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,
250 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
251 /* Self-export gets no proc entry */
254 rc = lprocfs_exp_setup(exp, client_nid, reconnect, &newnid);
256 /* Mask error for already created
264 struct nid_stat *tmp = exp->exp_nid_stats;
265 LASSERT(tmp != NULL);
267 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
268 if (tmp->nid_brw_stats == NULL)
269 GOTO(clean, rc = -ENOMEM);
271 init_brw_stats(tmp->nid_brw_stats);
272 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
273 0644, &filter_per_nid_stats_fops,
276 CWARN("Error adding the brw_stats file\n");
278 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
282 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
286 rc = lprocfs_nid_ldlm_stats_init(tmp);
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, 0, 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 obd->obd_max_recoverable_clients++;
884 class_export_put(exp);
886 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
887 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
891 obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
893 obd->u.obt.obt_mount_count = mount_count + 1;
894 lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
896 /* save it, so mount count and last_transno is current */
897 rc = filter_update_server_data(obd);
899 GOTO(err_client, rc);
904 class_disconnect_exports(obd);
906 filter_free_server_data(&obd->u.obt);
910 static int filter_cleanup_groups(struct obd_device *obd)
912 struct filter_obd *filter = &obd->u.filter;
914 struct dentry *dentry;
918 if (filter->fo_dentry_O_groups != NULL) {
919 for (i = 0; i < filter->fo_group_count; i++) {
920 dentry = filter->fo_dentry_O_groups[i];
924 OBD_FREE(filter->fo_dentry_O_groups,
925 filter->fo_group_count *
926 sizeof(*filter->fo_dentry_O_groups));
927 filter->fo_dentry_O_groups = NULL;
929 if (filter->fo_last_objid_files != NULL) {
930 for (i = 0; i < filter->fo_group_count; i++) {
931 filp = filter->fo_last_objid_files[i];
935 OBD_FREE(filter->fo_last_objid_files,
936 filter->fo_group_count *
937 sizeof(*filter->fo_last_objid_files));
938 filter->fo_last_objid_files = NULL;
940 if (filter->fo_dentry_O_sub != NULL) {
941 for (i = 0; i < filter->fo_group_count; i++) {
942 for (j = 0; j < filter->fo_subdir_count; j++) {
943 dentry = filter->fo_dentry_O_sub[i].dentry[j];
948 OBD_FREE(filter->fo_dentry_O_sub,
949 filter->fo_group_count *
950 sizeof(*filter->fo_dentry_O_sub));
951 filter->fo_dentry_O_sub = NULL;
953 if (filter->fo_last_objids != NULL) {
954 OBD_FREE(filter->fo_last_objids,
955 filter->fo_group_count *
956 sizeof(*filter->fo_last_objids));
957 filter->fo_last_objids = NULL;
959 if (filter->fo_dentry_O != NULL) {
960 f_dput(filter->fo_dentry_O);
961 filter->fo_dentry_O = NULL;
966 static int filter_update_last_group(struct obd_device *obd, int group)
968 struct filter_obd *filter = &obd->u.filter;
969 struct file *filp = NULL;
970 int last_group = 0, rc;
974 if (group <= filter->fo_committed_group)
977 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
981 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
985 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
987 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
991 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
992 obd->obd_name, last_group, group);
996 /* must be sync: bXXXX */
997 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
999 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1003 filter->fo_committed_group = group;
1006 filp_close(filp, 0);
1010 static int filter_read_group_internal(struct obd_device *obd, int group,
1013 struct filter_obd *filter = &obd->u.filter;
1014 __u64 *new_objids = NULL;
1015 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1016 struct dentry **new_groups = NULL;
1017 struct file **new_files = NULL;
1018 struct dentry *dentry;
1020 int old_count = filter->fo_group_count, rc, stage = 0, i;
1024 int len = group + 1;
1026 snprintf(name, 24, "%d", group);
1030 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1032 if (IS_ERR(dentry)) {
1033 CERROR("Cannot lookup expected object group %d: %ld\n",
1034 group, PTR_ERR(dentry));
1035 RETURN(PTR_ERR(dentry));
1038 dentry = simple_mkdir(filter->fo_dentry_O,
1039 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1040 if (IS_ERR(dentry)) {
1041 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1043 RETURN(PTR_ERR(dentry));
1048 snprintf(name, 24, "O/%d/LAST_ID", group);
1050 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1052 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1053 GOTO(cleanup, rc = PTR_ERR(filp));
1057 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1059 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1063 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1064 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1065 if (tmp_subdirs == NULL)
1066 GOTO(cleanup, rc = -ENOMEM);
1069 for (i = 0; i < filter->fo_subdir_count; i++) {
1071 snprintf(dir, sizeof(dir), "d%u", i);
1073 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1074 obd->u.obt.obt_vfsmnt,
1076 if (IS_ERR(tmp_subdirs->dentry[i])) {
1077 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1078 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1083 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1084 tmp_subdirs->dentry[i]);
1088 /* 'group' is an index; we need an array of length 'group + 1' */
1089 if (group + 1 > old_count) {
1090 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1091 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1092 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1093 OBD_ALLOC(new_files, len * sizeof(*new_files));
1095 if (new_objids == NULL || new_subdirs == NULL ||
1096 new_groups == NULL || new_files == NULL)
1097 GOTO(cleanup, rc = -ENOMEM);
1100 memcpy(new_objids, filter->fo_last_objids,
1101 old_count * sizeof(*new_objids));
1102 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1103 old_count * sizeof(*new_subdirs));
1104 memcpy(new_groups, filter->fo_dentry_O_groups,
1105 old_count * sizeof(*new_groups));
1106 memcpy(new_files, filter->fo_last_objid_files,
1107 old_count * sizeof(*new_files));
1109 OBD_FREE(filter->fo_last_objids,
1110 old_count * sizeof(*new_objids));
1111 OBD_FREE(filter->fo_dentry_O_sub,
1112 old_count * sizeof(*new_subdirs));
1113 OBD_FREE(filter->fo_dentry_O_groups,
1114 old_count * sizeof(*new_groups));
1115 OBD_FREE(filter->fo_last_objid_files,
1116 old_count * sizeof(*new_files));
1118 filter->fo_last_objids = new_objids;
1119 filter->fo_dentry_O_sub = new_subdirs;
1120 filter->fo_dentry_O_groups = new_groups;
1121 filter->fo_last_objid_files = new_files;
1122 filter->fo_group_count = len;
1125 filter->fo_dentry_O_groups[group] = dentry;
1126 filter->fo_last_objid_files[group] = filp;
1127 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1128 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1129 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1132 filter_update_last_group(obd, group);
1134 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1135 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1136 rc = filter_update_last_objid(obd, group, 1);
1140 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1141 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1142 obd->obd_name, group, last_objid);
1147 if (new_objids != NULL)
1148 OBD_FREE(new_objids, len * sizeof(*new_objids));
1149 if (new_subdirs != NULL)
1150 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1151 if (new_groups != NULL)
1152 OBD_FREE(new_groups, len * sizeof(*new_groups));
1153 if (new_files != NULL)
1154 OBD_FREE(new_files, len * sizeof(*new_files));
1156 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1157 for (i = 0; i < filter->fo_subdir_count; i++) {
1158 if (tmp_subdirs->dentry[i] != NULL)
1159 dput(tmp_subdirs->dentry[i]);
1161 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1164 filp_close(filp, 0);
1171 static int filter_read_groups(struct obd_device *obd, int last_group,
1174 struct filter_obd *filter = &obd->u.filter;
1175 int old_count, group, rc = 0;
1177 cfs_down(&filter->fo_init_lock);
1178 old_count = filter->fo_group_count;
1179 for (group = old_count; group <= last_group; group++) {
1180 rc = filter_read_group_internal(obd, group, create);
1184 cfs_up(&filter->fo_init_lock);
1188 /* FIXME: object groups */
1189 static int filter_prep_groups(struct obd_device *obd)
1191 struct filter_obd *filter = &obd->u.filter;
1192 struct dentry *O_dentry;
1194 int last_group, rc = 0, cleanup_phase = 0;
1198 O_dentry = simple_mkdir(current->fs->pwd, obd->u.obt.obt_vfsmnt,
1200 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1201 if (IS_ERR(O_dentry)) {
1202 rc = PTR_ERR(O_dentry);
1203 CERROR("cannot open/create O: rc = %d\n", rc);
1206 filter->fo_dentry_O = O_dentry;
1207 cleanup_phase = 1; /* O_dentry */
1209 /* we have to initialize all groups before first connections from
1210 * clients because they may send create/destroy for any group -bzzz */
1211 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1213 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1214 GOTO(cleanup, rc = PTR_ERR(filp));
1216 cleanup_phase = 2; /* filp */
1218 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1220 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1225 last_group = FID_SEQ_OST_MDT0;
1227 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1228 FID_SEQ_OST_MDT0, last_group);
1229 filter->fo_committed_group = last_group;
1230 rc = filter_read_groups(obd, last_group, 1);
1234 filp_close(filp, 0);
1238 switch (cleanup_phase) {
1240 filp_close(filp, 0);
1242 filter_cleanup_groups(obd);
1243 f_dput(filter->fo_dentry_O);
1244 filter->fo_dentry_O = NULL;
1252 /* setup the object store with correct subdirectories */
1253 static int filter_prep(struct obd_device *obd)
1255 struct lvfs_run_ctxt saved;
1256 struct filter_obd *filter = &obd->u.filter;
1258 struct inode *inode;
1262 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1263 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1264 if (!file || IS_ERR(file)) {
1266 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1270 obd->u.obt.obt_rcvd_filp = file;
1271 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1272 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1273 file->f_dentry->d_inode->i_mode);
1274 GOTO(err_filp, rc = -ENOENT);
1277 inode = file->f_dentry->d_parent->d_inode;
1278 /* We use i_op->unlink directly in filter_vfs_unlink() */
1279 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1280 CERROR("%s: filesystem does not support create/unlink ops\n",
1282 GOTO(err_filp, rc = -EOPNOTSUPP);
1285 rc = filter_init_server_data(obd, file);
1287 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1290 LASSERT(obd->u.obt.obt_lut);
1291 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1293 /* open and create health check io file*/
1294 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1297 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1299 GOTO(err_server_data, rc);
1301 filter->fo_obt.obt_health_check_filp = file;
1302 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1303 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1304 file->f_dentry->d_inode->i_mode);
1305 GOTO(err_health_check, rc = -ENOENT);
1307 rc = lvfs_check_io_health(obd, file);
1309 GOTO(err_health_check, rc);
1311 rc = filter_prep_groups(obd);
1313 GOTO(err_health_check, rc);
1315 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1320 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1321 CERROR("can't close %s after error\n", HEALTH_CHECK);
1322 filter->fo_obt.obt_health_check_filp = NULL;
1324 target_recovery_fini(obd);
1325 filter_free_server_data(&obd->u.obt);
1327 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1328 CERROR("can't close %s after error\n", LAST_RCVD);
1329 obd->u.obt.obt_rcvd_filp = NULL;
1333 /* cleanup the filter: write last used object id to status file */
1334 static void filter_post(struct obd_device *obd)
1336 struct lvfs_run_ctxt saved;
1337 struct filter_obd *filter = &obd->u.filter;
1340 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1341 * best to start a transaction with h_sync, because we removed this
1344 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1345 rc = filter_update_server_data(obd);
1347 CERROR("error writing server data: rc = %d\n", rc);
1349 for (i = 0; i < filter->fo_group_count; i++) {
1350 rc = filter_update_last_objid(obd, i,
1351 (i == filter->fo_group_count - 1));
1353 CERROR("error writing group %d lastobjid: rc = %d\n",
1357 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1358 obd->u.obt.obt_rcvd_filp = NULL;
1360 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1362 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1363 filter->fo_obt.obt_health_check_filp = NULL;
1365 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1367 filter_cleanup_groups(obd);
1368 filter_free_server_data(&obd->u.obt);
1369 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1371 filter_free_capa_keys(filter);
1372 cleanup_capa_hash(filter->fo_capa_hash);
1375 static void filter_set_last_id(struct filter_obd *filter,
1376 obd_id id, obd_seq group)
1378 LASSERT(group <= filter->fo_group_count);
1380 cfs_spin_lock(&filter->fo_objidlock);
1381 filter->fo_last_objids[group] = id;
1382 cfs_spin_unlock(&filter->fo_objidlock);
1385 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1388 LASSERT(group <= filter->fo_group_count);
1389 LASSERT(filter->fo_last_objids != NULL);
1391 /* FIXME: object groups */
1392 cfs_spin_lock(&filter->fo_objidlock);
1393 id = filter->fo_last_objids[group];
1394 cfs_spin_unlock(&filter->fo_objidlock);
1398 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1400 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1404 /* We never dget the object parent, so DON'T dput it either */
1405 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1407 struct filter_obd *filter = &obd->u.filter;
1408 struct filter_subdirs *subdirs;
1409 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1411 if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1412 return filter->fo_dentry_O_groups[group];
1414 subdirs = &filter->fo_dentry_O_sub[group];
1415 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1418 /* We never dget the object parent, so DON'T dput it either */
1419 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1422 unsigned long now = jiffies;
1423 struct dentry *dparent = filter_parent(obd, group, objid);
1426 if (IS_ERR(dparent))
1428 if (dparent == NULL)
1429 return ERR_PTR(-ENOENT);
1431 rc = filter_lock_dentry(obd, dparent);
1432 fsfilt_check_slow(obd, now, "parent lock");
1433 return rc ? ERR_PTR(rc) : dparent;
1436 /* We never dget the object parent, so DON'T dput it either */
1437 static void filter_parent_unlock(struct dentry *dparent)
1439 UNLOCK_INODE_MUTEX(dparent->d_inode);
1442 /* How to get files, dentries, inodes from object id's.
