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);
293 lprocfs_exp_cleanup(exp);
297 /* Add client data to the FILTER. We use a bitmap to locate a free space
298 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
299 * Otherwise, we have just read the data from the last_rcvd file and
300 * we know its offset. */
301 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
304 struct obd_device_target *obt = &obd->u.obt;
305 struct tg_export_data *ted = &exp->exp_target_data;
306 struct lr_server_data *lsd = class_server_data(obd);
307 unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
308 int new_client = (cl_idx == -1);
312 LASSERT(bitmap != NULL);
313 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
316 if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
319 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
320 * there's no need for extra complication here
323 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
325 if (cl_idx >= LR_MAX_CLIENTS) {
326 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
330 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
331 cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
336 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
337 CERROR("FILTER client %d: bit already set in bitmap!\n",
343 ted->ted_lr_idx = cl_idx;
344 ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
345 cl_idx * le16_to_cpu(lsd->lsd_client_size);
346 cfs_init_mutex(&ted->ted_lcd_lock);
347 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
349 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
350 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
353 struct lvfs_run_ctxt saved;
354 loff_t off = ted->ted_lr_off;
358 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
359 ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
361 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
364 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
365 /* Transaction needed to fix bug 1403 */
366 handle = fsfilt_start(obd,
367 obt->obt_rcvd_filp->f_dentry->d_inode,
368 FSFILT_OP_SETATTR, NULL);
369 if (IS_ERR(handle)) {
370 rc = PTR_ERR(handle);
371 CERROR("unable to start transaction: rc %d\n", rc);
373 ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
374 exp->exp_last_request_time = cfs_time_current_sec();
375 rc = fsfilt_add_journal_cb(obd, 0, handle,
376 target_client_add_cb,
377 class_export_cb_get(exp));
379 cfs_spin_lock(&exp->exp_lock);
380 exp->exp_need_sync = 1;
381 cfs_spin_unlock(&exp->exp_lock);
383 rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
385 sizeof(*ted->ted_lcd),
386 &off, rc /* sync if no cb */);
388 obt->obt_rcvd_filp->f_dentry->d_inode,
391 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
394 CERROR("error writing %s client idx %u: rc %d\n",
395 LAST_RCVD, ted->ted_lr_idx, rc);
402 static int filter_client_del(struct obd_export *exp)
404 struct tg_export_data *ted = &exp->exp_target_data;
405 struct obd_device_target *obt = &exp->exp_obd->u.obt;
406 struct lvfs_run_ctxt saved;
411 if (ted->ted_lcd == NULL)
414 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
415 if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
418 LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
420 off = ted->ted_lr_off;
422 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
423 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
425 /* Don't clear ted_lr_idx here as it is likely also unset. At worst
426 * we leak a client slot that will be cleaned on the next recovery. */
428 CERROR("%s: client idx %d has med_off %lld\n",
429 exp->exp_obd->obd_name, ted->ted_lr_idx, off);
430 GOTO(free, rc = -EINVAL);
433 /* Clear the bit _after_ zeroing out the client so we don't
434 race with filter_client_add and zero out new clients.*/
435 if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
436 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
441 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
442 /* Make sure the server's last_transno is up to date.
443 * This should be done before zeroing client slot so last_transno will
444 * be in server data or in client data in case of failure */
445 filter_update_server_data(exp->exp_obd);
447 cfs_mutex_down(&ted->ted_lcd_lock);
448 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
449 rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
451 sizeof(*ted->ted_lcd), &off, 0);
452 cfs_mutex_up(&ted->ted_lcd_lock);
453 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
455 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
456 "zero out client %s at idx %u/%llu in %s, rc %d\n",
457 ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
464 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
465 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
466 struct filter_mod_data *fmd)
468 LASSERT_SPIN_LOCKED(&fed->fed_lock);
469 if (--fmd->fmd_refcount == 0) {
470 /* XXX when we have persistent reservations and the handle
471 * is stored herein we need to drop it here. */
472 fed->fed_mod_count--;
473 cfs_list_del(&fmd->fmd_list);
474 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
478 /* drop fmd reference, free it if last ref */
479 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
481 struct filter_export_data *fed;
486 fed = &exp->exp_filter_data;
487 cfs_spin_lock(&fed->fed_lock);
488 filter_fmd_put_nolock(fed, fmd); /* caller reference */
489 cfs_spin_unlock(&fed->fed_lock);
492 /* expire entries from the end of the list if there are too many
493 * or they are too old */
494 static void filter_fmd_expire_nolock(struct filter_obd *filter,
495 struct filter_export_data *fed,
496 struct filter_mod_data *keep)
498 struct filter_mod_data *fmd, *tmp;
500 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
504 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
505 fed->fed_mod_count < filter->fo_fmd_max_num)
508 cfs_list_del_init(&fmd->fmd_list);
509 filter_fmd_put_nolock(fed, fmd); /* list reference */
513 void filter_fmd_expire(struct obd_export *exp)
515 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
516 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
517 &exp->exp_filter_data, NULL);
518 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
521 /* find specified objid, group in export fmd list.
522 * caller must hold fed_lock and take fmd reference itself */
523 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
524 struct filter_export_data *fed,
525 obd_id objid, obd_seq group)
527 struct filter_mod_data *found = NULL, *fmd;
529 LASSERT_SPIN_LOCKED(&fed->fed_lock);
531 cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
532 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
534 cfs_list_del(&fmd->fmd_list);
535 cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
536 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
541 filter_fmd_expire_nolock(filter, fed, found);
546 /* Find fmd based on objid and group, or return NULL if not found. */
547 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
548 obd_id objid, obd_seq group)
550 struct filter_mod_data *fmd;
552 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
553 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
554 &exp->exp_filter_data, objid, group);
556 fmd->fmd_refcount++; /* caller reference */
557 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
562 /* Find fmd based on objid and group, or create a new one if none is found.
563 * It is possible for this function to return NULL under memory pressure,
564 * or if objid = 0 is passed (which will only cause old entries to expire).
565 * Currently this is not fatal because any fmd state is transient and
566 * may also be freed when it gets sufficiently old. */
567 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
568 obd_id objid, obd_seq group)
570 struct filter_export_data *fed = &exp->exp_filter_data;
571 struct filter_mod_data *found = NULL, *fmd_new = NULL;
573 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
575 cfs_spin_lock(&fed->fed_lock);
576 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
579 cfs_list_add_tail(&fmd_new->fmd_list,
581 fmd_new->fmd_id = objid;
582 fmd_new->fmd_gr = group;
583 fmd_new->fmd_refcount++; /* list reference */
585 fed->fed_mod_count++;
587 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
591 found->fmd_refcount++; /* caller reference */
592 found->fmd_expire = jiffies +
593 exp->exp_obd->u.filter.fo_fmd_max_age;
596 cfs_spin_unlock(&fed->fed_lock);
602 /* drop fmd list reference so it will disappear when last reference is put.
603 * This isn't so critical because it would in fact only affect the one client
604 * that is doing the unlink and at worst we have an stale entry referencing
605 * an object that should never be used again. */
606 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
608 struct filter_mod_data *found = NULL;
610 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
611 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
613 cfs_list_del_init(&found->fmd_list);
614 filter_fmd_put_nolock(&exp->exp_filter_data, found);
616 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
619 #define filter_fmd_drop(exp, objid, group)
622 /* remove all entries from fmd list */
623 static void filter_fmd_cleanup(struct obd_export *exp)
625 struct filter_export_data *fed = &exp->exp_filter_data;
626 struct filter_mod_data *fmd = NULL, *tmp;
628 cfs_spin_lock(&fed->fed_lock);
629 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
630 cfs_list_del_init(&fmd->fmd_list);
631 filter_fmd_put_nolock(fed, fmd);
633 cfs_spin_unlock(&fed->fed_lock);
636 static int filter_init_export(struct obd_export *exp)
639 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
640 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
642 cfs_spin_lock(&exp->exp_lock);
643 exp->exp_connecting = 1;
644 cfs_spin_unlock(&exp->exp_lock);
645 rc = lut_client_alloc(exp);
647 rc = ldlm_init_export(exp);
652 static int filter_free_server_data(struct obd_device_target *obt)
654 lut_fini(NULL, obt->obt_lut);
655 OBD_FREE_PTR(obt->obt_lut);
659 /* assumes caller is already in kernel ctxt */
660 int filter_update_server_data(struct obd_device *obd)
662 struct file *filp = obd->u.obt.obt_rcvd_filp;
663 struct lr_server_data *lsd = class_server_data(obd);
668 CDEBUG(D_INODE, "server uuid : %s\n", lsd->lsd_uuid);
669 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
670 le64_to_cpu(lsd->lsd_last_transno));
671 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
672 le64_to_cpu(lsd->lsd_mount_count));
674 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
676 CERROR("error writing lr_server_data: rc = %d\n", rc);
681 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
684 struct filter_obd *filter = &obd->u.filter;
690 if (filter->fo_last_objid_files[group] == NULL) {
691 CERROR("Object seq "LPU64" not fully setup; not updating "
692 "last_objid\n", group);
696 CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
697 obd->obd_name, filter->fo_last_objids[group], group);
699 tmp = cpu_to_le64(filter->fo_last_objids[group]);
700 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
701 &tmp, sizeof(tmp), &off, force_sync);
703 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
707 extern int ost_handle(struct ptlrpc_request *req);
708 /* assumes caller has already in kernel ctxt */
709 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
711 struct filter_obd *filter = &obd->u.filter;
712 struct lr_server_data *lsd;
713 struct lsd_client_data *lcd = NULL;
714 struct inode *inode = filp->f_dentry->d_inode;
715 unsigned long last_rcvd_size = i_size_read(inode);
716 struct lu_target *lut;
723 /* ensure padding in the struct is the correct size */
724 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
725 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
726 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
727 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
729 /* allocate and initialize lu_target */
733 rc = lut_init(NULL, lut, obd, NULL);
736 lsd = class_server_data(obd);
737 if (last_rcvd_size == 0) {
738 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
740 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
741 lsd->lsd_last_transno = 0;
742 mount_count = lsd->lsd_mount_count = 0;
743 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
744 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
745 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
746 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
747 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
748 /* OBD_COMPAT_OST is set in filter_connect_internal when the
749 * MDS first connects and assigns the OST index number. */
750 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
753 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
755 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
759 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
760 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
761 "using the wrong disk %s. Were the "
762 "/dev/ assignments rearranged?\n",
763 obd->obd_uuid.uuid, lsd->lsd_uuid);
764 GOTO(err_lut, rc = -EINVAL);
766 mount_count = le64_to_cpu(lsd->lsd_mount_count);
767 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
769 /* Assume old last_rcvd format unless I_C_LR is set */
770 if (!(lsd->lsd_feature_incompat &
771 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
772 lsd->lsd_last_transno = lsd->lsd_compat14;
774 /* OBD_COMPAT_OST is set in filter_connect_internal when the
775 * MDS first connects and assigns the OST index number. */
776 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
780 if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
781 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
782 obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
783 ~FILTER_INCOMPAT_SUPP);
784 GOTO(err_lut, rc = -EINVAL);
786 if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
787 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
788 obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
789 ~FILTER_ROCOMPAT_SUPP);
790 /* Do something like remount filesystem read-only */
791 GOTO(err_lut, rc = -EINVAL);
794 start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
796 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
797 obd->obd_name, start_epoch);
798 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
799 obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
800 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
801 obd->obd_name, mount_count + 1);
802 CDEBUG(D_INODE, "%s: server data size: %u\n",
803 obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
804 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
805 obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
806 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
807 obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
808 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
809 obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
810 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
811 last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
812 (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
813 le16_to_cpu(lsd->lsd_client_size));
815 if (!obd->obd_replayable) {
816 CWARN("%s: recovery support OFF\n", obd->obd_name);
822 GOTO(err_client, rc = -ENOMEM);
824 for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
825 off < last_rcvd_size; cl_idx++) {
827 struct obd_export *exp;
828 struct filter_export_data *fed;
830 /* Don't assume off is incremented properly by
831 * fsfilt_read_record(), in case sizeof(*lcd)
832 * isn't the same as lsd->lsd_client_size. */
833 off = le32_to_cpu(lsd->lsd_client_start) +
834 cl_idx * le16_to_cpu(lsd->lsd_client_size);
835 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
837 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
838 LAST_RCVD, cl_idx, off, rc);
839 break; /* read error shouldn't cause startup to fail */
842 if (lcd->lcd_uuid[0] == '\0') {
843 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
848 check_lcd(obd->obd_name, cl_idx, lcd);
850 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
852 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
853 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
854 last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
856 /* These exports are cleaned up by filter_disconnect(), so they
857 * need to be set up like real exports as filter_connect() does.
