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.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdfilter/filter.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Andreas Dilger <adilger@clusterfs.com>
43 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44 * (which need to get journal_lock, may block if journal full).
46 * Invariant: Call filter_start_transno() before any journal ops to avoid the
47 * same deadlock problem. We can (and want) to get rid of the
48 * transno sem in favour of the dir/inode i_mutex to avoid single
49 * threaded operation on the OST.
52 #define DEBUG_SUBSYSTEM S_FILTER
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
57 #include <linux/module.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
66 #include <obd_cksum.h>
67 #include <obd_class.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
80 #include "filter_internal.h"
82 static struct lvfs_callback_ops filter_lvfs_ops;
83 cfs_mem_cache_t *ll_fmd_cachep;
85 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
86 void *cb_data, int error)
88 struct obd_export *exp = cb_data;
89 LASSERT(exp->exp_obd == obd);
90 obd_transno_commit_cb(obd, transno, exp, error);
91 class_export_cb_put(exp);
94 int filter_version_get_check(struct obd_export *exp,
95 struct obd_trans_info *oti, struct inode *inode)
99 if (inode == NULL || oti == NULL)
102 curr_version = fsfilt_get_version(exp->exp_obd, inode);
103 if ((__s64)curr_version == -EOPNOTSUPP)
105 /* VBR: version is checked always because costs nothing */
106 if (oti->oti_pre_version != 0 &&
107 oti->oti_pre_version != curr_version) {
108 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
109 oti->oti_pre_version, curr_version);
110 cfs_spin_lock(&exp->exp_lock);
111 exp->exp_vbr_failed = 1;
112 cfs_spin_unlock(&exp->exp_lock);
115 oti->oti_pre_version = curr_version;
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121 struct obd_trans_info *oti, int rc, int force_sync)
123 struct obd_device_target *obt = &exp->exp_obd->u.obt;
124 struct tg_export_data *ted = &exp->exp_target_data;
125 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
126 struct lsd_client_data *lcd;
129 int err, log_pri = D_RPCTRACE;
131 /* Propagate error code. */
135 if (!exp->exp_obd->obd_replayable || oti == NULL)
138 cfs_mutex_down(&ted->ted_lcd_lock);
140 /* if the export has already been disconnected, we have no last_rcvd slot,
141 * update server data with latest transno then */
143 cfs_mutex_up(&ted->ted_lcd_lock);
144 CWARN("commit transaction for disconnected client %s: rc %d\n",
145 exp->exp_client_uuid.uuid, rc);
146 err = filter_update_server_data(exp->exp_obd);
150 /* we don't allocate new transnos for replayed requests */
151 cfs_spin_lock(&obt->obt_lut->lut_translock);
152 if (oti->oti_transno == 0) {
153 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
154 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
156 last_rcvd = oti->oti_transno;
157 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
158 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
160 oti->oti_transno = last_rcvd;
162 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
163 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
164 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
165 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
166 cfs_spin_unlock(&obt->obt_lut->lut_translock);
169 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
171 off = ted->ted_lr_off;
173 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
174 ted->ted_lr_idx, ted->ted_lr_off);
177 class_export_cb_get(exp); /* released when the cb is called */
179 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
185 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
186 lcd, sizeof(*lcd), &off,
187 force_sync | exp->exp_need_sync);
189 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
197 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
198 last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
199 cfs_mutex_up(&ted->ted_lcd_lock);
203 void f_dput(struct dentry *dentry)
205 /* Can't go inside filter_ddelete because it can block */
206 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
207 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
208 LASSERT(atomic_read(&dentry->d_count) > 0);
213 static void init_brw_stats(struct brw_stats *brw_stats)
216 for (i = 0; i < BRW_LAST; i++)
217 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
220 static int lprocfs_init_rw_stats(struct obd_device *obd,
221 struct lprocfs_stats **stats)
225 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
226 LPROC_FILTER_LAST - 1;
227 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
231 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
232 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
233 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
234 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
235 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
240 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
241 plus the procfs overhead :( */
242 static int filter_export_stats_init(struct obd_device *obd,
243 struct obd_export *exp,
249 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
250 /* Self-export gets no proc entry */
253 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
255 /* Mask error for already created
263 struct nid_stat *tmp = exp->exp_nid_stats;
264 LASSERT(tmp != NULL);
266 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
267 if (tmp->nid_brw_stats == NULL)
268 GOTO(clean, rc = -ENOMEM);
270 init_brw_stats(tmp->nid_brw_stats);
271 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
272 0644, &filter_per_nid_stats_fops,
275 CWARN("Error adding the brw_stats file\n");
277 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
281 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
285 rc = lprocfs_nid_ldlm_stats_init(tmp);
292 lprocfs_exp_cleanup(exp);
296 /* Add client data to the FILTER. We use a bitmap to locate a free space
297 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
298 * Otherwise, we have just read the data from the last_rcvd file and
299 * we know its offset. */
300 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
303 struct obd_device_target *obt = &obd->u.obt;
304 struct tg_export_data *ted = &exp->exp_target_data;
305 struct lr_server_data *lsd = class_server_data(obd);
306 unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
307 int new_client = (cl_idx == -1);
311 LASSERT(bitmap != NULL);
312 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
315 if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
318 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
319 * there's no need for extra complication here
322 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
324 if (cl_idx >= LR_MAX_CLIENTS) {
325 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
329 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
330 cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
335 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
336 CERROR("FILTER client %d: bit already set in bitmap!\n",
342 ted->ted_lr_idx = cl_idx;
343 ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
344 cl_idx * le16_to_cpu(lsd->lsd_client_size);
345 cfs_init_mutex(&ted->ted_lcd_lock);
346 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
348 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
349 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
352 struct lvfs_run_ctxt saved;
353 loff_t off = ted->ted_lr_off;
357 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
358 ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
360 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
363 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
364 /* Transaction needed to fix bug 1403 */
365 handle = fsfilt_start(obd,
366 obt->obt_rcvd_filp->f_dentry->d_inode,
367 FSFILT_OP_SETATTR, NULL);
368 if (IS_ERR(handle)) {
369 rc = PTR_ERR(handle);
370 CERROR("unable to start transaction: rc %d\n", rc);
372 ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
373 exp->exp_last_request_time = cfs_time_current_sec();
374 rc = fsfilt_add_journal_cb(obd, 0, handle,
375 target_client_add_cb,
376 class_export_cb_get(exp));
378 cfs_spin_lock(&exp->exp_lock);
379 exp->exp_need_sync = 1;
380 cfs_spin_unlock(&exp->exp_lock);
382 rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
384 sizeof(*ted->ted_lcd),
385 &off, rc /* sync if no cb */);
387 obt->obt_rcvd_filp->f_dentry->d_inode,
390 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
393 CERROR("error writing %s client idx %u: rc %d\n",
394 LAST_RCVD, ted->ted_lr_idx, rc);
401 static int filter_client_del(struct obd_export *exp)
403 struct tg_export_data *ted = &exp->exp_target_data;
404 struct obd_device_target *obt = &exp->exp_obd->u.obt;
405 struct lvfs_run_ctxt saved;
410 if (ted->ted_lcd == NULL)
413 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
414 if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
417 LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
419 off = ted->ted_lr_off;
421 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
422 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
424 /* Don't clear ted_lr_idx here as it is likely also unset. At worst
425 * we leak a client slot that will be cleaned on the next recovery. */
427 CERROR("%s: client idx %d has med_off %lld\n",
428 exp->exp_obd->obd_name, ted->ted_lr_idx, off);
429 GOTO(free, rc = -EINVAL);
432 /* Clear the bit _after_ zeroing out the client so we don't
433 race with filter_client_add and zero out new clients.*/
434 if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
435 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
440 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
441 /* Make sure the server's last_transno is up to date.
442 * This should be done before zeroing client slot so last_transno will
443 * be in server data or in client data in case of failure */
444 filter_update_server_data(exp->exp_obd);
446 cfs_mutex_down(&ted->ted_lcd_lock);
447 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
448 rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
450 sizeof(*ted->ted_lcd), &off, 0);
451 cfs_mutex_up(&ted->ted_lcd_lock);
452 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
454 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
455 "zero out client %s at idx %u/%llu in %s, rc %d\n",
456 ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
463 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
464 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
465 struct filter_mod_data *fmd)
467 LASSERT_SPIN_LOCKED(&fed->fed_lock);
468 if (--fmd->fmd_refcount == 0) {
469 /* XXX when we have persistent reservations and the handle
470 * is stored herein we need to drop it here. */
471 fed->fed_mod_count--;
472 cfs_list_del(&fmd->fmd_list);
473 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
477 /* drop fmd reference, free it if last ref */
478 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
480 struct filter_export_data *fed;
485 fed = &exp->exp_filter_data;
486 cfs_spin_lock(&fed->fed_lock);
487 filter_fmd_put_nolock(fed, fmd); /* caller reference */
488 cfs_spin_unlock(&fed->fed_lock);
491 /* expire entries from the end of the list if there are too many
492 * or they are too old */
493 static void filter_fmd_expire_nolock(struct filter_obd *filter,
494 struct filter_export_data *fed,
495 struct filter_mod_data *keep)
497 struct filter_mod_data *fmd, *tmp;
499 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
503 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
504 fed->fed_mod_count < filter->fo_fmd_max_num)
507 cfs_list_del_init(&fmd->fmd_list);
508 filter_fmd_put_nolock(fed, fmd); /* list reference */
512 void filter_fmd_expire(struct obd_export *exp)
514 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
515 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
516 &exp->exp_filter_data, NULL);
517 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
520 /* find specified objid, group in export fmd list.
521 * caller must hold fed_lock and take fmd reference itself */
522 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
523 struct filter_export_data *fed,
524 obd_id objid, obd_seq group)
526 struct filter_mod_data *found = NULL, *fmd;
528 LASSERT_SPIN_LOCKED(&fed->fed_lock);
530 cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
531 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
533 cfs_list_del(&fmd->fmd_list);
534 cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
535 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
540 filter_fmd_expire_nolock(filter, fed, found);
545 /* Find fmd based on objid and group, or return NULL if not found. */
546 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
547 obd_id objid, obd_seq group)
549 struct filter_mod_data *fmd;
551 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
552 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
553 &exp->exp_filter_data, objid, group);
555 fmd->fmd_refcount++; /* caller reference */
556 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
561 /* Find fmd based on objid and group, or create a new one if none is found.
562 * It is possible for this function to return NULL under memory pressure,
563 * or if objid = 0 is passed (which will only cause old entries to expire).
564 * Currently this is not fatal because any fmd state is transient and
565 * may also be freed when it gets sufficiently old. */
566 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
567 obd_id objid, obd_seq group)
569 struct filter_export_data *fed = &exp->exp_filter_data;
570 struct filter_mod_data *found = NULL, *fmd_new = NULL;
572 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
574 cfs_spin_lock(&fed->fed_lock);
575 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
578 cfs_list_add_tail(&fmd_new->fmd_list,
580 fmd_new->fmd_id = objid;
581 fmd_new->fmd_gr = group;
582 fmd_new->fmd_refcount++; /* list reference */
584 fed->fed_mod_count++;
586 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
590 found->fmd_refcount++; /* caller reference */
591 found->fmd_expire = jiffies +
592 exp->exp_obd->u.filter.fo_fmd_max_age;
595 cfs_spin_unlock(&fed->fed_lock);
601 /* drop fmd list reference so it will disappear when last reference is put.