1444 * If dir_dentry is passed, the caller has already locked the parent
1445 * appropriately for this operation (normally a write lock). If
1446 * dir_dentry is NULL, we do a read lock while we do the lookup to
1447 * avoid races with create/destroy and such changing the directory
1448 * internal to the filesystem code. */
1449 struct dentry *filter_fid2dentry(struct obd_device *obd,
1450 struct dentry *dir_dentry,
1451 obd_seq group, obd_id id)
1453 struct dentry *dparent = dir_dentry;
1454 struct dentry *dchild;
1459 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1460 obd->u.filter.fo_destroys_in_progress == 0) {
1461 /* don't fail lookups for orphan recovery, it causes
1462 * later LBUGs when objects still exist during precreate */
1463 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1464 RETURN(ERR_PTR(-ENOENT));
1467 CERROR("fatal: invalid object id 0\n");
1468 RETURN(ERR_PTR(-ESTALE));
1471 len = sprintf(name, LPU64, id);
1472 if (dir_dentry == NULL) {
1473 dparent = filter_parent_lock(obd, group, id);
1474 if (IS_ERR(dparent)) {
1475 CERROR("%s: error getting object "POSTID
1476 " parent: rc %ld\n", obd->obd_name,
1477 id, group, PTR_ERR(dparent));
1481 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1482 dparent->d_name.len, dparent->d_name.name, name);
1483 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1484 if (dir_dentry == NULL)
1485 filter_parent_unlock(dparent);
1486 if (IS_ERR(dchild)) {
1487 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1488 obd->obd_name, id, group, PTR_ERR(dchild));
1492 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1493 CERROR("%s: got bad object "LPU64" inode %lu\n",
1494 obd->obd_name, id, dchild->d_inode->i_ino);
1496 RETURN(ERR_PTR(-ENOENT));
1499 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1500 name, dchild, atomic_read(&dchild->d_count));
1502 LASSERT(atomic_read(&dchild->d_count) > 0);
1507 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1508 obd_id group, struct lustre_handle *lockh)
1510 int flags = LDLM_AST_DISCARD_DATA, rc;
1511 struct ldlm_res_id res_id;
1512 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1515 osc_build_res_name(objid, group, &res_id);
1516 /* Tell the clients that the object is gone now and that they should
1517 * throw away any cached pages. */
1518 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1519 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1520 ldlm_completion_ast, NULL, NULL, 0, NULL,
1527 static void filter_fini_destroy(struct obd_device *obd,
1528 struct lustre_handle *lockh)
1530 if (lustre_handle_is_used(lockh))
1531 ldlm_lock_decref(lockh, LCK_PW);
1534 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1535 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1536 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1537 * reverse. Caller must take i_sem before starting the transaction and we
1538 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1539 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1540 struct vfsmount *mnt)
1545 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1546 * itself we already hold dir->i_mutex for child create/unlink ops */
1547 LASSERT(dentry->d_inode != NULL);
1548 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1549 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1553 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1554 GOTO(out, rc = -ENOENT);
1556 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1561 GOTO(out, rc = -EPERM);
1563 /* check_sticky() */
1564 if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1565 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1566 IS_IMMUTABLE(dentry->d_inode))
1567 GOTO(out, rc = -EPERM);
1569 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1570 * that was done because of journal_start (which is already done
1571 * here) or some other ordering issue. */
1572 ll_vfs_dq_init(dir);
1574 rc = ll_security_inode_unlink(dir, dentry, mnt);
1578 rc = dir->i_op->unlink(dir, dentry);
1580 /* need to drop i_mutex before we lose inode reference */
1581 UNLOCK_INODE_MUTEX(dentry->d_inode);
1588 /* Caller must hold LCK_PW on parent and push us into kernel context.
1589 * Caller must hold child i_mutex, we drop it always.
1590 * Caller is also required to ensure that dchild->d_inode exists. */
1591 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1592 obd_seq group, struct dentry *dparent,
1593 struct dentry *dchild)
1595 struct inode *inode = dchild->d_inode;
1598 /* There should be 2 references to the inode:
1599 * 1) taken by filter_prepare_destroy
1600 * 2) taken by filter_destroy */
1601 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1602 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1603 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1604 (unsigned long)inode->i_nlink,
1605 atomic_read(&inode->i_count));
1608 rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1610 CERROR("error unlinking objid %.*s: rc %d\n",
1611 dchild->d_name.len, dchild->d_name.name, rc);
1615 struct filter_intent_args {
1616 struct ldlm_lock **victim;
1621 static enum interval_iter filter_intent_cb(struct interval_node *n,
1624 struct ldlm_interval *node = (struct ldlm_interval *)n;
1625 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1626 __u64 size = arg->size;
1627 struct ldlm_lock **v = arg->victim;
1628 struct ldlm_lock *lck;
1630 /* If the interval is lower than the current file size,
1632 if (interval_high(n) <= size)
1633 return INTERVAL_ITER_STOP;
1635 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1636 /* Don't send glimpse ASTs to liblustre clients.
1637 * They aren't listening for them, and they do
1638 * entirely synchronous I/O anyways. */
1639 if (lck->l_export == NULL ||
1640 lck->l_export->exp_libclient == 1)
1643 if (*arg->liblustre)
1644 *arg->liblustre = 0;
1647 *v = LDLM_LOCK_GET(lck);
1648 } else if ((*v)->l_policy_data.l_extent.start <
1649 lck->l_policy_data.l_extent.start) {
1650 LDLM_LOCK_RELEASE(*v);
1651 *v = LDLM_LOCK_GET(lck);
1654 /* the same policy group - every lock has the
1655 * same extent, so needn't do it any more */
1659 return INTERVAL_ITER_CONT;
1662 static int filter_intent_policy(struct ldlm_namespace *ns,
1663 struct ldlm_lock **lockp, void *req_cookie,
1664 ldlm_mode_t mode, int flags, void *data)
1666 CFS_LIST_HEAD(rpc_list);
1667 struct ptlrpc_request *req = req_cookie;
1668 struct ldlm_lock *lock = *lockp, *l = NULL;
1669 struct ldlm_resource *res = lock->l_resource;
1670 ldlm_processing_policy policy;
1671 struct ost_lvb *res_lvb, *reply_lvb;
1672 struct ldlm_reply *rep;
1674 int idx, rc, tmpflags = 0, only_liblustre = 1;
1675 struct ldlm_interval_tree *tree;
1676 struct filter_intent_args arg;
1677 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1678 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1679 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1682 policy = ldlm_get_processing_policy(res);
1683 LASSERT(policy != NULL);
1684 LASSERT(req != NULL);
1686 rc = lustre_pack_reply(req, 3, repsize, NULL);
1688 RETURN(req->rq_status = rc);
1690 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1691 LASSERT(rep != NULL);
1693 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1694 sizeof(*reply_lvb));
1695 LASSERT(reply_lvb != NULL);
1697 //fixup_handle_for_resent_req(req, lock, &lockh);
1699 /* Call the extent policy function to see if our request can be
1700 * granted, or is blocked.
1701 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1702 * lock, and should not be granted if the lock will be blocked.
1705 LASSERT(ns == ldlm_res_to_ns(res));
1707 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1708 check_res_locked(res);
1710 /* FIXME: we should change the policy function slightly, to not make
1711 * this list at all, since we just turn around and free it */
1712 while (!cfs_list_empty(&rpc_list)) {
1713 struct ldlm_lock *wlock =
1714 cfs_list_entry(rpc_list.next, struct ldlm_lock,
1716 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1717 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1718 lock->l_flags &= ~LDLM_FL_CP_REQD;
1719 cfs_list_del_init(&wlock->l_cp_ast);
1720 LDLM_LOCK_RELEASE(wlock);
1723 /* The lock met with no resistance; we're finished. */
1724 if (rc == LDLM_ITER_CONTINUE) {
1725 /* do not grant locks to the liblustre clients: they cannot
1726 * handle ASTs robustly. We need to do this while still
1727 * holding lr_lock to avoid the lock remaining on the res_link
1728 * list (and potentially being added to l_pending_list by an
1729 * AST) when we are going to drop this lock ASAP. */
1730 if (lock->l_export->exp_libclient ||
1731 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1732 ldlm_resource_unlink_lock(lock);
1733 err = ELDLM_LOCK_ABORTED;
1735 err = ELDLM_LOCK_REPLACED;
1741 /* Do not grant any lock, but instead send GL callbacks. The extent
1742 * policy nicely created a list of all PW locks for us. We will choose
1743 * the highest of those which are larger than the size in the LVB, if
1744 * any, and perform a glimpse callback. */
1745 res_lvb = res->lr_lvb_data;
1746 LASSERT(res_lvb != NULL);
1747 *reply_lvb = *res_lvb;
1750 * lr_lock guarantees that no new locks are granted, and,
1751 * therefore, that res->lr_lvb_data cannot increase beyond the
1752 * end of already granted lock. As a result, it is safe to
1753 * check against "stale" reply_lvb->lvb_size value without
1756 arg.size = reply_lvb->lvb_size;
1758 arg.liblustre = &only_liblustre;
1759 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1760 tree = &res->lr_itree[idx];
1761 if (tree->lit_mode == LCK_PR)
1764 interval_iterate_reverse(tree->lit_root,
1765 filter_intent_cb, &arg);
1769 /* There were no PW locks beyond the size in the LVB; finished. */
1771 if (only_liblustre) {
1772 /* If we discovered a liblustre client with a PW lock,
1773 * however, the LVB may be out of date! The LVB is
1774 * updated only on glimpse (which we don't do for
1775 * liblustre clients) and cancel (which the client
1776 * obviously has not yet done). So if it has written
1777 * data but kept the lock, the LVB is stale and needs
1778 * to be updated from disk.
