1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdfilter/filter.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Andreas Dilger <adilger@clusterfs.com>
43 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
44 * (which need to get journal_lock, may block if journal full).
46 * Invariant: Call filter_start_transno() before any journal ops to avoid the
47 * same deadlock problem. We can (and want) to get rid of the
48 * transno sem in favour of the dir/inode i_mutex to avoid single
49 * threaded operation on the OST.
52 #define DEBUG_SUBSYSTEM S_FILTER
54 #ifndef AUTOCONF_INCLUDED
55 #include <linux/config.h>
57 #include <linux/module.h>
59 #include <linux/dcache.h>
60 #include <linux/init.h>
61 #include <linux/version.h>
62 #include <linux/sched.h>
63 #include <linux/mount.h>
64 #include <linux/buffer_head.h>
66 #include <obd_cksum.h>
67 #include <obd_class.h>
69 #include <lustre_dlm.h>
70 #include <lustre_fsfilt.h>
71 #include <lprocfs_status.h>
72 #include <lustre_log.h>
73 #include <libcfs/list.h>
74 #include <lustre_disk.h>
75 #include <lustre_quota.h>
76 #include <linux/slab.h>
77 #include <lustre_param.h>
78 #include <lustre/ll_fiemap.h>
80 #include "filter_internal.h"
82 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
83 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS1_N_BASE
84 static struct lvfs_callback_ops filter_lvfs_ops;
85 cfs_mem_cache_t *ll_fmd_cachep;
87 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
88 void *cb_data, int error)
90 obd_transno_commit_cb(obd, transno, error);
93 /* Assumes caller has already pushed us into the kernel context. */
94 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
95 int rc, int force_sync)
97 struct filter_obd *filter = &exp->exp_obd->u.filter;
98 struct filter_export_data *fed = &exp->exp_filter_data;
99 struct lsd_client_data *lcd = fed->fed_lcd;
102 int err, log_pri = D_RPCTRACE;
104 /* Propagate error code. */
108 if (!exp->exp_obd->obd_replayable || oti == NULL)
111 /* we don't allocate new transnos for replayed requests */
112 if (oti->oti_transno == 0) {
113 spin_lock(&filter->fo_translock);
114 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
115 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
116 spin_unlock(&filter->fo_translock);
117 oti->oti_transno = last_rcvd;
119 spin_lock(&filter->fo_translock);
120 last_rcvd = oti->oti_transno;
121 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
122 filter->fo_fsd->lsd_last_transno =
123 cpu_to_le64(last_rcvd);
124 spin_unlock(&filter->fo_translock);
126 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
128 /* could get xid from oti, if it's ever needed */
129 lcd->lcd_last_xid = 0;
131 off = fed->fed_lr_off;
133 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
134 fed->fed_lr_idx, fed->fed_lr_off);
138 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
144 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
145 lcd, sizeof(*lcd), &off,
146 force_sync | exp->exp_need_sync);
148 filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
156 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
157 last_rcvd, lcd->lcd_uuid, fed->fed_lr_idx, err);
162 void f_dput(struct dentry *dentry)
164 /* Can't go inside filter_ddelete because it can block */
165 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
166 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
167 LASSERT(atomic_read(&dentry->d_count) > 0);
172 static void init_brw_stats(struct brw_stats *brw_stats)
175 for (i = 0; i < BRW_LAST; i++)
176 spin_lock_init(&brw_stats->hist[i].oh_lock);
179 static int lprocfs_init_rw_stats(struct obd_device *obd,
180 struct lprocfs_stats **stats)
184 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
185 LPROC_FILTER_LAST - 1;
186 *stats = lprocfs_alloc_stats(num_stats, 0);
190 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
191 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
192 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
193 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
194 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
199 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
200 plus the procfs overhead :( */
201 static int filter_export_stats_init(struct obd_device *obd,
202 struct obd_export *exp,
205 struct filter_export_data *fed = &exp->exp_filter_data;
209 init_brw_stats(&fed->fed_brw_stats);
211 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
212 /* Self-export gets no proc entry */
215 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
217 /* Mask error for already created
225 struct nid_stat *tmp = exp->exp_nid_stats;
226 LASSERT(tmp != NULL);
228 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
229 if (tmp->nid_brw_stats == NULL)
232 init_brw_stats(tmp->nid_brw_stats);
233 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
234 0644, &filter_per_nid_stats_fops,
237 CWARN("Error adding the brw_stats file\n");
239 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
243 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
247 /* Always add in ldlm_stats */
248 tmp->nid_ldlm_stats = lprocfs_alloc_stats(LDLM_LAST_OPC -
250 if (tmp->nid_ldlm_stats == NULL)
253 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
254 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
255 tmp->nid_ldlm_stats);
263 /* Add client data to the FILTER. We use a bitmap to locate a free space
264 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
265 * Otherwise, we have just read the data from the last_rcvd file and
266 * we know its offset. */
267 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
270 struct filter_obd *filter = &obd->u.filter;
271 struct filter_export_data *fed = &exp->exp_filter_data;
272 unsigned long *bitmap = filter->fo_last_rcvd_slots;
273 int new_client = (cl_idx == -1);
277 LASSERT(bitmap != NULL);
278 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
281 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
284 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
285 * there's no need for extra complication here
288 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
290 if (cl_idx >= LR_MAX_CLIENTS) {
291 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
295 if (test_and_set_bit(cl_idx, bitmap)) {
296 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
301 if (test_and_set_bit(cl_idx, bitmap)) {
302 CERROR("FILTER client %d: bit already set in bitmap!\n",
308 fed->fed_lr_idx = cl_idx;
309 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
310 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
311 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
313 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
314 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
317 struct lvfs_run_ctxt saved;
318 loff_t off = fed->fed_lr_off;
322 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
323 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