602 * This isn't so critical because it would in fact only affect the one client
603 * that is doing the unlink and at worst we have an stale entry referencing
604 * an object that should never be used again. */
605 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
607 struct filter_mod_data *found = NULL;
609 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
610 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
612 cfs_list_del_init(&found->fmd_list);
613 filter_fmd_put_nolock(&exp->exp_filter_data, found);
615 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
618 #define filter_fmd_drop(exp, objid, group)
621 /* remove all entries from fmd list */
622 static void filter_fmd_cleanup(struct obd_export *exp)
624 struct filter_export_data *fed = &exp->exp_filter_data;
625 struct filter_mod_data *fmd = NULL, *tmp;
627 cfs_spin_lock(&fed->fed_lock);
628 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
629 cfs_list_del_init(&fmd->fmd_list);
630 filter_fmd_put_nolock(fed, fmd);
632 cfs_spin_unlock(&fed->fed_lock);
635 static int filter_init_export(struct obd_export *exp)
638 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
639 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
641 cfs_spin_lock(&exp->exp_lock);
642 exp->exp_connecting = 1;
643 cfs_spin_unlock(&exp->exp_lock);
644 rc = lut_client_alloc(exp);
646 rc = ldlm_init_export(exp);
651 static int filter_free_server_data(struct obd_device_target *obt)
653 lut_fini(NULL, obt->obt_lut);
654 OBD_FREE_PTR(obt->obt_lut);
658 /* assumes caller is already in kernel ctxt */
659 int filter_update_server_data(struct obd_device *obd)
661 struct file *filp = obd->u.obt.obt_rcvd_filp;
662 struct lr_server_data *lsd = class_server_data(obd);
667 CDEBUG(D_INODE, "server uuid : %s\n", lsd->lsd_uuid);
668 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
669 le64_to_cpu(lsd->lsd_last_transno));
670 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
671 le64_to_cpu(lsd->lsd_mount_count));
673 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
675 CERROR("error writing lr_server_data: rc = %d\n", rc);
680 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
683 struct filter_obd *filter = &obd->u.filter;
689 if (filter->fo_last_objid_files[group] == NULL) {
690 CERROR("Object seq "LPU64" not fully setup; not updating "
691 "last_objid\n", group);
695 CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
696 obd->obd_name, filter->fo_last_objids[group], group);
698 tmp = cpu_to_le64(filter->fo_last_objids[group]);
699 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
700 &tmp, sizeof(tmp), &off, force_sync);
702 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
706 extern int ost_handle(struct ptlrpc_request *req);
707 /* assumes caller has already in kernel ctxt */
708 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
710 struct filter_obd *filter = &obd->u.filter;
711 struct lr_server_data *lsd;
712 struct lsd_client_data *lcd = NULL;
713 struct inode *inode = filp->f_dentry->d_inode;
714 unsigned long last_rcvd_size = i_size_read(inode);
715 struct lu_target *lut;
722 /* ensure padding in the struct is the correct size */
723 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
724 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
725 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
726 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
728 /* allocate and initialize lu_target */
732 rc = lut_init(NULL, lut, obd, NULL);
735 lsd = class_server_data(obd);
736 if (last_rcvd_size == 0) {
737 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
739 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
740 lsd->lsd_last_transno = 0;
741 mount_count = lsd->lsd_mount_count = 0;
742 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
743 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
744 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
745 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
746 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
747 /* OBD_COMPAT_OST is set in filter_connect_internal when the
748 * MDS first connects and assigns the OST index number. */
749 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
752 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
754 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
758 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
759 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
760 "using the wrong disk %s. Were the "
761 "/dev/ assignments rearranged?\n",
762 obd->obd_uuid.uuid, lsd->lsd_uuid);
763 GOTO(err_lut, rc = -EINVAL);
765 mount_count = le64_to_cpu(lsd->lsd_mount_count);
766 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
768 /* Assume old last_rcvd format unless I_C_LR is set */
769 if (!(lsd->lsd_feature_incompat &
770 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
771 lsd->lsd_last_transno = lsd->lsd_compat14;
773 /* OBD_COMPAT_OST is set in filter_connect_internal when the
774 * MDS first connects and assigns the OST index number. */
775 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
779 if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
780 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
781 obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
782 ~FILTER_INCOMPAT_SUPP);
783 GOTO(err_lut, rc = -EINVAL);
785 if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
786 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
787 obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
788 ~FILTER_ROCOMPAT_SUPP);
789 /* Do something like remount filesystem read-only */
790 GOTO(err_lut, rc = -EINVAL);
793 start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
795 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
796 obd->obd_name, start_epoch);
797 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
798 obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
799 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
800 obd->obd_name, mount_count + 1);
801 CDEBUG(D_INODE, "%s: server data size: %u\n",
802 obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
803 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
804 obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
805 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
806 obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
807 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
808 obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
809 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
810 last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
811 (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
812 le16_to_cpu(lsd->lsd_client_size));
814 if (!obd->obd_replayable) {
815 CWARN("%s: recovery support OFF\n", obd->obd_name);
821 GOTO(err_client, rc = -ENOMEM);
823 for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
824 off < last_rcvd_size; cl_idx++) {
826 struct obd_export *exp;
827 struct filter_export_data *fed;
829 /* Don't assume off is incremented properly by
830 * fsfilt_read_record(), in case sizeof(*lcd)
831 * isn't the same as lsd->lsd_client_size. */
832 off = le32_to_cpu(lsd->lsd_client_start) +
833 cl_idx * le16_to_cpu(lsd->lsd_client_size);
834 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
836 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
837 LAST_RCVD, cl_idx, off, rc);
838 break; /* read error shouldn't cause startup to fail */
841 if (lcd->lcd_uuid[0] == '\0') {
842 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
847 check_lcd(obd->obd_name, cl_idx, lcd);
849 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
851 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
852 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
853 last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
855 /* These exports are cleaned up by filter_disconnect(), so they
856 * need to be set up like real exports as filter_connect() does.
858 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
860 if (PTR_ERR(exp) == -EALREADY) {
861 /* export already exists, zero out this one */
862 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
866 GOTO(err_client, rc = PTR_ERR(exp));
869 fed = &exp->exp_filter_data;
870 *fed->fed_ted.ted_lcd = *lcd;
871 fed->fed_group = 0; /* will be assigned at connect */
872 filter_export_stats_init(obd, exp, NULL);
873 rc = filter_client_add(obd, exp, cl_idx);
874 /* can't fail for existing client */
875 LASSERTF(rc == 0, "rc = %d\n", rc);
877 /* VBR: set export last committed */
878 exp->exp_last_committed = last_rcvd;
879 cfs_spin_lock(&exp->exp_lock);
880 exp->exp_connecting = 0;
881 exp->exp_in_recovery = 0;
882 cfs_spin_unlock(&exp->exp_lock);
883 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
884 obd->obd_max_recoverable_clients++;
885 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
886 class_export_put(exp);
888 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
889 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
893 obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
895 obd->u.obt.obt_mount_count = mount_count + 1;
896 lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
898 /* save it, so mount count and last_transno is current */
899 rc = filter_update_server_data(obd);
901 GOTO(err_client, rc);
906 class_disconnect_exports(obd);
908 filter_free_server_data(&obd->u.obt);
912 static int filter_cleanup_groups(struct obd_device *obd)
914 struct filter_obd *filter = &obd->u.filter;
916 struct dentry *dentry;
920 if (filter->fo_dentry_O_groups != NULL) {
921 for (i = 0; i < filter->fo_group_count; i++) {
922 dentry = filter->fo_dentry_O_groups[i];
926 OBD_FREE(filter->fo_dentry_O_groups,
927 filter->fo_group_count *
928 sizeof(*filter->fo_dentry_O_groups));
929 filter->fo_dentry_O_groups = NULL;
931 if (filter->fo_last_objid_files != NULL) {
932 for (i = 0; i < filter->fo_group_count; i++) {
933 filp = filter->fo_last_objid_files[i];
937 OBD_FREE(filter->fo_last_objid_files,
938 filter->fo_group_count *
939 sizeof(*filter->fo_last_objid_files));
940 filter->fo_last_objid_files = NULL;
942 if (filter->fo_dentry_O_sub != NULL) {
943 for (i = 0; i < filter->fo_group_count; i++) {
944 for (j = 0; j < filter->fo_subdir_count; j++) {
945 dentry = filter->fo_dentry_O_sub[i].dentry[j];
950 OBD_FREE(filter->fo_dentry_O_sub,
951 filter->fo_group_count *
952 sizeof(*filter->fo_dentry_O_sub));
953 filter->fo_dentry_O_sub = NULL;
955 if (filter->fo_last_objids != NULL) {
956 OBD_FREE(filter->fo_last_objids,
957 filter->fo_group_count *
958 sizeof(*filter->fo_last_objids));
959 filter->fo_last_objids = NULL;
961 if (filter->fo_dentry_O != NULL) {
962 f_dput(filter->fo_dentry_O);
963 filter->fo_dentry_O = NULL;
968 static int filter_update_last_group(struct obd_device *obd, int group)
970 struct filter_obd *filter = &obd->u.filter;
971 struct file *filp = NULL;
972 int last_group = 0, rc;
976 if (group <= filter->fo_committed_group)
979 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
983 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
987 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
989 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
993 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
994 obd->obd_name, last_group, group);
998 /* must be sync: bXXXX */
999 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1001 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1005 filter->fo_committed_group = group;
1008 filp_close(filp, 0);
1012 static int filter_read_group_internal(struct obd_device *obd, int group,
1015 struct filter_obd *filter = &obd->u.filter;
1016 __u64 *new_objids = NULL;
1017 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1018 struct dentry **new_groups = NULL;
1019 struct file **new_files = NULL;
1020 struct dentry *dentry;
1022 int old_count = filter->fo_group_count, rc, stage = 0, i;
1026 int len = group + 1;
1028 snprintf(name, 24, "%d", group);
1032 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1034 if (IS_ERR(dentry)) {
1035 CERROR("Cannot lookup expected object group %d: %ld\n",
1036 group, PTR_ERR(dentry));
1037 RETURN(PTR_ERR(dentry));
1040 dentry = simple_mkdir(filter->fo_dentry_O,
1041 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1042 if (IS_ERR(dentry)) {
1043 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1045 RETURN(PTR_ERR(dentry));
1050 snprintf(name, 24, "O/%d/LAST_ID", group);
1052 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1054 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1055 GOTO(cleanup, rc = PTR_ERR(filp));
1059 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1061 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1065 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1066 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1067 if (tmp_subdirs == NULL)
1068 GOTO(cleanup, rc = -ENOMEM);
1071 for (i = 0; i < filter->fo_subdir_count; i++) {
1073 snprintf(dir, sizeof(dir), "d%u", i);
1075 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1076 obd->u.obt.obt_vfsmnt,
1078 if (IS_ERR(tmp_subdirs->dentry[i])) {
1079 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1080 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1085 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1086 tmp_subdirs->dentry[i]);
1090 /* 'group' is an index; we need an array of length 'group + 1' */
1091 if (group + 1 > old_count) {
1092 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1093 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1094 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1095 OBD_ALLOC(new_files, len * sizeof(*new_files));
1097 if (new_objids == NULL || new_subdirs == NULL ||
1098 new_groups == NULL || new_files == NULL)
1099 GOTO(cleanup, rc = -ENOMEM);
1102 memcpy(new_objids, filter->fo_last_objids,
1103 old_count * sizeof(*new_objids));
1104 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1105 old_count * sizeof(*new_subdirs));
1106 memcpy(new_groups, filter->fo_dentry_O_groups,
1107 old_count * sizeof(*new_groups));
1108 memcpy(new_files, filter->fo_last_objid_files,
1109 old_count * sizeof(*new_files));
1111 OBD_FREE(filter->fo_last_objids,
1112 old_count * sizeof(*new_objids));
1113 OBD_FREE(filter->fo_dentry_O_sub,
1114 old_count * sizeof(*new_subdirs));
1115 OBD_FREE(filter->fo_dentry_O_groups,
1116 old_count * sizeof(*new_groups));
1117 OBD_FREE(filter->fo_last_objid_files,
1118 old_count * sizeof(*new_files));
1120 filter->fo_last_objids = new_objids;
1121 filter->fo_dentry_O_sub = new_subdirs;
1122 filter->fo_dentry_O_groups = new_groups;
1123 filter->fo_last_objid_files = new_files;
1124 filter->fo_group_count = len;
1127 filter->fo_dentry_O_groups[group] = dentry;
1128 filter->fo_last_objid_files[group] = filp;
1129 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1130 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1131 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1134 filter_update_last_group(obd, group);
1136 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1137 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1138 rc = filter_update_last_objid(obd, group, 1);
1142 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1143 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1144 obd->obd_name, group, last_objid);
1149 if (new_objids != NULL)
1150 OBD_FREE(new_objids, len * sizeof(*new_objids));
1151 if (new_subdirs != NULL)
1152 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1153 if (new_groups != NULL)
1154 OBD_FREE(new_groups, len * sizeof(*new_groups));
1155 if (new_files != NULL)
1156 OBD_FREE(new_files, len * sizeof(*new_files));
1158 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1159 for (i = 0; i < filter->fo_subdir_count; i++) {
1160 if (tmp_subdirs->dentry[i] != NULL)
1161 dput(tmp_subdirs->dentry[i]);
1163 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1166 filp_close(filp, 0);
1173 static int filter_read_groups(struct obd_device *obd, int last_group,
1176 struct filter_obd *filter = &obd->u.filter;
1177 int old_count, group, rc = 0;
1179 cfs_down(&filter->fo_init_lock);
1180 old_count = filter->fo_group_count;
1181 for (group = old_count; group <= last_group; group++) {
1182 rc = filter_read_group_internal(obd, group, create);
1186 cfs_up(&filter->fo_init_lock);
1190 /* FIXME: object groups */
1191 static int filter_prep_groups(struct obd_device *obd)
1193 struct filter_obd *filter = &obd->u.filter;
1194 struct dentry *O_dentry;
1196 int last_group, rc = 0, cleanup_phase = 0;
1200 O_dentry = simple_mkdir(current->fs->pwd, obd->u.obt.obt_vfsmnt,
1202 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1203 if (IS_ERR(O_dentry)) {
1204 rc = PTR_ERR(O_dentry);
1205 CERROR("cannot open/create O: rc = %d\n", rc);
1208 filter->fo_dentry_O = O_dentry;
1209 cleanup_phase = 1; /* O_dentry */
1211 /* we have to initialize all groups before first connections from
1212 * clients because they may send create/destroy for any group -bzzz */
1213 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1215 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1216 GOTO(cleanup, rc = PTR_ERR(filp));
1218 cleanup_phase = 2; /* filp */
1220 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1222 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1227 last_group = FID_SEQ_OST_MDT0;
1229 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1230 FID_SEQ_OST_MDT0, last_group);
1231 filter->fo_committed_group = last_group;
1232 rc = filter_read_groups(obd, last_group, 1);
1236 filp_close(filp, 0);
1240 switch (cleanup_phase) {
1242 filp_close(filp, 0);
1244 filter_cleanup_groups(obd);
1245 f_dput(filter->fo_dentry_O);
1246 filter->fo_dentry_O = NULL;
1254 /* setup the object store with correct subdirectories */
1255 static int filter_prep(struct obd_device *obd)
1257 struct lvfs_run_ctxt saved;
1258 struct filter_obd *filter = &obd->u.filter;
1260 struct inode *inode;
1264 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1265 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1266 if (!file || IS_ERR(file)) {
1268 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1272 obd->u.obt.obt_rcvd_filp = file;
1273 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1274 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1275 file->f_dentry->d_inode->i_mode);
1276 GOTO(err_filp, rc = -ENOENT);
1279 inode = file->f_dentry->d_parent->d_inode;
1280 /* We use i_op->unlink directly in filter_vfs_unlink() */
1281 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1282 CERROR("%s: filesystem does not support create/unlink ops\n",
1284 GOTO(err_filp, rc = -EOPNOTSUPP);
1287 rc = filter_init_server_data(obd, file);
1289 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1292 LASSERT(obd->u.obt.obt_lut);
1293 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1295 /* open and create health check io file*/
1296 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1299 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1301 GOTO(err_server_data, rc);
1303 filter->fo_obt.obt_health_check_filp = file;
1304 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1305 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1306 file->f_dentry->d_inode->i_mode);
1307 GOTO(err_health_check, rc = -ENOENT);
1309 rc = lvfs_check_io_health(obd, file);
1311 GOTO(err_health_check, rc);
1313 rc = filter_prep_groups(obd);
1315 GOTO(err_health_check, rc);
1317 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1322 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1323 CERROR("can't close %s after error\n", HEALTH_CHECK);
1324 filter->fo_obt.obt_health_check_filp = NULL;
1326 target_recovery_fini(obd);
1327 filter_free_server_data(&obd->u.obt);
1329 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1330 CERROR("can't close %s after error\n", LAST_RCVD);
1331 obd->u.obt.obt_rcvd_filp = NULL;
1335 /* cleanup the filter: write last used object id to status file */
1336 static void filter_post(struct obd_device *obd)
1338 struct lvfs_run_ctxt saved;
1339 struct filter_obd *filter = &obd->u.filter;
1342 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1343 * best to start a transaction with h_sync, because we removed this
1346 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1347 rc = filter_update_server_data(obd);
1349 CERROR("error writing server data: rc = %d\n", rc);
1351 for (i = 0; i < filter->fo_group_count; i++) {
1352 rc = filter_update_last_objid(obd, i,
1353 (i == filter->fo_group_count - 1));
1355 CERROR("error writing group %d lastobjid: rc = %d\n",
1359 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1360 obd->u.obt.obt_rcvd_filp = NULL;
1362 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1364 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1365 filter->fo_obt.obt_health_check_filp = NULL;
1367 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1369 filter_cleanup_groups(obd);
1370 filter_free_server_data(&obd->u.obt);
1371 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1373 filter_free_capa_keys(filter);
1374 cleanup_capa_hash(filter->fo_capa_hash);
1377 static void filter_set_last_id(struct filter_obd *filter,
1378 obd_id id, obd_seq group)
1380 LASSERT(group <= filter->fo_group_count);
1382 cfs_spin_lock(&filter->fo_objidlock);
1383 filter->fo_last_objids[group] = id;
1384 cfs_spin_unlock(&filter->fo_objidlock);
1387 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1390 LASSERT(group <= filter->fo_group_count);
1391 LASSERT(filter->fo_last_objids != NULL);
1393 /* FIXME: object groups */
1394 cfs_spin_lock(&filter->fo_objidlock);
1395 id = filter->fo_last_objids[group];
1396 cfs_spin_unlock(&filter->fo_objidlock);
1400 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1402 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1406 /* We never dget the object parent, so DON'T dput it either */
1407 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1409 struct filter_obd *filter = &obd->u.filter;
1410 struct filter_subdirs *subdirs;
1411 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1413 if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1414 return filter->fo_dentry_O_groups[group];
1416 subdirs = &filter->fo_dentry_O_sub[group];
1417 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1420 /* We never dget the object parent, so DON'T dput it either */
1421 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1424 unsigned long now = jiffies;
1425 struct dentry *dparent = filter_parent(obd, group, objid);
1428 if (IS_ERR(dparent))
1430 if (dparent == NULL)
1431 return ERR_PTR(-ENOENT);
1433 rc = filter_lock_dentry(obd, dparent);
1434 fsfilt_check_slow(obd, now, "parent lock");
1435 return rc ? ERR_PTR(rc) : dparent;
1438 /* We never dget the object parent, so DON'T dput it either */
1439 static void filter_parent_unlock(struct dentry *dparent)
1441 UNLOCK_INODE_MUTEX(dparent->d_inode);
1444 /* How to get files, dentries, inodes from object id's.