1780 * Of course, this will all disappear when we switch to
1781 * taking liblustre locks on the OST. */
1782 ldlm_res_lvbo_update(res, NULL, 1);
1784 RETURN(ELDLM_LOCK_ABORTED);
1788 * This check is for lock taken in filter_prepare_destroy() that does
1789 * not have l_glimpse_ast set. So the logic is: if there is a lock
1790 * with no l_glimpse_ast set, this object is being destroyed already.
1792 * Hence, if you are grabbing DLM locks on the server, always set
1793 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1795 if (l->l_glimpse_ast == NULL) {
1796 /* We are racing with unlink(); just return -ENOENT */
1797 rep->lock_policy_res1 = -ENOENT;
1801 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1802 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1805 *reply_lvb = *res_lvb;
1809 LDLM_LOCK_RELEASE(l);
1811 RETURN(ELDLM_LOCK_ABORTED);
1815 * per-obd_device iobuf pool.
1817 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1818 * allocations in write-path has to be minimized (see bug 5137).
1820 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1821 * OST threads (see ost_thread_{init,done}()).
1823 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1824 * at the OST layer there are only (potentially) multiple obd_device of type
1825 * unknown at the time of OST thread creation.
1827 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1828 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1829 * it's very own iobuf.
1833 * filter_kiobuf_pool_init()
1835 * filter_kiobuf_pool_done()
1837 * filter_iobuf_get()
1839 * operate on this array. They are "generic" in a sense that they don't depend
1840 * on actual type of iobuf's (the latter depending on Linux kernel version).
1844 * destroy pool created by filter_iobuf_pool_init
1846 static void filter_iobuf_pool_done(struct filter_obd *filter)
1848 struct filter_iobuf **pool;
1853 pool = filter->fo_iobuf_pool;
1855 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1856 if (pool[i] != NULL)
1857 filter_free_iobuf(pool[i]);
1859 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1860 filter->fo_iobuf_pool = NULL;
1865 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1867 struct filter_obd *filter = &obd->u.filter;
1868 struct sptlrpc_rule_set tmp_rset;
1871 sptlrpc_rule_set_init(&tmp_rset);
1872 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1874 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1879 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1881 cfs_write_lock(&filter->fo_sptlrpc_lock);
1882 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1883 filter->fo_sptlrpc_rset = tmp_rset;
1884 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1890 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1892 static int filter_iobuf_pool_init(struct filter_obd *filter)
1899 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1901 if (filter->fo_iobuf_pool == NULL)
1904 filter->fo_iobuf_count = OSS_THREADS_MAX;
1909 /* Return iobuf allocated for @thread_id. We don't know in advance how
1910 * many threads there will be so we allocate a large empty array and only
1911 * fill in those slots that are actually in use.
1912 * If we haven't allocated a pool entry for this thread before, do so now. */
1913 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1915 int thread_id = (oti && oti->oti_thread) ?
1916 oti->oti_thread->t_id : -1;
1917 struct filter_iobuf *pool = NULL;
1918 struct filter_iobuf **pool_place = NULL;
1920 if (thread_id >= 0) {
1921 LASSERT(thread_id < filter->fo_iobuf_count);
1922 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1925 if (unlikely(pool == NULL)) {
1926 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1927 PTLRPC_MAX_BRW_PAGES);
1928 if (pool_place != NULL)
1935 /* mount the file system (secretly). lustre_cfg parameters are:
1938 * 3 = flags: failover=f, failout=n
1941 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1944 struct filter_obd *filter = &obd->u.filter;
1945 struct vfsmount *mnt;
1946 struct lustre_mount_info *lmi;
1947 struct obd_uuid uuid;
1955 if (lcfg->lcfg_bufcount < 3 ||
1956 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1957 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1960 lmi = server_get_mount(obd->obd_name);
1962 /* We already mounted in lustre_fill_super.
1963 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1964 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1966 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1968 /* gets recovery timeouts from mount data */
1969 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1970 obd->obd_recovery_timeout =
1971 lsi->lsi_lmd->lmd_recovery_time_soft;
1972 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1973 obd->obd_recovery_time_hard =
1974 lsi->lsi_lmd->lmd_recovery_time_hard;
1976 /* old path - used by lctl */
1977 CERROR("Using old MDS mount method\n");
1978 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1979 MS_NOATIME|MS_NODIRATIME,
1980 lustre_cfg_string(lcfg, 1), option);
1983 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1984 lustre_cfg_string(lcfg, 1), rc);
1988 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1990 if (IS_ERR(obd->obd_fsops))
1991 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1993 rc = filter_iobuf_pool_init(filter);
1997 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
1998 CERROR("%s: Underlying device is marked as read-only. "
1999 "Setup failed\n", obd->obd_name);
2000 GOTO(err_ops, rc = -EROFS);
2003 /* failover is the default */
2004 obd->obd_replayable = 1;
2006 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2007 str = lustre_cfg_string(lcfg, 3);
2008 if (strchr(str, 'n')) {
2009 CWARN("%s: recovery disabled\n", obd->obd_name);
2010 obd->obd_replayable = 0;
2014 obd->u.obt.obt_vfsmnt = mnt;
2015 obd->u.obt.obt_sb = mnt->mnt_sb;
2016 obd->u.obt.obt_magic = OBT_MAGIC;
2017 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2018 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2020 fsfilt_setup(obd, obd->u.obt.obt_sb);
2022 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2023 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2024 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2025 obd->obd_lvfs_ctxt.fs = get_ds();
2026 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2028 cfs_init_mutex(&filter->fo_init_lock);
2029 filter->fo_committed_group = 0;
2030 filter->fo_destroys_in_progress = 0;
2031 for (i = 0; i < 32; i++)
2032 cfs_sema_init(&filter->fo_create_locks[i], 1);
2034 cfs_spin_lock_init(&filter->fo_objidlock);
2035 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2036 cfs_sema_init(&filter->fo_alloc_lock, 1);
2037 init_brw_stats(&filter->fo_filter_stats);
2038 cfs_spin_lock_init(&filter->fo_flags_lock);
2039 filter->fo_read_cache = 1; /* enable read-only cache by default */
2040 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2041 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2042 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2043 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2044 filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2045 filter_slc_set(filter); /* initialize sync on lock cancel */
2047 rc = filter_prep(obd);
2051 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2052 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2054 filter->fo_fl_oss_capa = 1;
2056 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2057 filter->fo_capa_hash = init_capa_hash();
2058 if (filter->fo_capa_hash == NULL)
2059 GOTO(err_post, rc = -ENOMEM);
2061 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2062 obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2063 LDLM_NAMESPACE_SERVER,
2064 LDLM_NAMESPACE_GREEDY,
2066 if (obd->obd_namespace == NULL)
2067 GOTO(err_post, rc = -ENOMEM);
2068 obd->obd_namespace->ns_lvbp = obd;
2069 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2070 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2072 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2073 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2075 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2077 CERROR("failed to setup llogging subsystems\n");
2081 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2082 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2083 /* do this after llog being initialized */
2084 filter_adapt_sptlrpc_conf(obd, 1);
2086 rc = lquota_setup(filter_quota_interface_ref, obd);
2090 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2091 if (q->max_sectors < q->max_hw_sectors &&
2092 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2093 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2094 "for larger I/O requests: max_sectors = %u "
2095 "could be up to max_hw_sectors=%u\n",
2096 obd->obd_name, mnt->mnt_sb->s_id,
2097 q->max_sectors, q->max_hw_sectors);
2099 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2100 if (uuid_ptr != NULL) {
2101 class_uuid_unparse(uuid_ptr, &uuid);
2107 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2108 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2109 obd->obd_name, label ?: str, lmi ? "on " : "",
2110 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2111 obd->obd_replayable ? "enabled" : "disabled");
2113 if (obd->obd_recovering)
2114 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2115 "or until %d client%s reconnect%s\n",
2117 obd->obd_recovery_timeout / 60,
2118 obd->obd_recovery_timeout % 60,
2119 obd->obd_max_recoverable_clients,
2120 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2121 (obd->obd_max_recoverable_clients == 1) ? "s": "");
2129 fsfilt_put_ops(obd->obd_fsops);
2130 filter_iobuf_pool_done(filter);
2132 server_put_mount(obd->obd_name, mnt);
2133 obd->u.obt.obt_sb = 0;
2137 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2139 struct lprocfs_static_vars lvars;
2144 CLASSERT(offsetof(struct obd_device, u.obt) ==
2145 offsetof(struct obd_device, u.filter.fo_obt));
2147 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2150 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2151 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2154 addr = (unsigned long)cfs_page_address(page);
2155 clear_page((void *)addr);
2157 /* lprocfs must be setup before the filter so state can be safely added
2158 * to /proc incrementally as the filter is setup */
2159 lprocfs_filter_init_vars(&lvars);
2160 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2161 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2162 /* Init obdfilter private stats here */
2163 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2164 LPROCFS_CNTR_AVGMINMAX,
2165 "read_bytes", "bytes");
2166 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2167 LPROCFS_CNTR_AVGMINMAX,
2168 "write_bytes", "bytes");
2169 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2170 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2171 "get_page", "usec");
2172 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2173 LPROCFS_CNTR_AVGMINMAX,
2174 "get_page_failures", "num");
2175 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2176 LPROCFS_CNTR_AVGMINMAX,
2177 "cache_access", "pages");
2178 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2179 LPROCFS_CNTR_AVGMINMAX,
2180 "cache_hit", "pages");
2181 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2182 LPROCFS_CNTR_AVGMINMAX,
2183 "cache_miss", "pages");
2185 lproc_filter_attach_seqstat(obd);
2186 obd->obd_proc_exports_entry = lprocfs_register("exports",
2187 obd->obd_proc_entry,
2189 if (IS_ERR(obd->obd_proc_exports_entry)) {
2190 rc = PTR_ERR(obd->obd_proc_exports_entry);
2191 CERROR("error %d setting up lprocfs for %s\n",
2193 obd->obd_proc_exports_entry = NULL;
2196 if (obd->obd_proc_exports_entry)
2197 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2198 lprocfs_nid_stats_clear_read,
2199 lprocfs_nid_stats_clear_write, obd, NULL);
2201 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2202 LUSTRE_CFG_BUFLEN(lcfg, 4));
2203 rc = filter_common_setup(obd, lcfg, (void *)addr);
2204 OBD_PAGE_FREE(page);
2207 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2208 lprocfs_free_per_client_stats(obd);
2209 lprocfs_free_obd_stats(obd);
2210 lprocfs_obd_cleanup(obd);
2216 static struct llog_operations filter_mds_ost_repl_logops;
2218 static struct llog_operations filter_size_orig_logops = {
2219 .lop_setup = llog_obd_origin_setup,
2220 .lop_cleanup = llog_obd_origin_cleanup,
2221 .lop_add = llog_obd_origin_add
2224 static int filter_olg_fini(struct obd_llog_group *olg)
2226 struct llog_ctxt *ctxt;
2227 int rc = 0, rc2 = 0;
2230 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2232 rc = llog_cleanup(ctxt);
2234 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2236 rc2 = llog_cleanup(ctxt);
2241 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2243 rc2 = llog_cleanup(ctxt);
2252 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2253 struct obd_device *tgt)
2258 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2259 &filter_mds_ost_repl_logops);
2263 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2264 &filter_size_orig_logops);
2270 filter_olg_fini(olg);
2275 * Init the default olg, which is embeded in the obd_device, for filter.
