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>
79 #include "filter_internal.h"
81 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
82 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS0
83 static struct lvfs_callback_ops filter_lvfs_ops;
84 cfs_mem_cache_t *ll_fmd_cachep;
86 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
87 void *cb_data, int error)
89 obd_transno_commit_cb(obd, transno, error);
92 /* Assumes caller has already pushed us into the kernel context. */
93 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
94 int rc, int force_sync)
96 struct filter_obd *filter = &exp->exp_obd->u.filter;
97 struct filter_export_data *fed = &exp->exp_filter_data;
98 struct lsd_client_data *lcd = fed->fed_lcd;
101 int err, log_pri = D_RPCTRACE;
103 /* Propagate error code. */
107 if (!exp->exp_obd->obd_replayable || oti == NULL)
110 /* we don't allocate new transnos for replayed requests */
111 if (oti->oti_transno == 0) {
112 spin_lock(&filter->fo_translock);
113 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
114 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
115 spin_unlock(&filter->fo_translock);
116 oti->oti_transno = last_rcvd;
118 spin_lock(&filter->fo_translock);
119 last_rcvd = oti->oti_transno;
120 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
121 filter->fo_fsd->lsd_last_transno =
122 cpu_to_le64(last_rcvd);
123 spin_unlock(&filter->fo_translock);
125 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
127 /* could get xid from oti, if it's ever needed */
128 lcd->lcd_last_xid = 0;
130 off = fed->fed_lr_off;
132 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
133 fed->fed_lr_idx, fed->fed_lr_off);
137 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
143 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
144 lcd, sizeof(*lcd), &off,
145 force_sync | exp->exp_need_sync);
147 filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
155 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
156 last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
161 void f_dput(struct dentry *dentry)
163 /* Can't go inside filter_ddelete because it can block */
164 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
165 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
166 LASSERT(atomic_read(&dentry->d_count) > 0);
171 static void init_brw_stats(struct brw_stats *brw_stats)
174 for (i = 0; i < BRW_LAST; i++)
175 spin_lock_init(&brw_stats->hist[i].oh_lock);
178 static int lprocfs_init_rw_stats(struct obd_device *obd,
179 struct lprocfs_stats **stats)
183 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
184 LPROC_FILTER_LAST - 1;
185 *stats = lprocfs_alloc_stats(num_stats, 0);
189 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
190 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
191 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
192 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
193 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
198 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
199 plus the procfs overhead :( */
200 static int filter_export_stats_init(struct obd_device *obd,
201 struct obd_export *exp,
204 struct filter_export_data *fed = &exp->exp_filter_data;
208 init_brw_stats(&fed->fed_brw_stats);
210 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
211 /* Self-export gets no proc entry */
214 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
219 struct nid_stat *tmp = exp->exp_nid_stats;
220 LASSERT(tmp != NULL);
222 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
223 if (tmp->nid_brw_stats == NULL)
226 init_brw_stats(tmp->nid_brw_stats);
227 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
228 0644, &filter_per_nid_stats_fops,
231 CWARN("Error adding the brw_stats file\n");
233 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
237 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
246 /* Add client data to the FILTER. We use a bitmap to locate a free space
247 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
248 * Otherwise, we have just read the data from the last_rcvd file and
249 * we know its offset. */
250 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
253 struct filter_obd *filter = &obd->u.filter;
254 struct filter_export_data *fed = &exp->exp_filter_data;
255 unsigned long *bitmap = filter->fo_last_rcvd_slots;
256 int new_client = (cl_idx == -1);
260 LASSERT(bitmap != NULL);
261 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
264 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
267 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
268 * there's no need for extra complication here
271 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
273 if (cl_idx >= LR_MAX_CLIENTS) {
274 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
278 if (test_and_set_bit(cl_idx, bitmap)) {
279 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
284 if (test_and_set_bit(cl_idx, bitmap)) {
285 CERROR("FILTER client %d: bit already set in bitmap!\n",
291 fed->fed_lr_idx = cl_idx;
292 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
293 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
294 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
296 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
297 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
300 struct lvfs_run_ctxt saved;
301 loff_t off = fed->fed_lr_off;
305 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
306 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
308 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
309 /* Transaction needed to fix bug 1403 */
310 handle = fsfilt_start(obd,
311 filter->fo_rcvd_filp->f_dentry->d_inode,
312 FSFILT_OP_SETATTR, NULL);
313 if (IS_ERR(handle)) {
314 rc = PTR_ERR(handle);
315 CERROR("unable to start transaction: rc %d\n", rc);
317 rc = fsfilt_add_journal_cb(obd, 0, handle,
318 target_client_add_cb, exp);
320 spin_lock(&exp->exp_lock);
321 exp->exp_need_sync = 1;
322 spin_unlock(&exp->exp_lock);
324 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
326 sizeof(*fed->fed_lcd),
327 &off, rc /* sync if no cb */);
329 filter->fo_rcvd_filp->f_dentry->d_inode,
332 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
335 CERROR("error writing %s client idx %u: rc %d\n",
336 LAST_RCVD, fed->fed_lr_idx, rc);
343 static int filter_client_free(struct obd_export *exp)
345 struct filter_export_data *fed = &exp->exp_filter_data;
346 struct filter_obd *filter = &exp->exp_obd->u.filter;
347 struct obd_device *obd = exp->exp_obd;
348 struct lsd_client_data zero_lcd;
349 struct lvfs_run_ctxt saved;
354 if (fed->fed_lcd == NULL)
357 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
358 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
361 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
362 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
364 LASSERT(filter->fo_last_rcvd_slots != NULL);
366 off = fed->fed_lr_off;
368 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
369 * we leak a client slot that will be cleaned on the next recovery. */
371 CERROR("%s: client idx %d has med_off %lld\n",
372 obd->obd_name, fed->fed_lr_idx, off);
373 GOTO(free, rc = -EINVAL);
376 /* Clear the bit _after_ zeroing out the client so we don't
377 race with filter_client_add and zero out new clients.*/
378 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
379 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
384 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
385 memset(&zero_lcd, 0, sizeof zero_lcd);
386 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
387 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
388 sizeof(zero_lcd), &off,
389 (!exp->exp_libclient ||
390 exp->exp_need_sync));
392 /* update server's transno */
393 filter_update_server_data(obd, filter->fo_rcvd_filp,
395 !exp->exp_libclient);
396 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
398 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
399 "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
400 fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
404 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
405 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
412 OBD_FREE_PTR(fed->fed_lcd);
418 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
419 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
420 struct filter_mod_data *fmd)
422 LASSERT_SPIN_LOCKED(&fed->fed_lock);
423 if (--fmd->fmd_refcount == 0) {
424 /* XXX when we have persistent reservations and the handle
425 * is stored herein we need to drop it here. */
426 fed->fed_mod_count--;
427 list_del(&fmd->fmd_list);
428 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
432 /* drop fmd reference, free it if last ref */
433 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
435 struct filter_export_data *fed;
440 fed = &exp->exp_filter_data;
441 spin_lock(&fed->fed_lock);
442 filter_fmd_put_nolock(fed, fmd); /* caller reference */
443 spin_unlock(&fed->fed_lock);
446 /* expire entries from the end of the list if there are too many
447 * or they are too old */
448 static void filter_fmd_expire_nolock(struct filter_obd *filter,
449 struct filter_export_data *fed,
450 struct filter_mod_data *keep)
452 struct filter_mod_data *fmd, *tmp;
454 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
458 if (time_before(jiffies, fmd->fmd_expire) &&
459 fed->fed_mod_count < filter->fo_fmd_max_num)
462 list_del_init(&fmd->fmd_list);
463 filter_fmd_put_nolock(fed, fmd); /* list reference */
467 void filter_fmd_expire(struct obd_export *exp)
469 spin_lock(&exp->exp_filter_data.fed_lock);
470 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
471 &exp->exp_filter_data, NULL);
472 spin_unlock(&exp->exp_filter_data.fed_lock);
475 /* find specified objid, group in export fmd list.
476 * caller must hold fed_lock and take fmd reference itself */
477 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
478 struct filter_export_data *fed,
479 obd_id objid, obd_gr group)
481 struct filter_mod_data *found = NULL, *fmd;
483 LASSERT_SPIN_LOCKED(&fed->fed_lock);
485 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
486 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
488 list_del(&fmd->fmd_list);
489 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
490 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
495 filter_fmd_expire_nolock(filter, fed, found);
500 /* Find fmd based on objid and group, or return NULL if not found. */
501 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
502 obd_id objid, obd_gr group)
504 struct filter_mod_data *fmd;
506 spin_lock(&exp->exp_filter_data.fed_lock);
507 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
508 &exp->exp_filter_data, objid, group);
510 fmd->fmd_refcount++; /* caller reference */
511 spin_unlock(&exp->exp_filter_data.fed_lock);
516 /* Find fmd based on objid and group, or create a new one if none is found.
517 * It is possible for this function to return NULL under memory pressure,
518 * or if objid = 0 is passed (which will only cause old entries to expire).
519 * Currently this is not fatal because any fmd state is transient and
520 * may also be freed when it gets sufficiently old. */
521 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
522 obd_id objid, obd_gr group)
524 struct filter_export_data *fed = &exp->exp_filter_data;
525 struct filter_mod_data *found = NULL, *fmd_new = NULL;
527 OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO, sizeof(*fmd_new));
529 spin_lock(&fed->fed_lock);
530 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
533 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
534 fmd_new->fmd_id = objid;
535 fmd_new->fmd_gr = group;
536 fmd_new->fmd_refcount++; /* list reference */
538 fed->fed_mod_count++;
540 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
544 found->fmd_refcount++; /* caller reference */
545 found->fmd_expire = jiffies +
546 exp->exp_obd->u.filter.fo_fmd_max_age;
549 spin_unlock(&fed->fed_lock);
555 /* drop fmd list reference so it will disappear when last reference is put.
556 * This isn't so critical because it would in fact only affect the one client
557 * that is doing the unlink and at worst we have an stale entry referencing
558 * an object that should never be used again. */
559 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
561 struct filter_mod_data *found = NULL;
563 spin_lock(&exp->exp_filter_data.fed_lock);
564 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
566 list_del_init(&found->fmd_list);
567 filter_fmd_put_nolock(&exp->exp_filter_data, found);
569 spin_unlock(&exp->exp_filter_data.fed_lock);
572 #define filter_fmd_drop(exp, objid, group)
575 /* remove all entries from fmd list */
576 static void filter_fmd_cleanup(struct obd_export *exp)
578 struct filter_export_data *fed = &exp->exp_filter_data;
579 struct filter_mod_data *fmd = NULL, *tmp;
581 spin_lock(&fed->fed_lock);
582 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
583 list_del_init(&fmd->fmd_list);
584 filter_fmd_put_nolock(fed, fmd);
586 spin_unlock(&fed->fed_lock);
589 static int filter_init_export(struct obd_export *exp)
591 spin_lock_init(&exp->exp_filter_data.fed_lock);
592 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
594 spin_lock(&exp->exp_lock);
595 exp->exp_connecting = 1;
596 spin_unlock(&exp->exp_lock);
601 static int filter_free_server_data(struct filter_obd *filter)
603 OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
604 filter->fo_fsd = NULL;
605 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
606 filter->fo_last_rcvd_slots = NULL;
610 /* assumes caller is already in kernel ctxt */
611 int filter_update_server_data(struct obd_device *obd, struct file *filp,
612 struct lr_server_data *fsd, int force_sync)
618 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
619 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
620 le64_to_cpu(fsd->lsd_last_transno));
621 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
622 le64_to_cpu(fsd->lsd_mount_count));
624 fsd->lsd_compat14 = fsd->lsd_last_transno;
625 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
627 CERROR("error writing lr_server_data: rc = %d\n", rc);
632 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
635 struct filter_obd *filter = &obd->u.filter;
641 if (filter->fo_last_objid_files[group] == NULL) {
642 CERROR("Object group "LPU64" not fully setup; not updating "
643 "last_objid\n", group);
647 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
648 obd->obd_name, group, filter->fo_last_objids[group]);
650 tmp = cpu_to_le64(filter->fo_last_objids[group]);
651 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
652 &tmp, sizeof(tmp), &off, force_sync);
654 CERROR("error writing group "LPU64" last objid: rc = %d\n",
658 extern int ost_handle(struct ptlrpc_request *req);
659 /* assumes caller has already in kernel ctxt */
660 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
662 struct filter_obd *filter = &obd->u.filter;
663 struct lr_server_data *fsd;
664 struct lsd_client_data *lcd = NULL;
665 struct inode *inode = filp->f_dentry->d_inode;
666 unsigned long last_rcvd_size = i_size_read(inode);
672 /* ensure padding in the struct is the correct size */
673 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
674 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
675 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
676 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
678 OBD_ALLOC(fsd, sizeof(*fsd));
681 filter->fo_fsd = fsd;
683 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
684 if (filter->fo_last_rcvd_slots == NULL) {
685 OBD_FREE(fsd, sizeof(*fsd));
689 if (last_rcvd_size == 0) {
690 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
692 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
693 fsd->lsd_last_transno = 0;
694 mount_count = fsd->lsd_mount_count = 0;
695 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
696 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
697 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
698 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
699 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
700 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
702 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
704 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
708 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
709 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
710 "using the wrong disk %s. Were the "
711 "/dev/ assignments rearranged?\n",
712 obd->obd_uuid.uuid, fsd->lsd_uuid);
713 GOTO(err_fsd, rc = -EINVAL);
715 mount_count = le64_to_cpu(fsd->lsd_mount_count);
716 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
718 /* Assume old last_rcvd format unless I_C_LR is set */
719 if (!(fsd->lsd_feature_incompat &
720 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
721 fsd->lsd_last_transno = fsd->lsd_compat14;
725 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
726 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
727 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
728 ~FILTER_INCOMPAT_SUPP);
729 GOTO(err_fsd, rc = -EINVAL);
731 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
732 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
733 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
734 ~FILTER_ROCOMPAT_SUPP);
735 /* Do something like remount filesystem read-only */
736 GOTO(err_fsd, rc = -EINVAL);
739 CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
740 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
741 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
742 obd->obd_name, mount_count + 1);
743 CDEBUG(D_INODE, "%s: server data size: %u\n",
744 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
745 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
746 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
747 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
748 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
749 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
750 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
751 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
752 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
753 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
754 le16_to_cpu(fsd->lsd_client_size));
756 if (!obd->obd_replayable) {
757 CWARN("%s: recovery support OFF\n", obd->obd_name);
761 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
762 off < last_rcvd_size; cl_idx++) {
764 struct obd_export *exp;
765 struct filter_export_data *fed;
770 GOTO(err_client, rc = -ENOMEM);
773 /* Don't assume off is incremented properly by
774 * fsfilt_read_record(), in case sizeof(*lcd)
775 * isn't the same as fsd->lsd_client_size. */
776 off = le32_to_cpu(fsd->lsd_client_start) +
777 cl_idx * le16_to_cpu(fsd->lsd_client_size);
778 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
780 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
781 LAST_RCVD, cl_idx, off, rc);
782 break; /* read error shouldn't cause startup to fail */
785 if (lcd->lcd_uuid[0] == '\0') {
786 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
791 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
793 /* These exports are cleaned up by filter_disconnect(), so they
794 * need to be set up like real exports as filter_connect() does.
