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 check_lcd(obd->obd_name, cl_idx, lcd);
869 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
871 /* These exports are cleaned up by filter_disconnect(), so they
872 * need to be set up like real exports as filter_connect() does.
874 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
876 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
877 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
878 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
880 if (PTR_ERR(exp) == -EALREADY) {
881 /* export already exists, zero out this one */
882 CERROR("Zeroing out duplicate export due to "
884 lcd->lcd_uuid[0] = '\0';
886 GOTO(err_client, rc = PTR_ERR(exp));
889 fed = &exp->exp_filter_data;
891 fed->fed_group = 0; /* will be assigned at connect */
892 filter_export_stats_init(obd, exp, NULL);
893 rc = filter_client_add(obd, exp, cl_idx);
894 /* can't fail for existing client */
895 LASSERTF(rc == 0, "rc = %d\n", rc);
897 /* VBR: set export last committed */
898 exp->exp_last_committed = last_rcvd;
899 spin_lock(&exp->exp_lock);
900 exp->exp_connecting = 0;
901 exp->exp_in_recovery = 0;
902 spin_unlock(&exp->exp_lock);
903 spin_lock_bh(&obd->obd_processing_task_lock);
904 obd->obd_max_recoverable_clients++;
905 spin_unlock_bh(&obd->obd_processing_task_lock);
907 class_export_put(exp);
910 /* Need to check last_rcvd even for duplicated exports. */
911 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
914 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
915 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
921 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
923 filter->fo_mount_count = mount_count + 1;
924 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
926 /* save it, so mount count and last_transno is current */
927 rc = filter_update_server_data(obd, filp, filter->fo_fsd);
929 GOTO(err_client, rc);
934 class_disconnect_exports(obd);
936 filter_free_server_data(filter);
940 static int filter_cleanup_groups(struct obd_device *obd)
942 struct filter_obd *filter = &obd->u.filter;
944 struct dentry *dentry;
948 if (filter->fo_dentry_O_groups != NULL) {
949 for (i = 0; i < filter->fo_group_count; i++) {
950 dentry = filter->fo_dentry_O_groups[i];
954 OBD_FREE(filter->fo_dentry_O_groups,
955 filter->fo_group_count *
956 sizeof(*filter->fo_dentry_O_groups));
957 filter->fo_dentry_O_groups = NULL;
959 if (filter->fo_last_objid_files != NULL) {
960 for (i = 0; i < filter->fo_group_count; i++) {
961 filp = filter->fo_last_objid_files[i];
965 OBD_FREE(filter->fo_last_objid_files,
966 filter->fo_group_count *
967 sizeof(*filter->fo_last_objid_files));
968 filter->fo_last_objid_files = NULL;
970 if (filter->fo_dentry_O_sub != NULL) {
971 for (i = 0; i < filter->fo_group_count; i++) {
972 for (j = 0; j < filter->fo_subdir_count; j++) {
973 dentry = filter->fo_dentry_O_sub[i].dentry[j];
978 OBD_FREE(filter->fo_dentry_O_sub,
979 filter->fo_group_count *
980 sizeof(*filter->fo_dentry_O_sub));
981 filter->fo_dentry_O_sub = NULL;
983 if (filter->fo_last_objids != NULL) {
984 OBD_FREE(filter->fo_last_objids,
985 filter->fo_group_count *
986 sizeof(*filter->fo_last_objids));
987 filter->fo_last_objids = NULL;
989 if (filter->fo_dentry_O != NULL) {
990 f_dput(filter->fo_dentry_O);
991 filter->fo_dentry_O = NULL;
996 static int filter_update_last_group(struct obd_device *obd, int group)
998 struct filter_obd *filter = &obd->u.filter;
999 struct file *filp = NULL;
1000 int last_group = 0, rc;
1004 if (group <= filter->fo_committed_group)
1007 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1011 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1015 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1017 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1021 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1022 obd->obd_name, last_group, group);
1026 /* must be sync: bXXXX */
1027 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1029 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1033 filter->fo_committed_group = group;
1036 filp_close(filp, 0);
1040 static int filter_read_group_internal(struct obd_device *obd, int group,
1043 struct filter_obd *filter = &obd->u.filter;
1044 __u64 *new_objids = NULL;
1045 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1046 struct dentry **new_groups = NULL;
1047 struct file **new_files = NULL;
1048 struct dentry *dentry;
1050 int old_count = filter->fo_group_count, rc, stage = 0, i;
1054 int len = group + 1;
1056 snprintf(name, 24, "%d", group);
1060 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1062 if (IS_ERR(dentry)) {
1063 CERROR("Cannot lookup expected object group %d: %ld\n",
1064 group, PTR_ERR(dentry));
1065 RETURN(PTR_ERR(dentry));
1068 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1070 if (IS_ERR(dentry)) {
1071 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1073 RETURN(PTR_ERR(dentry));
1078 snprintf(name, 24, "O/%d/LAST_ID", group);
1080 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1082 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1083 GOTO(cleanup, rc = PTR_ERR(filp));
1087 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1089 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1093 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1094 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1095 if (tmp_subdirs == NULL)
1096 GOTO(cleanup, rc = -ENOMEM);
1099 for (i = 0; i < filter->fo_subdir_count; i++) {
1101 snprintf(dir, sizeof(dir), "d%u", i);
1103 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1106 if (IS_ERR(tmp_subdirs->dentry[i])) {
1107 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1108 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1113 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1114 tmp_subdirs->dentry[i]);
1118 /* 'group' is an index; we need an array of length 'group + 1' */
1119 if (group + 1 > old_count) {
1120 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1121 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1122 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1123 OBD_ALLOC(new_files, len * sizeof(*new_files));
1125 if (new_objids == NULL || new_subdirs == NULL ||
1126 new_groups == NULL || new_files == NULL)
1127 GOTO(cleanup, rc = -ENOMEM);
1130 memcpy(new_objids, filter->fo_last_objids,
1131 old_count * sizeof(*new_objids));
1132 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1133 old_count * sizeof(*new_subdirs));
1134 memcpy(new_groups, filter->fo_dentry_O_groups,
1135 old_count * sizeof(*new_groups));
1136 memcpy(new_files, filter->fo_last_objid_files,
1137 old_count * sizeof(*new_files));
1139 OBD_FREE(filter->fo_last_objids,
1140 old_count * sizeof(*new_objids));
1141 OBD_FREE(filter->fo_dentry_O_sub,
1142 old_count * sizeof(*new_subdirs));
1143 OBD_FREE(filter->fo_dentry_O_groups,
1144 old_count * sizeof(*new_groups));
1145 OBD_FREE(filter->fo_last_objid_files,
1146 old_count * sizeof(*new_files));
1148 filter->fo_last_objids = new_objids;
1149 filter->fo_dentry_O_sub = new_subdirs;
1150 filter->fo_dentry_O_groups = new_groups;
1151 filter->fo_last_objid_files = new_files;
1152 filter->fo_group_count = len;
1155 filter->fo_dentry_O_groups[group] = dentry;
1156 filter->fo_last_objid_files[group] = filp;
1157 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1158 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1159 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1162 filter_update_last_group(obd, group);
1164 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1165 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1166 rc = filter_update_last_objid(obd, group, 1);
1170 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1171 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1172 obd->obd_name, group, last_objid);
1177 if (new_objids != NULL)
1178 OBD_FREE(new_objids, len * sizeof(*new_objids));
1179 if (new_subdirs != NULL)
1180 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1181 if (new_groups != NULL)
1182 OBD_FREE(new_groups, len * sizeof(*new_groups));
1183 if (new_files != NULL)
1184 OBD_FREE(new_files, len * sizeof(*new_files));
1186 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1187 for (i = 0; i < filter->fo_subdir_count; i++) {
1188 if (tmp_subdirs->dentry[i] != NULL)
1189 dput(tmp_subdirs->dentry[i]);
1191 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1194 filp_close(filp, 0);
1201 static int filter_read_groups(struct obd_device *obd, int last_group,
1204 struct filter_obd *filter = &obd->u.filter;
1205 int old_count, group, rc = 0;
1207 down(&filter->fo_init_lock);
1208 old_count = filter->fo_group_count;
1209 for (group = old_count; group <= last_group; group++) {
1210 rc = filter_read_group_internal(obd, group, create);
1214 up(&filter->fo_init_lock);
1218 /* FIXME: object groups */
1219 static int filter_prep_groups(struct obd_device *obd)
1221 struct filter_obd *filter = &obd->u.filter;
1222 struct dentry *O_dentry;
1224 int last_group, rc = 0, cleanup_phase = 0;
1228 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1230 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1231 if (IS_ERR(O_dentry)) {
1232 rc = PTR_ERR(O_dentry);
1233 CERROR("cannot open/create O: rc = %d\n", rc);
1236 filter->fo_dentry_O = O_dentry;
1237 cleanup_phase = 1; /* O_dentry */
1239 /* we have to initialize all groups before first connections from
1240 * clients because they may send create/destroy for any group -bzzz */
1241 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1243 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1244 GOTO(cleanup, rc = PTR_ERR(filp));
1246 cleanup_phase = 2; /* filp */
1248 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1250 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1255 last_group = FILTER_GROUP_MDS0;
1257 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1258 FILTER_GROUP_MDS0, last_group);
1259 filter->fo_committed_group = last_group;
1260 rc = filter_read_groups(obd, last_group, 1);
1264 filp_close(filp, 0);
1268 switch (cleanup_phase) {
1270 filp_close(filp, 0);
1272 filter_cleanup_groups(obd);
1273 f_dput(filter->fo_dentry_O);
1274 filter->fo_dentry_O = NULL;
1282 /* setup the object store with correct subdirectories */
1283 static int filter_prep(struct obd_device *obd)
1285 struct lvfs_run_ctxt saved;
1286 struct filter_obd *filter = &obd->u.filter;
1288 struct inode *inode;
1292 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1293 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1294 if (!file || IS_ERR(file)) {
1296 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1300 filter->fo_rcvd_filp = file;
1301 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1302 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1303 file->f_dentry->d_inode->i_mode);
1304 GOTO(err_filp, rc = -ENOENT);
1307 inode = file->f_dentry->d_parent->d_inode;
1308 /* We use i_op->unlink directly in filter_vfs_unlink() */
1309 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1310 CERROR("%s: filesystem does not support create/unlink ops\n",
1312 GOTO(err_filp, rc = -EOPNOTSUPP);
1315 /** lu_target has very limited use in filter now */
1316 lut_init(NULL, &filter->fo_lut, obd, NULL);
1318 rc = filter_init_server_data(obd, file);
1320 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1324 target_recovery_init(&filter->fo_lut, ost_handle);
1326 /* open and create health check io file*/
1327 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1330 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1332 GOTO(err_server_data, rc);
1334 filter->fo_obt.obt_health_check_filp = file;
1335 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1336 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1337 file->f_dentry->d_inode->i_mode);
1338 GOTO(err_health_check, rc = -ENOENT);
1340 rc = lvfs_check_io_health(obd, file);
1342 GOTO(err_health_check, rc);
1344 rc = filter_prep_groups(obd);
1346 GOTO(err_health_check, rc);
1348 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1353 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1354 CERROR("can't close %s after error\n", HEALTH_CHECK);
1355 filter->fo_obt.obt_health_check_filp = NULL;
1357 target_recovery_fini(obd);
1358 filter_free_server_data(filter);
1360 if (filp_close(filter->fo_rcvd_filp, 0))
1361 CERROR("can't close %s after error\n", LAST_RCVD);
1362 filter->fo_rcvd_filp = NULL;
1366 /* cleanup the filter: write last used object id to status file */
1367 static void filter_post(struct obd_device *obd)
1369 struct lvfs_run_ctxt saved;
1370 struct filter_obd *filter = &obd->u.filter;
1373 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1374 * best to start a transaction with h_sync, because we removed this
1377 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1378 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1381 CERROR("error writing server data: rc = %d\n", rc);
1383 for (i = 0; i < filter->fo_group_count; i++) {
1384 rc = filter_update_last_objid(obd, i,
1385 (i == filter->fo_group_count - 1));
1387 CERROR("error writing group %d lastobjid: rc = %d\n",
1391 rc = filp_close(filter->fo_rcvd_filp, 0);
1392 filter->fo_rcvd_filp = NULL;
1394 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1396 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1397 filter->fo_obt.obt_health_check_filp = NULL;
1399 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1401 filter_cleanup_groups(obd);
1402 filter_free_server_data(filter);
1403 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1405 filter_free_capa_keys(filter);
1406 cleanup_capa_hash(filter->fo_capa_hash);
1409 static void filter_set_last_id(struct filter_obd *filter,
1410 obd_id id, obd_gr group)
1412 LASSERT(filter->fo_fsd != NULL);
1413 LASSERT(group <= filter->fo_group_count);
1415 spin_lock(&filter->fo_objidlock);
1416 filter->fo_last_objids[group] = id;
1417 spin_unlock(&filter->fo_objidlock);
1420 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1423 LASSERT(filter->fo_fsd != NULL);
1424 LASSERT(group <= filter->fo_group_count);
1426 /* FIXME: object groups */
1427 spin_lock(&filter->fo_objidlock);
1428 id = filter->fo_last_objids[group];
1429 spin_unlock(&filter->fo_objidlock);
1433 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1435 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1439 /* We never dget the object parent, so DON'T dput it either */
1440 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1442 struct filter_obd *filter = &obd->u.filter;
1443 struct filter_subdirs *subdirs;
1444 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1446 if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1447 return filter->fo_dentry_O_groups[group];
1449 subdirs = &filter->fo_dentry_O_sub[group];
1450 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1453 /* We never dget the object parent, so DON'T dput it either */
1454 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1457 unsigned long now = jiffies;
1458 struct dentry *dparent = filter_parent(obd, group, objid);
1461 if (IS_ERR(dparent))
1463 if (dparent == NULL)
1464 return ERR_PTR(-ENOENT);
1466 rc = filter_lock_dentry(obd, dparent);
1467 fsfilt_check_slow(obd, now, "parent lock");
1468 return rc ? ERR_PTR(rc) : dparent;
1471 /* We never dget the object parent, so DON'T dput it either */
1472 static void filter_parent_unlock(struct dentry *dparent)
1474 UNLOCK_INODE_MUTEX(dparent->d_inode);
1477 /* How to get files, dentries, inodes from object id's.