325 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
326 /* Transaction needed to fix bug 1403 */
327 handle = fsfilt_start(obd,
328 filter->fo_rcvd_filp->f_dentry->d_inode,
329 FSFILT_OP_SETATTR, NULL);
330 if (IS_ERR(handle)) {
331 rc = PTR_ERR(handle);
332 CERROR("unable to start transaction: rc %d\n", rc);
334 rc = fsfilt_add_journal_cb(obd, 0, handle,
335 target_client_add_cb, exp);
337 spin_lock(&exp->exp_lock);
338 exp->exp_need_sync = 1;
339 spin_unlock(&exp->exp_lock);
341 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
343 sizeof(*fed->fed_lcd),
344 &off, rc /* sync if no cb */);
346 filter->fo_rcvd_filp->f_dentry->d_inode,
349 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
352 CERROR("error writing %s client idx %u: rc %d\n",
353 LAST_RCVD, fed->fed_lr_idx, rc);
360 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
362 static int filter_client_free(struct obd_export *exp)
364 struct filter_export_data *fed = &exp->exp_filter_data;
365 struct filter_obd *filter = &exp->exp_obd->u.filter;
366 struct obd_device *obd = exp->exp_obd;
367 struct lvfs_run_ctxt saved;
372 if (fed->fed_lcd == NULL)
375 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
376 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
379 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
380 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
382 LASSERT(filter->fo_last_rcvd_slots != NULL);
384 off = fed->fed_lr_off;
386 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
387 * we leak a client slot that will be cleaned on the next recovery. */
389 CERROR("%s: client idx %d has med_off %lld\n",
390 obd->obd_name, fed->fed_lr_idx, off);
391 GOTO(free, rc = -EINVAL);
394 /* Clear the bit _after_ zeroing out the client so we don't
395 race with filter_client_add and zero out new clients.*/
396 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
397 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
402 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
403 /* Don't force sync on disconnect if aborting recovery,
404 * or it does num_clients * num_osts. b=17194 */
405 int need_sync = (!exp->exp_libclient || exp->exp_need_sync) &&
406 !(exp->exp_flags&OBD_OPT_ABORT_RECOV);
407 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
408 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
409 sizeof(zero_lcd), &off, 0);
411 /* Make sure the server's last_transno is up to date. Do this
412 * after the client is freed so we know all the client's
413 * transactions have been committed. */
415 filter_update_server_data(obd, filter->fo_rcvd_filp,
416 filter->fo_fsd, need_sync);
417 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
419 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
420 "zero out client %s at idx %u/%llu in %s %ssync rc %d\n",
421 fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
422 LAST_RCVD, need_sync ? "" : "a", rc);
425 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
426 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
433 OBD_FREE_PTR(fed->fed_lcd);
439 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
440 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
441 struct filter_mod_data *fmd)
443 LASSERT_SPIN_LOCKED(&fed->fed_lock);
444 if (--fmd->fmd_refcount == 0) {
445 /* XXX when we have persistent reservations and the handle
446 * is stored herein we need to drop it here. */
447 fed->fed_mod_count--;
448 list_del(&fmd->fmd_list);
449 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
453 /* drop fmd reference, free it if last ref */
454 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
456 struct filter_export_data *fed;
461 fed = &exp->exp_filter_data;
462 spin_lock(&fed->fed_lock);
463 filter_fmd_put_nolock(fed, fmd); /* caller reference */
464 spin_unlock(&fed->fed_lock);
467 /* expire entries from the end of the list if there are too many
468 * or they are too old */
469 static void filter_fmd_expire_nolock(struct filter_obd *filter,
470 struct filter_export_data *fed,
471 struct filter_mod_data *keep)
473 struct filter_mod_data *fmd, *tmp;
475 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
479 if (time_before(jiffies, fmd->fmd_expire) &&
480 fed->fed_mod_count < filter->fo_fmd_max_num)
483 list_del_init(&fmd->fmd_list);
484 filter_fmd_put_nolock(fed, fmd); /* list reference */
488 void filter_fmd_expire(struct obd_export *exp)
490 spin_lock(&exp->exp_filter_data.fed_lock);
491 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
492 &exp->exp_filter_data, NULL);
493 spin_unlock(&exp->exp_filter_data.fed_lock);
496 /* find specified objid, group in export fmd list.
497 * caller must hold fed_lock and take fmd reference itself */
498 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
499 struct filter_export_data *fed,
500 obd_id objid, obd_gr group)
502 struct filter_mod_data *found = NULL, *fmd;
504 LASSERT_SPIN_LOCKED(&fed->fed_lock);
506 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
507 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
509 list_del(&fmd->fmd_list);
510 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
511 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
516 filter_fmd_expire_nolock(filter, fed, found);
521 /* Find fmd based on objid and group, or return NULL if not found. */
522 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
523 obd_id objid, obd_gr group)
525 struct filter_mod_data *fmd;
527 spin_lock(&exp->exp_filter_data.fed_lock);
528 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
529 &exp->exp_filter_data, objid, group);
531 fmd->fmd_refcount++; /* caller reference */
532 spin_unlock(&exp->exp_filter_data.fed_lock);
537 /* Find fmd based on objid and group, or create a new one if none is found.
538 * It is possible for this function to return NULL under memory pressure,
539 * or if objid = 0 is passed (which will only cause old entries to expire).
540 * Currently this is not fatal because any fmd state is transient and
541 * may also be freed when it gets sufficiently old. */
542 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
543 obd_id objid, obd_gr group)
545 struct filter_export_data *fed = &exp->exp_filter_data;
546 struct filter_mod_data *found = NULL, *fmd_new = NULL;
548 OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO, sizeof(*fmd_new));
550 spin_lock(&fed->fed_lock);
551 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
554 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
555 fmd_new->fmd_id = objid;
556 fmd_new->fmd_gr = group;
557 fmd_new->fmd_refcount++; /* list reference */
559 fed->fed_mod_count++;
561 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
565 found->fmd_refcount++; /* caller reference */
566 found->fmd_expire = jiffies +
567 exp->exp_obd->u.filter.fo_fmd_max_age;
570 spin_unlock(&fed->fed_lock);
576 /* drop fmd list reference so it will disappear when last reference is put.