796 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
798 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
799 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
800 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
802 if (PTR_ERR(exp) == -EALREADY) {
803 /* export already exists, zero out this one */
804 CERROR("Zeroing out duplicate export due to "
806 lcd->lcd_uuid[0] = '\0';
808 GOTO(err_client, rc = PTR_ERR(exp));
811 fed = &exp->exp_filter_data;
813 fed->fed_group = 0; /* will be assigned at connect */
814 filter_export_stats_init(obd, exp, NULL);
815 rc = filter_client_add(obd, exp, cl_idx);
816 /* can't fail for existing client */
817 LASSERTF(rc == 0, "rc = %d\n", rc);
820 spin_lock(&exp->exp_lock);
821 exp->exp_connecting = 0;
822 exp->exp_in_recovery = 0;
823 spin_unlock(&exp->exp_lock);
824 obd->obd_max_recoverable_clients++;
825 class_export_put(exp);
828 /* Need to check last_rcvd even for duplicated exports. */
829 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
832 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
833 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
839 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
841 target_recovery_init(obd, ost_handle);
844 filter->fo_mount_count = mount_count + 1;
845 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
847 /* save it, so mount count and last_transno is current */
848 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
850 GOTO(err_client, rc);
855 target_recovery_fini(obd);
857 filter_free_server_data(filter);
861 static int filter_cleanup_groups(struct obd_device *obd)
863 struct filter_obd *filter = &obd->u.filter;
865 struct dentry *dentry;
869 if (filter->fo_dentry_O_groups != NULL) {
870 for (i = 0; i < filter->fo_group_count; i++) {
871 dentry = filter->fo_dentry_O_groups[i];
875 OBD_FREE(filter->fo_dentry_O_groups,
876 filter->fo_group_count *
877 sizeof(*filter->fo_dentry_O_groups));
878 filter->fo_dentry_O_groups = NULL;
880 if (filter->fo_last_objid_files != NULL) {
881 for (i = 0; i < filter->fo_group_count; i++) {
882 filp = filter->fo_last_objid_files[i];
886 OBD_FREE(filter->fo_last_objid_files,
887 filter->fo_group_count *
888 sizeof(*filter->fo_last_objid_files));
889 filter->fo_last_objid_files = NULL;
891 if (filter->fo_dentry_O_sub != NULL) {
892 for (i = 0; i < filter->fo_group_count; i++) {
893 for (j = 0; j < filter->fo_subdir_count; j++) {
894 dentry = filter->fo_dentry_O_sub[i].dentry[j];
899 OBD_FREE(filter->fo_dentry_O_sub,
900 filter->fo_group_count *
901 sizeof(*filter->fo_dentry_O_sub));
902 filter->fo_dentry_O_sub = NULL;
904 if (filter->fo_last_objids != NULL) {
905 OBD_FREE(filter->fo_last_objids,
906 filter->fo_group_count *
907 sizeof(*filter->fo_last_objids));
908 filter->fo_last_objids = NULL;
910 if (filter->fo_dentry_O != NULL) {
911 f_dput(filter->fo_dentry_O);
912 filter->fo_dentry_O = NULL;
917 static int filter_update_last_group(struct obd_device *obd, int group)
919 struct filter_obd *filter = &obd->u.filter;
920 struct file *filp = NULL;
921 int last_group = 0, rc;
925 if (group <= filter->fo_committed_group)
928 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
932 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
936 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
938 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
941 LASSERT(off == 0 || last_group >= FILTER_MIN_GROUPS);
942 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
943 obd->obd_name, last_group, group);
947 /* must be sync: bXXXX */
948 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
950 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
954 filter->fo_committed_group = group;
961 static int filter_read_group_internal(struct obd_device *obd, int group,
964 struct filter_obd *filter = &obd->u.filter;
965 __u64 *new_objids = NULL;
966 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
967 struct dentry **new_groups = NULL;
968 struct file **new_files = NULL;
969 struct dentry *dentry;
971 int old_count = filter->fo_group_count, rc, stage = 0, i;
977 snprintf(name, 24, "%d", group);
981 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
983 if (IS_ERR(dentry)) {
984 CERROR("Cannot lookup expected object group %d: %ld\n",
985 group, PTR_ERR(dentry));
986 RETURN(PTR_ERR(dentry));
989 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
991 if (IS_ERR(dentry)) {
992 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
994 RETURN(PTR_ERR(dentry));
999 snprintf(name, 24, "O/%d/LAST_ID", group);
1001 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1003 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1004 GOTO(cleanup, rc = PTR_ERR(filp));
1008 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1010 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1014 if (filter->fo_subdir_count) {
1015 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1016 if (tmp_subdirs == NULL)
1017 GOTO(cleanup, rc = -ENOMEM);
1020 for (i = 0; i < filter->fo_subdir_count; i++) {
1022 snprintf(dir, sizeof(dir), "d%u", i);
1024 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1027 if (IS_ERR(tmp_subdirs->dentry[i])) {
1028 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1029 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1034 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1035 tmp_subdirs->dentry[i]);
1039 /* 'group' is an index; we need an array of length 'group + 1' */
1040 if (group + 1 > old_count) {
1041 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1042 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1043 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1044 OBD_ALLOC(new_files, len * sizeof(*new_files));
1046 if (new_objids == NULL || new_subdirs == NULL ||
1047 new_groups == NULL || new_files == NULL)
1048 GOTO(cleanup, rc = -ENOMEM);
1051 memcpy(new_objids, filter->fo_last_objids,
1052 old_count * sizeof(*new_objids));
1053 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1054 old_count * sizeof(*new_subdirs));
1055 memcpy(new_groups, filter->fo_dentry_O_groups,
1056 old_count * sizeof(*new_groups));
1057 memcpy(new_files, filter->fo_last_objid_files,
1058 old_count * sizeof(*new_files));
1060 OBD_FREE(filter->fo_last_objids,
1061 old_count * sizeof(*new_objids));
1062 OBD_FREE(filter->fo_dentry_O_sub,
1063 old_count * sizeof(*new_subdirs));
1064 OBD_FREE(filter->fo_dentry_O_groups,
1065 old_count * sizeof(*new_groups));
1066 OBD_FREE(filter->fo_last_objid_files,
1067 old_count * sizeof(*new_files));
1069 filter->fo_last_objids = new_objids;
1070 filter->fo_dentry_O_sub = new_subdirs;
1071 filter->fo_dentry_O_groups = new_groups;
1072 filter->fo_last_objid_files = new_files;
1073 filter->fo_group_count = len;
1076 filter->fo_dentry_O_groups[group] = dentry;
1077 filter->fo_last_objid_files[group] = filp;
1078 if (filter->fo_subdir_count) {
1079 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1080 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1083 filter_update_last_group(obd, group);
1085 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1086 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1087 rc = filter_update_last_objid(obd, group, 1);
1091 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1092 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1093 obd->obd_name, group, last_objid);
1098 if (new_objids != NULL)
1099 OBD_FREE(new_objids, len * sizeof(*new_objids));
1100 if (new_subdirs != NULL)
1101 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1102 if (new_groups != NULL)
1103 OBD_FREE(new_groups, len * sizeof(*new_groups));
1104 if (new_files != NULL)
1105 OBD_FREE(new_files, len * sizeof(*new_files));
1107 if (filter->fo_subdir_count) {
1108 for (i = 0; i < filter->fo_subdir_count; i++) {
1109 if (tmp_subdirs->dentry[i] != NULL)
1110 dput(tmp_subdirs->dentry[i]);
1112 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1115 filp_close(filp, 0);
1122 static int filter_read_groups(struct obd_device *obd, int last_group,
1125 struct filter_obd *filter = &obd->u.filter;
1126 int old_count, group, rc = 0;
1128 down(&filter->fo_init_lock);
1129 old_count = filter->fo_group_count;
1130 for (group = old_count; group <= last_group; group++) {
1132 continue; /* no group zero */
1134 rc = filter_read_group_internal(obd, group, create);
1138 up(&filter->fo_init_lock);
1142 /* FIXME: object groups */
1143 static int filter_prep_groups(struct obd_device *obd)
1145 struct filter_obd *filter = &obd->u.filter;
1146 struct dentry *dentry, *O_dentry;
1148 int last_group, rc = 0, cleanup_phase = 0;
1152 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1154 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1155 if (IS_ERR(O_dentry)) {
1156 rc = PTR_ERR(O_dentry);
1157 CERROR("cannot open/create O: rc = %d\n", rc);
1160 filter->fo_dentry_O = O_dentry;
1161 cleanup_phase = 1; /* O_dentry */
1163 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1164 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1165 * some time post 1.0 when all old-style OSTs have converted along
1166 * with the init_objid hack. */
1167 dentry = ll_lookup_one_len("R", O_dentry, 1);
1169 GOTO(cleanup, rc = PTR_ERR(dentry));
1170 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1171 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1174 CWARN("converting OST to new object layout\n");
1175 if (IS_ERR(O0_dentry)) {
1176 rc = PTR_ERR(O0_dentry);
1177 CERROR("error looking up O/0: rc %d\n", rc);
1178 GOTO(cleanup_R, rc);
1181 if (O0_dentry->d_inode) {
1182 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1183 GOTO(cleanup_O0, rc = -EEXIST);
1186 LOCK_INODE_MUTEX(O_dentry->d_inode);
1187 rc = ll_vfs_rename(O_dentry->d_inode, dentry, filter->fo_vfsmnt,
1188 O_dentry->d_inode, O0_dentry,
1190 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1193 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1194 GOTO(cleanup_O0, rc);
1196 filter->fo_fsd->lsd_feature_incompat |=
1197 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1198 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1200 GOTO(cleanup_O0, rc);
1212 cleanup_phase = 2; /* groups */
1214 /* we have to initialize all groups before first connections from
1215 * clients because they may send create/destroy for any group -bzzz */
1216 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1218 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1219 GOTO(cleanup, rc = PTR_ERR(filp));
1221 cleanup_phase = 3; /* filp */
1223 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1225 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1229 last_group = FILTER_MIN_GROUPS;
1231 LASSERT(last_group >= FILTER_MIN_GROUPS);
1234 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1235 FILTER_MIN_GROUPS, last_group);
1236 filter->fo_committed_group = last_group;
1237 rc = filter_read_groups(obd, last_group, 1);
1241 filp_close(filp, 0);
1245 switch (cleanup_phase) {
1247 filp_close(filp, 0);
1249 filter_cleanup_groups(obd);
1251 f_dput(filter->fo_dentry_O);
1252 filter->fo_dentry_O = NULL;
1260 /* setup the object store with correct subdirectories */
1261 static int filter_prep(struct obd_device *obd)
1263 struct lvfs_run_ctxt saved;
1264 struct filter_obd *filter = &obd->u.filter;
1266 struct inode *inode;
1270 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1271 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1272 if (!file || IS_ERR(file)) {
1274 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1278 filter->fo_rcvd_filp = file;
1279 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1280 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1281 file->f_dentry->d_inode->i_mode);
