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 cfs_spin_lock(&exp->exp_lock);
111 exp->exp_vbr_failed = 1;
112 cfs_spin_unlock(&exp->exp_lock);
115 oti->oti_pre_version = curr_version;
119 /* Assumes caller has already pushed us into the kernel context. */
120 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
121 struct obd_trans_info *oti, int rc, int force_sync)
123 struct 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 cfs_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 cfs_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 cfs_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 cfs_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 cfs_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 cfs_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)
270 GOTO(clean, rc = -ENOMEM);
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)
292 GOTO(clean, rc = -ENOMEM);
294 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
295 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
296 tmp->nid_ldlm_stats);
303 lprocfs_exp_cleanup(exp);
307 /* Add client data to the FILTER. We use a bitmap to locate a free space
308 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
309 * Otherwise, we have just read the data from the last_rcvd file and
310 * we know its offset. */
311 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
314 struct filter_obd *filter = &obd->u.filter;
315 struct filter_export_data *fed = &exp->exp_filter_data;
316 unsigned long *bitmap = filter->fo_last_rcvd_slots;
317 int new_client = (cl_idx == -1);
321 LASSERT(bitmap != NULL);
322 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
325 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
328 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
329 * there's no need for extra complication here
332 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
334 if (cl_idx >= LR_MAX_CLIENTS) {
335 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
339 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
340 cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
345 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
346 CERROR("FILTER client %d: bit already set in bitmap!\n",
352 fed->fed_lr_idx = cl_idx;
353 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
354 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
355 cfs_init_mutex(&fed->fed_lcd_lock);
356 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
358 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
359 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
362 struct lvfs_run_ctxt saved;
363 loff_t off = fed->fed_lr_off;
367 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
368 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
370 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
371 /* Transaction needed to fix bug 1403 */
372 handle = fsfilt_start(obd,
373 filter->fo_rcvd_filp->f_dentry->d_inode,
374 FSFILT_OP_SETATTR, NULL);
375 if (IS_ERR(handle)) {
376 rc = PTR_ERR(handle);
377 CERROR("unable to start transaction: rc %d\n", rc);
379 fed->fed_lcd->lcd_last_epoch =
380 filter->fo_fsd->lsd_start_epoch;
381 exp->exp_last_request_time = cfs_time_current_sec();
382 rc = fsfilt_add_journal_cb(obd, 0, handle,
383 target_client_add_cb,
384 class_export_cb_get(exp));
386 cfs_spin_lock(&exp->exp_lock);
387 exp->exp_need_sync = 1;
388 cfs_spin_unlock(&exp->exp_lock);
390 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
392 sizeof(*fed->fed_lcd),
393 &off, rc /* sync if no cb */);
395 filter->fo_rcvd_filp->f_dentry->d_inode,
398 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
401 CERROR("error writing %s client idx %u: rc %d\n",
402 LAST_RCVD, fed->fed_lr_idx, rc);
409 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
411 static int filter_client_free(struct obd_export *exp)
413 struct filter_export_data *fed = &exp->exp_filter_data;
414 struct filter_obd *filter = &exp->exp_obd->u.filter;
415 struct obd_device *obd = exp->exp_obd;
416 struct lvfs_run_ctxt saved;
417 struct lsd_client_data *lcd = fed->fed_lcd;
425 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
426 if (strcmp(lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
429 LASSERT(filter->fo_last_rcvd_slots != NULL);
431 off = fed->fed_lr_off;
433 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
434 fed->fed_lr_idx, fed->fed_lr_off, lcd->lcd_uuid);
436 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
437 * we leak a client slot that will be cleaned on the next recovery. */
439 CERROR("%s: client idx %d has med_off %lld\n",
440 obd->obd_name, fed->fed_lr_idx, off);
441 GOTO(free, rc = -EINVAL);
444 /* Clear the bit _after_ zeroing out the client so we don't
445 race with filter_client_add and zero out new clients.*/
446 if (!cfs_test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
447 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
452 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
453 /* Make sure the server's last_transno is up to date.
454 * This should be done before zeroing client slot so last_transno will
455 * be in server data or in client data in case of failure */
456 filter_update_server_data(obd, filter->fo_rcvd_filp, filter->fo_fsd);
458 cfs_mutex_down(&fed->fed_lcd_lock);
459 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
460 sizeof(zero_lcd), &off, 0);
462 cfs_mutex_up(&fed->fed_lcd_lock);
463 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
465 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
466 "zero out client %s at idx %u/%llu in %s, rc %d\n",
467 lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
470 if (!cfs_test_and_clear_bit(fed->fed_lr_idx,
471 filter->fo_last_rcvd_slots)) {
472 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
479 cfs_mutex_down(&fed->fed_lcd_lock);
481 cfs_mutex_up(&fed->fed_lcd_lock);
487 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
488 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
489 struct filter_mod_data *fmd)
491 LASSERT_SPIN_LOCKED(&fed->fed_lock);
492 if (--fmd->fmd_refcount == 0) {
493 /* XXX when we have persistent reservations and the handle
494 * is stored herein we need to drop it here. */
495 fed->fed_mod_count--;
496 cfs_list_del(&fmd->fmd_list);
497 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
501 /* drop fmd reference, free it if last ref */
502 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
504 struct filter_export_data *fed;
509 fed = &exp->exp_filter_data;
510 cfs_spin_lock(&fed->fed_lock);
511 filter_fmd_put_nolock(fed, fmd); /* caller reference */
512 cfs_spin_unlock(&fed->fed_lock);
515 /* expire entries from the end of the list if there are too many
516 * or they are too old */
517 static void filter_fmd_expire_nolock(struct filter_obd *filter,
518 struct filter_export_data *fed,
519 struct filter_mod_data *keep)
521 struct filter_mod_data *fmd, *tmp;
523 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
527 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
528 fed->fed_mod_count < filter->fo_fmd_max_num)
531 cfs_list_del_init(&fmd->fmd_list);
532 filter_fmd_put_nolock(fed, fmd); /* list reference */
536 void filter_fmd_expire(struct obd_export *exp)
538 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
539 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
540 &exp->exp_filter_data, NULL);
541 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
544 /* find specified objid, group in export fmd list.
545 * caller must hold fed_lock and take fmd reference itself */
546 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
547 struct filter_export_data *fed,
548 obd_id objid, obd_gr group)
550 struct filter_mod_data *found = NULL, *fmd;
552 LASSERT_SPIN_LOCKED(&fed->fed_lock);
554 cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
555 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
557 cfs_list_del(&fmd->fmd_list);
558 cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
559 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
564 filter_fmd_expire_nolock(filter, fed, found);
569 /* Find fmd based on objid and group, or return NULL if not found. */
570 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
571 obd_id objid, obd_gr group)
573 struct filter_mod_data *fmd;
575 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
576 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
577 &exp->exp_filter_data, objid, group);
579 fmd->fmd_refcount++; /* caller reference */
580 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
585 /* Find fmd based on objid and group, or create a new one if none is found.
586 * It is possible for this function to return NULL under memory pressure,
587 * or if objid = 0 is passed (which will only cause old entries to expire).
588 * Currently this is not fatal because any fmd state is transient and
589 * may also be freed when it gets sufficiently old. */
590 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
591 obd_id objid, obd_gr group)
593 struct filter_export_data *fed = &exp->exp_filter_data;
594 struct filter_mod_data *found = NULL, *fmd_new = NULL;
596 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
598 cfs_spin_lock(&fed->fed_lock);
599 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
602 cfs_list_add_tail(&fmd_new->fmd_list,
604 fmd_new->fmd_id = objid;
605 fmd_new->fmd_gr = group;
606 fmd_new->fmd_refcount++; /* list reference */
608 fed->fed_mod_count++;
610 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
614 found->fmd_refcount++; /* caller reference */
615 found->fmd_expire = jiffies +
616 exp->exp_obd->u.filter.fo_fmd_max_age;
619 cfs_spin_unlock(&fed->fed_lock);
625 /* drop fmd list reference so it will disappear when last reference is put.
626 * This isn't so critical because it would in fact only affect the one client
627 * that is doing the unlink and at worst we have an stale entry referencing
628 * an object that should never be used again. */
629 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
631 struct filter_mod_data *found = NULL;
633 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
634 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
636 cfs_list_del_init(&found->fmd_list);
637 filter_fmd_put_nolock(&exp->exp_filter_data, found);
639 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
642 #define filter_fmd_drop(exp, objid, group)
645 /* remove all entries from fmd list */
646 static void filter_fmd_cleanup(struct obd_export *exp)
648 struct filter_export_data *fed = &exp->exp_filter_data;
649 struct filter_mod_data *fmd = NULL, *tmp;
651 cfs_spin_lock(&fed->fed_lock);
652 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
653 cfs_list_del_init(&fmd->fmd_list);
654 filter_fmd_put_nolock(fed, fmd);
656 cfs_spin_unlock(&fed->fed_lock);
659 static int filter_init_export(struct obd_export *exp)
661 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
662 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
664 cfs_spin_lock(&exp->exp_lock);
665 exp->exp_connecting = 1;
666 cfs_spin_unlock(&exp->exp_lock);
668 return ldlm_init_export(exp);
671 static int filter_free_server_data(struct filter_obd *filter)
673 OBD_FREE_PTR(filter->fo_fsd);
674 filter->fo_fsd = NULL;
675 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
676 filter->fo_last_rcvd_slots = NULL;
680 /* assumes caller is already in kernel ctxt */
681 int filter_update_server_data(struct obd_device *obd, struct file *filp,
682 struct lr_server_data *fsd)
688 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
689 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
690 le64_to_cpu(fsd->lsd_last_transno));
691 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
692 le64_to_cpu(fsd->lsd_mount_count));
694 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, 0);
696 CERROR("error writing lr_server_data: rc = %d\n", rc);
701 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
704 struct filter_obd *filter = &obd->u.filter;
710 if (filter->fo_last_objid_files[group] == NULL) {
711 CERROR("Object group "LPU64" not fully setup; not updating "
712 "last_objid\n", group);
716 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
717 obd->obd_name, group, filter->fo_last_objids[group]);
719 tmp = cpu_to_le64(filter->fo_last_objids[group]);
720 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
721 &tmp, sizeof(tmp), &off, force_sync);
723 CERROR("error writing group "LPU64" last objid: rc = %d\n",
727 extern int ost_handle(struct ptlrpc_request *req);
728 /* assumes caller has already in kernel ctxt */
729 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
731 struct filter_obd *filter = &obd->u.filter;
732 struct lr_server_data *fsd;
733 struct lsd_client_data *lcd = NULL;
734 struct inode *inode = filp->f_dentry->d_inode;
735 unsigned long last_rcvd_size = i_size_read(inode);
742 /* ensure padding in the struct is the correct size */
743 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
744 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
745 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
746 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
748 OBD_ALLOC(fsd, sizeof(*fsd));
751 filter->fo_fsd = fsd;
753 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
754 if (filter->fo_last_rcvd_slots == NULL) {
755 OBD_FREE(fsd, sizeof(*fsd));
759 if (last_rcvd_size == 0) {
760 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
762 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
763 fsd->lsd_last_transno = 0;
764 mount_count = fsd->lsd_mount_count = 0;
765 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
766 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
767 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
768 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
769 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
770 /* OBD_COMPAT_OST is set in filter_connect_internal when the
771 * MDS first connects and assigns the OST index number. */
772 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
775 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
777 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
781 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
782 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
783 "using the wrong disk %s. Were the "
784 "/dev/ assignments rearranged?\n",
785 obd->obd_uuid.uuid, fsd->lsd_uuid);
786 GOTO(err_fsd, rc = -EINVAL);
788 mount_count = le64_to_cpu(fsd->lsd_mount_count);
789 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
791 /* Assume old last_rcvd format unless I_C_LR is set */
792 if (!(fsd->lsd_feature_incompat &
793 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
794 fsd->lsd_last_transno = fsd->lsd_compat14;
796 /* OBD_COMPAT_OST is set in filter_connect_internal when the
797 * MDS first connects and assigns the OST index number. */
798 fsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
802 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
803 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
804 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
805 ~FILTER_INCOMPAT_SUPP);
806 GOTO(err_fsd, rc = -EINVAL);
808 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
809 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
810 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
811 ~FILTER_ROCOMPAT_SUPP);
812 /* Do something like remount filesystem read-only */
813 GOTO(err_fsd, rc = -EINVAL);
816 start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
818 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
819 obd->obd_name, start_epoch);
820 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
821 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
822 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
823 obd->obd_name, mount_count + 1);
824 CDEBUG(D_INODE, "%s: server data size: %u\n",
825 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
826 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
827 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
828 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
829 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
830 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
831 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
832 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
833 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
834 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
835 le16_to_cpu(fsd->lsd_client_size));
837 if (!obd->obd_replayable) {
838 CWARN("%s: recovery support OFF\n", obd->obd_name);
842 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
843 off < last_rcvd_size; cl_idx++) {
845 struct obd_export *exp;
846 struct filter_export_data *fed;
851 GOTO(err_client, rc = -ENOMEM);
854 /* Don't assume off is incremented properly by
855 * fsfilt_read_record(), in case sizeof(*lcd)
856 * isn't the same as fsd->lsd_client_size. */
857 off = le32_to_cpu(fsd->lsd_client_start) +
858 cl_idx * le16_to_cpu(fsd->lsd_client_size);
859 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
861 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
862 LAST_RCVD, cl_idx, off, rc);
863 break; /* read error shouldn't cause startup to fail */
866 if (lcd->lcd_uuid[0] == '\0') {
867 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
872 check_lcd(obd->obd_name, cl_idx, lcd);
874 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
876 /* These exports are cleaned up by filter_disconnect(), so they
877 * need to be set up like real exports as filter_connect() does.