1446 * If dir_dentry is passed, the caller has already locked the parent
1447 * appropriately for this operation (normally a write lock). If
1448 * dir_dentry is NULL, we do a read lock while we do the lookup to
1449 * avoid races with create/destroy and such changing the directory
1450 * internal to the filesystem code. */
1451 struct dentry *filter_fid2dentry(struct obd_device *obd,
1452 struct dentry *dir_dentry,
1453 obd_seq group, obd_id id)
1455 struct dentry *dparent = dir_dentry;
1456 struct dentry *dchild;
1461 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1462 obd->u.filter.fo_destroys_in_progress == 0) {
1463 /* don't fail lookups for orphan recovery, it causes
1464 * later LBUGs when objects still exist during precreate */
1465 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1466 RETURN(ERR_PTR(-ENOENT));
1469 CERROR("fatal: invalid object id 0\n");
1470 RETURN(ERR_PTR(-ESTALE));
1473 len = sprintf(name, LPU64, id);
1474 if (dir_dentry == NULL) {
1475 dparent = filter_parent_lock(obd, group, id);
1476 if (IS_ERR(dparent)) {
1477 CERROR("%s: error getting object "POSTID
1478 " parent: rc %ld\n", obd->obd_name,
1479 id, group, PTR_ERR(dparent));
1483 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1484 dparent->d_name.len, dparent->d_name.name, name);
1485 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1486 if (dir_dentry == NULL)
1487 filter_parent_unlock(dparent);
1488 if (IS_ERR(dchild)) {
1489 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1494 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1495 CERROR("%s: got bad object "LPU64" inode %lu\n",
1496 obd->obd_name, id, dchild->d_inode->i_ino);
1498 RETURN(ERR_PTR(-ENOENT));
1501 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1502 name, dchild, atomic_read(&dchild->d_count));
1504 LASSERT(atomic_read(&dchild->d_count) > 0);
1509 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1510 obd_id group, struct lustre_handle *lockh)
1512 int flags = LDLM_AST_DISCARD_DATA, rc;
1513 struct ldlm_res_id res_id;
1514 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1517 osc_build_res_name(objid, group, &res_id);
1518 /* Tell the clients that the object is gone now and that they should
1519 * throw away any cached pages. */
1520 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1521 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1522 ldlm_completion_ast, NULL, NULL, 0, NULL,
1529 static void filter_fini_destroy(struct obd_device *obd,
1530 struct lustre_handle *lockh)
1532 if (lustre_handle_is_used(lockh))
1533 ldlm_lock_decref(lockh, LCK_PW);
1536 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1537 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1538 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1539 * reverse. Caller must take i_sem before starting the transaction and we
1540 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1541 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1542 struct vfsmount *mnt)
1547 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1548 * itself we already hold dir->i_mutex for child create/unlink ops */
1549 LASSERT(dentry->d_inode != NULL);
1550 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1551 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1555 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1556 GOTO(out, rc = -ENOENT);
1558 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1563 GOTO(out, rc = -EPERM);
1565 /* check_sticky() */
1566 if ((dentry->d_inode->i_uid != current->fsuid &&
1567 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1568 IS_IMMUTABLE(dentry->d_inode))
1569 GOTO(out, rc = -EPERM);
1571 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1572 * that was done because of journal_start (which is already done
1573 * here) or some other ordering issue. */
1576 rc = ll_security_inode_unlink(dir, dentry, mnt);
1580 rc = dir->i_op->unlink(dir, dentry);
1582 /* need to drop i_mutex before we lose inode reference */
1583 UNLOCK_INODE_MUTEX(dentry->d_inode);
1590 /* Caller must hold LCK_PW on parent and push us into kernel context.
1591 * Caller must hold child i_mutex, we drop it always.
1592 * Caller is also required to ensure that dchild->d_inode exists. */
1593 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1594 obd_seq group, struct dentry *dparent,
1595 struct dentry *dchild)
1597 struct inode *inode = dchild->d_inode;
1600 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1601 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1602 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1603 (unsigned long)inode->i_nlink,
1604 atomic_read(&inode->i_count));
1607 rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1609 CERROR("error unlinking objid %.*s: rc %d\n",
1610 dchild->d_name.len, dchild->d_name.name, rc);
1614 struct filter_intent_args {
1615 struct ldlm_lock **victim;
1620 static enum interval_iter filter_intent_cb(struct interval_node *n,
1623 struct ldlm_interval *node = (struct ldlm_interval *)n;
1624 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1625 __u64 size = arg->size;
1626 struct ldlm_lock **v = arg->victim;
1627 struct ldlm_lock *lck;
1629 /* If the interval is lower than the current file size,
1631 if (interval_high(n) <= size)
1632 return INTERVAL_ITER_STOP;
1634 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1635 /* Don't send glimpse ASTs to liblustre clients.
1636 * They aren't listening for them, and they do
1637 * entirely synchronous I/O anyways. */
1638 if (lck->l_export == NULL ||
1639 lck->l_export->exp_libclient == 1)
1642 if (*arg->liblustre)
1643 *arg->liblustre = 0;
1646 *v = LDLM_LOCK_GET(lck);
1647 } else if ((*v)->l_policy_data.l_extent.start <
1648 lck->l_policy_data.l_extent.start) {
1649 LDLM_LOCK_RELEASE(*v);
1650 *v = LDLM_LOCK_GET(lck);
1653 /* the same policy group - every lock has the
1654 * same extent, so needn't do it any more */
1658 return INTERVAL_ITER_CONT;
1661 static int filter_intent_policy(struct ldlm_namespace *ns,
1662 struct ldlm_lock **lockp, void *req_cookie,
1663 ldlm_mode_t mode, int flags, void *data)
1665 CFS_LIST_HEAD(rpc_list);
1666 struct ptlrpc_request *req = req_cookie;
1667 struct ldlm_lock *lock = *lockp, *l = NULL;
1668 struct ldlm_resource *res = lock->l_resource;
1669 ldlm_processing_policy policy;
1670 struct ost_lvb *res_lvb, *reply_lvb;
1671 struct ldlm_reply *rep;
1673 int idx, rc, tmpflags = 0, only_liblustre = 1;
1674 struct ldlm_interval_tree *tree;
1675 struct filter_intent_args arg;
1676 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1677 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1678 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1681 policy = ldlm_get_processing_policy(res);
1682 LASSERT(policy != NULL);
1683 LASSERT(req != NULL);
1685 rc = lustre_pack_reply(req, 3, repsize, NULL);
1687 RETURN(req->rq_status = rc);
1689 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1690 LASSERT(rep != NULL);
1692 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1693 sizeof(*reply_lvb));
1694 LASSERT(reply_lvb != NULL);
1696 //fixup_handle_for_resent_req(req, lock, &lockh);
1698 /* Call the extent policy function to see if our request can be
1699 * granted, or is blocked.
1700 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1701 * lock, and should not be granted if the lock will be blocked.
1704 LASSERT(ns == res->lr_namespace);
1706 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1707 check_res_locked(res);
1709 /* FIXME: we should change the policy function slightly, to not make
1710 * this list at all, since we just turn around and free it */
1711 while (!cfs_list_empty(&rpc_list)) {
1712 struct ldlm_lock *wlock =
1713 cfs_list_entry(rpc_list.next, struct ldlm_lock,
1715 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1716 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1717 lock->l_flags &= ~LDLM_FL_CP_REQD;
1718 cfs_list_del_init(&wlock->l_cp_ast);
1719 LDLM_LOCK_RELEASE(wlock);
1722 /* The lock met with no resistance; we're finished. */
1723 if (rc == LDLM_ITER_CONTINUE) {
1724 /* do not grant locks to the liblustre clients: they cannot
1725 * handle ASTs robustly. We need to do this while still
1726 * holding ns_lock to avoid the lock remaining on the res_link
1727 * list (and potentially being added to l_pending_list by an
1728 * AST) when we are going to drop this lock ASAP. */
1729 if (lock->l_export->exp_libclient ||
1730 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1731 ldlm_resource_unlink_lock(lock);
1732 err = ELDLM_LOCK_ABORTED;
1734 err = ELDLM_LOCK_REPLACED;
1740 /* Do not grant any lock, but instead send GL callbacks. The extent
1741 * policy nicely created a list of all PW locks for us. We will choose
1742 * the highest of those which are larger than the size in the LVB, if
1743 * any, and perform a glimpse callback. */
1744 res_lvb = res->lr_lvb_data;
1745 LASSERT(res_lvb != NULL);
1746 *reply_lvb = *res_lvb;
1749 * ->ns_lock guarantees that no new locks are granted, and,
1750 * therefore, that res->lr_lvb_data cannot increase beyond the
1751 * end of already granted lock. As a result, it is safe to
1752 * check against "stale" reply_lvb->lvb_size value without
1755 arg.size = reply_lvb->lvb_size;
1757 arg.liblustre = &only_liblustre;
1758 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1759 tree = &res->lr_itree[idx];
1760 if (tree->lit_mode == LCK_PR)
1763 interval_iterate_reverse(tree->lit_root,
1764 filter_intent_cb, &arg);
1768 /* There were no PW locks beyond the size in the LVB; finished. */
1770 if (only_liblustre) {
1771 /* If we discovered a liblustre client with a PW lock,
1772 * however, the LVB may be out of date! The LVB is
1773 * updated only on glimpse (which we don't do for
1774 * liblustre clients) and cancel (which the client
1775 * obviously has not yet done). So if it has written
1776 * data but kept the lock, the LVB is stale and needs
1777 * to be updated from disk.
1779 * Of course, this will all disappear when we switch to
1780 * taking liblustre locks on the OST. */
1781 ldlm_res_lvbo_update(res, NULL, 1);
1783 RETURN(ELDLM_LOCK_ABORTED);
1787 * This check is for lock taken in filter_prepare_destroy() that does
1788 * not have l_glimpse_ast set. So the logic is: if there is a lock
1789 * with no l_glimpse_ast set, this object is being destroyed already.