2278 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2279 struct obd_device *tgt)
2281 struct filter_obd *filter = &obd->u.filter;
2282 struct llog_ctxt *ctxt;
2286 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2287 if (!filter->fo_lcm)
2290 filter_mds_ost_repl_logops = llog_client_ops;
2291 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2292 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2293 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2295 rc = filter_olg_init(obd, olg, tgt);
2297 GOTO(cleanup_lcm, rc);
2299 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2302 GOTO(cleanup_olg, rc);
2304 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2306 CERROR("Can't get ctxt for %p:%x\n", olg,
2307 LLOG_MDS_OST_REPL_CTXT);
2308 GOTO(cleanup_olg, rc = -ENODEV);
2310 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2311 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2312 llog_ctxt_put(ctxt);
2316 filter_olg_fini(olg);
2318 llog_recov_thread_fini(filter->fo_lcm, 1);
2319 filter->fo_lcm = NULL;
2324 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2325 struct obd_device *tgt, int *index)
2327 struct filter_obd *filter = &obd->u.filter;
2328 struct llog_ctxt *ctxt;
2332 LASSERT(olg != NULL);
2333 if (olg == &obd->obd_olg)
2334 return filter_default_olg_init(obd, olg, tgt);
2336 LASSERT(filter->fo_lcm != NULL);
2337 rc = filter_olg_init(obd, olg, tgt);
2340 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2342 CERROR("Can't get ctxt for %p:%x\n", olg,
2343 LLOG_MDS_OST_REPL_CTXT);
2344 filter_olg_fini(olg);
2347 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2348 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2349 llog_ctxt_put(ctxt);
2353 static int filter_llog_finish(struct obd_device *obd, int count)
2355 struct filter_obd *filter = &obd->u.filter;
2356 struct llog_ctxt *ctxt;
2359 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2362 * Make sure that no cached llcds left in recov_thread.
2363 * We actually do sync in disconnect time, but disconnect
2364 * may not come being marked rq_no_resend = 1.
2366 llog_sync(ctxt, NULL);
2369 * Balance class_import_get() in llog_receptor_accept().
2370 * This is safe to do, as llog is already synchronized
2371 * and its import may go.
2373 cfs_mutex_down(&ctxt->loc_sem);
2374 if (ctxt->loc_imp) {
2375 class_import_put(ctxt->loc_imp);
2376 ctxt->loc_imp = NULL;
2378 cfs_mutex_up(&ctxt->loc_sem);
2379 llog_ctxt_put(ctxt);
2382 if (filter->fo_lcm) {
2383 cfs_mutex_down(&ctxt->loc_sem);
2384 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2385 filter->fo_lcm = NULL;
2386 cfs_mutex_up(&ctxt->loc_sem);
2388 RETURN(filter_olg_fini(&obd->obd_olg));
2392 * Find the group llog according to group index in the llog group list.
2394 static struct obd_llog_group *
2395 filter_find_olg_internal(struct filter_obd *filter, int group)
2397 struct obd_llog_group *olg;
2399 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2400 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2401 if (olg->olg_seq == group)
2408 * Find the group llog according to group index on the filter
2410 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2412 struct obd_llog_group *olg = NULL;
2413 struct filter_obd *filter;
2415 filter = &obd->u.filter;
2417 if (group == FID_SEQ_LLOG)
2418 RETURN(&obd->obd_olg);
2420 cfs_spin_lock(&filter->fo_llog_list_lock);
2421 olg = filter_find_olg_internal(filter, group);
2422 cfs_spin_unlock(&filter->fo_llog_list_lock);
2427 * Find the llog_group of the filter according to the group. If it can not
2428 * find, create the llog_group, which only happens when mds is being synced
2431 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2433 struct obd_llog_group *olg = NULL;
2434 struct filter_obd *filter;
2437 filter = &obd->u.filter;
2439 if (group == FID_SEQ_LLOG)
2440 RETURN(&obd->obd_olg);
2442 cfs_spin_lock(&filter->fo_llog_list_lock);
2443 olg = filter_find_olg_internal(filter, group);
2445 if (olg->olg_initializing) {
2446 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2448 GOTO(out_unlock, olg);
2453 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2455 llog_group_init(olg, group);
2456 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2457 olg->olg_initializing = 1;
2458 cfs_spin_unlock(&filter->fo_llog_list_lock);
2460 rc = obd_llog_init(obd, olg, obd, NULL);
2462 cfs_spin_lock(&filter->fo_llog_list_lock);
2463 cfs_list_del(&olg->olg_list);
2464 cfs_spin_unlock(&filter->fo_llog_list_lock);
2466 GOTO(out, olg = ERR_PTR(-ENOMEM));
2468 cfs_spin_lock(&filter->fo_llog_list_lock);
2469 olg->olg_initializing = 0;
2470 cfs_spin_unlock(&filter->fo_llog_list_lock);
2471 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2472 obd->obd_name, group, olg);
2477 cfs_spin_unlock(&filter->fo_llog_list_lock);
2481 static int filter_llog_connect(struct obd_export *exp,
2482 struct llogd_conn_body *body)
2484 struct obd_device *obd = exp->exp_obd;
2485 struct llog_ctxt *ctxt;
2486 struct obd_llog_group *olg;
2490 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2491 obd->obd_name, body->lgdc_logid.lgl_oid,
2492 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2494 olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2496 CERROR(" %s: can not find olg of group %d\n",
2497 obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2500 llog_group_set_export(olg, exp);
2502 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2503 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2504 body->lgdc_ctxt_idx);
2506 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2507 obd->obd_name, body->lgdc_logid.lgl_oid,
2508 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2510 cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2511 obd->u.filter.fo_mds_ost_sync = 1;
2512 cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2513 rc = llog_connect(ctxt, &body->lgdc_logid,
2514 &body->lgdc_gen, NULL);
2515 llog_ctxt_put(ctxt);
2517 CERROR("failed to connect rc %d idx %d\n", rc,
2518 body->lgdc_ctxt_idx);
2523 static int filter_llog_preclean(struct obd_device *obd)
2525 struct obd_llog_group *olg, *tmp;
2526 struct filter_obd *filter;
2527 cfs_list_t remove_list;
2531 rc = obd_llog_finish(obd, 0);
2533 CERROR("failed to cleanup llogging subsystem\n");
2535 filter = &obd->u.filter;
2536 CFS_INIT_LIST_HEAD(&remove_list);
2538 cfs_spin_lock(&filter->fo_llog_list_lock);
2539 while (!cfs_list_empty(&filter->fo_llog_list)) {
2540 olg = cfs_list_entry(filter->fo_llog_list.next,
2541 struct obd_llog_group, olg_list);
2542 cfs_list_del(&olg->olg_list);
2543 cfs_list_add(&olg->olg_list, &remove_list);
2545 cfs_spin_unlock(&filter->fo_llog_list_lock);
2547 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2548 cfs_list_del_init(&olg->olg_list);
2549 rc = filter_olg_fini(olg);
2551 CERROR("failed to cleanup llogging subsystem for %u\n",
2559 static int filter_precleanup(struct obd_device *obd,
2560 enum obd_cleanup_stage stage)
2566 case OBD_CLEANUP_EARLY:
2568 case OBD_CLEANUP_EXPORTS:
2569 /* Stop recovery before namespace cleanup. */
2570 target_recovery_fini(obd);
2571 rc = filter_llog_preclean(obd);
2577 static int filter_cleanup(struct obd_device *obd)
2579 struct filter_obd *filter = &obd->u.filter;
2583 LCONSOLE_WARN("%s: shutting down for failover; client state "
2584 "will be preserved.\n", obd->obd_name);
2586 obd_exports_barrier(obd);
2587 obd_zombie_barrier();
2589 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2590 lprocfs_free_per_client_stats(obd);
2591 lprocfs_free_obd_stats(obd);
2592 lprocfs_obd_cleanup(obd);
2593 lquota_cleanup(filter_quota_interface_ref, obd);
2595 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2596 obd->obd_namespace = NULL;
2598 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2600 if (obd->u.obt.obt_sb == NULL)
2605 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2606 shrink_dcache_sb(obd->u.obt.obt_sb);
2608 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2609 obd->u.obt.obt_sb = NULL;
2611 fsfilt_put_ops(obd->obd_fsops);
2613 filter_iobuf_pool_done(filter);
2615 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2620 static int filter_connect_internal(struct obd_export *exp,
2621 struct obd_connect_data *data,
2624 struct filter_export_data *fed = &exp->exp_filter_data;
2629 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2630 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2631 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2632 data->ocd_connect_flags, data->ocd_version,
2633 data->ocd_grant, data->ocd_index);
2635 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2636 CWARN("!!! This export (nid %s) used object group %d "
2637 "earlier; now it's trying to use group %d! This could "
2638 "be a bug in the MDS. Please report to "
2639 "http://bugzilla.lustre.org/\n",
2640 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2643 fed->fed_group = data->ocd_group;
2645 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2646 exp->exp_connect_flags = data->ocd_connect_flags;
2647 data->ocd_version = LUSTRE_VERSION_CODE;
2649 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2650 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2651 data->ocd_connect_flags & OBD_CONNECT_MDS))
2654 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2655 struct filter_obd *filter = &exp->exp_obd->u.filter;
2656 obd_size left, want;
2658 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2659 left = filter_grant_space_left(exp);
2660 want = data->ocd_grant;
2661 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2662 data->ocd_grant = fed->fed_grant;
2663 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2665 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2666 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2667 exp->exp_client_uuid.uuid, exp,
2668 data->ocd_grant, want, left);
2670 filter->fo_tot_granted_clients ++;
2673 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2674 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2675 int index = le32_to_cpu(lsd->lsd_ost_index);
2677 if (!(lsd->lsd_feature_compat &
2678 cpu_to_le32(OBD_COMPAT_OST))) {
2679 /* this will only happen on the first connect */
2680 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2681 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2682 /* sync is not needed here as filter_client_add will
2683 * set exp_need_sync flag */
2684 filter_update_server_data(exp->exp_obd);
2685 } else if (index != data->ocd_index) {
2686 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2687 " %u doesn't match actual OST index"
2688 " %u in last_rcvd file, bad "
2690 obd_export_nid2str(exp), index,
2694 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2695 * FIXME: We don't strictly need the COMPAT flag for that,
2696 * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2697 * FIXME: We needed it for the index, as index 0 is valid. */
2700 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2701 data->ocd_brw_size = 65536;
2702 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2703 data->ocd_brw_size = min(data->ocd_brw_size,
2704 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2705 LASSERT(data->ocd_brw_size);
2708 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2709 __u32 cksum_types = data->ocd_cksum_types;
2711 /* The client set in ocd_cksum_types the checksum types it
2712 * supports. We have to mask off the algorithms that we don't
2714 if (cksum_types & OBD_CKSUM_ALL)
2715 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2717 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2719 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2720 "%x\n", exp->exp_obd->obd_name,
2721 obd_export_nid2str(exp), cksum_types,
2722 data->ocd_cksum_types);
2724 /* This client does not support OBD_CONNECT_CKSUM
2725 * fall back to CRC32 */
2726 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2727 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2728 exp->exp_obd->obd_name,
2729 obd_export_nid2str(exp));
2735 static int filter_reconnect(const struct lu_env *env,
2736 struct obd_export *exp, struct obd_device *obd,
2737 struct obd_uuid *cluuid,
2738 struct obd_connect_data *data,
2744 if (exp == NULL || obd == NULL || cluuid == NULL)
2747 rc = filter_connect_internal(exp, data, 1);
2749 filter_export_stats_init(obd, exp, 1, localdata);
2754 /* nearly identical to mds_connect */
2755 static int filter_connect(const struct lu_env *env,
2756 struct obd_export **exp, struct obd_device *obd,
2757 struct obd_uuid *cluuid,
2758 struct obd_connect_data *data, void *localdata)
2760 struct lvfs_run_ctxt saved;
2761 struct lustre_handle conn = { 0 };
2762 struct obd_export *lexp;
2767 if (exp == NULL || obd == NULL || cluuid == NULL)
2770 rc = class_connect(&conn, obd, cluuid);
2773 lexp = class_conn2export(&conn);
2774 LASSERT(lexp != NULL);
2776 rc = filter_connect_internal(lexp, data, 0);
2780 filter_export_stats_init(obd, lexp, 0, localdata);
2781 if (obd->obd_replayable) {
2782 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2784 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2785 rc = filter_client_add(obd, lexp, -1);
2790 group = data->ocd_group;
2792 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2793 obd->obd_name, lexp->exp_handle.h_cookie, group);
2795 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2796 rc = filter_read_groups(obd, group, 1);
2797 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2799 CERROR("can't read group %u\n", group);
2807 class_disconnect(lexp);
2816 /* Do extra sanity checks for grant accounting. We do this at connect,
2817 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2818 * always get rid of it or turn it off when we know accounting is good. */
2819 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2821 struct filter_export_data *fed;
2822 struct obd_export *exp;
2823 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2824 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2825 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2827 if (cfs_list_empty(&obd->obd_exports))
2830 /* We don't want to do this for large machines that do lots of
2831 mounts or unmounts. It burns... */
2832 if (obd->obd_num_exports > 100)
2835 cfs_spin_lock(&obd->obd_osfs_lock);
2836 cfs_spin_lock(&obd->obd_dev_lock);
2837 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2839 fed = &exp->exp_filter_data;
2840 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2843 if (maxsize > 0) { /* we may not have done a statfs yet */
2844 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2845 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2846 exp->exp_client_uuid.uuid, exp,
2847 fed->fed_grant, fed->fed_pending, maxsize);
2848 LASSERTF(fed->fed_dirty <= maxsize,
2849 "%s: cli %s/%p %ld > "LPU64"\n", func,
2850 exp->exp_client_uuid.uuid, exp,
2851 fed->fed_dirty, maxsize);
2854 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2855 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2856 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2858 CDEBUG(D_CACHE, "%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);
2861 tot_granted += fed->fed_grant + fed->fed_pending;
2862 tot_pending += fed->fed_pending;
2863 tot_dirty += fed->fed_dirty;
2865 fo_tot_granted = obd->u.filter.fo_tot_granted;
2866 fo_tot_pending = obd->u.filter.fo_tot_pending;
2867 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2868 cfs_spin_unlock(&obd->obd_dev_lock);
2869 cfs_spin_unlock(&obd->obd_osfs_lock);
2871 /* Do these assertions outside the spinlocks so we don't kill system */
2872 if (tot_granted != fo_tot_granted)
2873 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2874 func, tot_granted, fo_tot_granted);
2875 if (tot_pending != fo_tot_pending)
2876 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2877 func, tot_pending, fo_tot_pending);
2878 if (tot_dirty != fo_tot_dirty)
2879 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2880 func, tot_dirty, fo_tot_dirty);
2881 if (tot_pending > tot_granted)
2882 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2883 func, tot_pending, tot_granted);
2884 if (tot_granted > maxsize)
2885 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2886 func, tot_granted, maxsize);
2887 if (tot_dirty > maxsize)
2888 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2889 func, tot_dirty, maxsize);
2892 /* Remove this client from the grant accounting totals. We also remove
2893 * the export from the obd device under the osfs and dev locks to ensure
2894 * that the filter_grant_sanity_check() calculations are always valid.
