1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * linux/fs/obdfilter/filter.c
6 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
7 * Author: Peter Braam <braam@clusterfs.com>
8 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
28 * (which need to get journal_lock, may block if journal full).
30 * Invariant: Call filter_start_transno() before any journal ops to avoid the
31 * same deadlock problem. We can (and want) to get rid of the
32 * transno sem in favour of the dir/inode i_mutex to avoid single
33 * threaded operation on the OST.
36 #define DEBUG_SUBSYSTEM S_FILTER
38 #ifndef AUTOCONF_INCLUDED
39 #include <linux/config.h>
41 #include <linux/module.h>
43 #include <linux/dcache.h>
44 #include <linux/init.h>
45 #include <linux/version.h>
46 #include <linux/sched.h>
47 #include <linux/mount.h>
48 #include <linux/buffer_head.h>
50 #include <obd_cksum.h>
51 #include <obd_class.h>
53 #include <lustre_dlm.h>
54 #include <lustre_fsfilt.h>
55 #include <lprocfs_status.h>
56 #include <lustre_log.h>
57 #include <lustre_commit_confd.h>
58 #include <libcfs/list.h>
59 #include <lustre_disk.h>
60 #include <lustre_quota.h>
61 #include <linux/slab.h>
62 #include <lustre_param.h>
64 #include "filter_internal.h"
66 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
67 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS0
68 static struct lvfs_callback_ops filter_lvfs_ops;
69 cfs_mem_cache_t *ll_fmd_cachep;
71 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
72 void *cb_data, int error)
74 obd_transno_commit_cb(obd, transno, error);
77 /* Assumes caller has already pushed us into the kernel context. */
78 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
79 int rc, int force_sync)
81 struct filter_obd *filter = &exp->exp_obd->u.filter;
82 struct filter_export_data *fed = &exp->exp_filter_data;
83 struct lsd_client_data *lcd = fed->fed_lcd;
86 int err, log_pri = D_RPCTRACE;
88 /* Propagate error code. */
92 if (!exp->exp_obd->obd_replayable || oti == NULL)
95 /* we don't allocate new transnos for replayed requests */
96 if (oti->oti_transno == 0) {
97 spin_lock(&filter->fo_translock);
98 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
99 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
100 spin_unlock(&filter->fo_translock);
101 oti->oti_transno = last_rcvd;
103 spin_lock(&filter->fo_translock);
104 last_rcvd = oti->oti_transno;
105 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
106 filter->fo_fsd->lsd_last_transno =
107 cpu_to_le64(last_rcvd);
108 spin_unlock(&filter->fo_translock);
110 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
112 /* could get xid from oti, if it's ever needed */
113 lcd->lcd_last_xid = 0;
115 off = fed->fed_lr_off;
117 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
118 fed->fed_lr_idx, fed->fed_lr_off);
122 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
128 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
129 lcd, sizeof(*lcd), &off,
130 force_sync | exp->exp_need_sync);
132 filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
140 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
141 last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
146 void f_dput(struct dentry *dentry)
148 /* Can't go inside filter_ddelete because it can block */
149 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
150 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
151 LASSERT(atomic_read(&dentry->d_count) > 0);
156 static void init_brw_stats(struct brw_stats *brw_stats)
159 for (i = 0; i < BRW_LAST; i++)
160 spin_lock_init(&brw_stats->hist[i].oh_lock);
163 static int lprocfs_init_rw_stats(struct obd_device *obd,
164 struct lprocfs_stats **stats)
168 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
169 LPROC_FILTER_LAST - 1;
170 *stats = lprocfs_alloc_stats(num_stats, 0);
174 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
175 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
176 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
177 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
178 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
183 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
184 plus the procfs overhead :( */
185 static int filter_export_stats_init(struct obd_device *obd,
186 struct obd_export *exp,
189 struct filter_export_data *fed = &exp->exp_filter_data;
193 init_brw_stats(&fed->fed_brw_stats);
195 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
196 /* Self-export gets no proc entry */
199 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
204 struct nid_stat *tmp = exp->exp_nid_stats;
205 LASSERT(tmp != NULL);
207 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
208 if (tmp->nid_brw_stats == NULL)
211 init_brw_stats(tmp->nid_brw_stats);
212 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
213 0644, &filter_per_nid_stats_fops,
216 CWARN("Error adding the brw_stats file\n");
218 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
222 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
231 /* Add client data to the FILTER. We use a bitmap to locate a free space
232 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
233 * Otherwise, we have just read the data from the last_rcvd file and
234 * we know its offset. */
235 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
238 struct filter_obd *filter = &obd->u.filter;
239 struct filter_export_data *fed = &exp->exp_filter_data;
240 unsigned long *bitmap = filter->fo_last_rcvd_slots;
241 int new_client = (cl_idx == -1);
245 LASSERT(bitmap != NULL);
246 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
249 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
252 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
253 * there's no need for extra complication here
256 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
258 if (cl_idx >= LR_MAX_CLIENTS) {
259 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
263 if (test_and_set_bit(cl_idx, bitmap)) {
264 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
269 if (test_and_set_bit(cl_idx, bitmap)) {
270 CERROR("FILTER client %d: bit already set in bitmap!\n",
276 fed->fed_lr_idx = cl_idx;
277 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
278 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
279 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
281 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
282 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
285 struct lvfs_run_ctxt saved;
286 loff_t off = fed->fed_lr_off;
290 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
291 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
293 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
294 /* Transaction needed to fix bug 1403 */
295 handle = fsfilt_start(obd,
296 filter->fo_rcvd_filp->f_dentry->d_inode,
297 FSFILT_OP_SETATTR, NULL);
298 if (IS_ERR(handle)) {
299 rc = PTR_ERR(handle);
300 CERROR("unable to start transaction: rc %d\n", rc);
302 rc = fsfilt_add_journal_cb(obd, 0, handle,
303 target_client_add_cb, exp);
305 spin_lock(&exp->exp_lock);
306 exp->exp_need_sync = 1;
307 spin_unlock(&exp->exp_lock);
309 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
311 sizeof(*fed->fed_lcd),
312 &off, rc /* sync if no cb */);
314 filter->fo_rcvd_filp->f_dentry->d_inode,
317 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
320 CERROR("error writing %s client idx %u: rc %d\n",
321 LAST_RCVD, fed->fed_lr_idx, rc);
328 static int filter_client_free(struct obd_export *exp)
330 struct filter_export_data *fed = &exp->exp_filter_data;
331 struct filter_obd *filter = &exp->exp_obd->u.filter;
332 struct obd_device *obd = exp->exp_obd;
333 struct lsd_client_data zero_lcd;
334 struct lvfs_run_ctxt saved;
339 if (fed->fed_lcd == NULL)
342 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
343 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
346 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
347 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
349 LASSERT(filter->fo_last_rcvd_slots != NULL);
351 off = fed->fed_lr_off;
353 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
354 * we leak a client slot that will be cleaned on the next recovery. */
356 CERROR("%s: client idx %d has med_off %lld\n",
357 obd->obd_name, fed->fed_lr_idx, off);
358 GOTO(free, rc = -EINVAL);
361 /* Clear the bit _after_ zeroing out the client so we don't
362 race with filter_client_add and zero out new clients.*/
363 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
364 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
369 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
370 memset(&zero_lcd, 0, sizeof zero_lcd);
371 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
372 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
373 sizeof(zero_lcd), &off,
374 (!exp->exp_libclient ||
375 exp->exp_need_sync));
377 /* update server's transno */
378 filter_update_server_data(obd, filter->fo_rcvd_filp,
380 !exp->exp_libclient);
381 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
383 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
384 "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
385 fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
389 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
390 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
397 OBD_FREE_PTR(fed->fed_lcd);
403 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
404 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
405 struct filter_mod_data *fmd)
407 LASSERT_SPIN_LOCKED(&fed->fed_lock);
408 if (--fmd->fmd_refcount == 0) {
409 /* XXX when we have persistent reservations and the handle
410 * is stored herein we need to drop it here. */
411 fed->fed_mod_count--;
412 list_del(&fmd->fmd_list);
413 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
417 /* drop fmd reference, free it if last ref */
418 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
420 struct filter_export_data *fed;
425 fed = &exp->exp_filter_data;
426 spin_lock(&fed->fed_lock);
427 filter_fmd_put_nolock(fed, fmd); /* caller reference */
428 spin_unlock(&fed->fed_lock);
431 /* expire entries from the end of the list if there are too many
432 * or they are too old */
433 static void filter_fmd_expire_nolock(struct filter_obd *filter,
434 struct filter_export_data *fed,
435 struct filter_mod_data *keep)
437 struct filter_mod_data *fmd, *tmp;
439 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
443 if (time_before(jiffies, fmd->fmd_expire) &&
444 fed->fed_mod_count < filter->fo_fmd_max_num)
447 list_del_init(&fmd->fmd_list);
448 filter_fmd_put_nolock(fed, fmd); /* list reference */
452 void filter_fmd_expire(struct obd_export *exp)
454 spin_lock(&exp->exp_filter_data.fed_lock);
455 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
456 &exp->exp_filter_data, NULL);
457 spin_unlock(&exp->exp_filter_data.fed_lock);
460 /* find specified objid, group in export fmd list.
461 * caller must hold fed_lock and take fmd reference itself */
462 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
463 struct filter_export_data *fed,
464 obd_id objid, obd_gr group)
466 struct filter_mod_data *found = NULL, *fmd;
468 LASSERT_SPIN_LOCKED(&fed->fed_lock);
470 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
471 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
473 list_del(&fmd->fmd_list);
474 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
475 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
480 filter_fmd_expire_nolock(filter, fed, found);
485 /* Find fmd based on objid and group, or return NULL if not found. */
486 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
487 obd_id objid, obd_gr group)
489 struct filter_mod_data *fmd;
491 spin_lock(&exp->exp_filter_data.fed_lock);
492 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
493 &exp->exp_filter_data, objid, group);
495 fmd->fmd_refcount++; /* caller reference */
496 spin_unlock(&exp->exp_filter_data.fed_lock);
501 /* Find fmd based on objid and group, or create a new one if none is found.
502 * It is possible for this function to return NULL under memory pressure,
503 * or if objid = 0 is passed (which will only cause old entries to expire).
504 * Currently this is not fatal because any fmd state is transient and
505 * may also be freed when it gets sufficiently old. */
506 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
507 obd_id objid, obd_gr group)
509 struct filter_export_data *fed = &exp->exp_filter_data;
510 struct filter_mod_data *found = NULL, *fmd_new = NULL;
512 OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO, sizeof(*fmd_new));
514 spin_lock(&fed->fed_lock);
515 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
518 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
519 fmd_new->fmd_id = objid;
520 fmd_new->fmd_gr = group;
521 fmd_new->fmd_refcount++; /* list reference */
523 fed->fed_mod_count++;
525 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
529 found->fmd_refcount++; /* caller reference */
530 found->fmd_expire = jiffies +
531 exp->exp_obd->u.filter.fo_fmd_max_age;
534 spin_unlock(&fed->fed_lock);
540 /* drop fmd list reference so it will disappear when last reference is put.
