4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
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
27 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdfilter/filter.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Andreas Dilger <adilger@clusterfs.com>
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).
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.
52 #define DEBUG_SUBSYSTEM S_FILTER
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
57 #include <linux/module.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
66 #include <obd_cksum.h>
67 #include <obd_class.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
80 #include "filter_internal.h"
82 static struct lvfs_callback_ops filter_lvfs_ops;
83 cfs_mem_cache_t *ll_fmd_cachep;
85 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
86 void *cb_data, int error)
88 struct obd_export *exp = cb_data;
89 LASSERT(exp->exp_obd == obd);
90 obd_transno_commit_cb(obd, transno, exp, error);
91 class_export_cb_put(exp);
94 int filter_version_get_check(struct obd_export *exp,
95 struct obd_trans_info *oti, struct inode *inode)
99 if (inode == NULL || oti == NULL)
102 curr_version = fsfilt_get_version(exp->exp_obd, inode);
103 if ((__s64)curr_version == -EOPNOTSUPP)
105 /* VBR: version is checked always because costs nothing */
106 if (oti->oti_pre_version != 0 &&
107 oti->oti_pre_version != curr_version) {
108 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
109 oti->oti_pre_version, curr_version);
110 cfs_spin_lock(&exp->exp_lock);
111 exp->exp_vbr_failed = 1;
112 cfs_spin_unlock(&exp->exp_lock);
115 oti->oti_pre_version = curr_version;
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121 struct obd_trans_info *oti, int rc, int force_sync)
123 struct obd_device_target *obt = &exp->exp_obd->u.obt;
124 struct tg_export_data *ted = &exp->exp_target_data;
125 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
126 struct lsd_client_data *lcd;
129 int err, log_pri = D_RPCTRACE;
131 /* Propagate error code. */
135 if (!exp->exp_obd->obd_replayable || oti == NULL)
138 cfs_mutex_lock(&ted->ted_lcd_lock);
140 /* if the export has already been disconnected, we have no last_rcvd slot,
141 * update server data with latest transno then */
143 cfs_mutex_unlock(&ted->ted_lcd_lock);
144 CWARN("commit transaction for disconnected client %s: rc %d\n",
145 exp->exp_client_uuid.uuid, rc);
146 err = filter_update_server_data(exp->exp_obd);
150 /* we don't allocate new transnos for replayed requests */
151 cfs_spin_lock(&obt->obt_lut->lut_translock);
152 if (oti->oti_transno == 0) {
153 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
154 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
155 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
157 last_rcvd = oti->oti_transno;
158 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
159 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
160 if (unlikely(last_rcvd < le64_to_cpu(lcd->lcd_last_transno))) {
161 CERROR("Trying to overwrite bigger transno, on-disk: "
162 LPU64", new: "LPU64"\n",
163 le64_to_cpu(lcd->lcd_last_transno), last_rcvd);
164 cfs_spin_lock(&exp->exp_lock);
165 exp->exp_vbr_failed = 1;
166 cfs_spin_unlock(&exp->exp_lock);
167 cfs_spin_unlock(&obt->obt_lut->lut_translock);
168 cfs_mutex_unlock(&ted->ted_lcd_lock);
172 oti->oti_transno = last_rcvd;
174 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
175 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
176 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
177 cfs_spin_unlock(&obt->obt_lut->lut_translock);
180 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
182 off = ted->ted_lr_off;
184 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
185 ted->ted_lr_idx, ted->ted_lr_off);
188 class_export_cb_get(exp); /* released when the cb is called */
190 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
196 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
197 lcd, sizeof(*lcd), &off,
198 force_sync | exp->exp_need_sync);
200 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
208 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
209 last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
210 cfs_mutex_unlock(&ted->ted_lcd_lock);
214 void f_dput(struct dentry *dentry)
216 /* Can't go inside filter_ddelete because it can block */
217 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
218 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
219 LASSERT(atomic_read(&dentry->d_count) > 0);
224 static void init_brw_stats(struct brw_stats *brw_stats)
227 for (i = 0; i < BRW_LAST; i++)
228 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
231 static int lprocfs_init_rw_stats(struct obd_device *obd,
232 struct lprocfs_stats **stats)
236 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
237 LPROC_FILTER_LAST - 1;
238 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
242 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
243 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
244 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
245 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
246 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
251 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
252 plus the procfs overhead :( */
253 static int filter_export_stats_init(struct obd_device *obd,
254 struct obd_export *exp,
260 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
261 /* Self-export gets no proc entry */
264 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
266 /* Mask error for already created
274 struct nid_stat *tmp = exp->exp_nid_stats;
275 LASSERT(tmp != NULL);
277 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
278 if (tmp->nid_brw_stats == NULL)
279 GOTO(clean, rc = -ENOMEM);
281 init_brw_stats(tmp->nid_brw_stats);
282 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
283 0644, &filter_per_nid_stats_fops,
286 CWARN("Error adding the brw_stats file\n");
288 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
292 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
296 rc = lprocfs_nid_ldlm_stats_init(tmp);
306 /* Add client data to the FILTER. We use a bitmap to locate a free space
307 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
308 * Otherwise, we have just read the data from the last_rcvd file and
309 * we know its offset. */
310 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
313 struct obd_device_target *obt = &obd->u.obt;
314 struct tg_export_data *ted = &exp->exp_target_data;
315 struct lr_server_data *lsd = class_server_data(obd);
316 unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
317 int new_client = (cl_idx == -1);
321 LASSERT(bitmap != NULL);
322 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
325 if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
328 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
329 * there's no need for extra complication here
332 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
334 if (cl_idx >= LR_MAX_CLIENTS) {
335 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
339 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
340 cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
345 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
346 CERROR("FILTER client %d: bit already set in bitmap!\n",
352 ted->ted_lr_idx = cl_idx;
353 ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
354 cl_idx * le16_to_cpu(lsd->lsd_client_size);
355 cfs_mutex_init(&ted->ted_lcd_lock);
356 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
358 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
359 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
362 struct lvfs_run_ctxt saved;
363 loff_t off = ted->ted_lr_off;
367 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
368 ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
370 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
373 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
374 /* Transaction needed to fix bug 1403 */
375 handle = fsfilt_start(obd,
376 obt->obt_rcvd_filp->f_dentry->d_inode,
377 FSFILT_OP_SETATTR, NULL);
378 if (IS_ERR(handle)) {
379 rc = PTR_ERR(handle);
380 CERROR("unable to start transaction: rc %d\n", rc);
382 ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
383 exp->exp_last_request_time = cfs_time_current_sec();
384 rc = fsfilt_add_journal_cb(obd, 0, handle,
385 target_client_add_cb,
386 class_export_cb_get(exp));
388 cfs_spin_lock(&exp->exp_lock);
389 exp->exp_need_sync = 1;
390 cfs_spin_unlock(&exp->exp_lock);
392 rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
394 sizeof(*ted->ted_lcd),
395 &off, rc /* sync if no cb */);
397 obt->obt_rcvd_filp->f_dentry->d_inode,
400 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
403 CERROR("error writing %s client idx %u: rc %d\n",
404 LAST_RCVD, ted->ted_lr_idx, rc);
411 static int filter_client_del(struct obd_export *exp)
413 struct tg_export_data *ted = &exp->exp_target_data;
414 struct obd_device_target *obt = &exp->exp_obd->u.obt;
415 struct lvfs_run_ctxt saved;
420 if (ted->ted_lcd == NULL)
423 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
424 if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
427 LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
429 off = ted->ted_lr_off;
431 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
432 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
434 /* Don't clear ted_lr_idx here as it is likely also unset. At worst
435 * we leak a client slot that will be cleaned on the next recovery. */
437 CERROR("%s: client idx %d has med_off %lld\n",
438 exp->exp_obd->obd_name, ted->ted_lr_idx, off);
439 GOTO(free, rc = -EINVAL);
442 /* Clear the bit _after_ zeroing out the client so we don't
443 race with filter_client_add and zero out new clients.*/
444 if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
445 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
450 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
451 /* Make sure the server's last_transno is up to date.
452 * This should be done before zeroing client slot so last_transno will
453 * be in server data or in client data in case of failure */
454 filter_update_server_data(exp->exp_obd);
456 cfs_mutex_lock(&ted->ted_lcd_lock);
457 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
458 rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
460 sizeof(*ted->ted_lcd), &off, 0);
461 cfs_mutex_unlock(&ted->ted_lcd_lock);
462 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
464 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
465 "zero out client %s at idx %u/%llu in %s, rc %d\n",
466 ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
473 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
474 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
475 struct filter_mod_data *fmd)
477 LASSERT_SPIN_LOCKED(&fed->fed_lock);
478 if (--fmd->fmd_refcount == 0) {
479 /* XXX when we have persistent reservations and the handle
480 * is stored herein we need to drop it here. */
481 fed->fed_mod_count--;
482 cfs_list_del(&fmd->fmd_list);
483 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
487 /* drop fmd reference, free it if last ref */
488 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
490 struct filter_export_data *fed;
495 fed = &exp->exp_filter_data;
496 cfs_spin_lock(&fed->fed_lock);
497 filter_fmd_put_nolock(fed, fmd); /* caller reference */
498 cfs_spin_unlock(&fed->fed_lock);
501 /* expire entries from the end of the list if there are too many
502 * or they are too old */
503 static void filter_fmd_expire_nolock(struct filter_obd *filter,
504 struct filter_export_data *fed,
505 struct filter_mod_data *keep)
507 struct filter_mod_data *fmd, *tmp;
509 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
513 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
514 fed->fed_mod_count < filter->fo_fmd_max_num)
517 cfs_list_del_init(&fmd->fmd_list);
518 filter_fmd_put_nolock(fed, fmd); /* list reference */
522 void filter_fmd_expire(struct obd_export *exp)
524 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
525 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
526 &exp->exp_filter_data, NULL);
527 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
530 /* find specified objid, group in export fmd list.
531 * caller must hold fed_lock and take fmd reference itself */
532 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
533 struct filter_export_data *fed,
534 obd_id objid, obd_seq group)
536 struct filter_mod_data *found = NULL, *fmd;
538 LASSERT_SPIN_LOCKED(&fed->fed_lock);
540 cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
541 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
543 cfs_list_del(&fmd->fmd_list);
544 cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
545 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
550 filter_fmd_expire_nolock(filter, fed, found);
555 /* Find fmd based on objid and group, or return NULL if not found. */
556 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
557 obd_id objid, obd_seq group)
559 struct filter_mod_data *fmd;
561 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
562 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
563 &exp->exp_filter_data, objid, group);
565 fmd->fmd_refcount++; /* caller reference */
566 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
571 /* Find fmd based on objid and group, or create a new one if none is found.
572 * It is possible for this function to return NULL under memory pressure,
573 * or if objid = 0 is passed (which will only cause old entries to expire).
574 * Currently this is not fatal because any fmd state is transient and
575 * may also be freed when it gets sufficiently old. */
576 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
577 obd_id objid, obd_seq group)
579 struct filter_export_data *fed = &exp->exp_filter_data;
580 struct filter_mod_data *found = NULL, *fmd_new = NULL;
582 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
584 cfs_spin_lock(&fed->fed_lock);
585 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
588 cfs_list_add_tail(&fmd_new->fmd_list,
590 fmd_new->fmd_id = objid;
591 fmd_new->fmd_gr = group;
592 fmd_new->fmd_refcount++; /* list reference */
594 fed->fed_mod_count++;
596 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
600 found->fmd_refcount++; /* caller reference */
601 found->fmd_expire = jiffies +
602 exp->exp_obd->u.filter.fo_fmd_max_age;
605 cfs_spin_unlock(&fed->fed_lock);
611 /* drop fmd list reference so it will disappear when last reference is put.
612 * This isn't so critical because it would in fact only affect the one client
613 * that is doing the unlink and at worst we have an stale entry referencing
614 * an object that should never be used again. */
615 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
617 struct filter_mod_data *found = NULL;
619 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
620 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
622 cfs_list_del_init(&found->fmd_list);
623 filter_fmd_put_nolock(&exp->exp_filter_data, found);
625 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
628 #define filter_fmd_drop(exp, objid, group)
631 /* remove all entries from fmd list */
632 static void filter_fmd_cleanup(struct obd_export *exp)
634 struct filter_export_data *fed = &exp->exp_filter_data;
635 struct filter_mod_data *fmd = NULL, *tmp;
637 cfs_spin_lock(&fed->fed_lock);
638 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
639 cfs_list_del_init(&fmd->fmd_list);
640 filter_fmd_put_nolock(fed, fmd);
642 cfs_spin_unlock(&fed->fed_lock);
645 static int filter_init_export(struct obd_export *exp)
650 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
651 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
653 cfs_spin_lock(&exp->exp_lock);
654 exp->exp_connecting = 1;
655 cfs_spin_unlock(&exp->exp_lock);
657 /* self-export doesn't need client data and ldlm initialization */
658 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
659 &exp->exp_client_uuid)))
662 rc = lut_client_alloc(exp);
664 rc = ldlm_init_export(exp);
666 CERROR("%s: Can't initialize export: rc %d\n",
667 exp->exp_obd->obd_name, rc);
672 static int filter_free_server_data(struct obd_device_target *obt)
674 lut_fini(NULL, obt->obt_lut);
675 OBD_FREE_PTR(obt->obt_lut);
679 /* assumes caller is already in kernel ctxt */
680 int filter_update_server_data(struct obd_device *obd)
682 struct file *filp = obd->u.obt.obt_rcvd_filp;
683 struct lr_server_data *lsd = class_server_data(obd);
688 CDEBUG(D_INODE, "server uuid : %s\n", lsd->lsd_uuid);
689 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
690 le64_to_cpu(lsd->lsd_last_transno));
691 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
692 le64_to_cpu(lsd->lsd_mount_count));
694 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
696 CERROR("error writing lr_server_data: rc = %d\n", rc);
701 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
704 struct filter_obd *filter = &obd->u.filter;
710 if (filter->fo_last_objid_files[group] == NULL) {
711 CERROR("Object seq "LPU64" not fully setup; not updating "
712 "last_objid\n", group);
716 CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
717 obd->obd_name, filter->fo_last_objids[group], group);
719 tmp = cpu_to_le64(filter->fo_last_objids[group]);
720 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
721 &tmp, sizeof(tmp), &off, force_sync);
723 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
727 extern int ost_handle(struct ptlrpc_request *req);
728 /* assumes caller has already in kernel ctxt */
729 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
731 struct filter_obd *filter = &obd->u.filter;
732 struct lr_server_data *lsd;
733 struct lsd_client_data *lcd = NULL;
734 struct inode *inode = filp->f_dentry->d_inode;
735 unsigned long last_rcvd_size = i_size_read(inode);
736 struct lu_target *lut;
743 /* ensure padding in the struct is the correct size */
744 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
745 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
746 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
747 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
749 /* allocate and initialize lu_target */
753 rc = lut_init(NULL, lut, obd, NULL);
756 lsd = class_server_data(obd);
757 if (last_rcvd_size == 0) {
758 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
760 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
761 lsd->lsd_last_transno = 0;
762 mount_count = lsd->lsd_mount_count = 0;
763 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
764 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
765 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
766 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
767 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
768 /* OBD_COMPAT_OST is set in filter_connect_internal when the
769 * MDS first connects and assigns the OST index number. */
770 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
773 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
775 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
779 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
780 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
781 "using the wrong disk %s. Were the "
782 "/dev/ assignments rearranged?\n",
783 obd->obd_uuid.uuid, lsd->lsd_uuid);
784 GOTO(err_lut, rc = -EINVAL);
786 mount_count = le64_to_cpu(lsd->lsd_mount_count);
787 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
789 /* Assume old last_rcvd format unless I_C_LR is set */
790 if (!(lsd->lsd_feature_incompat &
791 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
792 lsd->lsd_last_transno = lsd->lsd_compat14;
794 /* OBD_COMPAT_OST is set in filter_connect_internal when the
795 * MDS first connects and assigns the OST index number. */
796 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
800 if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
801 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
802 obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
803 ~FILTER_INCOMPAT_SUPP);
804 GOTO(err_lut, rc = -EINVAL);
806 if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
807 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
808 obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
809 ~FILTER_ROCOMPAT_SUPP);
810 /* Do something like remount filesystem read-only */
811 GOTO(err_lut, rc = -EINVAL);
814 start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
816 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
817 obd->obd_name, start_epoch);
818 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
819 obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
820 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
821 obd->obd_name, mount_count + 1);
822 CDEBUG(D_INODE, "%s: server data size: %u\n",
823 obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
824 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
825 obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
826 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
827 obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
828 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
829 obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
830 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
831 last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
832 (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
833 le16_to_cpu(lsd->lsd_client_size));
835 if (!obd->obd_replayable) {
836 CWARN("%s: recovery support OFF\n", obd->obd_name);
842 GOTO(err_client, rc = -ENOMEM);
844 for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
845 off < last_rcvd_size; cl_idx++) {
847 struct obd_export *exp;
848 struct filter_export_data *fed;
850 /* Don't assume off is incremented properly by
851 * fsfilt_read_record(), in case sizeof(*lcd)
852 * isn't the same as lsd->lsd_client_size. */
853 off = le32_to_cpu(lsd->lsd_client_start) +
854 cl_idx * le16_to_cpu(lsd->lsd_client_size);
855 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
857 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
858 LAST_RCVD, cl_idx, off, rc);
859 break; /* read error shouldn't cause startup to fail */
862 if (lcd->lcd_uuid[0] == '\0') {
863 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
868 check_lcd(obd->obd_name, cl_idx, lcd);
870 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
872 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
873 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
874 last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
876 /* These exports are cleaned up by filter_disconnect(), so they
877 * need to be set up like real exports as filter_connect() does.
