1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdfilter/filter.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Andreas Dilger <adilger@clusterfs.com>
43 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44 * (which need to get journal_lock, may block if journal full).
46 * Invariant: Call filter_start_transno() before any journal ops to avoid the
47 * same deadlock problem. We can (and want) to get rid of the
48 * transno sem in favour of the dir/inode i_mutex to avoid single
49 * threaded operation on the OST.
52 #define DEBUG_SUBSYSTEM S_FILTER
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
57 #include <linux/module.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
66 #include <obd_cksum.h>
67 #include <obd_class.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
80 #include "filter_internal.h"
82 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
83 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS1_N_BASE
84 static struct lvfs_callback_ops filter_lvfs_ops;
85 cfs_mem_cache_t *ll_fmd_cachep;
87 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
88 void *cb_data, int error)
90 struct obd_export *exp = cb_data;
91 LASSERT(exp->exp_obd == obd);
92 obd_transno_commit_cb(obd, transno, exp, error);
93 class_export_put(exp);
96 int filter_version_get_check(struct obd_export *exp,
97 struct obd_trans_info *oti, struct inode *inode)
101 if (inode == NULL || oti == NULL)
104 curr_version = fsfilt_get_version(exp->exp_obd, inode);
105 if ((__s64)curr_version == -EOPNOTSUPP)
107 /* VBR: version is checked always because costs nothing */
108 if (oti->oti_pre_version != 0 &&
109 oti->oti_pre_version != curr_version) {
110 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
111 oti->oti_pre_version, curr_version);
112 spin_lock(&exp->exp_lock);
113 exp->exp_vbr_failed = 1;
114 spin_unlock(&exp->exp_lock);
117 oti->oti_pre_version = curr_version;
121 /* Assumes caller has already pushed us into the kernel context. */
122 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
123 struct obd_trans_info *oti, int rc, int force_sync)
125 struct filter_obd *filter = &exp->exp_obd->u.filter;
126 struct filter_export_data *fed = &exp->exp_filter_data;
127 struct lsd_client_data *lcd = fed->fed_lcd;
130 int err, log_pri = D_RPCTRACE;
132 /* Propagate error code. */
136 if (!exp->exp_obd->obd_replayable || oti == NULL)
139 /* we don't allocate new transnos for replayed requests */
140 spin_lock(&filter->fo_translock);
141 if (oti->oti_transno == 0) {
142 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
143 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
145 last_rcvd = oti->oti_transno;
146 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
147 filter->fo_fsd->lsd_last_transno =
148 cpu_to_le64(last_rcvd);
150 oti->oti_transno = last_rcvd;
151 if (last_rcvd <= le64_to_cpu(lcd->lcd_last_transno)) {
152 spin_unlock(&filter->fo_translock);
155 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
156 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
157 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
158 spin_unlock(&filter->fo_translock);
161 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
163 off = fed->fed_lr_off;
165 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
166 fed->fed_lr_idx, fed->fed_lr_off);
169 class_export_get(exp); /* released when the cb is called */
171 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
177 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
178 lcd, sizeof(*lcd), &off,
179 force_sync | exp->exp_need_sync);
181 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
189 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
190 last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
195 void f_dput(struct dentry *dentry)
197 /* Can't go inside filter_ddelete because it can block */
198 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
199 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
200 LASSERT(atomic_read(&dentry->d_count) > 0);
205 static void init_brw_stats(struct brw_stats *brw_stats)
208 for (i = 0; i < BRW_LAST; i++)
209 spin_lock_init(&brw_stats->hist[i].oh_lock);
212 static int lprocfs_init_rw_stats(struct obd_device *obd,
213 struct lprocfs_stats **stats)
217 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
218 LPROC_FILTER_LAST - 1;
219 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
223 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
224 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
225 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
226 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
227 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
232 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
233 plus the procfs overhead :( */
234 static int filter_export_stats_init(struct obd_device *obd,
235 struct obd_export *exp,
241 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
242 /* Self-export gets no proc entry */
245 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
247 /* Mask error for already created
255 struct nid_stat *tmp = exp->exp_nid_stats;
256 LASSERT(tmp != NULL);
258 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
259 if (tmp->nid_brw_stats == NULL)
262 init_brw_stats(tmp->nid_brw_stats);
263 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
264 0644, &filter_per_nid_stats_fops,
267 CWARN("Error adding the brw_stats file\n");
269 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
273 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
277 /* Always add in ldlm_stats */
278 tmp->nid_ldlm_stats =
279 lprocfs_alloc_stats(LDLM_LAST_OPC - LDLM_FIRST_OPC,
280 LPROCFS_STATS_FLAG_NOPERCPU);
281 if (tmp->nid_ldlm_stats == NULL)
284 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
285 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
286 tmp->nid_ldlm_stats);
294 /* Add client data to the FILTER. We use a bitmap to locate a free space
295 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
296 * Otherwise, we have just read the data from the last_rcvd file and
297 * we know its offset. */
298 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
301 struct filter_obd *filter = &obd->u.filter;
302 struct filter_export_data *fed = &exp->exp_filter_data;
303 unsigned long *bitmap = filter->fo_last_rcvd_slots;
304 int new_client = (cl_idx == -1);
308 LASSERT(bitmap != NULL);
309 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
312 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
315 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
316 * there's no need for extra complication here
319 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
321 if (cl_idx >= LR_MAX_CLIENTS) {
322 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
326 if (test_and_set_bit(cl_idx, bitmap)) {
327 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
332 if (test_and_set_bit(cl_idx, bitmap)) {
333 CERROR("FILTER client %d: bit already set in bitmap!\n",
339 fed->fed_lr_idx = cl_idx;
340 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
341 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
342 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
344 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
345 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
348 struct lvfs_run_ctxt saved;
349 loff_t off = fed->fed_lr_off;
353 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
354 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
356 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
357 /* Transaction needed to fix bug 1403 */
358 handle = fsfilt_start(obd,
359 filter->fo_rcvd_filp->f_dentry->d_inode,
360 FSFILT_OP_SETATTR, NULL);
361 if (IS_ERR(handle)) {
362 rc = PTR_ERR(handle);
363 CERROR("unable to start transaction: rc %d\n", rc);
365 fed->fed_lcd->lcd_last_epoch =
366 filter->fo_fsd->lsd_start_epoch;
367 exp->exp_last_request_time = cfs_time_current_sec();
368 rc = fsfilt_add_journal_cb(obd, 0, handle,
369 target_client_add_cb, exp);
371 spin_lock(&exp->exp_lock);
372 exp->exp_need_sync = 1;
373 spin_unlock(&exp->exp_lock);
375 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
377 sizeof(*fed->fed_lcd),
378 &off, rc /* sync if no cb */);
380 filter->fo_rcvd_filp->f_dentry->d_inode,
383 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
386 CERROR("error writing %s client idx %u: rc %d\n",
387 LAST_RCVD, fed->fed_lr_idx, rc);
394 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
396 static int filter_client_free(struct obd_export *exp)
398 struct filter_export_data *fed = &exp->exp_filter_data;
399 struct filter_obd *filter = &exp->exp_obd->u.filter;
400 struct obd_device *obd = exp->exp_obd;
401 struct lvfs_run_ctxt saved;
406 if (fed->fed_lcd == NULL)
409 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
410 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
413 LASSERT(filter->fo_last_rcvd_slots != NULL);
415 off = fed->fed_lr_off;
417 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
418 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
420 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
421 * we leak a client slot that will be cleaned on the next recovery. */
423 CERROR("%s: client idx %d has med_off %lld\n",
424 obd->obd_name, fed->fed_lr_idx, off);
425 GOTO(free, rc = -EINVAL);
428 /* Clear the bit _after_ zeroing out the client so we don't
429 race with filter_client_add and zero out new clients.*/
430 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
431 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
436 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
437 /* Don't force sync on disconnect if aborting recovery,
438 * or it does num_clients * num_osts. b=17194 */
439 int need_sync = exp->exp_need_sync &&
440 !(exp->exp_flags&OBD_OPT_ABORT_RECOV);
441 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
442 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
443 sizeof(zero_lcd), &off, 0);
445 /* Make sure the server's last_transno is up to date. Do this
446 * after the client is freed so we know all the client's
447 * transactions have been committed. */
449 filter_update_server_data(obd, filter->fo_rcvd_filp,
450 filter->fo_fsd, need_sync);
451 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
453 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
454 "zero out client %s at idx %u/%llu in %s %ssync rc %d\n",
455 fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
456 LAST_RCVD, need_sync ? "" : "a", rc);
459 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
460 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
467 OBD_FREE_PTR(fed->fed_lcd);
473 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
474 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
475 struct filter_mod_data *fmd)
477 LASSERT_SPIN_LOCKED(&fed->fed_lock);
478 if (--fmd->fmd_refcount == 0) {
479 /* XXX when we have persistent reservations and the handle
480 * is stored herein we need to drop it here. */
481 fed->fed_mod_count--;
482 list_del(&fmd->fmd_list);
483 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
487 /* drop fmd reference, free it if last ref */
488 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
490 struct filter_export_data *fed;
495 fed = &exp->exp_filter_data;
496 spin_lock(&fed->fed_lock);
497 filter_fmd_put_nolock(fed, fmd); /* caller reference */
498 spin_unlock(&fed->fed_lock);
501 /* expire entries from the end of the list if there are too many
502 * or they are too old */
503 static void filter_fmd_expire_nolock(struct filter_obd *filter,
504 struct filter_export_data *fed,
505 struct filter_mod_data *keep)
507 struct filter_mod_data *fmd, *tmp;
509 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
513 if (time_before(jiffies, fmd->fmd_expire) &&
514 fed->fed_mod_count < filter->fo_fmd_max_num)
517 list_del_init(&fmd->fmd_list);
518 filter_fmd_put_nolock(fed, fmd); /* list reference */
522 void filter_fmd_expire(struct obd_export *exp)
524 spin_lock(&exp->exp_filter_data.fed_lock);
525 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
526 &exp->exp_filter_data, NULL);
527 spin_unlock(&exp->exp_filter_data.fed_lock);
530 /* find specified objid, group in export fmd list.
531 * caller must hold fed_lock and take fmd reference itself */
532 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
533 struct filter_export_data *fed,
534 obd_id objid, obd_gr group)
536 struct filter_mod_data *found = NULL, *fmd;
538 LASSERT_SPIN_LOCKED(&fed->fed_lock);
540 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
541 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
543 list_del(&fmd->fmd_list);
544 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
545 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
550 filter_fmd_expire_nolock(filter, fed, found);
555 /* Find fmd based on objid and group, or return NULL if not found. */
556 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
557 obd_id objid, obd_gr group)
559 struct filter_mod_data *fmd;
561 spin_lock(&exp->exp_filter_data.fed_lock);
562 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
563 &exp->exp_filter_data, objid, group);
565 fmd->fmd_refcount++; /* caller reference */
566 spin_unlock(&exp->exp_filter_data.fed_lock);
571 /* Find fmd based on objid and group, or create a new one if none is found.
572 * It is possible for this function to return NULL under memory pressure,
573 * or if objid = 0 is passed (which will only cause old entries to expire).
574 * Currently this is not fatal because any fmd state is transient and
575 * may also be freed when it gets sufficiently old. */
576 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
577 obd_id objid, obd_gr group)
579 struct filter_export_data *fed = &exp->exp_filter_data;
580 struct filter_mod_data *found = NULL, *fmd_new = NULL;
582 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
584 spin_lock(&fed->fed_lock);
585 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
588 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
589 fmd_new->fmd_id = objid;
590 fmd_new->fmd_gr = group;
591 fmd_new->fmd_refcount++; /* list reference */
593 fed->fed_mod_count++;
595 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
599 found->fmd_refcount++; /* caller reference */
600 found->fmd_expire = jiffies +
601 exp->exp_obd->u.filter.fo_fmd_max_age;
604 spin_unlock(&fed->fed_lock);
610 /* drop fmd list reference so it will disappear when last reference is put.
