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 = { .name = { objid, 0, group, 0} };
1501 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1504 /* Tell the clients that the object is gone now and that they should
1505 * throw away any cached pages. */
1506 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1507 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1508 ldlm_completion_ast, NULL, NULL, 0, NULL,
1511 /* We only care about the side-effects, just drop the lock. */
1513 ldlm_lock_decref(&lockh, LCK_PW);
1518 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1519 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1520 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1521 * reverse. Caller must take i_sem before starting the transaction and we
1522 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1523 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry)
1528 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1529 * itself we already hold dir->i_mutex for child create/unlink ops */
1530 LASSERT(dentry->d_inode != NULL);
1531 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1532 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1536 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1537 GOTO(out, rc = -ENOENT);
1539 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1544 GOTO(out, rc = -EPERM);
1546 /* check_sticky() */
1547 if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1548 IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1549 GOTO(out, rc = -EPERM);
1551 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1552 * that was done because of journal_start (which is already done
1553 * here) or some other ordering issue. */
1556 rc = security_inode_unlink(dir, dentry);
1560 rc = dir->i_op->unlink(dir, dentry);
1562 /* need to drop i_mutex before we lose inode reference */
1563 UNLOCK_INODE_MUTEX(dentry->d_inode);
1570 /* Caller must hold LCK_PW on parent and push us into kernel context.
1571 * Caller must hold child i_mutex, we drop it always.
1572 * Caller is also required to ensure that dchild->d_inode exists. */
1573 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1574 obd_gr group, struct dentry *dparent,
1575 struct dentry *dchild)
1577 struct inode *inode = dchild->d_inode;
1580 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1581 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1582 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1583 (unsigned long)inode->i_nlink,
1584 atomic_read(&inode->i_count));
1587 rc = filter_vfs_unlink(dparent->d_inode, dchild);
1589 CERROR("error unlinking objid %.*s: rc %d\n",
1590 dchild->d_name.len, dchild->d_name.name, rc);
1594 struct filter_intent_args {
1595 struct ldlm_lock **victim;
1600 static enum interval_iter filter_intent_cb(struct interval_node *n,
1603 struct ldlm_interval *node = (struct ldlm_interval *)n;
1604 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1605 __u64 size = arg->size;
1606 struct ldlm_lock **v = arg->victim;
1607 struct ldlm_lock *lck;
1609 /* If the interval is lower than the current file size,
1611 if (interval_high(n) <= size)
1612 return INTERVAL_ITER_STOP;
1614 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1615 /* Don't send glimpse ASTs to liblustre clients.
1616 * They aren't listening for them, and they do
1617 * entirely synchronous I/O anyways. */
1618 if (lck->l_export == NULL ||
1619 lck->l_export->exp_libclient == 1)
1622 if (*arg->liblustre)
1623 *arg->liblustre = 0;
1626 *v = LDLM_LOCK_GET(lck);
1627 } else if ((*v)->l_policy_data.l_extent.start <
1628 lck->l_policy_data.l_extent.start) {
1630 *v = LDLM_LOCK_GET(lck);
1633 /* the same policy group - every lock has the
1634 * same extent, so needn't do it any more */
1638 return INTERVAL_ITER_CONT;
1641 static int filter_intent_policy(struct ldlm_namespace *ns,
1642 struct ldlm_lock **lockp, void *req_cookie,
1643 ldlm_mode_t mode, int flags, void *data)
1645 CFS_LIST_HEAD(rpc_list);
1646 struct ptlrpc_request *req = req_cookie;
1647 struct ldlm_lock *lock = *lockp, *l = NULL;
1648 struct ldlm_resource *res = lock->l_resource;
1649 ldlm_processing_policy policy;
1650 struct ost_lvb *res_lvb, *reply_lvb;
1651 struct ldlm_reply *rep;
1653 int idx, rc, tmpflags = 0, only_liblustre = 1;
1654 struct ldlm_interval_tree *tree;
1655 struct filter_intent_args arg;
1656 int repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1657 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1658 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1661 policy = ldlm_get_processing_policy(res);
1662 LASSERT(policy != NULL);
1663 LASSERT(req != NULL);
1665 rc = lustre_pack_reply(req, 3, repsize, NULL);
1667 RETURN(req->rq_status = rc);
1669 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1670 LASSERT(rep != NULL);
1672 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1673 sizeof(*reply_lvb));
1674 LASSERT(reply_lvb != NULL);
1676 //fixup_handle_for_resent_req(req, lock, &lockh);
1678 /* If we grant any lock at all, it will be a whole-file read lock.
1679 * Call the extent policy function to see if our request can be
1680 * granted, or is blocked.
1681 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse lock
1683 lock->l_policy_data.l_extent.start = 0;
1684 lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1685 lock->l_req_mode = LCK_PR;
1687 LASSERT(ns == res->lr_namespace);
1689 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1690 check_res_locked(res);
1692 /* FIXME: we should change the policy function slightly, to not make
1693 * this list at all, since we just turn around and free it */
1694 while (!list_empty(&rpc_list)) {
1695 struct ldlm_lock *wlock =
1696 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1697 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1698 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1699 lock->l_flags &= ~LDLM_FL_CP_REQD;
1700 list_del_init(&wlock->l_cp_ast);
1701 LDLM_LOCK_PUT(wlock);
1704 /* The lock met with no resistance; we're finished. */
1705 if (rc == LDLM_ITER_CONTINUE) {
1706 /* do not grant locks to the liblustre clients: they cannot
1707 * handle ASTs robustly. We need to do this while still
1708 * holding ns_lock to avoid the lock remaining on the res_link
1709 * list (and potentially being added to l_pending_list by an
1710 * AST) when we are going to drop this lock ASAP. */
1711 if (lock->l_export->exp_libclient ||
1712 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1713 ldlm_resource_unlink_lock(lock);
1714 err = ELDLM_LOCK_ABORTED;
1716 err = ELDLM_LOCK_REPLACED;
1722 /* Do not grant any lock, but instead send GL callbacks. The extent
1723 * policy nicely created a list of all PW locks for us. We will choose
1724 * the highest of those which are larger than the size in the LVB, if
1725 * any, and perform a glimpse callback. */
1726 res_lvb = res->lr_lvb_data;
1727 LASSERT(res_lvb != NULL);
1728 *reply_lvb = *res_lvb;
1731 * ->ns_lock guarantees that no new locks are granted, and,
1732 * therefore, that res->lr_lvb_data cannot increase beyond the
1733 * end of already granted lock. As a result, it is safe to
1734 * check against "stale" reply_lvb->lvb_size value without
1737 arg.size = reply_lvb->lvb_size;
1739 arg.liblustre = &only_liblustre;
1740 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1741 tree = &res->lr_itree[idx];
1742 if (tree->lit_mode == LCK_PR)
1745 interval_iterate_reverse(tree->lit_root,
1746 filter_intent_cb, &arg);
1750 /* There were no PW locks beyond the size in the LVB; finished. */
1752 if (only_liblustre) {
1753 /* If we discovered a liblustre client with a PW lock,
1754 * however, the LVB may be out of date! The LVB is
1755 * updated only on glimpse (which we don't do for
1756 * liblustre clients) and cancel (which the client
1757 * obviously has not yet done). So if it has written
1758 * data but kept the lock, the LVB is stale and needs
1759 * to be updated from disk.
1761 * Of course, this will all disappear when we switch to
1762 * taking liblustre locks on the OST. */
1763 ldlm_res_lvbo_update(res, NULL, 0, 1);
1765 RETURN(ELDLM_LOCK_ABORTED);
1769 * This check is for lock taken in filter_prepare_destroy() that does
1770 * not have l_glimpse_ast set. So the logic is: if there is a lock
1771 * with no l_glimpse_ast set, this object is being destroyed already.
1773 * Hence, if you are grabbing DLM locks on the server, always set
1774 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1776 if (l->l_glimpse_ast == NULL) {
1777 /* We are racing with unlink(); just return -ENOENT */
1778 rep->lock_policy_res1 = -ENOENT;
1782 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1783 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1784 /* Update the LVB from disk if the AST failed (this is a legal race) */
1786 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1787 * sending ast is not handled. This can result in lost client writes.
1790 ldlm_res_lvbo_update(res, NULL, 0, 1);
1793 *reply_lvb = *res_lvb;
1799 RETURN(ELDLM_LOCK_ABORTED);
1803 * per-obd_device iobuf pool.
1805 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1806 * allocations in write-path has to be minimized (see bug 5137).
1808 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1809 * OST threads (see ost_thread_{init,done}()).
1811 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1812 * at the OST layer there are only (potentially) multiple obd_device of type
1813 * unknown at the time of OST thread creation.
1815 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1816 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1817 * it's very own iobuf.
