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 static struct lvfs_callback_ops filter_lvfs_ops;
83 cfs_mem_cache_t *ll_fmd_cachep;
85 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
86 void *cb_data, int error)
88 struct obd_export *exp = cb_data;
89 LASSERT(exp->exp_obd == obd);
90 obd_transno_commit_cb(obd, transno, exp, error);
91 class_export_cb_put(exp);
94 int filter_version_get_check(struct obd_export *exp,
95 struct obd_trans_info *oti, struct inode *inode)
99 if (inode == NULL || oti == NULL)
102 curr_version = fsfilt_get_version(exp->exp_obd, inode);
103 if ((__s64)curr_version == -EOPNOTSUPP)
105 /* VBR: version is checked always because costs nothing */
106 if (oti->oti_pre_version != 0 &&
107 oti->oti_pre_version != curr_version) {
108 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
109 oti->oti_pre_version, curr_version);
110 spin_lock(&exp->exp_lock);
111 exp->exp_vbr_failed = 1;
112 spin_unlock(&exp->exp_lock);
115 oti->oti_pre_version = curr_version;
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121 struct obd_trans_info *oti, int rc, int force_sync)
123 struct filter_obd *filter = &exp->exp_obd->u.filter;
124 struct filter_export_data *fed = &exp->exp_filter_data;
125 struct lsd_client_data *lcd;
128 int err, log_pri = D_RPCTRACE;
130 /* Propagate error code. */
134 if (!exp->exp_obd->obd_replayable || oti == NULL)
137 mutex_down(&fed->fed_lcd_lock);
139 /* if the export has already been disconnected, we have no last_rcvd slot,
140 * update server data with latest transno then */
142 mutex_up(&fed->fed_lcd_lock);
143 CWARN("commit transaction for disconnected client %s: rc %d\n",
144 exp->exp_client_uuid.uuid, rc);
145 err = filter_update_server_data(exp->exp_obd,
146 filter->fo_rcvd_filp,
151 /* we don't allocate new transnos for replayed requests */
152 spin_lock(&filter->fo_translock);
153 if (oti->oti_transno == 0) {
154 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
155 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
157 last_rcvd = oti->oti_transno;
158 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
159 filter->fo_fsd->lsd_last_transno =
160 cpu_to_le64(last_rcvd);
162 oti->oti_transno = last_rcvd;
164 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
165 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
166 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
167 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
168 spin_unlock(&filter->fo_translock);
171 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
173 off = fed->fed_lr_off;
175 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
176 fed->fed_lr_idx, fed->fed_lr_off);
179 class_export_cb_get(exp); /* released when the cb is called */
181 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
187 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
188 lcd, sizeof(*lcd), &off,
189 force_sync | exp->exp_need_sync);
191 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
199 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
200 last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
201 mutex_up(&fed->fed_lcd_lock);
205 void f_dput(struct dentry *dentry)
207 /* Can't go inside filter_ddelete because it can block */
208 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
209 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
210 LASSERT(atomic_read(&dentry->d_count) > 0);
215 static void init_brw_stats(struct brw_stats *brw_stats)
218 for (i = 0; i < BRW_LAST; i++)
219 spin_lock_init(&brw_stats->hist[i].oh_lock);
222 static int lprocfs_init_rw_stats(struct obd_device *obd,
223 struct lprocfs_stats **stats)
227 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
228 LPROC_FILTER_LAST - 1;
229 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
233 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
234 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
235 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
236 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
237 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
242 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
243 plus the procfs overhead :( */
244 static int filter_export_stats_init(struct obd_device *obd,
245 struct obd_export *exp,
251 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
252 /* Self-export gets no proc entry */
255 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
257 /* Mask error for already created
265 struct nid_stat *tmp = exp->exp_nid_stats;
266 LASSERT(tmp != NULL);
268 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
269 if (tmp->nid_brw_stats == NULL)
272 init_brw_stats(tmp->nid_brw_stats);
273 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
274 0644, &filter_per_nid_stats_fops,
277 CWARN("Error adding the brw_stats file\n");
279 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
283 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
287 /* Always add in ldlm_stats */
288 tmp->nid_ldlm_stats =
289 lprocfs_alloc_stats(LDLM_LAST_OPC - LDLM_FIRST_OPC,
290 LPROCFS_STATS_FLAG_NOPERCPU);
291 if (tmp->nid_ldlm_stats == NULL)
294 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
295 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
296 tmp->nid_ldlm_stats);
304 /* Add client data to the FILTER. We use a bitmap to locate a free space
305 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
306 * Otherwise, we have just read the data from the last_rcvd file and
307 * we know its offset. */
308 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
311 struct filter_obd *filter = &obd->u.filter;
312 struct filter_export_data *fed = &exp->exp_filter_data;
313 unsigned long *bitmap = filter->fo_last_rcvd_slots;
314 int new_client = (cl_idx == -1);
318 LASSERT(bitmap != NULL);
319 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
322 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
325 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
326 * there's no need for extra complication here
329 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
331 if (cl_idx >= LR_MAX_CLIENTS) {
332 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
336 if (test_and_set_bit(cl_idx, bitmap)) {
337 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
342 if (test_and_set_bit(cl_idx, bitmap)) {
343 CERROR("FILTER client %d: bit already set in bitmap!\n",
349 fed->fed_lr_idx = cl_idx;
350 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
351 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
352 init_mutex(&fed->fed_lcd_lock);
353 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
355 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
356 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
359 struct lvfs_run_ctxt saved;
360 loff_t off = fed->fed_lr_off;
364 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
365 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
367 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
368 /* Transaction needed to fix bug 1403 */
369 handle = fsfilt_start(obd,
370 filter->fo_rcvd_filp->f_dentry->d_inode,
371 FSFILT_OP_SETATTR, NULL);
372 if (IS_ERR(handle)) {
373 rc = PTR_ERR(handle);
374 CERROR("unable to start transaction: rc %d\n", rc);
376 fed->fed_lcd->lcd_last_epoch =
377 filter->fo_fsd->lsd_start_epoch;
378 exp->exp_last_request_time = cfs_time_current_sec();
379 rc = fsfilt_add_journal_cb(obd, 0, handle,
380 target_client_add_cb,
381 class_export_cb_get(exp));
383 spin_lock(&exp->exp_lock);
384 exp->exp_need_sync = 1;
385 spin_unlock(&exp->exp_lock);
387 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
389 sizeof(*fed->fed_lcd),
390 &off, rc /* sync if no cb */);
392 filter->fo_rcvd_filp->f_dentry->d_inode,
395 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
398 CERROR("error writing %s client idx %u: rc %d\n",
399 LAST_RCVD, fed->fed_lr_idx, rc);
406 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
408 static int filter_client_free(struct obd_export *exp)
410 struct filter_export_data *fed = &exp->exp_filter_data;
411 struct filter_obd *filter = &exp->exp_obd->u.filter;
412 struct obd_device *obd = exp->exp_obd;
413 struct lvfs_run_ctxt saved;
414 struct lsd_client_data *lcd = fed->fed_lcd;
422 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
423 if (strcmp(lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
426 LASSERT(filter->fo_last_rcvd_slots != NULL);
428 off = fed->fed_lr_off;
430 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
431 fed->fed_lr_idx, fed->fed_lr_off, lcd->lcd_uuid);
433 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
434 * we leak a client slot that will be cleaned on the next recovery. */
436 CERROR("%s: client idx %d has med_off %lld\n",
437 obd->obd_name, fed->fed_lr_idx, off);
438 GOTO(free, rc = -EINVAL);
441 /* Clear the bit _after_ zeroing out the client so we don't
442 race with filter_client_add and zero out new clients.*/
443 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
444 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
449 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
450 /* Make sure the server's last_transno is up to date.
451 * This should be done before zeroing client slot so last_transno will
452 * be in server data or in client data in case of failure */
453 filter_update_server_data(obd, filter->fo_rcvd_filp, filter->fo_fsd);
455 mutex_down(&fed->fed_lcd_lock);
456 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
457 sizeof(zero_lcd), &off, 0);
459 mutex_up(&fed->fed_lcd_lock);
460 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
462 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
463 "zero out client %s at idx %u/%llu in %s, rc %d\n",
464 lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
467 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
468 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
475 mutex_down(&fed->fed_lcd_lock);
477 mutex_up(&fed->fed_lcd_lock);
483 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
484 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
485 struct filter_mod_data *fmd)
487 LASSERT_SPIN_LOCKED(&fed->fed_lock);
488 if (--fmd->fmd_refcount == 0) {
489 /* XXX when we have persistent reservations and the handle
490 * is stored herein we need to drop it here. */
491 fed->fed_mod_count--;
492 list_del(&fmd->fmd_list);
493 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
497 /* drop fmd reference, free it if last ref */
498 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
500 struct filter_export_data *fed;
505 fed = &exp->exp_filter_data;
506 spin_lock(&fed->fed_lock);
507 filter_fmd_put_nolock(fed, fmd); /* caller reference */
508 spin_unlock(&fed->fed_lock);
511 /* expire entries from the end of the list if there are too many
512 * or they are too old */
513 static void filter_fmd_expire_nolock(struct filter_obd *filter,
514 struct filter_export_data *fed,
515 struct filter_mod_data *keep)
517 struct filter_mod_data *fmd, *tmp;
519 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
523 if (time_before(jiffies, fmd->fmd_expire) &&
524 fed->fed_mod_count < filter->fo_fmd_max_num)
527 list_del_init(&fmd->fmd_list);
528 filter_fmd_put_nolock(fed, fmd); /* list reference */
532 void filter_fmd_expire(struct obd_export *exp)
534 spin_lock(&exp->exp_filter_data.fed_lock);
535 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
536 &exp->exp_filter_data, NULL);
537 spin_unlock(&exp->exp_filter_data.fed_lock);
540 /* find specified objid, group in export fmd list.
541 * caller must hold fed_lock and take fmd reference itself */
542 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
543 struct filter_export_data *fed,
544 obd_id objid, obd_gr group)
546 struct filter_mod_data *found = NULL, *fmd;
548 LASSERT_SPIN_LOCKED(&fed->fed_lock);
550 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
551 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
553 list_del(&fmd->fmd_list);
554 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
555 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
560 filter_fmd_expire_nolock(filter, fed, found);
565 /* Find fmd based on objid and group, or return NULL if not found. */
566 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
567 obd_id objid, obd_gr group)
569 struct filter_mod_data *fmd;
571 spin_lock(&exp->exp_filter_data.fed_lock);
572 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
573 &exp->exp_filter_data, objid, group);
575 fmd->fmd_refcount++; /* caller reference */
576 spin_unlock(&exp->exp_filter_data.fed_lock);
581 /* Find fmd based on objid and group, or create a new one if none is found.
582 * It is possible for this function to return NULL under memory pressure,
583 * or if objid = 0 is passed (which will only cause old entries to expire).
584 * Currently this is not fatal because any fmd state is transient and
585 * may also be freed when it gets sufficiently old. */
586 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
587 obd_id objid, obd_gr group)
589 struct filter_export_data *fed = &exp->exp_filter_data;
590 struct filter_mod_data *found = NULL, *fmd_new = NULL;
592 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
594 spin_lock(&fed->fed_lock);
595 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
598 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
599 fmd_new->fmd_id = objid;
600 fmd_new->fmd_gr = group;
601 fmd_new->fmd_refcount++; /* list reference */
603 fed->fed_mod_count++;
605 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
609 found->fmd_refcount++; /* caller reference */
610 found->fmd_expire = jiffies +
611 exp->exp_obd->u.filter.fo_fmd_max_age;
614 spin_unlock(&fed->fed_lock);
620 /* drop fmd list reference so it will disappear when last reference is put.