2895 * The client should do something similar when it invalidates its import. */
2896 static void filter_grant_discard(struct obd_export *exp)
2898 struct obd_device *obd = exp->exp_obd;
2899 struct filter_obd *filter = &obd->u.filter;
2900 struct filter_export_data *fed = &exp->exp_filter_data;
2902 cfs_spin_lock(&obd->obd_osfs_lock);
2903 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2904 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2905 obd->obd_name, filter->fo_tot_granted,
2906 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2907 filter->fo_tot_granted -= fed->fed_grant;
2908 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2909 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2910 obd->obd_name, filter->fo_tot_pending,
2911 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2912 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2913 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2914 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2915 obd->obd_name, filter->fo_tot_dirty,
2916 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2917 filter->fo_tot_dirty -= fed->fed_dirty;
2921 cfs_spin_unlock(&obd->obd_osfs_lock);
2924 static int filter_destroy_export(struct obd_export *exp)
2926 struct filter_export_data *fed = &exp->exp_filter_data;
2929 if (fed->fed_pending)
2930 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2931 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2932 exp, fed->fed_pending);
2934 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2936 target_destroy_export(exp);
2937 ldlm_destroy_export(exp);
2938 lut_client_free(exp);
2940 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2943 if (!exp->exp_obd->obd_replayable)
2944 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2946 filter_grant_discard(exp);
2947 filter_fmd_cleanup(exp);
2949 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2950 struct filter_obd *filter = &exp->exp_obd->u.filter;
2951 if (filter->fo_tot_granted_clients > 0)
2952 filter->fo_tot_granted_clients --;
2955 if (!(exp->exp_flags & OBD_OPT_FORCE))
2956 filter_grant_sanity_check(exp->exp_obd, __func__);
2961 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2963 struct obd_llog_group *olg_min, *olg;
2964 struct filter_obd *filter;
2965 int worked = -1, group;
2966 struct llog_ctxt *ctxt;
2969 filter = &obd->u.filter;
2971 /* we can't sync log holding spinlock. also, we do not want to get
2972 * into livelock. so we do following: loop over MDS's exports in
2973 * group order and skip already synced llogs -bzzz */
2975 /* look for group with min. number, but > worked */
2978 cfs_spin_lock(&filter->fo_llog_list_lock);
2979 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2980 if (olg->olg_seq <= worked) {
2981 /* this group is already synced */
2984 if (group < olg->olg_seq) {
2985 /* we have group with smaller number to sync */
2988 /* store current minimal group */
2990 group = olg->olg_seq;
2992 cfs_spin_unlock(&filter->fo_llog_list_lock);
2994 if (olg_min == NULL)
2997 worked = olg_min->olg_seq;
2998 if (olg_min->olg_exp &&
2999 (dexp == olg_min->olg_exp || dexp == NULL)) {
3001 ctxt = llog_group_get_ctxt(olg_min,
3002 LLOG_MDS_OST_REPL_CTXT);
3004 err = llog_sync(ctxt, olg_min->olg_exp);
3005 llog_ctxt_put(ctxt);
3007 CERROR("error flushing logs to MDS: "
3012 } while (olg_min != NULL);
3015 /* Also incredibly similar to mds_disconnect */
3016 static int filter_disconnect(struct obd_export *exp)
3018 struct obd_device *obd = exp->exp_obd;
3023 class_export_get(exp);
3025 if (!(exp->exp_flags & OBD_OPT_FORCE))
3026 filter_grant_sanity_check(obd, __func__);
3027 filter_grant_discard(exp);
3029 /* Flush any remaining cancel messages out to the target */
3030 filter_sync_llogs(obd, exp);
3032 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3034 rc = server_disconnect_export(exp);
3036 /* Do not erase record for recoverable client. */
3037 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3038 filter_client_del(exp);
3040 fsfilt_sync(obd, obd->u.obt.obt_sb);
3042 class_export_put(exp);
3046 /* reverse import is changed, sync all cancels */
3047 static void filter_revimp_update(struct obd_export *exp)
3052 class_export_get(exp);
3054 /* flush any remaining cancel messages out to the target */
3055 filter_sync_llogs(exp->exp_obd, exp);
3056 class_export_put(exp);
3060 static int filter_ping(struct obd_export *exp)
3062 filter_fmd_expire(exp);
3066 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3067 const char *what, int quiet)
3069 struct dentry *dchild = NULL;
3071 dchild = filter_fid2dentry(obd, NULL, ostid->oi_seq, ostid->oi_id);
3073 if (IS_ERR(dchild)) {
3074 CERROR("%s error looking up object: "POSTID"\n",
3075 what, ostid->oi_id, ostid->oi_seq);
3079 if (dchild->d_inode == NULL) {
3081 CERROR("%s: %s on non-existent object: "POSTID" \n",
3082 obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3084 RETURN(ERR_PTR(-ENOENT));
3090 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3092 struct dentry *dentry = NULL;
3093 struct obd_device *obd;
3097 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3098 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3102 obd = class_exp2obd(exp);
3104 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3108 dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3110 RETURN(PTR_ERR(dentry));
3112 /* Limit the valid bits in the return data to what we actually use */
3113 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3114 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3120 /* this should be enabled/disabled in condition to enabled/disabled large
3121 * inodes (fast EAs) in backing store FS. */
3122 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3123 void *handle, struct obdo *oa)
3125 struct obd_device *obd = exp->exp_obd;
3129 if (oa->o_valid & OBD_MD_FLFID) {
3130 struct filter_fid ff;
3132 if (!(oa->o_valid & OBD_MD_FLGROUP))
3134 /* packing fid and converting it to LE for storing into EA.