1821 * filter_kiobuf_pool_init()
1823 * filter_kiobuf_pool_done()
1825 * filter_iobuf_get()
1827 * operate on this array. They are "generic" in a sense that they don't depend
1828 * on actual type of iobuf's (the latter depending on Linux kernel version).
1832 * destroy pool created by filter_iobuf_pool_init
1834 static void filter_iobuf_pool_done(struct filter_obd *filter)
1836 struct filter_iobuf **pool;
1841 pool = filter->fo_iobuf_pool;
1843 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1844 if (pool[i] != NULL)
1845 filter_free_iobuf(pool[i]);
1847 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1848 filter->fo_iobuf_pool = NULL;
1854 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1856 static int filter_iobuf_pool_init(struct filter_obd *filter)
1863 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1865 if (filter->fo_iobuf_pool == NULL)
1868 filter->fo_iobuf_count = OSS_THREADS_MAX;
1873 /* Return iobuf allocated for @thread_id. We don't know in advance how
1874 * many threads there will be so we allocate a large empty array and only
1875 * fill in those slots that are actually in use.
1876 * If we haven't allocated a pool entry for this thread before, do so now. */
1877 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1879 int thread_id = oti ? oti->oti_thread_id : -1;
1880 struct filter_iobuf *pool = NULL;
1881 struct filter_iobuf **pool_place = NULL;
1883 if (thread_id >= 0) {
1884 LASSERT(thread_id < filter->fo_iobuf_count);
1885 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1888 if (unlikely(pool == NULL)) {
1889 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1890 PTLRPC_MAX_BRW_PAGES);
1891 if (pool_place != NULL)
1898 /* mount the file system (secretly). lustre_cfg parameters are:
1901 * 3 = flags: failover=f, failout=n
1904 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1907 struct filter_obd *filter = &obd->u.filter;
1908 struct vfsmount *mnt;
1909 struct lustre_mount_info *lmi;
1910 struct obd_uuid uuid;
1917 if (lcfg->lcfg_bufcount < 3 ||
1918 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1919 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1922 lmi = server_get_mount(obd->obd_name);
1924 /* We already mounted in lustre_fill_super.
1925 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1926 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1928 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1930 /* old path - used by lctl */
1931 CERROR("Using old MDS mount method\n");
1932 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1933 MS_NOATIME|MS_NODIRATIME,
1934 lustre_cfg_string(lcfg, 1), option);
1937 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1938 lustre_cfg_string(lcfg, 1), rc);
1942 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1944 if (IS_ERR(obd->obd_fsops))
1945 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1947 rc = filter_iobuf_pool_init(filter);
1951 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1953 /* failover is the default */
1954 obd->obd_replayable = 1;
1956 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1957 str = lustre_cfg_string(lcfg, 3);
1958 if (strchr(str, 'n')) {
1959 CWARN("%s: recovery disabled\n", obd->obd_name);
1960 obd->obd_replayable = 0;
1964 filter->fo_vfsmnt = mnt;
1965 obd->u.obt.obt_sb = mnt->mnt_sb;
1966 filter->fo_fstype = mnt->mnt_sb->s_type->name;
1967 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1969 fsfilt_setup(obd, obd->u.obt.obt_sb);
1971 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1972 obd->obd_lvfs_ctxt.pwdmnt = mnt;
1973 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1974 obd->obd_lvfs_ctxt.fs = get_ds();
1975 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
1977 init_mutex(&filter->fo_init_lock);
1978 filter->fo_committed_group = 0;
1980 rc = filter_prep(obd);
1984 filter->fo_destroys_in_progress = 0;
1985 for (i = 0; i < 32; i++)
1986 sema_init(&filter->fo_create_locks[i], 1);
1988 spin_lock_init(&filter->fo_translock);
1989 spin_lock_init(&filter->fo_objidlock);
1990 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
1991 sema_init(&filter->fo_alloc_lock, 1);
1992 init_brw_stats(&filter->fo_filter_stats);
1993 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1994 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
1995 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
1997 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
1998 spin_lock_init(&filter->fo_llog_list_lock);
2000 filter->fo_sptlrpc_lock = RW_LOCK_UNLOCKED;
2001 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2003 filter->fo_fl_oss_capa = 0;
2004 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2005 filter->fo_capa_hash = init_capa_hash();
2006 if (filter->fo_capa_hash == NULL)
2007 GOTO(err_ops, rc = -ENOMEM);
2009 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2010 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2011 LDLM_NAMESPACE_GREEDY);
2012 if (obd->obd_namespace == NULL)
2013 GOTO(err_post, rc = -ENOMEM);
2014 obd->obd_namespace->ns_lvbp = obd;
2015 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2016 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2018 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2019 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2021 rc = llog_cat_initialize(obd, &obd->obd_olg, 1, NULL);
2023 CERROR("failed to setup llogging subsystems\n");
2027 rc = lquota_setup(filter_quota_interface_ref, obd);
2031 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2032 if (uuid_ptr != NULL) {
2033 class_uuid_unparse(uuid_ptr, &uuid);
2039 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2041 if (obd->obd_recovering) {
2042 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2043 "recovery for at least %d:%.02d, or until %d "
2044 "client%s reconnect. During this time new clients"
2045 " will not be allowed to connect. "
2046 "Recovery progress can be monitored by watching "
2047 "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
2048 obd->obd_name, lustre_cfg_string(lcfg, 1),
2049 label ?: "", label ? "/" : "", str,
2050 obd->obd_recovery_timeout / 60,
2051 obd->obd_recovery_timeout % 60,
2052 obd->obd_max_recoverable_clients,
2053 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2056 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2057 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2058 label ?: "", label ? "/" : "", str,
2059 obd->obd_replayable ? "enabled" : "disabled");
2067 fsfilt_put_ops(obd->obd_fsops);
2068 filter_iobuf_pool_done(filter);
2070 server_put_mount(obd->obd_name, mnt);
2071 obd->u.obt.obt_sb = 0;
2075 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2077 struct lprocfs_static_vars lvars;
2082 CLASSERT(offsetof(struct obd_device, u.obt) ==
2083 offsetof(struct obd_device, u.filter.fo_obt));
2085 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2088 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2089 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2092 addr = (unsigned long)cfs_page_address(page);
2093 clear_page((void *)addr);
2095 /* lprocfs must be setup before the filter so state can be safely added
2096 * to /proc incrementally as the filter is setup */
2097 lprocfs_filter_init_vars(&lvars);
2098 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2099 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2100 /* Init obdfilter private stats here */
2101 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2102 LPROCFS_CNTR_AVGMINMAX,
2103 "read_bytes", "bytes");
2104 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2105 LPROCFS_CNTR_AVGMINMAX,
2106 "write_bytes", "bytes");
2108 lproc_filter_attach_seqstat(obd);
2109 obd->obd_proc_exports_entry = lprocfs_register("exports",
2110 obd->obd_proc_entry,
2112 if (IS_ERR(obd->obd_proc_exports_entry)) {
2113 rc = PTR_ERR(obd->obd_proc_exports_entry);
2114 CERROR("error %d setting up lprocfs for %s\n",
2116 obd->obd_proc_exports_entry = NULL;
2119 if (obd->obd_proc_exports_entry)
2120 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2121 lprocfs_nid_stats_clear_read,
2122 lprocfs_nid_stats_clear_write, obd);
2124 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2125 LUSTRE_CFG_BUFLEN(lcfg, 4));
2126 rc = filter_common_setup(obd, lcfg, (void *)addr);
2127 OBD_PAGE_FREE(page);
2130 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2131 lprocfs_free_per_client_stats(obd);
2132 lprocfs_free_obd_stats(obd);
2133 lprocfs_obd_cleanup(obd);
2139 static int filter_group_llog_finish(struct obd_llog_group *olg)
2141 struct llog_ctxt *ctxt;
2142 int rc = 0, rc2 = 0;
2145 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2147 rc = llog_cleanup(ctxt);
2149 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2151 rc2 = llog_cleanup(ctxt);
2158 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
2159 static struct llog_operations filter_size_orig_logops = {
2160 lop_setup: llog_obd_origin_setup,
2161 lop_cleanup: llog_obd_origin_cleanup,
2162 lop_add: llog_obd_origin_add
2166 * Init obd_llog_group of the filter
2168 static int filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2169 struct obd_device *tgt)
2173 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2174 &filter_mds_ost_repl_logops);
2178 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2179 &filter_size_orig_logops);
2184 filter_group_llog_finish(olg);
2189 * Init the default olg, which is embeded in the obd_device, for filter.