879 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
881 if (PTR_ERR(exp) == -EALREADY) {
882 /* export already exists, zero out this one */
883 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
887 GOTO(err_client, rc = PTR_ERR(exp));
890 fed = &exp->exp_filter_data;
891 *fed->fed_ted.ted_lcd = *lcd;
892 fed->fed_group = 0; /* will be assigned at connect */
893 filter_export_stats_init(obd, exp, NULL);
894 rc = filter_client_add(obd, exp, cl_idx);
895 /* can't fail for existing client */
896 LASSERTF(rc == 0, "rc = %d\n", rc);
898 /* VBR: set export last committed */
899 exp->exp_last_committed = last_rcvd;
900 cfs_spin_lock(&exp->exp_lock);
901 exp->exp_connecting = 0;
902 exp->exp_in_recovery = 0;
903 cfs_spin_unlock(&exp->exp_lock);
904 obd->obd_max_recoverable_clients++;
905 class_export_put(exp);
907 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
908 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
912 obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
914 obd->u.obt.obt_mount_count = mount_count + 1;
915 obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
916 lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
918 /* save it, so mount count and last_transno is current */
919 rc = filter_update_server_data(obd);
921 GOTO(err_client, rc);
926 class_disconnect_exports(obd);
928 filter_free_server_data(&obd->u.obt);
932 static int filter_cleanup_groups(struct obd_device *obd)
934 struct filter_obd *filter = &obd->u.filter;
936 struct dentry *dentry;
940 if (filter->fo_dentry_O_groups != NULL) {
941 for (i = 0; i < filter->fo_group_count; i++) {
942 dentry = filter->fo_dentry_O_groups[i];
946 OBD_FREE(filter->fo_dentry_O_groups,
947 filter->fo_group_count *
948 sizeof(*filter->fo_dentry_O_groups));
949 filter->fo_dentry_O_groups = NULL;
951 if (filter->fo_last_objid_files != NULL) {
952 for (i = 0; i < filter->fo_group_count; i++) {
953 filp = filter->fo_last_objid_files[i];
957 OBD_FREE(filter->fo_last_objid_files,
958 filter->fo_group_count *
959 sizeof(*filter->fo_last_objid_files));
960 filter->fo_last_objid_files = NULL;
962 if (filter->fo_dentry_O_sub != NULL) {
963 for (i = 0; i < filter->fo_group_count; i++) {
964 for (j = 0; j < filter->fo_subdir_count; j++) {
965 dentry = filter->fo_dentry_O_sub[i].dentry[j];
970 OBD_FREE(filter->fo_dentry_O_sub,
971 filter->fo_group_count *
972 sizeof(*filter->fo_dentry_O_sub));
973 filter->fo_dentry_O_sub = NULL;
975 if (filter->fo_last_objids != NULL) {
976 OBD_FREE(filter->fo_last_objids,
977 filter->fo_group_count *
978 sizeof(*filter->fo_last_objids));
979 filter->fo_last_objids = NULL;
981 if (filter->fo_dentry_O != NULL) {
982 f_dput(filter->fo_dentry_O);
983 filter->fo_dentry_O = NULL;
988 static int filter_update_last_group(struct obd_device *obd, int group)
990 struct filter_obd *filter = &obd->u.filter;
991 struct file *filp = NULL;
992 int last_group = 0, rc;
996 if (group <= filter->fo_committed_group)
999 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1003 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1007 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1009 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1013 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1014 obd->obd_name, last_group, group);
1018 /* must be sync: bXXXX */
1019 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1021 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1025 filter->fo_committed_group = group;
1028 filp_close(filp, 0);
1032 static int filter_read_group_internal(struct obd_device *obd, int group,
1035 struct filter_obd *filter = &obd->u.filter;
1036 __u64 *new_objids = NULL;
1037 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1038 struct dentry **new_groups = NULL;
1039 struct file **new_files = NULL;
1040 struct dentry *dentry;
1042 int old_count = filter->fo_group_count, rc, stage = 0, i;
1046 int len = group + 1;
1048 snprintf(name, 24, "%d", group);
1052 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1054 if (IS_ERR(dentry)) {
1055 CERROR("Cannot lookup expected object group %d: %ld\n",
1056 group, PTR_ERR(dentry));
1057 RETURN(PTR_ERR(dentry));
1060 dentry = simple_mkdir(filter->fo_dentry_O,
1061 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1062 if (IS_ERR(dentry)) {
1063 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1065 RETURN(PTR_ERR(dentry));
1070 snprintf(name, 24, "O/%d/LAST_ID", group);
1072 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1074 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1075 GOTO(cleanup, rc = PTR_ERR(filp));
1079 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1081 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1085 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1086 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1087 if (tmp_subdirs == NULL)
1088 GOTO(cleanup, rc = -ENOMEM);
1091 for (i = 0; i < filter->fo_subdir_count; i++) {
1093 snprintf(dir, sizeof(dir), "d%u", i);
1095 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1096 obd->u.obt.obt_vfsmnt,
1098 if (IS_ERR(tmp_subdirs->dentry[i])) {
1099 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1100 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1105 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1106 tmp_subdirs->dentry[i]);
1110 /* 'group' is an index; we need an array of length 'group + 1' */
1111 if (group + 1 > old_count) {
1112 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1113 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1114 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1115 OBD_ALLOC(new_files, len * sizeof(*new_files));
1117 if (new_objids == NULL || new_subdirs == NULL ||
1118 new_groups == NULL || new_files == NULL)
1119 GOTO(cleanup, rc = -ENOMEM);
1122 memcpy(new_objids, filter->fo_last_objids,
1123 old_count * sizeof(*new_objids));
1124 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1125 old_count * sizeof(*new_subdirs));
1126 memcpy(new_groups, filter->fo_dentry_O_groups,
1127 old_count * sizeof(*new_groups));
1128 memcpy(new_files, filter->fo_last_objid_files,
1129 old_count * sizeof(*new_files));
1131 OBD_FREE(filter->fo_last_objids,
1132 old_count * sizeof(*new_objids));
1133 OBD_FREE(filter->fo_dentry_O_sub,
1134 old_count * sizeof(*new_subdirs));
1135 OBD_FREE(filter->fo_dentry_O_groups,
1136 old_count * sizeof(*new_groups));
1137 OBD_FREE(filter->fo_last_objid_files,
1138 old_count * sizeof(*new_files));
1140 filter->fo_last_objids = new_objids;
1141 filter->fo_dentry_O_sub = new_subdirs;
1142 filter->fo_dentry_O_groups = new_groups;
1143 filter->fo_last_objid_files = new_files;
1144 filter->fo_group_count = len;
1147 filter->fo_dentry_O_groups[group] = dentry;
1148 filter->fo_last_objid_files[group] = filp;
1149 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1150 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1151 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1154 filter_update_last_group(obd, group);
1156 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1157 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1158 rc = filter_update_last_objid(obd, group, 1);
1162 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1163 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1164 obd->obd_name, group, last_objid);
1169 if (new_objids != NULL)
1170 OBD_FREE(new_objids, len * sizeof(*new_objids));
1171 if (new_subdirs != NULL)
1172 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1173 if (new_groups != NULL)
1174 OBD_FREE(new_groups, len * sizeof(*new_groups));
1175 if (new_files != NULL)
1176 OBD_FREE(new_files, len * sizeof(*new_files));
1178 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1179 for (i = 0; i < filter->fo_subdir_count; i++) {
1180 if (tmp_subdirs->dentry[i] != NULL)
1181 dput(tmp_subdirs->dentry[i]);
1183 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1186 filp_close(filp, 0);
1193 static int filter_read_groups(struct obd_device *obd, int last_group,
1196 struct filter_obd *filter = &obd->u.filter;
1197 int old_count, group, rc = 0;
1199 cfs_mutex_lock(&filter->fo_init_lock);
1200 old_count = filter->fo_group_count;
1201 for (group = old_count; group <= last_group; group++) {
1202 rc = filter_read_group_internal(obd, group, create);
1206 cfs_mutex_unlock(&filter->fo_init_lock);
1210 /* FIXME: object groups */
1211 static int filter_prep_groups(struct obd_device *obd)
1213 struct filter_obd *filter = &obd->u.filter;
1214 struct dentry *O_dentry;
1216 int last_group, rc = 0, cleanup_phase = 0;
1220 O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1222 CDEBUG(D_INODE, "%s: got/created O: %p\n", obd->obd_name, O_dentry);
1223 if (IS_ERR(O_dentry)) {
1224 rc = PTR_ERR(O_dentry);
1225 CERROR("%s: cannot open/create O: rc = %d\n", obd->obd_name,rc);
1228 filter->fo_dentry_O = O_dentry;
1229 cleanup_phase = 1; /* O_dentry */
1231 /* we have to initialize all groups before first connections from
1232 * clients because they may send create/destroy for any group -bzzz */
1233 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1235 CERROR("%s: cannot create LAST_GROUP: rc = %ld\n",
1236 obd->obd_name, PTR_ERR(filp));
1237 GOTO(cleanup, rc = PTR_ERR(filp));
1239 cleanup_phase = 2; /* filp */
1241 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1243 CERROR("%s: error reading LAST_GROUP: rc %d\n",
1249 last_group = FID_SEQ_OST_MDT0;
1251 CDEBUG(D_INODE, "%s: initialize group %u (max %u)\n", obd->obd_name,
1252 FID_SEQ_OST_MDT0, last_group);
1253 filter->fo_committed_group = last_group;
1254 rc = filter_read_groups(obd, last_group, 1);
1258 filp_close(filp, 0);
1262 switch (cleanup_phase) {
1264 filp_close(filp, 0);
1266 filter_cleanup_groups(obd);
1267 f_dput(filter->fo_dentry_O);
1268 filter->fo_dentry_O = NULL;
1276 /* setup the object store with correct subdirectories */
1277 static int filter_prep(struct obd_device *obd)
1279 struct lvfs_run_ctxt saved;
1280 struct filter_obd *filter = &obd->u.filter;
1282 struct inode *inode;
1286 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1287 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1288 if (!file || IS_ERR(file)) {
1290 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1294 obd->u.obt.obt_rcvd_filp = file;
1295 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1296 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1297 file->f_dentry->d_inode->i_mode);
1298 GOTO(err_filp, rc = -ENOENT);
1301 inode = file->f_dentry->d_parent->d_inode;
1302 /* We use i_op->unlink directly in filter_vfs_unlink() */
1303 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1304 CERROR("%s: filesystem does not support create/unlink ops\n",
1306 GOTO(err_filp, rc = -EOPNOTSUPP);
1309 rc = filter_init_server_data(obd, file);
1311 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1314 LASSERT(obd->u.obt.obt_lut);
1315 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1317 /* open and create health check io file*/
1318 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1321 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1323 GOTO(err_server_data, rc);
1325 filter->fo_obt.obt_health_check_filp = file;
1326 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1327 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1328 file->f_dentry->d_inode->i_mode);
1329 GOTO(err_health_check, rc = -ENOENT);
1331 rc = lvfs_check_io_health(obd, file);
1333 GOTO(err_health_check, rc);
1335 rc = filter_prep_groups(obd);
1337 GOTO(err_health_check, rc);
1339 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1344 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1345 CERROR("can't close %s after error\n", HEALTH_CHECK);
1346 filter->fo_obt.obt_health_check_filp = NULL;
1348 target_recovery_fini(obd);
1349 filter_free_server_data(&obd->u.obt);
1351 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1352 CERROR("can't close %s after error\n", LAST_RCVD);
1353 obd->u.obt.obt_rcvd_filp = NULL;
1357 /* cleanup the filter: write last used object id to status file */
1358 static void filter_post(struct obd_device *obd)
1360 struct lvfs_run_ctxt saved;
1361 struct filter_obd *filter = &obd->u.filter;
1364 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1365 * best to start a transaction with h_sync, because we removed this
1368 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1369 rc = filter_update_server_data(obd);
1371 CERROR("error writing server data: rc = %d\n", rc);
1373 for (i = 0; i < filter->fo_group_count; i++) {
1374 rc = filter_update_last_objid(obd, i,
1375 (i == filter->fo_group_count - 1));
1377 CERROR("error writing group %d lastobjid: rc = %d\n",
1381 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1382 obd->u.obt.obt_rcvd_filp = NULL;
1384 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1386 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1387 filter->fo_obt.obt_health_check_filp = NULL;
1389 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1391 filter_cleanup_groups(obd);
1392 filter_free_server_data(&obd->u.obt);
1393 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1395 filter_free_capa_keys(filter);
1396 cleanup_capa_hash(filter->fo_capa_hash);
1399 static void filter_set_last_id(struct filter_obd *filter,
1400 obd_id id, obd_seq group)
1402 LASSERT(group <= filter->fo_group_count);
1404 cfs_spin_lock(&filter->fo_objidlock);
1405 filter->fo_last_objids[group] = id;
1406 cfs_spin_unlock(&filter->fo_objidlock);
1409 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1412 LASSERT(group <= filter->fo_group_count);
1413 LASSERT(filter->fo_last_objids != NULL);
1415 /* FIXME: object groups */
1416 cfs_spin_lock(&filter->fo_objidlock);
1417 id = filter->fo_last_objids[group];
1418 cfs_spin_unlock(&filter->fo_objidlock);
1422 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1424 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1428 /* We never dget the object parent, so DON'T dput it either */
1429 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1431 struct filter_obd *filter = &obd->u.filter;
1432 struct filter_subdirs *subdirs;
1434 if (group >= filter->fo_group_count) /* FIXME: object groups */
1435 return ERR_PTR(-EBADF);
1437 if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1438 return filter->fo_dentry_O_groups[group];
1440 subdirs = &filter->fo_dentry_O_sub[group];
1441 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1444 /* We never dget the object parent, so DON'T dput it either */
1445 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1448 unsigned long now = jiffies;
1449 struct dentry *dparent = filter_parent(obd, group, objid);
1452 if (IS_ERR(dparent))
1454 if (dparent == NULL)
1455 return ERR_PTR(-ENOENT);
1457 rc = filter_lock_dentry(obd, dparent);
1458 fsfilt_check_slow(obd, now, "parent lock");
1459 return rc ? ERR_PTR(rc) : dparent;
1462 /* We never dget the object parent, so DON'T dput it either */
1463 static void filter_parent_unlock(struct dentry *dparent)
1465 UNLOCK_INODE_MUTEX(dparent->d_inode);
1468 /* How to get files, dentries, inodes from object id's.