879 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
881 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
882 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
883 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
885 if (PTR_ERR(exp) == -EALREADY) {
886 /* export already exists, zero out this one */
887 CERROR("Zeroing out duplicate export due to "
889 lcd->lcd_uuid[0] = '\0';
891 GOTO(err_client, rc = PTR_ERR(exp));
894 fed = &exp->exp_filter_data;
896 fed->fed_group = 0; /* will be assigned at connect */
897 filter_export_stats_init(obd, exp, NULL);
898 rc = filter_client_add(obd, exp, cl_idx);
899 /* can't fail for existing client */
900 LASSERTF(rc == 0, "rc = %d\n", rc);
902 /* VBR: set export last committed */
903 exp->exp_last_committed = last_rcvd;
904 cfs_spin_lock(&exp->exp_lock);
905 exp->exp_connecting = 0;
906 exp->exp_in_recovery = 0;
907 cfs_spin_unlock(&exp->exp_lock);
908 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
909 obd->obd_max_recoverable_clients++;
910 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
912 class_export_put(exp);
915 /* Need to check last_rcvd even for duplicated exports. */
916 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
919 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
920 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
926 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
928 filter->fo_mount_count = mount_count + 1;
929 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
931 /* save it, so mount count and last_transno is current */
932 rc = filter_update_server_data(obd, filp, filter->fo_fsd);
934 GOTO(err_client, rc);
939 class_disconnect_exports(obd);
941 filter_free_server_data(filter);
945 static int filter_cleanup_groups(struct obd_device *obd)
947 struct filter_obd *filter = &obd->u.filter;
949 struct dentry *dentry;
953 if (filter->fo_dentry_O_groups != NULL) {
954 for (i = 0; i < filter->fo_group_count; i++) {
955 dentry = filter->fo_dentry_O_groups[i];
959 OBD_FREE(filter->fo_dentry_O_groups,
960 filter->fo_group_count *
961 sizeof(*filter->fo_dentry_O_groups));
962 filter->fo_dentry_O_groups = NULL;
964 if (filter->fo_last_objid_files != NULL) {
965 for (i = 0; i < filter->fo_group_count; i++) {
966 filp = filter->fo_last_objid_files[i];
970 OBD_FREE(filter->fo_last_objid_files,
971 filter->fo_group_count *
972 sizeof(*filter->fo_last_objid_files));
973 filter->fo_last_objid_files = NULL;
975 if (filter->fo_dentry_O_sub != NULL) {
976 for (i = 0; i < filter->fo_group_count; i++) {
977 for (j = 0; j < filter->fo_subdir_count; j++) {
978 dentry = filter->fo_dentry_O_sub[i].dentry[j];
983 OBD_FREE(filter->fo_dentry_O_sub,
984 filter->fo_group_count *
985 sizeof(*filter->fo_dentry_O_sub));
986 filter->fo_dentry_O_sub = NULL;
988 if (filter->fo_last_objids != NULL) {
989 OBD_FREE(filter->fo_last_objids,
990 filter->fo_group_count *
991 sizeof(*filter->fo_last_objids));
992 filter->fo_last_objids = NULL;
994 if (filter->fo_dentry_O != NULL) {
995 f_dput(filter->fo_dentry_O);
996 filter->fo_dentry_O = NULL;
1001 static int filter_update_last_group(struct obd_device *obd, int group)
1003 struct filter_obd *filter = &obd->u.filter;
1004 struct file *filp = NULL;
1005 int last_group = 0, rc;
1009 if (group <= filter->fo_committed_group)
1012 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1016 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1020 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1022 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1026 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1027 obd->obd_name, last_group, group);
1031 /* must be sync: bXXXX */
1032 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1034 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1038 filter->fo_committed_group = group;
1041 filp_close(filp, 0);
1045 static int filter_read_group_internal(struct obd_device *obd, int group,
1048 struct filter_obd *filter = &obd->u.filter;
1049 __u64 *new_objids = NULL;
1050 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1051 struct dentry **new_groups = NULL;
1052 struct file **new_files = NULL;
1053 struct dentry *dentry;
1055 int old_count = filter->fo_group_count, rc, stage = 0, i;
1059 int len = group + 1;
1061 snprintf(name, 24, "%d", group);
1065 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1067 if (IS_ERR(dentry)) {
1068 CERROR("Cannot lookup expected object group %d: %ld\n",
1069 group, PTR_ERR(dentry));
1070 RETURN(PTR_ERR(dentry));
1073 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1075 if (IS_ERR(dentry)) {
1076 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1078 RETURN(PTR_ERR(dentry));
1083 snprintf(name, 24, "O/%d/LAST_ID", group);
1085 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1087 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1088 GOTO(cleanup, rc = PTR_ERR(filp));
1092 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1094 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1098 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1099 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1100 if (tmp_subdirs == NULL)
1101 GOTO(cleanup, rc = -ENOMEM);
1104 for (i = 0; i < filter->fo_subdir_count; i++) {
1106 snprintf(dir, sizeof(dir), "d%u", i);
1108 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1111 if (IS_ERR(tmp_subdirs->dentry[i])) {
1112 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1113 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1118 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1119 tmp_subdirs->dentry[i]);
1123 /* 'group' is an index; we need an array of length 'group + 1' */
1124 if (group + 1 > old_count) {
1125 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1126 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1127 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1128 OBD_ALLOC(new_files, len * sizeof(*new_files));
1130 if (new_objids == NULL || new_subdirs == NULL ||
1131 new_groups == NULL || new_files == NULL)
1132 GOTO(cleanup, rc = -ENOMEM);
1135 memcpy(new_objids, filter->fo_last_objids,
1136 old_count * sizeof(*new_objids));
1137 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1138 old_count * sizeof(*new_subdirs));
1139 memcpy(new_groups, filter->fo_dentry_O_groups,
1140 old_count * sizeof(*new_groups));
1141 memcpy(new_files, filter->fo_last_objid_files,
1142 old_count * sizeof(*new_files));
1144 OBD_FREE(filter->fo_last_objids,
1145 old_count * sizeof(*new_objids));
1146 OBD_FREE(filter->fo_dentry_O_sub,
1147 old_count * sizeof(*new_subdirs));
1148 OBD_FREE(filter->fo_dentry_O_groups,
1149 old_count * sizeof(*new_groups));
1150 OBD_FREE(filter->fo_last_objid_files,
1151 old_count * sizeof(*new_files));
1153 filter->fo_last_objids = new_objids;
1154 filter->fo_dentry_O_sub = new_subdirs;
1155 filter->fo_dentry_O_groups = new_groups;
1156 filter->fo_last_objid_files = new_files;
1157 filter->fo_group_count = len;
1160 filter->fo_dentry_O_groups[group] = dentry;
1161 filter->fo_last_objid_files[group] = filp;
1162 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1163 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1164 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1167 filter_update_last_group(obd, group);
1169 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1170 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1171 rc = filter_update_last_objid(obd, group, 1);
1175 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1176 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1177 obd->obd_name, group, last_objid);
1182 if (new_objids != NULL)
1183 OBD_FREE(new_objids, len * sizeof(*new_objids));
1184 if (new_subdirs != NULL)
1185 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1186 if (new_groups != NULL)
1187 OBD_FREE(new_groups, len * sizeof(*new_groups));
1188 if (new_files != NULL)
1189 OBD_FREE(new_files, len * sizeof(*new_files));
1191 if (filter->fo_subdir_count && filter_group_is_mds(group)) {
1192 for (i = 0; i < filter->fo_subdir_count; i++) {
1193 if (tmp_subdirs->dentry[i] != NULL)
1194 dput(tmp_subdirs->dentry[i]);
1196 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1199 filp_close(filp, 0);
1206 static int filter_read_groups(struct obd_device *obd, int last_group,
1209 struct filter_obd *filter = &obd->u.filter;
1210 int old_count, group, rc = 0;
1212 cfs_down(&filter->fo_init_lock);
1213 old_count = filter->fo_group_count;
1214 for (group = old_count; group <= last_group; group++) {
1215 rc = filter_read_group_internal(obd, group, create);
1219 cfs_up(&filter->fo_init_lock);
1223 /* FIXME: object groups */
1224 static int filter_prep_groups(struct obd_device *obd)
1226 struct filter_obd *filter = &obd->u.filter;
1227 struct dentry *O_dentry;
1229 int last_group, rc = 0, cleanup_phase = 0;
1233 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1235 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1236 if (IS_ERR(O_dentry)) {
1237 rc = PTR_ERR(O_dentry);
1238 CERROR("cannot open/create O: rc = %d\n", rc);
1241 filter->fo_dentry_O = O_dentry;
1242 cleanup_phase = 1; /* O_dentry */
1244 /* we have to initialize all groups before first connections from
1245 * clients because they may send create/destroy for any group -bzzz */
1246 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1248 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1249 GOTO(cleanup, rc = PTR_ERR(filp));
1251 cleanup_phase = 2; /* filp */
1253 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1255 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1260 last_group = FILTER_GROUP_MDS0;
1262 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1263 FILTER_GROUP_MDS0, last_group);
1264 filter->fo_committed_group = last_group;
1265 rc = filter_read_groups(obd, last_group, 1);
1269 filp_close(filp, 0);
1273 switch (cleanup_phase) {
1275 filp_close(filp, 0);
1277 filter_cleanup_groups(obd);
1278 f_dput(filter->fo_dentry_O);
1279 filter->fo_dentry_O = NULL;
1287 /* setup the object store with correct subdirectories */
1288 static int filter_prep(struct obd_device *obd)
1290 struct lvfs_run_ctxt saved;
1291 struct filter_obd *filter = &obd->u.filter;
1293 struct inode *inode;
1297 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1298 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1299 if (!file || IS_ERR(file)) {
1301 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1305 filter->fo_rcvd_filp = file;
1306 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1307 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1308 file->f_dentry->d_inode->i_mode);
1309 GOTO(err_filp, rc = -ENOENT);
1312 inode = file->f_dentry->d_parent->d_inode;
1313 /* We use i_op->unlink directly in filter_vfs_unlink() */
1314 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1315 CERROR("%s: filesystem does not support create/unlink ops\n",
1317 GOTO(err_filp, rc = -EOPNOTSUPP);
1320 /** lu_target has very limited use in filter now */
1321 lut_init(NULL, &filter->fo_lut, obd, NULL);
1323 rc = filter_init_server_data(obd, file);
1325 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1329 target_recovery_init(&filter->fo_lut, ost_handle);
1331 /* open and create health check io file*/
1332 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1335 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1337 GOTO(err_server_data, rc);
1339 filter->fo_obt.obt_health_check_filp = file;
1340 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1341 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1342 file->f_dentry->d_inode->i_mode);
1343 GOTO(err_health_check, rc = -ENOENT);
1345 rc = lvfs_check_io_health(obd, file);
1347 GOTO(err_health_check, rc);
1349 rc = filter_prep_groups(obd);
1351 GOTO(err_health_check, rc);
1353 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1358 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1359 CERROR("can't close %s after error\n", HEALTH_CHECK);
1360 filter->fo_obt.obt_health_check_filp = NULL;
1362 target_recovery_fini(obd);
1363 filter_free_server_data(filter);
1365 if (filp_close(filter->fo_rcvd_filp, 0))
1366 CERROR("can't close %s after error\n", LAST_RCVD);
1367 filter->fo_rcvd_filp = NULL;
1371 /* cleanup the filter: write last used object id to status file */
1372 static void filter_post(struct obd_device *obd)
1374 struct lvfs_run_ctxt saved;
1375 struct filter_obd *filter = &obd->u.filter;
1378 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1379 * best to start a transaction with h_sync, because we removed this
1382 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1383 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1386 CERROR("error writing server data: rc = %d\n", rc);
1388 for (i = 0; i < filter->fo_group_count; i++) {
1389 rc = filter_update_last_objid(obd, i,
1390 (i == filter->fo_group_count - 1));
1392 CERROR("error writing group %d lastobjid: rc = %d\n",
1396 rc = filp_close(filter->fo_rcvd_filp, 0);
1397 filter->fo_rcvd_filp = NULL;
1399 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1401 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1402 filter->fo_obt.obt_health_check_filp = NULL;
1404 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1406 filter_cleanup_groups(obd);
1407 filter_free_server_data(filter);
1408 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1410 filter_free_capa_keys(filter);
1411 cleanup_capa_hash(filter->fo_capa_hash);
1414 static void filter_set_last_id(struct filter_obd *filter,
1415 obd_id id, obd_gr group)
1417 LASSERT(filter->fo_fsd != NULL);
1418 LASSERT(group <= filter->fo_group_count);
1420 cfs_spin_lock(&filter->fo_objidlock);
1421 filter->fo_last_objids[group] = id;
1422 cfs_spin_unlock(&filter->fo_objidlock);
1425 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1428 LASSERT(filter->fo_fsd != NULL);
1429 LASSERT(group <= filter->fo_group_count);
1430 LASSERT(filter->fo_last_objids != NULL);
1432 /* FIXME: object groups */
1433 cfs_spin_lock(&filter->fo_objidlock);
1434 id = filter->fo_last_objids[group];
1435 cfs_spin_unlock(&filter->fo_objidlock);
1439 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1441 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1445 /* We never dget the object parent, so DON'T dput it either */
1446 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1448 struct filter_obd *filter = &obd->u.filter;
1449 struct filter_subdirs *subdirs;
1450 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1452 if (!filter_group_is_mds(group) || filter->fo_subdir_count == 0)
1453 return filter->fo_dentry_O_groups[group];
1455 subdirs = &filter->fo_dentry_O_sub[group];
1456 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1459 /* We never dget the object parent, so DON'T dput it either */
1460 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1463 unsigned long now = jiffies;
1464 struct dentry *dparent = filter_parent(obd, group, objid);
1467 if (IS_ERR(dparent))
1469 if (dparent == NULL)
1470 return ERR_PTR(-ENOENT);
1472 rc = filter_lock_dentry(obd, dparent);
1473 fsfilt_check_slow(obd, now, "parent lock");
1474 return rc ? ERR_PTR(rc) : dparent;
1477 /* We never dget the object parent, so DON'T dput it either */
1478 static void filter_parent_unlock(struct dentry *dparent)
1480 UNLOCK_INODE_MUTEX(dparent->d_inode);
1483 /* How to get files, dentries, inodes from object id's.