2192 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2193 struct obd_device *tgt)
2195 struct filter_obd *filter = &obd->u.filter;
2196 struct llog_ctxt *ctxt;
2199 LASSERT(filter->fo_lcm == NULL);
2200 OBD_ALLOC(filter->fo_lcm, sizeof(struct llog_commit_master));
2201 if (!filter->fo_lcm)
2204 rc = llog_init_commit_master((struct llog_commit_master *)
2209 filter_mds_ost_repl_logops = llog_client_ops;
2210 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2211 filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
2212 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2214 rc = filter_olg_init(obd, olg, tgt);
2218 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2219 LASSERT(ctxt != NULL);
2220 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2221 ctxt->loc_lcm = obd->u.filter.fo_lcm;
2222 /* Only start log commit thread for default llog ctxt*/
2223 rc = llog_start_commit_thread(ctxt->loc_lcm);
2224 llog_ctxt_put(ctxt);
2229 llog_cleanup_commit_master(filter->fo_lcm, 0);
2230 OBD_FREE(filter->fo_lcm, sizeof(struct llog_commit_master));
2231 filter->fo_lcm = NULL;
2236 static int filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2237 struct obd_device *tgt, int count,
2238 struct llog_catid *catid,
2239 struct obd_uuid *uuid)
2241 struct filter_obd *filter = &obd->u.filter;
2242 struct llog_ctxt *ctxt;
2246 LASSERT(olg != NULL);
2247 if (olg == &obd->obd_olg)
2248 return filter_default_olg_init(obd, olg, tgt);
2250 /* For other group llog, which will be initialized in
2251 * llog_connect(after default_olg_init, so fo_lcm must
2252 * already be initialized */
2253 LASSERT(filter->fo_lcm != NULL);
2254 rc = filter_olg_init(obd, olg, tgt);
2257 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2258 LASSERT(ctxt != NULL);
2259 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2260 ctxt->loc_lcm = obd->u.filter.fo_lcm;
2261 llog_ctxt_put(ctxt);
2266 static int filter_llog_finish(struct obd_device *obd, int count)
2271 if (obd->u.filter.fo_lcm) {
2272 llog_cleanup_commit_master((struct llog_commit_master *)
2273 obd->u.filter.fo_lcm, 1);
2274 OBD_FREE(obd->u.filter.fo_lcm,
2275 sizeof(struct llog_commit_master));
2276 obd->u.filter.fo_lcm = NULL;
2278 /* finish obd llog group */
2279 rc = filter_group_llog_finish(&obd->obd_olg);
2285 * Find the group llog according to group index in the llog group list.
2287 static struct obd_llog_group *
2288 filter_find_olg_internal(struct filter_obd *filter, int group)
2290 struct obd_llog_group *olg;
2292 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2293 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2294 if (olg->olg_group == group)
2301 * Find the group llog according to group index on the filter
2303 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2305 struct obd_llog_group *olg = NULL;
2306 struct filter_obd *filter;
2308 filter = &obd->u.filter;
2310 if (group == OBD_LLOG_GROUP)
2311 RETURN(&obd->obd_olg);
2313 spin_lock(&filter->fo_llog_list_lock);
2314 olg = filter_find_olg_internal(filter, group);
2315 spin_unlock(&filter->fo_llog_list_lock);
2320 * Find the llog_group of the filter according to the group. If it can not
2321 * find, create the llog_group, which only happens when mds is being synced
2324 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2326 struct obd_llog_group *olg = NULL;
2327 struct filter_obd *filter;
2330 filter = &obd->u.filter;
2332 if (group == OBD_LLOG_GROUP)
2333 RETURN(&obd->obd_olg);
2335 spin_lock(&filter->fo_llog_list_lock);
2336 olg = filter_find_olg_internal(filter, group);
2338 if (olg->olg_initializing) {
2339 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2341 GOTO(out_unlock, olg);
2346 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2348 llog_group_init(olg, group);
2349 list_add(&olg->olg_list, &filter->fo_llog_list);
2350 olg->olg_initializing = 1;
2351 spin_unlock(&filter->fo_llog_list_lock);
2353 rc = llog_cat_initialize(obd, olg, 1, NULL);
2355 spin_lock(&filter->fo_llog_list_lock);
2356 list_del(&olg->olg_list);
2357 spin_unlock(&filter->fo_llog_list_lock);
2359 GOTO(out, olg = ERR_PTR(-ENOMEM));
2361 spin_lock(&filter->fo_llog_list_lock);
2362 olg->olg_initializing = 0;
2363 spin_unlock(&filter->fo_llog_list_lock);
2364 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2365 obd->obd_name, group, olg);
2370 spin_unlock(&filter->fo_llog_list_lock);
2374 static int filter_llog_connect(struct obd_export *exp,
2375 struct llogd_conn_body *body)
2377 struct obd_device *obd = exp->exp_obd;
2378 struct llog_ctxt *ctxt;
2379 struct obd_llog_group *olg;
2383 CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2384 (unsigned) body->lgdc_logid.lgl_ogr,
2385 (unsigned) body->lgdc_logid.lgl_oid,
2386 (unsigned) body->lgdc_logid.lgl_ogen);
2388 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2390 CERROR(" %s: can not find olg of group %d\n",
2391 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2394 llog_group_set_export(olg, exp);
2396 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2397 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2398 body->lgdc_ctxt_idx);
2399 rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2400 &body->lgdc_gen, NULL);
2401 llog_ctxt_put(ctxt);
2403 CERROR("failed to connect rc %d idx %d\n", rc,
2404 body->lgdc_ctxt_idx);
2409 static int filter_llog_preclean (struct obd_device *obd)
2411 struct obd_llog_group *olg, *tmp;
2412 struct filter_obd *filter;
2413 struct list_head remove_list;
2417 rc = obd_llog_finish(obd, 0);
2419 CERROR("failed to cleanup llogging subsystem\n");
2421 filter = &obd->u.filter;
2422 CFS_INIT_LIST_HEAD(&remove_list);
2424 spin_lock(&filter->fo_llog_list_lock);
2425 while (!list_empty(&filter->fo_llog_list)) {
2426 olg = list_entry(filter->fo_llog_list.next,
2427 struct obd_llog_group, olg_list);
2428 list_del(&olg->olg_list);
2429 list_add(&olg->olg_list, &remove_list);
2431 spin_unlock(&filter->fo_llog_list_lock);
2433 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2434 list_del_init(&olg->olg_list);
2435 rc = filter_group_llog_finish(olg);
2437 CERROR("failed to cleanup llogging subsystem for %u\n",
2445 static int filter_precleanup(struct obd_device *obd,
2446 enum obd_cleanup_stage stage)
2452 case OBD_CLEANUP_EARLY:
2454 case OBD_CLEANUP_EXPORTS:
2455 target_cleanup_recovery(obd);
2456 rc = filter_llog_preclean(obd);
2462 static int filter_cleanup(struct obd_device *obd)
2464 struct filter_obd *filter = &obd->u.filter;
2468 LCONSOLE_WARN("%s: shutting down for failover; client state "
2469 "will be preserved.\n", obd->obd_name);
2471 if (!list_empty(&obd->obd_exports)) {
2472 CERROR("%s: still has clients!\n", obd->obd_name);
2473 class_disconnect_exports(obd);
2474 if (!list_empty(&obd->obd_exports)) {
2475 CERROR("still has exports after forced cleanup?\n");
2480 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2481 lprocfs_free_per_client_stats(obd);
2482 lprocfs_free_obd_stats(obd);
2483 lprocfs_obd_cleanup(obd);
2484 lquota_cleanup(filter_quota_interface_ref, obd);
2486 /* Stop recovery before namespace cleanup. */
2487 target_stop_recovery_thread(obd);
2488 target_cleanup_recovery(obd);
2490 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2491 obd->obd_namespace = NULL;
2493 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2495 if (obd->u.obt.obt_sb == NULL)
2500 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2501 shrink_dcache_sb(obd->u.obt.obt_sb);
2503 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2504 obd->u.obt.obt_sb = NULL;
2506 fsfilt_put_ops(obd->obd_fsops);
2508 filter_iobuf_pool_done(filter);
2510 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2515 static int filter_connect_internal(struct obd_export *exp,
2516 struct obd_connect_data *data)
2518 struct filter_export_data *fed = &exp->exp_filter_data;
2523 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2524 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2525 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2526 data->ocd_connect_flags, data->ocd_version,
