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, 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 *index)
2404 struct filter_obd *filter = &obd->u.filter;
2405 struct llog_ctxt *ctxt;
2409 LASSERT(olg != NULL);
2410 if (olg == &obd->obd_olg)
2411 return filter_default_olg_init(obd, olg, tgt);
2413 LASSERT(filter->fo_lcm != NULL);
2414 rc = filter_olg_init(obd, olg, tgt);
2417 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2419 CERROR("Can't get ctxt for %p:%x\n", olg,
2420 LLOG_MDS_OST_REPL_CTXT);
2421 filter_olg_fini(olg);
2424 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2425 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2426 llog_ctxt_put(ctxt);
2430 static int filter_llog_finish(struct obd_device *obd, int count)
2432 struct filter_obd *filter = &obd->u.filter;
2433 struct llog_ctxt *ctxt;
2436 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2439 * Make sure that no cached llcds left in recov_thread.
2440 * We actually do sync in disconnect time, but disconnect
2441 * may not come being marked rq_no_resend = 1.
2443 llog_sync(ctxt, NULL);
2446 * Balance class_import_get() in llog_receptor_accept().
2447 * This is safe to do, as llog is already synchronized
2448 * and its import may go.
2450 mutex_down(&ctxt->loc_sem);
2451 if (ctxt->loc_imp) {
2452 class_import_put(ctxt->loc_imp);
2453 ctxt->loc_imp = NULL;
2455 mutex_up(&ctxt->loc_sem);
2456 llog_ctxt_put(ctxt);
2459 if (filter->fo_lcm) {
2460 mutex_down(&ctxt->loc_sem);
2461 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2462 filter->fo_lcm = NULL;
2463 mutex_up(&ctxt->loc_sem);
2465 RETURN(filter_olg_fini(&obd->obd_olg));
2469 * Find the group llog according to group index in the llog group list.
2471 static struct obd_llog_group *
2472 filter_find_olg_internal(struct filter_obd *filter, int group)
2474 struct obd_llog_group *olg;
2476 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2477 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2478 if (olg->olg_group == group)
2485 * Find the group llog according to group index on the filter
2487 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2489 struct obd_llog_group *olg = NULL;
2490 struct filter_obd *filter;
2492 filter = &obd->u.filter;
2494 if (group == FILTER_GROUP_LLOG)
2495 RETURN(&obd->obd_olg);
2497 spin_lock(&filter->fo_llog_list_lock);
2498 olg = filter_find_olg_internal(filter, group);
2499 spin_unlock(&filter->fo_llog_list_lock);
2504 * Find the llog_group of the filter according to the group. If it can not
2505 * find, create the llog_group, which only happens when mds is being synced
2508 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2510 struct obd_llog_group *olg = NULL;
2511 struct filter_obd *filter;
2514 filter = &obd->u.filter;
2516 if (group == FILTER_GROUP_LLOG)
2517 RETURN(&obd->obd_olg);
2519 spin_lock(&filter->fo_llog_list_lock);
2520 olg = filter_find_olg_internal(filter, group);
2522 if (olg->olg_initializing) {
2523 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2525 GOTO(out_unlock, olg);
2530 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2532 llog_group_init(olg, group);
2533 list_add(&olg->olg_list, &filter->fo_llog_list);
2534 olg->olg_initializing = 1;
2535 spin_unlock(&filter->fo_llog_list_lock);
2537 rc = obd_llog_init(obd, olg, obd, NULL);
2539 spin_lock(&filter->fo_llog_list_lock);
2540 list_del(&olg->olg_list);
2541 spin_unlock(&filter->fo_llog_list_lock);
2543 GOTO(out, olg = ERR_PTR(-ENOMEM));
2545 spin_lock(&filter->fo_llog_list_lock);
2546 olg->olg_initializing = 0;
2547 spin_unlock(&filter->fo_llog_list_lock);
2548 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2549 obd->obd_name, group, olg);
2554 spin_unlock(&filter->fo_llog_list_lock);
2558 static int filter_llog_connect(struct obd_export *exp,
2559 struct llogd_conn_body *body)
2561 struct obd_device *obd = exp->exp_obd;
2562 struct llog_ctxt *ctxt;
2563 struct obd_llog_group *olg;
2567 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2568 obd->obd_name, body->lgdc_logid.lgl_oid,
2569 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2571 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2573 CERROR(" %s: can not find olg of group %d\n",
2574 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2577 llog_group_set_export(olg, exp);
2579 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2580 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2581 body->lgdc_ctxt_idx);
2583 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2584 obd->obd_name, body->lgdc_logid.lgl_oid,
2585 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2587 spin_lock_bh(&obd->obd_processing_task_lock);
2588 obd->u.filter.fo_mds_ost_sync = 1;
2589 spin_unlock_bh(&obd->obd_processing_task_lock);
2590 rc = llog_connect(ctxt, &body->lgdc_logid,
2591 &body->lgdc_gen, NULL);
2592 llog_ctxt_put(ctxt);
2594 CERROR("failed to connect rc %d idx %d\n", rc,
2595 body->lgdc_ctxt_idx);
2600 static int filter_llog_preclean(struct obd_device *obd)
2602 struct obd_llog_group *olg, *tmp;
2603 struct filter_obd *filter;
2604 struct list_head remove_list;
2608 rc = obd_llog_finish(obd, 0);
2610 CERROR("failed to cleanup llogging subsystem\n");
2612 filter = &obd->u.filter;
2613 CFS_INIT_LIST_HEAD(&remove_list);
2615 spin_lock(&filter->fo_llog_list_lock);
2616 while (!list_empty(&filter->fo_llog_list)) {
2617 olg = list_entry(filter->fo_llog_list.next,
2618 struct obd_llog_group, olg_list);
2619 list_del(&olg->olg_list);
2620 list_add(&olg->olg_list, &remove_list);
2622 spin_unlock(&filter->fo_llog_list_lock);
2624 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2625 list_del_init(&olg->olg_list);
2626 rc = filter_olg_fini(olg);
2628 CERROR("failed to cleanup llogging subsystem for %u\n",
2636 static int filter_precleanup(struct obd_device *obd,
2637 enum obd_cleanup_stage stage)
2643 case OBD_CLEANUP_EARLY:
2645 case OBD_CLEANUP_EXPORTS:
2646 /* Stop recovery before namespace cleanup. */
2647 target_recovery_fini(obd);
2648 rc = filter_llog_preclean(obd);
2654 static int filter_cleanup(struct obd_device *obd)
2656 struct filter_obd *filter = &obd->u.filter;
2660 LCONSOLE_WARN("%s: shutting down for failover; client state "
2661 "will be preserved.\n", obd->obd_name);
2663 obd_exports_barrier(obd);
2664 obd_zombie_barrier();
2666 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2667 lprocfs_free_per_client_stats(obd);
2668 lprocfs_free_obd_stats(obd);
2669 lprocfs_obd_cleanup(obd);
2670 lquota_cleanup(filter_quota_interface_ref, obd);
2672 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2673 obd->obd_namespace = NULL;
2675 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2677 if (obd->u.obt.obt_sb == NULL)
2682 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2683 shrink_dcache_sb(obd->u.obt.obt_sb);
2685 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2686 obd->u.obt.obt_sb = NULL;
2688 fsfilt_put_ops(obd->obd_fsops);
2690 filter_iobuf_pool_done(filter);
2692 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2697 static int filter_connect_internal(struct obd_export *exp,
2698 struct obd_connect_data *data)
2700 struct filter_export_data *fed = &exp->exp_filter_data;
2705 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2706 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2707 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2708 data->ocd_connect_flags, data->ocd_version,
2709 data->ocd_grant, data->ocd_index);
2711 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2712 CWARN("!!! This export (nid %s) used object group %d "
2713 "earlier; now it's trying to use group %d! This could "
2714 "be a bug in the MDS. Please report to "
2715 "http://bugzilla.lustre.org/\n",
2716 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2719 fed->fed_group = data->ocd_group;
2721 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2722 exp->exp_connect_flags = data->ocd_connect_flags;
