1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
38 * lustre/obdfilter/filter.c
40 * Author: Peter Braam <braam@clusterfs.com>
41 * Author: Andreas Dilger <adilger@clusterfs.com>
45 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
46 * (which need to get journal_lock, may block if journal full).
48 * Invariant: Call filter_start_transno() before any journal ops to avoid the
49 * same deadlock problem. We can (and want) to get rid of the
50 * transno sem in favour of the dir/inode i_mutex to avoid single
51 * threaded operation on the OST.
54 #define DEBUG_SUBSYSTEM S_FILTER
56 #ifndef AUTOCONF_INCLUDED
57 #include <linux/config.h>
59 #include <linux/module.h>
61 #include <linux/dcache.h>
62 #include <linux/init.h>
63 #include <linux/version.h>
64 #include <linux/sched.h>
65 #include <linux/mount.h>
66 #include <linux/buffer_head.h>
68 #include <obd_cksum.h>
69 #include <obd_class.h>
71 #include <lustre_dlm.h>
72 #include <lustre_fsfilt.h>
73 #include <lprocfs_status.h>
74 #include <lustre_log.h>
75 #include <libcfs/list.h>
76 #include <lustre_disk.h>
77 #include <lustre_quota.h>
78 #include <linux/slab.h>
79 #include <lustre_param.h>
80 #include <lustre/ll_fiemap.h>
82 #include "filter_internal.h"
84 static struct lvfs_callback_ops filter_lvfs_ops;
85 cfs_mem_cache_t *ll_fmd_cachep;
87 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
88 void *cb_data, int error)
90 struct obd_export *exp = cb_data;
91 LASSERT(exp->exp_obd == obd);
92 obd_transno_commit_cb(obd, transno, exp, error);
93 class_export_cb_put(exp);
96 int filter_version_get_check(struct obd_export *exp,
97 struct obd_trans_info *oti, struct inode *inode)
101 if (inode == NULL || oti == NULL)
104 curr_version = fsfilt_get_version(exp->exp_obd, inode);
105 if ((__s64)curr_version == -EOPNOTSUPP)
107 /* VBR: version is checked always because costs nothing */
108 if (oti->oti_pre_version != 0 &&
109 oti->oti_pre_version != curr_version) {
110 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
111 oti->oti_pre_version, curr_version);
112 cfs_spin_lock(&exp->exp_lock);
113 exp->exp_vbr_failed = 1;
114 cfs_spin_unlock(&exp->exp_lock);
117 oti->oti_pre_version = curr_version;
121 /* Assumes caller has already pushed us into the kernel context. */
122 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
123 struct obd_trans_info *oti, int rc, int force_sync)
125 struct obd_device_target *obt = &exp->exp_obd->u.obt;
126 struct tg_export_data *ted = &exp->exp_target_data;
127 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
128 struct lsd_client_data *lcd;
131 int err, log_pri = D_RPCTRACE;
133 /* Propagate error code. */
137 if (!exp->exp_obd->obd_replayable || oti == NULL)
140 cfs_mutex_lock(&ted->ted_lcd_lock);
142 /* if the export has already been disconnected, we have no last_rcvd slot,
143 * update server data with latest transno then */
145 cfs_mutex_unlock(&ted->ted_lcd_lock);
146 CWARN("commit transaction for disconnected client %s: rc %d\n",
147 exp->exp_client_uuid.uuid, rc);
148 err = filter_update_server_data(exp->exp_obd);
152 /* we don't allocate new transnos for replayed requests */
153 cfs_spin_lock(&obt->obt_lut->lut_translock);
154 if (oti->oti_transno == 0) {
155 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
156 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
157 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
159 last_rcvd = oti->oti_transno;
160 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
161 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
162 if (unlikely(last_rcvd < le64_to_cpu(lcd->lcd_last_transno))) {
163 CERROR("Trying to overwrite bigger transno, on-disk: "
164 LPU64", new: "LPU64"\n",
165 le64_to_cpu(lcd->lcd_last_transno), last_rcvd);
166 cfs_spin_lock(&exp->exp_lock);
167 exp->exp_vbr_failed = 1;
168 cfs_spin_unlock(&exp->exp_lock);
169 cfs_spin_unlock(&obt->obt_lut->lut_translock);
170 cfs_mutex_unlock(&ted->ted_lcd_lock);
174 oti->oti_transno = last_rcvd;
176 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
177 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
178 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
179 cfs_spin_unlock(&obt->obt_lut->lut_translock);
182 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
184 off = ted->ted_lr_off;
186 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
187 ted->ted_lr_idx, ted->ted_lr_off);
190 class_export_cb_get(exp); /* released when the cb is called */
192 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
198 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
199 lcd, sizeof(*lcd), &off,
200 force_sync | exp->exp_need_sync);
202 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
210 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
211 last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
212 cfs_mutex_unlock(&ted->ted_lcd_lock);
216 void f_dput(struct dentry *dentry)
218 /* Can't go inside filter_ddelete because it can block */
219 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
220 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
221 LASSERT(atomic_read(&dentry->d_count) > 0);
226 static void init_brw_stats(struct brw_stats *brw_stats)
229 for (i = 0; i < BRW_LAST; i++)
230 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
233 static int lprocfs_init_rw_stats(struct obd_device *obd,
234 struct lprocfs_stats **stats)
238 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
239 LPROC_FILTER_LAST - 1;
240 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
244 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
245 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
246 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
247 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
248 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
253 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
254 plus the procfs overhead :( */
255 static int filter_export_stats_init(struct obd_device *obd,
256 struct obd_export *exp,
262 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
263 /* Self-export gets no proc entry */
266 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
268 /* Mask error for already created
276 struct nid_stat *tmp = exp->exp_nid_stats;
277 LASSERT(tmp != NULL);
279 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
280 if (tmp->nid_brw_stats == NULL)
281 GOTO(clean, rc = -ENOMEM);
283 init_brw_stats(tmp->nid_brw_stats);
284 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
285 0644, &filter_per_nid_stats_fops,
288 CWARN("Error adding the brw_stats file\n");
290 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
294 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
298 rc = lprocfs_nid_ldlm_stats_init(tmp);
308 /* Add client data to the FILTER. We use a bitmap to locate a free space
309 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
310 * Otherwise, we have just read the data from the last_rcvd file and
311 * we know its offset. */
312 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
315 struct obd_device_target *obt = &obd->u.obt;
316 struct tg_export_data *ted = &exp->exp_target_data;
317 struct lr_server_data *lsd = class_server_data(obd);
318 unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
319 int new_client = (cl_idx == -1);
323 LASSERT(bitmap != NULL);
324 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
327 if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
330 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
331 * there's no need for extra complication here
334 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
336 if (cl_idx >= LR_MAX_CLIENTS) {
337 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
341 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
342 cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
347 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
348 CERROR("FILTER client %d: bit already set in bitmap!\n",
354 ted->ted_lr_idx = cl_idx;
355 ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
356 cl_idx * le16_to_cpu(lsd->lsd_client_size);
357 cfs_mutex_init(&ted->ted_lcd_lock);
358 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
360 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
361 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
364 struct lvfs_run_ctxt saved;
365 loff_t off = ted->ted_lr_off;
369 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
370 ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
372 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
375 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
376 /* Transaction needed to fix bug 1403 */
377 handle = fsfilt_start(obd,
378 obt->obt_rcvd_filp->f_dentry->d_inode,
379 FSFILT_OP_SETATTR, NULL);
380 if (IS_ERR(handle)) {
381 rc = PTR_ERR(handle);
382 CERROR("unable to start transaction: rc %d\n", rc);
384 ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
385 exp->exp_last_request_time = cfs_time_current_sec();
386 rc = fsfilt_add_journal_cb(obd, 0, handle,
387 target_client_add_cb,
388 class_export_cb_get(exp));
390 cfs_spin_lock(&exp->exp_lock);
391 exp->exp_need_sync = 1;
392 cfs_spin_unlock(&exp->exp_lock);
394 rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
396 sizeof(*ted->ted_lcd),
397 &off, rc /* sync if no cb */);
399 obt->obt_rcvd_filp->f_dentry->d_inode,
402 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
405 CERROR("error writing %s client idx %u: rc %d\n",
406 LAST_RCVD, ted->ted_lr_idx, rc);
413 static int filter_client_del(struct obd_export *exp)
415 struct tg_export_data *ted = &exp->exp_target_data;
416 struct obd_device_target *obt = &exp->exp_obd->u.obt;
417 struct lvfs_run_ctxt saved;
422 if (ted->ted_lcd == NULL)
425 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
426 if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
429 LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
431 off = ted->ted_lr_off;
433 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
434 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
436 /* Don't clear ted_lr_idx here as it is likely also unset. At worst
437 * we leak a client slot that will be cleaned on the next recovery. */
439 CERROR("%s: client idx %d has med_off %lld\n",
440 exp->exp_obd->obd_name, ted->ted_lr_idx, off);
441 GOTO(free, rc = -EINVAL);
444 /* Clear the bit _after_ zeroing out the client so we don't
445 race with filter_client_add and zero out new clients.*/
446 if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
447 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
452 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
453 /* Make sure the server's last_transno is up to date.
454 * This should be done before zeroing client slot so last_transno will
455 * be in server data or in client data in case of failure */
456 filter_update_server_data(exp->exp_obd);
458 cfs_mutex_lock(&ted->ted_lcd_lock);
459 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
460 rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
462 sizeof(*ted->ted_lcd), &off, 0);
463 cfs_mutex_unlock(&ted->ted_lcd_lock);
464 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
466 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
467 "zero out client %s at idx %u/%llu in %s, rc %d\n",
468 ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
475 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
476 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
477 struct filter_mod_data *fmd)
479 LASSERT_SPIN_LOCKED(&fed->fed_lock);
480 if (--fmd->fmd_refcount == 0) {
481 /* XXX when we have persistent reservations and the handle
482 * is stored herein we need to drop it here. */
483 fed->fed_mod_count--;
484 cfs_list_del(&fmd->fmd_list);
485 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
489 /* drop fmd reference, free it if last ref */
490 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
492 struct filter_export_data *fed;
497 fed = &exp->exp_filter_data;
498 cfs_spin_lock(&fed->fed_lock);
499 filter_fmd_put_nolock(fed, fmd); /* caller reference */
500 cfs_spin_unlock(&fed->fed_lock);
503 /* expire entries from the end of the list if there are too many
504 * or they are too old */
505 static void filter_fmd_expire_nolock(struct filter_obd *filter,
506 struct filter_export_data *fed,
507 struct filter_mod_data *keep)
509 struct filter_mod_data *fmd, *tmp;
511 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
515 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
516 fed->fed_mod_count < filter->fo_fmd_max_num)
519 cfs_list_del_init(&fmd->fmd_list);
520 filter_fmd_put_nolock(fed, fmd); /* list reference */
524 void filter_fmd_expire(struct obd_export *exp)
526 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
527 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
528 &exp->exp_filter_data, NULL);
529 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
532 /* find specified objid, group in export fmd list.
533 * caller must hold fed_lock and take fmd reference itself */
534 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
535 struct filter_export_data *fed,
536 obd_id objid, obd_seq group)
538 struct filter_mod_data *found = NULL, *fmd;
540 LASSERT_SPIN_LOCKED(&fed->fed_lock);
542 cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
543 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
545 cfs_list_del(&fmd->fmd_list);
546 cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
547 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
552 filter_fmd_expire_nolock(filter, fed, found);
557 /* Find fmd based on objid and group, or return NULL if not found. */
558 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
559 obd_id objid, obd_seq group)
561 struct filter_mod_data *fmd;
563 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
564 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
565 &exp->exp_filter_data, objid, group);
567 fmd->fmd_refcount++; /* caller reference */
568 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
573 /* Find fmd based on objid and group, or create a new one if none is found.
574 * It is possible for this function to return NULL under memory pressure,
575 * or if objid = 0 is passed (which will only cause old entries to expire).
576 * Currently this is not fatal because any fmd state is transient and
577 * may also be freed when it gets sufficiently old. */
578 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
579 obd_id objid, obd_seq group)
581 struct filter_export_data *fed = &exp->exp_filter_data;
582 struct filter_mod_data *found = NULL, *fmd_new = NULL;
584 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
586 cfs_spin_lock(&fed->fed_lock);
587 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
590 cfs_list_add_tail(&fmd_new->fmd_list,
592 fmd_new->fmd_id = objid;
593 fmd_new->fmd_gr = group;
594 fmd_new->fmd_refcount++; /* list reference */
596 fed->fed_mod_count++;
598 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
602 found->fmd_refcount++; /* caller reference */
603 found->fmd_expire = jiffies +
604 exp->exp_obd->u.filter.fo_fmd_max_age;
607 cfs_spin_unlock(&fed->fed_lock);
613 /* drop fmd list reference so it will disappear when last reference is put.