2527 data->ocd_grant, data->ocd_index);
2529 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2530 CWARN("!!! This export (nid %s) used object group %d "
2531 "earlier; now it's trying to use group %d! This could "
2532 "be a bug in the MDS. Tell CFS.\n",
2533 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2536 fed->fed_group = data->ocd_group;
2538 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2539 exp->exp_connect_flags = data->ocd_connect_flags;
2540 data->ocd_version = LUSTRE_VERSION_CODE;
2542 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2543 obd_size left, want;
2545 spin_lock(&exp->exp_obd->obd_osfs_lock);
2546 left = filter_grant_space_left(exp);
2547 want = data->ocd_grant;
2548 filter_grant(exp, fed->fed_grant, want, left);
2549 data->ocd_grant = fed->fed_grant;
2550 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2552 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2553 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2554 exp->exp_client_uuid.uuid, exp,
2555 data->ocd_grant, want, left);
2558 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2559 struct filter_obd *filter = &exp->exp_obd->u.filter;
2560 struct lr_server_data *lsd = filter->fo_fsd;
2561 int index = le32_to_cpu(lsd->lsd_ost_index);
2563 if (!(lsd->lsd_feature_compat &
2564 cpu_to_le32(OBD_COMPAT_OST))) {
2565 /* this will only happen on the first connect */
2566 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2567 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2568 filter_update_server_data(exp->exp_obd,
2569 filter->fo_rcvd_filp, lsd, 1);
2570 } else if (index != data->ocd_index) {
2571 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2572 " %u doesn't match actual OST index"
2573 " %u in last_rcvd file, bad "
2575 obd_export_nid2str(exp), index,
2581 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2582 data->ocd_brw_size = 65536;
2583 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2584 data->ocd_brw_size = min(data->ocd_brw_size,
2585 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2586 LASSERT(data->ocd_brw_size);
2589 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2590 __u32 cksum_types = data->ocd_cksum_types;
2592 /* The client set in ocd_cksum_types the checksum types it
2593 * supports. We have to mask off the algorithms that we don't
2595 if (cksum_types & OBD_CKSUM_ALL)
2596 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2598 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2600 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2601 "%x\n", exp->exp_obd->obd_name,
2602 obd_export_nid2str(exp), cksum_types,
2603 data->ocd_cksum_types);
2605 /* This client does not support OBD_CONNECT_CKSUM
2606 * fall back to CRC32 */
2607 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2608 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2609 exp->exp_obd->obd_name,
2610 obd_export_nid2str(exp));
2613 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2614 * FIXME: We don't strictly need the COMPAT flag for that,
2615 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2616 * FIXME: We needed it for the index, as index 0 is valid. */
2621 static int filter_reconnect(const struct lu_env *env,
2622 struct obd_export *exp, struct obd_device *obd,
2623 struct obd_uuid *cluuid,
2624 struct obd_connect_data *data)
2629 if (exp == NULL || obd == NULL || cluuid == NULL)
2632 rc = filter_connect_internal(exp, data);
2637 /* nearly identical to mds_connect */
2638 static int filter_connect(const struct lu_env *env,
2639 struct lustre_handle *conn, struct obd_device *obd,
2640 struct obd_uuid *cluuid,
2641 struct obd_connect_data *data, void *localdata)
2643 struct lvfs_run_ctxt saved;
2644 struct obd_export *exp;
2645 struct filter_export_data *fed;
2646 struct lsd_client_data *lcd = NULL;
2651 if (conn == NULL || obd == NULL || cluuid == NULL)
2654 rc = class_connect(conn, obd, cluuid);
2657 exp = class_conn2export(conn);
2658 LASSERT(exp != NULL);
2660 fed = &exp->exp_filter_data;
2662 rc = filter_connect_internal(exp, data);
2666 filter_export_stats_init(obd, exp, localdata);
2667 if (obd->obd_replayable) {
2668 OBD_ALLOC(lcd, sizeof(*lcd));
2670 CERROR("filter: out of memory for client data\n");
2671 GOTO(cleanup, rc = -ENOMEM);
2674 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2676 rc = filter_client_add(obd, exp, -1);
2681 group = data->ocd_group;
2685 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2686 obd->obd_name, exp->exp_handle.h_cookie, group);
2688 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2689 rc = filter_read_groups(obd, group, 1);
2690 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2692 CERROR("can't read group %u\n", group);
2702 fed->fed_lcd = NULL;
2704 class_disconnect(exp);
2706 class_export_put(exp);
2712 /* Do extra sanity checks for grant accounting. We do this at connect,
2713 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2714 * always get rid of it or turn it off when we know accounting is good. */
2715 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2717 struct filter_export_data *fed;
2718 struct obd_export *exp;
2719 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2720 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2721 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2723 if (list_empty(&obd->obd_exports))
2726 /* We don't want to do this for large machines that do lots of
2727 mounts or unmounts. It burns... */
2728 if (obd->obd_num_exports > 100)
2731 spin_lock(&obd->obd_osfs_lock);
2732 spin_lock(&obd->obd_dev_lock);
2733 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2735 fed = &exp->exp_filter_data;
2736 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2739 if (maxsize > 0) { /* we may not have done a statfs yet */
2740 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2741 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2742 exp->exp_client_uuid.uuid, exp,
2743 fed->fed_grant, fed->fed_pending, maxsize);
2744 LASSERTF(fed->fed_dirty <= maxsize,
2745 "%s: cli %s/%p %ld > "LPU64"\n", func,
2746 exp->exp_client_uuid.uuid, exp,
2747 fed->fed_dirty, maxsize);
2750 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2751 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2752 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2754 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2755 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2756 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2757 tot_granted += fed->fed_grant + fed->fed_pending;
2758 tot_pending += fed->fed_pending;
2759 tot_dirty += fed->fed_dirty;
2761 fo_tot_granted = obd->u.filter.fo_tot_granted;
2762 fo_tot_pending = obd->u.filter.fo_tot_pending;
2763 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2764 spin_unlock(&obd->obd_dev_lock);
2765 spin_unlock(&obd->obd_osfs_lock);
2767 /* Do these assertions outside the spinlocks so we don't kill system */
2768 if (tot_granted != fo_tot_granted)
2769 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2770 func, tot_granted, fo_tot_granted);
2771 if (tot_pending != fo_tot_pending)
2772 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2773 func, tot_pending, fo_tot_pending);
2774 if (tot_dirty != fo_tot_dirty)
2775 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2776 func, tot_dirty, fo_tot_dirty);
2777 if (tot_pending > tot_granted)
2778 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2779 func, tot_pending, tot_granted);
2780 if (tot_granted > maxsize)
2781 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2782 func, tot_granted, maxsize);
2783 if (tot_dirty > maxsize)
2784 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2785 func, tot_dirty, maxsize);
2788 /* Remove this client from the grant accounting totals. We also remove
2789 * the export from the obd device under the osfs and dev locks to ensure
2790 * that the filter_grant_sanity_check() calculations are always valid.
