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_class.h>
52 #include <lustre_dlm.h>
53 #include <lustre_fsfilt.h>
54 #include <lprocfs_status.h>
55 #include <lustre_log.h>
56 #include <lustre_commit_confd.h>
57 #include <libcfs/list.h>
58 #include <lustre_disk.h>
59 #include <lustre_quota.h>
60 #include <linux/slab.h>
61 #include <lustre_param.h>
63 #include "filter_internal.h"
65 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
66 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS0
67 static struct lvfs_callback_ops filter_lvfs_ops;
68 cfs_mem_cache_t *ll_fmd_cachep;
70 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
71 void *cb_data, int error)
73 obd_transno_commit_cb(obd, transno, error);
76 /* Assumes caller has already pushed us into the kernel context. */
77 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
78 int rc, int force_sync)
80 struct filter_obd *filter = &exp->exp_obd->u.filter;
81 struct filter_export_data *fed = &exp->exp_filter_data;
82 struct filter_client_data *fcd = fed->fed_fcd;
85 int err, log_pri = D_RPCTRACE;
87 /* Propagate error code. */
91 if (!exp->exp_obd->obd_replayable || oti == NULL)
94 /* we don't allocate new transnos for replayed requests */
95 if (oti->oti_transno == 0) {
96 spin_lock(&filter->fo_translock);
97 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
98 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
99 spin_unlock(&filter->fo_translock);
100 oti->oti_transno = last_rcvd;
102 spin_lock(&filter->fo_translock);
103 last_rcvd = oti->oti_transno;
104 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
105 filter->fo_fsd->lsd_last_transno =
106 cpu_to_le64(last_rcvd);
107 spin_unlock(&filter->fo_translock);
109 fcd->fcd_last_rcvd = cpu_to_le64(last_rcvd);
111 /* could get xid from oti, if it's ever needed */
112 fcd->fcd_last_xid = 0;
114 off = fed->fed_lr_off;
116 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
117 fed->fed_lr_idx, fed->fed_lr_off);
121 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
127 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
128 fcd, sizeof(*fcd), &off,
129 force_sync | exp->exp_need_sync);
131 filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
139 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
140 last_rcvd, fcd->fcd_uuid, fed->fed_lr_idx, err);
145 void f_dput(struct dentry *dentry)
147 /* Can't go inside filter_ddelete because it can block */
148 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
149 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
150 LASSERT(atomic_read(&dentry->d_count) > 0);
155 static void init_brw_stats(struct brw_stats *brw_stats)
158 for (i = 0; i < BRW_LAST; i++)
159 spin_lock_init(&brw_stats->hist[i].oh_lock);
162 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
163 plus the procfs overhead :( */
164 static int filter_export_stats_init(struct obd_device *obd,
165 struct obd_export *exp)
167 struct filter_export_data *fed = &exp->exp_filter_data;
168 struct proc_dir_entry *brw_entry;
172 init_brw_stats(&fed->fed_brw_stats);
174 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
175 /* Self-export gets no proc entry */
178 rc = lprocfs_exp_setup(exp);
182 /* Create a per export proc entry for brw_stats */
183 brw_entry = create_proc_entry("brw_stats", 0644, exp->exp_proc);
184 if (brw_entry == NULL)
186 brw_entry->proc_fops = &filter_per_export_stats_fops;
187 brw_entry->data = fed;
189 /* Create a per export proc entry for ops stats */
190 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
191 LPROC_FILTER_LAST - 1;
192 exp->exp_ops_stats = lprocfs_alloc_stats(num_stats,
193 LPROCFS_STATS_FLAG_NOPERCPU);
194 if (exp->exp_ops_stats == NULL)
196 lprocfs_init_ops_stats(LPROC_FILTER_LAST, exp->exp_ops_stats);
197 lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_READ_BYTES,
198 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
199 lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_WRITE_BYTES,
200 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
201 lprocfs_register_stats(exp->exp_proc, "stats", exp->exp_ops_stats);
205 /* Add client data to the FILTER. We use a bitmap to locate a free space
206 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
207 * Otherwise, we have just read the data from the last_rcvd file and
208 * we know its offset. */
209 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
212 struct filter_obd *filter = &obd->u.filter;
213 struct filter_export_data *fed = &exp->exp_filter_data;
214 unsigned long *bitmap = filter->fo_last_rcvd_slots;
215 int new_client = (cl_idx == -1);
219 LASSERT(bitmap != NULL);
220 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
223 if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid) == 0)
226 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
227 * there's no need for extra complication here
230 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
232 if (cl_idx >= LR_MAX_CLIENTS) {
233 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
237 if (test_and_set_bit(cl_idx, bitmap)) {
238 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
243 if (test_and_set_bit(cl_idx, bitmap)) {
244 CERROR("FILTER client %d: bit already set in bitmap!\n",
250 fed->fed_lr_idx = cl_idx;
251 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
252 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
253 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
255 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
256 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
259 struct lvfs_run_ctxt saved;
260 loff_t off = fed->fed_lr_off;
264 CDEBUG(D_INFO, "writing client fcd at idx %u (%llu) (len %u)\n",
265 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_fcd));
267 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
268 /* Transaction needed to fix bug 1403 */
269 handle = fsfilt_start(obd,
270 filter->fo_rcvd_filp->f_dentry->d_inode,
271 FSFILT_OP_SETATTR, NULL);
272 if (IS_ERR(handle)) {
273 rc = PTR_ERR(handle);
274 CERROR("unable to start transaction: rc %d\n", rc);
276 rc = fsfilt_add_journal_cb(obd, 0, handle,
277 target_client_add_cb, exp);
279 spin_lock(&exp->exp_lock);
280 exp->exp_need_sync = 1;
281 spin_unlock(&exp->exp_lock);
283 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
285 sizeof(*fed->fed_fcd),
286 &off, rc /* sync if no cb */);
288 filter->fo_rcvd_filp->f_dentry->d_inode,
291 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
294 CERROR("error writing %s client idx %u: rc %d\n",
295 LAST_RCVD, fed->fed_lr_idx, rc);
302 static int filter_client_free(struct obd_export *exp)
304 struct filter_export_data *fed = &exp->exp_filter_data;
305 struct filter_obd *filter = &exp->exp_obd->u.filter;
306 struct obd_device *obd = exp->exp_obd;
307 struct filter_client_data zero_fcd;
308 struct lvfs_run_ctxt saved;
313 if (fed->fed_fcd == NULL)
316 /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
317 if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid ) == 0)
320 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
321 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
323 LASSERT(filter->fo_last_rcvd_slots != NULL);
325 off = fed->fed_lr_off;
327 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
328 * we leak a client slot that will be cleaned on the next recovery. */
330 CERROR("%s: client idx %d has med_off %lld\n",
331 obd->obd_name, fed->fed_lr_idx, off);
332 GOTO(free, rc = -EINVAL);
335 /* Clear the bit _after_ zeroing out the client so we don't
336 race with filter_client_add and zero out new clients.*/
337 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
338 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
343 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
344 memset(&zero_fcd, 0, sizeof zero_fcd);
345 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
346 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_fcd,
347 sizeof(zero_fcd), &off,
348 (!exp->exp_libclient ||
349 exp->exp_need_sync));
351 /* update server's transno */
352 filter_update_server_data(obd, filter->fo_rcvd_filp,
354 !exp->exp_libclient);
355 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
357 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
358 "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
359 fed->fed_fcd->fcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
363 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
364 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
371 OBD_FREE(fed->fed_fcd, sizeof(*fed->fed_fcd));
377 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
378 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
379 struct filter_mod_data *fmd)
381 LASSERT_SPIN_LOCKED(&fed->fed_lock);
382 if (--fmd->fmd_refcount == 0) {
383 /* XXX when we have persistent reservations and the handle
384 * is stored herein we need to drop it here. */
385 fed->fed_mod_count--;
386 list_del(&fmd->fmd_list);
387 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
391 /* drop fmd reference, free it if last ref */
392 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
394 struct filter_export_data *fed;
399 fed = &exp->exp_filter_data;
400 spin_lock(&fed->fed_lock);
401 filter_fmd_put_nolock(fed, fmd); /* caller reference */
402 spin_unlock(&fed->fed_lock);
405 /* expire entries from the end of the list if there are too many
406 * or they are too old */
407 static void filter_fmd_expire_nolock(struct filter_obd *filter,
408 struct filter_export_data *fed,
409 struct filter_mod_data *keep)
411 struct filter_mod_data *fmd, *tmp;
413 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
417 if (time_before(jiffies, fmd->fmd_expire) &&
418 fed->fed_mod_count < filter->fo_fmd_max_num)
421 list_del_init(&fmd->fmd_list);
422 filter_fmd_put_nolock(fed, fmd); /* list reference */
426 void filter_fmd_expire(struct obd_export *exp)
428 spin_lock(&exp->exp_filter_data.fed_lock);
429 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
430 &exp->exp_filter_data, NULL);
431 spin_unlock(&exp->exp_filter_data.fed_lock);
434 /* find specified objid, group in export fmd list.
435 * caller must hold fed_lock and take fmd reference itself */
436 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
437 struct filter_export_data *fed,
438 obd_id objid, obd_gr group)
440 struct filter_mod_data *found = NULL, *fmd;
442 LASSERT_SPIN_LOCKED(&fed->fed_lock);
444 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
445 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
447 list_del(&fmd->fmd_list);
448 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
449 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
454 filter_fmd_expire_nolock(filter, fed, found);
459 /* Find fmd based on objid and group, or return NULL if not found. */
460 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
461 obd_id objid, obd_gr group)
463 struct filter_mod_data *fmd;
465 spin_lock(&exp->exp_filter_data.fed_lock);
466 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
467 &exp->exp_filter_data, objid, group);
469 fmd->fmd_refcount++; /* caller reference */
470 spin_unlock(&exp->exp_filter_data.fed_lock);
475 /* Find fmd based on objid and group, or create a new one if none is found.
476 * It is possible for this function to return NULL under memory pressure,
477 * or if objid = 0 is passed (which will only cause old entries to expire).
478 * Currently this is not fatal because any fmd state is transient and
479 * may also be freed when it gets sufficiently old. */
480 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
481 obd_id objid, obd_gr group)
483 struct filter_export_data *fed = &exp->exp_filter_data;
484 struct filter_mod_data *found = NULL, *fmd_new = NULL;
486 OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO, sizeof(*fmd_new));
488 spin_lock(&fed->fed_lock);
489 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
492 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
493 fmd_new->fmd_id = objid;
494 fmd_new->fmd_gr = group;
495 fmd_new->fmd_refcount++; /* list reference */
497 fed->fed_mod_count++;
499 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
503 found->fmd_refcount++; /* caller reference */
504 found->fmd_expire = jiffies +
505 exp->exp_obd->u.filter.fo_fmd_max_age;
508 spin_unlock(&fed->fed_lock);
514 /* drop fmd list reference so it will disappear when last reference is put.