3135 * Here ->o_stripe_idx should be filled by LOV and rest of
3136 * fields - by client. */
3137 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3138 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3139 /* XXX: we are ignoring o_parent_ver here, since this should
3140 * be the same for all objects in this fileset. */
3141 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3142 ff.ff_objid = cpu_to_le64(oa->o_id);
3143 ff.ff_seq = cpu_to_le64(oa->o_seq);
3145 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3146 LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3149 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3151 CERROR("store fid in object failed! rc: %d\n", rc);
3153 CDEBUG(D_HA, "OSS object without fid info!\n");
3159 /* this is called from filter_truncate() until we have filter_punch() */
3160 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3161 struct obdo *oa, struct obd_trans_info *oti)
3163 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3164 struct llog_cookie *fcc = NULL;
3165 struct filter_obd *filter;
3166 int rc, err, sync = 0;
3167 loff_t old_size = 0;
3168 unsigned int ia_valid;
3169 struct inode *inode;
3170 struct page *page = NULL;
3175 LASSERT(dentry != NULL);
3176 LASSERT(!IS_ERR(dentry));
3178 inode = dentry->d_inode;
3179 LASSERT(inode != NULL);
3181 filter = &exp->exp_obd->u.filter;
3182 iattr_from_obdo(&iattr, oa, oa->o_valid);
3183 ia_valid = iattr.ia_valid;
3185 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3186 OBD_ALLOC(fcc, sizeof(*fcc));
3188 *fcc = oa->o_lcookie;
3190 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3191 unsigned long now = jiffies;
3192 ll_vfs_dq_init(inode);
3193 /* Filter truncates and writes are serialized by
3194 * i_alloc_sem, see the comment in
3195 * filter_preprw_write.*/
3196 if (ia_valid & ATTR_SIZE)
3197 down_write(&inode->i_alloc_sem);
3198 LOCK_INODE_MUTEX(inode);
3199 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
3200 old_size = i_size_read(inode);
3203 /* VBR: version recovery check */
3204 rc = filter_version_get_check(exp, oti, inode);
3206 GOTO(out_unlock, rc);
3208 /* Let's pin the last page so that ldiskfs_truncate
3209 * should not start GFP_FS allocation. */
3210 if (ia_valid & ATTR_SIZE) {
3211 page = grab_cache_page(inode->i_mapping,
3212 iattr.ia_size >> PAGE_CACHE_SHIFT);
3214 GOTO(out_unlock, rc = -ENOMEM);
3219 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3220 * then we will accept the UID+GID sent by the client during write for
3221 * initializing the ownership of this inode. We only allow this to
3222 * happen once so clear these bits in setattr. In 2.6 kernels it is
3223 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3224 * the flags that are actually being set. */
3225 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3226 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3227 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3229 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3230 if (!(ia_valid & ATTR_MODE)) {
3231 iattr.ia_mode = inode->i_mode;
3232 iattr.ia_valid |= ATTR_MODE;
3234 iattr.ia_mode &= ~S_ISUID;
3236 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3237 if (!(iattr.ia_valid & ATTR_MODE)) {
3238 iattr.ia_mode = inode->i_mode;
3239 iattr.ia_valid |= ATTR_MODE;
3241 iattr.ia_mode &= ~S_ISGID;
3244 orig_ids[USRQUOTA] = inode->i_uid;
3245 orig_ids[GRPQUOTA] = inode->i_gid;
3246 handle = fsfilt_start_log(exp->exp_obd, inode,
3247 FSFILT_OP_SETATTR, oti, 1);
3249 GOTO(out_unlock, rc = PTR_ERR(handle));
3251 /* update inode EA only once when inode is suid bit marked. As
3252 * on 2.6.x UID and GID may be set separately, we check here
3253 * only one of them to avoid double setting. */
3254 if (inode->i_mode & S_ISUID)
3255 filter_update_fidea(exp, inode, handle, oa);
3257 handle = fsfilt_start(exp->exp_obd, inode,
3258 FSFILT_OP_SETATTR, oti);
3260 GOTO(out_unlock, rc = PTR_ERR(handle));
3262 if (oa->o_valid & OBD_MD_FLFLAGS) {
3263 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3264 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3266 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3268 /* set cancel cookie callback function */
3269 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3270 filter_cancel_cookies_cb,
3274 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3275 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3277 /* The truncate might have used up our transaction credits. Make sure
3278 * we have two left for the last_rcvd and VBR inode version updates. */
3279 err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3281 /* Update inode version only if data has changed => size has changed */
3282 rc = filter_finish_transno(exp, ia_valid & ATTR_SIZE ? inode : NULL,
3286 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3290 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3292 CERROR("error on commit, err = %d\n", err);
3299 /* For a partial-page truncate flush the page to disk immediately
3300 * to avoid data corruption during direct disk write. b=17397 */
3301 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3302 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3303 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3313 page_cache_release(page);
3315 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3316 UNLOCK_INODE_MUTEX(inode);
3317 if (ia_valid & ATTR_SIZE)
3318 up_write(&inode->i_alloc_sem);
3320 OBD_FREE(fcc, sizeof(*fcc));
3322 /* trigger quota release */
3323 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3324 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3325 int rc2 = lquota_adjust(filter_quota_interface_ref,
3326 exp->exp_obd, cur_ids,
3327 orig_ids, rc, FSFILT_OP_SETATTR);
3328 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3329 "filter adjust qunit. (rc:%d)\n", rc2);
3334 /* this is called from filter_truncate() until we have filter_punch() */
3335 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3336 struct obd_trans_info *oti)
3338 struct obdo *oa = oinfo->oi_oa;
3339 struct lustre_capa *capa = oinfo_capa(oinfo);
3340 struct ldlm_res_id res_id;
3341 struct filter_mod_data *fmd;
3342 struct lvfs_run_ctxt saved;
3343 struct filter_obd *filter;
3344 struct ldlm_resource *res;
3345 struct dentry *dentry;
3346 __u64 opc = CAPA_OPC_META_WRITE;
3350 if (oa->o_valid & OBD_FL_TRUNC)
3351 opc |= CAPA_OPC_OSS_TRUNC;
3353 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3357 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3358 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3363 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3364 /* This would be very bad - accidentally truncating a file when
3365 * changing the time or similar - bug 12203. */
3366 if (oa->o_valid & OBD_MD_FLSIZE &&
3367 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3368 static char mdsinum[48];
3370 if (oa->o_valid & OBD_MD_FLFID)
3371 snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3372 oa->o_parent_seq, oa->o_parent_oid,
3377 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3378 "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3379 oa->o_id, oa->o_seq, mdsinum);
3383 dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3385 RETURN(PTR_ERR(dentry));
3387 filter = &exp->exp_obd->u.filter;
3388 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3391 * We need to be atomic against a concurrent write
3392 * (which takes the semaphore for reading). fmd_mactime_xid
3393 * checks will have no effect if a write request with lower
3394 * xid starts just before a setattr and finishes later than
3395 * the setattr (see bug 21489, comment 27).
3398 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3399 unsigned long now = jiffies;
3400 down_write(&dentry->d_inode->i_alloc_sem);
3401 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem");
3402 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3403 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3404 fmd->fmd_mactime_xid = oti->oti_xid;
3405 filter_fmd_put(exp, fmd);
3406 up_write(&dentry->d_inode->i_alloc_sem);
3409 /* setting objects attributes (including owner/group) */
3410 rc = filter_setattr_internal(exp, dentry, oa, oti);
3412 GOTO(out_unlock, rc);
3414 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3415 &res_id, LDLM_EXTENT, 0);
3418 LDLM_RESOURCE_ADDREF(res);
3419 rc = ldlm_res_lvbo_update(res, NULL, 0);
3420 LDLM_RESOURCE_DELREF(res);
3421 ldlm_resource_putref(res);
3424 oa->o_valid = OBD_MD_FLID;
3426 /* Quota release need uid/gid info */
3427 obdo_from_inode(oa, dentry->d_inode, NULL,
3428 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3433 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3437 /* XXX identical to osc_unpackmd */
3438 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3439 struct lov_mds_md *lmm, int lmm_bytes)
3445 if (lmm_bytes < sizeof (*lmm)) {
3446 CERROR("lov_mds_md too small: %d, need %d\n",
3447 lmm_bytes, (int)sizeof(*lmm));
3450 /* XXX LOV_MAGIC etc check? */
3452 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3453 CERROR("lov_mds_md: zero lmm_object_id\n");
3458 lsm_size = lov_stripe_md_size(1);
3462 if (*lsmp != NULL && lmm == NULL) {
3463 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3464 OBD_FREE(*lsmp, lsm_size);
3469 if (*lsmp == NULL) {
3470 OBD_ALLOC(*lsmp, lsm_size);
3474 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3475 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3476 OBD_FREE(*lsmp, lsm_size);
3479 loi_init((*lsmp)->lsm_oinfo[0]);
3483 /* XXX zero *lsmp? */
3484 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3485 LASSERT((*lsmp)->lsm_object_id);
3488 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3493 /* caller must hold fo_create_locks[oa->o_seq] */
3494 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3495 struct filter_obd *filter)
3497 struct obdo doa = { 0 }; /* XXX obdo on stack */
3503 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_seq]) != 0);
3505 memset(&doa, 0, sizeof(doa));
3507 doa.o_valid |= OBD_MD_FLGROUP;
3508 doa.o_seq = oa->o_seq;
3509 doa.o_mode = S_IFREG;
3511 if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3512 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3513 exp->exp_obd->obd_name, doa.o_seq);
3517 last = filter_last_id(filter, doa.o_seq);
3519 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3521 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3522 exp->exp_obd->obd_name, oa->o_id + 1, last,
3523 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3525 for (id = last; id > oa->o_id; id--) {
3527 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3528 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3529 CEMERG("error destroying precreate objid "LPU64": %d\n",
3532 /* update last_id on disk periodically so that if we restart
3533 * we don't need to re-scan all of the just-deleted objects. */
3534 if ((id & 511) == 0 && !skip_orphan) {
3535 filter_set_last_id(filter, id - 1, doa.o_seq);
3536 filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3540 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3541 exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3544 filter_set_last_id(filter, id, doa.o_seq);
3545 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3548 * We have destroyed orphan objects, but don't want to reuse
3549 * them. Therefore we don't reset last_id to the last created
3550 * objects. Instead, we report back to the MDS the object id
3551 * of the last orphan, so that the MDS can restart allocating
3552 * objects from this id + 1 and thus skip the whole orphan
3558 cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3563 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3564 obd_seq group, int *num);
3565 /* returns a negative error or a nonnegative number of files to create */
3566 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3567 obd_seq group, struct obd_trans_info *oti)
3569 struct obd_device *obd = exp->exp_obd;
3570 struct filter_obd *filter = &obd->u.filter;
3574 /* delete orphans request */
3575 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3576 obd_id last = filter_last_id(filter, group);
3578 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3579 CERROR("%s: dropping old orphan cleanup request\n",
3583 /* This causes inflight precreates to abort and drop lock */
3584 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3585 cfs_down(&filter->fo_create_locks[group]);
3586 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3587 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3588 exp->exp_obd->obd_name, group);
3589 cfs_up(&filter->fo_create_locks[group]);
3592 diff = oa->o_id - last;
3593 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3596 if (-diff > OST_MAX_PRECREATE) {
3597 CERROR("%s: ignoring bogus orphan destroy request: "
3598 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3600 /* FIXME: should reset precreate_next_id on MDS */
3601 GOTO(out, rc = -EINVAL);
3604 rc = filter_destroy_precreated(exp, oa, filter);
3606 CERROR("%s: unable to write lastobjid, but "
3607 "orphans were deleted\n", obd->obd_name);
3610 /* XXX: Used by MDS for the first time! */
3611 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3614 cfs_down(&filter->fo_create_locks[group]);
3615 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3616 CERROR("%s: dropping old precreate request\n",
3620 /* only precreate if group == 0 and o_id is specfied */
3621 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3624 diff = oa->o_id - filter_last_id(filter, group);
3625 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3626 filter_last_id(filter, group), diff);
3628 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3629 oa->o_id, filter_last_id(filter, group), diff);
3633 oa->o_id = filter_last_id(&obd->u.filter, group);
3634 rc = filter_precreate(obd, oa, group, &diff);
3635 oa->o_id = filter_last_id(&obd->u.filter, group);
3637 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3640 /* else diff == 0 */
3643 cfs_up(&filter->fo_create_locks[group]);
3647 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3648 __u64 max_age, __u32 flags)
3650 struct filter_obd *filter = &obd->u.filter;
3651 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3655 /* at least try to account for cached pages. its still racey and
3656 * might be under-reporting if clients haven't announced their
3657 * caches with brw recently */
3658 cfs_spin_lock(&obd->obd_osfs_lock);
3659 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3660 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3661 cfs_spin_unlock(&obd->obd_osfs_lock);
3663 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3664 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3665 filter->fo_tot_dirty, filter->fo_tot_granted,
3666 filter->fo_tot_pending,
3667 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3669 filter_grant_sanity_check(obd, __func__);
3671 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3672 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3673 osfs->os_bsize - 1) >> blockbits));
3675 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3676 struct lr_server_data *lsd = class_server_data(obd);
3677 int index = le32_to_cpu(lsd->lsd_ost_index);
3679 if (obd_fail_val == -1 ||
3680 index == obd_fail_val)
3681 osfs->os_bfree = osfs->os_bavail = 2;
3682 else if (obd_fail_loc & OBD_FAIL_ONCE)
3683 obd_fail_loc &= ~OBD_FAILED; /* reset flag */
3686 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3687 * stop creating files on MDS if OST is not good shape to create
3691 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3692 osfs->os_state = OS_STATE_READONLY;
3694 if (filter->fo_raid_degraded)
3695 osfs->os_state |= OS_STATE_DEGRADED;
3699 static int filter_use_existing_obj(struct obd_device *obd,
3700 struct dentry *dchild, void **handle,
3703 struct inode *inode = dchild->d_inode;
3707 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3710 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3711 if (IS_ERR(*handle))
3712 return PTR_ERR(*handle);
3714 iattr.ia_valid = ATTR_MODE;
3715 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3716 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3723 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3726 __u64 os_ffree = -1;
3728 cfs_spin_lock(&obd->obd_osfs_lock);
3729 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3731 os_ffree = obd->obd_osfs.os_ffree;
3732 cfs_spin_unlock(&obd->obd_osfs_lock);
3737 /* We rely on the fact that only one thread will be creating files in a given
3738 * group at a time, which is why we don't need an atomic filter_get_new_id.