1479 * If dir_dentry is passed, the caller has already locked the parent
1480 * appropriately for this operation (normally a write lock). If
1481 * dir_dentry is NULL, we do a read lock while we do the lookup to
1482 * avoid races with create/destroy and such changing the directory
1483 * internal to the filesystem code. */
1484 struct dentry *filter_fid2dentry(struct obd_device *obd,
1485 struct dentry *dir_dentry,
1486 obd_gr group, obd_id id)
1488 struct dentry *dparent = dir_dentry;
1489 struct dentry *dchild;
1494 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1495 obd->u.filter.fo_destroys_in_progress == 0) {
1496 /* don't fail lookups for orphan recovery, it causes
1497 * later LBUGs when objects still exist during precreate */
1498 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1499 RETURN(ERR_PTR(-ENOENT));
1502 CERROR("fatal: invalid object id 0\n");
1503 RETURN(ERR_PTR(-ESTALE));
1506 len = sprintf(name, LPU64, id);
1507 if (dir_dentry == NULL) {
1508 dparent = filter_parent_lock(obd, group, id);
1509 if (IS_ERR(dparent)) {
1510 CERROR("%s: error getting object "LPU64":"LPU64
1511 " parent: rc %ld\n", obd->obd_name,
1512 id, group, PTR_ERR(dparent));
1516 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1517 dparent->d_name.len, dparent->d_name.name, name);
1518 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1519 if (dir_dentry == NULL)
1520 filter_parent_unlock(dparent);
1521 if (IS_ERR(dchild)) {
1522 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1527 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1528 CERROR("%s: got bad object "LPU64" inode %lu\n",
1529 obd->obd_name, id, dchild->d_inode->i_ino);
1531 RETURN(ERR_PTR(-ENOENT));
1534 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1535 name, dchild, atomic_read(&dchild->d_count));
1537 LASSERT(atomic_read(&dchild->d_count) > 0);
1542 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1543 obd_id group, struct lustre_handle *lockh)
1545 int flags = LDLM_AST_DISCARD_DATA, rc;
1546 struct ldlm_res_id res_id;
1547 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1550 osc_build_res_name(objid, group, &res_id);
1551 /* Tell the clients that the object is gone now and that they should
1552 * throw away any cached pages. */
1553 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1554 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1555 ldlm_completion_ast, NULL, NULL, 0, NULL,
1562 static void filter_fini_destroy(struct obd_device *obd,
1563 struct lustre_handle *lockh)
1566 ldlm_lock_decref(lockh, LCK_PW);
1569 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1570 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1571 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1572 * reverse. Caller must take i_sem before starting the transaction and we
1573 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1574 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1575 struct vfsmount *mnt)
1580 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1581 * itself we already hold dir->i_mutex for child create/unlink ops */
1582 LASSERT(dentry->d_inode != NULL);
1583 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1584 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1588 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1589 GOTO(out, rc = -ENOENT);
1591 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1596 GOTO(out, rc = -EPERM);
1598 /* check_sticky() */
1599 if ((dentry->d_inode->i_uid != current->fsuid &&
1600 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1601 IS_IMMUTABLE(dentry->d_inode))
1602 GOTO(out, rc = -EPERM);
1604 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1605 * that was done because of journal_start (which is already done
1606 * here) or some other ordering issue. */
1609 rc = ll_security_inode_unlink(dir, dentry, mnt);
1613 rc = dir->i_op->unlink(dir, dentry);
1615 /* need to drop i_mutex before we lose inode reference */
1616 UNLOCK_INODE_MUTEX(dentry->d_inode);
1623 /* Caller must hold LCK_PW on parent and push us into kernel context.
1624 * Caller must hold child i_mutex, we drop it always.
1625 * Caller is also required to ensure that dchild->d_inode exists. */
1626 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1627 obd_gr group, struct dentry *dparent,
1628 struct dentry *dchild)
1630 struct inode *inode = dchild->d_inode;
1631 struct filter_obd *filter = &obd->u.filter;
1634 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1635 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1636 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1637 (unsigned long)inode->i_nlink,
1638 atomic_read(&inode->i_count));
1641 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1643 CERROR("error unlinking objid %.*s: rc %d\n",
1644 dchild->d_name.len, dchild->d_name.name, rc);
1648 struct filter_intent_args {
1649 struct ldlm_lock **victim;
1654 static enum interval_iter filter_intent_cb(struct interval_node *n,
1657 struct ldlm_interval *node = (struct ldlm_interval *)n;
1658 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1659 __u64 size = arg->size;
1660 struct ldlm_lock **v = arg->victim;
1661 struct ldlm_lock *lck;
1663 /* If the interval is lower than the current file size,
1665 if (interval_high(n) <= size)
1666 return INTERVAL_ITER_STOP;
1668 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1669 /* Don't send glimpse ASTs to liblustre clients.
1670 * They aren't listening for them, and they do
1671 * entirely synchronous I/O anyways. */
1672 if (lck->l_export == NULL ||
1673 lck->l_export->exp_libclient == 1)
1676 if (*arg->liblustre)
1677 *arg->liblustre = 0;
1680 *v = LDLM_LOCK_GET(lck);
1681 } else if ((*v)->l_policy_data.l_extent.start <
1682 lck->l_policy_data.l_extent.start) {
1683 LDLM_LOCK_RELEASE(*v);
1684 *v = LDLM_LOCK_GET(lck);
1687 /* the same policy group - every lock has the
1688 * same extent, so needn't do it any more */
1692 return INTERVAL_ITER_CONT;
1695 static int filter_intent_policy(struct ldlm_namespace *ns,
1696 struct ldlm_lock **lockp, void *req_cookie,
1697 ldlm_mode_t mode, int flags, void *data)
1699 CFS_LIST_HEAD(rpc_list);
1700 struct ptlrpc_request *req = req_cookie;
1701 struct ldlm_lock *lock = *lockp, *l = NULL;
1702 struct ldlm_resource *res = lock->l_resource;
1703 ldlm_processing_policy policy;
1704 struct ost_lvb *res_lvb, *reply_lvb;
1705 struct ldlm_reply *rep;
1707 int idx, rc, tmpflags = 0, only_liblustre = 1;
1708 struct ldlm_interval_tree *tree;
1709 struct filter_intent_args arg;
1710 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1711 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1712 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1715 policy = ldlm_get_processing_policy(res);
1716 LASSERT(policy != NULL);
1717 LASSERT(req != NULL);
1719 rc = lustre_pack_reply(req, 3, repsize, NULL);
1721 RETURN(req->rq_status = rc);
1723 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1724 LASSERT(rep != NULL);
1726 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1727 sizeof(*reply_lvb));
1728 LASSERT(reply_lvb != NULL);
1730 //fixup_handle_for_resent_req(req, lock, &lockh);
1732 /* Call the extent policy function to see if our request can be
1733 * granted, or is blocked.
1734 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1735 * lock, and should not be granted if the lock will be blocked.
1738 LASSERT(ns == res->lr_namespace);
1740 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1741 check_res_locked(res);
1743 /* FIXME: we should change the policy function slightly, to not make
1744 * this list at all, since we just turn around and free it */
1745 while (!list_empty(&rpc_list)) {
1746 struct ldlm_lock *wlock =
1747 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1748 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1749 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1750 lock->l_flags &= ~LDLM_FL_CP_REQD;
1751 list_del_init(&wlock->l_cp_ast);
1752 LDLM_LOCK_RELEASE(wlock);
1755 /* The lock met with no resistance; we're finished. */
1756 if (rc == LDLM_ITER_CONTINUE) {
1757 /* do not grant locks to the liblustre clients: they cannot
1758 * handle ASTs robustly. We need to do this while still
1759 * holding ns_lock to avoid the lock remaining on the res_link
1760 * list (and potentially being added to l_pending_list by an
1761 * AST) when we are going to drop this lock ASAP. */
1762 if (lock->l_export->exp_libclient ||
1763 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1764 ldlm_resource_unlink_lock(lock);
1765 err = ELDLM_LOCK_ABORTED;
1767 err = ELDLM_LOCK_REPLACED;
1773 /* Do not grant any lock, but instead send GL callbacks. The extent
1774 * policy nicely created a list of all PW locks for us. We will choose
1775 * the highest of those which are larger than the size in the LVB, if
1776 * any, and perform a glimpse callback. */
1777 res_lvb = res->lr_lvb_data;
1778 LASSERT(res_lvb != NULL);
1779 *reply_lvb = *res_lvb;
1782 * ->ns_lock guarantees that no new locks are granted, and,
1783 * therefore, that res->lr_lvb_data cannot increase beyond the
1784 * end of already granted lock. As a result, it is safe to
1785 * check against "stale" reply_lvb->lvb_size value without
1788 arg.size = reply_lvb->lvb_size;
1790 arg.liblustre = &only_liblustre;
1791 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1792 tree = &res->lr_itree[idx];
1793 if (tree->lit_mode == LCK_PR)
1796 interval_iterate_reverse(tree->lit_root,
1797 filter_intent_cb, &arg);
1801 /* There were no PW locks beyond the size in the LVB; finished. */
1803 if (only_liblustre) {
1804 /* If we discovered a liblustre client with a PW lock,
1805 * however, the LVB may be out of date! The LVB is
1806 * updated only on glimpse (which we don't do for
1807 * liblustre clients) and cancel (which the client
1808 * obviously has not yet done). So if it has written
1809 * data but kept the lock, the LVB is stale and needs
1810 * to be updated from disk.