2791 * The client should do something similar when it invalidates its import. */
2792 static void filter_grant_discard(struct obd_export *exp)
2794 struct obd_device *obd = exp->exp_obd;
2795 struct filter_obd *filter = &obd->u.filter;
2796 struct filter_export_data *fed = &exp->exp_filter_data;
2798 spin_lock(&obd->obd_osfs_lock);
2799 spin_lock(&obd->obd_dev_lock);
2800 list_del_init(&exp->exp_obd_chain);
2801 spin_unlock(&obd->obd_dev_lock);
2803 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2804 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2805 obd->obd_name, filter->fo_tot_granted,
2806 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2807 filter->fo_tot_granted -= fed->fed_grant;
2808 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2809 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2810 obd->obd_name, filter->fo_tot_pending,
2811 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2812 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2813 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2814 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2815 obd->obd_name, filter->fo_tot_dirty,
2816 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2817 filter->fo_tot_dirty -= fed->fed_dirty;
2821 spin_unlock(&obd->obd_osfs_lock);
2824 static int filter_destroy_export(struct obd_export *exp)
2828 if (exp->exp_filter_data.fed_pending)
2829 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2830 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2831 exp, exp->exp_filter_data.fed_pending);
2833 /* Not ported yet the b1_6 quota functionality
2834 * lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2837 target_destroy_export(exp);
2839 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2842 lprocfs_exp_cleanup(exp);
2844 if (exp->exp_obd->obd_replayable)
2845 filter_client_free(exp);
2847 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2849 filter_grant_discard(exp);
2850 filter_fmd_cleanup(exp);
2852 if (!(exp->exp_flags & OBD_OPT_FORCE))
2853 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2858 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2860 struct obd_llog_group *olg_min, *olg;
2861 struct filter_obd *filter;
2862 int worked = 0, group;
2863 struct llog_ctxt *ctxt;
2866 filter = &obd->u.filter;
2868 /* we can't sync log holding spinlock. also, we do not want to get
2869 * into livelock. so we do following: loop over MDS's exports in
2870 * group order and skip already synced llogs -bzzz */
2872 /* look for group with min. number, but > worked */
2875 spin_lock(&filter->fo_llog_list_lock);
2876 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2877 if (olg->olg_group <= worked) {
2878 /* this group is already synced */
2881 if (group < olg->olg_group) {
2882 /* we have group with smaller number to sync */
2885 /* store current minimal group */
2887 group = olg->olg_group;
2889 spin_unlock(&filter->fo_llog_list_lock);
2891 if (olg_min == NULL)
2894 worked = olg_min->olg_group;
2895 if (olg_min->olg_exp &&
2896 (dexp == olg_min->olg_exp || dexp == NULL)) {
2898 ctxt = llog_group_get_ctxt(olg_min,
2899 LLOG_MDS_OST_REPL_CTXT);
2900 LASSERT(ctxt != NULL);
2901 err = llog_sync(ctxt, olg_min->olg_exp);
2902 llog_ctxt_put(ctxt);
2904 CERROR("error flushing logs to MDS: rc %d\n",
2907 } while (olg_min != NULL);
2910 /* also incredibly similar to mds_disconnect */
2911 static int filter_disconnect(struct obd_export *exp)
2913 struct obd_device *obd = exp->exp_obd;
2918 class_export_get(exp);
2920 if (!(exp->exp_flags & OBD_OPT_FORCE))
2921 filter_grant_sanity_check(obd, __FUNCTION__);
2922 filter_grant_discard(exp);
2924 /* Disconnect early so that clients can't keep using export */
2925 rc = class_disconnect(exp);
2926 if (exp->exp_obd->obd_namespace != NULL)
2927 ldlm_cancel_locks_for_export(exp);
2929 fsfilt_sync(obd, obd->u.obt.obt_sb);
2931 /* flush any remaining cancel messages out to the target */
2932 filter_sync_llogs(obd, exp);
2933 class_export_put(exp);
2937 /* reverse import is changed, sync all cancels */
2938 static void filter_revimp_update(struct obd_export *exp)
2943 class_export_get(exp);
2945 /* flush any remaining cancel messages out to the target */
2946 filter_sync_llogs(exp->exp_obd, exp);
2947 class_export_put(exp);
2951 static int filter_ping(struct obd_export *exp)
2953 filter_fmd_expire(exp);
2958 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2959 const char *what, int quiet)
2961 struct dentry *dchild = NULL;
2964 if (oa->o_valid & OBD_MD_FLGROUP)
2967 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
2969 if (IS_ERR(dchild)) {
2970 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
2971 what, group, oa->o_id);
2975 if (dchild->d_inode == NULL) {
2977 CERROR("%s: %s on non-existent object: "LPU64"\n",
2978 obd->obd_name, what, oa->o_id);
2980 RETURN(ERR_PTR(-ENOENT));
2986 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2988 struct dentry *dentry = NULL;
2989 struct obd_device *obd;
2993 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
2994 oinfo_capa(oinfo), CAPA_OPC_META_READ);
2998 obd = class_exp2obd(exp);
3000 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3004 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3006 RETURN(PTR_ERR(dentry));
3008 /* Limit the valid bits in the return data to what we actually use */
3009 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3010 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3016 /* this should be enabled/disabled in condition to enabled/disabled large
3017 * inodes (fast EAs) in backing store FS. */
3018 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3019 void *handle, struct obdo *oa)
3021 struct obd_device *obd = exp->exp_obd;
3025 if (oa->o_valid & OBD_MD_FLFID) {
3026 struct filter_fid ff;
3028 if (!(oa->o_valid & OBD_MD_FLGROUP))
3030 /* packing fid and converting it to LE for storing into EA.
3031 * Here ->o_stripe_idx should be filled by LOV and rest of
3032 * fields - by client. */
3033 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3034 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3035 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3036 ff.ff_objid = cpu_to_le64(oa->o_id);
3037 ff.ff_group = cpu_to_le64(oa->o_gr);
3039 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3040 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3041 oa->o_generation, oa->o_id, oa->o_gr);
3043 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3045 CERROR("store fid in object failed! rc: %d\n", rc);
3047 CDEBUG(D_HA, "OSS object without fid info!\n");
3053 /* this is called from filter_truncate() until we have filter_punch() */
3054 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3055 struct obdo *oa, struct obd_trans_info *oti)
3057 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3058 struct llog_cookie *fcc = NULL;
3059 struct filter_obd *filter;
3060 int rc, err, locked = 0, sync = 0;
3061 unsigned int ia_valid;
3062 struct inode *inode;
3067 LASSERT(dentry != NULL);
3068 LASSERT(!IS_ERR(dentry));
3070 inode = dentry->d_inode;
3071 LASSERT(inode != NULL);
3073 filter = &exp->exp_obd->u.filter;
3074 iattr_from_obdo(&iattr, oa, oa->o_valid);
3075 ia_valid = iattr.ia_valid;
3077 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3078 OBD_ALLOC(fcc, sizeof(*fcc));
3080 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3083 if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
3085 LOCK_INODE_MUTEX(inode);
3089 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3090 * then we will accept the UID+GID sent by the client during write for
3091 * initializing the ownership of this inode. We only allow this to
3092 * happen once so clear these bits in setattr. In 2.6 kernels it is
3093 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3094 * the flags that are actually being set. */
3095 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3096 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3097 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3099 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3100 if (!(ia_valid & ATTR_MODE)) {
3101 iattr.ia_mode = inode->i_mode;
3102 iattr.ia_valid |= ATTR_MODE;
3104 iattr.ia_mode &= ~S_ISUID;
3106 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3107 if (!(iattr.ia_valid & ATTR_MODE)) {
3108 iattr.ia_mode = inode->i_mode;
3109 iattr.ia_valid |= ATTR_MODE;
3111 iattr.ia_mode &= ~S_ISGID;
3114 orig_ids[USRQUOTA] = inode->i_uid;
3115 orig_ids[GRPQUOTA] = inode->i_gid;
3116 handle = fsfilt_start_log(exp->exp_obd, inode,
3117 FSFILT_OP_SETATTR, oti, 1);
3119 GOTO(out_unlock, rc = PTR_ERR(handle));
3121 /* update inode EA only once when inode is suid bit marked. As
3122 * on 2.6.x UID and GID may be set separately, we check here
3123 * only one of them to avoid double setting. */
3124 if (inode->i_mode & S_ISUID)
3125 filter_update_fidea(exp, inode, handle, oa);
3127 handle = fsfilt_start(exp->exp_obd, inode,
3128 FSFILT_OP_SETATTR, oti);
3130 GOTO(out_unlock, rc = PTR_ERR(handle));
3132 if (oa->o_valid & OBD_MD_FLFLAGS) {
3133 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3134 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3136 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3138 /* set cancel cookie callback function */
3139 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3140 filter_cancel_cookies_cb,
3144 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3145 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3147 /* The truncate might have used up our transaction credits. Make
3148 * sure we have one left for the last_rcvd update. */
3149 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3151 rc = filter_finish_transno(exp, oti, rc, sync);
3153 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3157 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3159 CERROR("error on commit, err = %d\n", err);