2723 data->ocd_version = LUSTRE_VERSION_CODE;
2725 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2726 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2727 data->ocd_connect_flags & OBD_CONNECT_MDS))
2730 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2731 struct filter_obd *filter = &exp->exp_obd->u.filter;
2732 obd_size left, want;
2734 spin_lock(&exp->exp_obd->obd_osfs_lock);
2735 left = filter_grant_space_left(exp);
2736 want = data->ocd_grant;
2737 filter_grant(exp, fed->fed_grant, want, left);
2738 data->ocd_grant = fed->fed_grant;
2739 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2741 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2742 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2743 exp->exp_client_uuid.uuid, exp,
2744 data->ocd_grant, want, left);
2746 filter->fo_tot_granted_clients ++;
2749 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2750 struct filter_obd *filter = &exp->exp_obd->u.filter;
2751 struct lr_server_data *lsd = filter->fo_fsd;
2752 int index = le32_to_cpu(lsd->lsd_ost_index);
2754 if (!(lsd->lsd_feature_compat &
2755 cpu_to_le32(OBD_COMPAT_OST))) {
2756 /* this will only happen on the first connect */
2757 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2758 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2759 filter_update_server_data(exp->exp_obd,
2760 filter->fo_rcvd_filp, lsd, 1);
2761 } else if (index != data->ocd_index) {
2762 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2763 " %u doesn't match actual OST index"
2764 " %u in last_rcvd file, bad "
2766 obd_export_nid2str(exp), index,
2770 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2771 * FIXME: We don't strictly need the COMPAT flag for that,
2772 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2773 * FIXME: We needed it for the index, as index 0 is valid. */
2776 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2777 data->ocd_brw_size = 65536;
2778 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2779 data->ocd_brw_size = min(data->ocd_brw_size,
2780 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2781 LASSERT(data->ocd_brw_size);
2784 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2785 __u32 cksum_types = data->ocd_cksum_types;
2787 /* The client set in ocd_cksum_types the checksum types it
2788 * supports. We have to mask off the algorithms that we don't
2790 if (cksum_types & OBD_CKSUM_ALL)
2791 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2793 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2795 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2796 "%x\n", exp->exp_obd->obd_name,
2797 obd_export_nid2str(exp), cksum_types,
2798 data->ocd_cksum_types);
2800 /* This client does not support OBD_CONNECT_CKSUM
2801 * fall back to CRC32 */
2802 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2803 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2804 exp->exp_obd->obd_name,
2805 obd_export_nid2str(exp));
2811 static int filter_reconnect(const struct lu_env *env,
2812 struct obd_export *exp, struct obd_device *obd,
2813 struct obd_uuid *cluuid,
2814 struct obd_connect_data *data,
2820 if (exp == NULL || obd == NULL || cluuid == NULL)
2823 rc = filter_connect_internal(exp, data);
2825 filter_export_stats_init(obd, exp, localdata);
2830 /* nearly identical to mds_connect */
2831 static int filter_connect(const struct lu_env *env,
2832 struct obd_export **exp, struct obd_device *obd,
2833 struct obd_uuid *cluuid,
2834 struct obd_connect_data *data, void *localdata)
2836 struct lvfs_run_ctxt saved;
2837 struct lustre_handle conn = { 0 };
2838 struct obd_export *lexp;
2839 struct filter_export_data *fed;
2840 struct lsd_client_data *lcd = NULL;
2845 if (exp == NULL || obd == NULL || cluuid == NULL)
2848 rc = class_connect(&conn, obd, cluuid);
2851 lexp = class_conn2export(&conn);
2852 LASSERT(lexp != NULL);
2854 fed = &lexp->exp_filter_data;
2856 rc = filter_connect_internal(lexp, data);
2860 filter_export_stats_init(obd, lexp, localdata);
2861 if (obd->obd_replayable) {
2862 OBD_ALLOC(lcd, sizeof(*lcd));
2864 CERROR("filter: out of memory for client data\n");
2865 GOTO(cleanup, rc = -ENOMEM);
2868 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2870 rc = filter_client_add(obd, lexp, -1);
2875 group = data->ocd_group;
2877 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2878 obd->obd_name, lexp->exp_handle.h_cookie, group);
2880 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2881 rc = filter_read_groups(obd, group, 1);
2882 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2884 CERROR("can't read group %u\n", group);
2894 fed->fed_lcd = NULL;
2896 class_disconnect(lexp);
2905 /* Do extra sanity checks for grant accounting. We do this at connect,
2906 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2907 * always get rid of it or turn it off when we know accounting is good. */
2908 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2910 struct filter_export_data *fed;
2911 struct obd_export *exp;
2912 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2913 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2914 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2916 if (list_empty(&obd->obd_exports))
2919 /* We don't want to do this for large machines that do lots of
2920 mounts or unmounts. It burns... */
2921 if (obd->obd_num_exports > 100)
2924 spin_lock(&obd->obd_osfs_lock);
2925 spin_lock(&obd->obd_dev_lock);
2926 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2928 fed = &exp->exp_filter_data;
2929 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2932 if (maxsize > 0) { /* we may not have done a statfs yet */
2933 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2934 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2935 exp->exp_client_uuid.uuid, exp,
2936 fed->fed_grant, fed->fed_pending, maxsize);
2937 LASSERTF(fed->fed_dirty <= maxsize,
2938 "%s: cli %s/%p %ld > "LPU64"\n", func,
2939 exp->exp_client_uuid.uuid, exp,
2940 fed->fed_dirty, maxsize);
2943 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2944 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2945 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2947 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2948 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2949 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2950 tot_granted += fed->fed_grant + fed->fed_pending;
2951 tot_pending += fed->fed_pending;
2952 tot_dirty += fed->fed_dirty;
2954 fo_tot_granted = obd->u.filter.fo_tot_granted;
2955 fo_tot_pending = obd->u.filter.fo_tot_pending;
2956 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2957 spin_unlock(&obd->obd_dev_lock);
2958 spin_unlock(&obd->obd_osfs_lock);
2960 /* Do these assertions outside the spinlocks so we don't kill system */
2961 if (tot_granted != fo_tot_granted)
2962 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2963 func, tot_granted, fo_tot_granted);
2964 if (tot_pending != fo_tot_pending)
2965 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2966 func, tot_pending, fo_tot_pending);
2967 if (tot_dirty != fo_tot_dirty)
2968 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2969 func, tot_dirty, fo_tot_dirty);
2970 if (tot_pending > tot_granted)
2971 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2972 func, tot_pending, tot_granted);
2973 if (tot_granted > maxsize)
2974 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2975 func, tot_granted, maxsize);
2976 if (tot_dirty > maxsize)
2977 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2978 func, tot_dirty, maxsize);
2981 /* Remove this client from the grant accounting totals. We also remove
2982 * the export from the obd device under the osfs and dev locks to ensure
2983 * that the filter_grant_sanity_check() calculations are always valid.