1812 * Of course, this will all disappear when we switch to
1813 * taking liblustre locks on the OST. */
1814 ldlm_res_lvbo_update(res, NULL, 1);
1816 RETURN(ELDLM_LOCK_ABORTED);
1820 * This check is for lock taken in filter_prepare_destroy() that does
1821 * not have l_glimpse_ast set. So the logic is: if there is a lock
1822 * with no l_glimpse_ast set, this object is being destroyed already.
1824 * Hence, if you are grabbing DLM locks on the server, always set
1825 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1827 if (l->l_glimpse_ast == NULL) {
1828 /* We are racing with unlink(); just return -ENOENT */
1829 rep->lock_policy_res1 = -ENOENT;
1833 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1834 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1835 /* Update the LVB from disk if the AST failed (this is a legal race) */
1837 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1838 * sending ast is not handled. This can result in lost client writes.
1841 ldlm_res_lvbo_update(res, NULL, 1);
1844 *reply_lvb = *res_lvb;
1848 LDLM_LOCK_RELEASE(l);
1850 RETURN(ELDLM_LOCK_ABORTED);
1854 * per-obd_device iobuf pool.
1856 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1857 * allocations in write-path has to be minimized (see bug 5137).
1859 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1860 * OST threads (see ost_thread_{init,done}()).
1862 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1863 * at the OST layer there are only (potentially) multiple obd_device of type
1864 * unknown at the time of OST thread creation.
1866 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1867 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1868 * it's very own iobuf.
1872 * filter_kiobuf_pool_init()
1874 * filter_kiobuf_pool_done()
1876 * filter_iobuf_get()
1878 * operate on this array. They are "generic" in a sense that they don't depend
1879 * on actual type of iobuf's (the latter depending on Linux kernel version).
1883 * destroy pool created by filter_iobuf_pool_init
1885 static void filter_iobuf_pool_done(struct filter_obd *filter)
1887 struct filter_iobuf **pool;
1892 pool = filter->fo_iobuf_pool;
1894 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1895 if (pool[i] != NULL)
1896 filter_free_iobuf(pool[i]);
1898 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1899 filter->fo_iobuf_pool = NULL;
1904 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1906 struct filter_obd *filter = &obd->u.filter;
1907 struct sptlrpc_rule_set tmp_rset;
1910 sptlrpc_rule_set_init(&tmp_rset);
1911 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1913 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1918 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1920 write_lock(&filter->fo_sptlrpc_lock);
1921 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1922 filter->fo_sptlrpc_rset = tmp_rset;
1923 write_unlock(&filter->fo_sptlrpc_lock);
1929 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1931 static int filter_iobuf_pool_init(struct filter_obd *filter)
1938 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1940 if (filter->fo_iobuf_pool == NULL)
1943 filter->fo_iobuf_count = OSS_THREADS_MAX;
1948 /* Return iobuf allocated for @thread_id. We don't know in advance how
1949 * many threads there will be so we allocate a large empty array and only
1950 * fill in those slots that are actually in use.
1951 * If we haven't allocated a pool entry for this thread before, do so now. */
1952 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1954 int thread_id = (oti && oti->oti_thread) ?
1955 oti->oti_thread->t_id : -1;
1956 struct filter_iobuf *pool = NULL;
1957 struct filter_iobuf **pool_place = NULL;
1959 if (thread_id >= 0) {
1960 LASSERT(thread_id < filter->fo_iobuf_count);
1961 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1964 if (unlikely(pool == NULL)) {
1965 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1966 PTLRPC_MAX_BRW_PAGES);
1967 if (pool_place != NULL)
1974 /* mount the file system (secretly). lustre_cfg parameters are:
1977 * 3 = flags: failover=f, failout=n
1980 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1983 struct filter_obd *filter = &obd->u.filter;
1984 struct vfsmount *mnt;
1985 struct lustre_mount_info *lmi;
1986 struct obd_uuid uuid;
1994 if (lcfg->lcfg_bufcount < 3 ||
1995 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1996 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1999 lmi = server_get_mount(obd->obd_name);
2001 /* We already mounted in lustre_fill_super.
2002 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2003 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2005 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2007 /* old path - used by lctl */
2008 CERROR("Using old MDS mount method\n");
2009 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2010 MS_NOATIME|MS_NODIRATIME,
2011 lustre_cfg_string(lcfg, 1), option);
2014 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2015 lustre_cfg_string(lcfg, 1), rc);
2019 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2021 if (IS_ERR(obd->obd_fsops))
2022 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2024 rc = filter_iobuf_pool_init(filter);
2028 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2029 CERROR("%s: Underlying device is marked as read-only. "
2030 "Setup failed\n", obd->obd_name);
2031 GOTO(err_ops, rc = -EROFS);
2034 /* failover is the default */
2035 obd->obd_replayable = 1;
2037 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2038 str = lustre_cfg_string(lcfg, 3);
2039 if (strchr(str, 'n')) {
2040 CWARN("%s: recovery disabled\n", obd->obd_name);
2041 obd->obd_replayable = 0;
2045 filter->fo_vfsmnt = mnt;
2046 obd->u.obt.obt_sb = mnt->mnt_sb;
2047 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2048 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2050 fsfilt_setup(obd, obd->u.obt.obt_sb);
2052 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2053 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2054 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2055 obd->obd_lvfs_ctxt.fs = get_ds();
2056 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2058 init_mutex(&filter->fo_init_lock);
2059 filter->fo_committed_group = 0;
2060 filter->fo_destroys_in_progress = 0;
2061 for (i = 0; i < 32; i++)
2062 sema_init(&filter->fo_create_locks[i], 1);
2064 spin_lock_init(&filter->fo_translock);
2065 spin_lock_init(&filter->fo_objidlock);
2066 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2067 sema_init(&filter->fo_alloc_lock, 1);
2068 init_brw_stats(&filter->fo_filter_stats);
2069 filter->fo_read_cache = 1; /* enable read-only cache by default */
2070 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2071 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2072 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2073 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2075 rc = filter_prep(obd);
2079 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2080 spin_lock_init(&filter->fo_llog_list_lock);
2082 filter->fo_fl_oss_capa = 1;
2084 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2085 filter->fo_capa_hash = init_capa_hash();
2086 if (filter->fo_capa_hash == NULL)
2087 GOTO(err_post, rc = -ENOMEM);
2089 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2090 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2091 LDLM_NAMESPACE_GREEDY);
2092 if (obd->obd_namespace == NULL)
2093 GOTO(err_post, rc = -ENOMEM);
2094 obd->obd_namespace->ns_lvbp = obd;
2095 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2096 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2098 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2099 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2101 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2103 CERROR("failed to setup llogging subsystems\n");
2107 rwlock_init(&filter->fo_sptlrpc_lock);
2108 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2109 /* do this after llog being initialized */
2110 filter_adapt_sptlrpc_conf(obd, 1);
2112 rc = lquota_setup(filter_quota_interface_ref, obd);
2116 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2117 if (q->max_sectors < q->max_hw_sectors &&
2118 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2119 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2120 "for larger I/O requests: max_sectors = %u "
2121 "could be up to max_hw_sectors=%u\n",
2122 obd->obd_name, mnt->mnt_sb->s_id,
2123 q->max_sectors, q->max_hw_sectors);
2125 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2126 if (uuid_ptr != NULL) {
2127 class_uuid_unparse(uuid_ptr, &uuid);
2133 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2135 if (obd->obd_recovering) {
2136 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2137 "recovery for at least %d:%.02d, or until %d "
2138 "client%s reconnect%s.\n",
2139 obd->obd_name, lustre_cfg_string(lcfg, 1),
2140 label ?: "", label ? "/" : "", str,
2141 obd->obd_recovery_timeout / 60,
2142 obd->obd_recovery_timeout % 60,
2143 obd->obd_max_recoverable_clients,
2144 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2145 (obd->obd_max_recoverable_clients == 1) ? "s":"");
2147 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2148 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2149 label ?: "", label ? "/" : "", str,
2150 obd->obd_replayable ? "enabled" : "disabled");
2158 fsfilt_put_ops(obd->obd_fsops);
2159 filter_iobuf_pool_done(filter);
2161 server_put_mount(obd->obd_name, mnt);
2162 obd->u.obt.obt_sb = 0;
2166 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2168 struct lprocfs_static_vars lvars;
2173 CLASSERT(offsetof(struct obd_device, u.obt) ==
2174 offsetof(struct obd_device, u.filter.fo_obt));
2176 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2179 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2180 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2183 addr = (unsigned long)cfs_page_address(page);
2184 clear_page((void *)addr);
2186 /* lprocfs must be setup before the filter so state can be safely added
2187 * to /proc incrementally as the filter is setup */
2188 lprocfs_filter_init_vars(&lvars);
2189 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2190 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2191 /* Init obdfilter private stats here */
2192 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2193 LPROCFS_CNTR_AVGMINMAX,
2194 "read_bytes", "bytes");
2195 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2196 LPROCFS_CNTR_AVGMINMAX,
2197 "write_bytes", "bytes");
2198 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2199 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2200 "get_page", "usec");
2201 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2202 LPROCFS_CNTR_AVGMINMAX,
2203 "get_page_failures", "num");
2204 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2205 LPROCFS_CNTR_AVGMINMAX,
2206 "cache_access", "pages");
2207 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2208 LPROCFS_CNTR_AVGMINMAX,
2209 "cache_hit", "pages");
2210 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2211 LPROCFS_CNTR_AVGMINMAX,
2212 "cache_miss", "pages");
2214 lproc_filter_attach_seqstat(obd);
2215 obd->obd_proc_exports_entry = lprocfs_register("exports",
2216 obd->obd_proc_entry,
2218 if (IS_ERR(obd->obd_proc_exports_entry)) {
2219 rc = PTR_ERR(obd->obd_proc_exports_entry);
2220 CERROR("error %d setting up lprocfs for %s\n",
2222 obd->obd_proc_exports_entry = NULL;
2225 if (obd->obd_proc_exports_entry)
2226 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2227 lprocfs_nid_stats_clear_read,
2228 lprocfs_nid_stats_clear_write, obd, NULL);
2230 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2231 LUSTRE_CFG_BUFLEN(lcfg, 4));
2232 rc = filter_common_setup(obd, lcfg, (void *)addr);
2233 OBD_PAGE_FREE(page);
2236 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2237 lprocfs_free_per_client_stats(obd);
2238 lprocfs_free_obd_stats(obd);
2239 lprocfs_obd_cleanup(obd);
2245 static struct llog_operations filter_mds_ost_repl_logops;
2247 static struct llog_operations filter_size_orig_logops = {
2248 .lop_setup = llog_obd_origin_setup,
2249 .lop_cleanup = llog_obd_origin_cleanup,
2250 .lop_add = llog_obd_origin_add
2253 static int filter_olg_fini(struct obd_llog_group *olg)
2255 struct llog_ctxt *ctxt;
2256 int rc = 0, rc2 = 0;
2259 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2261 rc = llog_cleanup(ctxt);
2263 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2265 rc2 = llog_cleanup(ctxt);
2270 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2272 rc2 = llog_cleanup(ctxt);
2281 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2282 struct obd_device *tgt)
2287 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2288 &filter_mds_ost_repl_logops);
2292 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2293 &filter_size_orig_logops);
2299 filter_olg_fini(olg);
2304 * Init the default olg, which is embeded in the obd_device, for filter.