3167 /* Let's flush truncated page on disk immediately, then we can
3168 * avoid need to search for page aliases before directio writes
3169 * and this sort of stuff at expense of somewhat slower
3170 * truncates not on a page boundary. I believe this is the only
3171 * place in filter code that can lead to pages getting to
3172 * pagecache so far. */
3173 filter_clear_truncated_page(inode);
3174 UNLOCK_INODE_MUTEX(inode);
3181 UNLOCK_INODE_MUTEX(inode);
3184 OBD_FREE(fcc, sizeof(*fcc));
3186 /* trigger quota release */
3187 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3188 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3189 int rc2 = lquota_adjust(filter_quota_interface_ref,
3190 exp->exp_obd, cur_ids,
3191 orig_ids, rc, FSFILT_OP_SETATTR);
3192 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3193 "filter adjust qunit. (rc:%d)\n", rc2);
3198 /* this is called from filter_truncate() until we have filter_punch() */
3199 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3200 struct obd_trans_info *oti)
3202 struct ldlm_res_id res_id = { .name = { oinfo->oi_oa->o_id, 0,
3203 oinfo->oi_oa->o_gr, 0 } };
3204 struct filter_mod_data *fmd;
3205 struct lvfs_run_ctxt saved;
3206 struct filter_obd *filter;
3207 struct ldlm_resource *res;
3208 struct dentry *dentry;
3212 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3213 oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
3217 /* This would be very bad - accidentally truncating a file when
3218 * changing the time or similar - bug 12203. */
3219 if (oinfo->oi_oa->o_valid & OBD_MD_FLSIZE &&
3220 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3221 static char mdsinum[48];
3223 if (oinfo->oi_oa->o_valid & OBD_MD_FLFID)
3224 snprintf(mdsinum, sizeof(mdsinum) - 1,
3225 " of inode "LPU64"/%u", oinfo->oi_oa->o_fid,
3226 oinfo->oi_oa->o_generation);
3230 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3232 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3233 oinfo->oi_oa->o_id, mdsinum);
3237 dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
3240 RETURN(PTR_ERR(dentry));
3242 filter = &exp->exp_obd->u.filter;
3243 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3246 if (oinfo->oi_oa->o_valid &
3247 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3248 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
3249 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3250 fmd->fmd_mactime_xid = oti->oti_xid;
3251 filter_fmd_put(exp, fmd);
3254 /* setting objects attributes (including owner/group) */
3255 rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
3257 GOTO(out_unlock, rc);
3259 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3260 &res_id, LDLM_EXTENT, 0);
3263 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3264 ldlm_resource_putref(res);
3267 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3269 /* Quota release need uid/gid info */
3270 obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
3271 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3277 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3281 /* XXX identical to osc_unpackmd */
3282 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3283 struct lov_mds_md *lmm, int lmm_bytes)
3289 if (lmm_bytes < sizeof (*lmm)) {
3290 CERROR("lov_mds_md too small: %d, need %d\n",
3291 lmm_bytes, (int)sizeof(*lmm));
3294 /* XXX LOV_MAGIC etc check? */
3296 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3297 CERROR("lov_mds_md: zero lmm_object_id\n");
3302 lsm_size = lov_stripe_md_size(1);
3306 if (*lsmp != NULL && lmm == NULL) {
3307 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3308 OBD_FREE(*lsmp, lsm_size);
3313 if (*lsmp == NULL) {
3314 OBD_ALLOC(*lsmp, lsm_size);
3318 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3319 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3320 OBD_FREE(*lsmp, lsm_size);
3323 loi_init((*lsmp)->lsm_oinfo[0]);
3327 /* XXX zero *lsmp? */
3328 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3329 LASSERT((*lsmp)->lsm_object_id);
3332 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3337 /* caller must hold fo_create_locks[oa->o_gr] */
3338 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3339 struct filter_obd *filter)
3341 struct obdo doa; /* XXX obdo on stack */
3347 LASSERT(oa->o_gr != 0);
3348 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3349 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3351 memset(&doa, 0, sizeof(doa));
3353 doa.o_valid |= OBD_MD_FLGROUP;
3354 doa.o_gr = oa->o_gr;
3355 doa.o_mode = S_IFREG;
3357 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3358 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3359 exp->exp_obd->obd_name, doa.o_gr);
3363 last = filter_last_id(filter, doa.o_gr);
3365 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3366 exp->exp_obd->obd_name, oa->o_id + 1, last);
3368 for (id = last; id > oa->o_id; id--) {
3370 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
3371 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3372 CEMERG("error destroying precreate objid "LPU64": %d\n",
3374 filter_set_last_id(filter, id - 1, doa.o_gr);
3375 /* update last_id on disk periodically so that if we restart
3376 * we don't need to re-scan all of the just-deleted objects. */
3377 if ((id & 511) == 0)
3378 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3381 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3382 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3384 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3385 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3390 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3391 obd_gr group, int *num);
3392 /* returns a negative error or a nonnegative number of files to create */
3393 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3394 obd_gr group, struct obd_trans_info *oti)
3396 struct obd_device *obd = exp->exp_obd;
3397 struct filter_obd *filter = &obd->u.filter;
3401 /* delete orphans request */
3402 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3403 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3404 CERROR("%s: dropping old orphan cleanup request\n",
3408 /* This causes inflight precreates to abort and drop lock */
3409 set_bit(group, &filter->fo_destroys_in_progress);
3410 down(&filter->fo_create_locks[group]);
3411 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3412 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3413 exp->exp_obd->obd_name, group);
3414 up(&filter->fo_create_locks[group]);
3417 diff = oa->o_id - filter_last_id(filter, group);
3418 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3419 filter_last_id(filter, group), diff);
3421 if (-diff > OST_MAX_PRECREATE) {
3422 CERROR("%s: ignoring bogus orphan destroy request: "
3423 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3424 oa->o_id, filter_last_id(filter, group));
3425 /* FIXME: should reset precreate_next_id on MDS */
3426 GOTO(out, rc = -EINVAL);
3429 rc = filter_destroy_precreated(exp, oa, filter);
3431 CERROR("%s: unable to write lastobjid, but "
3432 "orphans were deleted\n", obd->obd_name);
3435 /* XXX: Used by MDS for the first time! */
3436 clear_bit(group, &filter->fo_destroys_in_progress);
3439 down(&filter->fo_create_locks[group]);
3440 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3441 CERROR("%s: dropping old precreate request\n",
3445 /* only precreate if group == 0 and o_id is specfied */
3446 if (group < FILTER_GROUP_MDS0 || oa->o_id == 0)
3449 diff = oa->o_id - filter_last_id(filter, group);
3450 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3451 filter_last_id(filter, group), diff);
3453 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3454 oa->o_id, filter_last_id(filter, group), diff);
3458 oa->o_id = filter_last_id(&obd->u.filter, group);
3459 rc = filter_precreate(obd, oa, group, &diff);
3460 oa->o_id = filter_last_id(&obd->u.filter, group);
3462 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3465 /* else diff == 0 */
3468 up(&filter->fo_create_locks[group]);
3472 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3473 __u64 max_age, __u32 flags)
3475 struct filter_obd *filter = &obd->u.filter;
3476 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3480 /* at least try to account for cached pages. its still racey and
3481 * might be under-reporting if clients haven't announced their
3482 * caches with brw recently */
3483 spin_lock(&obd->obd_osfs_lock);
3484 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3485 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3486 spin_unlock(&obd->obd_osfs_lock);
3488 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3489 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3490 filter->fo_tot_dirty, filter->fo_tot_granted,
3491 filter->fo_tot_pending,
3492 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3494 filter_grant_sanity_check(obd, __FUNCTION__);
3496 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3497 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3498 osfs->os_bsize - 1) >> blockbits));
3500 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3501 * stop creating files on MDS if OST is not good shape to create
3503 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3508 static int filter_use_existing_obj(struct obd_device *obd,
3509 struct dentry *dchild, void **handle,
3512 struct inode *inode = dchild->d_inode;
3516 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3519 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3520 if (IS_ERR(*handle))
3521 return PTR_ERR(*handle);
3523 iattr.ia_valid = ATTR_MODE;
3524 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3525 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3533 /* We rely on the fact that only one thread will be creating files in a given
3534 * group at a time, which is why we don't need an atomic filter_get_new_id.
