rset);
1904         rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1905         if (rc) {
1906                 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1907                        obd->obd_name, rc);
1908                 return rc;
1909         }
1910
1911         sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1912
1913         cfs_write_lock(&filter->fo_sptlrpc_lock);
1914         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1915         filter->fo_sptlrpc_rset = tmp_rset;
1916         cfs_write_unlock(&filter->fo_sptlrpc_lock);
1917
1918         return 0;
1919 }
1920
1921 static unsigned
1922 filter_iobuf_hop_hash(cfs_hash_t *hs, const void *key, unsigned mask)
1923 {
1924         __u64   val = *((__u64 *)key);
1925
1926         return cfs_hash_long(val, hs->hs_cur_bits);
1927 }
1928
1929 static void *
1930 filter_iobuf_hop_key(cfs_hlist_node_t *hnode)
1931 {
1932         struct filter_iobuf     *pool;
1933
1934         pool = cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist);
1935         return &pool->dr_hkey;
1936 }
1937
1938 static int
1939 filter_iobuf_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
1940 {
1941         struct filter_iobuf     *pool;
1942
1943         pool = cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist);
1944         return pool->dr_hkey == *((__u64 *)key);
1945 }
1946
1947 static void *
1948 filter_iobuf_hop_object(cfs_hlist_node_t *hnode)
1949 {
1950         return cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist);
1951 }
1952
1953 static void
1954 filter_iobuf_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1955 {
1956         /* dummy, required by cfs_hash */
1957 }
1958
1959 static void
1960 filter_iobuf_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1961 {
1962         /* dummy, required by cfs_hash */
1963 }
1964
1965 static void
1966 filter_iobuf_hop_exit(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1967 {
1968         struct filter_iobuf     *pool;
1969
1970         pool = cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist);
1971         filter_free_iobuf(pool);
1972 }
1973
1974 static struct cfs_hash_ops filter_iobuf_hops = {
1975         .hs_hash        = filter_iobuf_hop_hash,
1976         .hs_key         = filter_iobuf_hop_key,
1977         .hs_keycmp      = filter_iobuf_hop_keycmp,
1978         .hs_object      = filter_iobuf_hop_object,
1979         .hs_get         = filter_iobuf_hop_get,
1980         .hs_put_locked  = filter_iobuf_hop_put_locked,
1981         .hs_exit        = filter_iobuf_hop_exit
1982 };
1983
1984 #define FILTER_IOBUF_HASH_BITS  9
1985 #define FILTER_IOBUF_HBKT_BITS  4
1986
1987 /*
1988  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1989  */
1990 static int filter_iobuf_pool_init(struct filter_obd *filter)
1991 {
1992         filter->fo_iobuf_hash = cfs_hash_create("filter_iobuf",
1993                                                 FILTER_IOBUF_HASH_BITS,
1994                                                 FILTER_IOBUF_HASH_BITS,
1995                                                 FILTER_IOBUF_HBKT_BITS, 0,
1996                                                 CFS_HASH_MIN_THETA,
1997                                                 CFS_HASH_MAX_THETA,
1998                                                 &filter_iobuf_hops,
1999                                                 CFS_HASH_RW_BKTLOCK |
2000                                                 CFS_HASH_NO_ITEMREF);
2001
2002         return filter->fo_iobuf_hash != NULL ? 0 : -ENOMEM;
2003 }
2004
2005 /* Return iobuf allocated for @thread_id.
2006  * If we haven't allocated a pool entry for this thread before, do so now and
2007  * insert it into fo_iobuf_hash, otherwise we can find it from fo_iobuf_hash */
2008 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
2009 {
2010         struct filter_iobuf     *pool = NULL;
2011         __u64                   key = 0;
2012         int                     thread_id;
2013         int                     rc;
2014
2015         thread_id = (oti && oti->oti_thread) ? oti->oti_thread->t_id : -1;
2016         if (thread_id >= 0) {
2017                 struct ptlrpc_service_part *svcpt;
2018
2019                 svcpt = oti->oti_thread->t_svcpt;
2020                 LASSERT(svcpt != NULL);
2021
2022                 key = (__u64)(svcpt->scp_cpt) << 32 | thread_id;
2023                 pool = cfs_hash_lookup(filter->fo_iobuf_hash, &key);
2024                 if (pool != NULL)
2025                         return pool;
2026         }
2027
2028         pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE, PTLRPC_MAX_BRW_PAGES);
2029         if (pool == NULL)
2030                 return NULL;
2031
2032         if (thread_id >= 0) {
2033                 pool->dr_hkey = key;
2034                 rc = cfs_hash_add_unique(filter->fo_iobuf_hash,
2035                                          &key, &pool->dr_hlist);
2036                 /* ptlrpc service thould guarantee thread ID is unique */
2037                 LASSERT(rc != -EALREADY);
2038         }
2039
2040         return pool;
2041 }
2042
2043 /* mount the file system (secretly).  lustre_cfg parameters are:
2044  * 1 = device
2045  * 2 = fstype
2046  * 3 = flags: failover=f, failout=n
2047  * 4 = mount options
2048  */
2049 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
2050                         void *option)
2051 {
2052         struct filter_obd *filter = &obd->u.filter;
2053         struct vfsmount *mnt;
2054         struct file_system_type *type;
2055         struct lustre_mount_info *lmi;
2056         struct obd_uuid uuid;
2057         __u8 *uuid_ptr;
2058         char *str, *label;
2059         char ns_name[48];
2060         struct request_queue *q;
2061         int rc, i;
2062         ENTRY;
2063
2064         if (lcfg->lcfg_bufcount < 3 ||
2065             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
2066             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
2067                 RETURN(-EINVAL);
2068
2069         lmi = server_get_mount(obd->obd_name);
2070         if (lmi) {
2071                 /* We already mounted in lustre_fill_super.
