1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/fs/obdfilter/filter.c
6 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
7 * Author: Peter Braam <braam@clusterfs.com>
8 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
28 * (which need to get journal_lock, may block if journal full).
30 * Invariant: Call filter_start_transno() before any journal ops to avoid the
31 * same deadlock problem. We can (and want) to get rid of the
32 * transno sem in favour of the dir/inode i_mutex to avoid single
33 * threaded operation on the OST.
36 #define DEBUG_SUBSYSTEM S_FILTER
38 #include <linux/config.h>
39 #include <linux/module.h>
41 #include <linux/dcache.h>
42 #include <linux/init.h>
43 #include <linux/version.h>
44 #include <linux/sched.h>
45 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
46 # include <linux/mount.h>
47 # include <linux/buffer_head.h>
50 #include <obd_class.h>
52 #include <lustre_dlm.h>
53 #include <lustre_fsfilt.h>
54 #include <lprocfs_status.h>
55 #include <lustre_log.h>
56 #include <lustre_commit_confd.h>
57 #include <libcfs/list.h>
58 #include <lustre_disk.h>
59 #include <lustre_quota.h>
60 #include <linux/slab.h>
61 #include <lustre_param.h>
63 #include "filter_internal.h"
65 static struct lvfs_callback_ops filter_lvfs_ops;
66 kmem_cache_t *ll_fmd_cachep;
68 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
69 void *cb_data, int error)
71 obd_transno_commit_cb(obd, transno, error);
74 /* Assumes caller has already pushed us into the kernel context. */
75 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
78 struct filter_obd *filter = &exp->exp_obd->u.filter;
79 struct filter_export_data *fed = &exp->exp_filter_data;
80 struct filter_client_data *fcd = fed->fed_fcd;
83 int err, log_pri = D_HA;
85 /* Propagate error code. */
89 if (!exp->exp_obd->obd_replayable || oti == NULL)
92 /* we don't allocate new transnos for replayed requests */
93 if (oti->oti_transno == 0) {
94 spin_lock(&filter->fo_translock);
95 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
96 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
97 spin_unlock(&filter->fo_translock);
98 oti->oti_transno = last_rcvd;
100 spin_lock(&filter->fo_translock);
101 last_rcvd = oti->oti_transno;
102 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
103 filter->fo_fsd->lsd_last_transno =
104 cpu_to_le64(last_rcvd);
105 spin_unlock(&filter->fo_translock);
107 fcd->fcd_last_rcvd = cpu_to_le64(last_rcvd);
109 /* could get xid from oti, if it's ever needed */
110 fcd->fcd_last_xid = 0;
112 off = fed->fed_lr_off;
114 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
115 fed->fed_lr_idx, fed->fed_lr_off);
118 fsfilt_add_journal_cb(exp->exp_obd, last_rcvd, oti->oti_handle,
119 filter_commit_cb, NULL);
120 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
121 fcd, sizeof(*fcd), &off,
130 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
131 last_rcvd, fcd->fcd_uuid, fed->fed_lr_idx, err);
136 void f_dput(struct dentry *dentry)
138 /* Can't go inside filter_ddelete because it can block */
139 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
140 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
141 LASSERT(atomic_read(&dentry->d_count) > 0);
146 static void init_brw_stats(struct brw_stats *brw_stats)
149 for (i = 0; i < BRW_LAST; i++)
150 spin_lock_init(&brw_stats->hist[i].oh_lock);
153 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
154 plus the procfs overhead :( */
155 static int filter_export_stats_init(struct obd_device *obd,
156 struct obd_export *exp)
158 struct filter_export_data *fed = &exp->exp_filter_data;
159 struct proc_dir_entry *brw_entry;
163 init_brw_stats(&fed->fed_brw_stats);
165 rc = lprocfs_exp_setup(exp);
169 /* Create a per export proc entry for brw_stats */
170 brw_entry = create_proc_entry("brw_stats", 0644, exp->exp_proc);
171 if (brw_entry == NULL)
173 brw_entry->proc_fops = &filter_per_export_stats_fops;
174 brw_entry->data = fed;
176 /* Create a per export proc entry for ops stats */
177 num_stats = (sizeof(*obd->obd_type->typ_ops) / sizeof(void *)) +
178 LPROC_FILTER_LAST - 1;
179 exp->exp_ops_stats = lprocfs_alloc_stats(num_stats);
180 if (exp->exp_ops_stats == NULL)
182 lprocfs_init_ops_stats(LPROC_FILTER_LAST, exp->exp_ops_stats);
183 lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_READ_BYTES,
184 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
185 lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_WRITE_BYTES,
186 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
187 lprocfs_register_stats(exp->exp_proc, "stats", exp->exp_ops_stats);
191 /* Add client data to the FILTER. We use a bitmap to locate a free space
192 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
193 * Otherwise, we have just read the data from the last_rcvd file and
194 * we know its offset. */
195 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
198 struct filter_obd *filter = &obd->u.filter;
199 struct filter_export_data *fed = &exp->exp_filter_data;
200 unsigned long *bitmap = filter->fo_last_rcvd_slots;
201 int new_client = (cl_idx == -1);
205 LASSERT(bitmap != NULL);
206 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
209 if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid) == 0) {
210 init_brw_stats(&fed->fed_brw_stats);
214 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
215 * there's no need for extra complication here
218 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
220 if (cl_idx >= LR_MAX_CLIENTS) {
221 CERROR("no room for %u clients - fix LR_MAX_CLIENTS\n",
225 if (test_and_set_bit(cl_idx, bitmap)) {
226 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
231 if (test_and_set_bit(cl_idx, bitmap)) {
232 CERROR("FILTER client %d: bit already set in bitmap!\n",
238 fed->fed_lr_idx = cl_idx;
239 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
240 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
241 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
242 filter_export_stats_init(obd, exp);
244 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
245 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
248 struct lvfs_run_ctxt saved;
249 loff_t off = fed->fed_lr_off;
253 CDEBUG(D_INFO, "writing client fcd at idx %u (%llu) (len %u)\n",
254 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_fcd));
256 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
257 /* Transaction needed to fix bug 1403 */
258 handle = fsfilt_start(obd,
259 filter->fo_rcvd_filp->f_dentry->d_inode,
260 FSFILT_OP_SETATTR, NULL);
261 if (IS_ERR(handle)) {
262 rc = PTR_ERR(handle);
263 CERROR("unable to start transaction: rc %d\n", rc);
265 rc = fsfilt_add_journal_cb(obd, 0, handle,
266 target_client_add_cb, exp);
268 exp->exp_need_sync = 1;
269 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
271 sizeof(*fed->fed_fcd),
272 &off, rc /* sync if no cb */);
274 filter->fo_rcvd_filp->f_dentry->d_inode,
277 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
280 CERROR("error writing %s client idx %u: rc %d\n",
281 LAST_RCVD, fed->fed_lr_idx, rc);
288 static int filter_client_free(struct obd_export *exp)
290 struct filter_export_data *fed = &exp->exp_filter_data;
291 struct filter_obd *filter = &exp->exp_obd->u.filter;
292 struct obd_device *obd = exp->exp_obd;
293 struct filter_client_data zero_fcd;
294 struct lvfs_run_ctxt saved;
299 if (fed->fed_fcd == NULL)
302 /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
303 if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid ) == 0)
306 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
307 fed->fed_lr_idx, off, fed->fed_fcd->fcd_uuid);
309 lprocfs_exp_cleanup(exp);
311 LASSERT(filter->fo_last_rcvd_slots != NULL);
313 off = fed->fed_lr_off;
315 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
316 * we leak a client slot that will be cleaned on the next recovery. */
318 CERROR("%s: client idx %d has med_off %lld\n",
319 obd->obd_name, fed->fed_lr_idx, off);
320 GOTO(free, rc = -EINVAL);
323 /* Clear the bit _after_ zeroing out the client so we don't
324 race with filter_client_add and zero out new clients.*/
325 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
326 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
331 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
332 memset(&zero_fcd, 0, sizeof zero_fcd);
333 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
334 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_fcd,
335 sizeof(zero_fcd), &off,
336 !exp->exp_libclient);
339 /* update server's transno */
340 filter_update_server_data(obd, filter->fo_rcvd_filp,
342 !exp->exp_libclient);
343 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
345 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
346 "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
347 fed->fed_fcd->fcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
351 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
352 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
359 OBD_FREE(fed->fed_fcd, sizeof(*fed->fed_fcd));
365 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
366 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
367 struct filter_mod_data *fmd)
369 LASSERT_SPIN_LOCKED(&fed->fed_lock);
370 if (--fmd->fmd_refcount == 0) {
371 /* XXX when we have persistent reservations and the handle
372 * is stored herein we need to drop it here. */
373 fed->fed_mod_count--;
374 list_del(&fmd->fmd_list);
375 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
379 /* drop fmd reference, free it if last ref */
380 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
382 struct filter_export_data *fed;
387 fed = &exp->exp_filter_data;
388 spin_lock(&fed->fed_lock);
389 filter_fmd_put_nolock(fed, fmd); /* caller reference */
390 spin_unlock(&fed->fed_lock);
393 /* expire entries from the end of the list if there are too many
394 * or they are too old */
395 static void filter_fmd_expire_nolock(struct filter_obd *filter,
396 struct filter_export_data *fed,
397 struct filter_mod_data *keep)
399 struct filter_mod_data *fmd, *tmp;
401 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
405 if (time_before(jiffies, fmd->fmd_expire) &&
406 fed->fed_mod_count < filter->fo_fmd_max_num)
409 list_del_init(&fmd->fmd_list);
410 filter_fmd_put_nolock(fed, fmd); /* list reference */
414 void filter_fmd_expire(struct obd_export *exp)
416 spin_lock(&exp->exp_filter_data.fed_lock);
417 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
418 &exp->exp_filter_data, NULL);
419 spin_unlock(&exp->exp_filter_data.fed_lock);
422 /* find specified objid, group in export fmd list.
423 * caller must hold fed_lock and take fmd reference itself */
424 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
425 struct filter_export_data *fed,
426 obd_id objid, obd_gr group)
428 struct filter_mod_data *found = NULL, *fmd;
430 LASSERT_SPIN_LOCKED(&fed->fed_lock);
432 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
433 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
435 list_del(&fmd->fmd_list);
436 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
437 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
442 filter_fmd_expire_nolock(filter, fed, found);
447 /* Find fmd based on objid and group, or return NULL if not found. */
448 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
449 obd_id objid, obd_gr group)
451 struct filter_mod_data *fmd;
453 spin_lock(&exp->exp_filter_data.fed_lock);
454 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
455 &exp->exp_filter_data, objid, group);
457 fmd->fmd_refcount++; /* caller reference */
458 spin_unlock(&exp->exp_filter_data.fed_lock);
463 /* Find fmd based on objid and group, or create a new one if none is found.
464 * It is possible for this function to return NULL under memory pressure,
465 * or if objid = 0 is passed (which will only cause old entries to expire).