541 * This isn't so critical because it would in fact only affect the one client
542 * that is doing the unlink and at worst we have an stale entry referencing
543 * an object that should never be used again. */
544 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
546 struct filter_mod_data *found = NULL;
548 spin_lock(&exp->exp_filter_data.fed_lock);
549 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
551 list_del_init(&found->fmd_list);
552 filter_fmd_put_nolock(&exp->exp_filter_data, found);
554 spin_unlock(&exp->exp_filter_data.fed_lock);
557 #define filter_fmd_drop(exp, objid, group)
560 /* remove all entries from fmd list */
561 static void filter_fmd_cleanup(struct obd_export *exp)
563 struct filter_export_data *fed = &exp->exp_filter_data;
564 struct filter_mod_data *fmd = NULL, *tmp;
566 spin_lock(&fed->fed_lock);
567 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
568 list_del_init(&fmd->fmd_list);
569 filter_fmd_put_nolock(fed, fmd);
571 spin_unlock(&fed->fed_lock);
574 static int filter_init_export(struct obd_export *exp)
576 spin_lock_init(&exp->exp_filter_data.fed_lock);
577 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
579 spin_lock(&exp->exp_lock);
580 exp->exp_connecting = 1;
581 spin_unlock(&exp->exp_lock);
586 static int filter_free_server_data(struct filter_obd *filter)
588 OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
589 filter->fo_fsd = NULL;
590 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
591 filter->fo_last_rcvd_slots = NULL;
595 /* assumes caller is already in kernel ctxt */
596 int filter_update_server_data(struct obd_device *obd, struct file *filp,
597 struct lr_server_data *fsd, int force_sync)
603 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
604 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
605 le64_to_cpu(fsd->lsd_last_transno));
606 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
607 le64_to_cpu(fsd->lsd_mount_count));
609 fsd->lsd_compat14 = fsd->lsd_last_transno;
610 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
612 CERROR("error writing lr_server_data: rc = %d\n", rc);
617 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
620 struct filter_obd *filter = &obd->u.filter;
626 if (filter->fo_last_objid_files[group] == NULL) {
627 CERROR("Object group "LPU64" not fully setup; not updating "
628 "last_objid\n", group);
632 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
633 obd->obd_name, group, filter->fo_last_objids[group]);
635 tmp = cpu_to_le64(filter->fo_last_objids[group]);
636 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
637 &tmp, sizeof(tmp), &off, force_sync);
639 CERROR("error writing group "LPU64" last objid: rc = %d\n",
643 extern int ost_handle(struct ptlrpc_request *req);
644 /* assumes caller has already in kernel ctxt */
645 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
647 struct filter_obd *filter = &obd->u.filter;
648 struct lr_server_data *fsd;
649 struct lsd_client_data *lcd = NULL;
650 struct inode *inode = filp->f_dentry->d_inode;
651 unsigned long last_rcvd_size = i_size_read(inode);
657 /* ensure padding in the struct is the correct size */
658 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
659 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
660 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
661 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
663 OBD_ALLOC(fsd, sizeof(*fsd));
666 filter->fo_fsd = fsd;
668 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
669 if (filter->fo_last_rcvd_slots == NULL) {
670 OBD_FREE(fsd, sizeof(*fsd));
674 if (last_rcvd_size == 0) {
675 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
677 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
678 fsd->lsd_last_transno = 0;
679 mount_count = fsd->lsd_mount_count = 0;
680 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
681 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
682 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
683 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
684 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
685 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
687 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
689 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
693 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
694 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
695 "using the wrong disk %s. Were the "
696 "/dev/ assignments rearranged?\n",
697 obd->obd_uuid.uuid, fsd->lsd_uuid);
698 GOTO(err_fsd, rc = -EINVAL);
700 mount_count = le64_to_cpu(fsd->lsd_mount_count);
701 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
703 /* Assume old last_rcvd format unless I_C_LR is set */
704 if (!(fsd->lsd_feature_incompat &
705 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
706 fsd->lsd_last_transno = fsd->lsd_compat14;
710 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
711 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
712 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
713 ~FILTER_INCOMPAT_SUPP);
714 GOTO(err_fsd, rc = -EINVAL);
716 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
717 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
718 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
719 ~FILTER_ROCOMPAT_SUPP);
720 /* Do something like remount filesystem read-only */
721 GOTO(err_fsd, rc = -EINVAL);
724 CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
725 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
726 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
727 obd->obd_name, mount_count + 1);
728 CDEBUG(D_INODE, "%s: server data size: %u\n",
729 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
730 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
731 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
732 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
733 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
734 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
735 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
736 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
737 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
738 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
739 le16_to_cpu(fsd->lsd_client_size));
741 if (!obd->obd_replayable) {
742 CWARN("%s: recovery support OFF\n", obd->obd_name);
746 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
747 off < last_rcvd_size; cl_idx++) {
749 struct obd_export *exp;
750 struct filter_export_data *fed;
755 GOTO(err_client, rc = -ENOMEM);
758 /* Don't assume off is incremented properly by
759 * fsfilt_read_record(), in case sizeof(*lcd)
760 * isn't the same as fsd->lsd_client_size. */
761 off = le32_to_cpu(fsd->lsd_client_start) +
762 cl_idx * le16_to_cpu(fsd->lsd_client_size);
763 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
765 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
766 LAST_RCVD, cl_idx, off, rc);
767 break; /* read error shouldn't cause startup to fail */
770 if (lcd->lcd_uuid[0] == '\0') {
771 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
776 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
778 /* These exports are cleaned up by filter_disconnect(), so they
779 * need to be set up like real exports as filter_connect() does.
781 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
783 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
784 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
785 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
787 if (PTR_ERR(exp) == -EALREADY) {
788 /* export already exists, zero out this one */
789 CERROR("Zeroing out duplicate export due to "
791 lcd->lcd_uuid[0] = '\0';
793 GOTO(err_client, rc = PTR_ERR(exp));
796 fed = &exp->exp_filter_data;
798 fed->fed_group = 0; /* will be assigned at connect */
799 filter_export_stats_init(obd, exp, NULL);
800 rc = filter_client_add(obd, exp, cl_idx);
801 /* can't fail for existing client */
802 LASSERTF(rc == 0, "rc = %d\n", rc);
805 spin_lock(&exp->exp_lock);
806 exp->exp_connecting = 0;
807 exp->exp_in_recovery = 0;
808 spin_unlock(&exp->exp_lock);
809 obd->obd_max_recoverable_clients++;
810 class_export_put(exp);
813 /* Need to check last_rcvd even for duplicated exports. */
814 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
817 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
818 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
824 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
826 target_recovery_init(obd, ost_handle);
829 filter->fo_mount_count = mount_count + 1;
830 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
832 /* save it, so mount count and last_transno is current */
833 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
835 GOTO(err_client, rc);
840 target_recovery_fini(obd);
842 filter_free_server_data(filter);
846 static int filter_cleanup_groups(struct obd_device *obd)
848 struct filter_obd *filter = &obd->u.filter;
850 struct dentry *dentry;
854 if (filter->fo_dentry_O_groups != NULL) {
855 for (i = 0; i < filter->fo_group_count; i++) {
856 dentry = filter->fo_dentry_O_groups[i];
860 OBD_FREE(filter->fo_dentry_O_groups,
861 filter->fo_group_count *
862 sizeof(*filter->fo_dentry_O_groups));
863 filter->fo_dentry_O_groups = NULL;
865 if (filter->fo_last_objid_files != NULL) {
866 for (i = 0; i < filter->fo_group_count; i++) {
867 filp = filter->fo_last_objid_files[i];
871 OBD_FREE(filter->fo_last_objid_files,
872 filter->fo_group_count *
873 sizeof(*filter->fo_last_objid_files));
874 filter->fo_last_objid_files = NULL;
876 if (filter->fo_dentry_O_sub != NULL) {
877 for (i = 0; i < filter->fo_group_count; i++) {
878 for (j = 0; j < filter->fo_subdir_count; j++) {
879 dentry = filter->fo_dentry_O_sub[i].dentry[j];
884 OBD_FREE(filter->fo_dentry_O_sub,
885 filter->fo_group_count *
886 sizeof(*filter->fo_dentry_O_sub));
887 filter->fo_dentry_O_sub = NULL;
889 if (filter->fo_last_objids != NULL) {
890 OBD_FREE(filter->fo_last_objids,
891 filter->fo_group_count *
892 sizeof(*filter->fo_last_objids));
893 filter->fo_last_objids = NULL;
895 if (filter->fo_dentry_O != NULL) {
896 f_dput(filter->fo_dentry_O);
897 filter->fo_dentry_O = NULL;
902 static int filter_update_last_group(struct obd_device *obd, int group)
904 struct filter_obd *filter = &obd->u.filter;
905 struct file *filp = NULL;
906 int last_group = 0, rc;
910 if (group <= filter->fo_committed_group)
913 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
917 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
921 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
923 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
926 LASSERT(off == 0 || last_group >= FILTER_MIN_GROUPS);
927 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
928 obd->obd_name, last_group, group);
932 /* must be sync: bXXXX */
933 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
935 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
939 filter->fo_committed_group = group;
946 static int filter_read_group_internal(struct obd_device *obd, int group,
949 struct filter_obd *filter = &obd->u.filter;
950 __u64 *new_objids = NULL;
951 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
952 struct dentry **new_groups = NULL;
953 struct file **new_files = NULL;
954 struct dentry *dentry;
956 int old_count = filter->fo_group_count, rc, stage = 0, i;
962 snprintf(name, 24, "%d", group);
966 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
968 if (IS_ERR(dentry)) {
969 CERROR("Cannot lookup expected object group %d: %ld\n",
970 group, PTR_ERR(dentry));
971 RETURN(PTR_ERR(dentry));
974 dentry = simple_mkdir(filter->fo_dentry_O, name, 0700, 1);
975 if (IS_ERR(dentry)) {
976 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
978 RETURN(PTR_ERR(dentry));
983 snprintf(name, 24, "O/%d/LAST_ID", group);
985 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
987 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
988 GOTO(cleanup, rc = PTR_ERR(filp));
992 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
994 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
998 if (filter->fo_subdir_count) {
999 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1000 if (tmp_subdirs == NULL)
1001 GOTO(cleanup, rc = -ENOMEM);
1004 for (i = 0; i < filter->fo_subdir_count; i++) {
1006 snprintf(dir, sizeof(dir), "d%u", i);
1008 tmp_subdirs->dentry[i] = simple_mkdir(dentry, dir, 0700, 1);
1009 if (IS_ERR(tmp_subdirs->dentry[i])) {
1010 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1011 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1016 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1017 tmp_subdirs->dentry[i]);
1021 /* 'group' is an index; we need an array of length 'group + 1' */
1022 if (group + 1 > old_count) {
1023 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1024 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1025 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1026 OBD_ALLOC(new_files, len * sizeof(*new_files));
1028 if (new_objids == NULL || new_subdirs == NULL ||
1029 new_groups == NULL || new_files == NULL)
1030 GOTO(cleanup, rc = -ENOMEM);
1033 memcpy(new_objids, filter->fo_last_objids,
1034 old_count * sizeof(*new_objids));
1035 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1036 old_count * sizeof(*new_subdirs));
1037 memcpy(new_groups, filter->fo_dentry_O_groups,
1038 old_count * sizeof(*new_groups));
1039 memcpy(new_files, filter->fo_last_objid_files,
1040 old_count * sizeof(*new_files));
1042 OBD_FREE(filter->fo_last_objids,
1043 old_count * sizeof(*new_objids));
1044 OBD_FREE(filter->fo_dentry_O_sub,
1045 old_count * sizeof(*new_subdirs));
1046 OBD_FREE(filter->fo_dentry_O_groups,
1047 old_count * sizeof(*new_groups));
1048 OBD_FREE(filter->fo_last_objid_files,
1049 old_count * sizeof(*new_files));
1051 filter->fo_last_objids = new_objids;
1052 filter->fo_dentry_O_sub = new_subdirs;
1053 filter->fo_dentry_O_groups = new_groups;
1054 filter->fo_last_objid_files = new_files;
1055 filter->fo_group_count = len;
1058 filter->fo_dentry_O_groups[group] = dentry;
1059 filter->fo_last_objid_files[group] = filp;
1060 if (filter->fo_subdir_count) {
1061 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1062 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1065 filter_update_last_group(obd, group);
1067 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1068 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1069 rc = filter_update_last_objid(obd, group, 1);
1073 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1074 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1075 obd->obd_name, group, last_objid);
1080 if (new_objids != NULL)
1081 OBD_FREE(new_objids, len * sizeof(*new_objids));
1082 if (new_subdirs != NULL)
1083 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1084 if (new_groups != NULL)
1085 OBD_FREE(new_groups, len * sizeof(*new_groups));
1086 if (new_files != NULL)
1087 OBD_FREE(new_files, len * sizeof(*new_files));
1089 if (filter->fo_subdir_count) {
1090 for (i = 0; i < filter->fo_subdir_count; i++) {
1091 if (tmp_subdirs->dentry[i] != NULL)
1092 dput(tmp_subdirs->dentry[i]);
1094 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1097 filp_close(filp, 0);
1104 static int filter_read_groups(struct obd_device *obd, int last_group,