2984 * The client should do something similar when it invalidates its import. */
2985 static void filter_grant_discard(struct obd_export *exp)
2987 struct obd_device *obd = exp->exp_obd;
2988 struct filter_obd *filter = &obd->u.filter;
2989 struct filter_export_data *fed = &exp->exp_filter_data;
2991 spin_lock(&obd->obd_osfs_lock);
2992 spin_lock(&obd->obd_dev_lock);
2993 list_del_init(&exp->exp_obd_chain);
2994 spin_unlock(&obd->obd_dev_lock);
2996 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2997 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2998 obd->obd_name, filter->fo_tot_granted,
2999 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
3000 filter->fo_tot_granted -= fed->fed_grant;
3001 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
3002 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
3003 obd->obd_name, filter->fo_tot_pending,
3004 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
3005 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
3006 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
3007 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
3008 obd->obd_name, filter->fo_tot_dirty,
3009 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
3010 filter->fo_tot_dirty -= fed->fed_dirty;
3014 spin_unlock(&obd->obd_osfs_lock);
3017 static int filter_destroy_export(struct obd_export *exp)
3021 if (exp->exp_filter_data.fed_pending)
3022 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
3023 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
3024 exp, exp->exp_filter_data.fed_pending);
3026 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3028 target_destroy_export(exp);
3029 ldlm_destroy_export(exp);
3031 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
3035 if (exp->exp_obd->obd_replayable)
3036 filter_client_free(exp);
3038 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3040 filter_grant_discard(exp);
3041 filter_fmd_cleanup(exp);
3043 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3044 struct filter_obd *filter = &exp->exp_obd->u.filter;
3045 if (filter->fo_tot_granted_clients > 0)
3046 filter->fo_tot_granted_clients --;
3049 if (!(exp->exp_flags & OBD_OPT_FORCE))
3050 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3055 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3057 struct obd_llog_group *olg_min, *olg;
3058 struct filter_obd *filter;
3059 int worked = -1, group;
3060 struct llog_ctxt *ctxt;
3063 filter = &obd->u.filter;
3065 /* we can't sync log holding spinlock. also, we do not want to get
3066 * into livelock. so we do following: loop over MDS's exports in
3067 * group order and skip already synced llogs -bzzz */
3069 /* look for group with min. number, but > worked */
3072 spin_lock(&filter->fo_llog_list_lock);
3073 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3074 if (olg->olg_group <= worked) {
3075 /* this group is already synced */
3078 if (group < olg->olg_group) {
3079 /* we have group with smaller number to sync */
3082 /* store current minimal group */
3084 group = olg->olg_group;
3086 spin_unlock(&filter->fo_llog_list_lock);
3088 if (olg_min == NULL)
3091 worked = olg_min->olg_group;
3092 if (olg_min->olg_exp &&
3093 (dexp == olg_min->olg_exp || dexp == NULL)) {
3095 ctxt = llog_group_get_ctxt(olg_min,
3096 LLOG_MDS_OST_REPL_CTXT);
3098 err = llog_sync(ctxt, olg_min->olg_exp);
3099 llog_ctxt_put(ctxt);
3101 CERROR("error flushing logs to MDS: "
3106 } while (olg_min != NULL);
3109 /* Also incredibly similar to mds_disconnect */
3110 static int filter_disconnect(struct obd_export *exp)
3112 struct obd_device *obd = exp->exp_obd;
3117 class_export_get(exp);
3119 if (!(exp->exp_flags & OBD_OPT_FORCE))
3120 filter_grant_sanity_check(obd, __FUNCTION__);
3121 filter_grant_discard(exp);
3123 /* Flush any remaining cancel messages out to the target */
3124 filter_sync_llogs(obd, exp);
3126 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3128 /* Disconnect early so that clients can't keep using export */
3129 rc = class_disconnect(exp);
3130 if (exp->exp_obd->obd_namespace != NULL)
3131 ldlm_cancel_locks_for_export(exp);
3133 fsfilt_sync(obd, obd->u.obt.obt_sb);
3135 lprocfs_exp_cleanup(exp);
3137 class_export_put(exp);
3141 /* reverse import is changed, sync all cancels */
3142 static void filter_revimp_update(struct obd_export *exp)
3147 class_export_get(exp);
3149 /* flush any remaining cancel messages out to the target */
3150 filter_sync_llogs(exp->exp_obd, exp);
3151 class_export_put(exp);
3155 static int filter_ping(struct obd_export *exp)
3157 filter_fmd_expire(exp);
3161 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3162 const char *what, int quiet)
3164 struct dentry *dchild = NULL;
3167 if (oa->o_valid & OBD_MD_FLGROUP)
3170 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3172 if (IS_ERR(dchild)) {
3173 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3174 what, group, oa->o_id);
3178 if (dchild->d_inode == NULL) {
3180 CERROR("%s: %s on non-existent object: "LPU64"\n",
3181 obd->obd_name, what, oa->o_id);
3183 RETURN(ERR_PTR(-ENOENT));
3189 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3191 struct dentry *dentry = NULL;
3192 struct obd_device *obd;
3196 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3197 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3201 obd = class_exp2obd(exp);
3203 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3207 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3209 RETURN(PTR_ERR(dentry));
3211 /* Limit the valid bits in the return data to what we actually use */
3212 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3213 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3219 /* this should be enabled/disabled in condition to enabled/disabled large
3220 * inodes (fast EAs) in backing store FS. */
3221 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3222 void *handle, struct obdo *oa)
3224 struct obd_device *obd = exp->exp_obd;
3228 if (oa->o_valid & OBD_MD_FLFID) {
3229 struct filter_fid ff;
3231 if (!(oa->o_valid & OBD_MD_FLGROUP))
3233 /* packing fid and converting it to LE for storing into EA.