1470 * If dir_dentry is passed, the caller has already locked the parent
1471 * appropriately for this operation (normally a write lock). If
1472 * dir_dentry is NULL, we do a read lock while we do the lookup to
1473 * avoid races with create/destroy and such changing the directory
1474 * internal to the filesystem code. */
1475 struct dentry *filter_fid2dentry(struct obd_device *obd,
1476 struct dentry *dir_dentry,
1477 obd_seq group, obd_id id)
1479 struct dentry *dparent = dir_dentry;
1480 struct dentry *dchild;
1485 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1486 obd->u.filter.fo_destroys_in_progress == 0) {
1487 /* don't fail lookups for orphan recovery, it causes
1488 * later LBUGs when objects still exist during precreate */
1489 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1490 RETURN(ERR_PTR(-ENOENT));
1493 CERROR("fatal: invalid object id 0\n");
1494 RETURN(ERR_PTR(-ESTALE));
1497 len = sprintf(name, LPU64, id);
1498 if (dir_dentry == NULL) {
1499 dparent = filter_parent_lock(obd, group, id);
1500 if (IS_ERR(dparent)) {
1501 CERROR("%s: error getting object "POSTID
1502 " parent: rc %ld\n", obd->obd_name,
1503 id, group, PTR_ERR(dparent));
1507 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1508 dparent->d_name.len, dparent->d_name.name, name);
1509 /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1510 dchild = lookup_one_len(name, dparent, len);
1511 if (dir_dentry == NULL)
1512 filter_parent_unlock(dparent);
1513 if (IS_ERR(dchild)) {
1514 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1515 obd->obd_name, id, group, PTR_ERR(dchild));
1519 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1520 CERROR("%s: got bad object "LPU64" inode %lu\n",
1521 obd->obd_name, id, dchild->d_inode->i_ino);
1523 RETURN(ERR_PTR(-ENOENT));
1526 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1527 name, dchild, atomic_read(&dchild->d_count));
1529 LASSERT(atomic_read(&dchild->d_count) > 0);
1534 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1535 obd_id group, struct lustre_handle *lockh)
1537 int flags = LDLM_AST_DISCARD_DATA, rc;
1538 struct ldlm_res_id res_id;
1539 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1542 osc_build_res_name(objid, group, &res_id);
1543 /* Tell the clients that the object is gone now and that they should
1544 * throw away any cached pages. */
1545 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1546 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1547 ldlm_completion_ast, NULL, NULL, 0, NULL,
1554 static void filter_fini_destroy(struct obd_device *obd,
1555 struct lustre_handle *lockh)
1557 if (lustre_handle_is_used(lockh))
1558 ldlm_lock_decref(lockh, LCK_PW);
1561 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1562 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1563 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1564 * reverse. Caller must take i_sem before starting the transaction and we
1565 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1566 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1567 struct vfsmount *mnt)
1572 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1573 * itself we already hold dir->i_mutex for child create/unlink ops */
1574 LASSERT(dentry->d_inode != NULL);
1575 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1576 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1580 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1581 GOTO(out, rc = -ENOENT);
1583 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1588 GOTO(out, rc = -EPERM);
1590 /* check_sticky() */
1591 if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1592 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1593 IS_IMMUTABLE(dentry->d_inode))
1594 GOTO(out, rc = -EPERM);
1596 /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
1597 ll_vfs_dq_init(dir);
1599 rc = ll_security_inode_unlink(dir, dentry, mnt);
1603 rc = dir->i_op->unlink(dir, dentry);
1605 /* need to drop i_mutex before we lose inode reference */
1606 UNLOCK_INODE_MUTEX(dentry->d_inode);
1613 /* Caller must hold LCK_PW on parent and push us into kernel context.
1614 * Caller must hold child i_mutex, we drop it always.
1615 * Caller is also required to ensure that dchild->d_inode exists. */
1616 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1617 obd_seq group, struct dentry *dparent,
1618 struct dentry *dchild)
1620 struct inode *inode = dchild->d_inode;
1623 /* There should be 2 references to the inode:
1624 * 1) taken by filter_prepare_destroy
1625 * 2) taken by filter_destroy */
1626 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1627 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1628 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1629 (unsigned long)inode->i_nlink,
1630 atomic_read(&inode->i_count));
1633 rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1635 CERROR("error unlinking objid %.*s: rc %d\n",
1636 dchild->d_name.len, dchild->d_name.name, rc);
1640 struct filter_intent_args {
1641 struct ldlm_lock **victim;
1646 static enum interval_iter filter_intent_cb(struct interval_node *n,
1649 struct ldlm_interval *node = (struct ldlm_interval *)n;
1650 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1651 __u64 size = arg->size;
1652 struct ldlm_lock **v = arg->victim;
1653 struct ldlm_lock *lck;
1655 /* If the interval is lower than the current file size,
1657 if (interval_high(n) <= size)
1658 return INTERVAL_ITER_STOP;
1660 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1661 /* Don't send glimpse ASTs to liblustre clients.
1662 * They aren't listening for them, and they do
1663 * entirely synchronous I/O anyways. */
1664 if (lck->l_export == NULL ||
1665 lck->l_export->exp_libclient == 1)
1668 if (*arg->liblustre)
1669 *arg->liblustre = 0;
1672 *v = LDLM_LOCK_GET(lck);
1673 } else if ((*v)->l_policy_data.l_extent.start <
1674 lck->l_policy_data.l_extent.start) {
1675 LDLM_LOCK_RELEASE(*v);
1676 *v = LDLM_LOCK_GET(lck);
1679 /* the same policy group - every lock has the
1680 * same extent, so needn't do it any more */
1684 return INTERVAL_ITER_CONT;
1687 static int filter_intent_policy(struct ldlm_namespace *ns,
1688 struct ldlm_lock **lockp, void *req_cookie,
1689 ldlm_mode_t mode, int flags, void *data)
1691 struct ptlrpc_request *req = req_cookie;
1692 struct ldlm_lock *lock = *lockp, *l = NULL;
1693 struct ldlm_resource *res = lock->l_resource;
1694 ldlm_processing_policy policy;
1695 struct ost_lvb *res_lvb, *reply_lvb;
1696 struct ldlm_reply *rep;
1698 int idx, rc, tmpflags = 0, only_liblustre = 1;
1699 struct ldlm_interval_tree *tree;
1700 struct filter_intent_args arg;
1701 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1702 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1703 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1706 policy = ldlm_get_processing_policy(res);
1707 LASSERT(policy != NULL);
1708 LASSERT(req != NULL);
1710 rc = lustre_pack_reply(req, 3, repsize, NULL);
1712 RETURN(req->rq_status = rc);
1714 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1715 LASSERT(rep != NULL);
1717 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1718 sizeof(*reply_lvb));
1719 LASSERT(reply_lvb != NULL);
1721 //fixup_handle_for_resent_req(req, lock, &lockh);
1723 /* Call the extent policy function to see if our request can be
1724 * granted, or is blocked.
1725 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1726 * lock, and should not be granted if the lock will be blocked.
1729 if (flags & LDLM_FL_BLOCK_NOWAIT) {
1730 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_AGL_DELAY, 5);
1732 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_AGL_NOLOCK))
1733 RETURN(ELDLM_LOCK_ABORTED);
1736 LASSERT(ns == ldlm_res_to_ns(res));
1738 rc = policy(lock, &tmpflags, 0, &err, NULL);
1739 check_res_locked(res);
1741 /* The lock met with no resistance; we're finished. */
1742 if (rc == LDLM_ITER_CONTINUE) {
1743 /* do not grant locks to the liblustre clients: they cannot
1744 * handle ASTs robustly. We need to do this while still
1745 * holding lr_lock to avoid the lock remaining on the res_link
1746 * list (and potentially being added to l_pending_list by an
1747 * AST) when we are going to drop this lock ASAP. */
1748 if (lock->l_export->exp_libclient ||
1749 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1750 ldlm_resource_unlink_lock(lock);
1751 err = ELDLM_LOCK_ABORTED;
1753 err = ELDLM_LOCK_REPLACED;
1757 } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1758 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1759 * callback for glimpse size. The real size user will trigger
1760 * the glimpse callback when necessary. */
1762 RETURN(ELDLM_LOCK_ABORTED);
1765 /* Do not grant any lock, but instead send GL callbacks. The extent
1766 * policy nicely created a list of all PW locks for us. We will choose
1767 * the highest of those which are larger than the size in the LVB, if
1768 * any, and perform a glimpse callback. */
1769 res_lvb = res->lr_lvb_data;
1770 LASSERT(res_lvb != NULL);
1771 *reply_lvb = *res_lvb;
1774 * lr_lock guarantees that no new locks are granted, and,
1775 * therefore, that res->lr_lvb_data cannot increase beyond the
1776 * end of already granted lock. As a result, it is safe to
1777 * check against "stale" reply_lvb->lvb_size value without
1778 * res->lr_lvb_mutex.
1780 arg.size = reply_lvb->lvb_size;
1782 arg.liblustre = &only_liblustre;
1783 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1784 tree = &res->lr_itree[idx];
1785 if (tree->lit_mode == LCK_PR)
1788 interval_iterate_reverse(tree->lit_root,
1789 filter_intent_cb, &arg);
1793 /* There were no PW locks beyond the size in the LVB; finished. */
1795 if (only_liblustre) {
1796 /* If we discovered a liblustre client with a PW lock,
1797 * however, the LVB may be out of date! The LVB is
1798 * updated only on glimpse (which we don't do for
1799 * liblustre clients) and cancel (which the client
1800 * obviously has not yet done). So if it has written
1801 * data but kept the lock, the LVB is stale and needs
1802 * to be updated from disk.
1804 * Of course, this will all disappear when we switch to
1805 * taking liblustre locks on the OST. */
1806 ldlm_res_lvbo_update(res, NULL, 1);
1808 RETURN(ELDLM_LOCK_ABORTED);
1812 * This check is for lock taken in filter_prepare_destroy() that does
1813 * not have l_glimpse_ast set. So the logic is: if there is a lock
1814 * with no l_glimpse_ast set, this object is being destroyed already.
1816 * Hence, if you are grabbing DLM locks on the server, always set
1817 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1819 if (l->l_glimpse_ast == NULL) {
1820 /* We are racing with unlink(); just return -ENOENT */
1821 rep->lock_policy_res1 = -ENOENT;
1825 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1826 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1829 *reply_lvb = *res_lvb;
1833 LDLM_LOCK_RELEASE(l);
1835 RETURN(ELDLM_LOCK_ABORTED);
1839 * per-obd_device iobuf pool.
1841 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1842 * allocations in write-path has to be minimized (see bug 5137).
1844 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1845 * OST threads (see ost_thread_{init,done}()).
1847 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1848 * at the OST layer there are only (potentially) multiple obd_device of type
1849 * unknown at the time of OST thread creation.
1851 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1852 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1853 * it's very own iobuf.
1857 * filter_kiobuf_pool_init()
1859 * filter_kiobuf_pool_done()
1861 * filter_iobuf_get()
1863 * operate on this array. They are "generic" in a sense that they don't depend
1864 * on actual type of iobuf's (the latter depending on Linux kernel version).
1868 * destroy pool created by filter_iobuf_pool_init
1870 static void filter_iobuf_pool_done(struct filter_obd *filter)
1872 struct filter_iobuf **pool;
1877 pool = filter->fo_iobuf_pool;
1879 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1880 if (pool[i] != NULL)
1881 filter_free_iobuf(pool[i]);
1883 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1884 filter->fo_iobuf_pool = NULL;
1889 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1891 struct filter_obd *filter = &obd->u.filter;
1892 struct sptlrpc_rule_set tmp_rset;
1895 sptlrpc_rule_set_init(&tmp_rset);
1896 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1898 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1903 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1905 cfs_write_lock(&filter->fo_sptlrpc_lock);
1906 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1907 filter->fo_sptlrpc_rset = tmp_rset;
1908 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1914 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1916 static int filter_iobuf_pool_init(struct filter_obd *filter)
1923 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1925 if (filter->fo_iobuf_pool == NULL)
1928 filter->fo_iobuf_count = OSS_THREADS_MAX;
1933 /* Return iobuf allocated for @thread_id. We don't know in advance how
1934 * many threads there will be so we allocate a large empty array and only
1935 * fill in those slots that are actually in use.