1485 * If dir_dentry is passed, the caller has already locked the parent
1486 * appropriately for this operation (normally a write lock). If
1487 * dir_dentry is NULL, we do a read lock while we do the lookup to
1488 * avoid races with create/destroy and such changing the directory
1489 * internal to the filesystem code. */
1490 struct dentry *filter_fid2dentry(struct obd_device *obd,
1491 struct dentry *dir_dentry,
1492 obd_gr group, obd_id id)
1494 struct dentry *dparent = dir_dentry;
1495 struct dentry *dchild;
1500 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1501 obd->u.filter.fo_destroys_in_progress == 0) {
1502 /* don't fail lookups for orphan recovery, it causes
1503 * later LBUGs when objects still exist during precreate */
1504 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1505 RETURN(ERR_PTR(-ENOENT));
1508 CERROR("fatal: invalid object id 0\n");
1509 RETURN(ERR_PTR(-ESTALE));
1512 len = sprintf(name, LPU64, id);
1513 if (dir_dentry == NULL) {
1514 dparent = filter_parent_lock(obd, group, id);
1515 if (IS_ERR(dparent)) {
1516 CERROR("%s: error getting object "LPU64":"LPU64
1517 " parent: rc %ld\n", obd->obd_name,
1518 id, group, PTR_ERR(dparent));
1522 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1523 dparent->d_name.len, dparent->d_name.name, name);
1524 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1525 if (dir_dentry == NULL)
1526 filter_parent_unlock(dparent);
1527 if (IS_ERR(dchild)) {
1528 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1533 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1534 CERROR("%s: got bad object "LPU64" inode %lu\n",
1535 obd->obd_name, id, dchild->d_inode->i_ino);
1537 RETURN(ERR_PTR(-ENOENT));
1540 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1541 name, dchild, atomic_read(&dchild->d_count));
1543 LASSERT(atomic_read(&dchild->d_count) > 0);
1548 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1549 obd_id group, struct lustre_handle *lockh)
1551 int flags = LDLM_AST_DISCARD_DATA, rc;
1552 struct ldlm_res_id res_id;
1553 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1556 osc_build_res_name(objid, group, &res_id);
1557 /* Tell the clients that the object is gone now and that they should
1558 * throw away any cached pages. */
1559 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1560 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1561 ldlm_completion_ast, NULL, NULL, 0, NULL,
1568 static void filter_fini_destroy(struct obd_device *obd,
1569 struct lustre_handle *lockh)
1571 if (lustre_handle_is_used(lockh))
1572 ldlm_lock_decref(lockh, LCK_PW);
1575 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1576 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1577 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1578 * reverse. Caller must take i_sem before starting the transaction and we
1579 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1580 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1581 struct vfsmount *mnt)
1586 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1587 * itself we already hold dir->i_mutex for child create/unlink ops */
1588 LASSERT(dentry->d_inode != NULL);
1589 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1590 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1594 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1595 GOTO(out, rc = -ENOENT);
1597 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1602 GOTO(out, rc = -EPERM);
1604 /* check_sticky() */
1605 if ((dentry->d_inode->i_uid != current->fsuid &&
1606 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1607 IS_IMMUTABLE(dentry->d_inode))
1608 GOTO(out, rc = -EPERM);
1610 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1611 * that was done because of journal_start (which is already done
1612 * here) or some other ordering issue. */
1615 rc = ll_security_inode_unlink(dir, dentry, mnt);
1619 rc = dir->i_op->unlink(dir, dentry);
1621 /* need to drop i_mutex before we lose inode reference */
1622 UNLOCK_INODE_MUTEX(dentry->d_inode);
1629 /* Caller must hold LCK_PW on parent and push us into kernel context.
1630 * Caller must hold child i_mutex, we drop it always.
1631 * Caller is also required to ensure that dchild->d_inode exists. */
1632 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1633 obd_gr group, struct dentry *dparent,
1634 struct dentry *dchild)
1636 struct inode *inode = dchild->d_inode;
1637 struct filter_obd *filter = &obd->u.filter;
1640 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1641 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1642 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1643 (unsigned long)inode->i_nlink,
1644 atomic_read(&inode->i_count));
1647 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1649 CERROR("error unlinking objid %.*s: rc %d\n",
1650 dchild->d_name.len, dchild->d_name.name, rc);
1654 struct filter_intent_args {
1655 struct ldlm_lock **victim;
1660 static enum interval_iter filter_intent_cb(struct interval_node *n,
1663 struct ldlm_interval *node = (struct ldlm_interval *)n;
1664 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1665 __u64 size = arg->size;
1666 struct ldlm_lock **v = arg->victim;
1667 struct ldlm_lock *lck;
1669 /* If the interval is lower than the current file size,
1671 if (interval_high(n) <= size)
1672 return INTERVAL_ITER_STOP;
1674 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1675 /* Don't send glimpse ASTs to liblustre clients.
1676 * They aren't listening for them, and they do
1677 * entirely synchronous I/O anyways. */
1678 if (lck->l_export == NULL ||
1679 lck->l_export->exp_libclient == 1)
1682 if (*arg->liblustre)
1683 *arg->liblustre = 0;
1686 *v = LDLM_LOCK_GET(lck);
1687 } else if ((*v)->l_policy_data.l_extent.start <
1688 lck->l_policy_data.l_extent.start) {
1689 LDLM_LOCK_RELEASE(*v);
1690 *v = LDLM_LOCK_GET(lck);
1693 /* the same policy group - every lock has the
1694 * same extent, so needn't do it any more */
1698 return INTERVAL_ITER_CONT;
1701 static int filter_intent_policy(struct ldlm_namespace *ns,
1702 struct ldlm_lock **lockp, void *req_cookie,
1703 ldlm_mode_t mode, int flags, void *data)
1705 CFS_LIST_HEAD(rpc_list);
1706 struct ptlrpc_request *req = req_cookie;
1707 struct ldlm_lock *lock = *lockp, *l = NULL;
1708 struct ldlm_resource *res = lock->l_resource;
1709 ldlm_processing_policy policy;
1710 struct ost_lvb *res_lvb, *reply_lvb;
1711 struct ldlm_reply *rep;
1713 int idx, rc, tmpflags = 0, only_liblustre = 1;
1714 struct ldlm_interval_tree *tree;
1715 struct filter_intent_args arg;
1716 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1717 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1718 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1721 policy = ldlm_get_processing_policy(res);
1722 LASSERT(policy != NULL);
1723 LASSERT(req != NULL);
1725 rc = lustre_pack_reply(req, 3, repsize, NULL);
1727 RETURN(req->rq_status = rc);
1729 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1730 LASSERT(rep != NULL);
1732 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1733 sizeof(*reply_lvb));
1734 LASSERT(reply_lvb != NULL);
1736 //fixup_handle_for_resent_req(req, lock, &lockh);
1738 /* Call the extent policy function to see if our request can be
1739 * granted, or is blocked.
1740 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1741 * lock, and should not be granted if the lock will be blocked.
1744 LASSERT(ns == res->lr_namespace);
1746 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1747 check_res_locked(res);
1749 /* FIXME: we should change the policy function slightly, to not make
1750 * this list at all, since we just turn around and free it */
1751 while (!cfs_list_empty(&rpc_list)) {
1752 struct ldlm_lock *wlock =
1753 cfs_list_entry(rpc_list.next, struct ldlm_lock,
1755 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1756 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1757 lock->l_flags &= ~LDLM_FL_CP_REQD;
1758 cfs_list_del_init(&wlock->l_cp_ast);
1759 LDLM_LOCK_RELEASE(wlock);
1762 /* The lock met with no resistance; we're finished. */
1763 if (rc == LDLM_ITER_CONTINUE) {
1764 /* do not grant locks to the liblustre clients: they cannot
1765 * handle ASTs robustly. We need to do this while still
1766 * holding ns_lock to avoid the lock remaining on the res_link
1767 * list (and potentially being added to l_pending_list by an
1768 * AST) when we are going to drop this lock ASAP. */
1769 if (lock->l_export->exp_libclient ||
1770 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1771 ldlm_resource_unlink_lock(lock);
1772 err = ELDLM_LOCK_ABORTED;
1774 err = ELDLM_LOCK_REPLACED;
1780 /* Do not grant any lock, but instead send GL callbacks. The extent
1781 * policy nicely created a list of all PW locks for us. We will choose
1782 * the highest of those which are larger than the size in the LVB, if
1783 * any, and perform a glimpse callback. */
1784 res_lvb = res->lr_lvb_data;
1785 LASSERT(res_lvb != NULL);
1786 *reply_lvb = *res_lvb;
1789 * ->ns_lock guarantees that no new locks are granted, and,
1790 * therefore, that res->lr_lvb_data cannot increase beyond the
1791 * end of already granted lock. As a result, it is safe to
1792 * check against "stale" reply_lvb->lvb_size value without
1795 arg.size = reply_lvb->lvb_size;
1797 arg.liblustre = &only_liblustre;
1798 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1799 tree = &res->lr_itree[idx];
1800 if (tree->lit_mode == LCK_PR)
1803 interval_iterate_reverse(tree->lit_root,
1804 filter_intent_cb, &arg);
1808 /* There were no PW locks beyond the size in the LVB; finished. */
1810 if (only_liblustre) {
1811 /* If we discovered a liblustre client with a PW lock,
1812 * however, the LVB may be out of date! The LVB is
1813 * updated only on glimpse (which we don't do for
1814 * liblustre clients) and cancel (which the client
1815 * obviously has not yet done). So if it has written
1816 * data but kept the lock, the LVB is stale and needs
1817 * to be updated from disk.
1819 * Of course, this will all disappear when we switch to
1820 * taking liblustre locks on the OST. */
1821 ldlm_res_lvbo_update(res, NULL, 1);
1823 RETURN(ELDLM_LOCK_ABORTED);
1827 * This check is for lock taken in filter_prepare_destroy() that does
1828 * not have l_glimpse_ast set. So the logic is: if there is a lock
1829 * with no l_glimpse_ast set, this object is being destroyed already.
1831 * Hence, if you are grabbing DLM locks on the server, always set
1832 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1834 if (l->l_glimpse_ast == NULL) {
1835 /* We are racing with unlink(); just return -ENOENT */
1836 rep->lock_policy_res1 = -ENOENT;
1840 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1841 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1842 /* Update the LVB from disk if the AST failed (this is a legal race) */
1844 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1845 * sending ast is not handled. This can result in lost client writes.
1848 ldlm_res_lvbo_update(res, NULL, 1);
1851 *reply_lvb = *res_lvb;
1855 LDLM_LOCK_RELEASE(l);
1857 RETURN(ELDLM_LOCK_ABORTED);
1861 * per-obd_device iobuf pool.
1863 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1864 * allocations in write-path has to be minimized (see bug 5137).
1866 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1867 * OST threads (see ost_thread_{init,done}()).
1869 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1870 * at the OST layer there are only (potentially) multiple obd_device of type
1871 * unknown at the time of OST thread creation.
1873 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1874 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1875 * it's very own iobuf.
1879 * filter_kiobuf_pool_init()
1881 * filter_kiobuf_pool_done()
1883 * filter_iobuf_get()
1885 * operate on this array. They are "generic" in a sense that they don't depend
1886 * on actual type of iobuf's (the latter depending on Linux kernel version).
