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 2008 Sun Microsystems, Inc. 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 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
83 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS1_N_BASE
84 static struct lvfs_callback_ops filter_lvfs_ops;
85 cfs_mem_cache_t *ll_fmd_cachep;
87 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
88 void *cb_data, int error)
90 struct obd_export *exp = cb_data;
91 LASSERT(exp->exp_obd == obd);
92 obd_transno_commit_cb(obd, transno, exp, error);
93 atomic_dec(&exp->exp_cb_count);
94 class_export_put(exp);
97 int filter_version_get_check(struct obd_export *exp,
98 struct obd_trans_info *oti, struct inode *inode)
102 if (inode == NULL || oti == NULL)
105 curr_version = fsfilt_get_version(exp->exp_obd, inode);
106 if ((__s64)curr_version == -EOPNOTSUPP)
108 /* VBR: version is checked always because costs nothing */
109 if (oti->oti_pre_version != 0 &&
110 oti->oti_pre_version != curr_version) {
111 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
112 oti->oti_pre_version, curr_version);
113 spin_lock(&exp->exp_lock);
114 exp->exp_vbr_failed = 1;
115 spin_unlock(&exp->exp_lock);
118 oti->oti_pre_version = curr_version;
122 /* Assumes caller has already pushed us into the kernel context. */
123 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
124 struct obd_trans_info *oti, int rc, int force_sync)
126 struct filter_obd *filter = &exp->exp_obd->u.filter;
127 struct filter_export_data *fed = &exp->exp_filter_data;
128 struct lsd_client_data *lcd = fed->fed_lcd;
131 int err, log_pri = D_RPCTRACE;
133 /* Propagate error code. */
137 if (!exp->exp_obd->obd_replayable || oti == NULL)
140 /* we don't allocate new transnos for replayed requests */
141 spin_lock(&filter->fo_translock);
142 if (oti->oti_transno == 0) {
143 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
144 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
146 last_rcvd = oti->oti_transno;
147 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
148 filter->fo_fsd->lsd_last_transno =
149 cpu_to_le64(last_rcvd);
151 oti->oti_transno = last_rcvd;
152 if (last_rcvd <= le64_to_cpu(lcd->lcd_last_transno)) {
153 spin_unlock(&filter->fo_translock);
156 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
157 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
158 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
159 spin_unlock(&filter->fo_translock);
162 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
164 off = fed->fed_lr_off;
166 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
167 fed->fed_lr_idx, fed->fed_lr_off);
170 class_export_get(exp); /* released when the cb is called */
171 atomic_inc(&exp->exp_cb_count);
173 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
179 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
180 lcd, sizeof(*lcd), &off,
181 force_sync | exp->exp_need_sync);
183 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
191 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
192 last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
197 void f_dput(struct dentry *dentry)
199 /* Can't go inside filter_ddelete because it can block */
200 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
201 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
202 LASSERT(atomic_read(&dentry->d_count) > 0);
207 static void init_brw_stats(struct brw_stats *brw_stats)
210 for (i = 0; i < BRW_LAST; i++)
211 spin_lock_init(&brw_stats->hist[i].oh_lock);
214 static int lprocfs_init_rw_stats(struct obd_device *obd,
215 struct lprocfs_stats **stats)
219 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
220 LPROC_FILTER_LAST - 1;
221 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
225 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
226 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
227 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
228 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
229 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
234 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
235 plus the procfs overhead :( */
236 static int filter_export_stats_init(struct obd_device *obd,
237 struct obd_export *exp,
243 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
244 /* Self-export gets no proc entry */
247 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
249 /* Mask error for already created
257 struct nid_stat *tmp = exp->exp_nid_stats;
258 LASSERT(tmp != NULL);
260 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
261 if (tmp->nid_brw_stats == NULL)
264 init_brw_stats(tmp->nid_brw_stats);
265 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
266 0644, &filter_per_nid_stats_fops,
269 CWARN("Error adding the brw_stats file\n");
271 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
275 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
279 /* Always add in ldlm_stats */
280 tmp->nid_ldlm_stats =
281 lprocfs_alloc_stats(LDLM_LAST_OPC - LDLM_FIRST_OPC,
282 LPROCFS_STATS_FLAG_NOPERCPU);
283 if (tmp->nid_ldlm_stats == NULL)
286 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
287 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
288 tmp->nid_ldlm_stats);
296 /* Add client data to the FILTER. We use a bitmap to locate a free space
297 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
298 * Otherwise, we have just read the data from the last_rcvd file and
299 * we know its offset. */
300 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
303 struct filter_obd *filter = &obd->u.filter;
304 struct filter_export_data *fed = &exp->exp_filter_data;
305 unsigned long *bitmap = filter->fo_last_rcvd_slots;
306 int new_client = (cl_idx == -1);
310 LASSERT(bitmap != NULL);
311 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
314 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
317 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
318 * there's no need for extra complication here
321 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
323 if (cl_idx >= LR_MAX_CLIENTS) {
324 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
328 if (test_and_set_bit(cl_idx, bitmap)) {
329 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
334 if (test_and_set_bit(cl_idx, bitmap)) {
335 CERROR("FILTER client %d: bit already set in bitmap!\n",
341 fed->fed_lr_idx = cl_idx;
342 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
343 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
344 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
346 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
347 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
350 struct lvfs_run_ctxt saved;
351 loff_t off = fed->fed_lr_off;
355 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
356 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
358 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
359 /* Transaction needed to fix bug 1403 */
360 handle = fsfilt_start(obd,
361 filter->fo_rcvd_filp->f_dentry->d_inode,
362 FSFILT_OP_SETATTR, NULL);
363 if (IS_ERR(handle)) {
364 rc = PTR_ERR(handle);
365 CERROR("unable to start transaction: rc %d\n", rc);
367 fed->fed_lcd->lcd_last_epoch =
368 filter->fo_fsd->lsd_start_epoch;
369 exp->exp_last_request_time = cfs_time_current_sec();
370 rc = fsfilt_add_journal_cb(obd, 0, handle,
371 target_client_add_cb, exp);
373 spin_lock(&exp->exp_lock);
374 exp->exp_need_sync = 1;
375 spin_unlock(&exp->exp_lock);
377 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
379 sizeof(*fed->fed_lcd),
380 &off, rc /* sync if no cb */);
382 filter->fo_rcvd_filp->f_dentry->d_inode,
385 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
388 CERROR("error writing %s client idx %u: rc %d\n",
389 LAST_RCVD, fed->fed_lr_idx, rc);
396 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
398 static int filter_client_free(struct obd_export *exp)
400 struct filter_export_data *fed = &exp->exp_filter_data;
401 struct filter_obd *filter = &exp->exp_obd->u.filter;
402 struct obd_device *obd = exp->exp_obd;
403 struct lvfs_run_ctxt saved;
408 if (fed->fed_lcd == NULL)
411 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
412 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
415 LASSERT(filter->fo_last_rcvd_slots != NULL);
417 off = fed->fed_lr_off;
419 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
420 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
422 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
423 * we leak a client slot that will be cleaned on the next recovery. */
425 CERROR("%s: client idx %d has med_off %lld\n",
426 obd->obd_name, fed->fed_lr_idx, off);
427 GOTO(free, rc = -EINVAL);
430 /* Clear the bit _after_ zeroing out the client so we don't
431 race with filter_client_add and zero out new clients.*/
432 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
433 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
438 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
439 /* Don't force sync on disconnect if aborting recovery,
440 * or it does num_clients * num_osts. b=17194 */
441 int need_sync = exp->exp_need_sync &&
442 !(exp->exp_flags&OBD_OPT_ABORT_RECOV);
443 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
444 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
445 sizeof(zero_lcd), &off, 0);
447 /* Make sure the server's last_transno is up to date. Do this
448 * after the client is freed so we know all the client's
449 * transactions have been committed. */
451 filter_update_server_data(obd, filter->fo_rcvd_filp,
452 filter->fo_fsd, need_sync);
453 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
455 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
456 "zero out client %s at idx %u/%llu in %s %ssync rc %d\n",
457 fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
458 LAST_RCVD, need_sync ? "" : "a", rc);
461 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
462 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
469 OBD_FREE_PTR(fed->fed_lcd);
475 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
476 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
477 struct filter_mod_data *fmd)
479 LASSERT_SPIN_LOCKED(&fed->fed_lock);
480 if (--fmd->fmd_refcount == 0) {
481 /* XXX when we have persistent reservations and the handle
482 * is stored herein we need to drop it here. */
483 fed->fed_mod_count--;
484 list_del(&fmd->fmd_list);
485 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
489 /* drop fmd reference, free it if last ref */
490 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
492 struct filter_export_data *fed;
497 fed = &exp->exp_filter_data;
498 spin_lock(&fed->fed_lock);
499 filter_fmd_put_nolock(fed, fmd); /* caller reference */
500 spin_unlock(&fed->fed_lock);
503 /* expire entries from the end of the list if there are too many
504 * or they are too old */
505 static void filter_fmd_expire_nolock(struct filter_obd *filter,
506 struct filter_export_data *fed,
507 struct filter_mod_data *keep)
509 struct filter_mod_data *fmd, *tmp;
511 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
515 if (time_before(jiffies, fmd->fmd_expire) &&
516 fed->fed_mod_count < filter->fo_fmd_max_num)
519 list_del_init(&fmd->fmd_list);
520 filter_fmd_put_nolock(fed, fmd); /* list reference */
524 void filter_fmd_expire(struct obd_export *exp)
526 spin_lock(&exp->exp_filter_data.fed_lock);
527 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
528 &exp->exp_filter_data, NULL);
529 spin_unlock(&exp->exp_filter_data.fed_lock);
532 /* find specified objid, group in export fmd list.
533 * caller must hold fed_lock and take fmd reference itself */
534 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
535 struct filter_export_data *fed,
536 obd_id objid, obd_gr group)
538 struct filter_mod_data *found = NULL, *fmd;
540 LASSERT_SPIN_LOCKED(&fed->fed_lock);
542 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
543 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
545 list_del(&fmd->fmd_list);
546 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
547 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
552 filter_fmd_expire_nolock(filter, fed, found);
557 /* Find fmd based on objid and group, or return NULL if not found. */
558 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
559 obd_id objid, obd_gr group)
561 struct filter_mod_data *fmd;
563 spin_lock(&exp->exp_filter_data.fed_lock);
564 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
565 &exp->exp_filter_data, objid, group);
567 fmd->fmd_refcount++; /* caller reference */
568 spin_unlock(&exp->exp_filter_data.fed_lock);
573 /* Find fmd based on objid and group, or create a new one if none is found.
574 * It is possible for this function to return NULL under memory pressure,
575 * or if objid = 0 is passed (which will only cause old entries to expire).
576 * Currently this is not fatal because any fmd state is transient and
577 * may also be freed when it gets sufficiently old. */
578 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
579 obd_id objid, obd_gr group)
581 struct filter_export_data *fed = &exp->exp_filter_data;
582 struct filter_mod_data *found = NULL, *fmd_new = NULL;
584 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
586 spin_lock(&fed->fed_lock);
587 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
590 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
591 fmd_new->fmd_id = objid;
592 fmd_new->fmd_gr = group;
593 fmd_new->fmd_refcount++; /* list reference */
595 fed->fed_mod_count++;
597 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
601 found->fmd_refcount++; /* caller reference */
602 found->fmd_expire = jiffies +
603 exp->exp_obd->u.filter.fo_fmd_max_age;
606 spin_unlock(&fed->fed_lock);
612 /* drop fmd list reference so it will disappear when last reference is put.