1936 * If we haven't allocated a pool entry for this thread before, do so now. */
1937 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1939 int thread_id = (oti && oti->oti_thread) ?
1940 oti->oti_thread->t_id : -1;
1941 struct filter_iobuf *pool = NULL;
1942 struct filter_iobuf **pool_place = NULL;
1944 if (thread_id >= 0) {
1945 LASSERT(thread_id < filter->fo_iobuf_count);
1946 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1949 if (unlikely(pool == NULL)) {
1950 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1951 PTLRPC_MAX_BRW_PAGES);
1952 if (pool_place != NULL)
1959 /* mount the file system (secretly). lustre_cfg parameters are:
1962 * 3 = flags: failover=f, failout=n
1965 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1968 struct filter_obd *filter = &obd->u.filter;
1969 struct vfsmount *mnt;
1970 struct file_system_type *type;
1971 struct lustre_mount_info *lmi;
1972 struct obd_uuid uuid;
1976 struct request_queue *q;
1980 if (lcfg->lcfg_bufcount < 3 ||
1981 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1982 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1985 lmi = server_get_mount(obd->obd_name);
1987 /* We already mounted in lustre_fill_super.
1988 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1989 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1991 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1993 /* old path - used by lctl */
1994 CERROR("Using old MDS mount method\n");
1995 type = get_fs_type(lustre_cfg_string(lcfg, 2));
1997 CERROR("get_fs_type failed\n");
2000 mnt = vfs_kern_mount(type, MS_NOATIME|MS_NODIRATIME,
2001 lustre_cfg_string(lcfg, 1), option);
2002 cfs_module_put(type->owner);
2005 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2006 lustre_cfg_string(lcfg, 1), rc);
2010 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2012 if (IS_ERR(obd->obd_fsops))
2013 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2015 rc = filter_iobuf_pool_init(filter);
2019 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2020 CERROR("%s: Underlying device is marked as read-only. "
2021 "Setup failed\n", obd->obd_name);
2022 GOTO(err_ops, rc = -EROFS);
2025 /* failover is the default */
2026 obd->obd_replayable = 1;
2028 /* disable connection until configuration finishes */
2029 obd->obd_no_conn = 1;
2031 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2032 str = lustre_cfg_string(lcfg, 3);
2033 if (strchr(str, 'n')) {
2034 CWARN("%s: recovery disabled\n", obd->obd_name);
2035 obd->obd_replayable = 0;
2039 obd->u.obt.obt_magic = OBT_MAGIC;
2040 obd->u.obt.obt_vfsmnt = mnt;
2041 obd->u.obt.obt_sb = mnt->mnt_sb;
2042 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2043 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2045 rc = fsfilt_setup(obd, obd->u.obt.obt_sb);
2049 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2050 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2051 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2052 obd->obd_lvfs_ctxt.fs = get_ds();
2053 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2055 cfs_mutex_init(&filter->fo_init_lock);
2056 filter->fo_committed_group = 0;
2057 filter->fo_destroys_in_progress = 0;
2058 for (i = 0; i < 32; i++)
2059 cfs_mutex_init(&filter->fo_create_locks[i]);
2061 cfs_spin_lock_init(&filter->fo_objidlock);
2062 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2063 cfs_mutex_init(&filter->fo_alloc_lock);
2064 init_brw_stats(&filter->fo_filter_stats);
2065 cfs_spin_lock_init(&filter->fo_flags_lock);
2066 filter->fo_read_cache = 1; /* enable read-only cache by default */
2067 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2068 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2069 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2070 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2071 filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2072 filter_slc_set(filter); /* initialize sync on lock cancel */
2074 rc = filter_prep(obd);
2078 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2079 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2081 filter->fo_fl_oss_capa = 1;
2083 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2084 filter->fo_capa_hash = init_capa_hash();
2085 if (filter->fo_capa_hash == NULL)
2086 GOTO(err_post, rc = -ENOMEM);
2088 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2089 obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2090 LDLM_NAMESPACE_SERVER,
2091 LDLM_NAMESPACE_GREEDY,
2093 if (obd->obd_namespace == NULL)
2094 GOTO(err_post, rc = -ENOMEM);
2095 obd->obd_namespace->ns_lvbp = obd;
2096 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2097 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2099 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2100 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2102 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2104 CERROR("failed to setup llogging subsystems\n");
2108 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2109 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2110 /* do this after llog being initialized */
2111 filter_adapt_sptlrpc_conf(obd, 1);
2113 rc = lquota_setup(filter_quota_interface_ref, obd);
2117 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2118 if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2119 queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2120 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2121 "for larger I/O requests: max_sectors = %u "
2122 "could be up to max_hw_sectors=%u\n",
2123 obd->obd_name, mnt->mnt_sb->s_id,
2124 queue_max_sectors(q), queue_max_hw_sectors(q));
2126 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2127 if (uuid_ptr != NULL) {
2128 class_uuid_unparse(uuid_ptr, &uuid);
2134 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2135 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2136 obd->obd_name, label ?: str, lmi ? "on " : "",
2137 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2138 obd->obd_replayable ? "enabled" : "disabled");
2145 fsfilt_put_ops(obd->obd_fsops);
2146 filter_iobuf_pool_done(filter);
2148 server_put_mount(obd->obd_name, mnt);
2149 obd->u.obt.obt_sb = 0;
2153 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2155 struct lprocfs_static_vars lvars;
2156 cfs_proc_dir_entry_t *entry;
2162 CLASSERT(offsetof(struct obd_device, u.obt) ==
2163 offsetof(struct obd_device, u.filter.fo_obt));
2165 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2168 /* lprocfs must be setup before the filter so state can be safely added
2169 * to /proc incrementally as the filter is setup */
2170 lprocfs_filter_init_vars(&lvars);
2171 rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2173 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2178 rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2180 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2182 GOTO(obd_cleanup, rc);
2185 /* Init obdfilter private stats here */
2186 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2187 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2188 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2189 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2190 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2191 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2192 "get_page", "usec");
2193 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2194 LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2195 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2196 LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2197 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2198 LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2199 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2200 LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2202 rc = lproc_filter_attach_seqstat(obd);
2204 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2205 GOTO(free_obd_stats, rc);
2208 entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2209 if (IS_ERR(entry)) {
2210 rc = PTR_ERR(entry);
2211 CERROR("%s: error %d setting up lprocfs for %s\n",
2212 obd->obd_name, rc, "exports");
2213 GOTO(free_obd_stats, rc);
2215 obd->obd_proc_exports_entry = entry;
2217 entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2218 lprocfs_nid_stats_clear_read,
2219 lprocfs_nid_stats_clear_write, obd, NULL);
2220 if (IS_ERR(entry)) {
2221 rc = PTR_ERR(entry);
2222 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2224 GOTO(free_obd_stats, rc);
2227 rc = lprocfs_job_stats_init(obd, LPROC_FILTER_STATS_LAST,
2228 filter_stats_counter_init);
2230 GOTO(remove_entry_clear, rc);
2232 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2233 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2235 GOTO(job_stats_fini, rc = -ENOMEM);
2236 addr = (unsigned long)cfs_page_address(page);
2237 clear_page((void *)addr);
2238 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2239 LUSTRE_CFG_BUFLEN(lcfg, 4));
2240 rc = filter_common_setup(obd, lcfg, (void *)addr);
2241 OBD_PAGE_FREE(page);
2243 CERROR("%s: filter_common_setup failed: %d.\n",
2245 GOTO(job_stats_fini, rc);
2251 lprocfs_job_stats_fini(obd);
2253 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2255 lprocfs_free_obd_stats(obd);
2257 lprocfs_obd_cleanup(obd);
2261 static struct llog_operations filter_mds_ost_repl_logops;
2263 static struct llog_operations filter_size_orig_logops = {
2264 .lop_setup = llog_obd_origin_setup,
2265 .lop_cleanup = llog_obd_origin_cleanup,
2266 .lop_add = llog_obd_origin_add
2269 static int filter_olg_fini(struct obd_llog_group *olg)
2271 struct llog_ctxt *ctxt;
2272 int rc = 0, rc2 = 0;
2275 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2277 rc = llog_cleanup(ctxt);
2279 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2281 rc2 = llog_cleanup(ctxt);
2286 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2288 rc2 = llog_cleanup(ctxt);
2297 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2298 struct obd_device *tgt)
2303 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2304 &filter_mds_ost_repl_logops);
2308 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2309 &filter_size_orig_logops);
2315 filter_olg_fini(olg);
2320 * Init the default olg, which is embeded in the obd_device, for filter.
2323 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2324 struct obd_device *tgt)
2326 struct filter_obd *filter = &obd->u.filter;
2327 struct llog_ctxt *ctxt;
2331 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2332 if (!filter->fo_lcm)
2335 filter_mds_ost_repl_logops = llog_client_ops;
2336 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2337 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2338 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2340 rc = filter_olg_init(obd, olg, tgt);
2342 GOTO(cleanup_lcm, rc);
2344 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2347 GOTO(cleanup_olg, rc);
2349 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2351 CERROR("Can't get ctxt for %p:%x\n", olg,
2352 LLOG_MDS_OST_REPL_CTXT);
2353 GOTO(cleanup_olg, rc = -ENODEV);
2355 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2356 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2357 llog_ctxt_put(ctxt);
2361 filter_olg_fini(olg);
2363 llog_recov_thread_fini(filter->fo_lcm, 1);
2364 filter->fo_lcm = NULL;
2369 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2370 struct obd_device *tgt, int *index)
2372 struct filter_obd *filter = &obd->u.filter;
2373 struct llog_ctxt *ctxt;
2377 LASSERT(olg != NULL);
2378 if (olg == &obd->obd_olg)
2379 return filter_default_olg_init(obd, olg, tgt);
2381 LASSERT(filter->fo_lcm != NULL);
2382 rc = filter_olg_init(obd, olg, tgt);
2385 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2387 CERROR("Can't get ctxt for %p:%x\n", olg,
2388 LLOG_MDS_OST_REPL_CTXT);
2389 filter_olg_fini(olg);
2392 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2393 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2394 llog_ctxt_put(ctxt);
2398 static int filter_llog_finish(struct obd_device *obd, int count)
2400 struct filter_obd *filter = &obd->u.filter;
2401 struct llog_ctxt *ctxt;
2404 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2407 * Make sure that no cached llcds left in recov_thread.
2408 * We actually do sync in disconnect time, but disconnect
2409 * may not come being marked rq_no_resend = 1.
2411 llog_sync(ctxt, NULL, OBD_LLOG_FL_EXIT);
2414 * Balance class_import_get() in llog_receptor_accept().
2415 * This is safe to do, as llog is already synchronized
2416 * and its import may go.
2418 cfs_mutex_lock(&ctxt->loc_mutex);
2419 if (ctxt->loc_imp) {
2420 class_import_put(ctxt->loc_imp);
2421 ctxt->loc_imp = NULL;
2424 if (filter->fo_lcm) {
2425 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2426 filter->fo_lcm = NULL;
2429 cfs_mutex_unlock(&ctxt->loc_mutex);
2430 llog_ctxt_put(ctxt);
2433 RETURN(filter_olg_fini(&obd->obd_olg));
2437 * Find the group llog according to group index in the llog group list.