2072                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2073                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2074                 mnt = lmi->lmi_mnt;
2075                 obd->obd_fsops = fsfilt_get_ops(lsi->lsi_fstype);
2076         } else {
2077                 /* old path - used by lctl */
2078                 CERROR("Using old MDS mount method\n");
2079                 type = get_fs_type(lustre_cfg_string(lcfg, 2));
2080                 if (!type) {
2081                         CERROR("get_fs_type failed\n");
2082                         RETURN(-ENODEV);
2083                 }
2084                 mnt = vfs_kern_mount(type, MS_NOATIME|MS_NODIRATIME,
2085                                      lustre_cfg_string(lcfg, 1), option);
2086                 cfs_module_put(type->owner);
2087                 if (IS_ERR(mnt)) {
2088                         rc = PTR_ERR(mnt);
2089                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2090                                            lustre_cfg_string(lcfg, 1), rc);
2091                         RETURN(rc);
2092                 }
2093
2094                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2095         }
2096         if (IS_ERR(obd->obd_fsops))
2097                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2098
2099         rc = filter_iobuf_pool_init(filter);
2100         if (rc != 0)
2101                 GOTO(err_ops, rc);
2102
2103         if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2104                 CERROR("%s: Underlying device is marked as read-only. "
2105                        "Setup failed\n", obd->obd_name);
2106                 GOTO(err_ops, rc = -EROFS);
2107         }
2108
2109         /* failover is the default */
2110         obd->obd_replayable = 1;
2111
2112         /* disable connection until configuration finishes */
2113         obd->obd_no_conn = 1;
2114
2115         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2116                 str = lustre_cfg_string(lcfg, 3);
2117                 if (strchr(str, 'n')) {
2118                         CWARN("%s: recovery disabled\n", obd->obd_name);
2119                         obd->obd_replayable = 0;
2120                 }
2121         }
2122
2123         obd->u.obt.obt_magic = OBT_MAGIC;
2124         obd->u.obt.obt_vfsmnt = mnt;
2125         obd->u.obt.obt_sb = mnt->mnt_sb;
2126         filter->fo_fstype = mnt->mnt_sb->s_type->name;
2127         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2128
2129         rc = fsfilt_setup(obd, obd->u.obt.obt_sb);
2130         if (rc)
2131                 GOTO(err_ops, rc);
2132
2133         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2134         obd->obd_lvfs_ctxt.pwdmnt = mnt;
2135         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2136         obd->obd_lvfs_ctxt.fs = get_ds();
2137         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2138
2139         cfs_mutex_init(&filter->fo_init_lock);
2140         filter->fo_committed_group = 0;
2141         filter->fo_destroys_in_progress = 0;
2142         for (i = 0; i < 32; i++)
2143                 cfs_mutex_init(&filter->fo_create_locks[i]);
2144
2145         cfs_spin_lock_init(&filter->fo_objidlock);
2146         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2147         cfs_mutex_init(&filter->fo_alloc_lock);
2148         init_brw_stats(&filter->fo_filter_stats);
2149         cfs_spin_lock_init(&filter->fo_flags_lock);
2150         filter->fo_read_cache = 1; /* enable read-only cache by default */
2151         filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2152         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2153         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2154         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2155         filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2156         filter_slc_set(filter); /* initialize sync on lock cancel */
2157
2158         rc = filter_prep(obd);
2159         if (rc)
2160                 GOTO(err_ops, rc);
2161
2162         CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2163         cfs_spin_lock_init(&filter->fo_llog_list_lock);
2164
2165         filter->fo_fl_oss_capa = 1;
2166
2167         CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2168         filter->fo_capa_hash = init_capa_hash();
2169         if (filter->fo_capa_hash == NULL)
2170                 GOTO(err_post, rc = -ENOMEM);
2171
2172         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2173         obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2174                                                 LDLM_NAMESPACE_SERVER,
2175                                                 LDLM_NAMESPACE_GREEDY,
2176                                                 LDLM_NS_TYPE_OST);
2177         if (obd->obd_namespace == NULL)
2178                 GOTO(err_post, rc = -ENOMEM);
2179         obd->obd_namespace->ns_lvbp = obd;
2180         obd->obd_namespace->ns_lvbo = &filter_lvbo;
2181         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2182
2183         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2184                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2185
2186         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2187         if (rc) {
2188                 CERROR("failed to setup llogging subsystems\n");
2189                 GOTO(err_post, rc);
2190         }
2191
2192         cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2193         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2194         /* do this after llog being initialized */
2195         filter_adapt_sptlrpc_conf(obd, 1);
2196
2197         q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2198         if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2199             queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2200                 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2201                               "for larger I/O requests: max_sectors = %u "
2202                               "could be up to max_hw_sectors=%u\n",
2203                               obd->obd_name, mnt->mnt_sb->s_id,
2204                               queue_max_sectors(q), queue_max_hw_sectors(q));
2205
2206         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2207         if (uuid_ptr != NULL) {
2208                 class_uuid_unparse(uuid_ptr, &uuid);
2209                 str = uuid.uuid;
2210         } else {
2211                 str = "no UUID";
2212         }
2213
2214         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2215         LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2216                       obd->obd_name, label ?: str, lmi ? "on " : "",
2217                       lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2218                       obd->obd_replayable ? "enabled" : "disabled");
2219
2220         RETURN(0);
2221
2222 err_post:
2223         filter_post(obd);
2224 err_ops:
2225         fsfilt_put_ops(obd->obd_fsops);
2226         filter_iobuf_pool_done(filter);
2227 err_mntput:
2228         server_put_mount(obd->obd_name, mnt);
2229         obd->u.obt.obt_sb = 0;
2230         return rc;
2231 }
2232
2233 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2234 {
2235         struct lprocfs_static_vars lvars;
2236         cfs_proc_dir_entry_t *entry;
2237         unsigned long addr;
2238         struct page *page;
2239         int rc;
2240         ENTRY;
2241
2242         CLASSERT(offsetof(struct obd_device, u.obt) ==
2243                  offsetof(struct obd_device, u.filter.fo_obt));
2244
2245         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2246                 RETURN(-EINVAL);
2247
2248         /* lprocfs must be setup before the filter so state can be safely added
2249          * to /proc incrementally as the filter is setup */
2250         lprocfs_filter_init_vars(&lvars);
2251         rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2252         if (rc) {
2253                 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2254                        obd->obd_name, rc);
2255                 RETURN(rc);
2256         }
2257
2258         rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2259         if (rc) {
2260                 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2261                        obd->obd_name, rc);
2262                 GOTO(obd_cleanup, rc);
2263         }
2264
2265         /* Init obdfilter private stats here */
2266         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2267                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2268         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2269                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2270         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2271                              LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2272                              "get_page", "usec");
2273         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2274                              LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2275         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2276                              LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2277         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2278                              LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2279         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2280                              LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2281
2282         rc = lproc_filter_attach_seqstat(obd);
2283         if (rc) {
2284                 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2285                 GOTO(free_obd_stats, rc);
2286         }
2287
2288         entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2289         if (IS_ERR(entry)) {
2290                 rc = PTR_ERR(entry);
2291                 CERROR("%s: error %d setting up lprocfs for %s\n",
2292                        obd->obd_name, rc, "exports");
2293                 GOTO(free_obd_stats, rc);
2294         }
2295         obd->obd_proc_exports_entry = entry;
2296
2297         entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2298                                    lprocfs_nid_stats_clear_read,
2299                                    lprocfs_nid_stats_clear_write, obd, NULL);
2300         if (IS_ERR(entry)) {
2301                 rc = PTR_ERR(entry);
2302                 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2303                        obd->obd_name, rc);
2304                 GOTO(free_obd_stats, rc);
2305         }
2306
2307         rc = lprocfs_job_stats_init(obd, LPROC_FILTER_STATS_LAST,
2308                                     filter_stats_counter_init);
2309         if (rc)
2310                 GOTO(remove_entry_clear, rc);
2311
2312         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2313         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2314         if (!page)
2315                 GOTO(job_stats_fini, rc = -ENOMEM);
2316         addr = (unsigned long)cfs_page_address(page);
2317         clear_page((void *)addr);
2318         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2319                LUSTRE_CFG_BUFLEN(lcfg, 4));
2320         rc = filter_common_setup(obd, lcfg, (void *)addr);
2321         OBD_PAGE_FREE(page);
2322         if (rc) {
2323                 CERROR("%s: filter_common_setup failed: %d.\n",
2324                        obd->obd_name, rc);
2325                 GOTO(job_stats_fini, rc);
2326         }
2327
2328         RETURN(0);
2329
2330 job_stats_fini:
2331         lprocfs_job_stats_fini(obd);
2332 remove_entry_clear:
2333         lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2334 free_obd_stats:
2335         lprocfs_free_obd_stats(obd);
2336 obd_cleanup:
2337         lprocfs_obd_cleanup(obd);
2338         return rc;
2339 }
2340
2341 static struct llog_operations filter_mds_ost_repl_logops;
2342
2343 static struct llog_operations filter_size_orig_logops = {};
2344
2345 static int filter_olg_fini(struct obd_llog_group *olg)
2346 {
2347         struct llog_ctxt *ctxt;
2348
2349         ENTRY;
2350
2351         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2352         if (ctxt)
2353                 llog_cleanup(NULL, ctxt);
2354
2355         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2356         if (ctxt)
2357                 llog_cleanup(NULL, ctxt);
2358
2359         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2360         if (ctxt)
2361                 llog_cleanup(NULL, ctxt);
2362
2363         RETURN(0);
2364 }
2365
2366 static int
2367 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2368                 struct obd_device *tgt)
2369 {
2370         struct llog_ctxt        *ctxt = NULL;
2371         int                      rc;
2372
2373         ENTRY;
2374
2375         filter_mds_ost_repl_logops = llog_client_ops;
2376         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2377         filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2378         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2379
2380         rc = llog_setup(NULL, obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt,
2381                         &filter_mds_ost_repl_logops);
2382         if (rc)
2383                 RETURN(rc);
2384
2385         rc = llog_setup(NULL, obd, olg, LLOG_SIZE_ORIG_CTXT, tgt,
2386                         &filter_size_orig_logops);
2387         if (rc)
2388                 GOTO(out, rc);
2389
2390         RETURN(0);
2391 out:
2392         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2393         llog_cleanup(NULL, ctxt);
2394         return rc;
2395 }
2396
2397 /**
2398  * Init the default olg, which is embeded in the obd_device, for filter.