613 * This isn't so critical because it would in fact only affect the one client
614 * that is doing the unlink and at worst we have an stale entry referencing
615 * an object that should never be used again. */
616 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
618 struct filter_mod_data *found = NULL;
620 spin_lock(&exp->exp_filter_data.fed_lock);
621 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
623 list_del_init(&found->fmd_list);
624 filter_fmd_put_nolock(&exp->exp_filter_data, found);
626 spin_unlock(&exp->exp_filter_data.fed_lock);
629 #define filter_fmd_drop(exp, objid, group)
632 /* remove all entries from fmd list */
633 static void filter_fmd_cleanup(struct obd_export *exp)
635 struct filter_export_data *fed = &exp->exp_filter_data;
636 struct filter_mod_data *fmd = NULL, *tmp;
638 spin_lock(&fed->fed_lock);
639 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
640 list_del_init(&fmd->fmd_list);
641 filter_fmd_put_nolock(fed, fmd);
643 spin_unlock(&fed->fed_lock);
646 static int filter_init_export(struct obd_export *exp)
648 spin_lock_init(&exp->exp_filter_data.fed_lock);
649 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
651 spin_lock(&exp->exp_lock);
652 exp->exp_connecting = 1;
653 spin_unlock(&exp->exp_lock);
655 return ldlm_init_export(exp);
658 static int filter_free_server_data(struct filter_obd *filter)
660 OBD_FREE_PTR(filter->fo_fsd);
661 filter->fo_fsd = NULL;
662 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
663 filter->fo_last_rcvd_slots = NULL;
667 /* assumes caller is already in kernel ctxt */
668 int filter_update_server_data(struct obd_device *obd, struct file *filp,
669 struct lr_server_data *fsd, int force_sync)
675 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
676 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
677 le64_to_cpu(fsd->lsd_last_transno));
678 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
679 le64_to_cpu(fsd->lsd_mount_count));
681 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
683 CERROR("error writing lr_server_data: rc = %d\n", rc);
688 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
691 struct filter_obd *filter = &obd->u.filter;
697 if (filter->fo_last_objid_files[group] == NULL) {
698 CERROR("Object group "LPU64" not fully setup; not updating "
699 "last_objid\n", group);
703 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
704 obd->obd_name, group, filter->fo_last_objids[group]);
706 tmp = cpu_to_le64(filter->fo_last_objids[group]);
707 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
708 &tmp, sizeof(tmp), &off, force_sync);
710 CERROR("error writing group "LPU64" last objid: rc = %d\n",
714 extern int ost_handle(struct ptlrpc_request *req);
715 /* assumes caller has already in kernel ctxt */
716 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
718 struct filter_obd *filter = &obd->u.filter;
719 struct lr_server_data *fsd;
720 struct lsd_client_data *lcd = NULL;
721 struct inode *inode = filp->f_dentry->d_inode;
722 unsigned long last_rcvd_size = i_size_read(inode);
729 /* ensure padding in the struct is the correct size */
730 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
731 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
732 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
733 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
735 OBD_ALLOC(fsd, sizeof(*fsd));
738 filter->fo_fsd = fsd;
740 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
741 if (filter->fo_last_rcvd_slots == NULL) {
742 OBD_FREE(fsd, sizeof(*fsd));
746 if (last_rcvd_size == 0) {
747 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
749 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
750 fsd->lsd_last_transno = 0;
751 mount_count = fsd->lsd_mount_count = 0;
752 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
753 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
754 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
755 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
756 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
757 /* OBD_COMPAT_OST is set in filter_connect_internal when the
758 * MDS first connects and assigns the OST index number. */
759 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
762 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
764 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
768 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
769 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
770 "using the wrong disk %s. Were the "
771 "/dev/ assignments rearranged?\n",
772 obd->obd_uuid.uuid, fsd->lsd_uuid);
773 GOTO(err_fsd, rc = -EINVAL);
775 mount_count = le64_to_cpu(fsd->lsd_mount_count);
776 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
778 /* Assume old last_rcvd format unless I_C_LR is set */
779 if (!(fsd->lsd_feature_incompat &
780 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
781 fsd->lsd_last_transno = fsd->lsd_compat14;
783 /* OBD_COMPAT_OST is set in filter_connect_internal when the
784 * MDS first connects and assigns the OST index number. */
785 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
789 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
790 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
791 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
792 ~FILTER_INCOMPAT_SUPP);
793 GOTO(err_fsd, rc = -EINVAL);
795 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
796 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
797 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
798 ~FILTER_ROCOMPAT_SUPP);
799 /* Do something like remount filesystem read-only */
800 GOTO(err_fsd, rc = -EINVAL);
803 start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
805 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
806 obd->obd_name, start_epoch);
807 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
808 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
809 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
810 obd->obd_name, mount_count + 1);
811 CDEBUG(D_INODE, "%s: server data size: %u\n",
812 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
813 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
814 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
815 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
816 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
817 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
818 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
819 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
820 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
821 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
822 le16_to_cpu(fsd->lsd_client_size));
824 if (!obd->obd_replayable) {
825 CWARN("%s: recovery support OFF\n", obd->obd_name);
829 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
830 off < last_rcvd_size; cl_idx++) {
832 struct obd_export *exp;
833 struct filter_export_data *fed;
838 GOTO(err_client, rc = -ENOMEM);
841 /* Don't assume off is incremented properly by
842 * fsfilt_read_record(), in case sizeof(*lcd)
843 * isn't the same as fsd->lsd_client_size. */
844 off = le32_to_cpu(fsd->lsd_client_start) +
845 cl_idx * le16_to_cpu(fsd->lsd_client_size);
846 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
848 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
849 LAST_RCVD, cl_idx, off, rc);
850 break; /* read error shouldn't cause startup to fail */
853 if (lcd->lcd_uuid[0] == '\0') {
854 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
859 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
861 /* These exports are cleaned up by filter_disconnect(), so they
862 * need to be set up like real exports as filter_connect() does.
864 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
866 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
867 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
868 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
870 if (PTR_ERR(exp) == -EALREADY) {
871 /* export already exists, zero out this one */
872 CERROR("Zeroing out duplicate export due to "
874 lcd->lcd_uuid[0] = '\0';
876 GOTO(err_client, rc = PTR_ERR(exp));
879 fed = &exp->exp_filter_data;
881 fed->fed_group = 0; /* will be assigned at connect */
882 filter_export_stats_init(obd, exp, NULL);
883 rc = filter_client_add(obd, exp, cl_idx);
884 /* can't fail for existing client */
885 LASSERTF(rc == 0, "rc = %d\n", rc);
887 /* VBR: set export last committed */
888 exp->exp_last_committed = last_rcvd;
889 spin_lock(&exp->exp_lock);
890 exp->exp_connecting = 0;
891 exp->exp_in_recovery = 0;
892 spin_unlock(&exp->exp_lock);
893 spin_lock_bh(&obd->obd_processing_task_lock);
894 obd->obd_max_recoverable_clients++;
895 spin_unlock_bh(&obd->obd_processing_task_lock);
897 class_export_put(exp);
900 /* Need to check last_rcvd even for duplicated exports. */
901 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
904 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
905 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
911 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
913 filter->fo_mount_count = mount_count + 1;
914 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
916 /* save it, so mount count and last_transno is current */
917 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
919 GOTO(err_client, rc);
924 class_disconnect_exports(obd);
926 filter_free_server_data(filter);
930 static int filter_cleanup_groups(struct obd_device *obd)
932 struct filter_obd *filter = &obd->u.filter;
934 struct dentry *dentry;
938 if (filter->fo_dentry_O_groups != NULL) {
939 for (i = 0; i < filter->fo_group_count; i++) {
940 dentry = filter->fo_dentry_O_groups[i];
944 OBD_FREE(filter->fo_dentry_O_groups,
945 filter->fo_group_count *
946 sizeof(*filter->fo_dentry_O_groups));
947 filter->fo_dentry_O_groups = NULL;
949 if (filter->fo_last_objid_files != NULL) {
950 for (i = 0; i < filter->fo_group_count; i++) {
951 filp = filter->fo_last_objid_files[i];
955 OBD_FREE(filter->fo_last_objid_files,
956 filter->fo_group_count *
957 sizeof(*filter->fo_last_objid_files));
958 filter->fo_last_objid_files = NULL;
960 if (filter->fo_dentry_O_sub != NULL) {
961 for (i = 0; i < filter->fo_group_count; i++) {
962 for (j = 0; j < filter->fo_subdir_count; j++) {
963 dentry = filter->fo_dentry_O_sub[i].dentry[j];
968 OBD_FREE(filter->fo_dentry_O_sub,
969 filter->fo_group_count *
970 sizeof(*filter->fo_dentry_O_sub));
971 filter->fo_dentry_O_sub = NULL;
973 if (filter->fo_last_objids != NULL) {
974 OBD_FREE(filter->fo_last_objids,
975 filter->fo_group_count *
976 sizeof(*filter->fo_last_objids));
977 filter->fo_last_objids = NULL;
979 if (filter->fo_dentry_O != NULL) {
980 f_dput(filter->fo_dentry_O);
981 filter->fo_dentry_O = NULL;
986 static int filter_update_last_group(struct obd_device *obd, int group)
988 struct filter_obd *filter = &obd->u.filter;
989 struct file *filp = NULL;
990 int last_group = 0, rc;
994 if (group <= filter->fo_committed_group)
997 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1001 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1005 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1007 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1010 LASSERTF(off == 0 || CHECK_MDS_GROUP(last_group),
1011 "off = %llu and last_group = %d\n", off, last_group);
1013 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1014 obd->obd_name, last_group, group);
1018 /* must be sync: bXXXX */
1019 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1021 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1025 filter->fo_committed_group = group;
1028 filp_close(filp, 0);
1032 static int filter_read_group_internal(struct obd_device *obd, int group,
1035 struct filter_obd *filter = &obd->u.filter;
1036 __u64 *new_objids = NULL;
1037 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1038 struct dentry **new_groups = NULL;
1039 struct file **new_files = NULL;
1040 struct dentry *dentry;
1042 int old_count = filter->fo_group_count, rc, stage = 0, i;
1046 int len = group + 1;
1048 snprintf(name, 24, "%d", group);
1052 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1054 if (IS_ERR(dentry)) {
1055 CERROR("Cannot lookup expected object group %d: %ld\n",
1056 group, PTR_ERR(dentry));
1057 RETURN(PTR_ERR(dentry));
1060 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1062 if (IS_ERR(dentry)) {
1063 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1065 RETURN(PTR_ERR(dentry));
1070 snprintf(name, 24, "O/%d/LAST_ID", group);
1072 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1074 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1075 GOTO(cleanup, rc = PTR_ERR(filp));
1079 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1081 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1085 if (filter->fo_subdir_count) {
1086 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1087 if (tmp_subdirs == NULL)
1088 GOTO(cleanup, rc = -ENOMEM);
1091 for (i = 0; i < filter->fo_subdir_count; i++) {
1093 snprintf(dir, sizeof(dir), "d%u", i);
1095 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1098 if (IS_ERR(tmp_subdirs->dentry[i])) {
1099 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1100 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1105 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1106 tmp_subdirs->dentry[i]);
1110 /* 'group' is an index; we need an array of length 'group + 1' */
1111 if (group + 1 > old_count) {
1112 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1113 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1114 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1115 OBD_ALLOC(new_files, len * sizeof(*new_files));
1117 if (new_objids == NULL || new_subdirs == NULL ||
1118 new_groups == NULL || new_files == NULL)
1119 GOTO(cleanup, rc = -ENOMEM);
1122 memcpy(new_objids, filter->fo_last_objids,
1123 old_count * sizeof(*new_objids));
1124 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1125 old_count * sizeof(*new_subdirs));
1126 memcpy(new_groups, filter->fo_dentry_O_groups,
1127 old_count * sizeof(*new_groups));
1128 memcpy(new_files, filter->fo_last_objid_files,
1129 old_count * sizeof(*new_files));
1131 OBD_FREE(filter->fo_last_objids,
1132 old_count * sizeof(*new_objids));
1133 OBD_FREE(filter->fo_dentry_O_sub,
1134 old_count * sizeof(*new_subdirs));
1135 OBD_FREE(filter->fo_dentry_O_groups,
1136 old_count * sizeof(*new_groups));
1137 OBD_FREE(filter->fo_last_objid_files,
1138 old_count * sizeof(*new_files));
1140 filter->fo_last_objids = new_objids;
1141 filter->fo_dentry_O_sub = new_subdirs;
1142 filter->fo_dentry_O_groups = new_groups;
1143 filter->fo_last_objid_files = new_files;
1144 filter->fo_group_count = len;
1147 filter->fo_dentry_O_groups[group] = dentry;
1148 filter->fo_last_objid_files[group] = filp;
1149 if (filter->fo_subdir_count) {
1150 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1151 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1154 filter_update_last_group(obd, group);
1156 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1157 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1158 rc = filter_update_last_objid(obd, group, 1);
1162 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1163 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1164 obd->obd_name, group, last_objid);
1169 if (new_objids != NULL)
1170 OBD_FREE(new_objids, len * sizeof(*new_objids));
1171 if (new_subdirs != NULL)
1172 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1173 if (new_groups != NULL)
1174 OBD_FREE(new_groups, len * sizeof(*new_groups));
1175 if (new_files != NULL)
1176 OBD_FREE(new_files, len * sizeof(*new_files));
1178 if (filter->fo_subdir_count) {
1179 for (i = 0; i < filter->fo_subdir_count; i++) {
1180 if (tmp_subdirs->dentry[i] != NULL)
1181 dput(tmp_subdirs->dentry[i]);