611 * This isn't so critical because it would in fact only affect the one client
612 * that is doing the unlink and at worst we have an stale entry referencing
613 * an object that should never be used again. */
614 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
616 struct filter_mod_data *found = NULL;
618 spin_lock(&exp->exp_filter_data.fed_lock);
619 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
621 list_del_init(&found->fmd_list);
622 filter_fmd_put_nolock(&exp->exp_filter_data, found);
624 spin_unlock(&exp->exp_filter_data.fed_lock);
627 #define filter_fmd_drop(exp, objid, group)
630 /* remove all entries from fmd list */
631 static void filter_fmd_cleanup(struct obd_export *exp)
633 struct filter_export_data *fed = &exp->exp_filter_data;
634 struct filter_mod_data *fmd = NULL, *tmp;
636 spin_lock(&fed->fed_lock);
637 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
638 list_del_init(&fmd->fmd_list);
639 filter_fmd_put_nolock(fed, fmd);
641 spin_unlock(&fed->fed_lock);
644 static int filter_init_export(struct obd_export *exp)
646 spin_lock_init(&exp->exp_filter_data.fed_lock);
647 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
649 spin_lock(&exp->exp_lock);
650 exp->exp_connecting = 1;
651 spin_unlock(&exp->exp_lock);
653 return ldlm_init_export(exp);
656 static int filter_free_server_data(struct filter_obd *filter)
658 OBD_FREE_PTR(filter->fo_fsd);
659 filter->fo_fsd = NULL;
660 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
661 filter->fo_last_rcvd_slots = NULL;
665 /* assumes caller is already in kernel ctxt */
666 int filter_update_server_data(struct obd_device *obd, struct file *filp,
667 struct lr_server_data *fsd, int force_sync)
673 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
674 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
675 le64_to_cpu(fsd->lsd_last_transno));
676 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
677 le64_to_cpu(fsd->lsd_mount_count));
679 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
681 CERROR("error writing lr_server_data: rc = %d\n", rc);
686 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
689 struct filter_obd *filter = &obd->u.filter;
695 if (filter->fo_last_objid_files[group] == NULL) {
696 CERROR("Object group "LPU64" not fully setup; not updating "
697 "last_objid\n", group);
701 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
702 obd->obd_name, group, filter->fo_last_objids[group]);
704 tmp = cpu_to_le64(filter->fo_last_objids[group]);
705 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
706 &tmp, sizeof(tmp), &off, force_sync);
708 CERROR("error writing group "LPU64" last objid: rc = %d\n",
712 extern int ost_handle(struct ptlrpc_request *req);
713 /* assumes caller has already in kernel ctxt */
714 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
716 struct filter_obd *filter = &obd->u.filter;
717 struct lr_server_data *fsd;
718 struct lsd_client_data *lcd = NULL;
719 struct inode *inode = filp->f_dentry->d_inode;
720 unsigned long last_rcvd_size = i_size_read(inode);
727 /* ensure padding in the struct is the correct size */
728 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
729 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
730 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
731 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
733 OBD_ALLOC(fsd, sizeof(*fsd));
736 filter->fo_fsd = fsd;
738 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
739 if (filter->fo_last_rcvd_slots == NULL) {
740 OBD_FREE(fsd, sizeof(*fsd));
744 if (last_rcvd_size == 0) {
745 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
747 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
748 fsd->lsd_last_transno = 0;
749 mount_count = fsd->lsd_mount_count = 0;
750 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
751 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
752 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
753 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
754 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
755 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
757 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
759 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
763 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
764 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
765 "using the wrong disk %s. Were the "
766 "/dev/ assignments rearranged?\n",
767 obd->obd_uuid.uuid, fsd->lsd_uuid);
768 GOTO(err_fsd, rc = -EINVAL);
770 mount_count = le64_to_cpu(fsd->lsd_mount_count);
771 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
773 /* Assume old last_rcvd format unless I_C_LR is set */
774 if (!(fsd->lsd_feature_incompat &
775 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
776 fsd->lsd_last_transno = fsd->lsd_compat14;
780 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
781 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
782 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
783 ~FILTER_INCOMPAT_SUPP);
784 GOTO(err_fsd, rc = -EINVAL);
786 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
787 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
788 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
789 ~FILTER_ROCOMPAT_SUPP);
790 /* Do something like remount filesystem read-only */
791 GOTO(err_fsd, rc = -EINVAL);
794 start_epoch = le32_to_cpu(fsd->lsd_start_epoch);
796 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
797 obd->obd_name, start_epoch);
798 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
799 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
800 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
801 obd->obd_name, mount_count + 1);
802 CDEBUG(D_INODE, "%s: server data size: %u\n",
803 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
804 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
805 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
806 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
807 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
808 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
809 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
810 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
811 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
812 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
813 le16_to_cpu(fsd->lsd_client_size));
815 if (!obd->obd_replayable) {
816 CWARN("%s: recovery support OFF\n", obd->obd_name);
820 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
821 off < last_rcvd_size; cl_idx++) {
823 struct obd_export *exp;
824 struct filter_export_data *fed;
829 GOTO(err_client, rc = -ENOMEM);
832 /* Don't assume off is incremented properly by
833 * fsfilt_read_record(), in case sizeof(*lcd)
834 * isn't the same as fsd->lsd_client_size. */
835 off = le32_to_cpu(fsd->lsd_client_start) +
836 cl_idx * le16_to_cpu(fsd->lsd_client_size);
837 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
839 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
840 LAST_RCVD, cl_idx, off, rc);
841 break; /* read error shouldn't cause startup to fail */
844 if (lcd->lcd_uuid[0] == '\0') {
845 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
850 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
852 /* These exports are cleaned up by filter_disconnect(), so they
853 * need to be set up like real exports as filter_connect() does.
855 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
857 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
858 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
859 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
861 if (PTR_ERR(exp) == -EALREADY) {
862 /* export already exists, zero out this one */
863 CERROR("Zeroing out duplicate export due to "
865 lcd->lcd_uuid[0] = '\0';
867 GOTO(err_client, rc = PTR_ERR(exp));
870 fed = &exp->exp_filter_data;
872 fed->fed_group = 0; /* will be assigned at connect */
873 filter_export_stats_init(obd, exp, NULL);
874 rc = filter_client_add(obd, exp, cl_idx);
875 /* can't fail for existing client */
876 LASSERTF(rc == 0, "rc = %d\n", rc);
878 /* VBR: set export last committed */
879 exp->exp_last_committed = last_rcvd;
880 spin_lock(&exp->exp_lock);
881 exp->exp_connecting = 0;
882 exp->exp_in_recovery = 0;
883 spin_unlock(&exp->exp_lock);
884 spin_lock_bh(&obd->obd_processing_task_lock);
885 obd->obd_max_recoverable_clients++;
886 spin_unlock_bh(&obd->obd_processing_task_lock);
888 class_export_put(exp);
891 /* Need to check last_rcvd even for duplicated exports. */
892 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
895 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
896 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
902 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
904 target_recovery_init(&filter->fo_lut, ost_handle);
907 filter->fo_mount_count = mount_count + 1;
908 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
910 /* save it, so mount count and last_transno is current */
911 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
913 GOTO(err_client, rc);
918 target_recovery_fini(obd);
920 filter_free_server_data(filter);
924 static int filter_cleanup_groups(struct obd_device *obd)
926 struct filter_obd *filter = &obd->u.filter;
928 struct dentry *dentry;
932 if (filter->fo_dentry_O_groups != NULL) {
933 for (i = 0; i < filter->fo_group_count; i++) {
934 dentry = filter->fo_dentry_O_groups[i];
938 OBD_FREE(filter->fo_dentry_O_groups,
939 filter->fo_group_count *
940 sizeof(*filter->fo_dentry_O_groups));
941 filter->fo_dentry_O_groups = NULL;
943 if (filter->fo_last_objid_files != NULL) {
944 for (i = 0; i < filter->fo_group_count; i++) {
945 filp = filter->fo_last_objid_files[i];
949 OBD_FREE(filter->fo_last_objid_files,
950 filter->fo_group_count *
951 sizeof(*filter->fo_last_objid_files));
952 filter->fo_last_objid_files = NULL;
954 if (filter->fo_dentry_O_sub != NULL) {
955 for (i = 0; i < filter->fo_group_count; i++) {
956 for (j = 0; j < filter->fo_subdir_count; j++) {
957 dentry = filter->fo_dentry_O_sub[i].dentry[j];
962 OBD_FREE(filter->fo_dentry_O_sub,
963 filter->fo_group_count *
964 sizeof(*filter->fo_dentry_O_sub));
965 filter->fo_dentry_O_sub = NULL;
967 if (filter->fo_last_objids != NULL) {
968 OBD_FREE(filter->fo_last_objids,
969 filter->fo_group_count *
970 sizeof(*filter->fo_last_objids));
971 filter->fo_last_objids = NULL;
973 if (filter->fo_dentry_O != NULL) {
974 f_dput(filter->fo_dentry_O);
975 filter->fo_dentry_O = NULL;
980 static int filter_update_last_group(struct obd_device *obd, int group)
982 struct filter_obd *filter = &obd->u.filter;
983 struct file *filp = NULL;
984 int last_group = 0, rc;
988 if (group <= filter->fo_committed_group)
991 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
995 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
999 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1001 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1004 LASSERTF(off == 0 || CHECK_MDS_GROUP(last_group),
1005 "off = %llu and last_group = %d\n", off, last_group);
1007 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1008 obd->obd_name, last_group, group);
1012 /* must be sync: bXXXX */
1013 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1015 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1019 filter->fo_committed_group = group;
1022 filp_close(filp, 0);
1026 static int filter_read_group_internal(struct obd_device *obd, int group,
1029 struct filter_obd *filter = &obd->u.filter;
1030 __u64 *new_objids = NULL;
1031 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1032 struct dentry **new_groups = NULL;
1033 struct file **new_files = NULL;
1034 struct dentry *dentry;
1036 int old_count = filter->fo_group_count, rc, stage = 0, i;
1040 int len = group + 1;
1042 snprintf(name, 24, "%d", group);
1046 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1048 if (IS_ERR(dentry)) {
1049 CERROR("Cannot lookup expected object group %d: %ld\n",
1050 group, PTR_ERR(dentry));
1051 RETURN(PTR_ERR(dentry));
1054 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1056 if (IS_ERR(dentry)) {
1057 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1059 RETURN(PTR_ERR(dentry));
1064 snprintf(name, 24, "O/%d/LAST_ID", group);
1066 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1068 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1069 GOTO(cleanup, rc = PTR_ERR(filp));
1073 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1075 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1079 if (filter->fo_subdir_count) {
1080 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1081 if (tmp_subdirs == NULL)
1082 GOTO(cleanup, rc = -ENOMEM);
1085 for (i = 0; i < filter->fo_subdir_count; i++) {
1087 snprintf(dir, sizeof(dir), "d%u", i);
1089 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1092 if (IS_ERR(tmp_subdirs->dentry[i])) {
1093 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1094 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1099 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1100 tmp_subdirs->dentry[i]);
1104 /* 'group' is an index; we need an array of length 'group + 1' */
1105 if (group + 1 > old_count) {
1106 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1107 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1108 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1109 OBD_ALLOC(new_files, len * sizeof(*new_files));
1111 if (new_objids == NULL || new_subdirs == NULL ||
1112 new_groups == NULL || new_files == NULL)
1113 GOTO(cleanup, rc = -ENOMEM);
1116 memcpy(new_objids, filter->fo_last_objids,
1117 old_count * sizeof(*new_objids));
1118 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1119 old_count * sizeof(*new_subdirs));
1120 memcpy(new_groups, filter->fo_dentry_O_groups,
1121 old_count * sizeof(*new_groups));
1122 memcpy(new_files, filter->fo_last_objid_files,
1123 old_count * sizeof(*new_files));
1125 OBD_FREE(filter->fo_last_objids,
1126 old_count * sizeof(*new_objids));
1127 OBD_FREE(filter->fo_dentry_O_sub,
1128 old_count * sizeof(*new_subdirs));
1129 OBD_FREE(filter->fo_dentry_O_groups,
1130 old_count * sizeof(*new_groups));
1131 OBD_FREE(filter->fo_last_objid_files,
1132 old_count * sizeof(*new_files));
1134 filter->fo_last_objids = new_objids;
1135 filter->fo_dentry_O_sub = new_subdirs;
1136 filter->fo_dentry_O_groups = new_groups;
1137 filter->fo_last_objid_files = new_files;
1138 filter->fo_group_count = len;
1141 filter->fo_dentry_O_groups[group] = dentry;
1142 filter->fo_last_objid_files[group] = filp;
1143 if (filter->fo_subdir_count) {
1144 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1145 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1148 filter_update_last_group(obd, group);
1150 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1151 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1152 rc = filter_update_last_objid(obd, group, 1);
1156 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1157 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1158 obd->obd_name, group, last_objid);
1163 if (new_objids != NULL)
1164 OBD_FREE(new_objids, len * sizeof(*new_objids));
1165 if (new_subdirs != NULL)
1166 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1167 if (new_groups != NULL)
1168 OBD_FREE(new_groups, len * sizeof(*new_groups));
1169 if (new_files != NULL)
1170 OBD_FREE(new_files, len * sizeof(*new_files));
1172 if (filter->fo_subdir_count) {
1173 for (i = 0; i < filter->fo_subdir_count; i++) {
1174 if (tmp_subdirs->dentry[i] != NULL)
1175 dput(tmp_subdirs->dentry[i]);
1177 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1180 filp_close(filp, 0);
1187 static int filter_read_groups(struct obd_device *obd, int last_group,