614 * This isn't so critical because it would in fact only affect the one client
615 * that is doing the unlink and at worst we have an stale entry referencing
616 * an object that should never be used again. */
617 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
619 struct filter_mod_data *found = NULL;
621 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
622 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
624 cfs_list_del_init(&found->fmd_list);
625 filter_fmd_put_nolock(&exp->exp_filter_data, found);
627 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
630 #define filter_fmd_drop(exp, objid, group)
633 /* remove all entries from fmd list */
634 static void filter_fmd_cleanup(struct obd_export *exp)
636 struct filter_export_data *fed = &exp->exp_filter_data;
637 struct filter_mod_data *fmd = NULL, *tmp;
639 cfs_spin_lock(&fed->fed_lock);
640 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
641 cfs_list_del_init(&fmd->fmd_list);
642 filter_fmd_put_nolock(fed, fmd);
644 cfs_spin_unlock(&fed->fed_lock);
647 static int filter_init_export(struct obd_export *exp)
652 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
653 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
655 cfs_spin_lock(&exp->exp_lock);
656 exp->exp_connecting = 1;
657 cfs_spin_unlock(&exp->exp_lock);
659 /* self-export doesn't need client data and ldlm initialization */
660 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
661 &exp->exp_client_uuid)))
664 rc = lut_client_alloc(exp);
666 rc = ldlm_init_export(exp);
668 CERROR("%s: Can't initialize export: rc %d\n",
669 exp->exp_obd->obd_name, rc);
674 static int filter_free_server_data(struct obd_device_target *obt)
676 lut_fini(NULL, obt->obt_lut);
677 OBD_FREE_PTR(obt->obt_lut);
681 /* assumes caller is already in kernel ctxt */
682 int filter_update_server_data(struct obd_device *obd)
684 struct file *filp = obd->u.obt.obt_rcvd_filp;
685 struct lr_server_data *lsd = class_server_data(obd);
690 CDEBUG(D_INODE, "server uuid : %s\n", lsd->lsd_uuid);
691 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
692 le64_to_cpu(lsd->lsd_last_transno));
693 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
694 le64_to_cpu(lsd->lsd_mount_count));
696 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
698 CERROR("error writing lr_server_data: rc = %d\n", rc);
703 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
706 struct filter_obd *filter = &obd->u.filter;
712 if (filter->fo_last_objid_files[group] == NULL) {
713 CERROR("Object seq "LPU64" not fully setup; not updating "
714 "last_objid\n", group);
718 CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
719 obd->obd_name, filter->fo_last_objids[group], group);
721 tmp = cpu_to_le64(filter->fo_last_objids[group]);
722 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
723 &tmp, sizeof(tmp), &off, force_sync);
725 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
729 extern int ost_handle(struct ptlrpc_request *req);
730 /* assumes caller has already in kernel ctxt */
731 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
733 struct filter_obd *filter = &obd->u.filter;
734 struct lr_server_data *lsd;
735 struct lsd_client_data *lcd = NULL;
736 struct inode *inode = filp->f_dentry->d_inode;
737 unsigned long last_rcvd_size = i_size_read(inode);
738 struct lu_target *lut;
745 /* ensure padding in the struct is the correct size */
746 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
747 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
748 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
749 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
751 /* allocate and initialize lu_target */
755 rc = lut_init(NULL, lut, obd, NULL);
758 lsd = class_server_data(obd);
759 if (last_rcvd_size == 0) {
760 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
762 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
763 lsd->lsd_last_transno = 0;
764 mount_count = lsd->lsd_mount_count = 0;
765 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
766 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
767 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
768 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
769 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
770 /* OBD_COMPAT_OST is set in filter_connect_internal when the
771 * MDS first connects and assigns the OST index number. */
772 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
775 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
777 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
781 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
782 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
783 "using the wrong disk %s. Were the "
784 "/dev/ assignments rearranged?\n",
785 obd->obd_uuid.uuid, lsd->lsd_uuid);
786 GOTO(err_lut, rc = -EINVAL);
788 mount_count = le64_to_cpu(lsd->lsd_mount_count);
789 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
791 /* Assume old last_rcvd format unless I_C_LR is set */
792 if (!(lsd->lsd_feature_incompat &
793 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
794 lsd->lsd_last_transno = lsd->lsd_compat14;
796 /* OBD_COMPAT_OST is set in filter_connect_internal when the
797 * MDS first connects and assigns the OST index number. */
798 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
802 if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
803 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
804 obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
805 ~FILTER_INCOMPAT_SUPP);
806 GOTO(err_lut, rc = -EINVAL);
808 if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
809 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
810 obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
811 ~FILTER_ROCOMPAT_SUPP);
812 /* Do something like remount filesystem read-only */
813 GOTO(err_lut, rc = -EINVAL);
816 start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
818 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
819 obd->obd_name, start_epoch);
820 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
821 obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
822 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
823 obd->obd_name, mount_count + 1);
824 CDEBUG(D_INODE, "%s: server data size: %u\n",
825 obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
826 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
827 obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
828 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
829 obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
830 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
831 obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
832 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
833 last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
834 (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
835 le16_to_cpu(lsd->lsd_client_size));
837 if (!obd->obd_replayable) {
838 CWARN("%s: recovery support OFF\n", obd->obd_name);
844 GOTO(err_client, rc = -ENOMEM);
846 for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
847 off < last_rcvd_size; cl_idx++) {
849 struct obd_export *exp;
850 struct filter_export_data *fed;
852 /* Don't assume off is incremented properly by
853 * fsfilt_read_record(), in case sizeof(*lcd)
854 * isn't the same as lsd->lsd_client_size. */
855 off = le32_to_cpu(lsd->lsd_client_start) +
856 cl_idx * le16_to_cpu(lsd->lsd_client_size);
857 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
859 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
860 LAST_RCVD, cl_idx, off, rc);
861 break; /* read error shouldn't cause startup to fail */
864 if (lcd->lcd_uuid[0] == '\0') {
865 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
870 check_lcd(obd->obd_name, cl_idx, lcd);
872 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
874 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
875 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
876 last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
878 /* These exports are cleaned up by filter_disconnect(), so they
879 * need to be set up like real exports as filter_connect() does.
881 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
883 if (PTR_ERR(exp) == -EALREADY) {
884 /* export already exists, zero out this one */
885 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
889 GOTO(err_client, rc = PTR_ERR(exp));
892 fed = &exp->exp_filter_data;
893 *fed->fed_ted.ted_lcd = *lcd;
894 fed->fed_group = 0; /* will be assigned at connect */
895 filter_export_stats_init(obd, exp, NULL);
896 rc = filter_client_add(obd, exp, cl_idx);
897 /* can't fail for existing client */
898 LASSERTF(rc == 0, "rc = %d\n", rc);
900 /* VBR: set export last committed */
901 exp->exp_last_committed = last_rcvd;
902 cfs_spin_lock(&exp->exp_lock);
903 exp->exp_connecting = 0;
904 exp->exp_in_recovery = 0;
905 cfs_spin_unlock(&exp->exp_lock);
906 obd->obd_max_recoverable_clients++;
907 class_export_put(exp);
909 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
910 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
914 obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
916 obd->u.obt.obt_mount_count = mount_count + 1;
917 obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
918 lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
920 /* save it, so mount count and last_transno is current */
921 rc = filter_update_server_data(obd);
923 GOTO(err_client, rc);
928 class_disconnect_exports(obd);
930 filter_free_server_data(&obd->u.obt);
934 static int filter_cleanup_groups(struct obd_device *obd)
936 struct filter_obd *filter = &obd->u.filter;
938 struct dentry *dentry;
942 if (filter->fo_dentry_O_groups != NULL) {
943 for (i = 0; i < filter->fo_group_count; i++) {
944 dentry = filter->fo_dentry_O_groups[i];
948 OBD_FREE(filter->fo_dentry_O_groups,
949 filter->fo_group_count *
950 sizeof(*filter->fo_dentry_O_groups));
951 filter->fo_dentry_O_groups = NULL;
953 if (filter->fo_last_objid_files != NULL) {
954 for (i = 0; i < filter->fo_group_count; i++) {
955 filp = filter->fo_last_objid_files[i];
959 OBD_FREE(filter->fo_last_objid_files,
960 filter->fo_group_count *
961 sizeof(*filter->fo_last_objid_files));
962 filter->fo_last_objid_files = NULL;
964 if (filter->fo_dentry_O_sub != NULL) {
965 for (i = 0; i < filter->fo_group_count; i++) {
966 for (j = 0; j < filter->fo_subdir_count; j++) {
967 dentry = filter->fo_dentry_O_sub[i].dentry[j];
972 OBD_FREE(filter->fo_dentry_O_sub,
973 filter->fo_group_count *
974 sizeof(*filter->fo_dentry_O_sub));
975 filter->fo_dentry_O_sub = NULL;
977 if (filter->fo_last_objids != NULL) {
978 OBD_FREE(filter->fo_last_objids,
979 filter->fo_group_count *
980 sizeof(*filter->fo_last_objids));
981 filter->fo_last_objids = NULL;
983 if (filter->fo_dentry_O != NULL) {
984 f_dput(filter->fo_dentry_O);
985 filter->fo_dentry_O = NULL;
990 static int filter_update_last_group(struct obd_device *obd, int group)
992 struct filter_obd *filter = &obd->u.filter;
993 struct file *filp = NULL;
994 int last_group = 0, rc;
998 if (group <= filter->fo_committed_group)
1001 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1005 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1009 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1011 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1015 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1016 obd->obd_name, last_group, group);
1020 /* must be sync: bXXXX */
1021 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1023 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1027 filter->fo_committed_group = group;
1030 filp_close(filp, 0);
1034 static int filter_read_group_internal(struct obd_device *obd, int group,
1037 struct filter_obd *filter = &obd->u.filter;
1038 __u64 *new_objids = NULL;
1039 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1040 struct dentry **new_groups = NULL;
1041 struct file **new_files = NULL;
1042 struct dentry *dentry;
1044 int old_count = filter->fo_group_count, rc, stage = 0, i;
1048 int len = group + 1;
1050 snprintf(name, 24, "%d", group);
1054 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1056 if (IS_ERR(dentry)) {
1057 CERROR("Cannot lookup expected object group %d: %ld\n",
1058 group, PTR_ERR(dentry));
1059 RETURN(PTR_ERR(dentry));
1062 dentry = simple_mkdir(filter->fo_dentry_O,
1063 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1064 if (IS_ERR(dentry)) {
1065 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1067 RETURN(PTR_ERR(dentry));
1072 snprintf(name, 24, "O/%d/LAST_ID", group);
1074 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1076 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1077 GOTO(cleanup, rc = PTR_ERR(filp));
1081 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1083 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1087 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1088 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1089 if (tmp_subdirs == NULL)
1090 GOTO(cleanup, rc = -ENOMEM);
1093 for (i = 0; i < filter->fo_subdir_count; i++) {
1095 snprintf(dir, sizeof(dir), "d%u", i);
1097 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1098 obd->u.obt.obt_vfsmnt,
1100 if (IS_ERR(tmp_subdirs->dentry[i])) {
1101 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1102 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1107 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1108 tmp_subdirs->dentry[i]);
1112 /* 'group' is an index; we need an array of length 'group + 1' */
1113 if (group + 1 > old_count) {
1114 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1115 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1116 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1117 OBD_ALLOC(new_files, len * sizeof(*new_files));
1119 if (new_objids == NULL || new_subdirs == NULL ||
1120 new_groups == NULL || new_files == NULL)
1121 GOTO(cleanup, rc = -ENOMEM);
1124 memcpy(new_objids, filter->fo_last_objids,
1125 old_count * sizeof(*new_objids));
1126 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1127 old_count * sizeof(*new_subdirs));
1128 memcpy(new_groups, filter->fo_dentry_O_groups,
1129 old_count * sizeof(*new_groups));
1130 memcpy(new_files, filter->fo_last_objid_files,
1131 old_count * sizeof(*new_files));
1133 OBD_FREE(filter->fo_last_objids,
1134 old_count * sizeof(*new_objids));
1135 OBD_FREE(filter->fo_dentry_O_sub,
1136 old_count * sizeof(*new_subdirs));
1137 OBD_FREE(filter->fo_dentry_O_groups,
1138 old_count * sizeof(*new_groups));
1139 OBD_FREE(filter->fo_last_objid_files,
1140 old_count * sizeof(*new_files));
1142 filter->fo_last_objids = new_objids;
1143 filter->fo_dentry_O_sub = new_subdirs;
1144 filter->fo_dentry_O_groups = new_groups;
1145 filter->fo_last_objid_files = new_files;
1146 filter->fo_group_count = len;
1149 filter->fo_dentry_O_groups[group] = dentry;
1150 filter->fo_last_objid_files[group] = filp;
1151 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1152 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1153 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1156 filter_update_last_group(obd, group);
1158 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1159 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1160 rc = filter_update_last_objid(obd, group, 1);
1164 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1165 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1166 obd->obd_name, group, last_objid);
1171 if (new_objids != NULL)
1172 OBD_FREE(new_objids, len * sizeof(*new_objids));
1173 if (new_subdirs != NULL)
1174 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1175 if (new_groups != NULL)
1176 OBD_FREE(new_groups, len * sizeof(*new_groups));
1177 if (new_files != NULL)
1178 OBD_FREE(new_files, len * sizeof(*new_files));
1180 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1181 for (i = 0; i < filter->fo_subdir_count; i++) {
1182 if (tmp_subdirs->dentry[i] != NULL)
1183 dput(tmp_subdirs->dentry[i]);
1185 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1188 filp_close(filp, 0);
1195 static int filter_read_groups(struct obd_device *obd, int last_group,
1198 struct filter_obd *filter = &obd->u.filter;
1199 int old_count, group, rc = 0;
1201 cfs_mutex_lock(&filter->fo_init_lock);
1202 old_count = filter->fo_group_count;
1203 for (group = old_count; group <= last_group; group++) {
1204 rc = filter_read_group_internal(obd, group, create);
1208 cfs_mutex_unlock(&filter->fo_init_lock);
1212 /* FIXME: object groups */
1213 static int filter_prep_groups(struct obd_device *obd)
1215 struct filter_obd *filter = &obd->u.filter;
1216 struct dentry *O_dentry;
1218 int last_group, rc = 0, cleanup_phase = 0;
1222 O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1224 CDEBUG(D_INODE, "%s: got/created O: %p\n", obd->obd_name, O_dentry);
1225 if (IS_ERR(O_dentry)) {
1226 rc = PTR_ERR(O_dentry);
1227 CERROR("%s: cannot open/create O: rc = %d\n", obd->obd_name,rc);
1230 filter->fo_dentry_O = O_dentry;
1231 cleanup_phase = 1; /* O_dentry */
1233 /* we have to initialize all groups before first connections from
1234 * clients because they may send create/destroy for any group -bzzz */
1235 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1237 CERROR("%s: cannot create LAST_GROUP: rc = %ld\n",
1238 obd->obd_name, PTR_ERR(filp));
1239 GOTO(cleanup, rc = PTR_ERR(filp));
1241 cleanup_phase = 2; /* filp */
1243 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1245 CERROR("%s: error reading LAST_GROUP: rc %d\n",
1251 last_group = FID_SEQ_OST_MDT0;
1253 CDEBUG(D_INODE, "%s: initialize group %u (max %u)\n", obd->obd_name,
1254 FID_SEQ_OST_MDT0, last_group);
1255 filter->fo_committed_group = last_group;
1256 rc = filter_read_groups(obd, last_group, 1);
1260 filp_close(filp, 0);
1264 switch (cleanup_phase) {
1266 filp_close(filp, 0);
1268 filter_cleanup_groups(obd);
1269 f_dput(filter->fo_dentry_O);
1270 filter->fo_dentry_O = NULL;
1278 /* setup the object store with correct subdirectories */
1279 static int filter_prep(struct obd_device *obd)
1281 struct lvfs_run_ctxt saved;
1282 struct filter_obd *filter = &obd->u.filter;
1284 struct inode *inode;
1288 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1289 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1290 if (!file || IS_ERR(file)) {
1292 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1296 obd->u.obt.obt_rcvd_filp = file;
1297 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1298 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1299 file->f_dentry->d_inode->i_mode);
1300 GOTO(err_filp, rc = -ENOENT);
1303 inode = file->f_dentry->d_parent->d_inode;
1304 /* We use i_op->unlink directly in filter_vfs_unlink() */
1305 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1306 CERROR("%s: filesystem does not support create/unlink ops\n",
1308 GOTO(err_filp, rc = -EOPNOTSUPP);
1311 rc = filter_init_server_data(obd, file);
1313 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1316 LASSERT(obd->u.obt.obt_lut);
1317 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1319 /* open and create health check io file*/
1320 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1323 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1325 GOTO(err_server_data, rc);
1327 filter->fo_obt.obt_health_check_filp = file;
1328 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1329 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1330 file->f_dentry->d_inode->i_mode);
1331 GOTO(err_health_check, rc = -ENOENT);
1333 rc = lvfs_check_io_health(obd, file);
1335 GOTO(err_health_check, rc);
1337 rc = filter_prep_groups(obd);
1339 GOTO(err_health_check, rc);
1341 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1346 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1347 CERROR("can't close %s after error\n", HEALTH_CHECK);
1348 filter->fo_obt.obt_health_check_filp = NULL;
1350 target_recovery_fini(obd);
1351 filter_free_server_data(&obd->u.obt);
1353 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1354 CERROR("can't close %s after error\n", LAST_RCVD);
1355 obd->u.obt.obt_rcvd_filp = NULL;
1359 /* cleanup the filter: write last used object id to status file */
1360 static void filter_post(struct obd_device *obd)
1362 struct lvfs_run_ctxt saved;
1363 struct filter_obd *filter = &obd->u.filter;
1366 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1367 * best to start a transaction with h_sync, because we removed this
1370 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1371 rc = filter_update_server_data(obd);
1373 CERROR("error writing server data: rc = %d\n", rc);
1375 for (i = 0; i < filter->fo_group_count; i++) {
1376 rc = filter_update_last_objid(obd, i,
1377 (i == filter->fo_group_count - 1));
1379 CERROR("error writing group %d lastobjid: rc = %d\n",
1383 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1384 obd->u.obt.obt_rcvd_filp = NULL;
1386 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1388 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1389 filter->fo_obt.obt_health_check_filp = NULL;
1391 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1393 filter_cleanup_groups(obd);
1394 filter_free_server_data(&obd->u.obt);
1395 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1397 filter_free_capa_keys(filter);
1398 cleanup_capa_hash(filter->fo_capa_hash);
1401 static void filter_set_last_id(struct filter_obd *filter,
1402 obd_id id, obd_seq group)
1404 LASSERT(group <= filter->fo_group_count);
1406 cfs_spin_lock(&filter->fo_objidlock);
1407 filter->fo_last_objids[group] = id;
1408 cfs_spin_unlock(&filter->fo_objidlock);
1411 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1414 LASSERT(group <= filter->fo_group_count);
1415 LASSERT(filter->fo_last_objids != NULL);
1417 /* FIXME: object groups */
1418 cfs_spin_lock(&filter->fo_objidlock);
1419 id = filter->fo_last_objids[group];
1420 cfs_spin_unlock(&filter->fo_objidlock);
1424 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1426 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1430 /* We never dget the object parent, so DON'T dput it either */
1431 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1433 struct filter_obd *filter = &obd->u.filter;
1434 struct filter_subdirs *subdirs;
1436 if (group >= filter->fo_group_count) /* FIXME: object groups */
1437 return ERR_PTR(-EBADF);
1439 if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1440 return filter->fo_dentry_O_groups[group];
1442 subdirs = &filter->fo_dentry_O_sub[group];
1443 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1446 /* We never dget the object parent, so DON'T dput it either */
1447 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1450 unsigned long now = jiffies;
1451 struct dentry *dparent = filter_parent(obd, group, objid);
1454 if (IS_ERR(dparent))
1456 if (dparent == NULL)
1457 return ERR_PTR(-ENOENT);
1459 rc = filter_lock_dentry(obd, dparent);
1460 fsfilt_check_slow(obd, now, "parent lock");
1461 return rc ? ERR_PTR(rc) : dparent;
1464 /* We never dget the object parent, so DON'T dput it either */
1465 static void filter_parent_unlock(struct dentry *dparent)
1467 UNLOCK_INODE_MUTEX(dparent->d_inode);
1470 /* How to get files, dentries, inodes from object id's.