1890 * destroy pool created by filter_iobuf_pool_init
1892 static void filter_iobuf_pool_done(struct filter_obd *filter)
1894 struct filter_iobuf **pool;
1899 pool = filter->fo_iobuf_pool;
1901 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1902 if (pool[i] != NULL)
1903 filter_free_iobuf(pool[i]);
1905 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1906 filter->fo_iobuf_pool = NULL;
1911 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1913 struct filter_obd *filter = &obd->u.filter;
1914 struct sptlrpc_rule_set tmp_rset;
1917 sptlrpc_rule_set_init(&tmp_rset);
1918 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1920 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1925 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1927 cfs_write_lock(&filter->fo_sptlrpc_lock);
1928 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1929 filter->fo_sptlrpc_rset = tmp_rset;
1930 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1936 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1938 static int filter_iobuf_pool_init(struct filter_obd *filter)
1945 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1947 if (filter->fo_iobuf_pool == NULL)
1950 filter->fo_iobuf_count = OSS_THREADS_MAX;
1955 /* Return iobuf allocated for @thread_id. We don't know in advance how
1956 * many threads there will be so we allocate a large empty array and only
1957 * fill in those slots that are actually in use.
1958 * If we haven't allocated a pool entry for this thread before, do so now. */
1959 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1961 int thread_id = (oti && oti->oti_thread) ?
1962 oti->oti_thread->t_id : -1;
1963 struct filter_iobuf *pool = NULL;
1964 struct filter_iobuf **pool_place = NULL;
1966 if (thread_id >= 0) {
1967 LASSERT(thread_id < filter->fo_iobuf_count);
1968 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1971 if (unlikely(pool == NULL)) {
1972 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1973 PTLRPC_MAX_BRW_PAGES);
1974 if (pool_place != NULL)
1981 /* mount the file system (secretly). lustre_cfg parameters are:
1984 * 3 = flags: failover=f, failout=n
1987 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1990 struct filter_obd *filter = &obd->u.filter;
1991 struct vfsmount *mnt;
1992 struct lustre_mount_info *lmi;
1993 struct obd_uuid uuid;
2001 if (lcfg->lcfg_bufcount < 3 ||
2002 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
2003 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
2006 lmi = server_get_mount(obd->obd_name);
2008 /* We already mounted in lustre_fill_super.
2009 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2010 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2012 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2014 /* old path - used by lctl */
2015 CERROR("Using old MDS mount method\n");
2016 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2017 MS_NOATIME|MS_NODIRATIME,
2018 lustre_cfg_string(lcfg, 1), option);
2021 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2022 lustre_cfg_string(lcfg, 1), rc);
2026 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2028 if (IS_ERR(obd->obd_fsops))
2029 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2031 rc = filter_iobuf_pool_init(filter);
2035 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2036 CERROR("%s: Underlying device is marked as read-only. "
2037 "Setup failed\n", obd->obd_name);
2038 GOTO(err_ops, rc = -EROFS);
2041 /* failover is the default */
2042 obd->obd_replayable = 1;
2044 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2045 str = lustre_cfg_string(lcfg, 3);
2046 if (strchr(str, 'n')) {
2047 CWARN("%s: recovery disabled\n", obd->obd_name);
2048 obd->obd_replayable = 0;
2052 filter->fo_vfsmnt = mnt;
2053 obd->u.obt.obt_sb = mnt->mnt_sb;
2054 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2055 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2057 fsfilt_setup(obd, obd->u.obt.obt_sb);
2059 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2060 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2061 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2062 obd->obd_lvfs_ctxt.fs = get_ds();
2063 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2065 cfs_init_mutex(&filter->fo_init_lock);
2066 filter->fo_committed_group = 0;
2067 filter->fo_destroys_in_progress = 0;
2068 for (i = 0; i < 32; i++)
2069 cfs_sema_init(&filter->fo_create_locks[i], 1);
2071 cfs_spin_lock_init(&filter->fo_translock);
2072 cfs_spin_lock_init(&filter->fo_objidlock);
2073 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2074 cfs_sema_init(&filter->fo_alloc_lock, 1);
2075 init_brw_stats(&filter->fo_filter_stats);
2076 filter->fo_read_cache = 1; /* enable read-only cache by default */
2077 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2078 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2079 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2080 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2082 rc = filter_prep(obd);
2086 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2087 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2089 filter->fo_fl_oss_capa = 1;
2091 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2092 filter->fo_capa_hash = init_capa_hash();
2093 if (filter->fo_capa_hash == NULL)
2094 GOTO(err_post, rc = -ENOMEM);
2096 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2097 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2098 LDLM_NAMESPACE_GREEDY);
2099 if (obd->obd_namespace == NULL)
2100 GOTO(err_post, rc = -ENOMEM);
2101 obd->obd_namespace->ns_lvbp = obd;
2102 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2103 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2105 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2106 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2108 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2110 CERROR("failed to setup llogging subsystems\n");
2114 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2115 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2116 /* do this after llog being initialized */
2117 filter_adapt_sptlrpc_conf(obd, 1);
2119 rc = lquota_setup(filter_quota_interface_ref, obd);
2123 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2124 if (q->max_sectors < q->max_hw_sectors &&
2125 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2126 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2127 "for larger I/O requests: max_sectors = %u "
2128 "could be up to max_hw_sectors=%u\n",
2129 obd->obd_name, mnt->mnt_sb->s_id,
2130 q->max_sectors, q->max_hw_sectors);
2132 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2133 if (uuid_ptr != NULL) {
2134 class_uuid_unparse(uuid_ptr, &uuid);
2140 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2142 if (obd->obd_recovering) {
2143 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2144 "recovery for at least %d:%.02d, or until %d "
2145 "client%s reconnect%s.\n",
2146 obd->obd_name, lustre_cfg_string(lcfg, 1),
2147 label ?: "", label ? "/" : "", str,
2148 obd->obd_recovery_timeout / 60,
2149 obd->obd_recovery_timeout % 60,
2150 obd->obd_max_recoverable_clients,
2151 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2152 (obd->obd_max_recoverable_clients == 1) ? "s":"");
2154 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2155 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2156 label ?: "", label ? "/" : "", str,
2157 obd->obd_replayable ? "enabled" : "disabled");
2165 fsfilt_put_ops(obd->obd_fsops);
2166 filter_iobuf_pool_done(filter);
2168 server_put_mount(obd->obd_name, mnt);
2169 obd->u.obt.obt_sb = 0;
2173 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2175 struct lprocfs_static_vars lvars;
2180 CLASSERT(offsetof(struct obd_device, u.obt) ==
2181 offsetof(struct obd_device, u.filter.fo_obt));
2183 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2186 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2187 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2190 addr = (unsigned long)cfs_page_address(page);
2191 clear_page((void *)addr);
2193 /* lprocfs must be setup before the filter so state can be safely added
2194 * to /proc incrementally as the filter is setup */
2195 lprocfs_filter_init_vars(&lvars);
2196 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2197 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2198 /* Init obdfilter private stats here */
2199 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2200 LPROCFS_CNTR_AVGMINMAX,
2201 "read_bytes", "bytes");
2202 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2203 LPROCFS_CNTR_AVGMINMAX,
2204 "write_bytes", "bytes");
2205 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2206 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2207 "get_page", "usec");
2208 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2209 LPROCFS_CNTR_AVGMINMAX,
2210 "get_page_failures", "num");
2211 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2212 LPROCFS_CNTR_AVGMINMAX,
2213 "cache_access", "pages");
2214 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2215 LPROCFS_CNTR_AVGMINMAX,
2216 "cache_hit", "pages");
2217 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2218 LPROCFS_CNTR_AVGMINMAX,
2219 "cache_miss", "pages");
2221 lproc_filter_attach_seqstat(obd);
2222 obd->obd_proc_exports_entry = lprocfs_register("exports",
2223 obd->obd_proc_entry,
2225 if (IS_ERR(obd->obd_proc_exports_entry)) {
2226 rc = PTR_ERR(obd->obd_proc_exports_entry);
2227 CERROR("error %d setting up lprocfs for %s\n",
2229 obd->obd_proc_exports_entry = NULL;
2232 if (obd->obd_proc_exports_entry)
2233 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2234 lprocfs_nid_stats_clear_read,
2235 lprocfs_nid_stats_clear_write, obd, NULL);
2237 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2238 LUSTRE_CFG_BUFLEN(lcfg, 4));
2239 rc = filter_common_setup(obd, lcfg, (void *)addr);
2240 OBD_PAGE_FREE(page);
2243 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2244 lprocfs_free_per_client_stats(obd);
2245 lprocfs_free_obd_stats(obd);
2246 lprocfs_obd_cleanup(obd);
2252 static struct llog_operations filter_mds_ost_repl_logops;
2254 static struct llog_operations filter_size_orig_logops = {
2255 .lop_setup = llog_obd_origin_setup,
2256 .lop_cleanup = llog_obd_origin_cleanup,
2257 .lop_add = llog_obd_origin_add
2260 static int filter_olg_fini(struct obd_llog_group *olg)
2262 struct llog_ctxt *ctxt;
2263 int rc = 0, rc2 = 0;
2266 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2268 rc = llog_cleanup(ctxt);
2270 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2272 rc2 = llog_cleanup(ctxt);
2277 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2279 rc2 = llog_cleanup(ctxt);
2288 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2289 struct obd_device *tgt)
2294 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2295 &filter_mds_ost_repl_logops);
2299 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2300 &filter_size_orig_logops);
2306 filter_olg_fini(olg);
2311 * Init the default olg, which is embeded in the obd_device, for filter.
2314 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2315 struct obd_device *tgt)
2317 struct filter_obd *filter = &obd->u.filter;
2318 struct llog_ctxt *ctxt;
2322 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2323 if (!filter->fo_lcm)
2326 filter_mds_ost_repl_logops = llog_client_ops;
2327 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2328 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2329 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2331 rc = filter_olg_init(obd, olg, tgt);
2333 GOTO(cleanup_lcm, rc);
2335 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2338 GOTO(cleanup_olg, rc);
2340 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2342 CERROR("Can't get ctxt for %p:%x\n", olg,
2343 LLOG_MDS_OST_REPL_CTXT);
2344 GOTO(cleanup_olg, rc = -ENODEV);
2346 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2347 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2348 llog_ctxt_put(ctxt);
2352 filter_olg_fini(olg);
2354 llog_recov_thread_fini(filter->fo_lcm, 1);
2355 filter->fo_lcm = NULL;
2360 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2361 struct obd_device *tgt, int *index)
2363 struct filter_obd *filter = &obd->u.filter;
2364 struct llog_ctxt *ctxt;
2368 LASSERT(olg != NULL);
2369 if (olg == &obd->obd_olg)
2370 return filter_default_olg_init(obd, olg, tgt);
2372 LASSERT(filter->fo_lcm != NULL);
2373 rc = filter_olg_init(obd, olg, tgt);
2376 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2378 CERROR("Can't get ctxt for %p:%x\n", olg,
2379 LLOG_MDS_OST_REPL_CTXT);
2380 filter_olg_fini(olg);
2383 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2384 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2385 llog_ctxt_put(ctxt);
2389 static int filter_llog_finish(struct obd_device *obd, int count)
2391 struct filter_obd *filter = &obd->u.filter;
2392 struct llog_ctxt *ctxt;
2395 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2398 * Make sure that no cached llcds left in recov_thread.
2399 * We actually do sync in disconnect time, but disconnect
2400 * may not come being marked rq_no_resend = 1.
2402 llog_sync(ctxt, NULL);
2405 * Balance class_import_get() in llog_receptor_accept().
2406 * This is safe to do, as llog is already synchronized
2407 * and its import may go.
2409 cfs_mutex_down(&ctxt->loc_sem);
2410 if (ctxt->loc_imp) {
2411 class_import_put(ctxt->loc_imp);
2412 ctxt->loc_imp = NULL;
2414 cfs_mutex_up(&ctxt->loc_sem);
2415 llog_ctxt_put(ctxt);
2418 if (filter->fo_lcm) {
2419 cfs_mutex_down(&ctxt->loc_sem);
2420 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2421 filter->fo_lcm = NULL;
2422 cfs_mutex_up(&ctxt->loc_sem);
2424 RETURN(filter_olg_fini(&obd->obd_olg));
2428 * Find the group llog according to group index in the llog group list.
