1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/fs/obdfilter/filter.c
6 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
7 * Author: Peter Braam <braam@clusterfs.com>
8 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
28 * (which need to get journal_lock, may block if journal full).
30 * Invariant: Call filter_start_transno() before any journal ops to avoid the
31 * same deadlock problem. We can (and want) to get rid of the
32 * transno sem in favour of the dir/inode i_mutex to avoid single
33 * threaded operation on the OST.
36 #define DEBUG_SUBSYSTEM S_FILTER
38 #ifndef AUTOCONF_INCLUDED
39 #include <linux/config.h>
41 #include <linux/module.h>
43 #include <linux/dcache.h>
44 #include <linux/init.h>
45 #include <linux/version.h>
46 #include <linux/sched.h>
47 #include <linux/mount.h>
48 #include <linux/buffer_head.h>
50 #include <obd_cksum.h>
51 #include <obd_class.h>
53 #include <lustre_dlm.h>
54 #include <lustre_fsfilt.h>
55 #include <lprocfs_status.h>
56 #include <lustre_log.h>
57 #include <lustre_commit_confd.h>
58 #include <libcfs/list.h>
59 #include <lustre_disk.h>
60 #include <lustre_quota.h>
61 #include <linux/slab.h>
62 #include <lustre_param.h>
64 #include "filter_internal.h"
66 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
67 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS0
68 static struct lvfs_callback_ops filter_lvfs_ops;
69 cfs_mem_cache_t *ll_fmd_cachep;
71 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
72 void *cb_data, int error)
74 obd_transno_commit_cb(obd, transno, error);
77 /* Assumes caller has already pushed us into the kernel context. */
78 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
79 int rc, int force_sync)
81 struct filter_obd *filter = &exp->exp_obd->u.filter;
82 struct filter_export_data *fed = &exp->exp_filter_data;
83 struct filter_client_data *fcd = fed->fed_fcd;
86 int err, log_pri = D_RPCTRACE;
88 /* Propagate error code. */
92 if (!exp->exp_obd->obd_replayable || oti == NULL)
95 /* we don't allocate new transnos for replayed requests */
96 if (oti->oti_transno == 0) {
97 spin_lock(&filter->fo_translock);
98 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
99 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
100 spin_unlock(&filter->fo_translock);
101 oti->oti_transno = last_rcvd;
103 spin_lock(&filter->fo_translock);
104 last_rcvd = oti->oti_transno;
105 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
106 filter->fo_fsd->lsd_last_transno =
107 cpu_to_le64(last_rcvd);
108 spin_unlock(&filter->fo_translock);
110 fcd->fcd_last_rcvd = cpu_to_le64(last_rcvd);
112 /* could get xid from oti, if it's ever needed */
113 fcd->fcd_last_xid = 0;
115 off = fed->fed_lr_off;
117 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
118 fed->fed_lr_idx, fed->fed_lr_off);
122 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
128 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
129 fcd, sizeof(*fcd), &off,
130 force_sync | exp->exp_need_sync);
132 filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
140 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
141 last_rcvd, fcd->fcd_uuid, fed->fed_lr_idx, err);
146 void f_dput(struct dentry *dentry)
148 /* Can't go inside filter_ddelete because it can block */
149 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
150 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
151 LASSERT(atomic_read(&dentry->d_count) > 0);
156 static void init_brw_stats(struct brw_stats *brw_stats)
159 for (i = 0; i < BRW_LAST; i++)
160 spin_lock_init(&brw_stats->hist[i].oh_lock);
163 static int lprocfs_init_rw_stats(struct obd_device *obd,
164 struct lprocfs_stats **stats)
168 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
169 LPROC_FILTER_LAST - 1;
170 *stats = lprocfs_alloc_stats(num_stats, 0);
174 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
175 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
176 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
177 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
178 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
183 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
184 plus the procfs overhead :( */
185 static int filter_export_stats_init(struct obd_device *obd,
186 struct obd_export *exp,
189 struct filter_export_data *fed = &exp->exp_filter_data;
193 init_brw_stats(&fed->fed_brw_stats);
195 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
196 /* Self-export gets no proc entry */
199 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
204 struct nid_stat *tmp = exp->exp_nid_stats;
205 LASSERT(tmp != NULL);
207 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
208 if (tmp->nid_brw_stats == NULL)
211 init_brw_stats(tmp->nid_brw_stats);
212 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
213 0644, &filter_per_nid_stats_fops,
216 CWARN("Error adding the brw_stats file\n");
218 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
222 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
231 /* Add client data to the FILTER. We use a bitmap to locate a free space
232 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
233 * Otherwise, we have just read the data from the last_rcvd file and
234 * we know its offset. */
235 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
238 struct filter_obd *filter = &obd->u.filter;
239 struct filter_export_data *fed = &exp->exp_filter_data;
240 unsigned long *bitmap = filter->fo_last_rcvd_slots;
241 int new_client = (cl_idx == -1);
245 LASSERT(bitmap != NULL);
246 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
249 if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid) == 0)
252 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
253 * there's no need for extra complication here
256 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
258 if (cl_idx >= LR_MAX_CLIENTS) {
259 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
263 if (test_and_set_bit(cl_idx, bitmap)) {
264 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
269 if (test_and_set_bit(cl_idx, bitmap)) {
270 CERROR("FILTER client %d: bit already set in bitmap!\n",
276 fed->fed_lr_idx = cl_idx;
277 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
278 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
279 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
281 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
282 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
285 struct lvfs_run_ctxt saved;
286 loff_t off = fed->fed_lr_off;
290 CDEBUG(D_INFO, "writing client fcd at idx %u (%llu) (len %u)\n",
291 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_fcd));
293 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
294 /* Transaction needed to fix bug 1403 */
295 handle = fsfilt_start(obd,
296 filter->fo_rcvd_filp->f_dentry->d_inode,
297 FSFILT_OP_SETATTR, NULL);
298 if (IS_ERR(handle)) {
299 rc = PTR_ERR(handle);
300 CERROR("unable to start transaction: rc %d\n", rc);
302 rc = fsfilt_add_journal_cb(obd, 0, handle,
303 target_client_add_cb, exp);
305 spin_lock(&exp->exp_lock);
306 exp->exp_need_sync = 1;
307 spin_unlock(&exp->exp_lock);
309 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
311 sizeof(*fed->fed_fcd),
312 &off, rc /* sync if no cb */);
314 filter->fo_rcvd_filp->f_dentry->d_inode,
317 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
320 CERROR("error writing %s client idx %u: rc %d\n",
321 LAST_RCVD, fed->fed_lr_idx, rc);
328 static int filter_client_free(struct obd_export *exp)
330 struct filter_export_data *fed = &exp->exp_filter_data;
331 struct filter_obd *filter = &exp->exp_obd->u.filter;
332 struct obd_device *obd = exp->exp_obd;
333 struct filter_client_data zero_fcd;
334 struct lvfs_run_ctxt saved;
339 if (fed->fed_fcd == NULL)
342 /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
343 if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid ) == 0)
346 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
347 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
349 LASSERT(filter->fo_last_rcvd_slots != NULL);
351 off = fed->fed_lr_off;
353 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
354 * we leak a client slot that will be cleaned on the next recovery. */
356 CERROR("%s: client idx %d has med_off %lld\n",
357 obd->obd_name, fed->fed_lr_idx, off);
358 GOTO(free, rc = -EINVAL);
361 /* Clear the bit _after_ zeroing out the client so we don't
362 race with filter_client_add and zero out new clients.*/
363 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
364 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
369 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
370 memset(&zero_fcd, 0, sizeof zero_fcd);
371 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
372 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_fcd,
373 sizeof(zero_fcd), &off,
374 (!exp->exp_libclient ||
375 exp->exp_need_sync));
377 /* update server's transno */
378 filter_update_server_data(obd, filter->fo_rcvd_filp,
380 !exp->exp_libclient);
381 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
383 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
384 "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
385 fed->fed_fcd->fcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
389 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
390 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
397 OBD_FREE(fed->fed_fcd, sizeof(*fed->fed_fcd));
403 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
404 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
405 struct filter_mod_data *fmd)
407 LASSERT_SPIN_LOCKED(&fed->fed_lock);
408 if (--fmd->fmd_refcount == 0) {
409 /* XXX when we have persistent reservations and the handle
410 * is stored herein we need to drop it here. */
411 fed->fed_mod_count--;
412 list_del(&fmd->fmd_list);
413 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
417 /* drop fmd reference, free it if last ref */
418 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
420 struct filter_export_data *fed;
425 fed = &exp->exp_filter_data;
426 spin_lock(&fed->fed_lock);
427 filter_fmd_put_nolock(fed, fmd); /* caller reference */
428 spin_unlock(&fed->fed_lock);
431 /* expire entries from the end of the list if there are too many
432 * or they are too old */
433 static void filter_fmd_expire_nolock(struct filter_obd *filter,
434 struct filter_export_data *fed,
435 struct filter_mod_data *keep)
437 struct filter_mod_data *fmd, *tmp;
439 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
443 if (time_before(jiffies, fmd->fmd_expire) &&
444 fed->fed_mod_count < filter->fo_fmd_max_num)
447 list_del_init(&fmd->fmd_list);
448 filter_fmd_put_nolock(fed, fmd); /* list reference */
452 void filter_fmd_expire(struct obd_export *exp)
454 spin_lock(&exp->exp_filter_data.fed_lock);
455 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
456 &exp->exp_filter_data, NULL);
457 spin_unlock(&exp->exp_filter_data.fed_lock);
460 /* find specified objid, group in export fmd list.
461 * caller must hold fed_lock and take fmd reference itself */
462 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
463 struct filter_export_data *fed,
464 obd_id objid, obd_gr group)
466 struct filter_mod_data *found = NULL, *fmd;
468 LASSERT_SPIN_LOCKED(&fed->fed_lock);
470 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
471 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
473 list_del(&fmd->fmd_list);
474 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
475 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
480 filter_fmd_expire_nolock(filter, fed, found);
485 /* Find fmd based on objid and group, or return NULL if not found. */
486 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
487 obd_id objid, obd_gr group)
489 struct filter_mod_data *fmd;
491 spin_lock(&exp->exp_filter_data.fed_lock);
492 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
493 &exp->exp_filter_data, objid, group);
495 fmd->fmd_refcount++; /* caller reference */
496 spin_unlock(&exp->exp_filter_data.fed_lock);
501 /* Find fmd based on objid and group, or create a new one if none is found.
502 * It is possible for this function to return NULL under memory pressure,
503 * or if objid = 0 is passed (which will only cause old entries to expire).
504 * Currently this is not fatal because any fmd state is transient and
505 * may also be freed when it gets sufficiently old. */
506 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
507 obd_id objid, obd_gr group)
509 struct filter_export_data *fed = &exp->exp_filter_data;
510 struct filter_mod_data *found = NULL, *fmd_new = NULL;
512 OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO, sizeof(*fmd_new));
514 spin_lock(&fed->fed_lock);
515 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
518 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
519 fmd_new->fmd_id = objid;
520 fmd_new->fmd_gr = group;
521 fmd_new->fmd_refcount++; /* list reference */
523 fed->fed_mod_count++;
525 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
529 found->fmd_refcount++; /* caller reference */
530 found->fmd_expire = jiffies +
531 exp->exp_obd->u.filter.fo_fmd_max_age;
534 spin_unlock(&fed->fed_lock);
540 /* drop fmd list reference so it will disappear when last reference is put.