577 * This isn't so critical because it would in fact only affect the one client
578 * that is doing the unlink and at worst we have an stale entry referencing
579 * an object that should never be used again. */
580 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
582 struct filter_mod_data *found = NULL;
584 spin_lock(&exp->exp_filter_data.fed_lock);
585 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
587 list_del_init(&found->fmd_list);
588 filter_fmd_put_nolock(&exp->exp_filter_data, found);
590 spin_unlock(&exp->exp_filter_data.fed_lock);
593 #define filter_fmd_drop(exp, objid, group)
596 /* remove all entries from fmd list */
597 static void filter_fmd_cleanup(struct obd_export *exp)
599 struct filter_export_data *fed = &exp->exp_filter_data;
600 struct filter_mod_data *fmd = NULL, *tmp;
602 spin_lock(&fed->fed_lock);
603 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
604 list_del_init(&fmd->fmd_list);
605 filter_fmd_put_nolock(fed, fmd);
607 spin_unlock(&fed->fed_lock);
610 static int filter_init_export(struct obd_export *exp)
612 spin_lock_init(&exp->exp_filter_data.fed_lock);
613 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
615 spin_lock(&exp->exp_lock);
616 exp->exp_connecting = 1;
617 spin_unlock(&exp->exp_lock);
619 return ldlm_init_export(exp);
622 static int filter_free_server_data(struct filter_obd *filter)
624 OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
625 filter->fo_fsd = NULL;
626 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
627 filter->fo_last_rcvd_slots = NULL;
631 /* assumes caller is already in kernel ctxt */
632 int filter_update_server_data(struct obd_device *obd, struct file *filp,
633 struct lr_server_data *fsd, int force_sync)
639 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
640 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
641 le64_to_cpu(fsd->lsd_last_transno));
642 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
643 le64_to_cpu(fsd->lsd_mount_count));
645 fsd->lsd_compat14 = fsd->lsd_last_transno;
646 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
648 CERROR("error writing lr_server_data: rc = %d\n", rc);
653 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
656 struct filter_obd *filter = &obd->u.filter;
662 if (filter->fo_last_objid_files[group] == NULL) {
663 CERROR("Object group "LPU64" not fully setup; not updating "
664 "last_objid\n", group);
668 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
669 obd->obd_name, group, filter->fo_last_objids[group]);
671 tmp = cpu_to_le64(filter->fo_last_objids[group]);
672 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
673 &tmp, sizeof(tmp), &off, force_sync);
675 CERROR("error writing group "LPU64" last objid: rc = %d\n",
679 extern int ost_handle(struct ptlrpc_request *req);
680 /* assumes caller has already in kernel ctxt */
681 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
683 struct filter_obd *filter = &obd->u.filter;
684 struct lr_server_data *fsd;
685 struct lsd_client_data *lcd = NULL;
686 struct inode *inode = filp->f_dentry->d_inode;
687 unsigned long last_rcvd_size = i_size_read(inode);
693 /* ensure padding in the struct is the correct size */
694 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
695 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
696 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
697 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
699 OBD_ALLOC(fsd, sizeof(*fsd));
702 filter->fo_fsd = fsd;
704 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
705 if (filter->fo_last_rcvd_slots == NULL) {
706 OBD_FREE(fsd, sizeof(*fsd));
710 if (last_rcvd_size == 0) {
711 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
713 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
714 fsd->lsd_last_transno = 0;
715 mount_count = fsd->lsd_mount_count = 0;
716 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
717 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
718 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
719 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
720 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
721 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
723 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
725 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
729 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
730 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
731 "using the wrong disk %s. Were the "
732 "/dev/ assignments rearranged?\n",
733 obd->obd_uuid.uuid, fsd->lsd_uuid);
734 GOTO(err_fsd, rc = -EINVAL);
736 mount_count = le64_to_cpu(fsd->lsd_mount_count);
737 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
739 /* Assume old last_rcvd format unless I_C_LR is set */
740 if (!(fsd->lsd_feature_incompat &
741 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
742 fsd->lsd_last_transno = fsd->lsd_compat14;
746 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
747 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
748 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
749 ~FILTER_INCOMPAT_SUPP);
750 GOTO(err_fsd, rc = -EINVAL);
752 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
753 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
754 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
755 ~FILTER_ROCOMPAT_SUPP);
756 /* Do something like remount filesystem read-only */
757 GOTO(err_fsd, rc = -EINVAL);
760 CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
761 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
762 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
763 obd->obd_name, mount_count + 1);
764 CDEBUG(D_INODE, "%s: server data size: %u\n",
765 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
766 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
767 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
768 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
769 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
770 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
771 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
772 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
773 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
774 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
775 le16_to_cpu(fsd->lsd_client_size));
777 if (!obd->obd_replayable) {
778 CWARN("%s: recovery support OFF\n", obd->obd_name);
782 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
783 off < last_rcvd_size; cl_idx++) {
785 struct obd_export *exp;
786 struct filter_export_data *fed;
791 GOTO(err_client, rc = -ENOMEM);
794 /* Don't assume off is incremented properly by
795 * fsfilt_read_record(), in case sizeof(*lcd)
796 * isn't the same as fsd->lsd_client_size. */
797 off = le32_to_cpu(fsd->lsd_client_start) +
798 cl_idx * le16_to_cpu(fsd->lsd_client_size);
799 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
801 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
802 LAST_RCVD, cl_idx, off, rc);
803 break; /* read error shouldn't cause startup to fail */
806 if (lcd->lcd_uuid[0] == '\0') {
807 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
812 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
814 /* These exports are cleaned up by filter_disconnect(), so they
815 * need to be set up like real exports as filter_connect() does.