466 * Currently this is not fatal because any fmd state is transient and
467 * may also be freed when it gets sufficiently old. */
468 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
469 obd_id objid, obd_gr group)
471 struct filter_export_data *fed = &exp->exp_filter_data;
472 struct filter_mod_data *found = NULL, *fmd_new = NULL;
474 OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, SLAB_NOFS, sizeof(*fmd_new));
476 spin_lock(&fed->fed_lock);
477 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
480 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
481 fmd_new->fmd_id = objid;
482 fmd_new->fmd_gr = group;
483 fmd_new->fmd_refcount++; /* list reference */
485 fed->fed_mod_count++;
487 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
491 found->fmd_refcount++; /* caller reference */
492 found->fmd_expire = jiffies +
493 exp->exp_obd->u.filter.fo_fmd_max_age;
496 spin_unlock(&fed->fed_lock);
502 /* drop fmd list reference so it will disappear when last reference is put.
503 * This isn't so critical because it would in fact only affect the one client
504 * that is doing the unlink and at worst we have an stale entry referencing
505 * an object that should never be used again. */
506 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
508 struct filter_mod_data *found = NULL;
510 spin_lock(&exp->exp_filter_data.fed_lock);
511 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
513 list_del_init(&found->fmd_list);
514 filter_fmd_put_nolock(&exp->exp_filter_data, found);
516 spin_unlock(&exp->exp_filter_data.fed_lock);
519 #define filter_fmd_drop(exp, objid, group)
522 /* remove all entries from fmd list */
523 static void filter_fmd_cleanup(struct obd_export *exp)
525 struct filter_export_data *fed = &exp->exp_filter_data;
526 struct filter_mod_data *fmd = NULL, *tmp;
528 spin_lock(&fed->fed_lock);
529 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
530 list_del_init(&fmd->fmd_list);
531 filter_fmd_put_nolock(fed, fmd);
533 spin_unlock(&fed->fed_lock);
536 static int filter_init_export(struct obd_export *exp)
538 spin_lock_init(&exp->exp_filter_data.fed_lock);
539 INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
540 exp->exp_connecting = 1;
545 static int filter_free_server_data(struct filter_obd *filter)
547 OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
548 filter->fo_fsd = NULL;
549 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
550 filter->fo_last_rcvd_slots = NULL;
554 /* assumes caller is already in kernel ctxt */
555 int filter_update_server_data(struct obd_device *obd, struct file *filp,
556 struct lr_server_data *fsd, int force_sync)
562 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
563 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
564 le64_to_cpu(fsd->lsd_last_transno));
565 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
566 le64_to_cpu(fsd->lsd_mount_count));
568 fsd->lsd_compat14 = fsd->lsd_last_transno;
569 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off,force_sync);
571 CERROR("error writing lr_server_data: rc = %d\n", rc);
576 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
579 struct filter_obd *filter = &obd->u.filter;
585 if (filter->fo_last_objid_files[group] == NULL) {
586 CERROR("Object group "LPU64" not fully setup; not updating "
587 "last_objid\n", group);
591 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
592 obd->obd_name, group, filter->fo_last_objids[group]);
594 tmp = cpu_to_le64(filter->fo_last_objids[group]);
595 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
596 &tmp, sizeof(tmp), &off, force_sync);
598 CERROR("error writing group "LPU64" last objid: rc = %d\n",
603 /* assumes caller has already in kernel ctxt */
604 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
606 struct filter_obd *filter = &obd->u.filter;
607 struct lr_server_data *fsd;
608 struct filter_client_data *fcd = NULL;
609 struct inode *inode = filp->f_dentry->d_inode;
610 unsigned long last_rcvd_size = inode->i_size;
616 /* ensure padding in the struct is the correct size */
617 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
618 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
619 CLASSERT (offsetof(struct filter_client_data, fcd_padding) +
620 sizeof(fcd->fcd_padding) == LR_CLIENT_SIZE);
622 OBD_ALLOC(fsd, sizeof(*fsd));
625 filter->fo_fsd = fsd;
627 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
628 if (filter->fo_last_rcvd_slots == NULL) {
629 OBD_FREE(fsd, sizeof(*fsd));
633 if (last_rcvd_size == 0) {
634 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
636 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
637 fsd->lsd_last_transno = 0;
638 mount_count = fsd->lsd_mount_count = 0;
639 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
640 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
641 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
642 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
643 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
644 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
646 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
648 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
652 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
653 LCONSOLE_ERROR("Trying to start OBD %s using the wrong"
654 " disk %s. Were the /dev/ assignments "
656 obd->obd_uuid.uuid, fsd->lsd_uuid);
657 GOTO(err_fsd, rc = -EINVAL);
659 mount_count = le64_to_cpu(fsd->lsd_mount_count);
660 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
662 /* Assume old last_rcvd format unless I_C_LR is set */
663 if (!(fsd->lsd_feature_incompat &
664 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
665 fsd->lsd_last_transno = fsd->lsd_compat14;
669 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
670 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
671 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
672 ~FILTER_INCOMPAT_SUPP);
673 GOTO(err_fsd, rc = -EINVAL);
675 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
676 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
677 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
678 ~FILTER_ROCOMPAT_SUPP);
679 /* Do something like remount filesystem read-only */
680 GOTO(err_fsd, rc = -EINVAL);
683 CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
684 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
685 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
686 obd->obd_name, mount_count + 1);
687 CDEBUG(D_INODE, "%s: server data size: %u\n",
688 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
689 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
690 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
691 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
692 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
693 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
694 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
695 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
696 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
697 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
698 le16_to_cpu(fsd->lsd_client_size));
700 if (!obd->obd_replayable) {
701 CWARN("%s: recovery support OFF\n", obd->obd_name);
705 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
706 off < last_rcvd_size; cl_idx++) {
708 struct obd_export *exp;
709 struct filter_export_data *fed;
712 OBD_ALLOC(fcd, sizeof(*fcd));
714 GOTO(err_client, rc = -ENOMEM);
717 /* Don't assume off is incremented properly by
718 * fsfilt_read_record(), in case sizeof(*fcd)
719 * isn't the same as fsd->lsd_client_size. */
720 off = le32_to_cpu(fsd->lsd_client_start) +
721 cl_idx * le16_to_cpu(fsd->lsd_client_size);
722 rc = fsfilt_read_record(obd, filp, fcd, sizeof(*fcd), &off);
724 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
725 LAST_RCVD, cl_idx, off, rc);
726 break; /* read error shouldn't cause startup to fail */
729 if (fcd->fcd_uuid[0] == '\0') {
730 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
735 last_rcvd = le64_to_cpu(fcd->fcd_last_rcvd);
737 /* These exports are cleaned up by filter_disconnect(), so they
738 * need to be set up like real exports as filter_connect() does.
740 exp = class_new_export(obd, (struct obd_uuid *)fcd->fcd_uuid);
741 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
742 " srv lr: "LPU64"\n", fcd->fcd_uuid, cl_idx,
743 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
745 if (PTR_ERR(exp) == -EALREADY) {
746 /* export already exists, zero out this one */
747 CERROR("Zeroing out duplicate export due to "
749 fcd->fcd_uuid[0] = '\0';
751 GOTO(err_client, rc = PTR_ERR(exp));
754 fed = &exp->exp_filter_data;
756 rc = filter_client_add(obd, exp, cl_idx);
757 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
760 exp->exp_replay_needed = 1;
761 exp->exp_connecting = 0;
762 obd->obd_recoverable_clients++;
763 obd->obd_max_recoverable_clients++;
764 class_export_put(exp);
767 /* Need to check last_rcvd even for duplicated exports. */
768 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
771 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
772 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
776 OBD_FREE(fcd, sizeof(*fcd));
778 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
780 if (obd->obd_recoverable_clients) {
781 CWARN("RECOVERY: service %s, %d recoverable clients, "
782 "last_rcvd "LPU64"\n", obd->obd_name,
783 obd->obd_recoverable_clients,
784 le64_to_cpu(fsd->lsd_last_transno));
785 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
786 obd->obd_recovering = 1;
787 obd->obd_recovery_start = CURRENT_SECONDS;
788 /* Only used for lprocfs_status */
789 obd->obd_recovery_end = obd->obd_recovery_start +
790 OBD_RECOVERY_TIMEOUT;
794 filter->fo_mount_count = mount_count + 1;
795 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
797 /* save it, so mount count and last_transno is current */
798 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
800 GOTO(err_client, rc);
805 class_disconnect_exports(obd);
807 filter_free_server_data(filter);
811 static int filter_cleanup_groups(struct obd_device *obd)
813 struct filter_obd *filter = &obd->u.filter;
815 struct dentry *dentry;
819 if (filter->fo_dentry_O_groups != NULL) {
820 for (i = 0; i < FILTER_GROUPS; i++) {
821 dentry = filter->fo_dentry_O_groups[i];
825 OBD_FREE(filter->fo_dentry_O_groups,
826 FILTER_GROUPS * sizeof(*filter->fo_dentry_O_groups));
827 filter->fo_dentry_O_groups = NULL;
829 if (filter->fo_last_objid_files != NULL) {
830 for (i = 0; i < FILTER_GROUPS; i++) {
831 filp = filter->fo_last_objid_files[i];
835 OBD_FREE(filter->fo_last_objid_files,
836 FILTER_GROUPS * sizeof(*filter->fo_last_objid_files));
837 filter->fo_last_objid_files = NULL;
839 if (filter->fo_dentry_O_sub != NULL) {
840 for (i = 0; i < filter->fo_subdir_count; i++) {
841 dentry = filter->fo_dentry_O_sub[i];
845 OBD_FREE(filter->fo_dentry_O_sub,
846 filter->fo_subdir_count *
847 sizeof(*filter->fo_dentry_O_sub));
848 filter->fo_dentry_O_sub = NULL;
850 if (filter->fo_last_objids != NULL) {
851 OBD_FREE(filter->fo_last_objids,
852 FILTER_GROUPS * sizeof(*filter->fo_last_objids));
853 filter->fo_last_objids = NULL;
855 if (filter->fo_dentry_O != NULL) {
856 f_dput(filter->fo_dentry_O);
857 filter->fo_dentry_O = NULL;
862 /* FIXME: object groups */
863 static int filter_prep_groups(struct obd_device *obd)