541 * This isn't so critical because it would in fact only affect the one client
542 * that is doing the unlink and at worst we have an stale entry referencing
543 * an object that should never be used again. */
544 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
546 struct filter_mod_data *found = NULL;
548 spin_lock(&exp->exp_filter_data.fed_lock);
549 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
551 list_del_init(&found->fmd_list);
552 filter_fmd_put_nolock(&exp->exp_filter_data, found);
554 spin_unlock(&exp->exp_filter_data.fed_lock);
557 #define filter_fmd_drop(exp, objid, group)
560 /* remove all entries from fmd list */
561 static void filter_fmd_cleanup(struct obd_export *exp)
563 struct filter_export_data *fed = &exp->exp_filter_data;
564 struct filter_mod_data *fmd = NULL, *tmp;
566 spin_lock(&fed->fed_lock);
567 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
568 list_del_init(&fmd->fmd_list);
569 filter_fmd_put_nolock(fed, fmd);
571 spin_unlock(&fed->fed_lock);
574 static int filter_init_export(struct obd_export *exp)
576 spin_lock_init(&exp->exp_filter_data.fed_lock);
577 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
579 spin_lock(&exp->exp_lock);
580 exp->exp_connecting = 1;
581 spin_unlock(&exp->exp_lock);
586 static int filter_free_server_data(struct filter_obd *filter)
588 OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
589 filter->fo_fsd = NULL;
590 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
591 filter->fo_last_rcvd_slots = NULL;
595 /* assumes caller is already in kernel ctxt */
596 int filter_update_server_data(struct obd_device *obd, struct file *filp,
597 struct lr_server_data *fsd, int force_sync)
603 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
604 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
605 le64_to_cpu(fsd->lsd_last_transno));
606 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
607 le64_to_cpu(fsd->lsd_mount_count));
609 fsd->lsd_compat14 = fsd->lsd_last_transno;
610 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
612 CERROR("error writing lr_server_data: rc = %d\n", rc);
617 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
620 struct filter_obd *filter = &obd->u.filter;
626 if (filter->fo_last_objid_files[group] == NULL) {
627 CERROR("Object group "LPU64" not fully setup; not updating "
628 "last_objid\n", group);
632 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
633 obd->obd_name, group, filter->fo_last_objids[group]);
635 tmp = cpu_to_le64(filter->fo_last_objids[group]);
636 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
637 &tmp, sizeof(tmp), &off, force_sync);
639 CERROR("error writing group "LPU64" last objid: rc = %d\n",
643 extern int ost_handle(struct ptlrpc_request *req);
644 /* assumes caller has already in kernel ctxt */
645 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
647 struct filter_obd *filter = &obd->u.filter;
648 struct lr_server_data *fsd;
649 struct filter_client_data *fcd = NULL;
650 struct inode *inode = filp->f_dentry->d_inode;
651 unsigned long last_rcvd_size = i_size_read(inode);
657 /* ensure padding in the struct is the correct size */
658 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
659 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
660 CLASSERT (offsetof(struct filter_client_data, fcd_padding) +
661 sizeof(fcd->fcd_padding) == LR_CLIENT_SIZE);
663 OBD_ALLOC(fsd, sizeof(*fsd));
666 filter->fo_fsd = fsd;
668 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
669 if (filter->fo_last_rcvd_slots == NULL) {
670 OBD_FREE(fsd, sizeof(*fsd));
674 if (last_rcvd_size == 0) {
675 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
677 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
678 fsd->lsd_last_transno = 0;
679 mount_count = fsd->lsd_mount_count = 0;
680 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
681 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
682 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
683 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
684 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
685 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
687 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
689 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
693 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
694 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
695 "using the wrong disk %s. Were the "
696 "/dev/ assignments rearranged?\n",
697 obd->obd_uuid.uuid, fsd->lsd_uuid);
698 GOTO(err_fsd, rc = -EINVAL);
700 mount_count = le64_to_cpu(fsd->lsd_mount_count);
701 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
703 /* Assume old last_rcvd format unless I_C_LR is set */
704 if (!(fsd->lsd_feature_incompat &
705 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
706 fsd->lsd_last_transno = fsd->lsd_compat14;
710 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
711 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
712 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
713 ~FILTER_INCOMPAT_SUPP);
714 GOTO(err_fsd, rc = -EINVAL);
716 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
717 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
718 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
719 ~FILTER_ROCOMPAT_SUPP);
720 /* Do something like remount filesystem read-only */
721 GOTO(err_fsd, rc = -EINVAL);
724 CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
725 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
726 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
727 obd->obd_name, mount_count + 1);
728 CDEBUG(D_INODE, "%s: server data size: %u\n",
729 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
730 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
731 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
732 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
733 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
734 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
735 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
736 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
737 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
738 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
739 le16_to_cpu(fsd->lsd_client_size));
741 if (!obd->obd_replayable) {
742 CWARN("%s: recovery support OFF\n", obd->obd_name);
746 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
747 off < last_rcvd_size; cl_idx++) {
749 struct obd_export *exp;
750 struct filter_export_data *fed;
753 OBD_ALLOC(fcd, sizeof(*fcd));
755 GOTO(err_client, rc = -ENOMEM);
758 /* Don't assume off is incremented properly by
759 * fsfilt_read_record(), in case sizeof(*fcd)
760 * isn't the same as fsd->lsd_client_size. */
761 off = le32_to_cpu(fsd->lsd_client_start) +
762 cl_idx * le16_to_cpu(fsd->lsd_client_size);
763 rc = fsfilt_read_record(obd, filp, fcd, sizeof(*fcd), &off);
765 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
766 LAST_RCVD, cl_idx, off, rc);
767 break; /* read error shouldn't cause startup to fail */
770 if (fcd->fcd_uuid[0] == '\0') {
771 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
776 last_rcvd = le64_to_cpu(fcd->fcd_last_rcvd);
778 /* These exports are cleaned up by filter_disconnect(), so they
779 * need to be set up like real exports as filter_connect() does.
781 exp = class_new_export(obd, (struct obd_uuid *)fcd->fcd_uuid);
783 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
784 " srv lr: "LPU64" fcd_group %d\n", fcd->fcd_uuid, cl_idx,
785 last_rcvd, le64_to_cpu(fsd->lsd_last_transno),
786 le32_to_cpu(fcd->fcd_group));
788 if (PTR_ERR(exp) == -EALREADY) {
789 /* export already exists, zero out this one */
790 CERROR("Zeroing out duplicate export due to "
792 fcd->fcd_uuid[0] = '\0';
794 GOTO(err_client, rc = PTR_ERR(exp));
797 fed = &exp->exp_filter_data;
799 fed->fed_group = le32_to_cpu(fcd->fcd_group);
800 filter_export_stats_init(obd, exp, NULL);
801 rc = filter_client_add(obd, exp, cl_idx);
802 /* can't fail for existing client */
803 LASSERTF(rc == 0, "rc = %d\n", rc);
806 spin_lock(&exp->exp_lock);
807 exp->exp_connecting = 0;
808 exp->exp_in_recovery = 0;
809 spin_unlock(&exp->exp_lock);
810 obd->obd_max_recoverable_clients++;
811 class_export_put(exp);
814 /* Need to check last_rcvd even for duplicated exports. */
815 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
818 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
819 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
823 OBD_FREE(fcd, sizeof(*fcd));
825 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
827 target_recovery_init(obd, ost_handle);
830 filter->fo_mount_count = mount_count + 1;
831 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
833 /* save it, so mount count and last_transno is current */
834 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
836 GOTO(err_client, rc);
841 target_recovery_fini(obd);
843 filter_free_server_data(filter);
847 static int filter_cleanup_groups(struct obd_device *obd)
849 struct filter_obd *filter = &obd->u.filter;
851 struct dentry *dentry;
855 if (filter->fo_dentry_O_groups != NULL) {
856 for (i = 0; i < filter->fo_group_count; i++) {
857 dentry = filter->fo_dentry_O_groups[i];
861 OBD_FREE(filter->fo_dentry_O_groups,
862 filter->fo_group_count *
863 sizeof(*filter->fo_dentry_O_groups));
864 filter->fo_dentry_O_groups = NULL;
866 if (filter->fo_last_objid_files != NULL) {
867 for (i = 0; i < filter->fo_group_count; i++) {
868 filp = filter->fo_last_objid_files[i];
872 OBD_FREE(filter->fo_last_objid_files,
873 filter->fo_group_count *
874 sizeof(*filter->fo_last_objid_files));
875 filter->fo_last_objid_files = NULL;
877 if (filter->fo_dentry_O_sub != NULL) {
878 for (i = 0; i < filter->fo_group_count; i++) {
879 for (j = 0; j < filter->fo_subdir_count; j++) {
880 dentry = filter->fo_dentry_O_sub[i].dentry[j];
885 OBD_FREE(filter->fo_dentry_O_sub,
886 filter->fo_group_count *
887 sizeof(*filter->fo_dentry_O_sub));
888 filter->fo_dentry_O_sub = NULL;
890 if (filter->fo_last_objids != NULL) {
891 OBD_FREE(filter->fo_last_objids,
892 filter->fo_group_count *
893 sizeof(*filter->fo_last_objids));
894 filter->fo_last_objids = NULL;
896 if (filter->fo_dentry_O != NULL) {
897 f_dput(filter->fo_dentry_O);
898 filter->fo_dentry_O = NULL;
903 static int filter_update_last_group(struct obd_device *obd, int group)
905 struct filter_obd *filter = &obd->u.filter;
906 struct file *filp = NULL;
907 int last_group = 0, rc;
911 if (group <= filter->fo_committed_group)
914 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
918 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
922 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
924 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
927 LASSERT(off == 0 || last_group >= FILTER_MIN_GROUPS);
928 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
929 obd->obd_name, last_group, group);
933 /* must be sync: bXXXX */
934 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
936 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
940 filter->fo_committed_group = group;
947 static int filter_read_group_internal(struct obd_device *obd, int group,
950 struct filter_obd *filter = &obd->u.filter;
951 __u64 *new_objids = NULL;
952 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
953 struct dentry **new_groups = NULL;
954 struct file **new_files = NULL;
955 struct dentry *dentry;
957 int old_count = filter->fo_group_count, rc, stage = 0, i;
963 snprintf(name, 24, "%d", group);
967 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
969 if (IS_ERR(dentry)) {
970 CERROR("Cannot lookup expected object group %d: %ld\n",
971 group, PTR_ERR(dentry));
972 RETURN(PTR_ERR(dentry));
975 dentry = simple_mkdir(filter->fo_dentry_O, name, 0700, 1);
976 if (IS_ERR(dentry)) {
977 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
979 RETURN(PTR_ERR(dentry));
984 snprintf(name, 24, "O/%d/LAST_ID", group);
986 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
988 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
989 GOTO(cleanup, rc = PTR_ERR(filp));
993 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
995 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
999 if (filter->fo_subdir_count) {
1000 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1001 if (tmp_subdirs == NULL)
1002 GOTO(cleanup, rc = -ENOMEM);
1005 for (i = 0; i < filter->fo_subdir_count; i++) {
1007 snprintf(dir, sizeof(dir), "d%u", i);
1009 tmp_subdirs->dentry[i] = simple_mkdir(dentry, dir, 0700, 1);
1010 if (IS_ERR(tmp_subdirs->dentry[i])) {
1011 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1012 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1017 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1018 tmp_subdirs->dentry[i]);
1022 /* 'group' is an index; we need an array of length 'group + 1' */
1023 if (group + 1 > old_count) {
1024 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1025 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1026 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1027 OBD_ALLOC(new_files, len * sizeof(*new_files));
1029 if (new_objids == NULL || new_subdirs == NULL ||
1030 new_groups == NULL || new_files == NULL)
1031 GOTO(cleanup, rc = -ENOMEM);
1034 memcpy(new_objids, filter->fo_last_objids,
1035 old_count * sizeof(*new_objids));
1036 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1037 old_count * sizeof(*new_subdirs));
1038 memcpy(new_groups, filter->fo_dentry_O_groups,
1039 old_count * sizeof(*new_groups));
1040 memcpy(new_files, filter->fo_last_objid_files,
1041 old_count * sizeof(*new_files));
1043 OBD_FREE(filter->fo_last_objids,
1044 old_count * sizeof(*new_objids));
1045 OBD_FREE(filter->fo_dentry_O_sub,
1046 old_count * sizeof(*new_subdirs));
1047 OBD_FREE(filter->fo_dentry_O_groups,
1048 old_count * sizeof(*new_groups));
1049 OBD_FREE(filter->fo_last_objid_files,
1050 old_count * sizeof(*new_files));
1052 filter->fo_last_objids = new_objids;
1053 filter->fo_dentry_O_sub = new_subdirs;
1054 filter->fo_dentry_O_groups = new_groups;
1055 filter->fo_last_objid_files = new_files;
1056 filter->fo_group_count = len;
1059 filter->fo_dentry_O_groups[group] = dentry;
1060 filter->fo_last_objid_files[group] = filp;
1061 if (filter->fo_subdir_count) {
1062 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1063 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1066 filter_update_last_group(obd, group);
1068 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1069 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1070 rc = filter_update_last_objid(obd, group, 1);
1074 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1075 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1076 obd->obd_name, group, last_objid);
1081 if (new_objids != NULL)
1082 OBD_FREE(new_objids, len * sizeof(*new_objids));
1083 if (new_subdirs != NULL)
1084 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1085 if (new_groups != NULL)
1086 OBD_FREE(new_groups, len * sizeof(*new_groups));
1087 if (new_files != NULL)
1088 OBD_FREE(new_files, len * sizeof(*new_files));
1090 if (filter->fo_subdir_count) {
1091 for (i = 0; i < filter->fo_subdir_count; i++) {
1092 if (tmp_subdirs->dentry[i] != NULL)
1093 dput(tmp_subdirs->dentry[i]);
1095 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1098 filp_close(filp, 0);