2399  */
2400 static int
2401 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2402                         struct obd_device *tgt)
2403 {
2404         struct filter_obd *filter = &obd->u.filter;
2405         struct llog_ctxt *ctxt;
2406         int rc;
2407         ENTRY;
2408
2409         filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2410         if (!filter->fo_lcm)
2411                 RETURN(-ENOMEM);
2412
2413         rc = filter_olg_init(obd, olg, tgt);
2414         if (rc)
2415                 GOTO(cleanup_lcm, rc);
2416
2417         rc = llog_setup(NULL, obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt,
2418                         &llog_lvfs_ops);
2419         if (rc)
2420                 GOTO(cleanup_olg, rc);
2421
2422         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2423         if (!ctxt) {
2424                 CERROR("Can't get ctxt for %p:%x\n", olg,
2425                        LLOG_MDS_OST_REPL_CTXT);
2426                 GOTO(cleanup_olg, rc = -ENODEV);
2427         }
2428         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2429         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2430         llog_ctxt_put(ctxt);
2431
2432         RETURN(0);
2433 cleanup_olg:
2434         filter_olg_fini(olg);
2435 cleanup_lcm:
2436         llog_recov_thread_fini(filter->fo_lcm, 1);
2437         filter->fo_lcm = NULL;
2438         return rc;
2439 }
2440
2441 static int
2442 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2443                  struct obd_device *tgt, int *index)
2444 {
2445         struct filter_obd *filter = &obd->u.filter;
2446         struct llog_ctxt *ctxt;
2447         int rc;
2448         ENTRY;
2449
2450         LASSERT(olg != NULL);
2451         if (olg == &obd->obd_olg)
2452                 return filter_default_olg_init(obd, olg, tgt);
2453
2454         LASSERT(filter->fo_lcm != NULL);
2455         rc = filter_olg_init(obd, olg, tgt);
2456         if (rc)
2457                 RETURN(rc);
2458         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2459         if (!ctxt) {
2460                 CERROR("Can't get ctxt for %p:%x\n", olg,
2461                        LLOG_MDS_OST_REPL_CTXT);
2462                 filter_olg_fini(olg);
2463                 RETURN(-ENODEV);
2464         }
2465         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2466         ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2467         llog_ctxt_put(ctxt);
2468         RETURN(rc);
2469 }
2470
2471 static int filter_llog_finish(struct obd_device *obd, int count)
2472 {
2473         struct filter_obd *filter = &obd->u.filter;
2474         struct llog_ctxt *ctxt;
2475         ENTRY;
2476
2477         ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2478         if (ctxt) {
2479                 /*
2480                  * Make sure that no cached llcds left in recov_thread.
2481                  * We actually do sync in disconnect time, but disconnect
2482                  * may not come being marked rq_no_resend = 1.
2483                  */
2484                 llog_sync(ctxt, NULL, OBD_LLOG_FL_EXIT);
2485
2486                 /*
2487                  * Balance class_import_get() in llog_receptor_accept().
2488                  * This is safe to do, as llog is already synchronized
2489                  * and its import may go.
2490                  */
2491                 cfs_mutex_lock(&ctxt->loc_mutex);
2492                 if (ctxt->loc_imp) {
2493                         class_import_put(ctxt->loc_imp);
2494                         ctxt->loc_imp = NULL;
2495                 }
2496
2497                 if (filter->fo_lcm) {
2498                         llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2499                         filter->fo_lcm = NULL;
2500                 }
2501
2502                 cfs_mutex_unlock(&ctxt->loc_mutex);
2503                 llog_ctxt_put(ctxt);
2504         }
2505
2506         RETURN(filter_olg_fini(&obd->obd_olg));
2507 }
2508
2509 /**
2510  * Find the group llog according to group index in the llog group list.
2511  */
2512 static struct obd_llog_group *
2513 filter_find_olg_internal(struct filter_obd *filter, int group)
2514 {
2515         struct obd_llog_group *olg;
2516
2517         LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2518         cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2519                 if (olg->olg_seq == group)
2520                         RETURN(olg);
2521         }
2522         RETURN(NULL);
2523 }
2524
2525 /**
2526  * Find the group llog according to group index on the filter
2527  */
2528 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2529 {
2530         struct obd_llog_group *olg = NULL;
2531         struct filter_obd *filter;
2532
2533         filter = &obd->u.filter;
2534
2535         if (group == FID_SEQ_LLOG)
2536                 RETURN(&obd->obd_olg);
2537
2538         cfs_spin_lock(&filter->fo_llog_list_lock);
2539         olg = filter_find_olg_internal(filter, group);
2540         cfs_spin_unlock(&filter->fo_llog_list_lock);
2541
2542         RETURN(olg);
2543 }
2544 /**
2545  * Find the llog_group of the filter according to the group. If it can not
2546  * find, create the llog_group, which only happens when mds is being synced
2547  * with OST.