3234 * Here ->o_stripe_idx should be filled by LOV and rest of
3235 * fields - by client. */
3236 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3237 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3238 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3239 ff.ff_objid = cpu_to_le64(oa->o_id);
3240 ff.ff_group = cpu_to_le64(oa->o_gr);
3242 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3243 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3244 oa->o_generation, oa->o_id, oa->o_gr);
3246 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3248 CERROR("store fid in object failed! rc: %d\n", rc);
3250 CDEBUG(D_HA, "OSS object without fid info!\n");
3256 /* this is called from filter_truncate() until we have filter_punch() */
3257 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3258 struct obdo *oa, struct obd_trans_info *oti)
3260 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3261 struct llog_cookie *fcc = NULL;
3262 struct filter_obd *filter;
3263 int rc, err, sync = 0;
3264 loff_t old_size = 0;
3265 unsigned int ia_valid;
3266 struct inode *inode;
3271 LASSERT(dentry != NULL);
3272 LASSERT(!IS_ERR(dentry));
3274 inode = dentry->d_inode;
3275 LASSERT(inode != NULL);
3277 filter = &exp->exp_obd->u.filter;
3278 iattr_from_obdo(&iattr, oa, oa->o_valid);
3279 ia_valid = iattr.ia_valid;
3281 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3282 OBD_ALLOC(fcc, sizeof(*fcc));
3284 *fcc = oa->o_lcookie;
3286 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3288 /* Filter truncates and writes are serialized by
3289 * i_alloc_sem, see the comment in
3290 * filter_preprw_write.*/
3291 if (ia_valid & ATTR_SIZE)
3292 down_write(&inode->i_alloc_sem);
3293 LOCK_INODE_MUTEX(inode);
3294 old_size = i_size_read(inode);
3297 /* VBR: version recovery check */
3298 rc = filter_version_get_check(exp, oti, inode);
3300 GOTO(out_unlock, rc);
3302 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3303 * then we will accept the UID+GID sent by the client during write for
3304 * initializing the ownership of this inode. We only allow this to
3305 * happen once so clear these bits in setattr. In 2.6 kernels it is
3306 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3307 * the flags that are actually being set. */
3308 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3309 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3310 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3312 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3313 if (!(ia_valid & ATTR_MODE)) {
3314 iattr.ia_mode = inode->i_mode;
3315 iattr.ia_valid |= ATTR_MODE;
3317 iattr.ia_mode &= ~S_ISUID;
3319 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3320 if (!(iattr.ia_valid & ATTR_MODE)) {
3321 iattr.ia_mode = inode->i_mode;
3322 iattr.ia_valid |= ATTR_MODE;
3324 iattr.ia_mode &= ~S_ISGID;
3327 orig_ids[USRQUOTA] = inode->i_uid;
3328 orig_ids[GRPQUOTA] = inode->i_gid;
3329 handle = fsfilt_start_log(exp->exp_obd, inode,
3330 FSFILT_OP_SETATTR, oti, 1);
3332 GOTO(out_unlock, rc = PTR_ERR(handle));
3334 /* update inode EA only once when inode is suid bit marked. As
3335 * on 2.6.x UID and GID may be set separately, we check here
3336 * only one of them to avoid double setting. */
3337 if (inode->i_mode & S_ISUID)
3338 filter_update_fidea(exp, inode, handle, oa);
3340 handle = fsfilt_start(exp->exp_obd, inode,
3341 FSFILT_OP_SETATTR, oti);
3343 GOTO(out_unlock, rc = PTR_ERR(handle));
3345 if (oa->o_valid & OBD_MD_FLFLAGS) {
3346 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3347 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3349 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3351 /* set cancel cookie callback function */
3352 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3353 filter_cancel_cookies_cb,
3357 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3358 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3360 /* The truncate might have used up our transaction credits. Make
3361 * sure we have one left for the last_rcvd update. */
3362 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3364 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3366 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3370 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3372 CERROR("error on commit, err = %d\n", err);
3379 /* For a partial-page truncate flush the page to disk immediately
3380 * to avoid data corruption during direct disk write. b=17397 */
3381 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3382 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3383 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3391 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3392 UNLOCK_INODE_MUTEX(inode);
3393 if (ia_valid & ATTR_SIZE)
3394 up_write(&inode->i_alloc_sem);
3396 OBD_FREE(fcc, sizeof(*fcc));
3398 /* trigger quota release */
3399 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3400 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3401 int rc2 = lquota_adjust(filter_quota_interface_ref,
3402 exp->exp_obd, cur_ids,
3403 orig_ids, rc, FSFILT_OP_SETATTR);
3404 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3405 "filter adjust qunit. (rc:%d)\n", rc2);
3410 /* this is called from filter_truncate() until we have filter_punch() */
3411 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3412 struct obd_trans_info *oti)
3414 struct obdo *oa = oinfo->oi_oa;
3415 struct lustre_capa *capa = oinfo_capa(oinfo);
3416 struct ldlm_res_id res_id;
3417 struct filter_mod_data *fmd;
3418 struct lvfs_run_ctxt saved;
3419 struct filter_obd *filter;
3420 struct ldlm_resource *res;
3421 struct dentry *dentry;
3422 __u64 opc = CAPA_OPC_META_WRITE;
3426 if (oa->o_valid & OBD_FL_TRUNC)
3427 opc |= CAPA_OPC_OSS_TRUNC;
3428 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa, opc);
3432 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3433 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
3438 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3439 /* This would be very bad - accidentally truncating a file when
3440 * changing the time or similar - bug 12203. */
3441 if (oa->o_valid & OBD_MD_FLSIZE &&
3442 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3443 static char mdsinum[48];
3445 if (oa->o_valid & OBD_MD_FLFID)
3446 snprintf(mdsinum, sizeof(mdsinum) - 1,
3447 " of inode "LPU64"/%u", oa->o_fid,
3452 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3454 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3459 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3461 RETURN(PTR_ERR(dentry));
3463 filter = &exp->exp_obd->u.filter;
3464 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3468 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3469 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3470 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3471 fmd->fmd_mactime_xid = oti->oti_xid;
3472 filter_fmd_put(exp, fmd);
3475 /* setting objects attributes (including owner/group) */
3476 rc = filter_setattr_internal(exp, dentry, oa, oti);
3478 GOTO(out_unlock, rc);
3480 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3481 &res_id, LDLM_EXTENT, 0);
3484 LDLM_RESOURCE_ADDREF(res);
3485 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3486 LDLM_RESOURCE_DELREF(res);
3487 ldlm_resource_putref(res);
3490 oa->o_valid = OBD_MD_FLID;
3492 /* Quota release need uid/gid info */
3493 obdo_from_inode(oa, dentry->d_inode,