1105 static int filter_read_groups(struct obd_device *obd, int last_group,
1108 struct filter_obd *filter = &obd->u.filter;
1109 int old_count, group, rc = 0;
1111 down(&filter->fo_init_lock);
1112 old_count = filter->fo_group_count;
1113 for (group = old_count; group <= last_group; group++) {
1115 continue; /* no group zero */
1117 rc = filter_read_group_internal(obd, group, create);
1121 up(&filter->fo_init_lock);
1125 /* FIXME: object groups */
1126 static int filter_prep_groups(struct obd_device *obd)
1128 struct filter_obd *filter = &obd->u.filter;
1129 struct dentry *dentry, *O_dentry;
1131 int last_group, rc = 0, cleanup_phase = 0;
1135 O_dentry = simple_mkdir(current->fs->pwd, "O", 0700, 1);
1136 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1137 if (IS_ERR(O_dentry)) {
1138 rc = PTR_ERR(O_dentry);
1139 CERROR("cannot open/create O: rc = %d\n", rc);
1142 filter->fo_dentry_O = O_dentry;
1143 cleanup_phase = 1; /* O_dentry */
1145 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1146 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1147 * some time post 1.0 when all old-style OSTs have converted along
1148 * with the init_objid hack. */
1149 dentry = ll_lookup_one_len("R", O_dentry, 1);
1151 GOTO(cleanup, rc = PTR_ERR(dentry));
1152 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1153 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1156 CWARN("converting OST to new object layout\n");
1157 if (IS_ERR(O0_dentry)) {
1158 rc = PTR_ERR(O0_dentry);
1159 CERROR("error looking up O/0: rc %d\n", rc);
1160 GOTO(cleanup_R, rc);
1163 if (O0_dentry->d_inode) {
1164 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1165 GOTO(cleanup_O0, rc = -EEXIST);
1168 LOCK_INODE_MUTEX(O_dentry->d_inode);
1169 rc = vfs_rename(O_dentry->d_inode, dentry,
1170 O_dentry->d_inode, O0_dentry);
1171 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1174 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1175 GOTO(cleanup_O0, rc);
1177 filter->fo_fsd->lsd_feature_incompat |=
1178 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1179 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1181 GOTO(cleanup_O0, rc);
1193 cleanup_phase = 2; /* groups */
1195 /* we have to initialize all groups before first connections from
1196 * clients because they may send create/destroy for any group -bzzz */
1197 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1199 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1200 GOTO(cleanup, rc = PTR_ERR(filp));
1202 cleanup_phase = 3; /* filp */
1204 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1206 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1210 last_group = FILTER_MIN_GROUPS;
1212 LASSERT(last_group >= FILTER_MIN_GROUPS);
1215 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1216 FILTER_MIN_GROUPS, last_group);
1217 filter->fo_committed_group = last_group;
1218 rc = filter_read_groups(obd, last_group, 1);
1222 filp_close(filp, 0);
1226 switch (cleanup_phase) {
1228 filp_close(filp, 0);
1230 filter_cleanup_groups(obd);
1232 f_dput(filter->fo_dentry_O);
1233 filter->fo_dentry_O = NULL;
1241 /* setup the object store with correct subdirectories */
1242 static int filter_prep(struct obd_device *obd)
1244 struct lvfs_run_ctxt saved;
1245 struct filter_obd *filter = &obd->u.filter;
1247 struct inode *inode;
1251 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1252 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1253 if (!file || IS_ERR(file)) {
1255 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1259 filter->fo_rcvd_filp = file;
1260 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1261 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1262 file->f_dentry->d_inode->i_mode);
1263 GOTO(err_filp, rc = -ENOENT);
1266 inode = file->f_dentry->d_parent->d_inode;
1267 /* We use i_op->unlink directly in filter_vfs_unlink() */
1268 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1269 CERROR("%s: filesystem does not support create/unlink ops\n",
1271 GOTO(err_filp, rc = -EOPNOTSUPP);
1274 rc = filter_init_server_data(obd, file);
1276 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1279 /* open and create health check io file*/
1280 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1283 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1285 GOTO(err_server_data, rc);
1287 filter->fo_health_check_filp = file;
1288 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1289 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1290 file->f_dentry->d_inode->i_mode);
1291 GOTO(err_health_check, rc = -ENOENT);
1293 rc = lvfs_check_io_health(obd, file);
1295 GOTO(err_health_check, rc);
1297 rc = filter_prep_groups(obd);
1299 GOTO(err_health_check, rc);
1301 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1306 if (filp_close(filter->fo_health_check_filp, 0))
1307 CERROR("can't close %s after error\n", HEALTH_CHECK);
1308 filter->fo_health_check_filp = NULL;
1310 target_recovery_fini(obd);
1311 filter_free_server_data(filter);
1313 if (filp_close(filter->fo_rcvd_filp, 0))
1314 CERROR("can't close %s after error\n", LAST_RCVD);
1315 filter->fo_rcvd_filp = NULL;
1319 /* cleanup the filter: write last used object id to status file */
1320 static void filter_post(struct obd_device *obd)
1322 struct lvfs_run_ctxt saved;
1323 struct filter_obd *filter = &obd->u.filter;
1326 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1327 * best to start a transaction with h_sync, because we removed this
1330 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1331 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1334 CERROR("error writing server data: rc = %d\n", rc);
1336 for (i = 1; i < filter->fo_group_count; i++) {
1337 rc = filter_update_last_objid(obd, i,
1338 (i == filter->fo_group_count - 1));
1340 CERROR("error writing group %d lastobjid: rc = %d\n",
1344 rc = filp_close(filter->fo_rcvd_filp, 0);
1345 filter->fo_rcvd_filp = NULL;
1347 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1349 rc = filp_close(filter->fo_health_check_filp, 0);
1350 filter->fo_health_check_filp = NULL;
1352 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1354 filter_cleanup_groups(obd);
1355 filter_free_server_data(filter);
1356 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1358 filter_free_capa_keys(filter);
1359 cleanup_capa_hash(filter->fo_capa_hash);
1362 static void filter_set_last_id(struct filter_obd *filter,
1363 obd_id id, obd_gr group)
1365 LASSERT(filter->fo_fsd != NULL);
1366 LASSERT(group <= filter->fo_group_count);
1368 spin_lock(&filter->fo_objidlock);
1369 filter->fo_last_objids[group] = id;
1370 spin_unlock(&filter->fo_objidlock);
1373 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1376 LASSERT(filter->fo_fsd != NULL);
1377 LASSERT(group <= filter->fo_group_count);
1379 /* FIXME: object groups */
1380 spin_lock(&filter->fo_objidlock);
1381 id = filter->fo_last_objids[group];
1382 spin_unlock(&filter->fo_objidlock);
1387 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1389 LOCK_INODE_MUTEX(dparent->d_inode);
1393 /* We never dget the object parent, so DON'T dput it either */
1394 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1396 struct filter_obd *filter = &obd->u.filter;
1397 struct filter_subdirs *subdirs;
1398 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1400 if (group > 0 || filter->fo_subdir_count == 0)
1401 return filter->fo_dentry_O_groups[group];
1403 subdirs = &filter->fo_dentry_O_sub[group];
1404 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1407 /* We never dget the object parent, so DON'T dput it either */
1408 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1411 unsigned long now = jiffies;
1412 struct dentry *dparent = filter_parent(obd, group, objid);
1415 if (IS_ERR(dparent))
1417 if (dparent == NULL)
1418 return ERR_PTR(-ENOENT);
1420 rc = filter_lock_dentry(obd, dparent);
1421 fsfilt_check_slow(obd, now, obd_timeout, "parent lock");
1422 return rc ? ERR_PTR(rc) : dparent;
1425 /* We never dget the object parent, so DON'T dput it either */
1426 static void filter_parent_unlock(struct dentry *dparent)
1428 UNLOCK_INODE_MUTEX(dparent->d_inode);
1431 /* How to get files, dentries, inodes from object id's.
1433 * If dir_dentry is passed, the caller has already locked the parent
1434 * appropriately for this operation (normally a write lock). If
1435 * dir_dentry is NULL, we do a read lock while we do the lookup to
1436 * avoid races with create/destroy and such changing the directory
1437 * internal to the filesystem code. */
1438 struct dentry *filter_fid2dentry(struct obd_device *obd,
1439 struct dentry *dir_dentry,
1440 obd_gr group, obd_id id)
1442 struct dentry *dparent = dir_dentry;
1443 struct dentry *dchild;
1448 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1449 obd->u.filter.fo_destroys_in_progress == 0) {
1450 /* don't fail lookups for orphan recovery, it causes
1451 * later LBUGs when objects still exist during precreate */
1452 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1453 RETURN(ERR_PTR(-ENOENT));
1456 CERROR("fatal: invalid object id 0\n");
1457 RETURN(ERR_PTR(-ESTALE));
1460 len = sprintf(name, LPU64, id);
1461 if (dir_dentry == NULL) {
1462 dparent = filter_parent_lock(obd, group, id);
1463 if (IS_ERR(dparent)) {
1464 CERROR("%s: error getting object "LPU64":"LPU64
1465 " parent: rc %ld\n", obd->obd_name,
1466 id, group, PTR_ERR(dparent));
1470 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1471 dparent->d_name.len, dparent->d_name.name, name);
1472 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1473 if (dir_dentry == NULL)
1474 filter_parent_unlock(dparent);
1475 if (IS_ERR(dchild)) {
1476 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1481 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1482 CERROR("%s: got bad object "LPU64" inode %lu\n",
1483 obd->obd_name, id, dchild->d_inode->i_ino);
1485 RETURN(ERR_PTR(-ENOENT));
1488 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1489 name, dchild, atomic_read(&dchild->d_count));
1491 LASSERT(atomic_read(&dchild->d_count) > 0);
1496 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1499 struct lustre_handle lockh;
1500 int flags = LDLM_AST_DISCARD_DATA, rc;
1501 struct ldlm_res_id res_id = { .name = { objid, 0, group, 0} };
1502 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1505 /* Tell the clients that the object is gone now and that they should
1506 * throw away any cached pages. */
1507 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1508 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1509 ldlm_completion_ast, NULL, NULL, 0, NULL,
1512 /* We only care about the side-effects, just drop the lock. */
1514 ldlm_lock_decref(&lockh, LCK_PW);
1519 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1520 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1521 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1522 * reverse. Caller must take i_sem before starting the transaction and we
1523 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1524 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry)
1529 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1530 * itself we already hold dir->i_mutex for child create/unlink ops */
1531 LASSERT(dentry->d_inode != NULL);
1532 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1533 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1537 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1538 GOTO(out, rc = -ENOENT);
1540 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1545 GOTO(out, rc = -EPERM);
1547 /* check_sticky() */
1548 if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1549 IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1550 GOTO(out, rc = -EPERM);
1552 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1553 * that was done because of journal_start (which is already done
1554 * here) or some other ordering issue. */
1557 rc = security_inode_unlink(dir, dentry);
1561 rc = dir->i_op->unlink(dir, dentry);
1563 /* need to drop i_mutex before we lose inode reference */
1564 UNLOCK_INODE_MUTEX(dentry->d_inode);
1571 /* Caller must hold LCK_PW on parent and push us into kernel context.
1572 * Caller must hold child i_mutex, we drop it always.
1573 * Caller is also required to ensure that dchild->d_inode exists. */
1574 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1575 obd_gr group, struct dentry *dparent,
1576 struct dentry *dchild)
1578 struct inode *inode = dchild->d_inode;
1581 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1582 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1583 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1584 (unsigned long)inode->i_nlink,
1585 atomic_read(&inode->i_count));
1588 rc = filter_vfs_unlink(dparent->d_inode, dchild);
1590 CERROR("error unlinking objid %.*s: rc %d\n",
1591 dchild->d_name.len, dchild->d_name.name, rc);
1595 struct filter_intent_args {
1596 struct ldlm_lock **victim;
1601 static enum interval_iter filter_intent_cb(struct interval_node *n,
1604 struct ldlm_interval *node = (struct ldlm_interval *)n;
1605 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1606 __u64 size = arg->size;
1607 struct ldlm_lock **v = arg->victim;
1608 struct ldlm_lock *lck;
1610 /* If the interval is lower than the current file size,
1612 if (interval_high(n) <= size)
1613 return INTERVAL_ITER_STOP;
1615 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1616 /* Don't send glimpse ASTs to liblustre clients.
1617 * They aren't listening for them, and they do
1618 * entirely synchronous I/O anyways. */
1619 if (lck->l_export == NULL ||
1620 lck->l_export->exp_libclient == 1)
1623 if (*arg->liblustre)
1624 *arg->liblustre = 0;
1627 *v = LDLM_LOCK_GET(lck);
1628 } else if ((*v)->l_policy_data.l_extent.start <
1629 lck->l_policy_data.l_extent.start) {
1631 *v = LDLM_LOCK_GET(lck);
1634 /* the same policy group - every lock has the
1635 * same extent, so needn't do it any more */
1639 return INTERVAL_ITER_CONT;
1642 static int filter_intent_policy(struct ldlm_namespace *ns,
1643 struct ldlm_lock **lockp, void *req_cookie,
1644 ldlm_mode_t mode, int flags, void *data)
1646 CFS_LIST_HEAD(rpc_list);
1647 struct ptlrpc_request *req = req_cookie;
1648 struct ldlm_lock *lock = *lockp, *l = NULL;
1649 struct ldlm_resource *res = lock->l_resource;
1650 ldlm_processing_policy policy;
1651 struct ost_lvb *res_lvb, *reply_lvb;
1652 struct ldlm_reply *rep;
1654 int idx, rc, tmpflags = 0, only_liblustre = 1;
1655 struct ldlm_interval_tree *tree;
1656 struct filter_intent_args arg;
1657 int repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1658 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1659 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1662 policy = ldlm_get_processing_policy(res);
1663 LASSERT(policy != NULL);
1664 LASSERT(req != NULL);
1666 rc = lustre_pack_reply(req, 3, repsize, NULL);
1668 RETURN(req->rq_status = rc);
1670 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1671 LASSERT(rep != NULL);
1673 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1674 sizeof(*reply_lvb));
1675 LASSERT(reply_lvb != NULL);
1677 //fixup_handle_for_resent_req(req, lock, &lockh);
1679 /* If we grant any lock at all, it will be a whole-file read lock.