859 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
861 if (PTR_ERR(exp) == -EALREADY) {
862 /* export already exists, zero out this one */
863 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
867 GOTO(err_client, rc = PTR_ERR(exp));
870 fed = &exp->exp_filter_data;
871 *fed->fed_ted.ted_lcd = *lcd;
872 fed->fed_group = 0; /* will be assigned at connect */
873 filter_export_stats_init(obd, exp, 0, NULL);
874 rc = filter_client_add(obd, exp, cl_idx);
875 /* can't fail for existing client */
876 LASSERTF(rc == 0, "rc = %d\n", rc);
878 /* VBR: set export last committed */
879 exp->exp_last_committed = last_rcvd;
880 cfs_spin_lock(&exp->exp_lock);
881 exp->exp_connecting = 0;
882 exp->exp_in_recovery = 0;
883 cfs_spin_unlock(&exp->exp_lock);
884 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
885 obd->obd_max_recoverable_clients++;
886 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
887 class_export_put(exp);
889 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
890 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
894 obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
896 obd->u.obt.obt_mount_count = mount_count + 1;
897 lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
899 /* save it, so mount count and last_transno is current */
900 rc = filter_update_server_data(obd);
902 GOTO(err_client, rc);
907 class_disconnect_exports(obd);
909 filter_free_server_data(&obd->u.obt);
913 static int filter_cleanup_groups(struct obd_device *obd)
915 struct filter_obd *filter = &obd->u.filter;
917 struct dentry *dentry;
921 if (filter->fo_dentry_O_groups != NULL) {
922 for (i = 0; i < filter->fo_group_count; i++) {
923 dentry = filter->fo_dentry_O_groups[i];
927 OBD_FREE(filter->fo_dentry_O_groups,
928 filter->fo_group_count *
929 sizeof(*filter->fo_dentry_O_groups));
930 filter->fo_dentry_O_groups = NULL;
932 if (filter->fo_last_objid_files != NULL) {
933 for (i = 0; i < filter->fo_group_count; i++) {
934 filp = filter->fo_last_objid_files[i];
938 OBD_FREE(filter->fo_last_objid_files,
939 filter->fo_group_count *
940 sizeof(*filter->fo_last_objid_files));
941 filter->fo_last_objid_files = NULL;
943 if (filter->fo_dentry_O_sub != NULL) {
944 for (i = 0; i < filter->fo_group_count; i++) {
945 for (j = 0; j < filter->fo_subdir_count; j++) {
946 dentry = filter->fo_dentry_O_sub[i].dentry[j];
951 OBD_FREE(filter->fo_dentry_O_sub,
952 filter->fo_group_count *
953 sizeof(*filter->fo_dentry_O_sub));
954 filter->fo_dentry_O_sub = NULL;
956 if (filter->fo_last_objids != NULL) {
957 OBD_FREE(filter->fo_last_objids,
958 filter->fo_group_count *
959 sizeof(*filter->fo_last_objids));
960 filter->fo_last_objids = NULL;
962 if (filter->fo_dentry_O != NULL) {
963 f_dput(filter->fo_dentry_O);
964 filter->fo_dentry_O = NULL;
969 static int filter_update_last_group(struct obd_device *obd, int group)
971 struct filter_obd *filter = &obd->u.filter;
972 struct file *filp = NULL;
973 int last_group = 0, rc;
977 if (group <= filter->fo_committed_group)
980 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
984 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
988 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
990 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
994 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
995 obd->obd_name, last_group, group);
999 /* must be sync: bXXXX */
1000 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1002 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1006 filter->fo_committed_group = group;
1009 filp_close(filp, 0);
1013 static int filter_read_group_internal(struct obd_device *obd, int group,
1016 struct filter_obd *filter = &obd->u.filter;
1017 __u64 *new_objids = NULL;
1018 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1019 struct dentry **new_groups = NULL;
1020 struct file **new_files = NULL;
1021 struct dentry *dentry;
1023 int old_count = filter->fo_group_count, rc, stage = 0, i;
1027 int len = group + 1;
1029 snprintf(name, 24, "%d", group);
1033 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1035 if (IS_ERR(dentry)) {
1036 CERROR("Cannot lookup expected object group %d: %ld\n",
1037 group, PTR_ERR(dentry));
1038 RETURN(PTR_ERR(dentry));
1041 dentry = simple_mkdir(filter->fo_dentry_O,
1042 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1043 if (IS_ERR(dentry)) {
1044 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1046 RETURN(PTR_ERR(dentry));
1051 snprintf(name, 24, "O/%d/LAST_ID", group);
1053 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1055 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1056 GOTO(cleanup, rc = PTR_ERR(filp));
1060 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1062 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1066 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1067 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1068 if (tmp_subdirs == NULL)
1069 GOTO(cleanup, rc = -ENOMEM);
1072 for (i = 0; i < filter->fo_subdir_count; i++) {
1074 snprintf(dir, sizeof(dir), "d%u", i);
1076 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1077 obd->u.obt.obt_vfsmnt,
1079 if (IS_ERR(tmp_subdirs->dentry[i])) {
1080 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1081 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1086 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1087 tmp_subdirs->dentry[i]);
1091 /* 'group' is an index; we need an array of length 'group + 1' */
1092 if (group + 1 > old_count) {
1093 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1094 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1095 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1096 OBD_ALLOC(new_files, len * sizeof(*new_files));
1098 if (new_objids == NULL || new_subdirs == NULL ||
1099 new_groups == NULL || new_files == NULL)
1100 GOTO(cleanup, rc = -ENOMEM);
1103 memcpy(new_objids, filter->fo_last_objids,
1104 old_count * sizeof(*new_objids));
1105 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1106 old_count * sizeof(*new_subdirs));
1107 memcpy(new_groups, filter->fo_dentry_O_groups,
1108 old_count * sizeof(*new_groups));
1109 memcpy(new_files, filter->fo_last_objid_files,
1110 old_count * sizeof(*new_files));
1112 OBD_FREE(filter->fo_last_objids,
1113 old_count * sizeof(*new_objids));
1114 OBD_FREE(filter->fo_dentry_O_sub,
1115 old_count * sizeof(*new_subdirs));
1116 OBD_FREE(filter->fo_dentry_O_groups,
1117 old_count * sizeof(*new_groups));
1118 OBD_FREE(filter->fo_last_objid_files,
1119 old_count * sizeof(*new_files));
1121 filter->fo_last_objids = new_objids;
1122 filter->fo_dentry_O_sub = new_subdirs;
1123 filter->fo_dentry_O_groups = new_groups;
1124 filter->fo_last_objid_files = new_files;
1125 filter->fo_group_count = len;
1128 filter->fo_dentry_O_groups[group] = dentry;
1129 filter->fo_last_objid_files[group] = filp;
1130 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1131 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1132 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1135 filter_update_last_group(obd, group);
1137 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1138 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1139 rc = filter_update_last_objid(obd, group, 1);
1143 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1144 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1145 obd->obd_name, group, last_objid);
1150 if (new_objids != NULL)
1151 OBD_FREE(new_objids, len * sizeof(*new_objids));
1152 if (new_subdirs != NULL)
1153 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1154 if (new_groups != NULL)
1155 OBD_FREE(new_groups, len * sizeof(*new_groups));
1156 if (new_files != NULL)
1157 OBD_FREE(new_files, len * sizeof(*new_files));
1159 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1160 for (i = 0; i < filter->fo_subdir_count; i++) {
1161 if (tmp_subdirs->dentry[i] != NULL)
1162 dput(tmp_subdirs->dentry[i]);
1164 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1167 filp_close(filp, 0);
1174 static int filter_read_groups(struct obd_device *obd, int last_group,
1177 struct filter_obd *filter = &obd->u.filter;
1178 int old_count, group, rc = 0;
1180 cfs_down(&filter->fo_init_lock);
1181 old_count = filter->fo_group_count;
1182 for (group = old_count; group <= last_group; group++) {
1183 rc = filter_read_group_internal(obd, group, create);
1187 cfs_up(&filter->fo_init_lock);
1191 /* FIXME: object groups */
1192 static int filter_prep_groups(struct obd_device *obd)
1194 struct filter_obd *filter = &obd->u.filter;
1195 struct dentry *O_dentry;
1197 int last_group, rc = 0, cleanup_phase = 0;
1201 O_dentry = simple_mkdir(current->fs->pwd, obd->u.obt.obt_vfsmnt,
1203 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1204 if (IS_ERR(O_dentry)) {
1205 rc = PTR_ERR(O_dentry);
1206 CERROR("cannot open/create O: rc = %d\n", rc);
1209 filter->fo_dentry_O = O_dentry;
1210 cleanup_phase = 1; /* O_dentry */
1212 /* we have to initialize all groups before first connections from
1213 * clients because they may send create/destroy for any group -bzzz */
1214 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1216 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1217 GOTO(cleanup, rc = PTR_ERR(filp));
1219 cleanup_phase = 2; /* filp */
1221 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1223 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1228 last_group = FID_SEQ_OST_MDT0;
1230 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1231 FID_SEQ_OST_MDT0, last_group);
1232 filter->fo_committed_group = last_group;
1233 rc = filter_read_groups(obd, last_group, 1);
1237 filp_close(filp, 0);
1241 switch (cleanup_phase) {
1243 filp_close(filp, 0);
1245 filter_cleanup_groups(obd);
1246 f_dput(filter->fo_dentry_O);
1247 filter->fo_dentry_O = NULL;
1255 /* setup the object store with correct subdirectories */
1256 static int filter_prep(struct obd_device *obd)
1258 struct lvfs_run_ctxt saved;
1259 struct filter_obd *filter = &obd->u.filter;
1261 struct inode *inode;
1265 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1266 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1267 if (!file || IS_ERR(file)) {
1269 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1273 obd->u.obt.obt_rcvd_filp = file;
1274 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1275 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1276 file->f_dentry->d_inode->i_mode);
1277 GOTO(err_filp, rc = -ENOENT);
1280 inode = file->f_dentry->d_parent->d_inode;
1281 /* We use i_op->unlink directly in filter_vfs_unlink() */
1282 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1283 CERROR("%s: filesystem does not support create/unlink ops\n",
1285 GOTO(err_filp, rc = -EOPNOTSUPP);
1288 rc = filter_init_server_data(obd, file);
1290 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1293 LASSERT(obd->u.obt.obt_lut);
1294 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1296 /* open and create health check io file*/
1297 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1300 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1302 GOTO(err_server_data, rc);
1304 filter->fo_obt.obt_health_check_filp = file;
1305 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1306 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1307 file->f_dentry->d_inode->i_mode);
1308 GOTO(err_health_check, rc = -ENOENT);
1310 rc = lvfs_check_io_health(obd, file);
1312 GOTO(err_health_check, rc);
1314 rc = filter_prep_groups(obd);
1316 GOTO(err_health_check, rc);
1318 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1323 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1324 CERROR("can't close %s after error\n", HEALTH_CHECK);
1325 filter->fo_obt.obt_health_check_filp = NULL;
1327 target_recovery_fini(obd);
1328 filter_free_server_data(&obd->u.obt);
1330 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1331 CERROR("can't close %s after error\n", LAST_RCVD);
1332 obd->u.obt.obt_rcvd_filp = NULL;
1336 /* cleanup the filter: write last used object id to status file */
1337 static void filter_post(struct obd_device *obd)
1339 struct lvfs_run_ctxt saved;
1340 struct filter_obd *filter = &obd->u.filter;
1343 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1344 * best to start a transaction with h_sync, because we removed this
1347 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1348 rc = filter_update_server_data(obd);
1350 CERROR("error writing server data: rc = %d\n", rc);
1352 for (i = 0; i < filter->fo_group_count; i++) {
1353 rc = filter_update_last_objid(obd, i,
1354 (i == filter->fo_group_count - 1));
1356 CERROR("error writing group %d lastobjid: rc = %d\n",
1360 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1361 obd->u.obt.obt_rcvd_filp = NULL;
1363 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1365 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1366 filter->fo_obt.obt_health_check_filp = NULL;
1368 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1370 filter_cleanup_groups(obd);
1371 filter_free_server_data(&obd->u.obt);
1372 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1374 filter_free_capa_keys(filter);
1375 cleanup_capa_hash(filter->fo_capa_hash);
1378 static void filter_set_last_id(struct filter_obd *filter,
1379 obd_id id, obd_seq group)
1381 LASSERT(group <= filter->fo_group_count);
1383 cfs_spin_lock(&filter->fo_objidlock);
1384 filter->fo_last_objids[group] = id;
1385 cfs_spin_unlock(&filter->fo_objidlock);
1388 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1391 LASSERT(group <= filter->fo_group_count);
1392 LASSERT(filter->fo_last_objids != NULL);
1394 /* FIXME: object groups */
1395 cfs_spin_lock(&filter->fo_objidlock);
1396 id = filter->fo_last_objids[group];
1397 cfs_spin_unlock(&filter->fo_objidlock);
1401 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1403 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1407 /* We never dget the object parent, so DON'T dput it either */
1408 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1410 struct filter_obd *filter = &obd->u.filter;
1411 struct filter_subdirs *subdirs;
1412 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1414 if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1415 return filter->fo_dentry_O_groups[group];
1417 subdirs = &filter->fo_dentry_O_sub[group];
1418 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1421 /* We never dget the object parent, so DON'T dput it either */
1422 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1425 unsigned long now = jiffies;
1426 struct dentry *dparent = filter_parent(obd, group, objid);
1429 if (IS_ERR(dparent))
1431 if (dparent == NULL)
1432 return ERR_PTR(-ENOENT);
1434 rc = filter_lock_dentry(obd, dparent);
1435 fsfilt_check_slow(obd, now, "parent lock");
1436 return rc ? ERR_PTR(rc) : dparent;
1439 /* We never dget the object parent, so DON'T dput it either */
1440 static void filter_parent_unlock(struct dentry *dparent)
1442 UNLOCK_INODE_MUTEX(dparent->d_inode);
1445 /* How to get files, dentries, inodes from object id's.