2430 static struct obd_llog_group *
2431 filter_find_olg_internal(struct filter_obd *filter, int group)
2433 struct obd_llog_group *olg;
2435 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2436 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2437 if (olg->olg_group == group)
2444 * Find the group llog according to group index on the filter
2446 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2448 struct obd_llog_group *olg = NULL;
2449 struct filter_obd *filter;
2451 filter = &obd->u.filter;
2453 if (group == FILTER_GROUP_LLOG)
2454 RETURN(&obd->obd_olg);
2456 cfs_spin_lock(&filter->fo_llog_list_lock);
2457 olg = filter_find_olg_internal(filter, group);
2458 cfs_spin_unlock(&filter->fo_llog_list_lock);
2463 * Find the llog_group of the filter according to the group. If it can not
2464 * find, create the llog_group, which only happens when mds is being synced
2467 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2469 struct obd_llog_group *olg = NULL;
2470 struct filter_obd *filter;
2473 filter = &obd->u.filter;
2475 if (group == FILTER_GROUP_LLOG)
2476 RETURN(&obd->obd_olg);
2478 cfs_spin_lock(&filter->fo_llog_list_lock);
2479 olg = filter_find_olg_internal(filter, group);
2481 if (olg->olg_initializing) {
2482 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2484 GOTO(out_unlock, olg);
2489 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2491 llog_group_init(olg, group);
2492 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2493 olg->olg_initializing = 1;
2494 cfs_spin_unlock(&filter->fo_llog_list_lock);
2496 rc = obd_llog_init(obd, olg, obd, NULL);
2498 cfs_spin_lock(&filter->fo_llog_list_lock);
2499 cfs_list_del(&olg->olg_list);
2500 cfs_spin_unlock(&filter->fo_llog_list_lock);
2502 GOTO(out, olg = ERR_PTR(-ENOMEM));
2504 cfs_spin_lock(&filter->fo_llog_list_lock);
2505 olg->olg_initializing = 0;
2506 cfs_spin_unlock(&filter->fo_llog_list_lock);
2507 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2508 obd->obd_name, group, olg);
2513 cfs_spin_unlock(&filter->fo_llog_list_lock);
2517 static int filter_llog_connect(struct obd_export *exp,
2518 struct llogd_conn_body *body)
2520 struct obd_device *obd = exp->exp_obd;
2521 struct llog_ctxt *ctxt;
2522 struct obd_llog_group *olg;
2526 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2527 obd->obd_name, body->lgdc_logid.lgl_oid,
2528 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2530 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2532 CERROR(" %s: can not find olg of group %d\n",
2533 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2536 llog_group_set_export(olg, exp);
2538 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2539 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2540 body->lgdc_ctxt_idx);
2542 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2543 obd->obd_name, body->lgdc_logid.lgl_oid,
2544 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2546 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2547 obd->u.filter.fo_mds_ost_sync = 1;
2548 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2549 rc = llog_connect(ctxt, &body->lgdc_logid,
2550 &body->lgdc_gen, NULL);
2551 llog_ctxt_put(ctxt);
2553 CERROR("failed to connect rc %d idx %d\n", rc,
2554 body->lgdc_ctxt_idx);
2559 static int filter_llog_preclean(struct obd_device *obd)
2561 struct obd_llog_group *olg, *tmp;
2562 struct filter_obd *filter;
2563 cfs_list_t remove_list;
2567 rc = obd_llog_finish(obd, 0);
2569 CERROR("failed to cleanup llogging subsystem\n");
2571 filter = &obd->u.filter;
2572 CFS_INIT_LIST_HEAD(&remove_list);
2574 cfs_spin_lock(&filter->fo_llog_list_lock);
2575 while (!cfs_list_empty(&filter->fo_llog_list)) {
2576 olg = cfs_list_entry(filter->fo_llog_list.next,
2577 struct obd_llog_group, olg_list);
2578 cfs_list_del(&olg->olg_list);
2579 cfs_list_add(&olg->olg_list, &remove_list);
2581 cfs_spin_unlock(&filter->fo_llog_list_lock);
2583 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2584 cfs_list_del_init(&olg->olg_list);
2585 rc = filter_olg_fini(olg);
2587 CERROR("failed to cleanup llogging subsystem for %u\n",
2595 static int filter_precleanup(struct obd_device *obd,
2596 enum obd_cleanup_stage stage)
2602 case OBD_CLEANUP_EARLY:
2604 case OBD_CLEANUP_EXPORTS:
2605 /* Stop recovery before namespace cleanup. */
2606 target_recovery_fini(obd);
2607 rc = filter_llog_preclean(obd);
2613 static int filter_cleanup(struct obd_device *obd)
2615 struct filter_obd *filter = &obd->u.filter;
2619 LCONSOLE_WARN("%s: shutting down for failover; client state "
2620 "will be preserved.\n", obd->obd_name);
2622 obd_exports_barrier(obd);
2623 obd_zombie_barrier();
2625 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2626 lprocfs_free_per_client_stats(obd);
2627 lprocfs_free_obd_stats(obd);
2628 lprocfs_obd_cleanup(obd);
2629 lquota_cleanup(filter_quota_interface_ref, obd);
2631 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2632 obd->obd_namespace = NULL;
2634 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2636 if (obd->u.obt.obt_sb == NULL)
2641 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2642 shrink_dcache_sb(obd->u.obt.obt_sb);
2644 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2645 obd->u.obt.obt_sb = NULL;
2647 fsfilt_put_ops(obd->obd_fsops);
2649 filter_iobuf_pool_done(filter);
2651 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2656 static int filter_connect_internal(struct obd_export *exp,
2657 struct obd_connect_data *data,
2660 struct filter_export_data *fed = &exp->exp_filter_data;
2665 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2666 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2667 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2668 data->ocd_connect_flags, data->ocd_version,
2669 data->ocd_grant, data->ocd_index);
2671 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2672 CWARN("!!! This export (nid %s) used object group %d "
2673 "earlier; now it's trying to use group %d! This could "
2674 "be a bug in the MDS. Please report to "
2675 "http://bugzilla.lustre.org/\n",
2676 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2679 fed->fed_group = data->ocd_group;
2681 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2682 exp->exp_connect_flags = data->ocd_connect_flags;
2683 data->ocd_version = LUSTRE_VERSION_CODE;
2685 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2686 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2687 data->ocd_connect_flags & OBD_CONNECT_MDS))
2690 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2691 struct filter_obd *filter = &exp->exp_obd->u.filter;
2692 obd_size left, want;
2694 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2695 left = filter_grant_space_left(exp);
2696 want = data->ocd_grant;
2697 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2698 data->ocd_grant = fed->fed_grant;
2699 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2701 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2702 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2703 exp->exp_client_uuid.uuid, exp,
2704 data->ocd_grant, want, left);
2706 filter->fo_tot_granted_clients ++;
2709 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2710 struct filter_obd *filter = &exp->exp_obd->u.filter;
2711 struct lr_server_data *lsd = filter->fo_fsd;
2712 int index = le32_to_cpu(lsd->lsd_ost_index);
2714 if (!(lsd->lsd_feature_compat &
2715 cpu_to_le32(OBD_COMPAT_OST))) {
2716 /* this will only happen on the first connect */
2717 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2718 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2719 /* sync is not needed here as filter_client_add will
2720 * set exp_need_sync flag */
2721 filter_update_server_data(exp->exp_obd,
2722 filter->fo_rcvd_filp, lsd);
2723 } else if (index != data->ocd_index) {
2724 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2725 " %u doesn't match actual OST index"
2726 " %u in last_rcvd file, bad "
2728 obd_export_nid2str(exp), index,
2732 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2733 * FIXME: We don't strictly need the COMPAT flag for that,
2734 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2735 * FIXME: We needed it for the index, as index 0 is valid. */
2738 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2739 data->ocd_brw_size = 65536;
2740 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2741 data->ocd_brw_size = min(data->ocd_brw_size,
2742 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2743 LASSERT(data->ocd_brw_size);
2746 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2747 __u32 cksum_types = data->ocd_cksum_types;
2749 /* The client set in ocd_cksum_types the checksum types it
2750 * supports. We have to mask off the algorithms that we don't
2752 if (cksum_types & OBD_CKSUM_ALL)
2753 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2755 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2757 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2758 "%x\n", exp->exp_obd->obd_name,
2759 obd_export_nid2str(exp), cksum_types,
2760 data->ocd_cksum_types);
2762 /* This client does not support OBD_CONNECT_CKSUM
2763 * fall back to CRC32 */
2764 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2765 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2766 exp->exp_obd->obd_name,
2767 obd_export_nid2str(exp));
2773 static int filter_reconnect(const struct lu_env *env,
2774 struct obd_export *exp, struct obd_device *obd,
2775 struct obd_uuid *cluuid,
2776 struct obd_connect_data *data,
2782 if (exp == NULL || obd == NULL || cluuid == NULL)
2785 rc = filter_connect_internal(exp, data, 1);
2787 filter_export_stats_init(obd, exp, localdata);
2792 /* nearly identical to mds_connect */
2793 static int filter_connect(const struct lu_env *env,
2794 struct obd_export **exp, struct obd_device *obd,
2795 struct obd_uuid *cluuid,
2796 struct obd_connect_data *data, void *localdata)
2798 struct lvfs_run_ctxt saved;
2799 struct lustre_handle conn = { 0 };
2800 struct obd_export *lexp;
2801 struct filter_export_data *fed;
2802 struct lsd_client_data *lcd = NULL;
2807 if (exp == NULL || obd == NULL || cluuid == NULL)
2810 rc = class_connect(&conn, obd, cluuid);
2813 lexp = class_conn2export(&conn);
2814 LASSERT(lexp != NULL);
2816 fed = &lexp->exp_filter_data;
2818 rc = filter_connect_internal(lexp, data, 0);
2822 filter_export_stats_init(obd, lexp, localdata);
2823 if (obd->obd_replayable) {
2824 OBD_ALLOC(lcd, sizeof(*lcd));
2826 CERROR("filter: out of memory for client data\n");
2827 GOTO(cleanup, rc = -ENOMEM);
2830 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2832 rc = filter_client_add(obd, lexp, -1);
2837 group = data->ocd_group;
2839 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2840 obd->obd_name, lexp->exp_handle.h_cookie, group);
2842 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2843 rc = filter_read_groups(obd, group, 1);
2844 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2846 CERROR("can't read group %u\n", group);
2856 fed->fed_lcd = NULL;
2858 class_disconnect(lexp);
2859 lprocfs_exp_cleanup(lexp);
2868 /* Do extra sanity checks for grant accounting. We do this at connect,
2869 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2870 * always get rid of it or turn it off when we know accounting is good. */
2871 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2873 struct filter_export_data *fed;
2874 struct obd_export *exp;
2875 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2876 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2877 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2879 if (cfs_list_empty(&obd->obd_exports))
2882 /* We don't want to do this for large machines that do lots of
2883 mounts or unmounts. It burns... */
2884 if (obd->obd_num_exports > 100)
2887 cfs_spin_lock(&obd->obd_osfs_lock);
2888 cfs_spin_lock(&obd->obd_dev_lock);
2889 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2891 fed = &exp->exp_filter_data;
2892 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2895 if (maxsize > 0) { /* we may not have done a statfs yet */
2896 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2897 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2898 exp->exp_client_uuid.uuid, exp,
2899 fed->fed_grant, fed->fed_pending, maxsize);
2900 LASSERTF(fed->fed_dirty <= maxsize,
2901 "%s: cli %s/%p %ld > "LPU64"\n", func,
2902 exp->exp_client_uuid.uuid, exp,
2903 fed->fed_dirty, maxsize);
2906 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2907 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2908 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2910 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2911 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2912 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2913 tot_granted += fed->fed_grant + fed->fed_pending;
2914 tot_pending += fed->fed_pending;
2915 tot_dirty += fed->fed_dirty;
2917 fo_tot_granted = obd->u.filter.fo_tot_granted;
2918 fo_tot_pending = obd->u.filter.fo_tot_pending;
2919 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2920 cfs_spin_unlock(&obd->obd_dev_lock);
2921 cfs_spin_unlock(&obd->obd_osfs_lock);
2923 /* Do these assertions outside the spinlocks so we don't kill system */
2924 if (tot_granted != fo_tot_granted)
2925 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2926 func, tot_granted, fo_tot_granted);
2927 if (tot_pending != fo_tot_pending)
2928 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2929 func, tot_pending, fo_tot_pending);
2930 if (tot_dirty != fo_tot_dirty)
2931 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2932 func, tot_dirty, fo_tot_dirty);
2933 if (tot_pending > tot_granted)
2934 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2935 func, tot_pending, tot_granted);
2936 if (tot_granted > maxsize)
2937 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2938 func, tot_granted, maxsize);
2939 if (tot_dirty > maxsize)
2940 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2941 func, tot_dirty, maxsize);
2944 /* Remove this client from the grant accounting totals. We also remove
2945 * the export from the obd device under the osfs and dev locks to ensure
2946 * that the filter_grant_sanity_check() calculations are always valid.