515 * This isn't so critical because it would in fact only affect the one client
516 * that is doing the unlink and at worst we have an stale entry referencing
517 * an object that should never be used again. */
518 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
520 struct filter_mod_data *found = NULL;
522 spin_lock(&exp->exp_filter_data.fed_lock);
523 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
525 list_del_init(&found->fmd_list);
526 filter_fmd_put_nolock(&exp->exp_filter_data, found);
528 spin_unlock(&exp->exp_filter_data.fed_lock);
531 #define filter_fmd_drop(exp, objid, group)
534 /* remove all entries from fmd list */
535 static void filter_fmd_cleanup(struct obd_export *exp)
537 struct filter_export_data *fed = &exp->exp_filter_data;
538 struct filter_mod_data *fmd = NULL, *tmp;
540 spin_lock(&fed->fed_lock);
541 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
542 list_del_init(&fmd->fmd_list);
543 filter_fmd_put_nolock(fed, fmd);
545 spin_unlock(&fed->fed_lock);
548 static int filter_init_export(struct obd_export *exp)
550 spin_lock_init(&exp->exp_filter_data.fed_lock);
551 INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
553 spin_lock(&exp->exp_lock);
554 exp->exp_connecting = 1;
555 spin_unlock(&exp->exp_lock);
560 static int filter_free_server_data(struct filter_obd *filter)
562 OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
563 filter->fo_fsd = NULL;
564 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
565 filter->fo_last_rcvd_slots = NULL;
569 /* assumes caller is already in kernel ctxt */
570 int filter_update_server_data(struct obd_device *obd, struct file *filp,
571 struct lr_server_data *fsd, int force_sync)
577 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
578 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
579 le64_to_cpu(fsd->lsd_last_transno));
580 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
581 le64_to_cpu(fsd->lsd_mount_count));
583 fsd->lsd_compat14 = fsd->lsd_last_transno;
584 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
586 CERROR("error writing lr_server_data: rc = %d\n", rc);
591 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
594 struct filter_obd *filter = &obd->u.filter;
600 if (filter->fo_last_objid_files[group] == NULL) {
601 CERROR("Object group "LPU64" not fully setup; not updating "
602 "last_objid\n", group);
606 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
607 obd->obd_name, group, filter->fo_last_objids[group]);
609 tmp = cpu_to_le64(filter->fo_last_objids[group]);
610 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
611 &tmp, sizeof(tmp), &off, force_sync);
613 CERROR("error writing group "LPU64" last objid: rc = %d\n",
617 extern int ost_handle(struct ptlrpc_request *req);
618 /* assumes caller has already in kernel ctxt */
619 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
621 struct filter_obd *filter = &obd->u.filter;
622 struct lr_server_data *fsd;
623 struct filter_client_data *fcd = NULL;
624 struct inode *inode = filp->f_dentry->d_inode;
625 unsigned long last_rcvd_size = i_size_read(inode);
631 /* ensure padding in the struct is the correct size */
632 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
633 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
634 CLASSERT (offsetof(struct filter_client_data, fcd_padding) +
635 sizeof(fcd->fcd_padding) == LR_CLIENT_SIZE);
637 OBD_ALLOC(fsd, sizeof(*fsd));
640 filter->fo_fsd = fsd;
642 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
643 if (filter->fo_last_rcvd_slots == NULL) {
644 OBD_FREE(fsd, sizeof(*fsd));
648 if (last_rcvd_size == 0) {
649 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
651 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
652 fsd->lsd_last_transno = 0;
653 mount_count = fsd->lsd_mount_count = 0;
654 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
655 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
656 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
657 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
658 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
659 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
661 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
663 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
667 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
668 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
669 "using the wrong disk %s. Were the "
670 "/dev/ assignments rearranged?\n",
671 obd->obd_uuid.uuid, fsd->lsd_uuid);
672 GOTO(err_fsd, rc = -EINVAL);
674 mount_count = le64_to_cpu(fsd->lsd_mount_count);
675 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
677 /* Assume old last_rcvd format unless I_C_LR is set */
678 if (!(fsd->lsd_feature_incompat &
679 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
680 fsd->lsd_last_transno = fsd->lsd_compat14;
684 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
685 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
686 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
687 ~FILTER_INCOMPAT_SUPP);
688 GOTO(err_fsd, rc = -EINVAL);
690 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
691 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
692 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
693 ~FILTER_ROCOMPAT_SUPP);
694 /* Do something like remount filesystem read-only */
695 GOTO(err_fsd, rc = -EINVAL);
698 CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
699 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
700 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
701 obd->obd_name, mount_count + 1);
702 CDEBUG(D_INODE, "%s: server data size: %u\n",
703 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
704 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
705 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
706 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
707 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
708 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
709 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
710 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
711 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
712 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
713 le16_to_cpu(fsd->lsd_client_size));
715 if (!obd->obd_replayable) {
716 CWARN("%s: recovery support OFF\n", obd->obd_name);
720 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
721 off < last_rcvd_size; cl_idx++) {
723 struct obd_export *exp;
724 struct filter_export_data *fed;
727 OBD_ALLOC(fcd, sizeof(*fcd));
729 GOTO(err_client, rc = -ENOMEM);
732 /* Don't assume off is incremented properly by
733 * fsfilt_read_record(), in case sizeof(*fcd)
734 * isn't the same as fsd->lsd_client_size. */
735 off = le32_to_cpu(fsd->lsd_client_start) +
736 cl_idx * le16_to_cpu(fsd->lsd_client_size);
737 rc = fsfilt_read_record(obd, filp, fcd, sizeof(*fcd), &off);
739 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
740 LAST_RCVD, cl_idx, off, rc);
741 break; /* read error shouldn't cause startup to fail */
744 if (fcd->fcd_uuid[0] == '\0') {
745 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
750 last_rcvd = le64_to_cpu(fcd->fcd_last_rcvd);
752 /* These exports are cleaned up by filter_disconnect(), so they
753 * need to be set up like real exports as filter_connect() does.
755 exp = class_new_export(obd, (struct obd_uuid *)fcd->fcd_uuid);
757 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
758 " srv lr: "LPU64" fcd_group %d\n", fcd->fcd_uuid, cl_idx,
759 last_rcvd, le64_to_cpu(fsd->lsd_last_transno),
760 le32_to_cpu(fcd->fcd_group));
762 if (PTR_ERR(exp) == -EALREADY) {
763 /* export already exists, zero out this one */
764 CERROR("Zeroing out duplicate export due to "
766 fcd->fcd_uuid[0] = '\0';
768 GOTO(err_client, rc = PTR_ERR(exp));
771 fed = &exp->exp_filter_data;
773 fed->fed_group = le32_to_cpu(fcd->fcd_group);
774 filter_export_stats_init(obd, exp);
775 rc = filter_client_add(obd, exp, cl_idx);
776 /* can't fail for existing client */
777 LASSERTF(rc == 0, "rc = %d\n", rc);
780 spin_lock(&exp->exp_lock);
781 exp->exp_connecting = 0;
782 exp->exp_in_recovery = 0;
783 spin_unlock(&exp->exp_lock);
784 obd->obd_max_recoverable_clients++;
785 class_export_put(exp);
788 /* Need to check last_rcvd even for duplicated exports. */
789 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
792 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
793 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
797 OBD_FREE(fcd, sizeof(*fcd));
799 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
801 target_recovery_init(obd, ost_handle);
804 filter->fo_mount_count = mount_count + 1;
805 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
807 /* save it, so mount count and last_transno is current */
808 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
810 GOTO(err_client, rc);
815 target_recovery_fini(obd);
817 filter_free_server_data(filter);
821 static int filter_cleanup_groups(struct obd_device *obd)
823 struct filter_obd *filter = &obd->u.filter;
825 struct dentry *dentry;
829 if (filter->fo_dentry_O_groups != NULL) {
830 for (i = 0; i < filter->fo_group_count; i++) {
831 dentry = filter->fo_dentry_O_groups[i];
835 OBD_FREE(filter->fo_dentry_O_groups,
836 filter->fo_group_count *
837 sizeof(*filter->fo_dentry_O_groups));
838 filter->fo_dentry_O_groups = NULL;
840 if (filter->fo_last_objid_files != NULL) {
841 for (i = 0; i < filter->fo_group_count; i++) {
842 filp = filter->fo_last_objid_files[i];
846 OBD_FREE(filter->fo_last_objid_files,
847 filter->fo_group_count *
848 sizeof(*filter->fo_last_objid_files));
849 filter->fo_last_objid_files = NULL;
851 if (filter->fo_dentry_O_sub != NULL) {
852 for (i = 0; i < filter->fo_group_count; i++) {
853 for (j = 0; j < filter->fo_subdir_count; j++) {
854 dentry = filter->fo_dentry_O_sub[i].dentry[j];
859 OBD_FREE(filter->fo_dentry_O_sub,
860 filter->fo_group_count *
861 sizeof(*filter->fo_dentry_O_sub));
862 filter->fo_dentry_O_sub = NULL;
864 if (filter->fo_last_objids != NULL) {
865 OBD_FREE(filter->fo_last_objids,
866 filter->fo_group_count *
867 sizeof(*filter->fo_last_objids));
868 filter->fo_last_objids = NULL;
870 if (filter->fo_dentry_O != NULL) {
871 f_dput(filter->fo_dentry_O);
872 filter->fo_dentry_O = NULL;
877 static int filter_update_last_group(struct obd_device *obd, int group)
879 struct filter_obd *filter = &obd->u.filter;
880 struct file *filp = NULL;
881 int last_group = 0, rc;
885 if (group <= filter->fo_committed_group)
888 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
892 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
896 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
898 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
901 LASSERT(off == 0 || last_group >= FILTER_MIN_GROUPS);
902 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
903 obd->obd_name, last_group, group);
907 /* must be sync: bXXXX */
908 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
910 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
914 filter->fo_committed_group = group;
921 static int filter_read_group_internal(struct obd_device *obd, int group,
924 struct filter_obd *filter = &obd->u.filter;
925 __u64 *new_objids = NULL;
926 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
927 struct dentry **new_groups = NULL;
928 struct file **new_files = NULL;
929 struct dentry *dentry;
931 int old_count = filter->fo_group_count, rc, stage = 0, i;
937 snprintf(name, 24, "%d", group);
941 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
943 if (IS_ERR(dentry)) {
944 CERROR("Cannot lookup expected object group %d: %ld\n",
945 group, PTR_ERR(dentry));
946 RETURN(PTR_ERR(dentry));
949 dentry = simple_mkdir(filter->fo_dentry_O, name, 0700, 1);
950 if (IS_ERR(dentry)) {
951 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
953 RETURN(PTR_ERR(dentry));
958 snprintf(name, 24, "O/%d/LAST_ID", group);
960 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
962 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
963 GOTO(cleanup, rc = PTR_ERR(filp));
967 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
969 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
973 if (filter->fo_subdir_count) {
974 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
975 if (tmp_subdirs == NULL)
976 GOTO(cleanup, rc = -ENOMEM);
979 for (i = 0; i < filter->fo_subdir_count; i++) {
981 snprintf(dir, sizeof(dir), "d%u", i);
983 tmp_subdirs->dentry[i] = simple_mkdir(dentry, dir, 0700, 1);
984 if (IS_ERR(tmp_subdirs->dentry[i])) {
985 rc = PTR_ERR(tmp_subdirs->dentry[i]);
986 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
991 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
992 tmp_subdirs->dentry[i]);
996 /* 'group' is an index; we need an array of length 'group + 1' */
997 if (group + 1 > old_count) {
998 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
999 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1000 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1001 OBD_ALLOC(new_files, len * sizeof(*new_files));
1003 if (new_objids == NULL || new_subdirs == NULL ||
1004 new_groups == NULL || new_files == NULL)
1005 GOTO(cleanup, rc = -ENOMEM);
1008 memcpy(new_objids, filter->fo_last_objids,
1009 old_count * sizeof(*new_objids));
1010 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1011 old_count * sizeof(*new_subdirs));
1012 memcpy(new_groups, filter->fo_dentry_O_groups,
1013 old_count * sizeof(*new_groups));
1014 memcpy(new_files, filter->fo_last_objid_files,
1015 old_count * sizeof(*new_files));
1017 OBD_FREE(filter->fo_last_objids,
1018 old_count * sizeof(*new_objids));
1019 OBD_FREE(filter->fo_dentry_O_sub,
1020 old_count * sizeof(*new_subdirs));
1021 OBD_FREE(filter->fo_dentry_O_groups,
1022 old_count * sizeof(*new_groups));
1023 OBD_FREE(filter->fo_last_objid_files,
1024 old_count * sizeof(*new_files));
1026 filter->fo_last_objids = new_objids;
1027 filter->fo_dentry_O_sub = new_subdirs;
1028 filter->fo_dentry_O_groups = new_groups;
1029 filter->fo_last_objid_files = new_files;
1030 filter->fo_group_count = len;
1033 filter->fo_dentry_O_groups[group] = dentry;
1034 filter->fo_last_objid_files[group] = filp;
1035 if (filter->fo_subdir_count) {
1036 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1037 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1040 filter_update_last_group(obd, group);
1042 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1043 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1044 rc = filter_update_last_objid(obd, group, 1);
1048 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1049 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1050 obd->obd_name, group, last_objid);
1055 if (new_objids != NULL)
1056 OBD_FREE(new_objids, len * sizeof(*new_objids));
1057 if (new_subdirs != NULL)
1058 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1059 if (new_groups != NULL)
1060 OBD_FREE(new_groups, len * sizeof(*new_groups));
1061 if (new_files != NULL)
1062 OBD_FREE(new_files, len * sizeof(*new_files));
1064 if (filter->fo_subdir_count) {
1065 for (i = 0; i < filter->fo_subdir_count; i++) {
1066 if (tmp_subdirs->dentry[i] != NULL)
1067 dput(tmp_subdirs->dentry[i]);
1069 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1072 filp_close(filp, 0);