2307 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2308 struct obd_device *tgt)
2310 struct filter_obd *filter = &obd->u.filter;
2311 struct llog_ctxt *ctxt;
2315 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2316 if (!filter->fo_lcm)
2319 filter_mds_ost_repl_logops = llog_client_ops;
2320 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2321 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2322 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2324 rc = filter_olg_init(obd, olg, tgt);
2326 GOTO(cleanup_lcm, rc);
2328 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2331 GOTO(cleanup_olg, rc);
2333 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2335 CERROR("Can't get ctxt for %p:%x\n", olg,
2336 LLOG_MDS_OST_REPL_CTXT);
2337 GOTO(cleanup_olg, rc = -ENODEV);
2339 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2340 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2341 llog_ctxt_put(ctxt);
2345 filter_olg_fini(olg);
2347 llog_recov_thread_fini(filter->fo_lcm, 1);
2348 filter->fo_lcm = NULL;
2353 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2354 struct obd_device *tgt, int *index)
2356 struct filter_obd *filter = &obd->u.filter;
2357 struct llog_ctxt *ctxt;
2361 LASSERT(olg != NULL);
2362 if (olg == &obd->obd_olg)
2363 return filter_default_olg_init(obd, olg, tgt);
2365 LASSERT(filter->fo_lcm != NULL);
2366 rc = filter_olg_init(obd, olg, tgt);
2369 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2371 CERROR("Can't get ctxt for %p:%x\n", olg,
2372 LLOG_MDS_OST_REPL_CTXT);
2373 filter_olg_fini(olg);
2376 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2377 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2378 llog_ctxt_put(ctxt);
2382 static int filter_llog_finish(struct obd_device *obd, int count)
2384 struct filter_obd *filter = &obd->u.filter;
2385 struct llog_ctxt *ctxt;
2388 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2391 * Make sure that no cached llcds left in recov_thread.
2392 * We actually do sync in disconnect time, but disconnect
2393 * may not come being marked rq_no_resend = 1.
2395 llog_sync(ctxt, NULL);
2398 * Balance class_import_get() in llog_receptor_accept().
2399 * This is safe to do, as llog is already synchronized
2400 * and its import may go.
2402 mutex_down(&ctxt->loc_sem);
2403 if (ctxt->loc_imp) {
2404 class_import_put(ctxt->loc_imp);
2405 ctxt->loc_imp = NULL;
2407 mutex_up(&ctxt->loc_sem);
2408 llog_ctxt_put(ctxt);
2411 if (filter->fo_lcm) {
2412 mutex_down(&ctxt->loc_sem);
2413 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2414 filter->fo_lcm = NULL;
2415 mutex_up(&ctxt->loc_sem);
2417 RETURN(filter_olg_fini(&obd->obd_olg));
2421 * Find the group llog according to group index in the llog group list.
2423 static struct obd_llog_group *
2424 filter_find_olg_internal(struct filter_obd *filter, int group)
2426 struct obd_llog_group *olg;
2428 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2429 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2430 if (olg->olg_group == group)
2437 * Find the group llog according to group index on the filter
2439 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2441 struct obd_llog_group *olg = NULL;
2442 struct filter_obd *filter;
2444 filter = &obd->u.filter;
2446 if (group == FILTER_GROUP_LLOG)
2447 RETURN(&obd->obd_olg);
2449 spin_lock(&filter->fo_llog_list_lock);
2450 olg = filter_find_olg_internal(filter, group);
2451 spin_unlock(&filter->fo_llog_list_lock);
2456 * Find the llog_group of the filter according to the group. If it can not
2457 * find, create the llog_group, which only happens when mds is being synced
2460 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2462 struct obd_llog_group *olg = NULL;
2463 struct filter_obd *filter;
2466 filter = &obd->u.filter;
2468 if (group == FILTER_GROUP_LLOG)
2469 RETURN(&obd->obd_olg);
2471 spin_lock(&filter->fo_llog_list_lock);
2472 olg = filter_find_olg_internal(filter, group);
2474 if (olg->olg_initializing) {
2475 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2477 GOTO(out_unlock, olg);
2482 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2484 llog_group_init(olg, group);
2485 list_add(&olg->olg_list, &filter->fo_llog_list);
2486 olg->olg_initializing = 1;
2487 spin_unlock(&filter->fo_llog_list_lock);
2489 rc = obd_llog_init(obd, olg, obd, NULL);
2491 spin_lock(&filter->fo_llog_list_lock);
2492 list_del(&olg->olg_list);
2493 spin_unlock(&filter->fo_llog_list_lock);
2495 GOTO(out, olg = ERR_PTR(-ENOMEM));
2497 spin_lock(&filter->fo_llog_list_lock);
2498 olg->olg_initializing = 0;
2499 spin_unlock(&filter->fo_llog_list_lock);
2500 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2501 obd->obd_name, group, olg);
2506 spin_unlock(&filter->fo_llog_list_lock);
2510 static int filter_llog_connect(struct obd_export *exp,
2511 struct llogd_conn_body *body)
2513 struct obd_device *obd = exp->exp_obd;
2514 struct llog_ctxt *ctxt;
2515 struct obd_llog_group *olg;
2519 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2520 obd->obd_name, body->lgdc_logid.lgl_oid,
2521 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2523 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2525 CERROR(" %s: can not find olg of group %d\n",
2526 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2529 llog_group_set_export(olg, exp);
2531 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2532 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2533 body->lgdc_ctxt_idx);
2535 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2536 obd->obd_name, body->lgdc_logid.lgl_oid,
2537 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2539 spin_lock_bh(&obd->obd_processing_task_lock);
2540 obd->u.filter.fo_mds_ost_sync = 1;
2541 spin_unlock_bh(&obd->obd_processing_task_lock);
2542 rc = llog_connect(ctxt, &body->lgdc_logid,
2543 &body->lgdc_gen, NULL);
2544 llog_ctxt_put(ctxt);
2546 CERROR("failed to connect rc %d idx %d\n", rc,
2547 body->lgdc_ctxt_idx);
2552 static int filter_llog_preclean(struct obd_device *obd)
2554 struct obd_llog_group *olg, *tmp;
2555 struct filter_obd *filter;
2556 struct list_head remove_list;
2560 rc = obd_llog_finish(obd, 0);
2562 CERROR("failed to cleanup llogging subsystem\n");
2564 filter = &obd->u.filter;
2565 CFS_INIT_LIST_HEAD(&remove_list);
2567 spin_lock(&filter->fo_llog_list_lock);
2568 while (!list_empty(&filter->fo_llog_list)) {
2569 olg = list_entry(filter->fo_llog_list.next,
2570 struct obd_llog_group, olg_list);
2571 list_del(&olg->olg_list);
2572 list_add(&olg->olg_list, &remove_list);
2574 spin_unlock(&filter->fo_llog_list_lock);
2576 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2577 list_del_init(&olg->olg_list);
2578 rc = filter_olg_fini(olg);
2580 CERROR("failed to cleanup llogging subsystem for %u\n",
2588 static int filter_precleanup(struct obd_device *obd,
2589 enum obd_cleanup_stage stage)
2595 case OBD_CLEANUP_EARLY:
2597 case OBD_CLEANUP_EXPORTS:
2598 /* Stop recovery before namespace cleanup. */
2599 target_recovery_fini(obd);
2600 rc = filter_llog_preclean(obd);
2606 static int filter_cleanup(struct obd_device *obd)
2608 struct filter_obd *filter = &obd->u.filter;
2612 LCONSOLE_WARN("%s: shutting down for failover; client state "
2613 "will be preserved.\n", obd->obd_name);
2615 obd_exports_barrier(obd);
2616 obd_zombie_barrier();
2618 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2619 lprocfs_free_per_client_stats(obd);
2620 lprocfs_free_obd_stats(obd);
2621 lprocfs_obd_cleanup(obd);
2622 lquota_cleanup(filter_quota_interface_ref, obd);
2624 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2625 obd->obd_namespace = NULL;
2627 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2629 if (obd->u.obt.obt_sb == NULL)
2634 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2635 shrink_dcache_sb(obd->u.obt.obt_sb);
2637 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2638 obd->u.obt.obt_sb = NULL;
2640 fsfilt_put_ops(obd->obd_fsops);
2642 filter_iobuf_pool_done(filter);
2644 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2649 static int filter_connect_internal(struct obd_export *exp,
2650 struct obd_connect_data *data,
2653 struct filter_export_data *fed = &exp->exp_filter_data;
2658 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2659 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2660 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2661 data->ocd_connect_flags, data->ocd_version,
2662 data->ocd_grant, data->ocd_index);
2664 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2665 CWARN("!!! This export (nid %s) used object group %d "
2666 "earlier; now it's trying to use group %d! This could "
2667 "be a bug in the MDS. Please report to "
2668 "http://bugzilla.lustre.org/\n",
2669 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2672 fed->fed_group = data->ocd_group;
2674 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2675 exp->exp_connect_flags = data->ocd_connect_flags;
2676 data->ocd_version = LUSTRE_VERSION_CODE;
2678 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2679 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2680 data->ocd_connect_flags & OBD_CONNECT_MDS))
2683 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2684 struct filter_obd *filter = &exp->exp_obd->u.filter;
2685 obd_size left, want;
2687 spin_lock(&exp->exp_obd->obd_osfs_lock);
2688 left = filter_grant_space_left(exp);
2689 want = data->ocd_grant;
2690 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2691 data->ocd_grant = fed->fed_grant;
2692 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2694 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2695 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2696 exp->exp_client_uuid.uuid, exp,
2697 data->ocd_grant, want, left);
2699 filter->fo_tot_granted_clients ++;
2702 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2703 struct filter_obd *filter = &exp->exp_obd->u.filter;
2704 struct lr_server_data *lsd = filter->fo_fsd;
2705 int index = le32_to_cpu(lsd->lsd_ost_index);
2707 if (!(lsd->lsd_feature_compat &
2708 cpu_to_le32(OBD_COMPAT_OST))) {
2709 /* this will only happen on the first connect */
2710 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2711 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2712 /* sync is not needed here as filter_client_add will
2713 * set exp_need_sync flag */
2714 filter_update_server_data(exp->exp_obd,
2715 filter->fo_rcvd_filp, lsd);
2716 } else if (index != data->ocd_index) {
2717 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2718 " %u doesn't match actual OST index"
2719 " %u in last_rcvd file, bad "
2721 obd_export_nid2str(exp), index,
2725 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2726 * FIXME: We don't strictly need the COMPAT flag for that,
2727 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2728 * FIXME: We needed it for the index, as index 0 is valid. */
2731 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2732 data->ocd_brw_size = 65536;
2733 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2734 data->ocd_brw_size = min(data->ocd_brw_size,
2735 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2736 LASSERT(data->ocd_brw_size);
2739 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2740 __u32 cksum_types = data->ocd_cksum_types;
2742 /* The client set in ocd_cksum_types the checksum types it