865 struct filter_obd *filter = &obd->u.filter;
866 struct dentry *dentry, *O_dentry;
868 int i, rc = 0, cleanup_phase = 0;
871 O_dentry = simple_mkdir(current->fs->pwd, "O", 0700, 1);
872 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
873 if (IS_ERR(O_dentry)) {
874 rc = PTR_ERR(O_dentry);
875 CERROR("cannot open/create O: rc = %d\n", rc);
878 filter->fo_dentry_O = O_dentry;
879 cleanup_phase = 1; /* O_dentry */
881 OBD_ALLOC(filter->fo_last_objids, FILTER_GROUPS * sizeof(__u64));
882 if (filter->fo_last_objids == NULL)
883 GOTO(cleanup, rc = -ENOMEM);
884 cleanup_phase = 2; /* groups */
886 OBD_ALLOC(filter->fo_dentry_O_groups, FILTER_GROUPS * sizeof(dentry));
887 if (filter->fo_dentry_O_groups == NULL)
888 GOTO(cleanup, rc = -ENOMEM);
889 OBD_ALLOC(filter->fo_last_objid_files, FILTER_GROUPS * sizeof(filp));
890 if (filter->fo_last_objid_files == NULL)
891 GOTO(cleanup, rc = -ENOMEM);
893 for (i = 0; i < FILTER_GROUPS; i++) {
897 sprintf(name, "%d", i);
898 dentry = simple_mkdir(O_dentry, name, 0700, 1);
899 CDEBUG(D_INODE, "got/created O/%s: %p\n", name, dentry);
900 if (IS_ERR(dentry)) {
901 rc = PTR_ERR(dentry);
902 CERROR("cannot lookup/create O/%s: rc = %d\n",
906 filter->fo_dentry_O_groups[i] = dentry;
908 sprintf(name, "O/%d/LAST_ID", i);
909 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
912 CERROR("cannot create %s: rc = %d\n", name, rc);
915 filter->fo_last_objid_files[i] = filp;
917 if (filp->f_dentry->d_inode->i_size == 0) {
918 filter->fo_last_objids[i] = FILTER_INIT_OBJID;
919 rc = filter_update_last_objid(obd, i, 1);
925 rc = fsfilt_read_record(obd, filp, &filter->fo_last_objids[i],
926 sizeof(__u64), &off);
928 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
932 filter->fo_last_objids[i] =
933 le64_to_cpu(filter->fo_last_objids[i]);
934 CDEBUG(D_HA, "%s: server last_objid group %d: "LPU64"\n",
935 obd->obd_name, i, filter->fo_last_objids[i]);
938 if (filter->fo_subdir_count) {
939 O_dentry = filter->fo_dentry_O_groups[0];
940 OBD_ALLOC(filter->fo_dentry_O_sub,
941 filter->fo_subdir_count * sizeof(dentry));
942 if (filter->fo_dentry_O_sub == NULL)
943 GOTO(cleanup, rc = -ENOMEM);
945 for (i = 0; i < filter->fo_subdir_count; i++) {
947 snprintf(dir, sizeof(dir), "d%u", i);
949 dentry = simple_mkdir(O_dentry, dir, 0700, 1);
950 CDEBUG(D_INODE, "got/created O/0/%s: %p\n", dir,dentry);
951 if (IS_ERR(dentry)) {
952 rc = PTR_ERR(dentry);
953 CERROR("can't lookup/create O/0/%s: rc = %d\n",
957 filter->fo_dentry_O_sub[i] = dentry;
963 filter_cleanup_groups(obd);
967 /* setup the object store with correct subdirectories */
968 static int filter_prep(struct obd_device *obd)
970 struct lvfs_run_ctxt saved;
971 struct filter_obd *filter = &obd->u.filter;
977 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
978 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
979 if (!file || IS_ERR(file)) {
981 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
985 filter->fo_rcvd_filp = file;
986 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
987 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
988 file->f_dentry->d_inode->i_mode);
989 GOTO(err_filp, rc = -ENOENT);
992 inode = file->f_dentry->d_parent->d_inode;
993 /* We use i_op->unlink directly in filter_vfs_unlink() */
994 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
995 CERROR("%s: filesystem does not support create/unlink ops\n",
997 GOTO(err_filp, rc = -EOPNOTSUPP);
1000 rc = filter_init_server_data(obd, file);
1002 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1005 /* open and create health check io file*/
1006 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1009 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1013 filter->fo_health_check_filp = file;
1014 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1015 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1016 file->f_dentry->d_inode->i_mode);
1017 GOTO(err_health_check, rc = -ENOENT);
1019 rc = lvfs_check_io_health(obd, file);
1021 GOTO(err_health_check, rc);
1023 rc = filter_prep_groups(obd);
1025 GOTO(err_server_data, rc);
1027 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1032 //class_disconnect_exports(obd, 0);
1033 filter_free_server_data(filter);
1035 if (filp_close(filter->fo_health_check_filp, 0))
1036 CERROR("can't close %s after error\n", HEALTH_CHECK);
1037 filter->fo_health_check_filp = NULL;
1039 if (filp_close(filter->fo_rcvd_filp, 0))
1040 CERROR("can't close %s after error\n", LAST_RCVD);
1041 filter->fo_rcvd_filp = NULL;
1045 /* cleanup the filter: write last used object id to status file */
1046 static void filter_post(struct obd_device *obd)
1048 struct lvfs_run_ctxt saved;
1049 struct filter_obd *filter = &obd->u.filter;
1052 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1053 * best to start a transaction with h_sync, because we removed this
1056 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1057 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1060 CERROR("error writing server data: rc = %d\n", rc);
1062 for (i = 0; i < FILTER_GROUPS; i++) {
1063 rc = filter_update_last_objid(obd, i, (i == FILTER_GROUPS - 1));
1065 CERROR("error writing group %d lastobjid: rc = %d\n",
1069 rc = filp_close(filter->fo_rcvd_filp, 0);
1070 filter->fo_rcvd_filp = NULL;
1072 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1074 rc = filp_close(filter->fo_health_check_filp, 0);
1075 filter->fo_health_check_filp = NULL;
1077 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1079 filter_cleanup_groups(obd);
1080 filter_free_server_data(filter);
1081 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1084 static void filter_set_last_id(struct filter_obd *filter,
1085 obd_id id, obd_gr group)
1087 LASSERT(filter->fo_fsd != NULL);
1088 LASSERT(group <= FILTER_GROUPS);
1090 spin_lock(&filter->fo_objidlock);
1091 filter->fo_last_objids[group] = id;
1092 spin_unlock(&filter->fo_objidlock);
1095 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1098 LASSERT(filter->fo_fsd != NULL);
1099 LASSERT(group <= FILTER_GROUPS);
1101 /* FIXME: object groups */
1102 spin_lock(&filter->fo_objidlock);
1103 id = filter->fo_last_objids[group];
1104 spin_unlock(&filter->fo_objidlock);
1109 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1111 LOCK_INODE_MUTEX(dparent->d_inode);
1115 /* We never dget the object parent, so DON'T dput it either */
1116 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1118 struct filter_obd *filter = &obd->u.filter;
1119 LASSERT(group < FILTER_GROUPS); /* FIXME: object groups */
1121 if (group > 0 || filter->fo_subdir_count == 0)
1122 return filter->fo_dentry_O_groups[group];
1124 return filter->fo_dentry_O_sub[objid & (filter->fo_subdir_count - 1)];
1127 /* We never dget the object parent, so DON'T dput it either */
1128 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1131 unsigned long now = jiffies;
1132 struct dentry *dparent = filter_parent(obd, group, objid);
1135 if (IS_ERR(dparent))
1138 rc = filter_lock_dentry(obd, dparent);
1139 fsfilt_check_slow(obd, now, obd_timeout, "parent lock");
1140 return rc ? ERR_PTR(rc) : dparent;
1143 /* We never dget the object parent, so DON'T dput it either */
1144 static void filter_parent_unlock(struct dentry *dparent)
1146 UNLOCK_INODE_MUTEX(dparent->d_inode);
1149 /* How to get files, dentries, inodes from object id's.
1151 * If dir_dentry is passed, the caller has already locked the parent
1152 * appropriately for this operation (normally a write lock). If
1153 * dir_dentry is NULL, we do a read lock while we do the lookup to
1154 * avoid races with create/destroy and such changing the directory
1155 * internal to the filesystem code. */
1156 struct dentry *filter_fid2dentry(struct obd_device *obd,
1157 struct dentry *dir_dentry,
1158 obd_gr group, obd_id id)
1160 struct dentry *dparent = dir_dentry;
1161 struct dentry *dchild;
1166 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1167 !obd->u.filter.fo_destroy_in_progress) {
1168 /* don't fail lookups for orphan recovery, it causes
1169 * later LBUGs when objects still exist during precreate */
1170 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1171 RETURN(ERR_PTR(-ENOENT));
1175 CERROR("fatal: invalid object id 0\n");
1176 RETURN(ERR_PTR(-ESTALE));
1179 len = sprintf(name, LPU64, id);
1180 if (dir_dentry == NULL) {
1181 dparent = filter_parent_lock(obd, group, id);
1182 if (IS_ERR(dparent)) {
1183 CERROR("%s: error getting object "LPU64":"LPU64
1184 " parent: rc %ld\n", obd->obd_name,
1185 id, group, PTR_ERR(dparent));
1189 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1190 dparent->d_name.len, dparent->d_name.name, name);
1191 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1192 if (dir_dentry == NULL)
1193 filter_parent_unlock(dparent);
1194 if (IS_ERR(dchild)) {
1195 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1200 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1201 CERROR("%s: got bad object "LPU64" inode %lu\n",
1202 obd->obd_name, id, dchild->d_inode->i_ino);
1204 RETURN(ERR_PTR(-ENOENT));
1207 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1208 name, dchild, atomic_read(&dchild->d_count));
1210 LASSERT(atomic_read(&dchild->d_count) > 0);
1215 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid)
1217 struct lustre_handle lockh;
1218 int flags = LDLM_AST_DISCARD_DATA, rc;
1219 struct ldlm_res_id res_id = { .name = { objid } };
1220 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1223 /* Tell the clients that the object is gone now and that they should
1224 * throw away any cached pages. */
1225 rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id, LDLM_EXTENT,
1226 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1227 ldlm_completion_ast, NULL, NULL, 0, NULL,
1230 /* We only care about the side-effects, just drop the lock. */
1232 ldlm_lock_decref(&lockh, LCK_PW);
1237 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1238 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1239 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1240 * reverse. Caller must take i_sem before starting the transaction and we
1241 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1242 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry)
1247 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1248 * itself we already hold dir->i_mutex for child create/unlink ops */
1249 LASSERT(dentry->d_inode != NULL);
1250 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1251 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1255 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1256 GOTO(out, rc = -ENOENT);
1258 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1263 GOTO(out, rc = -EPERM);
1265 /* check_sticky() */
1266 if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1267 IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1268 GOTO(out, rc = -EPERM);
1270 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1271 * that was done because of journal_start (which is already done
1272 * here) or some other ordering issue. */
1275 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1276 rc = security_inode_unlink(dir, dentry);
1281 rc = dir->i_op->unlink(dir, dentry);
1283 /* need to drop i_mutex before we lose inode reference */
1284 UNLOCK_INODE_MUTEX(dentry->d_inode);
1291 /* Caller must hold LCK_PW on parent and push us into kernel context.
1292 * Caller must hold child i_mutex, we drop it always.