1107 struct filter_obd *filter = &obd->u.filter;
1108 int old_count, group, rc = 0;
1110 down(&filter->fo_init_lock);
1111 old_count = filter->fo_group_count;
1112 for (group = old_count; group <= last_group; group++) {
1114 continue; /* no group zero */
1116 rc = filter_read_group_internal(obd, group, create);
1120 up(&filter->fo_init_lock);
1124 /* FIXME: object groups */
1125 static int filter_prep_groups(struct obd_device *obd)
1127 struct filter_obd *filter = &obd->u.filter;
1128 struct dentry *dentry, *O_dentry;
1130 int last_group, rc = 0, cleanup_phase = 0;
1134 O_dentry = simple_mkdir(current->fs->pwd, "O", 0700, 1);
1135 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1136 if (IS_ERR(O_dentry)) {
1137 rc = PTR_ERR(O_dentry);
1138 CERROR("cannot open/create O: rc = %d\n", rc);
1141 filter->fo_dentry_O = O_dentry;
1142 cleanup_phase = 1; /* O_dentry */
1144 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1145 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1146 * some time post 1.0 when all old-style OSTs have converted along
1147 * with the init_objid hack. */
1148 dentry = ll_lookup_one_len("R", O_dentry, 1);
1150 GOTO(cleanup, rc = PTR_ERR(dentry));
1151 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1152 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1155 CWARN("converting OST to new object layout\n");
1156 if (IS_ERR(O0_dentry)) {
1157 rc = PTR_ERR(O0_dentry);
1158 CERROR("error looking up O/0: rc %d\n", rc);
1159 GOTO(cleanup_R, rc);
1162 if (O0_dentry->d_inode) {
1163 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1164 GOTO(cleanup_O0, rc = -EEXIST);
1167 LOCK_INODE_MUTEX(O_dentry->d_inode);
1168 rc = vfs_rename(O_dentry->d_inode, dentry,
1169 O_dentry->d_inode, O0_dentry);
1170 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1173 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1174 GOTO(cleanup_O0, rc);
1176 filter->fo_fsd->lsd_feature_incompat |=
1177 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1178 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1180 GOTO(cleanup_O0, rc);
1192 cleanup_phase = 2; /* groups */
1194 /* we have to initialize all groups before first connections from
1195 * clients because they may send create/destroy for any group -bzzz */
1196 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1198 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1199 GOTO(cleanup, rc = PTR_ERR(filp));
1201 cleanup_phase = 3; /* filp */
1203 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1205 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1209 last_group = FILTER_MIN_GROUPS;
1211 LASSERT(last_group >= FILTER_MIN_GROUPS);
1214 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1215 FILTER_MIN_GROUPS, last_group);
1216 filter->fo_committed_group = last_group;
1217 rc = filter_read_groups(obd, last_group, 1);
1221 filp_close(filp, 0);
1225 switch (cleanup_phase) {
1227 filp_close(filp, 0);
1229 filter_cleanup_groups(obd);
1231 f_dput(filter->fo_dentry_O);
1232 filter->fo_dentry_O = NULL;
1240 /* setup the object store with correct subdirectories */
1241 static int filter_prep(struct obd_device *obd)
1243 struct lvfs_run_ctxt saved;
1244 struct filter_obd *filter = &obd->u.filter;
1246 struct inode *inode;
1250 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1251 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1252 if (!file || IS_ERR(file)) {
1254 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1258 filter->fo_rcvd_filp = file;
1259 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1260 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1261 file->f_dentry->d_inode->i_mode);
1262 GOTO(err_filp, rc = -ENOENT);
1265 inode = file->f_dentry->d_parent->d_inode;
1266 /* We use i_op->unlink directly in filter_vfs_unlink() */
1267 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1268 CERROR("%s: filesystem does not support create/unlink ops\n",
1270 GOTO(err_filp, rc = -EOPNOTSUPP);
1273 rc = filter_init_server_data(obd, file);
1275 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1278 /* open and create health check io file*/
1279 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1282 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1284 GOTO(err_server_data, rc);
1286 filter->fo_health_check_filp = file;
1287 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1288 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1289 file->f_dentry->d_inode->i_mode);
1290 GOTO(err_health_check, rc = -ENOENT);
1292 rc = lvfs_check_io_health(obd, file);
1294 GOTO(err_health_check, rc);
1296 rc = filter_prep_groups(obd);
1298 GOTO(err_health_check, rc);
1300 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1305 if (filp_close(filter->fo_health_check_filp, 0))
1306 CERROR("can't close %s after error\n", HEALTH_CHECK);
1307 filter->fo_health_check_filp = NULL;
1309 target_recovery_fini(obd);
1310 filter_free_server_data(filter);
1312 if (filp_close(filter->fo_rcvd_filp, 0))
1313 CERROR("can't close %s after error\n", LAST_RCVD);
1314 filter->fo_rcvd_filp = NULL;
1318 /* cleanup the filter: write last used object id to status file */
1319 static void filter_post(struct obd_device *obd)
1321 struct lvfs_run_ctxt saved;
1322 struct filter_obd *filter = &obd->u.filter;
1325 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1326 * best to start a transaction with h_sync, because we removed this
1329 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1330 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1333 CERROR("error writing server data: rc = %d\n", rc);
1335 for (i = 1; i < filter->fo_group_count; i++) {
1336 rc = filter_update_last_objid(obd, i,
1337 (i == filter->fo_group_count - 1));
1339 CERROR("error writing group %d lastobjid: rc = %d\n",
1343 rc = filp_close(filter->fo_rcvd_filp, 0);
1344 filter->fo_rcvd_filp = NULL;
1346 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1348 rc = filp_close(filter->fo_health_check_filp, 0);
1349 filter->fo_health_check_filp = NULL;
1351 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1353 filter_cleanup_groups(obd);
1354 filter_free_server_data(filter);
1355 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1357 filter_free_capa_keys(filter);
1358 cleanup_capa_hash(filter->fo_capa_hash);
1361 static void filter_set_last_id(struct filter_obd *filter,
1362 obd_id id, obd_gr group)
1364 LASSERT(filter->fo_fsd != NULL);
1365 LASSERT(group <= filter->fo_group_count);
1367 spin_lock(&filter->fo_objidlock);
1368 filter->fo_last_objids[group] = id;
1369 spin_unlock(&filter->fo_objidlock);
1372 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1375 LASSERT(filter->fo_fsd != NULL);
1376 LASSERT(group <= filter->fo_group_count);
1378 /* FIXME: object groups */
1379 spin_lock(&filter->fo_objidlock);
1380 id = filter->fo_last_objids[group];
1381 spin_unlock(&filter->fo_objidlock);
1386 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1388 LOCK_INODE_MUTEX(dparent->d_inode);
1392 /* We never dget the object parent, so DON'T dput it either */
1393 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1395 struct filter_obd *filter = &obd->u.filter;
1396 struct filter_subdirs *subdirs;
1397 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1399 if (group > 0 || filter->fo_subdir_count == 0)
1400 return filter->fo_dentry_O_groups[group];
1402 subdirs = &filter->fo_dentry_O_sub[group];
1403 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1406 /* We never dget the object parent, so DON'T dput it either */
1407 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1410 unsigned long now = jiffies;
1411 struct dentry *dparent = filter_parent(obd, group, objid);
1414 if (IS_ERR(dparent))
1416 if (dparent == NULL)
1417 return ERR_PTR(-ENOENT);
1419 rc = filter_lock_dentry(obd, dparent);
1420 fsfilt_check_slow(obd, now, "parent lock");
1421 return rc ? ERR_PTR(rc) : dparent;
1424 /* We never dget the object parent, so DON'T dput it either */
1425 static void filter_parent_unlock(struct dentry *dparent)
1427 UNLOCK_INODE_MUTEX(dparent->d_inode);
1430 /* How to get files, dentries, inodes from object id's.
1432 * If dir_dentry is passed, the caller has already locked the parent
1433 * appropriately for this operation (normally a write lock). If
1434 * dir_dentry is NULL, we do a read lock while we do the lookup to
1435 * avoid races with create/destroy and such changing the directory
1436 * internal to the filesystem code. */
1437 struct dentry *filter_fid2dentry(struct obd_device *obd,
1438 struct dentry *dir_dentry,
1439 obd_gr group, obd_id id)
1441 struct dentry *dparent = dir_dentry;
1442 struct dentry *dchild;
1447 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1448 obd->u.filter.fo_destroys_in_progress == 0) {
1449 /* don't fail lookups for orphan recovery, it causes
1450 * later LBUGs when objects still exist during precreate */
1451 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1452 RETURN(ERR_PTR(-ENOENT));
1455 CERROR("fatal: invalid object id 0\n");
1456 RETURN(ERR_PTR(-ESTALE));
1459 len = sprintf(name, LPU64, id);
1460 if (dir_dentry == NULL) {
1461 dparent = filter_parent_lock(obd, group, id);
1462 if (IS_ERR(dparent)) {
1463 CERROR("%s: error getting object "LPU64":"LPU64
1464 " parent: rc %ld\n", obd->obd_name,
1465 id, group, PTR_ERR(dparent));
1469 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1470 dparent->d_name.len, dparent->d_name.name, name);
1471 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1472 if (dir_dentry == NULL)
1473 filter_parent_unlock(dparent);
1474 if (IS_ERR(dchild)) {
1475 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1480 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1481 CERROR("%s: got bad object "LPU64" inode %lu\n",
1482 obd->obd_name, id, dchild->d_inode->i_ino);
1484 RETURN(ERR_PTR(-ENOENT));
1487 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1488 name, dchild, atomic_read(&dchild->d_count));
1490 LASSERT(atomic_read(&dchild->d_count) > 0);
1495 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1498 struct lustre_handle lockh;
1499 int flags = LDLM_AST_DISCARD_DATA, rc;
1500 struct ldlm_res_id res_id;
1501 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1504 osc_build_res_name(objid, group, &res_id);
1505 /* Tell the clients that the object is gone now and that they should
1506 * throw away any cached pages. */
1507 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1508 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1509 ldlm_completion_ast, NULL, NULL, 0, NULL,
1512 /* We only care about the side-effects, just drop the lock. */
1514 ldlm_lock_decref(&lockh, LCK_PW);
1519 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1520 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1521 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1522 * reverse. Caller must take i_sem before starting the transaction and we
1523 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1524 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry)
1529 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1530 * itself we already hold dir->i_mutex for child create/unlink ops */
1531 LASSERT(dentry->d_inode != NULL);
1532 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1533 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1537 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1538 GOTO(out, rc = -ENOENT);
1540 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1545 GOTO(out, rc = -EPERM);
1547 /* check_sticky() */
1548 if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1549 IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1550 GOTO(out, rc = -EPERM);
1552 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1553 * that was done because of journal_start (which is already done
1554 * here) or some other ordering issue. */
1557 rc = security_inode_unlink(dir, dentry);
1561 rc = dir->i_op->unlink(dir, dentry);
1563 /* need to drop i_mutex before we lose inode reference */
1564 UNLOCK_INODE_MUTEX(dentry->d_inode);
1571 /* Caller must hold LCK_PW on parent and push us into kernel context.
1572 * Caller must hold child i_mutex, we drop it always.
1573 * Caller is also required to ensure that dchild->d_inode exists. */
1574 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1575 obd_gr group, struct dentry *dparent,
1576 struct dentry *dchild)
1578 struct inode *inode = dchild->d_inode;
1581 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1582 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1583 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1584 (unsigned long)inode->i_nlink,
1585 atomic_read(&inode->i_count));
1588 rc = filter_vfs_unlink(dparent->d_inode, dchild);
1590 CERROR("error unlinking objid %.*s: rc %d\n",
1591 dchild->d_name.len, dchild->d_name.name, rc);
1595 struct filter_intent_args {
1596 struct ldlm_lock **victim;
1601 static enum interval_iter filter_intent_cb(struct interval_node *n,
1604 struct ldlm_interval *node = (struct ldlm_interval *)n;
1605 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1606 __u64 size = arg->size;
1607 struct ldlm_lock **v = arg->victim;
1608 struct ldlm_lock *lck;
1610 /* If the interval is lower than the current file size,
1612 if (interval_high(n) <= size)
1613 return INTERVAL_ITER_STOP;
1615 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1616 /* Don't send glimpse ASTs to liblustre clients.
1617 * They aren't listening for them, and they do
1618 * entirely synchronous I/O anyways. */
1619 if (lck->l_export == NULL ||
1620 lck->l_export->exp_libclient == 1)
1623 if (*arg->liblustre)
1624 *arg->liblustre = 0;
1627 *v = LDLM_LOCK_GET(lck);
1628 } else if ((*v)->l_policy_data.l_extent.start <
1629 lck->l_policy_data.l_extent.start) {
1631 *v = LDLM_LOCK_GET(lck);
1634 /* the same policy group - every lock has the
1635 * same extent, so needn't do it any more */
1639 return INTERVAL_ITER_CONT;
1642 static int filter_intent_policy(struct ldlm_namespace *ns,
1643 struct ldlm_lock **lockp, void *req_cookie,
1644 ldlm_mode_t mode, int flags, void *data)
1646 CFS_LIST_HEAD(rpc_list);
1647 struct ptlrpc_request *req = req_cookie;
1648 struct ldlm_lock *lock = *lockp, *l = NULL;
1649 struct ldlm_resource *res = lock->l_resource;
1650 ldlm_processing_policy policy;
1651 struct ost_lvb *res_lvb, *reply_lvb;
1652 struct ldlm_reply *rep;
1654 int idx, rc, tmpflags = 0, only_liblustre = 1;
1655 struct ldlm_interval_tree *tree;
1656 struct filter_intent_args arg;
1657 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1658 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1659 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1662 policy = ldlm_get_processing_policy(res);
1663 LASSERT(policy != NULL);
1664 LASSERT(req != NULL);
1666 rc = lustre_pack_reply(req, 3, repsize, NULL);
1668 RETURN(req->rq_status = rc);
1670 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1671 LASSERT(rep != NULL);
1673 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1674 sizeof(*reply_lvb));
1675 LASSERT(reply_lvb != NULL);
1677 //fixup_handle_for_resent_req(req, lock, &lockh);
1679 /* If we grant any lock at all, it will be a whole-file read lock.
1680 * Call the extent policy function to see if our request can be
1681 * granted, or is blocked.