3739 * Even if we had that atomic function, the following race would exist:
3741 * thread 1: gets id x from filter_next_id
3742 * thread 2: gets id (x + 1) from filter_next_id
3743 * thread 2: creates object (x + 1)
3744 * thread 1: tries to create object x, gets -ENOSPC
3746 * Caller must hold fo_create_locks[group]
3748 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3749 obd_seq group, int *num)
3751 struct dentry *dchild = NULL, *dparent = NULL;
3752 struct filter_obd *filter;
3753 struct obd_statfs *osfs;
3754 int err = 0, rc = 0, recreate_obj = 0, i;
3755 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3758 void *handle = NULL;
3761 filter = &obd->u.filter;
3763 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3765 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3767 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3768 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3771 OBD_ALLOC(osfs, sizeof(*osfs));
3774 rc = filter_statfs(obd, osfs,
3775 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
3777 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3778 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3779 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3780 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3784 OBD_FREE(osfs, sizeof(*osfs));
3789 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3790 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3792 for (i = 0; i < *num && err == 0; i++) {
3793 int cleanup_phase = 0;
3795 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3796 CWARN("%s: create aborted by destroy\n",
3805 last_id = filter_last_id(filter, group);
3806 if (next_id > last_id) {
3807 CERROR("Error: Trying to recreate obj greater"
3808 "than last id "LPD64" > "LPD64"\n",
3810 GOTO(cleanup, rc = -EINVAL);
3813 next_id = filter_last_id(filter, group) + 1;
3815 /* Don't create objects beyond the valid range for this SEQ */
3816 if (unlikely(fid_seq_is_mdt0(group) &&
3817 next_id >= IDIF_MAX_OID)) {
3818 CERROR("%s:"POSTID" hit the IDIF_MAX_OID (1<<48)!\n",
3819 obd->obd_name, next_id, group);
3820 GOTO(cleanup, rc = -ENOSPC);
3821 } else if (unlikely(!fid_seq_is_mdt0(group) &&
3822 next_id >= OBIF_MAX_OID)) {
3823 CERROR("%s:"POSTID" hit the OBIF_MAX_OID (1<<32)!\n",
3824 obd->obd_name, next_id, group);
3825 GOTO(cleanup, rc = -ENOSPC);
3828 dparent = filter_parent_lock(obd, group, next_id);
3829 if (IS_ERR(dparent))
3830 GOTO(cleanup, rc = PTR_ERR(dparent));
3831 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3833 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3835 GOTO(cleanup, rc = PTR_ERR(dchild));
3836 cleanup_phase = 2; /* f_dput(dchild) */
3838 if (dchild->d_inode != NULL) {
3839 /* This would only happen if lastobjid was bad on disk*/
3840 /* Could also happen if recreating missing obj but it
3841 * already exists. */
3843 CERROR("%s: recreating existing object %.*s?\n",
3844 obd->obd_name, dchild->d_name.len,
3845 dchild->d_name.name);
3847 /* Use these existing objects if they are
3849 if (dchild->d_inode->i_size == 0) {
3850 rc = filter_use_existing_obj(obd,dchild,
3851 &handle, &cleanup_phase);
3858 CERROR("%s: Serious error: objid %.*s already "
3859 "exists; is this filesystem corrupt?\n",
3860 obd->obd_name, dchild->d_name.len,
3861 dchild->d_name.name);
3864 GOTO(cleanup, rc = -EEXIST);
3867 handle = fsfilt_start_log(obd, dparent->d_inode,
3868 FSFILT_OP_CREATE, NULL, 1);
3870 GOTO(cleanup, rc = PTR_ERR(handle));
3873 CDEBUG(D_INODE, "%s: filter_precreate(od->o_seq="LPU64
3874 ",od->o_id="LPU64")\n", obd->obd_name, group,
3877 /* We mark object SUID+SGID to flag it for accepting UID+GID
3878 * from client on first write. Currently the permission bits
3879 * on the OST are never used, so this is OK. */
3880 rc = ll_vfs_create(dparent->d_inode, dchild,
3881 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3883 CERROR("create failed rc = %d\n", rc);
3884 if (rc == -ENOSPC) {
3885 os_ffree = filter_calc_free_inodes(obd);
3887 CERROR("%s: free inode "LPU64"\n",
3888 obd->obd_name, os_ffree);
3893 if (dchild->d_inode)
3894 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3895 dchild->d_inode->i_ino);
3898 if (!recreate_obj) {
3899 filter_set_last_id(filter, next_id, group);
3900 err = filter_update_last_objid(obd, group, 0);
3902 CERROR("unable to write lastobjid "
3903 "but file created\n");
3907 switch(cleanup_phase) {
3909 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3911 CERROR("error on commit, err = %d\n", err);
3918 filter_parent_unlock(dparent);
3925 if (cfs_time_after(jiffies, enough_time)) {
3928 "%s: precreate slow - want %d got %d \n",
3929 obd->obd_name, *num, i);
3936 "%s: created %d objects for group "POSTID" rc %d\n",
3937 obd->obd_name, i, filter->fo_last_objids[group], group, rc);
3942 int filter_create(struct obd_export *exp, struct obdo *oa,
3943 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3945 struct obd_device *obd = exp->exp_obd;
3946 struct filter_export_data *fed;
3947 struct filter_obd *filter;
3948 struct lvfs_run_ctxt saved;
3949 struct lov_stripe_md *lsm = NULL;
3953 CDEBUG(D_INODE, "%s: filter_create(group="LPU64",id="
3954 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
3956 fed = &exp->exp_filter_data;
3957 filter = &obd->u.filter;
3959 if (fed->fed_group != oa->o_seq) {
3960 CERROR("%s: this export (nid %s) used object group %d "
3961 "earlier; now it's trying to use group "LPU64"!"
3962 " This could be a bug in the MDS. Please report to "
3963 "http://bugzilla.lustre.org/\n", obd->obd_name,
3964 obd_export_nid2str(exp), fed->fed_group, oa->o_seq);
3971 rc = obd_alloc_memmd(exp, &lsm);
3978 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3980 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3981 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3982 if (!obd->obd_recovering ||
3983 oa->o_id > filter_last_id(filter, oa->o_seq)) {
3984 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3985 oa->o_id, filter_last_id(filter, oa->o_seq));
3989 cfs_down(&filter->fo_create_locks[oa->o_seq]);
3990 rc = filter_precreate(obd, oa, oa->o_seq, &diff);
3991 cfs_up(&filter->fo_create_locks[oa->o_seq]);
3994 rc = filter_handle_precreate(exp, oa, oa->o_seq, oti);
3997 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3998 if (rc && ea != NULL && *ea != lsm) {
3999 obd_free_memmd(exp, &lsm);
4000 } else if (rc == 0 && ea != NULL) {
4001 /* XXX LOV STACKING: the lsm that is passed to us from
4002 * LOV does not have valid lsm_oinfo data structs, so
4003 * don't go touching that. This needs to be fixed in a
4005 lsm->lsm_object_id = oa->o_id;
4012 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4013 struct lov_stripe_md *md, struct obd_trans_info *oti,
4014 struct obd_export *md_exp, void *capa)
4016 unsigned int qcids[MAXQUOTAS] = {0, 0};
4017 struct obd_device *obd;
4018 struct filter_obd *filter;
4019 struct dentry *dchild = NULL, *dparent = NULL;
4020 struct lustre_handle lockh = { 0 };
4021 struct lvfs_run_ctxt saved;
4022 void *handle = NULL;
4023 struct llog_cookie *fcc = NULL;
4024 int rc, rc2, cleanup_phase = 0, sync = 0;
4029 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4030 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4035 filter = &obd->u.filter;
4037 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4040 CDEBUG(D_INODE, "%s: filter_destroy(group="LPU64",oid="
4041 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4043 dchild = filter_fid2dentry(obd, NULL, oa->o_seq, oa->o_id);
4045 GOTO(cleanup, rc = PTR_ERR(dchild));
4048 if (dchild->d_inode == NULL) {
4049 CDEBUG(D_INODE, "destroying non-existent object "POSTID"\n",
4050 oa->o_id, oa->o_seq);
4051 /* If object already gone, cancel cookie right now */
4052 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4053 struct llog_ctxt *ctxt;
4054 struct obd_llog_group *olg;
4056 olg = filter_find_olg(obd, oa->o_seq);
4058 CERROR(" %s: can not find olg of group %d\n",
4059 obd->obd_name, (int)oa->o_seq);
4060 GOTO(cleanup, rc = PTR_ERR(olg));
4062 fcc = &oa->o_lcookie;
4063 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4064 llog_cancel(ctxt, NULL, 1, fcc, 0);
4065 llog_ctxt_put(ctxt);
4066 fcc = NULL; /* we didn't allocate fcc, don't free it */
4068 GOTO(cleanup, rc = -ENOENT);
4071 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_seq, &lockh);
4075 /* Our MDC connection is established by the MDS to us */
4076 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4077 OBD_ALLOC(fcc, sizeof(*fcc));
4079 *fcc = oa->o_lcookie;
4081 ll_vfs_dq_init(dchild->d_inode);
4083 /* we're gonna truncate it first in order to avoid possible deadlock:
4085 * open trasaction open transaction
4086 * down(i_zombie) down(i_zombie)
4087 * restart transaction
4088 * (see BUG 4180) -bzzz
4090 * take i_alloc_sem too to prevent other threads from writing to the
4091 * file while we are truncating it. This can cause lock ordering issue
4092 * between page lock, i_mutex & starting new journal handle.