621 * This isn't so critical because it would in fact only affect the one client
622 * that is doing the unlink and at worst we have an stale entry referencing
623 * an object that should never be used again. */
624 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
626 struct filter_mod_data *found = NULL;
628 spin_lock(&exp->exp_filter_data.fed_lock);
629 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
631 list_del_init(&found->fmd_list);
632 filter_fmd_put_nolock(&exp->exp_filter_data, found);
634 spin_unlock(&exp->exp_filter_data.fed_lock);
637 #define filter_fmd_drop(exp, objid, group)
640 /* remove all entries from fmd list */
641 static void filter_fmd_cleanup(struct obd_export *exp)
643 struct filter_export_data *fed = &exp->exp_filter_data;
644 struct filter_mod_data *fmd = NULL, *tmp;
646 spin_lock(&fed->fed_lock);
647 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
648 list_del_init(&fmd->fmd_list);
649 filter_fmd_put_nolock(fed, fmd);
651 spin_unlock(&fed->fed_lock);
654 static int filter_init_export(struct obd_export *exp)
656 spin_lock_init(&exp->exp_filter_data.fed_lock);
657 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
659 spin_lock(&exp->exp_lock);
660 exp->exp_connecting = 1;
661 spin_unlock(&exp->exp_lock);
663 return ldlm_init_export(exp);
666 static int filter_free_server_data(struct filter_obd *filter)
668 OBD_FREE_PTR(filter->fo_fsd);
669 filter->fo_fsd = NULL;
670 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
671 filter->fo_last_rcvd_slots = NULL;
675 /* assumes caller is already in kernel ctxt */
676 int filter_update_server_data(struct obd_device *obd, struct file *filp,
677 struct lr_server_data *fsd)
683 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
684 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
685 le64_to_cpu(fsd->lsd_last_transno));
686 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
687 le64_to_cpu(fsd->lsd_mount_count));
689 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, 0);
691 CERROR("error writing lr_server_data: rc = %d\n", rc);
696 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
699 struct filter_obd *filter = &obd->u.filter;
705 if (filter->fo_last_objid_files[group] == NULL) {
706 CERROR("Object group "LPU64" not fully setup; not updating "
707 "last_objid\n", group);
711 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
712 obd->obd_name, group, filter->fo_last_objids[group]);
714 tmp = cpu_to_le64(filter->fo_last_objids[group]);
715 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
716 &tmp, sizeof(tmp), &off, force_sync);
718 CERROR("error writing group "LPU64" last objid: rc = %d\n",
722 extern int ost_handle(struct ptlrpc_request *req);
723 /* assumes caller has already in kernel ctxt */
724 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
726 struct filter_obd *filter = &obd->u.filter;
727 struct lr_server_data *fsd;
728 struct lsd_client_data *lcd = NULL;
729 struct inode *inode = filp->f_dentry->d_inode;
730 unsigned long last_rcvd_size = i_size_read(inode);
737 /* ensure padding in the struct is the correct size */
738 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
739 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
740 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
741 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
743 OBD_ALLOC(fsd, sizeof(*fsd));
746 filter->fo_fsd = fsd;
748 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
749 if (filter->fo_last_rcvd_slots == NULL) {
750 OBD_FREE(fsd, sizeof(*fsd));
754 if (last_rcvd_size == 0) {
755 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
757 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
758 fsd->lsd_last_transno = 0;
759 mount_count = fsd->lsd_mount_count = 0;
760 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
761 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
762 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
763 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
764 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
765 /* OBD_COMPAT_OST is set in filter_connect_internal when the
766 * MDS first connects and assigns the OST index number. */
767 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
770 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
772 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
776 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
777 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
778 "using the wrong disk %s. Were the "
779 "/dev/ assignments rearranged?\n",
780 obd->obd_uuid.uuid, fsd->lsd_uuid);
781 GOTO(err_fsd, rc = -EINVAL);
783 mount_count = le64_to_cpu(fsd->lsd_mount_count);
784 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
786 /* Assume old last_rcvd format unless I_C_LR is set */
787 if (!(fsd->lsd_feature_incompat &
788 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
789 fsd->lsd_last_transno = fsd->lsd_compat14;
791 /* OBD_COMPAT_OST is set in filter_connect_internal when the
792 * MDS first connects and assigns the OST index number. */
793 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
797 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
798 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
799 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
800 ~FILTER_INCOMPAT_SUPP);
801 GOTO(err_fsd, rc = -EINVAL);
803 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
804 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
805 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
806 ~FILTER_ROCOMPAT_SUPP);
807 /* Do something like remount filesystem read-only */
808 GOTO(err_fsd, rc = -EINVAL);
811 start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
813 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
814 obd->obd_name, start_epoch);
815 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
816 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
817 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
818 obd->obd_name, mount_count + 1);
819 CDEBUG(D_INODE, "%s: server data size: %u\n",
820 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
821 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
822 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
823 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
824 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
825 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
826 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
827 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
828 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
829 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
830 le16_to_cpu(fsd->lsd_client_size));
832 if (!obd->obd_replayable) {
833 CWARN("%s: recovery support OFF\n", obd->obd_name);
837 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
838 off < last_rcvd_size; cl_idx++) {
840 struct obd_export *exp;
841 struct filter_export_data *fed;
846 GOTO(err_client, rc = -ENOMEM);
849 /* Don't assume off is incremented properly by
850 * fsfilt_read_record(), in case sizeof(*lcd)
851 * isn't the same as fsd->lsd_client_size. */
852 off = le32_to_cpu(fsd->lsd_client_start) +
853 cl_idx * le16_to_cpu(fsd->lsd_client_size);
854 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
856 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
857 LAST_RCVD, cl_idx, off, rc);
858 break; /* read error shouldn't cause startup to fail */
861 if (lcd->lcd_uuid[0] == '\0') {
862 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
867 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
869 /* These exports are cleaned up by filter_disconnect(), so they
870 * need to be set up like real exports as filter_connect() does.
872 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
874 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
875 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
876 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
878 if (PTR_ERR(exp) == -EALREADY) {
879 /* export already exists, zero out this one */
880 CERROR("Zeroing out duplicate export due to "
882 lcd->lcd_uuid[0] = '\0';
884 GOTO(err_client, rc = PTR_ERR(exp));
887 fed = &exp->exp_filter_data;
889 fed->fed_group = 0; /* will be assigned at connect */
890 filter_export_stats_init(obd, exp, NULL);
891 rc = filter_client_add(obd, exp, cl_idx);
892 /* can't fail for existing client */
893 LASSERTF(rc == 0, "rc = %d\n", rc);
895 /* VBR: set export last committed */
896 exp->exp_last_committed = last_rcvd;
897 spin_lock(&exp->exp_lock);
898 exp->exp_connecting = 0;
899 exp->exp_in_recovery = 0;
900 spin_unlock(&exp->exp_lock);
901 spin_lock_bh(&obd->obd_processing_task_lock);
902 obd->obd_max_recoverable_clients++;
903 spin_unlock_bh(&obd->obd_processing_task_lock);
905 class_export_put(exp);
908 /* Need to check last_rcvd even for duplicated exports. */
909 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
912 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
913 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
919 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
921 filter->fo_mount_count = mount_count + 1;
922 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
924 /* save it, so mount count and last_transno is current */
925 rc = filter_update_server_data(obd, filp, filter->fo_fsd);
927 GOTO(err_client, rc);
932 class_disconnect_exports(obd);
934 filter_free_server_data(filter);
938 static int filter_cleanup_groups(struct obd_device *obd)
940 struct filter_obd *filter = &obd->u.filter;
942 struct dentry *dentry;
946 if (filter->fo_dentry_O_groups != NULL) {
947 for (i = 0; i < filter->fo_group_count; i++) {
948 dentry = filter->fo_dentry_O_groups[i];
952 OBD_FREE(filter->fo_dentry_O_groups,
953 filter->fo_group_count *
954 sizeof(*filter->fo_dentry_O_groups));
955 filter->fo_dentry_O_groups = NULL;
957 if (filter->fo_last_objid_files != NULL) {
958 for (i = 0; i < filter->fo_group_count; i++) {
959 filp = filter->fo_last_objid_files[i];
963 OBD_FREE(filter->fo_last_objid_files,
964 filter->fo_group_count *
965 sizeof(*filter->fo_last_objid_files));
966 filter->fo_last_objid_files = NULL;
968 if (filter->fo_dentry_O_sub != NULL) {
969 for (i = 0; i < filter->fo_group_count; i++) {
970 for (j = 0; j < filter->fo_subdir_count; j++) {
971 dentry = filter->fo_dentry_O_sub[i].dentry[j];
976 OBD_FREE(filter->fo_dentry_O_sub,
977 filter->fo_group_count *
978 sizeof(*filter->fo_dentry_O_sub));
979 filter->fo_dentry_O_sub = NULL;
981 if (filter->fo_last_objids != NULL) {
982 OBD_FREE(filter->fo_last_objids,
983 filter->fo_group_count *
984 sizeof(*filter->fo_last_objids));
985 filter->fo_last_objids = NULL;
987 if (filter->fo_dentry_O != NULL) {
988 f_dput(filter->fo_dentry_O);
989 filter->fo_dentry_O = NULL;
994 static int filter_update_last_group(struct obd_device *obd, int group)
996 struct filter_obd *filter = &obd->u.filter;
997 struct file *filp = NULL;
998 int last_group = 0, rc;
1002 if (group <= filter->fo_committed_group)
1005 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1009 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1013 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1015 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1019 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1020 obd->obd_name, last_group, group);
1024 /* must be sync: bXXXX */
1025 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1027 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1031 filter->fo_committed_group = group;
1034 filp_close(filp, 0);
1038 static int filter_read_group_internal(struct obd_device *obd, int group,
1041 struct filter_obd *filter = &obd->u.filter;
1042 __u64 *new_objids = NULL;
1043 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1044 struct dentry **new_groups = NULL;
1045 struct file **new_files = NULL;
1046 struct dentry *dentry;
1048 int old_count = filter->fo_group_count, rc, stage = 0, i;
1052 int len = group + 1;
1054 snprintf(name, 24, "%d", group);
1058 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1060 if (IS_ERR(dentry)) {
1061 CERROR("Cannot lookup expected object group %d: %ld\n",
1062 group, PTR_ERR(dentry));
1063 RETURN(PTR_ERR(dentry));
1066 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1068 if (IS_ERR(dentry)) {
1069 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1071 RETURN(PTR_ERR(dentry));
1076 snprintf(name, 24, "O/%d/LAST_ID", group);
1078 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1080 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1081 GOTO(cleanup, rc = PTR_ERR(filp));
1085 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1087 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1091 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1092 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1093 if (tmp_subdirs == NULL)
1094 GOTO(cleanup, rc = -ENOMEM);
1097 for (i = 0; i < filter->fo_subdir_count; i++) {
1099 snprintf(dir, sizeof(dir), "d%u", i);
1101 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1104 if (IS_ERR(tmp_subdirs->dentry[i])) {
1105 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1106 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1111 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1112 tmp_subdirs->dentry[i]);
1116 /* 'group' is an index; we need an array of length 'group + 1' */
1117 if (group + 1 > old_count) {
1118 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1119 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1120 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1121 OBD_ALLOC(new_files, len * sizeof(*new_files));
1123 if (new_objids == NULL || new_subdirs == NULL ||
1124 new_groups == NULL || new_files == NULL)
1125 GOTO(cleanup, rc = -ENOMEM);
1128 memcpy(new_objids, filter->fo_last_objids,
1129 old_count * sizeof(*new_objids));
1130 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1131 old_count * sizeof(*new_subdirs));
1132 memcpy(new_groups, filter->fo_dentry_O_groups,
1133 old_count * sizeof(*new_groups));
1134 memcpy(new_files, filter->fo_last_objid_files,
1135 old_count * sizeof(*new_files));
1137 OBD_FREE(filter->fo_last_objids,
1138 old_count * sizeof(*new_objids));
1139 OBD_FREE(filter->fo_dentry_O_sub,
1140 old_count * sizeof(*new_subdirs));
1141 OBD_FREE(filter->fo_dentry_O_groups,
1142 old_count * sizeof(*new_groups));
1143 OBD_FREE(filter->fo_last_objid_files,
1144 old_count * sizeof(*new_files));
1146 filter->fo_last_objids = new_objids;
1147 filter->fo_dentry_O_sub = new_subdirs;
1148 filter->fo_dentry_O_groups = new_groups;
1149 filter->fo_last_objid_files = new_files;
1150 filter->fo_group_count = len;
1153 filter->fo_dentry_O_groups[group] = dentry;
1154 filter->fo_last_objid_files[group] = filp;
1155 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1156 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1157 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1160 filter_update_last_group(obd, group);
1162 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1163 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1164 rc = filter_update_last_objid(obd, group, 1);
1168 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1169 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1170 obd->obd_name, group, last_objid);
1175 if (new_objids != NULL)
1176 OBD_FREE(new_objids, len * sizeof(*new_objids));
1177 if (new_subdirs != NULL)
1178 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1179 if (new_groups != NULL)
1180 OBD_FREE(new_groups, len * sizeof(*new_groups));
1181 if (new_files != NULL)
1182 OBD_FREE(new_files, len * sizeof(*new_files));
1184 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1185 for (i = 0; i < filter->fo_subdir_count; i++) {
1186 if (tmp_subdirs->dentry[i] != NULL)
1187 dput(tmp_subdirs->dentry[i]);
1189 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1192 filp_close(filp, 0);
1199 static int filter_read_groups(struct obd_device *obd, int last_group,
1202 struct filter_obd *filter = &obd->u.filter;
1203 int old_count, group, rc = 0;
1205 down(&filter->fo_init_lock);
1206 old_count = filter->fo_group_count;
1207 for (group = old_count; group <= last_group; group++) {
1208 rc = filter_read_group_internal(obd, group, create);
1212 up(&filter->fo_init_lock);
1216 /* FIXME: object groups */
1217 static int filter_prep_groups(struct obd_device *obd)
1219 struct filter_obd *filter = &obd->u.filter;
1220 struct dentry *O_dentry;
1222 int last_group, rc = 0, cleanup_phase = 0;