817 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
819 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
820 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
821 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
823 if (PTR_ERR(exp) == -EALREADY) {
824 /* export already exists, zero out this one */
825 CERROR("Zeroing out duplicate export due to "
827 lcd->lcd_uuid[0] = '\0';
829 GOTO(err_client, rc = PTR_ERR(exp));
832 fed = &exp->exp_filter_data;
834 fed->fed_group = 0; /* will be assigned at connect */
835 filter_export_stats_init(obd, exp, NULL);
836 rc = filter_client_add(obd, exp, cl_idx);
837 /* can't fail for existing client */
838 LASSERTF(rc == 0, "rc = %d\n", rc);
841 spin_lock(&exp->exp_lock);
842 exp->exp_connecting = 0;
843 exp->exp_in_recovery = 0;
844 spin_unlock(&exp->exp_lock);
845 obd->obd_max_recoverable_clients++;
846 class_export_put(exp);
849 /* Need to check last_rcvd even for duplicated exports. */
850 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
853 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
854 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
860 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
862 target_recovery_init(obd, ost_handle);
865 filter->fo_mount_count = mount_count + 1;
866 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
868 /* save it, so mount count and last_transno is current */
869 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
871 GOTO(err_client, rc);
876 target_recovery_fini(obd);
878 filter_free_server_data(filter);
882 static int filter_cleanup_groups(struct obd_device *obd)
884 struct filter_obd *filter = &obd->u.filter;
886 struct dentry *dentry;
890 if (filter->fo_dentry_O_groups != NULL) {
891 for (i = 0; i < filter->fo_group_count; i++) {
892 dentry = filter->fo_dentry_O_groups[i];
896 OBD_FREE(filter->fo_dentry_O_groups,
897 filter->fo_group_count *
898 sizeof(*filter->fo_dentry_O_groups));
899 filter->fo_dentry_O_groups = NULL;
901 if (filter->fo_last_objid_files != NULL) {
902 for (i = 0; i < filter->fo_group_count; i++) {
903 filp = filter->fo_last_objid_files[i];
907 OBD_FREE(filter->fo_last_objid_files,
908 filter->fo_group_count *
909 sizeof(*filter->fo_last_objid_files));
910 filter->fo_last_objid_files = NULL;
912 if (filter->fo_dentry_O_sub != NULL) {
913 for (i = 0; i < filter->fo_group_count; i++) {
914 for (j = 0; j < filter->fo_subdir_count; j++) {
915 dentry = filter->fo_dentry_O_sub[i].dentry[j];
920 OBD_FREE(filter->fo_dentry_O_sub,
921 filter->fo_group_count *
922 sizeof(*filter->fo_dentry_O_sub));
923 filter->fo_dentry_O_sub = NULL;
925 if (filter->fo_last_objids != NULL) {
926 OBD_FREE(filter->fo_last_objids,
927 filter->fo_group_count *
928 sizeof(*filter->fo_last_objids));
929 filter->fo_last_objids = NULL;
931 if (filter->fo_dentry_O != NULL) {
932 f_dput(filter->fo_dentry_O);
933 filter->fo_dentry_O = NULL;
938 static int filter_update_last_group(struct obd_device *obd, int group)
940 struct filter_obd *filter = &obd->u.filter;
941 struct file *filp = NULL;
942 int last_group = 0, rc;
946 if (group <= filter->fo_committed_group)
949 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
953 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
957 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
959 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
962 LASSERTF(off == 0 || CHECK_MDS_GROUP(last_group),
963 "off = %llu and last_group = %d\n", off, last_group);
965 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
966 obd->obd_name, last_group, group);
970 /* must be sync: bXXXX */
971 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
973 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
977 filter->fo_committed_group = group;
984 static int filter_read_group_internal(struct obd_device *obd, int group,
987 struct filter_obd *filter = &obd->u.filter;
988 __u64 *new_objids = NULL;
989 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
990 struct dentry **new_groups = NULL;
991 struct file **new_files = NULL;
992 struct dentry *dentry;
994 int old_count = filter->fo_group_count, rc, stage = 0, i;
1000 snprintf(name, 24, "%d", group);
1004 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1006 if (IS_ERR(dentry)) {
1007 CERROR("Cannot lookup expected object group %d: %ld\n",
1008 group, PTR_ERR(dentry));
1009 RETURN(PTR_ERR(dentry));
1012 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1014 if (IS_ERR(dentry)) {
1015 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1017 RETURN(PTR_ERR(dentry));
1022 snprintf(name, 24, "O/%d/LAST_ID", group);
1024 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1026 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1027 GOTO(cleanup, rc = PTR_ERR(filp));
1031 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1033 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1037 if (filter->fo_subdir_count) {
1038 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1039 if (tmp_subdirs == NULL)
1040 GOTO(cleanup, rc = -ENOMEM);
1043 for (i = 0; i < filter->fo_subdir_count; i++) {
1045 snprintf(dir, sizeof(dir), "d%u", i);
1047 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1050 if (IS_ERR(tmp_subdirs->dentry[i])) {
1051 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1052 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1057 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1058 tmp_subdirs->dentry[i]);
1062 /* 'group' is an index; we need an array of length 'group + 1' */
1063 if (group + 1 > old_count) {
1064 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1065 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1066 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1067 OBD_ALLOC(new_files, len * sizeof(*new_files));
1069 if (new_objids == NULL || new_subdirs == NULL ||
1070 new_groups == NULL || new_files == NULL)
1071 GOTO(cleanup, rc = -ENOMEM);
1074 memcpy(new_objids, filter->fo_last_objids,
1075 old_count * sizeof(*new_objids));
1076 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1077 old_count * sizeof(*new_subdirs));
1078 memcpy(new_groups, filter->fo_dentry_O_groups,
1079 old_count * sizeof(*new_groups));
1080 memcpy(new_files, filter->fo_last_objid_files,
1081 old_count * sizeof(*new_files));
1083 OBD_FREE(filter->fo_last_objids,
1084 old_count * sizeof(*new_objids));
1085 OBD_FREE(filter->fo_dentry_O_sub,
1086 old_count * sizeof(*new_subdirs));
1087 OBD_FREE(filter->fo_dentry_O_groups,
1088 old_count * sizeof(*new_groups));
1089 OBD_FREE(filter->fo_last_objid_files,
1090 old_count * sizeof(*new_files));
1092 filter->fo_last_objids = new_objids;
1093 filter->fo_dentry_O_sub = new_subdirs;
1094 filter->fo_dentry_O_groups = new_groups;
1095 filter->fo_last_objid_files = new_files;