2947 * The client should do something similar when it invalidates its import. */
2948 static void filter_grant_discard(struct obd_export *exp)
2950 struct obd_device *obd = exp->exp_obd;
2951 struct filter_obd *filter = &obd->u.filter;
2952 struct filter_export_data *fed = &exp->exp_filter_data;
2954 cfs_spin_lock(&obd->obd_osfs_lock);
2955 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2956 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2957 obd->obd_name, filter->fo_tot_granted,
2958 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2959 filter->fo_tot_granted -= fed->fed_grant;
2960 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2961 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2962 obd->obd_name, filter->fo_tot_pending,
2963 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2964 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2965 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2966 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2967 obd->obd_name, filter->fo_tot_dirty,
2968 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2969 filter->fo_tot_dirty -= fed->fed_dirty;
2973 cfs_spin_unlock(&obd->obd_osfs_lock);
2976 static int filter_destroy_export(struct obd_export *exp)
2980 if (exp->exp_filter_data.fed_pending)
2981 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2982 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2983 exp, exp->exp_filter_data.fed_pending);
2985 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2987 target_destroy_export(exp);
2988 ldlm_destroy_export(exp);
2990 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2994 if (!exp->exp_obd->obd_replayable)
2995 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2997 filter_grant_discard(exp);
2998 filter_fmd_cleanup(exp);
3000 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3001 struct filter_obd *filter = &exp->exp_obd->u.filter;
3002 if (filter->fo_tot_granted_clients > 0)
3003 filter->fo_tot_granted_clients --;
3006 if (!(exp->exp_flags & OBD_OPT_FORCE))
3007 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3012 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3014 struct obd_llog_group *olg_min, *olg;
3015 struct filter_obd *filter;
3016 int worked = -1, group;
3017 struct llog_ctxt *ctxt;
3020 filter = &obd->u.filter;
3022 /* we can't sync log holding spinlock. also, we do not want to get
3023 * into livelock. so we do following: loop over MDS's exports in
3024 * group order and skip already synced llogs -bzzz */
3026 /* look for group with min. number, but > worked */
3029 cfs_spin_lock(&filter->fo_llog_list_lock);
3030 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3031 if (olg->olg_group <= worked) {
3032 /* this group is already synced */
3035 if (group < olg->olg_group) {
3036 /* we have group with smaller number to sync */
3039 /* store current minimal group */
3041 group = olg->olg_group;
3043 cfs_spin_unlock(&filter->fo_llog_list_lock);
3045 if (olg_min == NULL)
3048 worked = olg_min->olg_group;
3049 if (olg_min->olg_exp &&
3050 (dexp == olg_min->olg_exp || dexp == NULL)) {
3052 ctxt = llog_group_get_ctxt(olg_min,
3053 LLOG_MDS_OST_REPL_CTXT);
3055 err = llog_sync(ctxt, olg_min->olg_exp);
3056 llog_ctxt_put(ctxt);
3058 CERROR("error flushing logs to MDS: "
3063 } while (olg_min != NULL);
3066 /* Also incredibly similar to mds_disconnect */
3067 static int filter_disconnect(struct obd_export *exp)
3069 struct obd_device *obd = exp->exp_obd;
3074 class_export_get(exp);
3076 if (!(exp->exp_flags & OBD_OPT_FORCE))
3077 filter_grant_sanity_check(obd, __FUNCTION__);
3078 filter_grant_discard(exp);
3080 /* Flush any remaining cancel messages out to the target */
3081 filter_sync_llogs(obd, exp);
3083 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3085 rc = server_disconnect_export(exp);
3087 if (exp->exp_obd->obd_replayable)
3088 filter_client_free(exp);
3090 fsfilt_sync(obd, obd->u.obt.obt_sb);
3092 class_export_put(exp);
3096 /* reverse import is changed, sync all cancels */
3097 static void filter_revimp_update(struct obd_export *exp)
3102 class_export_get(exp);
3104 /* flush any remaining cancel messages out to the target */
3105 filter_sync_llogs(exp->exp_obd, exp);
3106 class_export_put(exp);
3110 static int filter_ping(struct obd_export *exp)
3112 filter_fmd_expire(exp);
3116 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3117 const char *what, int quiet)
3119 struct dentry *dchild = NULL;
3122 if (oa->o_valid & OBD_MD_FLGROUP)
3125 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3127 if (IS_ERR(dchild)) {
3128 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3129 what, group, oa->o_id);
3133 if (dchild->d_inode == NULL) {
3135 CERROR("%s: %s on non-existent object: "LPU64"\n",
3136 obd->obd_name, what, oa->o_id);
3138 RETURN(ERR_PTR(-ENOENT));
3144 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3146 struct dentry *dentry = NULL;
3147 struct obd_device *obd;
3151 LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
3152 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_gr,
3153 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3157 obd = class_exp2obd(exp);
3159 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3163 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3165 RETURN(PTR_ERR(dentry));
3167 /* Limit the valid bits in the return data to what we actually use */
3168 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3169 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3175 /* this should be enabled/disabled in condition to enabled/disabled large
3176 * inodes (fast EAs) in backing store FS. */
3177 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3178 void *handle, struct obdo *oa)
3180 struct obd_device *obd = exp->exp_obd;
3184 if (oa->o_valid & OBD_MD_FLFID) {
3185 struct filter_fid ff;
3187 if (!(oa->o_valid & OBD_MD_FLGROUP))
3189 /* packing fid and converting it to LE for storing into EA.
3190 * Here ->o_stripe_idx should be filled by LOV and rest of
3191 * fields - by client. */
3192 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3193 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3194 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3195 ff.ff_objid = cpu_to_le64(oa->o_id);
3196 ff.ff_group = cpu_to_le64(oa->o_gr);
3198 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3199 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3200 oa->o_generation, oa->o_id, oa->o_gr);
3202 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3204 CERROR("store fid in object failed! rc: %d\n", rc);
3206 CDEBUG(D_HA, "OSS object without fid info!\n");
3212 /* this is called from filter_truncate() until we have filter_punch() */
3213 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3214 struct obdo *oa, struct obd_trans_info *oti)
3216 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3217 struct llog_cookie *fcc = NULL;
3218 struct filter_obd *filter;
3219 int rc, err, sync = 0;
3220 loff_t old_size = 0;
3221 unsigned int ia_valid;
3222 struct inode *inode;
3223 struct page *page = NULL;
3228 LASSERT(dentry != NULL);
3229 LASSERT(!IS_ERR(dentry));
3231 inode = dentry->d_inode;
3232 LASSERT(inode != NULL);
3234 filter = &exp->exp_obd->u.filter;
3235 iattr_from_obdo(&iattr, oa, oa->o_valid);
3236 ia_valid = iattr.ia_valid;
3238 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3239 OBD_ALLOC(fcc, sizeof(*fcc));
3241 *fcc = oa->o_lcookie;
3243 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3245 /* Filter truncates and writes are serialized by
3246 * i_alloc_sem, see the comment in
3247 * filter_preprw_write.*/
3248 if (ia_valid & ATTR_SIZE)
3249 down_write(&inode->i_alloc_sem);
3250 LOCK_INODE_MUTEX(inode);
3251 old_size = i_size_read(inode);
3254 /* VBR: version recovery check */
3255 rc = filter_version_get_check(exp, oti, inode);
3257 GOTO(out_unlock, rc);
3259 /* Let's pin the last page so that ldiskfs_truncate
3260 * should not start GFP_FS allocation. */
3261 if (ia_valid & ATTR_SIZE) {
3262 page = grab_cache_page(inode->i_mapping,
3263 iattr.ia_size >> PAGE_CACHE_SHIFT);
3265 GOTO(out_unlock, rc = -ENOMEM);
3270 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3271 * then we will accept the UID+GID sent by the client during write for
3272 * initializing the ownership of this inode. We only allow this to
3273 * happen once so clear these bits in setattr. In 2.6 kernels it is
3274 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3275 * the flags that are actually being set. */
3276 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3277 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3278 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3280 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3281 if (!(ia_valid & ATTR_MODE)) {
3282 iattr.ia_mode = inode->i_mode;
3283 iattr.ia_valid |= ATTR_MODE;
3285 iattr.ia_mode &= ~S_ISUID;
3287 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3288 if (!(iattr.ia_valid & ATTR_MODE)) {
3289 iattr.ia_mode = inode->i_mode;
3290 iattr.ia_valid |= ATTR_MODE;
3292 iattr.ia_mode &= ~S_ISGID;
3295 orig_ids[USRQUOTA] = inode->i_uid;
3296 orig_ids[GRPQUOTA] = inode->i_gid;
3297 handle = fsfilt_start_log(exp->exp_obd, inode,
3298 FSFILT_OP_SETATTR, oti, 1);
3300 GOTO(out_unlock, rc = PTR_ERR(handle));
3302 /* update inode EA only once when inode is suid bit marked. As
3303 * on 2.6.x UID and GID may be set separately, we check here
3304 * only one of them to avoid double setting. */
3305 if (inode->i_mode & S_ISUID)
3306 filter_update_fidea(exp, inode, handle, oa);
3308 handle = fsfilt_start(exp->exp_obd, inode,
3309 FSFILT_OP_SETATTR, oti);
3311 GOTO(out_unlock, rc = PTR_ERR(handle));
3313 if (oa->o_valid & OBD_MD_FLFLAGS) {
3314 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3315 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3317 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3319 /* set cancel cookie callback function */
3320 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3321 filter_cancel_cookies_cb,
3325 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3326 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3328 /* The truncate might have used up our transaction credits. Make
3329 * sure we have one left for the last_rcvd update. */
3330 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3332 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3334 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3338 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3340 CERROR("error on commit, err = %d\n", err);
3347 /* For a partial-page truncate flush the page to disk immediately
3348 * to avoid data corruption during direct disk write. b=17397 */
3349 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3350 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3351 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3361 page_cache_release(page);
3363 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3364 UNLOCK_INODE_MUTEX(inode);
3365 if (ia_valid & ATTR_SIZE)
3366 up_write(&inode->i_alloc_sem);
3368 OBD_FREE(fcc, sizeof(*fcc));
3370 /* trigger quota release */
3371 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3372 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3373 int rc2 = lquota_adjust(filter_quota_interface_ref,
3374 exp->exp_obd, cur_ids,
3375 orig_ids, rc, FSFILT_OP_SETATTR);
3376 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3377 "filter adjust qunit. (rc:%d)\n", rc2);
3382 /* this is called from filter_truncate() until we have filter_punch() */
3383 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3384 struct obd_trans_info *oti)
3386 struct obdo *oa = oinfo->oi_oa;
3387 struct lustre_capa *capa = oinfo_capa(oinfo);
3388 struct ldlm_res_id res_id;
3389 struct filter_mod_data *fmd;
3390 struct lvfs_run_ctxt saved;
3391 struct filter_obd *filter;
3392 struct ldlm_resource *res;
3393 struct dentry *dentry;
3394 __u64 opc = CAPA_OPC_META_WRITE;
3398 if (oa->o_valid & OBD_FL_TRUNC)
3399 opc |= CAPA_OPC_OSS_TRUNC;
3401 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3402 rc = filter_auth_capa(exp, NULL, oa->o_gr, capa, opc);
3406 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3407 rc = filter_capa_fixoa(exp, oa, oa->o_gr, capa);
3412 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3413 /* This would be very bad - accidentally truncating a file when
3414 * changing the time or similar - bug 12203. */
3415 if (oa->o_valid & OBD_MD_FLSIZE &&
3416 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3417 static char mdsinum[48];
3419 if (oa->o_valid & OBD_MD_FLFID)
3420 snprintf(mdsinum, sizeof(mdsinum) - 1,
3421 " of inode "LPU64"/%u", oa->o_fid,
3426 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3428 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3433 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3435 RETURN(PTR_ERR(dentry));
3437 filter = &exp->exp_obd->u.filter;
3438 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3441 * We need to be atomic against a concurrent write
3442 * (which takes the semaphore for reading). fmd_mactime_xid
3443 * checks will have no effect if a write request with lower
3444 * xid starts just before a setattr and finishes later than
3445 * the setattr (see bug 21489, comment 27).