1079 static int filter_read_groups(struct obd_device *obd, int last_group,
1082 struct filter_obd *filter = &obd->u.filter;
1083 int old_count, group, rc = 0;
1085 down(&filter->fo_init_lock);
1086 old_count = filter->fo_group_count;
1087 for (group = old_count; group <= last_group; group++) {
1089 continue; /* no group zero */
1091 rc = filter_read_group_internal(obd, group, create);
1095 up(&filter->fo_init_lock);
1099 /* FIXME: object groups */
1100 static int filter_prep_groups(struct obd_device *obd)
1102 struct filter_obd *filter = &obd->u.filter;
1103 struct dentry *dentry, *O_dentry;
1105 int last_group, rc = 0, cleanup_phase = 0;
1109 O_dentry = simple_mkdir(current->fs->pwd, "O", 0700, 1);
1110 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1111 if (IS_ERR(O_dentry)) {
1112 rc = PTR_ERR(O_dentry);
1113 CERROR("cannot open/create O: rc = %d\n", rc);
1116 filter->fo_dentry_O = O_dentry;
1117 cleanup_phase = 1; /* O_dentry */
1119 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1120 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1121 * some time post 1.0 when all old-style OSTs have converted along
1122 * with the init_objid hack. */
1123 dentry = ll_lookup_one_len("R", O_dentry, 1);
1125 GOTO(cleanup, rc = PTR_ERR(dentry));
1126 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1127 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1130 CWARN("converting OST to new object layout\n");
1131 if (IS_ERR(O0_dentry)) {
1132 rc = PTR_ERR(O0_dentry);
1133 CERROR("error looking up O/0: rc %d\n", rc);
1134 GOTO(cleanup_R, rc);
1137 if (O0_dentry->d_inode) {
1138 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1139 GOTO(cleanup_O0, rc = -EEXIST);
1142 LOCK_INODE_MUTEX(O_dentry->d_inode);
1143 rc = vfs_rename(O_dentry->d_inode, dentry,
1144 O_dentry->d_inode, O0_dentry);
1145 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1148 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1149 GOTO(cleanup_O0, rc);
1151 filter->fo_fsd->lsd_feature_incompat |=
1152 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1153 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1155 GOTO(cleanup_O0, rc);
1167 cleanup_phase = 2; /* groups */
1169 /* we have to initialize all groups before first connections from
1170 * clients because they may send create/destroy for any group -bzzz */
1171 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1173 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1174 GOTO(cleanup, rc = PTR_ERR(filp));
1176 cleanup_phase = 3; /* filp */
1178 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1180 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1184 last_group = FILTER_MIN_GROUPS;
1186 LASSERT(last_group >= FILTER_MIN_GROUPS);
1189 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1190 FILTER_MIN_GROUPS, last_group);
1191 filter->fo_committed_group = last_group;
1192 rc = filter_read_groups(obd, last_group, 1);
1196 filp_close(filp, 0);
1200 switch (cleanup_phase) {
1202 filp_close(filp, 0);
1204 filter_cleanup_groups(obd);
1206 f_dput(filter->fo_dentry_O);
1207 filter->fo_dentry_O = NULL;
1215 /* setup the object store with correct subdirectories */
1216 static int filter_prep(struct obd_device *obd)
1218 struct lvfs_run_ctxt saved;
1219 struct filter_obd *filter = &obd->u.filter;
1221 struct inode *inode;
1225 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1226 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1227 if (!file || IS_ERR(file)) {
1229 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1233 filter->fo_rcvd_filp = file;
1234 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1235 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1236 file->f_dentry->d_inode->i_mode);
1237 GOTO(err_filp, rc = -ENOENT);
1240 inode = file->f_dentry->d_parent->d_inode;
1241 /* We use i_op->unlink directly in filter_vfs_unlink() */
1242 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1243 CERROR("%s: filesystem does not support create/unlink ops\n",
1245 GOTO(err_filp, rc = -EOPNOTSUPP);
1248 rc = filter_init_server_data(obd, file);
1250 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1253 /* open and create health check io file*/
1254 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1257 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1259 GOTO(err_server_data, rc);
1261 filter->fo_health_check_filp = file;
1262 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1263 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1264 file->f_dentry->d_inode->i_mode);
1265 GOTO(err_health_check, rc = -ENOENT);
1267 rc = lvfs_check_io_health(obd, file);
1269 GOTO(err_health_check, rc);
1271 rc = filter_prep_groups(obd);
1273 GOTO(err_health_check, rc);
1275 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1280 if (filp_close(filter->fo_health_check_filp, 0))
1281 CERROR("can't close %s after error\n", HEALTH_CHECK);
1282 filter->fo_health_check_filp = NULL;
1284 target_recovery_fini(obd);
1285 filter_free_server_data(filter);
1287 if (filp_close(filter->fo_rcvd_filp, 0))
1288 CERROR("can't close %s after error\n", LAST_RCVD);
1289 filter->fo_rcvd_filp = NULL;
1293 /* cleanup the filter: write last used object id to status file */
1294 static void filter_post(struct obd_device *obd)
1296 struct lvfs_run_ctxt saved;
1297 struct filter_obd *filter = &obd->u.filter;
1300 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1301 * best to start a transaction with h_sync, because we removed this
1304 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1305 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1308 CERROR("error writing server data: rc = %d\n", rc);
1310 for (i = 1; i < filter->fo_group_count; i++) {
1311 rc = filter_update_last_objid(obd, i,
1312 (i == filter->fo_group_count - 1));
1314 CERROR("error writing group %d lastobjid: rc = %d\n",
1318 rc = filp_close(filter->fo_rcvd_filp, 0);
1319 filter->fo_rcvd_filp = NULL;
1321 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1323 rc = filp_close(filter->fo_health_check_filp, 0);
1324 filter->fo_health_check_filp = NULL;
1326 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1328 filter_cleanup_groups(obd);
1329 filter_free_server_data(filter);
1330 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1332 filter_free_capa_keys(filter);
1333 cleanup_capa_hash(filter->fo_capa_hash);
1336 static void filter_set_last_id(struct filter_obd *filter,
1337 obd_id id, obd_gr group)
1339 LASSERT(filter->fo_fsd != NULL);
1340 LASSERT(group <= filter->fo_group_count);
1342 spin_lock(&filter->fo_objidlock);
1343 filter->fo_last_objids[group] = id;
1344 spin_unlock(&filter->fo_objidlock);
1347 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1350 LASSERT(filter->fo_fsd != NULL);
1351 LASSERT(group <= filter->fo_group_count);
1353 /* FIXME: object groups */
1354 spin_lock(&filter->fo_objidlock);
1355 id = filter->fo_last_objids[group];
1356 spin_unlock(&filter->fo_objidlock);
1361 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1363 LOCK_INODE_MUTEX(dparent->d_inode);
1367 /* We never dget the object parent, so DON'T dput it either */
1368 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1370 struct filter_obd *filter = &obd->u.filter;
1371 struct filter_subdirs *subdirs;
1372 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1374 if (group > 0 || filter->fo_subdir_count == 0)
1375 return filter->fo_dentry_O_groups[group];
1377 subdirs = &filter->fo_dentry_O_sub[group];
1378 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1381 /* We never dget the object parent, so DON'T dput it either */
1382 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1385 unsigned long now = jiffies;
1386 struct dentry *dparent = filter_parent(obd, group, objid);
1389 if (IS_ERR(dparent))
1391 if (dparent == NULL)
1392 return ERR_PTR(-ENOENT);
1394 rc = filter_lock_dentry(obd, dparent);
1395 fsfilt_check_slow(obd, now, obd_timeout, "parent lock");
1396 return rc ? ERR_PTR(rc) : dparent;
1399 /* We never dget the object parent, so DON'T dput it either */
1400 static void filter_parent_unlock(struct dentry *dparent)
1402 UNLOCK_INODE_MUTEX(dparent->d_inode);
1405 /* How to get files, dentries, inodes from object id's.
1407 * If dir_dentry is passed, the caller has already locked the parent
1408 * appropriately for this operation (normally a write lock). If
1409 * dir_dentry is NULL, we do a read lock while we do the lookup to
1410 * avoid races with create/destroy and such changing the directory
1411 * internal to the filesystem code. */
1412 struct dentry *filter_fid2dentry(struct obd_device *obd,
1413 struct dentry *dir_dentry,
1414 obd_gr group, obd_id id)
1416 struct dentry *dparent = dir_dentry;
1417 struct dentry *dchild;
1422 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1423 obd->u.filter.fo_destroys_in_progress == 0) {
1424 /* don't fail lookups for orphan recovery, it causes
1425 * later LBUGs when objects still exist during precreate */
1426 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1427 RETURN(ERR_PTR(-ENOENT));
1430 CERROR("fatal: invalid object id 0\n");
1431 RETURN(ERR_PTR(-ESTALE));
1434 len = sprintf(name, LPU64, id);
1435 if (dir_dentry == NULL) {
1436 dparent = filter_parent_lock(obd, group, id);
1437 if (IS_ERR(dparent)) {
1438 CERROR("%s: error getting object "LPU64":"LPU64
1439 " parent: rc %ld\n", obd->obd_name,
1440 id, group, PTR_ERR(dparent));
1444 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1445 dparent->d_name.len, dparent->d_name.name, name);
1446 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1447 if (dir_dentry == NULL)
1448 filter_parent_unlock(dparent);
1449 if (IS_ERR(dchild)) {
1450 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1455 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1456 CERROR("%s: got bad object "LPU64" inode %lu\n",
1457 obd->obd_name, id, dchild->d_inode->i_ino);
1459 RETURN(ERR_PTR(-ENOENT));
1462 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1463 name, dchild, atomic_read(&dchild->d_count));
1465 LASSERT(atomic_read(&dchild->d_count) > 0);
1470 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1473 struct lustre_handle lockh;
1474 int flags = LDLM_AST_DISCARD_DATA, rc;
1475 struct ldlm_res_id res_id = { .name = { objid, 0, group, 0} };
1476 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1479 /* Tell the clients that the object is gone now and that they should
1480 * throw away any cached pages. */
1481 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1482 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1483 ldlm_completion_ast, NULL, NULL, 0, NULL,
1486 /* We only care about the side-effects, just drop the lock. */
1488 ldlm_lock_decref(&lockh, LCK_PW);
1493 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1494 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1495 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1496 * reverse. Caller must take i_sem before starting the transaction and we
1497 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1498 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry)
1503 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1504 * itself we already hold dir->i_mutex for child create/unlink ops */
1505 LASSERT(dentry->d_inode != NULL);
1506 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1507 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1511 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1512 GOTO(out, rc = -ENOENT);
1514 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1519 GOTO(out, rc = -EPERM);
1521 /* check_sticky() */
1522 if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1523 IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1524 GOTO(out, rc = -EPERM);
1526 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1527 * that was done because of journal_start (which is already done
1528 * here) or some other ordering issue. */
1531 rc = security_inode_unlink(dir, dentry);
1535 rc = dir->i_op->unlink(dir, dentry);
1537 /* need to drop i_mutex before we lose inode reference */
1538 UNLOCK_INODE_MUTEX(dentry->d_inode);
1545 /* Caller must hold LCK_PW on parent and push us into kernel context.
1546 * Caller must hold child i_mutex, we drop it always.
1547 * Caller is also required to ensure that dchild->d_inode exists. */
1548 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1549 obd_gr group, struct dentry *dparent,
1550 struct dentry *dchild)
1552 struct inode *inode = dchild->d_inode;
1555 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1556 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1557 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1558 (unsigned long)inode->i_nlink,
1559 atomic_read(&inode->i_count));
1562 rc = filter_vfs_unlink(dparent->d_inode, dchild);
1564 CERROR("error unlinking objid %.*s: rc %d\n",
1565 dchild->d_name.len, dchild->d_name.name, rc);
1569 struct filter_intent_args {
1570 struct ldlm_lock **victim;
1575 static enum interval_iter filter_intent_cb(struct interval_node *n,
1578 struct ldlm_interval *node = (struct ldlm_interval *)n;
1579 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1580 __u64 size = arg->size;
1581 struct ldlm_lock **v = arg->victim;
1582 struct ldlm_lock *lck;
1584 /* If the interval is lower than the current file size,
1586 if (interval_high(n) <= size)
1587 return INTERVAL_ITER_STOP;
1589 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1590 /* Don't send glimpse ASTs to liblustre clients.
1591 * They aren't listening for them, and they do
1592 * entirely synchronous I/O anyways. */
1593 if (lck->l_export == NULL ||
1594 lck->l_export->exp_libclient == 1)
1597 if (*arg->liblustre)
1598 *arg->liblustre = 0;
1601 *v = LDLM_LOCK_GET(lck);
1602 } else if ((*v)->l_policy_data.l_extent.start <
1603 lck->l_policy_data.l_extent.start) {
1605 *v = LDLM_LOCK_GET(lck);
1608 /* the same policy group - every lock has the
1609 * same extent, so needn't do it any more */
1613 return INTERVAL_ITER_CONT;
1616 static int filter_intent_policy(struct ldlm_namespace *ns,
1617 struct ldlm_lock **lockp, void *req_cookie,
1618 ldlm_mode_t mode, int flags, void *data)
1620 struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1621 struct ptlrpc_request *req = req_cookie;
1622 struct ldlm_lock *lock = *lockp, *l = NULL;
1623 struct ldlm_resource *res = lock->l_resource;
1624 ldlm_processing_policy policy;
1625 struct ost_lvb *res_lvb, *reply_lvb;
1626 struct ldlm_reply *rep;
1628 int idx, rc, tmpflags = 0, only_liblustre = 1;
1629 struct ldlm_interval_tree *tree;
1630 struct filter_intent_args arg;
1631 int repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1632 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1633 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1636 policy = ldlm_get_processing_policy(res);
1637 LASSERT(policy != NULL);
1638 LASSERT(req != NULL);
1640 rc = lustre_pack_reply(req, 3, repsize, NULL);
1642 RETURN(req->rq_status = rc);
1644 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1645 LASSERT(rep != NULL);
1647 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1648 sizeof(*reply_lvb));
1649 LASSERT(reply_lvb != NULL);
1651 //fixup_handle_for_resent_req(req, lock, &lockh);
1653 /* If we grant any lock at all, it will be a whole-file read lock.