1226 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1228 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1229 if (IS_ERR(O_dentry)) {
1230 rc = PTR_ERR(O_dentry);
1231 CERROR("cannot open/create O: rc = %d\n", rc);
1234 filter->fo_dentry_O = O_dentry;
1235 cleanup_phase = 1; /* O_dentry */
1237 /* we have to initialize all groups before first connections from
1238 * clients because they may send create/destroy for any group -bzzz */
1239 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1241 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1242 GOTO(cleanup, rc = PTR_ERR(filp));
1244 cleanup_phase = 2; /* filp */
1246 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1248 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1253 last_group = FILTER_GROUP_MDS0;
1255 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1256 FILTER_GROUP_MDS0, last_group);
1257 filter->fo_committed_group = last_group;
1258 rc = filter_read_groups(obd, last_group, 1);
1262 filp_close(filp, 0);
1266 switch (cleanup_phase) {
1268 filp_close(filp, 0);
1270 filter_cleanup_groups(obd);
1271 f_dput(filter->fo_dentry_O);
1272 filter->fo_dentry_O = NULL;
1280 /* setup the object store with correct subdirectories */
1281 static int filter_prep(struct obd_device *obd)
1283 struct lvfs_run_ctxt saved;
1284 struct filter_obd *filter = &obd->u.filter;
1286 struct inode *inode;
1290 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1291 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1292 if (!file || IS_ERR(file)) {
1294 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1298 filter->fo_rcvd_filp = file;
1299 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1300 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1301 file->f_dentry->d_inode->i_mode);
1302 GOTO(err_filp, rc = -ENOENT);
1305 inode = file->f_dentry->d_parent->d_inode;
1306 /* We use i_op->unlink directly in filter_vfs_unlink() */
1307 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1308 CERROR("%s: filesystem does not support create/unlink ops\n",
1310 GOTO(err_filp, rc = -EOPNOTSUPP);
1313 /** lu_target has very limited use in filter now */
1314 lut_init(NULL, &filter->fo_lut, obd, NULL);
1316 rc = filter_init_server_data(obd, file);
1318 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1322 target_recovery_init(&filter->fo_lut, ost_handle);
1324 /* open and create health check io file*/
1325 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1328 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1330 GOTO(err_server_data, rc);
1332 filter->fo_obt.obt_health_check_filp = file;
1333 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1334 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1335 file->f_dentry->d_inode->i_mode);
1336 GOTO(err_health_check, rc = -ENOENT);
1338 rc = lvfs_check_io_health(obd, file);
1340 GOTO(err_health_check, rc);
1342 rc = filter_prep_groups(obd);
1344 GOTO(err_health_check, rc);
1346 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1351 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1352 CERROR("can't close %s after error\n", HEALTH_CHECK);
1353 filter->fo_obt.obt_health_check_filp = NULL;
1355 target_recovery_fini(obd);
1356 filter_free_server_data(filter);
1358 if (filp_close(filter->fo_rcvd_filp, 0))
1359 CERROR("can't close %s after error\n", LAST_RCVD);
1360 filter->fo_rcvd_filp = NULL;
1364 /* cleanup the filter: write last used object id to status file */
1365 static void filter_post(struct obd_device *obd)
1367 struct lvfs_run_ctxt saved;
1368 struct filter_obd *filter = &obd->u.filter;
1371 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1372 * best to start a transaction with h_sync, because we removed this
1375 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1376 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1379 CERROR("error writing server data: rc = %d\n", rc);
1381 for (i = 0; i < filter->fo_group_count; i++) {
1382 rc = filter_update_last_objid(obd, i,
1383 (i == filter->fo_group_count - 1));
1385 CERROR("error writing group %d lastobjid: rc = %d\n",
1389 rc = filp_close(filter->fo_rcvd_filp, 0);
1390 filter->fo_rcvd_filp = NULL;
1392 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1394 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1395 filter->fo_obt.obt_health_check_filp = NULL;
1397 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1399 filter_cleanup_groups(obd);
1400 filter_free_server_data(filter);
1401 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1403 filter_free_capa_keys(filter);
1404 cleanup_capa_hash(filter->fo_capa_hash);
1407 static void filter_set_last_id(struct filter_obd *filter,
1408 obd_id id, obd_gr group)
1410 LASSERT(filter->fo_fsd != NULL);
1411 LASSERT(group <= filter->fo_group_count);
1413 spin_lock(&filter->fo_objidlock);
1414 filter->fo_last_objids[group] = id;
1415 spin_unlock(&filter->fo_objidlock);
1418 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1421 LASSERT(filter->fo_fsd != NULL);
1422 LASSERT(group <= filter->fo_group_count);
1424 /* FIXME: object groups */
1425 spin_lock(&filter->fo_objidlock);
1426 id = filter->fo_last_objids[group];
1427 spin_unlock(&filter->fo_objidlock);
1431 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1433 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1437 /* We never dget the object parent, so DON'T dput it either */
1438 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1440 struct filter_obd *filter = &obd->u.filter;
1441 struct filter_subdirs *subdirs;
1442 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1444 if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1445 return filter->fo_dentry_O_groups[group];
1447 subdirs = &filter->fo_dentry_O_sub[group];
1448 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1451 /* We never dget the object parent, so DON'T dput it either */
1452 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1455 unsigned long now = jiffies;
1456 struct dentry *dparent = filter_parent(obd, group, objid);
1459 if (IS_ERR(dparent))
1461 if (dparent == NULL)
1462 return ERR_PTR(-ENOENT);
1464 rc = filter_lock_dentry(obd, dparent);
1465 fsfilt_check_slow(obd, now, "parent lock");
1466 return rc ? ERR_PTR(rc) : dparent;
1469 /* We never dget the object parent, so DON'T dput it either */
1470 static void filter_parent_unlock(struct dentry *dparent)
1472 UNLOCK_INODE_MUTEX(dparent->d_inode);
1475 /* How to get files, dentries, inodes from object id's.
1477 * If dir_dentry is passed, the caller has already locked the parent
1478 * appropriately for this operation (normally a write lock). If
1479 * dir_dentry is NULL, we do a read lock while we do the lookup to
1480 * avoid races with create/destroy and such changing the directory
1481 * internal to the filesystem code. */
1482 struct dentry *filter_fid2dentry(struct obd_device *obd,
1483 struct dentry *dir_dentry,
1484 obd_gr group, obd_id id)
1486 struct dentry *dparent = dir_dentry;
1487 struct dentry *dchild;
1492 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1493 obd->u.filter.fo_destroys_in_progress == 0) {
1494 /* don't fail lookups for orphan recovery, it causes
1495 * later LBUGs when objects still exist during precreate */
1496 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1497 RETURN(ERR_PTR(-ENOENT));
1500 CERROR("fatal: invalid object id 0\n");
1501 RETURN(ERR_PTR(-ESTALE));
1504 len = sprintf(name, LPU64, id);
1505 if (dir_dentry == NULL) {
1506 dparent = filter_parent_lock(obd, group, id);
1507 if (IS_ERR(dparent)) {
1508 CERROR("%s: error getting object "LPU64":"LPU64
1509 " parent: rc %ld\n", obd->obd_name,
1510 id, group, PTR_ERR(dparent));
1514 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1515 dparent->d_name.len, dparent->d_name.name, name);
1516 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1517 if (dir_dentry == NULL)
1518 filter_parent_unlock(dparent);
1519 if (IS_ERR(dchild)) {
1520 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1525 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1526 CERROR("%s: got bad object "LPU64" inode %lu\n",
1527 obd->obd_name, id, dchild->d_inode->i_ino);
1529 RETURN(ERR_PTR(-ENOENT));
1532 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1533 name, dchild, atomic_read(&dchild->d_count));
1535 LASSERT(atomic_read(&dchild->d_count) > 0);
1540 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1541 obd_id group, struct lustre_handle *lockh)
1543 int flags = LDLM_AST_DISCARD_DATA, rc;
1544 struct ldlm_res_id res_id;
1545 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1548 osc_build_res_name(objid, group, &res_id);
1549 /* Tell the clients that the object is gone now and that they should
1550 * throw away any cached pages. */
1551 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1552 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1553 ldlm_completion_ast, NULL, NULL, 0, NULL,
1560 static void filter_fini_destroy(struct obd_device *obd,
1561 struct lustre_handle *lockh)
1564 ldlm_lock_decref(lockh, LCK_PW);
1567 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1568 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1569 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1570 * reverse. Caller must take i_sem before starting the transaction and we
1571 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1572 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1573 struct vfsmount *mnt)
1578 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1579 * itself we already hold dir->i_mutex for child create/unlink ops */
1580 LASSERT(dentry->d_inode != NULL);
1581 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1582 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1586 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1587 GOTO(out, rc = -ENOENT);
1589 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1594 GOTO(out, rc = -EPERM);
1596 /* check_sticky() */
1597 if ((dentry->d_inode->i_uid != current->fsuid &&
1598 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1599 IS_IMMUTABLE(dentry->d_inode))
1600 GOTO(out, rc = -EPERM);
1602 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1603 * that was done because of journal_start (which is already done
1604 * here) or some other ordering issue. */
1607 rc = ll_security_inode_unlink(dir, dentry, mnt);
1611 rc = dir->i_op->unlink(dir, dentry);
1613 /* need to drop i_mutex before we lose inode reference */
1614 UNLOCK_INODE_MUTEX(dentry->d_inode);
1621 /* Caller must hold LCK_PW on parent and push us into kernel context.
1622 * Caller must hold child i_mutex, we drop it always.
1623 * Caller is also required to ensure that dchild->d_inode exists. */
1624 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1625 obd_gr group, struct dentry *dparent,
1626 struct dentry *dchild)
1628 struct inode *inode = dchild->d_inode;
1629 struct filter_obd *filter = &obd->u.filter;
1632 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1633 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1634 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1635 (unsigned long)inode->i_nlink,
1636 atomic_read(&inode->i_count));
1639 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1641 CERROR("error unlinking objid %.*s: rc %d\n",
1642 dchild->d_name.len, dchild->d_name.name, rc);
1646 struct filter_intent_args {
1647 struct ldlm_lock **victim;
1652 static enum interval_iter filter_intent_cb(struct interval_node *n,
1655 struct ldlm_interval *node = (struct ldlm_interval *)n;
1656 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1657 __u64 size = arg->size;
1658 struct ldlm_lock **v = arg->victim;
1659 struct ldlm_lock *lck;
1661 /* If the interval is lower than the current file size,
1663 if (interval_high(n) <= size)
1664 return INTERVAL_ITER_STOP;
1666 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1667 /* Don't send glimpse ASTs to liblustre clients.
1668 * They aren't listening for them, and they do
1669 * entirely synchronous I/O anyways. */
1670 if (lck->l_export == NULL ||
1671 lck->l_export->exp_libclient == 1)
1674 if (*arg->liblustre)
1675 *arg->liblustre = 0;
1678 *v = LDLM_LOCK_GET(lck);
1679 } else if ((*v)->l_policy_data.l_extent.start <
1680 lck->l_policy_data.l_extent.start) {
1681 LDLM_LOCK_RELEASE(*v);
1682 *v = LDLM_LOCK_GET(lck);
1685 /* the same policy group - every lock has the
1686 * same extent, so needn't do it any more */
1690 return INTERVAL_ITER_CONT;
1693 static int filter_intent_policy(struct ldlm_namespace *ns,
1694 struct ldlm_lock **lockp, void *req_cookie,
1695 ldlm_mode_t mode, int flags, void *data)
1697 CFS_LIST_HEAD(rpc_list);
1698 struct ptlrpc_request *req = req_cookie;
1699 struct ldlm_lock *lock = *lockp, *l = NULL;
1700 struct ldlm_resource *res = lock->l_resource;
1701 ldlm_processing_policy policy;
1702 struct ost_lvb *res_lvb, *reply_lvb;
1703 struct ldlm_reply *rep;
1705 int idx, rc, tmpflags = 0, only_liblustre = 1;
1706 struct ldlm_interval_tree *tree;
1707 struct filter_intent_args arg;
1708 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1709 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1710 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1713 policy = ldlm_get_processing_policy(res);
1714 LASSERT(policy != NULL);
1715 LASSERT(req != NULL);
1717 rc = lustre_pack_reply(req, 3, repsize, NULL);
1719 RETURN(req->rq_status = rc);
1721 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1722 LASSERT(rep != NULL);
1724 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1725 sizeof(*reply_lvb));
1726 LASSERT(reply_lvb != NULL);
1728 //fixup_handle_for_resent_req(req, lock, &lockh);
1730 /* Call the extent policy function to see if our request can be
1731 * granted, or is blocked.
1732 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1733 * lock, and should not be granted if the lock will be blocked.
1736 LASSERT(ns == res->lr_namespace);
1738 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1739 check_res_locked(res);
1741 /* FIXME: we should change the policy function slightly, to not make
1742 * this list at all, since we just turn around and free it */
1743 while (!list_empty(&rpc_list)) {
1744 struct ldlm_lock *wlock =
1745 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1746 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1747 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1748 lock->l_flags &= ~LDLM_FL_CP_REQD;
1749 list_del_init(&wlock->l_cp_ast);
1750 LDLM_LOCK_RELEASE(wlock);
1753 /* The lock met with no resistance; we're finished. */
1754 if (rc == LDLM_ITER_CONTINUE) {
1755 /* do not grant locks to the liblustre clients: they cannot
1756 * handle ASTs robustly. We need to do this while still
1757 * holding ns_lock to avoid the lock remaining on the res_link
1758 * list (and potentially being added to l_pending_list by an
1759 * AST) when we are going to drop this lock ASAP. */
1760 if (lock->l_export->exp_libclient ||
1761 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1762 ldlm_resource_unlink_lock(lock);
1763 err = ELDLM_LOCK_ABORTED;
1765 err = ELDLM_LOCK_REPLACED;
1771 /* Do not grant any lock, but instead send GL callbacks. The extent
1772 * policy nicely created a list of all PW locks for us. We will choose
1773 * the highest of those which are larger than the size in the LVB, if
1774 * any, and perform a glimpse callback. */
1775 res_lvb = res->lr_lvb_data;
1776 LASSERT(res_lvb != NULL);
1777 *reply_lvb = *res_lvb;
1780 * ->ns_lock guarantees that no new locks are granted, and,
1781 * therefore, that res->lr_lvb_data cannot increase beyond the
1782 * end of already granted lock. As a result, it is safe to
1783 * check against "stale" reply_lvb->lvb_size value without
1786 arg.size = reply_lvb->lvb_size;
1788 arg.liblustre = &only_liblustre;
1789 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1790 tree = &res->lr_itree[idx];
1791 if (tree->lit_mode == LCK_PR)
1794 interval_iterate_reverse(tree->lit_root,
1795 filter_intent_cb, &arg);
1799 /* There were no PW locks beyond the size in the LVB; finished. */
1801 if (only_liblustre) {
1802 /* If we discovered a liblustre client with a PW lock,
1803 * however, the LVB may be out of date! The LVB is
1804 * updated only on glimpse (which we don't do for
1805 * liblustre clients) and cancel (which the client
1806 * obviously has not yet done). So if it has written
1807 * data but kept the lock, the LVB is stale and needs
1808 * to be updated from disk.