3494 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3500 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3504 /* XXX identical to osc_unpackmd */
3505 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3506 struct lov_mds_md *lmm, int lmm_bytes)
3512 if (lmm_bytes < sizeof (*lmm)) {
3513 CERROR("lov_mds_md too small: %d, need %d\n",
3514 lmm_bytes, (int)sizeof(*lmm));
3517 /* XXX LOV_MAGIC etc check? */
3519 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3520 CERROR("lov_mds_md: zero lmm_object_id\n");
3525 lsm_size = lov_stripe_md_size(1);
3529 if (*lsmp != NULL && lmm == NULL) {
3530 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3531 OBD_FREE(*lsmp, lsm_size);
3536 if (*lsmp == NULL) {
3537 OBD_ALLOC(*lsmp, lsm_size);
3541 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3542 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3543 OBD_FREE(*lsmp, lsm_size);
3546 loi_init((*lsmp)->lsm_oinfo[0]);
3550 /* XXX zero *lsmp? */
3551 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3552 LASSERT((*lsmp)->lsm_object_id);
3555 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3560 /* caller must hold fo_create_locks[oa->o_gr] */
3561 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3562 struct filter_obd *filter)
3564 struct obdo doa = { 0 }; /* XXX obdo on stack */
3571 LASSERT_MDS_GROUP(oa->o_gr);
3572 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3573 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3575 memset(&doa, 0, sizeof(doa));
3577 doa.o_valid |= OBD_MD_FLGROUP;
3578 doa.o_gr = oa->o_gr;
3579 doa.o_mode = S_IFREG;
3581 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3582 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3583 exp->exp_obd->obd_name, doa.o_gr);
3587 last = filter_last_id(filter, doa.o_gr);
3589 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3591 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3592 exp->exp_obd->obd_name, oa->o_id + 1, last,
3593 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3595 for (id = last; id > oa->o_id; id--) {
3597 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3598 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3599 CEMERG("error destroying precreate objid "LPU64": %d\n",
3602 /* update last_id on disk periodically so that if we restart
3603 * we don't need to re-scan all of the just-deleted objects. */
3604 if ((id & 511) == 0 && !skip_orphan) {
3605 filter_set_last_id(filter, id - 1, doa.o_gr);
3606 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3610 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3611 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3614 filter_set_last_id(filter, id, doa.o_gr);
3615 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3617 /* don't reuse orphan object, return last used objid */
3621 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3626 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3627 obd_gr group, int *num);
3628 /* returns a negative error or a nonnegative number of files to create */
3629 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3630 obd_gr group, struct obd_trans_info *oti)
3632 struct obd_device *obd = exp->exp_obd;
3633 struct filter_obd *filter = &obd->u.filter;
3637 /* delete orphans request */
3638 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3639 obd_id last = filter_last_id(filter, group);
3641 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3642 CERROR("%s: dropping old orphan cleanup request\n",
3646 /* This causes inflight precreates to abort and drop lock */
3647 set_bit(group, &filter->fo_destroys_in_progress);
3648 down(&filter->fo_create_locks[group]);
3649 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3650 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3651 exp->exp_obd->obd_name, group);
3652 up(&filter->fo_create_locks[group]);
3655 diff = oa->o_id - last;
3656 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3659 if (-diff > OST_MAX_PRECREATE) {
3660 CERROR("%s: ignoring bogus orphan destroy request: "
3661 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3663 /* FIXME: should reset precreate_next_id on MDS */
3664 GOTO(out, rc = -EINVAL);
3667 rc = filter_destroy_precreated(exp, oa, filter);
3669 CERROR("%s: unable to write lastobjid, but "
3670 "orphans were deleted\n", obd->obd_name);
3673 /* XXX: Used by MDS for the first time! */
3674 clear_bit(group, &filter->fo_destroys_in_progress);
3677 down(&filter->fo_create_locks[group]);
3678 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3679 CERROR("%s: dropping old precreate request\n",
3683 /* only precreate if group == 0 and o_id is specfied */
3684 if (group == FILTER_GROUP_LLOG || oa->o_id == 0)
3687 diff = oa->o_id - filter_last_id(filter, group);
3688 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3689 filter_last_id(filter, group), diff);
3691 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3692 oa->o_id, filter_last_id(filter, group), diff);
3696 oa->o_id = filter_last_id(&obd->u.filter, group);
3697 rc = filter_precreate(obd, oa, group, &diff);
3698 oa->o_id = filter_last_id(&obd->u.filter, group);
3700 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3703 /* else diff == 0 */
3706 up(&filter->fo_create_locks[group]);
3710 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3711 __u64 max_age, __u32 flags)
3713 struct filter_obd *filter = &obd->u.filter;
3714 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3718 /* at least try to account for cached pages. its still racey and
3719 * might be under-reporting if clients haven't announced their
3720 * caches with brw recently */
3721 spin_lock(&obd->obd_osfs_lock);
3722 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3723 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3724 spin_unlock(&obd->obd_osfs_lock);
3726 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3727 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3728 filter->fo_tot_dirty, filter->fo_tot_granted,
3729 filter->fo_tot_pending,
3730 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3732 filter_grant_sanity_check(obd, __FUNCTION__);
3734 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3735 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3736 osfs->os_bsize - 1) >> blockbits));
3738 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3739 * stop creating files on MDS if OST is not good shape to create
3743 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3744 osfs->os_state = OS_STATE_READONLY;
3746 if (filter->fo_raid_degraded)
3747 osfs->os_state |= OS_STATE_DEGRADED;
3751 static int filter_use_existing_obj(struct obd_device *obd,
3752 struct dentry *dchild, void **handle,
3755 struct inode *inode = dchild->d_inode;
3759 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3762 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3763 if (IS_ERR(*handle))
3764 return PTR_ERR(*handle);
3766 iattr.ia_valid = ATTR_MODE;
3767 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3768 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3775 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3778 __u64 os_ffree = -1;
3780 spin_lock(&obd->obd_osfs_lock);
3781 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3783 os_ffree = obd->obd_osfs.os_ffree;
3784 spin_unlock(&obd->obd_osfs_lock);
3789 /* We rely on the fact that only one thread will be creating files in a given
3790 * group at a time, which is why we don't need an atomic filter_get_new_id.