1282 GOTO(err_filp, rc = -ENOENT);
1285 inode = file->f_dentry->d_parent->d_inode;
1286 /* We use i_op->unlink directly in filter_vfs_unlink() */
1287 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1288 CERROR("%s: filesystem does not support create/unlink ops\n",
1290 GOTO(err_filp, rc = -EOPNOTSUPP);
1293 rc = filter_init_server_data(obd, file);
1295 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1298 /* open and create health check io file*/
1299 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1302 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1304 GOTO(err_server_data, rc);
1306 filter->fo_health_check_filp = file;
1307 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1308 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1309 file->f_dentry->d_inode->i_mode);
1310 GOTO(err_health_check, rc = -ENOENT);
1312 rc = lvfs_check_io_health(obd, file);
1314 GOTO(err_health_check, rc);
1316 rc = filter_prep_groups(obd);
1318 GOTO(err_health_check, rc);
1320 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1325 if (filp_close(filter->fo_health_check_filp, 0))
1326 CERROR("can't close %s after error\n", HEALTH_CHECK);
1327 filter->fo_health_check_filp = NULL;
1329 target_recovery_fini(obd);
1330 filter_free_server_data(filter);
1332 if (filp_close(filter->fo_rcvd_filp, 0))
1333 CERROR("can't close %s after error\n", LAST_RCVD);
1334 filter->fo_rcvd_filp = NULL;
1338 /* cleanup the filter: write last used object id to status file */
1339 static void filter_post(struct obd_device *obd)
1341 struct lvfs_run_ctxt saved;
1342 struct filter_obd *filter = &obd->u.filter;
1345 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1346 * best to start a transaction with h_sync, because we removed this
1349 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1350 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1353 CERROR("error writing server data: rc = %d\n", rc);
1355 for (i = 1; i < filter->fo_group_count; i++) {
1356 rc = filter_update_last_objid(obd, i,
1357 (i == filter->fo_group_count - 1));
1359 CERROR("error writing group %d lastobjid: rc = %d\n",
1363 rc = filp_close(filter->fo_rcvd_filp, 0);
1364 filter->fo_rcvd_filp = NULL;
1366 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1368 rc = filp_close(filter->fo_health_check_filp, 0);
1369 filter->fo_health_check_filp = NULL;
1371 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1373 filter_cleanup_groups(obd);
1374 filter_free_server_data(filter);
1375 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1377 filter_free_capa_keys(filter);
1378 cleanup_capa_hash(filter->fo_capa_hash);
1381 static void filter_set_last_id(struct filter_obd *filter,
1382 obd_id id, obd_gr group)
1384 LASSERT(filter->fo_fsd != NULL);
1385 LASSERT(group <= filter->fo_group_count);
1387 spin_lock(&filter->fo_objidlock);
1388 filter->fo_last_objids[group] = id;
1389 spin_unlock(&filter->fo_objidlock);
1392 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1395 LASSERT(filter->fo_fsd != NULL);
1396 LASSERT(group <= filter->fo_group_count);
1398 /* FIXME: object groups */
1399 spin_lock(&filter->fo_objidlock);
1400 id = filter->fo_last_objids[group];
1401 spin_unlock(&filter->fo_objidlock);
1406 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1408 LOCK_INODE_MUTEX(dparent->d_inode);
1412 /* We never dget the object parent, so DON'T dput it either */
1413 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1415 struct filter_obd *filter = &obd->u.filter;
1416 struct filter_subdirs *subdirs;
1417 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1419 if (group > 0 || filter->fo_subdir_count == 0)
1420 return filter->fo_dentry_O_groups[group];
1422 subdirs = &filter->fo_dentry_O_sub[group];
1423 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1426 /* We never dget the object parent, so DON'T dput it either */
1427 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1430 unsigned long now = jiffies;
1431 struct dentry *dparent = filter_parent(obd, group, objid);
1434 if (IS_ERR(dparent))
1436 if (dparent == NULL)
1437 return ERR_PTR(-ENOENT);
1439 rc = filter_lock_dentry(obd, dparent);
1440 fsfilt_check_slow(obd, now, "parent lock");
1441 return rc ? ERR_PTR(rc) : dparent;
1444 /* We never dget the object parent, so DON'T dput it either */
1445 static void filter_parent_unlock(struct dentry *dparent)
1447 UNLOCK_INODE_MUTEX(dparent->d_inode);
1450 /* How to get files, dentries, inodes from object id's.
1452 * If dir_dentry is passed, the caller has already locked the parent
1453 * appropriately for this operation (normally a write lock). If
1454 * dir_dentry is NULL, we do a read lock while we do the lookup to
1455 * avoid races with create/destroy and such changing the directory
1456 * internal to the filesystem code. */
1457 struct dentry *filter_fid2dentry(struct obd_device *obd,
1458 struct dentry *dir_dentry,
1459 obd_gr group, obd_id id)
1461 struct dentry *dparent = dir_dentry;
1462 struct dentry *dchild;
1467 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1468 obd->u.filter.fo_destroys_in_progress == 0) {
1469 /* don't fail lookups for orphan recovery, it causes
1470 * later LBUGs when objects still exist during precreate */
1471 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1472 RETURN(ERR_PTR(-ENOENT));
1475 CERROR("fatal: invalid object id 0\n");
1476 RETURN(ERR_PTR(-ESTALE));
1479 len = sprintf(name, LPU64, id);
1480 if (dir_dentry == NULL) {
1481 dparent = filter_parent_lock(obd, group, id);
1482 if (IS_ERR(dparent)) {
1483 CERROR("%s: error getting object "LPU64":"LPU64
1484 " parent: rc %ld\n", obd->obd_name,
1485 id, group, PTR_ERR(dparent));
1489 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1490 dparent->d_name.len, dparent->d_name.name, name);
1491 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1492 if (dir_dentry == NULL)
1493 filter_parent_unlock(dparent);
1494 if (IS_ERR(dchild)) {
1495 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1500 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1501 CERROR("%s: got bad object "LPU64" inode %lu\n",
1502 obd->obd_name, id, dchild->d_inode->i_ino);
1504 RETURN(ERR_PTR(-ENOENT));
1507 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1508 name, dchild, atomic_read(&dchild->d_count));
1510 LASSERT(atomic_read(&dchild->d_count) > 0);
1515 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1518 struct lustre_handle lockh;
1519 int flags = LDLM_AST_DISCARD_DATA, rc;
1520 struct ldlm_res_id res_id;
1521 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1524 osc_build_res_name(objid, group, &res_id);
1525 /* Tell the clients that the object is gone now and that they should
1526 * throw away any cached pages. */
1527 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1528 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1529 ldlm_completion_ast, NULL, NULL, 0, NULL,
1532 /* We only care about the side-effects, just drop the lock. */
1534 ldlm_lock_decref(&lockh, LCK_PW);
1539 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1540 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1541 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1542 * reverse. Caller must take i_sem before starting the transaction and we
1543 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1544 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1545 struct vfsmount *mnt)
1550 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1551 * itself we already hold dir->i_mutex for child create/unlink ops */
1552 LASSERT(dentry->d_inode != NULL);
1553 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1554 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1558 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1559 GOTO(out, rc = -ENOENT);
1561 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1566 GOTO(out, rc = -EPERM);
1568 /* check_sticky() */
1569 if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1570 IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1571 GOTO(out, rc = -EPERM);
1573 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1574 * that was done because of journal_start (which is already done
1575 * here) or some other ordering issue. */
1578 rc = ll_security_inode_unlink(dir, dentry, mnt);
1582 rc = dir->i_op->unlink(dir, dentry);
1584 /* need to drop i_mutex before we lose inode reference */
1585 UNLOCK_INODE_MUTEX(dentry->d_inode);
1592 /* Caller must hold LCK_PW on parent and push us into kernel context.
1593 * Caller must hold child i_mutex, we drop it always.
1594 * Caller is also required to ensure that dchild->d_inode exists. */
1595 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1596 obd_gr group, struct dentry *dparent,
1597 struct dentry *dchild)
1599 struct inode *inode = dchild->d_inode;
1600 struct filter_obd *filter = &obd->u.filter;
1603 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1604 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1605 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1606 (unsigned long)inode->i_nlink,
1607 atomic_read(&inode->i_count));
1610 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1612 CERROR("error unlinking objid %.*s: rc %d\n",
1613 dchild->d_name.len, dchild->d_name.name, rc);
1617 struct filter_intent_args {
1618 struct ldlm_lock **victim;
1623 static enum interval_iter filter_intent_cb(struct interval_node *n,
1626 struct ldlm_interval *node = (struct ldlm_interval *)n;
1627 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1628 __u64 size = arg->size;
1629 struct ldlm_lock **v = arg->victim;
1630 struct ldlm_lock *lck;
1632 /* If the interval is lower than the current file size,
1634 if (interval_high(n) <= size)
1635 return INTERVAL_ITER_STOP;
1637 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1638 /* Don't send glimpse ASTs to liblustre clients.
1639 * They aren't listening for them, and they do
1640 * entirely synchronous I/O anyways. */
1641 if (lck->l_export == NULL ||
1642 lck->l_export->exp_libclient == 1)
1645 if (*arg->liblustre)
1646 *arg->liblustre = 0;
1649 *v = LDLM_LOCK_GET(lck);
1650 } else if ((*v)->l_policy_data.l_extent.start <
1651 lck->l_policy_data.l_extent.start) {
1653 *v = LDLM_LOCK_GET(lck);
1656 /* the same policy group - every lock has the
1657 * same extent, so needn't do it any more */
1661 return INTERVAL_ITER_CONT;
1664 static int filter_intent_policy(struct ldlm_namespace *ns,
1665 struct ldlm_lock **lockp, void *req_cookie,
1666 ldlm_mode_t mode, int flags, void *data)
1668 CFS_LIST_HEAD(rpc_list);
1669 struct ptlrpc_request *req = req_cookie;
1670 struct ldlm_lock *lock = *lockp, *l = NULL;
1671 struct ldlm_resource *res = lock->l_resource;
1672 ldlm_processing_policy policy;
1673 struct ost_lvb *res_lvb, *reply_lvb;
1674 struct ldlm_reply *rep;
1676 int idx, rc, tmpflags = 0, only_liblustre = 1;
1677 struct ldlm_interval_tree *tree;
1678 struct filter_intent_args arg;
1679 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1680 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1681 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1684 policy = ldlm_get_processing_policy(res);
1685 LASSERT(policy != NULL);
1686 LASSERT(req != NULL);
1688 rc = lustre_pack_reply(req, 3, repsize, NULL);
1690 RETURN(req->rq_status = rc);
1692 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1693 LASSERT(rep != NULL);
1695 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1696 sizeof(*reply_lvb));
1697 LASSERT(reply_lvb != NULL);
1699 //fixup_handle_for_resent_req(req, lock, &lockh);
1701 /* If we grant any lock at all, it will be a whole-file read lock.
1702 * Call the extent policy function to see if our request can be
1703 * granted, or is blocked.