1096 filter->fo_group_count = len;
1099 filter->fo_dentry_O_groups[group] = dentry;
1100 filter->fo_last_objid_files[group] = filp;
1101 if (filter->fo_subdir_count) {
1102 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1103 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1106 filter_update_last_group(obd, group);
1108 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1109 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1110 rc = filter_update_last_objid(obd, group, 1);
1114 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1115 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1116 obd->obd_name, group, last_objid);
1121 if (new_objids != NULL)
1122 OBD_FREE(new_objids, len * sizeof(*new_objids));
1123 if (new_subdirs != NULL)
1124 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1125 if (new_groups != NULL)
1126 OBD_FREE(new_groups, len * sizeof(*new_groups));
1127 if (new_files != NULL)
1128 OBD_FREE(new_files, len * sizeof(*new_files));
1130 if (filter->fo_subdir_count) {
1131 for (i = 0; i < filter->fo_subdir_count; i++) {
1132 if (tmp_subdirs->dentry[i] != NULL)
1133 dput(tmp_subdirs->dentry[i]);
1135 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1138 filp_close(filp, 0);
1145 static int filter_read_groups(struct obd_device *obd, int last_group,
1148 struct filter_obd *filter = &obd->u.filter;
1149 int old_count, group, rc = 0;
1151 down(&filter->fo_init_lock);
1152 old_count = filter->fo_group_count;
1153 for (group = old_count; group <= last_group; group++) {
1155 rc = filter_read_group_internal(obd, group, create);
1159 up(&filter->fo_init_lock);
1163 /* FIXME: object groups */
1164 static int filter_prep_groups(struct obd_device *obd)
1166 struct filter_obd *filter = &obd->u.filter;
1167 struct dentry *dentry, *O_dentry;
1169 int last_group, rc = 0, cleanup_phase = 0;
1173 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1175 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1176 if (IS_ERR(O_dentry)) {
1177 rc = PTR_ERR(O_dentry);
1178 CERROR("cannot open/create O: rc = %d\n", rc);
1181 filter->fo_dentry_O = O_dentry;
1182 cleanup_phase = 1; /* O_dentry */
1184 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1185 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1186 * some time post 1.0 when all old-style OSTs have converted along
1187 * with the init_objid hack. */
1188 dentry = ll_lookup_one_len("R", O_dentry, 1);
1190 GOTO(cleanup, rc = PTR_ERR(dentry));
1191 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1192 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1195 CWARN("converting OST to new object layout\n");
1196 if (IS_ERR(O0_dentry)) {
1197 rc = PTR_ERR(O0_dentry);
1198 CERROR("error looking up O/0: rc %d\n", rc);
1199 GOTO(cleanup_R, rc);
1202 if (O0_dentry->d_inode) {
1203 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1204 GOTO(cleanup_O0, rc = -EEXIST);
1207 LOCK_INODE_MUTEX(O_dentry->d_inode);
1208 rc = ll_vfs_rename(O_dentry->d_inode, dentry, filter->fo_vfsmnt,
1209 O_dentry->d_inode, O0_dentry,
1211 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1214 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1215 GOTO(cleanup_O0, rc);
1217 filter->fo_fsd->lsd_feature_incompat |=
1218 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1219 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1221 GOTO(cleanup_O0, rc);
1233 cleanup_phase = 2; /* groups */
1235 /* we have to initialize all groups before first connections from
1236 * clients because they may send create/destroy for any group -bzzz */
1237 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1239 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1240 GOTO(cleanup, rc = PTR_ERR(filp));
1242 cleanup_phase = 3; /* filp */
1244 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1246 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1250 last_group = FILTER_MIN_GROUPS;
1252 LASSERT_MDS_GROUP(last_group);
1255 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1256 FILTER_MIN_GROUPS, last_group);
1257 filter->fo_committed_group = last_group;
1258 rc = filter_read_groups(obd, last_group, 1);
1262 filp_close(filp, 0);
1266 switch (cleanup_phase) {
1268 filp_close(filp, 0);
1270 filter_cleanup_groups(obd);
1272 f_dput(filter->fo_dentry_O);
1273 filter->fo_dentry_O = NULL;
1281 /* setup the object store with correct subdirectories */
1282 static int filter_prep(struct obd_device *obd)
1284 struct lvfs_run_ctxt saved;
1285 struct filter_obd *filter = &obd->u.filter;
1287 struct inode *inode;
1291 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1292 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1293 if (!file || IS_ERR(file)) {
1295 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1299 filter->fo_rcvd_filp = file;
1300 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1301 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1302 file->f_dentry->d_inode->i_mode);
1303 GOTO(err_filp, rc = -ENOENT);
1306 inode = file->f_dentry->d_parent->d_inode;
1307 /* We use i_op->unlink directly in filter_vfs_unlink() */
1308 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1309 CERROR("%s: filesystem does not support create/unlink ops\n",
1311 GOTO(err_filp, rc = -EOPNOTSUPP);
1314 rc = filter_init_server_data(obd, file);
1316 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1319 /* open and create health check io file*/
1320 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1323 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1325 GOTO(err_server_data, rc);
1327 filter->fo_health_check_filp = file;
1328 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1329 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1330 file->f_dentry->d_inode->i_mode);
1331 GOTO(err_health_check, rc = -ENOENT);
1333 rc = lvfs_check_io_health(obd, file);
1335 GOTO(err_health_check, rc);
1337 rc = filter_prep_groups(obd);
1339 GOTO(err_health_check, rc);
1341 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1346 if (filp_close(filter->fo_health_check_filp, 0))
1347 CERROR("can't close %s after error\n", HEALTH_CHECK);
1348 filter->fo_health_check_filp = NULL;
1350 target_recovery_fini(obd);
1351 filter_free_server_data(filter);
1353 if (filp_close(filter->fo_rcvd_filp, 0))
1354 CERROR("can't close %s after error\n", LAST_RCVD);
1355 filter->fo_rcvd_filp = NULL;
1359 /* cleanup the filter: write last used object id to status file */
1360 static void filter_post(struct obd_device *obd)
1362 struct lvfs_run_ctxt saved;
1363 struct filter_obd *filter = &obd->u.filter;
1366 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1367 * best to start a transaction with h_sync, because we removed this