2548  */
2549 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2550 {
2551         struct obd_llog_group *olg = NULL, *olg_new = NULL;
2552         struct filter_obd *filter;
2553         int rc;
2554
2555         filter = &obd->u.filter;
2556
2557         if (group == FID_SEQ_LLOG)
2558                 RETURN(&obd->obd_olg);
2559
2560         OBD_ALLOC_PTR(olg_new);
2561         if (olg_new == NULL)
2562                RETURN(ERR_PTR(-ENOMEM));
2563
2564         cfs_spin_lock(&filter->fo_llog_list_lock);
2565         olg = filter_find_olg_internal(filter, group);
2566         if (olg) {
2567                 if (olg->olg_initializing) {
2568                         GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2569                 } else {
2570                         GOTO(out_unlock, olg);
2571                 }
2572         } else {
2573                 /* set as the newly allocated one */
2574                 olg = olg_new;
2575                 olg_new = NULL;
2576         }
2577
2578         llog_group_init(olg, group);
2579         cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2580         olg->olg_initializing = 1;
2581         cfs_spin_unlock(&filter->fo_llog_list_lock);
2582
2583         rc = obd_llog_init(obd, olg, obd, NULL);
2584         if (rc) {
2585                cfs_spin_lock(&filter->fo_llog_list_lock);
2586                cfs_list_del(&olg->olg_list);
2587                cfs_spin_unlock(&filter->fo_llog_list_lock);
2588                OBD_FREE_PTR(olg);
2589                GOTO(out, olg = ERR_PTR(-ENOMEM));
2590         }
2591         cfs_spin_lock(&filter->fo_llog_list_lock);
2592         olg->olg_initializing = 0;
2593         cfs_spin_unlock(&filter->fo_llog_list_lock);
2594         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2595               obd->obd_name, group, olg);
2596 out:
2597         RETURN(olg);
2598
2599 out_unlock:
2600         cfs_spin_unlock(&filter->fo_llog_list_lock);
2601         if (olg_new)
2602                OBD_FREE_PTR(olg_new);
2603         goto out;
2604 }
2605
2606 static int filter_llog_connect(struct obd_export *exp,
2607                                struct llogd_conn_body *body)
2608 {
2609         struct obd_device *obd = exp->exp_obd;
2610         struct llog_ctxt *ctxt;
2611         struct obd_llog_group *olg;
2612         int rc;
2613         ENTRY;
2614
2615         CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2616                obd->obd_name, body->lgdc_logid.lgl_oid,
2617                body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2618
2619         olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2620         if (!olg) {
2621                 CERROR(" %s: can not find olg of group %d\n",
2622                        obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2623                 RETURN(-ENOENT);
2624         }
2625         llog_group_set_export(olg, exp);
2626
2627         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2628         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2629                  body->lgdc_ctxt_idx);
2630
2631         CDEBUG(D_HA, "%s: Recovery from log "LPX64"/"LPX64":%x\n",
2632                obd->obd_name, body->lgdc_logid.lgl_oid,
2633                body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2634
2635         cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2636         obd->u.filter.fo_mds_ost_sync = 1;
2637         cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2638         rc = llog_connect(ctxt, &body->lgdc_logid,
2639                           &body->lgdc_gen, NULL);
2640         llog_ctxt_put(ctxt);
2641         if (rc != 0)
2642                 CERROR("failed to connect rc %d idx %d\n", rc,
2643                                 body->lgdc_ctxt_idx);
2644
2645         RETURN(rc);
2646 }
2647
2648 static int filter_llog_preclean(struct obd_device *obd)
2649 {
2650         struct obd_llog_group *olg, *tmp;
2651         struct filter_obd *filter;
2652         cfs_list_t  remove_list;
2653         int rc = 0;
2654         ENTRY;
2655
2656         rc = obd_llog_finish(obd, 0);
2657         if (rc)
2658                 CERROR("failed to cleanup llogging subsystem\n");
2659
2660         filter = &obd->u.filter;
2661         CFS_INIT_LIST_HEAD(&remove_list);
2662
2663         cfs_spin_lock(&filter->fo_llog_list_lock);
2664         while (!cfs_list_empty(&filter->fo_llog_list)) {
2665                 olg = cfs_list_entry(filter->fo_llog_list.next,
2666                                      struct obd_llog_group, olg_list);
2667                 cfs_list_del(&olg->olg_list);
2668                 cfs_list_add(&olg->olg_list, &remove_list);
2669         }
2670         cfs_spin_unlock(&filter->fo_llog_list_lock);
2671
2672         cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2673                 cfs_list_del_init(&olg->olg_list);
2674                 rc = filter_olg_fini(olg);
2675                 if (rc)
2676                         CERROR("failed to cleanup llogging subsystem for %u\n",
2677                                olg->olg_seq);
2678                 OBD_FREE_PTR(olg);
2679         }
2680
2681         RETURN(rc);
2682 }
2683
2684 static int filter_precleanup(struct obd_device *obd,
2685                              enum obd_cleanup_stage stage)
2686 {
2687         int rc = 0;
2688         ENTRY;
2689
2690         switch(stage) {
2691         case OBD_CLEANUP_EARLY:
2692                 break;
2693         case OBD_CLEANUP_EXPORTS:
2694                 /* Stop recovery before namespace cleanup. */
2695                 target_recovery_fini(obd);
2696
2697                 obd_exports_barrier(obd);
2698                 obd_zombie_barrier();
2699
2700                 rc = filter_llog_preclean(obd);
2701                 lprocfs_job_stats_fini(obd);
2702                 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2703                 lprocfs_free_per_client_stats(obd);
2704                 lprocfs_obd_cleanup(obd);
2705                 lprocfs_free_obd_stats(obd);
2706                 break;
2707         }
2708         RETURN(rc);
2709 }
2710
2711 static int filter_cleanup(struct obd_device *obd)
2712 {
2713         struct filter_obd *filter = &obd->u.filter;
2714         ENTRY;
2715
2716         if (obd->obd_fail)
2717                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2718                               "will be preserved.\n", obd->obd_name);
2719
2720         ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2721         obd->obd_namespace = NULL;
2722
2723         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2724
2725         if (obd->u.obt.obt_sb == NULL)
2726                 RETURN(0);
2727
2728         filter_post(obd);
2729
2730         ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2731         shrink_dcache_sb(obd->u.obt.obt_sb);
2732
2733         server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2734         obd->u.obt.obt_sb = NULL;
2735
2736         fsfilt_put_ops(obd->obd_fsops);
2737
2738         filter_iobuf_pool_done(filter);
2739
2740         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2741
2742         RETURN(0);
2743 }
2744
2745 static int filter_connect_internal(struct obd_export *exp,
2746                                    struct obd_connect_data *data,
2747                                    int reconnect)
2748 {
2749         struct filter_export_data *fed = &exp->exp_filter_data;
2750
2751         if (!data)
2752                 RETURN(0);
2753
2754         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2755                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2756                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2757                data->ocd_connect_flags, data->ocd_version,
2758                data->ocd_grant, data->ocd_index);
2759
2760         if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2761                 CWARN("!!! This export (nid %s) used object group %d "
2762                        "earlier; now it's trying to use group %d!  This could "
2763                        "be a bug in the MDS. Please report to "
2764                        "http://bugs.whamcloud.com/\n",
2765                        obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2766                 RETURN(-EPROTO);
2767         }
2768         fed->fed_group = data->ocd_group;
2769
2770         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2771         exp->exp_connect_flags = data->ocd_connect_flags;
2772         data->ocd_version = LUSTRE_VERSION_CODE;
2773
2774         /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2775         if (data->ocd_connect_flags & OBD_CONNECT_MDS)
2776                 CDEBUG(D_HA, "%s: Received MDS connection for group %u\n",
2777                        exp->exp_obd->obd_name, data->ocd_group);
2778         else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
2779                 RETURN(-EPROTO);
2780
2781         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2782                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2783                 obd_size left, want;
2784
2785                 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2786                 left = filter_grant_space_left(exp);
2787                 want = data->ocd_grant;
2788                 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2789                 data->ocd_grant = fed->fed_grant;
2790                 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2791
2792                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2793                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2794                        exp->exp_client_uuid.uuid, exp,
2795                        data->ocd_grant, want, left);
2796
2797                 filter->fo_tot_granted_clients ++;
2798         }
2799
2800         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2801                 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2802                 int index = le32_to_cpu(lsd->lsd_ost_index);
2803
2804                 if (!(lsd->lsd_feature_compat &
2805                       cpu_to_le32(OBD_COMPAT_OST))) {
2806                         /* this will only happen on the first connect */
2807                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2808                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2809                         /* sync is not needed here as filter_client_add will
2810                          * set exp_need_sync flag */
2811                         filter_update_server_data(exp->exp_obd);
2812                 } else if (index != data->ocd_index) {
2813                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2814                                            " %u doesn't match actual OST index"
2815                                            " %u in last_rcvd file, bad "
2816                                            "configuration?\n",
2817                                            obd_export_nid2str(exp), index,
2818                                            data->ocd_index);
2819                         RETURN(-EBADF);
2820                 }
2821                 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2822                  * FIXME: We don't strictly need the COMPAT flag for that,
2823                  * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2824                  * FIXME: We needed it for the index, as index 0 is valid. */
2825         }
2826
2827         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2828                 data->ocd_brw_size = 65536;
2829         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2830                 data->ocd_brw_size = min(data->ocd_brw_size,
2831                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2832                 if (data->ocd_brw_size == 0) {
2833                         CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2834                                " ocd_version: %x ocd_grant: %d ocd_index: %u "
2835                                "ocd_brw_size is unexpectedly zero, "
2836                                "network data corruption?"