1704 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse lock
1706 lock->l_policy_data.l_extent.start = 0;
1707 lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1708 lock->l_req_mode = LCK_PR;
1710 LASSERT(ns == res->lr_namespace);
1712 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1713 check_res_locked(res);
1715 /* FIXME: we should change the policy function slightly, to not make
1716 * this list at all, since we just turn around and free it */
1717 while (!list_empty(&rpc_list)) {
1718 struct ldlm_lock *wlock =
1719 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1720 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1721 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1722 lock->l_flags &= ~LDLM_FL_CP_REQD;
1723 list_del_init(&wlock->l_cp_ast);
1724 LDLM_LOCK_PUT(wlock);
1727 /* The lock met with no resistance; we're finished. */
1728 if (rc == LDLM_ITER_CONTINUE) {
1729 /* do not grant locks to the liblustre clients: they cannot
1730 * handle ASTs robustly. We need to do this while still
1731 * holding ns_lock to avoid the lock remaining on the res_link
1732 * list (and potentially being added to l_pending_list by an
1733 * AST) when we are going to drop this lock ASAP. */
1734 if (lock->l_export->exp_libclient ||
1735 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1736 ldlm_resource_unlink_lock(lock);
1737 err = ELDLM_LOCK_ABORTED;
1739 err = ELDLM_LOCK_REPLACED;
1745 /* Do not grant any lock, but instead send GL callbacks. The extent
1746 * policy nicely created a list of all PW locks for us. We will choose
1747 * the highest of those which are larger than the size in the LVB, if
1748 * any, and perform a glimpse callback. */
1749 res_lvb = res->lr_lvb_data;
1750 LASSERT(res_lvb != NULL);
1751 *reply_lvb = *res_lvb;
1754 * ->ns_lock guarantees that no new locks are granted, and,
1755 * therefore, that res->lr_lvb_data cannot increase beyond the
1756 * end of already granted lock. As a result, it is safe to
1757 * check against "stale" reply_lvb->lvb_size value without
1760 arg.size = reply_lvb->lvb_size;
1762 arg.liblustre = &only_liblustre;
1763 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1764 tree = &res->lr_itree[idx];
1765 if (tree->lit_mode == LCK_PR)
1768 interval_iterate_reverse(tree->lit_root,
1769 filter_intent_cb, &arg);
1773 /* There were no PW locks beyond the size in the LVB; finished. */
1775 if (only_liblustre) {
1776 /* If we discovered a liblustre client with a PW lock,
1777 * however, the LVB may be out of date! The LVB is
1778 * updated only on glimpse (which we don't do for
1779 * liblustre clients) and cancel (which the client
1780 * obviously has not yet done). So if it has written
1781 * data but kept the lock, the LVB is stale and needs
1782 * to be updated from disk.
1784 * Of course, this will all disappear when we switch to
1785 * taking liblustre locks on the OST. */
1786 ldlm_res_lvbo_update(res, NULL, 0, 1);
1788 RETURN(ELDLM_LOCK_ABORTED);
1792 * This check is for lock taken in filter_prepare_destroy() that does
1793 * not have l_glimpse_ast set. So the logic is: if there is a lock
1794 * with no l_glimpse_ast set, this object is being destroyed already.
1796 * Hence, if you are grabbing DLM locks on the server, always set
1797 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1799 if (l->l_glimpse_ast == NULL) {
1800 /* We are racing with unlink(); just return -ENOENT */
1801 rep->lock_policy_res1 = -ENOENT;
1805 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1806 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1807 /* Update the LVB from disk if the AST failed (this is a legal race) */
1809 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1810 * sending ast is not handled. This can result in lost client writes.
1813 ldlm_res_lvbo_update(res, NULL, 0, 1);
1816 *reply_lvb = *res_lvb;
1822 RETURN(ELDLM_LOCK_ABORTED);
1826 * per-obd_device iobuf pool.
1828 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1829 * allocations in write-path has to be minimized (see bug 5137).
1831 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1832 * OST threads (see ost_thread_{init,done}()).
1834 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1835 * at the OST layer there are only (potentially) multiple obd_device of type
1836 * unknown at the time of OST thread creation.
1838 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1839 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1840 * it's very own iobuf.
1844 * filter_kiobuf_pool_init()
1846 * filter_kiobuf_pool_done()
1848 * filter_iobuf_get()
1850 * operate on this array. They are "generic" in a sense that they don't depend
1851 * on actual type of iobuf's (the latter depending on Linux kernel version).
1855 * destroy pool created by filter_iobuf_pool_init
1857 static void filter_iobuf_pool_done(struct filter_obd *filter)
1859 struct filter_iobuf **pool;
1864 pool = filter->fo_iobuf_pool;
1866 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1867 if (pool[i] != NULL)
1868 filter_free_iobuf(pool[i]);
1870 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1871 filter->fo_iobuf_pool = NULL;
1877 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1879 static int filter_iobuf_pool_init(struct filter_obd *filter)
1886 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1888 if (filter->fo_iobuf_pool == NULL)
1891 filter->fo_iobuf_count = OSS_THREADS_MAX;
1896 /* Return iobuf allocated for @thread_id. We don't know in advance how
1897 * many threads there will be so we allocate a large empty array and only
1898 * fill in those slots that are actually in use.
1899 * If we haven't allocated a pool entry for this thread before, do so now. */
1900 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1902 int thread_id = oti ? oti->oti_thread_id : -1;
1903 struct filter_iobuf *pool = NULL;
1904 struct filter_iobuf **pool_place = NULL;
1906 if (thread_id >= 0) {
1907 LASSERT(thread_id < filter->fo_iobuf_count);
1908 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1911 if (unlikely(pool == NULL)) {
1912 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1913 PTLRPC_MAX_BRW_PAGES);
1914 if (pool_place != NULL)
1921 /* mount the file system (secretly). lustre_cfg parameters are:
1924 * 3 = flags: failover=f, failout=n
1927 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1930 struct filter_obd *filter = &obd->u.filter;
1931 struct vfsmount *mnt;
1932 struct lustre_mount_info *lmi;
1933 struct obd_uuid uuid;
1941 if (lcfg->lcfg_bufcount < 3 ||
1942 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1943 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1946 lmi = server_get_mount(obd->obd_name);
1948 /* We already mounted in lustre_fill_super.
1949 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1950 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1952 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1954 /* old path - used by lctl */
1955 CERROR("Using old MDS mount method\n");
1956 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1957 MS_NOATIME|MS_NODIRATIME,
1958 lustre_cfg_string(lcfg, 1), option);
1961 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1962 lustre_cfg_string(lcfg, 1), rc);
1966 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1968 if (IS_ERR(obd->obd_fsops))
1969 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1971 rc = filter_iobuf_pool_init(filter);
1975 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1977 /* failover is the default */
1978 obd->obd_replayable = 1;
1980 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1981 str = lustre_cfg_string(lcfg, 3);
1982 if (strchr(str, 'n')) {
1983 CWARN("%s: recovery disabled\n", obd->obd_name);
1984 obd->obd_replayable = 0;
1988 filter->fo_vfsmnt = mnt;
1989 obd->u.obt.obt_sb = mnt->mnt_sb;
1990 filter->fo_fstype = mnt->mnt_sb->s_type->name;
1991 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1993 fsfilt_setup(obd, obd->u.obt.obt_sb);
1995 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1996 obd->obd_lvfs_ctxt.pwdmnt = mnt;
1997 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1998 obd->obd_lvfs_ctxt.fs = get_ds();
1999 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2001 init_mutex(&filter->fo_init_lock);
2002 filter->fo_committed_group = 0;
2004 rc = filter_prep(obd);
2008 filter->fo_destroys_in_progress = 0;
2009 for (i = 0; i < 32; i++)
2010 sema_init(&filter->fo_create_locks[i], 1);
2012 spin_lock_init(&filter->fo_translock);
2013 spin_lock_init(&filter->fo_objidlock);
2014 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2015 sema_init(&filter->fo_alloc_lock, 1);
2016 init_brw_stats(&filter->fo_filter_stats);
2017 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2018 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2019 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2021 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2022 spin_lock_init(&filter->fo_llog_list_lock);
2024 filter->fo_sptlrpc_lock = RW_LOCK_UNLOCKED;
2025 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2027 filter->fo_fl_oss_capa = 0;
2028 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2029 filter->fo_capa_hash = init_capa_hash();
2030 if (filter->fo_capa_hash == NULL)
2031 GOTO(err_ops, rc = -ENOMEM);
2033 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2034 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2035 LDLM_NAMESPACE_GREEDY);
2036 if (obd->obd_namespace == NULL)
2037 GOTO(err_post, rc = -ENOMEM);
2038 obd->obd_namespace->ns_lvbp = obd;
2039 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2040 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2042 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2043 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2045 rc = llog_cat_initialize(obd, &obd->obd_olg, 1, NULL);
2047 CERROR("failed to setup llogging subsystems\n");
2051 rc = lquota_setup(filter_quota_interface_ref, obd);
2055 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2056 if (q->max_sectors < q->max_hw_sectors &&
2057 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2058 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2059 "for larger I/O requests: max_sectors = %u "
2060 "could be up to max_hw_sectors=%u\n",
2061 obd->obd_name, mnt->mnt_sb->s_id,
2062 q->max_sectors, q->max_hw_sectors);
2064 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2065 if (uuid_ptr != NULL) {
2066 class_uuid_unparse(uuid_ptr, &uuid);
2072 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2074 if (obd->obd_recovering) {
2075 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2076 "recovery for at least %d:%.02d, or until %d "
2077 "client%s reconnect. During this time new clients"
2078 " will not be allowed to connect. "
2079 "Recovery progress can be monitored by watching "
2080 "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
2081 obd->obd_name, lustre_cfg_string(lcfg, 1),
2082 label ?: "", label ? "/" : "", str,
2083 obd->obd_recovery_timeout / 60,
2084 obd->obd_recovery_timeout % 60,
2085 obd->obd_max_recoverable_clients,
2086 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2089 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2090 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2091 label ?: "", label ? "/" : "", str,
2092 obd->obd_replayable ? "enabled" : "disabled");
2100 fsfilt_put_ops(obd->obd_fsops);
2101 filter_iobuf_pool_done(filter);
2103 server_put_mount(obd->obd_name, mnt);
2104 obd->u.obt.obt_sb = 0;
2108 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2110 struct lprocfs_static_vars lvars;
2115 CLASSERT(offsetof(struct obd_device, u.obt) ==
2116 offsetof(struct obd_device, u.filter.fo_obt));
2118 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2121 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2122 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2125 addr = (unsigned long)cfs_page_address(page);
2126 clear_page((void *)addr);
2128 /* lprocfs must be setup before the filter so state can be safely added
2129 * to /proc incrementally as the filter is setup */
2130 lprocfs_filter_init_vars(&lvars);
2131 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2132 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2133 /* Init obdfilter private stats here */
2134 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2135 LPROCFS_CNTR_AVGMINMAX,
2136 "read_bytes", "bytes");
2137 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2138 LPROCFS_CNTR_AVGMINMAX,
2139 "write_bytes", "bytes");
2141 lproc_filter_attach_seqstat(obd);
2142 obd->obd_proc_exports_entry = lprocfs_register("exports",
2143 obd->obd_proc_entry,
2145 if (IS_ERR(obd->obd_proc_exports_entry)) {
2146 rc = PTR_ERR(obd->obd_proc_exports_entry);
2147 CERROR("error %d setting up lprocfs for %s\n",
2149 obd->obd_proc_exports_entry = NULL;
2152 if (obd->obd_proc_exports_entry)
2153 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2154 lprocfs_nid_stats_clear_read,
2155 lprocfs_nid_stats_clear_write, obd);
2157 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2158 LUSTRE_CFG_BUFLEN(lcfg, 4));
2159 rc = filter_common_setup(obd, lcfg, (void *)addr);
2160 OBD_PAGE_FREE(page);
2163 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2164 lprocfs_free_per_client_stats(obd);
2165 lprocfs_free_obd_stats(obd);
2166 lprocfs_obd_cleanup(obd);
2172 static struct llog_operations filter_mds_ost_repl_logops;
2174 static struct llog_operations filter_size_orig_logops = {
2175 .lop_setup = llog_obd_origin_setup,
2176 .lop_cleanup = llog_obd_origin_cleanup,
2177 .lop_add = llog_obd_origin_add
2180 static int filter_olg_fini(struct obd_llog_group *olg)
2182 struct llog_ctxt *ctxt;
2183 int rc = 0, rc2 = 0;
2186 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2188 rc = llog_cleanup(ctxt);
2190 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2192 rc2 = llog_cleanup(ctxt);
2200 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2201 struct obd_device *tgt)
2206 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2207 &filter_mds_ost_repl_logops);
2211 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2212 &filter_size_orig_logops);
2218 filter_olg_fini(olg);
2223 * Init the default olg, which is embeded in the obd_device, for filter.