1654 * Call the extent policy function to see if our request can be
1655 * granted, or is blocked.
1656 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse lock
1658 lock->l_policy_data.l_extent.start = 0;
1659 lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1660 lock->l_req_mode = LCK_PR;
1662 LASSERT(ns == res->lr_namespace);
1664 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1665 check_res_locked(res);
1667 /* FIXME: we should change the policy function slightly, to not make
1668 * this list at all, since we just turn around and free it */
1669 while (!list_empty(&rpc_list)) {
1670 struct ldlm_lock *wlock =
1671 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1672 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1673 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1674 lock->l_flags &= ~LDLM_FL_CP_REQD;
1675 list_del_init(&wlock->l_cp_ast);
1676 LDLM_LOCK_PUT(wlock);
1679 /* The lock met with no resistance; we're finished. */
1680 if (rc == LDLM_ITER_CONTINUE) {
1681 /* do not grant locks to the liblustre clients: they cannot
1682 * handle ASTs robustly. We need to do this while still
1683 * holding ns_lock to avoid the lock remaining on the res_link
1684 * list (and potentially being added to l_pending_list by an
1685 * AST) when we are going to drop this lock ASAP. */
1686 if (lock->l_export->exp_libclient ||
1687 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1688 ldlm_resource_unlink_lock(lock);
1689 err = ELDLM_LOCK_ABORTED;
1691 err = ELDLM_LOCK_REPLACED;
1697 /* Do not grant any lock, but instead send GL callbacks. The extent
1698 * policy nicely created a list of all PW locks for us. We will choose
1699 * the highest of those which are larger than the size in the LVB, if
1700 * any, and perform a glimpse callback. */
1701 res_lvb = res->lr_lvb_data;
1702 LASSERT(res_lvb != NULL);
1703 *reply_lvb = *res_lvb;
1706 * ->ns_lock guarantees that no new locks are granted, and,
1707 * therefore, that res->lr_lvb_data cannot increase beyond the
1708 * end of already granted lock. As a result, it is safe to
1709 * check against "stale" reply_lvb->lvb_size value without
1712 arg.size = reply_lvb->lvb_size;
1714 arg.liblustre = &only_liblustre;
1715 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1716 tree = &res->lr_itree[idx];
1717 if (tree->lit_mode == LCK_PR)
1720 interval_iterate_reverse(tree->lit_root,
1721 filter_intent_cb, &arg);
1725 /* There were no PW locks beyond the size in the LVB; finished. */
1727 if (only_liblustre) {
1728 /* If we discovered a liblustre client with a PW lock,
1729 * however, the LVB may be out of date! The LVB is
1730 * updated only on glimpse (which we don't do for
1731 * liblustre clients) and cancel (which the client
1732 * obviously has not yet done). So if it has written
1733 * data but kept the lock, the LVB is stale and needs
1734 * to be updated from disk.
1736 * Of course, this will all disappear when we switch to
1737 * taking liblustre locks on the OST. */
1738 ldlm_res_lvbo_update(res, NULL, 0, 1);
1740 RETURN(ELDLM_LOCK_ABORTED);
1744 * This check is for lock taken in filter_prepare_destroy() that does
1745 * not have l_glimpse_ast set. So the logic is: if there is a lock
1746 * with no l_glimpse_ast set, this object is being destroyed already.
1748 * Hence, if you are grabbing DLM locks on the server, always set
1749 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1751 if (l->l_glimpse_ast == NULL) {
1752 /* We are racing with unlink(); just return -ENOENT */
1753 rep->lock_policy_res1 = -ENOENT;
1757 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1758 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1759 /* Update the LVB from disk if the AST failed (this is a legal race) */
1761 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1762 * sending ast is not handled. This can result in lost client writes.
1765 ldlm_res_lvbo_update(res, NULL, 0, 1);
1768 *reply_lvb = *res_lvb;
1774 RETURN(ELDLM_LOCK_ABORTED);
1778 * per-obd_device iobuf pool.
1780 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1781 * allocations in write-path has to be minimized (see bug 5137).
1783 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1784 * OST threads (see ost_thread_{init,done}()).
1786 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1787 * at the OST layer there are only (potentially) multiple obd_device of type
1788 * unknown at the time of OST thread creation.
1790 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1791 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1792 * it's very own iobuf.
1796 * filter_kiobuf_pool_init()
1798 * filter_kiobuf_pool_done()
1800 * filter_iobuf_get()
1802 * operate on this array. They are "generic" in a sense that they don't depend
1803 * on actual type of iobuf's (the latter depending on Linux kernel version).
1807 * destroy pool created by filter_iobuf_pool_init
1809 static void filter_iobuf_pool_done(struct filter_obd *filter)
1811 struct filter_iobuf **pool;
1816 pool = filter->fo_iobuf_pool;
1818 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1819 if (pool[i] != NULL)
1820 filter_free_iobuf(pool[i]);
1822 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1823 filter->fo_iobuf_pool = NULL;
1829 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1831 static int filter_iobuf_pool_init(struct filter_obd *filter)
1838 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1840 if (filter->fo_iobuf_pool == NULL)
1843 filter->fo_iobuf_count = OSS_THREADS_MAX;
1848 /* Return iobuf allocated for @thread_id. We don't know in advance how
1849 * many threads there will be so we allocate a large empty array and only
1850 * fill in those slots that are actually in use.
1851 * If we haven't allocated a pool entry for this thread before, do so now. */
1852 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1854 int thread_id = oti ? oti->oti_thread_id : -1;
1855 struct filter_iobuf *pool = NULL;
1856 struct filter_iobuf **pool_place = NULL;
1858 if (thread_id >= 0) {
1859 LASSERT(thread_id < filter->fo_iobuf_count);
1860 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1863 if (unlikely(pool == NULL)) {
1864 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1865 PTLRPC_MAX_BRW_PAGES);
1866 if (pool_place != NULL)
1873 /* mount the file system (secretly). lustre_cfg parameters are:
1876 * 3 = flags: failover=f, failout=n
1879 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1882 struct filter_obd *filter = &obd->u.filter;
1883 struct vfsmount *mnt;
1884 struct lustre_mount_info *lmi;
1885 struct obd_uuid uuid;
1892 if (lcfg->lcfg_bufcount < 3 ||
1893 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1894 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1897 lmi = server_get_mount(obd->obd_name);
1899 /* We already mounted in lustre_fill_super.
1900 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1901 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1903 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1905 /* old path - used by lctl */
1906 CERROR("Using old MDS mount method\n");
1907 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1908 MS_NOATIME|MS_NODIRATIME,
1909 lustre_cfg_string(lcfg, 1), option);
1912 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1913 lustre_cfg_string(lcfg, 1), rc);
1917 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1919 if (IS_ERR(obd->obd_fsops))
1920 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1922 rc = filter_iobuf_pool_init(filter);
1926 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1928 /* failover is the default */
1929 obd->obd_replayable = 1;
1931 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1932 str = lustre_cfg_string(lcfg, 3);
1933 if (strchr(str, 'n')) {
1934 CWARN("%s: recovery disabled\n", obd->obd_name);
1935 obd->obd_replayable = 0;
1939 filter->fo_vfsmnt = mnt;
1940 obd->u.obt.obt_sb = mnt->mnt_sb;
1941 filter->fo_fstype = mnt->mnt_sb->s_type->name;
1942 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1944 fsfilt_setup(obd, obd->u.obt.obt_sb);
1946 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1947 obd->obd_lvfs_ctxt.pwdmnt = mnt;
1948 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1949 obd->obd_lvfs_ctxt.fs = get_ds();
1950 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
1952 sema_init(&filter->fo_init_lock, 1);
1953 filter->fo_committed_group = 0;
1955 rc = filter_prep(obd);
1959 filter->fo_destroys_in_progress = 0;
1960 for (i = 0; i < 32; i++)
1961 sema_init(&filter->fo_create_locks[i], 1);
1963 spin_lock_init(&filter->fo_translock);
1964 spin_lock_init(&filter->fo_objidlock);
1965 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
1966 sema_init(&filter->fo_alloc_lock, 1);
1967 init_brw_stats(&filter->fo_filter_stats);
1968 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1969 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
1970 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
1972 INIT_LIST_HEAD(&filter->fo_llog_list);
1973 spin_lock_init(&filter->fo_llog_list_lock);
1975 filter->fo_sptlrpc_lock = RW_LOCK_UNLOCKED;
1976 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
1978 filter->fo_fl_oss_capa = 0;
1979 INIT_LIST_HEAD(&filter->fo_capa_keys);
1980 filter->fo_capa_hash = init_capa_hash();
1981 if (filter->fo_capa_hash == NULL)
1982 GOTO(err_ops, rc = -ENOMEM);
1984 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
1985 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER,
1986 LDLM_NAMESPACE_GREEDY);
1987 if (obd->obd_namespace == NULL)
1988 GOTO(err_post, rc = -ENOMEM);
1989 obd->obd_namespace->ns_lvbp = obd;
1990 obd->obd_namespace->ns_lvbo = &filter_lvbo;
1991 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
1993 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1994 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1996 rc = llog_cat_initialize(obd, &obd->obd_olg, 1, NULL);
1998 CERROR("failed to setup llogging subsystems\n");
2002 rc = lquota_setup(filter_quota_interface_ref, obd);
2006 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2007 if (uuid_ptr != NULL) {
2008 class_uuid_unparse(uuid_ptr, &uuid);
2014 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2016 if (obd->obd_recovering) {
2017 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2018 "recovery until %d %s reconnect, or if no clients"
2019 " reconnect for %d:%.02d; during that time new "
2020 "clients will not be allowed to connect. "
2021 "Recovery progress can be monitored by watching "
2022 "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
2023 obd->obd_name, lustre_cfg_string(lcfg, 1),
2024 label ?: "", label ? "/" : "", str,
2025 obd->obd_max_recoverable_clients,
2026 (obd->obd_max_recoverable_clients == 1)
2027 ? "client" : "clients",
2028 (int)(OBD_RECOVERY_TIMEOUT) / 60,
2029 (int)(OBD_RECOVERY_TIMEOUT) % 60,
2032 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2033 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2034 label ?: "", label ? "/" : "", str,
2035 obd->obd_replayable ? "enabled" : "disabled");
2043 fsfilt_put_ops(obd->obd_fsops);
2044 filter_iobuf_pool_done(filter);
2046 server_put_mount(obd->obd_name, mnt);
2047 obd->u.obt.obt_sb = 0;
2051 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2053 struct lprocfs_static_vars lvars;
2058 CLASSERT(offsetof(struct obd_device, u.obt) ==
2059 offsetof(struct obd_device, u.filter.fo_obt));
2061 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2064 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2065 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2068 addr = (unsigned long)cfs_page_address(page);
2069 clear_page((void *)addr);
2071 /* lprocfs must be setup before the filter so state can be safely added
2072 * to /proc incrementally as the filter is setup */
2073 lprocfs_filter_init_vars(&lvars);
2074 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2075 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2076 /* Init obdfilter private stats here */
2077 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2078 LPROCFS_CNTR_AVGMINMAX,
2079 "read_bytes", "bytes");
2080 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2081 LPROCFS_CNTR_AVGMINMAX,
2082 "write_bytes", "bytes");
2084 lproc_filter_attach_seqstat(obd);
2085 obd->obd_proc_exports = proc_mkdir("exports",
2086 obd->obd_proc_entry);
2089 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2090 LUSTRE_CFG_BUFLEN(lcfg, 4));
2091 rc = filter_common_setup(obd, lcfg, (void *)addr);
2092 OBD_PAGE_FREE(page);
2095 lprocfs_obd_cleanup(obd);
2096 lprocfs_free_obd_stats(obd);
2102 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
2103 static struct llog_operations filter_size_orig_logops = {
2104 lop_setup: llog_obd_origin_setup,
2105 lop_cleanup: llog_obd_origin_cleanup,
2106 lop_add: llog_obd_origin_add
2109 static int filter_llog_init(struct obd_device *obd, int group,
2110 struct obd_device *tgt, int count,
2111 struct llog_catid *catid,
2112 struct obd_uuid *uuid)
2114 struct filter_obd *filter = &obd->u.filter;
2115 struct obd_llog_group *olg;
2116 struct llog_ctxt *ctxt;
2120 olg = filter_find_olg(obd, group);
2122 RETURN(PTR_ERR(olg));
2124 if (group == OBD_LLOG_GROUP) {
2125 LASSERT(filter->fo_lcm == NULL);
2126 OBD_ALLOC(filter->fo_lcm, sizeof(struct llog_commit_master));
2127 if (!filter->fo_lcm)
2130 rc = llog_init_commit_master((struct llog_commit_master *)
2135 filter_mds_ost_repl_logops = llog_client_ops;
2136 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2137 filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
2138 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2140 LASSERT(filter->fo_lcm != NULL);
2142 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2143 &filter_mds_ost_repl_logops);
2147 /* FIXME - assign unlink_cb for filter's recovery */