1370 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1371 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1374 CERROR("error writing server data: rc = %d\n", rc);
1376 for (i = 0; i < filter->fo_group_count; i++) {
1377 rc = filter_update_last_objid(obd, i,
1378 (i == filter->fo_group_count - 1));
1380 CERROR("error writing group %d lastobjid: rc = %d\n",
1384 rc = filp_close(filter->fo_rcvd_filp, 0);
1385 filter->fo_rcvd_filp = NULL;
1387 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1389 rc = filp_close(filter->fo_health_check_filp, 0);
1390 filter->fo_health_check_filp = NULL;
1392 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1394 filter_cleanup_groups(obd);
1395 filter_free_server_data(filter);
1396 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1398 filter_free_capa_keys(filter);
1399 cleanup_capa_hash(filter->fo_capa_hash);
1402 static void filter_set_last_id(struct filter_obd *filter,
1403 obd_id id, obd_gr group)
1405 LASSERT(filter->fo_fsd != NULL);
1406 LASSERT(group <= filter->fo_group_count);
1408 spin_lock(&filter->fo_objidlock);
1409 filter->fo_last_objids[group] = id;
1410 spin_unlock(&filter->fo_objidlock);
1413 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1416 LASSERT(filter->fo_fsd != NULL);
1417 LASSERT(group <= filter->fo_group_count);
1419 /* FIXME: object groups */
1420 spin_lock(&filter->fo_objidlock);
1421 id = filter->fo_last_objids[group];
1422 spin_unlock(&filter->fo_objidlock);
1426 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1428 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1432 /* We never dget the object parent, so DON'T dput it either */
1433 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1435 struct filter_obd *filter = &obd->u.filter;
1436 struct filter_subdirs *subdirs;
1437 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1439 if ((group > FILTER_GROUP_MDS0 && group < FILTER_GROUP_MDS1_N_BASE) ||
1440 filter->fo_subdir_count == 0)
1441 return filter->fo_dentry_O_groups[group];
1443 subdirs = &filter->fo_dentry_O_sub[group];
1444 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1447 /* We never dget the object parent, so DON'T dput it either */
1448 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1451 unsigned long now = jiffies;
1452 struct dentry *dparent = filter_parent(obd, group, objid);
1455 if (IS_ERR(dparent))
1457 if (dparent == NULL)
1458 return ERR_PTR(-ENOENT);
1460 rc = filter_lock_dentry(obd, dparent);
1461 fsfilt_check_slow(obd, now, "parent lock");
1462 return rc ? ERR_PTR(rc) : dparent;
1465 /* We never dget the object parent, so DON'T dput it either */
1466 static void filter_parent_unlock(struct dentry *dparent)
1468 UNLOCK_INODE_MUTEX(dparent->d_inode);
1471 /* How to get files, dentries, inodes from object id's.
1473 * If dir_dentry is passed, the caller has already locked the parent
1474 * appropriately for this operation (normally a write lock). If
1475 * dir_dentry is NULL, we do a read lock while we do the lookup to
1476 * avoid races with create/destroy and such changing the directory
1477 * internal to the filesystem code. */
1478 struct dentry *filter_fid2dentry(struct obd_device *obd,
1479 struct dentry *dir_dentry,
1480 obd_gr group, obd_id id)
1482 struct dentry *dparent = dir_dentry;
1483 struct dentry *dchild;
1488 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1489 obd->u.filter.fo_destroys_in_progress == 0) {
1490 /* don't fail lookups for orphan recovery, it causes
1491 * later LBUGs when objects still exist during precreate */
1492 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1493 RETURN(ERR_PTR(-ENOENT));
1496 CERROR("fatal: invalid object id 0\n");
1497 RETURN(ERR_PTR(-ESTALE));
1500 len = sprintf(name, LPU64, id);
1501 if (dir_dentry == NULL) {
1502 dparent = filter_parent_lock(obd, group, id);
1503 if (IS_ERR(dparent)) {
1504 CERROR("%s: error getting object "LPU64":"LPU64
1505 " parent: rc %ld\n", obd->obd_name,
1506 id, group, PTR_ERR(dparent));
1510 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1511 dparent->d_name.len, dparent->d_name.name, name);
1512 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1513 if (dir_dentry == NULL)
1514 filter_parent_unlock(dparent);
1515 if (IS_ERR(dchild)) {
1516 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1521 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1522 CERROR("%s: got bad object "LPU64" inode %lu\n",
1523 obd->obd_name, id, dchild->d_inode->i_ino);
1525 RETURN(ERR_PTR(-ENOENT));
1528 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1529 name, dchild, atomic_read(&dchild->d_count));
1531 LASSERT(atomic_read(&dchild->d_count) > 0);
1536 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1539 struct lustre_handle lockh;
1540 int flags = LDLM_AST_DISCARD_DATA, rc;
1541 struct ldlm_res_id res_id;
1542 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1545 osc_build_res_name(objid, group, &res_id);
1546 /* Tell the clients that the object is gone now and that they should
1547 * throw away any cached pages. */
1548 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1549 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1550 ldlm_completion_ast, NULL, NULL, 0, NULL,
1553 /* We only care about the side-effects, just drop the lock. */
1555 ldlm_lock_decref(&lockh, LCK_PW);
1560 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1561 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1562 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1563 * reverse. Caller must take i_sem before starting the transaction and we
1564 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1565 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1566 struct vfsmount *mnt)
1571 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1572 * itself we already hold dir->i_mutex for child create/unlink ops */
1573 LASSERT(dentry->d_inode != NULL);
1574 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1575 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1579 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1580 GOTO(out, rc = -ENOENT);
1582 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1587 GOTO(out, rc = -EPERM);
1589 /* check_sticky() */
1590 if ((dentry->d_inode->i_uid != current->fsuid &&
1591 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1592 IS_IMMUTABLE(dentry->d_inode))
1593 GOTO(out, rc = -EPERM);
1595 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1596 * that was done because of journal_start (which is already done
1597 * here) or some other ordering issue. */
1600 rc = ll_security_inode_unlink(dir, dentry, mnt);
1604 rc = dir->i_op->unlink(dir, dentry);
1606 /* need to drop i_mutex before we lose inode reference */
1607 UNLOCK_INODE_MUTEX(dentry->d_inode);
1614 /* Caller must hold LCK_PW on parent and push us into kernel context.