2226 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2227 struct obd_device *tgt)
2229 struct filter_obd *filter = &obd->u.filter;
2230 struct llog_ctxt *ctxt;
2234 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2235 if (!filter->fo_lcm)
2238 filter_mds_ost_repl_logops = llog_client_ops;
2239 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2240 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2241 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2243 rc = filter_olg_init(obd, olg, tgt);
2245 GOTO(cleanup_lcm, rc);
2247 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2249 CERROR("Can't get ctxt for %p:%x\n", olg,
2250 LLOG_MDS_OST_REPL_CTXT);
2251 GOTO(cleanup_olg, rc = -ENODEV);
2253 ctxt->loc_lcm = filter->fo_lcm;
2254 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2255 llog_ctxt_put(ctxt);
2259 filter_olg_fini(olg);
2261 llog_recov_thread_fini(filter->fo_lcm, 1);
2262 filter->fo_lcm = NULL;
2267 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2268 struct obd_device *tgt, int count, struct llog_catid *catid,
2269 struct obd_uuid *uuid)
2271 struct filter_obd *filter = &obd->u.filter;
2272 struct llog_ctxt *ctxt;
2276 LASSERT(olg != NULL);
2277 if (olg == &obd->obd_olg)
2278 return filter_default_olg_init(obd, olg, tgt);
2280 LASSERT(filter->fo_lcm != NULL);
2281 rc = filter_olg_init(obd, olg, tgt);
2284 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2286 CERROR("Can't get ctxt for %p:%x\n", olg,
2287 LLOG_MDS_OST_REPL_CTXT);
2288 filter_olg_fini(olg);
2291 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2292 ctxt->loc_lcm = filter->fo_lcm;
2293 llog_ctxt_put(ctxt);
2297 static int filter_llog_finish(struct obd_device *obd, int count)
2299 struct filter_obd *filter = &obd->u.filter;
2300 struct llog_ctxt *ctxt;
2303 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2304 LASSERT(ctxt != NULL);
2305 mutex_down(&ctxt->loc_sem);
2306 if (ctxt->loc_imp) {
2308 * Balance class_import_get() in llog_receptor_accept(). This
2309 * is safe to do here, as llog is already synchronized and its
2312 class_import_put(ctxt->loc_imp);
2313 ctxt->loc_imp = NULL;
2315 mutex_up(&ctxt->loc_sem);
2316 llog_ctxt_put(ctxt);
2318 if (filter->fo_lcm) {
2319 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2320 filter->fo_lcm = NULL;
2322 RETURN(filter_olg_fini(&obd->obd_olg));
2326 * Find the group llog according to group index in the llog group list.
2328 static struct obd_llog_group *
2329 filter_find_olg_internal(struct filter_obd *filter, int group)
2331 struct obd_llog_group *olg;
2333 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2334 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2335 if (olg->olg_group == group)
2342 * Find the group llog according to group index on the filter
2344 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2346 struct obd_llog_group *olg = NULL;
2347 struct filter_obd *filter;
2349 filter = &obd->u.filter;
2351 if (group == OBD_LLOG_GROUP)
2352 RETURN(&obd->obd_olg);
2354 spin_lock(&filter->fo_llog_list_lock);
2355 olg = filter_find_olg_internal(filter, group);
2356 spin_unlock(&filter->fo_llog_list_lock);
2361 * Find the llog_group of the filter according to the group. If it can not
2362 * find, create the llog_group, which only happens when mds is being synced
2365 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2367 struct obd_llog_group *olg = NULL;
2368 struct filter_obd *filter;
2371 filter = &obd->u.filter;
2373 if (group == OBD_LLOG_GROUP)
2374 RETURN(&obd->obd_olg);
2376 spin_lock(&filter->fo_llog_list_lock);
2377 olg = filter_find_olg_internal(filter, group);
2379 if (olg->olg_initializing) {
2380 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2382 GOTO(out_unlock, olg);
2387 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2389 llog_group_init(olg, group);
2390 list_add(&olg->olg_list, &filter->fo_llog_list);
2391 olg->olg_initializing = 1;
2392 spin_unlock(&filter->fo_llog_list_lock);
2394 rc = llog_cat_initialize(obd, olg, 1, NULL);
2396 spin_lock(&filter->fo_llog_list_lock);
2397 list_del(&olg->olg_list);
2398 spin_unlock(&filter->fo_llog_list_lock);
2400 GOTO(out, olg = ERR_PTR(-ENOMEM));
2402 spin_lock(&filter->fo_llog_list_lock);
2403 olg->olg_initializing = 0;
2404 spin_unlock(&filter->fo_llog_list_lock);
2405 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2406 obd->obd_name, group, olg);
2411 spin_unlock(&filter->fo_llog_list_lock);
2415 static int filter_llog_connect(struct obd_export *exp,
2416 struct llogd_conn_body *body)
2418 struct obd_device *obd = exp->exp_obd;
2419 struct llog_ctxt *ctxt;
2420 struct obd_llog_group *olg;
2424 CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2425 (unsigned) body->lgdc_logid.lgl_ogr,
2426 (unsigned) body->lgdc_logid.lgl_oid,
2427 (unsigned) body->lgdc_logid.lgl_ogen);
2429 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2431 CERROR(" %s: can not find olg of group %d\n",
2432 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2435 llog_group_set_export(olg, exp);
2437 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2438 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2439 body->lgdc_ctxt_idx);
2440 rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2441 &body->lgdc_gen, NULL);
2442 llog_ctxt_put(ctxt);
2444 CERROR("failed to connect rc %d idx %d\n", rc,
2445 body->lgdc_ctxt_idx);
2450 static int filter_llog_preclean(struct obd_device *obd)
2452 struct obd_llog_group *olg, *tmp;
2453 struct filter_obd *filter;
2454 struct list_head remove_list;
2458 rc = obd_llog_finish(obd, 0);
2460 CERROR("failed to cleanup llogging subsystem\n");
2462 filter = &obd->u.filter;
2463 CFS_INIT_LIST_HEAD(&remove_list);
2465 spin_lock(&filter->fo_llog_list_lock);
2466 while (!list_empty(&filter->fo_llog_list)) {
2467 olg = list_entry(filter->fo_llog_list.next,
2468 struct obd_llog_group, olg_list);
2469 list_del(&olg->olg_list);
2470 list_add(&olg->olg_list, &remove_list);
2472 spin_unlock(&filter->fo_llog_list_lock);
2474 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2475 list_del_init(&olg->olg_list);
2476 rc = filter_olg_fini(olg);
2478 CERROR("failed to cleanup llogging subsystem for %u\n",
2486 static int filter_precleanup(struct obd_device *obd,
2487 enum obd_cleanup_stage stage)
2493 case OBD_CLEANUP_EARLY:
2495 case OBD_CLEANUP_EXPORTS:
2496 target_cleanup_recovery(obd);
2497 rc = filter_llog_preclean(obd);
2503 static int filter_cleanup(struct obd_device *obd)
2505 struct filter_obd *filter = &obd->u.filter;
2509 LCONSOLE_WARN("%s: shutting down for failover; client state "
2510 "will be preserved.\n", obd->obd_name);
2512 if (!list_empty(&obd->obd_exports)) {
2513 CERROR("%s: still has clients!\n", obd->obd_name);
2514 class_disconnect_exports(obd);
2515 if (!list_empty(&obd->obd_exports)) {
2516 CERROR("still has exports after forced cleanup?\n");
2521 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2522 lprocfs_free_per_client_stats(obd);
2523 lprocfs_free_obd_stats(obd);
2524 lprocfs_obd_cleanup(obd);
2525 lquota_cleanup(filter_quota_interface_ref, obd);
2527 /* Stop recovery before namespace cleanup. */
2528 target_stop_recovery_thread(obd);
2529 target_cleanup_recovery(obd);
2531 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2532 obd->obd_namespace = NULL;
2534 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2536 if (obd->u.obt.obt_sb == NULL)
2541 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2542 shrink_dcache_sb(obd->u.obt.obt_sb);
2544 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2545 obd->u.obt.obt_sb = NULL;
2547 fsfilt_put_ops(obd->obd_fsops);
2549 filter_iobuf_pool_done(filter);
2551 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2556 static int filter_connect_internal(struct obd_export *exp,
2557 struct obd_connect_data *data)
2559 struct filter_export_data *fed = &exp->exp_filter_data;
2564 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2565 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2566 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2567 data->ocd_connect_flags, data->ocd_version,
2568 data->ocd_grant, data->ocd_index);
2570 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2571 CWARN("!!! This export (nid %s) used object group %d "
2572 "earlier; now it's trying to use group %d! This could "
2573 "be a bug in the MDS. Please report to "
2574 "http://bugzilla.lustre.org/\n",
2575 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2578 fed->fed_group = data->ocd_group;
2580 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2581 exp->exp_connect_flags = data->ocd_connect_flags;
2582 data->ocd_version = LUSTRE_VERSION_CODE;
2584 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2585 obd_size left, want;
2587 spin_lock(&exp->exp_obd->obd_osfs_lock);
2588 left = filter_grant_space_left(exp);
2589 want = data->ocd_grant;
2590 filter_grant(exp, fed->fed_grant, want, left);
2591 data->ocd_grant = fed->fed_grant;
2592 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2594 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2595 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2596 exp->exp_client_uuid.uuid, exp,
2597 data->ocd_grant, want, left);
2600 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2601 struct filter_obd *filter = &exp->exp_obd->u.filter;
2602 struct lr_server_data *lsd = filter->fo_fsd;
2603 int index = le32_to_cpu(lsd->lsd_ost_index);
2605 if (!(lsd->lsd_feature_compat &
2606 cpu_to_le32(OBD_COMPAT_OST))) {
2607 /* this will only happen on the first connect */
2608 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2609 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2610 filter_update_server_data(exp->exp_obd,
2611 filter->fo_rcvd_filp, lsd, 1);
2612 } else if (index != data->ocd_index) {
2613 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2614 " %u doesn't match actual OST index"
2615 " %u in last_rcvd file, bad "
2617 obd_export_nid2str(exp), index,
2623 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2624 data->ocd_brw_size = 65536;
2625 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2626 data->ocd_brw_size = min(data->ocd_brw_size,
2627 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2628 LASSERT(data->ocd_brw_size);
2631 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2632 __u32 cksum_types = data->ocd_cksum_types;
2634 /* The client set in ocd_cksum_types the checksum types it
2635 * supports. We have to mask off the algorithms that we don't
2637 if (cksum_types & OBD_CKSUM_ALL)
2638 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2640 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2642 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2643 "%x\n", exp->exp_obd->obd_name,
2644 obd_export_nid2str(exp), cksum_types,
2645 data->ocd_cksum_types);
2647 /* This client does not support OBD_CONNECT_CKSUM
2648 * fall back to CRC32 */
2649 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2650 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2651 exp->exp_obd->obd_name,
2652 obd_export_nid2str(exp));
2655 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2656 * FIXME: We don't strictly need the COMPAT flag for that,
2657 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2658 * FIXME: We needed it for the index, as index 0 is valid. */
2663 static int filter_reconnect(const struct lu_env *env,
2664 struct obd_export *exp, struct obd_device *obd,
2665 struct obd_uuid *cluuid,
2666 struct obd_connect_data *data)
2671 if (exp == NULL || obd == NULL || cluuid == NULL)
2674 rc = filter_connect_internal(exp, data);
2679 /* nearly identical to mds_connect */
2680 static int filter_connect(const struct lu_env *env,
2681 struct lustre_handle *conn, struct obd_device *obd,
2682 struct obd_uuid *cluuid,
2683 struct obd_connect_data *data, void *localdata)
2685 struct lvfs_run_ctxt saved;
2686 struct obd_export *exp;
2687 struct filter_export_data *fed;
2688 struct lsd_client_data *lcd = NULL;
2693 if (conn == NULL || obd == NULL || cluuid == NULL)
2696 rc = class_connect(conn, obd, cluuid);
2699 exp = class_conn2export(conn);
2700 LASSERT(exp != NULL);
2702 fed = &exp->exp_filter_data;
2704 rc = filter_connect_internal(exp, data);
2708 filter_export_stats_init(obd, exp, localdata);
2709 if (obd->obd_replayable) {
2710 OBD_ALLOC(lcd, sizeof(*lcd));
2712 CERROR("filter: out of memory for client data\n");
2713 GOTO(cleanup, rc = -ENOMEM);
2716 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2718 rc = filter_client_add(obd, exp, -1);
2723 group = data->ocd_group;
2727 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2728 obd->obd_name, exp->exp_handle.h_cookie, group);
2730 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2731 rc = filter_read_groups(obd, group, 1);
2732 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2734 CERROR("can't read group %u\n", group);
2744 fed->fed_lcd = NULL;
2746 class_disconnect(exp);
2748 class_export_put(exp);
2754 /* Do extra sanity checks for grant accounting. We do this at connect,
2755 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2756 * always get rid of it or turn it off when we know accounting is good. */
2757 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2759 struct filter_export_data *fed;
2760 struct obd_export *exp;
2761 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2762 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2763 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2765 if (list_empty(&obd->obd_exports))
2768 /* We don't want to do this for large machines that do lots of
2769 mounts or unmounts. It burns... */
2770 if (obd->obd_num_exports > 100)
2773 spin_lock(&obd->obd_osfs_lock);
2774 spin_lock(&obd->obd_dev_lock);
2775 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2777 fed = &exp->exp_filter_data;
2778 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2781 if (maxsize > 0) { /* we may not have done a statfs yet */
2782 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2783 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2784 exp->exp_client_uuid.uuid, exp,
2785 fed->fed_grant, fed->fed_pending, maxsize);
2786 LASSERTF(fed->fed_dirty <= maxsize,
2787 "%s: cli %s/%p %ld > "LPU64"\n", func,
2788 exp->exp_client_uuid.uuid, exp,
2789 fed->fed_dirty, maxsize);
2792 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2793 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2794 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2796 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2797 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2798 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2799 tot_granted += fed->fed_grant + fed->fed_pending;
2800 tot_pending += fed->fed_pending;
2801 tot_dirty += fed->fed_dirty;
2803 fo_tot_granted = obd->u.filter.fo_tot_granted;
2804 fo_tot_pending = obd->u.filter.fo_tot_pending;
2805 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2806 spin_unlock(&obd->obd_dev_lock);
2807 spin_unlock(&obd->obd_osfs_lock);
2809 /* Do these assertions outside the spinlocks so we don't kill system */
2810 if (tot_granted != fo_tot_granted)
2811 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2812 func, tot_granted, fo_tot_granted);
2813 if (tot_pending != fo_tot_pending)
2814 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2815 func, tot_pending, fo_tot_pending);
2816 if (tot_dirty != fo_tot_dirty)
2817 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2818 func, tot_dirty, fo_tot_dirty);
2819 if (tot_pending > tot_granted)
2820 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2821 func, tot_pending, tot_granted);
2822 if (tot_granted > maxsize)
2823 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2824 func, tot_granted, maxsize);
2825 if (tot_dirty > maxsize)
2826 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2827 func, tot_dirty, maxsize);
2830 /* Remove this client from the grant accounting totals. We also remove
2831 * the export from the obd device under the osfs and dev locks to ensure
2832 * that the filter_grant_sanity_check() calculations are always valid.