1447 * If dir_dentry is passed, the caller has already locked the parent
1448 * appropriately for this operation (normally a write lock). If
1449 * dir_dentry is NULL, we do a read lock while we do the lookup to
1450 * avoid races with create/destroy and such changing the directory
1451 * internal to the filesystem code. */
1452 struct dentry *filter_fid2dentry(struct obd_device *obd,
1453 struct dentry *dir_dentry,
1454 obd_seq group, obd_id id)
1456 struct dentry *dparent = dir_dentry;
1457 struct dentry *dchild;
1462 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1463 obd->u.filter.fo_destroys_in_progress == 0) {
1464 /* don't fail lookups for orphan recovery, it causes
1465 * later LBUGs when objects still exist during precreate */
1466 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1467 RETURN(ERR_PTR(-ENOENT));
1470 CERROR("fatal: invalid object id 0\n");
1471 RETURN(ERR_PTR(-ESTALE));
1474 len = sprintf(name, LPU64, id);
1475 if (dir_dentry == NULL) {
1476 dparent = filter_parent_lock(obd, group, id);
1477 if (IS_ERR(dparent)) {
1478 CERROR("%s: error getting object "POSTID
1479 " parent: rc %ld\n", obd->obd_name,
1480 id, group, PTR_ERR(dparent));
1484 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1485 dparent->d_name.len, dparent->d_name.name, name);
1486 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1487 if (dir_dentry == NULL)
1488 filter_parent_unlock(dparent);
1489 if (IS_ERR(dchild)) {
1490 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1495 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1496 CERROR("%s: got bad object "LPU64" inode %lu\n",
1497 obd->obd_name, id, dchild->d_inode->i_ino);
1499 RETURN(ERR_PTR(-ENOENT));
1502 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1503 name, dchild, atomic_read(&dchild->d_count));
1505 LASSERT(atomic_read(&dchild->d_count) > 0);
1510 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1511 obd_id group, struct lustre_handle *lockh)
1513 int flags = LDLM_AST_DISCARD_DATA, rc;
1514 struct ldlm_res_id res_id;
1515 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1518 osc_build_res_name(objid, group, &res_id);
1519 /* Tell the clients that the object is gone now and that they should
1520 * throw away any cached pages. */
1521 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1522 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1523 ldlm_completion_ast, NULL, NULL, 0, NULL,
1530 static void filter_fini_destroy(struct obd_device *obd,
1531 struct lustre_handle *lockh)
1533 if (lustre_handle_is_used(lockh))
1534 ldlm_lock_decref(lockh, LCK_PW);
1537 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1538 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1539 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1540 * reverse. Caller must take i_sem before starting the transaction and we
1541 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1542 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1543 struct vfsmount *mnt)
1548 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1549 * itself we already hold dir->i_mutex for child create/unlink ops */
1550 LASSERT(dentry->d_inode != NULL);
1551 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1552 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1556 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1557 GOTO(out, rc = -ENOENT);
1559 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1564 GOTO(out, rc = -EPERM);
1566 /* check_sticky() */
1567 if ((dentry->d_inode->i_uid != current->fsuid &&
1568 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1569 IS_IMMUTABLE(dentry->d_inode))
1570 GOTO(out, rc = -EPERM);
1572 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1573 * that was done because of journal_start (which is already done
1574 * here) or some other ordering issue. */
1577 rc = ll_security_inode_unlink(dir, dentry, mnt);
1581 rc = dir->i_op->unlink(dir, dentry);
1583 /* need to drop i_mutex before we lose inode reference */
1584 UNLOCK_INODE_MUTEX(dentry->d_inode);
1591 /* Caller must hold LCK_PW on parent and push us into kernel context.
1592 * Caller must hold child i_mutex, we drop it always.
1593 * Caller is also required to ensure that dchild->d_inode exists. */
1594 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1595 obd_seq group, struct dentry *dparent,
1596 struct dentry *dchild)
1598 struct inode *inode = dchild->d_inode;
1601 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
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 == res->lr_namespace);
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 ns_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 * ->ns_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 */
1803 /* Update the LVB from disk if the AST failed (this is a legal race) */
1805 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1806 * sending ast is not handled. This can result in lost client writes.
1809 ldlm_res_lvbo_update(res, NULL, 1);
1812 *reply_lvb = *res_lvb;
1816 LDLM_LOCK_RELEASE(l);
1818 RETURN(ELDLM_LOCK_ABORTED);
1822 * per-obd_device iobuf pool.
1824 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1825 * allocations in write-path has to be minimized (see bug 5137).
1827 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1828 * OST threads (see ost_thread_{init,done}()).
1830 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1831 * at the OST layer there are only (potentially) multiple obd_device of type
1832 * unknown at the time of OST thread creation.
1834 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1835 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1836 * it's very own iobuf.
1840 * filter_kiobuf_pool_init()
1842 * filter_kiobuf_pool_done()
1844 * filter_iobuf_get()
1846 * operate on this array. They are "generic" in a sense that they don't depend
1847 * on actual type of iobuf's (the latter depending on Linux kernel version).
1851 * destroy pool created by filter_iobuf_pool_init
1853 static void filter_iobuf_pool_done(struct filter_obd *filter)
1855 struct filter_iobuf **pool;
1860 pool = filter->fo_iobuf_pool;
1862 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1863 if (pool[i] != NULL)
1864 filter_free_iobuf(pool[i]);
1866 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1867 filter->fo_iobuf_pool = NULL;
1872 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1874 struct filter_obd *filter = &obd->u.filter;
1875 struct sptlrpc_rule_set tmp_rset;
1878 sptlrpc_rule_set_init(&tmp_rset);
1879 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1881 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1886 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1888 cfs_write_lock(&filter->fo_sptlrpc_lock);
1889 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1890 filter->fo_sptlrpc_rset = tmp_rset;
1891 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1897 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1899 static int filter_iobuf_pool_init(struct filter_obd *filter)
1906 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1908 if (filter->fo_iobuf_pool == NULL)
1911 filter->fo_iobuf_count = OSS_THREADS_MAX;
1916 /* Return iobuf allocated for @thread_id. We don't know in advance how
1917 * many threads there will be so we allocate a large empty array and only
1918 * fill in those slots that are actually in use.
1919 * If we haven't allocated a pool entry for this thread before, do so now. */
1920 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1922 int thread_id = (oti && oti->oti_thread) ?
1923 oti->oti_thread->t_id : -1;
1924 struct filter_iobuf *pool = NULL;
1925 struct filter_iobuf **pool_place = NULL;
1927 if (thread_id >= 0) {
1928 LASSERT(thread_id < filter->fo_iobuf_count);
1929 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1932 if (unlikely(pool == NULL)) {
1933 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1934 PTLRPC_MAX_BRW_PAGES);
1935 if (pool_place != NULL)
1942 /* mount the file system (secretly). lustre_cfg parameters are:
1945 * 3 = flags: failover=f, failout=n
1948 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1951 struct filter_obd *filter = &obd->u.filter;
1952 struct vfsmount *mnt;
1953 struct lustre_mount_info *lmi;
1954 struct obd_uuid uuid;
1962 if (lcfg->lcfg_bufcount < 3 ||
1963 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1964 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1967 lmi = server_get_mount(obd->obd_name);
1969 /* We already mounted in lustre_fill_super.