1472 * If dir_dentry is passed, the caller has already locked the parent
1473 * appropriately for this operation (normally a write lock). If
1474 * dir_dentry is NULL, we do a read lock while we do the lookup to
1475 * avoid races with create/destroy and such changing the directory
1476 * internal to the filesystem code. */
1477 struct dentry *filter_fid2dentry(struct obd_device *obd,
1478 struct dentry *dir_dentry,
1479 obd_seq group, obd_id id)
1481 struct dentry *dparent = dir_dentry;
1482 struct dentry *dchild;
1487 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1488 obd->u.filter.fo_destroys_in_progress == 0) {
1489 /* don't fail lookups for orphan recovery, it causes
1490 * later LBUGs when objects still exist during precreate */
1491 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1492 RETURN(ERR_PTR(-ENOENT));
1495 CERROR("fatal: invalid object id 0\n");
1496 RETURN(ERR_PTR(-ESTALE));
1499 len = sprintf(name, LPU64, id);
1500 if (dir_dentry == NULL) {
1501 dparent = filter_parent_lock(obd, group, id);
1502 if (IS_ERR(dparent)) {
1503 CERROR("%s: error getting object "POSTID
1504 " parent: rc %ld\n", obd->obd_name,
1505 id, group, PTR_ERR(dparent));
1509 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1510 dparent->d_name.len, dparent->d_name.name, name);
1511 /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1512 dchild = lookup_one_len(name, dparent, len);
1513 if (dir_dentry == NULL)
1514 filter_parent_unlock(dparent);
1515 if (IS_ERR(dchild)) {
1516 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1517 obd->obd_name, id, group, PTR_ERR(dchild));
1521 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1522 CERROR("%s: got bad object "LPU64" inode %lu\n",
1523 obd->obd_name, id, dchild->d_inode->i_ino);
1525 RETURN(ERR_PTR(-ENOENT));
1528 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1529 name, dchild, atomic_read(&dchild->d_count));
1531 LASSERT(atomic_read(&dchild->d_count) > 0);
1536 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1537 obd_id group, struct lustre_handle *lockh)
1539 int flags = LDLM_AST_DISCARD_DATA, rc;
1540 struct ldlm_res_id res_id;
1541 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1544 osc_build_res_name(objid, group, &res_id);
1545 /* Tell the clients that the object is gone now and that they should
1546 * throw away any cached pages. */
1547 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1548 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1549 ldlm_completion_ast, NULL, NULL, 0, NULL,
1556 static void filter_fini_destroy(struct obd_device *obd,
1557 struct lustre_handle *lockh)
1559 if (lustre_handle_is_used(lockh))
1560 ldlm_lock_decref(lockh, LCK_PW);
1563 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1564 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1565 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1566 * reverse. Caller must take i_sem before starting the transaction and we
1567 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1568 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1569 struct vfsmount *mnt)
1574 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1575 * itself we already hold dir->i_mutex for child create/unlink ops */
1576 LASSERT(dentry->d_inode != NULL);
1577 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1578 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1582 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1583 GOTO(out, rc = -ENOENT);
1585 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1590 GOTO(out, rc = -EPERM);
1592 /* check_sticky() */
1593 if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1594 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1595 IS_IMMUTABLE(dentry->d_inode))
1596 GOTO(out, rc = -EPERM);
1598 /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
1599 ll_vfs_dq_init(dir);
1601 rc = ll_security_inode_unlink(dir, dentry, mnt);
1605 rc = dir->i_op->unlink(dir, dentry);
1607 /* need to drop i_mutex before we lose inode reference */
1608 UNLOCK_INODE_MUTEX(dentry->d_inode);
1615 /* Caller must hold LCK_PW on parent and push us into kernel context.
1616 * Caller must hold child i_mutex, we drop it always.
1617 * Caller is also required to ensure that dchild->d_inode exists. */
1618 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1619 obd_seq group, struct dentry *dparent,
1620 struct dentry *dchild)
1622 struct inode *inode = dchild->d_inode;
1625 /* There should be 2 references to the inode:
1626 * 1) taken by filter_prepare_destroy
1627 * 2) taken by filter_destroy */
1628 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1629 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1630 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1631 (unsigned long)inode->i_nlink,
1632 atomic_read(&inode->i_count));
1635 rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1637 CERROR("error unlinking objid %.*s: rc %d\n",
1638 dchild->d_name.len, dchild->d_name.name, rc);
1642 struct filter_intent_args {
1643 struct ldlm_lock **victim;
1648 static enum interval_iter filter_intent_cb(struct interval_node *n,
1651 struct ldlm_interval *node = (struct ldlm_interval *)n;
1652 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1653 __u64 size = arg->size;
1654 struct ldlm_lock **v = arg->victim;
1655 struct ldlm_lock *lck;
1657 /* If the interval is lower than the current file size,
1659 if (interval_high(n) <= size)
1660 return INTERVAL_ITER_STOP;
1662 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1663 /* Don't send glimpse ASTs to liblustre clients.
1664 * They aren't listening for them, and they do
1665 * entirely synchronous I/O anyways. */
1666 if (lck->l_export == NULL ||
1667 lck->l_export->exp_libclient == 1)
1670 if (*arg->liblustre)
1671 *arg->liblustre = 0;
1674 *v = LDLM_LOCK_GET(lck);
1675 } else if ((*v)->l_policy_data.l_extent.start <
1676 lck->l_policy_data.l_extent.start) {
1677 LDLM_LOCK_RELEASE(*v);
1678 *v = LDLM_LOCK_GET(lck);
1681 /* the same policy group - every lock has the
1682 * same extent, so needn't do it any more */
1686 return INTERVAL_ITER_CONT;
1689 static int filter_intent_policy(struct ldlm_namespace *ns,
1690 struct ldlm_lock **lockp, void *req_cookie,
1691 ldlm_mode_t mode, int flags, void *data)
1693 struct ptlrpc_request *req = req_cookie;
1694 struct ldlm_lock *lock = *lockp, *l = NULL;
1695 struct ldlm_resource *res = lock->l_resource;
1696 ldlm_processing_policy policy;
1697 struct ost_lvb *res_lvb, *reply_lvb;
1698 struct ldlm_reply *rep;
1700 int idx, rc, tmpflags = 0, only_liblustre = 1;
1701 struct ldlm_interval_tree *tree;
1702 struct filter_intent_args arg;
1703 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1704 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1705 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1708 policy = ldlm_get_processing_policy(res);
1709 LASSERT(policy != NULL);
1710 LASSERT(req != NULL);
1712 rc = lustre_pack_reply(req, 3, repsize, NULL);
1714 RETURN(req->rq_status = rc);
1716 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1717 LASSERT(rep != NULL);
1719 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1720 sizeof(*reply_lvb));
1721 LASSERT(reply_lvb != NULL);
1723 //fixup_handle_for_resent_req(req, lock, &lockh);
1725 /* Call the extent policy function to see if our request can be
1726 * granted, or is blocked.
1727 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1728 * lock, and should not be granted if the lock will be blocked.
1731 if (flags & LDLM_FL_BLOCK_NOWAIT) {
1732 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_AGL_DELAY, 5);
1734 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_AGL_NOLOCK))
1735 RETURN(ELDLM_LOCK_ABORTED);
1738 LASSERT(ns == ldlm_res_to_ns(res));
1740 rc = policy(lock, &tmpflags, 0, &err, NULL);
1741 check_res_locked(res);
1743 /* The lock met with no resistance; we're finished. */
1744 if (rc == LDLM_ITER_CONTINUE) {
1745 /* do not grant locks to the liblustre clients: they cannot
1746 * handle ASTs robustly. We need to do this while still
1747 * holding lr_lock to avoid the lock remaining on the res_link
1748 * list (and potentially being added to l_pending_list by an
1749 * AST) when we are going to drop this lock ASAP. */
1750 if (lock->l_export->exp_libclient ||
1751 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1752 ldlm_resource_unlink_lock(lock);
1753 err = ELDLM_LOCK_ABORTED;
1755 err = ELDLM_LOCK_REPLACED;
1759 } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1760 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1761 * callback for glimpse size. The real size user will trigger
1762 * the glimpse callback when necessary. */
1764 RETURN(ELDLM_LOCK_ABORTED);
1767 /* Do not grant any lock, but instead send GL callbacks. The extent
1768 * policy nicely created a list of all PW locks for us. We will choose
1769 * the highest of those which are larger than the size in the LVB, if
1770 * any, and perform a glimpse callback. */
1771 res_lvb = res->lr_lvb_data;
1772 LASSERT(res_lvb != NULL);
1773 *reply_lvb = *res_lvb;
1776 * lr_lock guarantees that no new locks are granted, and,
1777 * therefore, that res->lr_lvb_data cannot increase beyond the
1778 * end of already granted lock. As a result, it is safe to
1779 * check against "stale" reply_lvb->lvb_size value without
1780 * res->lr_lvb_mutex.
1782 arg.size = reply_lvb->lvb_size;
1784 arg.liblustre = &only_liblustre;
1785 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1786 tree = &res->lr_itree[idx];
1787 if (tree->lit_mode == LCK_PR)
1790 interval_iterate_reverse(tree->lit_root,
1791 filter_intent_cb, &arg);
1795 /* There were no PW locks beyond the size in the LVB; finished. */
1797 if (only_liblustre) {
1798 /* If we discovered a liblustre client with a PW lock,
1799 * however, the LVB may be out of date! The LVB is
1800 * updated only on glimpse (which we don't do for
1801 * liblustre clients) and cancel (which the client
1802 * obviously has not yet done). So if it has written
1803 * data but kept the lock, the LVB is stale and needs
1804 * to be updated from disk.
1806 * Of course, this will all disappear when we switch to
1807 * taking liblustre locks on the OST. */
1808 ldlm_res_lvbo_update(res, NULL, 1);
1810 RETURN(ELDLM_LOCK_ABORTED);
1814 * This check is for lock taken in filter_prepare_destroy() that does
1815 * not have l_glimpse_ast set. So the logic is: if there is a lock
1816 * with no l_glimpse_ast set, this object is being destroyed already.
1818 * Hence, if you are grabbing DLM locks on the server, always set
1819 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1821 if (l->l_glimpse_ast == NULL) {
1822 /* We are racing with unlink(); just return -ENOENT */
1823 rep->lock_policy_res1 = -ENOENT;
1827 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1828 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1831 *reply_lvb = *res_lvb;
1835 LDLM_LOCK_RELEASE(l);
1837 RETURN(ELDLM_LOCK_ABORTED);
1841 * per-obd_device iobuf pool.
1843 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1844 * allocations in write-path has to be minimized (see bug 5137).
1846 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1847 * OST threads (see ost_thread_{init,done}()).
1849 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1850 * at the OST layer there are only (potentially) multiple obd_device of type
1851 * unknown at the time of OST thread creation.
1853 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1854 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1855 * it's very own iobuf.