1682 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse lock
1684 lock->l_policy_data.l_extent.start = 0;
1685 lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1686 lock->l_req_mode = LCK_PR;
1688 LASSERT(ns == res->lr_namespace);
1690 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1691 check_res_locked(res);
1693 /* FIXME: we should change the policy function slightly, to not make
1694 * this list at all, since we just turn around and free it */
1695 while (!list_empty(&rpc_list)) {
1696 struct ldlm_lock *wlock =
1697 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1698 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1699 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1700 lock->l_flags &= ~LDLM_FL_CP_REQD;
1701 list_del_init(&wlock->l_cp_ast);
1702 LDLM_LOCK_PUT(wlock);
1705 /* The lock met with no resistance; we're finished. */
1706 if (rc == LDLM_ITER_CONTINUE) {
1707 /* do not grant locks to the liblustre clients: they cannot
1708 * handle ASTs robustly. We need to do this while still
1709 * holding ns_lock to avoid the lock remaining on the res_link
1710 * list (and potentially being added to l_pending_list by an
1711 * AST) when we are going to drop this lock ASAP. */
1712 if (lock->l_export->exp_libclient ||
1713 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1714 ldlm_resource_unlink_lock(lock);
1715 err = ELDLM_LOCK_ABORTED;
1717 err = ELDLM_LOCK_REPLACED;
1723 /* Do not grant any lock, but instead send GL callbacks. The extent
1724 * policy nicely created a list of all PW locks for us. We will choose
1725 * the highest of those which are larger than the size in the LVB, if
1726 * any, and perform a glimpse callback. */
1727 res_lvb = res->lr_lvb_data;
1728 LASSERT(res_lvb != NULL);
1729 *reply_lvb = *res_lvb;
1732 * ->ns_lock guarantees that no new locks are granted, and,
1733 * therefore, that res->lr_lvb_data cannot increase beyond the
1734 * end of already granted lock. As a result, it is safe to
1735 * check against "stale" reply_lvb->lvb_size value without
1738 arg.size = reply_lvb->lvb_size;
1740 arg.liblustre = &only_liblustre;
1741 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1742 tree = &res->lr_itree[idx];
1743 if (tree->lit_mode == LCK_PR)
1746 interval_iterate_reverse(tree->lit_root,
1747 filter_intent_cb, &arg);
1751 /* There were no PW locks beyond the size in the LVB; finished. */
1753 if (only_liblustre) {
1754 /* If we discovered a liblustre client with a PW lock,
1755 * however, the LVB may be out of date! The LVB is
1756 * updated only on glimpse (which we don't do for
1757 * liblustre clients) and cancel (which the client
1758 * obviously has not yet done). So if it has written
1759 * data but kept the lock, the LVB is stale and needs
1760 * to be updated from disk.
1762 * Of course, this will all disappear when we switch to
1763 * taking liblustre locks on the OST. */
1764 ldlm_res_lvbo_update(res, NULL, 0, 1);
1766 RETURN(ELDLM_LOCK_ABORTED);
1770 * This check is for lock taken in filter_prepare_destroy() that does
1771 * not have l_glimpse_ast set. So the logic is: if there is a lock
1772 * with no l_glimpse_ast set, this object is being destroyed already.
1774 * Hence, if you are grabbing DLM locks on the server, always set
1775 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1777 if (l->l_glimpse_ast == NULL) {
1778 /* We are racing with unlink(); just return -ENOENT */
1779 rep->lock_policy_res1 = -ENOENT;
1783 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1784 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1785 /* Update the LVB from disk if the AST failed (this is a legal race) */
1787 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1788 * sending ast is not handled. This can result in lost client writes.
1791 ldlm_res_lvbo_update(res, NULL, 0, 1);
1794 *reply_lvb = *res_lvb;
1800 RETURN(ELDLM_LOCK_ABORTED);
1804 * per-obd_device iobuf pool.
1806 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1807 * allocations in write-path has to be minimized (see bug 5137).
1809 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1810 * OST threads (see ost_thread_{init,done}()).
1812 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1813 * at the OST layer there are only (potentially) multiple obd_device of type
1814 * unknown at the time of OST thread creation.
1816 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1817 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1818 * it's very own iobuf.
1822 * filter_kiobuf_pool_init()
1824 * filter_kiobuf_pool_done()
1826 * filter_iobuf_get()
1828 * operate on this array. They are "generic" in a sense that they don't depend
1829 * on actual type of iobuf's (the latter depending on Linux kernel version).
1833 * destroy pool created by filter_iobuf_pool_init
1835 static void filter_iobuf_pool_done(struct filter_obd *filter)
1837 struct filter_iobuf **pool;
1842 pool = filter->fo_iobuf_pool;
1844 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1845 if (pool[i] != NULL)
1846 filter_free_iobuf(pool[i]);
1848 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1849 filter->fo_iobuf_pool = NULL;
1855 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1857 static int filter_iobuf_pool_init(struct filter_obd *filter)
1864 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1866 if (filter->fo_iobuf_pool == NULL)
1869 filter->fo_iobuf_count = OSS_THREADS_MAX;
1874 /* Return iobuf allocated for @thread_id. We don't know in advance how
1875 * many threads there will be so we allocate a large empty array and only
1876 * fill in those slots that are actually in use.
1877 * If we haven't allocated a pool entry for this thread before, do so now. */
1878 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1880 int thread_id = oti ? oti->oti_thread_id : -1;
1881 struct filter_iobuf *pool = NULL;
1882 struct filter_iobuf **pool_place = NULL;
1884 if (thread_id >= 0) {
1885 LASSERT(thread_id < filter->fo_iobuf_count);
1886 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1889 if (unlikely(pool == NULL)) {
1890 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1891 PTLRPC_MAX_BRW_PAGES);
1892 if (pool_place != NULL)
1899 /* mount the file system (secretly). lustre_cfg parameters are:
1902 * 3 = flags: failover=f, failout=n
1905 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1908 struct filter_obd *filter = &obd->u.filter;
1909 struct vfsmount *mnt;
1910 struct lustre_mount_info *lmi;
1911 struct obd_uuid uuid;
1919 if (lcfg->lcfg_bufcount < 3 ||
1920 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1921 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1924 lmi = server_get_mount(obd->obd_name);
1926 /* We already mounted in lustre_fill_super.
1927 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1928 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1930 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1932 /* old path - used by lctl */
1933 CERROR("Using old MDS mount method\n");
1934 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1935 MS_NOATIME|MS_NODIRATIME,
1936 lustre_cfg_string(lcfg, 1), option);
1939 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1940 lustre_cfg_string(lcfg, 1), rc);
1944 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1946 if (IS_ERR(obd->obd_fsops))
1947 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1949 rc = filter_iobuf_pool_init(filter);
1953 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1955 /* failover is the default */
1956 obd->obd_replayable = 1;
1958 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1959 str = lustre_cfg_string(lcfg, 3);
1960 if (strchr(str, 'n')) {
1961 CWARN("%s: recovery disabled\n", obd->obd_name);
1962 obd->obd_replayable = 0;
1966 filter->fo_vfsmnt = mnt;
1967 obd->u.obt.obt_sb = mnt->mnt_sb;
1968 filter->fo_fstype = mnt->mnt_sb->s_type->name;
1969 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1971 fsfilt_setup(obd, obd->u.obt.obt_sb);
1973 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1974 obd->obd_lvfs_ctxt.pwdmnt = mnt;
1975 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1976 obd->obd_lvfs_ctxt.fs = get_ds();
1977 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
1979 init_mutex(&filter->fo_init_lock);
1980 filter->fo_committed_group = 0;
1982 rc = filter_prep(obd);
1986 filter->fo_destroys_in_progress = 0;
1987 for (i = 0; i < 32; i++)
1988 sema_init(&filter->fo_create_locks[i], 1);
1990 spin_lock_init(&filter->fo_translock);
1991 spin_lock_init(&filter->fo_objidlock);
1992 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
1993 sema_init(&filter->fo_alloc_lock, 1);
1994 init_brw_stats(&filter->fo_filter_stats);
1995 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1996 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
1997 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
1999 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2000 spin_lock_init(&filter->fo_llog_list_lock);
2002 filter->fo_sptlrpc_lock = RW_LOCK_UNLOCKED;
2003 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2005 filter->fo_fl_oss_capa = 0;
2006 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2007 filter->fo_capa_hash = init_capa_hash();
2008 if (filter->fo_capa_hash == NULL)
2009 GOTO(err_ops, rc = -ENOMEM);
2011 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2012 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2013 LDLM_NAMESPACE_GREEDY);
2014 if (obd->obd_namespace == NULL)
2015 GOTO(err_post, rc = -ENOMEM);
2016 obd->obd_namespace->ns_lvbp = obd;
2017 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2018 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2020 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2021 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2023 rc = llog_cat_initialize(obd, &obd->obd_olg, 1, NULL);
2025 CERROR("failed to setup llogging subsystems\n");
2029 rc = lquota_setup(filter_quota_interface_ref, obd);
2033 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2034 if (q->max_sectors < q->max_hw_sectors &&
2035 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2036 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2037 "for larger I/O requests: max_sectors = %u "
2038 "could be up to max_hw_sectors=%u\n",
2039 obd->obd_name, mnt->mnt_sb->s_id,
2040 q->max_sectors, q->max_hw_sectors);
2042 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2043 if (uuid_ptr != NULL) {
2044 class_uuid_unparse(uuid_ptr, &uuid);
2050 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2052 if (obd->obd_recovering) {
2053 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2054 "recovery for at least %d:%.02d, or until %d "
2055 "client%s reconnect. During this time new clients"
2056 " will not be allowed to connect. "
2057 "Recovery progress can be monitored by watching "
2058 "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
2059 obd->obd_name, lustre_cfg_string(lcfg, 1),
2060 label ?: "", label ? "/" : "", str,
2061 obd->obd_recovery_timeout / 60,
2062 obd->obd_recovery_timeout % 60,
2063 obd->obd_max_recoverable_clients,
2064 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2067 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2068 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2069 label ?: "", label ? "/" : "", str,
2070 obd->obd_replayable ? "enabled" : "disabled");
2078 fsfilt_put_ops(obd->obd_fsops);
2079 filter_iobuf_pool_done(filter);
2081 server_put_mount(obd->obd_name, mnt);
2082 obd->u.obt.obt_sb = 0;
2086 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2088 struct lprocfs_static_vars lvars;
2093 CLASSERT(offsetof(struct obd_device, u.obt) ==
2094 offsetof(struct obd_device, u.filter.fo_obt));
2096 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2099 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2100 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2103 addr = (unsigned long)cfs_page_address(page);
2104 clear_page((void *)addr);
2106 /* lprocfs must be setup before the filter so state can be safely added
2107 * to /proc incrementally as the filter is setup */
2108 lprocfs_filter_init_vars(&lvars);
2109 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2110 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2111 /* Init obdfilter private stats here */
2112 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2113 LPROCFS_CNTR_AVGMINMAX,
2114 "read_bytes", "bytes");
2115 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2116 LPROCFS_CNTR_AVGMINMAX,
2117 "write_bytes", "bytes");
2119 lproc_filter_attach_seqstat(obd);
2120 obd->obd_proc_exports_entry = lprocfs_register("exports",
2121 obd->obd_proc_entry,
2123 if (IS_ERR(obd->obd_proc_exports_entry)) {
2124 rc = PTR_ERR(obd->obd_proc_exports_entry);
2125 CERROR("error %d setting up lprocfs for %s\n",
2127 obd->obd_proc_exports_entry = NULL;
2130 if (obd->obd_proc_exports_entry)
2131 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2132 lprocfs_nid_stats_clear_read,
2133 lprocfs_nid_stats_clear_write, obd);
2135 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2136 LUSTRE_CFG_BUFLEN(lcfg, 4));
2137 rc = filter_common_setup(obd, lcfg, (void *)addr);
2138 OBD_PAGE_FREE(page);
2141 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2142 lprocfs_free_per_client_stats(obd);
2143 lprocfs_free_obd_stats(obd);
2144 lprocfs_obd_cleanup(obd);
2150 static int filter_group_llog_finish(struct obd_llog_group *olg)
2152 struct llog_ctxt *ctxt;
2153 int rc = 0, rc2 = 0;
2156 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2158 rc = llog_cleanup(ctxt);
2160 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2162 rc2 = llog_cleanup(ctxt);
2169 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
2170 static struct llog_operations filter_size_orig_logops = {
2171 lop_setup: llog_obd_origin_setup,
2172 lop_cleanup: llog_obd_origin_cleanup,
2173 lop_add: llog_obd_origin_add
2177 * Init obd_llog_group of the filter
2179 static int filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2180 struct obd_device *tgt)
2184 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2185 &filter_mds_ost_repl_logops);
2189 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2190 &filter_size_orig_logops);
2195 filter_group_llog_finish(olg);
2200 * Init the default olg, which is embeded in the obd_device, for filter.