2743 * supports. We have to mask off the algorithms that we don't
2745 if (cksum_types & OBD_CKSUM_ALL)
2746 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2748 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2750 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2751 "%x\n", exp->exp_obd->obd_name,
2752 obd_export_nid2str(exp), cksum_types,
2753 data->ocd_cksum_types);
2755 /* This client does not support OBD_CONNECT_CKSUM
2756 * fall back to CRC32 */
2757 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2758 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2759 exp->exp_obd->obd_name,
2760 obd_export_nid2str(exp));
2766 static int filter_reconnect(const struct lu_env *env,
2767 struct obd_export *exp, struct obd_device *obd,
2768 struct obd_uuid *cluuid,
2769 struct obd_connect_data *data,
2775 if (exp == NULL || obd == NULL || cluuid == NULL)
2778 rc = filter_connect_internal(exp, data, 1);
2780 filter_export_stats_init(obd, exp, localdata);
2785 /* nearly identical to mds_connect */
2786 static int filter_connect(const struct lu_env *env,
2787 struct obd_export **exp, struct obd_device *obd,
2788 struct obd_uuid *cluuid,
2789 struct obd_connect_data *data, void *localdata)
2791 struct lvfs_run_ctxt saved;
2792 struct lustre_handle conn = { 0 };
2793 struct obd_export *lexp;
2794 struct filter_export_data *fed;
2795 struct lsd_client_data *lcd = NULL;
2800 if (exp == NULL || obd == NULL || cluuid == NULL)
2803 rc = class_connect(&conn, obd, cluuid);
2806 lexp = class_conn2export(&conn);
2807 LASSERT(lexp != NULL);
2809 fed = &lexp->exp_filter_data;
2811 rc = filter_connect_internal(lexp, data, 0);
2815 filter_export_stats_init(obd, lexp, localdata);
2816 if (obd->obd_replayable) {
2817 OBD_ALLOC(lcd, sizeof(*lcd));
2819 CERROR("filter: out of memory for client data\n");
2820 GOTO(cleanup, rc = -ENOMEM);
2823 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2825 rc = filter_client_add(obd, lexp, -1);
2830 group = data->ocd_group;
2832 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2833 obd->obd_name, lexp->exp_handle.h_cookie, group);
2835 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2836 rc = filter_read_groups(obd, group, 1);
2837 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2839 CERROR("can't read group %u\n", group);
2849 fed->fed_lcd = NULL;
2851 class_disconnect(lexp);
2860 /* Do extra sanity checks for grant accounting. We do this at connect,
2861 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2862 * always get rid of it or turn it off when we know accounting is good. */
2863 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2865 struct filter_export_data *fed;
2866 struct obd_export *exp;
2867 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2868 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2869 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2871 if (list_empty(&obd->obd_exports))
2874 /* We don't want to do this for large machines that do lots of
2875 mounts or unmounts. It burns... */
2876 if (obd->obd_num_exports > 100)
2879 spin_lock(&obd->obd_osfs_lock);
2880 spin_lock(&obd->obd_dev_lock);
2881 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2883 fed = &exp->exp_filter_data;
2884 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2887 if (maxsize > 0) { /* we may not have done a statfs yet */
2888 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2889 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2890 exp->exp_client_uuid.uuid, exp,
2891 fed->fed_grant, fed->fed_pending, maxsize);
2892 LASSERTF(fed->fed_dirty <= maxsize,
2893 "%s: cli %s/%p %ld > "LPU64"\n", func,
2894 exp->exp_client_uuid.uuid, exp,
2895 fed->fed_dirty, maxsize);
2898 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2899 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2900 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2902 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2903 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2904 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2905 tot_granted += fed->fed_grant + fed->fed_pending;
2906 tot_pending += fed->fed_pending;
2907 tot_dirty += fed->fed_dirty;
2909 fo_tot_granted = obd->u.filter.fo_tot_granted;
2910 fo_tot_pending = obd->u.filter.fo_tot_pending;
2911 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2912 spin_unlock(&obd->obd_dev_lock);
2913 spin_unlock(&obd->obd_osfs_lock);
2915 /* Do these assertions outside the spinlocks so we don't kill system */
2916 if (tot_granted != fo_tot_granted)
2917 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2918 func, tot_granted, fo_tot_granted);
2919 if (tot_pending != fo_tot_pending)
2920 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2921 func, tot_pending, fo_tot_pending);
2922 if (tot_dirty != fo_tot_dirty)
2923 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2924 func, tot_dirty, fo_tot_dirty);
2925 if (tot_pending > tot_granted)
2926 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2927 func, tot_pending, tot_granted);
2928 if (tot_granted > maxsize)
2929 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2930 func, tot_granted, maxsize);
2931 if (tot_dirty > maxsize)
2932 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2933 func, tot_dirty, maxsize);
2936 /* Remove this client from the grant accounting totals. We also remove
2937 * the export from the obd device under the osfs and dev locks to ensure
2938 * that the filter_grant_sanity_check() calculations are always valid.
2939 * The client should do something similar when it invalidates its import. */
2940 static void filter_grant_discard(struct obd_export *exp)
2942 struct obd_device *obd = exp->exp_obd;
2943 struct filter_obd *filter = &obd->u.filter;
2944 struct filter_export_data *fed = &exp->exp_filter_data;
2946 spin_lock(&obd->obd_osfs_lock);
2947 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2948 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2949 obd->obd_name, filter->fo_tot_granted,
2950 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2951 filter->fo_tot_granted -= fed->fed_grant;
2952 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2953 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2954 obd->obd_name, filter->fo_tot_pending,
2955 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2956 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2957 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2958 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2959 obd->obd_name, filter->fo_tot_dirty,
2960 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2961 filter->fo_tot_dirty -= fed->fed_dirty;
2965 spin_unlock(&obd->obd_osfs_lock);
2968 static int filter_destroy_export(struct obd_export *exp)
2972 if (exp->exp_filter_data.fed_pending)
2973 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2974 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2975 exp, exp->exp_filter_data.fed_pending);
2977 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2979 target_destroy_export(exp);
2980 ldlm_destroy_export(exp);
2982 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2986 if (!exp->exp_obd->obd_replayable)
2987 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2989 filter_grant_discard(exp);
2990 filter_fmd_cleanup(exp);
2992 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2993 struct filter_obd *filter = &exp->exp_obd->u.filter;
2994 if (filter->fo_tot_granted_clients > 0)
2995 filter->fo_tot_granted_clients --;
2998 if (!(exp->exp_flags & OBD_OPT_FORCE))
2999 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3004 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3006 struct obd_llog_group *olg_min, *olg;
3007 struct filter_obd *filter;
3008 int worked = -1, group;
3009 struct llog_ctxt *ctxt;
3012 filter = &obd->u.filter;
3014 /* we can't sync log holding spinlock. also, we do not want to get
3015 * into livelock. so we do following: loop over MDS's exports in
3016 * group order and skip already synced llogs -bzzz */
3018 /* look for group with min. number, but > worked */
3021 spin_lock(&filter->fo_llog_list_lock);
3022 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3023 if (olg->olg_group <= worked) {
3024 /* this group is already synced */
3027 if (group < olg->olg_group) {
3028 /* we have group with smaller number to sync */
3031 /* store current minimal group */
3033 group = olg->olg_group;
3035 spin_unlock(&filter->fo_llog_list_lock);
3037 if (olg_min == NULL)
3040 worked = olg_min->olg_group;
3041 if (olg_min->olg_exp &&
3042 (dexp == olg_min->olg_exp || dexp == NULL)) {
3044 ctxt = llog_group_get_ctxt(olg_min,
3045 LLOG_MDS_OST_REPL_CTXT);
3047 err = llog_sync(ctxt, olg_min->olg_exp);
3048 llog_ctxt_put(ctxt);
3050 CERROR("error flushing logs to MDS: "
3055 } while (olg_min != NULL);
3058 /* Also incredibly similar to mds_disconnect */
3059 static int filter_disconnect(struct obd_export *exp)
3061 struct obd_device *obd = exp->exp_obd;
3066 class_export_get(exp);
3068 if (!(exp->exp_flags & OBD_OPT_FORCE))
3069 filter_grant_sanity_check(obd, __FUNCTION__);
3070 filter_grant_discard(exp);
3072 /* Flush any remaining cancel messages out to the target */
3073 filter_sync_llogs(obd, exp);
3075 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3077 rc = server_disconnect_export(exp);
3079 if (exp->exp_obd->obd_replayable)
3080 filter_client_free(exp);
3082 fsfilt_sync(obd, obd->u.obt.obt_sb);
3084 class_export_put(exp);
3088 /* reverse import is changed, sync all cancels */
3089 static void filter_revimp_update(struct obd_export *exp)
3094 class_export_get(exp);
3096 /* flush any remaining cancel messages out to the target */
3097 filter_sync_llogs(exp->exp_obd, exp);
3098 class_export_put(exp);
3102 static int filter_ping(struct obd_export *exp)
3104 filter_fmd_expire(exp);
3108 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3109 const char *what, int quiet)
3111 struct dentry *dchild = NULL;
3114 if (oa->o_valid & OBD_MD_FLGROUP)
3117 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3119 if (IS_ERR(dchild)) {
3120 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3121 what, group, oa->o_id);
3125 if (dchild->d_inode == NULL) {
3127 CERROR("%s: %s on non-existent object: "LPU64"\n",
3128 obd->obd_name, what, oa->o_id);
3130 RETURN(ERR_PTR(-ENOENT));
3136 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3138 struct dentry *dentry = NULL;
3139 struct obd_device *obd;
3143 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3144 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3145 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3149 obd = class_exp2obd(exp);
3151 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3155 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3157 RETURN(PTR_ERR(dentry));
3159 /* Limit the valid bits in the return data to what we actually use */
3160 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3161 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3167 /* this should be enabled/disabled in condition to enabled/disabled large
3168 * inodes (fast EAs) in backing store FS. */
3169 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3170 void *handle, struct obdo *oa)
3172 struct obd_device *obd = exp->exp_obd;
3176 if (oa->o_valid & OBD_MD_FLFID) {
3177 struct filter_fid ff;
3179 if (!(oa->o_valid & OBD_MD_FLGROUP))
3181 /* packing fid and converting it to LE for storing into EA.