1293 * Caller is also required to ensure that dchild->d_inode exists. */
1294 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1295 obd_gr group, struct dentry *dparent,
1296 struct dentry *dchild)
1298 struct inode *inode = dchild->d_inode;
1301 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1302 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1303 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1304 (unsigned long)inode->i_nlink,
1305 atomic_read(&inode->i_count));
1308 rc = filter_vfs_unlink(dparent->d_inode, dchild);
1310 CERROR("error unlinking objid %.*s: rc %d\n",
1311 dchild->d_name.len, dchild->d_name.name, rc);
1315 static int filter_intent_policy(struct ldlm_namespace *ns,
1316 struct ldlm_lock **lockp, void *req_cookie,
1317 ldlm_mode_t mode, int flags, void *data)
1319 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1320 struct ptlrpc_request *req = req_cookie;
1321 struct ldlm_lock *lock = *lockp, *l = NULL;
1322 struct ldlm_resource *res = lock->l_resource;
1323 ldlm_processing_policy policy;
1324 struct ost_lvb *res_lvb, *reply_lvb;
1325 struct ldlm_reply *rep;
1326 struct list_head *tmp;
1328 int rc, tmpflags = 0, only_liblustre = 0;
1329 int repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1330 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1331 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1334 policy = ldlm_get_processing_policy(res);
1335 LASSERT(policy != NULL);
1336 LASSERT(req != NULL);
1338 rc = lustre_pack_reply(req, 3, repsize, NULL);
1340 RETURN(req->rq_status = rc);
1342 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1343 LASSERT(rep != NULL);
1345 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1346 sizeof(*reply_lvb));
1347 LASSERT(reply_lvb != NULL);
1349 //fixup_handle_for_resent_req(req, lock, &lockh);
1351 /* If we grant any lock at all, it will be a whole-file read lock.
1352 * Call the extent policy function to see if our request can be
1353 * granted, or is blocked. */
1354 lock->l_policy_data.l_extent.start = 0;
1355 lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1356 lock->l_req_mode = LCK_PR;
1358 LASSERT(ns == res->lr_namespace);
1360 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1361 check_res_locked(res);
1363 /* FIXME: we should change the policy function slightly, to not make
1364 * this list at all, since we just turn around and free it */
1365 while (!list_empty(&rpc_list)) {
1366 struct ldlm_lock *wlock =
1367 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1368 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1369 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1370 lock->l_flags &= ~LDLM_FL_CP_REQD;
1371 list_del_init(&wlock->l_cp_ast);
1372 LDLM_LOCK_PUT(wlock);
1375 /* The lock met with no resistance; we're finished. */
1376 if (rc == LDLM_ITER_CONTINUE) {
1377 /* do not grant locks to the liblustre clients: they cannot
1378 * handle ASTs robustly. We need to do this while still
1379 * holding ns_lock to avoid the lock remaining on the res_link
1380 * list (and potentially being added to l_pending_list by an
1381 * AST) when we are going to drop this lock ASAP. */
1382 if (lock->l_export->exp_libclient ||
1383 OBD_FAIL_CHECK(OBD_FAIL_LDLM_GLIMPSE)) {
1384 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1385 ldlm_resource_unlink_lock(lock);
1386 err = ELDLM_LOCK_ABORTED;
1388 err = ELDLM_LOCK_REPLACED;
1394 /* Do not grant any lock, but instead send GL callbacks. The extent
1395 * policy nicely created a list of all PW locks for us. We will choose
1396 * the highest of those which are larger than the size in the LVB, if
1397 * any, and perform a glimpse callback. */
1398 res_lvb = res->lr_lvb_data;
1399 LASSERT(res_lvb != NULL);
1400 *reply_lvb = *res_lvb;
1402 list_for_each(tmp, &res->lr_granted) {
1403 struct ldlm_lock *tmplock =
1404 list_entry(tmp, struct ldlm_lock, l_res_link);
1406 if (tmplock->l_granted_mode == LCK_PR)
1409 * ->ns_lock guarantees that no new locks are granted, and,
1410 * therefore, that res->lr_lvb_data cannot increase beyond the
1411 * end of already granted lock. As a result, it is safe to
1412 * check against "stale" reply_lvb->lvb_size value without
1415 if (tmplock->l_policy_data.l_extent.end <= reply_lvb->lvb_size)
1418 /* Don't send glimpse ASTs to liblustre clients. They aren't
1419 * listening for them, and they do entirely synchronous I/O
1421 if (tmplock->l_export == NULL ||
1422 tmplock->l_export->exp_libclient == 1) {
1428 l = LDLM_LOCK_GET(tmplock);
1432 if (l->l_policy_data.l_extent.start >
1433 tmplock->l_policy_data.l_extent.start)
1437 l = LDLM_LOCK_GET(tmplock);
1441 /* There were no PW locks beyond the size in the LVB; finished. */
1443 if (only_liblustre) {
1444 /* If we discovered a liblustre client with a PW lock,
1445 * however, the LVB may be out of date! The LVB is
1446 * updated only on glimpse (which we don't do for
1447 * liblustre clients) and cancel (which the client
1448 * obviously has not yet done). So if it has written
1449 * data but kept the lock, the LVB is stale and needs
1450 * to be updated from disk.
1452 * Of course, this will all disappear when we switch to
1453 * taking liblustre locks on the OST. */
1454 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_update)
1455 ns->ns_lvbo->lvbo_update(res, NULL, 0, 1);
1457 RETURN(ELDLM_LOCK_ABORTED);
1461 * This check is for lock taken in filter_prepare_destroy() that does
1462 * not have l_glimpse_ast set. So the logic is: if there is a lock
1463 * with no l_glimpse_ast set, this object is being destroyed already.
1465 * Hence, if you are grabbing DLM locks on the server, always set
1466 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1468 if (l->l_glimpse_ast == NULL) {
1469 /* We are racing with unlink(); just return -ENOENT */
1470 rep->lock_policy_res1 = -ENOENT;
1474 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1475 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1476 /* Update the LVB from disk if the AST failed (this is a legal race) */
1478 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1479 * sending ast is not handled. This can result in lost client writes.
1481 if (rc != 0 && ns->ns_lvbo && ns->ns_lvbo->lvbo_update)
1482 ns->ns_lvbo->lvbo_update(res, NULL, 0, 1);
1485 *reply_lvb = *res_lvb;
1491 RETURN(ELDLM_LOCK_ABORTED);
1495 * per-obd_device iobuf pool.
1497 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1498 * allocations in write-path has to be minimized (see bug 5137).
1500 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1501 * OST threads (see ost_thread_{init,done}()).
1503 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1504 * at the OST layer there are only (potentially) multiple obd_device of type
1505 * unknown at the time of OST thread creation.
1507 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1508 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1509 * it's very own iobuf.
1513 * filter_kiobuf_pool_init()
1515 * filter_kiobuf_pool_done()
1517 * filter_iobuf_get()
1519 * operate on this array. They are "generic" in a sense that they don't depend
1520 * on actual type of iobuf's (the latter depending on Linux kernel version).
1524 * destroy pool created by filter_iobuf_pool_init
1526 static void filter_iobuf_pool_done(struct filter_obd *filter)
1528 struct filter_iobuf **pool;
1533 pool = filter->fo_iobuf_pool;
1535 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1536 if (pool[i] != NULL)
1537 filter_free_iobuf(pool[i]);
1539 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1540 filter->fo_iobuf_pool = NULL;
1546 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1548 static int filter_iobuf_pool_init(struct filter_obd *filter)
1554 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1556 if (filter->fo_iobuf_pool == NULL)
1559 filter->fo_iobuf_count = OSS_THREADS_MAX;
1564 /* Return iobuf allocated for @thread_id. We don't know in advance how
1565 * many threads there will be so we allocate a large empty array and only
1566 * fill in those slots that are actually in use.
1567 * If we haven't allocated a pool entry for this thread before, do so now. */
1568 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1570 int thread_id = oti ? oti->oti_thread_id : -1;
1571 struct filter_iobuf *pool = NULL;
1572 struct filter_iobuf **pool_place = NULL;
1574 if (thread_id >= 0) {
1575 LASSERT(thread_id < filter->fo_iobuf_count);
1576 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1579 if (unlikely(pool == NULL)) {
1580 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1581 PTLRPC_MAX_BRW_PAGES);
1582 if (pool_place != NULL)
1589 /* mount the file system (secretly). lustre_cfg parameters are:
1592 * 3 = flags: failover=f, failout=n
1595 int filter_common_setup(struct obd_device *obd, obd_count len, void *buf,
1598 struct lustre_cfg* lcfg = buf;
1599 struct filter_obd *filter = &obd->u.filter;
1600 struct vfsmount *mnt;
1601 struct lustre_mount_info *lmi;
1602 struct obd_uuid uuid;
1609 if (lcfg->lcfg_bufcount < 3 ||
1610 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1611 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1614 lmi = server_get_mount(obd->obd_name);
1616 /* We already mounted in lustre_fill_super.