3448 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3449 down_write(&dentry->d_inode->i_alloc_sem);
3450 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3451 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3452 fmd->fmd_mactime_xid = oti->oti_xid;
3453 filter_fmd_put(exp, fmd);
3454 up_write(&dentry->d_inode->i_alloc_sem);
3457 /* setting objects attributes (including owner/group) */
3458 rc = filter_setattr_internal(exp, dentry, oa, oti);
3460 GOTO(out_unlock, rc);
3462 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3463 &res_id, LDLM_EXTENT, 0);
3466 LDLM_RESOURCE_ADDREF(res);
3467 rc = ldlm_res_lvbo_update(res, NULL, 0);
3468 LDLM_RESOURCE_DELREF(res);
3469 ldlm_resource_putref(res);
3472 oa->o_valid = OBD_MD_FLID;
3474 /* Quota release need uid/gid info */
3475 obdo_from_inode(oa, dentry->d_inode,
3476 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3481 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3485 /* XXX identical to osc_unpackmd */
3486 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3487 struct lov_mds_md *lmm, int lmm_bytes)
3493 if (lmm_bytes < sizeof (*lmm)) {
3494 CERROR("lov_mds_md too small: %d, need %d\n",
3495 lmm_bytes, (int)sizeof(*lmm));
3498 /* XXX LOV_MAGIC etc check? */
3500 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3501 CERROR("lov_mds_md: zero lmm_object_id\n");
3506 lsm_size = lov_stripe_md_size(1);
3510 if (*lsmp != NULL && lmm == NULL) {
3511 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3512 OBD_FREE(*lsmp, lsm_size);
3517 if (*lsmp == NULL) {
3518 OBD_ALLOC(*lsmp, lsm_size);
3522 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3523 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3524 OBD_FREE(*lsmp, lsm_size);
3527 loi_init((*lsmp)->lsm_oinfo[0]);
3531 /* XXX zero *lsmp? */
3532 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3533 LASSERT((*lsmp)->lsm_object_id);
3536 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3541 /* caller must hold fo_create_locks[oa->o_gr] */
3542 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3543 struct filter_obd *filter)
3545 struct obdo doa = { 0 }; /* XXX obdo on stack */
3552 LASSERT_MDS_GROUP(oa->o_gr);
3553 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3554 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3556 memset(&doa, 0, sizeof(doa));
3558 doa.o_valid |= OBD_MD_FLGROUP;
3559 doa.o_gr = oa->o_gr;
3560 doa.o_mode = S_IFREG;
3562 if (!cfs_test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3563 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3564 exp->exp_obd->obd_name, doa.o_gr);
3568 last = filter_last_id(filter, doa.o_gr);
3570 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3572 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3573 exp->exp_obd->obd_name, oa->o_id + 1, last,
3574 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3576 for (id = last; id > oa->o_id; id--) {
3578 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3579 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3580 CEMERG("error destroying precreate objid "LPU64": %d\n",
3583 /* update last_id on disk periodically so that if we restart
3584 * we don't need to re-scan all of the just-deleted objects. */
3585 if ((id & 511) == 0 && !skip_orphan) {
3586 filter_set_last_id(filter, id - 1, doa.o_gr);
3587 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3591 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3592 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3595 filter_set_last_id(filter, id, doa.o_gr);
3596 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3598 /* don't reuse orphan object, return last used objid */
3602 cfs_clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3607 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3608 obd_gr group, int *num);
3609 /* returns a negative error or a nonnegative number of files to create */
3610 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3611 obd_gr group, struct obd_trans_info *oti)
3613 struct obd_device *obd = exp->exp_obd;
3614 struct filter_obd *filter = &obd->u.filter;
3618 /* delete orphans request */
3619 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3620 obd_id last = filter_last_id(filter, group);
3622 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3623 CERROR("%s: dropping old orphan cleanup request\n",
3627 /* This causes inflight precreates to abort and drop lock */
3628 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3629 cfs_down(&filter->fo_create_locks[group]);
3630 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3631 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3632 exp->exp_obd->obd_name, group);
3633 cfs_up(&filter->fo_create_locks[group]);
3636 diff = oa->o_id - last;
3637 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3640 if (-diff > OST_MAX_PRECREATE) {
3641 CERROR("%s: ignoring bogus orphan destroy request: "
3642 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3644 /* FIXME: should reset precreate_next_id on MDS */
3645 GOTO(out, rc = -EINVAL);
3648 rc = filter_destroy_precreated(exp, oa, filter);
3650 CERROR("%s: unable to write lastobjid, but "
3651 "orphans were deleted\n", obd->obd_name);
3654 /* XXX: Used by MDS for the first time! */
3655 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3658 cfs_down(&filter->fo_create_locks[group]);
3659 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3660 CERROR("%s: dropping old precreate request\n",
3664 /* only precreate if group == 0 and o_id is specfied */
3665 if (!filter_group_is_mds(group) || oa->o_id == 0)
3668 diff = oa->o_id - filter_last_id(filter, group);
3669 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3670 filter_last_id(filter, group), diff);
3672 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3673 oa->o_id, filter_last_id(filter, group), diff);
3677 oa->o_id = filter_last_id(&obd->u.filter, group);
3678 rc = filter_precreate(obd, oa, group, &diff);
3679 oa->o_id = filter_last_id(&obd->u.filter, group);
3681 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3684 /* else diff == 0 */
3687 cfs_up(&filter->fo_create_locks[group]);
3691 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3692 __u64 max_age, __u32 flags)
3694 struct filter_obd *filter = &obd->u.filter;
3695 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3699 /* at least try to account for cached pages. its still racey and
3700 * might be under-reporting if clients haven't announced their
3701 * caches with brw recently */
3702 cfs_spin_lock(&obd->obd_osfs_lock);
3703 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3704 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3705 cfs_spin_unlock(&obd->obd_osfs_lock);
3707 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3708 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3709 filter->fo_tot_dirty, filter->fo_tot_granted,
3710 filter->fo_tot_pending,
3711 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3713 filter_grant_sanity_check(obd, __FUNCTION__);
3715 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3716 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3717 osfs->os_bsize - 1) >> blockbits));
3719 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3720 struct lr_server_data *lsd = filter->fo_fsd;
3721 int index = le32_to_cpu(lsd->lsd_ost_index);
3723 if (obd_fail_val == -1 ||
3724 index == obd_fail_val)
3725 osfs->os_bfree = osfs->os_bavail = 2;
3726 else if (obd_fail_loc & OBD_FAIL_ONCE)
3727 obd_fail_loc &= ~OBD_FAILED; /* reset flag */
3730 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3731 * stop creating files on MDS if OST is not good shape to create
3735 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3736 osfs->os_state = OS_STATE_READONLY;
3738 if (filter->fo_raid_degraded)
3739 osfs->os_state |= OS_STATE_DEGRADED;
3743 static int filter_use_existing_obj(struct obd_device *obd,
3744 struct dentry *dchild, void **handle,
3747 struct inode *inode = dchild->d_inode;
3751 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3754 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3755 if (IS_ERR(*handle))
3756 return PTR_ERR(*handle);
3758 iattr.ia_valid = ATTR_MODE;
3759 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3760 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3767 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3770 __u64 os_ffree = -1;
3772 cfs_spin_lock(&obd->obd_osfs_lock);
3773 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3775 os_ffree = obd->obd_osfs.os_ffree;
3776 cfs_spin_unlock(&obd->obd_osfs_lock);
3781 /* We rely on the fact that only one thread will be creating files in a given
3782 * group at a time, which is why we don't need an atomic filter_get_new_id.
3783 * Even if we had that atomic function, the following race would exist:
3785 * thread 1: gets id x from filter_next_id
3786 * thread 2: gets id (x + 1) from filter_next_id
3787 * thread 2: creates object (x + 1)
3788 * thread 1: tries to create object x, gets -ENOSPC
3790 * Caller must hold fo_create_locks[group]
3792 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3793 obd_gr group, int *num)
3795 struct dentry *dchild = NULL, *dparent = NULL;
3796 struct filter_obd *filter;
3797 struct obd_statfs *osfs;
3798 int err = 0, rc = 0, recreate_obj = 0, i;
3799 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3802 void *handle = NULL;
3805 filter = &obd->u.filter;
3807 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3809 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3811 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3812 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3815 OBD_ALLOC(osfs, sizeof(*osfs));
3818 rc = filter_statfs(obd, osfs, cfs_time_current_64() - CFS_HZ,
3820 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3821 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3822 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3823 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3827 OBD_FREE(osfs, sizeof(*osfs));
3832 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3833 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3835 for (i = 0; i < *num && err == 0; i++) {
3836 int cleanup_phase = 0;
3838 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3839 CWARN("%s: create aborted by destroy\n",
3848 last_id = filter_last_id(filter, group);
3849 if (next_id > last_id) {
3850 CERROR("Error: Trying to recreate obj greater"
3851 "than last id "LPD64" > "LPD64"\n",
3853 GOTO(cleanup, rc = -EINVAL);
3856 next_id = filter_last_id(filter, group) + 1;
3858 dparent = filter_parent_lock(obd, group, next_id);
3859 if (IS_ERR(dparent))
3860 GOTO(cleanup, rc = PTR_ERR(dparent));
3861 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3863 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3865 GOTO(cleanup, rc = PTR_ERR(dchild));
3866 cleanup_phase = 2; /* f_dput(dchild) */
3868 if (dchild->d_inode != NULL) {
3869 /* This would only happen if lastobjid was bad on disk*/
3870 /* Could also happen if recreating missing obj but it
3871 * already exists. */
3873 CERROR("%s: recreating existing object %.*s?\n",
3874 obd->obd_name, dchild->d_name.len,
3875 dchild->d_name.name);
3877 /* Use these existing objects if they are
3879 if (dchild->d_inode->i_size == 0) {
3880 rc = filter_use_existing_obj(obd,dchild,
3881 &handle, &cleanup_phase);
3888 CERROR("%s: Serious error: objid %.*s already "
3889 "exists; is this filesystem corrupt?\n",
3890 obd->obd_name, dchild->d_name.len,
3891 dchild->d_name.name);
3894 GOTO(cleanup, rc = -EEXIST);
3897 handle = fsfilt_start_log(obd, dparent->d_inode,
3898 FSFILT_OP_CREATE, NULL, 1);
3900 GOTO(cleanup, rc = PTR_ERR(handle));
3903 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3904 ",od->o_id="LPU64")\n", obd->obd_name, group,
3907 /* We mark object SUID+SGID to flag it for accepting UID+GID
3908 * from client on first write. Currently the permission bits
3909 * on the OST are never used, so this is OK. */
3910 rc = ll_vfs_create(dparent->d_inode, dchild,
3911 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3913 CERROR("create failed rc = %d\n", rc);
3914 if (rc == -ENOSPC) {
3915 os_ffree = filter_calc_free_inodes(obd);
3917 CERROR("%s: free inode "LPU64"\n",
3918 obd->obd_name, os_ffree);
3923 if (dchild->d_inode)
3924 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3925 dchild->d_inode->i_ino);
3928 if (!recreate_obj) {
3929 filter_set_last_id(filter, next_id, group);
3930 err = filter_update_last_objid(obd, group, 0);
3932 CERROR("unable to write lastobjid "
3933 "but file created\n");
3937 switch(cleanup_phase) {
3939 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3941 CERROR("error on commit, err = %d\n", err);
3948 filter_parent_unlock(dparent);
3955 if (cfs_time_after(jiffies, enough_time)) {
3957 "%s: precreate slow - want %d got %d \n",
3958 obd->obd_name, *num, i);
3965 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3966 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3971 static int filter_create(struct obd_export *exp, struct obdo *oa,
3972 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3974 struct obd_device *obd = exp->exp_obd;
3975 struct filter_export_data *fed;
3976 struct filter_obd *filter;
3977 struct lvfs_run_ctxt saved;
3978 struct lov_stripe_md *lsm = NULL;
3979 int rc = 0, diff, group = oa->o_gr;
3982 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3983 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3985 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3986 CERROR("!!! nid %s sent invalid object group %d\n",
3987 obd_export_nid2str(exp), group);
3991 fed = &exp->exp_filter_data;
3992 filter = &obd->u.filter;
3994 if (fed->fed_group != group) {
3995 CERROR("!!! this export (nid %s) used object group %d "
3996 "earlier; now it's trying to use group %d! This could "
3997 "be a bug in the MDS. Please report to "
3998 "http://bugzilla.lustre.org/\n",
3999 obd_export_nid2str(exp), fed->fed_group, group);
4006 rc = obd_alloc_memmd(exp, &lsm);
4013 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4015 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4016 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4017 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
4018 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4019 oa->o_id, filter_last_id(filter,
4024 cfs_down(&filter->fo_create_locks[oa->o_gr]);
4025 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
4026 cfs_up(&filter->fo_create_locks[oa->o_gr]);
4029 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
4032 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4033 if (rc && ea != NULL && *ea != lsm) {
4034 obd_free_memmd(exp, &lsm);
4035 } else if (rc == 0 && ea != NULL) {
4036 /* XXX LOV STACKING: the lsm that is passed to us from
4037 * LOV does not have valid lsm_oinfo data structs, so
4038 * don't go touching that. This needs to be fixed in a
4040 lsm->lsm_object_id = oa->o_id;
4047 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4048 struct lov_stripe_md *md, struct obd_trans_info *oti,
4049 struct obd_export *md_exp, void *capa)
4051 unsigned int qcids[MAXQUOTAS] = {0, 0};
4052 struct obd_device *obd;
4053 struct filter_obd *filter;
4054 struct dentry *dchild = NULL, *dparent = NULL;
4055 struct lustre_handle lockh = { 0 };
4056 struct lvfs_run_ctxt saved;
4057 void *handle = NULL;
4058 struct llog_cookie *fcc = NULL;
4059 int rc, rc2, cleanup_phase = 0, sync = 0;
4063 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4064 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4065 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4070 filter = &obd->u.filter;
4072 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4075 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4076 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4078 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4080 GOTO(cleanup, rc = PTR_ERR(dchild));
4083 if (dchild->d_inode == NULL) {
4084 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4086 /* If object already gone, cancel cookie right now */
4087 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4088 struct llog_ctxt *ctxt;
4089 struct obd_llog_group *olg;
4091 olg = filter_find_olg(obd, oa->o_gr);
4093 CERROR(" %s: can not find olg of group %d\n",
4094 obd->obd_name, (int)oa->o_gr);
4095 GOTO(cleanup, rc = PTR_ERR(olg));
4097 fcc = &oa->o_lcookie;
4098 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4099 llog_cancel(ctxt, NULL, 1, fcc, 0);
4100 llog_ctxt_put(ctxt);
4101 fcc = NULL; /* we didn't allocate fcc, don't free it */
4103 GOTO(cleanup, rc = -ENOENT);
4106 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4110 /* Our MDC connection is established by the MDS to us */
4111 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4112 OBD_ALLOC(fcc, sizeof(*fcc));
4114 *fcc = oa->o_lcookie;
4116 DQUOT_INIT(dchild->d_inode);
4118 /* we're gonna truncate it first in order to avoid possible deadlock:
4120 * open trasaction open transaction
4121 * down(i_zombie) down(i_zombie)
4122 * restart transaction
4123 * (see BUG 4180) -bzzz
4125 * take i_alloc_sem too to prevent other threads from writing to the
4126 * file while we are truncating it. This can cause lock ordering issue
4127 * between page lock, i_mutex & starting new journal handle.