1810 * Of course, this will all disappear when we switch to
1811 * taking liblustre locks on the OST. */
1812 ldlm_res_lvbo_update(res, NULL, 1);
1814 RETURN(ELDLM_LOCK_ABORTED);
1818 * This check is for lock taken in filter_prepare_destroy() that does
1819 * not have l_glimpse_ast set. So the logic is: if there is a lock
1820 * with no l_glimpse_ast set, this object is being destroyed already.
1822 * Hence, if you are grabbing DLM locks on the server, always set
1823 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1825 if (l->l_glimpse_ast == NULL) {
1826 /* We are racing with unlink(); just return -ENOENT */
1827 rep->lock_policy_res1 = -ENOENT;
1831 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1832 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1833 /* Update the LVB from disk if the AST failed (this is a legal race) */
1835 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1836 * sending ast is not handled. This can result in lost client writes.
1839 ldlm_res_lvbo_update(res, NULL, 1);
1842 *reply_lvb = *res_lvb;
1846 LDLM_LOCK_RELEASE(l);
1848 RETURN(ELDLM_LOCK_ABORTED);
1852 * per-obd_device iobuf pool.
1854 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1855 * allocations in write-path has to be minimized (see bug 5137).
1857 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1858 * OST threads (see ost_thread_{init,done}()).
1860 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1861 * at the OST layer there are only (potentially) multiple obd_device of type
1862 * unknown at the time of OST thread creation.
1864 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1865 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1866 * it's very own iobuf.
1870 * filter_kiobuf_pool_init()
1872 * filter_kiobuf_pool_done()
1874 * filter_iobuf_get()
1876 * operate on this array. They are "generic" in a sense that they don't depend
1877 * on actual type of iobuf's (the latter depending on Linux kernel version).
1881 * destroy pool created by filter_iobuf_pool_init
1883 static void filter_iobuf_pool_done(struct filter_obd *filter)
1885 struct filter_iobuf **pool;
1890 pool = filter->fo_iobuf_pool;
1892 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1893 if (pool[i] != NULL)
1894 filter_free_iobuf(pool[i]);
1896 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1897 filter->fo_iobuf_pool = NULL;
1902 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1904 struct filter_obd *filter = &obd->u.filter;
1905 struct sptlrpc_rule_set tmp_rset;
1908 sptlrpc_rule_set_init(&tmp_rset);
1909 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1911 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1916 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1918 write_lock(&filter->fo_sptlrpc_lock);
1919 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1920 filter->fo_sptlrpc_rset = tmp_rset;
1921 write_unlock(&filter->fo_sptlrpc_lock);
1927 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1929 static int filter_iobuf_pool_init(struct filter_obd *filter)
1936 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1938 if (filter->fo_iobuf_pool == NULL)
1941 filter->fo_iobuf_count = OSS_THREADS_MAX;
1946 /* Return iobuf allocated for @thread_id. We don't know in advance how
1947 * many threads there will be so we allocate a large empty array and only
1948 * fill in those slots that are actually in use.
1949 * If we haven't allocated a pool entry for this thread before, do so now. */
1950 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1952 int thread_id = (oti && oti->oti_thread) ?
1953 oti->oti_thread->t_id : -1;
1954 struct filter_iobuf *pool = NULL;
1955 struct filter_iobuf **pool_place = NULL;
1957 if (thread_id >= 0) {
1958 LASSERT(thread_id < filter->fo_iobuf_count);
1959 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1962 if (unlikely(pool == NULL)) {
1963 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1964 PTLRPC_MAX_BRW_PAGES);
1965 if (pool_place != NULL)
1972 /* mount the file system (secretly). lustre_cfg parameters are:
1975 * 3 = flags: failover=f, failout=n
1978 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1981 struct filter_obd *filter = &obd->u.filter;
1982 struct vfsmount *mnt;
1983 struct lustre_mount_info *lmi;
1984 struct obd_uuid uuid;
1992 if (lcfg->lcfg_bufcount < 3 ||
1993 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1994 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1997 lmi = server_get_mount(obd->obd_name);
1999 /* We already mounted in lustre_fill_super.
2000 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2001 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2003 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2005 /* old path - used by lctl */
2006 CERROR("Using old MDS mount method\n");
2007 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2008 MS_NOATIME|MS_NODIRATIME,
2009 lustre_cfg_string(lcfg, 1), option);
2012 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2013 lustre_cfg_string(lcfg, 1), rc);
2017 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2019 if (IS_ERR(obd->obd_fsops))
2020 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2022 rc = filter_iobuf_pool_init(filter);
2026 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2027 CERROR("%s: Underlying device is marked as read-only. "
2028 "Setup failed\n", obd->obd_name);
2029 GOTO(err_ops, rc = -EROFS);
2032 /* failover is the default */
2033 obd->obd_replayable = 1;
2035 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2036 str = lustre_cfg_string(lcfg, 3);
2037 if (strchr(str, 'n')) {
2038 CWARN("%s: recovery disabled\n", obd->obd_name);
2039 obd->obd_replayable = 0;
2043 filter->fo_vfsmnt = mnt;
2044 obd->u.obt.obt_sb = mnt->mnt_sb;
2045 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2046 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2048 fsfilt_setup(obd, obd->u.obt.obt_sb);
2050 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2051 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2052 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2053 obd->obd_lvfs_ctxt.fs = get_ds();
2054 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2056 init_mutex(&filter->fo_init_lock);
2057 filter->fo_committed_group = 0;
2058 filter->fo_destroys_in_progress = 0;
2059 for (i = 0; i < 32; i++)
2060 sema_init(&filter->fo_create_locks[i], 1);
2062 spin_lock_init(&filter->fo_translock);
2063 spin_lock_init(&filter->fo_objidlock);
2064 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2065 sema_init(&filter->fo_alloc_lock, 1);
2066 init_brw_stats(&filter->fo_filter_stats);
2067 filter->fo_read_cache = 1; /* enable read-only cache by default */
2068 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2069 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2070 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2071 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2073 rc = filter_prep(obd);
2077 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2078 spin_lock_init(&filter->fo_llog_list_lock);
2080 filter->fo_fl_oss_capa = 1;
2082 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2083 filter->fo_capa_hash = init_capa_hash();
2084 if (filter->fo_capa_hash == NULL)
2085 GOTO(err_post, rc = -ENOMEM);
2087 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2088 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2089 LDLM_NAMESPACE_GREEDY);
2090 if (obd->obd_namespace == NULL)
2091 GOTO(err_post, rc = -ENOMEM);
2092 obd->obd_namespace->ns_lvbp = obd;
2093 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2094 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2096 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2097 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2099 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2101 CERROR("failed to setup llogging subsystems\n");
2105 rwlock_init(&filter->fo_sptlrpc_lock);
2106 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2107 /* do this after llog being initialized */
2108 filter_adapt_sptlrpc_conf(obd, 1);
2110 rc = lquota_setup(filter_quota_interface_ref, obd);
2114 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2115 if (q->max_sectors < q->max_hw_sectors &&
2116 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2117 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2118 "for larger I/O requests: max_sectors = %u "
2119 "could be up to max_hw_sectors=%u\n",
2120 obd->obd_name, mnt->mnt_sb->s_id,
2121 q->max_sectors, q->max_hw_sectors);
2123 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2124 if (uuid_ptr != NULL) {
2125 class_uuid_unparse(uuid_ptr, &uuid);
2131 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2133 if (obd->obd_recovering) {
2134 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2135 "recovery for at least %d:%.02d, or until %d "
2136 "client%s reconnect%s.\n",
2137 obd->obd_name, lustre_cfg_string(lcfg, 1),
2138 label ?: "", label ? "/" : "", str,
2139 obd->obd_recovery_timeout / 60,
2140 obd->obd_recovery_timeout % 60,
2141 obd->obd_max_recoverable_clients,
2142 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2143 (obd->obd_max_recoverable_clients == 1) ? "s":"");
2145 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2146 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2147 label ?: "", label ? "/" : "", str,
2148 obd->obd_replayable ? "enabled" : "disabled");
2156 fsfilt_put_ops(obd->obd_fsops);
2157 filter_iobuf_pool_done(filter);
2159 server_put_mount(obd->obd_name, mnt);
2160 obd->u.obt.obt_sb = 0;
2164 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2166 struct lprocfs_static_vars lvars;
2171 CLASSERT(offsetof(struct obd_device, u.obt) ==
2172 offsetof(struct obd_device, u.filter.fo_obt));
2174 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2177 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2178 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2181 addr = (unsigned long)cfs_page_address(page);
2182 clear_page((void *)addr);
2184 /* lprocfs must be setup before the filter so state can be safely added
2185 * to /proc incrementally as the filter is setup */
2186 lprocfs_filter_init_vars(&lvars);
2187 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2188 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2189 /* Init obdfilter private stats here */
2190 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2191 LPROCFS_CNTR_AVGMINMAX,
2192 "read_bytes", "bytes");
2193 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2194 LPROCFS_CNTR_AVGMINMAX,
2195 "write_bytes", "bytes");
2196 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2197 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2198 "get_page", "usec");
2199 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2200 LPROCFS_CNTR_AVGMINMAX,
2201 "get_page_failures", "num");
2202 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2203 LPROCFS_CNTR_AVGMINMAX,
2204 "cache_access", "pages");
2205 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2206 LPROCFS_CNTR_AVGMINMAX,
2207 "cache_hit", "pages");
2208 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2209 LPROCFS_CNTR_AVGMINMAX,
2210 "cache_miss", "pages");
2212 lproc_filter_attach_seqstat(obd);
2213 obd->obd_proc_exports_entry = lprocfs_register("exports",
2214 obd->obd_proc_entry,
2216 if (IS_ERR(obd->obd_proc_exports_entry)) {
2217 rc = PTR_ERR(obd->obd_proc_exports_entry);
2218 CERROR("error %d setting up lprocfs for %s\n",
2220 obd->obd_proc_exports_entry = NULL;
2223 if (obd->obd_proc_exports_entry)
2224 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2225 lprocfs_nid_stats_clear_read,
2226 lprocfs_nid_stats_clear_write, obd, NULL);
2228 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2229 LUSTRE_CFG_BUFLEN(lcfg, 4));
2230 rc = filter_common_setup(obd, lcfg, (void *)addr);
2231 OBD_PAGE_FREE(page);
2234 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2235 lprocfs_free_per_client_stats(obd);
2236 lprocfs_free_obd_stats(obd);
2237 lprocfs_obd_cleanup(obd);
2243 static struct llog_operations filter_mds_ost_repl_logops;
2245 static struct llog_operations filter_size_orig_logops = {
2246 .lop_setup = llog_obd_origin_setup,
2247 .lop_cleanup = llog_obd_origin_cleanup,
2248 .lop_add = llog_obd_origin_add
2251 static int filter_olg_fini(struct obd_llog_group *olg)
2253 struct llog_ctxt *ctxt;
2254 int rc = 0, rc2 = 0;
2257 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2259 rc = llog_cleanup(ctxt);
2261 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2263 rc2 = llog_cleanup(ctxt);
2268 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2270 rc2 = llog_cleanup(ctxt);
2279 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2280 struct obd_device *tgt)
2285 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2286 &filter_mds_ost_repl_logops);
2290 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2291 &filter_size_orig_logops);
2297 filter_olg_fini(olg);
2302 * Init the default olg, which is embeded in the obd_device, for filter.
2305 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2306 struct obd_device *tgt)
2308 struct filter_obd *filter = &obd->u.filter;
2309 struct llog_ctxt *ctxt;
2313 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2314 if (!filter->fo_lcm)
2317 filter_mds_ost_repl_logops = llog_client_ops;
2318 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2319 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2320 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2322 rc = filter_olg_init(obd, olg, tgt);
2324 GOTO(cleanup_lcm, rc);
2326 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2329 GOTO(cleanup_olg, rc);
2331 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2333 CERROR("Can't get ctxt for %p:%x\n", olg,
2334 LLOG_MDS_OST_REPL_CTXT);
2335 GOTO(cleanup_olg, rc = -ENODEV);
2337 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2338 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2339 llog_ctxt_put(ctxt);
2343 filter_olg_fini(olg);
2345 llog_recov_thread_fini(filter->fo_lcm, 1);
2346 filter->fo_lcm = NULL;
2351 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2352 struct obd_device *tgt, int *index)
2354 struct filter_obd *filter = &obd->u.filter;
2355 struct llog_ctxt *ctxt;
2359 LASSERT(olg != NULL);
2360 if (olg == &obd->obd_olg)
2361 return filter_default_olg_init(obd, olg, tgt);
2363 LASSERT(filter->fo_lcm != NULL);
2364 rc = filter_olg_init(obd, olg, tgt);
2367 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2369 CERROR("Can't get ctxt for %p:%x\n", olg,
2370 LLOG_MDS_OST_REPL_CTXT);
2371 filter_olg_fini(olg);
2374 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2375 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2376 llog_ctxt_put(ctxt);
2380 static int filter_llog_finish(struct obd_device *obd, int count)
2382 struct filter_obd *filter = &obd->u.filter;
2383 struct llog_ctxt *ctxt;
2386 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2389 * Make sure that no cached llcds left in recov_thread.