2439 static struct obd_llog_group *
2440 filter_find_olg_internal(struct filter_obd *filter, int group)
2442 struct obd_llog_group *olg;
2444 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2445 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2446 if (olg->olg_seq == group)
2453 * Find the group llog according to group index on the filter
2455 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2457 struct obd_llog_group *olg = NULL;
2458 struct filter_obd *filter;
2460 filter = &obd->u.filter;
2462 if (group == FID_SEQ_LLOG)
2463 RETURN(&obd->obd_olg);
2465 cfs_spin_lock(&filter->fo_llog_list_lock);
2466 olg = filter_find_olg_internal(filter, group);
2467 cfs_spin_unlock(&filter->fo_llog_list_lock);
2472 * Find the llog_group of the filter according to the group. If it can not
2473 * find, create the llog_group, which only happens when mds is being synced
2476 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2478 struct obd_llog_group *olg = NULL, *olg_new = NULL;
2479 struct filter_obd *filter;
2482 filter = &obd->u.filter;
2484 if (group == FID_SEQ_LLOG)
2485 RETURN(&obd->obd_olg);
2487 OBD_ALLOC_PTR(olg_new);
2488 if (olg_new == NULL)
2489 RETURN(ERR_PTR(-ENOMEM));
2491 cfs_spin_lock(&filter->fo_llog_list_lock);
2492 olg = filter_find_olg_internal(filter, group);
2494 if (olg->olg_initializing) {
2495 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2497 GOTO(out_unlock, olg);
2500 /* set as the newly allocated one */
2505 llog_group_init(olg, group);
2506 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2507 olg->olg_initializing = 1;
2508 cfs_spin_unlock(&filter->fo_llog_list_lock);
2510 rc = obd_llog_init(obd, olg, obd, NULL);
2512 cfs_spin_lock(&filter->fo_llog_list_lock);
2513 cfs_list_del(&olg->olg_list);
2514 cfs_spin_unlock(&filter->fo_llog_list_lock);
2516 GOTO(out, olg = ERR_PTR(-ENOMEM));
2518 cfs_spin_lock(&filter->fo_llog_list_lock);
2519 olg->olg_initializing = 0;
2520 cfs_spin_unlock(&filter->fo_llog_list_lock);
2521 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2522 obd->obd_name, group, olg);
2527 cfs_spin_unlock(&filter->fo_llog_list_lock);
2529 OBD_FREE_PTR(olg_new);
2533 static int filter_llog_connect(struct obd_export *exp,
2534 struct llogd_conn_body *body)
2536 struct obd_device *obd = exp->exp_obd;
2537 struct llog_ctxt *ctxt;
2538 struct obd_llog_group *olg;
2542 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2543 obd->obd_name, body->lgdc_logid.lgl_oid,
2544 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2546 olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2548 CERROR(" %s: can not find olg of group %d\n",
2549 obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2552 llog_group_set_export(olg, exp);
2554 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2555 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2556 body->lgdc_ctxt_idx);
2558 CDEBUG(D_HA, "%s: Recovery from log "LPX64"/"LPX64":%x\n",
2559 obd->obd_name, body->lgdc_logid.lgl_oid,
2560 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2562 cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2563 obd->u.filter.fo_mds_ost_sync = 1;
2564 cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2565 rc = llog_connect(ctxt, &body->lgdc_logid,
2566 &body->lgdc_gen, NULL);
2567 llog_ctxt_put(ctxt);
2569 CERROR("failed to connect rc %d idx %d\n", rc,
2570 body->lgdc_ctxt_idx);
2575 static int filter_llog_preclean(struct obd_device *obd)
2577 struct obd_llog_group *olg, *tmp;
2578 struct filter_obd *filter;
2579 cfs_list_t remove_list;
2583 rc = obd_llog_finish(obd, 0);
2585 CERROR("failed to cleanup llogging subsystem\n");
2587 filter = &obd->u.filter;
2588 CFS_INIT_LIST_HEAD(&remove_list);
2590 cfs_spin_lock(&filter->fo_llog_list_lock);
2591 while (!cfs_list_empty(&filter->fo_llog_list)) {
2592 olg = cfs_list_entry(filter->fo_llog_list.next,
2593 struct obd_llog_group, olg_list);
2594 cfs_list_del(&olg->olg_list);
2595 cfs_list_add(&olg->olg_list, &remove_list);
2597 cfs_spin_unlock(&filter->fo_llog_list_lock);
2599 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2600 cfs_list_del_init(&olg->olg_list);
2601 rc = filter_olg_fini(olg);
2603 CERROR("failed to cleanup llogging subsystem for %u\n",
2611 static int filter_precleanup(struct obd_device *obd,
2612 enum obd_cleanup_stage stage)
2618 case OBD_CLEANUP_EARLY:
2620 case OBD_CLEANUP_EXPORTS:
2621 /* Stop recovery before namespace cleanup. */
2622 target_recovery_fini(obd);
2624 obd_exports_barrier(obd);
2625 obd_zombie_barrier();
2627 rc = filter_llog_preclean(obd);
2628 lprocfs_job_stats_fini(obd);
2629 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2630 lprocfs_free_per_client_stats(obd);
2631 lprocfs_obd_cleanup(obd);
2632 lprocfs_free_obd_stats(obd);
2633 lquota_cleanup(filter_quota_interface_ref, obd);
2639 static int filter_cleanup(struct obd_device *obd)
2641 struct filter_obd *filter = &obd->u.filter;
2645 LCONSOLE_WARN("%s: shutting down for failover; client state "
2646 "will be preserved.\n", obd->obd_name);
2648 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2649 obd->obd_namespace = NULL;
2651 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2653 if (obd->u.obt.obt_sb == NULL)
2658 ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2659 shrink_dcache_sb(obd->u.obt.obt_sb);
2661 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2662 obd->u.obt.obt_sb = NULL;
2664 fsfilt_put_ops(obd->obd_fsops);
2666 filter_iobuf_pool_done(filter);
2668 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2673 static int filter_connect_internal(struct obd_export *exp,
2674 struct obd_connect_data *data,
2677 struct filter_export_data *fed = &exp->exp_filter_data;
2682 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2683 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2684 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2685 data->ocd_connect_flags, data->ocd_version,
2686 data->ocd_grant, data->ocd_index);
2688 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2689 CWARN("!!! This export (nid %s) used object group %d "
2690 "earlier; now it's trying to use group %d! This could "
2691 "be a bug in the MDS. Please report to "
2692 "http://bugs.whamcloud.com/\n",
2693 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2696 fed->fed_group = data->ocd_group;
2698 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2699 exp->exp_connect_flags = data->ocd_connect_flags;
2700 data->ocd_version = LUSTRE_VERSION_CODE;
2702 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2703 if (data->ocd_connect_flags & OBD_CONNECT_MDS)
2704 CDEBUG(D_HA, "%s: Received MDS connection for group %u\n",
2705 exp->exp_obd->obd_name, data->ocd_group);
2706 else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
2709 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2710 struct filter_obd *filter = &exp->exp_obd->u.filter;
2711 obd_size left, want;
2713 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2714 left = filter_grant_space_left(exp);
2715 want = data->ocd_grant;
2716 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2717 data->ocd_grant = fed->fed_grant;
2718 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2720 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2721 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2722 exp->exp_client_uuid.uuid, exp,
2723 data->ocd_grant, want, left);
2725 filter->fo_tot_granted_clients ++;
2728 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2729 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2730 int index = le32_to_cpu(lsd->lsd_ost_index);
2732 if (!(lsd->lsd_feature_compat &
2733 cpu_to_le32(OBD_COMPAT_OST))) {
2734 /* this will only happen on the first connect */
2735 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2736 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2737 /* sync is not needed here as filter_client_add will
2738 * set exp_need_sync flag */
2739 filter_update_server_data(exp->exp_obd);
2740 } else if (index != data->ocd_index) {
2741 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2742 " %u doesn't match actual OST index"
2743 " %u in last_rcvd file, bad "
2745 obd_export_nid2str(exp), index,
2749 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2750 * FIXME: We don't strictly need the COMPAT flag for that,
2751 * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2752 * FIXME: We needed it for the index, as index 0 is valid. */
2755 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2756 data->ocd_brw_size = 65536;
2757 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2758 data->ocd_brw_size = min(data->ocd_brw_size,
2759 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2760 if (data->ocd_brw_size == 0) {
2761 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2762 " ocd_version: %x ocd_grant: %d ocd_index: %u "
2763 "ocd_brw_size is unexpectedly zero, "
2764 "network data corruption?"
2765 "Refusing connection of this client\n",
2766 exp->exp_obd->obd_name,
2767 exp->exp_client_uuid.uuid,
2768 exp, data->ocd_connect_flags, data->ocd_version,
2769 data->ocd_grant, data->ocd_index);
2774 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2775 __u32 cksum_types = data->ocd_cksum_types;
2777 /* The client set in ocd_cksum_types the checksum types it
2778 * supports. We have to mask off the algorithms that we don't
2780 data->ocd_cksum_types &= cksum_types_supported_server();
2782 /* 1.6.4 clients are not supported any more */
2784 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2785 "%x\n", exp->exp_obd->obd_name,
2786 obd_export_nid2str(exp), cksum_types,
2787 data->ocd_cksum_types);
2789 /* This client does not support OBD_CONNECT_CKSUM
2790 * fall back to CRC32 */
2791 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2792 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2793 exp->exp_obd->obd_name,
2794 obd_export_nid2str(exp));
2797 if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2798 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2803 static int filter_reconnect(const struct lu_env *env,
2804 struct obd_export *exp, struct obd_device *obd,
2805 struct obd_uuid *cluuid,
2806 struct obd_connect_data *data,
2812 if (exp == NULL || obd == NULL || cluuid == NULL)
2815 rc = filter_connect_internal(exp, data, 1);
2817 filter_export_stats_init(obd, exp, localdata);
2822 static int filter_connect(const struct lu_env *env,
2823 struct obd_export **exp, struct obd_device *obd,
2824 struct obd_uuid *cluuid,
2825 struct obd_connect_data *data, void *localdata)
2827 struct lvfs_run_ctxt saved;
2828 struct lustre_handle conn = { 0 };
2829 struct obd_export *lexp;
2833 if (exp == NULL || obd == NULL || cluuid == NULL)
2836 rc = class_connect(&conn, obd, cluuid);
2839 lexp = class_conn2export(&conn);
2840 LASSERT(lexp != NULL);
2842 rc = filter_connect_internal(lexp, data, 0);
2846 filter_export_stats_init(obd, lexp, localdata);
2847 if (obd->obd_replayable) {
2848 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2850 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2851 rc = filter_client_add(obd, lexp, -1);
2856 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2857 rc = filter_read_groups(obd, data->ocd_group, 1);
2858 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2860 CERROR("can't read group %u\n", data->ocd_group);
2868 class_disconnect(lexp);
2877 /* Do extra sanity checks for grant accounting. We do this at connect,
2878 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2879 * always get rid of it or turn it off when we know accounting is good. */
2880 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2882 struct filter_export_data *fed;
2883 struct obd_export *exp;
2884 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2885 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2886 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2888 if (cfs_list_empty(&obd->obd_exports))
2891 /* We don't want to do this for large machines that do lots of
2892 mounts or unmounts. It burns... */
2893 if (obd->obd_num_exports > 100)
2896 cfs_spin_lock(&obd->obd_osfs_lock);
2897 cfs_spin_lock(&obd->obd_dev_lock);
2898 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2900 fed = &exp->exp_filter_data;
2901 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2904 if (maxsize > 0) { /* we may not have done a statfs yet */
2905 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2906 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2907 exp->exp_client_uuid.uuid, exp,
2908 fed->fed_grant, fed->fed_pending, maxsize);
2909 LASSERTF(fed->fed_dirty <= maxsize,
2910 "%s: cli %s/%p %ld > "LPU64"\n", func,
2911 exp->exp_client_uuid.uuid, exp,
2912 fed->fed_dirty, maxsize);
2915 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2916 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2917 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2919 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2920 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2921 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2922 tot_granted += fed->fed_grant + fed->fed_pending;
2923 tot_pending += fed->fed_pending;
2924 tot_dirty += fed->fed_dirty;
2926 fo_tot_granted = obd->u.filter.fo_tot_granted;
2927 fo_tot_pending = obd->u.filter.fo_tot_pending;
2928 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2929 cfs_spin_unlock(&obd->obd_dev_lock);
2930 cfs_spin_unlock(&obd->obd_osfs_lock);
2932 /* Do these assertions outside the spinlocks so we don't kill system */
2933 if (tot_granted != fo_tot_granted)
2934 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2935 func, tot_granted, fo_tot_granted);
2936 if (tot_pending != fo_tot_pending)
2937 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2938 func, tot_pending, fo_tot_pending);
2939 if (tot_dirty != fo_tot_dirty)
2940 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2941 func, tot_dirty, fo_tot_dirty);
2942 if (tot_pending > tot_granted)
2943 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2944 func, tot_pending, tot_granted);
2945 if (tot_granted > maxsize)
2946 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2947 func, tot_granted, maxsize);
2948 if (tot_dirty > maxsize)
2949 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2950 func, tot_dirty, maxsize);
2953 /* Remove this client from the grant accounting totals. We also remove
2954 * the export from the obd device under the osfs and dev locks to ensure
2955 * that the filter_grant_sanity_check() calculations are always valid.
2956 * The client should do something similar when it invalidates its import. */
2957 static void filter_grant_discard(struct obd_export *exp)
2959 struct obd_device *obd = exp->exp_obd;
2960 struct filter_obd *filter = &obd->u.filter;
2961 struct filter_export_data *fed = &exp->exp_filter_data;
2963 cfs_spin_lock(&obd->obd_osfs_lock);
2964 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2965 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2966 obd->obd_name, filter->fo_tot_granted,
2967 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2968 filter->fo_tot_granted -= fed->fed_grant;
2969 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2970 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2971 obd->obd_name, filter->fo_tot_pending,
2972 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2973 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2974 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2975 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2976 obd->obd_name, filter->fo_tot_dirty,
2977 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2978 filter->fo_tot_dirty -= fed->fed_dirty;
2982 cfs_spin_unlock(&obd->obd_osfs_lock);
2985 static int filter_destroy_export(struct obd_export *exp)
2987 struct filter_export_data *fed = &exp->exp_filter_data;
2990 if (fed->fed_pending)
2991 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2992 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2993 exp, fed->fed_pending);
2995 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2997 target_destroy_export(exp);
2999 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
3000 &exp->exp_client_uuid)))
3003 ldlm_destroy_export(exp);
3004 lut_client_free(exp);
3006 if (!exp->exp_obd->obd_replayable)
3007 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3009 filter_grant_discard(exp);
3010 filter_fmd_cleanup(exp);
3012 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3013 struct filter_obd *filter = &exp->exp_obd->u.filter;
3014 if (filter->fo_tot_granted_clients > 0)
3015 filter->fo_tot_granted_clients --;
3018 if (!(exp->exp_flags & OBD_OPT_FORCE))
3019 filter_grant_sanity_check(exp->exp_obd, __func__);
3024 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3026 struct obd_llog_group *olg_min, *olg;
3027 struct filter_obd *filter;
3028 int worked = -1, group;
3029 struct llog_ctxt *ctxt;
3032 filter = &obd->u.filter;
3034 /* we can't sync log holding spinlock. also, we do not want to get
3035 * into livelock. so we do following: loop over MDS's exports in
3036 * group order and skip already synced llogs -bzzz */
3038 /* look for group with min. number, but > worked */
3041 cfs_spin_lock(&filter->fo_llog_list_lock);
3042 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3043 if (olg->olg_seq <= worked) {
3044 /* this group is already synced */
3047 if (group < olg->olg_seq) {
3048 /* we have group with smaller number to sync */
3051 /* store current minimal group */
3053 group = olg->olg_seq;
3055 cfs_spin_unlock(&filter->fo_llog_list_lock);
3057 if (olg_min == NULL)
3060 worked = olg_min->olg_seq;
3061 if (olg_min->olg_exp &&
3062 (dexp == olg_min->olg_exp || dexp == NULL)) {
3064 ctxt = llog_group_get_ctxt(olg_min,
3065 LLOG_MDS_OST_REPL_CTXT);
3067 err = llog_sync(ctxt, olg_min->olg_exp, 0);
3068 llog_ctxt_put(ctxt);
3070 CERROR("error flushing logs to MDS: "
3075 } while (olg_min != NULL);
3078 /* Also incredibly similar to mds_disconnect */
3079 static int filter_disconnect(struct obd_export *exp)
3081 struct obd_device *obd = exp->exp_obd;
3086 class_export_get(exp);
3088 if (!(exp->exp_flags & OBD_OPT_FORCE))
3089 filter_grant_sanity_check(obd, __func__);
3090 filter_grant_discard(exp);
3092 /* Flush any remaining cancel messages out to the target */
3093 filter_sync_llogs(obd, exp);
3095 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3097 rc = server_disconnect_export(exp);
3099 /* Do not erase record for recoverable client. */
3100 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3101 filter_client_del(exp);
3103 fsfilt_sync(obd, obd->u.obt.obt_sb);
3105 class_export_put(exp);
3109 /* reverse import is changed, sync all cancels */
3110 static void filter_revimp_update(struct obd_export *exp)
3115 class_export_get(exp);
3117 /* flush any remaining cancel messages out to the target */
3118 filter_sync_llogs(exp->exp_obd, exp);
3119 class_export_put(exp);
3123 static int filter_ping(const struct lu_env *env, struct obd_export *exp)
3125 filter_fmd_expire(exp);
3129 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3130 const char *what, int quiet)
3132 struct dentry *dchild = NULL;
3134 dchild = filter_fid2dentry(obd, NULL, ostid->oi_seq, ostid->oi_id);
3136 if (IS_ERR(dchild)) {
3137 CERROR("%s error looking up object: "POSTID"\n",
3138 what, ostid->oi_id, ostid->oi_seq);
3142 if (dchild->d_inode == NULL) {
3144 CERROR("%s: %s on non-existent object: "POSTID" \n",
3145 obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3147 RETURN(ERR_PTR(-ENOENT));
3150 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0)
3151 /* Try to correct for a bug in 2.1.0 (LU-221) that caused negative
3152 * timestamps to appear to be in the far future, due old timestamp
3153 * being stored on disk as an unsigned value. This fixes up any
3154 * bad values stored on disk before returning them to the client,
3155 * and ensures any timestamp updates are correct. LU-1042 */
3156 if (unlikely(LTIME_S(dchild->d_inode->i_atime) == LU221_BAD_TIME))
3157 LTIME_S(dchild->d_inode->i_atime) = 0;
3158 if (unlikely(LTIME_S(dchild->d_inode->i_mtime) == LU221_BAD_TIME))
3159 LTIME_S(dchild->d_inode->i_mtime) = 0;
3160 if (unlikely(LTIME_S(dchild->d_inode->i_ctime) == LU221_BAD_TIME))
3161 LTIME_S(dchild->d_inode->i_ctime) = 0;
3163 #warning "remove old LU-221/LU-1042 workaround code"
3169 static int filter_getattr(const struct lu_env *env, struct obd_export *exp,
3170 struct obd_info *oinfo)
3172 struct dentry *dentry = NULL;
3173 struct obd_device *obd;
3178 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3179 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3183 obd = class_exp2obd(exp);
3185 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3189 dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3191 RETURN(PTR_ERR(dentry));
3193 /* Limit the valid bits in the return data to what we actually use */
3194 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3195 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3197 /* Store inode version in reply */
3198 curr_version = fsfilt_get_version(exp->exp_obd, dentry->d_inode);
3199 if ((__s64)curr_version != -EOPNOTSUPP) {
3200 oinfo->oi_oa->o_valid |= OBD_MD_FLDATAVERSION;
3201 oinfo->oi_oa->o_data_version = curr_version;
3208 /* this should be enabled/disabled in condition to enabled/disabled large
3209 * inodes (fast EAs) in backing store FS. */
3210 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3211 void *handle, struct obdo *oa)
3213 struct obd_device *obd = exp->exp_obd;
3217 if (oa->o_valid & OBD_MD_FLFID) {
3218 struct filter_fid ff;
3220 if (!(oa->o_valid & OBD_MD_FLGROUP))
3222 /* packing fid and converting it to LE for storing into EA.