2837                                "Refusing connection of this client\n",
2838                                 exp->exp_obd->obd_name,
2839                                 exp->exp_client_uuid.uuid,
2840                                 exp, data->ocd_connect_flags, data->ocd_version,
2841                                 data->ocd_grant, data->ocd_index);
2842                         RETURN(-EPROTO);
2843                 }
2844         }
2845
2846         if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2847                 __u32 cksum_types = data->ocd_cksum_types;
2848
2849                 /* The client set in ocd_cksum_types the checksum types it
2850                  * supports. We have to mask off the algorithms that we don't
2851                  * support */
2852                 data->ocd_cksum_types &= cksum_types_supported_server();
2853
2854                 /* 1.6.4 clients are not supported any more */
2855
2856                 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2857                                    "%x\n", exp->exp_obd->obd_name,
2858                                    obd_export_nid2str(exp), cksum_types,
2859                                    data->ocd_cksum_types);
2860         } else {
2861                 /* This client does not support OBD_CONNECT_CKSUM
2862                  * fall back to CRC32 */
2863                 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2864                                    "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2865                                    exp->exp_obd->obd_name,
2866                                    obd_export_nid2str(exp));
2867         }
2868
2869         if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2870                 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2871
2872         RETURN(0);
2873 }
2874
2875 static int filter_reconnect(const struct lu_env *env,
2876                             struct obd_export *exp, struct obd_device *obd,
2877                             struct obd_uuid *cluuid,
2878                             struct obd_connect_data *data,
2879                             void *localdata)
2880 {
2881         int rc;
2882         ENTRY;
2883
2884         if (exp == NULL || obd == NULL || cluuid == NULL)
2885                 RETURN(-EINVAL);
2886
2887         rc = filter_connect_internal(exp, data, 1);
2888         if (rc == 0)
2889                 filter_export_stats_init(obd, exp, localdata);
2890
2891         RETURN(rc);
2892 }
2893
2894 static int filter_connect(const struct lu_env *env,
2895                           struct obd_export **exp, struct obd_device *obd,
2896                           struct obd_uuid *cluuid,
2897                           struct obd_connect_data *data, void *localdata)
2898 {
2899         struct lvfs_run_ctxt saved;
2900         struct lustre_handle conn = { 0 };
2901         struct obd_export *lexp;
2902         int rc;
2903         ENTRY;
2904
2905         if (exp == NULL || obd == NULL || cluuid == NULL)
2906                 RETURN(-EINVAL);
2907
2908         rc = class_connect(&conn, obd, cluuid);
2909         if (rc)
2910                 RETURN(rc);
2911         lexp = class_conn2export(&conn);
2912         LASSERT(lexp != NULL);
2913
2914         rc = filter_connect_internal(lexp, data, 0);
2915         if (rc)
2916                 GOTO(cleanup, rc);
2917
2918         filter_export_stats_init(obd, lexp, localdata);
2919         if (obd->obd_replayable) {
2920                 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2921                 LASSERT(lcd);
2922                 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2923                 rc = filter_client_add(obd, lexp, -1);
2924                 if (rc)
2925                         GOTO(cleanup, rc);
2926         }
2927
2928         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2929         rc = filter_read_groups(obd, data->ocd_group, 1);
2930         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2931         if (rc != 0) {
2932                 CERROR("can't read group %u\n", data->ocd_group);
2933                 GOTO(cleanup, rc);
2934         }
2935
2936         GOTO(cleanup, rc);
2937
2938 cleanup:
2939         if (rc) {
2940                 class_disconnect(lexp);
2941                 *exp = NULL;
2942         } else {
2943                 *exp = lexp;
2944         }
2945
2946         RETURN(rc);
2947 }
2948
2949 /* Do extra sanity checks for grant accounting.  We do this at connect,
2950  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2951  * always get rid of it or turn it off when we know accounting is good. */
2952 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2953 {
2954         struct filter_export_data *fed;
2955         struct obd_export *exp;
2956         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2957         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2958         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2959
2960         if (cfs_list_empty(&obd->obd_exports))
2961                 return;
2962
2963         /* We don't want to do this for large machines that do lots of
2964            mounts or unmounts.  It burns... */
2965         if (obd->obd_num_exports > 100)
2966                 return;
2967
2968         cfs_spin_lock(&obd->obd_osfs_lock);
2969         cfs_spin_lock(&obd->obd_dev_lock);
2970         cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2971                 int error = 0;
2972                 fed = &exp->exp_filter_data;
2973                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2974                     fed->fed_dirty < 0)
2975                         error = 1;
2976                 if (maxsize > 0) { /* we may not have done a statfs yet */
2977                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2978                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2979                                  exp->exp_client_uuid.uuid, exp,
2980                                  fed->fed_grant, fed->fed_pending, maxsize);
2981                         LASSERTF(fed->fed_dirty <= maxsize,
2982                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2983                                  exp->exp_client_uuid.uuid, exp,
2984                                  fed->fed_dirty, maxsize);
2985                 }
2986                 if (error)
2987                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2988                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2989                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2990                 else
2991                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2992                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2993                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2994                 tot_granted += fed->fed_grant + fed->fed_pending;
2995                 tot_pending += fed->fed_pending;
2996                 tot_dirty += fed->fed_dirty;
2997         }
2998         fo_tot_granted = obd->u.filter.fo_tot_granted;
2999         fo_tot_pending = obd->u.filter.fo_tot_pending;
3000         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
3001         cfs_spin_unlock(&obd->obd_dev_lock);
3002         cfs_spin_unlock(&obd->obd_osfs_lock);
3003
3004         /* Do these assertions outside the spinlocks so we don't kill system */
3005         if (tot_granted != fo_tot_granted)
3006                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
3007                        func, tot_granted, fo_tot_granted);
3008         if (tot_pending != fo_tot_pending)
3009                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
3010                        func, tot_pending, fo_tot_pending);
3011         if (tot_dirty != fo_tot_dirty)
3012                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
3013                        func, tot_dirty, fo_tot_dirty);
3014         if (tot_pending > tot_granted)
3015                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
3016                        func, tot_pending, tot_granted);
3017         if (tot_granted > maxsize)
3018                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
3019                        func, tot_granted, maxsize);
3020         if (tot_dirty > maxsize)
3021                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
3022                        func, tot_dirty, maxsize);
3023 }
3024
3025 /* Remove this client from the grant accounting totals.  We also remove
3026  * the export from the obd device under the osfs and dev locks to ensure
3027  * that the filter_grant_sanity_check() calculations are always valid.