2390 * We actually do sync in disconnect time, but disconnect
2391 * may not come being marked rq_no_resend = 1.
2393 llog_sync(ctxt, NULL);
2396 * Balance class_import_get() in llog_receptor_accept().
2397 * This is safe to do, as llog is already synchronized
2398 * and its import may go.
2400 mutex_down(&ctxt->loc_sem);
2401 if (ctxt->loc_imp) {
2402 class_import_put(ctxt->loc_imp);
2403 ctxt->loc_imp = NULL;
2405 mutex_up(&ctxt->loc_sem);
2406 llog_ctxt_put(ctxt);
2409 if (filter->fo_lcm) {
2410 mutex_down(&ctxt->loc_sem);
2411 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2412 filter->fo_lcm = NULL;
2413 mutex_up(&ctxt->loc_sem);
2415 RETURN(filter_olg_fini(&obd->obd_olg));
2419 * Find the group llog according to group index in the llog group list.
2421 static struct obd_llog_group *
2422 filter_find_olg_internal(struct filter_obd *filter, int group)
2424 struct obd_llog_group *olg;
2426 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2427 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2428 if (olg->olg_group == group)
2435 * Find the group llog according to group index on the filter
2437 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2439 struct obd_llog_group *olg = NULL;
2440 struct filter_obd *filter;
2442 filter = &obd->u.filter;
2444 if (group == FILTER_GROUP_LLOG)
2445 RETURN(&obd->obd_olg);
2447 spin_lock(&filter->fo_llog_list_lock);
2448 olg = filter_find_olg_internal(filter, group);
2449 spin_unlock(&filter->fo_llog_list_lock);
2454 * Find the llog_group of the filter according to the group. If it can not
2455 * find, create the llog_group, which only happens when mds is being synced
2458 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2460 struct obd_llog_group *olg = NULL;
2461 struct filter_obd *filter;
2464 filter = &obd->u.filter;
2466 if (group == FILTER_GROUP_LLOG)
2467 RETURN(&obd->obd_olg);
2469 spin_lock(&filter->fo_llog_list_lock);
2470 olg = filter_find_olg_internal(filter, group);
2472 if (olg->olg_initializing) {
2473 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2475 GOTO(out_unlock, olg);
2480 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2482 llog_group_init(olg, group);
2483 list_add(&olg->olg_list, &filter->fo_llog_list);
2484 olg->olg_initializing = 1;
2485 spin_unlock(&filter->fo_llog_list_lock);
2487 rc = obd_llog_init(obd, olg, obd, NULL);
2489 spin_lock(&filter->fo_llog_list_lock);
2490 list_del(&olg->olg_list);
2491 spin_unlock(&filter->fo_llog_list_lock);
2493 GOTO(out, olg = ERR_PTR(-ENOMEM));
2495 spin_lock(&filter->fo_llog_list_lock);
2496 olg->olg_initializing = 0;
2497 spin_unlock(&filter->fo_llog_list_lock);
2498 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2499 obd->obd_name, group, olg);
2504 spin_unlock(&filter->fo_llog_list_lock);
2508 static int filter_llog_connect(struct obd_export *exp,
2509 struct llogd_conn_body *body)
2511 struct obd_device *obd = exp->exp_obd;
2512 struct llog_ctxt *ctxt;
2513 struct obd_llog_group *olg;
2517 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2518 obd->obd_name, body->lgdc_logid.lgl_oid,
2519 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2521 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2523 CERROR(" %s: can not find olg of group %d\n",
2524 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2527 llog_group_set_export(olg, exp);
2529 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2530 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2531 body->lgdc_ctxt_idx);
2533 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2534 obd->obd_name, body->lgdc_logid.lgl_oid,
2535 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2537 spin_lock_bh(&obd->obd_processing_task_lock);
2538 obd->u.filter.fo_mds_ost_sync = 1;
2539 spin_unlock_bh(&obd->obd_processing_task_lock);
2540 rc = llog_connect(ctxt, &body->lgdc_logid,
2541 &body->lgdc_gen, NULL);
2542 llog_ctxt_put(ctxt);
2544 CERROR("failed to connect rc %d idx %d\n", rc,
2545 body->lgdc_ctxt_idx);
2550 static int filter_llog_preclean(struct obd_device *obd)
2552 struct obd_llog_group *olg, *tmp;
2553 struct filter_obd *filter;
2554 struct list_head remove_list;
2558 rc = obd_llog_finish(obd, 0);
2560 CERROR("failed to cleanup llogging subsystem\n");
2562 filter = &obd->u.filter;
2563 CFS_INIT_LIST_HEAD(&remove_list);
2565 spin_lock(&filter->fo_llog_list_lock);
2566 while (!list_empty(&filter->fo_llog_list)) {
2567 olg = list_entry(filter->fo_llog_list.next,
2568 struct obd_llog_group, olg_list);
2569 list_del(&olg->olg_list);
2570 list_add(&olg->olg_list, &remove_list);
2572 spin_unlock(&filter->fo_llog_list_lock);
2574 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2575 list_del_init(&olg->olg_list);
2576 rc = filter_olg_fini(olg);
2578 CERROR("failed to cleanup llogging subsystem for %u\n",
2586 static int filter_precleanup(struct obd_device *obd,
2587 enum obd_cleanup_stage stage)
2593 case OBD_CLEANUP_EARLY:
2595 case OBD_CLEANUP_EXPORTS:
2596 /* Stop recovery before namespace cleanup. */
2597 target_recovery_fini(obd);
2598 rc = filter_llog_preclean(obd);
2604 static int filter_cleanup(struct obd_device *obd)
2606 struct filter_obd *filter = &obd->u.filter;
2610 LCONSOLE_WARN("%s: shutting down for failover; client state "
2611 "will be preserved.\n", obd->obd_name);
2613 obd_exports_barrier(obd);
2614 obd_zombie_barrier();
2616 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2617 lprocfs_free_per_client_stats(obd);
2618 lprocfs_free_obd_stats(obd);
2619 lprocfs_obd_cleanup(obd);
2620 lquota_cleanup(filter_quota_interface_ref, obd);
2622 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2623 obd->obd_namespace = NULL;
2625 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2627 if (obd->u.obt.obt_sb == NULL)
2632 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2633 shrink_dcache_sb(obd->u.obt.obt_sb);
2635 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2636 obd->u.obt.obt_sb = NULL;
2638 fsfilt_put_ops(obd->obd_fsops);
2640 filter_iobuf_pool_done(filter);
2642 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2647 static int filter_connect_internal(struct obd_export *exp,
2648 struct obd_connect_data *data,
2651 struct filter_export_data *fed = &exp->exp_filter_data;
2656 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2657 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2658 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2659 data->ocd_connect_flags, data->ocd_version,
2660 data->ocd_grant, data->ocd_index);
2662 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2663 CWARN("!!! This export (nid %s) used object group %d "
2664 "earlier; now it's trying to use group %d! This could "
2665 "be a bug in the MDS. Please report to "
2666 "http://bugzilla.lustre.org/\n",
2667 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2670 fed->fed_group = data->ocd_group;
2672 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2673 exp->exp_connect_flags = data->ocd_connect_flags;
2674 data->ocd_version = LUSTRE_VERSION_CODE;
2676 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2677 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2678 data->ocd_connect_flags & OBD_CONNECT_MDS))
2681 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2682 struct filter_obd *filter = &exp->exp_obd->u.filter;
2683 obd_size left, want;
2685 spin_lock(&exp->exp_obd->obd_osfs_lock);
2686 left = filter_grant_space_left(exp);
2687 want = data->ocd_grant;
2688 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2689 data->ocd_grant = fed->fed_grant;
2690 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2692 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2693 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2694 exp->exp_client_uuid.uuid, exp,
2695 data->ocd_grant, want, left);
2697 filter->fo_tot_granted_clients ++;
2700 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2701 struct filter_obd *filter = &exp->exp_obd->u.filter;
2702 struct lr_server_data *lsd = filter->fo_fsd;
2703 int index = le32_to_cpu(lsd->lsd_ost_index);
2705 if (!(lsd->lsd_feature_compat &
2706 cpu_to_le32(OBD_COMPAT_OST))) {
2707 /* this will only happen on the first connect */
2708 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2709 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2710 /* sync is not needed here as filter_client_add will
2711 * set exp_need_sync flag */
2712 filter_update_server_data(exp->exp_obd,
2713 filter->fo_rcvd_filp, lsd);
2714 } else if (index != data->ocd_index) {
2715 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2716 " %u doesn't match actual OST index"
2717 " %u in last_rcvd file, bad "
2719 obd_export_nid2str(exp), index,
2723 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2724 * FIXME: We don't strictly need the COMPAT flag for that,
2725 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2726 * FIXME: We needed it for the index, as index 0 is valid. */
2729 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2730 data->ocd_brw_size = 65536;
2731 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2732 data->ocd_brw_size = min(data->ocd_brw_size,
2733 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2734 LASSERT(data->ocd_brw_size);
2737 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2738 __u32 cksum_types = data->ocd_cksum_types;
2740 /* The client set in ocd_cksum_types the checksum types it
2741 * supports. We have to mask off the algorithms that we don't
2743 if (cksum_types & OBD_CKSUM_ALL)
2744 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2746 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2748 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2749 "%x\n", exp->exp_obd->obd_name,
2750 obd_export_nid2str(exp), cksum_types,
2751 data->ocd_cksum_types);
2753 /* This client does not support OBD_CONNECT_CKSUM
2754 * fall back to CRC32 */
2755 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2756 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2757 exp->exp_obd->obd_name,
2758 obd_export_nid2str(exp));
2764 static int filter_reconnect(const struct lu_env *env,
2765 struct obd_export *exp, struct obd_device *obd,
2766 struct obd_uuid *cluuid,
2767 struct obd_connect_data *data,
2773 if (exp == NULL || obd == NULL || cluuid == NULL)
2776 rc = filter_connect_internal(exp, data, 1);
2778 filter_export_stats_init(obd, exp, localdata);
2783 /* nearly identical to mds_connect */
2784 static int filter_connect(const struct lu_env *env,
2785 struct obd_export **exp, struct obd_device *obd,
2786 struct obd_uuid *cluuid,
2787 struct obd_connect_data *data, void *localdata)
2789 struct lvfs_run_ctxt saved;
2790 struct lustre_handle conn = { 0 };
2791 struct obd_export *lexp;
2792 struct filter_export_data *fed;
2793 struct lsd_client_data *lcd = NULL;
2798 if (exp == NULL || obd == NULL || cluuid == NULL)
2801 rc = class_connect(&conn, obd, cluuid);
2804 lexp = class_conn2export(&conn);
2805 LASSERT(lexp != NULL);
2807 fed = &lexp->exp_filter_data;
2809 rc = filter_connect_internal(lexp, data, 0);
2813 filter_export_stats_init(obd, lexp, localdata);
2814 if (obd->obd_replayable) {
2815 OBD_ALLOC(lcd, sizeof(*lcd));
2817 CERROR("filter: out of memory for client data\n");
2818 GOTO(cleanup, rc = -ENOMEM);
2821 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2823 rc = filter_client_add(obd, lexp, -1);
2828 group = data->ocd_group;
2830 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2831 obd->obd_name, lexp->exp_handle.h_cookie, group);
2833 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2834 rc = filter_read_groups(obd, group, 1);
2835 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2837 CERROR("can't read group %u\n", group);
2847 fed->fed_lcd = NULL;
2849 class_disconnect(lexp);
2858 /* Do extra sanity checks for grant accounting. We do this at connect,
2859 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2860 * always get rid of it or turn it off when we know accounting is good. */
2861 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2863 struct filter_export_data *fed;
2864 struct obd_export *exp;
2865 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2866 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2867 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2869 if (list_empty(&obd->obd_exports))
2872 /* We don't want to do this for large machines that do lots of
2873 mounts or unmounts. It burns... */
2874 if (obd->obd_num_exports > 100)
2877 spin_lock(&obd->obd_osfs_lock);
2878 spin_lock(&obd->obd_dev_lock);
2879 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2881 fed = &exp->exp_filter_data;
2882 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2885 if (maxsize > 0) { /* we may not have done a statfs yet */
2886 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2887 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2888 exp->exp_client_uuid.uuid, exp,
2889 fed->fed_grant, fed->fed_pending, maxsize);
2890 LASSERTF(fed->fed_dirty <= maxsize,
2891 "%s: cli %s/%p %ld > "LPU64"\n", func,
2892 exp->exp_client_uuid.uuid, exp,
2893 fed->fed_dirty, maxsize);
2896 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2897 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2898 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2900 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2901 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2902 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2903 tot_granted += fed->fed_grant + fed->fed_pending;
2904 tot_pending += fed->fed_pending;
2905 tot_dirty += fed->fed_dirty;
2907 fo_tot_granted = obd->u.filter.fo_tot_granted;
2908 fo_tot_pending = obd->u.filter.fo_tot_pending;
2909 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2910 spin_unlock(&obd->obd_dev_lock);
2911 spin_unlock(&obd->obd_osfs_lock);
2913 /* Do these assertions outside the spinlocks so we don't kill system */
2914 if (tot_granted != fo_tot_granted)
2915 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2916 func, tot_granted, fo_tot_granted);
2917 if (tot_pending != fo_tot_pending)
2918 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2919 func, tot_pending, fo_tot_pending);
2920 if (tot_dirty != fo_tot_dirty)
2921 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2922 func, tot_dirty, fo_tot_dirty);
2923 if (tot_pending > tot_granted)
2924 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2925 func, tot_pending, tot_granted);
2926 if (tot_granted > maxsize)
2927 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2928 func, tot_granted, maxsize);
2929 if (tot_dirty > maxsize)
2930 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2931 func, tot_dirty, maxsize);
2934 /* Remove this client from the grant accounting totals. We also remove
2935 * the export from the obd device under the osfs and dev locks to ensure
2936 * that the filter_grant_sanity_check() calculations are always valid.