2833 * The client should do something similar when it invalidates its import. */
2834 static void filter_grant_discard(struct obd_export *exp)
2836 struct obd_device *obd = exp->exp_obd;
2837 struct filter_obd *filter = &obd->u.filter;
2838 struct filter_export_data *fed = &exp->exp_filter_data;
2840 spin_lock(&obd->obd_osfs_lock);
2841 spin_lock(&obd->obd_dev_lock);
2842 list_del_init(&exp->exp_obd_chain);
2843 spin_unlock(&obd->obd_dev_lock);
2845 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2846 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2847 obd->obd_name, filter->fo_tot_granted,
2848 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2849 filter->fo_tot_granted -= fed->fed_grant;
2850 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2851 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2852 obd->obd_name, filter->fo_tot_pending,
2853 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2854 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2855 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2856 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2857 obd->obd_name, filter->fo_tot_dirty,
2858 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2859 filter->fo_tot_dirty -= fed->fed_dirty;
2863 spin_unlock(&obd->obd_osfs_lock);
2866 static int filter_destroy_export(struct obd_export *exp)
2870 if (exp->exp_filter_data.fed_pending)
2871 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2872 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2873 exp, exp->exp_filter_data.fed_pending);
2875 /* Not ported yet the b1_6 quota functionality
2876 * lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2879 target_destroy_export(exp);
2881 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2884 lprocfs_exp_cleanup(exp);
2886 if (exp->exp_obd->obd_replayable)
2887 filter_client_free(exp);
2889 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2891 filter_grant_discard(exp);
2892 filter_fmd_cleanup(exp);
2894 if (!(exp->exp_flags & OBD_OPT_FORCE))
2895 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2900 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2902 struct obd_llog_group *olg_min, *olg;
2903 struct filter_obd *filter;
2904 int worked = 0, group;
2905 struct llog_ctxt *ctxt;
2908 filter = &obd->u.filter;
2910 /* we can't sync log holding spinlock. also, we do not want to get
2911 * into livelock. so we do following: loop over MDS's exports in
2912 * group order and skip already synced llogs -bzzz */
2914 /* look for group with min. number, but > worked */
2917 spin_lock(&filter->fo_llog_list_lock);
2918 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2919 if (olg->olg_group <= worked) {
2920 /* this group is already synced */
2923 if (group < olg->olg_group) {
2924 /* we have group with smaller number to sync */
2927 /* store current minimal group */
2929 group = olg->olg_group;
2931 spin_unlock(&filter->fo_llog_list_lock);
2933 if (olg_min == NULL)
2936 worked = olg_min->olg_group;
2937 if (olg_min->olg_exp &&
2938 (dexp == olg_min->olg_exp || dexp == NULL)) {
2940 ctxt = llog_group_get_ctxt(olg_min,
2941 LLOG_MDS_OST_REPL_CTXT);
2942 LASSERT(ctxt != NULL);
2943 err = llog_sync(ctxt, olg_min->olg_exp);
2944 llog_ctxt_put(ctxt);
2946 CERROR("error flushing logs to MDS: rc %d\n",
2949 } while (olg_min != NULL);
2952 /* also incredibly similar to mds_disconnect */
2953 static int filter_disconnect(struct obd_export *exp)
2955 struct obd_device *obd = exp->exp_obd;
2960 class_export_get(exp);
2962 if (!(exp->exp_flags & OBD_OPT_FORCE))
2963 filter_grant_sanity_check(obd, __FUNCTION__);
2964 filter_grant_discard(exp);
2966 /* Disconnect early so that clients can't keep using export */
2967 rc = class_disconnect(exp);
2968 if (exp->exp_obd->obd_namespace != NULL)
2969 ldlm_cancel_locks_for_export(exp);
2971 fsfilt_sync(obd, obd->u.obt.obt_sb);
2973 /* flush any remaining cancel messages out to the target */
2974 filter_sync_llogs(obd, exp);
2975 class_export_put(exp);
2979 /* reverse import is changed, sync all cancels */
2980 static void filter_revimp_update(struct obd_export *exp)
2985 class_export_get(exp);
2987 /* flush any remaining cancel messages out to the target */
2988 filter_sync_llogs(exp->exp_obd, exp);
2989 class_export_put(exp);
2993 static int filter_ping(struct obd_export *exp)
2995 filter_fmd_expire(exp);
3000 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3001 const char *what, int quiet)
3003 struct dentry *dchild = NULL;
3006 if (oa->o_valid & OBD_MD_FLGROUP)
3009 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3011 if (IS_ERR(dchild)) {
3012 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3013 what, group, oa->o_id);
3017 if (dchild->d_inode == NULL) {
3019 CERROR("%s: %s on non-existent object: "LPU64"\n",
3020 obd->obd_name, what, oa->o_id);
3022 RETURN(ERR_PTR(-ENOENT));
3028 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3030 struct dentry *dentry = NULL;
3031 struct obd_device *obd;
3035 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3036 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3040 obd = class_exp2obd(exp);
3042 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3046 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3048 RETURN(PTR_ERR(dentry));
3050 /* Limit the valid bits in the return data to what we actually use */
3051 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3052 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3058 /* this should be enabled/disabled in condition to enabled/disabled large
3059 * inodes (fast EAs) in backing store FS. */
3060 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3061 void *handle, struct obdo *oa)
3063 struct obd_device *obd = exp->exp_obd;
3067 if (oa->o_valid & OBD_MD_FLFID) {
3068 struct filter_fid ff;
3070 if (!(oa->o_valid & OBD_MD_FLGROUP))
3072 /* packing fid and converting it to LE for storing into EA.
3073 * Here ->o_stripe_idx should be filled by LOV and rest of
3074 * fields - by client. */
3075 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3076 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3077 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3078 ff.ff_objid = cpu_to_le64(oa->o_id);
3079 ff.ff_group = cpu_to_le64(oa->o_gr);
3081 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3082 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3083 oa->o_generation, oa->o_id, oa->o_gr);
3085 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3087 CERROR("store fid in object failed! rc: %d\n", rc);
3089 CDEBUG(D_HA, "OSS object without fid info!\n");
3095 /* this is called from filter_truncate() until we have filter_punch() */
3096 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3097 struct obdo *oa, struct obd_trans_info *oti)
3099 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3100 struct llog_cookie *fcc = NULL;
3101 struct filter_obd *filter;
3102 int rc, err, locked = 0, sync = 0;
3103 unsigned int ia_valid;
3104 struct inode *inode;
3109 LASSERT(dentry != NULL);
3110 LASSERT(!IS_ERR(dentry));
3112 inode = dentry->d_inode;
3113 LASSERT(inode != NULL);
3115 filter = &exp->exp_obd->u.filter;
3116 iattr_from_obdo(&iattr, oa, oa->o_valid);
3117 ia_valid = iattr.ia_valid;
3119 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3120 OBD_ALLOC(fcc, sizeof(*fcc));
3122 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3125 if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
3127 LOCK_INODE_MUTEX(inode);
3131 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3132 * then we will accept the UID+GID sent by the client during write for
3133 * initializing the ownership of this inode. We only allow this to
3134 * happen once so clear these bits in setattr. In 2.6 kernels it is
3135 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3136 * the flags that are actually being set. */
3137 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3138 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3139 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3141 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3142 if (!(ia_valid & ATTR_MODE)) {
3143 iattr.ia_mode = inode->i_mode;
3144 iattr.ia_valid |= ATTR_MODE;
3146 iattr.ia_mode &= ~S_ISUID;
3148 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3149 if (!(iattr.ia_valid & ATTR_MODE)) {
3150 iattr.ia_mode = inode->i_mode;
3151 iattr.ia_valid |= ATTR_MODE;
3153 iattr.ia_mode &= ~S_ISGID;
3156 orig_ids[USRQUOTA] = inode->i_uid;
3157 orig_ids[GRPQUOTA] = inode->i_gid;
3158 handle = fsfilt_start_log(exp->exp_obd, inode,
3159 FSFILT_OP_SETATTR, oti, 1);
3161 GOTO(out_unlock, rc = PTR_ERR(handle));
3163 /* update inode EA only once when inode is suid bit marked. As
3164 * on 2.6.x UID and GID may be set separately, we check here
3165 * only one of them to avoid double setting. */
3166 if (inode->i_mode & S_ISUID)
3167 filter_update_fidea(exp, inode, handle, oa);
3169 handle = fsfilt_start(exp->exp_obd, inode,
3170 FSFILT_OP_SETATTR, oti);
3172 GOTO(out_unlock, rc = PTR_ERR(handle));
3174 if (oa->o_valid & OBD_MD_FLFLAGS) {
3175 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3176 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3178 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3180 /* set cancel cookie callback function */
3181 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3182 filter_cancel_cookies_cb,
3186 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3187 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3189 /* The truncate might have used up our transaction credits. Make
3190 * sure we have one left for the last_rcvd update. */
3191 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3193 rc = filter_finish_transno(exp, oti, rc, sync);
3195 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3199 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3201 CERROR("error on commit, err = %d\n", err);
3209 /* Let's flush truncated page on disk immediately, then we can
3210 * avoid need to search for page aliases before directio writes
3211 * and this sort of stuff at expense of somewhat slower
3212 * truncates not on a page boundary. I believe this is the only
3213 * place in filter code that can lead to pages getting to
3214 * pagecache so far. */
3215 filter_clear_truncated_page(inode);
3216 UNLOCK_INODE_MUTEX(inode);
3223 UNLOCK_INODE_MUTEX(inode);
3226 OBD_FREE(fcc, sizeof(*fcc));
3228 /* trigger quota release */
3229 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3230 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3231 int rc2 = lquota_adjust(filter_quota_interface_ref,
3232 exp->exp_obd, cur_ids,
3233 orig_ids, rc, FSFILT_OP_SETATTR);
3234 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3235 "filter adjust qunit. (rc:%d)\n", rc2);
3240 /* this is called from filter_truncate() until we have filter_punch() */
3241 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3242 struct obd_trans_info *oti)
3244 struct ldlm_res_id res_id;
3245 struct filter_mod_data *fmd;
3246 struct lvfs_run_ctxt saved;
3247 struct filter_obd *filter;
3248 struct ldlm_resource *res;
3249 struct dentry *dentry;
3253 osc_build_res_name(oinfo->oi_oa->o_id, oinfo->oi_oa->o_gr, &res_id);
3254 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3255 oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
3259 /* This would be very bad - accidentally truncating a file when
3260 * changing the time or similar - bug 12203. */
3261 if (oinfo->oi_oa->o_valid & OBD_MD_FLSIZE &&
3262 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3263 static char mdsinum[48];
3265 if (oinfo->oi_oa->o_valid & OBD_MD_FLFID)
3266 snprintf(mdsinum, sizeof(mdsinum) - 1,
3267 " of inode "LPU64"/%u", oinfo->oi_oa->o_fid,
3268 oinfo->oi_oa->o_generation);
3272 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3274 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3275 oinfo->oi_oa->o_id, mdsinum);
3279 dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
3282 RETURN(PTR_ERR(dentry));
3284 filter = &exp->exp_obd->u.filter;
3285 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3288 if (oinfo->oi_oa->o_valid &
3289 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3290 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
3291 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3292 fmd->fmd_mactime_xid = oti->oti_xid;
3293 filter_fmd_put(exp, fmd);
3296 /* setting objects attributes (including owner/group) */
3297 rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
3299 GOTO(out_unlock, rc);
3301 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3302 &res_id, LDLM_EXTENT, 0);
3305 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3306 ldlm_resource_putref(res);
3309 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3311 /* Quota release need uid/gid info */
3312 obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
3313 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3319 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3323 /* XXX identical to osc_unpackmd */
3324 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3325 struct lov_mds_md *lmm, int lmm_bytes)
3331 if (lmm_bytes < sizeof (*lmm)) {
3332 CERROR("lov_mds_md too small: %d, need %d\n",
3333 lmm_bytes, (int)sizeof(*lmm));
3336 /* XXX LOV_MAGIC etc check? */
3338 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3339 CERROR("lov_mds_md: zero lmm_object_id\n");
3344 lsm_size = lov_stripe_md_size(1);
3348 if (*lsmp != NULL && lmm == NULL) {
3349 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3350 OBD_FREE(*lsmp, lsm_size);
3355 if (*lsmp == NULL) {
3356 OBD_ALLOC(*lsmp, lsm_size);
3360 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3361 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3362 OBD_FREE(*lsmp, lsm_size);
3365 loi_init((*lsmp)->lsm_oinfo[0]);
3369 /* XXX zero *lsmp? */
3370 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3371 LASSERT((*lsmp)->lsm_object_id);
3374 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3379 /* caller must hold fo_create_locks[oa->o_gr] */
3380 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3381 struct filter_obd *filter)
3383 struct obdo doa; /* XXX obdo on stack */
3389 LASSERT(oa->o_gr != 0);
3390 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3391 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3393 memset(&doa, 0, sizeof(doa));
3395 doa.o_valid |= OBD_MD_FLGROUP;
3396 doa.o_gr = oa->o_gr;
3397 doa.o_mode = S_IFREG;
3399 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3400 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3401 exp->exp_obd->obd_name, doa.o_gr);