1190 struct filter_obd *filter = &obd->u.filter;
1191 int old_count, group, rc = 0;
1193 down(&filter->fo_init_lock);
1194 old_count = filter->fo_group_count;
1195 for (group = old_count; group <= last_group; group++) {
1197 rc = filter_read_group_internal(obd, group, create);
1201 up(&filter->fo_init_lock);
1205 /* FIXME: object groups */
1206 static int filter_prep_groups(struct obd_device *obd)
1208 struct filter_obd *filter = &obd->u.filter;
1209 struct dentry *dentry, *O_dentry;
1211 int last_group, rc = 0, cleanup_phase = 0;
1215 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1217 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1218 if (IS_ERR(O_dentry)) {
1219 rc = PTR_ERR(O_dentry);
1220 CERROR("cannot open/create O: rc = %d\n", rc);
1223 filter->fo_dentry_O = O_dentry;
1224 cleanup_phase = 1; /* O_dentry */
1226 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1227 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1228 * some time post 1.0 when all old-style OSTs have converted along
1229 * with the init_objid hack. */
1230 dentry = ll_lookup_one_len("R", O_dentry, 1);
1232 GOTO(cleanup, rc = PTR_ERR(dentry));
1233 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1234 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1237 CWARN("converting OST to new object layout\n");
1238 if (IS_ERR(O0_dentry)) {
1239 rc = PTR_ERR(O0_dentry);
1240 CERROR("error looking up O/0: rc %d\n", rc);
1241 GOTO(cleanup_R, rc);
1244 if (O0_dentry->d_inode) {
1245 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1246 GOTO(cleanup_O0, rc = -EEXIST);
1249 LOCK_INODE_MUTEX(O_dentry->d_inode);
1250 rc = ll_vfs_rename(O_dentry->d_inode, dentry, filter->fo_vfsmnt,
1251 O_dentry->d_inode, O0_dentry,
1253 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1256 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1257 GOTO(cleanup_O0, rc);
1259 filter->fo_fsd->lsd_feature_incompat |=
1260 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1261 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1263 GOTO(cleanup_O0, rc);
1275 cleanup_phase = 2; /* groups */
1277 /* we have to initialize all groups before first connections from
1278 * clients because they may send create/destroy for any group -bzzz */
1279 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1281 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1282 GOTO(cleanup, rc = PTR_ERR(filp));
1284 cleanup_phase = 3; /* filp */
1286 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1288 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1292 last_group = FILTER_MIN_GROUPS;
1294 LASSERT_MDS_GROUP(last_group);
1297 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1298 FILTER_MIN_GROUPS, last_group);
1299 filter->fo_committed_group = last_group;
1300 rc = filter_read_groups(obd, last_group, 1);
1304 filp_close(filp, 0);
1308 switch (cleanup_phase) {
1310 filp_close(filp, 0);
1312 filter_cleanup_groups(obd);
1314 f_dput(filter->fo_dentry_O);
1315 filter->fo_dentry_O = NULL;
1323 /* setup the object store with correct subdirectories */
1324 static int filter_prep(struct obd_device *obd)
1326 struct lvfs_run_ctxt saved;
1327 struct filter_obd *filter = &obd->u.filter;
1329 struct inode *inode;
1333 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1334 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1335 if (!file || IS_ERR(file)) {
1337 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1341 filter->fo_rcvd_filp = file;
1342 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1343 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1344 file->f_dentry->d_inode->i_mode);
1345 GOTO(err_filp, rc = -ENOENT);
1348 inode = file->f_dentry->d_parent->d_inode;
1349 /* We use i_op->unlink directly in filter_vfs_unlink() */
1350 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1351 CERROR("%s: filesystem does not support create/unlink ops\n",
1353 GOTO(err_filp, rc = -EOPNOTSUPP);
1356 /** lu_target has very limited use in filter now */
1357 lut_init(NULL, &filter->fo_lut, obd, NULL);
1359 rc = filter_init_server_data(obd, file);
1361 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1364 /* open and create health check io file*/
1365 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1368 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1370 GOTO(err_server_data, rc);
1372 filter->fo_health_check_filp = file;
1373 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1374 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1375 file->f_dentry->d_inode->i_mode);
1376 GOTO(err_health_check, rc = -ENOENT);
1378 rc = lvfs_check_io_health(obd, file);
1380 GOTO(err_health_check, rc);
1382 rc = filter_prep_groups(obd);
1384 GOTO(err_health_check, rc);
1386 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1391 if (filp_close(filter->fo_health_check_filp, 0))
1392 CERROR("can't close %s after error\n", HEALTH_CHECK);
1393 filter->fo_health_check_filp = NULL;
1395 target_recovery_fini(obd);
1396 filter_free_server_data(filter);
1398 if (filp_close(filter->fo_rcvd_filp, 0))
1399 CERROR("can't close %s after error\n", LAST_RCVD);
1400 filter->fo_rcvd_filp = NULL;
1404 /* cleanup the filter: write last used object id to status file */
1405 static void filter_post(struct obd_device *obd)
1407 struct lvfs_run_ctxt saved;
1408 struct filter_obd *filter = &obd->u.filter;
1411 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1412 * best to start a transaction with h_sync, because we removed this
1415 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1416 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1419 CERROR("error writing server data: rc = %d\n", rc);
1421 for (i = 0; i < filter->fo_group_count; i++) {
1422 rc = filter_update_last_objid(obd, i,
1423 (i == filter->fo_group_count - 1));
1425 CERROR("error writing group %d lastobjid: rc = %d\n",
1429 rc = filp_close(filter->fo_rcvd_filp, 0);
1430 filter->fo_rcvd_filp = NULL;
1432 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1434 rc = filp_close(filter->fo_health_check_filp, 0);
1435 filter->fo_health_check_filp = NULL;
1437 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1439 filter_cleanup_groups(obd);
1440 filter_free_server_data(filter);
1441 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1443 filter_free_capa_keys(filter);
1444 cleanup_capa_hash(filter->fo_capa_hash);
1447 static void filter_set_last_id(struct filter_obd *filter,
1448 obd_id id, obd_gr group)
1450 LASSERT(filter->fo_fsd != NULL);
1451 LASSERT(group <= filter->fo_group_count);
1453 spin_lock(&filter->fo_objidlock);
1454 filter->fo_last_objids[group] = id;
1455 spin_unlock(&filter->fo_objidlock);
1458 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1461 LASSERT(filter->fo_fsd != NULL);
1462 LASSERT(group <= filter->fo_group_count);
1464 /* FIXME: object groups */
1465 spin_lock(&filter->fo_objidlock);
1466 id = filter->fo_last_objids[group];
1467 spin_unlock(&filter->fo_objidlock);
1471 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1473 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1477 /* We never dget the object parent, so DON'T dput it either */
1478 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1480 struct filter_obd *filter = &obd->u.filter;
1481 struct filter_subdirs *subdirs;
1482 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1484 if ((group > FILTER_GROUP_MDS0 && group < FILTER_GROUP_MDS1_N_BASE) ||
1485 filter->fo_subdir_count == 0)
1486 return filter->fo_dentry_O_groups[group];
1488 subdirs = &filter->fo_dentry_O_sub[group];
1489 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1492 /* We never dget the object parent, so DON'T dput it either */
1493 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1496 unsigned long now = jiffies;
1497 struct dentry *dparent = filter_parent(obd, group, objid);
1500 if (IS_ERR(dparent))
1502 if (dparent == NULL)
1503 return ERR_PTR(-ENOENT);
1505 rc = filter_lock_dentry(obd, dparent);
1506 fsfilt_check_slow(obd, now, "parent lock");
1507 return rc ? ERR_PTR(rc) : dparent;
1510 /* We never dget the object parent, so DON'T dput it either */
1511 static void filter_parent_unlock(struct dentry *dparent)
1513 UNLOCK_INODE_MUTEX(dparent->d_inode);
1516 /* How to get files, dentries, inodes from object id's.
1518 * If dir_dentry is passed, the caller has already locked the parent
1519 * appropriately for this operation (normally a write lock). If
1520 * dir_dentry is NULL, we do a read lock while we do the lookup to
1521 * avoid races with create/destroy and such changing the directory
1522 * internal to the filesystem code. */
1523 struct dentry *filter_fid2dentry(struct obd_device *obd,
1524 struct dentry *dir_dentry,
1525 obd_gr group, obd_id id)
1527 struct dentry *dparent = dir_dentry;
1528 struct dentry *dchild;
1533 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1534 obd->u.filter.fo_destroys_in_progress == 0) {
1535 /* don't fail lookups for orphan recovery, it causes
1536 * later LBUGs when objects still exist during precreate */
1537 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1538 RETURN(ERR_PTR(-ENOENT));
1541 CERROR("fatal: invalid object id 0\n");
1542 RETURN(ERR_PTR(-ESTALE));
1545 len = sprintf(name, LPU64, id);
1546 if (dir_dentry == NULL) {
1547 dparent = filter_parent_lock(obd, group, id);
1548 if (IS_ERR(dparent)) {
1549 CERROR("%s: error getting object "LPU64":"LPU64
1550 " parent: rc %ld\n", obd->obd_name,
1551 id, group, PTR_ERR(dparent));
1555 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1556 dparent->d_name.len, dparent->d_name.name, name);
1557 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1558 if (dir_dentry == NULL)
1559 filter_parent_unlock(dparent);
1560 if (IS_ERR(dchild)) {
1561 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1566 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1567 CERROR("%s: got bad object "LPU64" inode %lu\n",
1568 obd->obd_name, id, dchild->d_inode->i_ino);
1570 RETURN(ERR_PTR(-ENOENT));
1573 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1574 name, dchild, atomic_read(&dchild->d_count));
1576 LASSERT(atomic_read(&dchild->d_count) > 0);
1581 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1582 obd_id group, struct lustre_handle *lockh)
1584 int flags = LDLM_AST_DISCARD_DATA, rc;
1585 struct ldlm_res_id res_id;
1586 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1589 osc_build_res_name(objid, group, &res_id);
1590 /* Tell the clients that the object is gone now and that they should
1591 * throw away any cached pages. */
1592 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1593 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1594 ldlm_completion_ast, NULL, NULL, 0, NULL,
1601 static void filter_fini_destroy(struct obd_device *obd,
1602 struct lustre_handle *lockh)
1605 ldlm_lock_decref(lockh, LCK_PW);
1608 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1609 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1610 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1611 * reverse. Caller must take i_sem before starting the transaction and we
1612 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1613 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1614 struct vfsmount *mnt)
1619 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1620 * itself we already hold dir->i_mutex for child create/unlink ops */
1621 LASSERT(dentry->d_inode != NULL);
1622 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1623 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1627 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1628 GOTO(out, rc = -ENOENT);
1630 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1635 GOTO(out, rc = -EPERM);
1637 /* check_sticky() */
1638 if ((dentry->d_inode->i_uid != current->fsuid &&
1639 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1640 IS_IMMUTABLE(dentry->d_inode))
1641 GOTO(out, rc = -EPERM);
1643 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1644 * that was done because of journal_start (which is already done
1645 * here) or some other ordering issue. */
1648 rc = ll_security_inode_unlink(dir, dentry, mnt);
1652 rc = dir->i_op->unlink(dir, dentry);
1654 /* need to drop i_mutex before we lose inode reference */
1655 UNLOCK_INODE_MUTEX(dentry->d_inode);
1662 /* Caller must hold LCK_PW on parent and push us into kernel context.
1663 * Caller must hold child i_mutex, we drop it always.
1664 * Caller is also required to ensure that dchild->d_inode exists. */
1665 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1666 obd_gr group, struct dentry *dparent,
1667 struct dentry *dchild)
1669 struct inode *inode = dchild->d_inode;
1670 struct filter_obd *filter = &obd->u.filter;
1673 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1674 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1675 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1676 (unsigned long)inode->i_nlink,
1677 atomic_read(&inode->i_count));
1680 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1682 CERROR("error unlinking objid %.*s: rc %d\n",
1683 dchild->d_name.len, dchild->d_name.name, rc);
1687 struct filter_intent_args {
1688 struct ldlm_lock **victim;
1693 static enum interval_iter filter_intent_cb(struct interval_node *n,
1696 struct ldlm_interval *node = (struct ldlm_interval *)n;
1697 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1698 __u64 size = arg->size;
1699 struct ldlm_lock **v = arg->victim;
1700 struct ldlm_lock *lck;
1702 /* If the interval is lower than the current file size,
1704 if (interval_high(n) <= size)
1705 return INTERVAL_ITER_STOP;
1707 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1708 /* Don't send glimpse ASTs to liblustre clients.
1709 * They aren't listening for them, and they do
1710 * entirely synchronous I/O anyways. */
1711 if (lck->l_export == NULL ||
1712 lck->l_export->exp_libclient == 1)
1715 if (*arg->liblustre)
1716 *arg->liblustre = 0;
1719 *v = LDLM_LOCK_GET(lck);
1720 } else if ((*v)->l_policy_data.l_extent.start <
1721 lck->l_policy_data.l_extent.start) {
1722 LDLM_LOCK_RELEASE(*v);
1723 *v = LDLM_LOCK_GET(lck);
1726 /* the same policy group - every lock has the
1727 * same extent, so needn't do it any more */
1731 return INTERVAL_ITER_CONT;
1734 static int filter_intent_policy(struct ldlm_namespace *ns,
1735 struct ldlm_lock **lockp, void *req_cookie,
1736 ldlm_mode_t mode, int flags, void *data)
1738 CFS_LIST_HEAD(rpc_list);
1739 struct ptlrpc_request *req = req_cookie;
1740 struct ldlm_lock *lock = *lockp, *l = NULL;
1741 struct ldlm_resource *res = lock->l_resource;
1742 ldlm_processing_policy policy;
1743 struct ost_lvb *res_lvb, *reply_lvb;
1744 struct ldlm_reply *rep;
1746 int idx, rc, tmpflags = 0, only_liblustre = 1;
1747 struct ldlm_interval_tree *tree;
1748 struct filter_intent_args arg;
1749 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1750 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1751 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1754 policy = ldlm_get_processing_policy(res);
1755 LASSERT(policy != NULL);
1756 LASSERT(req != NULL);
1758 rc = lustre_pack_reply(req, 3, repsize, NULL);
1760 RETURN(req->rq_status = rc);
1762 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1763 LASSERT(rep != NULL);
1765 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1766 sizeof(*reply_lvb));
1767 LASSERT(reply_lvb != NULL);
1769 //fixup_handle_for_resent_req(req, lock, &lockh);
1771 /* Call the extent policy function to see if our request can be
1772 * granted, or is blocked.
1773 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1774 * lock, and should not be granted if the lock will be blocked.