2937 * The client should do something similar when it invalidates its import. */
2938 static void filter_grant_discard(struct obd_export *exp)
2940 struct obd_device *obd = exp->exp_obd;
2941 struct filter_obd *filter = &obd->u.filter;
2942 struct filter_export_data *fed = &exp->exp_filter_data;
2944 spin_lock(&obd->obd_osfs_lock);
2945 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2946 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2947 obd->obd_name, filter->fo_tot_granted,
2948 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2949 filter->fo_tot_granted -= fed->fed_grant;
2950 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2951 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2952 obd->obd_name, filter->fo_tot_pending,
2953 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2954 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2955 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2956 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2957 obd->obd_name, filter->fo_tot_dirty,
2958 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2959 filter->fo_tot_dirty -= fed->fed_dirty;
2963 spin_unlock(&obd->obd_osfs_lock);
2966 static int filter_destroy_export(struct obd_export *exp)
2970 if (exp->exp_filter_data.fed_pending)
2971 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2972 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2973 exp, exp->exp_filter_data.fed_pending);
2975 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2977 target_destroy_export(exp);
2978 ldlm_destroy_export(exp);
2980 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2984 if (!exp->exp_obd->obd_replayable)
2985 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2987 filter_grant_discard(exp);
2988 filter_fmd_cleanup(exp);
2990 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2991 struct filter_obd *filter = &exp->exp_obd->u.filter;
2992 if (filter->fo_tot_granted_clients > 0)
2993 filter->fo_tot_granted_clients --;
2996 if (!(exp->exp_flags & OBD_OPT_FORCE))
2997 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3002 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3004 struct obd_llog_group *olg_min, *olg;
3005 struct filter_obd *filter;
3006 int worked = -1, group;
3007 struct llog_ctxt *ctxt;
3010 filter = &obd->u.filter;
3012 /* we can't sync log holding spinlock. also, we do not want to get
3013 * into livelock. so we do following: loop over MDS's exports in
3014 * group order and skip already synced llogs -bzzz */
3016 /* look for group with min. number, but > worked */
3019 spin_lock(&filter->fo_llog_list_lock);
3020 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3021 if (olg->olg_group <= worked) {
3022 /* this group is already synced */
3025 if (group < olg->olg_group) {
3026 /* we have group with smaller number to sync */
3029 /* store current minimal group */
3031 group = olg->olg_group;
3033 spin_unlock(&filter->fo_llog_list_lock);
3035 if (olg_min == NULL)
3038 worked = olg_min->olg_group;
3039 if (olg_min->olg_exp &&
3040 (dexp == olg_min->olg_exp || dexp == NULL)) {
3042 ctxt = llog_group_get_ctxt(olg_min,
3043 LLOG_MDS_OST_REPL_CTXT);
3045 err = llog_sync(ctxt, olg_min->olg_exp);
3046 llog_ctxt_put(ctxt);
3048 CERROR("error flushing logs to MDS: "
3053 } while (olg_min != NULL);
3056 /* Also incredibly similar to mds_disconnect */
3057 static int filter_disconnect(struct obd_export *exp)
3059 struct obd_device *obd = exp->exp_obd;
3064 class_export_get(exp);
3066 if (!(exp->exp_flags & OBD_OPT_FORCE))
3067 filter_grant_sanity_check(obd, __FUNCTION__);
3068 filter_grant_discard(exp);
3070 /* Flush any remaining cancel messages out to the target */
3071 filter_sync_llogs(obd, exp);
3073 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3075 rc = server_disconnect_export(exp);
3077 if (exp->exp_obd->obd_replayable)
3078 filter_client_free(exp);
3080 fsfilt_sync(obd, obd->u.obt.obt_sb);
3082 class_export_put(exp);
3086 /* reverse import is changed, sync all cancels */
3087 static void filter_revimp_update(struct obd_export *exp)
3092 class_export_get(exp);
3094 /* flush any remaining cancel messages out to the target */
3095 filter_sync_llogs(exp->exp_obd, exp);
3096 class_export_put(exp);
3100 static int filter_ping(struct obd_export *exp)
3102 filter_fmd_expire(exp);
3106 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3107 const char *what, int quiet)
3109 struct dentry *dchild = NULL;
3112 if (oa->o_valid & OBD_MD_FLGROUP)
3115 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3117 if (IS_ERR(dchild)) {
3118 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3119 what, group, oa->o_id);
3123 if (dchild->d_inode == NULL) {
3125 CERROR("%s: %s on non-existent object: "LPU64"\n",
3126 obd->obd_name, what, oa->o_id);
3128 RETURN(ERR_PTR(-ENOENT));
3134 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3136 struct dentry *dentry = NULL;
3137 struct obd_device *obd;
3141 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3142 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3143 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3147 obd = class_exp2obd(exp);
3149 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3153 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3155 RETURN(PTR_ERR(dentry));
3157 /* Limit the valid bits in the return data to what we actually use */
3158 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3159 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3165 /* this should be enabled/disabled in condition to enabled/disabled large
3166 * inodes (fast EAs) in backing store FS. */
3167 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3168 void *handle, struct obdo *oa)
3170 struct obd_device *obd = exp->exp_obd;
3174 if (oa->o_valid & OBD_MD_FLFID) {
3175 struct filter_fid ff;
3177 if (!(oa->o_valid & OBD_MD_FLGROUP))
3179 /* packing fid and converting it to LE for storing into EA.
3180 * Here ->o_stripe_idx should be filled by LOV and rest of
3181 * fields - by client. */
3182 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3183 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3184 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3185 ff.ff_objid = cpu_to_le64(oa->o_id);
3186 ff.ff_group = cpu_to_le64(oa->o_gr);
3188 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3189 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3190 oa->o_generation, oa->o_id, oa->o_gr);
3192 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3194 CERROR("store fid in object failed! rc: %d\n", rc);
3196 CDEBUG(D_HA, "OSS object without fid info!\n");
3202 /* this is called from filter_truncate() until we have filter_punch() */
3203 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3204 struct obdo *oa, struct obd_trans_info *oti)
3206 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3207 struct llog_cookie *fcc = NULL;
3208 struct filter_obd *filter;
3209 int rc, err, sync = 0;
3210 loff_t old_size = 0;
3211 unsigned int ia_valid;
3212 struct inode *inode;
3213 struct page *page = NULL;
3218 LASSERT(dentry != NULL);
3219 LASSERT(!IS_ERR(dentry));
3221 inode = dentry->d_inode;
3222 LASSERT(inode != NULL);
3224 filter = &exp->exp_obd->u.filter;
3225 iattr_from_obdo(&iattr, oa, oa->o_valid);
3226 ia_valid = iattr.ia_valid;
3228 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3229 OBD_ALLOC(fcc, sizeof(*fcc));
3231 *fcc = oa->o_lcookie;
3233 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3235 /* Filter truncates and writes are serialized by
3236 * i_alloc_sem, see the comment in
3237 * filter_preprw_write.*/
3238 if (ia_valid & ATTR_SIZE)
3239 down_write(&inode->i_alloc_sem);
3240 LOCK_INODE_MUTEX(inode);
3241 old_size = i_size_read(inode);
3244 /* VBR: version recovery check */
3245 rc = filter_version_get_check(exp, oti, inode);
3247 GOTO(out_unlock, rc);
3249 /* Let's pin the last page so that ldiskfs_truncate
3250 * should not start GFP_FS allocation. */
3251 if (ia_valid & ATTR_SIZE) {
3252 page = grab_cache_page(inode->i_mapping,
3253 iattr.ia_size >> PAGE_CACHE_SHIFT);
3255 GOTO(out_unlock, rc = -ENOMEM);
3260 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3261 * then we will accept the UID+GID sent by the client during write for
3262 * initializing the ownership of this inode. We only allow this to
3263 * happen once so clear these bits in setattr. In 2.6 kernels it is
3264 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3265 * the flags that are actually being set. */
3266 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3267 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3268 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3270 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3271 if (!(ia_valid & ATTR_MODE)) {
3272 iattr.ia_mode = inode->i_mode;
3273 iattr.ia_valid |= ATTR_MODE;
3275 iattr.ia_mode &= ~S_ISUID;
3277 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3278 if (!(iattr.ia_valid & ATTR_MODE)) {
3279 iattr.ia_mode = inode->i_mode;
3280 iattr.ia_valid |= ATTR_MODE;
3282 iattr.ia_mode &= ~S_ISGID;
3285 orig_ids[USRQUOTA] = inode->i_uid;
3286 orig_ids[GRPQUOTA] = inode->i_gid;
3287 handle = fsfilt_start_log(exp->exp_obd, inode,
3288 FSFILT_OP_SETATTR, oti, 1);
3290 GOTO(out_unlock, rc = PTR_ERR(handle));
3292 /* update inode EA only once when inode is suid bit marked. As
3293 * on 2.6.x UID and GID may be set separately, we check here
3294 * only one of them to avoid double setting. */
3295 if (inode->i_mode & S_ISUID)
3296 filter_update_fidea(exp, inode, handle, oa);
3298 handle = fsfilt_start(exp->exp_obd, inode,
3299 FSFILT_OP_SETATTR, oti);
3301 GOTO(out_unlock, rc = PTR_ERR(handle));
3303 if (oa->o_valid & OBD_MD_FLFLAGS) {
3304 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3305 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3307 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3309 /* set cancel cookie callback function */
3310 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3311 filter_cancel_cookies_cb,
3315 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3316 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3318 /* The truncate might have used up our transaction credits. Make
3319 * sure we have one left for the last_rcvd update. */
3320 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3322 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3324 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3328 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3330 CERROR("error on commit, err = %d\n", err);
3337 /* For a partial-page truncate flush the page to disk immediately
3338 * to avoid data corruption during direct disk write. b=17397 */
3339 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3340 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3341 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3351 page_cache_release(page);
3353 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3354 UNLOCK_INODE_MUTEX(inode);
3355 if (ia_valid & ATTR_SIZE)
3356 up_write(&inode->i_alloc_sem);
3358 OBD_FREE(fcc, sizeof(*fcc));
3360 /* trigger quota release */
3361 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3362 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3363 int rc2 = lquota_adjust(filter_quota_interface_ref,
3364 exp->exp_obd, cur_ids,
3365 orig_ids, rc, FSFILT_OP_SETATTR);
3366 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3367 "filter adjust qunit. (rc:%d)\n", rc2);
3372 /* this is called from filter_truncate() until we have filter_punch() */
3373 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3374 struct obd_trans_info *oti)
3376 struct obdo *oa = oinfo->oi_oa;
3377 struct lustre_capa *capa = oinfo_capa(oinfo);
3378 struct ldlm_res_id res_id;
3379 struct filter_mod_data *fmd;
3380 struct lvfs_run_ctxt saved;
3381 struct filter_obd *filter;
3382 struct ldlm_resource *res;
3383 struct dentry *dentry;
3384 __u64 opc = CAPA_OPC_META_WRITE;
3388 if (oa->o_valid & OBD_FL_TRUNC)
3389 opc |= CAPA_OPC_OSS_TRUNC;
3391 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3392 rc = filter_auth_capa(exp, NULL, oa->o_gr, capa, opc);
3396 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3397 rc = filter_capa_fixoa(exp, oa, oa->o_gr, capa);
3402 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3403 /* This would be very bad - accidentally truncating a file when
3404 * changing the time or similar - bug 12203. */
3405 if (oa->o_valid & OBD_MD_FLSIZE &&
3406 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3407 static char mdsinum[48];
3409 if (oa->o_valid & OBD_MD_FLFID)
3410 snprintf(mdsinum, sizeof(mdsinum) - 1,
3411 " of inode "LPU64"/%u", oa->o_fid,
3416 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3418 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3423 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3425 RETURN(PTR_ERR(dentry));
3427 filter = &exp->exp_obd->u.filter;
3428 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3431 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3432 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3433 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3434 fmd->fmd_mactime_xid = oti->oti_xid;
3435 filter_fmd_put(exp, fmd);
3438 /* setting objects attributes (including owner/group) */
3439 rc = filter_setattr_internal(exp, dentry, oa, oti);
3441 GOTO(out_unlock, rc);
3443 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3444 &res_id, LDLM_EXTENT, 0);
3447 LDLM_RESOURCE_ADDREF(res);
3448 rc = ldlm_res_lvbo_update(res, NULL, 0);
3449 LDLM_RESOURCE_DELREF(res);
3450 ldlm_resource_putref(res);
3453 oa->o_valid = OBD_MD_FLID;
3455 /* Quota release need uid/gid info */
3456 obdo_from_inode(oa, dentry->d_inode,
3457 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3462 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3466 /* XXX identical to osc_unpackmd */
3467 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3468 struct lov_mds_md *lmm, int lmm_bytes)
3474 if (lmm_bytes < sizeof (*lmm)) {
3475 CERROR("lov_mds_md too small: %d, need %d\n",
3476 lmm_bytes, (int)sizeof(*lmm));
3479 /* XXX LOV_MAGIC etc check? */
3481 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3482 CERROR("lov_mds_md: zero lmm_object_id\n");
3487 lsm_size = lov_stripe_md_size(1);
3491 if (*lsmp != NULL && lmm == NULL) {
3492 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3493 OBD_FREE(*lsmp, lsm_size);
3498 if (*lsmp == NULL) {
3499 OBD_ALLOC(*lsmp, lsm_size);
3503 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3504 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3505 OBD_FREE(*lsmp, lsm_size);
3508 loi_init((*lsmp)->lsm_oinfo[0]);
3512 /* XXX zero *lsmp? */
3513 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3514 LASSERT((*lsmp)->lsm_object_id);
3517 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3522 /* caller must hold fo_create_locks[oa->o_gr] */
3523 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3524 struct filter_obd *filter)
3526 struct obdo doa = { 0 }; /* XXX obdo on stack */
3533 LASSERT_MDS_GROUP(oa->o_gr);