1791 * Hence, if you are grabbing DLM locks on the server, always set
1792 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1794 if (l->l_glimpse_ast == NULL) {
1795 /* We are racing with unlink(); just return -ENOENT */
1796 rep->lock_policy_res1 = -ENOENT;
1800 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1801 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1802 /* Update the LVB from disk if the AST failed (this is a legal race) */
1804 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1805 * sending ast is not handled. This can result in lost client writes.
1808 ldlm_res_lvbo_update(res, NULL, 1);
1811 *reply_lvb = *res_lvb;
1815 LDLM_LOCK_RELEASE(l);
1817 RETURN(ELDLM_LOCK_ABORTED);
1821 * per-obd_device iobuf pool.
1823 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1824 * allocations in write-path has to be minimized (see bug 5137).
1826 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1827 * OST threads (see ost_thread_{init,done}()).
1829 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1830 * at the OST layer there are only (potentially) multiple obd_device of type
1831 * unknown at the time of OST thread creation.
1833 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1834 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1835 * it's very own iobuf.
1839 * filter_kiobuf_pool_init()
1841 * filter_kiobuf_pool_done()
1843 * filter_iobuf_get()
1845 * operate on this array. They are "generic" in a sense that they don't depend
1846 * on actual type of iobuf's (the latter depending on Linux kernel version).
1850 * destroy pool created by filter_iobuf_pool_init
1852 static void filter_iobuf_pool_done(struct filter_obd *filter)
1854 struct filter_iobuf **pool;
1859 pool = filter->fo_iobuf_pool;
1861 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1862 if (pool[i] != NULL)
1863 filter_free_iobuf(pool[i]);
1865 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1866 filter->fo_iobuf_pool = NULL;
1871 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1873 struct filter_obd *filter = &obd->u.filter;
1874 struct sptlrpc_rule_set tmp_rset;
1877 sptlrpc_rule_set_init(&tmp_rset);
1878 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1880 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1885 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1887 cfs_write_lock(&filter->fo_sptlrpc_lock);
1888 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1889 filter->fo_sptlrpc_rset = tmp_rset;
1890 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1896 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1898 static int filter_iobuf_pool_init(struct filter_obd *filter)
1905 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1907 if (filter->fo_iobuf_pool == NULL)
1910 filter->fo_iobuf_count = OSS_THREADS_MAX;
1915 /* Return iobuf allocated for @thread_id. We don't know in advance how
1916 * many threads there will be so we allocate a large empty array and only
1917 * fill in those slots that are actually in use.
1918 * If we haven't allocated a pool entry for this thread before, do so now. */
1919 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1921 int thread_id = (oti && oti->oti_thread) ?
1922 oti->oti_thread->t_id : -1;
1923 struct filter_iobuf *pool = NULL;
1924 struct filter_iobuf **pool_place = NULL;
1926 if (thread_id >= 0) {
1927 LASSERT(thread_id < filter->fo_iobuf_count);
1928 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1931 if (unlikely(pool == NULL)) {
1932 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1933 PTLRPC_MAX_BRW_PAGES);
1934 if (pool_place != NULL)
1941 /* mount the file system (secretly). lustre_cfg parameters are:
1944 * 3 = flags: failover=f, failout=n
1947 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1950 struct filter_obd *filter = &obd->u.filter;
1951 struct vfsmount *mnt;
1952 struct lustre_mount_info *lmi;
1953 struct obd_uuid uuid;
1961 if (lcfg->lcfg_bufcount < 3 ||
1962 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1963 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1966 lmi = server_get_mount(obd->obd_name);
1968 /* We already mounted in lustre_fill_super.
1969 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1970 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1972 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1974 /* gets recovery timeouts from mount data */
1975 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1976 obd->obd_recovery_timeout =
1977 lsi->lsi_lmd->lmd_recovery_time_soft;
1978 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1979 obd->obd_recovery_time_hard =
1980 lsi->lsi_lmd->lmd_recovery_time_hard;
1982 /* old path - used by lctl */
1983 CERROR("Using old MDS mount method\n");
1984 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1985 MS_NOATIME|MS_NODIRATIME,
1986 lustre_cfg_string(lcfg, 1), option);
1989 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1990 lustre_cfg_string(lcfg, 1), rc);
1994 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1996 if (IS_ERR(obd->obd_fsops))
1997 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1999 rc = filter_iobuf_pool_init(filter);
2003 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2004 CERROR("%s: Underlying device is marked as read-only. "
2005 "Setup failed\n", obd->obd_name);
2006 GOTO(err_ops, rc = -EROFS);
2009 /* failover is the default */
2010 obd->obd_replayable = 1;
2012 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2013 str = lustre_cfg_string(lcfg, 3);
2014 if (strchr(str, 'n')) {
2015 CWARN("%s: recovery disabled\n", obd->obd_name);
2016 obd->obd_replayable = 0;
2020 obd->u.obt.obt_vfsmnt = mnt;
2021 obd->u.obt.obt_sb = mnt->mnt_sb;
2022 obd->u.obt.obt_magic = OBT_MAGIC;
2023 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2024 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2026 fsfilt_setup(obd, obd->u.obt.obt_sb);
2028 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2029 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2030 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2031 obd->obd_lvfs_ctxt.fs = get_ds();
2032 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2034 cfs_init_mutex(&filter->fo_init_lock);
2035 filter->fo_committed_group = 0;
2036 filter->fo_destroys_in_progress = 0;
2037 for (i = 0; i < 32; i++)
2038 cfs_sema_init(&filter->fo_create_locks[i], 1);
2040 cfs_spin_lock_init(&filter->fo_objidlock);
2041 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2042 cfs_sema_init(&filter->fo_alloc_lock, 1);
2043 init_brw_stats(&filter->fo_filter_stats);
2044 filter->fo_read_cache = 1; /* enable read-only cache by default */
2045 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2046 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2047 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2048 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2049 filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2050 filter_slc_set(filter); /* initialize sync on lock cancel */
2052 rc = filter_prep(obd);
2056 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2057 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2059 filter->fo_fl_oss_capa = 1;
2061 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2062 filter->fo_capa_hash = init_capa_hash();
2063 if (filter->fo_capa_hash == NULL)
2064 GOTO(err_post, rc = -ENOMEM);
2066 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2067 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2068 LDLM_NAMESPACE_GREEDY);
2069 if (obd->obd_namespace == NULL)
2070 GOTO(err_post, rc = -ENOMEM);
2071 obd->obd_namespace->ns_lvbp = obd;
2072 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2073 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2075 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2076 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2078 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2080 CERROR("failed to setup llogging subsystems\n");
2084 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2085 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2086 /* do this after llog being initialized */
2087 filter_adapt_sptlrpc_conf(obd, 1);
2089 rc = lquota_setup(filter_quota_interface_ref, obd);
2093 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2094 if (q->max_sectors < q->max_hw_sectors &&
2095 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2096 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2097 "for larger I/O requests: max_sectors = %u "
2098 "could be up to max_hw_sectors=%u\n",
2099 obd->obd_name, mnt->mnt_sb->s_id,
2100 q->max_sectors, q->max_hw_sectors);
2102 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2103 if (uuid_ptr != NULL) {
2104 class_uuid_unparse(uuid_ptr, &uuid);
2110 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2111 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2112 obd->obd_name, label ?: str, lmi ? "on " : "",
2113 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2114 obd->obd_replayable ? "enabled" : "disabled");
2116 if (obd->obd_recovering)
2117 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2118 "or until %d client%s reconnect%s\n",
2120 obd->obd_recovery_timeout / 60,
2121 obd->obd_recovery_timeout % 60,
2122 obd->obd_max_recoverable_clients,
2123 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2124 (obd->obd_max_recoverable_clients == 1) ? "s": "");
2132 fsfilt_put_ops(obd->obd_fsops);
2133 filter_iobuf_pool_done(filter);
2135 server_put_mount(obd->obd_name, mnt);
2136 obd->u.obt.obt_sb = 0;
2140 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2142 struct lprocfs_static_vars lvars;
2147 CLASSERT(offsetof(struct obd_device, u.obt) ==
2148 offsetof(struct obd_device, u.filter.fo_obt));
2150 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2153 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2154 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2157 addr = (unsigned long)cfs_page_address(page);
2158 clear_page((void *)addr);
2160 /* lprocfs must be setup before the filter so state can be safely added
2161 * to /proc incrementally as the filter is setup */
2162 lprocfs_filter_init_vars(&lvars);
2163 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2164 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2165 /* Init obdfilter private stats here */
2166 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2167 LPROCFS_CNTR_AVGMINMAX,
2168 "read_bytes", "bytes");
2169 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2170 LPROCFS_CNTR_AVGMINMAX,
2171 "write_bytes", "bytes");
2172 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2173 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2174 "get_page", "usec");
2175 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2176 LPROCFS_CNTR_AVGMINMAX,
2177 "get_page_failures", "num");
2178 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2179 LPROCFS_CNTR_AVGMINMAX,
2180 "cache_access", "pages");
2181 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2182 LPROCFS_CNTR_AVGMINMAX,
2183 "cache_hit", "pages");
2184 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2185 LPROCFS_CNTR_AVGMINMAX,
2186 "cache_miss", "pages");
2188 lproc_filter_attach_seqstat(obd);
2189 obd->obd_proc_exports_entry = lprocfs_register("exports",
2190 obd->obd_proc_entry,
2192 if (IS_ERR(obd->obd_proc_exports_entry)) {
2193 rc = PTR_ERR(obd->obd_proc_exports_entry);
2194 CERROR("error %d setting up lprocfs for %s\n",
2196 obd->obd_proc_exports_entry = NULL;
2199 if (obd->obd_proc_exports_entry)
2200 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2201 lprocfs_nid_stats_clear_read,
2202 lprocfs_nid_stats_clear_write, obd, NULL);
2204 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2205 LUSTRE_CFG_BUFLEN(lcfg, 4));
2206 rc = filter_common_setup(obd, lcfg, (void *)addr);
2207 OBD_PAGE_FREE(page);
2210 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2211 lprocfs_free_per_client_stats(obd);
2212 lprocfs_free_obd_stats(obd);
2213 lprocfs_obd_cleanup(obd);
2219 static struct llog_operations filter_mds_ost_repl_logops;
2221 static struct llog_operations filter_size_orig_logops = {
2222 .lop_setup = llog_obd_origin_setup,
2223 .lop_cleanup = llog_obd_origin_cleanup,
2224 .lop_add = llog_obd_origin_add
2227 static int filter_olg_fini(struct obd_llog_group *olg)
2229 struct llog_ctxt *ctxt;
2230 int rc = 0, rc2 = 0;
2233 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2235 rc = llog_cleanup(ctxt);
2237 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2239 rc2 = llog_cleanup(ctxt);
2244 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2246 rc2 = llog_cleanup(ctxt);
2255 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2256 struct obd_device *tgt)
2261 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2262 &filter_mds_ost_repl_logops);
2266 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2267 &filter_size_orig_logops);
2273 filter_olg_fini(olg);
2278 * Init the default olg, which is embeded in the obd_device, for filter.
2281 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2282 struct obd_device *tgt)
2284 struct filter_obd *filter = &obd->u.filter;
2285 struct llog_ctxt *ctxt;
2289 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2290 if (!filter->fo_lcm)
2293 filter_mds_ost_repl_logops = llog_client_ops;
2294 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2295 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2296 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2298 rc = filter_olg_init(obd, olg, tgt);
2300 GOTO(cleanup_lcm, rc);
2302 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2305 GOTO(cleanup_olg, rc);
2307 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2309 CERROR("Can't get ctxt for %p:%x\n", olg,
2310 LLOG_MDS_OST_REPL_CTXT);
2311 GOTO(cleanup_olg, rc = -ENODEV);
2313 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2314 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2315 llog_ctxt_put(ctxt);
2319 filter_olg_fini(olg);
2321 llog_recov_thread_fini(filter->fo_lcm, 1);
2322 filter->fo_lcm = NULL;
2327 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2328 struct obd_device *tgt, int *index)
2330 struct filter_obd *filter = &obd->u.filter;
2331 struct llog_ctxt *ctxt;
2335 LASSERT(olg != NULL);
2336 if (olg == &obd->obd_olg)
2337 return filter_default_olg_init(obd, olg, tgt);
2339 LASSERT(filter->fo_lcm != NULL);
2340 rc = filter_olg_init(obd, olg, tgt);
2343 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2345 CERROR("Can't get ctxt for %p:%x\n", olg,
2346 LLOG_MDS_OST_REPL_CTXT);
2347 filter_olg_fini(olg);
2350 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2351 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2352 llog_ctxt_put(ctxt);
2356 static int filter_llog_finish(struct obd_device *obd, int count)
2358 struct filter_obd *filter = &obd->u.filter;
2359 struct llog_ctxt *ctxt;
2362 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2365 * Make sure that no cached llcds left in recov_thread.
2366 * We actually do sync in disconnect time, but disconnect
2367 * may not come being marked rq_no_resend = 1.
2369 llog_sync(ctxt, NULL);
2372 * Balance class_import_get() in llog_receptor_accept().
2373 * This is safe to do, as llog is already synchronized
2374 * and its import may go.
2376 cfs_mutex_down(&ctxt->loc_sem);
2377 if (ctxt->loc_imp) {
2378 class_import_put(ctxt->loc_imp);
2379 ctxt->loc_imp = NULL;
2381 cfs_mutex_up(&ctxt->loc_sem);
2382 llog_ctxt_put(ctxt);
2385 if (filter->fo_lcm) {
2386 cfs_mutex_down(&ctxt->loc_sem);
2387 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2388 filter->fo_lcm = NULL;
2389 cfs_mutex_up(&ctxt->loc_sem);
2391 RETURN(filter_olg_fini(&obd->obd_olg));
2395 * Find the group llog according to group index in the llog group list.