1970 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1971 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1973 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1975 /* gets recovery timeouts from mount data */
1976 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1977 obd->obd_recovery_timeout =
1978 lsi->lsi_lmd->lmd_recovery_time_soft;
1979 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1980 obd->obd_recovery_time_hard =
1981 lsi->lsi_lmd->lmd_recovery_time_hard;
1983 /* old path - used by lctl */
1984 CERROR("Using old MDS mount method\n");
1985 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1986 MS_NOATIME|MS_NODIRATIME,
1987 lustre_cfg_string(lcfg, 1), option);
1990 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1991 lustre_cfg_string(lcfg, 1), rc);
1995 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1997 if (IS_ERR(obd->obd_fsops))
1998 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2000 rc = filter_iobuf_pool_init(filter);
2004 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2005 CERROR("%s: Underlying device is marked as read-only. "
2006 "Setup failed\n", obd->obd_name);
2007 GOTO(err_ops, rc = -EROFS);
2010 /* failover is the default */
2011 obd->obd_replayable = 1;
2013 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2014 str = lustre_cfg_string(lcfg, 3);
2015 if (strchr(str, 'n')) {
2016 CWARN("%s: recovery disabled\n", obd->obd_name);
2017 obd->obd_replayable = 0;
2021 obd->u.obt.obt_vfsmnt = mnt;
2022 obd->u.obt.obt_sb = mnt->mnt_sb;
2023 obd->u.obt.obt_magic = OBT_MAGIC;
2024 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2025 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2027 fsfilt_setup(obd, obd->u.obt.obt_sb);
2029 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2030 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2031 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2032 obd->obd_lvfs_ctxt.fs = get_ds();
2033 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2035 cfs_init_mutex(&filter->fo_init_lock);
2036 filter->fo_committed_group = 0;
2037 filter->fo_destroys_in_progress = 0;
2038 for (i = 0; i < 32; i++)
2039 cfs_sema_init(&filter->fo_create_locks[i], 1);
2041 cfs_spin_lock_init(&filter->fo_objidlock);
2042 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2043 cfs_sema_init(&filter->fo_alloc_lock, 1);
2044 init_brw_stats(&filter->fo_filter_stats);
2045 filter->fo_read_cache = 1; /* enable read-only cache by default */
2046 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2047 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2048 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2049 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2050 filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2051 filter_slc_set(filter); /* initialize sync on lock cancel */
2053 rc = filter_prep(obd);
2057 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2058 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2060 filter->fo_fl_oss_capa = 1;
2062 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2063 filter->fo_capa_hash = init_capa_hash();
2064 if (filter->fo_capa_hash == NULL)
2065 GOTO(err_post, rc = -ENOMEM);
2067 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2068 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2069 LDLM_NAMESPACE_GREEDY);
2070 if (obd->obd_namespace == NULL)
2071 GOTO(err_post, rc = -ENOMEM);
2072 obd->obd_namespace->ns_lvbp = obd;
2073 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2074 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2076 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2077 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2079 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2081 CERROR("failed to setup llogging subsystems\n");
2085 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2086 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2087 /* do this after llog being initialized */
2088 filter_adapt_sptlrpc_conf(obd, 1);
2090 rc = lquota_setup(filter_quota_interface_ref, obd);
2094 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2095 if (q->max_sectors < q->max_hw_sectors &&
2096 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2097 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2098 "for larger I/O requests: max_sectors = %u "
2099 "could be up to max_hw_sectors=%u\n",
2100 obd->obd_name, mnt->mnt_sb->s_id,
2101 q->max_sectors, q->max_hw_sectors);
2103 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2104 if (uuid_ptr != NULL) {
2105 class_uuid_unparse(uuid_ptr, &uuid);
2111 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2112 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2113 obd->obd_name, label ?: str, lmi ? "on " : "",
2114 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2115 obd->obd_replayable ? "enabled" : "disabled");
2117 if (obd->obd_recovering)
2118 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2119 "or until %d client%s reconnect%s\n",
2121 obd->obd_recovery_timeout / 60,
2122 obd->obd_recovery_timeout % 60,
2123 obd->obd_max_recoverable_clients,
2124 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2125 (obd->obd_max_recoverable_clients == 1) ? "s": "");
2133 fsfilt_put_ops(obd->obd_fsops);
2134 filter_iobuf_pool_done(filter);
2136 server_put_mount(obd->obd_name, mnt);
2137 obd->u.obt.obt_sb = 0;
2141 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2143 struct lprocfs_static_vars lvars;
2148 CLASSERT(offsetof(struct obd_device, u.obt) ==
2149 offsetof(struct obd_device, u.filter.fo_obt));
2151 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2154 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2155 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2158 addr = (unsigned long)cfs_page_address(page);
2159 clear_page((void *)addr);
2161 /* lprocfs must be setup before the filter so state can be safely added
2162 * to /proc incrementally as the filter is setup */
2163 lprocfs_filter_init_vars(&lvars);
2164 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2165 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2166 /* Init obdfilter private stats here */
2167 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2168 LPROCFS_CNTR_AVGMINMAX,
2169 "read_bytes", "bytes");
2170 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2171 LPROCFS_CNTR_AVGMINMAX,
2172 "write_bytes", "bytes");
2173 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2174 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2175 "get_page", "usec");
2176 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2177 LPROCFS_CNTR_AVGMINMAX,
2178 "get_page_failures", "num");
2179 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2180 LPROCFS_CNTR_AVGMINMAX,
2181 "cache_access", "pages");
2182 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2183 LPROCFS_CNTR_AVGMINMAX,
2184 "cache_hit", "pages");
2185 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2186 LPROCFS_CNTR_AVGMINMAX,
2187 "cache_miss", "pages");
2189 lproc_filter_attach_seqstat(obd);
2190 obd->obd_proc_exports_entry = lprocfs_register("exports",
2191 obd->obd_proc_entry,
2193 if (IS_ERR(obd->obd_proc_exports_entry)) {
2194 rc = PTR_ERR(obd->obd_proc_exports_entry);
2195 CERROR("error %d setting up lprocfs for %s\n",
2197 obd->obd_proc_exports_entry = NULL;
2200 if (obd->obd_proc_exports_entry)
2201 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2202 lprocfs_nid_stats_clear_read,
2203 lprocfs_nid_stats_clear_write, obd, NULL);
2205 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2206 LUSTRE_CFG_BUFLEN(lcfg, 4));
2207 rc = filter_common_setup(obd, lcfg, (void *)addr);
2208 OBD_PAGE_FREE(page);
2211 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2212 lprocfs_free_per_client_stats(obd);
2213 lprocfs_free_obd_stats(obd);
2214 lprocfs_obd_cleanup(obd);
2220 static struct llog_operations filter_mds_ost_repl_logops;
2222 static struct llog_operations filter_size_orig_logops = {
2223 .lop_setup = llog_obd_origin_setup,
2224 .lop_cleanup = llog_obd_origin_cleanup,
2225 .lop_add = llog_obd_origin_add
2228 static int filter_olg_fini(struct obd_llog_group *olg)
2230 struct llog_ctxt *ctxt;
2231 int rc = 0, rc2 = 0;
2234 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2236 rc = llog_cleanup(ctxt);
2238 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2240 rc2 = llog_cleanup(ctxt);
2245 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2247 rc2 = llog_cleanup(ctxt);
2256 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2257 struct obd_device *tgt)
2262 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2263 &filter_mds_ost_repl_logops);
2267 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2268 &filter_size_orig_logops);
2274 filter_olg_fini(olg);
2279 * Init the default olg, which is embeded in the obd_device, for filter.
2282 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2283 struct obd_device *tgt)
2285 struct filter_obd *filter = &obd->u.filter;
2286 struct llog_ctxt *ctxt;
2290 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2291 if (!filter->fo_lcm)
2294 filter_mds_ost_repl_logops = llog_client_ops;
2295 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2296 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2297 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2299 rc = filter_olg_init(obd, olg, tgt);
2301 GOTO(cleanup_lcm, rc);
2303 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2306 GOTO(cleanup_olg, rc);
2308 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2310 CERROR("Can't get ctxt for %p:%x\n", olg,
2311 LLOG_MDS_OST_REPL_CTXT);
2312 GOTO(cleanup_olg, rc = -ENODEV);
2314 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2315 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2316 llog_ctxt_put(ctxt);
2320 filter_olg_fini(olg);
2322 llog_recov_thread_fini(filter->fo_lcm, 1);
2323 filter->fo_lcm = NULL;
2328 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2329 struct obd_device *tgt, int *index)
2331 struct filter_obd *filter = &obd->u.filter;
2332 struct llog_ctxt *ctxt;
2336 LASSERT(olg != NULL);
2337 if (olg == &obd->obd_olg)
2338 return filter_default_olg_init(obd, olg, tgt);
2340 LASSERT(filter->fo_lcm != NULL);
2341 rc = filter_olg_init(obd, olg, tgt);
2344 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2346 CERROR("Can't get ctxt for %p:%x\n", olg,
2347 LLOG_MDS_OST_REPL_CTXT);
2348 filter_olg_fini(olg);
2351 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2352 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2353 llog_ctxt_put(ctxt);
2357 static int filter_llog_finish(struct obd_device *obd, int count)
2359 struct filter_obd *filter = &obd->u.filter;
2360 struct llog_ctxt *ctxt;
2363 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2366 * Make sure that no cached llcds left in recov_thread.
2367 * We actually do sync in disconnect time, but disconnect
2368 * may not come being marked rq_no_resend = 1.
2370 llog_sync(ctxt, NULL);
2373 * Balance class_import_get() in llog_receptor_accept().
2374 * This is safe to do, as llog is already synchronized
2375 * and its import may go.
2377 cfs_mutex_down(&ctxt->loc_sem);
2378 if (ctxt->loc_imp) {
2379 class_import_put(ctxt->loc_imp);
2380 ctxt->loc_imp = NULL;
2382 cfs_mutex_up(&ctxt->loc_sem);
2383 llog_ctxt_put(ctxt);
2386 if (filter->fo_lcm) {
2387 cfs_mutex_down(&ctxt->loc_sem);
2388 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2389 filter->fo_lcm = NULL;
2390 cfs_mutex_up(&ctxt->loc_sem);
2392 RETURN(filter_olg_fini(&obd->obd_olg));
2396 * Find the group llog according to group index in the llog group list.
2398 static struct obd_llog_group *
2399 filter_find_olg_internal(struct filter_obd *filter, int group)
2401 struct obd_llog_group *olg;
2403 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2404 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2405 if (olg->olg_seq == group)
2412 * Find the group llog according to group index on the filter
2414 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2416 struct obd_llog_group *olg = NULL;
2417 struct filter_obd *filter;
2419 filter = &obd->u.filter;
2421 if (group == FID_SEQ_LLOG)
2422 RETURN(&obd->obd_olg);
2424 cfs_spin_lock(&filter->fo_llog_list_lock);
2425 olg = filter_find_olg_internal(filter, group);
2426 cfs_spin_unlock(&filter->fo_llog_list_lock);
2431 * Find the llog_group of the filter according to the group. If it can not
2432 * find, create the llog_group, which only happens when mds is being synced
2435 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2437 struct obd_llog_group *olg = NULL;
2438 struct filter_obd *filter;
2441 filter = &obd->u.filter;
2443 if (group == FID_SEQ_LLOG)
2444 RETURN(&obd->obd_olg);
2446 cfs_spin_lock(&filter->fo_llog_list_lock);
2447 olg = filter_find_olg_internal(filter, group);
2449 if (olg->olg_initializing) {
2450 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2452 GOTO(out_unlock, olg);
2457 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2459 llog_group_init(olg, group);
2460 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2461 olg->olg_initializing = 1;
2462 cfs_spin_unlock(&filter->fo_llog_list_lock);
2464 rc = obd_llog_init(obd, olg, obd, NULL);
2466 cfs_spin_lock(&filter->fo_llog_list_lock);
2467 cfs_list_del(&olg->olg_list);
2468 cfs_spin_unlock(&filter->fo_llog_list_lock);
2470 GOTO(out, olg = ERR_PTR(-ENOMEM));
2472 cfs_spin_lock(&filter->fo_llog_list_lock);
2473 olg->olg_initializing = 0;
2474 cfs_spin_unlock(&filter->fo_llog_list_lock);
2475 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2476 obd->obd_name, group, olg);
2481 cfs_spin_unlock(&filter->fo_llog_list_lock);
2485 static int filter_llog_connect(struct obd_export *exp,
2486 struct llogd_conn_body *body)
2488 struct obd_device *obd = exp->exp_obd;
2489 struct llog_ctxt *ctxt;
2490 struct obd_llog_group *olg;
2494 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2495 obd->obd_name, body->lgdc_logid.lgl_oid,
2496 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2498 olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2500 CERROR(" %s: can not find olg of group %d\n",
2501 obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2504 llog_group_set_export(olg, exp);
2506 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2507 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2508 body->lgdc_ctxt_idx);
2510 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2511 obd->obd_name, body->lgdc_logid.lgl_oid,
2512 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2514 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2515 obd->u.filter.fo_mds_ost_sync = 1;
2516 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2517 rc = llog_connect(ctxt, &body->lgdc_logid,
2518 &body->lgdc_gen, NULL);
2519 llog_ctxt_put(ctxt);
2521 CERROR("failed to connect rc %d idx %d\n", rc,
2522 body->lgdc_ctxt_idx);
2527 static int filter_llog_preclean(struct obd_device *obd)
2529 struct obd_llog_group *olg, *tmp;
2530 struct filter_obd *filter;
2531 cfs_list_t remove_list;
2535 rc = obd_llog_finish(obd, 0);
2537 CERROR("failed to cleanup llogging subsystem\n");
2539 filter = &obd->u.filter;
2540 CFS_INIT_LIST_HEAD(&remove_list);
2542 cfs_spin_lock(&filter->fo_llog_list_lock);
2543 while (!cfs_list_empty(&filter->fo_llog_list)) {
2544 olg = cfs_list_entry(filter->fo_llog_list.next,
2545 struct obd_llog_group, olg_list);
2546 cfs_list_del(&olg->olg_list);
2547 cfs_list_add(&olg->olg_list, &remove_list);
2549 cfs_spin_unlock(&filter->fo_llog_list_lock);
2551 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2552 cfs_list_del_init(&olg->olg_list);
2553 rc = filter_olg_fini(olg);
2555 CERROR("failed to cleanup llogging subsystem for %u\n",
2563 static int filter_precleanup(struct obd_device *obd,
2564 enum obd_cleanup_stage stage)