1859 * filter_kiobuf_pool_init()
1861 * filter_kiobuf_pool_done()
1863 * filter_iobuf_get()
1865 * operate on this array. They are "generic" in a sense that they don't depend
1866 * on actual type of iobuf's (the latter depending on Linux kernel version).
1870 * destroy pool created by filter_iobuf_pool_init
1872 static void filter_iobuf_pool_done(struct filter_obd *filter)
1874 struct filter_iobuf **pool;
1879 pool = filter->fo_iobuf_pool;
1881 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1882 if (pool[i] != NULL)
1883 filter_free_iobuf(pool[i]);
1885 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1886 filter->fo_iobuf_pool = NULL;
1891 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1893 struct filter_obd *filter = &obd->u.filter;
1894 struct sptlrpc_rule_set tmp_rset;
1897 sptlrpc_rule_set_init(&tmp_rset);
1898 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1900 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1905 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1907 cfs_write_lock(&filter->fo_sptlrpc_lock);
1908 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1909 filter->fo_sptlrpc_rset = tmp_rset;
1910 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1916 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1918 static int filter_iobuf_pool_init(struct filter_obd *filter)
1925 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1927 if (filter->fo_iobuf_pool == NULL)
1930 filter->fo_iobuf_count = OSS_THREADS_MAX;
1935 /* Return iobuf allocated for @thread_id. We don't know in advance how
1936 * many threads there will be so we allocate a large empty array and only
1937 * fill in those slots that are actually in use.
1938 * If we haven't allocated a pool entry for this thread before, do so now. */
1939 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1941 int thread_id = (oti && oti->oti_thread) ?
1942 oti->oti_thread->t_id : -1;
1943 struct filter_iobuf *pool = NULL;
1944 struct filter_iobuf **pool_place = NULL;
1946 if (thread_id >= 0) {
1947 LASSERT(thread_id < filter->fo_iobuf_count);
1948 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1951 if (unlikely(pool == NULL)) {
1952 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1953 PTLRPC_MAX_BRW_PAGES);
1954 if (pool_place != NULL)
1961 /* mount the file system (secretly). lustre_cfg parameters are:
1964 * 3 = flags: failover=f, failout=n
1967 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1970 struct filter_obd *filter = &obd->u.filter;
1971 struct vfsmount *mnt;
1972 struct lustre_mount_info *lmi;
1973 struct obd_uuid uuid;
1977 struct request_queue *q;
1981 if (lcfg->lcfg_bufcount < 3 ||
1982 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1983 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1986 lmi = server_get_mount(obd->obd_name);
1988 /* We already mounted in lustre_fill_super.
1989 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1990 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1992 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1994 /* old path - used by lctl */
1995 CERROR("Using old MDS mount method\n");
1996 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1997 MS_NOATIME|MS_NODIRATIME,
1998 lustre_cfg_string(lcfg, 1), option);
2001 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2002 lustre_cfg_string(lcfg, 1), rc);
2006 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2008 if (IS_ERR(obd->obd_fsops))
2009 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2011 rc = filter_iobuf_pool_init(filter);
2015 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2016 CERROR("%s: Underlying device is marked as read-only. "
2017 "Setup failed\n", obd->obd_name);
2018 GOTO(err_ops, rc = -EROFS);
2021 /* failover is the default */
2022 obd->obd_replayable = 1;
2024 /* disable connection until configuration finishes */
2025 obd->obd_no_conn = 1;
2027 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2028 str = lustre_cfg_string(lcfg, 3);
2029 if (strchr(str, 'n')) {
2030 CWARN("%s: recovery disabled\n", obd->obd_name);
2031 obd->obd_replayable = 0;
2035 obd->u.obt.obt_magic = OBT_MAGIC;
2036 obd->u.obt.obt_vfsmnt = mnt;
2037 obd->u.obt.obt_sb = mnt->mnt_sb;
2038 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2039 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2041 fsfilt_setup(obd, obd->u.obt.obt_sb);
2043 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2044 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2045 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2046 obd->obd_lvfs_ctxt.fs = get_ds();
2047 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2049 cfs_mutex_init(&filter->fo_init_lock);
2050 filter->fo_committed_group = 0;
2051 filter->fo_destroys_in_progress = 0;
2052 for (i = 0; i < 32; i++)
2053 cfs_mutex_init(&filter->fo_create_locks[i]);
2055 cfs_spin_lock_init(&filter->fo_objidlock);
2056 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2057 cfs_mutex_init(&filter->fo_alloc_lock);
2058 init_brw_stats(&filter->fo_filter_stats);
2059 cfs_spin_lock_init(&filter->fo_flags_lock);
2060 filter->fo_read_cache = 1; /* enable read-only cache by default */
2061 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2062 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2063 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2064 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2065 filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2066 filter_slc_set(filter); /* initialize sync on lock cancel */
2068 rc = filter_prep(obd);
2072 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2073 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2075 filter->fo_fl_oss_capa = 1;
2077 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2078 filter->fo_capa_hash = init_capa_hash();
2079 if (filter->fo_capa_hash == NULL)
2080 GOTO(err_post, rc = -ENOMEM);
2082 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2083 obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2084 LDLM_NAMESPACE_SERVER,
2085 LDLM_NAMESPACE_GREEDY,
2087 if (obd->obd_namespace == NULL)
2088 GOTO(err_post, rc = -ENOMEM);
2089 obd->obd_namespace->ns_lvbp = obd;
2090 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2091 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2093 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2094 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2096 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2098 CERROR("failed to setup llogging subsystems\n");
2102 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2103 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2104 /* do this after llog being initialized */
2105 filter_adapt_sptlrpc_conf(obd, 1);
2107 rc = lquota_setup(filter_quota_interface_ref, obd);
2111 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2112 if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2113 queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2114 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2115 "for larger I/O requests: max_sectors = %u "
2116 "could be up to max_hw_sectors=%u\n",
2117 obd->obd_name, mnt->mnt_sb->s_id,
2118 queue_max_sectors(q), queue_max_hw_sectors(q));
2120 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2121 if (uuid_ptr != NULL) {
2122 class_uuid_unparse(uuid_ptr, &uuid);
2128 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2129 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2130 obd->obd_name, label ?: str, lmi ? "on " : "",
2131 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2132 obd->obd_replayable ? "enabled" : "disabled");
2139 fsfilt_put_ops(obd->obd_fsops);
2140 filter_iobuf_pool_done(filter);
2142 server_put_mount(obd->obd_name, mnt);
2143 obd->u.obt.obt_sb = 0;
2147 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2149 struct lprocfs_static_vars lvars;
2150 cfs_proc_dir_entry_t *entry;
2156 CLASSERT(offsetof(struct obd_device, u.obt) ==
2157 offsetof(struct obd_device, u.filter.fo_obt));
2159 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2162 /* lprocfs must be setup before the filter so state can be safely added
2163 * to /proc incrementally as the filter is setup */
2164 lprocfs_filter_init_vars(&lvars);
2165 rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2167 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2172 rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2174 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2176 GOTO(obd_cleanup, rc);
2179 /* Init obdfilter private stats here */
2180 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2181 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2182 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2183 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2184 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2185 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2186 "get_page", "usec");
2187 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2188 LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2189 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2190 LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2191 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2192 LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2193 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2194 LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2196 rc = lproc_filter_attach_seqstat(obd);
2198 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2199 GOTO(free_obd_stats, rc);
2202 entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2203 if (IS_ERR(entry)) {
2204 rc = PTR_ERR(entry);
2205 CERROR("%s: error %d setting up lprocfs for %s\n",
2206 obd->obd_name, rc, "exports");
2207 GOTO(free_obd_stats, rc);
2209 obd->obd_proc_exports_entry = entry;
2211 entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2212 lprocfs_nid_stats_clear_read,
2213 lprocfs_nid_stats_clear_write, obd, NULL);
2214 if (IS_ERR(entry)) {
2215 rc = PTR_ERR(entry);
2216 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2218 GOTO(free_obd_stats, rc);
2221 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2222 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2224 GOTO(remove_entry_clear, rc = -ENOMEM);
2225 addr = (unsigned long)cfs_page_address(page);
2226 clear_page((void *)addr);
2227 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2228 LUSTRE_CFG_BUFLEN(lcfg, 4));
2229 rc = filter_common_setup(obd, lcfg, (void *)addr);
2230 OBD_PAGE_FREE(page);
2232 CERROR("%s: filter_common_setup failed: %d.\n",
2234 GOTO(remove_entry_clear, rc);
2240 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2242 lprocfs_free_obd_stats(obd);
2244 lprocfs_obd_cleanup(obd);
2248 static struct llog_operations filter_mds_ost_repl_logops;
2250 static struct llog_operations filter_size_orig_logops = {
2251 .lop_setup = llog_obd_origin_setup,
2252 .lop_cleanup = llog_obd_origin_cleanup,
2253 .lop_add = llog_obd_origin_add
2256 static int filter_olg_fini(struct obd_llog_group *olg)
2258 struct llog_ctxt *ctxt;
2259 int rc = 0, rc2 = 0;
2262 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2264 rc = llog_cleanup(ctxt);
2266 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2268 rc2 = llog_cleanup(ctxt);
2273 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2275 rc2 = llog_cleanup(ctxt);
2284 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2285 struct obd_device *tgt)
2290 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2291 &filter_mds_ost_repl_logops);
2295 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2296 &filter_size_orig_logops);
2302 filter_olg_fini(olg);
2307 * Init the default olg, which is embeded in the obd_device, for filter.
2310 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2311 struct obd_device *tgt)
2313 struct filter_obd *filter = &obd->u.filter;
2314 struct llog_ctxt *ctxt;
2318 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2319 if (!filter->fo_lcm)
2322 filter_mds_ost_repl_logops = llog_client_ops;
2323 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2324 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2325 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2327 rc = filter_olg_init(obd, olg, tgt);
2329 GOTO(cleanup_lcm, rc);
2331 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2334 GOTO(cleanup_olg, rc);
2336 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2338 CERROR("Can't get ctxt for %p:%x\n", olg,
2339 LLOG_MDS_OST_REPL_CTXT);
2340 GOTO(cleanup_olg, rc = -ENODEV);
2342 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2343 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2344 llog_ctxt_put(ctxt);
2348 filter_olg_fini(olg);
2350 llog_recov_thread_fini(filter->fo_lcm, 1);
2351 filter->fo_lcm = NULL;
2356 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2357 struct obd_device *tgt, int *index)
2359 struct filter_obd *filter = &obd->u.filter;
2360 struct llog_ctxt *ctxt;
2364 LASSERT(olg != NULL);
2365 if (olg == &obd->obd_olg)
2366 return filter_default_olg_init(obd, olg, tgt);
2368 LASSERT(filter->fo_lcm != NULL);
2369 rc = filter_olg_init(obd, olg, tgt);
2372 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2374 CERROR("Can't get ctxt for %p:%x\n", olg,
2375 LLOG_MDS_OST_REPL_CTXT);
2376 filter_olg_fini(olg);
2379 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2380 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2381 llog_ctxt_put(ctxt);
2385 static int filter_llog_finish(struct obd_device *obd, int count)
2387 struct filter_obd *filter = &obd->u.filter;
2388 struct llog_ctxt *ctxt;
2391 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2394 * Make sure that no cached llcds left in recov_thread.
2395 * We actually do sync in disconnect time, but disconnect
2396 * may not come being marked rq_no_resend = 1.
2398 llog_sync(ctxt, NULL);
2401 * Balance class_import_get() in llog_receptor_accept().
2402 * This is safe to do, as llog is already synchronized
2403 * and its import may go.
2405 cfs_mutex_lock(&ctxt->loc_mutex);
2406 if (ctxt->loc_imp) {
2407 class_import_put(ctxt->loc_imp);
2408 ctxt->loc_imp = NULL;
2410 cfs_mutex_unlock(&ctxt->loc_mutex);
2411 llog_ctxt_put(ctxt);
2414 if (filter->fo_lcm) {
2415 cfs_mutex_lock(&ctxt->loc_mutex);
2416 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2417 filter->fo_lcm = NULL;
2418 cfs_mutex_unlock(&ctxt->loc_mutex);
2420 RETURN(filter_olg_fini(&obd->obd_olg));
2424 * Find the group llog according to group index in the llog group list.