3182 * Here ->o_stripe_idx should be filled by LOV and rest of
3183 * fields - by client. */
3184 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3185 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3186 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3187 ff.ff_objid = cpu_to_le64(oa->o_id);
3188 ff.ff_group = cpu_to_le64(oa->o_gr);
3190 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3191 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3192 oa->o_generation, oa->o_id, oa->o_gr);
3194 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3196 CERROR("store fid in object failed! rc: %d\n", rc);
3198 CDEBUG(D_HA, "OSS object without fid info!\n");
3204 /* this is called from filter_truncate() until we have filter_punch() */
3205 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3206 struct obdo *oa, struct obd_trans_info *oti)
3208 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3209 struct llog_cookie *fcc = NULL;
3210 struct filter_obd *filter;
3211 int rc, err, sync = 0;
3212 loff_t old_size = 0;
3213 unsigned int ia_valid;
3214 struct inode *inode;
3215 struct page *page = NULL;
3220 LASSERT(dentry != NULL);
3221 LASSERT(!IS_ERR(dentry));
3223 inode = dentry->d_inode;
3224 LASSERT(inode != NULL);
3226 filter = &exp->exp_obd->u.filter;
3227 iattr_from_obdo(&iattr, oa, oa->o_valid);
3228 ia_valid = iattr.ia_valid;
3230 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3231 OBD_ALLOC(fcc, sizeof(*fcc));
3233 *fcc = oa->o_lcookie;
3235 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3237 /* Filter truncates and writes are serialized by
3238 * i_alloc_sem, see the comment in
3239 * filter_preprw_write.*/
3240 if (ia_valid & ATTR_SIZE)
3241 down_write(&inode->i_alloc_sem);
3242 LOCK_INODE_MUTEX(inode);
3243 old_size = i_size_read(inode);
3246 /* VBR: version recovery check */
3247 rc = filter_version_get_check(exp, oti, inode);
3249 GOTO(out_unlock, rc);
3251 /* Let's pin the last page so that ldiskfs_truncate
3252 * should not start GFP_FS allocation. */
3253 if (ia_valid & ATTR_SIZE) {
3254 page = grab_cache_page(inode->i_mapping,
3255 iattr.ia_size >> PAGE_CACHE_SHIFT);
3257 GOTO(out_unlock, rc = -ENOMEM);
3262 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3263 * then we will accept the UID+GID sent by the client during write for
3264 * initializing the ownership of this inode. We only allow this to
3265 * happen once so clear these bits in setattr. In 2.6 kernels it is
3266 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3267 * the flags that are actually being set. */
3268 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3269 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3270 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3272 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3273 if (!(ia_valid & ATTR_MODE)) {
3274 iattr.ia_mode = inode->i_mode;
3275 iattr.ia_valid |= ATTR_MODE;
3277 iattr.ia_mode &= ~S_ISUID;
3279 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3280 if (!(iattr.ia_valid & ATTR_MODE)) {
3281 iattr.ia_mode = inode->i_mode;
3282 iattr.ia_valid |= ATTR_MODE;
3284 iattr.ia_mode &= ~S_ISGID;
3287 orig_ids[USRQUOTA] = inode->i_uid;
3288 orig_ids[GRPQUOTA] = inode->i_gid;
3289 handle = fsfilt_start_log(exp->exp_obd, inode,
3290 FSFILT_OP_SETATTR, oti, 1);
3292 GOTO(out_unlock, rc = PTR_ERR(handle));
3294 /* update inode EA only once when inode is suid bit marked. As
3295 * on 2.6.x UID and GID may be set separately, we check here
3296 * only one of them to avoid double setting. */
3297 if (inode->i_mode & S_ISUID)
3298 filter_update_fidea(exp, inode, handle, oa);
3300 handle = fsfilt_start(exp->exp_obd, inode,
3301 FSFILT_OP_SETATTR, oti);
3303 GOTO(out_unlock, rc = PTR_ERR(handle));
3305 if (oa->o_valid & OBD_MD_FLFLAGS) {
3306 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3307 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3309 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3311 /* set cancel cookie callback function */
3312 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3313 filter_cancel_cookies_cb,
3317 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3318 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3320 /* The truncate might have used up our transaction credits. Make
3321 * sure we have one left for the last_rcvd update. */
3322 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3324 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3326 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3330 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3332 CERROR("error on commit, err = %d\n", err);
3339 /* For a partial-page truncate flush the page to disk immediately
3340 * to avoid data corruption during direct disk write. b=17397 */
3341 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3342 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3343 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3353 page_cache_release(page);
3355 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3356 UNLOCK_INODE_MUTEX(inode);
3357 if (ia_valid & ATTR_SIZE)
3358 up_write(&inode->i_alloc_sem);
3360 OBD_FREE(fcc, sizeof(*fcc));
3362 /* trigger quota release */
3363 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3364 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3365 int rc2 = lquota_adjust(filter_quota_interface_ref,
3366 exp->exp_obd, cur_ids,
3367 orig_ids, rc, FSFILT_OP_SETATTR);
3368 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3369 "filter adjust qunit. (rc:%d)\n", rc2);
3374 /* this is called from filter_truncate() until we have filter_punch() */
3375 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3376 struct obd_trans_info *oti)
3378 struct obdo *oa = oinfo->oi_oa;
3379 struct lustre_capa *capa = oinfo_capa(oinfo);
3380 struct ldlm_res_id res_id;
3381 struct filter_mod_data *fmd;
3382 struct lvfs_run_ctxt saved;
3383 struct filter_obd *filter;
3384 struct ldlm_resource *res;
3385 struct dentry *dentry;
3386 __u64 opc = CAPA_OPC_META_WRITE;
3390 if (oa->o_valid & OBD_FL_TRUNC)
3391 opc |= CAPA_OPC_OSS_TRUNC;
3393 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3394 rc = filter_auth_capa(exp, NULL, oa->o_gr, capa, opc);
3398 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3399 rc = filter_capa_fixoa(exp, oa, oa->o_gr, capa);
3404 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3405 /* This would be very bad - accidentally truncating a file when
3406 * changing the time or similar - bug 12203. */
3407 if (oa->o_valid & OBD_MD_FLSIZE &&
3408 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3409 static char mdsinum[48];
3411 if (oa->o_valid & OBD_MD_FLFID)
3412 snprintf(mdsinum, sizeof(mdsinum) - 1,
3413 " of inode "LPU64"/%u", oa->o_fid,
3418 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3420 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3425 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3427 RETURN(PTR_ERR(dentry));
3429 filter = &exp->exp_obd->u.filter;
3430 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3433 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3434 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3435 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3436 fmd->fmd_mactime_xid = oti->oti_xid;
3437 filter_fmd_put(exp, fmd);
3440 /* setting objects attributes (including owner/group) */
3441 rc = filter_setattr_internal(exp, dentry, oa, oti);
3443 GOTO(out_unlock, rc);
3445 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3446 &res_id, LDLM_EXTENT, 0);
3449 LDLM_RESOURCE_ADDREF(res);
3450 rc = ldlm_res_lvbo_update(res, NULL, 0);
3451 LDLM_RESOURCE_DELREF(res);
3452 ldlm_resource_putref(res);
3455 oa->o_valid = OBD_MD_FLID;
3457 /* Quota release need uid/gid info */
3458 obdo_from_inode(oa, dentry->d_inode,
3459 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3464 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3468 /* XXX identical to osc_unpackmd */
3469 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3470 struct lov_mds_md *lmm, int lmm_bytes)
3476 if (lmm_bytes < sizeof (*lmm)) {
3477 CERROR("lov_mds_md too small: %d, need %d\n",
3478 lmm_bytes, (int)sizeof(*lmm));
3481 /* XXX LOV_MAGIC etc check? */
3483 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3484 CERROR("lov_mds_md: zero lmm_object_id\n");
3489 lsm_size = lov_stripe_md_size(1);
3493 if (*lsmp != NULL && lmm == NULL) {
3494 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3495 OBD_FREE(*lsmp, lsm_size);
3500 if (*lsmp == NULL) {
3501 OBD_ALLOC(*lsmp, lsm_size);
3505 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3506 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3507 OBD_FREE(*lsmp, lsm_size);
3510 loi_init((*lsmp)->lsm_oinfo[0]);
3514 /* XXX zero *lsmp? */
3515 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3516 LASSERT((*lsmp)->lsm_object_id);
3519 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3524 /* caller must hold fo_create_locks[oa->o_gr] */
3525 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3526 struct filter_obd *filter)
3528 struct obdo doa = { 0 }; /* XXX obdo on stack */
3535 LASSERT_MDS_GROUP(oa->o_gr);
3536 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3537 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3539 memset(&doa, 0, sizeof(doa));
3541 doa.o_valid |= OBD_MD_FLGROUP;
3542 doa.o_gr = oa->o_gr;
3543 doa.o_mode = S_IFREG;
3545 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3546 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3547 exp->exp_obd->obd_name, doa.o_gr);
3551 last = filter_last_id(filter, doa.o_gr);
3553 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3555 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3556 exp->exp_obd->obd_name, oa->o_id + 1, last,
3557 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3559 for (id = last; id > oa->o_id; id--) {
3561 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3562 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3563 CEMERG("error destroying precreate objid "LPU64": %d\n",
3566 /* update last_id on disk periodically so that if we restart
3567 * we don't need to re-scan all of the just-deleted objects. */
3568 if ((id & 511) == 0 && !skip_orphan) {
3569 filter_set_last_id(filter, id - 1, doa.o_gr);
3570 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3574 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3575 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3578 filter_set_last_id(filter, id, doa.o_gr);
3579 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3581 /* don't reuse orphan object, return last used objid */
3585 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3590 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3591 obd_gr group, int *num);
3592 /* returns a negative error or a nonnegative number of files to create */
3593 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3594 obd_gr group, struct obd_trans_info *oti)
3596 struct obd_device *obd = exp->exp_obd;
3597 struct filter_obd *filter = &obd->u.filter;
3601 /* delete orphans request */
3602 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3603 obd_id last = filter_last_id(filter, group);
3605 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3606 CERROR("%s: dropping old orphan cleanup request\n",
3610 /* This causes inflight precreates to abort and drop lock */
3611 set_bit(group, &filter->fo_destroys_in_progress);
3612 down(&filter->fo_create_locks[group]);
3613 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3614 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3615 exp->exp_obd->obd_name, group);
3616 up(&filter->fo_create_locks[group]);
3619 diff = oa->o_id - last;
3620 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3623 if (-diff > OST_MAX_PRECREATE) {
3624 CERROR("%s: ignoring bogus orphan destroy request: "
3625 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3627 /* FIXME: should reset precreate_next_id on MDS */
3628 GOTO(out, rc = -EINVAL);
3631 rc = filter_destroy_precreated(exp, oa, filter);
3633 CERROR("%s: unable to write lastobjid, but "
3634 "orphans were deleted\n", obd->obd_name);
3637 /* XXX: Used by MDS for the first time! */
3638 clear_bit(group, &filter->fo_destroys_in_progress);
3641 down(&filter->fo_create_locks[group]);
3642 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3643 CERROR("%s: dropping old precreate request\n",
3647 /* only precreate if group == 0 and o_id is specfied */
3648 if (!filter_group_is_mds(group) || oa->o_id == 0)
3651 diff = oa->o_id - filter_last_id(filter, group);
3652 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3653 filter_last_id(filter, group), diff);
3655 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3656 oa->o_id, filter_last_id(filter, group), diff);
3660 oa->o_id = filter_last_id(&obd->u.filter, group);
3661 rc = filter_precreate(obd, oa, group, &diff);
3662 oa->o_id = filter_last_id(&obd->u.filter, group);
3664 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3667 /* else diff == 0 */
3670 up(&filter->fo_create_locks[group]);
3674 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3675 __u64 max_age, __u32 flags)
3677 struct filter_obd *filter = &obd->u.filter;
3678 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3682 /* at least try to account for cached pages. its still racey and
3683 * might be under-reporting if clients haven't announced their
3684 * caches with brw recently */
3685 spin_lock(&obd->obd_osfs_lock);
3686 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3687 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3688 spin_unlock(&obd->obd_osfs_lock);
3690 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3691 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3692 filter->fo_tot_dirty, filter->fo_tot_granted,
3693 filter->fo_tot_pending,
3694 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3696 filter_grant_sanity_check(obd, __FUNCTION__);
3698 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3699 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3700 osfs->os_bsize - 1) >> blockbits));
3702 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3703 struct lr_server_data *lsd = filter->fo_fsd;
3704 int index = le32_to_cpu(lsd->lsd_ost_index);
3706 if (obd_fail_val == -1 ||
3707 index == obd_fail_val)
3708 osfs->os_bfree = osfs->os_bavail = 2;
3709 else if (obd_fail_loc & OBD_FAIL_ONCE)
3710 obd_fail_loc &= ~OBD_FAILED; /* reset flag */
3713 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3714 * stop creating files on MDS if OST is not good shape to create
3718 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3719 osfs->os_state = OS_STATE_READONLY;
3721 if (filter->fo_raid_degraded)
3722 osfs->os_state |= OS_STATE_DEGRADED;
3726 static int filter_use_existing_obj(struct obd_device *obd,
3727 struct dentry *dchild, void **handle,
3730 struct inode *inode = dchild->d_inode;
3734 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3737 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3738 if (IS_ERR(*handle))
3739 return PTR_ERR(*handle);
3741 iattr.ia_valid = ATTR_MODE;
3742 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3743 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3750 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3753 __u64 os_ffree = -1;
3755 spin_lock(&obd->obd_osfs_lock);
3756 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3758 os_ffree = obd->obd_osfs.os_ffree;
3759 spin_unlock(&obd->obd_osfs_lock);
3764 /* We rely on the fact that only one thread will be creating files in a given
3765 * group at a time, which is why we don't need an atomic filter_get_new_id.