2149 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2151 LASSERT(ctxt != NULL);
2152 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2153 ctxt->loc_lcm = obd->u.filter.fo_lcm;
2154 rc = llog_start_commit_thread(ctxt->loc_lcm);
2155 llog_ctxt_put(ctxt);
2159 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2160 &filter_size_orig_logops);
2164 llog_cleanup_commit_master(filter->fo_lcm, 0);
2165 OBD_FREE(filter->fo_lcm, sizeof(struct llog_commit_master));
2166 filter->fo_lcm = NULL;
2171 static int filter_group_llog_finish(struct obd_llog_group *olg)
2173 struct llog_ctxt *ctxt;
2174 int rc = 0, rc2 = 0;
2177 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2179 rc = llog_cleanup(ctxt);
2181 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2183 rc2 = llog_cleanup(ctxt);
2190 static int filter_llog_finish(struct obd_device *obd, int count)
2195 if (obd->u.filter.fo_lcm) {
2196 llog_cleanup_commit_master((struct llog_commit_master *)
2197 obd->u.filter.fo_lcm, 0);
2198 OBD_FREE(obd->u.filter.fo_lcm,
2199 sizeof(struct llog_commit_master));
2200 obd->u.filter.fo_lcm = NULL;
2202 /* finish obd llog group */
2203 rc = filter_group_llog_finish(&obd->obd_olg);
2208 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2210 struct obd_llog_group *olg, *nolg;
2211 struct filter_obd *filter;
2214 filter = &obd->u.filter;
2216 if (group == OBD_LLOG_GROUP)
2217 RETURN(&obd->obd_olg);
2219 spin_lock(&filter->fo_llog_list_lock);
2220 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2221 if (olg->olg_group == group) {
2222 spin_unlock(&filter->fo_llog_list_lock);
2226 spin_unlock(&filter->fo_llog_list_lock);
2230 RETURN(ERR_PTR(-ENOMEM));
2232 llog_group_init(olg, group);
2233 spin_lock(&filter->fo_llog_list_lock);
2234 list_for_each_entry(nolg, &filter->fo_llog_list, olg_list) {
2235 LASSERT(nolg->olg_group != group);
2237 list_add(&olg->olg_list, &filter->fo_llog_list);
2238 spin_unlock(&filter->fo_llog_list_lock);
2240 rc = llog_cat_initialize(obd, olg, 1, NULL);
2242 spin_lock(&filter->fo_llog_list_lock);
2243 list_del(&olg->olg_list);
2244 spin_unlock(&filter->fo_llog_list_lock);
2246 RETURN(ERR_PTR(rc));
2248 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2249 obd->obd_name, group, olg);
2254 static int filter_llog_connect(struct obd_export *exp,
2255 struct llogd_conn_body *body)
2257 struct obd_device *obd = exp->exp_obd;
2258 struct llog_ctxt *ctxt;
2259 struct obd_llog_group *olg;
2263 CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2264 (unsigned) body->lgdc_logid.lgl_ogr,
2265 (unsigned) body->lgdc_logid.lgl_oid,
2266 (unsigned) body->lgdc_logid.lgl_ogen);
2268 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2270 RETURN(PTR_ERR(olg));
2271 llog_group_set_export(olg, exp);
2273 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2274 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2275 body->lgdc_ctxt_idx);
2276 rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2277 &body->lgdc_gen, NULL);
2278 llog_ctxt_put(ctxt);
2280 CERROR("failed to connect rc %d idx %d\n", rc,
2281 body->lgdc_ctxt_idx);
2286 static int filter_llog_preclean (struct obd_device *obd)
2288 struct obd_llog_group *olg;
2289 struct filter_obd *filter;
2293 rc = obd_llog_finish(obd, 0);
2295 CERROR("failed to cleanup llogging subsystem\n");
2297 filter = &obd->u.filter;
2298 spin_lock(&filter->fo_llog_list_lock);
2299 while (!list_empty(&filter->fo_llog_list)) {
2300 olg = list_entry(filter->fo_llog_list.next,
2301 struct obd_llog_group, olg_list);
2302 list_del(&olg->olg_list);
2303 spin_unlock(&filter->fo_llog_list_lock);
2305 rc = filter_group_llog_finish(olg);
2307 CERROR("failed to cleanup llogging subsystem for %u\n",
2310 spin_lock(&filter->fo_llog_list_lock);
2312 spin_unlock(&filter->fo_llog_list_lock);
2317 static int filter_precleanup(struct obd_device *obd,
2318 enum obd_cleanup_stage stage)
2324 case OBD_CLEANUP_EARLY:
2326 case OBD_CLEANUP_EXPORTS:
2327 target_cleanup_recovery(obd);
2328 rc = filter_llog_preclean(obd);
2330 case OBD_CLEANUP_SELF_EXP:
2332 case OBD_CLEANUP_OBD:
2338 static int filter_cleanup(struct obd_device *obd)
2340 struct filter_obd *filter = &obd->u.filter;
2344 LCONSOLE_WARN("%s: shutting down for failover; client state "
2345 "will be preserved.\n", obd->obd_name);
2347 if (!list_empty(&obd->obd_exports)) {
2348 CERROR("%s: still has clients!\n", obd->obd_name);
2349 class_disconnect_exports(obd);
2350 if (!list_empty(&obd->obd_exports)) {
2351 CERROR("still has exports after forced cleanup?\n");
2356 lprocfs_obd_cleanup(obd);
2357 lprocfs_free_obd_stats(obd);
2358 lquota_cleanup(filter_quota_interface_ref, obd);
2360 /* Stop recovery before namespace cleanup. */
2361 target_stop_recovery_thread(obd);
2362 target_cleanup_recovery(obd);
2364 ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2366 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2368 if (obd->u.obt.obt_sb == NULL)
2373 shrink_dcache_parent(obd->u.obt.obt_sb->s_root);
2375 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2377 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2378 obd->u.obt.obt_sb = NULL;
2380 fsfilt_put_ops(obd->obd_fsops);
2382 filter_iobuf_pool_done(filter);
2384 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2389 static int filter_connect_internal(struct obd_export *exp,
2390 struct obd_connect_data *data)
2395 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2396 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2397 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2398 data->ocd_connect_flags, data->ocd_version,
2399 data->ocd_grant, data->ocd_index);
2401 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2402 exp->exp_connect_flags = data->ocd_connect_flags;
2403 data->ocd_version = LUSTRE_VERSION_CODE;
2405 if ((exp->exp_connect_flags & OBD_CONNECT_FID) == 0) {
2406 CWARN("%s: OST requires FID support (flag="LPX64
2407 "), but client not\n",
2408 exp->exp_obd->obd_name,
2409 exp->exp_connect_flags);
2413 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2414 struct filter_export_data *fed = &exp->exp_filter_data;
2415 obd_size left, want;
2417 spin_lock(&exp->exp_obd->obd_osfs_lock);
2418 left = filter_grant_space_left(exp);
2419 want = data->ocd_grant;
2420 filter_grant(exp, fed->fed_grant, want, left);
2421 data->ocd_grant = fed->fed_grant;
2422 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2424 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2425 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2426 exp->exp_client_uuid.uuid, exp,
2427 data->ocd_grant, want, left);
2430 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2431 struct filter_obd *filter = &exp->exp_obd->u.filter;
2432 struct lr_server_data *lsd = filter->fo_fsd;
2433 int index = le32_to_cpu(lsd->lsd_ost_index);
2435 if (!(lsd->lsd_feature_compat &
2436 cpu_to_le32(OBD_COMPAT_OST))) {
2437 /* this will only happen on the first connect */
2438 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2439 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2440 filter_update_server_data(exp->exp_obd,
2441 filter->fo_rcvd_filp, lsd, 1);
2442 } else if (index != data->ocd_index) {
2443 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2444 " %u doesn't match actual OST index"
2445 " %u in last_rcvd file, bad "
2447 obd_export_nid2str(exp), index,
2453 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2454 data->ocd_brw_size = 65536;
2455 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2456 data->ocd_brw_size = min(data->ocd_brw_size,
2457 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2458 LASSERT(data->ocd_brw_size);
2461 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2462 __u32 cksum_types = data->ocd_cksum_types;
2464 /* The client set in ocd_cksum_types the checksum types it
2465 * supports. We have to mask off the algorithms that we don't
2467 if (cksum_types & OBD_CKSUM_ALL)
2468 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2470 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2472 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2473 "%x\n", exp->exp_obd->obd_name,
2474 obd_export_nid2str(exp), cksum_types,
2475 data->ocd_cksum_types);
2477 /* This client does not support OBD_CONNECT_CKSUM
2478 * fall back to CRC32 */
2479 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2480 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2481 exp->exp_obd->obd_name,
2482 obd_export_nid2str(exp));
2485 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2486 * FIXME: We don't strictly need the COMPAT flag for that,
2487 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2488 * FIXME: We needed it for the index, as index 0 is valid. */
2493 static int filter_reconnect(const struct lu_env *env,
2494 struct obd_export *exp, struct obd_device *obd,
2495 struct obd_uuid *cluuid,
2496 struct obd_connect_data *data)
2501 if (exp == NULL || obd == NULL || cluuid == NULL)
2504 rc = filter_connect_internal(exp, data);
2509 /* nearly identical to mds_connect */
2510 static int filter_connect(const struct lu_env *env,
2511 struct lustre_handle *conn, struct obd_device *obd,
2512 struct obd_uuid *cluuid,
2513 struct obd_connect_data *data)
2515 struct lvfs_run_ctxt saved;
2516 struct obd_export *exp;
2517 struct filter_export_data *fed;
2518 struct filter_client_data *fcd = NULL;
2523 if (conn == NULL || obd == NULL || cluuid == NULL)
2526 rc = class_connect(conn, obd, cluuid);
2529 exp = class_conn2export(conn);
2530 LASSERT(exp != NULL);
2532 fed = &exp->exp_filter_data;
2534 rc = filter_connect_internal(exp, data);
2538 filter_export_stats_init(obd, exp);
2539 group = data->ocd_group;
2540 if (obd->obd_replayable) {
2541 OBD_ALLOC(fcd, sizeof(*fcd));
2543 CERROR("filter: out of memory for client data\n");
2544 GOTO(cleanup, rc = -ENOMEM);
2547 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
2549 fed->fed_fcd->fcd_group = group;
2550 rc = filter_client_add(obd, exp, -1);
2554 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2555 obd->obd_name, exp->exp_handle.h_cookie, group);
2559 if (fed->fed_group != 0 && fed->fed_group != group) {
2560 CERROR("!!! This export (nid %s) used object group %d "
2561 "earlier; now it's trying to use group %d! This could "
2562 "be a bug in the MDS. Tell CFS.\n",
2563 obd_export_nid2str(exp), fed->fed_group, group);
2564 GOTO(cleanup, rc = -EPROTO);
2566 fed->fed_group = group;
2568 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2569 rc = filter_read_groups(obd, group, 1);
2570 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2572 CERROR("can't read group %u\n", group);
2581 OBD_FREE(fcd, sizeof(*fcd));
2582 fed->fed_fcd = NULL;
2584 class_disconnect(exp);
2586 class_export_put(exp);
2592 /* Do extra sanity checks for grant accounting. We do this at connect,
2593 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2594 * always get rid of it or turn it off when we know accounting is good. */
2595 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2597 struct filter_export_data *fed;
2598 struct obd_export *exp;
2599 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2600 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2601 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2603 if (list_empty(&obd->obd_exports))
2606 /* We don't want to do this for large machines that do lots of
2607 mounts or unmounts. It burns... */
2608 if (obd->obd_num_exports > 100)
2611 spin_lock(&obd->obd_osfs_lock);
2612 spin_lock(&obd->obd_dev_lock);
2613 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2615 fed = &exp->exp_filter_data;
2616 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2619 if (maxsize > 0) { /* we may not have done a statfs yet */
2620 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2621 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2622 exp->exp_client_uuid.uuid, exp,
2623 fed->fed_grant, fed->fed_pending, maxsize);
2624 LASSERTF(fed->fed_dirty <= maxsize,
2625 "%s: cli %s/%p %ld > "LPU64"\n", func,
2626 exp->exp_client_uuid.uuid, exp,
2627 fed->fed_dirty, maxsize);
2630 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2631 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2632 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2634 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2635 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2636 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2637 tot_granted += fed->fed_grant + fed->fed_pending;
2638 tot_pending += fed->fed_pending;
2639 tot_dirty += fed->fed_dirty;
2641 fo_tot_granted = obd->u.filter.fo_tot_granted;
2642 fo_tot_pending = obd->u.filter.fo_tot_pending;
2643 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2644 spin_unlock(&obd->obd_dev_lock);
2645 spin_unlock(&obd->obd_osfs_lock);
2647 /* Do these assertions outside the spinlocks so we don't kill system */
2648 if (tot_granted != fo_tot_granted)
2649 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2650 func, tot_granted, fo_tot_granted);
2651 if (tot_pending != fo_tot_pending)
2652 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2653 func, tot_pending, fo_tot_pending);
2654 if (tot_dirty != fo_tot_dirty)
2655 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2656 func, tot_dirty, fo_tot_dirty);
2657 if (tot_pending > tot_granted)
2658 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2659 func, tot_pending, tot_granted);
2660 if (tot_granted > maxsize)
2661 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2662 func, tot_granted, maxsize);
2663 if (tot_dirty > maxsize)
2664 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2665 func, tot_dirty, maxsize);
2668 /* Remove this client from the grant accounting totals. We also remove
2669 * the export from the obd device under the osfs and dev locks to ensure
2670 * that the filter_grant_sanity_check() calculations are always valid.