2397 static struct obd_llog_group *
2398 filter_find_olg_internal(struct filter_obd *filter, int group)
2400 struct obd_llog_group *olg;
2402 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2403 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2404 if (olg->olg_seq == group)
2411 * Find the group llog according to group index on the filter
2413 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2415 struct obd_llog_group *olg = NULL;
2416 struct filter_obd *filter;
2418 filter = &obd->u.filter;
2420 if (group == FID_SEQ_LLOG)
2421 RETURN(&obd->obd_olg);
2423 cfs_spin_lock(&filter->fo_llog_list_lock);
2424 olg = filter_find_olg_internal(filter, group);
2425 cfs_spin_unlock(&filter->fo_llog_list_lock);
2430 * Find the llog_group of the filter according to the group. If it can not
2431 * find, create the llog_group, which only happens when mds is being synced
2434 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2436 struct obd_llog_group *olg = NULL;
2437 struct filter_obd *filter;
2440 filter = &obd->u.filter;
2442 if (group == FID_SEQ_LLOG)
2443 RETURN(&obd->obd_olg);
2445 cfs_spin_lock(&filter->fo_llog_list_lock);
2446 olg = filter_find_olg_internal(filter, group);
2448 if (olg->olg_initializing) {
2449 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2451 GOTO(out_unlock, olg);
2456 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2458 llog_group_init(olg, group);
2459 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2460 olg->olg_initializing = 1;
2461 cfs_spin_unlock(&filter->fo_llog_list_lock);
2463 rc = obd_llog_init(obd, olg, obd, NULL);
2465 cfs_spin_lock(&filter->fo_llog_list_lock);
2466 cfs_list_del(&olg->olg_list);
2467 cfs_spin_unlock(&filter->fo_llog_list_lock);
2469 GOTO(out, olg = ERR_PTR(-ENOMEM));
2471 cfs_spin_lock(&filter->fo_llog_list_lock);
2472 olg->olg_initializing = 0;
2473 cfs_spin_unlock(&filter->fo_llog_list_lock);
2474 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2475 obd->obd_name, group, olg);
2480 cfs_spin_unlock(&filter->fo_llog_list_lock);
2484 static int filter_llog_connect(struct obd_export *exp,
2485 struct llogd_conn_body *body)
2487 struct obd_device *obd = exp->exp_obd;
2488 struct llog_ctxt *ctxt;
2489 struct obd_llog_group *olg;
2493 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2494 obd->obd_name, body->lgdc_logid.lgl_oid,
2495 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2497 olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2499 CERROR(" %s: can not find olg of group %d\n",
2500 obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2503 llog_group_set_export(olg, exp);
2505 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2506 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2507 body->lgdc_ctxt_idx);
2509 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2510 obd->obd_name, body->lgdc_logid.lgl_oid,
2511 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2513 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2514 obd->u.filter.fo_mds_ost_sync = 1;
2515 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2516 rc = llog_connect(ctxt, &body->lgdc_logid,
2517 &body->lgdc_gen, NULL);
2518 llog_ctxt_put(ctxt);
2520 CERROR("failed to connect rc %d idx %d\n", rc,
2521 body->lgdc_ctxt_idx);
2526 static int filter_llog_preclean(struct obd_device *obd)
2528 struct obd_llog_group *olg, *tmp;
2529 struct filter_obd *filter;
2530 cfs_list_t remove_list;
2534 rc = obd_llog_finish(obd, 0);
2536 CERROR("failed to cleanup llogging subsystem\n");
2538 filter = &obd->u.filter;
2539 CFS_INIT_LIST_HEAD(&remove_list);
2541 cfs_spin_lock(&filter->fo_llog_list_lock);
2542 while (!cfs_list_empty(&filter->fo_llog_list)) {
2543 olg = cfs_list_entry(filter->fo_llog_list.next,
2544 struct obd_llog_group, olg_list);
2545 cfs_list_del(&olg->olg_list);
2546 cfs_list_add(&olg->olg_list, &remove_list);
2548 cfs_spin_unlock(&filter->fo_llog_list_lock);
2550 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2551 cfs_list_del_init(&olg->olg_list);
2552 rc = filter_olg_fini(olg);
2554 CERROR("failed to cleanup llogging subsystem for %u\n",
2562 static int filter_precleanup(struct obd_device *obd,
2563 enum obd_cleanup_stage stage)
2569 case OBD_CLEANUP_EARLY:
2571 case OBD_CLEANUP_EXPORTS:
2572 /* Stop recovery before namespace cleanup. */
2573 target_recovery_fini(obd);
2574 rc = filter_llog_preclean(obd);
2580 static int filter_cleanup(struct obd_device *obd)
2582 struct filter_obd *filter = &obd->u.filter;
2586 LCONSOLE_WARN("%s: shutting down for failover; client state "
2587 "will be preserved.\n", obd->obd_name);
2589 obd_exports_barrier(obd);
2590 obd_zombie_barrier();
2592 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2593 lprocfs_free_per_client_stats(obd);
2594 lprocfs_free_obd_stats(obd);
2595 lprocfs_obd_cleanup(obd);
2596 lquota_cleanup(filter_quota_interface_ref, obd);
2598 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2599 obd->obd_namespace = NULL;
2601 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2603 if (obd->u.obt.obt_sb == NULL)
2608 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2609 shrink_dcache_sb(obd->u.obt.obt_sb);
2611 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2612 obd->u.obt.obt_sb = NULL;
2614 fsfilt_put_ops(obd->obd_fsops);
2616 filter_iobuf_pool_done(filter);
2618 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2623 static int filter_connect_internal(struct obd_export *exp,
2624 struct obd_connect_data *data,
2627 struct filter_export_data *fed = &exp->exp_filter_data;
2632 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2633 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2634 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2635 data->ocd_connect_flags, data->ocd_version,
2636 data->ocd_grant, data->ocd_index);
2638 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2639 CWARN("!!! This export (nid %s) used object group %d "
2640 "earlier; now it's trying to use group %d! This could "
2641 "be a bug in the MDS. Please report to "
2642 "http://bugzilla.lustre.org/\n",
2643 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2646 fed->fed_group = data->ocd_group;
2648 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2649 exp->exp_connect_flags = data->ocd_connect_flags;
2650 data->ocd_version = LUSTRE_VERSION_CODE;
2652 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2653 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2654 data->ocd_connect_flags & OBD_CONNECT_MDS))
2657 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2658 struct filter_obd *filter = &exp->exp_obd->u.filter;
2659 obd_size left, want;
2661 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2662 left = filter_grant_space_left(exp);
2663 want = data->ocd_grant;
2664 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2665 data->ocd_grant = fed->fed_grant;
2666 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2668 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2669 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2670 exp->exp_client_uuid.uuid, exp,
2671 data->ocd_grant, want, left);
2673 filter->fo_tot_granted_clients ++;
2676 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2677 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2678 int index = le32_to_cpu(lsd->lsd_ost_index);
2680 if (!(lsd->lsd_feature_compat &
2681 cpu_to_le32(OBD_COMPAT_OST))) {
2682 /* this will only happen on the first connect */
2683 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2684 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2685 /* sync is not needed here as filter_client_add will
2686 * set exp_need_sync flag */
2687 filter_update_server_data(exp->exp_obd);
2688 } else if (index != data->ocd_index) {
2689 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2690 " %u doesn't match actual OST index"
2691 " %u in last_rcvd file, bad "
2693 obd_export_nid2str(exp), index,
2697 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2698 * FIXME: We don't strictly need the COMPAT flag for that,
2699 * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2700 * FIXME: We needed it for the index, as index 0 is valid. */
2703 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2704 data->ocd_brw_size = 65536;
2705 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2706 data->ocd_brw_size = min(data->ocd_brw_size,
2707 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2708 LASSERT(data->ocd_brw_size);
2711 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2712 __u32 cksum_types = data->ocd_cksum_types;
2714 /* The client set in ocd_cksum_types the checksum types it
2715 * supports. We have to mask off the algorithms that we don't
2717 if (cksum_types & OBD_CKSUM_ALL)
2718 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2720 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2722 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2723 "%x\n", exp->exp_obd->obd_name,
2724 obd_export_nid2str(exp), cksum_types,
2725 data->ocd_cksum_types);
2727 /* This client does not support OBD_CONNECT_CKSUM
2728 * fall back to CRC32 */
2729 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2730 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2731 exp->exp_obd->obd_name,
2732 obd_export_nid2str(exp));
2738 static int filter_reconnect(const struct lu_env *env,
2739 struct obd_export *exp, struct obd_device *obd,
2740 struct obd_uuid *cluuid,
2741 struct obd_connect_data *data,
2747 if (exp == NULL || obd == NULL || cluuid == NULL)
2750 rc = filter_connect_internal(exp, data, 1);
2752 filter_export_stats_init(obd, exp, localdata);
2757 /* nearly identical to mds_connect */
2758 static int filter_connect(const struct lu_env *env,
2759 struct obd_export **exp, struct obd_device *obd,
2760 struct obd_uuid *cluuid,
2761 struct obd_connect_data *data, void *localdata)
2763 struct lvfs_run_ctxt saved;
2764 struct lustre_handle conn = { 0 };
2765 struct obd_export *lexp;
2770 if (exp == NULL || obd == NULL || cluuid == NULL)
2773 rc = class_connect(&conn, obd, cluuid);
2776 lexp = class_conn2export(&conn);
2777 LASSERT(lexp != NULL);
2779 rc = filter_connect_internal(lexp, data, 0);
2783 filter_export_stats_init(obd, lexp, localdata);
2784 if (obd->obd_replayable) {
2785 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2787 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2788 rc = filter_client_add(obd, lexp, -1);
2793 group = data->ocd_group;
2795 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2796 obd->obd_name, lexp->exp_handle.h_cookie, group);
2798 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2799 rc = filter_read_groups(obd, group, 1);
2800 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2802 CERROR("can't read group %u\n", group);
2810 class_disconnect(lexp);
2811 lprocfs_exp_cleanup(lexp);
2820 /* Do extra sanity checks for grant accounting. We do this at connect,
2821 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2822 * always get rid of it or turn it off when we know accounting is good. */
2823 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2825 struct filter_export_data *fed;
2826 struct obd_export *exp;
2827 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2828 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2829 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2831 if (cfs_list_empty(&obd->obd_exports))
2834 /* We don't want to do this for large machines that do lots of
2835 mounts or unmounts. It burns... */
2836 if (obd->obd_num_exports > 100)
2839 cfs_spin_lock(&obd->obd_osfs_lock);
2840 cfs_spin_lock(&obd->obd_dev_lock);
2841 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2843 fed = &exp->exp_filter_data;
2844 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2847 if (maxsize > 0) { /* we may not have done a statfs yet */
2848 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2849 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2850 exp->exp_client_uuid.uuid, exp,
2851 fed->fed_grant, fed->fed_pending, maxsize);
2852 LASSERTF(fed->fed_dirty <= maxsize,
2853 "%s: cli %s/%p %ld > "LPU64"\n", func,
2854 exp->exp_client_uuid.uuid, exp,
2855 fed->fed_dirty, maxsize);
2858 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2859 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2860 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2862 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2863 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2864 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2865 tot_granted += fed->fed_grant + fed->fed_pending;
2866 tot_pending += fed->fed_pending;
2867 tot_dirty += fed->fed_dirty;
2869 fo_tot_granted = obd->u.filter.fo_tot_granted;
2870 fo_tot_pending = obd->u.filter.fo_tot_pending;
2871 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2872 cfs_spin_unlock(&obd->obd_dev_lock);
2873 cfs_spin_unlock(&obd->obd_osfs_lock);
2875 /* Do these assertions outside the spinlocks so we don't kill system */
2876 if (tot_granted != fo_tot_granted)
2877 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2878 func, tot_granted, fo_tot_granted);
2879 if (tot_pending != fo_tot_pending)
2880 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2881 func, tot_pending, fo_tot_pending);
2882 if (tot_dirty != fo_tot_dirty)
2883 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2884 func, tot_dirty, fo_tot_dirty);
2885 if (tot_pending > tot_granted)
2886 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2887 func, tot_pending, tot_granted);
2888 if (tot_granted > maxsize)
2889 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2890 func, tot_granted, maxsize);
2891 if (tot_dirty > maxsize)
2892 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2893 func, tot_dirty, maxsize);
2896 /* Remove this client from the grant accounting totals. We also remove
2897 * the export from the obd device under the osfs and dev locks to ensure
2898 * that the filter_grant_sanity_check() calculations are always valid.