3766 * Even if we had that atomic function, the following race would exist:
3768 * thread 1: gets id x from filter_next_id
3769 * thread 2: gets id (x + 1) from filter_next_id
3770 * thread 2: creates object (x + 1)
3771 * thread 1: tries to create object x, gets -ENOSPC
3773 * Caller must hold fo_create_locks[group]
3775 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3776 obd_gr group, int *num)
3778 struct dentry *dchild = NULL, *dparent = NULL;
3779 struct filter_obd *filter;
3780 struct obd_statfs *osfs;
3781 int err = 0, rc = 0, recreate_obj = 0, i;
3782 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3785 void *handle = NULL;
3788 filter = &obd->u.filter;
3790 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3792 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3794 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3795 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3798 OBD_ALLOC(osfs, sizeof(*osfs));
3801 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3802 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3803 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3804 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3805 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3809 OBD_FREE(osfs, sizeof(*osfs));
3814 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3815 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3817 for (i = 0; i < *num && err == 0; i++) {
3818 int cleanup_phase = 0;
3820 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3821 CWARN("%s: create aborted by destroy\n",
3830 last_id = filter_last_id(filter, group);
3831 if (next_id > last_id) {
3832 CERROR("Error: Trying to recreate obj greater"
3833 "than last id "LPD64" > "LPD64"\n",
3835 GOTO(cleanup, rc = -EINVAL);
3838 next_id = filter_last_id(filter, group) + 1;
3840 dparent = filter_parent_lock(obd, group, next_id);
3841 if (IS_ERR(dparent))
3842 GOTO(cleanup, rc = PTR_ERR(dparent));
3843 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3845 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3847 GOTO(cleanup, rc = PTR_ERR(dchild));
3848 cleanup_phase = 2; /* f_dput(dchild) */
3850 if (dchild->d_inode != NULL) {
3851 /* This would only happen if lastobjid was bad on disk*/
3852 /* Could also happen if recreating missing obj but it
3853 * already exists. */
3855 CERROR("%s: recreating existing object %.*s?\n",
3856 obd->obd_name, dchild->d_name.len,
3857 dchild->d_name.name);
3859 /* Use these existing objects if they are
3861 if (dchild->d_inode->i_size == 0) {
3862 rc = filter_use_existing_obj(obd,dchild,
3863 &handle, &cleanup_phase);
3870 CERROR("%s: Serious error: objid %.*s already "
3871 "exists; is this filesystem corrupt?\n",
3872 obd->obd_name, dchild->d_name.len,
3873 dchild->d_name.name);
3876 GOTO(cleanup, rc = -EEXIST);
3879 handle = fsfilt_start_log(obd, dparent->d_inode,
3880 FSFILT_OP_CREATE, NULL, 1);
3882 GOTO(cleanup, rc = PTR_ERR(handle));
3885 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3886 ",od->o_id="LPU64")\n", obd->obd_name, group,
3889 /* We mark object SUID+SGID to flag it for accepting UID+GID
3890 * from client on first write. Currently the permission bits
3891 * on the OST are never used, so this is OK. */
3892 rc = ll_vfs_create(dparent->d_inode, dchild,
3893 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3895 CERROR("create failed rc = %d\n", rc);
3896 if (rc == -ENOSPC) {
3897 os_ffree = filter_calc_free_inodes(obd);
3899 CERROR("%s: free inode "LPU64"\n",
3900 obd->obd_name, os_ffree);
3905 if (dchild->d_inode)
3906 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3907 dchild->d_inode->i_ino);
3910 if (!recreate_obj) {
3911 filter_set_last_id(filter, next_id, group);
3912 err = filter_update_last_objid(obd, group, 0);
3914 CERROR("unable to write lastobjid "
3915 "but file created\n");
3919 switch(cleanup_phase) {
3921 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3923 CERROR("error on commit, err = %d\n", err);
3930 filter_parent_unlock(dparent);
3937 if (time_after(jiffies, enough_time)) {
3939 "%s: precreate slow - want %d got %d \n",
3940 obd->obd_name, *num, i);
3947 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3948 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3953 static int filter_create(struct obd_export *exp, struct obdo *oa,
3954 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3956 struct obd_device *obd = exp->exp_obd;
3957 struct filter_export_data *fed;
3958 struct filter_obd *filter;
3959 struct lvfs_run_ctxt saved;
3960 struct lov_stripe_md *lsm = NULL;
3961 int rc = 0, diff, group = oa->o_gr;
3964 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3965 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3967 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3968 CERROR("!!! nid %s sent invalid object group %d\n",
3969 obd_export_nid2str(exp), group);
3973 fed = &exp->exp_filter_data;
3974 filter = &obd->u.filter;
3976 if (fed->fed_group != group) {
3977 CERROR("!!! this export (nid %s) used object group %d "
3978 "earlier; now it's trying to use group %d! This could "
3979 "be a bug in the MDS. Please report to "
3980 "http://bugzilla.lustre.org/\n",
3981 obd_export_nid2str(exp), fed->fed_group, group);
3988 rc = obd_alloc_memmd(exp, &lsm);
3995 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3997 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3998 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3999 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
4000 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4001 oa->o_id, filter_last_id(filter,
4006 down(&filter->fo_create_locks[oa->o_gr]);
4007 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
4008 up(&filter->fo_create_locks[oa->o_gr]);
4011 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
4014 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4015 if (rc && ea != NULL && *ea != lsm) {
4016 obd_free_memmd(exp, &lsm);
4017 } else if (rc == 0 && ea != NULL) {
4018 /* XXX LOV STACKING: the lsm that is passed to us from
4019 * LOV does not have valid lsm_oinfo data structs, so
4020 * don't go touching that. This needs to be fixed in a
4022 lsm->lsm_object_id = oa->o_id;
4029 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4030 struct lov_stripe_md *md, struct obd_trans_info *oti,
4031 struct obd_export *md_exp, void *capa)
4033 unsigned int qcids[MAXQUOTAS] = {0, 0};
4034 struct obd_device *obd;
4035 struct filter_obd *filter;
4036 struct dentry *dchild = NULL, *dparent = NULL;
4037 struct lustre_handle lockh = { 0 };
4038 struct lvfs_run_ctxt saved;
4039 void *handle = NULL;
4040 struct llog_cookie *fcc = NULL;
4041 int rc, rc2, cleanup_phase = 0, sync = 0;
4045 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4046 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4047 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4052 filter = &obd->u.filter;
4054 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4057 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4058 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4060 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4062 GOTO(cleanup, rc = PTR_ERR(dchild));
4065 if (dchild->d_inode == NULL) {
4066 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4068 /* If object already gone, cancel cookie right now */
4069 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4070 struct llog_ctxt *ctxt;
4071 struct obd_llog_group *olg;
4073 olg = filter_find_olg(obd, oa->o_gr);
4075 CERROR(" %s: can not find olg of group %d\n",
4076 obd->obd_name, (int)oa->o_gr);
4077 GOTO(cleanup, rc = PTR_ERR(olg));
4079 fcc = &oa->o_lcookie;
4080 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4081 llog_cancel(ctxt, NULL, 1, fcc, 0);
4082 llog_ctxt_put(ctxt);
4083 fcc = NULL; /* we didn't allocate fcc, don't free it */
4085 GOTO(cleanup, rc = -ENOENT);
4088 filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4090 /* Our MDC connection is established by the MDS to us */
4091 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4092 OBD_ALLOC(fcc, sizeof(*fcc));
4094 *fcc = oa->o_lcookie;
4096 DQUOT_INIT(dchild->d_inode);
4098 /* we're gonna truncate it first in order to avoid possible deadlock:
4100 * open trasaction open transaction
4101 * down(i_zombie) down(i_zombie)
4102 * restart transaction
4103 * (see BUG 4180) -bzzz
4105 * take i_alloc_sem too to prevent other threads from writing to the
4106 * file while we are truncating it. This can cause lock ordering issue
4107 * between page lock, i_mutex & starting new journal handle.