2671 * The client should do something similar when it invalidates its import. */
2672 static void filter_grant_discard(struct obd_export *exp)
2674 struct obd_device *obd = exp->exp_obd;
2675 struct filter_obd *filter = &obd->u.filter;
2676 struct filter_export_data *fed = &exp->exp_filter_data;
2678 spin_lock(&obd->obd_osfs_lock);
2679 spin_lock(&obd->obd_dev_lock);
2680 list_del_init(&exp->exp_obd_chain);
2681 spin_unlock(&obd->obd_dev_lock);
2683 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2684 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2685 obd->obd_name, filter->fo_tot_granted,
2686 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2687 filter->fo_tot_granted -= fed->fed_grant;
2688 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2689 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2690 obd->obd_name, filter->fo_tot_pending,
2691 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2692 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2693 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2694 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2695 obd->obd_name, filter->fo_tot_dirty,
2696 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2697 filter->fo_tot_dirty -= fed->fed_dirty;
2701 spin_unlock(&obd->obd_osfs_lock);
2704 static int filter_destroy_export(struct obd_export *exp)
2708 if (exp->exp_filter_data.fed_pending)
2709 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2710 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2711 exp, exp->exp_filter_data.fed_pending);
2713 /* Not ported yet the b1_6 quota functionality
2714 * lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2717 target_destroy_export(exp);
2719 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2722 lprocfs_exp_cleanup(exp);
2724 if (exp->exp_obd->obd_replayable)
2725 filter_client_free(exp);
2727 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2729 filter_grant_discard(exp);
2730 filter_fmd_cleanup(exp);
2732 if (!(exp->exp_flags & OBD_OPT_FORCE))
2733 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2738 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2740 struct obd_llog_group *olg_min, *olg;
2741 struct filter_obd *filter;
2742 int worked = 0, group;
2743 struct llog_ctxt *ctxt;
2746 filter = &obd->u.filter;
2748 /* we can't sync log holding spinlock. also, we do not want to get
2749 * into livelock. so we do following: loop over MDS's exports in
2750 * group order and skip already synced llogs -bzzz */
2752 /* look for group with min. number, but > worked */
2755 spin_lock(&filter->fo_llog_list_lock);
2756 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2757 if (olg->olg_group <= worked) {
2758 /* this group is already synced */
2761 if (group < olg->olg_group) {
2762 /* we have group with smaller number to sync */
2765 /* store current minimal group */
2767 group = olg->olg_group;
2769 spin_unlock(&filter->fo_llog_list_lock);
2771 if (olg_min == NULL)
2774 worked = olg_min->olg_group;
2775 if (olg_min->olg_exp &&
2776 (dexp == olg_min->olg_exp || dexp == NULL)) {
2778 ctxt = llog_group_get_ctxt(olg_min,
2779 LLOG_MDS_OST_REPL_CTXT);
2780 LASSERT(ctxt != NULL);
2781 err = llog_sync(ctxt, olg_min->olg_exp);
2782 llog_ctxt_put(ctxt);
2784 CERROR("error flushing logs to MDS: rc %d\n",
2787 } while (olg_min != NULL);
2790 /* also incredibly similar to mds_disconnect */
2791 static int filter_disconnect(struct obd_export *exp)
2793 struct obd_device *obd = exp->exp_obd;
2798 class_export_get(exp);
2800 if (!(exp->exp_flags & OBD_OPT_FORCE))
2801 filter_grant_sanity_check(obd, __FUNCTION__);
2802 filter_grant_discard(exp);
2804 /* Disconnect early so that clients can't keep using export */
2805 rc = class_disconnect(exp);
2806 if (exp->exp_obd->obd_namespace != NULL)
2807 ldlm_cancel_locks_for_export(exp);
2809 fsfilt_sync(obd, obd->u.obt.obt_sb);
2811 /* flush any remaining cancel messages out to the target */
2812 filter_sync_llogs(obd, exp);
2813 class_export_put(exp);
2817 /* reverse import is changed, sync all cancels */
2818 static void filter_revimp_update(struct obd_export *exp)
2823 class_export_get(exp);
2825 /* flush any remaining cancel messages out to the target */
2826 filter_sync_llogs(exp->exp_obd, exp);
2827 class_export_put(exp);
2831 static int filter_ping(struct obd_export *exp)
2833 filter_fmd_expire(exp);
2838 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2839 const char *what, int quiet)
2841 struct dentry *dchild = NULL;
2844 if (oa->o_valid & OBD_MD_FLGROUP)
2847 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
2849 if (IS_ERR(dchild)) {
2850 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
2851 what, group, oa->o_id);
2855 if (dchild->d_inode == NULL) {
2857 CERROR("%s: %s on non-existent object: "LPU64"\n",
2858 obd->obd_name, what, oa->o_id);
2860 RETURN(ERR_PTR(-ENOENT));
2866 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2868 struct dentry *dentry = NULL;
2869 struct obd_device *obd;
2873 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
2874 oinfo_capa(oinfo), CAPA_OPC_META_READ);
2878 obd = class_exp2obd(exp);
2880 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
2884 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
2886 RETURN(PTR_ERR(dentry));
2888 /* Limit the valid bits in the return data to what we actually use */
2889 oinfo->oi_oa->o_valid = OBD_MD_FLID;
2890 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
2896 /* this should be enabled/disabled in condition to enabled/disabled large
2897 * inodes (fast EAs) in backing store FS. */
2898 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
2899 void *handle, struct obdo *oa)
2901 struct obd_device *obd = exp->exp_obd;
2905 if (oa->o_valid & OBD_MD_FLFID) {
2906 struct filter_fid ff;
2908 if (!(oa->o_valid & OBD_MD_FLGROUP))
2910 /* packing fid and converting it to LE for storing into EA.
2911 * Here ->o_stripe_idx should be filled by LOV and rest of
2912 * fields - by client. */
2913 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
2914 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
2915 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
2916 ff.ff_objid = cpu_to_le64(oa->o_id);
2917 ff.ff_group = cpu_to_le64(oa->o_gr);
2919 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
2920 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
2921 oa->o_generation, oa->o_id, oa->o_gr);
2923 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
2925 CERROR("store fid in object failed! rc: %d\n", rc);
2927 CDEBUG(D_HA, "OSS object without fid info!\n");
2933 /* this is called from filter_truncate() until we have filter_punch() */
2934 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
2935 struct obdo *oa, struct obd_trans_info *oti)
2937 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
2938 struct llog_cookie *fcc = NULL;
2939 struct filter_obd *filter;
2940 int rc, err, locked = 0, sync = 0;
2941 unsigned int ia_valid;
2942 struct inode *inode;
2947 LASSERT(dentry != NULL);
2948 LASSERT(!IS_ERR(dentry));
2950 inode = dentry->d_inode;
2951 LASSERT(inode != NULL);
2953 filter = &exp->exp_obd->u.filter;
2954 iattr_from_obdo(&iattr, oa, oa->o_valid);
2955 ia_valid = iattr.ia_valid;
2957 if (oa->o_valid & OBD_MD_FLCOOKIE) {
2958 OBD_ALLOC(fcc, sizeof(*fcc));
2960 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
2963 if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
2965 LOCK_INODE_MUTEX(inode);
2969 /* If the inode still has SUID+SGID bits set (see filter_precreate())
2970 * then we will accept the UID+GID sent by the client during write for
2971 * initializing the ownership of this inode. We only allow this to
2972 * happen once so clear these bits in setattr. In 2.6 kernels it is
2973 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
2974 * the flags that are actually being set. */
2975 if (ia_valid & (ATTR_UID | ATTR_GID)) {
2976 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
2977 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
2979 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
2980 if (!(ia_valid & ATTR_MODE)) {
2981 iattr.ia_mode = inode->i_mode;
2982 iattr.ia_valid |= ATTR_MODE;
2984 iattr.ia_mode &= ~S_ISUID;
2986 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
2987 if (!(iattr.ia_valid & ATTR_MODE)) {
2988 iattr.ia_mode = inode->i_mode;
2989 iattr.ia_valid |= ATTR_MODE;
2991 iattr.ia_mode &= ~S_ISGID;
2994 orig_ids[USRQUOTA] = inode->i_uid;
2995 orig_ids[GRPQUOTA] = inode->i_gid;
2996 handle = fsfilt_start_log(exp->exp_obd, inode,
2997 FSFILT_OP_SETATTR, oti, 1);
2999 GOTO(out_unlock, rc = PTR_ERR(handle));
3001 /* update inode EA only once when inode is suid bit marked. As
3002 * on 2.6.x UID and GID may be set separately, we check here
3003 * only one of them to avoid double setting. */
3004 if (inode->i_mode & S_ISUID)
3005 filter_update_fidea(exp, inode, handle, oa);
3007 handle = fsfilt_start(exp->exp_obd, inode,
3008 FSFILT_OP_SETATTR, oti);
3010 GOTO(out_unlock, rc = PTR_ERR(handle));
3012 if (oa->o_valid & OBD_MD_FLFLAGS) {
3013 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3014 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3016 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3018 /* set cancel cookie callback function */
3019 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3020 filter_cancel_cookies_cb,
3024 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3025 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3027 /* The truncate might have used up our transaction credits. Make
3028 * sure we have one left for the last_rcvd update. */
3029 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3031 rc = filter_finish_transno(exp, oti, rc, sync);
3033 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3037 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3039 CERROR("error on commit, err = %d\n", err);
3047 /* Let's flush truncated page on disk immediately, then we can
3048 * avoid need to search for page aliases before directio writes
3049 * and this sort of stuff at expense of somewhat slower
3050 * truncates not on a page boundary. I believe this is the only
3051 * place in filter code that can lead to pages getting to
3052 * pagecache so far. */
3053 filter_clear_truncated_page(inode);
3054 UNLOCK_INODE_MUTEX(inode);
3061 UNLOCK_INODE_MUTEX(inode);
3064 OBD_FREE(fcc, sizeof(*fcc));
3066 /* trigger quota release */
3067 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3068 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3069 int rc2 = lquota_adjust(filter_quota_interface_ref,
3070 exp->exp_obd, cur_ids,
3071 orig_ids, rc, FSFILT_OP_SETATTR);
3072 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3073 "filter adjust qunit. (rc:%d)\n", rc2);
3078 /* this is called from filter_truncate() until we have filter_punch() */
3079 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3080 struct obd_trans_info *oti)
3082 struct ldlm_res_id res_id = { .name = { oinfo->oi_oa->o_id, 0,
3083 oinfo->oi_oa->o_gr, 0 } };
3084 struct filter_mod_data *fmd;
3085 struct lvfs_run_ctxt saved;
3086 struct filter_obd *filter;
3087 struct ldlm_resource *res;
3088 struct dentry *dentry;
3092 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3093 oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
3097 dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
3100 RETURN(PTR_ERR(dentry));
3102 filter = &exp->exp_obd->u.filter;
3103 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3106 if (oinfo->oi_oa->o_valid &
3107 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3108 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
3109 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3110 fmd->fmd_mactime_xid = oti->oti_xid;
3111 filter_fmd_put(exp, fmd);
3114 /* setting objects attributes (including owner/group) */
3115 rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
3117 GOTO(out_unlock, rc);
3119 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3120 &res_id, LDLM_EXTENT, 0);
3123 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3124 ldlm_resource_putref(res);
3127 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3129 /* Quota release need uid/gid info */
3130 obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
3131 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3137 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3141 /* XXX identical to osc_unpackmd */
3142 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3143 struct lov_mds_md *lmm, int lmm_bytes)
3149 if (lmm_bytes < sizeof (*lmm)) {
3150 CERROR("lov_mds_md too small: %d, need %d\n",
3151 lmm_bytes, (int)sizeof(*lmm));
3154 /* XXX LOV_MAGIC etc check? */
3156 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3157 CERROR("lov_mds_md: zero lmm_object_id\n");
3162 lsm_size = lov_stripe_md_size(1);
3166 if (*lsmp != NULL && lmm == NULL) {
3167 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3168 OBD_FREE(*lsmp, lsm_size);
3173 if (*lsmp == NULL) {
3174 OBD_ALLOC(*lsmp, lsm_size);
3178 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3179 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3180 OBD_FREE(*lsmp, lsm_size);
3183 loi_init((*lsmp)->lsm_oinfo[0]);
3187 /* XXX zero *lsmp? */
3188 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3189 LASSERT((*lsmp)->lsm_object_id);
3192 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3197 /* caller must hold fo_create_locks[oa->o_gr] */
3198 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3199 struct filter_obd *filter)
3201 struct obdo doa; /* XXX obdo on stack */
3207 LASSERT(oa->o_gr != 0);
3208 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3209 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3211 memset(&doa, 0, sizeof(doa));
3213 doa.o_valid |= OBD_MD_FLGROUP;
3214 doa.o_gr = oa->o_gr;
3215 doa.o_mode = S_IFREG;
3217 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3218 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3219 exp->exp_obd->obd_name, doa.o_gr);
3223 last = filter_last_id(filter, doa.o_gr);
3225 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3226 exp->exp_obd->obd_name, oa->o_id + 1, last);
3228 for (id = last; id > oa->o_id; id--) {