1183 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1186 filp_close(filp, 0);
1193 static int filter_read_groups(struct obd_device *obd, int last_group,
1196 struct filter_obd *filter = &obd->u.filter;
1197 int old_count, group, rc = 0;
1199 down(&filter->fo_init_lock);
1200 old_count = filter->fo_group_count;
1201 for (group = old_count; group <= last_group; group++) {
1203 rc = filter_read_group_internal(obd, group, create);
1207 up(&filter->fo_init_lock);
1211 /* FIXME: object groups */
1212 static int filter_prep_groups(struct obd_device *obd)
1214 struct filter_obd *filter = &obd->u.filter;
1215 struct dentry *dentry, *O_dentry;
1217 int last_group, rc = 0, cleanup_phase = 0;
1221 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1223 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1224 if (IS_ERR(O_dentry)) {
1225 rc = PTR_ERR(O_dentry);
1226 CERROR("cannot open/create O: rc = %d\n", rc);
1229 filter->fo_dentry_O = O_dentry;
1230 cleanup_phase = 1; /* O_dentry */
1232 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1233 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1234 * some time post 1.0 when all old-style OSTs have converted along
1235 * with the init_objid hack. */
1236 dentry = ll_lookup_one_len("R", O_dentry, 1);
1238 GOTO(cleanup, rc = PTR_ERR(dentry));
1239 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1240 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1243 CWARN("converting OST to new object layout\n");
1244 if (IS_ERR(O0_dentry)) {
1245 rc = PTR_ERR(O0_dentry);
1246 CERROR("error looking up O/0: rc %d\n", rc);
1247 GOTO(cleanup_R, rc);
1250 if (O0_dentry->d_inode) {
1251 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1252 GOTO(cleanup_O0, rc = -EEXIST);
1255 LOCK_INODE_MUTEX(O_dentry->d_inode);
1256 rc = ll_vfs_rename(O_dentry->d_inode, dentry, filter->fo_vfsmnt,
1257 O_dentry->d_inode, O0_dentry,
1259 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1262 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1263 GOTO(cleanup_O0, rc);
1265 filter->fo_fsd->lsd_feature_incompat |=
1266 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1267 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1269 GOTO(cleanup_O0, rc);
1281 cleanup_phase = 2; /* groups */
1283 /* we have to initialize all groups before first connections from
1284 * clients because they may send create/destroy for any group -bzzz */
1285 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1287 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1288 GOTO(cleanup, rc = PTR_ERR(filp));
1290 cleanup_phase = 3; /* filp */
1292 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1294 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1298 last_group = FILTER_MIN_GROUPS;
1300 LASSERT_MDS_GROUP(last_group);
1303 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1304 FILTER_MIN_GROUPS, last_group);
1305 filter->fo_committed_group = last_group;
1306 rc = filter_read_groups(obd, last_group, 1);
1310 filp_close(filp, 0);
1314 switch (cleanup_phase) {
1316 filp_close(filp, 0);
1318 filter_cleanup_groups(obd);
1320 f_dput(filter->fo_dentry_O);
1321 filter->fo_dentry_O = NULL;
1329 /* setup the object store with correct subdirectories */
1330 static int filter_prep(struct obd_device *obd)
1332 struct lvfs_run_ctxt saved;
1333 struct filter_obd *filter = &obd->u.filter;
1335 struct inode *inode;
1339 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1340 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1341 if (!file || IS_ERR(file)) {
1343 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1347 filter->fo_rcvd_filp = file;
1348 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1349 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1350 file->f_dentry->d_inode->i_mode);
1351 GOTO(err_filp, rc = -ENOENT);
1354 inode = file->f_dentry->d_parent->d_inode;
1355 /* We use i_op->unlink directly in filter_vfs_unlink() */
1356 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1357 CERROR("%s: filesystem does not support create/unlink ops\n",
1359 GOTO(err_filp, rc = -EOPNOTSUPP);
1362 /** lu_target has very limited use in filter now */
1363 lut_init(NULL, &filter->fo_lut, obd, NULL);
1365 rc = filter_init_server_data(obd, file);
1367 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1371 target_recovery_init(&filter->fo_lut, ost_handle);
1373 /* open and create health check io file*/
1374 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1377 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1379 GOTO(err_server_data, rc);
1381 filter->fo_health_check_filp = file;
1382 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1383 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1384 file->f_dentry->d_inode->i_mode);
1385 GOTO(err_health_check, rc = -ENOENT);
1387 rc = lvfs_check_io_health(obd, file);
1389 GOTO(err_health_check, rc);
1391 rc = filter_prep_groups(obd);
1393 GOTO(err_health_check, rc);
1395 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1400 if (filp_close(filter->fo_health_check_filp, 0))
1401 CERROR("can't close %s after error\n", HEALTH_CHECK);
1402 filter->fo_health_check_filp = NULL;
1404 target_recovery_fini(obd);
1405 filter_free_server_data(filter);
1407 if (filp_close(filter->fo_rcvd_filp, 0))
1408 CERROR("can't close %s after error\n", LAST_RCVD);
1409 filter->fo_rcvd_filp = NULL;
1413 /* cleanup the filter: write last used object id to status file */
1414 static void filter_post(struct obd_device *obd)
1416 struct lvfs_run_ctxt saved;
1417 struct filter_obd *filter = &obd->u.filter;
1420 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1421 * best to start a transaction with h_sync, because we removed this
1424 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1425 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1428 CERROR("error writing server data: rc = %d\n", rc);
1430 for (i = 0; i < filter->fo_group_count; i++) {
1431 rc = filter_update_last_objid(obd, i,
1432 (i == filter->fo_group_count - 1));
1434 CERROR("error writing group %d lastobjid: rc = %d\n",
1438 rc = filp_close(filter->fo_rcvd_filp, 0);
1439 filter->fo_rcvd_filp = NULL;
1441 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1443 rc = filp_close(filter->fo_health_check_filp, 0);
1444 filter->fo_health_check_filp = NULL;
1446 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1448 filter_cleanup_groups(obd);
1449 filter_free_server_data(filter);
1450 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1452 filter_free_capa_keys(filter);
1453 cleanup_capa_hash(filter->fo_capa_hash);
1456 static void filter_set_last_id(struct filter_obd *filter,
1457 obd_id id, obd_gr group)
1459 LASSERT(filter->fo_fsd != NULL);
1460 LASSERT(group <= filter->fo_group_count);
1462 spin_lock(&filter->fo_objidlock);
1463 filter->fo_last_objids[group] = id;
1464 spin_unlock(&filter->fo_objidlock);
1467 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1470 LASSERT(filter->fo_fsd != NULL);
1471 LASSERT(group <= filter->fo_group_count);
1473 /* FIXME: object groups */
1474 spin_lock(&filter->fo_objidlock);
1475 id = filter->fo_last_objids[group];
1476 spin_unlock(&filter->fo_objidlock);
1480 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1482 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1486 /* We never dget the object parent, so DON'T dput it either */
1487 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1489 struct filter_obd *filter = &obd->u.filter;
1490 struct filter_subdirs *subdirs;
1491 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1493 if ((group > FILTER_GROUP_MDS0 && group < FILTER_GROUP_MDS1_N_BASE) ||
1494 filter->fo_subdir_count == 0)
1495 return filter->fo_dentry_O_groups[group];
1497 subdirs = &filter->fo_dentry_O_sub[group];
1498 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1501 /* We never dget the object parent, so DON'T dput it either */
1502 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1505 unsigned long now = jiffies;
1506 struct dentry *dparent = filter_parent(obd, group, objid);
1509 if (IS_ERR(dparent))
1511 if (dparent == NULL)
1512 return ERR_PTR(-ENOENT);
1514 rc = filter_lock_dentry(obd, dparent);
1515 fsfilt_check_slow(obd, now, "parent lock");
1516 return rc ? ERR_PTR(rc) : dparent;
1519 /* We never dget the object parent, so DON'T dput it either */
1520 static void filter_parent_unlock(struct dentry *dparent)
1522 UNLOCK_INODE_MUTEX(dparent->d_inode);
1525 /* How to get files, dentries, inodes from object id's.
1527 * If dir_dentry is passed, the caller has already locked the parent
1528 * appropriately for this operation (normally a write lock). If
1529 * dir_dentry is NULL, we do a read lock while we do the lookup to
1530 * avoid races with create/destroy and such changing the directory
1531 * internal to the filesystem code. */
1532 struct dentry *filter_fid2dentry(struct obd_device *obd,
1533 struct dentry *dir_dentry,
1534 obd_gr group, obd_id id)
1536 struct dentry *dparent = dir_dentry;
1537 struct dentry *dchild;
1542 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1543 obd->u.filter.fo_destroys_in_progress == 0) {
1544 /* don't fail lookups for orphan recovery, it causes
1545 * later LBUGs when objects still exist during precreate */
1546 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1547 RETURN(ERR_PTR(-ENOENT));
1550 CERROR("fatal: invalid object id 0\n");
1551 RETURN(ERR_PTR(-ESTALE));
1554 len = sprintf(name, LPU64, id);
1555 if (dir_dentry == NULL) {
1556 dparent = filter_parent_lock(obd, group, id);
1557 if (IS_ERR(dparent)) {
1558 CERROR("%s: error getting object "LPU64":"LPU64
1559 " parent: rc %ld\n", obd->obd_name,
1560 id, group, PTR_ERR(dparent));
1564 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1565 dparent->d_name.len, dparent->d_name.name, name);
1566 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1567 if (dir_dentry == NULL)
1568 filter_parent_unlock(dparent);
1569 if (IS_ERR(dchild)) {
1570 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1575 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1576 CERROR("%s: got bad object "LPU64" inode %lu\n",
1577 obd->obd_name, id, dchild->d_inode->i_ino);
1579 RETURN(ERR_PTR(-ENOENT));
1582 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1583 name, dchild, atomic_read(&dchild->d_count));
1585 LASSERT(atomic_read(&dchild->d_count) > 0);
1590 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1591 obd_id group, struct lustre_handle *lockh)
1593 int flags = LDLM_AST_DISCARD_DATA, rc;
1594 struct ldlm_res_id res_id;
1595 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1598 osc_build_res_name(objid, group, &res_id);
1599 /* Tell the clients that the object is gone now and that they should
1600 * throw away any cached pages. */
1601 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1602 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1603 ldlm_completion_ast, NULL, NULL, 0, NULL,
1610 static void filter_fini_destroy(struct obd_device *obd,
1611 struct lustre_handle *lockh)
1614 ldlm_lock_decref(lockh, LCK_PW);
1617 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1618 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1619 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1620 * reverse. Caller must take i_sem before starting the transaction and we
1621 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1622 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1623 struct vfsmount *mnt)
1628 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1629 * itself we already hold dir->i_mutex for child create/unlink ops */
1630 LASSERT(dentry->d_inode != NULL);
1631 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1632 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1636 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1637 GOTO(out, rc = -ENOENT);
1639 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1644 GOTO(out, rc = -EPERM);
1646 /* check_sticky() */
1647 if ((dentry->d_inode->i_uid != current->fsuid &&
1648 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1649 IS_IMMUTABLE(dentry->d_inode))
1650 GOTO(out, rc = -EPERM);
1652 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1653 * that was done because of journal_start (which is already done
1654 * here) or some other ordering issue. */
1657 rc = ll_security_inode_unlink(dir, dentry, mnt);
1661 rc = dir->i_op->unlink(dir, dentry);
1663 /* need to drop i_mutex before we lose inode reference */
1664 UNLOCK_INODE_MUTEX(dentry->d_inode);
1671 /* Caller must hold LCK_PW on parent and push us into kernel context.
1672 * Caller must hold child i_mutex, we drop it always.
1673 * Caller is also required to ensure that dchild->d_inode exists. */
1674 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1675 obd_gr group, struct dentry *dparent,
1676 struct dentry *dchild)
1678 struct inode *inode = dchild->d_inode;
1679 struct filter_obd *filter = &obd->u.filter;
1682 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1683 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1684 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1685 (unsigned long)inode->i_nlink,
1686 atomic_read(&inode->i_count));
1689 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1691 CERROR("error unlinking objid %.*s: rc %d\n",
1692 dchild->d_name.len, dchild->d_name.name, rc);
1696 struct filter_intent_args {
1697 struct ldlm_lock **victim;
1702 static enum interval_iter filter_intent_cb(struct interval_node *n,
1705 struct ldlm_interval *node = (struct ldlm_interval *)n;
1706 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1707 __u64 size = arg->size;
1708 struct ldlm_lock **v = arg->victim;
1709 struct ldlm_lock *lck;
1711 /* If the interval is lower than the current file size,
1713 if (interval_high(n) <= size)
1714 return INTERVAL_ITER_STOP;
1716 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1717 /* Don't send glimpse ASTs to liblustre clients.
1718 * They aren't listening for them, and they do
1719 * entirely synchronous I/O anyways. */
1720 if (lck->l_export == NULL ||
1721 lck->l_export->exp_libclient == 1)
1724 if (*arg->liblustre)
1725 *arg->liblustre = 0;
1728 *v = LDLM_LOCK_GET(lck);
1729 } else if ((*v)->l_policy_data.l_extent.start <
1730 lck->l_policy_data.l_extent.start) {
1731 LDLM_LOCK_RELEASE(*v);
1732 *v = LDLM_LOCK_GET(lck);
1735 /* the same policy group - every lock has the
1736 * same extent, so needn't do it any more */
1740 return INTERVAL_ITER_CONT;
1743 static int filter_intent_policy(struct ldlm_namespace *ns,
1744 struct ldlm_lock **lockp, void *req_cookie,
1745 ldlm_mode_t mode, int flags, void *data)
1747 CFS_LIST_HEAD(rpc_list);
1748 struct ptlrpc_request *req = req_cookie;
1749 struct ldlm_lock *lock = *lockp, *l = NULL;
1750 struct ldlm_resource *res = lock->l_resource;
1751 ldlm_processing_policy policy;
1752 struct ost_lvb *res_lvb, *reply_lvb;
1753 struct ldlm_reply *rep;
1755 int idx, rc, tmpflags = 0, only_liblustre = 1;
1756 struct ldlm_interval_tree *tree;
1757 struct filter_intent_args arg;
1758 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1759 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1760 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1763 policy = ldlm_get_processing_policy(res);
1764 LASSERT(policy != NULL);
1765 LASSERT(req != NULL);
1767 rc = lustre_pack_reply(req, 3, repsize, NULL);
1769 RETURN(req->rq_status = rc);
1771 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1772 LASSERT(rep != NULL);
1774 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1775 sizeof(*reply_lvb));
1776 LASSERT(reply_lvb != NULL);
1778 //fixup_handle_for_resent_req(req, lock, &lockh);
1780 /* Call the extent policy function to see if our request can be
1781 * granted, or is blocked.
1782 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1783 * lock, and should not be granted if the lock will be blocked.