3534 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3535 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3537 memset(&doa, 0, sizeof(doa));
3539 doa.o_valid |= OBD_MD_FLGROUP;
3540 doa.o_gr = oa->o_gr;
3541 doa.o_mode = S_IFREG;
3543 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3544 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3545 exp->exp_obd->obd_name, doa.o_gr);
3549 last = filter_last_id(filter, doa.o_gr);
3551 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3553 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3554 exp->exp_obd->obd_name, oa->o_id + 1, last,
3555 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3557 for (id = last; id > oa->o_id; id--) {
3559 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3560 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3561 CEMERG("error destroying precreate objid "LPU64": %d\n",
3564 /* update last_id on disk periodically so that if we restart
3565 * we don't need to re-scan all of the just-deleted objects. */
3566 if ((id & 511) == 0 && !skip_orphan) {
3567 filter_set_last_id(filter, id - 1, doa.o_gr);
3568 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3572 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3573 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3576 filter_set_last_id(filter, id, doa.o_gr);
3577 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3579 /* don't reuse orphan object, return last used objid */
3583 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3588 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3589 obd_gr group, int *num);
3590 /* returns a negative error or a nonnegative number of files to create */
3591 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3592 obd_gr group, struct obd_trans_info *oti)
3594 struct obd_device *obd = exp->exp_obd;
3595 struct filter_obd *filter = &obd->u.filter;
3599 /* delete orphans request */
3600 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3601 obd_id last = filter_last_id(filter, group);
3603 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3604 CERROR("%s: dropping old orphan cleanup request\n",
3608 /* This causes inflight precreates to abort and drop lock */
3609 set_bit(group, &filter->fo_destroys_in_progress);
3610 down(&filter->fo_create_locks[group]);
3611 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3612 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3613 exp->exp_obd->obd_name, group);
3614 up(&filter->fo_create_locks[group]);
3617 diff = oa->o_id - last;
3618 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3621 if (-diff > OST_MAX_PRECREATE) {
3622 CERROR("%s: ignoring bogus orphan destroy request: "
3623 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3625 /* FIXME: should reset precreate_next_id on MDS */
3626 GOTO(out, rc = -EINVAL);
3629 rc = filter_destroy_precreated(exp, oa, filter);
3631 CERROR("%s: unable to write lastobjid, but "
3632 "orphans were deleted\n", obd->obd_name);
3635 /* XXX: Used by MDS for the first time! */
3636 clear_bit(group, &filter->fo_destroys_in_progress);
3639 down(&filter->fo_create_locks[group]);
3640 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3641 CERROR("%s: dropping old precreate request\n",
3645 /* only precreate if group == 0 and o_id is specfied */
3646 if (!filter_group_is_mds(group) || oa->o_id == 0)
3649 diff = oa->o_id - filter_last_id(filter, group);
3650 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3651 filter_last_id(filter, group), diff);
3653 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3654 oa->o_id, filter_last_id(filter, group), diff);
3658 oa->o_id = filter_last_id(&obd->u.filter, group);
3659 rc = filter_precreate(obd, oa, group, &diff);
3660 oa->o_id = filter_last_id(&obd->u.filter, group);
3662 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3665 /* else diff == 0 */
3668 up(&filter->fo_create_locks[group]);
3672 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3673 __u64 max_age, __u32 flags)
3675 struct filter_obd *filter = &obd->u.filter;
3676 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3680 /* at least try to account for cached pages. its still racey and
3681 * might be under-reporting if clients haven't announced their
3682 * caches with brw recently */
3683 spin_lock(&obd->obd_osfs_lock);
3684 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3685 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3686 spin_unlock(&obd->obd_osfs_lock);
3688 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3689 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3690 filter->fo_tot_dirty, filter->fo_tot_granted,
3691 filter->fo_tot_pending,
3692 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3694 filter_grant_sanity_check(obd, __FUNCTION__);
3696 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3697 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3698 osfs->os_bsize - 1) >> blockbits));
3700 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3701 struct lr_server_data *lsd = filter->fo_fsd;
3702 int index = le32_to_cpu(lsd->lsd_ost_index);
3704 if (obd_fail_val == -1 ||
3705 index == obd_fail_val)
3706 osfs->os_bfree = osfs->os_bavail = 2;
3707 else if (obd_fail_loc & OBD_FAIL_ONCE)
3708 obd_fail_loc &= ~OBD_FAILED; /* reset flag */
3711 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3712 * stop creating files on MDS if OST is not good shape to create
3716 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3717 osfs->os_state = OS_STATE_READONLY;
3719 if (filter->fo_raid_degraded)
3720 osfs->os_state |= OS_STATE_DEGRADED;
3724 static int filter_use_existing_obj(struct obd_device *obd,
3725 struct dentry *dchild, void **handle,
3728 struct inode *inode = dchild->d_inode;
3732 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3735 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3736 if (IS_ERR(*handle))
3737 return PTR_ERR(*handle);
3739 iattr.ia_valid = ATTR_MODE;
3740 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3741 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3748 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3751 __u64 os_ffree = -1;
3753 spin_lock(&obd->obd_osfs_lock);
3754 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3756 os_ffree = obd->obd_osfs.os_ffree;
3757 spin_unlock(&obd->obd_osfs_lock);
3762 /* We rely on the fact that only one thread will be creating files in a given
3763 * group at a time, which is why we don't need an atomic filter_get_new_id.
3764 * Even if we had that atomic function, the following race would exist:
3766 * thread 1: gets id x from filter_next_id
3767 * thread 2: gets id (x + 1) from filter_next_id
3768 * thread 2: creates object (x + 1)
3769 * thread 1: tries to create object x, gets -ENOSPC
3771 * Caller must hold fo_create_locks[group]
3773 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3774 obd_gr group, int *num)
3776 struct dentry *dchild = NULL, *dparent = NULL;
3777 struct filter_obd *filter;
3778 struct obd_statfs *osfs;
3779 int err = 0, rc = 0, recreate_obj = 0, i;
3780 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3783 void *handle = NULL;
3786 filter = &obd->u.filter;
3788 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3790 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3792 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3793 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3796 OBD_ALLOC(osfs, sizeof(*osfs));
3799 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3800 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3801 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3802 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3803 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3807 OBD_FREE(osfs, sizeof(*osfs));
3812 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3813 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3815 for (i = 0; i < *num && err == 0; i++) {
3816 int cleanup_phase = 0;
3818 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3819 CWARN("%s: create aborted by destroy\n",
3828 last_id = filter_last_id(filter, group);
3829 if (next_id > last_id) {
3830 CERROR("Error: Trying to recreate obj greater"
3831 "than last id "LPD64" > "LPD64"\n",
3833 GOTO(cleanup, rc = -EINVAL);
3836 next_id = filter_last_id(filter, group) + 1;
3838 dparent = filter_parent_lock(obd, group, next_id);
3839 if (IS_ERR(dparent))
3840 GOTO(cleanup, rc = PTR_ERR(dparent));
3841 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3843 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3845 GOTO(cleanup, rc = PTR_ERR(dchild));
3846 cleanup_phase = 2; /* f_dput(dchild) */
3848 if (dchild->d_inode != NULL) {
3849 /* This would only happen if lastobjid was bad on disk*/
3850 /* Could also happen if recreating missing obj but it
3851 * already exists. */
3853 CERROR("%s: recreating existing object %.*s?\n",
3854 obd->obd_name, dchild->d_name.len,
3855 dchild->d_name.name);
3857 /* Use these existing objects if they are
3859 if (dchild->d_inode->i_size == 0) {
3860 rc = filter_use_existing_obj(obd,dchild,
3861 &handle, &cleanup_phase);
3868 CERROR("%s: Serious error: objid %.*s already "
3869 "exists; is this filesystem corrupt?\n",
3870 obd->obd_name, dchild->d_name.len,
3871 dchild->d_name.name);
3874 GOTO(cleanup, rc = -EEXIST);
3877 handle = fsfilt_start_log(obd, dparent->d_inode,
3878 FSFILT_OP_CREATE, NULL, 1);
3880 GOTO(cleanup, rc = PTR_ERR(handle));
3883 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3884 ",od->o_id="LPU64")\n", obd->obd_name, group,
3887 /* We mark object SUID+SGID to flag it for accepting UID+GID
3888 * from client on first write. Currently the permission bits
3889 * on the OST are never used, so this is OK. */
3890 rc = ll_vfs_create(dparent->d_inode, dchild,
3891 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3893 CERROR("create failed rc = %d\n", rc);
3894 if (rc == -ENOSPC) {
3895 os_ffree = filter_calc_free_inodes(obd);
3897 CERROR("%s: free inode "LPU64"\n",
3898 obd->obd_name, os_ffree);
3903 if (dchild->d_inode)
3904 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3905 dchild->d_inode->i_ino);
3908 if (!recreate_obj) {
3909 filter_set_last_id(filter, next_id, group);
3910 err = filter_update_last_objid(obd, group, 0);
3912 CERROR("unable to write lastobjid "
3913 "but file created\n");
3917 switch(cleanup_phase) {
3919 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3921 CERROR("error on commit, err = %d\n", err);
3928 filter_parent_unlock(dparent);
3935 if (time_after(jiffies, enough_time)) {
3937 "%s: precreate slow - want %d got %d \n",
3938 obd->obd_name, *num, i);
3945 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3946 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3951 static int filter_create(struct obd_export *exp, struct obdo *oa,
3952 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3954 struct obd_device *obd = exp->exp_obd;
3955 struct filter_export_data *fed;
3956 struct filter_obd *filter;
3957 struct lvfs_run_ctxt saved;
3958 struct lov_stripe_md *lsm = NULL;
3959 int rc = 0, diff, group = oa->o_gr;
3962 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3963 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3965 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3966 CERROR("!!! nid %s sent invalid object group %d\n",
3967 obd_export_nid2str(exp), group);
3971 fed = &exp->exp_filter_data;
3972 filter = &obd->u.filter;
3974 if (fed->fed_group != group) {
3975 CERROR("!!! this export (nid %s) used object group %d "
3976 "earlier; now it's trying to use group %d! This could "
3977 "be a bug in the MDS. Please report to "
3978 "http://bugzilla.lustre.org/\n",
3979 obd_export_nid2str(exp), fed->fed_group, group);
3986 rc = obd_alloc_memmd(exp, &lsm);
3993 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3995 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3996 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3997 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3998 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3999 oa->o_id, filter_last_id(filter,
4004 down(&filter->fo_create_locks[oa->o_gr]);
4005 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
4006 up(&filter->fo_create_locks[oa->o_gr]);
4009 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
4012 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4013 if (rc && ea != NULL && *ea != lsm) {
4014 obd_free_memmd(exp, &lsm);
4015 } else if (rc == 0 && ea != NULL) {
4016 /* XXX LOV STACKING: the lsm that is passed to us from
4017 * LOV does not have valid lsm_oinfo data structs, so
4018 * don't go touching that. This needs to be fixed in a
4020 lsm->lsm_object_id = oa->o_id;
4027 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4028 struct lov_stripe_md *md, struct obd_trans_info *oti,
4029 struct obd_export *md_exp, void *capa)
4031 unsigned int qcids[MAXQUOTAS] = {0, 0};
4032 struct obd_device *obd;
4033 struct filter_obd *filter;
4034 struct dentry *dchild = NULL, *dparent = NULL;
4035 struct lustre_handle lockh = { 0 };
4036 struct lvfs_run_ctxt saved;
4037 void *handle = NULL;
4038 struct llog_cookie *fcc = NULL;
4039 int rc, rc2, cleanup_phase = 0, sync = 0;
4043 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4044 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4045 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4050 filter = &obd->u.filter;
4052 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4055 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4056 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4058 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4060 GOTO(cleanup, rc = PTR_ERR(dchild));
4063 if (dchild->d_inode == NULL) {
4064 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4066 /* If object already gone, cancel cookie right now */
4067 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4068 struct llog_ctxt *ctxt;
4069 struct obd_llog_group *olg;
4071 olg = filter_find_olg(obd, oa->o_gr);
4073 CERROR(" %s: can not find olg of group %d\n",
4074 obd->obd_name, (int)oa->o_gr);
4075 GOTO(cleanup, rc = PTR_ERR(olg));
4077 fcc = &oa->o_lcookie;
4078 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4079 llog_cancel(ctxt, NULL, 1, fcc, 0);
4080 llog_ctxt_put(ctxt);
4081 fcc = NULL; /* we didn't allocate fcc, don't free it */
4083 GOTO(cleanup, rc = -ENOENT);
4086 filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4088 /* Our MDC connection is established by the MDS to us */
4089 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4090 OBD_ALLOC(fcc, sizeof(*fcc));
4092 *fcc = oa->o_lcookie;
4094 DQUOT_INIT(dchild->d_inode);
4096 /* we're gonna truncate it first in order to avoid possible deadlock:
4098 * open trasaction open transaction
4099 * down(i_zombie) down(i_zombie)
4100 * restart transaction
4101 * (see BUG 4180) -bzzz
4103 * take i_alloc_sem too to prevent other threads from writing to the
4104 * file while we are truncating it. This can cause lock ordering issue
4105 * between page lock, i_mutex & starting new journal handle.