3028  * The client should do something similar when it invalidates its import. */
3029 static void filter_grant_discard(struct obd_export *exp)
3030 {
3031         struct obd_device *obd = exp->exp_obd;
3032         struct filter_obd *filter = &obd->u.filter;
3033         struct filter_export_data *fed = &exp->exp_filter_data;
3034
3035         cfs_spin_lock(&obd->obd_osfs_lock);
3036         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
3037                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
3038                  obd->obd_name, filter->fo_tot_granted,
3039                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
3040         filter->fo_tot_granted -= fed->fed_grant;
3041         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
3042                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
3043                  obd->obd_name, filter->fo_tot_pending,
3044                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
3045         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
3046         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
3047                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
3048                  obd->obd_name, filter->fo_tot_dirty,
3049                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
3050         filter->fo_tot_dirty -= fed->fed_dirty;
3051         fed->fed_dirty = 0;
3052         fed->fed_grant = 0;
3053
3054         cfs_spin_unlock(&obd->obd_osfs_lock);
3055 }
3056
3057 static int filter_destroy_export(struct obd_export *exp)
3058 {
3059         struct filter_export_data *fed = &exp->exp_filter_data;
3060         ENTRY;
3061
3062         if (fed->fed_pending)
3063                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
3064                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
3065                        exp, fed->fed_pending);
3066
3067         target_destroy_export(exp);
3068
3069         if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
3070                                      &exp->exp_client_uuid)))
3071                RETURN(0);
3072
3073         ldlm_destroy_export(exp);
3074         lut_client_free(exp);
3075
3076         if (!exp->exp_obd->obd_replayable)
3077                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3078
3079         filter_grant_discard(exp);
3080         filter_fmd_cleanup(exp);
3081
3082         if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3083                 struct filter_obd *filter = &exp->exp_obd->u.filter;
3084                 if (filter->fo_tot_granted_clients > 0)
3085                         filter->fo_tot_granted_clients --;
3086         }
3087
3088         if (!(exp->exp_flags & OBD_OPT_FORCE))
3089                 filter_grant_sanity_check(exp->exp_obd, __func__);
3090
3091         RETURN(0);
3092 }
3093
3094 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3095 {
3096         struct obd_llog_group *olg_min, *olg;
3097         struct filter_obd *filter;
3098         int worked = -1, group;
3099         struct llog_ctxt *ctxt;
3100         ENTRY;
3101
3102         filter = &obd->u.filter;
3103
3104         /* we can't sync log holding spinlock. also, we do not want to get
3105          * into livelock. so we do following: loop over MDS's exports in
3106          * group order and skip already synced llogs -bzzz */
3107         do {
3108                 /* look for group with min. number, but > worked */
3109                 olg_min = NULL;
3110                 group = 1 << 30;
3111                 cfs_spin_lock(&filter->fo_llog_list_lock);
3112                 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3113                         if (olg->olg_seq <= worked) {
3114                                 /* this group is already synced */
3115                                 continue;
3116                         }
3117                         if (group < olg->olg_seq) {
3118                                 /* we have group with smaller number to sync */
3119                                 continue;
3120                         }
3121                         /* store current minimal group */
3122                         olg_min = olg;
3123                         group = olg->olg_seq;
3124                 }
3125                 cfs_spin_unlock(&filter->fo_llog_list_lock);
3126
3127                 if (olg_min == NULL)
3128                         break;
3129
3130                 worked = olg_min->olg_seq;
3131                 if (olg_min->olg_exp &&
3132                     (dexp == olg_min->olg_exp || dexp == NULL)) {
3133                         int err;
3134                         ctxt = llog_group_get_ctxt(olg_min,
3135                                                    LLOG_MDS_OST_REPL_CTXT);
3136                         if (ctxt) {
3137                                 err = llog_sync(ctxt, olg_min->olg_exp, 0);
3138                                 llog_ctxt_put(ctxt);
3139                                 if (err) {
3140                                         CERROR("error flushing logs to MDS: "
3141                                                "rc %d\n", err);
3142                                 }
3143                         }
3144                 }
3145         } while (olg_min != NULL);
3146 }
3147
3148 /* Also incredibly similar to mds_disconnect */
3149 static int filter_disconnect(struct obd_export *exp)
3150 {
3151         struct obd_device *obd = exp->exp_obd;
3152         int rc;
3153         ENTRY;
3154
3155         LASSERT(exp);
3156         class_export_get(exp);
3157
3158         if (!(exp->exp_flags & OBD_OPT_FORCE))
3159                 filter_grant_sanity_check(obd, __func__);
3160         filter_grant_discard(exp);
3161
3162         /* Flush any remaining cancel messages out to the target */
3163         filter_sync_llogs(obd, exp);
3164
3165         rc = server_disconnect_export(exp);
3166
3167         /* Do not erase record for recoverable client. */
3168         if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3169                 filter_client_del(exp);
3170         else
3171                 fsfilt_sync(obd, obd->u.obt.obt_sb);
3172
3173         class_export_put(exp);
3174         RETURN(rc);
3175 }
3176
3177 /* reverse import is changed, sync all cancels */
3178 static void filter_revimp_update(struct obd_export *exp)
3179 {
3180         ENTRY;
3181
3182         LASSERT(exp);
3183         class_export_get(exp);
3184
3185         /* flush any remaining cancel messages out to the target */
3186         filter_sync_llogs(exp->exp_obd, exp);
3187         class_export_put(exp);
3188         EXIT;
3189 }
3190
3191 static int filter_ping(const struct lu_env *env, struct obd_export *exp)
3192 {
3193         filter_fmd_expire(exp);
3194         return 0;
3195 }
3196
3197 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3198                                   const char *what, int quiet)
3199 {
3200         struct dentry *dchild = NULL;
3201
3202         dchild = filter_fid2dentry(obd, NULL,  ostid->oi_seq, ostid->oi_id);
3203
3204         if (IS_ERR(dchild)) {
3205                 CERROR("%s error looking up object: "POSTID"\n",
3206                        what, ostid->oi_id, ostid->oi_seq);
3207                 RETURN(dchild);
3208         }
3209
3210         if (dchild->d_inode == NULL) {
3211                 if (!quiet)
3212                         CERROR("%s: %s on non-existent object: "POSTID" \n",
3213                                obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3214                 f_dput(dchild);
3215                 RETURN(ERR_PTR(-ENOENT));
3216         }
3217
3218 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
3219         /* Try to correct for a bug in 2.1.0 (LU-221) that caused negative
3220          * timestamps to appear to be in the far future, due old timestamp
3221          * being stored on disk as an unsigned value.  This fixes up any
3222          * bad values stored on disk before returning them to the client,
3223          * and ensures any timestamp updates are correct.  LU-1042 */
3224         if (unlikely(LTIME_S(dchild->d_inode->i_atime) == LU221_BAD_TIME))
3225                 LTIME_S(dchild->d_inode->i_atime) = 0;
3226         if (unlikely(LTIME_S(dchild->d_inode->i_mtime) == LU221_BAD_TIME))
3227                 LTIME_S(dchild->d_inode->i_mtime) = 0;
3228         if (unlikely(LTIME_S(dchild->d_inode->i_ctime) == LU221_BAD_TIME))
3229                 LTIME_S(dchild->d_inode->i_ctime) = 0;
3230 #else
3231 #warning "remove old LU-221/LU-1042 workaround code"
3232 #endif
3233
3234         return dchild;
3235 }
3236
3237 static int filter_getattr(const struct lu_env *env, struct obd_export *exp,
3238                           struct obd_info *oinfo)
3239 {
3240         struct dentry *dentry = NULL;
3241         struct obd_device *obd;
3242         __u64 curr_version;
3243         int rc = 0;
3244         ENTRY;
3245
3246         rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3247                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
3248         if (rc)
3249                 RETURN(rc);
3250
3251         obd = class_exp2obd(exp);
3252         if (obd == NULL) {
3253                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3254                 RETURN(-EINVAL);
3255         }
3256
3257         dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3258         if (IS_ERR(dentry))
3259                 RETURN(PTR_ERR(dentry));
3260
3261         /* Limit the valid bits in the return data to what we actually use */
3262         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3263         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3264
3265         /* Store inode version in reply */
3266         curr_version = fsfilt_get_version(exp->exp_obd, dentry->d_inode);
3267         if ((__s64)curr_version != -EOPNOTSUPP) {
3268                 oinfo->oi_oa->o_valid |= OBD_MD_FLDATAVERSION;
3269                 oinfo->oi_oa->o_data_version = curr_version;
3270         }
3271
3272         f_dput(dentry);
3273         RETURN(rc);
3274 }
3275
3276 /* this should be enabled/disabled in condition to enabled/disabled large
3277  * inodes (fast EAs) in backing store FS. */
3278 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3279                         void *handle, struct obdo *oa)
3280 {
3281         struct obd_device *obd = exp->exp_obd;
3282         int rc = 0;
3283         ENTRY;
3284
3285         if (oa->o_valid & OBD_MD_FLFID) {
3286                 struct filter_fid ff;
3287
3288                 if (!(oa->o_valid & OBD_MD_FLGROUP))
3289                         oa->o_seq = 0;
3290                 /* packing fid and converting it to LE for storing into EA.