2570 case OBD_CLEANUP_EARLY:
2572 case OBD_CLEANUP_EXPORTS:
2573 /* Stop recovery before namespace cleanup. */
2574 target_recovery_fini(obd);
2575 rc = filter_llog_preclean(obd);
2581 static int filter_cleanup(struct obd_device *obd)
2583 struct filter_obd *filter = &obd->u.filter;
2587 LCONSOLE_WARN("%s: shutting down for failover; client state "
2588 "will be preserved.\n", obd->obd_name);
2590 obd_exports_barrier(obd);
2591 obd_zombie_barrier();
2593 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2594 lprocfs_free_per_client_stats(obd);
2595 lprocfs_free_obd_stats(obd);
2596 lprocfs_obd_cleanup(obd);
2597 lquota_cleanup(filter_quota_interface_ref, obd);
2599 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2600 obd->obd_namespace = NULL;
2602 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2604 if (obd->u.obt.obt_sb == NULL)
2609 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2610 shrink_dcache_sb(obd->u.obt.obt_sb);
2612 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2613 obd->u.obt.obt_sb = NULL;
2615 fsfilt_put_ops(obd->obd_fsops);
2617 filter_iobuf_pool_done(filter);
2619 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2624 static int filter_connect_internal(struct obd_export *exp,
2625 struct obd_connect_data *data,
2628 struct filter_export_data *fed = &exp->exp_filter_data;
2633 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2634 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2635 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2636 data->ocd_connect_flags, data->ocd_version,
2637 data->ocd_grant, data->ocd_index);
2639 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2640 CWARN("!!! This export (nid %s) used object group %d "
2641 "earlier; now it's trying to use group %d! This could "
2642 "be a bug in the MDS. Please report to "
2643 "http://bugzilla.lustre.org/\n",
2644 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2647 fed->fed_group = data->ocd_group;
2649 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2650 exp->exp_connect_flags = data->ocd_connect_flags;
2651 data->ocd_version = LUSTRE_VERSION_CODE;
2653 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2654 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2655 data->ocd_connect_flags & OBD_CONNECT_MDS))
2658 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2659 struct filter_obd *filter = &exp->exp_obd->u.filter;
2660 obd_size left, want;
2662 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2663 left = filter_grant_space_left(exp);
2664 want = data->ocd_grant;
2665 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2666 data->ocd_grant = fed->fed_grant;
2667 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2669 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2670 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2671 exp->exp_client_uuid.uuid, exp,
2672 data->ocd_grant, want, left);
2674 filter->fo_tot_granted_clients ++;
2677 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2678 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2679 int index = le32_to_cpu(lsd->lsd_ost_index);
2681 if (!(lsd->lsd_feature_compat &
2682 cpu_to_le32(OBD_COMPAT_OST))) {
2683 /* this will only happen on the first connect */
2684 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2685 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2686 /* sync is not needed here as filter_client_add will
2687 * set exp_need_sync flag */
2688 filter_update_server_data(exp->exp_obd);
2689 } else if (index != data->ocd_index) {
2690 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2691 " %u doesn't match actual OST index"
2692 " %u in last_rcvd file, bad "
2694 obd_export_nid2str(exp), index,
2698 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2699 * FIXME: We don't strictly need the COMPAT flag for that,
2700 * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2701 * FIXME: We needed it for the index, as index 0 is valid. */
2704 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2705 data->ocd_brw_size = 65536;
2706 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2707 data->ocd_brw_size = min(data->ocd_brw_size,
2708 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2709 LASSERT(data->ocd_brw_size);
2712 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2713 __u32 cksum_types = data->ocd_cksum_types;
2715 /* The client set in ocd_cksum_types the checksum types it
2716 * supports. We have to mask off the algorithms that we don't
2718 if (cksum_types & OBD_CKSUM_ALL)
2719 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2721 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2723 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2724 "%x\n", exp->exp_obd->obd_name,
2725 obd_export_nid2str(exp), cksum_types,
2726 data->ocd_cksum_types);
2728 /* This client does not support OBD_CONNECT_CKSUM
2729 * fall back to CRC32 */
2730 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2731 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2732 exp->exp_obd->obd_name,
2733 obd_export_nid2str(exp));
2739 static int filter_reconnect(const struct lu_env *env,
2740 struct obd_export *exp, struct obd_device *obd,
2741 struct obd_uuid *cluuid,
2742 struct obd_connect_data *data,
2748 if (exp == NULL || obd == NULL || cluuid == NULL)
2751 rc = filter_connect_internal(exp, data, 1);
2753 filter_export_stats_init(obd, exp, 1, localdata);
2758 /* nearly identical to mds_connect */
2759 static int filter_connect(const struct lu_env *env,
2760 struct obd_export **exp, struct obd_device *obd,
2761 struct obd_uuid *cluuid,
2762 struct obd_connect_data *data, void *localdata)
2764 struct lvfs_run_ctxt saved;
2765 struct lustre_handle conn = { 0 };
2766 struct obd_export *lexp;
2771 if (exp == NULL || obd == NULL || cluuid == NULL)
2774 rc = class_connect(&conn, obd, cluuid);
2777 lexp = class_conn2export(&conn);
2778 LASSERT(lexp != NULL);
2780 rc = filter_connect_internal(lexp, data, 0);
2784 filter_export_stats_init(obd, lexp, 0, localdata);
2785 if (obd->obd_replayable) {
2786 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2788 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2789 rc = filter_client_add(obd, lexp, -1);
2794 group = data->ocd_group;
2796 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2797 obd->obd_name, lexp->exp_handle.h_cookie, group);
2799 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2800 rc = filter_read_groups(obd, group, 1);
2801 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2803 CERROR("can't read group %u\n", group);
2811 class_disconnect(lexp);
2812 lprocfs_exp_cleanup(lexp);
2821 /* Do extra sanity checks for grant accounting. We do this at connect,
2822 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2823 * always get rid of it or turn it off when we know accounting is good. */
2824 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2826 struct filter_export_data *fed;
2827 struct obd_export *exp;
2828 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2829 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2830 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2832 if (cfs_list_empty(&obd->obd_exports))
2835 /* We don't want to do this for large machines that do lots of
2836 mounts or unmounts. It burns... */
2837 if (obd->obd_num_exports > 100)
2840 cfs_spin_lock(&obd->obd_osfs_lock);
2841 cfs_spin_lock(&obd->obd_dev_lock);
2842 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2844 fed = &exp->exp_filter_data;
2845 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2848 if (maxsize > 0) { /* we may not have done a statfs yet */
2849 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2850 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2851 exp->exp_client_uuid.uuid, exp,
2852 fed->fed_grant, fed->fed_pending, maxsize);
2853 LASSERTF(fed->fed_dirty <= maxsize,
2854 "%s: cli %s/%p %ld > "LPU64"\n", func,
2855 exp->exp_client_uuid.uuid, exp,
2856 fed->fed_dirty, maxsize);
2859 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2860 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2861 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2863 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2864 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2865 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2866 tot_granted += fed->fed_grant + fed->fed_pending;
2867 tot_pending += fed->fed_pending;
2868 tot_dirty += fed->fed_dirty;
2870 fo_tot_granted = obd->u.filter.fo_tot_granted;
2871 fo_tot_pending = obd->u.filter.fo_tot_pending;
2872 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2873 cfs_spin_unlock(&obd->obd_dev_lock);
2874 cfs_spin_unlock(&obd->obd_osfs_lock);
2876 /* Do these assertions outside the spinlocks so we don't kill system */
2877 if (tot_granted != fo_tot_granted)
2878 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2879 func, tot_granted, fo_tot_granted);
2880 if (tot_pending != fo_tot_pending)
2881 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2882 func, tot_pending, fo_tot_pending);
2883 if (tot_dirty != fo_tot_dirty)
2884 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2885 func, tot_dirty, fo_tot_dirty);
2886 if (tot_pending > tot_granted)
2887 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2888 func, tot_pending, tot_granted);
2889 if (tot_granted > maxsize)
2890 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2891 func, tot_granted, maxsize);
2892 if (tot_dirty > maxsize)
2893 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2894 func, tot_dirty, maxsize);
2897 /* Remove this client from the grant accounting totals. We also remove
2898 * the export from the obd device under the osfs and dev locks to ensure
2899 * that the filter_grant_sanity_check() calculations are always valid.
2900 * The client should do something similar when it invalidates its import. */
2901 static void filter_grant_discard(struct obd_export *exp)
2903 struct obd_device *obd = exp->exp_obd;
2904 struct filter_obd *filter = &obd->u.filter;
2905 struct filter_export_data *fed = &exp->exp_filter_data;
2907 cfs_spin_lock(&obd->obd_osfs_lock);
2908 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2909 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2910 obd->obd_name, filter->fo_tot_granted,
2911 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2912 filter->fo_tot_granted -= fed->fed_grant;
2913 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2914 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2915 obd->obd_name, filter->fo_tot_pending,
2916 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2917 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2918 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2919 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2920 obd->obd_name, filter->fo_tot_dirty,
2921 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2922 filter->fo_tot_dirty -= fed->fed_dirty;
2926 cfs_spin_unlock(&obd->obd_osfs_lock);
2929 static int filter_destroy_export(struct obd_export *exp)
2931 struct filter_export_data *fed = &exp->exp_filter_data;
2934 if (fed->fed_pending)
2935 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2936 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2937 exp, fed->fed_pending);
2939 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2941 target_destroy_export(exp);
2942 ldlm_destroy_export(exp);
2943 lut_client_free(exp);
2945 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2948 if (!exp->exp_obd->obd_replayable)
2949 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2951 filter_grant_discard(exp);
2952 filter_fmd_cleanup(exp);
2954 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2955 struct filter_obd *filter = &exp->exp_obd->u.filter;
2956 if (filter->fo_tot_granted_clients > 0)
2957 filter->fo_tot_granted_clients --;
2960 if (!(exp->exp_flags & OBD_OPT_FORCE))
2961 filter_grant_sanity_check(exp->exp_obd, __func__);
2966 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2968 struct obd_llog_group *olg_min, *olg;
2969 struct filter_obd *filter;
2970 int worked = -1, group;
2971 struct llog_ctxt *ctxt;
2974 filter = &obd->u.filter;
2976 /* we can't sync log holding spinlock. also, we do not want to get
2977 * into livelock. so we do following: loop over MDS's exports in
2978 * group order and skip already synced llogs -bzzz */
2980 /* look for group with min. number, but > worked */
2983 cfs_spin_lock(&filter->fo_llog_list_lock);
2984 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2985 if (olg->olg_seq <= worked) {
2986 /* this group is already synced */
2989 if (group < olg->olg_seq) {
2990 /* we have group with smaller number to sync */
2993 /* store current minimal group */
2995 group = olg->olg_seq;
2997 cfs_spin_unlock(&filter->fo_llog_list_lock);
2999 if (olg_min == NULL)
3002 worked = olg_min->olg_seq;
3003 if (olg_min->olg_exp &&
3004 (dexp == olg_min->olg_exp || dexp == NULL)) {
3006 ctxt = llog_group_get_ctxt(olg_min,
3007 LLOG_MDS_OST_REPL_CTXT);
3009 err = llog_sync(ctxt, olg_min->olg_exp);
3010 llog_ctxt_put(ctxt);
3012 CERROR("error flushing logs to MDS: "
3017 } while (olg_min != NULL);
3020 /* Also incredibly similar to mds_disconnect */
3021 static int filter_disconnect(struct obd_export *exp)
3023 struct obd_device *obd = exp->exp_obd;
3028 class_export_get(exp);
3030 if (!(exp->exp_flags & OBD_OPT_FORCE))
3031 filter_grant_sanity_check(obd, __func__);
3032 filter_grant_discard(exp);
3034 /* Flush any remaining cancel messages out to the target */
3035 filter_sync_llogs(obd, exp);
3037 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3039 rc = server_disconnect_export(exp);
3041 /* Do not erase record for recoverable client. */
3042 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3043 filter_client_del(exp);
3045 fsfilt_sync(obd, obd->u.obt.obt_sb);
3047 class_export_put(exp);
3051 /* reverse import is changed, sync all cancels */
3052 static void filter_revimp_update(struct obd_export *exp)
3057 class_export_get(exp);
3059 /* flush any remaining cancel messages out to the target */
3060 filter_sync_llogs(exp->exp_obd, exp);
3061 class_export_put(exp);
3065 static int filter_ping(struct obd_export *exp)
3067 filter_fmd_expire(exp);
3071 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3072 const char *what, int quiet)
3074 struct dentry *dchild = NULL;
3076 dchild = filter_fid2dentry(obd, NULL, ostid->oi_seq, ostid->oi_id);
3078 if (IS_ERR(dchild)) {
3079 CERROR("%s error looking up object: "POSTID"\n",
3080 what, ostid->oi_id, ostid->oi_seq);
3084 if (dchild->d_inode == NULL) {
3086 CERROR("%s: %s on non-existent object: "POSTID" \n",
3087 obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3089 RETURN(ERR_PTR(-ENOENT));
3095 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3097 struct dentry *dentry = NULL;
3098 struct obd_device *obd;
3102 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3103 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3107 obd = class_exp2obd(exp);
3109 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3113 dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3115 RETURN(PTR_ERR(dentry));
3117 /* Limit the valid bits in the return data to what we actually use */
3118 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3119 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3125 /* this should be enabled/disabled in condition to enabled/disabled large
3126 * inodes (fast EAs) in backing store FS. */
3127 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3128 void *handle, struct obdo *oa)
3130 struct obd_device *obd = exp->exp_obd;
3134 if (oa->o_valid & OBD_MD_FLFID) {
3135 struct filter_fid ff;
3137 if (!(oa->o_valid & OBD_MD_FLGROUP))
3139 /* packing fid and converting it to LE for storing into EA.