4093 * (see bug 20321) -johann
4096 down_write(&dchild->d_inode->i_alloc_sem);
4097 LOCK_INODE_MUTEX(dchild->d_inode);
4098 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
4100 /* VBR: version recovery check */
4101 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4103 UNLOCK_INODE_MUTEX(dchild->d_inode);
4104 up_write(&dchild->d_inode->i_alloc_sem);
4108 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4110 if (IS_ERR(handle)) {
4111 UNLOCK_INODE_MUTEX(dchild->d_inode);
4112 up_write(&dchild->d_inode->i_alloc_sem);
4113 GOTO(cleanup, rc = PTR_ERR(handle));
4116 iattr.ia_valid = ATTR_SIZE;
4118 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4119 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4120 UNLOCK_INODE_MUTEX(dchild->d_inode);
4121 up_write(&dchild->d_inode->i_alloc_sem);
4125 GOTO(cleanup, rc = rc2);
4127 /* We don't actually need to lock the parent until we are unlinking
4128 * here, and not while truncating above. That avoids holding the
4129 * parent lock for a long time during truncate, which can block other
4130 * threads from doing anything to objects in that directory. bug 7171 */
4131 dparent = filter_parent_lock(obd, oa->o_seq, oa->o_id);
4132 if (IS_ERR(dparent))
4133 GOTO(cleanup, rc = PTR_ERR(dparent));
4134 cleanup_phase = 3; /* filter_parent_unlock */
4136 LOCK_INODE_MUTEX(dchild->d_inode);
4137 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4138 if (IS_ERR(handle)) {
4139 UNLOCK_INODE_MUTEX(dchild->d_inode);
4140 GOTO(cleanup, rc = PTR_ERR(handle));
4142 cleanup_phase = 4; /* fsfilt_commit */
4144 /* Quota release need uid/gid of inode */
4145 obdo_from_inode(oa, dchild->d_inode, NULL, OBD_MD_FLUID|OBD_MD_FLGID);
4147 filter_fmd_drop(exp, oa->o_id, oa->o_seq);
4149 /* this drops dchild->d_inode->i_mutex unconditionally */
4150 rc = filter_destroy_internal(obd, oa->o_id, oa->o_seq, dparent, dchild);
4154 switch(cleanup_phase) {
4157 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4158 oti->oti_handle : handle,
4159 filter_cancel_cookies_cb,
4161 /* If add_journal_cb failed, then filter_finish_transno
4162 * will commit the handle and we will do a sync
4163 * on commit. then we call callback directly to free
4166 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4168 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4171 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4173 CERROR("error on commit, err = %d\n", rc2);
4180 filter_parent_unlock(dparent);
4182 filter_fini_destroy(obd, &lockh);
4186 OBD_FREE(fcc, sizeof(*fcc));
4188 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4191 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4195 /* trigger quota release */
4196 qcids[USRQUOTA] = oa->o_uid;
4197 qcids[GRPQUOTA] = oa->o_gid;
4198 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4201 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4205 /* NB start and end are used for punch, but not truncate */
4206 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4207 struct obd_trans_info *oti,
4208 struct ptlrpc_request_set *rqset)
4213 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4214 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4215 oinfo->oi_policy.l_extent.end);
4219 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4220 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,oinfo->oi_oa->o_valid,
4221 oinfo->oi_policy.l_extent.start);
4223 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4224 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4225 rc = filter_setattr(exp, oinfo, oti);
4226 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4230 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4231 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4234 struct lvfs_run_ctxt saved;
4235 struct obd_device_target *obt;
4236 struct dentry *dentry;
4240 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4241 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4245 obt = &exp->exp_obd->u.obt;
4247 /* An objid of zero is taken to mean "sync whole filesystem" */
4248 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4249 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4250 /* Flush any remaining cancel messages out to the target */
4251 filter_sync_llogs(exp->exp_obd, exp);
4255 dentry = filter_oa2dentry(exp->exp_obd, &oa->o_oi);
4257 RETURN(PTR_ERR(dentry));
4259 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4261 LOCK_INODE_MUTEX(dentry->d_inode);
4263 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4265 /* just any file to grab fsync method - "file" arg unused */
4266 struct file *file = obt->obt_rcvd_filp;
4268 if (file->f_op && file->f_op->fsync)
4269 rc = file->f_op->fsync(NULL, dentry, 1);
4271 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4275 UNLOCK_INODE_MUTEX(dentry->d_inode);
4277 oa->o_valid = OBD_MD_FLID;
4278 obdo_from_inode(oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
4280 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4286 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4287 void *key, __u32 *vallen, void *val,
4288 struct lov_stripe_md *lsm)
4290 struct obd_device *obd;
4293 obd = class_exp2obd(exp);
4295 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4299 if (KEY_IS(KEY_BLOCKSIZE)) {
4300 __u32 *blocksize = val;
4302 if (*vallen < sizeof(*blocksize))
4304 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4306 *vallen = sizeof(*blocksize);
4310 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4311 __u32 *blocksize_bits = val;
4312 if (blocksize_bits) {
4313 if (*vallen < sizeof(*blocksize_bits))
4315 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4317 *vallen = sizeof(*blocksize_bits);
4321 if (KEY_IS(KEY_LAST_ID)) {
4322 obd_id *last_id = val;
4323 /* FIXME: object groups */
4325 if (*vallen < sizeof(*last_id))
4327 *last_id = filter_last_id(&obd->u.filter,
4328 exp->exp_filter_data.fed_group);
4330 *vallen = sizeof(*last_id);
4334 if (KEY_IS(KEY_FIEMAP)) {
4335 struct ll_fiemap_info_key *fm_key = key;
4336 struct dentry *dentry;
4337 struct ll_user_fiemap *fiemap = val;
4338 struct lvfs_run_ctxt saved;
4341 if (fiemap == NULL) {
4342 *vallen = fiemap_count_to_size(
4343 fm_key->fiemap.fm_extent_count);
4347 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa.o_oi,
4350 RETURN(PTR_ERR(dentry));
4352 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4353 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4354 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4356 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4362 if (KEY_IS(KEY_SYNC_LOCK_CANCEL)) {
4363 *((__u32 *) val) = obd->u.filter.fo_sync_lock_cancel;
4364 *vallen = sizeof(__u32);
4368 CDEBUG(D_IOCTL, "invalid key\n");
4372 static inline int filter_setup_llog_group(struct obd_export *exp,
4373 struct obd_device *obd,
4376 struct obd_llog_group *olg;
4377 struct llog_ctxt *ctxt;
4380 olg = filter_find_create_olg(obd, group);
4382 RETURN(PTR_ERR(olg));
4384 llog_group_set_export(olg, exp);
4386 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4387 LASSERTF(ctxt != NULL, "ctxt is null\n");
4389 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4390 llog_ctxt_put(ctxt);
4394 static int filter_set_grant_shrink(struct obd_export *exp,
4395 struct ost_body *body)
4397 /* handle shrink grant */
4398 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4399 filter_grant_incoming(exp, &body->oa);
4400 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4406 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4408 struct obd_device *obd;
4414 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4418 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4419 obd_export_nid2str(exp));
4420 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4422 /* setup llog imports */
4424 group = (int)(*(__u32 *)val);
4426 group = 0; /* default value */
4428 LASSERT_SEQ_IS_MDT(group);
4429 rc = filter_setup_llog_group(exp, obd, group);
4433 if (group == FID_SEQ_OST_MDT0) {
4434 /* setup llog group 1 for interop */
4435 filter_setup_llog_group(exp, obd, FID_SEQ_LLOG);
4438 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4443 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4444 void *key, __u32 vallen, void *val,
4445 struct ptlrpc_request_set *set)
4447 struct obd_device *obd;
4452 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4456 if (KEY_IS(KEY_CAPA_KEY)) {
4458 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4460 CERROR("filter update capability key failed: %d\n", rc);
4464 if (KEY_IS(KEY_REVIMP_UPD)) {
4465 filter_revimp_update(exp);
4466 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4470 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4471 filter_adapt_sptlrpc_conf(obd, 0);
4475 if (KEY_IS(KEY_MDS_CONN))
4476 RETURN(filter_set_mds_conn(exp, val));
4478 if (KEY_IS(KEY_GRANT_SHRINK))
4479 RETURN(filter_set_grant_shrink(exp, val));
4484 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4485 int len, void *karg, void *uarg)
4487 struct obd_device *obd = exp->exp_obd;
4488 struct obd_ioctl_data *data = karg;
4492 case OBD_IOC_ABORT_RECOVERY: {
4493 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4494 target_stop_recovery_thread(obd);
4498 case OBD_IOC_SYNC: {
4499 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4500 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4504 case OBD_IOC_SET_READONLY: {
4506 struct super_block *sb = obd->u.obt.obt_sb;
4507 struct inode *inode = sb->s_root->d_inode;
4508 BDEVNAME_DECLARE_STORAGE(tmp);
4509 CERROR("*** setting device %s read-only ***\n",
4510 ll_bdevname(sb, tmp));
4512 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4513 if (!IS_ERR(handle))
4514 rc = fsfilt_commit(obd, inode, handle, 1);
4516 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4517 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4519 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4523 case OBD_IOC_CATLOGLIST: {
4524 rc = llog_catalog_list(obd, 1, data);
4528 case OBD_IOC_LLOG_CANCEL:
4529 case OBD_IOC_LLOG_REMOVE:
4530 case OBD_IOC_LLOG_INFO:
4531 case OBD_IOC_LLOG_PRINT: {
4532 /* FIXME to be finished */
4533 RETURN(-EOPNOTSUPP);
4535 struct llog_ctxt *ctxt = NULL;
4537 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4538 rc = llog_ioctl(ctxt, cmd, data);
4539 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4552 static int filter_health_check(struct obd_device *obd)
4554 #ifdef USE_HEALTH_CHECK_WRITE
4555 struct filter_obd *filter = &obd->u.filter;
4560 * health_check to return 0 on healthy
4561 * and 1 on unhealthy.
4563 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4566 #ifdef USE_HEALTH_CHECK_WRITE
4567 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4568 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4573 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4576 return filter_fid2dentry(data, NULL, gr, id);
4579 static int filter_process_config(struct obd_device *obd, obd_count len,
4582 struct lustre_cfg *lcfg = buf;
4583 struct lprocfs_static_vars lvars;
4586 switch (lcfg->lcfg_command) {
4588 lprocfs_filter_init_vars(&lvars);
4590 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4600 static struct lvfs_callback_ops filter_lvfs_ops = {
4601 l_fid2dentry: filter_lvfs_fid2dentry,
4604 static struct obd_ops filter_obd_ops = {
4605 .o_owner = THIS_MODULE,
4606 .o_get_info = filter_get_info,
4607 .o_set_info_async = filter_set_info_async,
4608 .o_setup = filter_setup,
4609 .o_precleanup = filter_precleanup,
4610 .o_cleanup = filter_cleanup,
4611 .o_connect = filter_connect,
4612 .o_reconnect = filter_reconnect,
4613 .o_disconnect = filter_disconnect,
4614 .o_ping = filter_ping,
4615 .o_init_export = filter_init_export,
4616 .o_destroy_export = filter_destroy_export,
4617 .o_statfs = filter_statfs,
4618 .o_getattr = filter_getattr,
4619 .o_unpackmd = filter_unpackmd,
4620 .o_create = filter_create,
4621 .o_setattr = filter_setattr,
4622 .o_destroy = filter_destroy,
4623 .o_brw = filter_brw,
4624 .o_punch = filter_truncate,
4625 .o_sync = filter_sync,
4626 .o_preprw = filter_preprw,
4627 .o_commitrw = filter_commitrw,
4628 .o_llog_init = filter_llog_init,
4629 .o_llog_connect = filter_llog_connect,
4630 .o_llog_finish = filter_llog_finish,
4631 .o_iocontrol = filter_iocontrol,
4632 .o_health_check = filter_health_check,
4633 .o_process_config = filter_process_config,
4636 quota_interface_t *filter_quota_interface_ref;
4637 extern quota_interface_t filter_quota_interface;
4639 static int __init obdfilter_init(void)
4641 struct lprocfs_static_vars lvars;
4644 /** sanity check for group<->mdsno conversion */
4645 for (i = 0; i < MAX_MDT_COUNT; i++)
4646 LASSERT(objseq_to_mdsno(mdt_to_obd_objseq(i)) == i);
4648 lprocfs_filter_init_vars(&lvars);
4650 cfs_request_module("%s", "lquota");
4651 OBD_ALLOC(obdfilter_created_scratchpad,
4652 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4653 sizeof(*obdfilter_created_scratchpad));
4654 if (obdfilter_created_scratchpad == NULL)
4657 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4658 sizeof(struct filter_mod_data),
4661 GOTO(out, rc = -ENOMEM);
4663 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4664 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4666 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4667 LUSTRE_OST_NAME, NULL);
4671 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4672 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4673 ll_fmd_cachep = NULL;
4675 if (filter_quota_interface_ref)
4676 PORTAL_SYMBOL_PUT(filter_quota_interface);
4678 OBD_FREE(obdfilter_created_scratchpad,
4679 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4680 sizeof(*obdfilter_created_scratchpad));
4686 static void __exit obdfilter_exit(void)
4688 if (filter_quota_interface_ref)
4689 PORTAL_SYMBOL_PUT(filter_quota_interface);
4691 if (ll_fmd_cachep) {
4692 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4693 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4694 ll_fmd_cachep = NULL;
4697 class_unregister_type(LUSTRE_OST_NAME);
4698 OBD_FREE(obdfilter_created_scratchpad,
4699 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4700 sizeof(*obdfilter_created_scratchpad));
4703 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4704 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4705 MODULE_LICENSE("GPL");
4707 module_init(obdfilter_init);
4708 module_exit(obdfilter_exit);