1680 * Call the extent policy function to see if our request can be
1681 * granted, or is blocked.
1682 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse lock
1684 lock->l_policy_data.l_extent.start = 0;
1685 lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1686 lock->l_req_mode = LCK_PR;
1688 LASSERT(ns == res->lr_namespace);
1690 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1691 check_res_locked(res);
1693 /* FIXME: we should change the policy function slightly, to not make
1694 * this list at all, since we just turn around and free it */
1695 while (!list_empty(&rpc_list)) {
1696 struct ldlm_lock *wlock =
1697 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1698 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1699 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1700 lock->l_flags &= ~LDLM_FL_CP_REQD;
1701 list_del_init(&wlock->l_cp_ast);
1702 LDLM_LOCK_PUT(wlock);
1705 /* The lock met with no resistance; we're finished. */
1706 if (rc == LDLM_ITER_CONTINUE) {
1707 /* do not grant locks to the liblustre clients: they cannot
1708 * handle ASTs robustly. We need to do this while still
1709 * holding ns_lock to avoid the lock remaining on the res_link
1710 * list (and potentially being added to l_pending_list by an
1711 * AST) when we are going to drop this lock ASAP. */
1712 if (lock->l_export->exp_libclient ||
1713 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1714 ldlm_resource_unlink_lock(lock);
1715 err = ELDLM_LOCK_ABORTED;
1717 err = ELDLM_LOCK_REPLACED;
1723 /* Do not grant any lock, but instead send GL callbacks. The extent
1724 * policy nicely created a list of all PW locks for us. We will choose
1725 * the highest of those which are larger than the size in the LVB, if
1726 * any, and perform a glimpse callback. */
1727 res_lvb = res->lr_lvb_data;
1728 LASSERT(res_lvb != NULL);
1729 *reply_lvb = *res_lvb;
1732 * ->ns_lock guarantees that no new locks are granted, and,
1733 * therefore, that res->lr_lvb_data cannot increase beyond the
1734 * end of already granted lock. As a result, it is safe to
1735 * check against "stale" reply_lvb->lvb_size value without
1738 arg.size = reply_lvb->lvb_size;
1740 arg.liblustre = &only_liblustre;
1741 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1742 tree = &res->lr_itree[idx];
1743 if (tree->lit_mode == LCK_PR)
1746 interval_iterate_reverse(tree->lit_root,
1747 filter_intent_cb, &arg);
1751 /* There were no PW locks beyond the size in the LVB; finished. */
1753 if (only_liblustre) {
1754 /* If we discovered a liblustre client with a PW lock,
1755 * however, the LVB may be out of date! The LVB is
1756 * updated only on glimpse (which we don't do for
1757 * liblustre clients) and cancel (which the client
1758 * obviously has not yet done). So if it has written
1759 * data but kept the lock, the LVB is stale and needs
1760 * to be updated from disk.
1762 * Of course, this will all disappear when we switch to
1763 * taking liblustre locks on the OST. */
1764 ldlm_res_lvbo_update(res, NULL, 0, 1);
1766 RETURN(ELDLM_LOCK_ABORTED);
1770 * This check is for lock taken in filter_prepare_destroy() that does
1771 * not have l_glimpse_ast set. So the logic is: if there is a lock
1772 * with no l_glimpse_ast set, this object is being destroyed already.
1774 * Hence, if you are grabbing DLM locks on the server, always set
1775 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1777 if (l->l_glimpse_ast == NULL) {
1778 /* We are racing with unlink(); just return -ENOENT */
1779 rep->lock_policy_res1 = -ENOENT;
1783 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1784 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1785 /* Update the LVB from disk if the AST failed (this is a legal race) */
1787 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1788 * sending ast is not handled. This can result in lost client writes.
1791 ldlm_res_lvbo_update(res, NULL, 0, 1);
1794 *reply_lvb = *res_lvb;
1800 RETURN(ELDLM_LOCK_ABORTED);
1804 * per-obd_device iobuf pool.
1806 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1807 * allocations in write-path has to be minimized (see bug 5137).
1809 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1810 * OST threads (see ost_thread_{init,done}()).
1812 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1813 * at the OST layer there are only (potentially) multiple obd_device of type
1814 * unknown at the time of OST thread creation.
1816 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1817 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1818 * it's very own iobuf.
1822 * filter_kiobuf_pool_init()
1824 * filter_kiobuf_pool_done()
1826 * filter_iobuf_get()
1828 * operate on this array. They are "generic" in a sense that they don't depend
1829 * on actual type of iobuf's (the latter depending on Linux kernel version).
1833 * destroy pool created by filter_iobuf_pool_init
1835 static void filter_iobuf_pool_done(struct filter_obd *filter)
1837 struct filter_iobuf **pool;
1842 pool = filter->fo_iobuf_pool;
1844 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1845 if (pool[i] != NULL)
1846 filter_free_iobuf(pool[i]);
1848 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1849 filter->fo_iobuf_pool = NULL;
1855 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1857 static int filter_iobuf_pool_init(struct filter_obd *filter)
1864 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1866 if (filter->fo_iobuf_pool == NULL)
1869 filter->fo_iobuf_count = OSS_THREADS_MAX;
1874 /* Return iobuf allocated for @thread_id. We don't know in advance how
1875 * many threads there will be so we allocate a large empty array and only
1876 * fill in those slots that are actually in use.
1877 * If we haven't allocated a pool entry for this thread before, do so now. */
1878 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1880 int thread_id = oti ? oti->oti_thread_id : -1;
1881 struct filter_iobuf *pool = NULL;
1882 struct filter_iobuf **pool_place = NULL;
1884 if (thread_id >= 0) {
1885 LASSERT(thread_id < filter->fo_iobuf_count);
1886 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1889 if (unlikely(pool == NULL)) {
1890 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1891 PTLRPC_MAX_BRW_PAGES);
1892 if (pool_place != NULL)
1899 /* mount the file system (secretly). lustre_cfg parameters are:
1902 * 3 = flags: failover=f, failout=n
1905 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1908 struct filter_obd *filter = &obd->u.filter;
1909 struct vfsmount *mnt;
1910 struct lustre_mount_info *lmi;
1911 struct obd_uuid uuid;
1918 if (lcfg->lcfg_bufcount < 3 ||
1919 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1920 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1923 lmi = server_get_mount(obd->obd_name);
1925 /* We already mounted in lustre_fill_super.
1926 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1927 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1929 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1931 /* old path - used by lctl */
1932 CERROR("Using old MDS mount method\n");
1933 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1934 MS_NOATIME|MS_NODIRATIME,
1935 lustre_cfg_string(lcfg, 1), option);
1938 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1939 lustre_cfg_string(lcfg, 1), rc);
1943 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1945 if (IS_ERR(obd->obd_fsops))
1946 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1948 rc = filter_iobuf_pool_init(filter);
1952 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1954 /* failover is the default */
1955 obd->obd_replayable = 1;
1957 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1958 str = lustre_cfg_string(lcfg, 3);
1959 if (strchr(str, 'n')) {
1960 CWARN("%s: recovery disabled\n", obd->obd_name);
1961 obd->obd_replayable = 0;
1965 filter->fo_vfsmnt = mnt;
1966 obd->u.obt.obt_sb = mnt->mnt_sb;
1967 filter->fo_fstype = mnt->mnt_sb->s_type->name;
1968 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1970 fsfilt_setup(obd, obd->u.obt.obt_sb);
1972 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1973 obd->obd_lvfs_ctxt.pwdmnt = mnt;
1974 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1975 obd->obd_lvfs_ctxt.fs = get_ds();
1976 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
1978 sema_init(&filter->fo_init_lock, 1);
1979 filter->fo_committed_group = 0;
1981 rc = filter_prep(obd);
1985 filter->fo_destroys_in_progress = 0;
1986 for (i = 0; i < 32; i++)
1987 sema_init(&filter->fo_create_locks[i], 1);
1989 spin_lock_init(&filter->fo_translock);
1990 spin_lock_init(&filter->fo_objidlock);
1991 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
1992 sema_init(&filter->fo_alloc_lock, 1);
1993 init_brw_stats(&filter->fo_filter_stats);
1994 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1995 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
1996 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
1998 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
1999 spin_lock_init(&filter->fo_llog_list_lock);
2001 filter->fo_sptlrpc_lock = RW_LOCK_UNLOCKED;
2002 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2004 filter->fo_fl_oss_capa = 0;
2005 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2006 filter->fo_capa_hash = init_capa_hash();
2007 if (filter->fo_capa_hash == NULL)
2008 GOTO(err_ops, rc = -ENOMEM);
2010 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2011 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER,
2012 LDLM_NAMESPACE_GREEDY);
2013 if (obd->obd_namespace == NULL)
2014 GOTO(err_post, rc = -ENOMEM);
2015 obd->obd_namespace->ns_lvbp = obd;
2016 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2017 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2019 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2020 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2022 rc = llog_cat_initialize(obd, &obd->obd_olg, 1, NULL);
2024 CERROR("failed to setup llogging subsystems\n");
2028 rc = lquota_setup(filter_quota_interface_ref, obd);
2032 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2033 if (uuid_ptr != NULL) {
2034 class_uuid_unparse(uuid_ptr, &uuid);
2040 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2042 if (obd->obd_recovering) {
2043 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2044 "recovery until %d %s reconnect, or if no clients"
2045 " reconnect for %d:%.02d; during that time new "
2046 "clients will not be allowed to connect. "
2047 "Recovery progress can be monitored by watching "
2048 "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
2049 obd->obd_name, lustre_cfg_string(lcfg, 1),
2050 label ?: "", label ? "/" : "", str,
2051 obd->obd_max_recoverable_clients,
2052 (obd->obd_max_recoverable_clients == 1)
2053 ? "client" : "clients",
2054 (int)(OBD_RECOVERY_TIMEOUT) / 60,
2055 (int)(OBD_RECOVERY_TIMEOUT) % 60,
2058 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2059 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2060 label ?: "", label ? "/" : "", str,
2061 obd->obd_replayable ? "enabled" : "disabled");
2069 fsfilt_put_ops(obd->obd_fsops);
2070 filter_iobuf_pool_done(filter);
2072 server_put_mount(obd->obd_name, mnt);
2073 obd->u.obt.obt_sb = 0;
2077 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2079 struct lprocfs_static_vars lvars;
2084 CLASSERT(offsetof(struct obd_device, u.obt) ==
2085 offsetof(struct obd_device, u.filter.fo_obt));
2087 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2090 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2091 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2094 addr = (unsigned long)cfs_page_address(page);
2095 clear_page((void *)addr);
2097 /* lprocfs must be setup before the filter so state can be safely added
2098 * to /proc incrementally as the filter is setup */
2099 lprocfs_filter_init_vars(&lvars);
2100 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2101 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2102 /* Init obdfilter private stats here */
2103 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2104 LPROCFS_CNTR_AVGMINMAX,
2105 "read_bytes", "bytes");
2106 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2107 LPROCFS_CNTR_AVGMINMAX,
2108 "write_bytes", "bytes");
2110 lproc_filter_attach_seqstat(obd);
2111 obd->obd_proc_exports_entry = lprocfs_register("exports",
2112 obd->obd_proc_entry,
2114 if (IS_ERR(obd->obd_proc_exports_entry)) {
2115 rc = PTR_ERR(obd->obd_proc_exports_entry);
2116 CERROR("error %d setting up lprocfs for %s\n",
2118 obd->obd_proc_exports_entry = NULL;
2121 if (obd->obd_proc_exports_entry)
2122 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2123 lprocfs_nid_stats_clear_read,
2124 lprocfs_nid_stats_clear_write, obd);
2126 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2127 LUSTRE_CFG_BUFLEN(lcfg, 4));
2128 rc = filter_common_setup(obd, lcfg, (void *)addr);
2129 OBD_PAGE_FREE(page);
2132 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2133 lprocfs_free_per_client_stats(obd);
2134 lprocfs_free_obd_stats(obd);
2135 lprocfs_obd_cleanup(obd);
2141 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
2142 static struct llog_operations filter_size_orig_logops = {
2143 lop_setup: llog_obd_origin_setup,
2144 lop_cleanup: llog_obd_origin_cleanup,
2145 lop_add: llog_obd_origin_add
2148 static int filter_llog_init(struct obd_device *obd, int group,
2149 struct obd_device *tgt, int count,
2150 struct llog_catid *catid,
2151 struct obd_uuid *uuid)
2153 struct filter_obd *filter = &obd->u.filter;
2154 struct obd_llog_group *olg;
2155 struct llog_ctxt *ctxt;
2159 olg = filter_find_olg(obd, group);
2161 RETURN(PTR_ERR(olg));
2163 if (group == OBD_LLOG_GROUP) {
2164 LASSERT(filter->fo_lcm == NULL);
2165 OBD_ALLOC(filter->fo_lcm, sizeof(struct llog_commit_master));
2166 if (!filter->fo_lcm)
2169 rc = llog_init_commit_master((struct llog_commit_master *)
2174 filter_mds_ost_repl_logops = llog_client_ops;
2175 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2176 filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
2177 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2179 LASSERT(filter->fo_lcm != NULL);