3140 * Here ->o_stripe_idx should be filled by LOV and rest of
3141 * fields - by client. */
3142 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3143 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3144 /* XXX: we are ignoring o_parent_ver here, since this should
3145 * be the same for all objects in this fileset. */
3146 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3147 ff.ff_objid = cpu_to_le64(oa->o_id);
3148 ff.ff_seq = cpu_to_le64(oa->o_seq);
3150 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3151 LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3154 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3156 CERROR("store fid in object failed! rc: %d\n", rc);
3158 CDEBUG(D_HA, "OSS object without fid info!\n");
3164 /* this is called from filter_truncate() until we have filter_punch() */
3165 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3166 struct obdo *oa, struct obd_trans_info *oti)
3168 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3169 struct llog_cookie *fcc = NULL;
3170 struct filter_obd *filter;
3171 int rc, err, sync = 0;
3172 loff_t old_size = 0;
3173 unsigned int ia_valid;
3174 struct inode *inode;
3175 struct page *page = NULL;
3180 LASSERT(dentry != NULL);
3181 LASSERT(!IS_ERR(dentry));
3183 inode = dentry->d_inode;
3184 LASSERT(inode != NULL);
3186 filter = &exp->exp_obd->u.filter;
3187 iattr_from_obdo(&iattr, oa, oa->o_valid);
3188 ia_valid = iattr.ia_valid;
3190 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3191 OBD_ALLOC(fcc, sizeof(*fcc));
3193 *fcc = oa->o_lcookie;
3195 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3197 /* Filter truncates and writes are serialized by
3198 * i_alloc_sem, see the comment in
3199 * filter_preprw_write.*/
3200 if (ia_valid & ATTR_SIZE)
3201 down_write(&inode->i_alloc_sem);
3202 LOCK_INODE_MUTEX(inode);
3203 old_size = i_size_read(inode);
3206 /* VBR: version recovery check */
3207 rc = filter_version_get_check(exp, oti, inode);
3209 GOTO(out_unlock, rc);
3211 /* Let's pin the last page so that ldiskfs_truncate
3212 * should not start GFP_FS allocation. */
3213 if (ia_valid & ATTR_SIZE) {
3214 page = grab_cache_page(inode->i_mapping,
3215 iattr.ia_size >> PAGE_CACHE_SHIFT);
3217 GOTO(out_unlock, rc = -ENOMEM);
3222 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3223 * then we will accept the UID+GID sent by the client during write for
3224 * initializing the ownership of this inode. We only allow this to
3225 * happen once so clear these bits in setattr. In 2.6 kernels it is
3226 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3227 * the flags that are actually being set. */
3228 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3229 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3230 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3232 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3233 if (!(ia_valid & ATTR_MODE)) {
3234 iattr.ia_mode = inode->i_mode;
3235 iattr.ia_valid |= ATTR_MODE;
3237 iattr.ia_mode &= ~S_ISUID;
3239 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3240 if (!(iattr.ia_valid & ATTR_MODE)) {
3241 iattr.ia_mode = inode->i_mode;
3242 iattr.ia_valid |= ATTR_MODE;
3244 iattr.ia_mode &= ~S_ISGID;
3247 orig_ids[USRQUOTA] = inode->i_uid;
3248 orig_ids[GRPQUOTA] = inode->i_gid;
3249 handle = fsfilt_start_log(exp->exp_obd, inode,
3250 FSFILT_OP_SETATTR, oti, 1);
3252 GOTO(out_unlock, rc = PTR_ERR(handle));
3254 /* update inode EA only once when inode is suid bit marked. As
3255 * on 2.6.x UID and GID may be set separately, we check here
3256 * only one of them to avoid double setting. */
3257 if (inode->i_mode & S_ISUID)
3258 filter_update_fidea(exp, inode, handle, oa);
3260 handle = fsfilt_start(exp->exp_obd, inode,
3261 FSFILT_OP_SETATTR, oti);
3263 GOTO(out_unlock, rc = PTR_ERR(handle));
3265 if (oa->o_valid & OBD_MD_FLFLAGS) {
3266 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3267 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3269 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3271 /* set cancel cookie callback function */
3272 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3273 filter_cancel_cookies_cb,
3277 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3278 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3280 /* The truncate might have used up our transaction credits. Make sure
3281 * we have two left for the last_rcvd and VBR inode version updates. */
3282 err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3284 rc = filter_finish_transno(exp, inode, oti, rc, sync);
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 down_write(&dentry->d_inode->i_alloc_sem);
3400 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3401 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3402 fmd->fmd_mactime_xid = oti->oti_xid;
3403 filter_fmd_put(exp, fmd);
3404 up_write(&dentry->d_inode->i_alloc_sem);
3407 /* setting objects attributes (including owner/group) */
3408 rc = filter_setattr_internal(exp, dentry, oa, oti);
3410 GOTO(out_unlock, rc);
3412 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3413 &res_id, LDLM_EXTENT, 0);
3416 LDLM_RESOURCE_ADDREF(res);
3417 rc = ldlm_res_lvbo_update(res, NULL, 0);
3418 LDLM_RESOURCE_DELREF(res);
3419 ldlm_resource_putref(res);
3422 oa->o_valid = OBD_MD_FLID;
3424 /* Quota release need uid/gid info */
3425 obdo_from_inode(oa, dentry->d_inode, NULL,
3426 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3431 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3435 /* XXX identical to osc_unpackmd */
3436 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3437 struct lov_mds_md *lmm, int lmm_bytes)
3443 if (lmm_bytes < sizeof (*lmm)) {
3444 CERROR("lov_mds_md too small: %d, need %d\n",
3445 lmm_bytes, (int)sizeof(*lmm));
3448 /* XXX LOV_MAGIC etc check? */
3450 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3451 CERROR("lov_mds_md: zero lmm_object_id\n");
3456 lsm_size = lov_stripe_md_size(1);
3460 if (*lsmp != NULL && lmm == NULL) {
3461 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3462 OBD_FREE(*lsmp, lsm_size);
3467 if (*lsmp == NULL) {
3468 OBD_ALLOC(*lsmp, lsm_size);
3472 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3473 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3474 OBD_FREE(*lsmp, lsm_size);
3477 loi_init((*lsmp)->lsm_oinfo[0]);
3481 /* XXX zero *lsmp? */
3482 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3483 LASSERT((*lsmp)->lsm_object_id);
3486 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3491 /* caller must hold fo_create_locks[oa->o_seq] */
3492 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3493 struct filter_obd *filter)
3495 struct obdo doa = { 0 }; /* XXX obdo on stack */
3501 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_seq]) != 0);
3503 memset(&doa, 0, sizeof(doa));
3505 doa.o_valid |= OBD_MD_FLGROUP;
3506 doa.o_seq = oa->o_seq;
3507 doa.o_mode = S_IFREG;
3509 if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3510 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3511 exp->exp_obd->obd_name, doa.o_seq);
3515 last = filter_last_id(filter, doa.o_seq);
3517 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3519 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3520 exp->exp_obd->obd_name, oa->o_id + 1, last,
3521 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3523 for (id = last; id > oa->o_id; id--) {
3525 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3526 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3527 CEMERG("error destroying precreate objid "LPU64": %d\n",
3530 /* update last_id on disk periodically so that if we restart
3531 * we don't need to re-scan all of the just-deleted objects. */
3532 if ((id & 511) == 0 && !skip_orphan) {
3533 filter_set_last_id(filter, id - 1, doa.o_seq);
3534 filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3538 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3539 exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3542 filter_set_last_id(filter, id, doa.o_seq);
3543 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3545 /* don't reuse orphan object, return last used objid */
3549 cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3554 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3555 obd_seq group, int *num);
3556 /* returns a negative error or a nonnegative number of files to create */
3557 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3558 obd_seq group, struct obd_trans_info *oti)
3560 struct obd_device *obd = exp->exp_obd;
3561 struct filter_obd *filter = &obd->u.filter;
3565 /* delete orphans request */
3566 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3567 obd_id last = filter_last_id(filter, group);
3569 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3570 CERROR("%s: dropping old orphan cleanup request\n",
3574 /* This causes inflight precreates to abort and drop lock */
3575 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3576 cfs_down(&filter->fo_create_locks[group]);
3577 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3578 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3579 exp->exp_obd->obd_name, group);
3580 cfs_up(&filter->fo_create_locks[group]);
3583 diff = oa->o_id - last;
3584 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3587 if (-diff > OST_MAX_PRECREATE) {
3588 CERROR("%s: ignoring bogus orphan destroy request: "
3589 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3591 /* FIXME: should reset precreate_next_id on MDS */
3592 GOTO(out, rc = -EINVAL);
3595 rc = filter_destroy_precreated(exp, oa, filter);
3597 CERROR("%s: unable to write lastobjid, but "
3598 "orphans were deleted\n", obd->obd_name);
3601 /* XXX: Used by MDS for the first time! */
3602 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3605 cfs_down(&filter->fo_create_locks[group]);
3606 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3607 CERROR("%s: dropping old precreate request\n",
3611 /* only precreate if group == 0 and o_id is specfied */
3612 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3615 diff = oa->o_id - filter_last_id(filter, group);
3616 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3617 filter_last_id(filter, group), diff);
3619 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3620 oa->o_id, filter_last_id(filter, group), diff);
3624 oa->o_id = filter_last_id(&obd->u.filter, group);
3625 rc = filter_precreate(obd, oa, group, &diff);
3626 oa->o_id = filter_last_id(&obd->u.filter, group);
3628 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3631 /* else diff == 0 */
3634 cfs_up(&filter->fo_create_locks[group]);
3638 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3639 __u64 max_age, __u32 flags)
3641 struct filter_obd *filter = &obd->u.filter;
3642 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3646 /* at least try to account for cached pages. its still racey and
3647 * might be under-reporting if clients haven't announced their
3648 * caches with brw recently */
3649 cfs_spin_lock(&obd->obd_osfs_lock);
3650 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3651 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3652 cfs_spin_unlock(&obd->obd_osfs_lock);
3654 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3655 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3656 filter->fo_tot_dirty, filter->fo_tot_granted,
3657 filter->fo_tot_pending,
3658 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3660 filter_grant_sanity_check(obd, __func__);
3662 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3663 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3664 osfs->os_bsize - 1) >> blockbits));
3666 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3667 struct lr_server_data *lsd = class_server_data(obd);
3668 int index = le32_to_cpu(lsd->lsd_ost_index);
3670 if (obd_fail_val == -1 ||
3671 index == obd_fail_val)
3672 osfs->os_bfree = osfs->os_bavail = 2;
3673 else if (obd_fail_loc & OBD_FAIL_ONCE)
3674 obd_fail_loc &= ~OBD_FAILED; /* reset flag */
3677 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3678 * stop creating files on MDS if OST is not good shape to create
3682 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3683 osfs->os_state = OS_STATE_READONLY;
3685 if (filter->fo_raid_degraded)
3686 osfs->os_state |= OS_STATE_DEGRADED;
3690 static int filter_use_existing_obj(struct obd_device *obd,
3691 struct dentry *dchild, void **handle,
3694 struct inode *inode = dchild->d_inode;
3698 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3701 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3702 if (IS_ERR(*handle))
3703 return PTR_ERR(*handle);
3705 iattr.ia_valid = ATTR_MODE;
3706 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3707 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3714 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3717 __u64 os_ffree = -1;
3719 cfs_spin_lock(&obd->obd_osfs_lock);
3720 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3722 os_ffree = obd->obd_osfs.os_ffree;
3723 cfs_spin_unlock(&obd->obd_osfs_lock);
3728 /* We rely on the fact that only one thread will be creating files in a given
3729 * group at a time, which is why we don't need an atomic filter_get_new_id.