3223 * Here ->o_stripe_idx should be filled by LOV and rest of
3224 * fields - by client. */
3225 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3226 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3227 /* XXX: we are ignoring o_parent_ver here, since this should
3228 * be the same for all objects in this fileset. */
3229 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3230 ff.ff_objid = cpu_to_le64(oa->o_id);
3231 ff.ff_seq = cpu_to_le64(oa->o_seq);
3233 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3234 LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3237 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3239 CERROR("store fid in object failed! rc: %d\n", rc);
3241 CDEBUG(D_HA, "OSS object without fid info!\n");
3247 /* this is called from filter_truncate() until we have filter_punch() */
3248 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3249 struct obdo *oa, struct obd_trans_info *oti)
3251 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3252 struct llog_cookie *fcc = NULL;
3253 struct filter_obd *filter;
3254 int rc, err, sync = 0;
3255 loff_t old_size = 0;
3256 unsigned int ia_valid;
3257 struct inode *inode;
3258 struct page *page = NULL;
3263 LASSERT(dentry != NULL);
3264 LASSERT(!IS_ERR(dentry));
3266 inode = dentry->d_inode;
3267 LASSERT(inode != NULL);
3269 filter = &exp->exp_obd->u.filter;
3270 iattr_from_obdo(&iattr, oa, oa->o_valid);
3271 ia_valid = iattr.ia_valid;
3273 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3274 OBD_ALLOC(fcc, sizeof(*fcc));
3276 *fcc = oa->o_lcookie;
3278 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3279 unsigned long now = jiffies;
3280 /* Filter truncates and writes are serialized by
3281 * i_alloc_sem, see the comment in
3282 * filter_preprw_write.*/
3283 if (ia_valid & ATTR_SIZE)
3284 down_write(&inode->i_alloc_sem);
3285 LOCK_INODE_MUTEX(inode);
3286 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
3287 old_size = i_size_read(inode);
3290 /* VBR: version recovery check */
3291 rc = filter_version_get_check(exp, oti, inode);
3293 GOTO(out_unlock, rc);
3295 /* Let's pin the last page so that ldiskfs_truncate
3296 * should not start GFP_FS allocation. */
3297 if (ia_valid & ATTR_SIZE) {
3298 page = grab_cache_page(inode->i_mapping,
3299 iattr.ia_size >> PAGE_CACHE_SHIFT);
3301 GOTO(out_unlock, rc = -ENOMEM);
3306 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3307 * then we will accept the UID+GID sent by the client during write for
3308 * initializing the ownership of this inode. We only allow this to
3309 * happen once so clear these bits in setattr. In 2.6 kernels it is
3310 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3311 * the flags that are actually being set. */
3312 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3313 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3314 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3316 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3317 if (!(ia_valid & ATTR_MODE)) {
3318 iattr.ia_mode = inode->i_mode;
3319 iattr.ia_valid |= ATTR_MODE;
3321 iattr.ia_mode &= ~S_ISUID;
3323 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3324 if (!(iattr.ia_valid & ATTR_MODE)) {
3325 iattr.ia_mode = inode->i_mode;
3326 iattr.ia_valid |= ATTR_MODE;
3328 iattr.ia_mode &= ~S_ISGID;
3331 orig_ids[USRQUOTA] = inode->i_uid;
3332 orig_ids[GRPQUOTA] = inode->i_gid;
3333 handle = fsfilt_start_log(exp->exp_obd, inode,
3334 FSFILT_OP_SETATTR, oti, 1);
3336 GOTO(out_unlock, rc = PTR_ERR(handle));
3338 /* update inode EA only once when inode is suid bit marked. As
3339 * on 2.6.x UID and GID may be set separately, we check here
3340 * only one of them to avoid double setting. */
3341 if (inode->i_mode & S_ISUID)
3342 filter_update_fidea(exp, inode, handle, oa);
3344 handle = fsfilt_start(exp->exp_obd, inode,
3345 FSFILT_OP_SETATTR, oti);
3347 GOTO(out_unlock, rc = PTR_ERR(handle));
3350 /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
3351 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3352 ll_vfs_dq_init(inode);
3354 if (oa->o_valid & OBD_MD_FLFLAGS) {
3355 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3356 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3358 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3360 /* set cancel cookie callback function */
3361 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3362 filter_cancel_cookies_cb,
3366 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3367 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3369 /* The truncate might have used up our transaction credits. Make sure
3370 * we have two left for the last_rcvd and VBR inode version updates. */
3371 err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3373 /* Update inode version only if data has changed => size has changed */
3374 rc = filter_finish_transno(exp, ia_valid & ATTR_SIZE ? inode : NULL,
3378 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3382 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3384 CERROR("error on commit, err = %d\n", err);
3391 /* For a partial-page truncate flush the page to disk immediately
3392 * to avoid data corruption during direct disk write. b=17397 */
3393 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3394 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3395 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3405 page_cache_release(page);
3407 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3408 UNLOCK_INODE_MUTEX(inode);
3409 if (ia_valid & ATTR_SIZE)
3410 up_write(&inode->i_alloc_sem);
3412 OBD_FREE(fcc, sizeof(*fcc));
3414 /* trigger quota release */
3415 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3416 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3417 int rc2 = lquota_adjust(filter_quota_interface_ref,
3418 exp->exp_obd, cur_ids,
3419 orig_ids, rc, FSFILT_OP_SETATTR);
3420 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3421 "filter adjust qunit. (rc:%d)\n", rc2);
3426 /* this is called from filter_truncate() until we have filter_punch() */
3427 int filter_setattr(const struct lu_env *env, struct obd_export *exp,
3428 struct obd_info *oinfo, struct obd_trans_info *oti)
3430 struct obdo *oa = oinfo->oi_oa;
3431 struct lustre_capa *capa = oinfo_capa(oinfo);
3432 struct ldlm_res_id res_id;
3433 struct filter_mod_data *fmd;
3434 struct lvfs_run_ctxt saved;
3435 struct filter_obd *filter;
3436 struct ldlm_resource *res;
3437 struct dentry *dentry;
3438 __u64 opc = CAPA_OPC_META_WRITE;
3442 if (oinfo->oi_flags & OBD_FL_PUNCH)
3443 opc |= CAPA_OPC_OSS_TRUNC;
3445 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3449 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3450 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3455 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3456 /* This would be very bad - accidentally truncating a file when
3457 * changing the time or similar - bug 12203. */
3458 if (oa->o_valid & OBD_MD_FLSIZE &&
3459 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3460 static char mdsinum[48];
3462 if (oa->o_valid & OBD_MD_FLFID)
3463 snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3464 oa->o_parent_seq, oa->o_parent_oid,
3469 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3470 "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3471 oa->o_id, oa->o_seq, mdsinum);
3475 dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3477 RETURN(PTR_ERR(dentry));
3479 filter = &exp->exp_obd->u.filter;
3480 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3483 * We need to be atomic against a concurrent write
3484 * (which takes the semaphore for reading). fmd_mactime_xid
3485 * checks will have no effect if a write request with lower
3486 * xid starts just before a setattr and finishes later than
3487 * the setattr (see bug 21489, comment 27).
3490 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3491 unsigned long now = jiffies;
3492 down_write(&dentry->d_inode->i_alloc_sem);
3493 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem");
3494 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3495 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3496 fmd->fmd_mactime_xid = oti->oti_xid;
3497 filter_fmd_put(exp, fmd);
3498 up_write(&dentry->d_inode->i_alloc_sem);
3501 /* setting objects attributes (including owner/group) */
3502 rc = filter_setattr_internal(exp, dentry, oa, oti);
3504 GOTO(out_unlock, rc);
3506 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3507 &res_id, LDLM_EXTENT, 0);
3510 LDLM_RESOURCE_ADDREF(res);
3511 rc = ldlm_res_lvbo_update(res, NULL, 0);
3512 LDLM_RESOURCE_DELREF(res);
3513 ldlm_resource_putref(res);
3516 oa->o_valid = OBD_MD_FLID;
3518 /* Quota release need uid/gid info */
3519 obdo_from_inode(oa, dentry->d_inode,
3520 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3522 filter_counter_incr(exp, LPROC_FILTER_STATS_SETATTR,
3523 oti ? oti->oti_jobid : NULL, 1);
3527 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3531 /* XXX identical to osc_unpackmd */
3532 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3533 struct lov_mds_md *lmm, int lmm_bytes)
3539 if (lmm_bytes < sizeof (*lmm)) {
3540 CERROR("lov_mds_md too small: %d, need %d\n",
3541 lmm_bytes, (int)sizeof(*lmm));
3544 /* XXX LOV_MAGIC etc check? */
3546 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3547 CERROR("lov_mds_md: zero lmm_object_id\n");
3552 lsm_size = lov_stripe_md_size(1);
3556 if (*lsmp != NULL && lmm == NULL) {
3557 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3558 OBD_FREE(*lsmp, lsm_size);
3563 if (*lsmp == NULL) {
3564 OBD_ALLOC(*lsmp, lsm_size);
3568 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3569 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3570 OBD_FREE(*lsmp, lsm_size);
3573 loi_init((*lsmp)->lsm_oinfo[0]);
3577 /* XXX zero *lsmp? */
3578 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3579 LASSERT((*lsmp)->lsm_object_id);
3582 (*lsmp)->lsm_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
3587 /* caller must hold fo_create_locks[oa->o_seq] */
3588 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3589 struct filter_obd *filter)
3591 struct obdo doa = { 0 }; /* XXX obdo on stack */
3597 LASSERT_MUTEX_LOCKED(&filter->fo_create_locks[oa->o_seq]);
3599 memset(&doa, 0, sizeof(doa));
3601 doa.o_valid |= OBD_MD_FLGROUP;
3602 doa.o_seq = oa->o_seq;
3603 doa.o_mode = S_IFREG;
3605 if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3606 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3607 exp->exp_obd->obd_name, doa.o_seq);
3611 last = filter_last_id(filter, doa.o_seq);
3613 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3615 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3616 exp->exp_obd->obd_name, oa->o_id + 1, last,
3617 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3619 for (id = last; id > oa->o_id; id--) {
3621 rc = filter_destroy(NULL, exp, &doa, NULL, NULL, NULL, NULL);
3622 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3623 CEMERG("error destroying precreate objid "LPU64": %d\n",
3626 /* update last_id on disk periodically so that if we restart
3627 * we don't need to re-scan all of the just-deleted objects. */
3628 if ((id & 511) == 0 && !skip_orphan) {
3629 filter_set_last_id(filter, id - 1, doa.o_seq);
3630 filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3634 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3635 exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3638 filter_set_last_id(filter, id, doa.o_seq);
3639 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3642 * We have destroyed orphan objects, but don't want to reuse
3643 * them. Therefore we don't reset last_id to the last created
3644 * objects. Instead, we report back to the MDS the object id
3645 * of the last orphan, so that the MDS can restart allocating
3646 * objects from this id + 1 and thus skip the whole orphan
3652 cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3657 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3658 obd_seq group, int *num);
3659 /* returns a negative error or a nonnegative number of files to create */
3660 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3661 obd_seq group, struct obd_trans_info *oti)
3663 struct obd_device *obd = exp->exp_obd;
3664 struct filter_obd *filter = &obd->u.filter;
3668 /* delete orphans request */
3669 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3670 obd_id last = filter_last_id(filter, group);
3672 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3673 CERROR("%s: dropping old orphan cleanup request\n",
3677 /* This causes inflight precreates to abort and drop lock */
3678 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3679 cfs_mutex_lock(&filter->fo_create_locks[group]);
3680 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3681 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3682 exp->exp_obd->obd_name, group);
3683 cfs_mutex_unlock(&filter->fo_create_locks[group]);
3686 diff = oa->o_id - last;
3687 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3690 if (-diff > (OST_MAX_PRECREATE * 3) / 2) {
3691 CERROR("%s: ignoring bogus orphan destroy request: "
3692 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3694 /* FIXME: should reset precreate_next_id on MDS */
3695 GOTO(out, rc = -EINVAL);
3698 rc = filter_destroy_precreated(exp, oa, filter);
3700 CERROR("%s: unable to write lastobjid, but "
3701 "orphans were deleted\n", obd->obd_name);
3704 /* XXX: Used by MDS for the first time! */
3705 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3708 cfs_mutex_lock(&filter->fo_create_locks[group]);
3709 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3710 CERROR("%s: dropping old precreate request\n",
3714 /* only precreate if group == 0 and o_id is specfied */
3715 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3718 diff = oa->o_id - filter_last_id(filter, group);
3719 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3720 filter_last_id(filter, group), diff);
3722 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3723 oa->o_id, filter_last_id(filter, group), diff);
3727 oa->o_id = filter_last_id(&obd->u.filter, group);
3728 rc = filter_precreate(obd, oa, group, &diff);
3729 oa->o_id = filter_last_id(&obd->u.filter, group);
3731 oa->o_valid |= (OBD_MD_FLID | OBD_MD_FLGROUP);
3734 /* else diff == 0 */
3737 cfs_mutex_unlock(&filter->fo_create_locks[group]);
3741 static int filter_statfs(const struct lu_env *env, struct obd_export *exp,
3742 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
3744 struct obd_device *obd = class_exp2obd(exp);
3745 struct filter_obd *filter = &obd->u.filter;
3746 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3747 struct lr_server_data *lsd = class_server_data(obd);
3751 /* at least try to account for cached pages. its still racey and
3752 * might be under-reporting if clients haven't announced their
3753 * caches with brw recently */
3754 cfs_spin_lock(&obd->obd_osfs_lock);
3755 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3756 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3757 cfs_spin_unlock(&obd->obd_osfs_lock);
3759 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3760 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3761 filter->fo_tot_dirty, filter->fo_tot_granted,
3762 filter->fo_tot_pending,
3763 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3765 filter_grant_sanity_check(obd, __func__);
3767 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3768 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3769 osfs->os_bsize - 1) >> blockbits));
3771 if (OBD_FAIL_CHECK_VALUE(OBD_FAIL_OST_ENOSPC,
3772 le32_to_cpu(lsd->lsd_ost_index)))
3773 osfs->os_bfree = osfs->os_bavail = 2;
3775 if (OBD_FAIL_CHECK_VALUE(OBD_FAIL_OST_ENOINO,
3776 le32_to_cpu(lsd->lsd_ost_index)))
3779 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3780 * stop creating files on MDS if OST is not good shape to create
3784 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3785 osfs->os_state = OS_STATE_READONLY;
3787 if (filter->fo_raid_degraded)
3788 osfs->os_state |= OS_STATE_DEGRADED;
3792 static int filter_use_existing_obj(struct obd_device *obd,
3793 struct dentry *dchild, void **handle,
3796 struct inode *inode = dchild->d_inode;
3800 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3803 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3804 if (IS_ERR(*handle))
3805 return PTR_ERR(*handle);
3807 iattr.ia_valid = ATTR_MODE;
3808 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3809 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3816 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3819 __u64 os_ffree = -1;
3821 cfs_spin_lock(&obd->obd_osfs_lock);
3822 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3824 os_ffree = obd->obd_osfs.os_ffree;
3825 cfs_spin_unlock(&obd->obd_osfs_lock);
3830 /* We rely on the fact that only one thread will be creating files in a given
3831 * group at a time, which is why we don't need an atomic filter_get_new_id.