2181 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2182 &filter_mds_ost_repl_logops);
2186 /* FIXME - assign unlink_cb for filter's recovery */
2188 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2190 LASSERT(ctxt != NULL);
2191 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2192 ctxt->loc_lcm = obd->u.filter.fo_lcm;
2193 rc = llog_start_commit_thread(ctxt->loc_lcm);
2194 llog_ctxt_put(ctxt);
2198 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2199 &filter_size_orig_logops);
2203 llog_cleanup_commit_master(filter->fo_lcm, 0);
2204 OBD_FREE(filter->fo_lcm, sizeof(struct llog_commit_master));
2205 filter->fo_lcm = NULL;
2210 static int filter_group_llog_finish(struct obd_llog_group *olg)
2212 struct llog_ctxt *ctxt;
2213 int rc = 0, rc2 = 0;
2216 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2218 rc = llog_cleanup(ctxt);
2220 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2222 rc2 = llog_cleanup(ctxt);
2229 static int filter_llog_finish(struct obd_device *obd, int count)
2234 if (obd->u.filter.fo_lcm) {
2235 llog_cleanup_commit_master((struct llog_commit_master *)
2236 obd->u.filter.fo_lcm, 0);
2237 OBD_FREE(obd->u.filter.fo_lcm,
2238 sizeof(struct llog_commit_master));
2239 obd->u.filter.fo_lcm = NULL;
2241 /* finish obd llog group */
2242 rc = filter_group_llog_finish(&obd->obd_olg);
2247 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2249 struct obd_llog_group *olg, *nolg;
2250 struct filter_obd *filter;
2253 filter = &obd->u.filter;
2255 if (group == OBD_LLOG_GROUP)
2256 RETURN(&obd->obd_olg);
2258 spin_lock(&filter->fo_llog_list_lock);
2259 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2260 if (olg->olg_group == group) {
2261 spin_unlock(&filter->fo_llog_list_lock);
2265 spin_unlock(&filter->fo_llog_list_lock);
2269 RETURN(ERR_PTR(-ENOMEM));
2271 llog_group_init(olg, group);
2272 spin_lock(&filter->fo_llog_list_lock);
2273 list_for_each_entry(nolg, &filter->fo_llog_list, olg_list) {
2274 LASSERT(nolg->olg_group != group);
2276 list_add(&olg->olg_list, &filter->fo_llog_list);
2277 spin_unlock(&filter->fo_llog_list_lock);
2279 rc = llog_cat_initialize(obd, olg, 1, NULL);
2281 spin_lock(&filter->fo_llog_list_lock);
2282 list_del(&olg->olg_list);
2283 spin_unlock(&filter->fo_llog_list_lock);
2285 RETURN(ERR_PTR(rc));
2287 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2288 obd->obd_name, group, olg);
2293 static int filter_llog_connect(struct obd_export *exp,
2294 struct llogd_conn_body *body)
2296 struct obd_device *obd = exp->exp_obd;
2297 struct llog_ctxt *ctxt;
2298 struct obd_llog_group *olg;
2302 CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2303 (unsigned) body->lgdc_logid.lgl_ogr,
2304 (unsigned) body->lgdc_logid.lgl_oid,
2305 (unsigned) body->lgdc_logid.lgl_ogen);
2307 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2309 RETURN(PTR_ERR(olg));
2310 llog_group_set_export(olg, exp);
2312 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2313 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2314 body->lgdc_ctxt_idx);
2315 rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2316 &body->lgdc_gen, NULL);
2317 llog_ctxt_put(ctxt);
2319 CERROR("failed to connect rc %d idx %d\n", rc,
2320 body->lgdc_ctxt_idx);
2325 static int filter_llog_preclean (struct obd_device *obd)
2327 struct obd_llog_group *olg;
2328 struct filter_obd *filter;
2332 rc = obd_llog_finish(obd, 0);
2334 CERROR("failed to cleanup llogging subsystem\n");
2336 filter = &obd->u.filter;
2337 spin_lock(&filter->fo_llog_list_lock);
2338 while (!list_empty(&filter->fo_llog_list)) {
2339 olg = list_entry(filter->fo_llog_list.next,
2340 struct obd_llog_group, olg_list);
2341 list_del(&olg->olg_list);
2342 spin_unlock(&filter->fo_llog_list_lock);
2344 rc = filter_group_llog_finish(olg);
2346 CERROR("failed to cleanup llogging subsystem for %u\n",
2349 spin_lock(&filter->fo_llog_list_lock);
2351 spin_unlock(&filter->fo_llog_list_lock);
2356 static int filter_precleanup(struct obd_device *obd,
2357 enum obd_cleanup_stage stage)
2363 case OBD_CLEANUP_EARLY:
2365 case OBD_CLEANUP_EXPORTS:
2366 target_cleanup_recovery(obd);
2367 rc = filter_llog_preclean(obd);
2369 case OBD_CLEANUP_SELF_EXP:
2371 case OBD_CLEANUP_OBD:
2377 static int filter_cleanup(struct obd_device *obd)
2379 struct filter_obd *filter = &obd->u.filter;
2383 LCONSOLE_WARN("%s: shutting down for failover; client state "
2384 "will be preserved.\n", obd->obd_name);
2386 if (!list_empty(&obd->obd_exports)) {
2387 CERROR("%s: still has clients!\n", obd->obd_name);
2388 class_disconnect_exports(obd);
2389 if (!list_empty(&obd->obd_exports)) {
2390 CERROR("still has exports after forced cleanup?\n");
2395 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2396 lprocfs_free_per_client_stats(obd);
2397 lprocfs_free_obd_stats(obd);
2398 lprocfs_obd_cleanup(obd);
2399 lquota_cleanup(filter_quota_interface_ref, obd);
2401 /* Stop recovery before namespace cleanup. */
2402 target_stop_recovery_thread(obd);
2403 target_cleanup_recovery(obd);
2405 ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2407 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2409 if (obd->u.obt.obt_sb == NULL)
2414 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2415 shrink_dcache_sb(obd->u.obt.obt_sb);
2417 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2418 obd->u.obt.obt_sb = NULL;
2420 fsfilt_put_ops(obd->obd_fsops);
2422 filter_iobuf_pool_done(filter);
2424 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2429 static int filter_connect_internal(struct obd_export *exp,
2430 struct obd_connect_data *data)
2435 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2436 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2437 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2438 data->ocd_connect_flags, data->ocd_version,
2439 data->ocd_grant, data->ocd_index);
2441 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2442 exp->exp_connect_flags = data->ocd_connect_flags;
2443 data->ocd_version = LUSTRE_VERSION_CODE;
2445 if ((exp->exp_connect_flags & OBD_CONNECT_FID) == 0) {
2446 CWARN("%s: OST requires FID support (flag="LPX64
2447 "), but client not\n",
2448 exp->exp_obd->obd_name,
2449 exp->exp_connect_flags);
2453 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2454 struct filter_export_data *fed = &exp->exp_filter_data;
2455 obd_size left, want;
2457 spin_lock(&exp->exp_obd->obd_osfs_lock);
2458 left = filter_grant_space_left(exp);
2459 want = data->ocd_grant;
2460 filter_grant(exp, fed->fed_grant, want, left);
2461 data->ocd_grant = fed->fed_grant;
2462 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2464 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2465 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2466 exp->exp_client_uuid.uuid, exp,
2467 data->ocd_grant, want, left);
2470 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2471 struct filter_obd *filter = &exp->exp_obd->u.filter;
2472 struct lr_server_data *lsd = filter->fo_fsd;
2473 int index = le32_to_cpu(lsd->lsd_ost_index);
2475 if (!(lsd->lsd_feature_compat &
2476 cpu_to_le32(OBD_COMPAT_OST))) {
2477 /* this will only happen on the first connect */
2478 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2479 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2480 filter_update_server_data(exp->exp_obd,
2481 filter->fo_rcvd_filp, lsd, 1);
2482 } else if (index != data->ocd_index) {
2483 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2484 " %u doesn't match actual OST index"
2485 " %u in last_rcvd file, bad "
2487 obd_export_nid2str(exp), index,
2493 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2494 data->ocd_brw_size = 65536;
2495 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2496 data->ocd_brw_size = min(data->ocd_brw_size,
2497 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2498 LASSERT(data->ocd_brw_size);
2501 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2502 __u32 cksum_types = data->ocd_cksum_types;
2504 /* The client set in ocd_cksum_types the checksum types it
2505 * supports. We have to mask off the algorithms that we don't
2507 if (cksum_types & OBD_CKSUM_ALL)
2508 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2510 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2512 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2513 "%x\n", exp->exp_obd->obd_name,
2514 obd_export_nid2str(exp), cksum_types,
2515 data->ocd_cksum_types);
2517 /* This client does not support OBD_CONNECT_CKSUM
2518 * fall back to CRC32 */
2519 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2520 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2521 exp->exp_obd->obd_name,
2522 obd_export_nid2str(exp));
2525 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2526 * FIXME: We don't strictly need the COMPAT flag for that,
2527 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2528 * FIXME: We needed it for the index, as index 0 is valid. */
2533 static int filter_reconnect(const struct lu_env *env,
2534 struct obd_export *exp, struct obd_device *obd,
2535 struct obd_uuid *cluuid,
2536 struct obd_connect_data *data)
2541 if (exp == NULL || obd == NULL || cluuid == NULL)
2544 rc = filter_connect_internal(exp, data);
2549 /* nearly identical to mds_connect */
2550 static int filter_connect(const struct lu_env *env,
2551 struct lustre_handle *conn, struct obd_device *obd,
2552 struct obd_uuid *cluuid,
2553 struct obd_connect_data *data, void *localdata)
2555 struct lvfs_run_ctxt saved;
2556 struct obd_export *exp;
2557 struct filter_export_data *fed;
2558 struct filter_client_data *fcd = NULL;
2563 if (conn == NULL || obd == NULL || cluuid == NULL)
2566 rc = class_connect(conn, obd, cluuid);
2569 exp = class_conn2export(conn);
2570 LASSERT(exp != NULL);
2572 fed = &exp->exp_filter_data;
2574 rc = filter_connect_internal(exp, data);
2578 filter_export_stats_init(obd, exp, localdata);
2579 group = data->ocd_group;
2580 if (obd->obd_replayable) {
2581 OBD_ALLOC(fcd, sizeof(*fcd));
2583 CERROR("filter: out of memory for client data\n");
2584 GOTO(cleanup, rc = -ENOMEM);
2587 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
2589 fed->fed_fcd->fcd_group = group;
2590 rc = filter_client_add(obd, exp, -1);
2594 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2595 obd->obd_name, exp->exp_handle.h_cookie, group);
2599 if (fed->fed_group != 0 && fed->fed_group != group) {
2600 CERROR("!!! This export (nid %s) used object group %d "
2601 "earlier; now it's trying to use group %d! This could "
2602 "be a bug in the MDS. Tell CFS.\n",
2603 obd_export_nid2str(exp), fed->fed_group, group);
2604 GOTO(cleanup, rc = -EPROTO);
2606 fed->fed_group = group;
2608 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2609 rc = filter_read_groups(obd, group, 1);
2610 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2612 CERROR("can't read group %u\n", group);
2621 OBD_FREE(fcd, sizeof(*fcd));
2622 fed->fed_fcd = NULL;
2624 class_disconnect(exp);
2626 class_export_put(exp);
2632 /* Do extra sanity checks for grant accounting. We do this at connect,
2633 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2634 * always get rid of it or turn it off when we know accounting is good. */
2635 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2637 struct filter_export_data *fed;
2638 struct obd_export *exp;
2639 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2640 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2641 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2643 if (list_empty(&obd->obd_exports))
2646 /* We don't want to do this for large machines that do lots of
2647 mounts or unmounts. It burns... */
2648 if (obd->obd_num_exports > 100)
2651 spin_lock(&obd->obd_osfs_lock);
2652 spin_lock(&obd->obd_dev_lock);
2653 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2655 fed = &exp->exp_filter_data;
2656 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2659 if (maxsize > 0) { /* we may not have done a statfs yet */
2660 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2661 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2662 exp->exp_client_uuid.uuid, exp,
2663 fed->fed_grant, fed->fed_pending, maxsize);
2664 LASSERTF(fed->fed_dirty <= maxsize,
2665 "%s: cli %s/%p %ld > "LPU64"\n", func,
2666 exp->exp_client_uuid.uuid, exp,
2667 fed->fed_dirty, maxsize);
2670 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2671 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2672 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2674 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2675 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2676 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2677 tot_granted += fed->fed_grant + fed->fed_pending;
2678 tot_pending += fed->fed_pending;
2679 tot_dirty += fed->fed_dirty;
2681 fo_tot_granted = obd->u.filter.fo_tot_granted;
2682 fo_tot_pending = obd->u.filter.fo_tot_pending;
2683 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2684 spin_unlock(&obd->obd_dev_lock);
2685 spin_unlock(&obd->obd_osfs_lock);
2687 /* Do these assertions outside the spinlocks so we don't kill system */
2688 if (tot_granted != fo_tot_granted)
2689 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2690 func, tot_granted, fo_tot_granted);
2691 if (tot_pending != fo_tot_pending)
2692 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2693 func, tot_pending, fo_tot_pending);
2694 if (tot_dirty != fo_tot_dirty)
2695 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2696 func, tot_dirty, fo_tot_dirty);
2697 if (tot_pending > tot_granted)
2698 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2699 func, tot_pending, tot_granted);
2700 if (tot_granted > maxsize)
2701 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2702 func, tot_granted, maxsize);
2703 if (tot_dirty > maxsize)
2704 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2705 func, tot_dirty, maxsize);
2708 /* Remove this client from the grant accounting totals. We also remove
2709 * the export from the obd device under the osfs and dev locks to ensure
2710 * that the filter_grant_sanity_check() calculations are always valid.