1777 LASSERT(ns == res->lr_namespace);
1779 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1780 check_res_locked(res);
1782 /* FIXME: we should change the policy function slightly, to not make
1783 * this list at all, since we just turn around and free it */
1784 while (!list_empty(&rpc_list)) {
1785 struct ldlm_lock *wlock =
1786 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1787 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1788 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1789 lock->l_flags &= ~LDLM_FL_CP_REQD;
1790 list_del_init(&wlock->l_cp_ast);
1791 LDLM_LOCK_RELEASE(wlock);
1794 /* The lock met with no resistance; we're finished. */
1795 if (rc == LDLM_ITER_CONTINUE) {
1796 /* do not grant locks to the liblustre clients: they cannot
1797 * handle ASTs robustly. We need to do this while still
1798 * holding ns_lock to avoid the lock remaining on the res_link
1799 * list (and potentially being added to l_pending_list by an
1800 * AST) when we are going to drop this lock ASAP. */
1801 if (lock->l_export->exp_libclient ||
1802 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1803 ldlm_resource_unlink_lock(lock);
1804 err = ELDLM_LOCK_ABORTED;
1806 err = ELDLM_LOCK_REPLACED;
1812 /* Do not grant any lock, but instead send GL callbacks. The extent
1813 * policy nicely created a list of all PW locks for us. We will choose
1814 * the highest of those which are larger than the size in the LVB, if
1815 * any, and perform a glimpse callback. */
1816 res_lvb = res->lr_lvb_data;
1817 LASSERT(res_lvb != NULL);
1818 *reply_lvb = *res_lvb;
1821 * ->ns_lock guarantees that no new locks are granted, and,
1822 * therefore, that res->lr_lvb_data cannot increase beyond the
1823 * end of already granted lock. As a result, it is safe to
1824 * check against "stale" reply_lvb->lvb_size value without
1827 arg.size = reply_lvb->lvb_size;
1829 arg.liblustre = &only_liblustre;
1830 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1831 tree = &res->lr_itree[idx];
1832 if (tree->lit_mode == LCK_PR)
1835 interval_iterate_reverse(tree->lit_root,
1836 filter_intent_cb, &arg);
1840 /* There were no PW locks beyond the size in the LVB; finished. */
1842 if (only_liblustre) {
1843 /* If we discovered a liblustre client with a PW lock,
1844 * however, the LVB may be out of date! The LVB is
1845 * updated only on glimpse (which we don't do for
1846 * liblustre clients) and cancel (which the client
1847 * obviously has not yet done). So if it has written
1848 * data but kept the lock, the LVB is stale and needs
1849 * to be updated from disk.
1851 * Of course, this will all disappear when we switch to
1852 * taking liblustre locks on the OST. */
1853 ldlm_res_lvbo_update(res, NULL, 0, 1);
1855 RETURN(ELDLM_LOCK_ABORTED);
1859 * This check is for lock taken in filter_prepare_destroy() that does
1860 * not have l_glimpse_ast set. So the logic is: if there is a lock
1861 * with no l_glimpse_ast set, this object is being destroyed already.
1863 * Hence, if you are grabbing DLM locks on the server, always set
1864 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1866 if (l->l_glimpse_ast == NULL) {
1867 /* We are racing with unlink(); just return -ENOENT */
1868 rep->lock_policy_res1 = -ENOENT;
1872 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1873 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1874 /* Update the LVB from disk if the AST failed (this is a legal race) */
1876 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1877 * sending ast is not handled. This can result in lost client writes.
1880 ldlm_res_lvbo_update(res, NULL, 0, 1);
1883 *reply_lvb = *res_lvb;
1887 LDLM_LOCK_RELEASE(l);
1889 RETURN(ELDLM_LOCK_ABORTED);
1893 * per-obd_device iobuf pool.
1895 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1896 * allocations in write-path has to be minimized (see bug 5137).
1898 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1899 * OST threads (see ost_thread_{init,done}()).
1901 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1902 * at the OST layer there are only (potentially) multiple obd_device of type
1903 * unknown at the time of OST thread creation.
1905 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1906 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1907 * it's very own iobuf.
1911 * filter_kiobuf_pool_init()
1913 * filter_kiobuf_pool_done()
1915 * filter_iobuf_get()
1917 * operate on this array. They are "generic" in a sense that they don't depend
1918 * on actual type of iobuf's (the latter depending on Linux kernel version).
1922 * destroy pool created by filter_iobuf_pool_init
1924 static void filter_iobuf_pool_done(struct filter_obd *filter)
1926 struct filter_iobuf **pool;
1931 pool = filter->fo_iobuf_pool;
1933 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1934 if (pool[i] != NULL)
1935 filter_free_iobuf(pool[i]);
1937 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1938 filter->fo_iobuf_pool = NULL;
1943 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1945 struct filter_obd *filter = &obd->u.filter;
1946 struct sptlrpc_rule_set tmp_rset;
1949 sptlrpc_rule_set_init(&tmp_rset);
1950 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1952 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1957 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1959 write_lock(&filter->fo_sptlrpc_lock);
1960 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1961 filter->fo_sptlrpc_rset = tmp_rset;
1962 write_unlock(&filter->fo_sptlrpc_lock);
1968 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1970 static int filter_iobuf_pool_init(struct filter_obd *filter)
1977 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1979 if (filter->fo_iobuf_pool == NULL)
1982 filter->fo_iobuf_count = OSS_THREADS_MAX;
1987 /* Return iobuf allocated for @thread_id. We don't know in advance how
1988 * many threads there will be so we allocate a large empty array and only
1989 * fill in those slots that are actually in use.
1990 * If we haven't allocated a pool entry for this thread before, do so now. */
1991 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1993 int thread_id = (oti && oti->oti_thread) ?
1994 oti->oti_thread->t_id : -1;
1995 struct filter_iobuf *pool = NULL;
1996 struct filter_iobuf **pool_place = NULL;
1998 if (thread_id >= 0) {
1999 LASSERT(thread_id < filter->fo_iobuf_count);
2000 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
2003 if (unlikely(pool == NULL)) {
2004 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
2005 PTLRPC_MAX_BRW_PAGES);
2006 if (pool_place != NULL)
2013 /* mount the file system (secretly). lustre_cfg parameters are:
2016 * 3 = flags: failover=f, failout=n
2019 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
2022 struct filter_obd *filter = &obd->u.filter;
2023 struct vfsmount *mnt;
2024 struct lustre_mount_info *lmi;
2025 struct obd_uuid uuid;
2033 if (lcfg->lcfg_bufcount < 3 ||
2034 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
2035 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
2038 lmi = server_get_mount(obd->obd_name);
2040 /* We already mounted in lustre_fill_super.
2041 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2042 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2044 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2046 /* old path - used by lctl */
2047 CERROR("Using old MDS mount method\n");
2048 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2049 MS_NOATIME|MS_NODIRATIME,
2050 lustre_cfg_string(lcfg, 1), option);
2053 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2054 lustre_cfg_string(lcfg, 1), rc);
2058 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2060 if (IS_ERR(obd->obd_fsops))
2061 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2063 rc = filter_iobuf_pool_init(filter);
2067 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2068 CERROR("%s: Underlying device is marked as read-only. "
2069 "Setup failed\n", obd->obd_name);
2070 GOTO(err_ops, rc = -EROFS);
2073 /* failover is the default */
2074 obd->obd_replayable = 1;
2076 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2077 str = lustre_cfg_string(lcfg, 3);
2078 if (strchr(str, 'n')) {
2079 CWARN("%s: recovery disabled\n", obd->obd_name);
2080 obd->obd_replayable = 0;
2084 filter->fo_vfsmnt = mnt;
2085 obd->u.obt.obt_sb = mnt->mnt_sb;
2086 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2087 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2089 fsfilt_setup(obd, obd->u.obt.obt_sb);
2091 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2092 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2093 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2094 obd->obd_lvfs_ctxt.fs = get_ds();
2095 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2097 init_mutex(&filter->fo_init_lock);
2098 filter->fo_committed_group = 0;
2100 rc = filter_prep(obd);
2104 filter->fo_destroys_in_progress = 0;
2105 for (i = 0; i < 32; i++)
2106 sema_init(&filter->fo_create_locks[i], 1);
2108 spin_lock_init(&filter->fo_translock);
2109 spin_lock_init(&filter->fo_objidlock);
2110 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2111 sema_init(&filter->fo_alloc_lock, 1);
2112 init_brw_stats(&filter->fo_filter_stats);
2113 filter->fo_read_cache = 1; /* enable read-only cache by default */
2114 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2115 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2116 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2117 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2119 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2120 spin_lock_init(&filter->fo_llog_list_lock);
2122 filter->fo_fl_oss_capa = 1;
2124 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2125 filter->fo_capa_hash = init_capa_hash();
2126 if (filter->fo_capa_hash == NULL)
2127 GOTO(err_post, rc = -ENOMEM);
2129 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2130 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2131 LDLM_NAMESPACE_GREEDY);
2132 if (obd->obd_namespace == NULL)
2133 GOTO(err_post, rc = -ENOMEM);
2134 obd->obd_namespace->ns_lvbp = obd;
2135 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2136 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2138 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2139 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2141 rc = obd_llog_init(obd, &obd->obd_olg, obd, 1, NULL, NULL);
2143 CERROR("failed to setup llogging subsystems\n");
2147 rwlock_init(&filter->fo_sptlrpc_lock);
2148 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2149 /* do this after llog being initialized */
2150 filter_adapt_sptlrpc_conf(obd, 1);
2152 rc = lquota_setup(filter_quota_interface_ref, obd);
2156 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2157 if (q->max_sectors < q->max_hw_sectors &&
2158 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2159 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2160 "for larger I/O requests: max_sectors = %u "
2161 "could be up to max_hw_sectors=%u\n",
2162 obd->obd_name, mnt->mnt_sb->s_id,
2163 q->max_sectors, q->max_hw_sectors);
2165 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2166 if (uuid_ptr != NULL) {
2167 class_uuid_unparse(uuid_ptr, &uuid);
2173 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2175 if (obd->obd_recovering) {
2176 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2177 "recovery for at least %d:%.02d, or until %d "
2178 "client%s reconnect%s.\n",
2179 obd->obd_name, lustre_cfg_string(lcfg, 1),
2180 label ?: "", label ? "/" : "", str,
2181 obd->obd_recovery_timeout / 60,
2182 obd->obd_recovery_timeout % 60,
2183 obd->obd_max_recoverable_clients,
2184 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2185 (obd->obd_max_recoverable_clients == 1) ? "s":"");
2187 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2188 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2189 label ?: "", label ? "/" : "", str,
2190 obd->obd_replayable ? "enabled" : "disabled");
2198 fsfilt_put_ops(obd->obd_fsops);
2199 filter_iobuf_pool_done(filter);
2201 server_put_mount(obd->obd_name, mnt);
2202 obd->u.obt.obt_sb = 0;
2206 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2208 struct lprocfs_static_vars lvars;
2213 CLASSERT(offsetof(struct obd_device, u.obt) ==
2214 offsetof(struct obd_device, u.filter.fo_obt));
2216 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2219 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2220 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2223 addr = (unsigned long)cfs_page_address(page);
2224 clear_page((void *)addr);
2226 /* lprocfs must be setup before the filter so state can be safely added
2227 * to /proc incrementally as the filter is setup */
2228 lprocfs_filter_init_vars(&lvars);
2229 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2230 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2231 /* Init obdfilter private stats here */
2232 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2233 LPROCFS_CNTR_AVGMINMAX,
2234 "read_bytes", "bytes");
2235 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2236 LPROCFS_CNTR_AVGMINMAX,
2237 "write_bytes", "bytes");
2238 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2239 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2240 "get_page", "usec");
2241 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2242 LPROCFS_CNTR_AVGMINMAX,
2243 "get_page_failures", "num");
2244 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2245 LPROCFS_CNTR_AVGMINMAX,
2246 "cache_access", "pages");
2247 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2248 LPROCFS_CNTR_AVGMINMAX,
2249 "cache_hit", "pages");
2250 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2251 LPROCFS_CNTR_AVGMINMAX,
2252 "cache_miss", "pages");
2254 lproc_filter_attach_seqstat(obd);
2255 obd->obd_proc_exports_entry = lprocfs_register("exports",
2256 obd->obd_proc_entry,
2258 if (IS_ERR(obd->obd_proc_exports_entry)) {
2259 rc = PTR_ERR(obd->obd_proc_exports_entry);
2260 CERROR("error %d setting up lprocfs for %s\n",
2262 obd->obd_proc_exports_entry = NULL;
2265 if (obd->obd_proc_exports_entry)
2266 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2267 lprocfs_nid_stats_clear_read,
2268 lprocfs_nid_stats_clear_write, obd, NULL);
2270 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2271 LUSTRE_CFG_BUFLEN(lcfg, 4));
2272 rc = filter_common_setup(obd, lcfg, (void *)addr);
2273 OBD_PAGE_FREE(page);
2276 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2277 lprocfs_free_per_client_stats(obd);
2278 lprocfs_free_obd_stats(obd);
2279 lprocfs_obd_cleanup(obd);
2285 static struct llog_operations filter_mds_ost_repl_logops;
2287 static struct llog_operations filter_size_orig_logops = {
2288 .lop_setup = llog_obd_origin_setup,
2289 .lop_cleanup = llog_obd_origin_cleanup,
2290 .lop_add = llog_obd_origin_add
2293 static int filter_olg_fini(struct obd_llog_group *olg)
2295 struct llog_ctxt *ctxt;
2296 int rc = 0, rc2 = 0;
2299 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2301 rc = llog_cleanup(ctxt);
2303 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2305 rc2 = llog_cleanup(ctxt);
2310 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2312 rc2 = llog_cleanup(ctxt);
2321 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2322 struct obd_device *tgt)
2327 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2328 &filter_mds_ost_repl_logops);
2332 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2333 &filter_size_orig_logops);
2339 filter_olg_fini(olg);
2344 * Init the default olg, which is embeded in the obd_device, for filter.