1786 LASSERT(ns == res->lr_namespace);
1788 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1789 check_res_locked(res);
1791 /* FIXME: we should change the policy function slightly, to not make
1792 * this list at all, since we just turn around and free it */
1793 while (!list_empty(&rpc_list)) {
1794 struct ldlm_lock *wlock =
1795 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1796 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1797 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1798 lock->l_flags &= ~LDLM_FL_CP_REQD;
1799 list_del_init(&wlock->l_cp_ast);
1800 LDLM_LOCK_RELEASE(wlock);
1803 /* The lock met with no resistance; we're finished. */
1804 if (rc == LDLM_ITER_CONTINUE) {
1805 /* do not grant locks to the liblustre clients: they cannot
1806 * handle ASTs robustly. We need to do this while still
1807 * holding ns_lock to avoid the lock remaining on the res_link
1808 * list (and potentially being added to l_pending_list by an
1809 * AST) when we are going to drop this lock ASAP. */
1810 if (lock->l_export->exp_libclient ||
1811 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1812 ldlm_resource_unlink_lock(lock);
1813 err = ELDLM_LOCK_ABORTED;
1815 err = ELDLM_LOCK_REPLACED;
1821 /* Do not grant any lock, but instead send GL callbacks. The extent
1822 * policy nicely created a list of all PW locks for us. We will choose
1823 * the highest of those which are larger than the size in the LVB, if
1824 * any, and perform a glimpse callback. */
1825 res_lvb = res->lr_lvb_data;
1826 LASSERT(res_lvb != NULL);
1827 *reply_lvb = *res_lvb;
1830 * ->ns_lock guarantees that no new locks are granted, and,
1831 * therefore, that res->lr_lvb_data cannot increase beyond the
1832 * end of already granted lock. As a result, it is safe to
1833 * check against "stale" reply_lvb->lvb_size value without
1836 arg.size = reply_lvb->lvb_size;
1838 arg.liblustre = &only_liblustre;
1839 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1840 tree = &res->lr_itree[idx];
1841 if (tree->lit_mode == LCK_PR)
1844 interval_iterate_reverse(tree->lit_root,
1845 filter_intent_cb, &arg);
1849 /* There were no PW locks beyond the size in the LVB; finished. */
1851 if (only_liblustre) {
1852 /* If we discovered a liblustre client with a PW lock,
1853 * however, the LVB may be out of date! The LVB is
1854 * updated only on glimpse (which we don't do for
1855 * liblustre clients) and cancel (which the client
1856 * obviously has not yet done). So if it has written
1857 * data but kept the lock, the LVB is stale and needs
1858 * to be updated from disk.
1860 * Of course, this will all disappear when we switch to
1861 * taking liblustre locks on the OST. */
1862 ldlm_res_lvbo_update(res, NULL, 0, 1);
1864 RETURN(ELDLM_LOCK_ABORTED);
1868 * This check is for lock taken in filter_prepare_destroy() that does
1869 * not have l_glimpse_ast set. So the logic is: if there is a lock
1870 * with no l_glimpse_ast set, this object is being destroyed already.
1872 * Hence, if you are grabbing DLM locks on the server, always set
1873 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1875 if (l->l_glimpse_ast == NULL) {
1876 /* We are racing with unlink(); just return -ENOENT */
1877 rep->lock_policy_res1 = -ENOENT;
1881 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1882 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1883 /* Update the LVB from disk if the AST failed (this is a legal race) */
1885 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1886 * sending ast is not handled. This can result in lost client writes.
1889 ldlm_res_lvbo_update(res, NULL, 0, 1);
1892 *reply_lvb = *res_lvb;
1896 LDLM_LOCK_RELEASE(l);
1898 RETURN(ELDLM_LOCK_ABORTED);
1902 * per-obd_device iobuf pool.
1904 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1905 * allocations in write-path has to be minimized (see bug 5137).
1907 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1908 * OST threads (see ost_thread_{init,done}()).
1910 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1911 * at the OST layer there are only (potentially) multiple obd_device of type
1912 * unknown at the time of OST thread creation.
1914 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1915 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1916 * it's very own iobuf.
1920 * filter_kiobuf_pool_init()
1922 * filter_kiobuf_pool_done()
1924 * filter_iobuf_get()
1926 * operate on this array. They are "generic" in a sense that they don't depend
1927 * on actual type of iobuf's (the latter depending on Linux kernel version).
1931 * destroy pool created by filter_iobuf_pool_init
1933 static void filter_iobuf_pool_done(struct filter_obd *filter)
1935 struct filter_iobuf **pool;
1940 pool = filter->fo_iobuf_pool;
1942 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1943 if (pool[i] != NULL)
1944 filter_free_iobuf(pool[i]);
1946 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1947 filter->fo_iobuf_pool = NULL;
1952 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1954 struct filter_obd *filter = &obd->u.filter;
1955 struct sptlrpc_rule_set tmp_rset;
1958 sptlrpc_rule_set_init(&tmp_rset);
1959 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1961 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1966 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1968 write_lock(&filter->fo_sptlrpc_lock);
1969 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1970 filter->fo_sptlrpc_rset = tmp_rset;
1971 write_unlock(&filter->fo_sptlrpc_lock);
1977 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1979 static int filter_iobuf_pool_init(struct filter_obd *filter)
1986 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1988 if (filter->fo_iobuf_pool == NULL)
1991 filter->fo_iobuf_count = OSS_THREADS_MAX;
1996 /* Return iobuf allocated for @thread_id. We don't know in advance how
1997 * many threads there will be so we allocate a large empty array and only
1998 * fill in those slots that are actually in use.
1999 * If we haven't allocated a pool entry for this thread before, do so now. */
2000 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
2002 int thread_id = (oti && oti->oti_thread) ?
2003 oti->oti_thread->t_id : -1;
2004 struct filter_iobuf *pool = NULL;
2005 struct filter_iobuf **pool_place = NULL;
2007 if (thread_id >= 0) {
2008 LASSERT(thread_id < filter->fo_iobuf_count);
2009 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
2012 if (unlikely(pool == NULL)) {
2013 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
2014 PTLRPC_MAX_BRW_PAGES);
2015 if (pool_place != NULL)
2022 /* mount the file system (secretly). lustre_cfg parameters are:
2025 * 3 = flags: failover=f, failout=n
2028 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
2031 struct filter_obd *filter = &obd->u.filter;
2032 struct vfsmount *mnt;
2033 struct lustre_mount_info *lmi;
2034 struct obd_uuid uuid;
2042 if (lcfg->lcfg_bufcount < 3 ||
2043 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
2044 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
2047 lmi = server_get_mount(obd->obd_name);
2049 /* We already mounted in lustre_fill_super.
2050 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2051 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2053 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2055 /* old path - used by lctl */
2056 CERROR("Using old MDS mount method\n");
2057 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2058 MS_NOATIME|MS_NODIRATIME,
2059 lustre_cfg_string(lcfg, 1), option);
2062 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2063 lustre_cfg_string(lcfg, 1), rc);
2067 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2069 if (IS_ERR(obd->obd_fsops))
2070 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2072 rc = filter_iobuf_pool_init(filter);
2076 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2077 CERROR("%s: Underlying device is marked as read-only. "
2078 "Setup failed\n", obd->obd_name);
2079 GOTO(err_ops, rc = -EROFS);
2082 /* failover is the default */
2083 obd->obd_replayable = 1;
2085 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2086 str = lustre_cfg_string(lcfg, 3);
2087 if (strchr(str, 'n')) {
2088 CWARN("%s: recovery disabled\n", obd->obd_name);
2089 obd->obd_replayable = 0;
2093 filter->fo_vfsmnt = mnt;
2094 obd->u.obt.obt_sb = mnt->mnt_sb;
2095 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2096 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2098 fsfilt_setup(obd, obd->u.obt.obt_sb);
2100 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2101 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2102 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2103 obd->obd_lvfs_ctxt.fs = get_ds();
2104 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2106 init_mutex(&filter->fo_init_lock);
2107 filter->fo_committed_group = 0;
2108 filter->fo_destroys_in_progress = 0;
2109 for (i = 0; i < 32; i++)
2110 sema_init(&filter->fo_create_locks[i], 1);
2112 spin_lock_init(&filter->fo_translock);
2113 spin_lock_init(&filter->fo_objidlock);
2114 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2115 sema_init(&filter->fo_alloc_lock, 1);
2116 init_brw_stats(&filter->fo_filter_stats);
2117 filter->fo_read_cache = 1; /* enable read-only cache by default */
2118 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2119 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2120 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2121 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2123 rc = filter_prep(obd);
2127 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2128 spin_lock_init(&filter->fo_llog_list_lock);
2130 filter->fo_fl_oss_capa = 1;
2132 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2133 filter->fo_capa_hash = init_capa_hash();
2134 if (filter->fo_capa_hash == NULL)
2135 GOTO(err_post, rc = -ENOMEM);
2137 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2138 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2139 LDLM_NAMESPACE_GREEDY);
2140 if (obd->obd_namespace == NULL)
2141 GOTO(err_post, rc = -ENOMEM);
2142 obd->obd_namespace->ns_lvbp = obd;
2143 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2144 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2146 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2147 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2149 rc = obd_llog_init(obd, &obd->obd_olg, obd, 1, NULL, NULL);
2151 CERROR("failed to setup llogging subsystems\n");
2155 rwlock_init(&filter->fo_sptlrpc_lock);
2156 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2157 /* do this after llog being initialized */
2158 filter_adapt_sptlrpc_conf(obd, 1);
2160 rc = lquota_setup(filter_quota_interface_ref, obd);
2164 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2165 if (q->max_sectors < q->max_hw_sectors &&
2166 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2167 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2168 "for larger I/O requests: max_sectors = %u "
2169 "could be up to max_hw_sectors=%u\n",
2170 obd->obd_name, mnt->mnt_sb->s_id,
2171 q->max_sectors, q->max_hw_sectors);
2173 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2174 if (uuid_ptr != NULL) {
2175 class_uuid_unparse(uuid_ptr, &uuid);
2181 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2183 if (obd->obd_recovering) {
2184 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2185 "recovery for at least %d:%.02d, or until %d "
2186 "client%s reconnect%s.\n",
2187 obd->obd_name, lustre_cfg_string(lcfg, 1),
2188 label ?: "", label ? "/" : "", str,
2189 obd->obd_recovery_timeout / 60,
2190 obd->obd_recovery_timeout % 60,
2191 obd->obd_max_recoverable_clients,
2192 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2193 (obd->obd_max_recoverable_clients == 1) ? "s":"");
2195 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2196 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2197 label ?: "", label ? "/" : "", str,
2198 obd->obd_replayable ? "enabled" : "disabled");
2206 fsfilt_put_ops(obd->obd_fsops);
2207 filter_iobuf_pool_done(filter);
2209 server_put_mount(obd->obd_name, mnt);
2210 obd->u.obt.obt_sb = 0;
2214 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2216 struct lprocfs_static_vars lvars;
2221 CLASSERT(offsetof(struct obd_device, u.obt) ==
2222 offsetof(struct obd_device, u.filter.fo_obt));
2224 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2227 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2228 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2231 addr = (unsigned long)cfs_page_address(page);
2232 clear_page((void *)addr);
2234 /* lprocfs must be setup before the filter so state can be safely added
2235 * to /proc incrementally as the filter is setup */
2236 lprocfs_filter_init_vars(&lvars);
2237 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2238 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2239 /* Init obdfilter private stats here */
2240 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2241 LPROCFS_CNTR_AVGMINMAX,
2242 "read_bytes", "bytes");
2243 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2244 LPROCFS_CNTR_AVGMINMAX,
2245 "write_bytes", "bytes");
2246 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2247 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2248 "get_page", "usec");
2249 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2250 LPROCFS_CNTR_AVGMINMAX,
2251 "get_page_failures", "num");
2252 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2253 LPROCFS_CNTR_AVGMINMAX,
2254 "cache_access", "pages");
2255 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2256 LPROCFS_CNTR_AVGMINMAX,
2257 "cache_hit", "pages");
2258 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2259 LPROCFS_CNTR_AVGMINMAX,
2260 "cache_miss", "pages");
2262 lproc_filter_attach_seqstat(obd);
2263 obd->obd_proc_exports_entry = lprocfs_register("exports",
2264 obd->obd_proc_entry,
2266 if (IS_ERR(obd->obd_proc_exports_entry)) {
2267 rc = PTR_ERR(obd->obd_proc_exports_entry);
2268 CERROR("error %d setting up lprocfs for %s\n",
2270 obd->obd_proc_exports_entry = NULL;
2273 if (obd->obd_proc_exports_entry)
2274 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2275 lprocfs_nid_stats_clear_read,
2276 lprocfs_nid_stats_clear_write, obd, NULL);
2278 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2279 LUSTRE_CFG_BUFLEN(lcfg, 4));
2280 rc = filter_common_setup(obd, lcfg, (void *)addr);
2281 OBD_PAGE_FREE(page);
2284 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2285 lprocfs_free_per_client_stats(obd);
2286 lprocfs_free_obd_stats(obd);
2287 lprocfs_obd_cleanup(obd);
2293 static struct llog_operations filter_mds_ost_repl_logops;
2295 static struct llog_operations filter_size_orig_logops = {
2296 .lop_setup = llog_obd_origin_setup,
2297 .lop_cleanup = llog_obd_origin_cleanup,
2298 .lop_add = llog_obd_origin_add
2301 static int filter_olg_fini(struct obd_llog_group *olg)
2303 struct llog_ctxt *ctxt;
2304 int rc = 0, rc2 = 0;
2307 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2309 rc = llog_cleanup(ctxt);
2311 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2313 rc2 = llog_cleanup(ctxt);
2318 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2320 rc2 = llog_cleanup(ctxt);
2329 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2330 struct obd_device *tgt)
2335 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2336 &filter_mds_ost_repl_logops);
2340 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2341 &filter_size_orig_logops);
2347 filter_olg_fini(olg);
2352 * Init the default olg, which is embeded in the obd_device, for filter.