2711 * The client should do something similar when it invalidates its import. */
2712 static void filter_grant_discard(struct obd_export *exp)
2714 struct obd_device *obd = exp->exp_obd;
2715 struct filter_obd *filter = &obd->u.filter;
2716 struct filter_export_data *fed = &exp->exp_filter_data;
2718 spin_lock(&obd->obd_osfs_lock);
2719 spin_lock(&obd->obd_dev_lock);
2720 list_del_init(&exp->exp_obd_chain);
2721 spin_unlock(&obd->obd_dev_lock);
2723 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2724 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2725 obd->obd_name, filter->fo_tot_granted,
2726 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2727 filter->fo_tot_granted -= fed->fed_grant;
2728 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2729 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2730 obd->obd_name, filter->fo_tot_pending,
2731 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2732 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2733 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2734 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2735 obd->obd_name, filter->fo_tot_dirty,
2736 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2737 filter->fo_tot_dirty -= fed->fed_dirty;
2741 spin_unlock(&obd->obd_osfs_lock);
2744 static int filter_destroy_export(struct obd_export *exp)
2748 if (exp->exp_filter_data.fed_pending)
2749 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2750 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2751 exp, exp->exp_filter_data.fed_pending);
2753 /* Not ported yet the b1_6 quota functionality
2754 * lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2757 target_destroy_export(exp);
2759 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2762 lprocfs_exp_cleanup(exp);
2764 if (exp->exp_obd->obd_replayable)
2765 filter_client_free(exp);
2767 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2769 filter_grant_discard(exp);
2770 filter_fmd_cleanup(exp);
2772 if (!(exp->exp_flags & OBD_OPT_FORCE))
2773 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2778 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2780 struct obd_llog_group *olg_min, *olg;
2781 struct filter_obd *filter;
2782 int worked = 0, group;
2783 struct llog_ctxt *ctxt;
2786 filter = &obd->u.filter;
2788 /* we can't sync log holding spinlock. also, we do not want to get
2789 * into livelock. so we do following: loop over MDS's exports in
2790 * group order and skip already synced llogs -bzzz */
2792 /* look for group with min. number, but > worked */
2795 spin_lock(&filter->fo_llog_list_lock);
2796 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2797 if (olg->olg_group <= worked) {
2798 /* this group is already synced */
2801 if (group < olg->olg_group) {
2802 /* we have group with smaller number to sync */
2805 /* store current minimal group */
2807 group = olg->olg_group;
2809 spin_unlock(&filter->fo_llog_list_lock);
2811 if (olg_min == NULL)
2814 worked = olg_min->olg_group;
2815 if (olg_min->olg_exp &&
2816 (dexp == olg_min->olg_exp || dexp == NULL)) {
2818 ctxt = llog_group_get_ctxt(olg_min,
2819 LLOG_MDS_OST_REPL_CTXT);
2820 LASSERT(ctxt != NULL);
2821 err = llog_sync(ctxt, olg_min->olg_exp);
2822 llog_ctxt_put(ctxt);
2824 CERROR("error flushing logs to MDS: rc %d\n",
2827 } while (olg_min != NULL);
2830 /* also incredibly similar to mds_disconnect */
2831 static int filter_disconnect(struct obd_export *exp)
2833 struct obd_device *obd = exp->exp_obd;
2838 class_export_get(exp);
2840 if (!(exp->exp_flags & OBD_OPT_FORCE))
2841 filter_grant_sanity_check(obd, __FUNCTION__);
2842 filter_grant_discard(exp);
2844 /* Disconnect early so that clients can't keep using export */
2845 rc = class_disconnect(exp);
2846 if (exp->exp_obd->obd_namespace != NULL)
2847 ldlm_cancel_locks_for_export(exp);
2849 fsfilt_sync(obd, obd->u.obt.obt_sb);
2851 /* flush any remaining cancel messages out to the target */
2852 filter_sync_llogs(obd, exp);
2853 class_export_put(exp);
2857 /* reverse import is changed, sync all cancels */
2858 static void filter_revimp_update(struct obd_export *exp)
2863 class_export_get(exp);
2865 /* flush any remaining cancel messages out to the target */
2866 filter_sync_llogs(exp->exp_obd, exp);
2867 class_export_put(exp);
2871 static int filter_ping(struct obd_export *exp)
2873 filter_fmd_expire(exp);
2878 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2879 const char *what, int quiet)
2881 struct dentry *dchild = NULL;
2884 if (oa->o_valid & OBD_MD_FLGROUP)
2887 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
2889 if (IS_ERR(dchild)) {
2890 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
2891 what, group, oa->o_id);
2895 if (dchild->d_inode == NULL) {
2897 CERROR("%s: %s on non-existent object: "LPU64"\n",
2898 obd->obd_name, what, oa->o_id);
2900 RETURN(ERR_PTR(-ENOENT));
2906 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2908 struct dentry *dentry = NULL;
2909 struct obd_device *obd;
2913 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
2914 oinfo_capa(oinfo), CAPA_OPC_META_READ);
2918 obd = class_exp2obd(exp);
2920 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
2924 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
2926 RETURN(PTR_ERR(dentry));
2928 /* Limit the valid bits in the return data to what we actually use */
2929 oinfo->oi_oa->o_valid = OBD_MD_FLID;
2930 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
2936 /* this should be enabled/disabled in condition to enabled/disabled large
2937 * inodes (fast EAs) in backing store FS. */
2938 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
2939 void *handle, struct obdo *oa)
2941 struct obd_device *obd = exp->exp_obd;
2945 if (oa->o_valid & OBD_MD_FLFID) {
2946 struct filter_fid ff;
2948 if (!(oa->o_valid & OBD_MD_FLGROUP))
2950 /* packing fid and converting it to LE for storing into EA.
2951 * Here ->o_stripe_idx should be filled by LOV and rest of
2952 * fields - by client. */
2953 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
2954 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
2955 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
2956 ff.ff_objid = cpu_to_le64(oa->o_id);
2957 ff.ff_group = cpu_to_le64(oa->o_gr);
2959 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
2960 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
2961 oa->o_generation, oa->o_id, oa->o_gr);
2963 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
2965 CERROR("store fid in object failed! rc: %d\n", rc);
2967 CDEBUG(D_HA, "OSS object without fid info!\n");
2973 /* this is called from filter_truncate() until we have filter_punch() */
2974 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
2975 struct obdo *oa, struct obd_trans_info *oti)
2977 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
2978 struct llog_cookie *fcc = NULL;
2979 struct filter_obd *filter;
2980 int rc, err, locked = 0, sync = 0;
2981 unsigned int ia_valid;
2982 struct inode *inode;
2987 LASSERT(dentry != NULL);
2988 LASSERT(!IS_ERR(dentry));
2990 inode = dentry->d_inode;
2991 LASSERT(inode != NULL);
2993 filter = &exp->exp_obd->u.filter;
2994 iattr_from_obdo(&iattr, oa, oa->o_valid);
2995 ia_valid = iattr.ia_valid;
2997 if (oa->o_valid & OBD_MD_FLCOOKIE) {
2998 OBD_ALLOC(fcc, sizeof(*fcc));
3000 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3003 if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
3005 LOCK_INODE_MUTEX(inode);
3009 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3010 * then we will accept the UID+GID sent by the client during write for
3011 * initializing the ownership of this inode. We only allow this to
3012 * happen once so clear these bits in setattr. In 2.6 kernels it is
3013 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3014 * the flags that are actually being set. */
3015 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3016 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3017 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3019 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3020 if (!(ia_valid & ATTR_MODE)) {
3021 iattr.ia_mode = inode->i_mode;
3022 iattr.ia_valid |= ATTR_MODE;
3024 iattr.ia_mode &= ~S_ISUID;
3026 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3027 if (!(iattr.ia_valid & ATTR_MODE)) {
3028 iattr.ia_mode = inode->i_mode;
3029 iattr.ia_valid |= ATTR_MODE;
3031 iattr.ia_mode &= ~S_ISGID;
3034 orig_ids[USRQUOTA] = inode->i_uid;
3035 orig_ids[GRPQUOTA] = inode->i_gid;
3036 handle = fsfilt_start_log(exp->exp_obd, inode,
3037 FSFILT_OP_SETATTR, oti, 1);
3039 GOTO(out_unlock, rc = PTR_ERR(handle));
3041 /* update inode EA only once when inode is suid bit marked. As
3042 * on 2.6.x UID and GID may be set separately, we check here
3043 * only one of them to avoid double setting. */
3044 if (inode->i_mode & S_ISUID)
3045 filter_update_fidea(exp, inode, handle, oa);
3047 handle = fsfilt_start(exp->exp_obd, inode,
3048 FSFILT_OP_SETATTR, oti);
3050 GOTO(out_unlock, rc = PTR_ERR(handle));
3052 if (oa->o_valid & OBD_MD_FLFLAGS) {
3053 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3054 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3056 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3058 /* set cancel cookie callback function */
3059 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3060 filter_cancel_cookies_cb,
3064 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3065 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3067 /* The truncate might have used up our transaction credits. Make
3068 * sure we have one left for the last_rcvd update. */
3069 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3071 rc = filter_finish_transno(exp, oti, rc, sync);
3073 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3077 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3079 CERROR("error on commit, err = %d\n", err);
3087 /* Let's flush truncated page on disk immediately, then we can
3088 * avoid need to search for page aliases before directio writes
3089 * and this sort of stuff at expense of somewhat slower
3090 * truncates not on a page boundary. I believe this is the only
3091 * place in filter code that can lead to pages getting to
3092 * pagecache so far. */
3093 filter_clear_truncated_page(inode);
3094 UNLOCK_INODE_MUTEX(inode);
3101 UNLOCK_INODE_MUTEX(inode);
3104 OBD_FREE(fcc, sizeof(*fcc));
3106 /* trigger quota release */
3107 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3108 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3109 int rc2 = lquota_adjust(filter_quota_interface_ref,
3110 exp->exp_obd, cur_ids,
3111 orig_ids, rc, FSFILT_OP_SETATTR);
3112 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3113 "filter adjust qunit. (rc:%d)\n", rc2);
3118 /* this is called from filter_truncate() until we have filter_punch() */
3119 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3120 struct obd_trans_info *oti)
3122 struct ldlm_res_id res_id = { .name = { oinfo->oi_oa->o_id, 0,
3123 oinfo->oi_oa->o_gr, 0 } };
3124 struct filter_mod_data *fmd;
3125 struct lvfs_run_ctxt saved;
3126 struct filter_obd *filter;
3127 struct ldlm_resource *res;
3128 struct dentry *dentry;
3132 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3133 oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
3137 /* This would be very bad - accidentally truncating a file when
3138 * changing the time or similar - bug 12203. */
3139 if (oinfo->oi_oa->o_valid & OBD_MD_FLSIZE &&
3140 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3141 static char mdsinum[48];
3143 if (oinfo->oi_oa->o_valid & OBD_MD_FLFID)
3144 snprintf(mdsinum, sizeof(mdsinum) - 1,
3145 " of inode "LPU64"/%u", oinfo->oi_oa->o_fid,
3146 oinfo->oi_oa->o_generation);
3150 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3152 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3153 oinfo->oi_oa->o_id, mdsinum);
3157 dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
3160 RETURN(PTR_ERR(dentry));
3162 filter = &exp->exp_obd->u.filter;
3163 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3166 if (oinfo->oi_oa->o_valid &
3167 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3168 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
3169 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3170 fmd->fmd_mactime_xid = oti->oti_xid;
3171 filter_fmd_put(exp, fmd);
3174 /* setting objects attributes (including owner/group) */
3175 rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
3177 GOTO(out_unlock, rc);
3179 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3180 &res_id, LDLM_EXTENT, 0);
3183 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3184 ldlm_resource_putref(res);
3187 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3189 /* Quota release need uid/gid info */
3190 obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
3191 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3197 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3201 /* XXX identical to osc_unpackmd */
3202 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3203 struct lov_mds_md *lmm, int lmm_bytes)
3209 if (lmm_bytes < sizeof (*lmm)) {
3210 CERROR("lov_mds_md too small: %d, need %d\n",
3211 lmm_bytes, (int)sizeof(*lmm));
3214 /* XXX LOV_MAGIC etc check? */
3216 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3217 CERROR("lov_mds_md: zero lmm_object_id\n");
3222 lsm_size = lov_stripe_md_size(1);
3226 if (*lsmp != NULL && lmm == NULL) {
3227 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3228 OBD_FREE(*lsmp, lsm_size);
3233 if (*lsmp == NULL) {
3234 OBD_ALLOC(*lsmp, lsm_size);
3238 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3239 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3240 OBD_FREE(*lsmp, lsm_size);
3243 loi_init((*lsmp)->lsm_oinfo[0]);
3247 /* XXX zero *lsmp? */
3248 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3249 LASSERT((*lsmp)->lsm_object_id);
3252 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3257 /* caller must hold fo_create_locks[oa->o_gr] */
3258 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3259 struct filter_obd *filter)
3261 struct obdo doa; /* XXX obdo on stack */
3267 LASSERT(oa->o_gr != 0);
3268 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3269 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3271 memset(&doa, 0, sizeof(doa));
3273 doa.o_valid |= OBD_MD_FLGROUP;
3274 doa.o_gr = oa->o_gr;
3275 doa.o_mode = S_IFREG;