2347 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2348 struct obd_device *tgt)
2350 struct filter_obd *filter = &obd->u.filter;
2351 struct llog_ctxt *ctxt;
2355 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2356 if (!filter->fo_lcm)
2359 filter_mds_ost_repl_logops = llog_client_ops;
2360 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2361 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2362 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2364 rc = filter_olg_init(obd, olg, tgt);
2366 GOTO(cleanup_lcm, rc);
2368 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2371 GOTO(cleanup_olg, rc);
2373 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2375 CERROR("Can't get ctxt for %p:%x\n", olg,
2376 LLOG_MDS_OST_REPL_CTXT);
2377 GOTO(cleanup_olg, rc = -ENODEV);
2379 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2380 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2381 llog_ctxt_put(ctxt);
2385 filter_olg_fini(olg);
2387 llog_recov_thread_fini(filter->fo_lcm, 1);
2388 filter->fo_lcm = NULL;
2393 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2394 struct obd_device *tgt, int count, struct llog_catid *catid,
2395 struct obd_uuid *uuid)
2397 struct filter_obd *filter = &obd->u.filter;
2398 struct llog_ctxt *ctxt;
2402 LASSERT(olg != NULL);
2403 if (olg == &obd->obd_olg)
2404 return filter_default_olg_init(obd, olg, tgt);
2406 LASSERT(filter->fo_lcm != NULL);
2407 rc = filter_olg_init(obd, olg, tgt);
2410 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2412 CERROR("Can't get ctxt for %p:%x\n", olg,
2413 LLOG_MDS_OST_REPL_CTXT);
2414 filter_olg_fini(olg);
2417 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2418 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2419 llog_ctxt_put(ctxt);
2423 static int filter_llog_finish(struct obd_device *obd, int count)
2425 struct filter_obd *filter = &obd->u.filter;
2426 struct llog_ctxt *ctxt;
2429 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2432 * Make sure that no cached llcds left in recov_thread.
2433 * We actually do sync in disconnect time, but disconnect
2434 * may not come being marked rq_no_resend = 1.
2436 llog_sync(ctxt, NULL);
2439 * Balance class_import_get() in llog_receptor_accept().
2440 * This is safe to do, as llog is already synchronized
2441 * and its import may go.
2443 mutex_down(&ctxt->loc_sem);
2444 if (ctxt->loc_imp) {
2445 class_import_put(ctxt->loc_imp);
2446 ctxt->loc_imp = NULL;
2448 mutex_up(&ctxt->loc_sem);
2449 llog_ctxt_put(ctxt);
2452 if (filter->fo_lcm) {
2453 mutex_down(&ctxt->loc_sem);
2454 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2455 filter->fo_lcm = NULL;
2456 mutex_up(&ctxt->loc_sem);
2458 RETURN(filter_olg_fini(&obd->obd_olg));
2462 * Find the group llog according to group index in the llog group list.
2464 static struct obd_llog_group *
2465 filter_find_olg_internal(struct filter_obd *filter, int group)
2467 struct obd_llog_group *olg;
2469 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2470 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2471 if (olg->olg_group == group)
2478 * Find the group llog according to group index on the filter
2480 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2482 struct obd_llog_group *olg = NULL;
2483 struct filter_obd *filter;
2485 filter = &obd->u.filter;
2487 if (group == FILTER_GROUP_LLOG)
2488 RETURN(&obd->obd_olg);
2490 spin_lock(&filter->fo_llog_list_lock);
2491 olg = filter_find_olg_internal(filter, group);
2492 spin_unlock(&filter->fo_llog_list_lock);
2497 * Find the llog_group of the filter according to the group. If it can not
2498 * find, create the llog_group, which only happens when mds is being synced
2501 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2503 struct obd_llog_group *olg = NULL;
2504 struct filter_obd *filter;
2507 filter = &obd->u.filter;
2509 if (group == FILTER_GROUP_LLOG)
2510 RETURN(&obd->obd_olg);
2512 spin_lock(&filter->fo_llog_list_lock);
2513 olg = filter_find_olg_internal(filter, group);
2515 if (olg->olg_initializing) {
2516 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2518 GOTO(out_unlock, olg);
2523 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2525 llog_group_init(olg, group);
2526 list_add(&olg->olg_list, &filter->fo_llog_list);
2527 olg->olg_initializing = 1;
2528 spin_unlock(&filter->fo_llog_list_lock);
2530 rc = llog_cat_initialize(obd, olg, 1, NULL);
2532 spin_lock(&filter->fo_llog_list_lock);
2533 list_del(&olg->olg_list);
2534 spin_unlock(&filter->fo_llog_list_lock);
2536 GOTO(out, olg = ERR_PTR(-ENOMEM));
2538 spin_lock(&filter->fo_llog_list_lock);
2539 olg->olg_initializing = 0;
2540 spin_unlock(&filter->fo_llog_list_lock);
2541 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2542 obd->obd_name, group, olg);
2547 spin_unlock(&filter->fo_llog_list_lock);
2551 static int filter_llog_connect(struct obd_export *exp,
2552 struct llogd_conn_body *body)
2554 struct obd_device *obd = exp->exp_obd;
2555 struct llog_ctxt *ctxt;
2556 struct obd_llog_group *olg;
2560 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2561 obd->obd_name, body->lgdc_logid.lgl_oid,
2562 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2564 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2566 CERROR(" %s: can not find olg of group %d\n",
2567 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2570 llog_group_set_export(olg, exp);
2572 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2573 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2574 body->lgdc_ctxt_idx);
2576 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2577 obd->obd_name, body->lgdc_logid.lgl_oid,
2578 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2580 rc = llog_connect(ctxt, &body->lgdc_logid,
2581 &body->lgdc_gen, NULL);
2582 llog_ctxt_put(ctxt);
2584 CERROR("failed to connect rc %d idx %d\n", rc,
2585 body->lgdc_ctxt_idx);
2590 static int filter_llog_preclean(struct obd_device *obd)
2592 struct obd_llog_group *olg, *tmp;
2593 struct filter_obd *filter;
2594 struct list_head remove_list;
2598 rc = obd_llog_finish(obd, 0);
2600 CERROR("failed to cleanup llogging subsystem\n");
2602 filter = &obd->u.filter;
2603 CFS_INIT_LIST_HEAD(&remove_list);
2605 spin_lock(&filter->fo_llog_list_lock);
2606 while (!list_empty(&filter->fo_llog_list)) {
2607 olg = list_entry(filter->fo_llog_list.next,
2608 struct obd_llog_group, olg_list);
2609 list_del(&olg->olg_list);
2610 list_add(&olg->olg_list, &remove_list);
2612 spin_unlock(&filter->fo_llog_list_lock);
2614 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2615 list_del_init(&olg->olg_list);
2616 rc = filter_olg_fini(olg);
2618 CERROR("failed to cleanup llogging subsystem for %u\n",
2626 static int filter_precleanup(struct obd_device *obd,
2627 enum obd_cleanup_stage stage)
2633 case OBD_CLEANUP_EARLY:
2635 case OBD_CLEANUP_EXPORTS:
2636 /* Stop recovery before namespace cleanup. */
2637 target_stop_recovery_thread(obd);
2638 target_cleanup_recovery(obd);
2639 rc = filter_llog_preclean(obd);
2645 static int filter_cleanup(struct obd_device *obd)
2647 struct filter_obd *filter = &obd->u.filter;
2651 LCONSOLE_WARN("%s: shutting down for failover; client state "
2652 "will be preserved.\n", obd->obd_name);
2654 if (!list_empty(&obd->obd_exports)) {
2655 CERROR("%s: still has clients!\n", obd->obd_name);
2656 class_disconnect_exports(obd);
2657 if (!list_empty(&obd->obd_exports)) {
2658 CERROR("still has exports after forced cleanup?\n");
2663 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2664 lprocfs_free_per_client_stats(obd);
2665 lprocfs_free_obd_stats(obd);
2666 lprocfs_obd_cleanup(obd);
2667 lquota_cleanup(filter_quota_interface_ref, obd);
2669 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2670 obd->obd_namespace = NULL;
2672 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2674 if (obd->u.obt.obt_sb == NULL)
2679 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2680 shrink_dcache_sb(obd->u.obt.obt_sb);
2682 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2683 obd->u.obt.obt_sb = NULL;
2685 fsfilt_put_ops(obd->obd_fsops);
2687 filter_iobuf_pool_done(filter);
2689 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2694 static int filter_connect_internal(struct obd_export *exp,
2695 struct obd_connect_data *data)
2697 struct filter_export_data *fed = &exp->exp_filter_data;
2702 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2703 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2704 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2705 data->ocd_connect_flags, data->ocd_version,
2706 data->ocd_grant, data->ocd_index);
2708 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2709 CWARN("!!! This export (nid %s) used object group %d "
2710 "earlier; now it's trying to use group %d! This could "
2711 "be a bug in the MDS. Please report to "
2712 "http://bugzilla.lustre.org/\n",
2713 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2716 fed->fed_group = data->ocd_group;
2718 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2719 exp->exp_connect_flags = data->ocd_connect_flags;
2720 data->ocd_version = LUSTRE_VERSION_CODE;
2722 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2723 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2724 data->ocd_connect_flags & OBD_CONNECT_MDS))
2727 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2728 struct filter_obd *filter = &exp->exp_obd->u.filter;
2729 obd_size left, want;
2731 spin_lock(&exp->exp_obd->obd_osfs_lock);
2732 left = filter_grant_space_left(exp);
2733 want = data->ocd_grant;
2734 filter_grant(exp, fed->fed_grant, want, left);
2735 data->ocd_grant = fed->fed_grant;
2736 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2738 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2739 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2740 exp->exp_client_uuid.uuid, exp,
2741 data->ocd_grant, want, left);
2743 filter->fo_tot_granted_clients ++;
2746 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2747 struct filter_obd *filter = &exp->exp_obd->u.filter;
2748 struct lr_server_data *lsd = filter->fo_fsd;
2749 int index = le32_to_cpu(lsd->lsd_ost_index);
2751 if (!(lsd->lsd_feature_compat &
2752 cpu_to_le32(OBD_COMPAT_OST))) {
2753 /* this will only happen on the first connect */
2754 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2755 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2756 filter_update_server_data(exp->exp_obd,
2757 filter->fo_rcvd_filp, lsd, 1);
2758 } else if (index != data->ocd_index) {
2759 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2760 " %u doesn't match actual OST index"
2761 " %u in last_rcvd file, bad "
2763 obd_export_nid2str(exp), index,
2767 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2768 * FIXME: We don't strictly need the COMPAT flag for that,
2769 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2770 * FIXME: We needed it for the index, as index 0 is valid. */
2773 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2774 data->ocd_brw_size = 65536;
2775 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2776 data->ocd_brw_size = min(data->ocd_brw_size,
2777 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2778 LASSERT(data->ocd_brw_size);
2781 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2782 __u32 cksum_types = data->ocd_cksum_types;
2784 /* The client set in ocd_cksum_types the checksum types it
2785 * supports. We have to mask off the algorithms that we don't
2787 if (cksum_types & OBD_CKSUM_ALL)
2788 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2790 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2792 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2793 "%x\n", exp->exp_obd->obd_name,
2794 obd_export_nid2str(exp), cksum_types,
2795 data->ocd_cksum_types);
2797 /* This client does not support OBD_CONNECT_CKSUM
2798 * fall back to CRC32 */
2799 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2800 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2801 exp->exp_obd->obd_name,
2802 obd_export_nid2str(exp));
2808 static int filter_reconnect(const struct lu_env *env,
2809 struct obd_export *exp, struct obd_device *obd,
2810 struct obd_uuid *cluuid,
2811 struct obd_connect_data *data,
2817 if (exp == NULL || obd == NULL || cluuid == NULL)
2820 rc = filter_connect_internal(exp, data);
2822 filter_export_stats_init(obd, exp, localdata);
2827 /* nearly identical to mds_connect */
2828 static int filter_connect(const struct lu_env *env,
2829 struct obd_export **exp, struct obd_device *obd,
2830 struct obd_uuid *cluuid,
2831 struct obd_connect_data *data, void *localdata)
2833 struct lvfs_run_ctxt saved;
2834 struct lustre_handle conn = { 0 };
2835 struct obd_export *lexp;
2836 struct filter_export_data *fed;
2837 struct lsd_client_data *lcd = NULL;
2842 if (exp == NULL || obd == NULL || cluuid == NULL)
2845 rc = class_connect(&conn, obd, cluuid);
2848 lexp = class_conn2export(&conn);
2849 LASSERT(lexp != NULL);
2851 fed = &lexp->exp_filter_data;
2853 rc = filter_connect_internal(lexp, data);
2857 filter_export_stats_init(obd, lexp, localdata);
2858 if (obd->obd_replayable) {
2859 OBD_ALLOC(lcd, sizeof(*lcd));
2861 CERROR("filter: out of memory for client data\n");
2862 GOTO(cleanup, rc = -ENOMEM);
2865 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2867 rc = filter_client_add(obd, lexp, -1);
2872 group = data->ocd_group;
2874 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2875 obd->obd_name, lexp->exp_handle.h_cookie, group);
2877 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2878 rc = filter_read_groups(obd, group, 1);
2879 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2881 CERROR("can't read group %u\n", group);
2891 fed->fed_lcd = NULL;
2893 class_disconnect(lexp);
2902 /* Do extra sanity checks for grant accounting. We do this at connect,
2903 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2904 * always get rid of it or turn it off when we know accounting is good. */
2905 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2907 struct filter_export_data *fed;
2908 struct obd_export *exp;
2909 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2910 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2911 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2913 if (list_empty(&obd->obd_exports))
2916 /* We don't want to do this for large machines that do lots of
2917 mounts or unmounts. It burns... */
2918 if (obd->obd_num_exports > 100)
2921 spin_lock(&obd->obd_osfs_lock);
2922 spin_lock(&obd->obd_dev_lock);
2923 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2925 fed = &exp->exp_filter_data;
2926 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2929 if (maxsize > 0) { /* we may not have done a statfs yet */
2930 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2931 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2932 exp->exp_client_uuid.uuid, exp,
2933 fed->fed_grant, fed->fed_pending, maxsize);
2934 LASSERTF(fed->fed_dirty <= maxsize,
2935 "%s: cli %s/%p %ld > "LPU64"\n", func,
2936 exp->exp_client_uuid.uuid, exp,
2937 fed->fed_dirty, maxsize);
2940 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2941 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2942 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2944 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2945 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2946 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2947 tot_granted += fed->fed_grant + fed->fed_pending;
2948 tot_pending += fed->fed_pending;
2949 tot_dirty += fed->fed_dirty;
2951 fo_tot_granted = obd->u.filter.fo_tot_granted;
2952 fo_tot_pending = obd->u.filter.fo_tot_pending;
2953 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2954 spin_unlock(&obd->obd_dev_lock);
2955 spin_unlock(&obd->obd_osfs_lock);
2957 /* Do these assertions outside the spinlocks so we don't kill system */
2958 if (tot_granted != fo_tot_granted)
2959 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2960 func, tot_granted, fo_tot_granted);
2961 if (tot_pending != fo_tot_pending)
2962 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2963 func, tot_pending, fo_tot_pending);
2964 if (tot_dirty != fo_tot_dirty)
2965 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2966 func, tot_dirty, fo_tot_dirty);
2967 if (tot_pending > tot_granted)
2968 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2969 func, tot_pending, tot_granted);
2970 if (tot_granted > maxsize)
2971 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2972 func, tot_granted, maxsize);
2973 if (tot_dirty > maxsize)
2974 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2975 func, tot_dirty, maxsize);
2978 /* Remove this client from the grant accounting totals. We also remove
2979 * the export from the obd device under the osfs and dev locks to ensure
2980 * that the filter_grant_sanity_check() calculations are always valid.