3535 * Even if we had that atomic function, the following race would exist:
3537 * thread 1: gets id x from filter_next_id
3538 * thread 2: gets id (x + 1) from filter_next_id
3539 * thread 2: creates object (x + 1)
3540 * thread 1: tries to create object x, gets -ENOSPC
3542 * Caller must hold fo_create_locks[group]
3544 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3545 obd_gr group, int *num)
3547 struct dentry *dchild = NULL, *dparent = NULL;
3548 struct filter_obd *filter;
3549 struct obd_statfs *osfs;
3550 int err = 0, rc = 0, recreate_obj = 0, i;
3551 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3553 void *handle = NULL;
3556 filter = &obd->u.filter;
3558 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3560 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3562 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3563 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3566 OBD_ALLOC(osfs, sizeof(*osfs));
3569 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3570 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3571 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3572 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3573 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3577 OBD_FREE(osfs, sizeof(*osfs));
3582 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3583 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3585 for (i = 0; i < *num && err == 0; i++) {
3586 int cleanup_phase = 0;
3588 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3589 CWARN("%s: create aborted by destroy\n",
3598 last_id = filter_last_id(filter, group);
3599 if (next_id > last_id) {
3600 CERROR("Error: Trying to recreate obj greater"
3601 "than last id "LPD64" > "LPD64"\n",
3603 GOTO(cleanup, rc = -EINVAL);
3606 next_id = filter_last_id(filter, group) + 1;
3608 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
3610 dparent = filter_parent_lock(obd, group, next_id);
3611 if (IS_ERR(dparent))
3612 GOTO(cleanup, rc = PTR_ERR(dparent));
3613 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3615 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3617 GOTO(cleanup, rc = PTR_ERR(dchild));
3618 cleanup_phase = 2; /* f_dput(dchild) */
3620 if (dchild->d_inode != NULL) {
3621 /* This would only happen if lastobjid was bad on disk*/
3622 /* Could also happen if recreating missing obj but it
3623 * already exists. */
3625 CERROR("%s: recreating existing object %.*s?\n",
3626 obd->obd_name, dchild->d_name.len,
3627 dchild->d_name.name);
3629 /* Use these existing objects if they are
3631 if (dchild->d_inode->i_size == 0) {
3632 rc = filter_use_existing_obj(obd,dchild,
3633 &handle, &cleanup_phase);
3640 CERROR("%s: Serious error: objid %.*s already "
3641 "exists; is this filesystem corrupt?\n",
3642 obd->obd_name, dchild->d_name.len,
3643 dchild->d_name.name);
3646 GOTO(cleanup, rc = -EEXIST);
3649 handle = fsfilt_start_log(obd, dparent->d_inode,
3650 FSFILT_OP_CREATE, NULL, 1);
3652 GOTO(cleanup, rc = PTR_ERR(handle));
3655 /* We mark object SUID+SGID to flag it for accepting UID+GID
3656 * from client on first write. Currently the permission bits
3657 * on the OST are never used, so this is OK. */
3658 rc = ll_vfs_create(dparent->d_inode, dchild,
3659 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3661 CERROR("create failed rc = %d\n", rc);
3666 if (!recreate_obj) {
3667 filter_set_last_id(filter, next_id, group);
3668 err = filter_update_last_objid(obd, group, 0);
3670 CERROR("unable to write lastobjid "
3671 "but file created\n");
3675 switch(cleanup_phase) {
3677 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3679 CERROR("error on commit, err = %d\n", err);
3686 filter_parent_unlock(dparent);
3693 if (time_after(jiffies, enough_time)) {
3695 "%s: precreate slow - want %d got %d \n",
3696 obd->obd_name, *num, i);
3703 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3704 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3709 static int filter_create(struct obd_export *exp, struct obdo *oa,
3710 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3712 struct filter_export_data *fed;
3713 struct obd_device *obd = NULL;
3714 struct filter_obd *filter;
3715 struct lvfs_run_ctxt saved;
3716 struct lov_stripe_md *lsm = NULL;
3717 int rc = 0, diff, group = oa->o_gr;
3720 if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
3721 CERROR("!!! nid %s sent invalid object group %d\n",
3722 obd_export_nid2str(exp), group);
3727 fed = &exp->exp_filter_data;
3728 filter = &obd->u.filter;
3730 if (fed->fed_group != group) {
3731 CERROR("!!! this export (nid %s) used object group %d "
3732 "earlier; now it's trying to use group %d! This could "
3733 "be a bug in the MDS. Tell CFS.\n",
3734 obd_export_nid2str(exp), fed->fed_group, group);
3738 CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
3739 oa->o_gr, oa->o_id);
3743 rc = obd_alloc_memmd(exp, &lsm);
3750 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3752 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3753 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3754 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3755 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3756 oa->o_id, filter_last_id(filter,
3761 down(&filter->fo_create_locks[oa->o_gr]);
3762 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3763 up(&filter->fo_create_locks[oa->o_gr]);
3766 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3769 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3770 if (rc && ea != NULL && *ea != lsm) {
3771 obd_free_memmd(exp, &lsm);
3772 } else if (rc == 0 && ea != NULL) {
3773 /* XXX LOV STACKING: the lsm that is passed to us from
3774 * LOV does not have valid lsm_oinfo data structs, so
3775 * don't go touching that. This needs to be fixed in a
3777 lsm->lsm_object_id = oa->o_id;
3784 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3785 struct lov_stripe_md *md, struct obd_trans_info *oti,
3786 struct obd_export *md_exp)
3788 unsigned int qcids[MAXQUOTAS] = {0, 0};
3789 struct obd_device *obd;
3790 struct filter_obd *filter;
3791 struct dentry *dchild = NULL, *dparent = NULL;
3792 struct lvfs_run_ctxt saved;
3793 void *handle = NULL;
3794 struct llog_cookie *fcc = NULL;
3795 int rc, rc2, cleanup_phase = 0, sync = 0;
3799 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3802 filter = &obd->u.filter;
3804 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3807 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3809 GOTO(cleanup, rc = PTR_ERR(dchild));
3812 if (dchild->d_inode == NULL) {
3813 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3815 /* If object already gone, cancel cookie right now */
3816 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3817 struct llog_ctxt *ctxt;
3818 struct obd_llog_group *olg;
3819 fcc = obdo_logcookie(oa);
3820 olg = filter_find_olg(obd, oa->o_gr);
3822 CERROR(" %s: can not find olg of group %d\n",
3823 obd->obd_name, (int)oa->o_gr);
3824 GOTO(cleanup, rc = PTR_ERR(olg));
3826 llog_group_set_export(olg, exp);
3828 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
3829 llog_cancel(ctxt, NULL, 1, fcc, 0);
3830 llog_ctxt_put(ctxt);
3831 fcc = NULL; /* we didn't allocate fcc, don't free it */
3833 GOTO(cleanup, rc = -ENOENT);
3836 filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3838 /* Our MDC connection is established by the MDS to us */
3839 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3840 OBD_ALLOC(fcc, sizeof(*fcc));
3842 memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3844 DQUOT_INIT(dchild->d_inode);
3846 /* we're gonna truncate it first in order to avoid possible deadlock:
3848 * open trasaction open transaction
3849 * down(i_zombie) down(i_zombie)
3850 * restart transaction
3851 * (see BUG 4180) -bzzz
3853 LOCK_INODE_MUTEX(dchild->d_inode);
3854 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3856 if (IS_ERR(handle)) {
3857 UNLOCK_INODE_MUTEX(dchild->d_inode);
3858 GOTO(cleanup, rc = PTR_ERR(handle));
3861 iattr.ia_valid = ATTR_SIZE;
3863 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3864 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3865 UNLOCK_INODE_MUTEX(dchild->d_inode);
3869 GOTO(cleanup, rc = rc2);
3871 /* We don't actually need to lock the parent until we are unlinking
3872 * here, and not while truncating above. That avoids holding the
3873 * parent lock for a long time during truncate, which can block other
3874 * threads from doing anything to objects in that directory. bug 7171 */
3875 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3876 if (IS_ERR(dparent))
3877 GOTO(cleanup, rc = PTR_ERR(dparent));
3878 cleanup_phase = 3; /* filter_parent_unlock */
3880 LOCK_INODE_MUTEX(dchild->d_inode);
3881 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3882 if (IS_ERR(handle)) {
3883 UNLOCK_INODE_MUTEX(dchild->d_inode);
3884 GOTO(cleanup, rc = PTR_ERR(handle));
3886 cleanup_phase = 4; /* fsfilt_commit */
3888 /* Quota release need uid/gid of inode */
3889 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3891 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3893 /* this drops dchild->d_inode->i_mutex unconditionally */
3894 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3898 switch(cleanup_phase) {
3901 sync = fsfilt_add_journal_cb(obd, 0, oti ?