3230 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
3231 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3232 CEMERG("error destroying precreate objid "LPU64": %d\n",
3234 filter_set_last_id(filter, id - 1, doa.o_gr);
3235 /* update last_id on disk periodically so that if we restart
3236 * we don't need to re-scan all of the just-deleted objects. */
3237 if ((id & 511) == 0)
3238 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3241 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3242 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3244 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3245 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3250 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3251 obd_gr group, int *num);
3252 /* returns a negative error or a nonnegative number of files to create */
3253 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3254 obd_gr group, struct obd_trans_info *oti)
3256 struct obd_device *obd = exp->exp_obd;
3257 struct filter_obd *filter = &obd->u.filter;
3261 /* delete orphans request */
3262 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3263 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3264 CERROR("%s: dropping old orphan cleanup request\n",
3268 /* This causes inflight precreates to abort and drop lock */
3269 set_bit(group, &filter->fo_destroys_in_progress);
3270 down(&filter->fo_create_locks[group]);
3271 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3272 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3273 exp->exp_obd->obd_name, group);
3274 up(&filter->fo_create_locks[group]);
3277 diff = oa->o_id - filter_last_id(filter, group);
3278 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3279 filter_last_id(filter, group), diff);
3281 if (-diff > OST_MAX_PRECREATE) {
3282 CERROR("%s: ignoring bogus orphan destroy request: "
3283 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3284 oa->o_id, filter_last_id(filter, group));
3285 /* FIXME: should reset precreate_next_id on MDS */
3286 GOTO(out, rc = -EINVAL);
3289 rc = filter_destroy_precreated(exp, oa, filter);
3291 CERROR("%s: unable to write lastobjid, but "
3292 "orphans were deleted\n", obd->obd_name);
3295 /* XXX: Used by MDS for the first time! */
3296 clear_bit(group, &filter->fo_destroys_in_progress);
3299 down(&filter->fo_create_locks[group]);
3300 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3301 CERROR("%s: dropping old precreate request\n",
3305 /* only precreate if group == 0 and o_id is specfied */
3306 if (group < FILTER_GROUP_MDS0 || oa->o_id == 0)
3309 diff = oa->o_id - filter_last_id(filter, group);
3310 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3311 filter_last_id(filter, group), diff);
3313 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3314 oa->o_id, filter_last_id(filter, group), diff);
3318 oa->o_id = filter_last_id(&obd->u.filter, group);
3319 rc = filter_precreate(obd, oa, group, &diff);
3320 oa->o_id = filter_last_id(&obd->u.filter, group);
3322 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3325 /* else diff == 0 */
3328 up(&filter->fo_create_locks[group]);
3332 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3335 struct filter_obd *filter = &obd->u.filter;
3336 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3340 /* at least try to account for cached pages. its still racey and
3341 * might be under-reporting if clients haven't announced their
3342 * caches with brw recently */
3343 spin_lock(&obd->obd_osfs_lock);
3344 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3345 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3346 spin_unlock(&obd->obd_osfs_lock);
3348 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3349 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3350 filter->fo_tot_dirty, filter->fo_tot_granted,
3351 filter->fo_tot_pending,
3352 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3354 filter_grant_sanity_check(obd, __FUNCTION__);
3356 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3357 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3358 osfs->os_bsize - 1) >> blockbits));
3360 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3361 * stop creating files on MDS if OST is not good shape to create
3363 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3368 static int filter_use_existing_obj(struct obd_device *obd,
3369 struct dentry *dchild, void **handle,
3372 struct inode *inode = dchild->d_inode;
3376 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3379 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3380 if (IS_ERR(*handle))
3381 return PTR_ERR(*handle);
3383 iattr.ia_valid = ATTR_MODE;
3384 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3385 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3393 /* We rely on the fact that only one thread will be creating files in a given
3394 * group at a time, which is why we don't need an atomic filter_get_new_id.
3395 * Even if we had that atomic function, the following race would exist:
3397 * thread 1: gets id x from filter_next_id
3398 * thread 2: gets id (x + 1) from filter_next_id
3399 * thread 2: creates object (x + 1)
3400 * thread 1: tries to create object x, gets -ENOSPC
3402 * Caller must hold fo_create_locks[group]
3404 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3405 obd_gr group, int *num)
3407 struct dentry *dchild = NULL, *dparent = NULL;
3408 struct filter_obd *filter;
3409 struct obd_statfs *osfs;
3410 int err = 0, rc = 0, recreate_obj = 0, i;
3411 unsigned long enough_time = jiffies + min(obd_timeout * HZ / 4, 10U*HZ);
3413 void *handle = NULL;
3416 filter = &obd->u.filter;
3418 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3420 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3422 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3423 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3426 OBD_ALLOC(osfs, sizeof(*osfs));
3429 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ);
3430 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3431 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3432 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3433 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3437 OBD_FREE(osfs, sizeof(*osfs));
3442 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3443 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3445 for (i = 0; i < *num && err == 0; i++) {
3446 int cleanup_phase = 0;
3448 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3449 CWARN("%s: create aborted by destroy\n",
3458 last_id = filter_last_id(filter, group);
3459 if (next_id > last_id) {
3460 CERROR("Error: Trying to recreate obj greater"
3461 "than last id "LPD64" > "LPD64"\n",
3463 GOTO(cleanup, rc = -EINVAL);
3466 next_id = filter_last_id(filter, group) + 1;
3468 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
3470 dparent = filter_parent_lock(obd, group, next_id);
3471 if (IS_ERR(dparent))
3472 GOTO(cleanup, rc = PTR_ERR(dparent));
3473 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3475 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3477 GOTO(cleanup, rc = PTR_ERR(dchild));
3478 cleanup_phase = 2; /* f_dput(dchild) */
3480 if (dchild->d_inode != NULL) {
3481 /* This would only happen if lastobjid was bad on disk*/
3482 /* Could also happen if recreating missing obj but it
3483 * already exists. */
3485 CERROR("%s: recreating existing object %.*s?\n",
3486 obd->obd_name, dchild->d_name.len,
3487 dchild->d_name.name);
3489 /* Use these existing objects if they are
3491 if (dchild->d_inode->i_size == 0) {
3492 rc = filter_use_existing_obj(obd,dchild,
3493 &handle, &cleanup_phase);
3500 CERROR("%s: Serious error: objid %.*s already "
3501 "exists; is this filesystem corrupt?\n",
3502 obd->obd_name, dchild->d_name.len,
3503 dchild->d_name.name);
3506 GOTO(cleanup, rc = -EEXIST);
3509 handle = fsfilt_start_log(obd, dparent->d_inode,
3510 FSFILT_OP_CREATE, NULL, 1);
3512 GOTO(cleanup, rc = PTR_ERR(handle));
3515 /* We mark object SUID+SGID to flag it for accepting UID+GID
3516 * from client on first write. Currently the permission bits
3517 * on the OST are never used, so this is OK. */
3518 rc = ll_vfs_create(dparent->d_inode, dchild,
3519 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3521 CERROR("create failed rc = %d\n", rc);
3526 if (!recreate_obj) {
3527 filter_set_last_id(filter, next_id, group);
3528 err = filter_update_last_objid(obd, group, 0);
3530 CERROR("unable to write lastobjid "
3531 "but file created\n");
3535 switch(cleanup_phase) {
3537 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3539 CERROR("error on commit, err = %d\n", err);
3546 filter_parent_unlock(dparent);
3553 if (time_after(jiffies, enough_time)) {
3555 "%s: precreate slow - want %d got %d \n",
3556 obd->obd_name, *num, i);
3563 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3564 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3569 static int filter_create(struct obd_export *exp, struct obdo *oa,
3570 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3572 struct filter_export_data *fed;
3573 struct obd_device *obd = NULL;
3574 struct filter_obd *filter;
3575 struct lvfs_run_ctxt saved;
3576 struct lov_stripe_md *lsm = NULL;
3577 int rc = 0, diff, group = oa->o_gr;
3580 if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
3581 CERROR("!!! nid %s sent invalid object group %d\n",
3582 obd_export_nid2str(exp), group);
3587 fed = &exp->exp_filter_data;
3588 filter = &obd->u.filter;
3590 if (fed->fed_group != group) {
3591 CERROR("!!! this export (nid %s) used object group %d "
3592 "earlier; now it's trying to use group %d! This could "
3593 "be a bug in the MDS. Tell CFS.\n",
3594 obd_export_nid2str(exp), fed->fed_group, group);
3598 CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
3599 oa->o_gr, oa->o_id);
3603 rc = obd_alloc_memmd(exp, &lsm);
3610 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3612 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3613 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3614 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3615 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3616 oa->o_id, filter_last_id(filter,
3621 down(&filter->fo_create_locks[oa->o_gr]);
3622 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3623 up(&filter->fo_create_locks[oa->o_gr]);
3626 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3629 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3630 if (rc && ea != NULL && *ea != lsm) {
3631 obd_free_memmd(exp, &lsm);
3632 } else if (rc == 0 && ea != NULL) {
3633 /* XXX LOV STACKING: the lsm that is passed to us from
3634 * LOV does not have valid lsm_oinfo data structs, so
3635 * don't go touching that. This needs to be fixed in a
3637 lsm->lsm_object_id = oa->o_id;
3644 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3645 struct lov_stripe_md *md, struct obd_trans_info *oti,
3646 struct obd_export *md_exp)
3648 unsigned int qcids[MAXQUOTAS] = {0, 0};
3649 struct obd_device *obd;
3650 struct filter_obd *filter;
3651 struct dentry *dchild = NULL, *dparent = NULL;
3652 struct lvfs_run_ctxt saved;
3653 void *handle = NULL;
3654 struct llog_cookie *fcc = NULL;
3655 int rc, rc2, cleanup_phase = 0, sync = 0;
3659 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3662 filter = &obd->u.filter;
3664 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3667 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3669 GOTO(cleanup, rc = PTR_ERR(dchild));
3672 if (dchild->d_inode == NULL) {
3673 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3675 /* If object already gone, cancel cookie right now */
3676 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3677 struct llog_ctxt *ctxt;
3678 struct obd_llog_group *olg;
3679 fcc = obdo_logcookie(oa);
3680 olg = filter_find_olg(obd, oa->o_gr);
3682 GOTO(cleanup, rc = PTR_ERR(olg));
3683 llog_group_set_export(olg, exp);
3685 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
3686 llog_cancel(ctxt, NULL, 1, fcc, 0);
3687 llog_ctxt_put(ctxt);
3688 fcc = NULL; /* we didn't allocate fcc, don't free it */
3690 GOTO(cleanup, rc = -ENOENT);
3693 filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3695 /* Our MDC connection is established by the MDS to us */
3696 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3697 OBD_ALLOC(fcc, sizeof(*fcc));
3699 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3701 DQUOT_INIT(dchild->d_inode);
3703 /* we're gonna truncate it first in order to avoid possible deadlock:
3705 * open trasaction open transaction
3706 * down(i_zombie) down(i_zombie)
3707 * restart transaction
3708 * (see BUG 4180) -bzzz
3710 LOCK_INODE_MUTEX(dchild->d_inode);
3711 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3713 if (IS_ERR(handle)) {
3714 UNLOCK_INODE_MUTEX(dchild->d_inode);
3715 GOTO(cleanup, rc = PTR_ERR(handle));
3718 iattr.ia_valid = ATTR_SIZE;
3720 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3721 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3722 UNLOCK_INODE_MUTEX(dchild->d_inode);
3726 GOTO(cleanup, rc = rc2);
3728 /* We don't actually need to lock the parent until we are unlinking
3729 * here, and not while truncating above. That avoids holding the
3730 * parent lock for a long time during truncate, which can block other
3731 * threads from doing anything to objects in that directory. bug 7171 */
3732 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3733 if (IS_ERR(dparent))
3734 GOTO(cleanup, rc = PTR_ERR(dparent));
3735 cleanup_phase = 3; /* filter_parent_unlock */
3737 LOCK_INODE_MUTEX(dchild->d_inode);
3738 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3739 if (IS_ERR(handle)) {
3740 UNLOCK_INODE_MUTEX(dchild->d_inode);
3741 GOTO(cleanup, rc = PTR_ERR(handle));
3743 cleanup_phase = 4; /* fsfilt_commit */
3745 /* Quota release need uid/gid of inode */
3746 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3748 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3750 /* this drops dchild->d_inode->i_mutex unconditionally */
3751 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3755 switch(cleanup_phase) {
3758 sync = fsfilt_add_journal_cb(obd, 0, oti ?