3791 * Even if we had that atomic function, the following race would exist:
3793 * thread 1: gets id x from filter_next_id
3794 * thread 2: gets id (x + 1) from filter_next_id
3795 * thread 2: creates object (x + 1)
3796 * thread 1: tries to create object x, gets -ENOSPC
3798 * Caller must hold fo_create_locks[group]
3800 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3801 obd_gr group, int *num)
3803 struct dentry *dchild = NULL, *dparent = NULL;
3804 struct filter_obd *filter;
3805 struct obd_statfs *osfs;
3806 int err = 0, rc = 0, recreate_obj = 0, i;
3807 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3810 void *handle = NULL;
3813 filter = &obd->u.filter;
3815 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3817 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3819 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3820 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3823 OBD_ALLOC(osfs, sizeof(*osfs));
3826 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3827 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3828 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3829 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3830 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3834 OBD_FREE(osfs, sizeof(*osfs));
3839 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3840 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3842 for (i = 0; i < *num && err == 0; i++) {
3843 int cleanup_phase = 0;
3845 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3846 CWARN("%s: create aborted by destroy\n",
3855 last_id = filter_last_id(filter, group);
3856 if (next_id > last_id) {
3857 CERROR("Error: Trying to recreate obj greater"
3858 "than last id "LPD64" > "LPD64"\n",
3860 GOTO(cleanup, rc = -EINVAL);
3863 next_id = filter_last_id(filter, group) + 1;
3865 dparent = filter_parent_lock(obd, group, next_id);
3866 if (IS_ERR(dparent))
3867 GOTO(cleanup, rc = PTR_ERR(dparent));
3868 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3870 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3872 GOTO(cleanup, rc = PTR_ERR(dchild));
3873 cleanup_phase = 2; /* f_dput(dchild) */
3875 if (dchild->d_inode != NULL) {
3876 /* This would only happen if lastobjid was bad on disk*/
3877 /* Could also happen if recreating missing obj but it
3878 * already exists. */
3880 CERROR("%s: recreating existing object %.*s?\n",
3881 obd->obd_name, dchild->d_name.len,
3882 dchild->d_name.name);
3884 /* Use these existing objects if they are
3886 if (dchild->d_inode->i_size == 0) {
3887 rc = filter_use_existing_obj(obd,dchild,
3888 &handle, &cleanup_phase);
3895 CERROR("%s: Serious error: objid %.*s already "
3896 "exists; is this filesystem corrupt?\n",
3897 obd->obd_name, dchild->d_name.len,
3898 dchild->d_name.name);
3901 GOTO(cleanup, rc = -EEXIST);
3904 handle = fsfilt_start_log(obd, dparent->d_inode,
3905 FSFILT_OP_CREATE, NULL, 1);
3907 GOTO(cleanup, rc = PTR_ERR(handle));
3910 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3911 ",od->o_id="LPU64")\n", obd->obd_name, group,
3914 /* We mark object SUID+SGID to flag it for accepting UID+GID
3915 * from client on first write. Currently the permission bits
3916 * on the OST are never used, so this is OK. */
3917 rc = ll_vfs_create(dparent->d_inode, dchild,
3918 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3920 CERROR("create failed rc = %d\n", rc);
3921 if (rc == -ENOSPC) {
3922 os_ffree = filter_calc_free_inodes(obd);
3924 CERROR("%s: free inode "LPU64"\n",
3925 obd->obd_name, os_ffree);
3930 if (dchild->d_inode)
3931 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3932 dchild->d_inode->i_ino);
3935 if (!recreate_obj) {
3936 filter_set_last_id(filter, next_id, group);
3937 err = filter_update_last_objid(obd, group, 0);
3939 CERROR("unable to write lastobjid "
3940 "but file created\n");
3944 switch(cleanup_phase) {
3946 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3948 CERROR("error on commit, err = %d\n", err);
3955 filter_parent_unlock(dparent);
3962 if (time_after(jiffies, enough_time)) {
3964 "%s: precreate slow - want %d got %d \n",
3965 obd->obd_name, *num, i);
3972 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3973 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3978 static int filter_create(struct obd_export *exp, struct obdo *oa,
3979 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3981 struct obd_device *obd = exp->exp_obd;
3982 struct filter_export_data *fed;
3983 struct filter_obd *filter;
3984 struct lvfs_run_ctxt saved;
3985 struct lov_stripe_md *lsm = NULL;
3986 int rc = 0, diff, group = oa->o_gr;
3989 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3990 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3992 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3993 CERROR("!!! nid %s sent invalid object group %d\n",
3994 obd_export_nid2str(exp), group);
3998 fed = &exp->exp_filter_data;
3999 filter = &obd->u.filter;
4001 if (fed->fed_group != group) {
4002 CERROR("!!! this export (nid %s) used object group %d "
4003 "earlier; now it's trying to use group %d! This could "
4004 "be a bug in the MDS. Please report to "
4005 "http://bugzilla.lustre.org/\n",
4006 obd_export_nid2str(exp), fed->fed_group, group);
4013 rc = obd_alloc_memmd(exp, &lsm);
4020 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4022 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4023 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4024 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
4025 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4026 oa->o_id, filter_last_id(filter,
4031 down(&filter->fo_create_locks[oa->o_gr]);
4032 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
4033 up(&filter->fo_create_locks[oa->o_gr]);
4036 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
4039 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4040 if (rc && ea != NULL && *ea != lsm) {
4041 obd_free_memmd(exp, &lsm);
4042 } else if (rc == 0 && ea != NULL) {
4043 /* XXX LOV STACKING: the lsm that is passed to us from
4044 * LOV does not have valid lsm_oinfo data structs, so
4045 * don't go touching that. This needs to be fixed in a
4047 lsm->lsm_object_id = oa->o_id;
4054 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4055 struct lov_stripe_md *md, struct obd_trans_info *oti,
4056 struct obd_export *md_exp, void *capa)
4058 unsigned int qcids[MAXQUOTAS] = {0, 0};
4059 struct obd_device *obd;
4060 struct filter_obd *filter;
4061 struct dentry *dchild = NULL, *dparent = NULL;
4062 struct lustre_handle lockh = { 0 };
4063 struct lvfs_run_ctxt saved;
4064 void *handle = NULL;
4065 struct llog_cookie *fcc = NULL;
4066 int rc, rc2, cleanup_phase = 0, sync = 0;
4070 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4072 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
4073 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4078 filter = &obd->u.filter;
4080 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4083 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4084 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4086 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4088 GOTO(cleanup, rc = PTR_ERR(dchild));
4091 if (dchild->d_inode == NULL) {
4092 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4094 /* If object already gone, cancel cookie right now */