2981 * The client should do something similar when it invalidates its import. */
2982 static void filter_grant_discard(struct obd_export *exp)
2984 struct obd_device *obd = exp->exp_obd;
2985 struct filter_obd *filter = &obd->u.filter;
2986 struct filter_export_data *fed = &exp->exp_filter_data;
2988 spin_lock(&obd->obd_osfs_lock);
2989 spin_lock(&obd->obd_dev_lock);
2990 list_del_init(&exp->exp_obd_chain);
2991 spin_unlock(&obd->obd_dev_lock);
2993 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2994 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2995 obd->obd_name, filter->fo_tot_granted,
2996 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2997 filter->fo_tot_granted -= fed->fed_grant;
2998 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2999 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
3000 obd->obd_name, filter->fo_tot_pending,
3001 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
3002 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
3003 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
3004 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
3005 obd->obd_name, filter->fo_tot_dirty,
3006 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
3007 filter->fo_tot_dirty -= fed->fed_dirty;
3011 spin_unlock(&obd->obd_osfs_lock);
3014 static int filter_destroy_export(struct obd_export *exp)
3018 if (exp->exp_filter_data.fed_pending)
3019 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
3020 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
3021 exp, exp->exp_filter_data.fed_pending);
3023 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3025 target_destroy_export(exp);
3026 ldlm_destroy_export(exp);
3028 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
3032 if (exp->exp_obd->obd_replayable)
3033 filter_client_free(exp);
3035 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3037 filter_grant_discard(exp);
3038 filter_fmd_cleanup(exp);
3040 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3041 struct filter_obd *filter = &exp->exp_obd->u.filter;
3042 if (filter->fo_tot_granted_clients > 0)
3043 filter->fo_tot_granted_clients --;
3046 if (!(exp->exp_flags & OBD_OPT_FORCE))
3047 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3052 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3054 struct obd_llog_group *olg_min, *olg;
3055 struct filter_obd *filter;
3056 int worked = -1, group;
3057 struct llog_ctxt *ctxt;
3060 filter = &obd->u.filter;
3062 /* we can't sync log holding spinlock. also, we do not want to get
3063 * into livelock. so we do following: loop over MDS's exports in
3064 * group order and skip already synced llogs -bzzz */
3066 /* look for group with min. number, but > worked */
3069 spin_lock(&filter->fo_llog_list_lock);
3070 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3071 if (olg->olg_group <= worked) {
3072 /* this group is already synced */
3075 if (group < olg->olg_group) {
3076 /* we have group with smaller number to sync */
3079 /* store current minimal group */
3081 group = olg->olg_group;
3083 spin_unlock(&filter->fo_llog_list_lock);
3085 if (olg_min == NULL)
3088 worked = olg_min->olg_group;
3089 if (olg_min->olg_exp &&
3090 (dexp == olg_min->olg_exp || dexp == NULL)) {
3092 ctxt = llog_group_get_ctxt(olg_min,
3093 LLOG_MDS_OST_REPL_CTXT);
3095 err = llog_sync(ctxt, olg_min->olg_exp);
3096 llog_ctxt_put(ctxt);
3098 CERROR("error flushing logs to MDS: "
3103 } while (olg_min != NULL);
3106 /* Also incredibly similar to mds_disconnect */
3107 static int filter_disconnect(struct obd_export *exp)
3109 struct obd_device *obd = exp->exp_obd;
3114 class_export_get(exp);
3116 if (!(exp->exp_flags & OBD_OPT_FORCE))
3117 filter_grant_sanity_check(obd, __FUNCTION__);
3118 filter_grant_discard(exp);
3120 /* Flush any remaining cancel messages out to the target */
3121 filter_sync_llogs(obd, exp);
3123 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3125 /* Disconnect early so that clients can't keep using export */
3126 rc = class_disconnect(exp);
3127 if (exp->exp_obd->obd_namespace != NULL)
3128 ldlm_cancel_locks_for_export(exp);
3130 fsfilt_sync(obd, obd->u.obt.obt_sb);
3132 lprocfs_exp_cleanup(exp);
3134 class_export_put(exp);
3138 /* reverse import is changed, sync all cancels */
3139 static void filter_revimp_update(struct obd_export *exp)
3144 class_export_get(exp);
3146 /* flush any remaining cancel messages out to the target */
3147 filter_sync_llogs(exp->exp_obd, exp);
3148 class_export_put(exp);
3152 static int filter_ping(struct obd_export *exp)
3154 filter_fmd_expire(exp);
3158 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3159 const char *what, int quiet)
3161 struct dentry *dchild = NULL;
3164 if (oa->o_valid & OBD_MD_FLGROUP)
3167 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3169 if (IS_ERR(dchild)) {
3170 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3171 what, group, oa->o_id);
3175 if (dchild->d_inode == NULL) {
3177 CERROR("%s: %s on non-existent object: "LPU64"\n",
3178 obd->obd_name, what, oa->o_id);
3180 RETURN(ERR_PTR(-ENOENT));
3186 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3188 struct dentry *dentry = NULL;
3189 struct obd_device *obd;
3193 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3194 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3198 obd = class_exp2obd(exp);
3200 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3204 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3206 RETURN(PTR_ERR(dentry));
3208 /* Limit the valid bits in the return data to what we actually use */
3209 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3210 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3216 /* this should be enabled/disabled in condition to enabled/disabled large
3217 * inodes (fast EAs) in backing store FS. */
3218 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3219 void *handle, struct obdo *oa)
3221 struct obd_device *obd = exp->exp_obd;
3225 if (oa->o_valid & OBD_MD_FLFID) {
3226 struct filter_fid ff;
3228 if (!(oa->o_valid & OBD_MD_FLGROUP))
3230 /* packing fid and converting it to LE for storing into EA.
3231 * Here ->o_stripe_idx should be filled by LOV and rest of
3232 * fields - by client. */
3233 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3234 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3235 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3236 ff.ff_objid = cpu_to_le64(oa->o_id);
3237 ff.ff_group = cpu_to_le64(oa->o_gr);
3239 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3240 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3241 oa->o_generation, oa->o_id, oa->o_gr);
3243 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3245 CERROR("store fid in object failed! rc: %d\n", rc);
3247 CDEBUG(D_HA, "OSS object without fid info!\n");
3253 /* this is called from filter_truncate() until we have filter_punch() */
3254 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3255 struct obdo *oa, struct obd_trans_info *oti)
3257 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3258 struct llog_cookie *fcc = NULL;
3259 struct filter_obd *filter;
3260 int rc, err, sync = 0;
3261 loff_t old_size = 0;
3262 unsigned int ia_valid;
3263 struct inode *inode;
3268 LASSERT(dentry != NULL);
3269 LASSERT(!IS_ERR(dentry));
3271 inode = dentry->d_inode;
3272 LASSERT(inode != NULL);
3274 filter = &exp->exp_obd->u.filter;
3275 iattr_from_obdo(&iattr, oa, oa->o_valid);
3276 ia_valid = iattr.ia_valid;
3278 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3279 OBD_ALLOC(fcc, sizeof(*fcc));
3281 *fcc = oa->o_lcookie;
3283 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3285 /* Filter truncates and writes are serialized by
3286 * i_alloc_sem, see the comment in
3287 * filter_preprw_write.*/
3288 if (ia_valid & ATTR_SIZE)
3289 down_write(&inode->i_alloc_sem);
3290 LOCK_INODE_MUTEX(inode);
3291 old_size = i_size_read(inode);
3294 /* VBR: version recovery check */
3295 rc = filter_version_get_check(exp, oti, inode);
3297 GOTO(out_unlock, rc);
3299 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3300 * then we will accept the UID+GID sent by the client during write for
3301 * initializing the ownership of this inode. We only allow this to
3302 * happen once so clear these bits in setattr. In 2.6 kernels it is
3303 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3304 * the flags that are actually being set. */
3305 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3306 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3307 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3309 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3310 if (!(ia_valid & ATTR_MODE)) {
3311 iattr.ia_mode = inode->i_mode;
3312 iattr.ia_valid |= ATTR_MODE;
3314 iattr.ia_mode &= ~S_ISUID;
3316 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3317 if (!(iattr.ia_valid & ATTR_MODE)) {
3318 iattr.ia_mode = inode->i_mode;
3319 iattr.ia_valid |= ATTR_MODE;
3321 iattr.ia_mode &= ~S_ISGID;
3324 orig_ids[USRQUOTA] = inode->i_uid;
3325 orig_ids[GRPQUOTA] = inode->i_gid;
3326 handle = fsfilt_start_log(exp->exp_obd, inode,
3327 FSFILT_OP_SETATTR, oti, 1);
3329 GOTO(out_unlock, rc = PTR_ERR(handle));
3331 /* update inode EA only once when inode is suid bit marked. As
3332 * on 2.6.x UID and GID may be set separately, we check here
3333 * only one of them to avoid double setting. */
3334 if (inode->i_mode & S_ISUID)
3335 filter_update_fidea(exp, inode, handle, oa);
3337 handle = fsfilt_start(exp->exp_obd, inode,
3338 FSFILT_OP_SETATTR, oti);
3340 GOTO(out_unlock, rc = PTR_ERR(handle));
3342 if (oa->o_valid & OBD_MD_FLFLAGS) {
3343 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3344 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3346 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3348 /* set cancel cookie callback function */
3349 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3350 filter_cancel_cookies_cb,
3354 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3355 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3357 /* The truncate might have used up our transaction credits. Make
3358 * sure we have one left for the last_rcvd update. */
3359 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3361 rc = filter_finish_transno(exp, inode, oti, rc, sync);
3363 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3367 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3369 CERROR("error on commit, err = %d\n", err);
3376 /* For a partial-page truncate flush the page to disk immediately
3377 * to avoid data corruption during direct disk write. b=17397 */
3378 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3379 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3380 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3388 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3389 UNLOCK_INODE_MUTEX(inode);
3390 if (ia_valid & ATTR_SIZE)
3391 up_write(&inode->i_alloc_sem);
3393 OBD_FREE(fcc, sizeof(*fcc));
3395 /* trigger quota release */
3396 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3397 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3398 int rc2 = lquota_adjust(filter_quota_interface_ref,
3399 exp->exp_obd, cur_ids,
3400 orig_ids, rc, FSFILT_OP_SETATTR);
3401 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3402 "filter adjust qunit. (rc:%d)\n", rc2);
3407 /* this is called from filter_truncate() until we have filter_punch() */
3408 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3409 struct obd_trans_info *oti)
3411 struct obdo *oa = oinfo->oi_oa;
3412 struct lustre_capa *capa = oinfo_capa(oinfo);
3413 struct ldlm_res_id res_id;
3414 struct filter_mod_data *fmd;
3415 struct lvfs_run_ctxt saved;
3416 struct filter_obd *filter;
3417 struct ldlm_resource *res;
3418 struct dentry *dentry;
3419 __u64 opc = CAPA_OPC_META_WRITE;
3423 if (oa->o_valid & OBD_FL_TRUNC)
3424 opc |= CAPA_OPC_OSS_TRUNC;
3425 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa, opc);
3429 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3430 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
3435 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3436 /* This would be very bad - accidentally truncating a file when
3437 * changing the time or similar - bug 12203. */
3438 if (oa->o_valid & OBD_MD_FLSIZE &&
3439 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3440 static char mdsinum[48];
3442 if (oa->o_valid & OBD_MD_FLFID)
3443 snprintf(mdsinum, sizeof(mdsinum) - 1,
3444 " of inode "LPU64"/%u", oa->o_fid,
3449 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3451 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3456 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3458 RETURN(PTR_ERR(dentry));
3460 filter = &exp->exp_obd->u.filter;
3461 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3465 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3466 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3467 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3468 fmd->fmd_mactime_xid = oti->oti_xid;
3469 filter_fmd_put(exp, fmd);
3472 /* setting objects attributes (including owner/group) */
3473 rc = filter_setattr_internal(exp, dentry, oa, oti);
3475 GOTO(out_unlock, rc);
3477 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3478 &res_id, LDLM_EXTENT, 0);
3481 LDLM_RESOURCE_ADDREF(res);
3482 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3483 LDLM_RESOURCE_DELREF(res);
3484 ldlm_resource_putref(res);
3487 oa->o_valid = OBD_MD_FLID;
3489 /* Quota release need uid/gid info */
3490 obdo_from_inode(oa, dentry->d_inode,
3491 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3497 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3501 /* XXX identical to osc_unpackmd */