4106 * (see bug 20321) -johann
4108 down_write(&dchild->d_inode->i_alloc_sem);
4109 LOCK_INODE_MUTEX(dchild->d_inode);
4111 /* VBR: version recovery check */
4112 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4114 UNLOCK_INODE_MUTEX(dchild->d_inode);
4115 up_write(&dchild->d_inode->i_alloc_sem);
4119 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4121 if (IS_ERR(handle)) {
4122 UNLOCK_INODE_MUTEX(dchild->d_inode);
4123 up_write(&dchild->d_inode->i_alloc_sem);
4124 GOTO(cleanup, rc = PTR_ERR(handle));
4127 iattr.ia_valid = ATTR_SIZE;
4129 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4130 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4131 UNLOCK_INODE_MUTEX(dchild->d_inode);
4132 up_write(&dchild->d_inode->i_alloc_sem);
4136 GOTO(cleanup, rc = rc2);
4138 /* We don't actually need to lock the parent until we are unlinking
4139 * here, and not while truncating above. That avoids holding the
4140 * parent lock for a long time during truncate, which can block other
4141 * threads from doing anything to objects in that directory. bug 7171 */
4142 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4143 if (IS_ERR(dparent))
4144 GOTO(cleanup, rc = PTR_ERR(dparent));
4145 cleanup_phase = 3; /* filter_parent_unlock */
4147 LOCK_INODE_MUTEX(dchild->d_inode);
4148 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4149 if (IS_ERR(handle)) {
4150 UNLOCK_INODE_MUTEX(dchild->d_inode);
4151 GOTO(cleanup, rc = PTR_ERR(handle));
4153 cleanup_phase = 4; /* fsfilt_commit */
4155 /* Quota release need uid/gid of inode */
4156 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4158 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4160 /* this drops dchild->d_inode->i_mutex unconditionally */
4161 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4165 switch(cleanup_phase) {
4168 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4169 oti->oti_handle : handle,
4170 filter_cancel_cookies_cb,
4172 /* If add_journal_cb failed, then filter_finish_transno
4173 * will commit the handle and we will do a sync
4174 * on commit. then we call callback directly to free
4177 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4179 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4182 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4184 CERROR("error on commit, err = %d\n", rc2);
4191 filter_parent_unlock(dparent);
4193 filter_fini_destroy(obd, &lockh);
4197 OBD_FREE(fcc, sizeof(*fcc));
4199 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4202 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4206 /* trigger quota release */
4207 qcids[USRQUOTA] = oa->o_uid;
4208 qcids[GRPQUOTA] = oa->o_gid;
4209 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4212 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4216 /* NB start and end are used for punch, but not truncate */
4217 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4218 struct obd_trans_info *oti,
4219 struct ptlrpc_request_set *rqset)
4224 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4225 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4226 oinfo->oi_policy.l_extent.end);
4230 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4231 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4232 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4234 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4235 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4236 rc = filter_setattr(exp, oinfo, oti);
4237 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4241 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4242 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4245 struct lvfs_run_ctxt saved;
4246 struct filter_obd *filter;
4247 struct dentry *dentry;
4251 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4252 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4253 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4257 filter = &exp->exp_obd->u.filter;
4259 /* An objid of zero is taken to mean "sync whole filesystem" */
4260 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4261 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4262 /* Flush any remaining cancel messages out to the target */
4263 filter_sync_llogs(exp->exp_obd, exp);
4267 dentry = filter_oa2dentry(exp->exp_obd, oa);
4269 RETURN(PTR_ERR(dentry));
4271 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4273 LOCK_INODE_MUTEX(dentry->d_inode);
4275 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4277 /* just any file to grab fsync method - "file" arg unused */
4278 struct file *file = filter->fo_rcvd_filp;
4280 if (file->f_op && file->f_op->fsync)
4281 rc = file->f_op->fsync(NULL, dentry, 1);
4283 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4287 UNLOCK_INODE_MUTEX(dentry->d_inode);
4289 oa->o_valid = OBD_MD_FLID;
4290 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4292 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4298 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4299 void *key, __u32 *vallen, void *val,
4300 struct lov_stripe_md *lsm)
4302 struct obd_device *obd;
4305 obd = class_exp2obd(exp);
4307 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4311 if (KEY_IS(KEY_BLOCKSIZE)) {
4312 __u32 *blocksize = val;
4314 if (*vallen < sizeof(*blocksize))
4316 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4318 *vallen = sizeof(*blocksize);
4322 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4323 __u32 *blocksize_bits = val;
4324 if (blocksize_bits) {
4325 if (*vallen < sizeof(*blocksize_bits))
4327 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4329 *vallen = sizeof(*blocksize_bits);
4333 if (KEY_IS(KEY_LAST_ID)) {
4334 obd_id *last_id = val;
4335 /* FIXME: object groups */
4337 if (*vallen < sizeof(*last_id))
4339 *last_id = filter_last_id(&obd->u.filter,
4340 exp->exp_filter_data.fed_group);
4342 *vallen = sizeof(*last_id);
4346 if (KEY_IS(KEY_FIEMAP)) {
4347 struct ll_fiemap_info_key *fm_key = key;
4348 struct dentry *dentry;
4349 struct ll_user_fiemap *fiemap = val;
4350 struct lvfs_run_ctxt saved;
4353 if (fiemap == NULL) {
4354 *vallen = fiemap_count_to_size(
4355 fm_key->fiemap.fm_extent_count);
4359 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4362 RETURN(PTR_ERR(dentry));
4364 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4365 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4366 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4368 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4374 CDEBUG(D_IOCTL, "invalid key\n");
4378 static inline int filter_setup_llog_group(struct obd_export *exp,
4379 struct obd_device *obd,
4382 struct obd_llog_group *olg;
4383 struct llog_ctxt *ctxt;
4386 olg = filter_find_create_olg(obd, group);
4388 RETURN(PTR_ERR(olg));
4390 llog_group_set_export(olg, exp);
4392 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4393 LASSERTF(ctxt != NULL, "ctxt is null\n");
4395 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4396 llog_ctxt_put(ctxt);
4400 static int filter_set_grant_shrink(struct obd_export *exp,
4401 struct ost_body *body)
4403 /* handle shrink grant */
4404 spin_lock(&exp->exp_obd->obd_osfs_lock);
4405 filter_grant_incoming(exp, &body->oa);
4406 spin_unlock(&exp->exp_obd->obd_osfs_lock);
4412 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4414 struct obd_device *obd;
4420 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4424 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4425 obd_export_nid2str(exp));
4426 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4428 /* setup llog imports */
4430 group = (int)(*(__u32 *)val);
4432 group = 0; /* default value */
4434 LASSERT_MDS_GROUP(group);
4435 rc = filter_setup_llog_group(exp, obd, group);
4439 if (group == FILTER_GROUP_MDS0) {
4440 /* setup llog group 1 for interop */
4441 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4444 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4449 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4450 void *key, __u32 vallen, void *val,
4451 struct ptlrpc_request_set *set)
4453 struct obd_device *obd;
4458 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4462 if (KEY_IS(KEY_CAPA_KEY)) {
4464 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4466 CERROR("filter update capability key failed: %d\n", rc);
4470 if (KEY_IS(KEY_REVIMP_UPD)) {
4471 filter_revimp_update(exp);
4472 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4476 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4477 filter_adapt_sptlrpc_conf(obd, 0);
4481 if (KEY_IS(KEY_MDS_CONN))
4482 RETURN(filter_set_mds_conn(exp, val));
4484 if (KEY_IS(KEY_GRANT_SHRINK))
4485 RETURN(filter_set_grant_shrink(exp, val));
4490 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4491 int len, void *karg, void *uarg)
4493 struct obd_device *obd = exp->exp_obd;
4494 struct obd_ioctl_data *data = karg;
4498 case OBD_IOC_ABORT_RECOVERY: {
4499 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4500 target_stop_recovery_thread(obd);
4504 case OBD_IOC_SYNC: {
4505 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4506 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4510 case OBD_IOC_SET_READONLY: {
4512 struct super_block *sb = obd->u.obt.obt_sb;
4513 struct inode *inode = sb->s_root->d_inode;
4514 BDEVNAME_DECLARE_STORAGE(tmp);
4515 CERROR("*** setting device %s read-only ***\n",
4516 ll_bdevname(sb, tmp));
4518 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4519 if (!IS_ERR(handle))
4520 rc = fsfilt_commit(obd, inode, handle, 1);
4522 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4523 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4525 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4529 case OBD_IOC_CATLOGLIST: {
4530 rc = llog_catalog_list(obd, 1, data);
4534 case OBD_IOC_LLOG_CANCEL:
4535 case OBD_IOC_LLOG_REMOVE:
4536 case OBD_IOC_LLOG_INFO:
4537 case OBD_IOC_LLOG_PRINT: {
4538 /* FIXME to be finished */
4539 RETURN(-EOPNOTSUPP);
4541 struct llog_ctxt *ctxt = NULL;
4543 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4544 rc = llog_ioctl(ctxt, cmd, data);
4545 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4558 static int filter_health_check(struct obd_device *obd)
4560 #ifdef USE_HEALTH_CHECK_WRITE
4561 struct filter_obd *filter = &obd->u.filter;
4566 * health_check to return 0 on healthy
4567 * and 1 on unhealthy.
4569 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4572 #ifdef USE_HEALTH_CHECK_WRITE
4573 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4574 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4579 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4582 return filter_fid2dentry(data, NULL, gr, id);
4585 static int filter_process_config(struct obd_device *obd, obd_count len,
4588 struct lustre_cfg *lcfg = buf;
4589 struct lprocfs_static_vars lvars;
4592 switch (lcfg->lcfg_command) {
4594 lprocfs_filter_init_vars(&lvars);
4596 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4606 static struct lvfs_callback_ops filter_lvfs_ops = {
4607 l_fid2dentry: filter_lvfs_fid2dentry,
4610 static struct obd_ops filter_obd_ops = {
4611 .o_owner = THIS_MODULE,
4612 .o_get_info = filter_get_info,
4613 .o_set_info_async = filter_set_info_async,
4614 .o_setup = filter_setup,
4615 .o_precleanup = filter_precleanup,
4616 .o_cleanup = filter_cleanup,
4617 .o_connect = filter_connect,
4618 .o_reconnect = filter_reconnect,
4619 .o_disconnect = filter_disconnect,
4620 .o_ping = filter_ping,
4621 .o_init_export = filter_init_export,
4622 .o_destroy_export = filter_destroy_export,
4623 .o_statfs = filter_statfs,
4624 .o_getattr = filter_getattr,
4625 .o_unpackmd = filter_unpackmd,
4626 .o_create = filter_create,
4627 .o_setattr = filter_setattr,
4628 .o_destroy = filter_destroy,
4629 .o_brw = filter_brw,
4630 .o_punch = filter_truncate,
4631 .o_sync = filter_sync,
4632 .o_preprw = filter_preprw,
4633 .o_commitrw = filter_commitrw,
4634 .o_llog_init = filter_llog_init,
4635 .o_llog_connect = filter_llog_connect,
4636 .o_llog_finish = filter_llog_finish,
4637 .o_iocontrol = filter_iocontrol,
4638 .o_health_check = filter_health_check,
4639 .o_process_config = filter_process_config,
4642 quota_interface_t *filter_quota_interface_ref;
4643 extern quota_interface_t filter_quota_interface;
4645 static int __init obdfilter_init(void)
4647 struct lprocfs_static_vars lvars;
4650 /** sanity check for group<->mdsno conversion */
4651 for (i = 0; i < 32; i++)
4652 LASSERT(objgrp_to_mdsno(mdt_to_obd_objgrp(i)) == i);
4654 lprocfs_filter_init_vars(&lvars);
4656 request_module("%s", "lquota");
4657 OBD_ALLOC(obdfilter_created_scratchpad,
4658 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4659 sizeof(*obdfilter_created_scratchpad));
4660 if (obdfilter_created_scratchpad == NULL)
4663 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4664 sizeof(struct filter_mod_data),
4667 GOTO(out, rc = -ENOMEM);
4669 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4670 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4672 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4673 LUSTRE_OST_NAME, NULL);
4677 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4678 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4679 ll_fmd_cachep = NULL;
4681 if (filter_quota_interface_ref)
4682 PORTAL_SYMBOL_PUT(filter_quota_interface);
4684 OBD_FREE(obdfilter_created_scratchpad,
4685 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4686 sizeof(*obdfilter_created_scratchpad));
4692 static void __exit obdfilter_exit(void)
4694 if (filter_quota_interface_ref)
4695 PORTAL_SYMBOL_PUT(filter_quota_interface);
4697 if (ll_fmd_cachep) {
4698 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4699 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4700 ll_fmd_cachep = NULL;
4703 class_unregister_type(LUSTRE_OST_NAME);
4704 OBD_FREE(obdfilter_created_scratchpad,
4705 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4706 sizeof(*obdfilter_created_scratchpad));
4709 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4710 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4711 MODULE_LICENSE("GPL");
4713 module_init(obdfilter_init);
4714 module_exit(obdfilter_exit);