4095 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4096 struct llog_ctxt *ctxt;
4097 struct obd_llog_group *olg;
4098 fcc = &oa->o_lcookie;
4099 olg = filter_find_olg(obd, oa->o_gr);
4101 CERROR(" %s: can not find olg of group %d\n",
4102 obd->obd_name, (int)oa->o_gr);
4103 GOTO(cleanup, rc = PTR_ERR(olg));
4105 llog_group_set_export(olg, exp);
4107 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4108 llog_cancel(ctxt, NULL, 1, fcc, 0);
4109 llog_ctxt_put(ctxt);
4110 fcc = NULL; /* we didn't allocate fcc, don't free it */
4112 GOTO(cleanup, rc = -ENOENT);
4115 filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4117 /* Our MDC connection is established by the MDS to us */
4118 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4119 OBD_ALLOC(fcc, sizeof(*fcc));
4121 *fcc = oa->o_lcookie;
4123 DQUOT_INIT(dchild->d_inode);
4125 /* we're gonna truncate it first in order to avoid possible deadlock:
4127 * open trasaction open transaction
4128 * down(i_zombie) down(i_zombie)
4129 * restart transaction
4130 * (see BUG 4180) -bzzz
4132 LOCK_INODE_MUTEX(dchild->d_inode);
4134 /* VBR: version recovery check */
4135 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4139 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4141 if (IS_ERR(handle)) {
4142 UNLOCK_INODE_MUTEX(dchild->d_inode);
4143 GOTO(cleanup, rc = PTR_ERR(handle));
4146 iattr.ia_valid = ATTR_SIZE;
4148 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4149 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4150 UNLOCK_INODE_MUTEX(dchild->d_inode);
4154 GOTO(cleanup, rc = rc2);
4156 /* We don't actually need to lock the parent until we are unlinking
4157 * here, and not while truncating above. That avoids holding the
4158 * parent lock for a long time during truncate, which can block other
4159 * threads from doing anything to objects in that directory. bug 7171 */
4160 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4161 if (IS_ERR(dparent))
4162 GOTO(cleanup, rc = PTR_ERR(dparent));
4163 cleanup_phase = 3; /* filter_parent_unlock */
4165 LOCK_INODE_MUTEX(dchild->d_inode);
4166 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4167 if (IS_ERR(handle)) {
4168 UNLOCK_INODE_MUTEX(dchild->d_inode);
4169 GOTO(cleanup, rc = PTR_ERR(handle));
4171 cleanup_phase = 4; /* fsfilt_commit */
4173 /* Quota release need uid/gid of inode */
4174 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4176 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4178 /* this drops dchild->d_inode->i_mutex unconditionally */
4179 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4183 switch(cleanup_phase) {
4186 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4187 oti->oti_handle : handle,
4188 filter_cancel_cookies_cb,
4190 /* If add_journal_cb failed, then filter_finish_transno
4191 * will commit the handle and we will do a sync
4192 * on commit. then we call callback directly to free
4195 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4197 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4200 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4202 CERROR("error on commit, err = %d\n", rc2);
4209 filter_parent_unlock(dparent);
4211 filter_fini_destroy(obd, &lockh);
4215 OBD_FREE(fcc, sizeof(*fcc));
4217 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4220 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4224 /* trigger quota release */
4225 qcids[USRQUOTA] = oa->o_uid;
4226 qcids[GRPQUOTA] = oa->o_gid;
4227 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4230 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4234 /* NB start and end are used for punch, but not truncate */
4235 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4236 struct obd_trans_info *oti,
4237 struct ptlrpc_request_set *rqset)
4242 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4243 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4244 oinfo->oi_policy.l_extent.end);
4248 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4249 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4250 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4252 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4253 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4254 rc = filter_setattr(exp, oinfo, oti);
4255 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4259 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4260 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4263 struct lvfs_run_ctxt saved;
4264 struct filter_obd *filter;
4265 struct dentry *dentry;
4269 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
4270 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4274 filter = &exp->exp_obd->u.filter;
4276 /* An objid of zero is taken to mean "sync whole filesystem" */
4277 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4278 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4279 /* Flush any remaining cancel messages out to the target */
4280 filter_sync_llogs(exp->exp_obd, exp);
4284 dentry = filter_oa2dentry(exp->exp_obd, oa);
4286 RETURN(PTR_ERR(dentry));
4288 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4290 LOCK_INODE_MUTEX(dentry->d_inode);
4292 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4294 /* just any file to grab fsync method - "file" arg unused */
4295 struct file *file = filter->fo_rcvd_filp;
4297 if (file->f_op && file->f_op->fsync)
4298 rc = file->f_op->fsync(NULL, dentry, 1);
4300 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4304 UNLOCK_INODE_MUTEX(dentry->d_inode);
4306 oa->o_valid = OBD_MD_FLID;
4307 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4309 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4315 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4316 void *key, __u32 *vallen, void *val,
4317 struct lov_stripe_md *lsm)
4319 struct obd_device *obd;
4322 obd = class_exp2obd(exp);
4324 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4328 if (KEY_IS(KEY_BLOCKSIZE)) {
4329 __u32 *blocksize = val;
4331 if (*vallen < sizeof(*blocksize))
4333 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4335 *vallen = sizeof(*blocksize);
4339 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4340 __u32 *blocksize_bits = val;
4341 if (blocksize_bits) {
4342 if (*vallen < sizeof(*blocksize_bits))
4344 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4346 *vallen = sizeof(*blocksize_bits);
4350 if (KEY_IS(KEY_LAST_ID)) {
4351 obd_id *last_id = val;
4352 /* FIXME: object groups */
4354 if (*vallen < sizeof(*last_id))
4356 *last_id = filter_last_id(&obd->u.filter,
4357 exp->exp_filter_data.fed_group);
4359 *vallen = sizeof(*last_id);
4363 if (KEY_IS(KEY_FIEMAP)) {
4364 struct ll_fiemap_info_key *fm_key = key;
4365 struct dentry *dentry;
4366 struct ll_user_fiemap *fiemap = val;
4367 struct lvfs_run_ctxt saved;
4370 if (fiemap == NULL) {
4371 *vallen = fiemap_count_to_size(
4372 fm_key->fiemap.fm_extent_count);
4376 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4379 RETURN(PTR_ERR(dentry));
4381 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4382 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4383 rc = fsfilt_iocontrol(obd, dentry->d_inode, NULL,
4384 EXT3_IOC_FIEMAP, (long)fiemap);
4385 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4391 CDEBUG(D_IOCTL, "invalid key\n");
4395 static inline int filter_setup_llog_group(struct obd_export *exp,
4396 struct obd_device *obd,