3902 oti->oti_handle : handle,
3903 filter_cancel_cookies_cb,
3905 /* If add_journal_cb failed, then filter_finish_transno
3906 * will commit the handle and we will do a sync
3907 * on commit. then we call callback directly to free
3910 rc = filter_finish_transno(exp, oti, rc, sync);
3912 filter_cancel_cookies_cb(obd, 0, fcc, rc);
3915 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3917 CERROR("error on commit, err = %d\n", rc2);
3924 filter_parent_unlock(dparent);
3928 OBD_FREE(fcc, sizeof(*fcc));
3930 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3933 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3937 /* trigger quota release */
3938 qcids[USRQUOTA] = oa->o_uid;
3939 qcids[GRPQUOTA] = oa->o_gid;
3940 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
3944 CDEBUG(D_QUOTA, "filter adjust qunit! (rc:%d)\n", rc2);
3948 /* NB start and end are used for punch, but not truncate */
3949 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3950 struct obd_trans_info *oti,
3951 struct ptlrpc_request_set *rqset)
3956 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3957 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
3958 oinfo->oi_policy.l_extent.end);
3962 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
3963 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
3964 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
3966 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3967 oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
3971 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
3972 rc = filter_setattr(exp, oinfo, oti);
3976 static int filter_sync(struct obd_export *exp, struct obdo *oa,
3977 struct lov_stripe_md *lsm, obd_off start, obd_off end,
3980 struct lvfs_run_ctxt saved;
3981 struct filter_obd *filter;
3982 struct dentry *dentry;
3986 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
3987 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
3991 filter = &exp->exp_obd->u.filter;
3993 /* an objid of zero is taken to mean "sync whole filesystem" */
3994 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
3995 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
3996 /* flush any remaining cancel messages out to the target */
3997 filter_sync_llogs(exp->exp_obd, exp);
4001 dentry = filter_oa2dentry(exp->exp_obd, oa);
4003 RETURN(PTR_ERR(dentry));
4005 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4007 LOCK_INODE_MUTEX(dentry->d_inode);
4009 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4011 /* just any file to grab fsync method - "file" arg unused */
4012 struct file *file = filter->fo_rcvd_filp;
4014 if (file->f_op && file->f_op->fsync)
4015 rc = file->f_op->fsync(NULL, dentry, 1);
4017 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4021 UNLOCK_INODE_MUTEX(dentry->d_inode);
4023 oa->o_valid = OBD_MD_FLID;
4024 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4026 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4032 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4033 void *key, __u32 *vallen, void *val)
4035 struct obd_device *obd;
4038 obd = class_exp2obd(exp);
4040 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4044 if (KEY_IS(KEY_BLOCKSIZE)) {
4045 __u32 *blocksize = val;
4047 if (*vallen < sizeof(*blocksize))
4049 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4051 *vallen = sizeof(*blocksize);
4055 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4056 __u32 *blocksize_bits = val;
4057 if (blocksize_bits) {
4058 if (*vallen < sizeof(*blocksize_bits))
4060 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4062 *vallen = sizeof(*blocksize_bits);
4066 if (KEY_IS(KEY_LAST_ID)) {
4067 obd_id *last_id = val;
4068 /* FIXME: object groups */
4070 if (*vallen < sizeof(*last_id))
4072 *last_id = filter_last_id(&obd->u.filter,
4073 exp->exp_filter_data.fed_group);
4075 *vallen = sizeof(*last_id);
4079 CDEBUG(D_IOCTL, "invalid key\n");
4083 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4084 void *key, __u32 vallen, void *val,
4085 struct ptlrpc_request_set *set)
4087 struct obd_device *obd;
4088 struct obd_llog_group *olg;
4089 struct llog_ctxt *ctxt;
4095 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4099 if (KEY_IS(KEY_CAPA_KEY)) {
4100 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4102 CERROR("filter update capability key failed: %d\n", rc);
4106 if (KEY_IS(KEY_REVIMP_UPD)) {
4107 filter_revimp_update(exp);
4111 if (!KEY_IS(KEY_MDS_CONN))
4114 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4115 obd_export_nid2str(exp));
4116 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4118 /* setup llog imports */
4119 LASSERT(val != NULL);
4120 group = (int)(*(__u32 *)val);
4121 LASSERT(group >= FILTER_GROUP_MDS0);
4123 olg = filter_find_create_olg(obd, group);
4125 RETURN(PTR_ERR(olg));
4127 llog_group_set_export(olg, exp);
4129 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4130 LASSERTF(ctxt != NULL, "ctxt is null\n"),
4132 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4133 llog_ctxt_put(ctxt);
4135 lquota_setinfo(filter_quota_interface_ref, exp, obd);
4140 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4141 int len, void *karg, void *uarg)
4143 struct obd_device *obd = exp->exp_obd;
4144 struct obd_ioctl_data *data = karg;
4148 case OBD_IOC_ABORT_RECOVERY: {
4149 CERROR("aborting recovery for device %s\n", obd->obd_name);
4150 target_stop_recovery_thread(obd);
4154 case OBD_IOC_SYNC: {
4155 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4156 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4160 case OBD_IOC_SET_READONLY: {
4162 struct super_block *sb = obd->u.obt.obt_sb;
4163 struct inode *inode = sb->s_root->d_inode;
4164 BDEVNAME_DECLARE_STORAGE(tmp);
4165 CERROR("*** setting device %s read-only ***\n",
4166 ll_bdevname(sb, tmp));
4168 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4169 if (!IS_ERR(handle))
4170 rc = fsfilt_commit(obd, inode, handle, 1);
4172 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4173 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4175 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4179 case OBD_IOC_CATLOGLIST: {
4180 rc = llog_catalog_list(obd, 1, data);
4184 case OBD_IOC_LLOG_CANCEL:
4185 case OBD_IOC_LLOG_REMOVE:
4186 case OBD_IOC_LLOG_INFO:
4187 case OBD_IOC_LLOG_PRINT: {
4188 /* FIXME to be finished */
4189 RETURN(-EOPNOTSUPP);
4191 struct llog_ctxt *ctxt = NULL;
4193 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4194 rc = llog_ioctl(ctxt, cmd, data);
4195 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4208 static int filter_health_check(struct obd_device *obd)
4210 #ifdef USE_HEALTH_CHECK_WRITE
4211 struct filter_obd *filter = &obd->u.filter;
4216 * health_check to return 0 on healthy
4217 * and 1 on unhealthy.
4219 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4222 #ifdef USE_HEALTH_CHECK_WRITE
4223 LASSERT(filter->fo_health_check_filp != NULL);
4224 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4229 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4232 return filter_fid2dentry(data, NULL, gr, id);
4235 static int filter_process_config(struct obd_device *obd, obd_count len,
4238 struct lustre_cfg *lcfg = buf;
4239 struct lprocfs_static_vars lvars;
4242 switch (lcfg->lcfg_command) {
4243 case LCFG_SPTLRPC_CONF: {
4244 struct filter_obd *filter = &obd->u.filter;
4245 struct sptlrpc_conf_log *log;
4246 struct sptlrpc_rule_set tmp_rset;
4248 log = sptlrpc_conf_log_extract(lcfg);
4254 sptlrpc_rule_set_init(&tmp_rset);
4256 rc = sptlrpc_rule_set_from_log(&tmp_rset, log);
4258 CERROR("obd %s: failed get sptlrpc rules: %d\n",
4263 write_lock(&filter->fo_sptlrpc_lock);
4264 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
4265 filter->fo_sptlrpc_rset = tmp_rset;
4266 write_unlock(&filter->fo_sptlrpc_lock);
4268 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
4272 lprocfs_filter_init_vars(&lvars);
4274 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4282 static struct lvfs_callback_ops filter_lvfs_ops = {
4283 l_fid2dentry: filter_lvfs_fid2dentry,
4286 static struct obd_ops filter_obd_ops = {
4287 .o_owner = THIS_MODULE,
4288 .o_get_info = filter_get_info,
4289 .o_set_info_async = filter_set_info_async,
4290 .o_setup = filter_setup,
4291 .o_precleanup = filter_precleanup,
4292 .o_cleanup = filter_cleanup,
4293 .o_connect = filter_connect,
4294 .o_reconnect = filter_reconnect,
4295 .o_disconnect = filter_disconnect,
4296 .o_ping = filter_ping,
4297 .o_init_export = filter_init_export,
4298 .o_destroy_export = filter_destroy_export,
4299 .o_statfs = filter_statfs,
4300 .o_getattr = filter_getattr,
4301 .o_unpackmd = filter_unpackmd,
4302 .o_create = filter_create,
4303 .o_setattr = filter_setattr,
4304 .o_destroy = filter_destroy,
4305 .o_brw = filter_brw,
4306 .o_punch = filter_truncate,
4307 .o_sync = filter_sync,
4308 .o_preprw = filter_preprw,
4309 .o_commitrw = filter_commitrw,
4310 .o_llog_init = filter_llog_init,
4311 .o_llog_connect = filter_llog_connect,
4312 .o_llog_finish = filter_llog_finish,
4313 .o_iocontrol = filter_iocontrol,
4314 .o_health_check = filter_health_check,
4315 .o_process_config = filter_process_config,
4318 quota_interface_t *filter_quota_interface_ref;
4319 extern quota_interface_t filter_quota_interface;
4321 static int __init obdfilter_init(void)
4323 struct lprocfs_static_vars lvars;
4326 lprocfs_filter_init_vars(&lvars);
4328 request_module("lquota");
4329 OBD_ALLOC(obdfilter_created_scratchpad,
4330 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4331 sizeof(*obdfilter_created_scratchpad));
4332 if (obdfilter_created_scratchpad == NULL)
4335 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4336 sizeof(struct filter_mod_data),
4339 GOTO(out, rc = -ENOMEM);
4341 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4342 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4344 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4345 LUSTRE_OST_NAME, NULL);
4349 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4350 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4351 ll_fmd_cachep = NULL;
4353 if (filter_quota_interface_ref)
4354 PORTAL_SYMBOL_PUT(filter_quota_interface);
4356 OBD_FREE(obdfilter_created_scratchpad,
4357 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4358 sizeof(*obdfilter_created_scratchpad));
4364 static void __exit obdfilter_exit(void)
4366 if (filter_quota_interface_ref)
4367 PORTAL_SYMBOL_PUT(filter_quota_interface);
4369 if (ll_fmd_cachep) {
4370 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4371 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4372 ll_fmd_cachep = NULL;
4375 class_unregister_type(LUSTRE_OST_NAME);
4376 OBD_FREE(obdfilter_created_scratchpad,
4377 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4378 sizeof(*obdfilter_created_scratchpad));
4381 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4382 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4383 MODULE_LICENSE("GPL");
4385 module_init(obdfilter_init);
4386 module_exit(obdfilter_exit);