3277 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3278 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3279 exp->exp_obd->obd_name, doa.o_gr);
3283 last = filter_last_id(filter, doa.o_gr);
3285 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3286 exp->exp_obd->obd_name, oa->o_id + 1, last);
3288 for (id = last; id > oa->o_id; id--) {
3290 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
3291 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3292 CEMERG("error destroying precreate objid "LPU64": %d\n",
3294 filter_set_last_id(filter, id - 1, doa.o_gr);
3295 /* update last_id on disk periodically so that if we restart
3296 * we don't need to re-scan all of the just-deleted objects. */
3297 if ((id & 511) == 0)
3298 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3301 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3302 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3304 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3305 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3310 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3311 obd_gr group, int *num);
3312 /* returns a negative error or a nonnegative number of files to create */
3313 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3314 obd_gr group, struct obd_trans_info *oti)
3316 struct obd_device *obd = exp->exp_obd;
3317 struct filter_obd *filter = &obd->u.filter;
3321 /* delete orphans request */
3322 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3323 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3324 CERROR("%s: dropping old orphan cleanup request\n",
3328 /* This causes inflight precreates to abort and drop lock */
3329 set_bit(group, &filter->fo_destroys_in_progress);
3330 down(&filter->fo_create_locks[group]);
3331 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3332 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3333 exp->exp_obd->obd_name, group);
3334 up(&filter->fo_create_locks[group]);
3337 diff = oa->o_id - filter_last_id(filter, group);
3338 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3339 filter_last_id(filter, group), diff);
3341 if (-diff > OST_MAX_PRECREATE) {
3342 CERROR("%s: ignoring bogus orphan destroy request: "
3343 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3344 oa->o_id, filter_last_id(filter, group));
3345 /* FIXME: should reset precreate_next_id on MDS */
3346 GOTO(out, rc = -EINVAL);
3349 rc = filter_destroy_precreated(exp, oa, filter);
3351 CERROR("%s: unable to write lastobjid, but "
3352 "orphans were deleted\n", obd->obd_name);
3355 /* XXX: Used by MDS for the first time! */
3356 clear_bit(group, &filter->fo_destroys_in_progress);
3359 down(&filter->fo_create_locks[group]);
3360 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3361 CERROR("%s: dropping old precreate request\n",
3365 /* only precreate if group == 0 and o_id is specfied */
3366 if (group < FILTER_GROUP_MDS0 || oa->o_id == 0)
3369 diff = oa->o_id - filter_last_id(filter, group);
3370 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3371 filter_last_id(filter, group), diff);
3373 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3374 oa->o_id, filter_last_id(filter, group), diff);
3378 oa->o_id = filter_last_id(&obd->u.filter, group);
3379 rc = filter_precreate(obd, oa, group, &diff);
3380 oa->o_id = filter_last_id(&obd->u.filter, group);
3382 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3385 /* else diff == 0 */
3388 up(&filter->fo_create_locks[group]);
3392 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3393 __u64 max_age, __u32 flags)
3395 struct filter_obd *filter = &obd->u.filter;
3396 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3400 /* at least try to account for cached pages. its still racey and
3401 * might be under-reporting if clients haven't announced their
3402 * caches with brw recently */
3403 spin_lock(&obd->obd_osfs_lock);
3404 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3405 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3406 spin_unlock(&obd->obd_osfs_lock);
3408 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3409 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3410 filter->fo_tot_dirty, filter->fo_tot_granted,
3411 filter->fo_tot_pending,
3412 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3414 filter_grant_sanity_check(obd, __FUNCTION__);
3416 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3417 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3418 osfs->os_bsize - 1) >> blockbits));
3420 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3421 * stop creating files on MDS if OST is not good shape to create
3423 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3428 static int filter_use_existing_obj(struct obd_device *obd,
3429 struct dentry *dchild, void **handle,
3432 struct inode *inode = dchild->d_inode;
3436 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3439 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3440 if (IS_ERR(*handle))
3441 return PTR_ERR(*handle);
3443 iattr.ia_valid = ATTR_MODE;
3444 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3445 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3453 /* We rely on the fact that only one thread will be creating files in a given
3454 * group at a time, which is why we don't need an atomic filter_get_new_id.
3455 * Even if we had that atomic function, the following race would exist:
3457 * thread 1: gets id x from filter_next_id
3458 * thread 2: gets id (x + 1) from filter_next_id
3459 * thread 2: creates object (x + 1)
3460 * thread 1: tries to create object x, gets -ENOSPC
3462 * Caller must hold fo_create_locks[group]
3464 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3465 obd_gr group, int *num)
3467 struct dentry *dchild = NULL, *dparent = NULL;
3468 struct filter_obd *filter;
3469 struct obd_statfs *osfs;
3470 int err = 0, rc = 0, recreate_obj = 0, i;
3471 unsigned long enough_time = jiffies + min(obd_timeout * HZ / 4, 10U*HZ);
3473 void *handle = NULL;
3476 filter = &obd->u.filter;
3478 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3480 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3482 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3483 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3486 OBD_ALLOC(osfs, sizeof(*osfs));
3489 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3490 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3491 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3492 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3493 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3497 OBD_FREE(osfs, sizeof(*osfs));
3502 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3503 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3505 for (i = 0; i < *num && err == 0; i++) {
3506 int cleanup_phase = 0;
3508 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3509 CWARN("%s: create aborted by destroy\n",
3518 last_id = filter_last_id(filter, group);
3519 if (next_id > last_id) {
3520 CERROR("Error: Trying to recreate obj greater"
3521 "than last id "LPD64" > "LPD64"\n",
3523 GOTO(cleanup, rc = -EINVAL);
3526 next_id = filter_last_id(filter, group) + 1;
3528 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
3530 dparent = filter_parent_lock(obd, group, next_id);
3531 if (IS_ERR(dparent))
3532 GOTO(cleanup, rc = PTR_ERR(dparent));
3533 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3535 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3537 GOTO(cleanup, rc = PTR_ERR(dchild));
3538 cleanup_phase = 2; /* f_dput(dchild) */
3540 if (dchild->d_inode != NULL) {
3541 /* This would only happen if lastobjid was bad on disk*/
3542 /* Could also happen if recreating missing obj but it
3543 * already exists. */
3545 CERROR("%s: recreating existing object %.*s?\n",
3546 obd->obd_name, dchild->d_name.len,
3547 dchild->d_name.name);
3549 /* Use these existing objects if they are
3551 if (dchild->d_inode->i_size == 0) {
3552 rc = filter_use_existing_obj(obd,dchild,
3553 &handle, &cleanup_phase);
3560 CERROR("%s: Serious error: objid %.*s already "
3561 "exists; is this filesystem corrupt?\n",
3562 obd->obd_name, dchild->d_name.len,
3563 dchild->d_name.name);
3566 GOTO(cleanup, rc = -EEXIST);
3569 handle = fsfilt_start_log(obd, dparent->d_inode,
3570 FSFILT_OP_CREATE, NULL, 1);
3572 GOTO(cleanup, rc = PTR_ERR(handle));
3575 /* We mark object SUID+SGID to flag it for accepting UID+GID
3576 * from client on first write. Currently the permission bits
3577 * on the OST are never used, so this is OK. */
3578 rc = ll_vfs_create(dparent->d_inode, dchild,
3579 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3581 CERROR("create failed rc = %d\n", rc);
3586 if (!recreate_obj) {
3587 filter_set_last_id(filter, next_id, group);
3588 err = filter_update_last_objid(obd, group, 0);
3590 CERROR("unable to write lastobjid "
3591 "but file created\n");
3595 switch(cleanup_phase) {
3597 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3599 CERROR("error on commit, err = %d\n", err);
3606 filter_parent_unlock(dparent);
3613 if (time_after(jiffies, enough_time)) {
3615 "%s: precreate slow - want %d got %d \n",
3616 obd->obd_name, *num, i);
3623 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3624 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3629 static int filter_create(struct obd_export *exp, struct obdo *oa,
3630 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3632 struct filter_export_data *fed;
3633 struct obd_device *obd = NULL;
3634 struct filter_obd *filter;
3635 struct lvfs_run_ctxt saved;
3636 struct lov_stripe_md *lsm = NULL;
3637 int rc = 0, diff, group = oa->o_gr;
3640 if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
3641 CERROR("!!! nid %s sent invalid object group %d\n",
3642 obd_export_nid2str(exp), group);
3647 fed = &exp->exp_filter_data;
3648 filter = &obd->u.filter;
3650 if (fed->fed_group != group) {
3651 CERROR("!!! this export (nid %s) used object group %d "
3652 "earlier; now it's trying to use group %d! This could "
3653 "be a bug in the MDS. Tell CFS.\n",
3654 obd_export_nid2str(exp), fed->fed_group, group);
3658 CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
3659 oa->o_gr, oa->o_id);
3663 rc = obd_alloc_memmd(exp, &lsm);
3670 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3672 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3673 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3674 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3675 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3676 oa->o_id, filter_last_id(filter,
3681 down(&filter->fo_create_locks[oa->o_gr]);
3682 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3683 up(&filter->fo_create_locks[oa->o_gr]);
3686 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3689 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3690 if (rc && ea != NULL && *ea != lsm) {
3691 obd_free_memmd(exp, &lsm);
3692 } else if (rc == 0 && ea != NULL) {
3693 /* XXX LOV STACKING: the lsm that is passed to us from
3694 * LOV does not have valid lsm_oinfo data structs, so
3695 * don't go touching that. This needs to be fixed in a
3697 lsm->lsm_object_id = oa->o_id;
3704 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3705 struct lov_stripe_md *md, struct obd_trans_info *oti,
3706 struct obd_export *md_exp)
3708 unsigned int qcids[MAXQUOTAS] = {0, 0};
3709 struct obd_device *obd;
3710 struct filter_obd *filter;
3711 struct dentry *dchild = NULL, *dparent = NULL;
3712 struct lvfs_run_ctxt saved;
3713 void *handle = NULL;
3714 struct llog_cookie *fcc = NULL;
3715 int rc, rc2, cleanup_phase = 0, sync = 0;
3719 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3722 filter = &obd->u.filter;
3724 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3727 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3729 GOTO(cleanup, rc = PTR_ERR(dchild));
3732 if (dchild->d_inode == NULL) {
3733 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3735 /* If object already gone, cancel cookie right now */
3736 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3737 struct llog_ctxt *ctxt;
3738 struct obd_llog_group *olg;
3739 fcc = obdo_logcookie(oa);
3740 olg = filter_find_olg(obd, oa->o_gr);
3742 GOTO(cleanup, rc = PTR_ERR(olg));
3743 llog_group_set_export(olg, exp);
3745 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
3746 llog_cancel(ctxt, NULL, 1, fcc, 0);
3747 llog_ctxt_put(ctxt);
3748 fcc = NULL; /* we didn't allocate fcc, don't free it */
3750 GOTO(cleanup, rc = -ENOENT);
3753 filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3755 /* Our MDC connection is established by the MDS to us */
3756 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3757 OBD_ALLOC(fcc, sizeof(*fcc));
3759 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3761 DQUOT_INIT(dchild->d_inode);
3763 /* we're gonna truncate it first in order to avoid possible deadlock:
3765 * open trasaction open transaction
3766 * down(i_zombie) down(i_zombie)
3767 * restart transaction
3768 * (see BUG 4180) -bzzz
3770 LOCK_INODE_MUTEX(dchild->d_inode);
3771 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3773 if (IS_ERR(handle)) {
3774 UNLOCK_INODE_MUTEX(dchild->d_inode);
3775 GOTO(cleanup, rc = PTR_ERR(handle));
3778 iattr.ia_valid = ATTR_SIZE;
3780 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3781 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3782 UNLOCK_INODE_MUTEX(dchild->d_inode);
3786 GOTO(cleanup, rc = rc2);
3788 /* We don't actually need to lock the parent until we are unlinking
3789 * here, and not while truncating above. That avoids holding the
3790 * parent lock for a long time during truncate, which can block other
3791 * threads from doing anything to objects in that directory. bug 7171 */
3792 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3793 if (IS_ERR(dparent))
3794 GOTO(cleanup, rc = PTR_ERR(dparent));
3795 cleanup_phase = 3; /* filter_parent_unlock */
3797 LOCK_INODE_MUTEX(dchild->d_inode);
3798 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3799 if (IS_ERR(handle)) {
3800 UNLOCK_INODE_MUTEX(dchild->d_inode);
3801 GOTO(cleanup, rc = PTR_ERR(handle));
3803 cleanup_phase = 4; /* fsfilt_commit */
3805 /* Quota release need uid/gid of inode */
3806 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3808 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3810 /* this drops dchild->d_inode->i_mutex unconditionally */
3811 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3815 switch(cleanup_phase) {
3818 sync = fsfilt_add_journal_cb(obd, 0, oti ?