4399 struct obd_llog_group *olg;
4400 struct llog_ctxt *ctxt;
4403 olg = filter_find_create_olg(obd, group);
4405 RETURN(PTR_ERR(olg));
4407 llog_group_set_export(olg, exp);
4409 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4410 LASSERTF(ctxt != NULL, "ctxt is null\n");
4412 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4413 llog_ctxt_put(ctxt);
4417 static int filter_set_grant_shrink(struct obd_export *exp,
4418 struct ost_body *body)
4420 /* handle shrink grant */
4421 spin_lock(&exp->exp_obd->obd_osfs_lock);
4422 filter_grant_incoming(exp, &body->oa);
4423 spin_unlock(&exp->exp_obd->obd_osfs_lock);
4429 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4431 struct obd_device *obd;
4437 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4441 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4442 obd_export_nid2str(exp));
4443 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4445 /* setup llog imports */
4447 group = (int)(*(__u32 *)val);
4449 group = 0; /* default value */
4451 LASSERT_MDS_GROUP(group);
4452 rc = filter_setup_llog_group(exp, obd, group);
4456 if (group == FILTER_GROUP_MDS0) {
4457 /* setup llog group 1 for interop */
4458 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4461 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4466 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4467 void *key, __u32 vallen, void *val,
4468 struct ptlrpc_request_set *set)
4470 struct obd_device *obd;
4475 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4479 if (KEY_IS(KEY_CAPA_KEY)) {
4481 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4483 CERROR("filter update capability key failed: %d\n", rc);
4487 if (KEY_IS(KEY_REVIMP_UPD)) {
4488 filter_revimp_update(exp);
4489 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4493 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4494 filter_adapt_sptlrpc_conf(obd, 0);
4498 if (KEY_IS(KEY_MDS_CONN))
4499 RETURN(filter_set_mds_conn(exp, val));
4501 if (KEY_IS(KEY_GRANT_SHRINK))
4502 RETURN(filter_set_grant_shrink(exp, val));
4507 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4508 int len, void *karg, void *uarg)
4510 struct obd_device *obd = exp->exp_obd;
4511 struct obd_ioctl_data *data = karg;
4515 case OBD_IOC_ABORT_RECOVERY: {
4516 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4517 target_stop_recovery_thread(obd);
4521 case OBD_IOC_SYNC: {
4522 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4523 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4527 case OBD_IOC_SET_READONLY: {
4529 struct super_block *sb = obd->u.obt.obt_sb;
4530 struct inode *inode = sb->s_root->d_inode;
4531 BDEVNAME_DECLARE_STORAGE(tmp);
4532 CERROR("*** setting device %s read-only ***\n",
4533 ll_bdevname(sb, tmp));
4535 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4536 if (!IS_ERR(handle))
4537 rc = fsfilt_commit(obd, inode, handle, 1);
4539 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4540 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4542 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4546 case OBD_IOC_CATLOGLIST: {
4547 rc = llog_catalog_list(obd, 1, data);
4551 case OBD_IOC_LLOG_CANCEL:
4552 case OBD_IOC_LLOG_REMOVE:
4553 case OBD_IOC_LLOG_INFO:
4554 case OBD_IOC_LLOG_PRINT: {
4555 /* FIXME to be finished */
4556 RETURN(-EOPNOTSUPP);
4558 struct llog_ctxt *ctxt = NULL;
4560 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4561 rc = llog_ioctl(ctxt, cmd, data);
4562 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4575 static int filter_health_check(struct obd_device *obd)
4577 #ifdef USE_HEALTH_CHECK_WRITE
4578 struct filter_obd *filter = &obd->u.filter;
4583 * health_check to return 0 on healthy
4584 * and 1 on unhealthy.
4586 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4589 #ifdef USE_HEALTH_CHECK_WRITE
4590 LASSERT(filter->fo_health_check_filp != NULL);
4591 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4596 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4599 return filter_fid2dentry(data, NULL, gr, id);
4602 static int filter_process_config(struct obd_device *obd, obd_count len,
4605 struct lustre_cfg *lcfg = buf;
4606 struct lprocfs_static_vars lvars;
4609 switch (lcfg->lcfg_command) {
4611 lprocfs_filter_init_vars(&lvars);
4613 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4623 static struct lvfs_callback_ops filter_lvfs_ops = {
4624 l_fid2dentry: filter_lvfs_fid2dentry,
4627 static struct obd_ops filter_obd_ops = {
4628 .o_owner = THIS_MODULE,
4629 .o_get_info = filter_get_info,
4630 .o_set_info_async = filter_set_info_async,
4631 .o_setup = filter_setup,
4632 .o_precleanup = filter_precleanup,
4633 .o_cleanup = filter_cleanup,
4634 .o_connect = filter_connect,
4635 .o_reconnect = filter_reconnect,
4636 .o_disconnect = filter_disconnect,
4637 .o_ping = filter_ping,
4638 .o_init_export = filter_init_export,
4639 .o_destroy_export = filter_destroy_export,
4640 .o_statfs = filter_statfs,
4641 .o_getattr = filter_getattr,
4642 .o_unpackmd = filter_unpackmd,
4643 .o_create = filter_create,
4644 .o_setattr = filter_setattr,
4645 .o_destroy = filter_destroy,
4646 .o_brw = filter_brw,
4647 .o_punch = filter_truncate,
4648 .o_sync = filter_sync,
4649 .o_preprw = filter_preprw,
4650 .o_commitrw = filter_commitrw,
4651 .o_llog_init = filter_llog_init,
4652 .o_llog_connect = filter_llog_connect,
4653 .o_llog_finish = filter_llog_finish,
4654 .o_iocontrol = filter_iocontrol,
4655 .o_health_check = filter_health_check,
4656 .o_process_config = filter_process_config,
4659 quota_interface_t *filter_quota_interface_ref;
4660 extern quota_interface_t filter_quota_interface;
4662 static int __init obdfilter_init(void)
4664 struct lprocfs_static_vars lvars;
4667 lprocfs_filter_init_vars(&lvars);
4669 request_module("lquota");
4670 OBD_ALLOC(obdfilter_created_scratchpad,
4671 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4672 sizeof(*obdfilter_created_scratchpad));
4673 if (obdfilter_created_scratchpad == NULL)
4676 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4677 sizeof(struct filter_mod_data),
4680 GOTO(out, rc = -ENOMEM);
4682 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4683 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4685 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4686 LUSTRE_OST_NAME, NULL);
4690 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4691 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4692 ll_fmd_cachep = NULL;
4694 if (filter_quota_interface_ref)
4695 PORTAL_SYMBOL_PUT(filter_quota_interface);
4697 OBD_FREE(obdfilter_created_scratchpad,
4698 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4699 sizeof(*obdfilter_created_scratchpad));
4705 static void __exit obdfilter_exit(void)
4707 if (filter_quota_interface_ref)
4708 PORTAL_SYMBOL_PUT(filter_quota_interface);
4710 if (ll_fmd_cachep) {
4711 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4712 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4713 ll_fmd_cachep = NULL;
4716 class_unregister_type(LUSTRE_OST_NAME);
4717 OBD_FREE(obdfilter_created_scratchpad,
4718 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4719 sizeof(*obdfilter_created_scratchpad));
4722 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4723 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4724 MODULE_LICENSE("GPL");
4726 module_init(obdfilter_init);
4727 module_exit(obdfilter_exit);