3832 * Even if we had that atomic function, the following race would exist:
3834 * thread 1: gets id x from filter_next_id
3835 * thread 2: gets id (x + 1) from filter_next_id
3836 * thread 2: creates object (x + 1)
3837 * thread 1: tries to create object x, gets -ENOSPC
3839 * Caller must hold fo_create_locks[group]
3841 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3842 obd_seq group, int *num)
3844 struct dentry *dchild = NULL, *dparent = NULL;
3845 struct filter_obd *filter;
3846 struct obd_statfs *osfs;
3848 int err = 0, rc = 0, recreate_obj = 0, i;
3849 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3852 void *handle = NULL;
3855 filter = &obd->u.filter;
3857 LASSERT_MUTEX_LOCKED(&filter->fo_create_locks[group]);
3859 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3861 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3862 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3865 OBD_ALLOC(osfs, sizeof(*osfs));
3868 rc = filter_statfs(NULL, obd->obd_self_export, osfs,
3869 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
3871 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3872 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3873 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3874 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3876 if (oa->o_valid & OBD_MD_FLFLAGS)
3877 oa->o_flags |= OBD_FL_NOSPC_BLK;
3879 oa->o_valid |= OBD_MD_FLFLAGS;
3880 oa->o_flags = OBD_FL_NOSPC_BLK;
3885 OBD_FREE(osfs, sizeof(*osfs));
3890 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3891 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3893 for (i = 0; i < *num && err == 0; i++) {
3894 int cleanup_phase = 0;
3896 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3897 CWARN("%s: create aborted by destroy\n",
3906 last_id = filter_last_id(filter, group);
3907 if (next_id > last_id) {
3908 CERROR("Error: Trying to recreate obj greater"
3909 "than last id "LPD64" > "LPD64"\n",
3911 GOTO(cleanup, rc = -EINVAL);
3914 next_id = filter_last_id(filter, group) + 1;
3916 /* Don't create objects beyond the valid range for this SEQ */
3917 if (unlikely(fid_seq_is_mdt0(group) &&
3918 next_id >= IDIF_MAX_OID)) {
3919 CERROR("%s:"POSTID" hit the IDIF_MAX_OID (1<<48)!\n",
3920 obd->obd_name, next_id, group);
3921 GOTO(cleanup, rc = -ENOSPC);
3922 } else if (unlikely(!fid_seq_is_mdt0(group) &&
3923 next_id >= OBIF_MAX_OID)) {
3924 CERROR("%s:"POSTID" hit the OBIF_MAX_OID (1<<32)!\n",
3925 obd->obd_name, next_id, group);
3926 GOTO(cleanup, rc = -ENOSPC);
3929 dparent = filter_parent_lock(obd, group, next_id);
3930 if (IS_ERR(dparent))
3931 GOTO(cleanup, rc = PTR_ERR(dparent));
3932 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3934 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3936 GOTO(cleanup, rc = PTR_ERR(dchild));
3937 cleanup_phase = 2; /* f_dput(dchild) */
3939 if (dchild->d_inode != NULL) {
3940 /* This would only happen if lastobjid was bad on disk*/
3941 /* Could also happen if recreating missing obj but it
3942 * already exists. */
3944 CERROR("%s: recreating existing object %.*s?\n",
3945 obd->obd_name, dchild->d_name.len,
3946 dchild->d_name.name);
3948 /* Use these existing objects if they are
3950 if (i_size_read(dchild->d_inode) == 0) {
3951 rc = filter_use_existing_obj(obd,dchild,
3952 &handle, &cleanup_phase);
3959 CERROR("%s: Serious error: objid %.*s already "
3960 "exists; is this filesystem corrupt?\n",
3961 obd->obd_name, dchild->d_name.len,
3962 dchild->d_name.name);
3965 GOTO(cleanup, rc = -EEXIST);
3968 handle = fsfilt_start_log(obd, dparent->d_inode,
3969 FSFILT_OP_CREATE, NULL, 1);
3971 GOTO(cleanup, rc = PTR_ERR(handle));
3974 CDEBUG(D_INODE, "%s: filter_precreate(od->o_seq="LPU64
3975 ",od->o_id="LPU64")\n", obd->obd_name, group,
3978 /* We mark object SUID+SGID to flag it for accepting UID+GID
3979 * from client on first write. Currently the permission bits
3980 * on the OST are never used, so this is OK. */
3981 rc = ll_vfs_create(dparent->d_inode, dchild,
3982 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3984 CWARN("%s: create failed: rc = %d\n", obd->obd_name,rc);
3985 if (rc == -ENOSPC) {
3986 os_ffree = filter_calc_free_inodes(obd);
3990 if (obd->obd_osfs.os_bavail <
3991 (obd->obd_osfs.os_blocks >> 10)) {
3992 if (oa->o_valid & OBD_MD_FLFLAGS) {
3993 oa->o_flags |= OBD_FL_NOSPC_BLK;
3995 oa->o_valid |= OBD_MD_FLFLAGS;
3996 oa->o_flags = OBD_FL_NOSPC_BLK;
3999 CWARN("%s: free inode "LPU64"\n",
4000 obd->obd_name, os_ffree);
4006 if (dchild->d_inode)
4007 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
4008 dchild->d_inode->i_ino);
4011 /* Initialize a/c/m time so any client timestamp will always
4012 * be newer and update the inode. ctime = 0 is also handled
4013 * specially in fsfilt_ext3_setattr(). See LU-221, LU-1042 */
4014 iattr.ia_valid = ATTR_ATIME | ATTR_MTIME | ATTR_CTIME;
4015 LTIME_S(iattr.ia_atime) = 0;
4016 LTIME_S(iattr.ia_mtime) = 0;
4017 LTIME_S(iattr.ia_ctime) = 0;
4018 err = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
4020 CWARN("%s: unable to initialize a/c/m time of newly "
4021 "created object %.*s: rc = %d\n",
4022 obd->obd_name, dchild->d_name.len,
4023 dchild->d_name.name, err);
4025 if (!recreate_obj) {
4026 filter_set_last_id(filter, next_id, group);
4027 err = filter_update_last_objid(obd, group, 0);
4029 CERROR("%s: unable to write lastobjid "
4030 "but file created: rc = %d\n",
4031 obd->obd_name, err);
4035 switch(cleanup_phase) {
4037 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4039 CERROR("error on commit, err = %d\n", err);
4046 filter_parent_unlock(dparent);
4053 if (cfs_time_after(jiffies, enough_time)) {
4056 "%s: precreate slow - want %d got %d \n",
4057 obd->obd_name, *num, i);
4064 "%s: created %d objects for group "POSTID" rc %d\n",
4065 obd->obd_name, i, filter->fo_last_objids[group], group, rc);
4070 int filter_create(const struct lu_env *env, struct obd_export *exp,
4071 struct obdo *oa, struct lov_stripe_md **ea,
4072 struct obd_trans_info *oti)
4074 struct obd_device *obd = exp->exp_obd;
4075 struct filter_export_data *fed;
4076 struct filter_obd *filter;
4077 struct lvfs_run_ctxt saved;
4078 struct lov_stripe_md *lsm = NULL;
4082 CDEBUG(D_INODE, "%s: filter_create(group="LPU64",id="
4083 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4085 fed = &exp->exp_filter_data;
4086 filter = &obd->u.filter;
4088 /* 1.8 client doesn't carry the ocd_group with connect request,
4089 * so the fed_group will always be zero for 1.8 client. */
4090 if (!(exp->exp_connect_flags & OBD_CONNECT_FULL20)) {
4091 if (oa->o_seq != FID_SEQ_OST_MDT0 &&
4092 oa->o_seq != FID_SEQ_LLOG &&
4093 oa->o_seq != FID_SEQ_ECHO) {
4094 CERROR("The request from older client has invalid"
4095 " group "LPU64"!\n", oa->o_seq);
4098 } else if (fed->fed_group != oa->o_seq) {
4099 CERROR("%s: this export (nid %s) used object group %d "
4100 "earlier; now it's trying to use group "LPU64"!"
4101 " This could be a bug in the MDS. Please report to "
4102 "http://bugzilla.lustre.org/\n", obd->obd_name,
4103 obd_export_nid2str(exp), fed->fed_group, oa->o_seq);
4110 rc = obd_alloc_memmd(exp, &lsm);
4117 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4119 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4120 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4121 if (!obd->obd_recovering ||
4122 oa->o_id > filter_last_id(filter, oa->o_seq)) {
4123 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4124 oa->o_id, filter_last_id(filter, oa->o_seq));
4128 cfs_mutex_lock(&filter->fo_create_locks[oa->o_seq]);
4129 rc = filter_precreate(obd, oa, oa->o_seq, &diff);
4130 cfs_mutex_unlock(&filter->fo_create_locks[oa->o_seq]);
4133 rc = filter_handle_precreate(exp, oa, oa->o_seq, oti);
4136 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4137 if (rc && ea != NULL && *ea != lsm) {
4138 obd_free_memmd(exp, &lsm);
4139 } else if (rc == 0 && ea != NULL) {
4140 /* XXX LOV STACKING: the lsm that is passed to us from
4141 * LOV does not have valid lsm_oinfo data structs, so
4142 * don't go touching that. This needs to be fixed in a
4144 lsm->lsm_object_id = oa->o_id;
4151 int filter_destroy(const struct lu_env *env, struct obd_export *exp,
4152 struct obdo *oa, struct lov_stripe_md *md,
4153 struct obd_trans_info *oti, struct obd_export *md_exp,
4156 unsigned int qcids[MAXQUOTAS] = {0, 0};
4157 struct obd_device *obd;
4158 struct filter_obd *filter;
4159 struct dentry *dchild = NULL, *dparent = NULL;
4160 struct lustre_handle lockh = { 0 };
4161 struct lvfs_run_ctxt saved;
4162 void *handle = NULL;
4163 struct llog_cookie *fcc = NULL;
4164 int rc, rc2, cleanup_phase = 0, sync = 0;
4169 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4170 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4175 filter = &obd->u.filter;
4177 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4180 CDEBUG(D_INODE, "%s: filter_destroy(group="LPU64",oid="
4181 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4183 dchild = filter_fid2dentry(obd, NULL, oa->o_seq, oa->o_id);
4185 GOTO(cleanup, rc = PTR_ERR(dchild));
4188 if (dchild->d_inode == NULL) {
4189 CDEBUG(D_INODE, "destroying non-existent object "POSTID"\n",
4190 oa->o_id, oa->o_seq);
4191 /* If object already gone, cancel cookie right now */
4192 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4193 struct llog_ctxt *ctxt;
4194 struct obd_llog_group *olg;
4196 olg = filter_find_olg(obd, oa->o_seq);
4198 CERROR(" %s: can not find olg of group %d\n",
4199 obd->obd_name, (int)oa->o_seq);
4200 GOTO(cleanup, rc = PTR_ERR(olg));
4202 fcc = &oa->o_lcookie;
4203 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4204 llog_cancel(ctxt, NULL, 1, fcc, 0);
4205 llog_ctxt_put(ctxt);
4206 fcc = NULL; /* we didn't allocate fcc, don't free it */
4208 GOTO(cleanup, rc = -ENOENT);
4211 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_seq, &lockh);
4215 /* Our MDC connection is established by the MDS to us */
4216 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4217 OBD_ALLOC(fcc, sizeof(*fcc));
4219 *fcc = oa->o_lcookie;
4222 /* we're gonna truncate it first in order to avoid possible deadlock:
4224 * open trasaction open transaction
4225 * down(i_zombie) down(i_zombie)
4226 * restart transaction
4227 * (see BUG 4180) -bzzz
4229 * take i_alloc_sem too to prevent other threads from writing to the
4230 * file while we are truncating it. This can cause lock ordering issue
4231 * between page lock, i_mutex & starting new journal handle.