3502 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3503 struct lov_mds_md *lmm, int lmm_bytes)
3509 if (lmm_bytes < sizeof (*lmm)) {
3510 CERROR("lov_mds_md too small: %d, need %d\n",
3511 lmm_bytes, (int)sizeof(*lmm));
3514 /* XXX LOV_MAGIC etc check? */
3516 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3517 CERROR("lov_mds_md: zero lmm_object_id\n");
3522 lsm_size = lov_stripe_md_size(1);
3526 if (*lsmp != NULL && lmm == NULL) {
3527 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3528 OBD_FREE(*lsmp, lsm_size);
3533 if (*lsmp == NULL) {
3534 OBD_ALLOC(*lsmp, lsm_size);
3538 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3539 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3540 OBD_FREE(*lsmp, lsm_size);
3543 loi_init((*lsmp)->lsm_oinfo[0]);
3547 /* XXX zero *lsmp? */
3548 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3549 LASSERT((*lsmp)->lsm_object_id);
3552 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3557 /* caller must hold fo_create_locks[oa->o_gr] */
3558 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3559 struct filter_obd *filter)
3561 struct obdo doa = { 0 }; /* XXX obdo on stack */
3568 LASSERT_MDS_GROUP(oa->o_gr);
3569 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3570 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3572 memset(&doa, 0, sizeof(doa));
3574 doa.o_valid |= OBD_MD_FLGROUP;
3575 doa.o_gr = oa->o_gr;
3576 doa.o_mode = S_IFREG;
3578 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3579 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3580 exp->exp_obd->obd_name, doa.o_gr);
3584 last = filter_last_id(filter, doa.o_gr);
3586 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3588 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3589 exp->exp_obd->obd_name, oa->o_id + 1, last,
3590 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3592 for (id = last; id > oa->o_id; id--) {
3594 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3595 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3596 CEMERG("error destroying precreate objid "LPU64": %d\n",
3599 /* update last_id on disk periodically so that if we restart
3600 * we don't need to re-scan all of the just-deleted objects. */
3601 if ((id & 511) == 0 && !skip_orphan) {
3602 filter_set_last_id(filter, id - 1, doa.o_gr);
3603 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3607 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3608 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3611 filter_set_last_id(filter, id, doa.o_gr);
3612 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3614 /* don't reuse orphan object, return last used objid */
3618 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3623 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3624 obd_gr group, int *num);
3625 /* returns a negative error or a nonnegative number of files to create */
3626 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3627 obd_gr group, struct obd_trans_info *oti)
3629 struct obd_device *obd = exp->exp_obd;
3630 struct filter_obd *filter = &obd->u.filter;
3634 /* delete orphans request */
3635 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3636 obd_id last = filter_last_id(filter, group);
3638 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3639 CERROR("%s: dropping old orphan cleanup request\n",
3643 /* This causes inflight precreates to abort and drop lock */
3644 set_bit(group, &filter->fo_destroys_in_progress);
3645 down(&filter->fo_create_locks[group]);
3646 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3647 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3648 exp->exp_obd->obd_name, group);
3649 up(&filter->fo_create_locks[group]);
3652 diff = oa->o_id - last;
3653 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3656 if (-diff > OST_MAX_PRECREATE) {
3657 CERROR("%s: ignoring bogus orphan destroy request: "
3658 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3660 /* FIXME: should reset precreate_next_id on MDS */
3661 GOTO(out, rc = -EINVAL);
3664 rc = filter_destroy_precreated(exp, oa, filter);
3666 CERROR("%s: unable to write lastobjid, but "
3667 "orphans were deleted\n", obd->obd_name);
3670 /* XXX: Used by MDS for the first time! */
3671 clear_bit(group, &filter->fo_destroys_in_progress);
3674 down(&filter->fo_create_locks[group]);
3675 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3676 CERROR("%s: dropping old precreate request\n",
3680 /* only precreate if group == 0 and o_id is specfied */
3681 if (group == FILTER_GROUP_LLOG || oa->o_id == 0)
3684 diff = oa->o_id - filter_last_id(filter, group);
3685 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3686 filter_last_id(filter, group), diff);
3688 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3689 oa->o_id, filter_last_id(filter, group), diff);
3693 oa->o_id = filter_last_id(&obd->u.filter, group);
3694 rc = filter_precreate(obd, oa, group, &diff);
3695 oa->o_id = filter_last_id(&obd->u.filter, group);
3697 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3700 /* else diff == 0 */
3703 up(&filter->fo_create_locks[group]);
3707 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3708 __u64 max_age, __u32 flags)
3710 struct filter_obd *filter = &obd->u.filter;
3711 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3715 /* at least try to account for cached pages. its still racey and
3716 * might be under-reporting if clients haven't announced their
3717 * caches with brw recently */
3718 spin_lock(&obd->obd_osfs_lock);
3719 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3720 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3721 spin_unlock(&obd->obd_osfs_lock);
3723 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3724 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3725 filter->fo_tot_dirty, filter->fo_tot_granted,
3726 filter->fo_tot_pending,
3727 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3729 filter_grant_sanity_check(obd, __FUNCTION__);
3731 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3732 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3733 osfs->os_bsize - 1) >> blockbits));
3735 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3736 * stop creating files on MDS if OST is not good shape to create
3738 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3739 OS_STATE_READONLY : 0;
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 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 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() - HZ, 0);
3819 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3820 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3821 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3822 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3826 OBD_FREE(osfs, sizeof(*osfs));
3831 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3832 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3834 for (i = 0; i < *num && err == 0; i++) {
3835 int cleanup_phase = 0;
3837 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3838 CWARN("%s: create aborted by destroy\n",
3847 last_id = filter_last_id(filter, group);
3848 if (next_id > last_id) {
3849 CERROR("Error: Trying to recreate obj greater"
3850 "than last id "LPD64" > "LPD64"\n",
3852 GOTO(cleanup, rc = -EINVAL);
3855 next_id = filter_last_id(filter, group) + 1;
3857 dparent = filter_parent_lock(obd, group, next_id);
3858 if (IS_ERR(dparent))
3859 GOTO(cleanup, rc = PTR_ERR(dparent));
3860 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3862 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3864 GOTO(cleanup, rc = PTR_ERR(dchild));
3865 cleanup_phase = 2; /* f_dput(dchild) */
3867 if (dchild->d_inode != NULL) {
3868 /* This would only happen if lastobjid was bad on disk*/
3869 /* Could also happen if recreating missing obj but it
3870 * already exists. */
3872 CERROR("%s: recreating existing object %.*s?\n",
3873 obd->obd_name, dchild->d_name.len,
3874 dchild->d_name.name);
3876 /* Use these existing objects if they are
3878 if (dchild->d_inode->i_size == 0) {
3879 rc = filter_use_existing_obj(obd,dchild,
3880 &handle, &cleanup_phase);
3887 CERROR("%s: Serious error: objid %.*s already "
3888 "exists; is this filesystem corrupt?\n",
3889 obd->obd_name, dchild->d_name.len,
3890 dchild->d_name.name);
3893 GOTO(cleanup, rc = -EEXIST);
3896 handle = fsfilt_start_log(obd, dparent->d_inode,
3897 FSFILT_OP_CREATE, NULL, 1);
3899 GOTO(cleanup, rc = PTR_ERR(handle));
3902 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3903 ",od->o_id="LPU64")\n", obd->obd_name, group,
3906 /* We mark object SUID+SGID to flag it for accepting UID+GID
3907 * from client on first write. Currently the permission bits
3908 * on the OST are never used, so this is OK. */
3909 rc = ll_vfs_create(dparent->d_inode, dchild,
3910 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3912 CERROR("create failed rc = %d\n", rc);
3913 if (rc == -ENOSPC) {
3914 os_ffree = filter_calc_free_inodes(obd);
3916 CERROR("%s: free inode "LPU64"\n",
3917 obd->obd_name, os_ffree);
3923 if (!recreate_obj) {
3924 filter_set_last_id(filter, next_id, group);
3925 err = filter_update_last_objid(obd, group, 0);
3927 CERROR("unable to write lastobjid "
3928 "but file created\n");
3932 switch(cleanup_phase) {
3934 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3936 CERROR("error on commit, err = %d\n", err);
3943 filter_parent_unlock(dparent);
3950 if (time_after(jiffies, enough_time)) {
3952 "%s: precreate slow - want %d got %d \n",
3953 obd->obd_name, *num, i);
3960 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3961 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3966 static int filter_create(struct obd_export *exp, struct obdo *oa,
3967 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3969 struct obd_device *obd = exp->exp_obd;
3970 struct filter_export_data *fed;
3971 struct filter_obd *filter;
3972 struct lvfs_run_ctxt saved;
3973 struct lov_stripe_md *lsm = NULL;
3974 int rc = 0, diff, group = oa->o_gr;
3977 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3978 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3980 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3981 CERROR("!!! nid %s sent invalid object group %d\n",
3982 obd_export_nid2str(exp), group);
3986 fed = &exp->exp_filter_data;
3987 filter = &obd->u.filter;
3989 if (fed->fed_group != group) {
3990 CERROR("!!! this export (nid %s) used object group %d "
3991 "earlier; now it's trying to use group %d! This could "
3992 "be a bug in the MDS. Please report to "
3993 "http://bugzilla.lustre.org/\n",
3994 obd_export_nid2str(exp), fed->fed_group, group);
4001 rc = obd_alloc_memmd(exp, &lsm);
4008 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4010 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4011 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4012 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
4013 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4014 oa->o_id, filter_last_id(filter,
4019 down(&filter->fo_create_locks[oa->o_gr]);
4020 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
4021 up(&filter->fo_create_locks[oa->o_gr]);
4024 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
4027 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4028 if (rc && ea != NULL && *ea != lsm) {
4029 obd_free_memmd(exp, &lsm);
4030 } else if (rc == 0 && ea != NULL) {
4031 /* XXX LOV STACKING: the lsm that is passed to us from
4032 * LOV does not have valid lsm_oinfo data structs, so
4033 * don't go touching that. This needs to be fixed in a
4035 lsm->lsm_object_id = oa->o_id;
4042 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4043 struct lov_stripe_md *md, struct obd_trans_info *oti,
4044 struct obd_export *md_exp, void *capa)
4046 unsigned int qcids[MAXQUOTAS] = {0, 0};
4047 struct obd_device *obd;
4048 struct filter_obd *filter;
4049 struct dentry *dchild = NULL, *dparent = NULL;
4050 struct lustre_handle lockh = { 0 };
4051 struct lvfs_run_ctxt saved;
4052 void *handle = NULL;
4053 struct llog_cookie *fcc = NULL;
4054 int rc, rc2, cleanup_phase = 0, sync = 0;
4058 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
4060 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
4061 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4066 filter = &obd->u.filter;
4068 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4071 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4072 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4074 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4076 GOTO(cleanup, rc = PTR_ERR(dchild));
4079 if (dchild->d_inode == NULL) {
4080 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4082 /* If object already gone, cancel cookie right now */
4083 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4084 struct llog_ctxt *ctxt;
4085 struct obd_llog_group *olg;
4086 fcc = &oa->o_lcookie;
4087 olg = filter_find_olg(obd, oa->o_gr);
4089 CERROR(" %s: can not find olg of group %d\n",
4090 obd->obd_name, (int)oa->o_gr);
4091 GOTO(cleanup, rc = PTR_ERR(olg));
4093 llog_group_set_export(olg, exp);
4095 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4096 llog_cancel(ctxt, NULL, 1, fcc, 0);
4097 llog_ctxt_put(ctxt);
4098 fcc = NULL; /* we didn't allocate fcc, don't free it */
4100 GOTO(cleanup, rc = -ENOENT);
4103 filter_prepare_destroy(obd, oa->o_id, oa->o_gr, &lockh);
4105 /* Our MDC connection is established by the MDS to us */
4106 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4107 OBD_ALLOC(fcc, sizeof(*fcc));
4109 *fcc = oa->o_lcookie;
4111 DQUOT_INIT(dchild->d_inode);
4113 /* we're gonna truncate it first in order to avoid possible deadlock:
4115 * open trasaction open transaction
4116 * down(i_zombie) down(i_zombie)
4117 * restart transaction