2355 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2356 struct obd_device *tgt)
2358 struct filter_obd *filter = &obd->u.filter;
2359 struct llog_ctxt *ctxt;
2363 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2364 if (!filter->fo_lcm)
2367 filter_mds_ost_repl_logops = llog_client_ops;
2368 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2369 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2370 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2372 rc = filter_olg_init(obd, olg, tgt);
2374 GOTO(cleanup_lcm, rc);
2376 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2379 GOTO(cleanup_olg, rc);
2381 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2383 CERROR("Can't get ctxt for %p:%x\n", olg,
2384 LLOG_MDS_OST_REPL_CTXT);
2385 GOTO(cleanup_olg, rc = -ENODEV);
2387 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2388 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2389 llog_ctxt_put(ctxt);
2393 filter_olg_fini(olg);
2395 llog_recov_thread_fini(filter->fo_lcm, 1);
2396 filter->fo_lcm = NULL;
2401 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2402 struct obd_device *tgt, int count, struct llog_catid *catid,
2403 struct obd_uuid *uuid)
2405 struct filter_obd *filter = &obd->u.filter;
2406 struct llog_ctxt *ctxt;
2410 LASSERT(olg != NULL);
2411 if (olg == &obd->obd_olg)
2412 return filter_default_olg_init(obd, olg, tgt);
2414 LASSERT(filter->fo_lcm != NULL);
2415 rc = filter_olg_init(obd, olg, tgt);
2418 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2420 CERROR("Can't get ctxt for %p:%x\n", olg,
2421 LLOG_MDS_OST_REPL_CTXT);
2422 filter_olg_fini(olg);
2425 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2426 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2427 llog_ctxt_put(ctxt);
2431 static int filter_llog_finish(struct obd_device *obd, int count)
2433 struct filter_obd *filter = &obd->u.filter;
2434 struct llog_ctxt *ctxt;
2437 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2440 * Make sure that no cached llcds left in recov_thread.
2441 * We actually do sync in disconnect time, but disconnect
2442 * may not come being marked rq_no_resend = 1.
2444 llog_sync(ctxt, NULL);
2447 * Balance class_import_get() in llog_receptor_accept().
2448 * This is safe to do, as llog is already synchronized
2449 * and its import may go.
2451 mutex_down(&ctxt->loc_sem);
2452 if (ctxt->loc_imp) {
2453 class_import_put(ctxt->loc_imp);
2454 ctxt->loc_imp = NULL;
2456 mutex_up(&ctxt->loc_sem);
2457 llog_ctxt_put(ctxt);
2460 if (filter->fo_lcm) {
2461 mutex_down(&ctxt->loc_sem);
2462 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2463 filter->fo_lcm = NULL;
2464 mutex_up(&ctxt->loc_sem);
2466 RETURN(filter_olg_fini(&obd->obd_olg));
2470 * Find the group llog according to group index in the llog group list.
2472 static struct obd_llog_group *
2473 filter_find_olg_internal(struct filter_obd *filter, int group)
2475 struct obd_llog_group *olg;
2477 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2478 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2479 if (olg->olg_group == group)
2486 * Find the group llog according to group index on the filter
2488 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2490 struct obd_llog_group *olg = NULL;
2491 struct filter_obd *filter;
2493 filter = &obd->u.filter;
2495 if (group == FILTER_GROUP_LLOG)
2496 RETURN(&obd->obd_olg);
2498 spin_lock(&filter->fo_llog_list_lock);
2499 olg = filter_find_olg_internal(filter, group);
2500 spin_unlock(&filter->fo_llog_list_lock);
2505 * Find the llog_group of the filter according to the group. If it can not
2506 * find, create the llog_group, which only happens when mds is being synced
2509 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2511 struct obd_llog_group *olg = NULL;
2512 struct filter_obd *filter;
2515 filter = &obd->u.filter;
2517 if (group == FILTER_GROUP_LLOG)
2518 RETURN(&obd->obd_olg);
2520 spin_lock(&filter->fo_llog_list_lock);
2521 olg = filter_find_olg_internal(filter, group);
2523 if (olg->olg_initializing) {
2524 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2526 GOTO(out_unlock, olg);
2531 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2533 llog_group_init(olg, group);
2534 list_add(&olg->olg_list, &filter->fo_llog_list);
2535 olg->olg_initializing = 1;
2536 spin_unlock(&filter->fo_llog_list_lock);
2538 rc = llog_cat_initialize(obd, olg, 1, NULL);
2540 spin_lock(&filter->fo_llog_list_lock);
2541 list_del(&olg->olg_list);
2542 spin_unlock(&filter->fo_llog_list_lock);
2544 GOTO(out, olg = ERR_PTR(-ENOMEM));
2546 spin_lock(&filter->fo_llog_list_lock);
2547 olg->olg_initializing = 0;
2548 spin_unlock(&filter->fo_llog_list_lock);
2549 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2550 obd->obd_name, group, olg);
2555 spin_unlock(&filter->fo_llog_list_lock);
2559 static int filter_llog_connect(struct obd_export *exp,
2560 struct llogd_conn_body *body)
2562 struct obd_device *obd = exp->exp_obd;
2563 struct llog_ctxt *ctxt;
2564 struct obd_llog_group *olg;
2568 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2569 obd->obd_name, body->lgdc_logid.lgl_oid,
2570 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2572 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2574 CERROR(" %s: can not find olg of group %d\n",
2575 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2578 llog_group_set_export(olg, exp);
2580 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2581 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2582 body->lgdc_ctxt_idx);
2584 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2585 obd->obd_name, body->lgdc_logid.lgl_oid,
2586 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2588 spin_lock_bh(&obd->obd_processing_task_lock);
2589 obd->u.filter.fo_mds_ost_sync = 1;
2590 spin_unlock_bh(&obd->obd_processing_task_lock);
2591 rc = llog_connect(ctxt, &body->lgdc_logid,
2592 &body->lgdc_gen, NULL);
2593 llog_ctxt_put(ctxt);
2595 CERROR("failed to connect rc %d idx %d\n", rc,
2596 body->lgdc_ctxt_idx);
2601 static int filter_llog_preclean(struct obd_device *obd)
2603 struct obd_llog_group *olg, *tmp;
2604 struct filter_obd *filter;
2605 struct list_head remove_list;
2609 rc = obd_llog_finish(obd, 0);
2611 CERROR("failed to cleanup llogging subsystem\n");
2613 filter = &obd->u.filter;
2614 CFS_INIT_LIST_HEAD(&remove_list);
2616 spin_lock(&filter->fo_llog_list_lock);
2617 while (!list_empty(&filter->fo_llog_list)) {
2618 olg = list_entry(filter->fo_llog_list.next,
2619 struct obd_llog_group, olg_list);
2620 list_del(&olg->olg_list);
2621 list_add(&olg->olg_list, &remove_list);
2623 spin_unlock(&filter->fo_llog_list_lock);
2625 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2626 list_del_init(&olg->olg_list);
2627 rc = filter_olg_fini(olg);
2629 CERROR("failed to cleanup llogging subsystem for %u\n",
2637 static int filter_precleanup(struct obd_device *obd,
2638 enum obd_cleanup_stage stage)
2644 case OBD_CLEANUP_EARLY:
2646 case OBD_CLEANUP_EXPORTS:
2647 /* Stop recovery before namespace cleanup. */
2648 target_recovery_fini(obd);
2649 rc = filter_llog_preclean(obd);
2655 static int filter_cleanup(struct obd_device *obd)
2657 struct filter_obd *filter = &obd->u.filter;
2661 LCONSOLE_WARN("%s: shutting down for failover; client state "
2662 "will be preserved.\n", obd->obd_name);
2664 obd_exports_barrier(obd);
2665 obd_zombie_barrier();
2667 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2668 lprocfs_free_per_client_stats(obd);
2669 lprocfs_free_obd_stats(obd);
2670 lprocfs_obd_cleanup(obd);
2671 lquota_cleanup(filter_quota_interface_ref, obd);
2673 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2674 obd->obd_namespace = NULL;
2676 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2678 if (obd->u.obt.obt_sb == NULL)
2683 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2684 shrink_dcache_sb(obd->u.obt.obt_sb);
2686 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2687 obd->u.obt.obt_sb = NULL;
2689 fsfilt_put_ops(obd->obd_fsops);
2691 filter_iobuf_pool_done(filter);
2693 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2698 static int filter_connect_internal(struct obd_export *exp,
2699 struct obd_connect_data *data)
2701 struct filter_export_data *fed = &exp->exp_filter_data;
2706 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2707 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2708 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2709 data->ocd_connect_flags, data->ocd_version,
2710 data->ocd_grant, data->ocd_index);
2712 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2713 CWARN("!!! This export (nid %s) used object group %d "
2714 "earlier; now it's trying to use group %d! This could "
2715 "be a bug in the MDS. Please report to "
2716 "http://bugzilla.lustre.org/\n",
2717 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2720 fed->fed_group = data->ocd_group;
2722 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2723 exp->exp_connect_flags = data->ocd_connect_flags;
2724 data->ocd_version = LUSTRE_VERSION_CODE;
2726 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2727 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2728 data->ocd_connect_flags & OBD_CONNECT_MDS))
2731 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2732 struct filter_obd *filter = &exp->exp_obd->u.filter;
2733 obd_size left, want;
2735 spin_lock(&exp->exp_obd->obd_osfs_lock);
2736 left = filter_grant_space_left(exp);
2737 want = data->ocd_grant;
2738 filter_grant(exp, fed->fed_grant, want, left);
2739 data->ocd_grant = fed->fed_grant;
2740 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2742 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2743 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2744 exp->exp_client_uuid.uuid, exp,
2745 data->ocd_grant, want, left);
2747 filter->fo_tot_granted_clients ++;
2750 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2751 struct filter_obd *filter = &exp->exp_obd->u.filter;
2752 struct lr_server_data *lsd = filter->fo_fsd;
2753 int index = le32_to_cpu(lsd->lsd_ost_index);
2755 if (!(lsd->lsd_feature_compat &
2756 cpu_to_le32(OBD_COMPAT_OST))) {
2757 /* this will only happen on the first connect */
2758 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2759 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2760 filter_update_server_data(exp->exp_obd,
2761 filter->fo_rcvd_filp, lsd, 1);
2762 } else if (index != data->ocd_index) {
2763 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2764 " %u doesn't match actual OST index"
2765 " %u in last_rcvd file, bad "
2767 obd_export_nid2str(exp), index,
2771 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2772 * FIXME: We don't strictly need the COMPAT flag for that,
2773 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2774 * FIXME: We needed it for the index, as index 0 is valid. */
2777 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2778 data->ocd_brw_size = 65536;
2779 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2780 data->ocd_brw_size = min(data->ocd_brw_size,
2781 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2782 LASSERT(data->ocd_brw_size);
2785 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2786 __u32 cksum_types = data->ocd_cksum_types;
2788 /* The client set in ocd_cksum_types the checksum types it
2789 * supports. We have to mask off the algorithms that we don't
2791 if (cksum_types & OBD_CKSUM_ALL)
2792 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2794 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2796 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2797 "%x\n", exp->exp_obd->obd_name,
2798 obd_export_nid2str(exp), cksum_types,
2799 data->ocd_cksum_types);
2801 /* This client does not support OBD_CONNECT_CKSUM
2802 * fall back to CRC32 */
2803 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2804 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2805 exp->exp_obd->obd_name,
2806 obd_export_nid2str(exp));
2812 static int filter_reconnect(const struct lu_env *env,
2813 struct obd_export *exp, struct obd_device *obd,
2814 struct obd_uuid *cluuid,
2815 struct obd_connect_data *data,
2821 if (exp == NULL || obd == NULL || cluuid == NULL)
2824 rc = filter_connect_internal(exp, data);
2826 filter_export_stats_init(obd, exp, localdata);
2831 /* nearly identical to mds_connect */
2832 static int filter_connect(const struct lu_env *env,
2833 struct obd_export **exp, struct obd_device *obd,
2834 struct obd_uuid *cluuid,
2835 struct obd_connect_data *data, void *localdata)
2837 struct lvfs_run_ctxt saved;
2838 struct lustre_handle conn = { 0 };
2839 struct obd_export *lexp;
2840 struct filter_export_data *fed;
2841 struct lsd_client_data *lcd = NULL;
2846 if (exp == NULL || obd == NULL || cluuid == NULL)
2849 rc = class_connect(&conn, obd, cluuid);
2852 lexp = class_conn2export(&conn);
2853 LASSERT(lexp != NULL);
2855 fed = &lexp->exp_filter_data;
2857 rc = filter_connect_internal(lexp, data);
2861 filter_export_stats_init(obd, lexp, localdata);
2862 if (obd->obd_replayable) {
2863 OBD_ALLOC(lcd, sizeof(*lcd));
2865 CERROR("filter: out of memory for client data\n");
2866 GOTO(cleanup, rc = -ENOMEM);
2869 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2871 rc = filter_client_add(obd, lexp, -1);
2876 group = data->ocd_group;
2878 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2879 obd->obd_name, lexp->exp_handle.h_cookie, group);
2881 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2882 rc = filter_read_groups(obd, group, 1);
2883 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2885 CERROR("can't read group %u\n", group);
2895 fed->fed_lcd = NULL;
2897 class_disconnect(lexp);
2906 /* Do extra sanity checks for grant accounting. We do this at connect,
2907 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2908 * always get rid of it or turn it off when we know accounting is good. */
2909 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2911 struct filter_export_data *fed;
2912 struct obd_export *exp;
2913 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2914 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2915 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2917 if (list_empty(&obd->obd_exports))
2920 /* We don't want to do this for large machines that do lots of
2921 mounts or unmounts. It burns... */
2922 if (obd->obd_num_exports > 100)
2925 spin_lock(&obd->obd_osfs_lock);
2926 spin_lock(&obd->obd_dev_lock);
2927 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2929 fed = &exp->exp_filter_data;
2930 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2933 if (maxsize > 0) { /* we may not have done a statfs yet */
2934 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2935 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2936 exp->exp_client_uuid.uuid, exp,
2937 fed->fed_grant, fed->fed_pending, maxsize);
2938 LASSERTF(fed->fed_dirty <= maxsize,
2939 "%s: cli %s/%p %ld > "LPU64"\n", func,
2940 exp->exp_client_uuid.uuid, exp,
2941 fed->fed_dirty, maxsize);
2944 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2945 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2946 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2948 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2949 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2950 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2951 tot_granted += fed->fed_grant + fed->fed_pending;
2952 tot_pending += fed->fed_pending;
2953 tot_dirty += fed->fed_dirty;
2955 fo_tot_granted = obd->u.filter.fo_tot_granted;
2956 fo_tot_pending = obd->u.filter.fo_tot_pending;
2957 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2958 spin_unlock(&obd->obd_dev_lock);
2959 spin_unlock(&obd->obd_osfs_lock);
2961 /* Do these assertions outside the spinlocks so we don't kill system */
2962 if (tot_granted != fo_tot_granted)
2963 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2964 func, tot_granted, fo_tot_granted);
2965 if (tot_pending != fo_tot_pending)
2966 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2967 func, tot_pending, fo_tot_pending);
2968 if (tot_dirty != fo_tot_dirty)
2969 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2970 func, tot_dirty, fo_tot_dirty);
2971 if (tot_pending > tot_granted)
2972 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2973 func, tot_pending, tot_granted);
2974 if (tot_granted > maxsize)
2975 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2976 func, tot_granted, maxsize);
2977 if (tot_dirty > maxsize)
2978 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2979 func, tot_dirty, maxsize);
2982 /* Remove this client from the grant accounting totals. We also remove
2983 * the export from the obd device under the osfs and dev locks to ensure
2984 * that the filter_grant_sanity_check() calculations are always valid.