3405 last = filter_last_id(filter, doa.o_gr);
3407 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3408 exp->exp_obd->obd_name, oa->o_id + 1, last);
3410 for (id = last; id > oa->o_id; id--) {
3412 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
3413 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3414 CEMERG("error destroying precreate objid "LPU64": %d\n",
3416 filter_set_last_id(filter, id - 1, doa.o_gr);
3417 /* update last_id on disk periodically so that if we restart
3418 * we don't need to re-scan all of the just-deleted objects. */
3419 if ((id & 511) == 0)
3420 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3423 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3424 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3426 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3427 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3432 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3433 obd_gr group, int *num);
3434 /* returns a negative error or a nonnegative number of files to create */
3435 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3436 obd_gr group, struct obd_trans_info *oti)
3438 struct obd_device *obd = exp->exp_obd;
3439 struct filter_obd *filter = &obd->u.filter;
3443 /* delete orphans request */
3444 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3445 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3446 CERROR("%s: dropping old orphan cleanup request\n",
3450 /* This causes inflight precreates to abort and drop lock */
3451 set_bit(group, &filter->fo_destroys_in_progress);
3452 down(&filter->fo_create_locks[group]);
3453 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3454 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3455 exp->exp_obd->obd_name, group);
3456 up(&filter->fo_create_locks[group]);
3459 diff = oa->o_id - filter_last_id(filter, group);
3460 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3461 filter_last_id(filter, group), diff);
3463 if (-diff > OST_MAX_PRECREATE) {
3464 CERROR("%s: ignoring bogus orphan destroy request: "
3465 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3466 oa->o_id, filter_last_id(filter, group));
3467 /* FIXME: should reset precreate_next_id on MDS */
3468 GOTO(out, rc = -EINVAL);
3471 rc = filter_destroy_precreated(exp, oa, filter);
3473 CERROR("%s: unable to write lastobjid, but "
3474 "orphans were deleted\n", obd->obd_name);
3477 /* XXX: Used by MDS for the first time! */
3478 clear_bit(group, &filter->fo_destroys_in_progress);
3481 down(&filter->fo_create_locks[group]);
3482 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3483 CERROR("%s: dropping old precreate request\n",
3487 /* only precreate if group == 0 and o_id is specfied */
3488 if (group < FILTER_GROUP_MDS0 || oa->o_id == 0)
3491 diff = oa->o_id - filter_last_id(filter, group);
3492 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3493 filter_last_id(filter, group), diff);
3495 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3496 oa->o_id, filter_last_id(filter, group), diff);
3500 oa->o_id = filter_last_id(&obd->u.filter, group);
3501 rc = filter_precreate(obd, oa, group, &diff);
3502 oa->o_id = filter_last_id(&obd->u.filter, group);
3504 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3507 /* else diff == 0 */
3510 up(&filter->fo_create_locks[group]);
3514 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3515 __u64 max_age, __u32 flags)
3517 struct filter_obd *filter = &obd->u.filter;
3518 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3522 /* at least try to account for cached pages. its still racey and
3523 * might be under-reporting if clients haven't announced their
3524 * caches with brw recently */
3525 spin_lock(&obd->obd_osfs_lock);
3526 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3527 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3528 spin_unlock(&obd->obd_osfs_lock);
3530 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3531 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3532 filter->fo_tot_dirty, filter->fo_tot_granted,
3533 filter->fo_tot_pending,
3534 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3536 filter_grant_sanity_check(obd, __FUNCTION__);
3538 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3539 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3540 osfs->os_bsize - 1) >> blockbits));
3542 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3543 * stop creating files on MDS if OST is not good shape to create
3545 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3550 static int filter_use_existing_obj(struct obd_device *obd,
3551 struct dentry *dchild, void **handle,
3554 struct inode *inode = dchild->d_inode;
3558 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3561 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3562 if (IS_ERR(*handle))
3563 return PTR_ERR(*handle);
3565 iattr.ia_valid = ATTR_MODE;
3566 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3567 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3575 /* We rely on the fact that only one thread will be creating files in a given
3576 * group at a time, which is why we don't need an atomic filter_get_new_id.
3577 * Even if we had that atomic function, the following race would exist:
3579 * thread 1: gets id x from filter_next_id
3580 * thread 2: gets id (x + 1) from filter_next_id
3581 * thread 2: creates object (x + 1)
3582 * thread 1: tries to create object x, gets -ENOSPC
3584 * Caller must hold fo_create_locks[group]
3586 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3587 obd_gr group, int *num)
3589 struct dentry *dchild = NULL, *dparent = NULL;
3590 struct filter_obd *filter;
3591 struct obd_statfs *osfs;
3592 int err = 0, rc = 0, recreate_obj = 0, i;
3593 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3595 void *handle = NULL;
3598 filter = &obd->u.filter;
3600 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3602 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3604 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3605 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3608 OBD_ALLOC(osfs, sizeof(*osfs));
3611 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3612 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3613 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3614 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3615 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3619 OBD_FREE(osfs, sizeof(*osfs));
3624 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3625 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3627 for (i = 0; i < *num && err == 0; i++) {
3628 int cleanup_phase = 0;
3630 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3631 CWARN("%s: create aborted by destroy\n",
3640 last_id = filter_last_id(filter, group);
3641 if (next_id > last_id) {
3642 CERROR("Error: Trying to recreate obj greater"
3643 "than last id "LPD64" > "LPD64"\n",
3645 GOTO(cleanup, rc = -EINVAL);
3648 next_id = filter_last_id(filter, group) + 1;
3650 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
3652 dparent = filter_parent_lock(obd, group, next_id);
3653 if (IS_ERR(dparent))
3654 GOTO(cleanup, rc = PTR_ERR(dparent));
3655 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3657 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3659 GOTO(cleanup, rc = PTR_ERR(dchild));
3660 cleanup_phase = 2; /* f_dput(dchild) */
3662 if (dchild->d_inode != NULL) {
3663 /* This would only happen if lastobjid was bad on disk*/
3664 /* Could also happen if recreating missing obj but it
3665 * already exists. */
3667 CERROR("%s: recreating existing object %.*s?\n",
3668 obd->obd_name, dchild->d_name.len,
3669 dchild->d_name.name);
3671 /* Use these existing objects if they are
3673 if (dchild->d_inode->i_size == 0) {
3674 rc = filter_use_existing_obj(obd,dchild,
3675 &handle, &cleanup_phase);
3682 CERROR("%s: Serious error: objid %.*s already "
3683 "exists; is this filesystem corrupt?\n",
3684 obd->obd_name, dchild->d_name.len,
3685 dchild->d_name.name);
3688 GOTO(cleanup, rc = -EEXIST);
3691 handle = fsfilt_start_log(obd, dparent->d_inode,
3692 FSFILT_OP_CREATE, NULL, 1);
3694 GOTO(cleanup, rc = PTR_ERR(handle));
3697 /* We mark object SUID+SGID to flag it for accepting UID+GID
3698 * from client on first write. Currently the permission bits
3699 * on the OST are never used, so this is OK. */
3700 rc = ll_vfs_create(dparent->d_inode, dchild,
3701 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3703 CERROR("create failed rc = %d\n", rc);
3708 if (!recreate_obj) {
3709 filter_set_last_id(filter, next_id, group);
3710 err = filter_update_last_objid(obd, group, 0);
3712 CERROR("unable to write lastobjid "
3713 "but file created\n");
3717 switch(cleanup_phase) {
3719 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3721 CERROR("error on commit, err = %d\n", err);
3728 filter_parent_unlock(dparent);
3735 if (time_after(jiffies, enough_time)) {
3737 "%s: precreate slow - want %d got %d \n",
3738 obd->obd_name, *num, i);
3745 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3746 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3751 static int filter_create(struct obd_export *exp, struct obdo *oa,
3752 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3754 struct filter_export_data *fed;
3755 struct obd_device *obd = NULL;
3756 struct filter_obd *filter;
3757 struct lvfs_run_ctxt saved;
3758 struct lov_stripe_md *lsm = NULL;
3759 int rc = 0, diff, group = oa->o_gr;
3762 if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
3763 CERROR("!!! nid %s sent invalid object group %d\n",
3764 obd_export_nid2str(exp), group);
3769 fed = &exp->exp_filter_data;
3770 filter = &obd->u.filter;
3772 if (fed->fed_group != group) {
3773 CERROR("!!! this export (nid %s) used object group %d "
3774 "earlier; now it's trying to use group %d! This could "
3775 "be a bug in the MDS. Please report to "
3776 "http://bugzilla.lustre.org/\n",
3777 obd_export_nid2str(exp), fed->fed_group, group);
3781 CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
3782 oa->o_gr, oa->o_id);
3786 rc = obd_alloc_memmd(exp, &lsm);
3793 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3795 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3796 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3797 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3798 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3799 oa->o_id, filter_last_id(filter,
3804 down(&filter->fo_create_locks[oa->o_gr]);
3805 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3806 up(&filter->fo_create_locks[oa->o_gr]);
3809 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3812 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3813 if (rc && ea != NULL && *ea != lsm) {
3814 obd_free_memmd(exp, &lsm);
3815 } else if (rc == 0 && ea != NULL) {
3816 /* XXX LOV STACKING: the lsm that is passed to us from
3817 * LOV does not have valid lsm_oinfo data structs, so
3818 * don't go touching that. This needs to be fixed in a
3820 lsm->lsm_object_id = oa->o_id;
3827 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3828 struct lov_stripe_md *md, struct obd_trans_info *oti,
3829 struct obd_export *md_exp)
3831 unsigned int qcids[MAXQUOTAS] = {0, 0};
3832 struct obd_device *obd;
3833 struct filter_obd *filter;
3834 struct dentry *dchild = NULL, *dparent = NULL;
3835 struct lvfs_run_ctxt saved;
3836 void *handle = NULL;
3837 struct llog_cookie *fcc = NULL;
3838 int rc, rc2, cleanup_phase = 0, sync = 0;
3842 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3845 filter = &obd->u.filter;
3847 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3850 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3852 GOTO(cleanup, rc = PTR_ERR(dchild));
3855 if (dchild->d_inode == NULL) {
3856 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3858 /* If object already gone, cancel cookie right now */
3859 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3860 struct llog_ctxt *ctxt;
3861 struct obd_llog_group *olg;
3862 fcc = obdo_logcookie(oa);
3863 olg = filter_find_olg(obd, oa->o_gr);
3865 CERROR(" %s: can not find olg of group %d\n",
3866 obd->obd_name, (int)oa->o_gr);
3867 GOTO(cleanup, rc = PTR_ERR(olg));
3869 llog_group_set_export(olg, exp);
3871 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
3872 llog_cancel(ctxt, NULL, 1, fcc, 0);
3873 llog_ctxt_put(ctxt);
3874 fcc = NULL; /* we didn't allocate fcc, don't free it */
3876 GOTO(cleanup, rc = -ENOENT);
3879 filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3881 /* Our MDC connection is established by the MDS to us */
3882 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3883 OBD_ALLOC(fcc, sizeof(*fcc));
3885 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3887 DQUOT_INIT(dchild->d_inode);
3889 /* we're gonna truncate it first in order to avoid possible deadlock:
3891 * open trasaction open transaction
3892 * down(i_zombie) down(i_zombie)
3893 * restart transaction
3894 * (see BUG 4180) -bzzz
3896 LOCK_INODE_MUTEX(dchild->d_inode);
3897 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3899 if (IS_ERR(handle)) {
3900 UNLOCK_INODE_MUTEX(dchild->d_inode);
3901 GOTO(cleanup, rc = PTR_ERR(handle));
3904 iattr.ia_valid = ATTR_SIZE;
3906 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3907 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3908 UNLOCK_INODE_MUTEX(dchild->d_inode);
3912 GOTO(cleanup, rc = rc2);
3914 /* We don't actually need to lock the parent until we are unlinking
3915 * here, and not while truncating above. That avoids holding the
3916 * parent lock for a long time during truncate, which can block other
3917 * threads from doing anything to objects in that directory. bug 7171 */
3918 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3919 if (IS_ERR(dparent))
3920 GOTO(cleanup, rc = PTR_ERR(dparent));
3921 cleanup_phase = 3; /* filter_parent_unlock */
3923 LOCK_INODE_MUTEX(dchild->d_inode);
3924 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3925 if (IS_ERR(handle)) {
3926 UNLOCK_INODE_MUTEX(dchild->d_inode);
3927 GOTO(cleanup, rc = PTR_ERR(handle));
3929 cleanup_phase = 4; /* fsfilt_commit */
3931 /* Quota release need uid/gid of inode */
3932 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3934 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3936 /* this drops dchild->d_inode->i_mutex unconditionally */
3937 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3941 switch(cleanup_phase) {
3944 sync = fsfilt_add_journal_cb(obd, 0, oti ?