1617 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1618 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1620 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1622 /* old path - used by lctl */
1623 CERROR("Using old MDS mount method\n");
1624 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1625 MS_NOATIME|MS_NODIRATIME,
1626 lustre_cfg_string(lcfg, 1), option);
1629 LCONSOLE_ERROR("Can't mount disk %s (%d)\n",
1630 lustre_cfg_string(lcfg, 1), rc);
1634 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1636 if (IS_ERR(obd->obd_fsops))
1637 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1639 rc = filter_iobuf_pool_init(filter);
1643 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1645 /* failover is the default */
1646 obd->obd_replayable = 1;
1648 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1649 str = lustre_cfg_string(lcfg, 3);
1650 if (strchr(str, 'n')) {
1651 CWARN("%s: recovery disabled\n", obd->obd_name);
1652 obd->obd_replayable = 0;
1656 filter->fo_vfsmnt = mnt;
1657 obd->u.obt.obt_sb = mnt->mnt_sb;
1658 filter->fo_fstype = mnt->mnt_sb->s_type->name;
1659 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1661 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1662 obd->obd_lvfs_ctxt.pwdmnt = mnt;
1663 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1664 obd->obd_lvfs_ctxt.fs = get_ds();
1665 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
1667 rc = filter_prep(obd);
1671 filter->fo_destroy_in_progress = 0;
1672 sema_init(&filter->fo_create_lock, 1);
1673 spin_lock_init(&filter->fo_translock);
1674 spin_lock_init(&filter->fo_objidlock);
1675 INIT_LIST_HEAD(&filter->fo_export_list);
1676 sema_init(&filter->fo_alloc_lock, 1);
1677 init_brw_stats(&filter->fo_filter_stats);
1678 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1679 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
1680 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
1682 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
1683 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
1684 if (obd->obd_namespace == NULL)
1685 GOTO(err_post, rc = -ENOMEM);
1686 obd->obd_namespace->ns_lvbp = obd;
1687 obd->obd_namespace->ns_lvbo = &filter_lvbo;
1688 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
1690 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1691 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1693 rc = llog_cat_initialize(obd, 1, NULL);
1695 CERROR("failed to setup llogging subsystems\n");
1699 rc = lquota_setup(filter_quota_interface_ref, obd);
1703 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
1704 if (uuid_ptr != NULL) {
1705 class_uuid_unparse(uuid_ptr, &uuid);
1711 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
1713 if (obd->obd_recovering) {
1714 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
1715 "recovery until %d %s reconnect, or if no clients"
1716 " reconnect for %d:%.02d; during that time new "
1717 "clients will not be allowed to connect. "
1718 "Recovery progress can be monitored by watching "
1719 "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
1720 obd->obd_name, lustre_cfg_string(lcfg, 1),
1721 label ?: "", label ? "/" : "", str,
1722 obd->obd_recoverable_clients,
1723 (obd->obd_recoverable_clients == 1)
1724 ? "client" : "clients",
1725 (int)(OBD_RECOVERY_TIMEOUT) / 60,
1726 (int)(OBD_RECOVERY_TIMEOUT) % 60,
1729 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
1730 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
1731 label ?: "", label ? "/" : "", str,
1732 obd->obd_replayable ? "enabled" : "disabled");
1740 fsfilt_put_ops(obd->obd_fsops);
1741 filter_iobuf_pool_done(filter);
1744 server_put_mount(obd->obd_name, mnt);
1751 obd->u.obt.obt_sb = 0;
1755 static int filter_setup(struct obd_device *obd, obd_count len, void *buf)
1757 struct lprocfs_static_vars lvars;
1758 struct lustre_cfg* lcfg = buf;
1762 CLASSERT(offsetof(struct obd_device, u.obt) ==
1763 offsetof(struct obd_device, u.filter.fo_obt));
1765 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
1768 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
1769 page = get_zeroed_page(GFP_KERNEL);
1773 memcpy((void *)page, lustre_cfg_buf(lcfg, 4),
1774 LUSTRE_CFG_BUFLEN(lcfg, 4));
1775 rc = filter_common_setup(obd, len, buf, (void *)page);
1778 lprocfs_init_vars(filter, &lvars);
1779 if (rc == 0 && lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
1780 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
1781 /* Init obdfilter private stats here */
1782 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
1783 LPROCFS_CNTR_AVGMINMAX,
1784 "read_bytes", "bytes");
1785 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
1786 LPROCFS_CNTR_AVGMINMAX,
1787 "write_bytes", "bytes");
1788 lproc_filter_attach_seqstat(obd);
1789 obd->obd_proc_exports = proc_mkdir("exports",
1790 obd->obd_proc_entry);
1796 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
1797 static struct llog_operations filter_size_orig_logops = {
1798 lop_setup: llog_obd_origin_setup,
1799 lop_cleanup: llog_obd_origin_cleanup,
1800 lop_add: llog_obd_origin_add
1803 static int filter_llog_init(struct obd_device *obd, struct obd_device *tgt,
1804 int count, struct llog_catid *catid,
1805 struct obd_uuid *uuid)
1807 struct llog_ctxt *ctxt;
1811 filter_mds_ost_repl_logops = llog_client_ops;
1812 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
1813 filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
1814 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
1816 rc = llog_setup(obd, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
1817 &filter_mds_ost_repl_logops);
1821 /* FIXME - assign unlink_cb for filter's recovery */
1822 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
1823 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
1825 rc = llog_setup(obd, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
1826 &filter_size_orig_logops);
1830 static int filter_llog_finish(struct obd_device *obd, int count)
1832 struct llog_ctxt *ctxt;
1833 int rc = 0, rc2 = 0;
1836 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
1838 rc = llog_cleanup(ctxt);
1840 ctxt = llog_get_context(obd, LLOG_SIZE_ORIG_CTXT);
1842 rc2 = llog_cleanup(ctxt);
1849 static int filter_precleanup(struct obd_device *obd,
1850 enum obd_cleanup_stage stage)
1856 case OBD_CLEANUP_EARLY:
1858 case OBD_CLEANUP_EXPORTS:
1859 target_cleanup_recovery(obd);
1861 case OBD_CLEANUP_SELF_EXP:
1862 rc = filter_llog_finish(obd, 0);
1864 case OBD_CLEANUP_OBD:
1870 static int filter_cleanup(struct obd_device *obd)
1872 struct filter_obd *filter = &obd->u.filter;
1873 lvfs_sbdev_type save_dev;
1874 int must_relock = 0, must_put = 0;
1878 LCONSOLE_WARN("%s: shutting down for failover; client state "
1879 "will be preserved.\n", obd->obd_name);
1881 if (!list_empty(&obd->obd_exports)) {
1882 CERROR("%s: still has clients!\n", obd->obd_name);
1883 class_disconnect_exports(obd);
1884 if (!list_empty(&obd->obd_exports)) {
1885 CERROR("still has exports after forced cleanup?\n");
1890 lprocfs_obd_cleanup(obd);
1891 lprocfs_free_obd_stats(obd);
1893 lquota_cleanup(filter_quota_interface_ref, obd);
1895 ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
1897 if (obd->u.obt.obt_sb == NULL)
1899 save_dev = lvfs_sbdev(obd->u.obt.obt_sb);
1903 shrink_dcache_parent(obd->u.obt.obt_sb->s_root);
1905 LL_DQUOT_OFF(obd->u.obt.obt_sb);
1907 must_put = server_put_mount(obd->obd_name, filter->fo_vfsmnt);
1908 /* must_put is for old method (l_p_m returns non-0 on err) */
1910 /* We can only unlock kernel if we are in the context of sys_ioctl,
1911 otherwise we never called lock_kernel */
1912 if (ll_kernel_locked()) {
1918 /* In case we didn't mount with lustre_get_mount -- old method*/
1919 mntput(filter->fo_vfsmnt);
1920 lvfs_clear_rdonly(save_dev);
1922 obd->u.obt.obt_sb = NULL;
1927 fsfilt_put_ops(obd->obd_fsops);
1929 filter_iobuf_pool_done(filter);
1931 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
1936 static int filter_connect_internal(struct obd_export *exp,
1937 struct obd_connect_data *data)
1942 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
1943 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
1944 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1945 data->ocd_connect_flags, data->ocd_version,
1946 data->ocd_grant, data->ocd_index);
1948 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
1949 exp->exp_connect_flags = data->ocd_connect_flags;
1950 data->ocd_version = LUSTRE_VERSION_CODE;
1952 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
1953 obd_size left, want;
1955 spin_lock(&exp->exp_obd->obd_osfs_lock);
1956 left = filter_grant_space_left(exp);
1957 want = data->ocd_grant;
1958 data->ocd_grant = filter_grant(exp, 0, want, left);
1959 spin_unlock(&exp->exp_obd->obd_osfs_lock);
1961 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
1962 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
1963 exp->exp_client_uuid.uuid, exp,
1964 data->ocd_grant, want, left);
1967 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
1968 struct filter_obd *filter = &exp->exp_obd->u.filter;
1969 struct lr_server_data *lsd = filter->fo_fsd;
1970 int index = le32_to_cpu(lsd->lsd_ost_index);
1972 if (!(lsd->lsd_feature_compat &
1973 cpu_to_le32(OBD_COMPAT_OST))) {
1974 /* this will only happen on the first connect */
1975 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
1976 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
1977 filter_update_server_data(exp->exp_obd,
1978 filter->fo_rcvd_filp, lsd, 1);
1979 } else if (index != data->ocd_index) {
1980 LCONSOLE_ERROR("Connection from %s to index %u doesn't "
1981 "match actual OST index %u in last_rcvd "
1982 "file, bad configuration?\n",
1983 obd_export_nid2str(exp), index,
1989 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
1990 data->ocd_brw_size = 65536;
1991 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
1992 data->ocd_brw_size = min(data->ocd_brw_size,
1993 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
1994 LASSERT(data->ocd_brw_size);
1997 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
1998 * FIXME: We don't strictly need the COMPAT flag for that,
1999 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2000 * FIXME: We needed it for the index, as index 0 is valid. */
2005 static int filter_reconnect(struct obd_export *exp, struct obd_device *obd,
2006 struct obd_uuid *cluuid,
2007 struct obd_connect_data *data)
2012 if (exp == NULL || obd == NULL || cluuid == NULL)
2015 rc = filter_connect_internal(exp, data);
2020 /* nearly identical to mds_connect */
2021 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
2022 struct obd_uuid *cluuid,
2023 struct obd_connect_data *data)
2025 struct obd_export *exp;
2026 struct filter_export_data *fed;
2027 struct filter_client_data *fcd = NULL;
2031 if (conn == NULL || obd == NULL || cluuid == NULL)
2034 rc = class_connect(conn, obd, cluuid);
2037 exp = class_conn2export(conn);
2038 LASSERT(exp != NULL);
2040 fed = &exp->exp_filter_data;
2042 rc = filter_connect_internal(exp, data);
2046 if (!obd->obd_replayable)
2047 GOTO(cleanup, rc = 0);
2049 OBD_ALLOC(fcd, sizeof(*fcd));
2051 CERROR("filter: out of memory for client data\n");
2052 GOTO(cleanup, rc = -ENOMEM);
2055 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
2058 rc = filter_client_add(obd, exp, -1);
2065 OBD_FREE(fcd, sizeof(*fcd));
2066 fed->fed_fcd = NULL;
2068 class_disconnect(exp);
2070 class_export_put(exp);
2076 /* Do extra sanity checks for grant accounting. We do this at connect,
2077 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2078 * always get rid of it or turn it off when we know accounting is good. */
2079 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2081 struct filter_export_data *fed;
2082 struct obd_export *exp;
2083 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2084 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2085 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2087 if (list_empty(&obd->obd_exports))
2090 spin_lock(&obd->obd_osfs_lock);
2091 spin_lock(&obd->obd_dev_lock);
2092 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2094 fed = &exp->exp_filter_data;
2095 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2098 if (maxsize > 0) { /* we may not have done a statfs yet */
2099 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2100 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2101 exp->exp_client_uuid.uuid, exp,
2102 fed->fed_grant, fed->fed_pending, maxsize);
2103 LASSERTF(fed->fed_dirty <= maxsize,
2104 "%s: cli %s/%p %ld > "LPU64"\n", func,
2105 exp->exp_client_uuid.uuid, exp,
2106 fed->fed_dirty, maxsize);
2109 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2110 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2111 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2113 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2114 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2115 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2116 tot_granted += fed->fed_grant + fed->fed_pending;
2117 tot_pending += fed->fed_pending;
2118 tot_dirty += fed->fed_dirty;
2120 fo_tot_granted = obd->u.filter.fo_tot_granted;
2121 fo_tot_pending = obd->u.filter.fo_tot_pending;
2122 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2123 spin_unlock(&obd->obd_dev_lock);
2124 spin_unlock(&obd->obd_osfs_lock);
2126 /* Do these assertions outside the spinlocks so we don't kill system */
2127 if (tot_granted != fo_tot_granted)
2128 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2129 func, tot_granted, fo_tot_granted);
2130 if (tot_pending != fo_tot_pending)
2131 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2132 func, tot_pending, fo_tot_pending);
2133 if (tot_dirty != fo_tot_dirty)
2134 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2135 func, tot_dirty, fo_tot_dirty);
2136 if (tot_pending > tot_granted)
2137 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2138 func, tot_pending, tot_granted);
2139 if (tot_granted > maxsize)
2140 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2141 func, tot_granted, maxsize);
2142 if (tot_dirty > maxsize)
2143 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2144 func, tot_dirty, maxsize);
2147 /* Remove this client from the grant accounting totals. We also remove
2148 * the export from the obd device under the osfs and dev locks to ensure
2149 * that the filter_grant_sanity_check() calculations are always valid.