2203 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2204 struct obd_device *tgt)
2206 struct filter_obd *filter = &obd->u.filter;
2207 struct llog_ctxt *ctxt;
2210 LASSERT(filter->fo_lcm == NULL);
2211 OBD_ALLOC(filter->fo_lcm, sizeof(struct llog_commit_master));
2212 if (!filter->fo_lcm)
2215 rc = llog_init_commit_master((struct llog_commit_master *)
2220 filter_mds_ost_repl_logops = llog_client_ops;
2221 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2222 filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
2223 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2225 rc = filter_olg_init(obd, olg, tgt);
2229 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2230 LASSERT(ctxt != NULL);
2231 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2232 ctxt->loc_lcm = obd->u.filter.fo_lcm;
2233 /* Only start log commit thread for default llog ctxt*/
2234 rc = llog_start_commit_thread(ctxt->loc_lcm);
2235 llog_ctxt_put(ctxt);
2240 llog_cleanup_commit_master(filter->fo_lcm, 0);
2241 OBD_FREE(filter->fo_lcm, sizeof(struct llog_commit_master));
2242 filter->fo_lcm = NULL;
2247 static int filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2248 struct obd_device *tgt, int count,
2249 struct llog_catid *catid,
2250 struct obd_uuid *uuid)
2252 struct filter_obd *filter = &obd->u.filter;
2253 struct llog_ctxt *ctxt;
2257 LASSERT(olg != NULL);
2258 if (olg == &obd->obd_olg)
2259 return filter_default_olg_init(obd, olg, tgt);
2261 /* For other group llog, which will be initialized in
2262 * llog_connect(after default_olg_init, so fo_lcm must
2263 * already be initialized */
2264 LASSERT(filter->fo_lcm != NULL);
2265 rc = filter_olg_init(obd, olg, tgt);
2268 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2269 LASSERT(ctxt != NULL);
2270 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2271 ctxt->loc_lcm = obd->u.filter.fo_lcm;
2272 llog_ctxt_put(ctxt);
2277 static int filter_llog_finish(struct obd_device *obd, int count)
2282 if (obd->u.filter.fo_lcm) {
2283 llog_cleanup_commit_master((struct llog_commit_master *)
2284 obd->u.filter.fo_lcm, 1);
2285 OBD_FREE(obd->u.filter.fo_lcm,
2286 sizeof(struct llog_commit_master));
2287 obd->u.filter.fo_lcm = NULL;
2289 /* finish obd llog group */
2290 rc = filter_group_llog_finish(&obd->obd_olg);
2296 * Find the group llog according to group index in the llog group list.
2298 static struct obd_llog_group *
2299 filter_find_olg_internal(struct filter_obd *filter, int group)
2301 struct obd_llog_group *olg;
2303 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2304 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2305 if (olg->olg_group == group)
2312 * Find the group llog according to group index on the filter
2314 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2316 struct obd_llog_group *olg = NULL;
2317 struct filter_obd *filter;
2319 filter = &obd->u.filter;
2321 if (group == OBD_LLOG_GROUP)
2322 RETURN(&obd->obd_olg);
2324 spin_lock(&filter->fo_llog_list_lock);
2325 olg = filter_find_olg_internal(filter, group);
2326 spin_unlock(&filter->fo_llog_list_lock);
2331 * Find the llog_group of the filter according to the group. If it can not
2332 * find, create the llog_group, which only happens when mds is being synced
2335 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2337 struct obd_llog_group *olg = NULL;
2338 struct filter_obd *filter;
2341 filter = &obd->u.filter;
2343 if (group == OBD_LLOG_GROUP)
2344 RETURN(&obd->obd_olg);
2346 spin_lock(&filter->fo_llog_list_lock);
2347 olg = filter_find_olg_internal(filter, group);
2349 if (olg->olg_initializing) {
2350 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2352 GOTO(out_unlock, olg);
2357 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2359 llog_group_init(olg, group);
2360 list_add(&olg->olg_list, &filter->fo_llog_list);
2361 olg->olg_initializing = 1;
2362 spin_unlock(&filter->fo_llog_list_lock);
2364 rc = llog_cat_initialize(obd, olg, 1, NULL);
2366 spin_lock(&filter->fo_llog_list_lock);
2367 list_del(&olg->olg_list);
2368 spin_unlock(&filter->fo_llog_list_lock);
2370 GOTO(out, olg = ERR_PTR(-ENOMEM));
2372 spin_lock(&filter->fo_llog_list_lock);
2373 olg->olg_initializing = 0;
2374 spin_unlock(&filter->fo_llog_list_lock);
2375 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2376 obd->obd_name, group, olg);
2381 spin_unlock(&filter->fo_llog_list_lock);
2385 static int filter_llog_connect(struct obd_export *exp,
2386 struct llogd_conn_body *body)
2388 struct obd_device *obd = exp->exp_obd;
2389 struct llog_ctxt *ctxt;
2390 struct obd_llog_group *olg;
2394 CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2395 (unsigned) body->lgdc_logid.lgl_ogr,
2396 (unsigned) body->lgdc_logid.lgl_oid,
2397 (unsigned) body->lgdc_logid.lgl_ogen);
2399 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2401 CERROR(" %s: can not find olg of group %d\n",
2402 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2405 llog_group_set_export(olg, exp);
2407 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2408 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2409 body->lgdc_ctxt_idx);
2410 rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2411 &body->lgdc_gen, NULL);
2412 llog_ctxt_put(ctxt);
2414 CERROR("failed to connect rc %d idx %d\n", rc,
2415 body->lgdc_ctxt_idx);
2420 static int filter_llog_preclean (struct obd_device *obd)
2422 struct obd_llog_group *olg, *tmp;
2423 struct filter_obd *filter;
2424 struct list_head remove_list;
2428 rc = obd_llog_finish(obd, 0);
2430 CERROR("failed to cleanup llogging subsystem\n");
2432 filter = &obd->u.filter;
2433 CFS_INIT_LIST_HEAD(&remove_list);
2435 spin_lock(&filter->fo_llog_list_lock);
2436 while (!list_empty(&filter->fo_llog_list)) {
2437 olg = list_entry(filter->fo_llog_list.next,
2438 struct obd_llog_group, olg_list);
2439 list_del(&olg->olg_list);
2440 list_add(&olg->olg_list, &remove_list);
2442 spin_unlock(&filter->fo_llog_list_lock);
2444 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2445 list_del_init(&olg->olg_list);
2446 rc = filter_group_llog_finish(olg);
2448 CERROR("failed to cleanup llogging subsystem for %u\n",
2456 static int filter_precleanup(struct obd_device *obd,
2457 enum obd_cleanup_stage stage)
2463 case OBD_CLEANUP_EARLY:
2465 case OBD_CLEANUP_EXPORTS:
2466 target_cleanup_recovery(obd);
2467 rc = filter_llog_preclean(obd);
2473 static int filter_cleanup(struct obd_device *obd)
2475 struct filter_obd *filter = &obd->u.filter;
2479 LCONSOLE_WARN("%s: shutting down for failover; client state "
2480 "will be preserved.\n", obd->obd_name);
2482 if (!list_empty(&obd->obd_exports)) {
2483 CERROR("%s: still has clients!\n", obd->obd_name);
2484 class_disconnect_exports(obd);
2485 if (!list_empty(&obd->obd_exports)) {
2486 CERROR("still has exports after forced cleanup?\n");
2491 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2492 lprocfs_free_per_client_stats(obd);
2493 lprocfs_free_obd_stats(obd);
2494 lprocfs_obd_cleanup(obd);
2495 lquota_cleanup(filter_quota_interface_ref, obd);
2497 /* Stop recovery before namespace cleanup. */
2498 target_stop_recovery_thread(obd);
2499 target_cleanup_recovery(obd);
2501 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2502 obd->obd_namespace = NULL;
2504 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2506 if (obd->u.obt.obt_sb == NULL)
2511 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2512 shrink_dcache_sb(obd->u.obt.obt_sb);
2514 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2515 obd->u.obt.obt_sb = NULL;
2517 fsfilt_put_ops(obd->obd_fsops);
2519 filter_iobuf_pool_done(filter);
2521 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2526 static int filter_connect_internal(struct obd_export *exp,
2527 struct obd_connect_data *data)
2529 struct filter_export_data *fed = &exp->exp_filter_data;
2534 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2535 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2536 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2537 data->ocd_connect_flags, data->ocd_version,
2538 data->ocd_grant, data->ocd_index);
2540 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2541 CWARN("!!! This export (nid %s) used object group %d "
2542 "earlier; now it's trying to use group %d! This could "
2543 "be a bug in the MDS. Tell CFS.\n",
2544 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2547 fed->fed_group = data->ocd_group;
2549 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2550 exp->exp_connect_flags = data->ocd_connect_flags;
2551 data->ocd_version = LUSTRE_VERSION_CODE;
2553 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2554 obd_size left, want;
2556 spin_lock(&exp->exp_obd->obd_osfs_lock);
2557 left = filter_grant_space_left(exp);
2558 want = data->ocd_grant;
2559 filter_grant(exp, fed->fed_grant, want, left);
2560 data->ocd_grant = fed->fed_grant;
2561 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2563 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2564 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2565 exp->exp_client_uuid.uuid, exp,
2566 data->ocd_grant, want, left);
2569 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2570 struct filter_obd *filter = &exp->exp_obd->u.filter;
2571 struct lr_server_data *lsd = filter->fo_fsd;
2572 int index = le32_to_cpu(lsd->lsd_ost_index);
2574 if (!(lsd->lsd_feature_compat &
2575 cpu_to_le32(OBD_COMPAT_OST))) {
2576 /* this will only happen on the first connect */
2577 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2578 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2579 filter_update_server_data(exp->exp_obd,
2580 filter->fo_rcvd_filp, lsd, 1);
2581 } else if (index != data->ocd_index) {
2582 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2583 " %u doesn't match actual OST index"
2584 " %u in last_rcvd file, bad "
2586 obd_export_nid2str(exp), index,
2592 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2593 data->ocd_brw_size = 65536;
2594 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2595 data->ocd_brw_size = min(data->ocd_brw_size,
2596 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2597 LASSERT(data->ocd_brw_size);
2600 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2601 __u32 cksum_types = data->ocd_cksum_types;
2603 /* The client set in ocd_cksum_types the checksum types it
2604 * supports. We have to mask off the algorithms that we don't
2606 if (cksum_types & OBD_CKSUM_ALL)
2607 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2609 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2611 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2612 "%x\n", exp->exp_obd->obd_name,
2613 obd_export_nid2str(exp), cksum_types,
2614 data->ocd_cksum_types);
2616 /* This client does not support OBD_CONNECT_CKSUM
2617 * fall back to CRC32 */
2618 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2619 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2620 exp->exp_obd->obd_name,
2621 obd_export_nid2str(exp));
2624 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2625 * FIXME: We don't strictly need the COMPAT flag for that,
2626 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2627 * FIXME: We needed it for the index, as index 0 is valid. */
2632 static int filter_reconnect(const struct lu_env *env,
2633 struct obd_export *exp, struct obd_device *obd,
2634 struct obd_uuid *cluuid,
2635 struct obd_connect_data *data)
2640 if (exp == NULL || obd == NULL || cluuid == NULL)
2643 rc = filter_connect_internal(exp, data);
2648 /* nearly identical to mds_connect */
2649 static int filter_connect(const struct lu_env *env,
2650 struct lustre_handle *conn, struct obd_device *obd,
2651 struct obd_uuid *cluuid,
2652 struct obd_connect_data *data, void *localdata)
2654 struct lvfs_run_ctxt saved;
2655 struct obd_export *exp;
2656 struct filter_export_data *fed;
2657 struct lsd_client_data *lcd = NULL;
2662 if (conn == NULL || obd == NULL || cluuid == NULL)
2665 rc = class_connect(conn, obd, cluuid);
2668 exp = class_conn2export(conn);
2669 LASSERT(exp != NULL);
2671 fed = &exp->exp_filter_data;
2673 rc = filter_connect_internal(exp, data);
2677 filter_export_stats_init(obd, exp, localdata);
2678 if (obd->obd_replayable) {
2679 OBD_ALLOC(lcd, sizeof(*lcd));
2681 CERROR("filter: out of memory for client data\n");
2682 GOTO(cleanup, rc = -ENOMEM);
2685 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2687 rc = filter_client_add(obd, exp, -1);
2692 group = data->ocd_group;
2696 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2697 obd->obd_name, exp->exp_handle.h_cookie, group);
2699 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2700 rc = filter_read_groups(obd, group, 1);
2701 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2703 CERROR("can't read group %u\n", group);
2713 fed->fed_lcd = NULL;
2715 class_disconnect(exp);
2717 class_export_put(exp);
2723 /* Do extra sanity checks for grant accounting. We do this at connect,
2724 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2725 * always get rid of it or turn it off when we know accounting is good. */
2726 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2728 struct filter_export_data *fed;
2729 struct obd_export *exp;
2730 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2731 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2732 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2734 if (list_empty(&obd->obd_exports))
2737 /* We don't want to do this for large machines that do lots of
2738 mounts or unmounts. It burns... */
2739 if (obd->obd_num_exports > 100)
2742 spin_lock(&obd->obd_osfs_lock);
2743 spin_lock(&obd->obd_dev_lock);
2744 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2746 fed = &exp->exp_filter_data;
2747 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2750 if (maxsize > 0) { /* we may not have done a statfs yet */
2751 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2752 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2753 exp->exp_client_uuid.uuid, exp,
2754 fed->fed_grant, fed->fed_pending, maxsize);
2755 LASSERTF(fed->fed_dirty <= maxsize,
2756 "%s: cli %s/%p %ld > "LPU64"\n", func,
2757 exp->exp_client_uuid.uuid, exp,
2758 fed->fed_dirty, maxsize);
2761 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2762 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2763 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2765 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2766 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2767 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2768 tot_granted += fed->fed_grant + fed->fed_pending;
2769 tot_pending += fed->fed_pending;
2770 tot_dirty += fed->fed_dirty;
2772 fo_tot_granted = obd->u.filter.fo_tot_granted;
2773 fo_tot_pending = obd->u.filter.fo_tot_pending;
2774 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2775 spin_unlock(&obd->obd_dev_lock);
2776 spin_unlock(&obd->obd_osfs_lock);
2778 /* Do these assertions outside the spinlocks so we don't kill system */
2779 if (tot_granted != fo_tot_granted)
2780 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2781 func, tot_granted, fo_tot_granted);
2782 if (tot_pending != fo_tot_pending)
2783 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2784 func, tot_pending, fo_tot_pending);
2785 if (tot_dirty != fo_tot_dirty)
2786 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2787 func, tot_dirty, fo_tot_dirty);
2788 if (tot_pending > tot_granted)
2789 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2790 func, tot_pending, tot_granted);
2791 if (tot_granted > maxsize)
2792 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2793 func, tot_granted, maxsize);
2794 if (tot_dirty > maxsize)
2795 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2796 func, tot_dirty, maxsize);
2799 /* Remove this client from the grant accounting totals. We also remove
2800 * the export from the obd device under the osfs and dev locks to ensure
2801 * that the filter_grant_sanity_check() calculations are always valid.