3730 * Even if we had that atomic function, the following race would exist:
3732 * thread 1: gets id x from filter_next_id
3733 * thread 2: gets id (x + 1) from filter_next_id
3734 * thread 2: creates object (x + 1)
3735 * thread 1: tries to create object x, gets -ENOSPC
3737 * Caller must hold fo_create_locks[group]
3739 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3740 obd_seq group, int *num)
3742 struct dentry *dchild = NULL, *dparent = NULL;
3743 struct filter_obd *filter;
3744 struct obd_statfs *osfs;
3745 int err = 0, rc = 0, recreate_obj = 0, i;
3746 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3749 void *handle = NULL;
3752 filter = &obd->u.filter;
3754 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3756 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3758 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3759 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3762 OBD_ALLOC(osfs, sizeof(*osfs));
3765 rc = filter_statfs(obd, osfs,
3766 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
3768 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3769 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3770 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3771 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3775 OBD_FREE(osfs, sizeof(*osfs));
3780 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3781 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3783 for (i = 0; i < *num && err == 0; i++) {
3784 int cleanup_phase = 0;
3786 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3787 CWARN("%s: create aborted by destroy\n",
3796 last_id = filter_last_id(filter, group);
3797 if (next_id > last_id) {
3798 CERROR("Error: Trying to recreate obj greater"
3799 "than last id "LPD64" > "LPD64"\n",
3801 GOTO(cleanup, rc = -EINVAL);
3804 next_id = filter_last_id(filter, group) + 1;
3806 /* Temporary solution for oid in CMD before fid-on-OST */
3807 if ((fid_seq_is_mdt0(oa->o_seq) && next_id >= IDIF_MAX_OID) &&
3808 (fid_seq_is_cmd(oa->o_seq) && next_id >= OBIF_MAX_OID)) {
3809 CERROR("%s:"POSTID" hit the max IDIF_MAX_OID(1<<48)!\n",
3810 obd->obd_name, next_id, group);
3811 GOTO(cleanup, rc = -ENOSPC);
3814 dparent = filter_parent_lock(obd, group, next_id);
3815 if (IS_ERR(dparent))
3816 GOTO(cleanup, rc = PTR_ERR(dparent));
3817 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3819 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3821 GOTO(cleanup, rc = PTR_ERR(dchild));
3822 cleanup_phase = 2; /* f_dput(dchild) */
3824 if (dchild->d_inode != NULL) {
3825 /* This would only happen if lastobjid was bad on disk*/
3826 /* Could also happen if recreating missing obj but it
3827 * already exists. */
3829 CERROR("%s: recreating existing object %.*s?\n",
3830 obd->obd_name, dchild->d_name.len,
3831 dchild->d_name.name);
3833 /* Use these existing objects if they are
3835 if (dchild->d_inode->i_size == 0) {
3836 rc = filter_use_existing_obj(obd,dchild,
3837 &handle, &cleanup_phase);
3844 CERROR("%s: Serious error: objid %.*s already "
3845 "exists; is this filesystem corrupt?\n",
3846 obd->obd_name, dchild->d_name.len,
3847 dchild->d_name.name);
3850 GOTO(cleanup, rc = -EEXIST);
3853 handle = fsfilt_start_log(obd, dparent->d_inode,
3854 FSFILT_OP_CREATE, NULL, 1);
3856 GOTO(cleanup, rc = PTR_ERR(handle));
3859 CDEBUG(D_INODE, "%s: filter_precreate(od->o_seq="LPU64
3860 ",od->o_id="LPU64")\n", obd->obd_name, group,
3863 /* We mark object SUID+SGID to flag it for accepting UID+GID
3864 * from client on first write. Currently the permission bits
3865 * on the OST are never used, so this is OK. */
3866 rc = ll_vfs_create(dparent->d_inode, dchild,
3867 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3869 CERROR("create failed rc = %d\n", rc);
3870 if (rc == -ENOSPC) {
3871 os_ffree = filter_calc_free_inodes(obd);
3873 CERROR("%s: free inode "LPU64"\n",
3874 obd->obd_name, os_ffree);
3879 if (dchild->d_inode)
3880 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3881 dchild->d_inode->i_ino);
3884 if (!recreate_obj) {
3885 filter_set_last_id(filter, next_id, group);
3886 err = filter_update_last_objid(obd, group, 0);
3888 CERROR("unable to write lastobjid "
3889 "but file created\n");
3893 switch(cleanup_phase) {
3895 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3897 CERROR("error on commit, err = %d\n", err);
3904 filter_parent_unlock(dparent);
3911 if (cfs_time_after(jiffies, enough_time)) {
3913 "%s: precreate slow - want %d got %d \n",
3914 obd->obd_name, *num, i);
3921 "%s: created %d objects for group "POSTID" rc %d\n",
3922 obd->obd_name, i, filter->fo_last_objids[group], group, rc);
3927 int filter_create(struct obd_export *exp, struct obdo *oa,
3928 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3930 struct obd_device *obd = exp->exp_obd;
3931 struct filter_export_data *fed;
3932 struct filter_obd *filter;
3933 struct lvfs_run_ctxt saved;
3934 struct lov_stripe_md *lsm = NULL;
3938 CDEBUG(D_INODE, "%s: filter_create(group="LPU64",id="
3939 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
3941 fed = &exp->exp_filter_data;
3942 filter = &obd->u.filter;
3944 if (fed->fed_group != oa->o_seq) {
3945 CERROR("%s: this export (nid %s) used object group %d "
3946 "earlier; now it's trying to use group "LPU64"!"
3947 " This could be a bug in the MDS. Please report to "
3948 "http://bugzilla.lustre.org/\n", obd->obd_name,
3949 obd_export_nid2str(exp), fed->fed_group, oa->o_seq);
3956 rc = obd_alloc_memmd(exp, &lsm);
3963 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3965 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3966 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3967 if (!obd->obd_recovering ||
3968 oa->o_id > filter_last_id(filter, oa->o_seq)) {
3969 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3970 oa->o_id, filter_last_id(filter, oa->o_seq));
3974 cfs_down(&filter->fo_create_locks[oa->o_seq]);
3975 rc = filter_precreate(obd, oa, oa->o_seq, &diff);
3976 cfs_up(&filter->fo_create_locks[oa->o_seq]);
3979 rc = filter_handle_precreate(exp, oa, oa->o_seq, oti);
3982 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3983 if (rc && ea != NULL && *ea != lsm) {
3984 obd_free_memmd(exp, &lsm);
3985 } else if (rc == 0 && ea != NULL) {
3986 /* XXX LOV STACKING: the lsm that is passed to us from
3987 * LOV does not have valid lsm_oinfo data structs, so
3988 * don't go touching that. This needs to be fixed in a
3990 lsm->lsm_object_id = oa->o_id;
3997 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3998 struct lov_stripe_md *md, struct obd_trans_info *oti,
3999 struct obd_export *md_exp, void *capa)
4001 unsigned int qcids[MAXQUOTAS] = {0, 0};
4002 struct obd_device *obd;
4003 struct filter_obd *filter;
4004 struct dentry *dchild = NULL, *dparent = NULL;
4005 struct lustre_handle lockh = { 0 };
4006 struct lvfs_run_ctxt saved;
4007 void *handle = NULL;
4008 struct llog_cookie *fcc = NULL;
4009 int rc, rc2, cleanup_phase = 0, sync = 0;
4013 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4014 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4019 filter = &obd->u.filter;
4021 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4024 CDEBUG(D_INODE, "%s: filter_destroy(group="LPU64",oid="
4025 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4027 dchild = filter_fid2dentry(obd, NULL, oa->o_seq, oa->o_id);
4029 GOTO(cleanup, rc = PTR_ERR(dchild));
4032 if (dchild->d_inode == NULL) {
4033 CDEBUG(D_INODE, "destroying non-existent object "POSTID"\n",
4034 oa->o_id, oa->o_seq);
4035 /* If object already gone, cancel cookie right now */
4036 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4037 struct llog_ctxt *ctxt;
4038 struct obd_llog_group *olg;
4040 olg = filter_find_olg(obd, oa->o_seq);
4042 CERROR(" %s: can not find olg of group %d\n",
4043 obd->obd_name, (int)oa->o_seq);
4044 GOTO(cleanup, rc = PTR_ERR(olg));
4046 fcc = &oa->o_lcookie;
4047 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4048 llog_cancel(ctxt, NULL, 1, fcc, 0);
4049 llog_ctxt_put(ctxt);
4050 fcc = NULL; /* we didn't allocate fcc, don't free it */
4052 GOTO(cleanup, rc = -ENOENT);
4055 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_seq, &lockh);
4059 /* Our MDC connection is established by the MDS to us */
4060 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4061 OBD_ALLOC(fcc, sizeof(*fcc));
4063 *fcc = oa->o_lcookie;
4065 DQUOT_INIT(dchild->d_inode);
4067 /* we're gonna truncate it first in order to avoid possible deadlock:
4069 * open trasaction open transaction
4070 * down(i_zombie) down(i_zombie)
4071 * restart transaction
4072 * (see BUG 4180) -bzzz
4074 * take i_alloc_sem too to prevent other threads from writing to the
4075 * file while we are truncating it. This can cause lock ordering issue
4076 * between page lock, i_mutex & starting new journal handle.