2426 static struct obd_llog_group *
2427 filter_find_olg_internal(struct filter_obd *filter, int group)
2429 struct obd_llog_group *olg;
2431 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2432 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2433 if (olg->olg_seq == group)
2440 * Find the group llog according to group index on the filter
2442 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2444 struct obd_llog_group *olg = NULL;
2445 struct filter_obd *filter;
2447 filter = &obd->u.filter;
2449 if (group == FID_SEQ_LLOG)
2450 RETURN(&obd->obd_olg);
2452 cfs_spin_lock(&filter->fo_llog_list_lock);
2453 olg = filter_find_olg_internal(filter, group);
2454 cfs_spin_unlock(&filter->fo_llog_list_lock);
2459 * Find the llog_group of the filter according to the group. If it can not
2460 * find, create the llog_group, which only happens when mds is being synced
2463 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2465 struct obd_llog_group *olg = NULL, *olg_new = NULL;
2466 struct filter_obd *filter;
2469 filter = &obd->u.filter;
2471 if (group == FID_SEQ_LLOG)
2472 RETURN(&obd->obd_olg);
2474 OBD_ALLOC_PTR(olg_new);
2475 if (olg_new == NULL)
2476 RETURN(ERR_PTR(-ENOMEM));
2478 cfs_spin_lock(&filter->fo_llog_list_lock);
2479 olg = filter_find_olg_internal(filter, group);
2481 if (olg->olg_initializing) {
2482 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2484 GOTO(out_unlock, olg);
2487 /* set as the newly allocated one */
2492 llog_group_init(olg, group);
2493 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2494 olg->olg_initializing = 1;
2495 cfs_spin_unlock(&filter->fo_llog_list_lock);
2497 rc = obd_llog_init(obd, olg, obd, NULL);
2499 cfs_spin_lock(&filter->fo_llog_list_lock);
2500 cfs_list_del(&olg->olg_list);
2501 cfs_spin_unlock(&filter->fo_llog_list_lock);
2503 GOTO(out, olg = ERR_PTR(-ENOMEM));
2505 cfs_spin_lock(&filter->fo_llog_list_lock);
2506 olg->olg_initializing = 0;
2507 cfs_spin_unlock(&filter->fo_llog_list_lock);
2508 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2509 obd->obd_name, group, olg);
2514 cfs_spin_unlock(&filter->fo_llog_list_lock);
2516 OBD_FREE_PTR(olg_new);
2520 static int filter_llog_connect(struct obd_export *exp,
2521 struct llogd_conn_body *body)
2523 struct obd_device *obd = exp->exp_obd;
2524 struct llog_ctxt *ctxt;
2525 struct obd_llog_group *olg;
2529 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2530 obd->obd_name, body->lgdc_logid.lgl_oid,
2531 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2533 olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2535 CERROR(" %s: can not find olg of group %d\n",
2536 obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2539 llog_group_set_export(olg, exp);
2541 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2542 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2543 body->lgdc_ctxt_idx);
2545 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2546 obd->obd_name, body->lgdc_logid.lgl_oid,
2547 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2549 cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2550 obd->u.filter.fo_mds_ost_sync = 1;
2551 cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2552 rc = llog_connect(ctxt, &body->lgdc_logid,
2553 &body->lgdc_gen, NULL);
2554 llog_ctxt_put(ctxt);
2556 CERROR("failed to connect rc %d idx %d\n", rc,
2557 body->lgdc_ctxt_idx);
2562 static int filter_llog_preclean(struct obd_device *obd)
2564 struct obd_llog_group *olg, *tmp;
2565 struct filter_obd *filter;
2566 cfs_list_t remove_list;
2570 rc = obd_llog_finish(obd, 0);
2572 CERROR("failed to cleanup llogging subsystem\n");
2574 filter = &obd->u.filter;
2575 CFS_INIT_LIST_HEAD(&remove_list);
2577 cfs_spin_lock(&filter->fo_llog_list_lock);
2578 while (!cfs_list_empty(&filter->fo_llog_list)) {
2579 olg = cfs_list_entry(filter->fo_llog_list.next,
2580 struct obd_llog_group, olg_list);
2581 cfs_list_del(&olg->olg_list);
2582 cfs_list_add(&olg->olg_list, &remove_list);
2584 cfs_spin_unlock(&filter->fo_llog_list_lock);
2586 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2587 cfs_list_del_init(&olg->olg_list);
2588 rc = filter_olg_fini(olg);
2590 CERROR("failed to cleanup llogging subsystem for %u\n",
2598 static int filter_precleanup(struct obd_device *obd,
2599 enum obd_cleanup_stage stage)
2605 case OBD_CLEANUP_EARLY:
2607 case OBD_CLEANUP_EXPORTS:
2608 /* Stop recovery before namespace cleanup. */
2609 target_recovery_fini(obd);
2611 obd_exports_barrier(obd);
2612 obd_zombie_barrier();
2614 rc = filter_llog_preclean(obd);
2615 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2616 lprocfs_free_per_client_stats(obd);
2617 lprocfs_obd_cleanup(obd);
2618 lprocfs_free_obd_stats(obd);
2619 lquota_cleanup(filter_quota_interface_ref, obd);
2625 static int filter_cleanup(struct obd_device *obd)
2627 struct filter_obd *filter = &obd->u.filter;
2631 LCONSOLE_WARN("%s: shutting down for failover; client state "
2632 "will be preserved.\n", obd->obd_name);
2634 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2635 obd->obd_namespace = NULL;
2637 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2639 if (obd->u.obt.obt_sb == NULL)
2644 ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2645 shrink_dcache_sb(obd->u.obt.obt_sb);
2647 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2648 obd->u.obt.obt_sb = NULL;
2650 fsfilt_put_ops(obd->obd_fsops);
2652 filter_iobuf_pool_done(filter);
2654 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2659 static int filter_connect_internal(struct obd_export *exp,
2660 struct obd_connect_data *data,
2663 struct filter_export_data *fed = &exp->exp_filter_data;
2668 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2669 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2670 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2671 data->ocd_connect_flags, data->ocd_version,
2672 data->ocd_grant, data->ocd_index);
2674 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2675 CWARN("!!! This export (nid %s) used object group %d "
2676 "earlier; now it's trying to use group %d! This could "
2677 "be a bug in the MDS. Please report to "
2678 "http://bugs.whamcloud.com/\n",
2679 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2682 fed->fed_group = data->ocd_group;
2684 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2685 exp->exp_connect_flags = data->ocd_connect_flags;
2686 data->ocd_version = LUSTRE_VERSION_CODE;
2688 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2689 if (data->ocd_connect_flags & OBD_CONNECT_MDS)
2690 CWARN("%s: Received MDS connection for group %u\n",
2691 exp->exp_obd->obd_name, data->ocd_group);
2692 else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
2695 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2696 struct filter_obd *filter = &exp->exp_obd->u.filter;
2697 obd_size left, want;
2699 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2700 left = filter_grant_space_left(exp);
2701 want = data->ocd_grant;
2702 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2703 data->ocd_grant = fed->fed_grant;
2704 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2706 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2707 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2708 exp->exp_client_uuid.uuid, exp,
2709 data->ocd_grant, want, left);
2711 filter->fo_tot_granted_clients ++;
2714 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2715 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2716 int index = le32_to_cpu(lsd->lsd_ost_index);
2718 if (!(lsd->lsd_feature_compat &
2719 cpu_to_le32(OBD_COMPAT_OST))) {
2720 /* this will only happen on the first connect */
2721 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2722 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2723 /* sync is not needed here as filter_client_add will
2724 * set exp_need_sync flag */
2725 filter_update_server_data(exp->exp_obd);
2726 } else if (index != data->ocd_index) {
2727 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2728 " %u doesn't match actual OST index"
2729 " %u in last_rcvd file, bad "
2731 obd_export_nid2str(exp), index,
2735 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2736 * FIXME: We don't strictly need the COMPAT flag for that,
2737 * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2738 * FIXME: We needed it for the index, as index 0 is valid. */
2741 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2742 data->ocd_brw_size = 65536;
2743 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2744 data->ocd_brw_size = min(data->ocd_brw_size,
2745 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2746 if (data->ocd_brw_size == 0) {
2747 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2748 " ocd_version: %x ocd_grant: %d ocd_index: %u "
2749 "ocd_brw_size is unexpectedly zero, "
2750 "network data corruption?"
2751 "Refusing connection of this client\n",
2752 exp->exp_obd->obd_name,
2753 exp->exp_client_uuid.uuid,
2754 exp, data->ocd_connect_flags, data->ocd_version,
2755 data->ocd_grant, data->ocd_index);
2760 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2761 __u32 cksum_types = data->ocd_cksum_types;
2763 /* The client set in ocd_cksum_types the checksum types it
2764 * supports. We have to mask off the algorithms that we don't
2766 data->ocd_cksum_types &= cksum_types_supported();
2768 /* 1.6.4- only support CRC32 and didn't set ocd_cksum_types */
2769 if (unlikely(data->ocd_cksum_types == 0))
2770 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2772 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2773 "%x\n", exp->exp_obd->obd_name,
2774 obd_export_nid2str(exp), cksum_types,
2775 data->ocd_cksum_types);
2777 /* This client does not support OBD_CONNECT_CKSUM
2778 * fall back to CRC32 */
2779 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2780 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2781 exp->exp_obd->obd_name,
2782 obd_export_nid2str(exp));
2785 if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2786 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2791 static int filter_reconnect(const struct lu_env *env,
2792 struct obd_export *exp, struct obd_device *obd,
2793 struct obd_uuid *cluuid,
2794 struct obd_connect_data *data,
2800 if (exp == NULL || obd == NULL || cluuid == NULL)
2803 rc = filter_connect_internal(exp, data, 1);
2805 filter_export_stats_init(obd, exp, localdata);
2810 static int filter_connect(const struct lu_env *env,
2811 struct obd_export **exp, struct obd_device *obd,
2812 struct obd_uuid *cluuid,
2813 struct obd_connect_data *data, void *localdata)
2815 struct lvfs_run_ctxt saved;
2816 struct lustre_handle conn = { 0 };
2817 struct obd_export *lexp;
2821 if (exp == NULL || obd == NULL || cluuid == NULL)
2824 rc = class_connect(&conn, obd, cluuid);
2827 lexp = class_conn2export(&conn);
2828 LASSERT(lexp != NULL);
2830 rc = filter_connect_internal(lexp, data, 0);
2834 filter_export_stats_init(obd, lexp, localdata);
2835 if (obd->obd_replayable) {
2836 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2838 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2839 rc = filter_client_add(obd, lexp, -1);
2844 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2845 rc = filter_read_groups(obd, data->ocd_group, 1);
2846 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2848 CERROR("can't read group %u\n", data->ocd_group);
2856 class_disconnect(lexp);
2865 /* Do extra sanity checks for grant accounting. We do this at connect,
2866 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2867 * always get rid of it or turn it off when we know accounting is good. */
2868 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2870 struct filter_export_data *fed;
2871 struct obd_export *exp;
2872 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2873 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2874 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2876 if (cfs_list_empty(&obd->obd_exports))
2879 /* We don't want to do this for large machines that do lots of
2880 mounts or unmounts. It burns... */
2881 if (obd->obd_num_exports > 100)
2884 cfs_spin_lock(&obd->obd_osfs_lock);
2885 cfs_spin_lock(&obd->obd_dev_lock);
2886 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2888 fed = &exp->exp_filter_data;
2889 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2892 if (maxsize > 0) { /* we may not have done a statfs yet */
2893 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2894 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2895 exp->exp_client_uuid.uuid, exp,
2896 fed->fed_grant, fed->fed_pending, maxsize);
2897 LASSERTF(fed->fed_dirty <= maxsize,
2898 "%s: cli %s/%p %ld > "LPU64"\n", func,
2899 exp->exp_client_uuid.uuid, exp,
2900 fed->fed_dirty, maxsize);
2903 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2904 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2905 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2907 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2908 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2909 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2910 tot_granted += fed->fed_grant + fed->fed_pending;
2911 tot_pending += fed->fed_pending;
2912 tot_dirty += fed->fed_dirty;
2914 fo_tot_granted = obd->u.filter.fo_tot_granted;
2915 fo_tot_pending = obd->u.filter.fo_tot_pending;
2916 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2917 cfs_spin_unlock(&obd->obd_dev_lock);
2918 cfs_spin_unlock(&obd->obd_osfs_lock);
2920 /* Do these assertions outside the spinlocks so we don't kill system */
2921 if (tot_granted != fo_tot_granted)
2922 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2923 func, tot_granted, fo_tot_granted);
2924 if (tot_pending != fo_tot_pending)
2925 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2926 func, tot_pending, fo_tot_pending);
2927 if (tot_dirty != fo_tot_dirty)
2928 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2929 func, tot_dirty, fo_tot_dirty);
2930 if (tot_pending > tot_granted)
2931 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2932 func, tot_pending, tot_granted);
2933 if (tot_granted > maxsize)
2934 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2935 func, tot_granted, maxsize);
2936 if (tot_dirty > maxsize)
2937 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2938 func, tot_dirty, maxsize);
2941 /* Remove this client from the grant accounting totals. We also remove
2942 * the export from the obd device under the osfs and dev locks to ensure
2943 * that the filter_grant_sanity_check() calculations are always valid.
2944 * The client should do something similar when it invalidates its import. */
2945 static void filter_grant_discard(struct obd_export *exp)
2947 struct obd_device *obd = exp->exp_obd;
2948 struct filter_obd *filter = &obd->u.filter;
2949 struct filter_export_data *fed = &exp->exp_filter_data;
2951 cfs_spin_lock(&obd->obd_osfs_lock);
2952 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2953 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2954 obd->obd_name, filter->fo_tot_granted,
2955 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2956 filter->fo_tot_granted -= fed->fed_grant;
2957 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2958 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2959 obd->obd_name, filter->fo_tot_pending,
2960 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2961 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2962 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2963 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2964 obd->obd_name, filter->fo_tot_dirty,
2965 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2966 filter->fo_tot_dirty -= fed->fed_dirty;
2970 cfs_spin_unlock(&obd->obd_osfs_lock);
2973 static int filter_destroy_export(struct obd_export *exp)
2975 struct filter_export_data *fed = &exp->exp_filter_data;
2978 if (fed->fed_pending)
2979 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2980 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2981 exp, fed->fed_pending);
2983 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2985 target_destroy_export(exp);
2987 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
2988 &exp->exp_client_uuid)))
2991 ldlm_destroy_export(exp);
2992 lut_client_free(exp);
2994 if (!exp->exp_obd->obd_replayable)
2995 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2997 filter_grant_discard(exp);
2998 filter_fmd_cleanup(exp);
3000 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3001 struct filter_obd *filter = &exp->exp_obd->u.filter;
3002 if (filter->fo_tot_granted_clients > 0)
3003 filter->fo_tot_granted_clients --;
3006 if (!(exp->exp_flags & OBD_OPT_FORCE))
3007 filter_grant_sanity_check(exp->exp_obd, __func__);
3012 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3014 struct obd_llog_group *olg_min, *olg;
3015 struct filter_obd *filter;
3016 int worked = -1, group;
3017 struct llog_ctxt *ctxt;
3020 filter = &obd->u.filter;
3022 /* we can't sync log holding spinlock. also, we do not want to get
3023 * into livelock. so we do following: loop over MDS's exports in
3024 * group order and skip already synced llogs -bzzz */
3026 /* look for group with min. number, but > worked */
3029 cfs_spin_lock(&filter->fo_llog_list_lock);
3030 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3031 if (olg->olg_seq <= worked) {
3032 /* this group is already synced */
3035 if (group < olg->olg_seq) {
3036 /* we have group with smaller number to sync */
3039 /* store current minimal group */
3041 group = olg->olg_seq;
3043 cfs_spin_unlock(&filter->fo_llog_list_lock);
3045 if (olg_min == NULL)
3048 worked = olg_min->olg_seq;
3049 if (olg_min->olg_exp &&
3050 (dexp == olg_min->olg_exp || dexp == NULL)) {
3052 ctxt = llog_group_get_ctxt(olg_min,
3053 LLOG_MDS_OST_REPL_CTXT);
3055 err = llog_sync(ctxt, olg_min->olg_exp);
3056 llog_ctxt_put(ctxt);
3058 CERROR("error flushing logs to MDS: "
3063 } while (olg_min != NULL);
3066 /* Also incredibly similar to mds_disconnect */
3067 static int filter_disconnect(struct obd_export *exp)
3069 struct obd_device *obd = exp->exp_obd;
3074 class_export_get(exp);
3076 if (!(exp->exp_flags & OBD_OPT_FORCE))
3077 filter_grant_sanity_check(obd, __func__);
3078 filter_grant_discard(exp);
3080 /* Flush any remaining cancel messages out to the target */
3081 filter_sync_llogs(obd, exp);
3083 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3085 rc = server_disconnect_export(exp);
3087 /* Do not erase record for recoverable client. */
3088 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3089 filter_client_del(exp);
3091 fsfilt_sync(obd, obd->u.obt.obt_sb);
3093 class_export_put(exp);
3097 /* reverse import is changed, sync all cancels */
3098 static void filter_revimp_update(struct obd_export *exp)
3103 class_export_get(exp);
3105 /* flush any remaining cancel messages out to the target */
3106 filter_sync_llogs(exp->exp_obd, exp);
3107 class_export_put(exp);
3111 static int filter_ping(struct obd_export *exp)
3113 filter_fmd_expire(exp);
3117 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3118 const char *what, int quiet)
3120 struct dentry *dchild = NULL;
3122 dchild = filter_fid2dentry(obd, NULL, ostid->oi_seq, ostid->oi_id);
3124 if (IS_ERR(dchild)) {
3125 CERROR("%s error looking up object: "POSTID"\n",
3126 what, ostid->oi_id, ostid->oi_seq);
3130 if (dchild->d_inode == NULL) {
3132 CERROR("%s: %s on non-existent object: "POSTID" \n",
3133 obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3135 RETURN(ERR_PTR(-ENOENT));
3138 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0)
3139 /* Try to correct for a bug in 2.1.0 (LU-221) that caused negative
3140 * timestamps to appear to be in the far future, due old timestamp
3141 * being stored on disk as an unsigned value. This fixes up any
3142 * bad values stored on disk before returning them to the client,
3143 * and ensures any timestamp updates are correct. LU-1042 */
3144 if (unlikely(LTIME_S(dchild->d_inode->i_atime) == LU221_BAD_TIME))
3145 LTIME_S(dchild->d_inode->i_atime) = 0;
3146 if (unlikely(LTIME_S(dchild->d_inode->i_mtime) == LU221_BAD_TIME))
3147 LTIME_S(dchild->d_inode->i_mtime) = 0;
3148 if (unlikely(LTIME_S(dchild->d_inode->i_ctime) == LU221_BAD_TIME))
3149 LTIME_S(dchild->d_inode->i_ctime) = 0;
3151 #warning "remove old LU-221/LU-1042 workaround code"
3157 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3159 struct dentry *dentry = NULL;
3160 struct obd_device *obd;
3165 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3166 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3170 obd = class_exp2obd(exp);
3172 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3176 dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3178 RETURN(PTR_ERR(dentry));
3180 /* Limit the valid bits in the return data to what we actually use */
3181 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3182 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3184 /* Store inode version in reply */
3185 curr_version = fsfilt_get_version(exp->exp_obd, dentry->d_inode);
3186 if ((__s64)curr_version != -EOPNOTSUPP) {
3187 oinfo->oi_oa->o_valid |= OBD_MD_FLDATAVERSION;
3188 oinfo->oi_oa->o_data_version = curr_version;
3195 /* this should be enabled/disabled in condition to enabled/disabled large
3196 * inodes (fast EAs) in backing store FS. */
3197 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3198 void *handle, struct obdo *oa)
3200 struct obd_device *obd = exp->exp_obd;
3204 if (oa->o_valid & OBD_MD_FLFID) {
3205 struct filter_fid ff;
3207 if (!(oa->o_valid & OBD_MD_FLGROUP))
3209 /* packing fid and converting it to LE for storing into EA.