2802 * The client should do something similar when it invalidates its import. */
2803 static void filter_grant_discard(struct obd_export *exp)
2805 struct obd_device *obd = exp->exp_obd;
2806 struct filter_obd *filter = &obd->u.filter;
2807 struct filter_export_data *fed = &exp->exp_filter_data;
2809 spin_lock(&obd->obd_osfs_lock);
2810 spin_lock(&obd->obd_dev_lock);
2811 list_del_init(&exp->exp_obd_chain);
2812 spin_unlock(&obd->obd_dev_lock);
2814 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2815 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2816 obd->obd_name, filter->fo_tot_granted,
2817 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2818 filter->fo_tot_granted -= fed->fed_grant;
2819 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2820 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2821 obd->obd_name, filter->fo_tot_pending,
2822 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2823 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2824 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2825 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2826 obd->obd_name, filter->fo_tot_dirty,
2827 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2828 filter->fo_tot_dirty -= fed->fed_dirty;
2832 spin_unlock(&obd->obd_osfs_lock);
2835 static int filter_destroy_export(struct obd_export *exp)
2839 if (exp->exp_filter_data.fed_pending)
2840 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2841 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2842 exp, exp->exp_filter_data.fed_pending);
2844 /* Not ported yet the b1_6 quota functionality
2845 * lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2848 target_destroy_export(exp);
2850 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2853 lprocfs_exp_cleanup(exp);
2855 if (exp->exp_obd->obd_replayable)
2856 filter_client_free(exp);
2858 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2860 filter_grant_discard(exp);
2861 filter_fmd_cleanup(exp);
2863 if (!(exp->exp_flags & OBD_OPT_FORCE))
2864 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2869 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2871 struct obd_llog_group *olg_min, *olg;
2872 struct filter_obd *filter;
2873 int worked = 0, group;
2874 struct llog_ctxt *ctxt;
2877 filter = &obd->u.filter;
2879 /* we can't sync log holding spinlock. also, we do not want to get
2880 * into livelock. so we do following: loop over MDS's exports in
2881 * group order and skip already synced llogs -bzzz */
2883 /* look for group with min. number, but > worked */
2886 spin_lock(&filter->fo_llog_list_lock);
2887 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2888 if (olg->olg_group <= worked) {
2889 /* this group is already synced */
2892 if (group < olg->olg_group) {
2893 /* we have group with smaller number to sync */
2896 /* store current minimal group */
2898 group = olg->olg_group;
2900 spin_unlock(&filter->fo_llog_list_lock);
2902 if (olg_min == NULL)
2905 worked = olg_min->olg_group;
2906 if (olg_min->olg_exp &&
2907 (dexp == olg_min->olg_exp || dexp == NULL)) {
2909 ctxt = llog_group_get_ctxt(olg_min,
2910 LLOG_MDS_OST_REPL_CTXT);
2911 LASSERT(ctxt != NULL);
2912 err = llog_sync(ctxt, olg_min->olg_exp);
2913 llog_ctxt_put(ctxt);
2915 CERROR("error flushing logs to MDS: rc %d\n",
2918 } while (olg_min != NULL);
2921 /* also incredibly similar to mds_disconnect */
2922 static int filter_disconnect(struct obd_export *exp)
2924 struct obd_device *obd = exp->exp_obd;
2929 class_export_get(exp);
2931 if (!(exp->exp_flags & OBD_OPT_FORCE))
2932 filter_grant_sanity_check(obd, __FUNCTION__);
2933 filter_grant_discard(exp);
2935 /* Disconnect early so that clients can't keep using export */
2936 rc = class_disconnect(exp);
2937 if (exp->exp_obd->obd_namespace != NULL)
2938 ldlm_cancel_locks_for_export(exp);
2940 fsfilt_sync(obd, obd->u.obt.obt_sb);
2942 /* flush any remaining cancel messages out to the target */
2943 filter_sync_llogs(obd, exp);
2944 class_export_put(exp);
2948 /* reverse import is changed, sync all cancels */
2949 static void filter_revimp_update(struct obd_export *exp)
2954 class_export_get(exp);
2956 /* flush any remaining cancel messages out to the target */
2957 filter_sync_llogs(exp->exp_obd, exp);
2958 class_export_put(exp);
2962 static int filter_ping(struct obd_export *exp)
2964 filter_fmd_expire(exp);
2969 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2970 const char *what, int quiet)
2972 struct dentry *dchild = NULL;
2975 if (oa->o_valid & OBD_MD_FLGROUP)
2978 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
2980 if (IS_ERR(dchild)) {
2981 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
2982 what, group, oa->o_id);
2986 if (dchild->d_inode == NULL) {
2988 CERROR("%s: %s on non-existent object: "LPU64"\n",
2989 obd->obd_name, what, oa->o_id);
2991 RETURN(ERR_PTR(-ENOENT));
2997 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2999 struct dentry *dentry = NULL;
3000 struct obd_device *obd;
3004 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3005 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3009 obd = class_exp2obd(exp);
3011 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3015 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3017 RETURN(PTR_ERR(dentry));
3019 /* Limit the valid bits in the return data to what we actually use */
3020 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3021 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3027 /* this should be enabled/disabled in condition to enabled/disabled large
3028 * inodes (fast EAs) in backing store FS. */
3029 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3030 void *handle, struct obdo *oa)
3032 struct obd_device *obd = exp->exp_obd;
3036 if (oa->o_valid & OBD_MD_FLFID) {
3037 struct filter_fid ff;
3039 if (!(oa->o_valid & OBD_MD_FLGROUP))
3041 /* packing fid and converting it to LE for storing into EA.
3042 * Here ->o_stripe_idx should be filled by LOV and rest of
3043 * fields - by client. */
3044 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3045 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3046 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3047 ff.ff_objid = cpu_to_le64(oa->o_id);
3048 ff.ff_group = cpu_to_le64(oa->o_gr);
3050 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3051 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3052 oa->o_generation, oa->o_id, oa->o_gr);
3054 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3056 CERROR("store fid in object failed! rc: %d\n", rc);
3058 CDEBUG(D_HA, "OSS object without fid info!\n");
3064 /* this is called from filter_truncate() until we have filter_punch() */
3065 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3066 struct obdo *oa, struct obd_trans_info *oti)
3068 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3069 struct llog_cookie *fcc = NULL;
3070 struct filter_obd *filter;
3071 int rc, err, locked = 0, sync = 0;
3072 unsigned int ia_valid;
3073 struct inode *inode;
3078 LASSERT(dentry != NULL);
3079 LASSERT(!IS_ERR(dentry));
3081 inode = dentry->d_inode;
3082 LASSERT(inode != NULL);
3084 filter = &exp->exp_obd->u.filter;
3085 iattr_from_obdo(&iattr, oa, oa->o_valid);
3086 ia_valid = iattr.ia_valid;
3088 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3089 OBD_ALLOC(fcc, sizeof(*fcc));
3091 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3094 if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
3096 LOCK_INODE_MUTEX(inode);
3100 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3101 * then we will accept the UID+GID sent by the client during write for
3102 * initializing the ownership of this inode. We only allow this to
3103 * happen once so clear these bits in setattr. In 2.6 kernels it is
3104 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3105 * the flags that are actually being set. */
3106 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3107 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3108 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3110 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3111 if (!(ia_valid & ATTR_MODE)) {
3112 iattr.ia_mode = inode->i_mode;
3113 iattr.ia_valid |= ATTR_MODE;
3115 iattr.ia_mode &= ~S_ISUID;
3117 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3118 if (!(iattr.ia_valid & ATTR_MODE)) {
3119 iattr.ia_mode = inode->i_mode;
3120 iattr.ia_valid |= ATTR_MODE;
3122 iattr.ia_mode &= ~S_ISGID;
3125 orig_ids[USRQUOTA] = inode->i_uid;
3126 orig_ids[GRPQUOTA] = inode->i_gid;
3127 handle = fsfilt_start_log(exp->exp_obd, inode,
3128 FSFILT_OP_SETATTR, oti, 1);
3130 GOTO(out_unlock, rc = PTR_ERR(handle));
3132 /* update inode EA only once when inode is suid bit marked. As
3133 * on 2.6.x UID and GID may be set separately, we check here
3134 * only one of them to avoid double setting. */
3135 if (inode->i_mode & S_ISUID)
3136 filter_update_fidea(exp, inode, handle, oa);
3138 handle = fsfilt_start(exp->exp_obd, inode,
3139 FSFILT_OP_SETATTR, oti);
3141 GOTO(out_unlock, rc = PTR_ERR(handle));
3143 if (oa->o_valid & OBD_MD_FLFLAGS) {
3144 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3145 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3147 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3149 /* set cancel cookie callback function */
3150 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3151 filter_cancel_cookies_cb,
3155 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3156 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3158 /* The truncate might have used up our transaction credits. Make
3159 * sure we have one left for the last_rcvd update. */
3160 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3162 rc = filter_finish_transno(exp, oti, rc, sync);
3164 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3168 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3170 CERROR("error on commit, err = %d\n", err);
3178 /* Let's flush truncated page on disk immediately, then we can
3179 * avoid need to search for page aliases before directio writes
3180 * and this sort of stuff at expense of somewhat slower
3181 * truncates not on a page boundary. I believe this is the only
3182 * place in filter code that can lead to pages getting to
3183 * pagecache so far. */
3184 filter_clear_truncated_page(inode);
3185 UNLOCK_INODE_MUTEX(inode);
3192 UNLOCK_INODE_MUTEX(inode);
3195 OBD_FREE(fcc, sizeof(*fcc));
3197 /* trigger quota release */
3198 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3199 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3200 int rc2 = lquota_adjust(filter_quota_interface_ref,
3201 exp->exp_obd, cur_ids,
3202 orig_ids, rc, FSFILT_OP_SETATTR);
3203 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3204 "filter adjust qunit. (rc:%d)\n", rc2);
3209 /* this is called from filter_truncate() until we have filter_punch() */
3210 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3211 struct obd_trans_info *oti)
3213 struct ldlm_res_id res_id;
3214 struct filter_mod_data *fmd;
3215 struct lvfs_run_ctxt saved;
3216 struct filter_obd *filter;
3217 struct ldlm_resource *res;
3218 struct dentry *dentry;
3222 osc_build_res_name(oinfo->oi_oa->o_id, oinfo->oi_oa->o_gr, &res_id);
3223 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3224 oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
3228 /* This would be very bad - accidentally truncating a file when
3229 * changing the time or similar - bug 12203. */
3230 if (oinfo->oi_oa->o_valid & OBD_MD_FLSIZE &&
3231 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3232 static char mdsinum[48];
3234 if (oinfo->oi_oa->o_valid & OBD_MD_FLFID)
3235 snprintf(mdsinum, sizeof(mdsinum) - 1,
3236 " of inode "LPU64"/%u", oinfo->oi_oa->o_fid,
3237 oinfo->oi_oa->o_generation);
3241 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3243 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3244 oinfo->oi_oa->o_id, mdsinum);
3248 dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
3251 RETURN(PTR_ERR(dentry));
3253 filter = &exp->exp_obd->u.filter;
3254 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3257 if (oinfo->oi_oa->o_valid &
3258 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3259 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
3260 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3261 fmd->fmd_mactime_xid = oti->oti_xid;
3262 filter_fmd_put(exp, fmd);
3265 /* setting objects attributes (including owner/group) */
3266 rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
3268 GOTO(out_unlock, rc);
3270 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3271 &res_id, LDLM_EXTENT, 0);
3274 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3275 ldlm_resource_putref(res);
3278 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3280 /* Quota release need uid/gid info */
3281 obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
3282 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3288 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3292 /* XXX identical to osc_unpackmd */
3293 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3294 struct lov_mds_md *lmm, int lmm_bytes)
3300 if (lmm_bytes < sizeof (*lmm)) {
3301 CERROR("lov_mds_md too small: %d, need %d\n",
3302 lmm_bytes, (int)sizeof(*lmm));
3305 /* XXX LOV_MAGIC etc check? */
3307 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3308 CERROR("lov_mds_md: zero lmm_object_id\n");
3313 lsm_size = lov_stripe_md_size(1);
3317 if (*lsmp != NULL && lmm == NULL) {
3318 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3319 OBD_FREE(*lsmp, lsm_size);
3324 if (*lsmp == NULL) {
3325 OBD_ALLOC(*lsmp, lsm_size);
3329 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3330 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3331 OBD_FREE(*lsmp, lsm_size);
3334 loi_init((*lsmp)->lsm_oinfo[0]);
3338 /* XXX zero *lsmp? */
3339 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3340 LASSERT((*lsmp)->lsm_object_id);
3343 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3348 /* caller must hold fo_create_locks[oa->o_gr] */
3349 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3350 struct filter_obd *filter)
3352 struct obdo doa; /* XXX obdo on stack */
3358 LASSERT(oa->o_gr != 0);
3359 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3360 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3362 memset(&doa, 0, sizeof(doa));
3364 doa.o_valid |= OBD_MD_FLGROUP;
3365 doa.o_gr = oa->o_gr;
3366 doa.o_mode = S_IFREG;
3368 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3369 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3370 exp->exp_obd->obd_name, doa.o_gr);