3819 oti->oti_handle : handle,
3820 filter_cancel_cookies_cb,
3822 /* If add_journal_cb failed, then filter_finish_transno
3823 * will commit the handle and we will do a sync
3824 * on commit. then we call callback directly to free
3827 rc = filter_finish_transno(exp, oti, rc, sync);
3829 filter_cancel_cookies_cb(obd, 0, fcc, rc);
3832 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3834 CERROR("error on commit, err = %d\n", rc2);
3841 filter_parent_unlock(dparent);
3845 OBD_FREE(fcc, sizeof(*fcc));
3847 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3850 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3854 /* trigger quota release */
3855 qcids[USRQUOTA] = oa->o_uid;
3856 qcids[GRPQUOTA] = oa->o_gid;
3857 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
3861 CDEBUG(D_QUOTA, "filter adjust qunit! (rc:%d)\n", rc2);
3865 /* NB start and end are used for punch, but not truncate */
3866 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3867 struct obd_trans_info *oti,
3868 struct ptlrpc_request_set *rqset)
3873 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3874 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
3875 oinfo->oi_policy.l_extent.end);
3879 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
3880 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
3881 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
3883 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3884 oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
3888 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
3889 rc = filter_setattr(exp, oinfo, oti);
3893 static int filter_sync(struct obd_export *exp, struct obdo *oa,
3894 struct lov_stripe_md *lsm, obd_off start, obd_off end,
3897 struct lvfs_run_ctxt saved;
3898 struct filter_obd *filter;
3899 struct dentry *dentry;
3903 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
3904 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
3908 filter = &exp->exp_obd->u.filter;
3910 /* an objid of zero is taken to mean "sync whole filesystem" */
3911 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
3912 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
3913 /* flush any remaining cancel messages out to the target */
3914 filter_sync_llogs(exp->exp_obd, exp);
3918 dentry = filter_oa2dentry(exp->exp_obd, oa);
3920 RETURN(PTR_ERR(dentry));
3922 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3924 LOCK_INODE_MUTEX(dentry->d_inode);
3926 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
3928 /* just any file to grab fsync method - "file" arg unused */
3929 struct file *file = filter->fo_rcvd_filp;
3931 if (file->f_op && file->f_op->fsync)
3932 rc = file->f_op->fsync(NULL, dentry, 1);
3934 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
3938 UNLOCK_INODE_MUTEX(dentry->d_inode);
3940 oa->o_valid = OBD_MD_FLID;
3941 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
3943 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3949 static int filter_get_info(struct obd_export *exp, __u32 keylen,
3950 void *key, __u32 *vallen, void *val)
3952 struct obd_device *obd;
3955 obd = class_exp2obd(exp);
3957 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3961 if (KEY_IS("blocksize")) {
3962 __u32 *blocksize = val;
3964 if (*vallen < sizeof(*blocksize))
3966 *blocksize = obd->u.obt.obt_sb->s_blocksize;
3968 *vallen = sizeof(*blocksize);
3972 if (KEY_IS("blocksize_bits")) {
3973 __u32 *blocksize_bits = val;
3974 if (blocksize_bits) {
3975 if (*vallen < sizeof(*blocksize_bits))
3977 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
3979 *vallen = sizeof(*blocksize_bits);
3983 if (KEY_IS("last_id")) {
3984 obd_id *last_id = val;
3985 /* FIXME: object groups */
3987 if (*vallen < sizeof(*last_id))
3989 *last_id = filter_last_id(&obd->u.filter,
3990 exp->exp_filter_data.fed_group);
3992 *vallen = sizeof(*last_id);
3996 CDEBUG(D_IOCTL, "invalid key\n");
4000 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4001 void *key, __u32 vallen, void *val,
4002 struct ptlrpc_request_set *set)
4004 struct obd_device *obd;
4005 struct obd_llog_group *olg;
4006 struct llog_ctxt *ctxt;
4012 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4016 if (KEY_IS(KEY_CAPA_KEY)) {
4017 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4019 CERROR("filter update capability key failed: %d\n", rc);
4023 if (KEY_IS(KEY_REVIMP_UPD)) {
4024 filter_revimp_update(exp);
4028 if (!KEY_IS(KEY_MDS_CONN))
4031 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4032 obd_export_nid2str(exp));
4033 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4035 /* setup llog imports */
4036 LASSERT(val != NULL);
4037 group = (int)(*(__u32 *)val);
4038 LASSERT(group >= FILTER_GROUP_MDS0);
4040 olg = filter_find_olg(obd, group);
4042 RETURN(PTR_ERR(olg));
4043 llog_group_set_export(olg, exp);
4045 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4046 LASSERTF(ctxt != NULL, "ctxt is null\n"),
4048 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4049 llog_ctxt_put(ctxt);
4051 lquota_setinfo(filter_quota_interface_ref, exp, obd);
4056 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4057 int len, void *karg, void *uarg)
4059 struct obd_device *obd = exp->exp_obd;
4060 struct obd_ioctl_data *data = karg;
4064 case OBD_IOC_ABORT_RECOVERY: {
4065 CERROR("aborting recovery for device %s\n", obd->obd_name);
4066 target_stop_recovery_thread(obd);
4070 case OBD_IOC_SYNC: {
4071 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4072 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4076 case OBD_IOC_SET_READONLY: {
4078 struct super_block *sb = obd->u.obt.obt_sb;
4079 struct inode *inode = sb->s_root->d_inode;
4080 BDEVNAME_DECLARE_STORAGE(tmp);
4081 CERROR("*** setting device %s read-only ***\n",
4082 ll_bdevname(sb, tmp));
4084 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4085 if (!IS_ERR(handle))
4086 rc = fsfilt_commit(obd, inode, handle, 1);
4088 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4089 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4091 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4095 case OBD_IOC_CATLOGLIST: {
4096 rc = llog_catalog_list(obd, 1, data);
4100 case OBD_IOC_LLOG_CANCEL:
4101 case OBD_IOC_LLOG_REMOVE:
4102 case OBD_IOC_LLOG_INFO:
4103 case OBD_IOC_LLOG_PRINT: {
4104 /* FIXME to be finished */
4105 RETURN(-EOPNOTSUPP);
4107 struct llog_ctxt *ctxt = NULL;
4109 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4110 rc = llog_ioctl(ctxt, cmd, data);
4111 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4124 static int filter_health_check(struct obd_device *obd)
4126 #ifdef USE_HEALTH_CHECK_WRITE
4127 struct filter_obd *filter = &obd->u.filter;
4132 * health_check to return 0 on healthy
4133 * and 1 on unhealthy.
4135 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4138 #ifdef USE_HEALTH_CHECK_WRITE
4139 LASSERT(filter->fo_health_check_filp != NULL);
4140 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4145 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4148 return filter_fid2dentry(data, NULL, gr, id);
4151 static int filter_process_config(struct obd_device *obd, obd_count len,
4154 struct lustre_cfg *lcfg = buf;
4155 struct lprocfs_static_vars lvars;
4158 switch (lcfg->lcfg_command) {
4159 case LCFG_SPTLRPC_CONF: {
4160 struct filter_obd *filter = &obd->u.filter;
4161 struct sptlrpc_conf_log *log;
4162 struct sptlrpc_rule_set tmp_rset;
4164 log = sptlrpc_conf_log_extract(lcfg);
4170 sptlrpc_rule_set_init(&tmp_rset);
4172 rc = sptlrpc_rule_set_from_log(&tmp_rset, log);
4174 CERROR("obd %s: failed get sptlrpc rules: %d\n",
4179 write_lock(&filter->fo_sptlrpc_lock);
4180 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
4181 filter->fo_sptlrpc_rset = tmp_rset;
4182 write_unlock(&filter->fo_sptlrpc_lock);
4184 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
4188 lprocfs_filter_init_vars(&lvars);
4190 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4198 static struct lvfs_callback_ops filter_lvfs_ops = {
4199 l_fid2dentry: filter_lvfs_fid2dentry,
4202 static struct obd_ops filter_obd_ops = {
4203 .o_owner = THIS_MODULE,
4204 .o_get_info = filter_get_info,
4205 .o_set_info_async = filter_set_info_async,
4206 .o_setup = filter_setup,
4207 .o_precleanup = filter_precleanup,
4208 .o_cleanup = filter_cleanup,
4209 .o_connect = filter_connect,
4210 .o_reconnect = filter_reconnect,
4211 .o_disconnect = filter_disconnect,
4212 .o_ping = filter_ping,
4213 .o_init_export = filter_init_export,
4214 .o_destroy_export = filter_destroy_export,
4215 .o_statfs = filter_statfs,
4216 .o_getattr = filter_getattr,
4217 .o_unpackmd = filter_unpackmd,
4218 .o_create = filter_create,
4219 .o_setattr = filter_setattr,
4220 .o_destroy = filter_destroy,
4221 .o_brw = filter_brw,
4222 .o_punch = filter_truncate,
4223 .o_sync = filter_sync,
4224 .o_preprw = filter_preprw,
4225 .o_commitrw = filter_commitrw,
4226 .o_llog_init = filter_llog_init,
4227 .o_llog_connect = filter_llog_connect,
4228 .o_llog_finish = filter_llog_finish,
4229 .o_iocontrol = filter_iocontrol,
4230 .o_health_check = filter_health_check,
4231 .o_process_config = filter_process_config,
4234 quota_interface_t *filter_quota_interface_ref;
4235 extern quota_interface_t filter_quota_interface;
4237 static int __init obdfilter_init(void)
4239 struct lprocfs_static_vars lvars;
4242 lprocfs_filter_init_vars(&lvars);
4244 request_module("lquota");
4245 OBD_ALLOC(obdfilter_created_scratchpad,
4246 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4247 sizeof(*obdfilter_created_scratchpad));
4248 if (obdfilter_created_scratchpad == NULL)
4251 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4252 sizeof(struct filter_mod_data),
4255 GOTO(out, rc = -ENOMEM);
4257 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4258 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4260 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4261 LUSTRE_OST_NAME, NULL);
4265 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4266 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4267 ll_fmd_cachep = NULL;
4269 if (filter_quota_interface_ref)
4270 PORTAL_SYMBOL_PUT(filter_quota_interface);
4272 OBD_FREE(obdfilter_created_scratchpad,
4273 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4274 sizeof(*obdfilter_created_scratchpad));
4280 static void __exit obdfilter_exit(void)
4282 if (filter_quota_interface_ref)
4283 PORTAL_SYMBOL_PUT(filter_quota_interface);
4285 if (ll_fmd_cachep) {
4286 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4287 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4288 ll_fmd_cachep = NULL;
4291 class_unregister_type(LUSTRE_OST_NAME);
4292 OBD_FREE(obdfilter_created_scratchpad,
4293 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4294 sizeof(*obdfilter_created_scratchpad));
4297 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4298 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4299 MODULE_LICENSE("GPL");
4301 module_init(obdfilter_init);
4302 module_exit(obdfilter_exit);