2985 * The client should do something similar when it invalidates its import. */
2986 static void filter_grant_discard(struct obd_export *exp)
2988 struct obd_device *obd = exp->exp_obd;
2989 struct filter_obd *filter = &obd->u.filter;
2990 struct filter_export_data *fed = &exp->exp_filter_data;
2992 spin_lock(&obd->obd_osfs_lock);
2993 spin_lock(&obd->obd_dev_lock);
2994 list_del_init(&exp->exp_obd_chain);
2995 spin_unlock(&obd->obd_dev_lock);
2997 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2998 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2999 obd->obd_name, filter->fo_tot_granted,
3000 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
3001 filter->fo_tot_granted -= fed->fed_grant;
3002 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
3003 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
3004 obd->obd_name, filter->fo_tot_pending,
3005 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
3006 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
3007 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
3008 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
3009 obd->obd_name, filter->fo_tot_dirty,
3010 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
3011 filter->fo_tot_dirty -= fed->fed_dirty;
3015 spin_unlock(&obd->obd_osfs_lock);
3018 static int filter_destroy_export(struct obd_export *exp)
3022 if (exp->exp_filter_data.fed_pending)
3023 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
3024 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
3025 exp, exp->exp_filter_data.fed_pending);
3027 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3029 target_destroy_export(exp);
3030 ldlm_destroy_export(exp);
3032 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
3036 if (exp->exp_obd->obd_replayable)
3037 filter_client_free(exp);
3039 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3041 filter_grant_discard(exp);
3042 filter_fmd_cleanup(exp);
3044 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3045 struct filter_obd *filter = &exp->exp_obd->u.filter;
3046 if (filter->fo_tot_granted_clients > 0)
3047 filter->fo_tot_granted_clients --;
3050 if (!(exp->exp_flags & OBD_OPT_FORCE))
3051 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3056 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3058 struct obd_llog_group *olg_min, *olg;
3059 struct filter_obd *filter;
3060 int worked = -1, group;
3061 struct llog_ctxt *ctxt;
3064 filter = &obd->u.filter;
3066 /* we can't sync log holding spinlock. also, we do not want to get
3067 * into livelock. so we do following: loop over MDS's exports in
3068 * group order and skip already synced llogs -bzzz */
3070 /* look for group with min. number, but > worked */
3073 spin_lock(&filter->fo_llog_list_lock);
3074 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3075 if (olg->olg_group <= worked) {
3076 /* this group is already synced */
3079 if (group < olg->olg_group) {
3080 /* we have group with smaller number to sync */
3083 /* store current minimal group */
3085 group = olg->olg_group;
3087 spin_unlock(&filter->fo_llog_list_lock);
3089 if (olg_min == NULL)
3092 worked = olg_min->olg_group;
3093 if (olg_min->olg_exp &&
3094 (dexp == olg_min->olg_exp || dexp == NULL)) {
3096 ctxt = llog_group_get_ctxt(olg_min,
3097 LLOG_MDS_OST_REPL_CTXT);
3099 err = llog_sync(ctxt, olg_min->olg_exp);
3100 llog_ctxt_put(ctxt);
3102 CERROR("error flushing logs to MDS: "
3107 } while (olg_min != NULL);
3110 /* Also incredibly similar to mds_disconnect */
3111 static int filter_disconnect(struct obd_export *exp)
3113 struct obd_device *obd = exp->exp_obd;
3118 class_export_get(exp);
3120 if (!(exp->exp_flags & OBD_OPT_FORCE))
3121 filter_grant_sanity_check(obd, __FUNCTION__);
3122 filter_grant_discard(exp);
3124 /* Flush any remaining cancel messages out to the target */
3125 filter_sync_llogs(obd, exp);
3127 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3129 /* Disconnect early so that clients can't keep using export */
3130 rc = class_disconnect(exp);
3131 if (exp->exp_obd->obd_namespace != NULL)
3132 ldlm_cancel_locks_for_export(exp);
3134 fsfilt_sync(obd, obd->u.obt.obt_sb);
3136 lprocfs_exp_cleanup(exp);
3138 class_export_put(exp);
3142 /* reverse import is changed, sync all cancels */
3143 static void filter_revimp_update(struct obd_export *exp)
3148 class_export_get(exp);
3150 /* flush any remaining cancel messages out to the target */
3151 filter_sync_llogs(exp->exp_obd, exp);
3152 class_export_put(exp);
3156 static int filter_ping(struct obd_export *exp)
3158 filter_fmd_expire(exp);
3162 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3163 const char *what, int quiet)
3165 struct dentry *dchild = NULL;
3168 if (oa->o_valid & OBD_MD_FLGROUP)
3171 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3173 if (IS_ERR(dchild)) {
3174 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3175 what, group, oa->o_id);
3179 if (dchild->d_inode == NULL) {
3181 CERROR("%s: %s on non-existent object: "LPU64"\n",
3182 obd->obd_name, what, oa->o_id);
3184 RETURN(ERR_PTR(-ENOENT));
3190 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3192 struct dentry *dentry = NULL;
3193 struct obd_device *obd;
3197 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3198 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3202 obd = class_exp2obd(exp);
3204 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3208 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3210 RETURN(PTR_ERR(dentry));
3212 /* Limit the valid bits in the return data to what we actually use */
3213 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3214 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3220 /* this should be enabled/disabled in condition to enabled/disabled large
3221 * inodes (fast EAs) in backing store FS. */
3222 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3223 void *handle, struct obdo *oa)
3225 struct obd_device *obd = exp->exp_obd;
3229 if (oa->o_valid & OBD_MD_FLFID) {
3230 struct filter_fid ff;
3232 if (!(oa->o_valid & OBD_MD_FLGROUP))
3234 /* packing fid and converting it to LE for storing into EA.
3235 * Here ->o_stripe_idx should be filled by LOV and rest of
3236 * fields - by client. */
3237 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3238 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3239 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3240 ff.ff_objid = cpu_to_le64(oa->o_id);
3241 ff.ff_group = cpu_to_le64(oa->o_gr);
3243 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3244 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3245 oa->o_generation, oa->o_id, oa->o_gr);
3247 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3249 CERROR("store fid in object failed! rc: %d\n", rc);
3251 CDEBUG(D_HA, "OSS object without fid info!\n");
3257 /* this is called from filter_truncate() until we have filter_punch() */
3258 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3259 struct obdo *oa, struct obd_trans_info *oti)
3261 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3262 struct llog_cookie *fcc = NULL;
3263 struct filter_obd *filter;
3264 int rc, err, sync = 0;
3265 loff_t old_size = 0;
3266 unsigned int ia_valid;
3267 struct inode *inode;
3272 LASSERT(dentry != NULL);
3273 LASSERT(!IS_ERR(dentry));
3275 inode = dentry->d_inode;
3276 LASSERT(inode != NULL);
3278 filter = &exp->exp_obd->u.filter;
3279 iattr_from_obdo(&iattr, oa, oa->o_valid);
3280 ia_valid = iattr.ia_valid;
3282 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3283 OBD_ALLOC(fcc, sizeof(*fcc));
3285 *fcc = oa->o_lcookie;
3287 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3289 /* Filter truncates and writes are serialized by
3290 * i_alloc_sem, see the comment in
3291 * filter_preprw_write.*/
3292 if (ia_valid & ATTR_SIZE)
3293 down_write(&inode->i_alloc_sem);
3294 LOCK_INODE_MUTEX(inode);
3295 old_size = i_size_read(inode);
3298 /* VBR: version recovery check */
3299 rc = filter_version_get_check(exp, oti, inode);
3301 GOTO(out_unlock, rc);
3303 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3304 * then we will accept the UID+GID sent by the client during write for
3305 * initializing the ownership of this inode. We only allow this to
3306 * happen once so clear these bits in setattr. In 2.6 kernels it is
3307 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3308 * the flags that are actually being set. */
3309 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3310 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3311 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3313 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3314 if (!(ia_valid & ATTR_MODE)) {
3315 iattr.ia_mode = inode->i_mode;
3316 iattr.ia_valid |= ATTR_MODE;
3318 iattr.ia_mode &= ~S_ISUID;
3320 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3321 if (!(iattr.ia_valid & ATTR_MODE)) {
3322 iattr.ia_mode = inode->i_mode;
3323 iattr.ia_valid |= ATTR_MODE;
3325 iattr.ia_mode &= ~S_ISGID;
3328 orig_ids[USRQUOTA] = inode->i_uid;
3329 orig_ids[GRPQUOTA] = inode->i_gid;
3330 handle = fsfilt_start_log(exp->exp_obd, inode,
3331 FSFILT_OP_SETATTR, oti, 1);
3333 GOTO(out_unlock, rc = PTR_ERR(handle));
3335 /* update inode EA only once when inode is suid bit marked. As
3336 * on 2.6.x UID and GID may be set separately, we check here
3337 * only one of them to avoid double setting. */
3338 if (inode->i_mode & S_ISUID)
3339 filter_update_fidea(exp, inode, handle, oa);
3341 handle = fsfilt_start(exp->exp_obd, inode,
3342 FSFILT_OP_SETATTR, oti);
3344 GOTO(out_unlock, rc = PTR_ERR(handle));
3346 if (oa->o_valid & OBD_MD_FLFLAGS) {
3347 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3348 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3350 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3352 /* set cancel cookie callback function */
3353 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3354 filter_cancel_cookies_cb,
3358 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3359 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3361 /* The truncate might have used up our transaction credits. Make
3362 * sure we have one left for the last_rcvd update. */
3363 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3365 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3367 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3371 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3373 CERROR("error on commit, err = %d\n", err);
3380 /* For a partial-page truncate flush the page to disk immediately
3381 * to avoid data corruption during direct disk write. b=17397 */
3382 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3383 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3384 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3392 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3393 UNLOCK_INODE_MUTEX(inode);
3394 if (ia_valid & ATTR_SIZE)
3395 up_write(&inode->i_alloc_sem);
3397 OBD_FREE(fcc, sizeof(*fcc));
3399 /* trigger quota release */
3400 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3401 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3402 int rc2 = lquota_adjust(filter_quota_interface_ref,
3403 exp->exp_obd, cur_ids,
3404 orig_ids, rc, FSFILT_OP_SETATTR);
3405 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3406 "filter adjust qunit. (rc:%d)\n", rc2);
3411 /* this is called from filter_truncate() until we have filter_punch() */
3412 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3413 struct obd_trans_info *oti)
3415 struct obdo *oa = oinfo->oi_oa;
3416 struct lustre_capa *capa = oinfo_capa(oinfo);
3417 struct ldlm_res_id res_id;
3418 struct filter_mod_data *fmd;
3419 struct lvfs_run_ctxt saved;
3420 struct filter_obd *filter;
3421 struct ldlm_resource *res;
3422 struct dentry *dentry;
3423 __u64 opc = CAPA_OPC_META_WRITE;
3427 if (oa->o_valid & OBD_FL_TRUNC)
3428 opc |= CAPA_OPC_OSS_TRUNC;
3429 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa, opc);
3433 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3434 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
3439 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3440 /* This would be very bad - accidentally truncating a file when
3441 * changing the time or similar - bug 12203. */
3442 if (oa->o_valid & OBD_MD_FLSIZE &&
3443 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3444 static char mdsinum[48];
3446 if (oa->o_valid & OBD_MD_FLFID)
3447 snprintf(mdsinum, sizeof(mdsinum) - 1,
3448 " of inode "LPU64"/%u", oa->o_fid,
3453 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3455 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3460 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3462 RETURN(PTR_ERR(dentry));
3464 filter = &exp->exp_obd->u.filter;
3465 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3469 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3470 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3471 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3472 fmd->fmd_mactime_xid = oti->oti_xid;
3473 filter_fmd_put(exp, fmd);
3476 /* setting objects attributes (including owner/group) */
3477 rc = filter_setattr_internal(exp, dentry, oa, oti);
3479 GOTO(out_unlock, rc);
3481 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3482 &res_id, LDLM_EXTENT, 0);
3485 LDLM_RESOURCE_ADDREF(res);
3486 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3487 LDLM_RESOURCE_DELREF(res);
3488 ldlm_resource_putref(res);
3491 oa->o_valid = OBD_MD_FLID;
3493 /* Quota release need uid/gid info */
3494 obdo_from_inode(oa, dentry->d_inode,
3495 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3501 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3505 /* XXX identical to osc_unpackmd */
3506 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3507 struct lov_mds_md *lmm, int lmm_bytes)
3513 if (lmm_bytes < sizeof (*lmm)) {
3514 CERROR("lov_mds_md too small: %d, need %d\n",
3515 lmm_bytes, (int)sizeof(*lmm));
3518 /* XXX LOV_MAGIC etc check? */
3520 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3521 CERROR("lov_mds_md: zero lmm_object_id\n");
3526 lsm_size = lov_stripe_md_size(1);
3530 if (*lsmp != NULL && lmm == NULL) {
3531 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3532 OBD_FREE(*lsmp, lsm_size);
3537 if (*lsmp == NULL) {
3538 OBD_ALLOC(*lsmp, lsm_size);
3542 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3543 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3544 OBD_FREE(*lsmp, lsm_size);