2150 * The client should do something similar when it invalidates its import. */
2151 static void filter_grant_discard(struct obd_export *exp)
2153 struct obd_device *obd = exp->exp_obd;
2154 struct filter_obd *filter = &obd->u.filter;
2155 struct filter_export_data *fed = &exp->exp_filter_data;
2157 spin_lock(&obd->obd_osfs_lock);
2158 spin_lock(&obd->obd_dev_lock);
2159 list_del_init(&exp->exp_obd_chain);
2160 spin_unlock(&obd->obd_dev_lock);
2162 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2163 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2164 obd->obd_name, filter->fo_tot_granted,
2165 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2166 filter->fo_tot_granted -= fed->fed_grant;
2167 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2168 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2169 obd->obd_name, filter->fo_tot_pending,
2170 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2171 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2172 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2173 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2174 obd->obd_name, filter->fo_tot_dirty,
2175 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2176 filter->fo_tot_dirty -= fed->fed_dirty;
2180 spin_unlock(&obd->obd_osfs_lock);
2183 static int filter_destroy_export(struct obd_export *exp)
2187 if (exp->exp_filter_data.fed_pending)
2188 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2189 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2190 exp, exp->exp_filter_data.fed_pending);
2192 target_destroy_export(exp);
2194 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2197 if (exp->exp_obd->obd_replayable)
2198 filter_client_free(exp);
2200 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2202 filter_grant_discard(exp);
2203 filter_fmd_cleanup(exp);
2205 if (!(exp->exp_flags & OBD_OPT_FORCE))
2206 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2211 /* also incredibly similar to mds_disconnect */
2212 static int filter_disconnect(struct obd_export *exp)
2214 struct obd_device *obd = exp->exp_obd;
2215 struct llog_ctxt *ctxt;
2220 class_export_get(exp);
2222 if (!(exp->exp_flags & OBD_OPT_FORCE))
2223 filter_grant_sanity_check(obd, __FUNCTION__);
2224 filter_grant_discard(exp);
2226 /* Disconnect early so that clients can't keep using export */
2227 rc = class_disconnect(exp);
2228 if (exp->exp_obd->obd_namespace != NULL)
2229 ldlm_cancel_locks_for_export(exp);
2231 /* flush any remaining cancel messages out to the target */
2232 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2233 err = llog_sync(ctxt, exp);
2235 CERROR("error flushing logs to MDS: rc %d\n", err);
2237 class_export_put(exp);
2241 static int filter_ping(struct obd_export *exp)
2243 filter_fmd_expire(exp);
2248 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2249 const char *what, int quiet)
2251 struct dentry *dchild = NULL;
2253 if (!(oa->o_valid & OBD_MD_FLGROUP))
2256 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
2258 if (IS_ERR(dchild)) {
2259 CERROR("%s error looking up object: "LPU64"\n",
2264 if (dchild->d_inode == NULL) {
2266 CERROR("%s: %s on non-existent object: "LPU64"\n",
2267 obd->obd_name, what, oa->o_id);
2269 RETURN(ERR_PTR(-ENOENT));
2275 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2277 struct dentry *dentry = NULL;
2278 struct obd_device *obd;
2282 obd = class_exp2obd(exp);
2284 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
2288 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
2290 RETURN(PTR_ERR(dentry));
2292 /* Limit the valid bits in the return data to what we actually use */
2293 oinfo->oi_oa->o_valid = OBD_MD_FLID;
2294 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
2300 /* this should be enabled/disabled in condition to enabled/disabled large
2301 * inodes (fast EAs) in backing store FS. */
2302 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
2303 void *handle, struct obdo *oa)
2305 struct obd_device *obd = exp->exp_obd;
2309 if (oa->o_valid & OBD_MD_FLFID) {
2310 struct filter_fid ff;
2312 if (!(oa->o_valid & OBD_MD_FLGROUP))
2315 /* packing fid and converting it to LE for storing into EA.
2316 * Here ->o_stripe_idx should be filled by LOV and rest of
2317 * fields - by client. */
2318 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
2319 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
2320 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
2321 ff.ff_objid = cpu_to_le64(oa->o_id);
2322 ff.ff_group = cpu_to_le64(oa->o_gr);
2324 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
2325 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
2326 oa->o_generation, oa->o_id, oa->o_gr);
2328 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
2330 CERROR("store fid in object failed! rc: %d\n", rc);
2332 CDEBUG(D_HA, "OSS object without fid info!\n");
2338 /* this is called from filter_truncate() until we have filter_punch() */
2339 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
2340 struct obdo *oa, struct obd_trans_info *oti)
2342 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
2343 struct llog_cookie *fcc = NULL;
2344 struct filter_obd *filter;
2345 int rc, err, locked = 0;
2346 unsigned int ia_valid;
2347 struct inode *inode;
2352 LASSERT(dentry != NULL);
2353 LASSERT(!IS_ERR(dentry));
2355 inode = dentry->d_inode;
2356 LASSERT(inode != NULL);
2358 filter = &exp->exp_obd->u.filter;
2359 iattr_from_obdo(&iattr, oa, oa->o_valid);
2360 ia_valid = iattr.ia_valid;
2362 if (oa->o_valid & OBD_MD_FLCOOKIE) {
2363 OBD_ALLOC(fcc, sizeof(*fcc));
2365 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
2368 if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
2370 LOCK_INODE_MUTEX(inode);
2374 /* If the inode still has SUID+SGID bits set (see filter_precreate())
2375 * then we will accept the UID+GID sent by the client during write for
2376 * initializing the ownership of this inode. We only allow this to
2377 * happen once so clear these bits in setattr. In 2.6 kernels it is
2378 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
2379 * the flags that are actually being set. */
2380 if (ia_valid & (ATTR_UID | ATTR_GID)) {
2381 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
2382 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
2384 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
2385 if (!(ia_valid & ATTR_MODE)) {
2386 iattr.ia_mode = inode->i_mode;
2387 iattr.ia_valid |= ATTR_MODE;
2389 iattr.ia_mode &= ~S_ISUID;
2391 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
2392 if (!(iattr.ia_valid & ATTR_MODE)) {
2393 iattr.ia_mode = inode->i_mode;
2394 iattr.ia_valid |= ATTR_MODE;
2396 iattr.ia_mode &= ~S_ISGID;
2399 orig_ids[USRQUOTA] = inode->i_uid;
2400 orig_ids[GRPQUOTA] = inode->i_gid;
2401 handle = fsfilt_start_log(exp->exp_obd, inode,
2402 FSFILT_OP_SETATTR, oti, 1);
2404 /* update inode EA only once when inode is suid bit marked. As
2405 * on 2.6.x UID and GID may be set separately, we check here
2406 * only one of them to avoid double setting. */
2407 if (inode->i_mode & S_ISUID)
2408 filter_update_fidea(exp, inode, handle, oa);
2410 handle = fsfilt_start(exp->exp_obd, inode,
2411 FSFILT_OP_SETATTR, oti);
2415 GOTO(out_unlock, rc = PTR_ERR(handle));
2417 if (oa->o_valid & OBD_MD_FLFLAGS) {
2418 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
2419 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
2421 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
2423 /* set cancel cookie callback function */
2424 if (fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
2425 filter_cancel_cookies_cb,
2427 exp->exp_need_sync = 1;
2433 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
2434 fsfilt_extend(exp->exp_obd, inode, 0, handle);
2436 /* The truncate might have used up our transaction credits. Make
2437 * sure we have one left for the last_rcvd update. */
2438 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
2439 rc = filter_finish_transno(exp, oti, rc);
2441 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
2443 CERROR("error on commit, err = %d\n", err);
2449 /* Let's flush truncated page on disk immediately, then we can
2450 * avoid need to search for page aliases before directio writes
2451 * and this sort of stuff at expense of somewhat slower
2452 * truncates not on a page boundary. I believe this is the only
2453 * place in filter code that can lead to pages getting to
2454 * pagecache so far. */
2455 filter_clear_truncated_page(inode);
2456 UNLOCK_INODE_MUTEX(inode);
2463 UNLOCK_INODE_MUTEX(inode);
2466 OBD_FREE(fcc, sizeof(*fcc));
2468 /* trigger quota release */
2469 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
2470 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
2471 int rc2 = lquota_adjust(filter_quota_interface_ref, exp->exp_obd, cur_ids,
2472 orig_ids, rc, FSFILT_OP_SETATTR);
2473 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
2474 "filter adjust qunit. (rc:%d)\n", rc2);
2479 /* this is called from filter_truncate() until we have filter_punch() */
2480 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
2481 struct obd_trans_info *oti)
2483 struct ldlm_res_id res_id = { .name = { oinfo->oi_oa->o_id } };
2484 struct ldlm_valblock_ops *ns_lvbo;
2485 struct filter_mod_data *fmd;
2486 struct lvfs_run_ctxt saved;
2487 struct filter_obd *filter;
2488 struct ldlm_resource *res;
2489 struct dentry *dentry;
2493 dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
2496 RETURN(PTR_ERR(dentry));
2498 filter = &exp->exp_obd->u.filter;
2499 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
2502 if (oinfo->oi_oa->o_valid &
2503 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
2504 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
2505 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
2506 fmd->fmd_mactime_xid = oti->oti_xid;
2507 filter_fmd_put(exp, fmd);
2510 /* setting objects attributes (including owner/group) */
2511 rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
2513 GOTO(out_unlock, rc);
2515 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
2516 res_id, LDLM_EXTENT, 0);
2519 ns_lvbo = res->lr_namespace->ns_lvbo;
2520 if (ns_lvbo && ns_lvbo->lvbo_update)
2521 rc = ns_lvbo->lvbo_update(res, NULL, 0, 0);
2522 ldlm_resource_putref(res);
2525 oinfo->oi_oa->o_valid = OBD_MD_FLID;
2527 /* Quota release need uid/gid info */
2528 obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
2529 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
2535 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
2539 /* XXX identical to osc_unpackmd */
2540 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2541 struct lov_mds_md *lmm, int lmm_bytes)
2547 if (lmm_bytes < sizeof (*lmm)) {
2548 CERROR("lov_mds_md too small: %d, need %d\n",
2549 lmm_bytes, (int)sizeof(*lmm));
2552 /* XXX LOV_MAGIC etc check? */
2554 if (lmm->lmm_object_id == cpu_to_le64(0)) {
2555 CERROR("lov_mds_md: zero lmm_object_id\n");
2560 lsm_size = lov_stripe_md_size(1);
2564 if (*lsmp != NULL && lmm == NULL) {
2565 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
2566 OBD_FREE(*lsmp, lsm_size);
2571 if (*lsmp == NULL) {
2572 OBD_ALLOC(*lsmp, lsm_size);
2575 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
2576 if ((*lsmp)->lsm_oinfo[0] == NULL) {
2577 OBD_FREE(*lsmp, lsm_size);
2580 loi_init((*lsmp)->lsm_oinfo[0]);
2584 /* XXX zero *lsmp? */
2585 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
2586 LASSERT((*lsmp)->lsm_object_id);
2589 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
2594 /* caller must hold fo_create_lock */
2595 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
2596 struct filter_obd *filter)
2598 struct obdo doa; /* XXX obdo on stack */
2604 LASSERT(down_trylock(&filter->fo_create_lock) != 0);
2606 memset(&doa, 0, sizeof(doa));
2607 if (oa->o_valid & OBD_MD_FLGROUP) {
2608 doa.o_valid |= OBD_MD_FLGROUP;
2609 doa.o_gr = oa->o_gr;
2613 doa.o_mode = S_IFREG;
2615 if (!filter->fo_destroy_in_progress) {
2616 CERROR("%s: destroy_in_progress already cleared\n",
2617 exp->exp_obd->obd_name);
2621 last = filter_last_id(filter, doa.o_gr);
2622 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
2623 exp->exp_obd->obd_name, oa->o_id + 1, last);
2624 for (id = last; id > oa->o_id; id--) {