3291                  * Here ->o_stripe_idx should be filled by LOV and rest of
3292                  * fields - by client. */
3293                 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3294                 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3295                 /* XXX: we are ignoring o_parent_ver here, since this should
3296                  *      be the same for all objects in this fileset. */
3297                 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3298                 ff.ff_objid = cpu_to_le64(oa->o_id);
3299                 ff.ff_seq = cpu_to_le64(oa->o_seq);
3300
3301                 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3302                        LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3303                        oa->o_seq);
3304
3305                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3306                 if (rc)
3307                         CERROR("store fid in object failed! rc: %d\n", rc);
3308         } else {
3309                 CDEBUG(D_HA, "OSS object without fid info!\n");
3310         }
3311
3312         RETURN(rc);
3313 }
3314
3315 /* this is called from filter_truncate() until we have filter_punch() */
3316 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3317                             struct obdo *oa, struct obd_trans_info *oti)
3318 {
3319         struct llog_cookie *fcc = NULL;
3320         struct filter_obd *filter;
3321         int rc, err, sync = 0;
3322         loff_t old_size = 0;
3323         unsigned int ia_valid;
3324         struct inode *inode;
3325         struct page *page = NULL;
3326         struct iattr iattr;
3327         void *handle;
3328         ENTRY;
3329
3330         LASSERT(dentry != NULL);
3331         LASSERT(!IS_ERR(dentry));
3332
3333         inode = dentry->d_inode;
3334         LASSERT(inode != NULL);
3335
3336         filter = &exp->exp_obd->u.filter;
3337         iattr_from_obdo(&iattr, oa, oa->o_valid);
3338         ia_valid = iattr.ia_valid;
3339
3340         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3341                 OBD_ALLOC(fcc, sizeof(*fcc));
3342                 if (fcc != NULL)
3343                         *fcc = oa->o_lcookie;
3344         }
3345         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3346                 unsigned long now = jiffies;
3347                 /* Filter truncates and writes are serialized by
3348                  * i_alloc_sem, see the comment in
3349                  * filter_preprw_write.*/
3350                 if (ia_valid & ATTR_SIZE)
3351                         down_write(&inode->i_alloc_sem);
3352                 mutex_lock(&inode->i_mutex);
3353                 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
3354                 old_size = i_size_read(inode);
3355         }
3356
3357         /* VBR: version recovery check */
3358         rc = filter_version_get_check(exp, oti, inode);
3359         if (rc)
3360                 GOTO(out_unlock, rc);
3361
3362         /* Let's pin the last page so that ldiskfs_truncate
3363          * should not start GFP_FS allocation. */
3364         if (ia_valid & ATTR_SIZE) {
3365                 page = grab_cache_page(inode->i_mapping,
3366                                        iattr.ia_size >> PAGE_CACHE_SHIFT);
3367                 if (page == NULL)
3368                         GOTO(out_unlock, rc = -ENOMEM);
3369
3370                 unlock_page(page);
3371         }
3372
3373         /* If the inode still has SUID+SGID bits set (see filter_precreate())
3374          * then we will accept the UID+GID sent by the client during write for
3375          * initializing the ownership of this inode.  We only allow this to
3376          * happen once so clear these bits in setattr. In 2.6 kernels it is
3377          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3378          * the flags that are actually being set. */
3379         if (ia_valid & (ATTR_UID | ATTR_GID)) {
3380                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3381                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3382
3383                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3384                         if (!(ia_valid & ATTR_MODE)) {
3385                                 iattr.ia_mode = inode->i_mode;
3386                                 iattr.ia_valid |= ATTR_MODE;
3387                         }
3388                         iattr.ia_mode &= ~S_ISUID;
3389                 }
3390                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3391                         if (!(iattr.ia_valid & ATTR_MODE)) {
3392                                 iattr.ia_mode = inode->i_mode;
3393                                 iattr.ia_valid |= ATTR_MODE;
3394                         }
3395                         iattr.ia_mode &= ~S_ISGID;
3396                 }
3397
3398                 handle = fsfilt_start_log(exp->exp_obd, inode,
3399                                           FSFILT_OP_SETATTR, oti, 1);
3400                 if (IS_ERR(handle))
3401                         GOTO(out_unlock, rc = PTR_ERR(handle));
3402
3403                 /* update inode EA only once when inode is suid bit marked. As
3404                  * on 2.6.x UID and GID may be set separately, we check here
3405                  * only one of them to avoid double setting. */
3406                 if (inode->i_mode & S_ISUID)
3407                         filter_update_fidea(exp, inode, handle, oa);
3408         } else {
3409                 handle = fsfilt_start(exp->exp_obd, inode,
3410                                       FSFILT_OP_SETATTR, oti);
3411                 if (IS_ERR(handle))
3412                         GOTO(out_unlock, rc = PTR_ERR(handle));
3413         }
3414
3415         /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
3416         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3417                 ll_vfs_dq_init(inode);
3418
3419         if (oa->o_valid & OBD_MD_FLFLAGS) {
3420                 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3421                                       FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3422         } else {
3423                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3424                 if (fcc != NULL)
3425                         /* set cancel cookie callback function */
3426                         sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3427                                                      filter_cancel_cookies_cb,
3428                                                      fcc);
3429         }
3430
3431         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3432                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3433
3434        /* The truncate might have used up our transaction credits.  Make sure
3435         * we have two left for the last_rcvd and VBR inode version updates. */
3436         err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3437
3438         /* Update inode version only if data has changed => size has changed */
3439         rc = filter_finish_transno(exp, ia_valid & ATTR_SIZE ? inode : NULL,
3440                                    oti, rc, sync);
3441
3442         if (sync) {
3443                 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3444                 fcc = NULL;
3445         }
3446
3447         err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3448         if (err) {
3449                 CERROR("error on commit, err = %d\n", err);
3450                 if (!rc)
3451                         rc = err;
3452         } else {
3453                 fcc = NULL;
3454         }
3455
3456         /* For a partial-page truncate flush the page to disk immediately
3457          * to avoid data corruption during direct disk write. b=17397 */
3458         if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3459             old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3460                 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3461                                                iattr.ia_size + 1);
3462                 if (!rc)
3463                         rc = err;
3464         }
3465
3466         EXIT;
3467
3468 out_unlock:
3469         if (page)
3470                 page_cache_release(page);
3471
3472         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3473                 mutex_unlock(&inode->i_mutex);
3474         if (ia_valid & ATTR_SIZE)
3475                 up_write(&inode->i_alloc_sem);
3476         if (fcc)
3477                 OBD_FREE(fcc, sizeof(*fcc));
3478         return rc;
3479 }
3480
3481 /* this is called from filter_truncate() until we have filter_punch() */
3482 int filter_setattr(const struct lu_env *env, struct obd_export *exp,
3483                    struct obd_info *oinfo, struct obd_trans_info *oti)
3484 {
3485         struct obdo *oa = oinfo->oi_oa;
3486         struct lustre_capa *capa = oinfo_capa(oinfo);
3487         struct ldlm_res_id res_id;
3488         struct filter_mod_data *fmd;
3489         struct lvfs_run_ctxt saved;
3490         struct filter_obd *filter;
3491         struct ldlm_resource *res;
3492         struct dentry *dentry;
3493         __u64 opc = CAPA_OPC_META_WRITE;
3494         int rc;
3495         ENTRY;
3496
3497         if (oinfo->oi_flags & OBD_FL_PUNCH)
3498                 opc |= CAPA_OPC_OSS_TRUNC;
3499
3500         rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3501         if (rc)
3502                 RETURN(rc);
3503
3504         if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3505                 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3506                 if (rc)
3507                         RETURN(rc);
3508         }
3509
3510         osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3511         /* This would be very bad - accidentally truncating a file when
3512          * changing the time or similar - bug 12203. */
3513         if (oa->o_valid & OBD_MD_FLSIZE &&
3514             oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3515                 static char mdsinum[48];
3516
3517                 if (oa->o_valid & OBD_MD_FLFID)
3518                         snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3519                                  oa->o_parent_seq, oa->o_parent_oid,
3520                                  oa->o_parent_ver);
3521                 else
3522                         mdsinum[0] = '\0';
3523
3524                 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3525                        "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3526                        oa->o_id, oa->o_seq, mdsinum);
3527                 RETURN(-EPERM);
3528         }
3529
3530         dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3531         if (IS_ERR(dentry))
3532                 RETURN(PTR_ERR(dentry));
3533
3534         filter = &exp->exp_obd->u.filter;
3535         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3536
3537         /*
3538          * We need to be atomic against a concurrent write
3539          * (which takes the semaphore for reading). fmd_mactime_xid
3540          * checks will have no effect if a write request with lower
3541          * xid starts just before a setattr and finishes later than
3542          * the setattr (see bug 21489, comment 27).