1615 * Caller must hold child i_mutex, we drop it always.
1616 * Caller is also required to ensure that dchild->d_inode exists. */
1617 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1618 obd_gr group, struct dentry *dparent,
1619 struct dentry *dchild)
1621 struct inode *inode = dchild->d_inode;
1622 struct filter_obd *filter = &obd->u.filter;
1625 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1626 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1627 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1628 (unsigned long)inode->i_nlink,
1629 atomic_read(&inode->i_count));
1632 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1634 CERROR("error unlinking objid %.*s: rc %d\n",
1635 dchild->d_name.len, dchild->d_name.name, rc);
1639 struct filter_intent_args {
1640 struct ldlm_lock **victim;
1645 static enum interval_iter filter_intent_cb(struct interval_node *n,
1648 struct ldlm_interval *node = (struct ldlm_interval *)n;
1649 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1650 __u64 size = arg->size;
1651 struct ldlm_lock **v = arg->victim;
1652 struct ldlm_lock *lck;
1654 /* If the interval is lower than the current file size,
1656 if (interval_high(n) <= size)
1657 return INTERVAL_ITER_STOP;
1659 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1660 /* Don't send glimpse ASTs to liblustre clients.
1661 * They aren't listening for them, and they do
1662 * entirely synchronous I/O anyways. */
1663 if (lck->l_export == NULL ||
1664 lck->l_export->exp_libclient == 1)
1667 if (*arg->liblustre)
1668 *arg->liblustre = 0;
1671 *v = LDLM_LOCK_GET(lck);
1672 } else if ((*v)->l_policy_data.l_extent.start <
1673 lck->l_policy_data.l_extent.start) {
1674 LDLM_LOCK_RELEASE(*v);
1675 *v = LDLM_LOCK_GET(lck);
1678 /* the same policy group - every lock has the
1679 * same extent, so needn't do it any more */
1683 return INTERVAL_ITER_CONT;
1686 static int filter_intent_policy(struct ldlm_namespace *ns,
1687 struct ldlm_lock **lockp, void *req_cookie,
1688 ldlm_mode_t mode, int flags, void *data)
1690 CFS_LIST_HEAD(rpc_list);
1691 struct ptlrpc_request *req = req_cookie;
1692 struct ldlm_lock *lock = *lockp, *l = NULL;
1693 struct ldlm_resource *res = lock->l_resource;
1694 ldlm_processing_policy policy;
1695 struct ost_lvb *res_lvb, *reply_lvb;
1696 struct ldlm_reply *rep;
1698 int idx, rc, tmpflags = 0, only_liblustre = 1;
1699 struct ldlm_interval_tree *tree;
1700 struct filter_intent_args arg;
1701 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1702 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1703 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1706 policy = ldlm_get_processing_policy(res);
1707 LASSERT(policy != NULL);
1708 LASSERT(req != NULL);
1710 rc = lustre_pack_reply(req, 3, repsize, NULL);
1712 RETURN(req->rq_status = rc);
1714 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1715 LASSERT(rep != NULL);
1717 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1718 sizeof(*reply_lvb));
1719 LASSERT(reply_lvb != NULL);
1721 //fixup_handle_for_resent_req(req, lock, &lockh);
1723 /* Call the extent policy function to see if our request can be
1724 * granted, or is blocked.
1725 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1726 * lock, and should not be granted if the lock will be blocked.
1729 LASSERT(ns == res->lr_namespace);
1731 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1732 check_res_locked(res);
1734 /* FIXME: we should change the policy function slightly, to not make
1735 * this list at all, since we just turn around and free it */
1736 while (!list_empty(&rpc_list)) {
1737 struct ldlm_lock *wlock =
1738 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1739 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1740 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1741 lock->l_flags &= ~LDLM_FL_CP_REQD;
1742 list_del_init(&wlock->l_cp_ast);
1743 LDLM_LOCK_RELEASE(wlock);
1746 /* The lock met with no resistance; we're finished. */
1747 if (rc == LDLM_ITER_CONTINUE) {
1748 /* do not grant locks to the liblustre clients: they cannot
1749 * handle ASTs robustly. We need to do this while still
1750 * holding ns_lock to avoid the lock remaining on the res_link
1751 * list (and potentially being added to l_pending_list by an
1752 * AST) when we are going to drop this lock ASAP. */
1753 if (lock->l_export->exp_libclient ||
1754 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1755 ldlm_resource_unlink_lock(lock);
1756 err = ELDLM_LOCK_ABORTED;
1758 err = ELDLM_LOCK_REPLACED;
1764 /* Do not grant any lock, but instead send GL callbacks. The extent
1765 * policy nicely created a list of all PW locks for us. We will choose
1766 * the highest of those which are larger than the size in the LVB, if
1767 * any, and perform a glimpse callback. */
1768 res_lvb = res->lr_lvb_data;
1769 LASSERT(res_lvb != NULL);
1770 *reply_lvb = *res_lvb;
1773 * ->ns_lock guarantees that no new locks are granted, and,
1774 * therefore, that res->lr_lvb_data cannot increase beyond the
1775 * end of already granted lock. As a result, it is safe to
1776 * check against "stale" reply_lvb->lvb_size value without
1779 arg.size = reply_lvb->lvb_size;
1781 arg.liblustre = &only_liblustre;
1782 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1783 tree = &res->lr_itree[idx];
1784 if (tree->lit_mode == LCK_PR)
1787 interval_iterate_reverse(tree->lit_root,
1788 filter_intent_cb, &arg);
1792 /* There were no PW locks beyond the size in the LVB; finished. */
1794 if (only_liblustre) {
1795 /* If we discovered a liblustre client with a PW lock,
1796 * however, the LVB may be out of date! The LVB is
1797 * updated only on glimpse (which we don't do for
1798 * liblustre clients) and cancel (which the client
1799 * obviously has not yet done). So if it has written
1800 * data but kept the lock, the LVB is stale and needs
1801 * to be updated from disk.