3210 * Here ->o_stripe_idx should be filled by LOV and rest of
3211 * fields - by client. */
3212 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3213 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3214 /* XXX: we are ignoring o_parent_ver here, since this should
3215 * be the same for all objects in this fileset. */
3216 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3217 ff.ff_objid = cpu_to_le64(oa->o_id);
3218 ff.ff_seq = cpu_to_le64(oa->o_seq);
3220 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3221 LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3224 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3226 CERROR("store fid in object failed! rc: %d\n", rc);
3228 CDEBUG(D_HA, "OSS object without fid info!\n");
3234 /* this is called from filter_truncate() until we have filter_punch() */
3235 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3236 struct obdo *oa, struct obd_trans_info *oti)
3238 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3239 struct llog_cookie *fcc = NULL;
3240 struct filter_obd *filter;
3241 int rc, err, sync = 0;
3242 loff_t old_size = 0;
3243 unsigned int ia_valid;
3244 struct inode *inode;
3245 struct page *page = NULL;
3250 LASSERT(dentry != NULL);
3251 LASSERT(!IS_ERR(dentry));
3253 inode = dentry->d_inode;
3254 LASSERT(inode != NULL);
3256 filter = &exp->exp_obd->u.filter;
3257 iattr_from_obdo(&iattr, oa, oa->o_valid);
3258 ia_valid = iattr.ia_valid;
3260 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3261 OBD_ALLOC(fcc, sizeof(*fcc));
3263 *fcc = oa->o_lcookie;
3265 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3266 unsigned long now = jiffies;
3267 /* Filter truncates and writes are serialized by
3268 * i_alloc_sem, see the comment in
3269 * filter_preprw_write.*/
3270 if (ia_valid & ATTR_SIZE)
3271 down_write(&inode->i_alloc_sem);
3272 LOCK_INODE_MUTEX(inode);
3273 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
3274 old_size = i_size_read(inode);
3277 /* VBR: version recovery check */
3278 rc = filter_version_get_check(exp, oti, inode);
3280 GOTO(out_unlock, rc);
3282 /* Let's pin the last page so that ldiskfs_truncate
3283 * should not start GFP_FS allocation. */
3284 if (ia_valid & ATTR_SIZE) {
3285 page = grab_cache_page(inode->i_mapping,
3286 iattr.ia_size >> PAGE_CACHE_SHIFT);
3288 GOTO(out_unlock, rc = -ENOMEM);
3293 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3294 * then we will accept the UID+GID sent by the client during write for
3295 * initializing the ownership of this inode. We only allow this to
3296 * happen once so clear these bits in setattr. In 2.6 kernels it is
3297 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3298 * the flags that are actually being set. */
3299 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3300 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3301 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3303 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3304 if (!(ia_valid & ATTR_MODE)) {
3305 iattr.ia_mode = inode->i_mode;
3306 iattr.ia_valid |= ATTR_MODE;
3308 iattr.ia_mode &= ~S_ISUID;
3310 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3311 if (!(iattr.ia_valid & ATTR_MODE)) {
3312 iattr.ia_mode = inode->i_mode;
3313 iattr.ia_valid |= ATTR_MODE;
3315 iattr.ia_mode &= ~S_ISGID;
3318 orig_ids[USRQUOTA] = inode->i_uid;
3319 orig_ids[GRPQUOTA] = inode->i_gid;
3320 handle = fsfilt_start_log(exp->exp_obd, inode,
3321 FSFILT_OP_SETATTR, oti, 1);
3323 GOTO(out_unlock, rc = PTR_ERR(handle));
3325 /* update inode EA only once when inode is suid bit marked. As
3326 * on 2.6.x UID and GID may be set separately, we check here
3327 * only one of them to avoid double setting. */
3328 if (inode->i_mode & S_ISUID)
3329 filter_update_fidea(exp, inode, handle, oa);
3331 handle = fsfilt_start(exp->exp_obd, inode,
3332 FSFILT_OP_SETATTR, oti);
3334 GOTO(out_unlock, rc = PTR_ERR(handle));
3337 /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
3338 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3339 ll_vfs_dq_init(inode);
3341 if (oa->o_valid & OBD_MD_FLFLAGS) {
3342 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3343 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3345 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3347 /* set cancel cookie callback function */
3348 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3349 filter_cancel_cookies_cb,
3353 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3354 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3356 /* The truncate might have used up our transaction credits. Make sure
3357 * we have two left for the last_rcvd and VBR inode version updates. */
3358 err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3360 /* Update inode version only if data has changed => size has changed */
3361 rc = filter_finish_transno(exp, ia_valid & ATTR_SIZE ? inode : NULL,
3365 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3369 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3371 CERROR("error on commit, err = %d\n", err);
3378 /* For a partial-page truncate flush the page to disk immediately
3379 * to avoid data corruption during direct disk write. b=17397 */
3380 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3381 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3382 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3392 page_cache_release(page);
3394 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3395 UNLOCK_INODE_MUTEX(inode);
3396 if (ia_valid & ATTR_SIZE)
3397 up_write(&inode->i_alloc_sem);
3399 OBD_FREE(fcc, sizeof(*fcc));
3401 /* trigger quota release */
3402 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3403 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3404 int rc2 = lquota_adjust(filter_quota_interface_ref,
3405 exp->exp_obd, cur_ids,
3406 orig_ids, rc, FSFILT_OP_SETATTR);
3407 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3408 "filter adjust qunit. (rc:%d)\n", rc2);
3413 /* this is called from filter_truncate() until we have filter_punch() */
3414 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3415 struct obd_trans_info *oti)
3417 struct obdo *oa = oinfo->oi_oa;
3418 struct lustre_capa *capa = oinfo_capa(oinfo);
3419 struct ldlm_res_id res_id;
3420 struct filter_mod_data *fmd;
3421 struct lvfs_run_ctxt saved;
3422 struct filter_obd *filter;
3423 struct ldlm_resource *res;
3424 struct dentry *dentry;
3425 __u64 opc = CAPA_OPC_META_WRITE;
3429 if (oinfo->oi_flags & OBD_FL_PUNCH)
3430 opc |= CAPA_OPC_OSS_TRUNC;
3432 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3436 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3437 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3442 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3443 /* This would be very bad - accidentally truncating a file when
3444 * changing the time or similar - bug 12203. */
3445 if (oa->o_valid & OBD_MD_FLSIZE &&
3446 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3447 static char mdsinum[48];
3449 if (oa->o_valid & OBD_MD_FLFID)
3450 snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3451 oa->o_parent_seq, oa->o_parent_oid,
3456 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3457 "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3458 oa->o_id, oa->o_seq, mdsinum);
3462 dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3464 RETURN(PTR_ERR(dentry));
3466 filter = &exp->exp_obd->u.filter;
3467 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3470 * We need to be atomic against a concurrent write
3471 * (which takes the semaphore for reading). fmd_mactime_xid
3472 * checks will have no effect if a write request with lower
3473 * xid starts just before a setattr and finishes later than
3474 * the setattr (see bug 21489, comment 27).
3477 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3478 unsigned long now = jiffies;
3479 down_write(&dentry->d_inode->i_alloc_sem);
3480 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem");
3481 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3482 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3483 fmd->fmd_mactime_xid = oti->oti_xid;
3484 filter_fmd_put(exp, fmd);
3485 up_write(&dentry->d_inode->i_alloc_sem);
3488 /* setting objects attributes (including owner/group) */
3489 rc = filter_setattr_internal(exp, dentry, oa, oti);
3491 GOTO(out_unlock, rc);
3493 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3494 &res_id, LDLM_EXTENT, 0);
3497 LDLM_RESOURCE_ADDREF(res);
3498 rc = ldlm_res_lvbo_update(res, NULL, 0);
3499 LDLM_RESOURCE_DELREF(res);
3500 ldlm_resource_putref(res);
3503 oa->o_valid = OBD_MD_FLID;
3505 /* Quota release need uid/gid info */
3506 obdo_from_inode(oa, dentry->d_inode, NULL,
3507 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3512 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3516 /* XXX identical to osc_unpackmd */
3517 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3518 struct lov_mds_md *lmm, int lmm_bytes)
3524 if (lmm_bytes < sizeof (*lmm)) {
3525 CERROR("lov_mds_md too small: %d, need %d\n",
3526 lmm_bytes, (int)sizeof(*lmm));
3529 /* XXX LOV_MAGIC etc check? */
3531 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3532 CERROR("lov_mds_md: zero lmm_object_id\n");
3537 lsm_size = lov_stripe_md_size(1);
3541 if (*lsmp != NULL && lmm == NULL) {
3542 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3543 OBD_FREE(*lsmp, lsm_size);
3548 if (*lsmp == NULL) {
3549 OBD_ALLOC(*lsmp, lsm_size);
3553 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3554 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3555 OBD_FREE(*lsmp, lsm_size);
3558 loi_init((*lsmp)->lsm_oinfo[0]);
3562 /* XXX zero *lsmp? */
3563 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3564 LASSERT((*lsmp)->lsm_object_id);
3567 (*lsmp)->lsm_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
3572 /* caller must hold fo_create_locks[oa->o_seq] */
3573 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3574 struct filter_obd *filter)
3576 struct obdo doa = { 0 }; /* XXX obdo on stack */
3582 LASSERT_MUTEX_LOCKED(&filter->fo_create_locks[oa->o_seq]);
3584 memset(&doa, 0, sizeof(doa));
3586 doa.o_valid |= OBD_MD_FLGROUP;
3587 doa.o_seq = oa->o_seq;
3588 doa.o_mode = S_IFREG;
3590 if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3591 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3592 exp->exp_obd->obd_name, doa.o_seq);
3596 last = filter_last_id(filter, doa.o_seq);
3598 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3600 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3601 exp->exp_obd->obd_name, oa->o_id + 1, last,
3602 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3604 for (id = last; id > oa->o_id; id--) {
3606 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3607 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3608 CEMERG("error destroying precreate objid "LPU64": %d\n",
3611 /* update last_id on disk periodically so that if we restart
3612 * we don't need to re-scan all of the just-deleted objects. */
3613 if ((id & 511) == 0 && !skip_orphan) {
3614 filter_set_last_id(filter, id - 1, doa.o_seq);
3615 filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3619 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3620 exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3623 filter_set_last_id(filter, id, doa.o_seq);
3624 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3627 * We have destroyed orphan objects, but don't want to reuse
3628 * them. Therefore we don't reset last_id to the last created
3629 * objects. Instead, we report back to the MDS the object id
3630 * of the last orphan, so that the MDS can restart allocating
3631 * objects from this id + 1 and thus skip the whole orphan
3637 cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3642 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3643 obd_seq group, int *num);
3644 /* returns a negative error or a nonnegative number of files to create */
3645 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3646 obd_seq group, struct obd_trans_info *oti)
3648 struct obd_device *obd = exp->exp_obd;
3649 struct filter_obd *filter = &obd->u.filter;
3653 /* delete orphans request */
3654 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3655 obd_id last = filter_last_id(filter, group);
3657 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3658 CERROR("%s: dropping old orphan cleanup request\n",
3662 /* This causes inflight precreates to abort and drop lock */
3663 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3664 cfs_mutex_lock(&filter->fo_create_locks[group]);
3665 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3666 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3667 exp->exp_obd->obd_name, group);
3668 cfs_mutex_unlock(&filter->fo_create_locks[group]);
3671 diff = oa->o_id - last;
3672 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3675 if (-diff > OST_MAX_PRECREATE) {
3676 CERROR("%s: ignoring bogus orphan destroy request: "
3677 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3679 /* FIXME: should reset precreate_next_id on MDS */
3680 GOTO(out, rc = -EINVAL);
3683 rc = filter_destroy_precreated(exp, oa, filter);
3685 CERROR("%s: unable to write lastobjid, but "
3686 "orphans were deleted\n", obd->obd_name);
3689 /* XXX: Used by MDS for the first time! */
3690 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3693 cfs_mutex_lock(&filter->fo_create_locks[group]);
3694 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3695 CERROR("%s: dropping old precreate request\n",
3699 /* only precreate if group == 0 and o_id is specfied */
3700 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3703 diff = oa->o_id - filter_last_id(filter, group);
3704 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3705 filter_last_id(filter, group), diff);
3707 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3708 oa->o_id, filter_last_id(filter, group), diff);
3712 oa->o_id = filter_last_id(&obd->u.filter, group);
3713 rc = filter_precreate(obd, oa, group, &diff);
3714 oa->o_id = filter_last_id(&obd->u.filter, group);
3716 oa->o_valid |= (OBD_MD_FLID | OBD_MD_FLGROUP);
3719 /* else diff == 0 */
3722 cfs_mutex_unlock(&filter->fo_create_locks[group]);
3726 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3727 __u64 max_age, __u32 flags)
3729 struct filter_obd *filter = &obd->u.filter;
3730 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3731 struct lr_server_data *lsd = class_server_data(obd);
3735 /* at least try to account for cached pages. its still racey and
3736 * might be under-reporting if clients haven't announced their
3737 * caches with brw recently */
3738 cfs_spin_lock(&obd->obd_osfs_lock);
3739 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3740 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3741 cfs_spin_unlock(&obd->obd_osfs_lock);
3743 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3744 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3745 filter->fo_tot_dirty, filter->fo_tot_granted,
3746 filter->fo_tot_pending,
3747 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3749 filter_grant_sanity_check(obd, __func__);
3751 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3752 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3753 osfs->os_bsize - 1) >> blockbits));
3755 if (OBD_FAIL_CHECK_VALUE(OBD_FAIL_OST_ENOSPC,
3756 le32_to_cpu(lsd->lsd_ost_index)))
3757 osfs->os_bfree = osfs->os_bavail = 2;
3759 if (OBD_FAIL_CHECK_VALUE(OBD_FAIL_OST_ENOINO,
3760 le32_to_cpu(lsd->lsd_ost_index)))
3763 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3764 * stop creating files on MDS if OST is not good shape to create
3768 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3769 osfs->os_state = OS_STATE_READONLY;
3771 if (filter->fo_raid_degraded)
3772 osfs->os_state |= OS_STATE_DEGRADED;
3776 static int filter_use_existing_obj(struct obd_device *obd,
3777 struct dentry *dchild, void **handle,
3780 struct inode *inode = dchild->d_inode;
3784 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3787 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3788 if (IS_ERR(*handle))
3789 return PTR_ERR(*handle);
3791 iattr.ia_valid = ATTR_MODE;
3792 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3793 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3800 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3803 __u64 os_ffree = -1;
3805 cfs_spin_lock(&obd->obd_osfs_lock);
3806 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3808 os_ffree = obd->obd_osfs.os_ffree;
3809 cfs_spin_unlock(&obd->obd_osfs_lock);
3814 /* We rely on the fact that only one thread will be creating files in a given
3815 * group at a time, which is why we don't need an atomic filter_get_new_id.