3374 last = filter_last_id(filter, doa.o_gr);
3376 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3377 exp->exp_obd->obd_name, oa->o_id + 1, last);
3379 for (id = last; id > oa->o_id; id--) {
3381 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
3382 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3383 CEMERG("error destroying precreate objid "LPU64": %d\n",
3385 filter_set_last_id(filter, id - 1, doa.o_gr);
3386 /* update last_id on disk periodically so that if we restart
3387 * we don't need to re-scan all of the just-deleted objects. */
3388 if ((id & 511) == 0)
3389 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3392 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3393 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3395 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3396 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3401 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3402 obd_gr group, int *num);
3403 /* returns a negative error or a nonnegative number of files to create */
3404 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3405 obd_gr group, struct obd_trans_info *oti)
3407 struct obd_device *obd = exp->exp_obd;
3408 struct filter_obd *filter = &obd->u.filter;
3412 /* delete orphans request */
3413 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3414 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3415 CERROR("%s: dropping old orphan cleanup request\n",
3419 /* This causes inflight precreates to abort and drop lock */
3420 set_bit(group, &filter->fo_destroys_in_progress);
3421 down(&filter->fo_create_locks[group]);
3422 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3423 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3424 exp->exp_obd->obd_name, group);
3425 up(&filter->fo_create_locks[group]);
3428 diff = oa->o_id - filter_last_id(filter, group);
3429 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3430 filter_last_id(filter, group), diff);
3432 if (-diff > OST_MAX_PRECREATE) {
3433 CERROR("%s: ignoring bogus orphan destroy request: "
3434 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3435 oa->o_id, filter_last_id(filter, group));
3436 /* FIXME: should reset precreate_next_id on MDS */
3437 GOTO(out, rc = -EINVAL);
3440 rc = filter_destroy_precreated(exp, oa, filter);
3442 CERROR("%s: unable to write lastobjid, but "
3443 "orphans were deleted\n", obd->obd_name);
3446 /* XXX: Used by MDS for the first time! */
3447 clear_bit(group, &filter->fo_destroys_in_progress);
3450 down(&filter->fo_create_locks[group]);
3451 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3452 CERROR("%s: dropping old precreate request\n",
3456 /* only precreate if group == 0 and o_id is specfied */
3457 if (group < FILTER_GROUP_MDS0 || oa->o_id == 0)
3460 diff = oa->o_id - filter_last_id(filter, group);
3461 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3462 filter_last_id(filter, group), diff);
3464 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3465 oa->o_id, filter_last_id(filter, group), diff);
3469 oa->o_id = filter_last_id(&obd->u.filter, group);
3470 rc = filter_precreate(obd, oa, group, &diff);
3471 oa->o_id = filter_last_id(&obd->u.filter, group);
3473 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3476 /* else diff == 0 */
3479 up(&filter->fo_create_locks[group]);
3483 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3484 __u64 max_age, __u32 flags)
3486 struct filter_obd *filter = &obd->u.filter;
3487 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3491 /* at least try to account for cached pages. its still racey and
3492 * might be under-reporting if clients haven't announced their
3493 * caches with brw recently */
3494 spin_lock(&obd->obd_osfs_lock);
3495 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3496 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3497 spin_unlock(&obd->obd_osfs_lock);
3499 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3500 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3501 filter->fo_tot_dirty, filter->fo_tot_granted,
3502 filter->fo_tot_pending,
3503 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3505 filter_grant_sanity_check(obd, __FUNCTION__);
3507 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3508 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3509 osfs->os_bsize - 1) >> blockbits));
3511 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3512 * stop creating files on MDS if OST is not good shape to create
3514 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3519 static int filter_use_existing_obj(struct obd_device *obd,
3520 struct dentry *dchild, void **handle,
3523 struct inode *inode = dchild->d_inode;
3527 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3530 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3531 if (IS_ERR(*handle))
3532 return PTR_ERR(*handle);
3534 iattr.ia_valid = ATTR_MODE;
3535 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3536 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3544 /* We rely on the fact that only one thread will be creating files in a given
3545 * group at a time, which is why we don't need an atomic filter_get_new_id.
3546 * Even if we had that atomic function, the following race would exist:
3548 * thread 1: gets id x from filter_next_id
3549 * thread 2: gets id (x + 1) from filter_next_id
3550 * thread 2: creates object (x + 1)
3551 * thread 1: tries to create object x, gets -ENOSPC
3553 * Caller must hold fo_create_locks[group]
3555 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3556 obd_gr group, int *num)
3558 struct dentry *dchild = NULL, *dparent = NULL;
3559 struct filter_obd *filter;
3560 struct obd_statfs *osfs;
3561 int err = 0, rc = 0, recreate_obj = 0, i;
3562 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3564 void *handle = NULL;
3567 filter = &obd->u.filter;
3569 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3571 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3573 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3574 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3577 OBD_ALLOC(osfs, sizeof(*osfs));
3580 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3581 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3582 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3583 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3584 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3588 OBD_FREE(osfs, sizeof(*osfs));
3593 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3594 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3596 for (i = 0; i < *num && err == 0; i++) {
3597 int cleanup_phase = 0;
3599 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3600 CWARN("%s: create aborted by destroy\n",
3609 last_id = filter_last_id(filter, group);
3610 if (next_id > last_id) {
3611 CERROR("Error: Trying to recreate obj greater"
3612 "than last id "LPD64" > "LPD64"\n",
3614 GOTO(cleanup, rc = -EINVAL);
3617 next_id = filter_last_id(filter, group) + 1;
3619 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
3621 dparent = filter_parent_lock(obd, group, next_id);
3622 if (IS_ERR(dparent))
3623 GOTO(cleanup, rc = PTR_ERR(dparent));
3624 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3626 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3628 GOTO(cleanup, rc = PTR_ERR(dchild));
3629 cleanup_phase = 2; /* f_dput(dchild) */
3631 if (dchild->d_inode != NULL) {
3632 /* This would only happen if lastobjid was bad on disk*/
3633 /* Could also happen if recreating missing obj but it
3634 * already exists. */
3636 CERROR("%s: recreating existing object %.*s?\n",
3637 obd->obd_name, dchild->d_name.len,
3638 dchild->d_name.name);
3640 /* Use these existing objects if they are
3642 if (dchild->d_inode->i_size == 0) {
3643 rc = filter_use_existing_obj(obd,dchild,
3644 &handle, &cleanup_phase);
3651 CERROR("%s: Serious error: objid %.*s already "
3652 "exists; is this filesystem corrupt?\n",
3653 obd->obd_name, dchild->d_name.len,
3654 dchild->d_name.name);
3657 GOTO(cleanup, rc = -EEXIST);
3660 handle = fsfilt_start_log(obd, dparent->d_inode,
3661 FSFILT_OP_CREATE, NULL, 1);
3663 GOTO(cleanup, rc = PTR_ERR(handle));
3666 /* We mark object SUID+SGID to flag it for accepting UID+GID
3667 * from client on first write. Currently the permission bits
3668 * on the OST are never used, so this is OK. */
3669 rc = ll_vfs_create(dparent->d_inode, dchild,
3670 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3672 CERROR("create failed rc = %d\n", rc);
3677 if (!recreate_obj) {
3678 filter_set_last_id(filter, next_id, group);
3679 err = filter_update_last_objid(obd, group, 0);
3681 CERROR("unable to write lastobjid "
3682 "but file created\n");
3686 switch(cleanup_phase) {
3688 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3690 CERROR("error on commit, err = %d\n", err);
3697 filter_parent_unlock(dparent);
3704 if (time_after(jiffies, enough_time)) {
3706 "%s: precreate slow - want %d got %d \n",
3707 obd->obd_name, *num, i);
3714 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3715 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3720 static int filter_create(struct obd_export *exp, struct obdo *oa,
3721 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3723 struct filter_export_data *fed;
3724 struct obd_device *obd = NULL;
3725 struct filter_obd *filter;
3726 struct lvfs_run_ctxt saved;
3727 struct lov_stripe_md *lsm = NULL;
3728 int rc = 0, diff, group = oa->o_gr;
3731 if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
3732 CERROR("!!! nid %s sent invalid object group %d\n",
3733 obd_export_nid2str(exp), group);
3738 fed = &exp->exp_filter_data;
3739 filter = &obd->u.filter;
3741 if (fed->fed_group != group) {
3742 CERROR("!!! this export (nid %s) used object group %d "
3743 "earlier; now it's trying to use group %d! This could "
3744 "be a bug in the MDS. Tell CFS.\n",
3745 obd_export_nid2str(exp), fed->fed_group, group);
3749 CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
3750 oa->o_gr, oa->o_id);
3754 rc = obd_alloc_memmd(exp, &lsm);
3761 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3763 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3764 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3765 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3766 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3767 oa->o_id, filter_last_id(filter,
3772 down(&filter->fo_create_locks[oa->o_gr]);
3773 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3774 up(&filter->fo_create_locks[oa->o_gr]);
3777 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3780 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3781 if (rc && ea != NULL && *ea != lsm) {
3782 obd_free_memmd(exp, &lsm);
3783 } else if (rc == 0 && ea != NULL) {
3784 /* XXX LOV STACKING: the lsm that is passed to us from
3785 * LOV does not have valid lsm_oinfo data structs, so
3786 * don't go touching that. This needs to be fixed in a
3788 lsm->lsm_object_id = oa->o_id;
3795 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3796 struct lov_stripe_md *md, struct obd_trans_info *oti,
3797 struct obd_export *md_exp)
3799 unsigned int qcids[MAXQUOTAS] = {0, 0};
3800 struct obd_device *obd;
3801 struct filter_obd *filter;
3802 struct dentry *dchild = NULL, *dparent = NULL;
3803 struct lvfs_run_ctxt saved;
3804 void *handle = NULL;
3805 struct llog_cookie *fcc = NULL;
3806 int rc, rc2, cleanup_phase = 0, sync = 0;
3810 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3813 filter = &obd->u.filter;
3815 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3818 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3820 GOTO(cleanup, rc = PTR_ERR(dchild));
3823 if (dchild->d_inode == NULL) {
3824 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3826 /* If object already gone, cancel cookie right now */
3827 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3828 struct llog_ctxt *ctxt;
3829 struct obd_llog_group *olg;
3830 fcc = obdo_logcookie(oa);
3831 olg = filter_find_olg(obd, oa->o_gr);
3833 CERROR(" %s: can not find olg of group %d\n",
3834 obd->obd_name, (int)oa->o_gr);
3835 GOTO(cleanup, rc = PTR_ERR(olg));
3837 llog_group_set_export(olg, exp);
3839 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
3840 llog_cancel(ctxt, NULL, 1, fcc, 0);
3841 llog_ctxt_put(ctxt);
3842 fcc = NULL; /* we didn't allocate fcc, don't free it */
3844 GOTO(cleanup, rc = -ENOENT);
3847 filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3849 /* Our MDC connection is established by the MDS to us */
3850 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3851 OBD_ALLOC(fcc, sizeof(*fcc));
3853 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3855 DQUOT_INIT(dchild->d_inode);
3857 /* we're gonna truncate it first in order to avoid possible deadlock:
3859 * open trasaction open transaction
3860 * down(i_zombie) down(i_zombie)
3861 * restart transaction
3862 * (see BUG 4180) -bzzz
3864 LOCK_INODE_MUTEX(dchild->d_inode);
3865 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3867 if (IS_ERR(handle)) {
3868 UNLOCK_INODE_MUTEX(dchild->d_inode);
3869 GOTO(cleanup, rc = PTR_ERR(handle));
3872 iattr.ia_valid = ATTR_SIZE;
3874 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3875 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3876 UNLOCK_INODE_MUTEX(dchild->d_inode);
3880 GOTO(cleanup, rc = rc2);
3882 /* We don't actually need to lock the parent until we are unlinking
3883 * here, and not while truncating above. That avoids holding the
3884 * parent lock for a long time during truncate, which can block other
3885 * threads from doing anything to objects in that directory. bug 7171 */
3886 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3887 if (IS_ERR(dparent))
3888 GOTO(cleanup, rc = PTR_ERR(dparent));
3889 cleanup_phase = 3; /* filter_parent_unlock */
3891 LOCK_INODE_MUTEX(dchild->d_inode);
3892 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3893 if (IS_ERR(handle)) {
3894 UNLOCK_INODE_MUTEX(dchild->d_inode);
3895 GOTO(cleanup, rc = PTR_ERR(handle));
3897 cleanup_phase = 4; /* fsfilt_commit */
3899 /* Quota release need uid/gid of inode */
3900 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3902 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3904 /* this drops dchild->d_inode->i_mutex unconditionally */
3905 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3909 switch(cleanup_phase) {
3912 sync = fsfilt_add_journal_cb(obd, 0, oti ?