4232 * (see bug 20321) -johann
4235 down_write(&dchild->d_inode->i_alloc_sem);
4236 LOCK_INODE_MUTEX(dchild->d_inode);
4237 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
4239 /* VBR: version recovery check */
4240 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4242 UNLOCK_INODE_MUTEX(dchild->d_inode);
4243 up_write(&dchild->d_inode->i_alloc_sem);
4247 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4249 if (IS_ERR(handle)) {
4250 UNLOCK_INODE_MUTEX(dchild->d_inode);
4251 up_write(&dchild->d_inode->i_alloc_sem);
4252 GOTO(cleanup, rc = PTR_ERR(handle));
4255 /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
4256 ll_vfs_dq_init(dchild->d_inode);
4258 iattr.ia_valid = ATTR_SIZE;
4260 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4261 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4262 UNLOCK_INODE_MUTEX(dchild->d_inode);
4263 up_write(&dchild->d_inode->i_alloc_sem);
4267 GOTO(cleanup, rc = rc2);
4269 /* We don't actually need to lock the parent until we are unlinking
4270 * here, and not while truncating above. That avoids holding the
4271 * parent lock for a long time during truncate, which can block other
4272 * threads from doing anything to objects in that directory. bug 7171 */
4273 dparent = filter_parent_lock(obd, oa->o_seq, oa->o_id);
4274 if (IS_ERR(dparent))
4275 GOTO(cleanup, rc = PTR_ERR(dparent));
4276 cleanup_phase = 3; /* filter_parent_unlock */
4278 LOCK_INODE_MUTEX(dchild->d_inode);
4279 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4280 if (IS_ERR(handle)) {
4281 UNLOCK_INODE_MUTEX(dchild->d_inode);
4282 GOTO(cleanup, rc = PTR_ERR(handle));
4284 cleanup_phase = 4; /* fsfilt_commit */
4286 /* Quota release need uid/gid of inode */
4287 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID | OBD_MD_FLGID);
4289 filter_fmd_drop(exp, oa->o_id, oa->o_seq);
4291 /* this drops dchild->d_inode->i_mutex unconditionally */
4292 rc = filter_destroy_internal(obd, oa->o_id, oa->o_seq, dparent, dchild);
4296 switch(cleanup_phase) {
4299 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4300 oti->oti_handle : handle,
4301 filter_cancel_cookies_cb,
4303 /* If add_journal_cb failed, then filter_finish_transno
4304 * will commit the handle and we will do a sync
4305 * on commit. then we call callback directly to free
4308 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4310 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4313 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4315 CERROR("error on commit, err = %d\n", rc2);
4322 filter_parent_unlock(dparent);
4324 filter_fini_destroy(obd, &lockh);
4328 OBD_FREE(fcc, sizeof(*fcc));
4330 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4333 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4337 /* trigger quota release */
4338 qcids[USRQUOTA] = oa->o_uid;
4339 qcids[GRPQUOTA] = oa->o_gid;
4340 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4343 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4347 /* NB start and end are used for punch, but not truncate */
4348 static int filter_truncate(const struct lu_env *env, struct obd_export *exp,
4349 struct obd_info *oinfo, struct obd_trans_info *oti,
4350 struct ptlrpc_request_set *rqset)
4355 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4356 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4357 oinfo->oi_policy.l_extent.end);
4361 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4362 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,oinfo->oi_oa->o_valid,
4363 oinfo->oi_policy.l_extent.start);
4365 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4366 rc = filter_setattr(env, exp, oinfo, oti);
4370 static int filter_sync(const struct lu_env *env, struct obd_export *exp,
4371 struct obd_info *oinfo, obd_off start, obd_off end,
4372 struct ptlrpc_request_set *set)
4374 struct lvfs_run_ctxt saved;
4375 struct obd_device_target *obt;
4376 struct dentry *dentry;
4380 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
4381 (struct lustre_capa *)oinfo->oi_capa,
4382 CAPA_OPC_OSS_WRITE);
4386 obt = &exp->exp_obd->u.obt;
4388 /* An objid of zero is taken to mean "sync whole filesystem" */
4389 if (!oinfo->oi_oa || !(oinfo->oi_oa->o_valid & OBD_MD_FLID)) {
4390 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4391 /* Flush any remaining cancel messages out to the target */
4392 filter_sync_llogs(exp->exp_obd, exp);
4396 dentry = filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi);
4398 RETURN(PTR_ERR(dentry));
4400 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4402 LOCK_INODE_MUTEX(dentry->d_inode);
4404 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4406 /* just any file to grab fsync method - "file" arg unused */
4407 struct file *file = obt->obt_rcvd_filp;
4409 if (file->f_op && file->f_op->fsync)
4410 rc = file->f_op->fsync(NULL, dentry, 1);
4412 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4416 UNLOCK_INODE_MUTEX(dentry->d_inode);
4418 oinfo->oi_oa->o_valid = OBD_MD_FLID;
4419 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
4421 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4423 filter_counter_incr(exp, LPROC_FILTER_STATS_SYNC, oinfo->oi_jobid, 1);
4428 static int filter_get_info(const struct lu_env *env, struct obd_export *exp,
4429 __u32 keylen, void *key, __u32 *vallen, void *val,
4430 struct lov_stripe_md *lsm)
4432 struct obd_device *obd;
4435 obd = class_exp2obd(exp);
4437 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4441 if (KEY_IS(KEY_BLOCKSIZE)) {
4442 __u32 *blocksize = val;
4444 if (*vallen < sizeof(*blocksize))
4446 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4448 *vallen = sizeof(*blocksize);
4452 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4453 __u32 *blocksize_bits = val;
4454 if (blocksize_bits) {
4455 if (*vallen < sizeof(*blocksize_bits))
4457 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4459 *vallen = sizeof(*blocksize_bits);
4463 if (KEY_IS(KEY_LAST_ID)) {
4464 obd_id *last_id = val;
4465 /* FIXME: object groups */
4467 if (*vallen < sizeof(*last_id))
4469 *last_id = filter_last_id(&obd->u.filter,
4470 exp->exp_filter_data.fed_group);
4472 *vallen = sizeof(*last_id);
4476 if (KEY_IS(KEY_FIEMAP)) {
4477 struct ll_fiemap_info_key *fm_key = key;
4478 struct dentry *dentry;
4479 struct ll_user_fiemap *fiemap = val;
4480 struct lvfs_run_ctxt saved;
4483 if (fiemap == NULL) {
4484 *vallen = fiemap_count_to_size(
4485 fm_key->fiemap.fm_extent_count);
4489 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa.o_oi,
4492 RETURN(PTR_ERR(dentry));
4494 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4495 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4496 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4498 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4504 if (KEY_IS(KEY_SYNC_LOCK_CANCEL)) {
4505 *((__u32 *) val) = obd->u.filter.fo_sync_lock_cancel;
4506 *vallen = sizeof(__u32);
4510 CDEBUG(D_IOCTL, "invalid key\n");
4514 static inline int filter_setup_llog_group(struct obd_export *exp,
4515 struct obd_device *obd,
4518 struct obd_llog_group *olg;
4519 struct llog_ctxt *ctxt;
4522 olg = filter_find_create_olg(obd, group);
4524 RETURN(PTR_ERR(olg));
4526 llog_group_set_export(olg, exp);
4528 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4529 LASSERTF(ctxt != NULL, "ctxt is null\n");
4531 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4532 llog_ctxt_put(ctxt);
4536 static int filter_set_grant_shrink(struct obd_export *exp,
4537 struct ost_body *body)
4539 /* handle shrink grant */
4540 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4541 filter_grant_incoming(exp, &body->oa);
4542 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4548 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4550 struct obd_device *obd;
4556 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4560 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4561 obd_export_nid2str(exp));
4562 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4564 /* setup llog imports */
4566 group = (int)(*(__u32 *)val);
4568 group = 0; /* default value */
4570 LASSERT_SEQ_IS_MDT(group);
4571 rc = filter_setup_llog_group(exp, obd, group);
4575 if (group == FID_SEQ_OST_MDT0) {
4576 /* setup llog group 1 for interop */
4577 filter_setup_llog_group(exp, obd, FID_SEQ_LLOG);
4580 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4585 static int filter_set_info_async(const struct lu_env *env,
4586 struct obd_export *exp, __u32 keylen,
4587 void *key, __u32 vallen, void *val,
4588 struct ptlrpc_request_set *set)
4590 struct obd_device *obd;
4595 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4599 if (KEY_IS(KEY_CAPA_KEY)) {
4601 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4603 CERROR("filter update capability key failed: %d\n", rc);
4607 if (KEY_IS(KEY_REVIMP_UPD)) {
4608 filter_revimp_update(exp);
4609 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4613 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4614 filter_adapt_sptlrpc_conf(obd, 0);
4618 if (KEY_IS(KEY_MDS_CONN))
4619 RETURN(filter_set_mds_conn(exp, val));
4621 if (KEY_IS(KEY_GRANT_SHRINK))
4622 RETURN(filter_set_grant_shrink(exp, val));
4627 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4628 int len, void *karg, void *uarg)
4630 struct obd_device *obd = exp->exp_obd;
4631 struct obd_ioctl_data *data = karg;
4635 case OBD_IOC_ABORT_RECOVERY: {
4636 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4637 target_stop_recovery_thread(obd);
4641 case OBD_IOC_SYNC: {
4642 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4643 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4647 case OBD_IOC_SET_READONLY: {
4649 struct super_block *sb = obd->u.obt.obt_sb;
4650 struct inode *inode = sb->s_root->d_inode;
4651 BDEVNAME_DECLARE_STORAGE(tmp);
4652 CERROR("*** setting device %s read-only ***\n",
4653 ll_bdevname(sb, tmp));
4655 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4656 if (!IS_ERR(handle))
4657 rc = fsfilt_commit(obd, inode, handle, 1);
4659 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4660 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4662 rc = lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4666 case OBD_IOC_CATLOGLIST: {
4667 rc = llog_catalog_list(obd, 1, data);
4671 case OBD_IOC_LLOG_CANCEL:
4672 case OBD_IOC_LLOG_REMOVE:
4673 case OBD_IOC_LLOG_INFO:
4674 case OBD_IOC_LLOG_PRINT: {
4675 /* FIXME to be finished */
4676 RETURN(-EOPNOTSUPP);
4678 struct llog_ctxt *ctxt = NULL;
4680 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4681 rc = llog_ioctl(ctxt, cmd, data);
4682 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4695 static int filter_health_check(const struct lu_env *env, struct obd_device *obd)
4697 #ifdef USE_HEALTH_CHECK_WRITE
4698 struct filter_obd *filter = &obd->u.filter;
4703 * health_check to return 0 on healthy
4704 * and 1 on unhealthy.
4706 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4709 #ifdef USE_HEALTH_CHECK_WRITE
4710 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4711 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4716 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4719 return filter_fid2dentry(data, NULL, gr, id);
4722 static int filter_process_config(struct obd_device *obd, obd_count len,
4725 struct lustre_cfg *lcfg = buf;
4726 struct lprocfs_static_vars lvars;
4729 switch (lcfg->lcfg_command) {
4731 lprocfs_filter_init_vars(&lvars);
4733 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4743 static int filter_notify(struct obd_device *obd,
4744 struct obd_device *unused,
4745 enum obd_notify_event ev, void *data)
4748 case OBD_NOTIFY_CONFIG:
4749 LASSERT(obd->obd_no_conn);
4750 cfs_spin_lock(&obd->obd_dev_lock);
4751 obd->obd_no_conn = 0;
4752 cfs_spin_unlock(&obd->obd_dev_lock);
4755 CDEBUG(D_INFO, "%s: Unhandled notification %#x\n",
4761 static struct lvfs_callback_ops filter_lvfs_ops = {
4762 l_fid2dentry: filter_lvfs_fid2dentry,
4765 static struct obd_ops filter_obd_ops = {
4766 .o_owner = THIS_MODULE,
4767 .o_get_info = filter_get_info,
4768 .o_set_info_async = filter_set_info_async,
4769 .o_setup = filter_setup,
4770 .o_precleanup = filter_precleanup,
4771 .o_cleanup = filter_cleanup,
4772 .o_connect = filter_connect,
4773 .o_reconnect = filter_reconnect,
4774 .o_disconnect = filter_disconnect,
4775 .o_ping = filter_ping,
4776 .o_init_export = filter_init_export,
4777 .o_destroy_export = filter_destroy_export,
4778 .o_statfs = filter_statfs,
4779 .o_getattr = filter_getattr,
4780 .o_unpackmd = filter_unpackmd,
4781 .o_create = filter_create,
4782 .o_setattr = filter_setattr,
4783 .o_destroy = filter_destroy,
4784 .o_punch = filter_truncate,
4785 .o_sync = filter_sync,
4786 .o_preprw = filter_preprw,
4787 .o_commitrw = filter_commitrw,
4788 .o_llog_init = filter_llog_init,
4789 .o_llog_connect = filter_llog_connect,
4790 .o_llog_finish = filter_llog_finish,
4791 .o_iocontrol = filter_iocontrol,
4792 .o_health_check = filter_health_check,
4793 .o_process_config = filter_process_config,
4794 .o_notify = filter_notify,
4797 quota_interface_t *filter_quota_interface_ref;
4798 extern quota_interface_t filter_quota_interface;
4800 static int __init obdfilter_init(void)
4802 struct lprocfs_static_vars lvars;
4805 /** sanity check for group<->mdsno conversion */
4806 for (i = 0; i < MAX_MDT_COUNT; i++)
4807 LASSERT(objseq_to_mdsno(mdt_to_obd_objseq(i)) == i);
4809 lprocfs_filter_init_vars(&lvars);
4811 cfs_request_module("%s", "lquota");
4812 OBD_ALLOC(obdfilter_created_scratchpad,
4813 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4814 sizeof(*obdfilter_created_scratchpad));
4815 if (obdfilter_created_scratchpad == NULL)
4818 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4819 sizeof(struct filter_mod_data),
4822 GOTO(out, rc = -ENOMEM);
4824 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4825 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4827 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4828 LUSTRE_OST_NAME, NULL);
4832 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4833 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4834 ll_fmd_cachep = NULL;
4836 if (filter_quota_interface_ref)
4837 PORTAL_SYMBOL_PUT(filter_quota_interface);
4839 OBD_FREE(obdfilter_created_scratchpad,
4840 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4841 sizeof(*obdfilter_created_scratchpad));
4847 static void __exit obdfilter_exit(void)
4849 if (filter_quota_interface_ref)
4850 PORTAL_SYMBOL_PUT(filter_quota_interface);
4852 if (ll_fmd_cachep) {
4853 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4854 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4855 ll_fmd_cachep = NULL;
4858 class_unregister_type(LUSTRE_OST_NAME);
4859 OBD_FREE(obdfilter_created_scratchpad,
4860 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4861 sizeof(*obdfilter_created_scratchpad));
4864 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4865 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4866 MODULE_LICENSE("GPL");
4868 module_init(obdfilter_init);
4869 module_exit(obdfilter_exit);