2899 * The client should do something similar when it invalidates its import. */
2900 static void filter_grant_discard(struct obd_export *exp)
2902 struct obd_device *obd = exp->exp_obd;
2903 struct filter_obd *filter = &obd->u.filter;
2904 struct filter_export_data *fed = &exp->exp_filter_data;
2906 cfs_spin_lock(&obd->obd_osfs_lock);
2907 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2908 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2909 obd->obd_name, filter->fo_tot_granted,
2910 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2911 filter->fo_tot_granted -= fed->fed_grant;
2912 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2913 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2914 obd->obd_name, filter->fo_tot_pending,
2915 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2916 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2917 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2918 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2919 obd->obd_name, filter->fo_tot_dirty,
2920 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2921 filter->fo_tot_dirty -= fed->fed_dirty;
2925 cfs_spin_unlock(&obd->obd_osfs_lock);
2928 static int filter_destroy_export(struct obd_export *exp)
2930 struct filter_export_data *fed = &exp->exp_filter_data;
2933 if (fed->fed_pending)
2934 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2935 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2936 exp, fed->fed_pending);
2938 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2940 target_destroy_export(exp);
2941 ldlm_destroy_export(exp);
2942 lut_client_free(exp);
2944 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2947 if (!exp->exp_obd->obd_replayable)
2948 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2950 filter_grant_discard(exp);
2951 filter_fmd_cleanup(exp);
2953 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2954 struct filter_obd *filter = &exp->exp_obd->u.filter;
2955 if (filter->fo_tot_granted_clients > 0)
2956 filter->fo_tot_granted_clients --;
2959 if (!(exp->exp_flags & OBD_OPT_FORCE))
2960 filter_grant_sanity_check(exp->exp_obd, __func__);
2965 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2967 struct obd_llog_group *olg_min, *olg;
2968 struct filter_obd *filter;
2969 int worked = -1, group;
2970 struct llog_ctxt *ctxt;
2973 filter = &obd->u.filter;
2975 /* we can't sync log holding spinlock. also, we do not want to get
2976 * into livelock. so we do following: loop over MDS's exports in
2977 * group order and skip already synced llogs -bzzz */
2979 /* look for group with min. number, but > worked */
2982 cfs_spin_lock(&filter->fo_llog_list_lock);
2983 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2984 if (olg->olg_seq <= worked) {
2985 /* this group is already synced */
2988 if (group < olg->olg_seq) {
2989 /* we have group with smaller number to sync */
2992 /* store current minimal group */
2994 group = olg->olg_seq;
2996 cfs_spin_unlock(&filter->fo_llog_list_lock);
2998 if (olg_min == NULL)
3001 worked = olg_min->olg_seq;
3002 if (olg_min->olg_exp &&
3003 (dexp == olg_min->olg_exp || dexp == NULL)) {
3005 ctxt = llog_group_get_ctxt(olg_min,
3006 LLOG_MDS_OST_REPL_CTXT);
3008 err = llog_sync(ctxt, olg_min->olg_exp);
3009 llog_ctxt_put(ctxt);
3011 CERROR("error flushing logs to MDS: "
3016 } while (olg_min != NULL);
3019 /* Also incredibly similar to mds_disconnect */
3020 static int filter_disconnect(struct obd_export *exp)
3022 struct obd_device *obd = exp->exp_obd;
3027 class_export_get(exp);
3029 if (!(exp->exp_flags & OBD_OPT_FORCE))
3030 filter_grant_sanity_check(obd, __func__);
3031 filter_grant_discard(exp);
3033 /* Flush any remaining cancel messages out to the target */
3034 filter_sync_llogs(obd, exp);
3036 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3038 rc = server_disconnect_export(exp);
3040 /* Do not erase record for recoverable client. */
3041 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3042 filter_client_del(exp);
3044 fsfilt_sync(obd, obd->u.obt.obt_sb);
3046 class_export_put(exp);
3050 /* reverse import is changed, sync all cancels */
3051 static void filter_revimp_update(struct obd_export *exp)
3056 class_export_get(exp);
3058 /* flush any remaining cancel messages out to the target */
3059 filter_sync_llogs(exp->exp_obd, exp);
3060 class_export_put(exp);
3064 static int filter_ping(struct obd_export *exp)
3066 filter_fmd_expire(exp);
3070 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3071 const char *what, int quiet)
3073 struct dentry *dchild = NULL;
3075 dchild = filter_fid2dentry(obd, NULL, ostid->oi_seq, ostid->oi_id);
3077 if (IS_ERR(dchild)) {
3078 CERROR("%s error looking up object: "POSTID"\n",
3079 what, ostid->oi_id, ostid->oi_seq);
3083 if (dchild->d_inode == NULL) {
3085 CERROR("%s: %s on non-existent object: "POSTID" \n",
3086 obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3088 RETURN(ERR_PTR(-ENOENT));
3094 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3096 struct dentry *dentry = NULL;
3097 struct obd_device *obd;
3101 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3102 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3106 obd = class_exp2obd(exp);
3108 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3112 dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3114 RETURN(PTR_ERR(dentry));
3116 /* Limit the valid bits in the return data to what we actually use */
3117 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3118 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3124 /* this should be enabled/disabled in condition to enabled/disabled large
3125 * inodes (fast EAs) in backing store FS. */
3126 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3127 void *handle, struct obdo *oa)
3129 struct obd_device *obd = exp->exp_obd;
3133 if (oa->o_valid & OBD_MD_FLFID) {
3134 struct filter_fid ff;
3136 if (!(oa->o_valid & OBD_MD_FLGROUP))
3138 /* packing fid and converting it to LE for storing into EA.
3139 * Here ->o_stripe_idx should be filled by LOV and rest of
3140 * fields - by client. */
3141 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3142 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3143 /* XXX: we are ignoring o_parent_ver here, since this should
3144 * be the same for all objects in this fileset. */
3145 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3146 ff.ff_objid = cpu_to_le64(oa->o_id);
3147 ff.ff_seq = cpu_to_le64(oa->o_seq);
3149 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3150 LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3153 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3155 CERROR("store fid in object failed! rc: %d\n", rc);
3157 CDEBUG(D_HA, "OSS object without fid info!\n");
3163 /* this is called from filter_truncate() until we have filter_punch() */
3164 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3165 struct obdo *oa, struct obd_trans_info *oti)
3167 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3168 struct llog_cookie *fcc = NULL;
3169 struct filter_obd *filter;
3170 int rc, err, sync = 0;
3171 loff_t old_size = 0;
3172 unsigned int ia_valid;
3173 struct inode *inode;
3174 struct page *page = NULL;
3179 LASSERT(dentry != NULL);
3180 LASSERT(!IS_ERR(dentry));
3182 inode = dentry->d_inode;
3183 LASSERT(inode != NULL);
3185 filter = &exp->exp_obd->u.filter;
3186 iattr_from_obdo(&iattr, oa, oa->o_valid);
3187 ia_valid = iattr.ia_valid;
3189 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3190 OBD_ALLOC(fcc, sizeof(*fcc));
3192 *fcc = oa->o_lcookie;
3194 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3196 /* Filter truncates and writes are serialized by
3197 * i_alloc_sem, see the comment in
3198 * filter_preprw_write.*/
3199 if (ia_valid & ATTR_SIZE)
3200 down_write(&inode->i_alloc_sem);
3201 LOCK_INODE_MUTEX(inode);
3202 old_size = i_size_read(inode);
3205 /* VBR: version recovery check */
3206 rc = filter_version_get_check(exp, oti, inode);
3208 GOTO(out_unlock, rc);
3210 /* Let's pin the last page so that ldiskfs_truncate
3211 * should not start GFP_FS allocation. */
3212 if (ia_valid & ATTR_SIZE) {
3213 page = grab_cache_page(inode->i_mapping,
3214 iattr.ia_size >> PAGE_CACHE_SHIFT);
3216 GOTO(out_unlock, rc = -ENOMEM);
3221 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3222 * then we will accept the UID+GID sent by the client during write for
3223 * initializing the ownership of this inode. We only allow this to
3224 * happen once so clear these bits in setattr. In 2.6 kernels it is
3225 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3226 * the flags that are actually being set. */
3227 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3228 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3229 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3231 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3232 if (!(ia_valid & ATTR_MODE)) {
3233 iattr.ia_mode = inode->i_mode;
3234 iattr.ia_valid |= ATTR_MODE;
3236 iattr.ia_mode &= ~S_ISUID;
3238 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3239 if (!(iattr.ia_valid & ATTR_MODE)) {
3240 iattr.ia_mode = inode->i_mode;
3241 iattr.ia_valid |= ATTR_MODE;
3243 iattr.ia_mode &= ~S_ISGID;
3246 orig_ids[USRQUOTA] = inode->i_uid;
3247 orig_ids[GRPQUOTA] = inode->i_gid;
3248 handle = fsfilt_start_log(exp->exp_obd, inode,
3249 FSFILT_OP_SETATTR, oti, 1);
3251 GOTO(out_unlock, rc = PTR_ERR(handle));
3253 /* update inode EA only once when inode is suid bit marked. As
3254 * on 2.6.x UID and GID may be set separately, we check here
3255 * only one of them to avoid double setting. */
3256 if (inode->i_mode & S_ISUID)
3257 filter_update_fidea(exp, inode, handle, oa);
3259 handle = fsfilt_start(exp->exp_obd, inode,
3260 FSFILT_OP_SETATTR, oti);
3262 GOTO(out_unlock, rc = PTR_ERR(handle));
3264 if (oa->o_valid & OBD_MD_FLFLAGS) {
3265 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3266 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3268 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3270 /* set cancel cookie callback function */
3271 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3272 filter_cancel_cookies_cb,
3276 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3277 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3279 /* The truncate might have used up our transaction credits. Make sure
3280 * we have two left for the last_rcvd and VBR inode version updates. */
3281 err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3283 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3285 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3289 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3291 CERROR("error on commit, err = %d\n", err);
3298 /* For a partial-page truncate flush the page to disk immediately
3299 * to avoid data corruption during direct disk write. b=17397 */
3300 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3301 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3302 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3312 page_cache_release(page);
3314 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3315 UNLOCK_INODE_MUTEX(inode);
3316 if (ia_valid & ATTR_SIZE)
3317 up_write(&inode->i_alloc_sem);
3319 OBD_FREE(fcc, sizeof(*fcc));
3321 /* trigger quota release */
3322 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3323 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3324 int rc2 = lquota_adjust(filter_quota_interface_ref,
3325 exp->exp_obd, cur_ids,
3326 orig_ids, rc, FSFILT_OP_SETATTR);
3327 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3328 "filter adjust qunit. (rc:%d)\n", rc2);
3333 /* this is called from filter_truncate() until we have filter_punch() */
3334 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3335 struct obd_trans_info *oti)
3337 struct obdo *oa = oinfo->oi_oa;
3338 struct lustre_capa *capa = oinfo_capa(oinfo);
3339 struct ldlm_res_id res_id;
3340 struct filter_mod_data *fmd;
3341 struct lvfs_run_ctxt saved;
3342 struct filter_obd *filter;
3343 struct ldlm_resource *res;
3344 struct dentry *dentry;
3345 __u64 opc = CAPA_OPC_META_WRITE;
3349 if (oa->o_valid & OBD_FL_TRUNC)
3350 opc |= CAPA_OPC_OSS_TRUNC;
3352 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3356 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3357 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3362 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3363 /* This would be very bad - accidentally truncating a file when
3364 * changing the time or similar - bug 12203. */
3365 if (oa->o_valid & OBD_MD_FLSIZE &&
3366 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3367 static char mdsinum[48];
3369 if (oa->o_valid & OBD_MD_FLFID)
3370 snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3371 oa->o_parent_seq, oa->o_parent_oid,
3376 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3377 "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3378 oa->o_id, oa->o_seq, mdsinum);
3382 dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3384 RETURN(PTR_ERR(dentry));
3386 filter = &exp->exp_obd->u.filter;
3387 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3390 * We need to be atomic against a concurrent write
3391 * (which takes the semaphore for reading). fmd_mactime_xid
3392 * checks will have no effect if a write request with lower
3393 * xid starts just before a setattr and finishes later than
3394 * the setattr (see bug 21489, comment 27).