3913 oti->oti_handle : handle,
3914 filter_cancel_cookies_cb,
3916 /* If add_journal_cb failed, then filter_finish_transno
3917 * will commit the handle and we will do a sync
3918 * on commit. then we call callback directly to free
3921 rc = filter_finish_transno(exp, oti, rc, sync);
3923 filter_cancel_cookies_cb(obd, 0, fcc, rc);
3926 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3928 CERROR("error on commit, err = %d\n", rc2);
3935 filter_parent_unlock(dparent);
3939 OBD_FREE(fcc, sizeof(*fcc));
3941 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3944 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3948 /* trigger quota release */
3949 qcids[USRQUOTA] = oa->o_uid;
3950 qcids[GRPQUOTA] = oa->o_gid;
3951 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
3955 CDEBUG(D_QUOTA, "filter adjust qunit! (rc:%d)\n", rc2);
3959 /* NB start and end are used for punch, but not truncate */
3960 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3961 struct obd_trans_info *oti,
3962 struct ptlrpc_request_set *rqset)
3967 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3968 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
3969 oinfo->oi_policy.l_extent.end);
3973 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
3974 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
3975 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
3977 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3978 oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
3982 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
3983 rc = filter_setattr(exp, oinfo, oti);
3987 static int filter_sync(struct obd_export *exp, struct obdo *oa,
3988 struct lov_stripe_md *lsm, obd_off start, obd_off end,
3991 struct lvfs_run_ctxt saved;
3992 struct filter_obd *filter;
3993 struct dentry *dentry;
3997 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
3998 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4002 filter = &exp->exp_obd->u.filter;
4004 /* an objid of zero is taken to mean "sync whole filesystem" */
4005 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4006 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4007 /* flush any remaining cancel messages out to the target */
4008 filter_sync_llogs(exp->exp_obd, exp);
4012 dentry = filter_oa2dentry(exp->exp_obd, oa);
4014 RETURN(PTR_ERR(dentry));
4016 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4018 LOCK_INODE_MUTEX(dentry->d_inode);
4020 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4022 /* just any file to grab fsync method - "file" arg unused */
4023 struct file *file = filter->fo_rcvd_filp;
4025 if (file->f_op && file->f_op->fsync)
4026 rc = file->f_op->fsync(NULL, dentry, 1);
4028 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4032 UNLOCK_INODE_MUTEX(dentry->d_inode);
4034 oa->o_valid = OBD_MD_FLID;
4035 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4037 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4043 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4044 void *key, __u32 *vallen, void *val)
4046 struct obd_device *obd;
4049 obd = class_exp2obd(exp);
4051 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4055 if (KEY_IS(KEY_BLOCKSIZE)) {
4056 __u32 *blocksize = val;
4058 if (*vallen < sizeof(*blocksize))
4060 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4062 *vallen = sizeof(*blocksize);
4066 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4067 __u32 *blocksize_bits = val;
4068 if (blocksize_bits) {
4069 if (*vallen < sizeof(*blocksize_bits))
4071 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4073 *vallen = sizeof(*blocksize_bits);
4077 if (KEY_IS(KEY_LAST_ID)) {
4078 obd_id *last_id = val;
4079 /* FIXME: object groups */
4081 if (*vallen < sizeof(*last_id))
4083 *last_id = filter_last_id(&obd->u.filter,
4084 exp->exp_filter_data.fed_group);
4086 *vallen = sizeof(*last_id);
4090 CDEBUG(D_IOCTL, "invalid key\n");
4094 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4095 void *key, __u32 vallen, void *val,
4096 struct ptlrpc_request_set *set)
4098 struct obd_device *obd;
4099 struct obd_llog_group *olg;
4100 struct llog_ctxt *ctxt;
4106 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4110 if (KEY_IS(KEY_CAPA_KEY)) {
4111 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4113 CERROR("filter update capability key failed: %d\n", rc);
4117 if (KEY_IS(KEY_REVIMP_UPD)) {
4118 filter_revimp_update(exp);
4122 if (!KEY_IS(KEY_MDS_CONN))
4125 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4126 obd_export_nid2str(exp));
4127 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4129 /* setup llog imports */
4130 LASSERT(val != NULL);
4131 group = (int)(*(__u32 *)val);
4132 LASSERT(group >= FILTER_GROUP_MDS0);
4134 olg = filter_find_create_olg(obd, group);
4136 RETURN(PTR_ERR(olg));
4138 llog_group_set_export(olg, exp);
4140 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4141 LASSERTF(ctxt != NULL, "ctxt is null\n"),
4143 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4144 llog_ctxt_put(ctxt);
4146 lquota_setinfo(filter_quota_interface_ref, exp, obd);
4151 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4152 int len, void *karg, void *uarg)
4154 struct obd_device *obd = exp->exp_obd;
4155 struct obd_ioctl_data *data = karg;
4159 case OBD_IOC_ABORT_RECOVERY: {
4160 CERROR("aborting recovery for device %s\n", obd->obd_name);
4161 target_stop_recovery_thread(obd);
4165 case OBD_IOC_SYNC: {
4166 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4167 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4171 case OBD_IOC_SET_READONLY: {
4173 struct super_block *sb = obd->u.obt.obt_sb;
4174 struct inode *inode = sb->s_root->d_inode;
4175 BDEVNAME_DECLARE_STORAGE(tmp);
4176 CERROR("*** setting device %s read-only ***\n",
4177 ll_bdevname(sb, tmp));
4179 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4180 if (!IS_ERR(handle))
4181 rc = fsfilt_commit(obd, inode, handle, 1);
4183 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4184 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4186 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4190 case OBD_IOC_CATLOGLIST: {
4191 rc = llog_catalog_list(obd, 1, data);
4195 case OBD_IOC_LLOG_CANCEL:
4196 case OBD_IOC_LLOG_REMOVE:
4197 case OBD_IOC_LLOG_INFO:
4198 case OBD_IOC_LLOG_PRINT: {
4199 /* FIXME to be finished */
4200 RETURN(-EOPNOTSUPP);
4202 struct llog_ctxt *ctxt = NULL;
4204 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4205 rc = llog_ioctl(ctxt, cmd, data);
4206 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4219 static int filter_health_check(struct obd_device *obd)
4221 #ifdef USE_HEALTH_CHECK_WRITE
4222 struct filter_obd *filter = &obd->u.filter;
4227 * health_check to return 0 on healthy
4228 * and 1 on unhealthy.
4230 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4233 #ifdef USE_HEALTH_CHECK_WRITE
4234 LASSERT(filter->fo_health_check_filp != NULL);
4235 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4240 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4243 return filter_fid2dentry(data, NULL, gr, id);
4246 static int filter_process_config(struct obd_device *obd, obd_count len,
4249 struct lustre_cfg *lcfg = buf;
4250 struct lprocfs_static_vars lvars;
4253 switch (lcfg->lcfg_command) {
4254 case LCFG_SPTLRPC_CONF: {
4255 struct filter_obd *filter = &obd->u.filter;
4256 struct sptlrpc_conf_log *log;
4257 struct sptlrpc_rule_set tmp_rset;
4259 log = sptlrpc_conf_log_extract(lcfg);
4265 sptlrpc_rule_set_init(&tmp_rset);
4267 rc = sptlrpc_rule_set_from_log(&tmp_rset, log);
4269 CERROR("obd %s: failed get sptlrpc rules: %d\n",
4274 write_lock(&filter->fo_sptlrpc_lock);
4275 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
4276 filter->fo_sptlrpc_rset = tmp_rset;
4277 write_unlock(&filter->fo_sptlrpc_lock);
4279 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
4283 lprocfs_filter_init_vars(&lvars);
4285 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4293 static struct lvfs_callback_ops filter_lvfs_ops = {
4294 l_fid2dentry: filter_lvfs_fid2dentry,
4297 static struct obd_ops filter_obd_ops = {
4298 .o_owner = THIS_MODULE,
4299 .o_get_info = filter_get_info,
4300 .o_set_info_async = filter_set_info_async,
4301 .o_setup = filter_setup,
4302 .o_precleanup = filter_precleanup,
4303 .o_cleanup = filter_cleanup,
4304 .o_connect = filter_connect,
4305 .o_reconnect = filter_reconnect,
4306 .o_disconnect = filter_disconnect,
4307 .o_ping = filter_ping,
4308 .o_init_export = filter_init_export,
4309 .o_destroy_export = filter_destroy_export,
4310 .o_statfs = filter_statfs,
4311 .o_getattr = filter_getattr,
4312 .o_unpackmd = filter_unpackmd,
4313 .o_create = filter_create,
4314 .o_setattr = filter_setattr,
4315 .o_destroy = filter_destroy,
4316 .o_brw = filter_brw,
4317 .o_punch = filter_truncate,
4318 .o_sync = filter_sync,
4319 .o_preprw = filter_preprw,
4320 .o_commitrw = filter_commitrw,
4321 .o_llog_init = filter_llog_init,
4322 .o_llog_connect = filter_llog_connect,
4323 .o_llog_finish = filter_llog_finish,
4324 .o_iocontrol = filter_iocontrol,
4325 .o_health_check = filter_health_check,
4326 .o_process_config = filter_process_config,
4329 quota_interface_t *filter_quota_interface_ref;
4330 extern quota_interface_t filter_quota_interface;
4332 static int __init obdfilter_init(void)
4334 struct lprocfs_static_vars lvars;
4337 lprocfs_filter_init_vars(&lvars);
4339 request_module("lquota");
4340 OBD_ALLOC(obdfilter_created_scratchpad,
4341 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4342 sizeof(*obdfilter_created_scratchpad));
4343 if (obdfilter_created_scratchpad == NULL)
4346 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4347 sizeof(struct filter_mod_data),
4350 GOTO(out, rc = -ENOMEM);
4352 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4353 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4355 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4356 LUSTRE_OST_NAME, NULL);
4360 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4361 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4362 ll_fmd_cachep = NULL;
4364 if (filter_quota_interface_ref)
4365 PORTAL_SYMBOL_PUT(filter_quota_interface);
4367 OBD_FREE(obdfilter_created_scratchpad,
4368 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4369 sizeof(*obdfilter_created_scratchpad));
4375 static void __exit obdfilter_exit(void)
4377 if (filter_quota_interface_ref)
4378 PORTAL_SYMBOL_PUT(filter_quota_interface);
4380 if (ll_fmd_cachep) {
4381 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4382 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4383 ll_fmd_cachep = NULL;
4386 class_unregister_type(LUSTRE_OST_NAME);
4387 OBD_FREE(obdfilter_created_scratchpad,
4388 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4389 sizeof(*obdfilter_created_scratchpad));
4392 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4393 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4394 MODULE_LICENSE("GPL");
4396 module_init(obdfilter_init);
4397 module_exit(obdfilter_exit);