3759 oti->oti_handle : handle,
3760 filter_cancel_cookies_cb,
3762 /* If add_journal_cb failed, then filter_finish_transno
3763 * will commit the handle and we will do a sync
3764 * on commit. then we call callback directly to free
3767 rc = filter_finish_transno(exp, oti, rc, sync);
3769 filter_cancel_cookies_cb(obd, 0, fcc, rc);
3772 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3774 CERROR("error on commit, err = %d\n", rc2);
3781 filter_parent_unlock(dparent);
3785 OBD_FREE(fcc, sizeof(*fcc));
3787 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3790 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3794 /* trigger quota release */
3795 qcids[USRQUOTA] = oa->o_uid;
3796 qcids[GRPQUOTA] = oa->o_gid;
3797 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
3801 CDEBUG(D_QUOTA, "filter adjust qunit! (rc:%d)\n", rc2);
3805 /* NB start and end are used for punch, but not truncate */
3806 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3807 struct obd_trans_info *oti,
3808 struct ptlrpc_request_set *rqset)
3813 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3814 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
3815 oinfo->oi_policy.l_extent.end);
3819 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
3820 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
3821 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
3823 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3824 oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
3828 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
3829 rc = filter_setattr(exp, oinfo, oti);
3833 static int filter_sync(struct obd_export *exp, struct obdo *oa,
3834 struct lov_stripe_md *lsm, obd_off start, obd_off end,
3837 struct lvfs_run_ctxt saved;
3838 struct filter_obd *filter;
3839 struct dentry *dentry;
3843 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
3844 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
3848 filter = &exp->exp_obd->u.filter;
3850 /* an objid of zero is taken to mean "sync whole filesystem" */
3851 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
3852 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
3853 /* flush any remaining cancel messages out to the target */
3854 filter_sync_llogs(exp->exp_obd, exp);
3858 dentry = filter_oa2dentry(exp->exp_obd, oa);
3860 RETURN(PTR_ERR(dentry));
3862 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3864 LOCK_INODE_MUTEX(dentry->d_inode);
3866 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
3868 /* just any file to grab fsync method - "file" arg unused */
3869 struct file *file = filter->fo_rcvd_filp;
3871 if (file->f_op && file->f_op->fsync)
3872 rc = file->f_op->fsync(NULL, dentry, 1);
3874 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
3878 UNLOCK_INODE_MUTEX(dentry->d_inode);
3880 oa->o_valid = OBD_MD_FLID;
3881 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
3883 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3889 static int filter_get_info(struct obd_export *exp, __u32 keylen,
3890 void *key, __u32 *vallen, void *val)
3892 struct obd_device *obd;
3895 obd = class_exp2obd(exp);
3897 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3901 if (KEY_IS("blocksize")) {
3902 __u32 *blocksize = val;
3904 if (*vallen < sizeof(*blocksize))
3906 *blocksize = obd->u.obt.obt_sb->s_blocksize;
3908 *vallen = sizeof(*blocksize);
3912 if (KEY_IS("blocksize_bits")) {
3913 __u32 *blocksize_bits = val;
3914 if (blocksize_bits) {
3915 if (*vallen < sizeof(*blocksize_bits))
3917 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
3919 *vallen = sizeof(*blocksize_bits);
3923 if (KEY_IS("last_id")) {
3924 obd_id *last_id = val;
3925 /* FIXME: object groups */
3927 if (*vallen < sizeof(*last_id))
3929 *last_id = filter_last_id(&obd->u.filter,
3930 exp->exp_filter_data.fed_group);
3932 *vallen = sizeof(*last_id);
3936 CDEBUG(D_IOCTL, "invalid key\n");
3940 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
3941 void *key, __u32 vallen, void *val,
3942 struct ptlrpc_request_set *set)
3944 struct obd_device *obd;
3945 struct obd_llog_group *olg;
3946 struct llog_ctxt *ctxt;
3952 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
3956 if (KEY_IS(KEY_CAPA_KEY)) {
3957 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
3959 CERROR("filter update capability key failed: %d\n", rc);
3963 if (KEY_IS(KEY_REVIMP_UPD)) {
3964 filter_revimp_update(exp);
3968 if (!KEY_IS(KEY_MDS_CONN))
3971 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
3972 obd_export_nid2str(exp));
3973 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
3975 /* setup llog imports */
3976 LASSERT(val != NULL);
3977 group = (int)(*(__u32 *)val);
3978 LASSERT(group >= FILTER_GROUP_MDS0);
3980 olg = filter_find_olg(obd, group);
3982 RETURN(PTR_ERR(olg));
3983 llog_group_set_export(olg, exp);
3985 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
3986 LASSERTF(ctxt != NULL, "ctxt is null\n"),
3988 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
3989 llog_ctxt_put(ctxt);
3991 lquota_setinfo(filter_quota_interface_ref, exp, obd);
3996 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
3997 int len, void *karg, void *uarg)
3999 struct obd_device *obd = exp->exp_obd;
4000 struct obd_ioctl_data *data = karg;
4004 case OBD_IOC_ABORT_RECOVERY: {
4005 CERROR("aborting recovery for device %s\n", obd->obd_name);
4006 target_stop_recovery_thread(obd);
4010 case OBD_IOC_SYNC: {
4011 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4012 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4016 case OBD_IOC_SET_READONLY: {
4018 struct super_block *sb = obd->u.obt.obt_sb;
4019 struct inode *inode = sb->s_root->d_inode;
4020 BDEVNAME_DECLARE_STORAGE(tmp);
4021 CERROR("*** setting device %s read-only ***\n",
4022 ll_bdevname(sb, tmp));
4024 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4025 if (!IS_ERR(handle))
4026 rc = fsfilt_commit(obd, inode, handle, 1);
4028 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4029 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4031 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4035 case OBD_IOC_CATLOGLIST: {
4036 rc = llog_catalog_list(obd, 1, data);
4040 case OBD_IOC_LLOG_CANCEL:
4041 case OBD_IOC_LLOG_REMOVE:
4042 case OBD_IOC_LLOG_INFO:
4043 case OBD_IOC_LLOG_PRINT: {
4044 /* FIXME to be finished */
4045 RETURN(-EOPNOTSUPP);
4047 struct llog_ctxt *ctxt = NULL;
4049 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4050 rc = llog_ioctl(ctxt, cmd, data);
4051 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4064 static int filter_health_check(struct obd_device *obd)
4066 #ifdef USE_HEALTH_CHECK_WRITE
4067 struct filter_obd *filter = &obd->u.filter;
4072 * health_check to return 0 on healthy
4073 * and 1 on unhealthy.
4075 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4078 #ifdef USE_HEALTH_CHECK_WRITE
4079 LASSERT(filter->fo_health_check_filp != NULL);
4080 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4085 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4088 return filter_fid2dentry(data, NULL, gr, id);
4091 static int filter_process_config(struct obd_device *obd, obd_count len,
4094 struct lustre_cfg *lcfg = buf;
4095 struct lprocfs_static_vars lvars;
4098 switch (lcfg->lcfg_command) {
4099 case LCFG_SPTLRPC_CONF: {
4100 struct filter_obd *filter = &obd->u.filter;
4101 struct sptlrpc_conf_log *log;
4102 struct sptlrpc_rule_set tmp_rset;
4104 log = sptlrpc_conf_log_extract(lcfg);
4110 sptlrpc_rule_set_init(&tmp_rset);
4112 rc = sptlrpc_rule_set_from_log(&tmp_rset, log);
4114 CERROR("obd %s: failed get sptlrpc rules: %d\n",
4119 write_lock(&filter->fo_sptlrpc_lock);
4120 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
4121 filter->fo_sptlrpc_rset = tmp_rset;
4122 write_unlock(&filter->fo_sptlrpc_lock);
4124 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
4128 lprocfs_filter_init_vars(&lvars);
4130 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4138 static struct lvfs_callback_ops filter_lvfs_ops = {
4139 l_fid2dentry: filter_lvfs_fid2dentry,
4142 static struct obd_ops filter_obd_ops = {
4143 .o_owner = THIS_MODULE,
4144 .o_get_info = filter_get_info,
4145 .o_set_info_async = filter_set_info_async,
4146 .o_setup = filter_setup,
4147 .o_precleanup = filter_precleanup,
4148 .o_cleanup = filter_cleanup,
4149 .o_connect = filter_connect,
4150 .o_reconnect = filter_reconnect,
4151 .o_disconnect = filter_disconnect,
4152 .o_ping = filter_ping,
4153 .o_init_export = filter_init_export,
4154 .o_destroy_export = filter_destroy_export,
4155 .o_statfs = filter_statfs,
4156 .o_getattr = filter_getattr,
4157 .o_unpackmd = filter_unpackmd,
4158 .o_create = filter_create,
4159 .o_setattr = filter_setattr,
4160 .o_destroy = filter_destroy,
4161 .o_brw = filter_brw,
4162 .o_punch = filter_truncate,
4163 .o_sync = filter_sync,
4164 .o_preprw = filter_preprw,
4165 .o_commitrw = filter_commitrw,
4166 .o_llog_init = filter_llog_init,
4167 .o_llog_connect = filter_llog_connect,
4168 .o_llog_finish = filter_llog_finish,
4169 .o_iocontrol = filter_iocontrol,
4170 .o_health_check = filter_health_check,
4171 .o_process_config = filter_process_config,
4174 quota_interface_t *filter_quota_interface_ref;
4175 extern quota_interface_t filter_quota_interface;
4177 static int __init obdfilter_init(void)
4179 struct lprocfs_static_vars lvars;
4182 lprocfs_filter_init_vars(&lvars);
4184 request_module("lquota");
4185 OBD_ALLOC(obdfilter_created_scratchpad,
4186 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4187 sizeof(*obdfilter_created_scratchpad));
4188 if (obdfilter_created_scratchpad == NULL)
4191 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4192 sizeof(struct filter_mod_data),
4195 GOTO(out, rc = -ENOMEM);
4197 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4198 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4200 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4201 LUSTRE_OST_NAME, NULL);
4205 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4206 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4207 ll_fmd_cachep = NULL;
4209 if (filter_quota_interface_ref)
4210 PORTAL_SYMBOL_PUT(filter_quota_interface);
4212 OBD_FREE(obdfilter_created_scratchpad,
4213 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4214 sizeof(*obdfilter_created_scratchpad));
4220 static void __exit obdfilter_exit(void)
4222 if (filter_quota_interface_ref)
4223 PORTAL_SYMBOL_PUT(filter_quota_interface);
4225 if (ll_fmd_cachep) {
4226 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4227 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4228 ll_fmd_cachep = NULL;
4231 class_unregister_type(LUSTRE_OST_NAME);
4232 OBD_FREE(obdfilter_created_scratchpad,
4233 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4234 sizeof(*obdfilter_created_scratchpad));
4237 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4238 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4239 MODULE_LICENSE("GPL");
4241 module_init(obdfilter_init);
4242 module_exit(obdfilter_exit);