3945 oti->oti_handle : handle,
3946 filter_cancel_cookies_cb,
3948 /* If add_journal_cb failed, then filter_finish_transno
3949 * will commit the handle and we will do a sync
3950 * on commit. then we call callback directly to free
3953 rc = filter_finish_transno(exp, oti, rc, sync);
3955 filter_cancel_cookies_cb(obd, 0, fcc, rc);
3958 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3960 CERROR("error on commit, err = %d\n", rc2);
3967 filter_parent_unlock(dparent);
3971 OBD_FREE(fcc, sizeof(*fcc));
3973 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3976 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3980 /* trigger quota release */
3981 qcids[USRQUOTA] = oa->o_uid;
3982 qcids[GRPQUOTA] = oa->o_gid;
3983 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
3987 CDEBUG(D_QUOTA, "filter adjust qunit! (rc:%d)\n", rc2);
3991 /* NB start and end are used for punch, but not truncate */
3992 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3993 struct obd_trans_info *oti,
3994 struct ptlrpc_request_set *rqset)
3999 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4000 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4001 oinfo->oi_policy.l_extent.end);
4005 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4006 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4007 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4009 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
4010 oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
4014 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4015 rc = filter_setattr(exp, oinfo, oti);
4019 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4020 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4023 struct lvfs_run_ctxt saved;
4024 struct filter_obd *filter;
4025 struct dentry *dentry;
4029 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
4030 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4034 filter = &exp->exp_obd->u.filter;
4036 /* an objid of zero is taken to mean "sync whole filesystem" */
4037 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4038 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4039 /* flush any remaining cancel messages out to the target */
4040 filter_sync_llogs(exp->exp_obd, exp);
4044 dentry = filter_oa2dentry(exp->exp_obd, oa);
4046 RETURN(PTR_ERR(dentry));
4048 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4050 LOCK_INODE_MUTEX(dentry->d_inode);
4052 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4054 /* just any file to grab fsync method - "file" arg unused */
4055 struct file *file = filter->fo_rcvd_filp;
4057 if (file->f_op && file->f_op->fsync)
4058 rc = file->f_op->fsync(NULL, dentry, 1);
4060 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4064 UNLOCK_INODE_MUTEX(dentry->d_inode);
4066 oa->o_valid = OBD_MD_FLID;
4067 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4069 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4075 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4076 void *key, __u32 *vallen, void *val,
4077 struct lov_stripe_md *lsm)
4079 struct obd_device *obd;
4082 obd = class_exp2obd(exp);
4084 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4088 if (KEY_IS(KEY_BLOCKSIZE)) {
4089 __u32 *blocksize = val;
4091 if (*vallen < sizeof(*blocksize))
4093 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4095 *vallen = sizeof(*blocksize);
4099 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4100 __u32 *blocksize_bits = val;
4101 if (blocksize_bits) {
4102 if (*vallen < sizeof(*blocksize_bits))
4104 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4106 *vallen = sizeof(*blocksize_bits);
4110 if (KEY_IS(KEY_LAST_ID)) {
4111 obd_id *last_id = val;
4112 /* FIXME: object groups */
4114 if (*vallen < sizeof(*last_id))
4116 *last_id = filter_last_id(&obd->u.filter,
4117 exp->exp_filter_data.fed_group);
4119 *vallen = sizeof(*last_id);
4123 CDEBUG(D_IOCTL, "invalid key\n");
4127 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4128 void *key, __u32 vallen, void *val,
4129 struct ptlrpc_request_set *set)
4131 struct obd_device *obd;
4132 struct obd_llog_group *olg;
4133 struct llog_ctxt *ctxt;
4139 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4143 if (KEY_IS(KEY_CAPA_KEY)) {
4144 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4146 CERROR("filter update capability key failed: %d\n", rc);
4150 if (KEY_IS(KEY_REVIMP_UPD)) {
4151 filter_revimp_update(exp);
4155 if (!KEY_IS(KEY_MDS_CONN))
4158 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4159 obd_export_nid2str(exp));
4160 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4162 /* setup llog imports */
4163 LASSERT(val != NULL);
4164 group = (int)(*(__u32 *)val);
4165 LASSERT(group >= FILTER_GROUP_MDS0);
4167 olg = filter_find_create_olg(obd, group);
4169 RETURN(PTR_ERR(olg));
4171 llog_group_set_export(olg, exp);
4173 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4174 LASSERTF(ctxt != NULL, "ctxt is null\n"),
4176 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4177 llog_ctxt_put(ctxt);
4179 lquota_setinfo(filter_quota_interface_ref, exp, obd);
4184 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4185 int len, void *karg, void *uarg)
4187 struct obd_device *obd = exp->exp_obd;
4188 struct obd_ioctl_data *data = karg;
4192 case OBD_IOC_ABORT_RECOVERY: {
4193 CERROR("aborting recovery for device %s\n", obd->obd_name);
4194 target_stop_recovery_thread(obd);
4198 case OBD_IOC_SYNC: {
4199 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4200 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4204 case OBD_IOC_SET_READONLY: {
4206 struct super_block *sb = obd->u.obt.obt_sb;
4207 struct inode *inode = sb->s_root->d_inode;
4208 BDEVNAME_DECLARE_STORAGE(tmp);
4209 CERROR("*** setting device %s read-only ***\n",
4210 ll_bdevname(sb, tmp));
4212 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4213 if (!IS_ERR(handle))
4214 rc = fsfilt_commit(obd, inode, handle, 1);
4216 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4217 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4219 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4223 case OBD_IOC_CATLOGLIST: {
4224 rc = llog_catalog_list(obd, 1, data);
4228 case OBD_IOC_LLOG_CANCEL:
4229 case OBD_IOC_LLOG_REMOVE:
4230 case OBD_IOC_LLOG_INFO:
4231 case OBD_IOC_LLOG_PRINT: {
4232 /* FIXME to be finished */
4233 RETURN(-EOPNOTSUPP);
4235 struct llog_ctxt *ctxt = NULL;
4237 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4238 rc = llog_ioctl(ctxt, cmd, data);
4239 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4252 static int filter_health_check(struct obd_device *obd)
4254 #ifdef USE_HEALTH_CHECK_WRITE
4255 struct filter_obd *filter = &obd->u.filter;
4260 * health_check to return 0 on healthy
4261 * and 1 on unhealthy.
4263 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4266 #ifdef USE_HEALTH_CHECK_WRITE
4267 LASSERT(filter->fo_health_check_filp != NULL);
4268 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4273 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4276 return filter_fid2dentry(data, NULL, gr, id);
4279 static int filter_process_config(struct obd_device *obd, obd_count len,
4282 struct lustre_cfg *lcfg = buf;
4283 struct lprocfs_static_vars lvars;
4286 switch (lcfg->lcfg_command) {
4287 case LCFG_SPTLRPC_CONF: {
4288 struct filter_obd *filter = &obd->u.filter;
4289 struct sptlrpc_conf_log *log;
4290 struct sptlrpc_rule_set tmp_rset;
4292 log = sptlrpc_conf_log_extract(lcfg);
4298 sptlrpc_rule_set_init(&tmp_rset);
4300 rc = sptlrpc_rule_set_from_log(&tmp_rset, log);
4302 CERROR("obd %s: failed get sptlrpc rules: %d\n",
4307 write_lock(&filter->fo_sptlrpc_lock);
4308 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
4309 filter->fo_sptlrpc_rset = tmp_rset;
4310 write_unlock(&filter->fo_sptlrpc_lock);
4312 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
4316 lprocfs_filter_init_vars(&lvars);
4318 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4326 static struct lvfs_callback_ops filter_lvfs_ops = {
4327 l_fid2dentry: filter_lvfs_fid2dentry,
4330 static struct obd_ops filter_obd_ops = {
4331 .o_owner = THIS_MODULE,
4332 .o_get_info = filter_get_info,
4333 .o_set_info_async = filter_set_info_async,
4334 .o_setup = filter_setup,
4335 .o_precleanup = filter_precleanup,
4336 .o_cleanup = filter_cleanup,
4337 .o_connect = filter_connect,
4338 .o_reconnect = filter_reconnect,
4339 .o_disconnect = filter_disconnect,
4340 .o_ping = filter_ping,
4341 .o_init_export = filter_init_export,
4342 .o_destroy_export = filter_destroy_export,
4343 .o_statfs = filter_statfs,
4344 .o_getattr = filter_getattr,
4345 .o_unpackmd = filter_unpackmd,
4346 .o_create = filter_create,
4347 .o_setattr = filter_setattr,
4348 .o_destroy = filter_destroy,
4349 .o_brw = filter_brw,
4350 .o_punch = filter_truncate,
4351 .o_sync = filter_sync,
4352 .o_preprw = filter_preprw,
4353 .o_commitrw = filter_commitrw,
4354 .o_llog_init = filter_llog_init,
4355 .o_llog_connect = filter_llog_connect,
4356 .o_llog_finish = filter_llog_finish,
4357 .o_iocontrol = filter_iocontrol,
4358 .o_health_check = filter_health_check,
4359 .o_process_config = filter_process_config,
4362 quota_interface_t *filter_quota_interface_ref;
4363 extern quota_interface_t filter_quota_interface;
4365 static int __init obdfilter_init(void)
4367 struct lprocfs_static_vars lvars;
4370 lprocfs_filter_init_vars(&lvars);
4372 request_module("lquota");
4373 OBD_ALLOC(obdfilter_created_scratchpad,
4374 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4375 sizeof(*obdfilter_created_scratchpad));
4376 if (obdfilter_created_scratchpad == NULL)
4379 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4380 sizeof(struct filter_mod_data),
4383 GOTO(out, rc = -ENOMEM);
4385 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4386 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4388 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4389 LUSTRE_OST_NAME, NULL);
4393 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4394 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4395 ll_fmd_cachep = NULL;
4397 if (filter_quota_interface_ref)
4398 PORTAL_SYMBOL_PUT(filter_quota_interface);
4400 OBD_FREE(obdfilter_created_scratchpad,
4401 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4402 sizeof(*obdfilter_created_scratchpad));
4408 static void __exit obdfilter_exit(void)
4410 if (filter_quota_interface_ref)
4411 PORTAL_SYMBOL_PUT(filter_quota_interface);
4413 if (ll_fmd_cachep) {
4414 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4415 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4416 ll_fmd_cachep = NULL;
4419 class_unregister_type(LUSTRE_OST_NAME);
4420 OBD_FREE(obdfilter_created_scratchpad,
4421 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4422 sizeof(*obdfilter_created_scratchpad));
4425 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4426 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4427 MODULE_LICENSE("GPL");
4429 module_init(obdfilter_init);
4430 module_exit(obdfilter_exit);