3547 loi_init((*lsmp)->lsm_oinfo[0]);
3551 /* XXX zero *lsmp? */
3552 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3553 LASSERT((*lsmp)->lsm_object_id);
3556 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3561 /* caller must hold fo_create_locks[oa->o_gr] */
3562 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3563 struct filter_obd *filter)
3565 struct obdo doa = { 0 }; /* XXX obdo on stack */
3572 LASSERT_MDS_GROUP(oa->o_gr);
3573 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3574 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3576 memset(&doa, 0, sizeof(doa));
3578 doa.o_valid |= OBD_MD_FLGROUP;
3579 doa.o_gr = oa->o_gr;
3580 doa.o_mode = S_IFREG;
3582 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3583 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3584 exp->exp_obd->obd_name, doa.o_gr);
3588 last = filter_last_id(filter, doa.o_gr);
3590 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3592 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3593 exp->exp_obd->obd_name, oa->o_id + 1, last,
3594 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3596 for (id = last; id > oa->o_id; id--) {
3598 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3599 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3600 CEMERG("error destroying precreate objid "LPU64": %d\n",
3603 /* update last_id on disk periodically so that if we restart
3604 * we don't need to re-scan all of the just-deleted objects. */
3605 if ((id & 511) == 0 && !skip_orphan) {
3606 filter_set_last_id(filter, id - 1, doa.o_gr);
3607 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3611 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3612 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3615 filter_set_last_id(filter, id, doa.o_gr);
3616 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3618 /* don't reuse orphan object, return last used objid */
3622 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3627 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3628 obd_gr group, int *num);
3629 /* returns a negative error or a nonnegative number of files to create */
3630 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3631 obd_gr group, struct obd_trans_info *oti)
3633 struct obd_device *obd = exp->exp_obd;
3634 struct filter_obd *filter = &obd->u.filter;
3638 /* delete orphans request */
3639 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3640 obd_id last = filter_last_id(filter, group);
3642 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3643 CERROR("%s: dropping old orphan cleanup request\n",
3647 /* This causes inflight precreates to abort and drop lock */
3648 set_bit(group, &filter->fo_destroys_in_progress);
3649 down(&filter->fo_create_locks[group]);
3650 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3651 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3652 exp->exp_obd->obd_name, group);
3653 up(&filter->fo_create_locks[group]);
3656 diff = oa->o_id - last;
3657 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3660 if (-diff > OST_MAX_PRECREATE) {
3661 CERROR("%s: ignoring bogus orphan destroy request: "
3662 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3664 /* FIXME: should reset precreate_next_id on MDS */
3665 GOTO(out, rc = -EINVAL);
3668 rc = filter_destroy_precreated(exp, oa, filter);
3670 CERROR("%s: unable to write lastobjid, but "
3671 "orphans were deleted\n", obd->obd_name);
3674 /* XXX: Used by MDS for the first time! */
3675 clear_bit(group, &filter->fo_destroys_in_progress);
3678 down(&filter->fo_create_locks[group]);
3679 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3680 CERROR("%s: dropping old precreate request\n",
3684 /* only precreate if group == 0 and o_id is specfied */
3685 if (group == FILTER_GROUP_LLOG || oa->o_id == 0)
3688 diff = oa->o_id - filter_last_id(filter, group);
3689 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3690 filter_last_id(filter, group), diff);
3692 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3693 oa->o_id, filter_last_id(filter, group), diff);
3697 oa->o_id = filter_last_id(&obd->u.filter, group);
3698 rc = filter_precreate(obd, oa, group, &diff);
3699 oa->o_id = filter_last_id(&obd->u.filter, group);
3701 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3704 /* else diff == 0 */
3707 up(&filter->fo_create_locks[group]);
3711 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3712 __u64 max_age, __u32 flags)
3714 struct filter_obd *filter = &obd->u.filter;
3715 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3719 /* at least try to account for cached pages. its still racey and
3720 * might be under-reporting if clients haven't announced their
3721 * caches with brw recently */
3722 spin_lock(&obd->obd_osfs_lock);
3723 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3724 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3725 spin_unlock(&obd->obd_osfs_lock);
3727 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3728 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3729 filter->fo_tot_dirty, filter->fo_tot_granted,
3730 filter->fo_tot_pending,
3731 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3733 filter_grant_sanity_check(obd, __FUNCTION__);
3735 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3736 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3737 osfs->os_bsize - 1) >> blockbits));
3739 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3740 * stop creating files on MDS if OST is not good shape to create
3744 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3745 osfs->os_state = OS_STATE_READONLY;
3747 if (filter->fo_raid_degraded)
3748 osfs->os_state |= OS_STATE_DEGRADED;
3752 static int filter_use_existing_obj(struct obd_device *obd,
3753 struct dentry *dchild, void **handle,
3756 struct inode *inode = dchild->d_inode;
3760 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3763 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3764 if (IS_ERR(*handle))
3765 return PTR_ERR(*handle);
3767 iattr.ia_valid = ATTR_MODE;
3768 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3769 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3776 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3779 __u64 os_ffree = -1;
3781 spin_lock(&obd->obd_osfs_lock);
3782 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3784 os_ffree = obd->obd_osfs.os_ffree;
3785 spin_unlock(&obd->obd_osfs_lock);
3790 /* We rely on the fact that only one thread will be creating files in a given
3791 * group at a time, which is why we don't need an atomic filter_get_new_id.
3792 * Even if we had that atomic function, the following race would exist:
3794 * thread 1: gets id x from filter_next_id
3795 * thread 2: gets id (x + 1) from filter_next_id
3796 * thread 2: creates object (x + 1)
3797 * thread 1: tries to create object x, gets -ENOSPC
3799 * Caller must hold fo_create_locks[group]
3801 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3802 obd_gr group, int *num)
3804 struct dentry *dchild = NULL, *dparent = NULL;
3805 struct filter_obd *filter;
3806 struct obd_statfs *osfs;
3807 int err = 0, rc = 0, recreate_obj = 0, i;
3808 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3811 void *handle = NULL;
3814 filter = &obd->u.filter;
3816 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3818 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3820 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3821 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3824 OBD_ALLOC(osfs, sizeof(*osfs));
3827 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3828 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3829 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3830 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3831 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3835 OBD_FREE(osfs, sizeof(*osfs));
3840 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3841 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3843 for (i = 0; i < *num && err == 0; i++) {
3844 int cleanup_phase = 0;
3846 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3847 CWARN("%s: create aborted by destroy\n",
3856 last_id = filter_last_id(filter, group);
3857 if (next_id > last_id) {
3858 CERROR("Error: Trying to recreate obj greater"
3859 "than last id "LPD64" > "LPD64"\n",
3861 GOTO(cleanup, rc = -EINVAL);
3864 next_id = filter_last_id(filter, group) + 1;
3866 dparent = filter_parent_lock(obd, group, next_id);
3867 if (IS_ERR(dparent))
3868 GOTO(cleanup, rc = PTR_ERR(dparent));
3869 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3871 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3873 GOTO(cleanup, rc = PTR_ERR(dchild));
3874 cleanup_phase = 2; /* f_dput(dchild) */
3876 if (dchild->d_inode != NULL) {
3877 /* This would only happen if lastobjid was bad on disk*/
3878 /* Could also happen if recreating missing obj but it
3879 * already exists. */
3881 CERROR("%s: recreating existing object %.*s?\n",
3882 obd->obd_name, dchild->d_name.len,
3883 dchild->d_name.name);
3885 /* Use these existing objects if they are
3887 if (dchild->d_inode->i_size == 0) {
3888 rc = filter_use_existing_obj(obd,dchild,
3889 &handle, &cleanup_phase);
3896 CERROR("%s: Serious error: objid %.*s already "
3897 "exists; is this filesystem corrupt?\n",
3898 obd->obd_name, dchild->d_name.len,
3899 dchild->d_name.name);
3902 GOTO(cleanup, rc = -EEXIST);
3905 handle = fsfilt_start_log(obd, dparent->d_inode,
3906 FSFILT_OP_CREATE, NULL, 1);
3908 GOTO(cleanup, rc = PTR_ERR(handle));
3911 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3912 ",od->o_id="LPU64")\n", obd->obd_name, group,
3915 /* We mark object SUID+SGID to flag it for accepting UID+GID
3916 * from client on first write. Currently the permission bits
3917 * on the OST are never used, so this is OK. */
3918 rc = ll_vfs_create(dparent->d_inode, dchild,
3919 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3921 CERROR("create failed rc = %d\n", rc);
3922 if (rc == -ENOSPC) {
3923 os_ffree = filter_calc_free_inodes(obd);
3925 CERROR("%s: free inode "LPU64"\n",
3926 obd->obd_name, os_ffree);
3931 if (dchild->d_inode)
3932 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3933 dchild->d_inode->i_ino);
3936 if (!recreate_obj) {
3937 filter_set_last_id(filter, next_id, group);
3938 err = filter_update_last_objid(obd, group, 0);
3940 CERROR("unable to write lastobjid "
3941 "but file created\n");
3945 switch(cleanup_phase) {
3947 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3949 CERROR("error on commit, err = %d\n", err);
3956 filter_parent_unlock(dparent);
3963 if (time_after(jiffies, enough_time)) {
3965 "%s: precreate slow - want %d got %d \n",
3966 obd->obd_name, *num, i);
3973 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3974 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3979 static int filter_create(struct obd_export *exp, struct obdo *oa,
3980 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3982 struct obd_device *obd = exp->exp_obd;
3983 struct filter_export_data *fed;
3984 struct filter_obd *filter;
3985 struct lvfs_run_ctxt saved;
3986 struct lov_stripe_md *lsm = NULL;
3987 int rc = 0, diff, group = oa->o_gr;
3990 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3991 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3993 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3994 CERROR("!!! nid %s sent invalid object group %d\n",
3995 obd_export_nid2str(exp), group);
3999 fed = &exp->exp_filter_data;
4000 filter = &obd->u.filter;
4002 if (fed->fed_group != group) {
4003 CERROR("!!! this export (nid %s) used object group %d "
4004 "earlier; now it's trying to use group %d! This could "
4005 "be a bug in the MDS. Please report to "
4006 "http://bugzilla.lustre.org/\n",
4007 obd_export_nid2str(exp), fed->fed_group, group);
4014 rc = obd_alloc_memmd(exp, &lsm);
4021 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4023 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4024 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4025 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
4026 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4027 oa->o_id, filter_last_id(filter,
4032 down(&filter->fo_create_locks[oa->o_gr]);
4033 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
4034 up(&filter->fo_create_locks[oa->o_gr]);
4037 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
4040 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4041 if (rc && ea != NULL && *ea != lsm) {
4042 obd_free_memmd(exp, &lsm);
4043 } else if (rc == 0 && ea != NULL) {
4044 /* XXX LOV STACKING: the lsm that is passed to us from
4045 * LOV does not have valid lsm_oinfo data structs, so
4046 * don't go touching that. This needs to be fixed in a
4048 lsm->lsm_object_id = oa->o_id;
4055 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4056 struct lov_stripe_md *md, struct obd_trans_info *oti,
4057 struct obd_export *md_exp, void *capa)
4059 unsigned int qcids[MAXQUOTAS] = {0, 0};
4060 struct obd_device *obd;
4061 struct filter_obd *filter;
4062 struct dentry *dchild = NULL, *dparent = NULL;
4063 struct lustre_handle lockh = { 0 };
4064 struct lvfs_run_ctxt saved;
4065 void *handle = NULL;
4066 struct llog_cookie *fcc = NULL;
4067 int rc, rc2, cleanup_phase = 0, sync = 0;
4071 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4073 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
4074 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4079 filter = &obd->u.filter;
4081 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4084 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4085 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4087 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4089 GOTO(cleanup, rc = PTR_ERR(dchild));
4092 if (dchild->d_inode == NULL) {
4093 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4095 /* If object already gone, cancel cookie right now */
4096 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4097 struct llog_ctxt *ctxt;
4098 struct obd_llog_group *olg;
4099 fcc = &oa->o_lcookie;
4100 olg = filter_find_olg(obd, oa->o_gr);
4102 CERROR(" %s: can not find olg of group %d\n",
4103 obd->obd_name, (int)oa->o_gr);
4104 GOTO(cleanup, rc = PTR_ERR(olg));
4106 llog_group_set_export(olg, exp);