3397 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3398 down_write(&dentry->d_inode->i_alloc_sem);
3399 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3400 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3401 fmd->fmd_mactime_xid = oti->oti_xid;
3402 filter_fmd_put(exp, fmd);
3403 up_write(&dentry->d_inode->i_alloc_sem);
3406 /* setting objects attributes (including owner/group) */
3407 rc = filter_setattr_internal(exp, dentry, oa, oti);
3409 GOTO(out_unlock, rc);
3411 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3412 &res_id, LDLM_EXTENT, 0);
3415 LDLM_RESOURCE_ADDREF(res);
3416 rc = ldlm_res_lvbo_update(res, NULL, 0);
3417 LDLM_RESOURCE_DELREF(res);
3418 ldlm_resource_putref(res);
3421 oa->o_valid = OBD_MD_FLID;
3423 /* Quota release need uid/gid info */
3424 obdo_from_inode(oa, dentry->d_inode, NULL,
3425 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3430 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3434 /* XXX identical to osc_unpackmd */
3435 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3436 struct lov_mds_md *lmm, int lmm_bytes)
3442 if (lmm_bytes < sizeof (*lmm)) {
3443 CERROR("lov_mds_md too small: %d, need %d\n",
3444 lmm_bytes, (int)sizeof(*lmm));
3447 /* XXX LOV_MAGIC etc check? */
3449 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3450 CERROR("lov_mds_md: zero lmm_object_id\n");
3455 lsm_size = lov_stripe_md_size(1);
3459 if (*lsmp != NULL && lmm == NULL) {
3460 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3461 OBD_FREE(*lsmp, lsm_size);
3466 if (*lsmp == NULL) {
3467 OBD_ALLOC(*lsmp, lsm_size);
3471 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3472 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3473 OBD_FREE(*lsmp, lsm_size);
3476 loi_init((*lsmp)->lsm_oinfo[0]);
3480 /* XXX zero *lsmp? */
3481 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3482 LASSERT((*lsmp)->lsm_object_id);
3485 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3490 /* caller must hold fo_create_locks[oa->o_seq] */
3491 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3492 struct filter_obd *filter)
3494 struct obdo doa = { 0 }; /* XXX obdo on stack */
3500 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_seq]) != 0);
3502 memset(&doa, 0, sizeof(doa));
3504 doa.o_valid |= OBD_MD_FLGROUP;
3505 doa.o_seq = oa->o_seq;
3506 doa.o_mode = S_IFREG;
3508 if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3509 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3510 exp->exp_obd->obd_name, doa.o_seq);
3514 last = filter_last_id(filter, doa.o_seq);
3516 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3518 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3519 exp->exp_obd->obd_name, oa->o_id + 1, last,
3520 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3522 for (id = last; id > oa->o_id; id--) {
3524 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3525 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3526 CEMERG("error destroying precreate objid "LPU64": %d\n",
3529 /* update last_id on disk periodically so that if we restart
3530 * we don't need to re-scan all of the just-deleted objects. */
3531 if ((id & 511) == 0 && !skip_orphan) {
3532 filter_set_last_id(filter, id - 1, doa.o_seq);
3533 filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3537 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3538 exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3541 filter_set_last_id(filter, id, doa.o_seq);
3542 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3544 /* don't reuse orphan object, return last used objid */
3548 cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3553 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3554 obd_seq group, int *num);
3555 /* returns a negative error or a nonnegative number of files to create */
3556 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3557 obd_seq group, struct obd_trans_info *oti)
3559 struct obd_device *obd = exp->exp_obd;
3560 struct filter_obd *filter = &obd->u.filter;
3564 /* delete orphans request */
3565 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3566 obd_id last = filter_last_id(filter, group);
3568 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3569 CERROR("%s: dropping old orphan cleanup request\n",
3573 /* This causes inflight precreates to abort and drop lock */
3574 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3575 cfs_down(&filter->fo_create_locks[group]);
3576 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3577 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3578 exp->exp_obd->obd_name, group);
3579 cfs_up(&filter->fo_create_locks[group]);
3582 diff = oa->o_id - last;
3583 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3586 if (-diff > OST_MAX_PRECREATE) {
3587 CERROR("%s: ignoring bogus orphan destroy request: "
3588 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3590 /* FIXME: should reset precreate_next_id on MDS */
3591 GOTO(out, rc = -EINVAL);
3594 rc = filter_destroy_precreated(exp, oa, filter);
3596 CERROR("%s: unable to write lastobjid, but "
3597 "orphans were deleted\n", obd->obd_name);
3600 /* XXX: Used by MDS for the first time! */
3601 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3604 cfs_down(&filter->fo_create_locks[group]);
3605 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3606 CERROR("%s: dropping old precreate request\n",
3610 /* only precreate if group == 0 and o_id is specfied */
3611 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3614 diff = oa->o_id - filter_last_id(filter, group);
3615 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3616 filter_last_id(filter, group), diff);
3618 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3619 oa->o_id, filter_last_id(filter, group), diff);
3623 oa->o_id = filter_last_id(&obd->u.filter, group);
3624 rc = filter_precreate(obd, oa, group, &diff);
3625 oa->o_id = filter_last_id(&obd->u.filter, group);
3627 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3630 /* else diff == 0 */
3633 cfs_up(&filter->fo_create_locks[group]);
3637 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3638 __u64 max_age, __u32 flags)
3640 struct filter_obd *filter = &obd->u.filter;
3641 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3645 /* at least try to account for cached pages. its still racey and
3646 * might be under-reporting if clients haven't announced their
3647 * caches with brw recently */
3648 cfs_spin_lock(&obd->obd_osfs_lock);
3649 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3650 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3651 cfs_spin_unlock(&obd->obd_osfs_lock);
3653 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3654 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3655 filter->fo_tot_dirty, filter->fo_tot_granted,
3656 filter->fo_tot_pending,
3657 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3659 filter_grant_sanity_check(obd, __func__);
3661 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3662 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3663 osfs->os_bsize - 1) >> blockbits));
3665 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3666 struct lr_server_data *lsd = class_server_data(obd);
3667 int index = le32_to_cpu(lsd->lsd_ost_index);
3669 if (obd_fail_val == -1 ||
3670 index == obd_fail_val)
3671 osfs->os_bfree = osfs->os_bavail = 2;
3672 else if (obd_fail_loc & OBD_FAIL_ONCE)
3673 obd_fail_loc &= ~OBD_FAILED; /* reset flag */
3676 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3677 * stop creating files on MDS if OST is not good shape to create
3681 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3682 osfs->os_state = OS_STATE_READONLY;
3684 if (filter->fo_raid_degraded)
3685 osfs->os_state |= OS_STATE_DEGRADED;
3689 static int filter_use_existing_obj(struct obd_device *obd,
3690 struct dentry *dchild, void **handle,
3693 struct inode *inode = dchild->d_inode;
3697 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3700 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3701 if (IS_ERR(*handle))
3702 return PTR_ERR(*handle);
3704 iattr.ia_valid = ATTR_MODE;
3705 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3706 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3713 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3716 __u64 os_ffree = -1;
3718 cfs_spin_lock(&obd->obd_osfs_lock);
3719 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3721 os_ffree = obd->obd_osfs.os_ffree;
3722 cfs_spin_unlock(&obd->obd_osfs_lock);
3727 /* We rely on the fact that only one thread will be creating files in a given
3728 * group at a time, which is why we don't need an atomic filter_get_new_id.
3729 * Even if we had that atomic function, the following race would exist:
3731 * thread 1: gets id x from filter_next_id
3732 * thread 2: gets id (x + 1) from filter_next_id
3733 * thread 2: creates object (x + 1)
3734 * thread 1: tries to create object x, gets -ENOSPC
3736 * Caller must hold fo_create_locks[group]
3738 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3739 obd_seq group, int *num)
3741 struct dentry *dchild = NULL, *dparent = NULL;
3742 struct filter_obd *filter;
3743 struct obd_statfs *osfs;
3744 int err = 0, rc = 0, recreate_obj = 0, i;
3745 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3748 void *handle = NULL;
3751 filter = &obd->u.filter;
3753 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3755 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3757 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3758 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3761 OBD_ALLOC(osfs, sizeof(*osfs));
3764 rc = filter_statfs(obd, osfs,
3765 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
3767 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3768 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3769 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3770 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3774 OBD_FREE(osfs, sizeof(*osfs));
3779 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3780 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3782 for (i = 0; i < *num && err == 0; i++) {
3783 int cleanup_phase = 0;
3785 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3786 CWARN("%s: create aborted by destroy\n",
3795 last_id = filter_last_id(filter, group);
3796 if (next_id > last_id) {
3797 CERROR("Error: Trying to recreate obj greater"
3798 "than last id "LPD64" > "LPD64"\n",
3800 GOTO(cleanup, rc = -EINVAL);
3803 next_id = filter_last_id(filter, group) + 1;
3805 /* Temporary solution for oid in CMD before fid-on-OST */
3806 if ((fid_seq_is_mdt0(oa->o_seq) && next_id >= IDIF_MAX_OID) &&
3807 (fid_seq_is_cmd(oa->o_seq) && next_id >= OBIF_MAX_OID)) {
3808 CERROR("%s:"POSTID" hit the max IDIF_MAX_OID(1<<48)!\n",
3809 obd->obd_name, next_id, group);
3810 GOTO(cleanup, rc = -ENOSPC);
3813 dparent = filter_parent_lock(obd, group, next_id);
3814 if (IS_ERR(dparent))
3815 GOTO(cleanup, rc = PTR_ERR(dparent));
3816 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3818 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3820 GOTO(cleanup, rc = PTR_ERR(dchild));
3821 cleanup_phase = 2; /* f_dput(dchild) */
3823 if (dchild->d_inode != NULL) {
3824 /* This would only happen if lastobjid was bad on disk*/
3825 /* Could also happen if recreating missing obj but it
3826 * already exists. */
3828 CERROR("%s: recreating existing object %.*s?\n",
3829 obd->obd_name, dchild->d_name.len,
3830 dchild->d_name.name);
3832 /* Use these existing objects if they are
3834 if (dchild->d_inode->i_size == 0) {
3835 rc = filter_use_existing_obj(obd,dchild,
3836 &handle, &cleanup_phase);
3843 CERROR("%s: Serious error: objid %.*s already "
3844 "exists; is this filesystem corrupt?\n",
3845 obd->obd_name, dchild->d_name.len,
3846 dchild->d_name.name);
3849 GOTO(cleanup, rc = -EEXIST);
3852 handle = fsfilt_start_log(obd, dparent->d_inode,
3853 FSFILT_OP_CREATE, NULL, 1);
3855 GOTO(cleanup, rc = PTR_ERR(handle));
3858 CDEBUG(D_INODE, "%s: filter_precreate(od->o_seq="LPU64
3859 ",od->o_id="LPU64")\n", obd->obd_name, group,
3862 /* We mark object SUID+SGID to flag it for accepting UID+GID
3863 * from client on first write. Currently the permission bits
3864 * on the OST are never used, so this is OK. */
3865 rc = ll_vfs_create(dparent->d_inode, dchild,
3866 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3868 CERROR("create failed rc = %d\n", rc);
3869 if (rc == -ENOSPC) {
3870 os_ffree = filter_calc_free_inodes(obd);
3872 CERROR("%s: free inode "LPU64"\n",
3873 obd->obd_name, os_ffree);
3878 if (dchild->d_inode)
3879 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3880 dchild->d_inode->i_ino);
3883 if (!recreate_obj) {
3884 filter_set_last_id(filter, next_id, group);
3885 err = filter_update_last_objid(obd, group, 0);
3887 CERROR("unable to write lastobjid "
3888 "but file created\n");
3892 switch(cleanup_phase) {
3894 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3896 CERROR("error on commit, err = %d\n", err);
3903 filter_parent_unlock(dparent);
3910 if (cfs_time_after(jiffies, enough_time)) {
3912 "%s: precreate slow - want %d got %d \n",
3913 obd->obd_name, *num, i);
3920 "%s: created %d objects for group "POSTID" rc %d\n",
3921 obd->obd_name, i, filter->fo_last_objids[group], group, rc);
3926 int filter_create(struct obd_export *exp, struct obdo *oa,
3927 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3929 struct obd_device *obd = exp->exp_obd;
3930 struct filter_export_data *fed;
3931 struct filter_obd *filter;
3932 struct lvfs_run_ctxt saved;
3933 struct lov_stripe_md *lsm = NULL;
3937 CDEBUG(D_INODE, "%s: filter_create(group="LPU64",id="
3938 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
3940 fed = &exp->exp_filter_data;
3941 filter = &obd->u.filter;
3943 if (fed->fed_group != oa->o_seq) {
3944 CERROR("%s: this export (nid %s) used object group %d "
3945 "earlier; now it's trying to use group "LPU64"!"
3946 " This could be a bug in the MDS. Please report to "
3947 "http://bugzilla.lustre.org/\n", obd->obd_name,
3948 obd_export_nid2str(exp), fed->fed_group, oa->o_seq);
3955 rc = obd_alloc_memmd(exp, &lsm);
3962 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3964 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3965 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3966 if (!obd->obd_recovering ||
3967 oa->o_id > filter_last_id(filter, oa->o_seq)) {
3968 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3969 oa->o_id, filter_last_id(filter, oa->o_seq));
3973 cfs_down(&filter->fo_create_locks[oa->o_seq]);
3974 rc = filter_precreate(obd, oa, oa->o_seq, &diff);
3975 cfs_up(&filter->fo_create_locks[oa->o_seq]);
3978 rc = filter_handle_precreate(exp, oa, oa->o_seq, oti);
3981 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3982 if (rc && ea != NULL && *ea != lsm) {
3983 obd_free_memmd(exp, &lsm);
3984 } else if (rc == 0 && ea != NULL) {
3985 /* XXX LOV STACKING: the lsm that is passed to us from
3986 * LOV does not have valid lsm_oinfo data structs, so
3987 * don't go touching that. This needs to be fixed in a
3989 lsm->lsm_object_id = oa->o_id;
3996 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3997 struct lov_stripe_md *md, struct obd_trans_info *oti,
3998 struct obd_export *md_exp, void *capa)
4000 unsigned int qcids[MAXQUOTAS] = {0, 0};
4001 struct obd_device *obd;
4002 struct filter_obd *filter;
4003 struct dentry *dchild = NULL, *dparent = NULL;
4004 struct lustre_handle lockh = { 0 };
4005 struct lvfs_run_ctxt saved;
4006 void *handle = NULL;
4007 struct llog_cookie *fcc = NULL;
4008 int rc, rc2, cleanup_phase = 0, sync = 0;
4012 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4013 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4018 filter = &obd->u.filter;
4020 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4023 CDEBUG(D_INODE, "%s: filter_destroy(group="LPU64",oid="
4024 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4026 dchild = filter_fid2dentry(obd, NULL, oa->o_seq, oa->o_id);
4028 GOTO(cleanup, rc = PTR_ERR(dchild));
4031 if (dchild->d_inode == NULL) {
4032 CDEBUG(D_INODE, "destroying non-existent object "POSTID"\n",
4033 oa->o_id, oa->o_seq);
4034 /* If object already gone, cancel cookie right now */
4035 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4036 struct llog_ctxt *ctxt;
4037 struct obd_llog_group *olg;
4039 olg = filter_find_olg(obd, oa->o_seq);
4041 CERROR(" %s: can not find olg of group %d\n",
4042 obd->obd_name, (int)oa->o_seq);
4043 GOTO(cleanup, rc = PTR_ERR(olg));
4045 fcc = &oa->o_lcookie;
4046 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4047 llog_cancel(ctxt, NULL, 1, fcc, 0);
4048 llog_ctxt_put(ctxt);
4049 fcc = NULL; /* we didn't allocate fcc, don't free it */
4051 GOTO(cleanup, rc = -ENOENT);
4054 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_seq, &lockh);
4058 /* Our MDC connection is established by the MDS to us */
4059 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4060 OBD_ALLOC(fcc, sizeof(*fcc));
4062 *fcc = oa->o_lcookie;
4064 DQUOT_INIT(dchild->d_inode);
4066 /* we're gonna truncate it first in order to avoid possible deadlock:
4068 * open trasaction open transaction
4069 * down(i_zombie) down(i_zombie)
4070 * restart transaction
4071 * (see BUG 4180) -bzzz
4073 * take i_alloc_sem too to prevent other threads from writing to the
4074 * file while we are truncating it. This can cause lock ordering issue
4075 * between page lock, i_mutex & starting new journal handle.