3543          */
3544         if (oa->o_valid &
3545             (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3546                 unsigned long now = jiffies;
3547                 down_write(&dentry->d_inode->i_alloc_sem);
3548                 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem");
3549                 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3550                 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3551                         fmd->fmd_mactime_xid = oti->oti_xid;
3552                 filter_fmd_put(exp, fmd);
3553                 up_write(&dentry->d_inode->i_alloc_sem);
3554         }
3555
3556         /* setting objects attributes (including owner/group) */
3557         rc = filter_setattr_internal(exp, dentry, oa, oti);
3558         if (rc)
3559                 GOTO(out_unlock, rc);
3560
3561         res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3562                                 &res_id, LDLM_EXTENT, 0);
3563
3564         if (res != NULL) {
3565                 LDLM_RESOURCE_ADDREF(res);
3566                 rc = ldlm_res_lvbo_update(res, NULL, 0);
3567                 LDLM_RESOURCE_DELREF(res);
3568                 ldlm_resource_putref(res);
3569         }
3570
3571         oa->o_valid = OBD_MD_FLID;
3572
3573         /* Quota release need uid/gid info */
3574         obdo_from_inode(oa, dentry->d_inode,
3575                         FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3576
3577         filter_counter_incr(exp, LPROC_FILTER_STATS_SETATTR,
3578                             oti ? oti->oti_jobid : NULL, 1);
3579         EXIT;
3580 out_unlock:
3581         f_dput(dentry);
3582         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3583         return rc;
3584 }
3585
3586 /* XXX identical to osc_unpackmd */
3587 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3588                            struct lov_mds_md *lmm, int lmm_bytes)
3589 {
3590         int lsm_size;
3591         ENTRY;
3592
3593         if (lmm != NULL) {
3594                 if (lmm_bytes < sizeof (*lmm)) {
3595                         CERROR("lov_mds_md too small: %d, need %d\n",
3596                                lmm_bytes, (int)sizeof(*lmm));
3597                         RETURN(-EINVAL);
3598                 }
3599                 /* XXX LOV_MAGIC etc check? */
3600
3601                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3602                         CERROR("lov_mds_md: zero lmm_object_id\n");
3603                         RETURN(-EINVAL);
3604                 }
3605         }
3606
3607         lsm_size = lov_stripe_md_size(1);
3608         if (lsmp == NULL)
3609                 RETURN(lsm_size);
3610
3611         if (*lsmp != NULL && lmm == NULL) {
3612                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3613                 OBD_FREE(*lsmp, lsm_size);
3614                 *lsmp = NULL;
3615                 RETURN(0);
3616         }
3617
3618         if (*lsmp == NULL) {
3619                 OBD_ALLOC(*lsmp, lsm_size);
3620                 if (*lsmp == NULL)
3621                         RETURN(-ENOMEM);
3622
3623                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3624                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3625                         OBD_FREE(*lsmp, lsm_size);
3626                         RETURN(-ENOMEM);
3627                 }
3628                 loi_init((*lsmp)->lsm_oinfo[0]);
3629         }
3630
3631         if (lmm != NULL) {
3632                 /* XXX zero *lsmp? */
3633                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3634                 LASSERT((*lsmp)->lsm_object_id);
3635         }
3636
3637         (*lsmp)->lsm_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
3638
3639         RETURN(lsm_size);
3640 }
3641
3642 /* caller must hold fo_create_locks[oa->o_seq] */
3643 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3644                                      struct filter_obd *filter)
3645 {
3646         struct obdo doa = { 0 }; /* XXX obdo on stack */
3647         obd_id last, id;
3648         int rc = 0;
3649         int skip_orphan;
3650         ENTRY;
3651
3652         LASSERT_MUTEX_LOCKED(&filter->fo_create_locks[oa->o_seq]);
3653
3654         memset(&doa, 0, sizeof(doa));
3655
3656         doa.o_valid |= OBD_MD_FLGROUP;
3657         doa.o_seq = oa->o_seq;
3658         doa.o_mode = S_IFREG;
3659
3660         if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3661                 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3662                        exp->exp_obd->obd_name, doa.o_seq);
3663                 RETURN(0);
3664         }
3665
3666         last = filter_last_id(filter, doa.o_seq);
3667
3668         skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3669
3670         CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3671                exp->exp_obd->obd_name, oa->o_id + 1, last,
3672                skip_orphan ? ", orphan objids won't be reused any more." : ".");
3673
3674         for (id = last; id > oa->o_id; id--) {
3675                 doa.o_id = id;
3676                 rc = filter_destroy(NULL, exp, &doa, NULL, NULL, NULL, NULL);
3677                 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3678                         CEMERG("error destroying precreate objid "LPU64": %d\n",
3679                                id, rc);
3680
3681                 /* update last_id on disk periodically so that if we restart
3682                  * we don't need to re-scan all of the just-deleted objects. */
3683                 if ((id & 511) == 0 && !skip_orphan) {
3684                         filter_set_last_id(filter, id - 1, doa.o_seq);
3685                         filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3686                 }
3687         }
3688
3689         CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3690                exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3691
3692         if (!skip_orphan) {
3693                 filter_set_last_id(filter, id, doa.o_seq);
3694                 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3695         } else {
3696                 /*
3697                  * We have destroyed orphan objects, but don't want to reuse
3698                  * them. Therefore we don't reset last_id to the last created
3699                  * objects. Instead, we report back to the MDS the object id
3700                  * of the last orphan, so that the MDS can restart allocating
3701                  * objects from this id + 1 and thus skip the whole orphan