4108 * (see bug 20321) -johann
4110 down_write(&dchild->d_inode->i_alloc_sem);
4111 LOCK_INODE_MUTEX(dchild->d_inode);
4113 /* VBR: version recovery check */
4114 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4116 UNLOCK_INODE_MUTEX(dchild->d_inode);
4117 up_write(&dchild->d_inode->i_alloc_sem);
4121 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4123 if (IS_ERR(handle)) {
4124 UNLOCK_INODE_MUTEX(dchild->d_inode);
4125 up_write(&dchild->d_inode->i_alloc_sem);
4126 GOTO(cleanup, rc = PTR_ERR(handle));
4129 iattr.ia_valid = ATTR_SIZE;
4131 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4132 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4133 UNLOCK_INODE_MUTEX(dchild->d_inode);
4134 up_write(&dchild->d_inode->i_alloc_sem);
4138 GOTO(cleanup, rc = rc2);
4140 /* We don't actually need to lock the parent until we are unlinking
4141 * here, and not while truncating above. That avoids holding the
4142 * parent lock for a long time during truncate, which can block other
4143 * threads from doing anything to objects in that directory. bug 7171 */
4144 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4145 if (IS_ERR(dparent))
4146 GOTO(cleanup, rc = PTR_ERR(dparent));
4147 cleanup_phase = 3; /* filter_parent_unlock */
4149 LOCK_INODE_MUTEX(dchild->d_inode);
4150 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4151 if (IS_ERR(handle)) {
4152 UNLOCK_INODE_MUTEX(dchild->d_inode);
4153 GOTO(cleanup, rc = PTR_ERR(handle));
4155 cleanup_phase = 4; /* fsfilt_commit */
4157 /* Quota release need uid/gid of inode */
4158 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4160 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4162 /* this drops dchild->d_inode->i_mutex unconditionally */
4163 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4167 switch(cleanup_phase) {
4170 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4171 oti->oti_handle : handle,
4172 filter_cancel_cookies_cb,
4174 /* If add_journal_cb failed, then filter_finish_transno
4175 * will commit the handle and we will do a sync
4176 * on commit. then we call callback directly to free
4179 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4181 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4184 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4186 CERROR("error on commit, err = %d\n", rc2);
4193 filter_parent_unlock(dparent);
4195 filter_fini_destroy(obd, &lockh);
4199 OBD_FREE(fcc, sizeof(*fcc));
4201 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4204 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4208 /* trigger quota release */
4209 qcids[USRQUOTA] = oa->o_uid;
4210 qcids[GRPQUOTA] = oa->o_gid;
4211 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4214 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4218 /* NB start and end are used for punch, but not truncate */
4219 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4220 struct obd_trans_info *oti,
4221 struct ptlrpc_request_set *rqset)
4226 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4227 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4228 oinfo->oi_policy.l_extent.end);
4232 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4233 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4234 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4236 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4237 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4238 rc = filter_setattr(exp, oinfo, oti);
4239 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4243 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4244 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4247 struct lvfs_run_ctxt saved;
4248 struct filter_obd *filter;
4249 struct dentry *dentry;
4253 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4254 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4255 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4259 filter = &exp->exp_obd->u.filter;
4261 /* An objid of zero is taken to mean "sync whole filesystem" */
4262 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4263 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4264 /* Flush any remaining cancel messages out to the target */
4265 filter_sync_llogs(exp->exp_obd, exp);
4269 dentry = filter_oa2dentry(exp->exp_obd, oa);
4271 RETURN(PTR_ERR(dentry));
4273 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4275 LOCK_INODE_MUTEX(dentry->d_inode);
4277 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4279 /* just any file to grab fsync method - "file" arg unused */
4280 struct file *file = filter->fo_rcvd_filp;
4282 if (file->f_op && file->f_op->fsync)
4283 rc = file->f_op->fsync(NULL, dentry, 1);
4285 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4289 UNLOCK_INODE_MUTEX(dentry->d_inode);
4291 oa->o_valid = OBD_MD_FLID;
4292 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4294 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4300 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4301 void *key, __u32 *vallen, void *val,
4302 struct lov_stripe_md *lsm)
4304 struct obd_device *obd;
4307 obd = class_exp2obd(exp);
4309 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4313 if (KEY_IS(KEY_BLOCKSIZE)) {
4314 __u32 *blocksize = val;
4316 if (*vallen < sizeof(*blocksize))
4318 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4320 *vallen = sizeof(*blocksize);
4324 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4325 __u32 *blocksize_bits = val;
4326 if (blocksize_bits) {
4327 if (*vallen < sizeof(*blocksize_bits))
4329 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4331 *vallen = sizeof(*blocksize_bits);
4335 if (KEY_IS(KEY_LAST_ID)) {
4336 obd_id *last_id = val;
4337 /* FIXME: object groups */
4339 if (*vallen < sizeof(*last_id))
4341 *last_id = filter_last_id(&obd->u.filter,
4342 exp->exp_filter_data.fed_group);
4344 *vallen = sizeof(*last_id);
4348 if (KEY_IS(KEY_FIEMAP)) {
4349 struct ll_fiemap_info_key *fm_key = key;
4350 struct dentry *dentry;
4351 struct ll_user_fiemap *fiemap = val;
4352 struct lvfs_run_ctxt saved;
4355 if (fiemap == NULL) {
4356 *vallen = fiemap_count_to_size(
4357 fm_key->fiemap.fm_extent_count);
4361 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4364 RETURN(PTR_ERR(dentry));
4366 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4367 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4368 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4370 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4376 CDEBUG(D_IOCTL, "invalid key\n");
4380 static inline int filter_setup_llog_group(struct obd_export *exp,
4381 struct obd_device *obd,
4384 struct obd_llog_group *olg;
4385 struct llog_ctxt *ctxt;
4388 olg = filter_find_create_olg(obd, group);
4390 RETURN(PTR_ERR(olg));
4392 llog_group_set_export(olg, exp);
4394 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4395 LASSERTF(ctxt != NULL, "ctxt is null\n");
4397 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4398 llog_ctxt_put(ctxt);
4402 static int filter_set_grant_shrink(struct obd_export *exp,
4403 struct ost_body *body)
4405 /* handle shrink grant */
4406 spin_lock(&exp->exp_obd->obd_osfs_lock);
4407 filter_grant_incoming(exp, &body->oa);
4408 spin_unlock(&exp->exp_obd->obd_osfs_lock);
4414 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4416 struct obd_device *obd;
4422 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4426 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4427 obd_export_nid2str(exp));
4428 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4430 /* setup llog imports */
4432 group = (int)(*(__u32 *)val);
4434 group = 0; /* default value */
4436 LASSERT_MDS_GROUP(group);
4437 rc = filter_setup_llog_group(exp, obd, group);
4441 if (group == FILTER_GROUP_MDS0) {
4442 /* setup llog group 1 for interop */
4443 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4446 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4451 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4452 void *key, __u32 vallen, void *val,
4453 struct ptlrpc_request_set *set)
4455 struct obd_device *obd;
4460 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4464 if (KEY_IS(KEY_CAPA_KEY)) {
4466 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4468 CERROR("filter update capability key failed: %d\n", rc);
4472 if (KEY_IS(KEY_REVIMP_UPD)) {
4473 filter_revimp_update(exp);
4474 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4478 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4479 filter_adapt_sptlrpc_conf(obd, 0);
4483 if (KEY_IS(KEY_MDS_CONN))
4484 RETURN(filter_set_mds_conn(exp, val));
4486 if (KEY_IS(KEY_GRANT_SHRINK))
4487 RETURN(filter_set_grant_shrink(exp, val));
4492 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4493 int len, void *karg, void *uarg)
4495 struct obd_device *obd = exp->exp_obd;
4496 struct obd_ioctl_data *data = karg;
4500 case OBD_IOC_ABORT_RECOVERY: {
4501 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4502 target_stop_recovery_thread(obd);
4506 case OBD_IOC_SYNC: {
4507 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4508 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4512 case OBD_IOC_SET_READONLY: {
4514 struct super_block *sb = obd->u.obt.obt_sb;
4515 struct inode *inode = sb->s_root->d_inode;
4516 BDEVNAME_DECLARE_STORAGE(tmp);
4517 CERROR("*** setting device %s read-only ***\n",
4518 ll_bdevname(sb, tmp));
4520 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4521 if (!IS_ERR(handle))
4522 rc = fsfilt_commit(obd, inode, handle, 1);
4524 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4525 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4527 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4531 case OBD_IOC_CATLOGLIST: {
4532 rc = llog_catalog_list(obd, 1, data);
4536 case OBD_IOC_LLOG_CANCEL:
4537 case OBD_IOC_LLOG_REMOVE:
4538 case OBD_IOC_LLOG_INFO:
4539 case OBD_IOC_LLOG_PRINT: {
4540 /* FIXME to be finished */
4541 RETURN(-EOPNOTSUPP);
4543 struct llog_ctxt *ctxt = NULL;
4545 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4546 rc = llog_ioctl(ctxt, cmd, data);
4547 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4560 static int filter_health_check(struct obd_device *obd)
4562 #ifdef USE_HEALTH_CHECK_WRITE
4563 struct filter_obd *filter = &obd->u.filter;
4568 * health_check to return 0 on healthy
4569 * and 1 on unhealthy.
4571 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4574 #ifdef USE_HEALTH_CHECK_WRITE
4575 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4576 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4581 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4584 return filter_fid2dentry(data, NULL, gr, id);
4587 static int filter_process_config(struct obd_device *obd, obd_count len,
4590 struct lustre_cfg *lcfg = buf;
4591 struct lprocfs_static_vars lvars;
4594 switch (lcfg->lcfg_command) {
4596 lprocfs_filter_init_vars(&lvars);
4598 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4608 static struct lvfs_callback_ops filter_lvfs_ops = {
4609 l_fid2dentry: filter_lvfs_fid2dentry,
4612 static struct obd_ops filter_obd_ops = {
4613 .o_owner = THIS_MODULE,
4614 .o_get_info = filter_get_info,
4615 .o_set_info_async = filter_set_info_async,
4616 .o_setup = filter_setup,
4617 .o_precleanup = filter_precleanup,
4618 .o_cleanup = filter_cleanup,
4619 .o_connect = filter_connect,
4620 .o_reconnect = filter_reconnect,
4621 .o_disconnect = filter_disconnect,
4622 .o_ping = filter_ping,
4623 .o_init_export = filter_init_export,
4624 .o_destroy_export = filter_destroy_export,
4625 .o_statfs = filter_statfs,
4626 .o_getattr = filter_getattr,
4627 .o_unpackmd = filter_unpackmd,
4628 .o_create = filter_create,
4629 .o_setattr = filter_setattr,
4630 .o_destroy = filter_destroy,
4631 .o_brw = filter_brw,
4632 .o_punch = filter_truncate,
4633 .o_sync = filter_sync,
4634 .o_preprw = filter_preprw,
4635 .o_commitrw = filter_commitrw,
4636 .o_llog_init = filter_llog_init,
4637 .o_llog_connect = filter_llog_connect,
4638 .o_llog_finish = filter_llog_finish,
4639 .o_iocontrol = filter_iocontrol,
4640 .o_health_check = filter_health_check,
4641 .o_process_config = filter_process_config,
4644 quota_interface_t *filter_quota_interface_ref;
4645 extern quota_interface_t filter_quota_interface;
4647 static int __init obdfilter_init(void)
4649 struct lprocfs_static_vars lvars;
4652 /** sanity check for group<->mdsno conversion */
4653 for (i = 0; i < 32; i++)
4654 LASSERT(objgrp_to_mdsno(mdt_to_obd_objgrp(i)) == i);
4656 lprocfs_filter_init_vars(&lvars);
4658 request_module("%s", "lquota");
4659 OBD_ALLOC(obdfilter_created_scratchpad,
4660 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4661 sizeof(*obdfilter_created_scratchpad));
4662 if (obdfilter_created_scratchpad == NULL)
4665 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4666 sizeof(struct filter_mod_data),
4669 GOTO(out, rc = -ENOMEM);
4671 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4672 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4674 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4675 LUSTRE_OST_NAME, NULL);
4679 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4680 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4681 ll_fmd_cachep = NULL;
4683 if (filter_quota_interface_ref)
4684 PORTAL_SYMBOL_PUT(filter_quota_interface);
4686 OBD_FREE(obdfilter_created_scratchpad,
4687 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4688 sizeof(*obdfilter_created_scratchpad));
4694 static void __exit obdfilter_exit(void)
4696 if (filter_quota_interface_ref)
4697 PORTAL_SYMBOL_PUT(filter_quota_interface);
4699 if (ll_fmd_cachep) {
4700 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4701 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4702 ll_fmd_cachep = NULL;
4705 class_unregister_type(LUSTRE_OST_NAME);
4706 OBD_FREE(obdfilter_created_scratchpad,
4707 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4708 sizeof(*obdfilter_created_scratchpad));
4711 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4712 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4713 MODULE_LICENSE("GPL");
4715 module_init(obdfilter_init);
4716 module_exit(obdfilter_exit);