4077 * (see bug 20321) -johann
4079 down_write(&dchild->d_inode->i_alloc_sem);
4080 LOCK_INODE_MUTEX(dchild->d_inode);
4082 /* VBR: version recovery check */
4083 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4085 UNLOCK_INODE_MUTEX(dchild->d_inode);
4086 up_write(&dchild->d_inode->i_alloc_sem);
4090 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4092 if (IS_ERR(handle)) {
4093 UNLOCK_INODE_MUTEX(dchild->d_inode);
4094 up_write(&dchild->d_inode->i_alloc_sem);
4095 GOTO(cleanup, rc = PTR_ERR(handle));
4098 iattr.ia_valid = ATTR_SIZE;
4100 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4101 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4102 UNLOCK_INODE_MUTEX(dchild->d_inode);
4103 up_write(&dchild->d_inode->i_alloc_sem);
4107 GOTO(cleanup, rc = rc2);
4109 /* We don't actually need to lock the parent until we are unlinking
4110 * here, and not while truncating above. That avoids holding the
4111 * parent lock for a long time during truncate, which can block other
4112 * threads from doing anything to objects in that directory. bug 7171 */
4113 dparent = filter_parent_lock(obd, oa->o_seq, oa->o_id);
4114 if (IS_ERR(dparent))
4115 GOTO(cleanup, rc = PTR_ERR(dparent));
4116 cleanup_phase = 3; /* filter_parent_unlock */
4118 LOCK_INODE_MUTEX(dchild->d_inode);
4119 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4120 if (IS_ERR(handle)) {
4121 UNLOCK_INODE_MUTEX(dchild->d_inode);
4122 GOTO(cleanup, rc = PTR_ERR(handle));
4124 cleanup_phase = 4; /* fsfilt_commit */
4126 /* Quota release need uid/gid of inode */
4127 obdo_from_inode(oa, dchild->d_inode, NULL, OBD_MD_FLUID|OBD_MD_FLGID);
4129 filter_fmd_drop(exp, oa->o_id, oa->o_seq);
4131 /* this drops dchild->d_inode->i_mutex unconditionally */
4132 rc = filter_destroy_internal(obd, oa->o_id, oa->o_seq, dparent, dchild);
4136 switch(cleanup_phase) {
4139 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4140 oti->oti_handle : handle,
4141 filter_cancel_cookies_cb,
4143 /* If add_journal_cb failed, then filter_finish_transno
4144 * will commit the handle and we will do a sync
4145 * on commit. then we call callback directly to free
4148 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4150 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4153 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4155 CERROR("error on commit, err = %d\n", rc2);
4162 filter_parent_unlock(dparent);
4164 filter_fini_destroy(obd, &lockh);
4168 OBD_FREE(fcc, sizeof(*fcc));
4170 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4173 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4177 /* trigger quota release */
4178 qcids[USRQUOTA] = oa->o_uid;
4179 qcids[GRPQUOTA] = oa->o_gid;
4180 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4183 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4187 /* NB start and end are used for punch, but not truncate */
4188 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4189 struct obd_trans_info *oti,
4190 struct ptlrpc_request_set *rqset)
4195 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4196 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4197 oinfo->oi_policy.l_extent.end);
4201 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4202 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,oinfo->oi_oa->o_valid,
4203 oinfo->oi_policy.l_extent.start);
4205 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4206 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4207 rc = filter_setattr(exp, oinfo, oti);
4208 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4212 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4213 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4216 struct lvfs_run_ctxt saved;
4217 struct obd_device_target *obt;
4218 struct dentry *dentry;
4222 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4223 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4227 obt = &exp->exp_obd->u.obt;
4229 /* An objid of zero is taken to mean "sync whole filesystem" */
4230 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4231 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4232 /* Flush any remaining cancel messages out to the target */
4233 filter_sync_llogs(exp->exp_obd, exp);
4237 dentry = filter_oa2dentry(exp->exp_obd, &oa->o_oi);
4239 RETURN(PTR_ERR(dentry));
4241 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4243 LOCK_INODE_MUTEX(dentry->d_inode);
4245 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4247 /* just any file to grab fsync method - "file" arg unused */
4248 struct file *file = obt->obt_rcvd_filp;
4250 if (file->f_op && file->f_op->fsync)
4251 rc = file->f_op->fsync(NULL, dentry, 1);
4253 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4257 UNLOCK_INODE_MUTEX(dentry->d_inode);
4259 oa->o_valid = OBD_MD_FLID;
4260 obdo_from_inode(oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
4262 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4268 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4269 void *key, __u32 *vallen, void *val,
4270 struct lov_stripe_md *lsm)
4272 struct obd_device *obd;
4275 obd = class_exp2obd(exp);
4277 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4281 if (KEY_IS(KEY_BLOCKSIZE)) {
4282 __u32 *blocksize = val;
4284 if (*vallen < sizeof(*blocksize))
4286 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4288 *vallen = sizeof(*blocksize);
4292 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4293 __u32 *blocksize_bits = val;
4294 if (blocksize_bits) {
4295 if (*vallen < sizeof(*blocksize_bits))
4297 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4299 *vallen = sizeof(*blocksize_bits);
4303 if (KEY_IS(KEY_LAST_ID)) {
4304 obd_id *last_id = val;
4305 /* FIXME: object groups */
4307 if (*vallen < sizeof(*last_id))
4309 *last_id = filter_last_id(&obd->u.filter,
4310 exp->exp_filter_data.fed_group);
4312 *vallen = sizeof(*last_id);
4316 if (KEY_IS(KEY_FIEMAP)) {
4317 struct ll_fiemap_info_key *fm_key = key;
4318 struct dentry *dentry;
4319 struct ll_user_fiemap *fiemap = val;
4320 struct lvfs_run_ctxt saved;
4323 if (fiemap == NULL) {
4324 *vallen = fiemap_count_to_size(
4325 fm_key->fiemap.fm_extent_count);
4329 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa.o_oi,
4332 RETURN(PTR_ERR(dentry));
4334 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4335 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4336 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4338 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4344 if (KEY_IS(KEY_SYNC_LOCK_CANCEL)) {
4345 *((__u32 *) val) = obd->u.filter.fo_sync_lock_cancel;
4346 *vallen = sizeof(__u32);
4350 CDEBUG(D_IOCTL, "invalid key\n");
4354 static inline int filter_setup_llog_group(struct obd_export *exp,
4355 struct obd_device *obd,
4358 struct obd_llog_group *olg;
4359 struct llog_ctxt *ctxt;
4362 olg = filter_find_create_olg(obd, group);
4364 RETURN(PTR_ERR(olg));
4366 llog_group_set_export(olg, exp);
4368 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4369 LASSERTF(ctxt != NULL, "ctxt is null\n");
4371 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4372 llog_ctxt_put(ctxt);
4376 static int filter_set_grant_shrink(struct obd_export *exp,
4377 struct ost_body *body)
4379 /* handle shrink grant */
4380 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4381 filter_grant_incoming(exp, &body->oa);
4382 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4388 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4390 struct obd_device *obd;
4396 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4400 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4401 obd_export_nid2str(exp));
4402 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4404 /* setup llog imports */
4406 group = (int)(*(__u32 *)val);
4408 group = 0; /* default value */
4410 LASSERT_SEQ_IS_MDT(group);
4411 rc = filter_setup_llog_group(exp, obd, group);
4415 if (group == FID_SEQ_OST_MDT0) {
4416 /* setup llog group 1 for interop */
4417 filter_setup_llog_group(exp, obd, FID_SEQ_LLOG);
4420 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4425 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4426 void *key, __u32 vallen, void *val,
4427 struct ptlrpc_request_set *set)
4429 struct obd_device *obd;
4434 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4438 if (KEY_IS(KEY_CAPA_KEY)) {
4440 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4442 CERROR("filter update capability key failed: %d\n", rc);
4446 if (KEY_IS(KEY_REVIMP_UPD)) {
4447 filter_revimp_update(exp);
4448 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4452 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4453 filter_adapt_sptlrpc_conf(obd, 0);
4457 if (KEY_IS(KEY_MDS_CONN))
4458 RETURN(filter_set_mds_conn(exp, val));
4460 if (KEY_IS(KEY_GRANT_SHRINK))
4461 RETURN(filter_set_grant_shrink(exp, val));
4466 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4467 int len, void *karg, void *uarg)
4469 struct obd_device *obd = exp->exp_obd;
4470 struct obd_ioctl_data *data = karg;
4474 case OBD_IOC_ABORT_RECOVERY: {
4475 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4476 target_stop_recovery_thread(obd);
4480 case OBD_IOC_SYNC: {
4481 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4482 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4486 case OBD_IOC_SET_READONLY: {
4488 struct super_block *sb = obd->u.obt.obt_sb;
4489 struct inode *inode = sb->s_root->d_inode;
4490 BDEVNAME_DECLARE_STORAGE(tmp);
4491 CERROR("*** setting device %s read-only ***\n",
4492 ll_bdevname(sb, tmp));
4494 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4495 if (!IS_ERR(handle))
4496 rc = fsfilt_commit(obd, inode, handle, 1);
4498 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4499 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4501 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4505 case OBD_IOC_CATLOGLIST: {
4506 rc = llog_catalog_list(obd, 1, data);
4510 case OBD_IOC_LLOG_CANCEL:
4511 case OBD_IOC_LLOG_REMOVE:
4512 case OBD_IOC_LLOG_INFO:
4513 case OBD_IOC_LLOG_PRINT: {
4514 /* FIXME to be finished */
4515 RETURN(-EOPNOTSUPP);
4517 struct llog_ctxt *ctxt = NULL;
4519 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4520 rc = llog_ioctl(ctxt, cmd, data);
4521 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4534 static int filter_health_check(struct obd_device *obd)
4536 #ifdef USE_HEALTH_CHECK_WRITE
4537 struct filter_obd *filter = &obd->u.filter;
4542 * health_check to return 0 on healthy
4543 * and 1 on unhealthy.
4545 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4548 #ifdef USE_HEALTH_CHECK_WRITE
4549 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4550 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4555 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4558 return filter_fid2dentry(data, NULL, gr, id);
4561 static int filter_process_config(struct obd_device *obd, obd_count len,
4564 struct lustre_cfg *lcfg = buf;
4565 struct lprocfs_static_vars lvars;
4568 switch (lcfg->lcfg_command) {
4570 lprocfs_filter_init_vars(&lvars);
4572 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4582 static struct lvfs_callback_ops filter_lvfs_ops = {
4583 l_fid2dentry: filter_lvfs_fid2dentry,
4586 static struct obd_ops filter_obd_ops = {
4587 .o_owner = THIS_MODULE,
4588 .o_get_info = filter_get_info,
4589 .o_set_info_async = filter_set_info_async,
4590 .o_setup = filter_setup,
4591 .o_precleanup = filter_precleanup,
4592 .o_cleanup = filter_cleanup,
4593 .o_connect = filter_connect,
4594 .o_reconnect = filter_reconnect,
4595 .o_disconnect = filter_disconnect,
4596 .o_ping = filter_ping,
4597 .o_init_export = filter_init_export,
4598 .o_destroy_export = filter_destroy_export,
4599 .o_statfs = filter_statfs,
4600 .o_getattr = filter_getattr,
4601 .o_unpackmd = filter_unpackmd,
4602 .o_create = filter_create,
4603 .o_setattr = filter_setattr,
4604 .o_destroy = filter_destroy,
4605 .o_brw = filter_brw,
4606 .o_punch = filter_truncate,
4607 .o_sync = filter_sync,
4608 .o_preprw = filter_preprw,
4609 .o_commitrw = filter_commitrw,
4610 .o_llog_init = filter_llog_init,
4611 .o_llog_connect = filter_llog_connect,
4612 .o_llog_finish = filter_llog_finish,
4613 .o_iocontrol = filter_iocontrol,
4614 .o_health_check = filter_health_check,
4615 .o_process_config = filter_process_config,
4618 quota_interface_t *filter_quota_interface_ref;
4619 extern quota_interface_t filter_quota_interface;
4621 static int __init obdfilter_init(void)
4623 struct lprocfs_static_vars lvars;
4626 /** sanity check for group<->mdsno conversion */
4627 for (i = 0; i < MAX_MDT_COUNT; i++)
4628 LASSERT(objseq_to_mdsno(mdt_to_obd_objseq(i)) == i);
4630 lprocfs_filter_init_vars(&lvars);
4632 cfs_request_module("%s", "lquota");
4633 OBD_ALLOC(obdfilter_created_scratchpad,
4634 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4635 sizeof(*obdfilter_created_scratchpad));
4636 if (obdfilter_created_scratchpad == NULL)
4639 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4640 sizeof(struct filter_mod_data),
4643 GOTO(out, rc = -ENOMEM);
4645 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4646 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4648 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4649 LUSTRE_OST_NAME, NULL);
4653 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4654 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4655 ll_fmd_cachep = NULL;
4657 if (filter_quota_interface_ref)
4658 PORTAL_SYMBOL_PUT(filter_quota_interface);
4660 OBD_FREE(obdfilter_created_scratchpad,
4661 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4662 sizeof(*obdfilter_created_scratchpad));
4668 static void __exit obdfilter_exit(void)
4670 if (filter_quota_interface_ref)
4671 PORTAL_SYMBOL_PUT(filter_quota_interface);
4673 if (ll_fmd_cachep) {
4674 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4675 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4676 ll_fmd_cachep = NULL;
4679 class_unregister_type(LUSTRE_OST_NAME);
4680 OBD_FREE(obdfilter_created_scratchpad,
4681 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4682 sizeof(*obdfilter_created_scratchpad));
4685 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4686 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4687 MODULE_LICENSE("GPL");
4689 module_init(obdfilter_init);
4690 module_exit(obdfilter_exit);