4076 * (see bug 20321) -johann
4078 down_write(&dchild->d_inode->i_alloc_sem);
4079 LOCK_INODE_MUTEX(dchild->d_inode);
4081 /* VBR: version recovery check */
4082 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4084 UNLOCK_INODE_MUTEX(dchild->d_inode);
4085 up_write(&dchild->d_inode->i_alloc_sem);
4089 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4091 if (IS_ERR(handle)) {
4092 UNLOCK_INODE_MUTEX(dchild->d_inode);
4093 up_write(&dchild->d_inode->i_alloc_sem);
4094 GOTO(cleanup, rc = PTR_ERR(handle));
4097 iattr.ia_valid = ATTR_SIZE;
4099 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4100 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4101 UNLOCK_INODE_MUTEX(dchild->d_inode);
4102 up_write(&dchild->d_inode->i_alloc_sem);
4106 GOTO(cleanup, rc = rc2);
4108 /* We don't actually need to lock the parent until we are unlinking
4109 * here, and not while truncating above. That avoids holding the
4110 * parent lock for a long time during truncate, which can block other
4111 * threads from doing anything to objects in that directory. bug 7171 */
4112 dparent = filter_parent_lock(obd, oa->o_seq, oa->o_id);
4113 if (IS_ERR(dparent))
4114 GOTO(cleanup, rc = PTR_ERR(dparent));
4115 cleanup_phase = 3; /* filter_parent_unlock */
4117 LOCK_INODE_MUTEX(dchild->d_inode);
4118 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4119 if (IS_ERR(handle)) {
4120 UNLOCK_INODE_MUTEX(dchild->d_inode);
4121 GOTO(cleanup, rc = PTR_ERR(handle));
4123 cleanup_phase = 4; /* fsfilt_commit */
4125 /* Quota release need uid/gid of inode */
4126 obdo_from_inode(oa, dchild->d_inode, NULL, OBD_MD_FLUID|OBD_MD_FLGID);
4128 filter_fmd_drop(exp, oa->o_id, oa->o_seq);
4130 /* this drops dchild->d_inode->i_mutex unconditionally */
4131 rc = filter_destroy_internal(obd, oa->o_id, oa->o_seq, dparent, dchild);
4135 switch(cleanup_phase) {
4138 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4139 oti->oti_handle : handle,
4140 filter_cancel_cookies_cb,
4142 /* If add_journal_cb failed, then filter_finish_transno
4143 * will commit the handle and we will do a sync
4144 * on commit. then we call callback directly to free
4147 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4149 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4152 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4154 CERROR("error on commit, err = %d\n", rc2);
4161 filter_parent_unlock(dparent);
4163 filter_fini_destroy(obd, &lockh);
4167 OBD_FREE(fcc, sizeof(*fcc));
4169 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4172 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4176 /* trigger quota release */
4177 qcids[USRQUOTA] = oa->o_uid;
4178 qcids[GRPQUOTA] = oa->o_gid;
4179 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4182 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4186 /* NB start and end are used for punch, but not truncate */
4187 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4188 struct obd_trans_info *oti,
4189 struct ptlrpc_request_set *rqset)
4194 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4195 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4196 oinfo->oi_policy.l_extent.end);
4200 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4201 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,oinfo->oi_oa->o_valid,
4202 oinfo->oi_policy.l_extent.start);
4204 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4205 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4206 rc = filter_setattr(exp, oinfo, oti);
4207 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4211 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4212 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4215 struct lvfs_run_ctxt saved;
4216 struct obd_device_target *obt;
4217 struct dentry *dentry;
4221 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4222 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4226 obt = &exp->exp_obd->u.obt;
4228 /* An objid of zero is taken to mean "sync whole filesystem" */
4229 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4230 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4231 /* Flush any remaining cancel messages out to the target */
4232 filter_sync_llogs(exp->exp_obd, exp);
4236 dentry = filter_oa2dentry(exp->exp_obd, &oa->o_oi);
4238 RETURN(PTR_ERR(dentry));
4240 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4242 LOCK_INODE_MUTEX(dentry->d_inode);
4244 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4246 /* just any file to grab fsync method - "file" arg unused */
4247 struct file *file = obt->obt_rcvd_filp;
4249 if (file->f_op && file->f_op->fsync)
4250 rc = file->f_op->fsync(NULL, dentry, 1);
4252 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4256 UNLOCK_INODE_MUTEX(dentry->d_inode);
4258 oa->o_valid = OBD_MD_FLID;
4259 obdo_from_inode(oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
4261 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4267 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4268 void *key, __u32 *vallen, void *val,
4269 struct lov_stripe_md *lsm)
4271 struct obd_device *obd;
4274 obd = class_exp2obd(exp);
4276 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4280 if (KEY_IS(KEY_BLOCKSIZE)) {
4281 __u32 *blocksize = val;
4283 if (*vallen < sizeof(*blocksize))
4285 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4287 *vallen = sizeof(*blocksize);
4291 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4292 __u32 *blocksize_bits = val;
4293 if (blocksize_bits) {
4294 if (*vallen < sizeof(*blocksize_bits))
4296 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4298 *vallen = sizeof(*blocksize_bits);
4302 if (KEY_IS(KEY_LAST_ID)) {
4303 obd_id *last_id = val;
4304 /* FIXME: object groups */
4306 if (*vallen < sizeof(*last_id))
4308 *last_id = filter_last_id(&obd->u.filter,
4309 exp->exp_filter_data.fed_group);
4311 *vallen = sizeof(*last_id);
4315 if (KEY_IS(KEY_FIEMAP)) {
4316 struct ll_fiemap_info_key *fm_key = key;
4317 struct dentry *dentry;
4318 struct ll_user_fiemap *fiemap = val;
4319 struct lvfs_run_ctxt saved;
4322 if (fiemap == NULL) {
4323 *vallen = fiemap_count_to_size(
4324 fm_key->fiemap.fm_extent_count);
4328 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa.o_oi,
4331 RETURN(PTR_ERR(dentry));
4333 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4334 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4335 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4337 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4343 if (KEY_IS(KEY_SYNC_LOCK_CANCEL)) {
4344 *((__u32 *) val) = obd->u.filter.fo_sync_lock_cancel;
4345 *vallen = sizeof(__u32);
4349 CDEBUG(D_IOCTL, "invalid key\n");
4353 static inline int filter_setup_llog_group(struct obd_export *exp,
4354 struct obd_device *obd,
4357 struct obd_llog_group *olg;
4358 struct llog_ctxt *ctxt;
4361 olg = filter_find_create_olg(obd, group);
4363 RETURN(PTR_ERR(olg));
4365 llog_group_set_export(olg, exp);
4367 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4368 LASSERTF(ctxt != NULL, "ctxt is null\n");
4370 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4371 llog_ctxt_put(ctxt);
4375 static int filter_set_grant_shrink(struct obd_export *exp,
4376 struct ost_body *body)
4378 /* handle shrink grant */
4379 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4380 filter_grant_incoming(exp, &body->oa);
4381 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4387 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4389 struct obd_device *obd;
4395 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4399 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4400 obd_export_nid2str(exp));
4401 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4403 /* setup llog imports */
4405 group = (int)(*(__u32 *)val);
4407 group = 0; /* default value */
4409 LASSERT_SEQ_IS_MDT(group);
4410 rc = filter_setup_llog_group(exp, obd, group);
4414 if (group == FID_SEQ_OST_MDT0) {
4415 /* setup llog group 1 for interop */
4416 filter_setup_llog_group(exp, obd, FID_SEQ_LLOG);
4419 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4424 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4425 void *key, __u32 vallen, void *val,
4426 struct ptlrpc_request_set *set)
4428 struct obd_device *obd;
4433 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4437 if (KEY_IS(KEY_CAPA_KEY)) {
4439 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4441 CERROR("filter update capability key failed: %d\n", rc);
4445 if (KEY_IS(KEY_REVIMP_UPD)) {
4446 filter_revimp_update(exp);
4447 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4451 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4452 filter_adapt_sptlrpc_conf(obd, 0);
4456 if (KEY_IS(KEY_MDS_CONN))
4457 RETURN(filter_set_mds_conn(exp, val));
4459 if (KEY_IS(KEY_GRANT_SHRINK))
4460 RETURN(filter_set_grant_shrink(exp, val));
4465 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4466 int len, void *karg, void *uarg)
4468 struct obd_device *obd = exp->exp_obd;
4469 struct obd_ioctl_data *data = karg;
4473 case OBD_IOC_ABORT_RECOVERY: {
4474 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4475 target_stop_recovery_thread(obd);
4479 case OBD_IOC_SYNC: {
4480 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4481 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4485 case OBD_IOC_SET_READONLY: {
4487 struct super_block *sb = obd->u.obt.obt_sb;
4488 struct inode *inode = sb->s_root->d_inode;
4489 BDEVNAME_DECLARE_STORAGE(tmp);
4490 CERROR("*** setting device %s read-only ***\n",
4491 ll_bdevname(sb, tmp));
4493 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4494 if (!IS_ERR(handle))
4495 rc = fsfilt_commit(obd, inode, handle, 1);
4497 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4498 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4500 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4504 case OBD_IOC_CATLOGLIST: {
4505 rc = llog_catalog_list(obd, 1, data);
4509 case OBD_IOC_LLOG_CANCEL:
4510 case OBD_IOC_LLOG_REMOVE:
4511 case OBD_IOC_LLOG_INFO:
4512 case OBD_IOC_LLOG_PRINT: {
4513 /* FIXME to be finished */
4514 RETURN(-EOPNOTSUPP);
4516 struct llog_ctxt *ctxt = NULL;
4518 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4519 rc = llog_ioctl(ctxt, cmd, data);
4520 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4533 static int filter_health_check(struct obd_device *obd)
4535 #ifdef USE_HEALTH_CHECK_WRITE
4536 struct filter_obd *filter = &obd->u.filter;
4541 * health_check to return 0 on healthy
4542 * and 1 on unhealthy.
4544 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4547 #ifdef USE_HEALTH_CHECK_WRITE
4548 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4549 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4554 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4557 return filter_fid2dentry(data, NULL, gr, id);
4560 static int filter_process_config(struct obd_device *obd, obd_count len,
4563 struct lustre_cfg *lcfg = buf;
4564 struct lprocfs_static_vars lvars;
4567 switch (lcfg->lcfg_command) {
4569 lprocfs_filter_init_vars(&lvars);
4571 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4581 static struct lvfs_callback_ops filter_lvfs_ops = {
4582 l_fid2dentry: filter_lvfs_fid2dentry,
4585 static struct obd_ops filter_obd_ops = {
4586 .o_owner = THIS_MODULE,
4587 .o_get_info = filter_get_info,
4588 .o_set_info_async = filter_set_info_async,
4589 .o_setup = filter_setup,
4590 .o_precleanup = filter_precleanup,
4591 .o_cleanup = filter_cleanup,
4592 .o_connect = filter_connect,
4593 .o_reconnect = filter_reconnect,
4594 .o_disconnect = filter_disconnect,
4595 .o_ping = filter_ping,
4596 .o_init_export = filter_init_export,
4597 .o_destroy_export = filter_destroy_export,
4598 .o_statfs = filter_statfs,
4599 .o_getattr = filter_getattr,
4600 .o_unpackmd = filter_unpackmd,
4601 .o_create = filter_create,
4602 .o_setattr = filter_setattr,
4603 .o_destroy = filter_destroy,
4604 .o_brw = filter_brw,
4605 .o_punch = filter_truncate,
4606 .o_sync = filter_sync,
4607 .o_preprw = filter_preprw,
4608 .o_commitrw = filter_commitrw,
4609 .o_llog_init = filter_llog_init,
4610 .o_llog_connect = filter_llog_connect,
4611 .o_llog_finish = filter_llog_finish,
4612 .o_iocontrol = filter_iocontrol,
4613 .o_health_check = filter_health_check,
4614 .o_process_config = filter_process_config,
4617 quota_interface_t *filter_quota_interface_ref;
4618 extern quota_interface_t filter_quota_interface;
4620 static int __init obdfilter_init(void)
4622 struct lprocfs_static_vars lvars;
4625 /** sanity check for group<->mdsno conversion */
4626 for (i = 0; i < MAX_MDT_COUNT; i++)
4627 LASSERT(objseq_to_mdsno(mdt_to_obd_objseq(i)) == i);
4629 lprocfs_filter_init_vars(&lvars);
4631 cfs_request_module("%s", "lquota");
4632 OBD_ALLOC(obdfilter_created_scratchpad,
4633 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4634 sizeof(*obdfilter_created_scratchpad));
4635 if (obdfilter_created_scratchpad == NULL)
4638 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4639 sizeof(struct filter_mod_data),
4642 GOTO(out, rc = -ENOMEM);
4644 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4645 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4647 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4648 LUSTRE_OST_NAME, NULL);
4652 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4653 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4654 ll_fmd_cachep = NULL;
4656 if (filter_quota_interface_ref)
4657 PORTAL_SYMBOL_PUT(filter_quota_interface);
4659 OBD_FREE(obdfilter_created_scratchpad,
4660 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4661 sizeof(*obdfilter_created_scratchpad));
4667 static void __exit obdfilter_exit(void)
4669 if (filter_quota_interface_ref)
4670 PORTAL_SYMBOL_PUT(filter_quota_interface);
4672 if (ll_fmd_cachep) {
4673 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4674 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4675 ll_fmd_cachep = NULL;
4678 class_unregister_type(LUSTRE_OST_NAME);
4679 OBD_FREE(obdfilter_created_scratchpad,
4680 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4681 sizeof(*obdfilter_created_scratchpad));
4684 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4685 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4686 MODULE_LICENSE("GPL");
4688 module_init(obdfilter_init);
4689 module_exit(obdfilter_exit);