4118 * (see BUG 4180) -bzzz
4120 LOCK_INODE_MUTEX(dchild->d_inode);
4122 /* VBR: version recovery check */
4123 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4127 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4129 if (IS_ERR(handle)) {
4130 UNLOCK_INODE_MUTEX(dchild->d_inode);
4131 GOTO(cleanup, rc = PTR_ERR(handle));
4134 iattr.ia_valid = ATTR_SIZE;
4136 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4137 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4138 UNLOCK_INODE_MUTEX(dchild->d_inode);
4142 GOTO(cleanup, rc = rc2);
4144 /* We don't actually need to lock the parent until we are unlinking
4145 * here, and not while truncating above. That avoids holding the
4146 * parent lock for a long time during truncate, which can block other
4147 * threads from doing anything to objects in that directory. bug 7171 */
4148 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4149 if (IS_ERR(dparent))
4150 GOTO(cleanup, rc = PTR_ERR(dparent));
4151 cleanup_phase = 3; /* filter_parent_unlock */
4153 LOCK_INODE_MUTEX(dchild->d_inode);
4154 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4155 if (IS_ERR(handle)) {
4156 UNLOCK_INODE_MUTEX(dchild->d_inode);
4157 GOTO(cleanup, rc = PTR_ERR(handle));
4159 cleanup_phase = 4; /* fsfilt_commit */
4161 /* Quota release need uid/gid of inode */
4162 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4164 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4166 /* this drops dchild->d_inode->i_mutex unconditionally */
4167 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4171 switch(cleanup_phase) {
4174 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4175 oti->oti_handle : handle,
4176 filter_cancel_cookies_cb,
4178 /* If add_journal_cb failed, then filter_finish_transno
4179 * will commit the handle and we will do a sync
4180 * on commit. then we call callback directly to free
4183 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4185 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4188 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4190 CERROR("error on commit, err = %d\n", rc2);
4197 filter_parent_unlock(dparent);
4199 filter_fini_destroy(obd, &lockh);
4203 OBD_FREE(fcc, sizeof(*fcc));
4205 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4208 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4212 /* trigger quota release */
4213 qcids[USRQUOTA] = oa->o_uid;
4214 qcids[GRPQUOTA] = oa->o_gid;
4215 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4218 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4222 /* NB start and end are used for punch, but not truncate */
4223 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4224 struct obd_trans_info *oti,
4225 struct ptlrpc_request_set *rqset)
4230 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4231 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4232 oinfo->oi_policy.l_extent.end);
4236 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4237 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4238 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4240 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4241 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4242 rc = filter_setattr(exp, oinfo, oti);
4243 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4247 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4248 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4251 struct lvfs_run_ctxt saved;
4252 struct filter_obd *filter;
4253 struct dentry *dentry;
4257 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
4258 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4262 filter = &exp->exp_obd->u.filter;
4264 /* An objid of zero is taken to mean "sync whole filesystem" */
4265 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4266 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4267 /* Flush any remaining cancel messages out to the target */
4268 filter_sync_llogs(exp->exp_obd, exp);
4272 dentry = filter_oa2dentry(exp->exp_obd, oa);
4274 RETURN(PTR_ERR(dentry));
4276 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4278 LOCK_INODE_MUTEX(dentry->d_inode);
4280 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4282 /* just any file to grab fsync method - "file" arg unused */
4283 struct file *file = filter->fo_rcvd_filp;
4285 if (file->f_op && file->f_op->fsync)
4286 rc = file->f_op->fsync(NULL, dentry, 1);
4288 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4292 UNLOCK_INODE_MUTEX(dentry->d_inode);
4294 oa->o_valid = OBD_MD_FLID;
4295 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4297 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4303 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4304 void *key, __u32 *vallen, void *val,
4305 struct lov_stripe_md *lsm)
4307 struct obd_device *obd;
4310 obd = class_exp2obd(exp);
4312 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4316 if (KEY_IS(KEY_BLOCKSIZE)) {
4317 __u32 *blocksize = val;
4319 if (*vallen < sizeof(*blocksize))
4321 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4323 *vallen = sizeof(*blocksize);
4327 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4328 __u32 *blocksize_bits = val;
4329 if (blocksize_bits) {
4330 if (*vallen < sizeof(*blocksize_bits))
4332 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4334 *vallen = sizeof(*blocksize_bits);
4338 if (KEY_IS(KEY_LAST_ID)) {
4339 obd_id *last_id = val;
4340 /* FIXME: object groups */
4342 if (*vallen < sizeof(*last_id))
4344 *last_id = filter_last_id(&obd->u.filter,
4345 exp->exp_filter_data.fed_group);
4347 *vallen = sizeof(*last_id);
4351 if (KEY_IS(KEY_FIEMAP)) {
4352 struct ll_fiemap_info_key *fm_key = key;
4353 struct dentry *dentry;
4354 struct ll_user_fiemap *fiemap = val;
4355 struct lvfs_run_ctxt saved;
4358 if (fiemap == NULL) {
4359 *vallen = fiemap_count_to_size(
4360 fm_key->fiemap.fm_extent_count);
4364 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4367 RETURN(PTR_ERR(dentry));
4369 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4370 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4371 rc = fsfilt_iocontrol(obd, dentry->d_inode, NULL,
4372 EXT3_IOC_FIEMAP, (long)fiemap);
4373 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4379 CDEBUG(D_IOCTL, "invalid key\n");
4383 static inline int filter_setup_llog_group(struct obd_export *exp,
4384 struct obd_device *obd,
4387 struct obd_llog_group *olg;
4388 struct llog_ctxt *ctxt;
4391 olg = filter_find_create_olg(obd, group);
4393 RETURN(PTR_ERR(olg));
4395 llog_group_set_export(olg, exp);
4397 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4398 LASSERTF(ctxt != NULL, "ctxt is null\n");
4400 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4401 llog_ctxt_put(ctxt);
4405 static int filter_set_grant_shrink(struct obd_export *exp,
4406 struct ost_body *body)
4408 /* handle shrink grant */
4409 spin_lock(&exp->exp_obd->obd_osfs_lock);
4410 filter_grant_incoming(exp, &body->oa);
4411 spin_unlock(&exp->exp_obd->obd_osfs_lock);
4417 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4419 struct obd_device *obd;
4425 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4429 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4430 obd_export_nid2str(exp));
4431 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4433 /* setup llog imports */
4435 group = (int)(*(__u32 *)val);
4437 group = 0; /* default value */
4439 LASSERT_MDS_GROUP(group);
4440 rc = filter_setup_llog_group(exp, obd, group);
4444 if (group == FILTER_GROUP_MDS0) {
4445 /* setup llog group 1 for interop */
4446 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4449 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4454 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4455 void *key, __u32 vallen, void *val,
4456 struct ptlrpc_request_set *set)
4458 struct obd_device *obd;
4463 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4467 if (KEY_IS(KEY_CAPA_KEY)) {
4469 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4471 CERROR("filter update capability key failed: %d\n", rc);
4475 if (KEY_IS(KEY_REVIMP_UPD)) {
4476 filter_revimp_update(exp);
4477 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4481 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4482 filter_adapt_sptlrpc_conf(obd, 0);
4486 if (KEY_IS(KEY_MDS_CONN))
4487 RETURN(filter_set_mds_conn(exp, val));
4489 if (KEY_IS(KEY_GRANT_SHRINK))
4490 RETURN(filter_set_grant_shrink(exp, val));
4495 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4496 int len, void *karg, void *uarg)
4498 struct obd_device *obd = exp->exp_obd;
4499 struct obd_ioctl_data *data = karg;
4503 case OBD_IOC_ABORT_RECOVERY: {
4504 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4505 target_stop_recovery_thread(obd);
4509 case OBD_IOC_SYNC: {
4510 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4511 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4515 case OBD_IOC_SET_READONLY: {
4517 struct super_block *sb = obd->u.obt.obt_sb;
4518 struct inode *inode = sb->s_root->d_inode;
4519 BDEVNAME_DECLARE_STORAGE(tmp);
4520 CERROR("*** setting device %s read-only ***\n",
4521 ll_bdevname(sb, tmp));
4523 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4524 if (!IS_ERR(handle))
4525 rc = fsfilt_commit(obd, inode, handle, 1);
4527 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4528 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4530 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4534 case OBD_IOC_CATLOGLIST: {
4535 rc = llog_catalog_list(obd, 1, data);
4539 case OBD_IOC_LLOG_CANCEL:
4540 case OBD_IOC_LLOG_REMOVE:
4541 case OBD_IOC_LLOG_INFO:
4542 case OBD_IOC_LLOG_PRINT: {
4543 /* FIXME to be finished */
4544 RETURN(-EOPNOTSUPP);
4546 struct llog_ctxt *ctxt = NULL;
4548 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4549 rc = llog_ioctl(ctxt, cmd, data);
4550 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4563 static int filter_health_check(struct obd_device *obd)
4565 #ifdef USE_HEALTH_CHECK_WRITE
4566 struct filter_obd *filter = &obd->u.filter;
4571 * health_check to return 0 on healthy
4572 * and 1 on unhealthy.
4574 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4577 #ifdef USE_HEALTH_CHECK_WRITE
4578 LASSERT(filter->fo_health_check_filp != NULL);
4579 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4584 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4587 return filter_fid2dentry(data, NULL, gr, id);
4590 static int filter_process_config(struct obd_device *obd, obd_count len,
4593 struct lustre_cfg *lcfg = buf;
4594 struct lprocfs_static_vars lvars;
4597 switch (lcfg->lcfg_command) {
4599 lprocfs_filter_init_vars(&lvars);
4601 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4611 static struct lvfs_callback_ops filter_lvfs_ops = {
4612 l_fid2dentry: filter_lvfs_fid2dentry,
4615 static struct obd_ops filter_obd_ops = {
4616 .o_owner = THIS_MODULE,
4617 .o_get_info = filter_get_info,
4618 .o_set_info_async = filter_set_info_async,
4619 .o_setup = filter_setup,
4620 .o_precleanup = filter_precleanup,
4621 .o_cleanup = filter_cleanup,
4622 .o_connect = filter_connect,
4623 .o_reconnect = filter_reconnect,
4624 .o_disconnect = filter_disconnect,
4625 .o_ping = filter_ping,
4626 .o_init_export = filter_init_export,
4627 .o_destroy_export = filter_destroy_export,
4628 .o_statfs = filter_statfs,
4629 .o_getattr = filter_getattr,
4630 .o_unpackmd = filter_unpackmd,
4631 .o_create = filter_create,
4632 .o_setattr = filter_setattr,
4633 .o_destroy = filter_destroy,
4634 .o_brw = filter_brw,
4635 .o_punch = filter_truncate,
4636 .o_sync = filter_sync,
4637 .o_preprw = filter_preprw,
4638 .o_commitrw = filter_commitrw,
4639 .o_llog_init = filter_llog_init,
4640 .o_llog_connect = filter_llog_connect,
4641 .o_llog_finish = filter_llog_finish,
4642 .o_iocontrol = filter_iocontrol,
4643 .o_health_check = filter_health_check,
4644 .o_process_config = filter_process_config,
4647 quota_interface_t *filter_quota_interface_ref;
4648 extern quota_interface_t filter_quota_interface;
4650 static int __init obdfilter_init(void)
4652 struct lprocfs_static_vars lvars;
4655 lprocfs_filter_init_vars(&lvars);
4657 request_module("lquota");
4658 OBD_ALLOC(obdfilter_created_scratchpad,
4659 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4660 sizeof(*obdfilter_created_scratchpad));
4661 if (obdfilter_created_scratchpad == NULL)
4664 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4665 sizeof(struct filter_mod_data),
4668 GOTO(out, rc = -ENOMEM);
4670 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4671 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4673 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4674 LUSTRE_OST_NAME, NULL);
4678 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4679 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4680 ll_fmd_cachep = NULL;
4682 if (filter_quota_interface_ref)
4683 PORTAL_SYMBOL_PUT(filter_quota_interface);
4685 OBD_FREE(obdfilter_created_scratchpad,
4686 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4687 sizeof(*obdfilter_created_scratchpad));
4693 static void __exit obdfilter_exit(void)
4695 if (filter_quota_interface_ref)
4696 PORTAL_SYMBOL_PUT(filter_quota_interface);
4698 if (ll_fmd_cachep) {
4699 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4700 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4701 ll_fmd_cachep = NULL;
4704 class_unregister_type(LUSTRE_OST_NAME);
4705 OBD_FREE(obdfilter_created_scratchpad,
4706 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4707 sizeof(*obdfilter_created_scratchpad));
4710 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4711 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4712 MODULE_LICENSE("GPL");
4714 module_init(obdfilter_init);
4715 module_exit(obdfilter_exit);