3816 * Even if we had that atomic function, the following race would exist:
3818 * thread 1: gets id x from filter_next_id
3819 * thread 2: gets id (x + 1) from filter_next_id
3820 * thread 2: creates object (x + 1)
3821 * thread 1: tries to create object x, gets -ENOSPC
3823 * Caller must hold fo_create_locks[group]
3825 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3826 obd_seq group, int *num)
3828 struct dentry *dchild = NULL, *dparent = NULL;
3829 struct filter_obd *filter;
3830 struct obd_statfs *osfs;
3832 int err = 0, rc = 0, recreate_obj = 0, i;
3833 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3836 void *handle = NULL;
3839 filter = &obd->u.filter;
3841 LASSERT_MUTEX_LOCKED(&filter->fo_create_locks[group]);
3843 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3845 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3846 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3849 OBD_ALLOC(osfs, sizeof(*osfs));
3852 rc = filter_statfs(obd, osfs,
3853 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
3855 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3856 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3857 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3858 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3860 if (oa->o_valid & OBD_MD_FLFLAGS)
3861 oa->o_flags |= OBD_FL_NOSPC_BLK;
3863 oa->o_valid |= OBD_MD_FLFLAGS;
3864 oa->o_flags = OBD_FL_NOSPC_BLK;
3869 OBD_FREE(osfs, sizeof(*osfs));
3874 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3875 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3877 for (i = 0; i < *num && err == 0; i++) {
3878 int cleanup_phase = 0;
3880 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3881 CWARN("%s: create aborted by destroy\n",
3890 last_id = filter_last_id(filter, group);
3891 if (next_id > last_id) {
3892 CERROR("Error: Trying to recreate obj greater"
3893 "than last id "LPD64" > "LPD64"\n",
3895 GOTO(cleanup, rc = -EINVAL);
3898 next_id = filter_last_id(filter, group) + 1;
3900 /* Don't create objects beyond the valid range for this SEQ */
3901 if (unlikely(fid_seq_is_mdt0(group) &&
3902 next_id >= IDIF_MAX_OID)) {
3903 CERROR("%s:"POSTID" hit the IDIF_MAX_OID (1<<48)!\n",
3904 obd->obd_name, next_id, group);
3905 GOTO(cleanup, rc = -ENOSPC);
3906 } else if (unlikely(!fid_seq_is_mdt0(group) &&
3907 next_id >= OBIF_MAX_OID)) {
3908 CERROR("%s:"POSTID" hit the OBIF_MAX_OID (1<<32)!\n",
3909 obd->obd_name, next_id, group);
3910 GOTO(cleanup, rc = -ENOSPC);
3913 dparent = filter_parent_lock(obd, group, next_id);
3914 if (IS_ERR(dparent))
3915 GOTO(cleanup, rc = PTR_ERR(dparent));
3916 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3918 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3920 GOTO(cleanup, rc = PTR_ERR(dchild));
3921 cleanup_phase = 2; /* f_dput(dchild) */
3923 if (dchild->d_inode != NULL) {
3924 /* This would only happen if lastobjid was bad on disk*/
3925 /* Could also happen if recreating missing obj but it
3926 * already exists. */
3928 CERROR("%s: recreating existing object %.*s?\n",
3929 obd->obd_name, dchild->d_name.len,
3930 dchild->d_name.name);
3932 /* Use these existing objects if they are
3934 if (i_size_read(dchild->d_inode) == 0) {
3935 rc = filter_use_existing_obj(obd,dchild,
3936 &handle, &cleanup_phase);
3943 CERROR("%s: Serious error: objid %.*s already "
3944 "exists; is this filesystem corrupt?\n",
3945 obd->obd_name, dchild->d_name.len,
3946 dchild->d_name.name);
3949 GOTO(cleanup, rc = -EEXIST);
3952 handle = fsfilt_start_log(obd, dparent->d_inode,
3953 FSFILT_OP_CREATE, NULL, 1);
3955 GOTO(cleanup, rc = PTR_ERR(handle));
3958 CDEBUG(D_INODE, "%s: filter_precreate(od->o_seq="LPU64
3959 ",od->o_id="LPU64")\n", obd->obd_name, group,
3962 /* We mark object SUID+SGID to flag it for accepting UID+GID
3963 * from client on first write. Currently the permission bits
3964 * on the OST are never used, so this is OK. */
3965 rc = ll_vfs_create(dparent->d_inode, dchild,
3966 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3968 CWARN("%s: create failed: rc = %d\n", obd->obd_name,rc);
3969 if (rc == -ENOSPC) {
3970 os_ffree = filter_calc_free_inodes(obd);
3974 if (obd->obd_osfs.os_bavail <
3975 (obd->obd_osfs.os_blocks >> 10)) {
3976 if (oa->o_valid & OBD_MD_FLFLAGS) {
3977 oa->o_flags |= OBD_FL_NOSPC_BLK;
3979 oa->o_valid |= OBD_MD_FLFLAGS;
3980 oa->o_flags = OBD_FL_NOSPC_BLK;
3983 CWARN("%s: free inode "LPU64"\n",
3984 obd->obd_name, os_ffree);
3990 if (dchild->d_inode)
3991 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3992 dchild->d_inode->i_ino);
3995 /* Initialize a/c/m time so any client timestamp will always
3996 * be newer and update the inode. ctime = 0 is also handled
3997 * specially in fsfilt_ext3_setattr(). See LU-221, LU-1042 */
3998 iattr.ia_valid = ATTR_ATIME | ATTR_MTIME | ATTR_CTIME;
3999 LTIME_S(iattr.ia_atime) = 0;
4000 LTIME_S(iattr.ia_mtime) = 0;
4001 LTIME_S(iattr.ia_ctime) = 0;
4002 err = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
4004 CWARN("%s: unable to initialize a/c/m time of newly "
4005 "created object %.*s: rc = %d\n",
4006 obd->obd_name, dchild->d_name.len,
4007 dchild->d_name.name, err);
4009 if (!recreate_obj) {
4010 filter_set_last_id(filter, next_id, group);
4011 err = filter_update_last_objid(obd, group, 0);
4013 CERROR("%s: unable to write lastobjid "
4014 "but file created: rc = %d\n",
4015 obd->obd_name, err);
4019 switch(cleanup_phase) {
4021 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4023 CERROR("error on commit, err = %d\n", err);
4030 filter_parent_unlock(dparent);
4037 if (cfs_time_after(jiffies, enough_time)) {
4040 "%s: precreate slow - want %d got %d \n",
4041 obd->obd_name, *num, i);
4048 "%s: created %d objects for group "POSTID" rc %d\n",
4049 obd->obd_name, i, filter->fo_last_objids[group], group, rc);
4054 int filter_create(struct obd_export *exp, struct obdo *oa,
4055 struct lov_stripe_md **ea, struct obd_trans_info *oti)
4057 struct obd_device *obd = exp->exp_obd;
4058 struct filter_export_data *fed;
4059 struct filter_obd *filter;
4060 struct lvfs_run_ctxt saved;
4061 struct lov_stripe_md *lsm = NULL;
4065 CDEBUG(D_INODE, "%s: filter_create(group="LPU64",id="
4066 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4068 fed = &exp->exp_filter_data;
4069 filter = &obd->u.filter;
4071 /* 1.8 client doesn't carry the ocd_group with connect request,
4072 * so the fed_group will always be zero for 1.8 client. */
4073 if (!(exp->exp_connect_flags & OBD_CONNECT_FULL20)) {
4074 if (oa->o_seq != FID_SEQ_OST_MDT0 &&
4075 oa->o_seq != FID_SEQ_LLOG &&
4076 oa->o_seq != FID_SEQ_ECHO) {
4077 CERROR("The request from older client has invalid"
4078 " group "LPU64"!\n", oa->o_seq);
4081 } else if (fed->fed_group != oa->o_seq) {
4082 CERROR("%s: this export (nid %s) used object group %d "
4083 "earlier; now it's trying to use group "LPU64"!"
4084 " This could be a bug in the MDS. Please report to "
4085 "http://bugzilla.lustre.org/\n", obd->obd_name,
4086 obd_export_nid2str(exp), fed->fed_group, oa->o_seq);
4093 rc = obd_alloc_memmd(exp, &lsm);
4100 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4102 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4103 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4104 if (!obd->obd_recovering ||
4105 oa->o_id > filter_last_id(filter, oa->o_seq)) {
4106 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4107 oa->o_id, filter_last_id(filter, oa->o_seq));
4111 cfs_mutex_lock(&filter->fo_create_locks[oa->o_seq]);
4112 rc = filter_precreate(obd, oa, oa->o_seq, &diff);
4113 cfs_mutex_unlock(&filter->fo_create_locks[oa->o_seq]);
4116 rc = filter_handle_precreate(exp, oa, oa->o_seq, oti);
4119 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4120 if (rc && ea != NULL && *ea != lsm) {
4121 obd_free_memmd(exp, &lsm);
4122 } else if (rc == 0 && ea != NULL) {
4123 /* XXX LOV STACKING: the lsm that is passed to us from
4124 * LOV does not have valid lsm_oinfo data structs, so
4125 * don't go touching that. This needs to be fixed in a
4127 lsm->lsm_object_id = oa->o_id;
4134 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4135 struct lov_stripe_md *md, struct obd_trans_info *oti,
4136 struct obd_export *md_exp, void *capa)
4138 unsigned int qcids[MAXQUOTAS] = {0, 0};
4139 struct obd_device *obd;
4140 struct filter_obd *filter;
4141 struct dentry *dchild = NULL, *dparent = NULL;
4142 struct lustre_handle lockh = { 0 };
4143 struct lvfs_run_ctxt saved;
4144 void *handle = NULL;
4145 struct llog_cookie *fcc = NULL;
4146 int rc, rc2, cleanup_phase = 0, sync = 0;
4151 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4152 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4157 filter = &obd->u.filter;
4159 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4162 CDEBUG(D_INODE, "%s: filter_destroy(group="LPU64",oid="
4163 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4165 dchild = filter_fid2dentry(obd, NULL, oa->o_seq, oa->o_id);
4167 GOTO(cleanup, rc = PTR_ERR(dchild));
4170 if (dchild->d_inode == NULL) {
4171 CDEBUG(D_INODE, "destroying non-existent object "POSTID"\n",
4172 oa->o_id, oa->o_seq);
4173 /* If object already gone, cancel cookie right now */
4174 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4175 struct llog_ctxt *ctxt;
4176 struct obd_llog_group *olg;
4178 olg = filter_find_olg(obd, oa->o_seq);
4180 CERROR(" %s: can not find olg of group %d\n",
4181 obd->obd_name, (int)oa->o_seq);
4182 GOTO(cleanup, rc = PTR_ERR(olg));
4184 fcc = &oa->o_lcookie;
4185 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4186 llog_cancel(ctxt, NULL, 1, fcc, 0);
4187 llog_ctxt_put(ctxt);
4188 fcc = NULL; /* we didn't allocate fcc, don't free it */
4190 GOTO(cleanup, rc = -ENOENT);
4193 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_seq, &lockh);
4197 /* Our MDC connection is established by the MDS to us */
4198 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4199 OBD_ALLOC(fcc, sizeof(*fcc));
4201 *fcc = oa->o_lcookie;
4204 /* we're gonna truncate it first in order to avoid possible deadlock:
4206 * open trasaction open transaction
4207 * down(i_zombie) down(i_zombie)
4208 * restart transaction
4209 * (see BUG 4180) -bzzz
4211 * take i_alloc_sem too to prevent other threads from writing to the
4212 * file while we are truncating it. This can cause lock ordering issue
4213 * between page lock, i_mutex & starting new journal handle.