4128 * (see bug 20321) -johann
4130 down_write(&dchild->d_inode->i_alloc_sem);
4131 LOCK_INODE_MUTEX(dchild->d_inode);
4133 /* VBR: version recovery check */
4134 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4136 UNLOCK_INODE_MUTEX(dchild->d_inode);
4137 up_write(&dchild->d_inode->i_alloc_sem);
4141 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4143 if (IS_ERR(handle)) {
4144 UNLOCK_INODE_MUTEX(dchild->d_inode);
4145 up_write(&dchild->d_inode->i_alloc_sem);
4146 GOTO(cleanup, rc = PTR_ERR(handle));
4149 iattr.ia_valid = ATTR_SIZE;
4151 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4152 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4153 UNLOCK_INODE_MUTEX(dchild->d_inode);
4154 up_write(&dchild->d_inode->i_alloc_sem);
4158 GOTO(cleanup, rc = rc2);
4160 /* We don't actually need to lock the parent until we are unlinking
4161 * here, and not while truncating above. That avoids holding the
4162 * parent lock for a long time during truncate, which can block other
4163 * threads from doing anything to objects in that directory. bug 7171 */
4164 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4165 if (IS_ERR(dparent))
4166 GOTO(cleanup, rc = PTR_ERR(dparent));
4167 cleanup_phase = 3; /* filter_parent_unlock */
4169 LOCK_INODE_MUTEX(dchild->d_inode);
4170 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4171 if (IS_ERR(handle)) {
4172 UNLOCK_INODE_MUTEX(dchild->d_inode);
4173 GOTO(cleanup, rc = PTR_ERR(handle));
4175 cleanup_phase = 4; /* fsfilt_commit */
4177 /* Quota release need uid/gid of inode */
4178 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4180 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4182 /* this drops dchild->d_inode->i_mutex unconditionally */
4183 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4187 switch(cleanup_phase) {
4190 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4191 oti->oti_handle : handle,
4192 filter_cancel_cookies_cb,
4194 /* If add_journal_cb failed, then filter_finish_transno
4195 * will commit the handle and we will do a sync
4196 * on commit. then we call callback directly to free
4199 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4201 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4204 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4206 CERROR("error on commit, err = %d\n", rc2);
4213 filter_parent_unlock(dparent);
4215 filter_fini_destroy(obd, &lockh);
4219 OBD_FREE(fcc, sizeof(*fcc));
4221 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4224 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4228 /* trigger quota release */
4229 qcids[USRQUOTA] = oa->o_uid;
4230 qcids[GRPQUOTA] = oa->o_gid;
4231 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4234 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4238 /* NB start and end are used for punch, but not truncate */
4239 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4240 struct obd_trans_info *oti,
4241 struct ptlrpc_request_set *rqset)
4246 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4247 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4248 oinfo->oi_policy.l_extent.end);
4252 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4253 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4254 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4256 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4257 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4258 rc = filter_setattr(exp, oinfo, oti);
4259 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4263 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4264 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4267 struct lvfs_run_ctxt saved;
4268 struct filter_obd *filter;
4269 struct dentry *dentry;
4273 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4274 rc = filter_auth_capa(exp, NULL, oa->o_gr,
4275 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4279 filter = &exp->exp_obd->u.filter;
4281 /* An objid of zero is taken to mean "sync whole filesystem" */
4282 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4283 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4284 /* Flush any remaining cancel messages out to the target */
4285 filter_sync_llogs(exp->exp_obd, exp);
4289 dentry = filter_oa2dentry(exp->exp_obd, oa);
4291 RETURN(PTR_ERR(dentry));
4293 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4295 LOCK_INODE_MUTEX(dentry->d_inode);
4297 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4299 /* just any file to grab fsync method - "file" arg unused */
4300 struct file *file = filter->fo_rcvd_filp;
4302 if (file->f_op && file->f_op->fsync)
4303 rc = file->f_op->fsync(NULL, dentry, 1);
4305 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4309 UNLOCK_INODE_MUTEX(dentry->d_inode);
4311 oa->o_valid = OBD_MD_FLID;
4312 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4314 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4320 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4321 void *key, __u32 *vallen, void *val,
4322 struct lov_stripe_md *lsm)
4324 struct obd_device *obd;
4327 obd = class_exp2obd(exp);
4329 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4333 if (KEY_IS(KEY_BLOCKSIZE)) {
4334 __u32 *blocksize = val;
4336 if (*vallen < sizeof(*blocksize))
4338 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4340 *vallen = sizeof(*blocksize);
4344 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4345 __u32 *blocksize_bits = val;
4346 if (blocksize_bits) {
4347 if (*vallen < sizeof(*blocksize_bits))
4349 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4351 *vallen = sizeof(*blocksize_bits);
4355 if (KEY_IS(KEY_LAST_ID)) {
4356 obd_id *last_id = val;
4357 /* FIXME: object groups */
4359 if (*vallen < sizeof(*last_id))
4361 *last_id = filter_last_id(&obd->u.filter,
4362 exp->exp_filter_data.fed_group);
4364 *vallen = sizeof(*last_id);
4368 if (KEY_IS(KEY_FIEMAP)) {
4369 struct ll_fiemap_info_key *fm_key = key;
4370 struct dentry *dentry;
4371 struct ll_user_fiemap *fiemap = val;
4372 struct lvfs_run_ctxt saved;
4375 if (fiemap == NULL) {
4376 *vallen = fiemap_count_to_size(
4377 fm_key->fiemap.fm_extent_count);
4381 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4384 RETURN(PTR_ERR(dentry));
4386 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4387 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4388 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4390 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4396 CDEBUG(D_IOCTL, "invalid key\n");
4400 static inline int filter_setup_llog_group(struct obd_export *exp,
4401 struct obd_device *obd,
4404 struct obd_llog_group *olg;
4405 struct llog_ctxt *ctxt;
4408 olg = filter_find_create_olg(obd, group);
4410 RETURN(PTR_ERR(olg));
4412 llog_group_set_export(olg, exp);
4414 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4415 LASSERTF(ctxt != NULL, "ctxt is null\n");
4417 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4418 llog_ctxt_put(ctxt);
4422 static int filter_set_grant_shrink(struct obd_export *exp,
4423 struct ost_body *body)
4425 /* handle shrink grant */
4426 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4427 filter_grant_incoming(exp, &body->oa);
4428 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4434 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4436 struct obd_device *obd;
4442 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4446 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4447 obd_export_nid2str(exp));
4448 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4450 /* setup llog imports */
4452 group = (int)(*(__u32 *)val);
4454 group = 0; /* default value */
4456 LASSERT_MDS_GROUP(group);
4457 rc = filter_setup_llog_group(exp, obd, group);
4461 if (group == FILTER_GROUP_MDS0) {
4462 /* setup llog group 1 for interop */
4463 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4466 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4471 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4472 void *key, __u32 vallen, void *val,
4473 struct ptlrpc_request_set *set)
4475 struct obd_device *obd;
4480 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4484 if (KEY_IS(KEY_CAPA_KEY)) {
4486 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4488 CERROR("filter update capability key failed: %d\n", rc);
4492 if (KEY_IS(KEY_REVIMP_UPD)) {
4493 filter_revimp_update(exp);
4494 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4498 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4499 filter_adapt_sptlrpc_conf(obd, 0);
4503 if (KEY_IS(KEY_MDS_CONN))
4504 RETURN(filter_set_mds_conn(exp, val));
4506 if (KEY_IS(KEY_GRANT_SHRINK))
4507 RETURN(filter_set_grant_shrink(exp, val));
4512 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4513 int len, void *karg, void *uarg)
4515 struct obd_device *obd = exp->exp_obd;
4516 struct obd_ioctl_data *data = karg;
4520 case OBD_IOC_ABORT_RECOVERY: {
4521 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4522 target_stop_recovery_thread(obd);
4526 case OBD_IOC_SYNC: {
4527 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4528 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4532 case OBD_IOC_SET_READONLY: {
4534 struct super_block *sb = obd->u.obt.obt_sb;
4535 struct inode *inode = sb->s_root->d_inode;
4536 BDEVNAME_DECLARE_STORAGE(tmp);
4537 CERROR("*** setting device %s read-only ***\n",
4538 ll_bdevname(sb, tmp));
4540 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4541 if (!IS_ERR(handle))
4542 rc = fsfilt_commit(obd, inode, handle, 1);
4544 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4545 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4547 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4551 case OBD_IOC_CATLOGLIST: {
4552 rc = llog_catalog_list(obd, 1, data);
4556 case OBD_IOC_LLOG_CANCEL:
4557 case OBD_IOC_LLOG_REMOVE:
4558 case OBD_IOC_LLOG_INFO:
4559 case OBD_IOC_LLOG_PRINT: {
4560 /* FIXME to be finished */
4561 RETURN(-EOPNOTSUPP);
4563 struct llog_ctxt *ctxt = NULL;
4565 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4566 rc = llog_ioctl(ctxt, cmd, data);
4567 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4580 static int filter_health_check(struct obd_device *obd)
4582 #ifdef USE_HEALTH_CHECK_WRITE
4583 struct filter_obd *filter = &obd->u.filter;
4588 * health_check to return 0 on healthy
4589 * and 1 on unhealthy.
4591 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4594 #ifdef USE_HEALTH_CHECK_WRITE
4595 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4596 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4601 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4604 return filter_fid2dentry(data, NULL, gr, id);
4607 static int filter_process_config(struct obd_device *obd, obd_count len,
4610 struct lustre_cfg *lcfg = buf;
4611 struct lprocfs_static_vars lvars;
4614 switch (lcfg->lcfg_command) {
4616 lprocfs_filter_init_vars(&lvars);
4618 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4628 static struct lvfs_callback_ops filter_lvfs_ops = {
4629 l_fid2dentry: filter_lvfs_fid2dentry,
4632 static struct obd_ops filter_obd_ops = {
4633 .o_owner = THIS_MODULE,
4634 .o_get_info = filter_get_info,
4635 .o_set_info_async = filter_set_info_async,
4636 .o_setup = filter_setup,
4637 .o_precleanup = filter_precleanup,
4638 .o_cleanup = filter_cleanup,
4639 .o_connect = filter_connect,
4640 .o_reconnect = filter_reconnect,
4641 .o_disconnect = filter_disconnect,
4642 .o_ping = filter_ping,
4643 .o_init_export = filter_init_export,
4644 .o_destroy_export = filter_destroy_export,
4645 .o_statfs = filter_statfs,
4646 .o_getattr = filter_getattr,
4647 .o_unpackmd = filter_unpackmd,
4648 .o_create = filter_create,
4649 .o_setattr = filter_setattr,
4650 .o_destroy = filter_destroy,
4651 .o_brw = filter_brw,
4652 .o_punch = filter_truncate,
4653 .o_sync = filter_sync,
4654 .o_preprw = filter_preprw,
4655 .o_commitrw = filter_commitrw,
4656 .o_llog_init = filter_llog_init,
4657 .o_llog_connect = filter_llog_connect,
4658 .o_llog_finish = filter_llog_finish,
4659 .o_iocontrol = filter_iocontrol,
4660 .o_health_check = filter_health_check,
4661 .o_process_config = filter_process_config,
4664 quota_interface_t *filter_quota_interface_ref;
4665 extern quota_interface_t filter_quota_interface;
4667 static int __init obdfilter_init(void)
4669 struct lprocfs_static_vars lvars;
4672 /** sanity check for group<->mdsno conversion */
4673 for (i = 0; i < 32; i++)
4674 LASSERT(objgrp_to_mdsno(mdt_to_obd_objgrp(i)) == i);
4676 lprocfs_filter_init_vars(&lvars);
4678 cfs_request_module("%s", "lquota");
4679 OBD_ALLOC(obdfilter_created_scratchpad,
4680 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4681 sizeof(*obdfilter_created_scratchpad));
4682 if (obdfilter_created_scratchpad == NULL)
4685 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4686 sizeof(struct filter_mod_data),
4689 GOTO(out, rc = -ENOMEM);
4691 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4692 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4694 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4695 LUSTRE_OST_NAME, NULL);
4699 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4700 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4701 ll_fmd_cachep = NULL;
4703 if (filter_quota_interface_ref)
4704 PORTAL_SYMBOL_PUT(filter_quota_interface);
4706 OBD_FREE(obdfilter_created_scratchpad,
4707 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4708 sizeof(*obdfilter_created_scratchpad));
4714 static void __exit obdfilter_exit(void)
4716 if (filter_quota_interface_ref)
4717 PORTAL_SYMBOL_PUT(filter_quota_interface);
4719 if (ll_fmd_cachep) {
4720 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4721 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4722 ll_fmd_cachep = NULL;
4725 class_unregister_type(LUSTRE_OST_NAME);
4726 OBD_FREE(obdfilter_created_scratchpad,
4727 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4728 sizeof(*obdfilter_created_scratchpad));
4731 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4732 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4733 MODULE_LICENSE("GPL");
4735 module_init(obdfilter_init);
4736 module_exit(obdfilter_exit);