2626 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
2627 if (rc && rc != -ENOENT) /* this is pretty fatal... */
2628 CEMERG("error destroying precreate objid "LPU64": %d\n",
2630 filter_set_last_id(filter, id - 1, doa.o_gr);
2631 /* update last_id on disk periodically so that if we restart
2632 * we don't need to re-scan all of the just-deleted objects. */
2633 if ((id & 511) == 0)
2634 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
2637 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
2638 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
2640 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
2641 filter->fo_destroy_in_progress = 0;
2646 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
2647 obd_gr group, int *num);
2648 /* returns a negative error or a nonnegative number of files to create */
2649 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
2650 obd_gr group, struct obd_trans_info *oti)
2652 struct obd_device *obd = exp->exp_obd;
2653 struct filter_obd *filter = &obd->u.filter;
2657 /* delete orphans request */
2658 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
2659 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
2660 CERROR("%s: dropping old orphan cleanup request\n",
2665 /* This causes inflight precreates to abort and drop lock */
2666 filter->fo_destroy_in_progress = 1;
2667 down(&filter->fo_create_lock);
2668 diff = oa->o_id - filter_last_id(filter, group);
2669 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
2670 filter_last_id(filter, group), diff);
2672 if (-diff > OST_MAX_PRECREATE) {
2673 CERROR("%s: ignoring bogus orphan destroy request: "
2674 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
2675 oa->o_id, filter_last_id(filter, group));
2676 /* FIXME: should reset precreate_next_id on MDS */
2677 GOTO(out, rc = -EINVAL);
2680 rc = filter_destroy_precreated(exp, oa, filter);
2682 CERROR("%s: unable to write lastobjid, but "
2683 "orphans were deleted\n", obd->obd_name);
2686 /*XXX used by MDS for the first time! */
2687 filter->fo_destroy_in_progress = 0;
2690 down(&filter->fo_create_lock);
2691 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
2692 CERROR("%s: dropping old precreate request\n",
2696 /* only precreate if group == 0 and o_id is specfied */
2697 if (group != 0 || oa->o_id == 0)
2700 diff = oa->o_id - filter_last_id(filter, group);
2701 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
2702 filter_last_id(filter, group), diff);
2704 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
2705 oa->o_id, filter_last_id(filter, group), diff);
2709 oa->o_id = filter_last_id(&obd->u.filter, group);
2710 rc = filter_precreate(obd, oa, group, &diff);
2711 oa->o_id = filter_last_id(&obd->u.filter, group);
2712 oa->o_valid = OBD_MD_FLID;
2715 /* else diff == 0 */
2718 up(&filter->fo_create_lock);
2722 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2725 struct filter_obd *filter = &obd->u.filter;
2726 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
2730 /* at least try to account for cached pages. its still racey and
2731 * might be under-reporting if clients haven't announced their
2732 * caches with brw recently */
2733 spin_lock(&obd->obd_osfs_lock);
2734 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
2735 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
2736 spin_unlock(&obd->obd_osfs_lock);
2738 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
2739 " pending "LPU64" free "LPU64" avail "LPU64"\n",
2740 filter->fo_tot_dirty, filter->fo_tot_granted,
2741 filter->fo_tot_pending,
2742 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
2744 filter_grant_sanity_check(obd, __FUNCTION__);
2746 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
2747 ((filter->fo_tot_dirty + filter->fo_tot_pending +
2748 osfs->os_bsize - 1) >> blockbits));
2750 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
2751 * stop creating files on MDS if OST is not good shape to create
2753 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
2758 /* We rely on the fact that only one thread will be creating files in a given
2759 * group at a time, which is why we don't need an atomic filter_get_new_id.
2760 * Even if we had that atomic function, the following race would exist:
2762 * thread 1: gets id x from filter_next_id
2763 * thread 2: gets id (x + 1) from filter_next_id
2764 * thread 2: creates object (x + 1)
2765 * thread 1: tries to create object x, gets -ENOSPC
2767 * Caller must hold fo_create_lock
2769 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
2770 obd_gr group, int *num)
2772 struct dentry *dchild = NULL, *dparent = NULL;
2773 struct filter_obd *filter;
2774 int err = 0, rc = 0, recreate_obj = 0, i;
2775 unsigned long enough_time = jiffies + min(obd_timeout * HZ / 4, 10U*HZ);
2777 void *handle = NULL;
2780 filter = &obd->u.filter;
2782 LASSERT(down_trylock(&filter->fo_create_lock) != 0);
2784 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
2785 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
2788 struct obd_statfs *osfs;
2790 OBD_ALLOC(osfs, sizeof(*osfs));
2793 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ);
2794 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
2795 CDEBUG(D_HA,"%s: not enough space for create "LPU64"\n",
2796 obd->obd_name, osfs->os_bavail <<
2797 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
2801 OBD_FREE(osfs, sizeof(*osfs));
2806 CDEBUG(D_HA, "%s: precreating %d objects in group "LPU64" at "LPU64"\n",
2807 obd->obd_name, *num, group, oa->o_id);
2809 for (i = 0; i < *num && err == 0; i++) {
2810 int cleanup_phase = 0;
2812 if (filter->fo_destroy_in_progress) {
2813 CWARN("%s: precreate aborted by destroy\n",
2822 last_id = filter_last_id(filter, group);
2823 if (next_id > last_id) {
2824 CERROR("Error: Trying to recreate obj greater"
2825 "than last id "LPD64" > "LPD64"\n",
2827 GOTO(cleanup, rc = -EINVAL);
2830 next_id = filter_last_id(filter, group) + 1;
2832 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
2834 dparent = filter_parent_lock(obd, group, next_id);
2835 if (IS_ERR(dparent))
2836 GOTO(cleanup, rc = PTR_ERR(dparent));
2839 dchild = filter_fid2dentry(obd, dparent, group, next_id);
2841 GOTO(cleanup, rc = PTR_ERR(dchild));
2844 if (dchild->d_inode != NULL) {
2845 /* This would only happen if lastobjid was bad on disk*/
2846 /* Could also happen if recreating missing obj but
2850 CERROR("%s: recreating existing object %.*s?\n",
2851 obd->obd_name, dchild->d_name.len,
2852 dchild->d_name.name);
2854 CERROR("%s: Serious error: objid %.*s already "
2855 "exists; is this filesystem corrupt?\n",
2856 obd->obd_name, dchild->d_name.len,
2857 dchild->d_name.name);
2860 GOTO(cleanup, rc = -EEXIST);
2863 handle = fsfilt_start_log(obd, dparent->d_inode,
2864 FSFILT_OP_CREATE, NULL, 1);
2866 GOTO(cleanup, rc = PTR_ERR(handle));
2869 /* We mark object SUID+SGID to flag it for accepting UID+GID
2870 * from client on first write. Currently the permission bits
2871 * on the OST are never used, so this is OK. */
2872 rc = ll_vfs_create(dparent->d_inode, dchild,
2873 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
2875 CERROR("create failed rc = %d\n", rc);
2879 if (!recreate_obj) {
2880 filter_set_last_id(filter, next_id, group);
2881 err = filter_update_last_objid(obd, group, 0);
2883 CERROR("unable to write lastobjid "
2884 "but file created\n");
2888 switch(cleanup_phase) {
2890 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
2892 CERROR("error on commit, err = %d\n", err);
2899 filter_parent_unlock(dparent);
2906 if (time_after(jiffies, enough_time)) {
2907 CDEBUG(D_HA, "%s: precreate slow - want %d got %d \n",
2908 obd->obd_name, *num, i);
2914 CDEBUG(D_HA, "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
2915 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
2920 static int filter_create(struct obd_export *exp, struct obdo *oa,
2921 struct lov_stripe_md **ea, struct obd_trans_info *oti)
2923 struct obd_device *obd = NULL;
2924 struct lvfs_run_ctxt saved;
2925 struct lov_stripe_md *lsm = NULL;
2929 if (!(oa->o_valid & OBD_MD_FLGROUP))
2932 CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
2933 oa->o_gr, oa->o_id);
2937 rc = obd_alloc_memmd(exp, &lsm);
2944 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2946 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
2947 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
2948 if (oa->o_id > filter_last_id(&obd->u.filter, oa->o_gr)) {
2949 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
2950 oa->o_id, filter_last_id(&obd->u.filter,
2954 struct filter_obd *filter = &obd->u.filter;
2957 down(&filter->fo_create_lock);
2958 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
2959 up(&filter->fo_create_lock);
2962 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
2965 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2966 if (rc && ea != NULL && *ea != lsm) {
2967 obd_free_memmd(exp, &lsm);
2968 } else if (rc == 0 && ea != NULL) {
2969 /* XXX LOV STACKING: the lsm that is passed to us from
2970 * LOV does not have valid lsm_oinfo data structs, so
2971 * don't go touching that. This needs to be fixed in a
2973 lsm->lsm_object_id = oa->o_id;
2980 int filter_destroy(struct obd_export *exp, struct obdo *oa,
2981 struct lov_stripe_md *md, struct obd_trans_info *oti,
2982 struct obd_export *md_exp)
2984 unsigned int qcids[MAXQUOTAS] = {0, 0};
2985 struct obd_device *obd;
2986 struct filter_obd *filter;
2987 struct dentry *dchild = NULL, *dparent = NULL;
2988 struct lvfs_run_ctxt saved;
2989 void *handle = NULL;
2990 struct llog_cookie *fcc = NULL;
2991 int rc, rc2, cleanup_phase = 0;
2995 if (!(oa->o_valid & OBD_MD_FLGROUP))
2999 filter = &obd->u.filter;
3001 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3004 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3006 GOTO(cleanup, rc = PTR_ERR(dchild));
3009 if (dchild->d_inode == NULL) {
3010 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3012 /* If object already gone, cancel cookie right now */
3013 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3014 fcc = obdo_logcookie(oa);
3015 llog_cancel(llog_get_context(obd, fcc->lgc_subsys + 1),
3017 fcc = NULL; /* we didn't allocate fcc, don't free it */
3019 GOTO(cleanup, rc = -ENOENT);
3022 filter_prepare_destroy(obd, oa->o_id);
3024 /* Our MDC connection is established by the MDS to us */
3025 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3026 OBD_ALLOC(fcc, sizeof(*fcc));
3028 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3030 DQUOT_INIT(dchild->d_inode);
3032 /* we're gonna truncate it first in order to avoid possible deadlock:
3034 * open trasaction open transaction
3035 * down(i_zombie) down(i_zombie)
3036 * restart transaction
3037 * (see BUG 4180) -bzzz
3039 LOCK_INODE_MUTEX(dchild->d_inode);
3040 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3042 if (IS_ERR(handle)) {
3043 UNLOCK_INODE_MUTEX(dchild->d_inode);
3044 GOTO(cleanup, rc = PTR_ERR(handle));
3047 iattr.ia_valid = ATTR_SIZE;
3049 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3050 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3051 UNLOCK_INODE_MUTEX(dchild->d_inode);
3055 GOTO(cleanup, rc = rc2);
3057 /* We don't actually need to lock the parent until we are unlinking
3058 * here, and not while truncating above. That avoids holding the
3059 * parent lock for a long time during truncate, which can block other
3060 * threads from doing anything to objects in that directory. bug 7171 */
3061 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3062 if (IS_ERR(dparent))
3063 GOTO(cleanup, rc = PTR_ERR(dparent));
3064 cleanup_phase = 3; /* filter_parent_unlock */
3066 LOCK_INODE_MUTEX(dchild->d_inode);
3067 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3068 if (IS_ERR(handle)) {
3069 UNLOCK_INODE_MUTEX(dchild->d_inode);
3070 GOTO(cleanup, rc = PTR_ERR(handle));
3072 cleanup_phase = 4; /* fsfilt_commit */
3074 /* Quota release need uid/gid of inode */
3075 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3077 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3079 /* this drops dchild->d_inode->i_mutex unconditionally */
3080 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3084 switch(cleanup_phase) {
3087 if (fsfilt_add_journal_cb(obd, 0, oti ?