4214 * (see bug 20321) -johann
4217 down_write(&dchild->d_inode->i_alloc_sem);
4218 LOCK_INODE_MUTEX(dchild->d_inode);
4219 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
4221 /* VBR: version recovery check */
4222 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4224 UNLOCK_INODE_MUTEX(dchild->d_inode);
4225 up_write(&dchild->d_inode->i_alloc_sem);
4229 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4231 if (IS_ERR(handle)) {
4232 UNLOCK_INODE_MUTEX(dchild->d_inode);
4233 up_write(&dchild->d_inode->i_alloc_sem);
4234 GOTO(cleanup, rc = PTR_ERR(handle));
4237 /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
4238 ll_vfs_dq_init(dchild->d_inode);
4240 iattr.ia_valid = ATTR_SIZE;
4242 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4243 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4244 UNLOCK_INODE_MUTEX(dchild->d_inode);
4245 up_write(&dchild->d_inode->i_alloc_sem);
4249 GOTO(cleanup, rc = rc2);
4251 /* We don't actually need to lock the parent until we are unlinking
4252 * here, and not while truncating above. That avoids holding the
4253 * parent lock for a long time during truncate, which can block other
4254 * threads from doing anything to objects in that directory. bug 7171 */
4255 dparent = filter_parent_lock(obd, oa->o_seq, oa->o_id);
4256 if (IS_ERR(dparent))
4257 GOTO(cleanup, rc = PTR_ERR(dparent));
4258 cleanup_phase = 3; /* filter_parent_unlock */
4260 LOCK_INODE_MUTEX(dchild->d_inode);
4261 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4262 if (IS_ERR(handle)) {
4263 UNLOCK_INODE_MUTEX(dchild->d_inode);
4264 GOTO(cleanup, rc = PTR_ERR(handle));
4266 cleanup_phase = 4; /* fsfilt_commit */
4268 /* Quota release need uid/gid of inode */
4269 obdo_from_inode(oa, dchild->d_inode, NULL, OBD_MD_FLUID|OBD_MD_FLGID);
4271 filter_fmd_drop(exp, oa->o_id, oa->o_seq);
4273 /* this drops dchild->d_inode->i_mutex unconditionally */
4274 rc = filter_destroy_internal(obd, oa->o_id, oa->o_seq, dparent, dchild);
4278 switch(cleanup_phase) {
4281 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4282 oti->oti_handle : handle,
4283 filter_cancel_cookies_cb,
4285 /* If add_journal_cb failed, then filter_finish_transno
4286 * will commit the handle and we will do a sync
4287 * on commit. then we call callback directly to free
4290 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4292 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4295 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4297 CERROR("error on commit, err = %d\n", rc2);
4304 filter_parent_unlock(dparent);
4306 filter_fini_destroy(obd, &lockh);
4310 OBD_FREE(fcc, sizeof(*fcc));
4312 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4315 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4319 /* trigger quota release */
4320 qcids[USRQUOTA] = oa->o_uid;
4321 qcids[GRPQUOTA] = oa->o_gid;
4322 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4325 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4329 /* NB start and end are used for punch, but not truncate */
4330 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4331 struct obd_trans_info *oti,
4332 struct ptlrpc_request_set *rqset)
4337 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4338 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4339 oinfo->oi_policy.l_extent.end);
4343 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4344 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,oinfo->oi_oa->o_valid,
4345 oinfo->oi_policy.l_extent.start);
4347 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4348 rc = filter_setattr(exp, oinfo, oti);
4352 static int filter_sync(struct obd_export *exp, struct obd_info *oinfo,
4353 obd_off start, obd_off end,
4354 struct ptlrpc_request_set *set)
4356 struct lvfs_run_ctxt saved;
4357 struct obd_device_target *obt;
4358 struct dentry *dentry;
4362 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
4363 (struct lustre_capa *)oinfo->oi_capa,
4364 CAPA_OPC_OSS_WRITE);
4368 obt = &exp->exp_obd->u.obt;
4370 /* An objid of zero is taken to mean "sync whole filesystem" */
4371 if (!oinfo->oi_oa || !(oinfo->oi_oa->o_valid & OBD_MD_FLID)) {
4372 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4373 /* Flush any remaining cancel messages out to the target */
4374 filter_sync_llogs(exp->exp_obd, exp);
4378 dentry = filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi);
4380 RETURN(PTR_ERR(dentry));
4382 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4384 LOCK_INODE_MUTEX(dentry->d_inode);
4386 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4388 /* just any file to grab fsync method - "file" arg unused */
4389 struct file *file = obt->obt_rcvd_filp;
4391 if (file->f_op && file->f_op->fsync)
4392 rc = file->f_op->fsync(NULL, dentry, 1);
4394 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4398 UNLOCK_INODE_MUTEX(dentry->d_inode);
4400 oinfo->oi_oa->o_valid = OBD_MD_FLID;
4401 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL,
4402 FILTER_VALID_FLAGS);
4404 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4410 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4411 void *key, __u32 *vallen, void *val,
4412 struct lov_stripe_md *lsm)
4414 struct obd_device *obd;
4417 obd = class_exp2obd(exp);
4419 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4423 if (KEY_IS(KEY_BLOCKSIZE)) {
4424 __u32 *blocksize = val;
4426 if (*vallen < sizeof(*blocksize))
4428 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4430 *vallen = sizeof(*blocksize);
4434 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4435 __u32 *blocksize_bits = val;
4436 if (blocksize_bits) {
4437 if (*vallen < sizeof(*blocksize_bits))
4439 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4441 *vallen = sizeof(*blocksize_bits);
4445 if (KEY_IS(KEY_LAST_ID)) {
4446 obd_id *last_id = val;
4447 /* FIXME: object groups */
4449 if (*vallen < sizeof(*last_id))
4451 *last_id = filter_last_id(&obd->u.filter,
4452 exp->exp_filter_data.fed_group);
4454 *vallen = sizeof(*last_id);
4458 if (KEY_IS(KEY_FIEMAP)) {
4459 struct ll_fiemap_info_key *fm_key = key;
4460 struct dentry *dentry;
4461 struct ll_user_fiemap *fiemap = val;
4462 struct lvfs_run_ctxt saved;
4465 if (fiemap == NULL) {
4466 *vallen = fiemap_count_to_size(
4467 fm_key->fiemap.fm_extent_count);
4471 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa.o_oi,
4474 RETURN(PTR_ERR(dentry));
4476 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4477 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4478 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4480 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4486 if (KEY_IS(KEY_SYNC_LOCK_CANCEL)) {
4487 *((__u32 *) val) = obd->u.filter.fo_sync_lock_cancel;
4488 *vallen = sizeof(__u32);
4492 CDEBUG(D_IOCTL, "invalid key\n");
4496 static inline int filter_setup_llog_group(struct obd_export *exp,
4497 struct obd_device *obd,
4500 struct obd_llog_group *olg;
4501 struct llog_ctxt *ctxt;
4504 olg = filter_find_create_olg(obd, group);
4506 RETURN(PTR_ERR(olg));
4508 llog_group_set_export(olg, exp);
4510 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4511 LASSERTF(ctxt != NULL, "ctxt is null\n");
4513 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4514 llog_ctxt_put(ctxt);
4518 static int filter_set_grant_shrink(struct obd_export *exp,
4519 struct ost_body *body)
4521 /* handle shrink grant */
4522 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4523 filter_grant_incoming(exp, &body->oa);
4524 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4530 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4532 struct obd_device *obd;
4538 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4542 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4543 obd_export_nid2str(exp));
4544 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4546 /* setup llog imports */
4548 group = (int)(*(__u32 *)val);
4550 group = 0; /* default value */
4552 LASSERT_SEQ_IS_MDT(group);
4553 rc = filter_setup_llog_group(exp, obd, group);
4557 if (group == FID_SEQ_OST_MDT0) {
4558 /* setup llog group 1 for interop */
4559 filter_setup_llog_group(exp, obd, FID_SEQ_LLOG);
4562 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4567 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4568 void *key, __u32 vallen, void *val,
4569 struct ptlrpc_request_set *set)
4571 struct obd_device *obd;
4576 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4580 if (KEY_IS(KEY_CAPA_KEY)) {
4582 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4584 CERROR("filter update capability key failed: %d\n", rc);
4588 if (KEY_IS(KEY_REVIMP_UPD)) {
4589 filter_revimp_update(exp);
4590 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4594 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4595 filter_adapt_sptlrpc_conf(obd, 0);
4599 if (KEY_IS(KEY_MDS_CONN))
4600 RETURN(filter_set_mds_conn(exp, val));
4602 if (KEY_IS(KEY_GRANT_SHRINK))
4603 RETURN(filter_set_grant_shrink(exp, val));
4608 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4609 int len, void *karg, void *uarg)
4611 struct obd_device *obd = exp->exp_obd;
4612 struct obd_ioctl_data *data = karg;
4616 case OBD_IOC_ABORT_RECOVERY: {
4617 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4618 target_stop_recovery_thread(obd);
4622 case OBD_IOC_SYNC: {
4623 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4624 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4628 case OBD_IOC_SET_READONLY: {
4630 struct super_block *sb = obd->u.obt.obt_sb;
4631 struct inode *inode = sb->s_root->d_inode;
4632 BDEVNAME_DECLARE_STORAGE(tmp);
4633 CERROR("*** setting device %s read-only ***\n",
4634 ll_bdevname(sb, tmp));
4636 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4637 if (!IS_ERR(handle))
4638 rc = fsfilt_commit(obd, inode, handle, 1);
4640 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4641 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4643 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4647 case OBD_IOC_CATLOGLIST: {
4648 rc = llog_catalog_list(obd, 1, data);
4652 case OBD_IOC_LLOG_CANCEL:
4653 case OBD_IOC_LLOG_REMOVE:
4654 case OBD_IOC_LLOG_INFO:
4655 case OBD_IOC_LLOG_PRINT: {
4656 /* FIXME to be finished */
4657 RETURN(-EOPNOTSUPP);
4659 struct llog_ctxt *ctxt = NULL;
4661 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4662 rc = llog_ioctl(ctxt, cmd, data);
4663 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4676 static int filter_health_check(struct obd_device *obd)
4678 #ifdef USE_HEALTH_CHECK_WRITE
4679 struct filter_obd *filter = &obd->u.filter;
4684 * health_check to return 0 on healthy
4685 * and 1 on unhealthy.
4687 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4690 #ifdef USE_HEALTH_CHECK_WRITE
4691 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4692 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4697 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4700 return filter_fid2dentry(data, NULL, gr, id);
4703 static int filter_process_config(struct obd_device *obd, obd_count len,
4706 struct lustre_cfg *lcfg = buf;
4707 struct lprocfs_static_vars lvars;
4710 switch (lcfg->lcfg_command) {
4712 lprocfs_filter_init_vars(&lvars);
4714 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4724 static int filter_notify(struct obd_device *obd,
4725 struct obd_device *unused,
4726 enum obd_notify_event ev, void *data)
4729 case OBD_NOTIFY_CONFIG:
4730 LASSERT(obd->obd_no_conn);
4731 cfs_spin_lock(&obd->obd_dev_lock);
4732 obd->obd_no_conn = 0;
4733 cfs_spin_unlock(&obd->obd_dev_lock);
4736 CDEBUG(D_INFO, "%s: Unhandled notification %#x\n",
4742 static struct lvfs_callback_ops filter_lvfs_ops = {
4743 l_fid2dentry: filter_lvfs_fid2dentry,
4746 static struct obd_ops filter_obd_ops = {
4747 .o_owner = THIS_MODULE,
4748 .o_get_info = filter_get_info,
4749 .o_set_info_async = filter_set_info_async,
4750 .o_setup = filter_setup,
4751 .o_precleanup = filter_precleanup,
4752 .o_cleanup = filter_cleanup,
4753 .o_connect = filter_connect,
4754 .o_reconnect = filter_reconnect,
4755 .o_disconnect = filter_disconnect,
4756 .o_ping = filter_ping,
4757 .o_init_export = filter_init_export,
4758 .o_destroy_export = filter_destroy_export,
4759 .o_statfs = filter_statfs,
4760 .o_getattr = filter_getattr,
4761 .o_unpackmd = filter_unpackmd,
4762 .o_create = filter_create,
4763 .o_setattr = filter_setattr,
4764 .o_destroy = filter_destroy,
4765 .o_brw = filter_brw,
4766 .o_punch = filter_truncate,
4767 .o_sync = filter_sync,
4768 .o_preprw = filter_preprw,
4769 .o_commitrw = filter_commitrw,
4770 .o_llog_init = filter_llog_init,
4771 .o_llog_connect = filter_llog_connect,
4772 .o_llog_finish = filter_llog_finish,
4773 .o_iocontrol = filter_iocontrol,
4774 .o_health_check = filter_health_check,
4775 .o_process_config = filter_process_config,
4776 .o_notify = filter_notify,
4779 quota_interface_t *filter_quota_interface_ref;
4780 extern quota_interface_t filter_quota_interface;
4782 static int __init obdfilter_init(void)
4784 struct lprocfs_static_vars lvars;
4787 /** sanity check for group<->mdsno conversion */
4788 for (i = 0; i < MAX_MDT_COUNT; i++)
4789 LASSERT(objseq_to_mdsno(mdt_to_obd_objseq(i)) == i);
4791 lprocfs_filter_init_vars(&lvars);
4793 cfs_request_module("%s", "lquota");
4794 OBD_ALLOC(obdfilter_created_scratchpad,
4795 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4796 sizeof(*obdfilter_created_scratchpad));
4797 if (obdfilter_created_scratchpad == NULL)
4800 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4801 sizeof(struct filter_mod_data),
4804 GOTO(out, rc = -ENOMEM);
4806 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4807 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4809 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4810 LUSTRE_OST_NAME, NULL);
4814 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4815 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4816 ll_fmd_cachep = NULL;
4818 if (filter_quota_interface_ref)
4819 PORTAL_SYMBOL_PUT(filter_quota_interface);
4821 OBD_FREE(obdfilter_created_scratchpad,
4822 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4823 sizeof(*obdfilter_created_scratchpad));
4829 static void __exit obdfilter_exit(void)
4831 if (filter_quota_interface_ref)
4832 PORTAL_SYMBOL_PUT(filter_quota_interface);
4834 if (ll_fmd_cachep) {
4835 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4836 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4837 ll_fmd_cachep = NULL;
4840 class_unregister_type(LUSTRE_OST_NAME);
4841 OBD_FREE(obdfilter_created_scratchpad,
4842 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4843 sizeof(*obdfilter_created_scratchpad));
4846 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4847 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4848 MODULE_LICENSE("GPL");
4850 module_init(obdfilter_init);
4851 module_exit(obdfilter_exit);