4108 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4109 llog_cancel(ctxt, NULL, 1, fcc, 0);
4110 llog_ctxt_put(ctxt);
4111 fcc = NULL; /* we didn't allocate fcc, don't free it */
4113 GOTO(cleanup, rc = -ENOENT);
4116 filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4118 /* Our MDC connection is established by the MDS to us */
4119 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4120 OBD_ALLOC(fcc, sizeof(*fcc));
4122 *fcc = oa->o_lcookie;
4124 DQUOT_INIT(dchild->d_inode);
4126 /* we're gonna truncate it first in order to avoid possible deadlock:
4128 * open trasaction open transaction
4129 * down(i_zombie) down(i_zombie)
4130 * restart transaction
4131 * (see BUG 4180) -bzzz
4133 LOCK_INODE_MUTEX(dchild->d_inode);
4135 /* VBR: version recovery check */
4136 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4140 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4142 if (IS_ERR(handle)) {
4143 UNLOCK_INODE_MUTEX(dchild->d_inode);
4144 GOTO(cleanup, rc = PTR_ERR(handle));
4147 iattr.ia_valid = ATTR_SIZE;
4149 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4150 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4151 UNLOCK_INODE_MUTEX(dchild->d_inode);
4155 GOTO(cleanup, rc = rc2);
4157 /* We don't actually need to lock the parent until we are unlinking
4158 * here, and not while truncating above. That avoids holding the
4159 * parent lock for a long time during truncate, which can block other
4160 * threads from doing anything to objects in that directory. bug 7171 */
4161 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4162 if (IS_ERR(dparent))
4163 GOTO(cleanup, rc = PTR_ERR(dparent));
4164 cleanup_phase = 3; /* filter_parent_unlock */
4166 LOCK_INODE_MUTEX(dchild->d_inode);
4167 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4168 if (IS_ERR(handle)) {
4169 UNLOCK_INODE_MUTEX(dchild->d_inode);
4170 GOTO(cleanup, rc = PTR_ERR(handle));
4172 cleanup_phase = 4; /* fsfilt_commit */
4174 /* Quota release need uid/gid of inode */
4175 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4177 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4179 /* this drops dchild->d_inode->i_mutex unconditionally */
4180 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4184 switch(cleanup_phase) {
4187 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4188 oti->oti_handle : handle,
4189 filter_cancel_cookies_cb,
4191 /* If add_journal_cb failed, then filter_finish_transno
4192 * will commit the handle and we will do a sync
4193 * on commit. then we call callback directly to free
4196 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4198 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4201 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4203 CERROR("error on commit, err = %d\n", rc2);
4210 filter_parent_unlock(dparent);
4212 filter_fini_destroy(obd, &lockh);
4216 OBD_FREE(fcc, sizeof(*fcc));
4218 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4221 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4225 /* trigger quota release */
4226 qcids[USRQUOTA] = oa->o_uid;
4227 qcids[GRPQUOTA] = oa->o_gid;
4228 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4231 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4235 /* NB start and end are used for punch, but not truncate */
4236 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4237 struct obd_trans_info *oti,
4238 struct ptlrpc_request_set *rqset)
4243 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4244 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4245 oinfo->oi_policy.l_extent.end);
4249 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4250 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4251 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4253 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4254 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4255 rc = filter_setattr(exp, oinfo, oti);
4256 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4260 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4261 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4264 struct lvfs_run_ctxt saved;
4265 struct filter_obd *filter;
4266 struct dentry *dentry;
4270 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
4271 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4275 filter = &exp->exp_obd->u.filter;
4277 /* An objid of zero is taken to mean "sync whole filesystem" */
4278 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4279 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4280 /* Flush any remaining cancel messages out to the target */
4281 filter_sync_llogs(exp->exp_obd, exp);
4285 dentry = filter_oa2dentry(exp->exp_obd, oa);
4287 RETURN(PTR_ERR(dentry));
4289 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4291 LOCK_INODE_MUTEX(dentry->d_inode);
4293 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4295 /* just any file to grab fsync method - "file" arg unused */
4296 struct file *file = filter->fo_rcvd_filp;
4298 if (file->f_op && file->f_op->fsync)
4299 rc = file->f_op->fsync(NULL, dentry, 1);
4301 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4305 UNLOCK_INODE_MUTEX(dentry->d_inode);
4307 oa->o_valid = OBD_MD_FLID;
4308 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4310 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4316 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4317 void *key, __u32 *vallen, void *val,
4318 struct lov_stripe_md *lsm)
4320 struct obd_device *obd;
4323 obd = class_exp2obd(exp);
4325 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4329 if (KEY_IS(KEY_BLOCKSIZE)) {
4330 __u32 *blocksize = val;
4332 if (*vallen < sizeof(*blocksize))
4334 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4336 *vallen = sizeof(*blocksize);
4340 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4341 __u32 *blocksize_bits = val;
4342 if (blocksize_bits) {
4343 if (*vallen < sizeof(*blocksize_bits))
4345 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4347 *vallen = sizeof(*blocksize_bits);
4351 if (KEY_IS(KEY_LAST_ID)) {
4352 obd_id *last_id = val;
4353 /* FIXME: object groups */
4355 if (*vallen < sizeof(*last_id))
4357 *last_id = filter_last_id(&obd->u.filter,
4358 exp->exp_filter_data.fed_group);
4360 *vallen = sizeof(*last_id);
4364 if (KEY_IS(KEY_FIEMAP)) {
4365 struct ll_fiemap_info_key *fm_key = key;
4366 struct dentry *dentry;
4367 struct ll_user_fiemap *fiemap = val;
4368 struct lvfs_run_ctxt saved;
4371 if (fiemap == NULL) {
4372 *vallen = fiemap_count_to_size(
4373 fm_key->fiemap.fm_extent_count);
4377 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4380 RETURN(PTR_ERR(dentry));
4382 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4383 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4384 rc = fsfilt_iocontrol(obd, dentry->d_inode, NULL,
4385 EXT3_IOC_FIEMAP, (long)fiemap);
4386 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4392 CDEBUG(D_IOCTL, "invalid key\n");
4396 static inline int filter_setup_llog_group(struct obd_export *exp,
4397 struct obd_device *obd,
4400 struct obd_llog_group *olg;
4401 struct llog_ctxt *ctxt;
4404 olg = filter_find_create_olg(obd, group);
4406 RETURN(PTR_ERR(olg));
4408 llog_group_set_export(olg, exp);
4410 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4411 LASSERTF(ctxt != NULL, "ctxt is null\n");
4413 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4414 llog_ctxt_put(ctxt);
4418 static int filter_set_grant_shrink(struct obd_export *exp,
4419 struct ost_body *body)
4421 /* handle shrink grant */
4422 spin_lock(&exp->exp_obd->obd_osfs_lock);
4423 filter_grant_incoming(exp, &body->oa);
4424 spin_unlock(&exp->exp_obd->obd_osfs_lock);
4430 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4432 struct obd_device *obd;
4438 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4442 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4443 obd_export_nid2str(exp));
4444 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4446 /* setup llog imports */
4448 group = (int)(*(__u32 *)val);
4450 group = 0; /* default value */
4452 LASSERT_MDS_GROUP(group);
4453 rc = filter_setup_llog_group(exp, obd, group);
4457 if (group == FILTER_GROUP_MDS0) {
4458 /* setup llog group 1 for interop */
4459 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4462 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4467 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4468 void *key, __u32 vallen, void *val,
4469 struct ptlrpc_request_set *set)
4471 struct obd_device *obd;
4476 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4480 if (KEY_IS(KEY_CAPA_KEY)) {
4482 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4484 CERROR("filter update capability key failed: %d\n", rc);
4488 if (KEY_IS(KEY_REVIMP_UPD)) {
4489 filter_revimp_update(exp);
4490 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4494 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4495 filter_adapt_sptlrpc_conf(obd, 0);
4499 if (KEY_IS(KEY_MDS_CONN))
4500 RETURN(filter_set_mds_conn(exp, val));
4502 if (KEY_IS(KEY_GRANT_SHRINK))
4503 RETURN(filter_set_grant_shrink(exp, val));
4508 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4509 int len, void *karg, void *uarg)
4511 struct obd_device *obd = exp->exp_obd;
4512 struct obd_ioctl_data *data = karg;
4516 case OBD_IOC_ABORT_RECOVERY: {
4517 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4518 target_stop_recovery_thread(obd);
4522 case OBD_IOC_SYNC: {
4523 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4524 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4528 case OBD_IOC_SET_READONLY: {
4530 struct super_block *sb = obd->u.obt.obt_sb;
4531 struct inode *inode = sb->s_root->d_inode;
4532 BDEVNAME_DECLARE_STORAGE(tmp);
4533 CERROR("*** setting device %s read-only ***\n",
4534 ll_bdevname(sb, tmp));
4536 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4537 if (!IS_ERR(handle))
4538 rc = fsfilt_commit(obd, inode, handle, 1);
4540 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4541 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4543 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4547 case OBD_IOC_CATLOGLIST: {
4548 rc = llog_catalog_list(obd, 1, data);
4552 case OBD_IOC_LLOG_CANCEL:
4553 case OBD_IOC_LLOG_REMOVE:
4554 case OBD_IOC_LLOG_INFO:
4555 case OBD_IOC_LLOG_PRINT: {
4556 /* FIXME to be finished */
4557 RETURN(-EOPNOTSUPP);
4559 struct llog_ctxt *ctxt = NULL;
4561 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4562 rc = llog_ioctl(ctxt, cmd, data);
4563 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4576 static int filter_health_check(struct obd_device *obd)
4578 #ifdef USE_HEALTH_CHECK_WRITE
4579 struct filter_obd *filter = &obd->u.filter;
4584 * health_check to return 0 on healthy
4585 * and 1 on unhealthy.
4587 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4590 #ifdef USE_HEALTH_CHECK_WRITE
4591 LASSERT(filter->fo_health_check_filp != NULL);
4592 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4597 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4600 return filter_fid2dentry(data, NULL, gr, id);
4603 static int filter_process_config(struct obd_device *obd, obd_count len,
4606 struct lustre_cfg *lcfg = buf;
4607 struct lprocfs_static_vars lvars;
4610 switch (lcfg->lcfg_command) {
4612 lprocfs_filter_init_vars(&lvars);
4614 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4624 static struct lvfs_callback_ops filter_lvfs_ops = {
4625 l_fid2dentry: filter_lvfs_fid2dentry,
4628 static struct obd_ops filter_obd_ops = {
4629 .o_owner = THIS_MODULE,
4630 .o_get_info = filter_get_info,
4631 .o_set_info_async = filter_set_info_async,
4632 .o_setup = filter_setup,
4633 .o_precleanup = filter_precleanup,
4634 .o_cleanup = filter_cleanup,
4635 .o_connect = filter_connect,
4636 .o_reconnect = filter_reconnect,
4637 .o_disconnect = filter_disconnect,
4638 .o_ping = filter_ping,
4639 .o_init_export = filter_init_export,
4640 .o_destroy_export = filter_destroy_export,
4641 .o_statfs = filter_statfs,
4642 .o_getattr = filter_getattr,
4643 .o_unpackmd = filter_unpackmd,
4644 .o_create = filter_create,
4645 .o_setattr = filter_setattr,
4646 .o_destroy = filter_destroy,
4647 .o_brw = filter_brw,
4648 .o_punch = filter_truncate,
4649 .o_sync = filter_sync,
4650 .o_preprw = filter_preprw,
4651 .o_commitrw = filter_commitrw,
4652 .o_llog_init = filter_llog_init,
4653 .o_llog_connect = filter_llog_connect,
4654 .o_llog_finish = filter_llog_finish,
4655 .o_iocontrol = filter_iocontrol,
4656 .o_health_check = filter_health_check,
4657 .o_process_config = filter_process_config,
4660 quota_interface_t *filter_quota_interface_ref;
4661 extern quota_interface_t filter_quota_interface;
4663 static int __init obdfilter_init(void)
4665 struct lprocfs_static_vars lvars;
4668 lprocfs_filter_init_vars(&lvars);
4670 request_module("lquota");
4671 OBD_ALLOC(obdfilter_created_scratchpad,
4672 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4673 sizeof(*obdfilter_created_scratchpad));
4674 if (obdfilter_created_scratchpad == NULL)
4677 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4678 sizeof(struct filter_mod_data),
4681 GOTO(out, rc = -ENOMEM);
4683 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4684 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4686 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4687 LUSTRE_OST_NAME, NULL);
4691 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4692 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4693 ll_fmd_cachep = NULL;
4695 if (filter_quota_interface_ref)
4696 PORTAL_SYMBOL_PUT(filter_quota_interface);
4698 OBD_FREE(obdfilter_created_scratchpad,
4699 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4700 sizeof(*obdfilter_created_scratchpad));
4706 static void __exit obdfilter_exit(void)
4708 if (filter_quota_interface_ref)
4709 PORTAL_SYMBOL_PUT(filter_quota_interface);
4711 if (ll_fmd_cachep) {
4712 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4713 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4714 ll_fmd_cachep = NULL;
4717 class_unregister_type(LUSTRE_OST_NAME);
4718 OBD_FREE(obdfilter_created_scratchpad,
4719 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4720 sizeof(*obdfilter_created_scratchpad));
4723 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4724 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4725 MODULE_LICENSE("GPL");
4727 module_init(obdfilter_init);
4728 module_exit(obdfilter_exit);