3088 oti->oti_handle : handle,
3089 filter_cancel_cookies_cb,
3093 rc = filter_finish_transno(exp, oti, rc);
3094 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3096 CERROR("error on commit, err = %d\n", rc2);
3101 filter_parent_unlock(dparent);
3105 OBD_FREE(fcc, sizeof(*fcc));
3107 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3110 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3114 /* trigger quota release */
3115 qcids[USRQUOTA] = oa->o_uid;
3116 qcids[GRPQUOTA] = oa->o_gid;
3117 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
3119 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3120 "filter adjust qunit! (rc:%d)\n", rc2);
3124 /* NB start and end are used for punch, but not truncate */
3125 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3126 struct obd_trans_info *oti,
3127 struct ptlrpc_request_set *rqset)
3132 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3133 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
3134 oinfo->oi_policy.l_extent.end);
3138 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
3139 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
3140 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
3142 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
3143 rc = filter_setattr(exp, oinfo, oti);
3147 static int filter_sync(struct obd_export *exp, struct obdo *oa,
3148 struct lov_stripe_md *lsm, obd_off start, obd_off end)
3150 struct lvfs_run_ctxt saved;
3151 struct filter_obd *filter;
3152 struct dentry *dentry;
3153 struct llog_ctxt *ctxt;
3157 filter = &exp->exp_obd->u.filter;
3159 /* an objid of zero is taken to mean "sync whole filesystem" */
3160 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
3161 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
3162 /* flush any remaining cancel messages out to the target */
3163 ctxt = llog_get_context(exp->exp_obd, LLOG_MDS_OST_REPL_CTXT);
3164 llog_sync(ctxt, exp);
3168 dentry = filter_oa2dentry(exp->exp_obd, oa);
3170 RETURN(PTR_ERR(dentry));
3172 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3174 LOCK_INODE_MUTEX(dentry->d_inode);
3176 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
3178 /* just any file to grab fsync method - "file" arg unused */
3179 struct file *file = filter->fo_rcvd_filp;
3181 if (file->f_op && file->f_op->fsync)
3182 rc = file->f_op->fsync(NULL, dentry, 1);
3184 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
3188 UNLOCK_INODE_MUTEX(dentry->d_inode);
3190 oa->o_valid = OBD_MD_FLID;
3191 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
3193 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3199 static int filter_get_info(struct obd_export *exp, __u32 keylen,
3200 void *key, __u32 *vallen, void *val)
3202 struct obd_device *obd;
3205 obd = class_exp2obd(exp);
3207 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3211 if (keylen == strlen("blocksize") &&
3212 memcmp(key, "blocksize", keylen) == 0) {
3213 __u32 *blocksize = val;
3214 *vallen = sizeof(*blocksize);
3215 *blocksize = obd->u.obt.obt_sb->s_blocksize;
3219 if (keylen == strlen("blocksize_bits") &&
3220 memcmp(key, "blocksize_bits", keylen) == 0) {
3221 __u32 *blocksize_bits = val;
3222 *vallen = sizeof(*blocksize_bits);
3223 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
3227 if (keylen >= strlen("last_id") && memcmp(key, "last_id", 7) == 0) {
3228 obd_id *last_id = val;
3229 /* FIXME: object groups */
3230 *last_id = filter_last_id(&obd->u.filter, 0);
3233 CDEBUG(D_IOCTL, "invalid key\n");
3237 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
3238 void *key, __u32 vallen, void *val,
3239 struct ptlrpc_request_set *set)
3241 struct obd_device *obd;
3242 struct llog_ctxt *ctxt;
3248 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
3252 if (keylen < strlen(KEY_MDS_CONN) ||
3253 memcmp(key, KEY_MDS_CONN, keylen) != 0)
3256 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
3257 obd_export_nid2str(exp));
3258 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
3260 /* setup llog imports */
3261 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
3262 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
3264 lquota_setinfo(filter_quota_interface_ref, exp, obd);
3269 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
3270 int len, void *karg, void *uarg)
3272 struct obd_device *obd = exp->exp_obd;
3273 struct obd_ioctl_data *data = karg;
3277 case OBD_IOC_ABORT_RECOVERY: {
3278 CERROR("aborting recovery for device %s\n", obd->obd_name);
3279 target_abort_recovery(obd);
3283 case OBD_IOC_SYNC: {
3284 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
3285 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3289 case OBD_IOC_SET_READONLY: {
3291 struct super_block *sb = obd->u.obt.obt_sb;
3292 struct inode *inode = sb->s_root->d_inode;
3293 BDEVNAME_DECLARE_STORAGE(tmp);
3294 CERROR("*** setting device %s read-only ***\n",
3295 ll_bdevname(sb, tmp));
3297 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
3298 if (!IS_ERR(handle))
3299 rc = fsfilt_commit(obd, inode, handle, 1);
3301 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
3302 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3304 lvfs_set_rdonly(lvfs_sbdev(obd->u.obt.obt_sb));
3308 case OBD_IOC_CATLOGLIST: {
3309 rc = llog_catalog_list(obd, 1, data);
3313 case OBD_IOC_LLOG_CANCEL:
3314 case OBD_IOC_LLOG_REMOVE:
3315 case OBD_IOC_LLOG_INFO:
3316 case OBD_IOC_LLOG_PRINT: {
3317 /* FIXME to be finished */
3318 RETURN(-EOPNOTSUPP);
3320 struct llog_ctxt *ctxt = NULL;
3322 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3323 rc = llog_ioctl(ctxt, cmd, data);
3324 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3337 static int filter_health_check(struct obd_device *obd)
3339 struct filter_obd *filter = &obd->u.filter;
3343 * health_check to return 0 on healthy
3344 * and 1 on unhealthy.
3346 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
3349 LASSERT(filter->fo_health_check_filp != NULL);
3350 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
3355 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
3358 return filter_fid2dentry(data, NULL, gr, id);
3361 static int filter_process_config(struct obd_device *obd, obd_count len, void *buf)
3363 struct lustre_cfg *lcfg = buf;
3364 struct lprocfs_static_vars lvars;
3367 lprocfs_init_vars(filter, &lvars);
3369 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars, lcfg, obd);
3374 static struct lvfs_callback_ops filter_lvfs_ops = {
3375 l_fid2dentry: filter_lvfs_fid2dentry,
3378 static struct obd_ops filter_obd_ops = {
3379 .o_owner = THIS_MODULE,
3380 .o_get_info = filter_get_info,
3381 .o_set_info_async = filter_set_info_async,
3382 .o_setup = filter_setup,
3383 .o_precleanup = filter_precleanup,
3384 .o_cleanup = filter_cleanup,
3385 .o_connect = filter_connect,
3386 .o_reconnect = filter_reconnect,
3387 .o_disconnect = filter_disconnect,
3388 .o_ping = filter_ping,
3389 .o_init_export = filter_init_export,
3390 .o_destroy_export = filter_destroy_export,
3391 .o_statfs = filter_statfs,
3392 .o_getattr = filter_getattr,
3393 .o_unpackmd = filter_unpackmd,
3394 .o_create = filter_create,
3395 .o_setattr = filter_setattr,
3396 .o_destroy = filter_destroy,
3397 .o_brw = filter_brw,
3398 .o_punch = filter_truncate,
3399 .o_sync = filter_sync,
3400 .o_preprw = filter_preprw,
3401 .o_commitrw = filter_commitrw,
3402 .o_llog_init = filter_llog_init,
3403 .o_llog_finish = filter_llog_finish,
3404 .o_iocontrol = filter_iocontrol,
3405 .o_health_check = filter_health_check,
3406 .o_process_config = filter_process_config,
3409 quota_interface_t *filter_quota_interface_ref;
3410 extern quota_interface_t filter_quota_interface;
3412 static int __init obdfilter_init(void)
3414 struct lprocfs_static_vars lvars;
3417 printk(KERN_INFO "Lustre: Filtering OBD driver; info@clusterfs.com\n");
3419 lprocfs_init_vars(filter, &lvars);
3421 request_module("lquota");
3422 OBD_ALLOC(obdfilter_created_scratchpad,
3423 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
3424 sizeof(*obdfilter_created_scratchpad));
3425 if (obdfilter_created_scratchpad == NULL)
3428 ll_fmd_cachep = kmem_cache_create("ll_fmd_cache",
3429 sizeof(struct filter_mod_data),
3432 GOTO(out, rc = -ENOMEM);
3434 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
3435 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
3437 rc = class_register_type(&filter_obd_ops, lvars.module_vars,
3442 err = kmem_cache_destroy(ll_fmd_cachep);
3443 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
3444 ll_fmd_cachep = NULL;
3446 if (filter_quota_interface_ref)
3447 PORTAL_SYMBOL_PUT(filter_quota_interface);
3449 OBD_FREE(obdfilter_created_scratchpad,
3450 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
3451 sizeof(*obdfilter_created_scratchpad));
3457 static void __exit obdfilter_exit(void)
3459 if (filter_quota_interface_ref)
3460 PORTAL_SYMBOL_PUT(filter_quota_interface);
3462 if (ll_fmd_cachep) {
3463 int rc = kmem_cache_destroy(ll_fmd_cachep);
3464 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
3465 ll_fmd_cachep = NULL;
3468 class_unregister_type(LUSTRE_OST_NAME);
3469 OBD_FREE(obdfilter_created_scratchpad,
3470 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
3471 sizeof(*obdfilter_created_scratchpad));
3474 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3475 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
3476 MODULE_LICENSE("GPL");
3478 module_init(obdfilter_init);
3479 module_exit(obdfilter_exit);