1803 * Of course, this will all disappear when we switch to
1804 * taking liblustre locks on the OST. */
1805 ldlm_res_lvbo_update(res, NULL, 0, 1);
1807 RETURN(ELDLM_LOCK_ABORTED);
1811 * This check is for lock taken in filter_prepare_destroy() that does
1812 * not have l_glimpse_ast set. So the logic is: if there is a lock
1813 * with no l_glimpse_ast set, this object is being destroyed already.
1815 * Hence, if you are grabbing DLM locks on the server, always set
1816 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1818 if (l->l_glimpse_ast == NULL) {
1819 /* We are racing with unlink(); just return -ENOENT */
1820 rep->lock_policy_res1 = -ENOENT;
1824 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1825 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1826 /* Update the LVB from disk if the AST failed (this is a legal race) */
1828 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1829 * sending ast is not handled. This can result in lost client writes.
1832 ldlm_res_lvbo_update(res, NULL, 0, 1);
1835 *reply_lvb = *res_lvb;
1839 LDLM_LOCK_RELEASE(l);
1841 RETURN(ELDLM_LOCK_ABORTED);
1845 * per-obd_device iobuf pool.
1847 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1848 * allocations in write-path has to be minimized (see bug 5137).
1850 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1851 * OST threads (see ost_thread_{init,done}()).
1853 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1854 * at the OST layer there are only (potentially) multiple obd_device of type
1855 * unknown at the time of OST thread creation.
1857 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1858 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1859 * it's very own iobuf.
1863 * filter_kiobuf_pool_init()
1865 * filter_kiobuf_pool_done()
1867 * filter_iobuf_get()
1869 * operate on this array. They are "generic" in a sense that they don't depend
1870 * on actual type of iobuf's (the latter depending on Linux kernel version).
1874 * destroy pool created by filter_iobuf_pool_init
1876 static void filter_iobuf_pool_done(struct filter_obd *filter)
1878 struct filter_iobuf **pool;
1883 pool = filter->fo_iobuf_pool;
1885 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1886 if (pool[i] != NULL)
1887 filter_free_iobuf(pool[i]);
1889 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1890 filter->fo_iobuf_pool = NULL;
1895 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1897 struct filter_obd *filter = &obd->u.filter;
1898 struct sptlrpc_rule_set tmp_rset;
1901 sptlrpc_rule_set_init(&tmp_rset);
1902 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1904 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1909 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1911 write_lock(&filter->fo_sptlrpc_lock);
1912 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1913 filter->fo_sptlrpc_rset = tmp_rset;
1914 write_unlock(&filter->fo_sptlrpc_lock);
1920 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1922 static int filter_iobuf_pool_init(struct filter_obd *filter)
1929 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1931 if (filter->fo_iobuf_pool == NULL)
1934 filter->fo_iobuf_count = OSS_THREADS_MAX;
1939 /* Return iobuf allocated for @thread_id. We don't know in advance how
1940 * many threads there will be so we allocate a large empty array and only
1941 * fill in those slots that are actually in use.
1942 * If we haven't allocated a pool entry for this thread before, do so now. */
1943 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1945 int thread_id = (oti && oti->oti_thread) ?
1946 oti->oti_thread->t_id : -1;
1947 struct filter_iobuf *pool = NULL;
1948 struct filter_iobuf **pool_place = NULL;
1950 if (thread_id >= 0) {
1951 LASSERT(thread_id < filter->fo_iobuf_count);
1952 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1955 if (unlikely(pool == NULL)) {
1956 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1957 PTLRPC_MAX_BRW_PAGES);
1958 if (pool_place != NULL)
1965 /* mount the file system (secretly). lustre_cfg parameters are:
1968 * 3 = flags: failover=f, failout=n
1971 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1974 struct filter_obd *filter = &obd->u.filter;
1975 struct vfsmount *mnt;
1976 struct lustre_mount_info *lmi;
1977 struct obd_uuid uuid;
1985 if (lcfg->lcfg_bufcount < 3 ||
1986 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1987 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1990 lmi = server_get_mount(obd->obd_name);
1992 /* We already mounted in lustre_fill_super.
1993 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1994 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1996 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1998 /* old path - used by lctl */
1999 CERROR("Using old MDS mount method\n");
2000 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2001 MS_NOATIME|MS_NODIRATIME,
2002 lustre_cfg_string(lcfg, 1), option);
2005 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2006 lustre_cfg_string(lcfg, 1), rc);
2010 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2012 if (IS_ERR(obd->obd_fsops))
2013 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2015 rc = filter_iobuf_pool_init(filter);
2019 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2020 CERROR("%s: Underlying device is marked as read-only. "
2021 "Setup failed\n", obd->obd_name);
2022 GOTO(err_ops, rc = -EROFS);
2025 /* failover is the default */
2026 obd->obd_replayable = 1;
2028 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2029 str = lustre_cfg_string(lcfg, 3);
2030 if (strchr(str, 'n')) {
2031 CWARN("%s: recovery disabled\n", obd->obd_name);
2032 obd->obd_replayable = 0;
2036 filter->fo_vfsmnt = mnt;
2037 obd->u.obt.obt_sb = mnt->mnt_sb;
2038 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2039 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2041 fsfilt_setup(obd, obd->u.obt.obt_sb);
2043 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2044 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2045 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2046 obd->obd_lvfs_ctxt.fs = get_ds();
2047 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2049 init_mutex(&filter->fo_init_lock);
2050 filter->fo_committed_group = 0;
2052 rc = filter_prep(obd);
2056 filter->fo_destroys_in_progress = 0;
2057 for (i = 0; i < 32; i++)
2058 sema_init(&filter->fo_create_locks[i], 1);
2060 spin_lock_init(&filter->fo_translock);
2061 spin_lock_init(&filter->fo_objidlock);
2062 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2063 sema_init(&filter->fo_alloc_lock, 1);
2064 init_brw_stats(&filter->fo_filter_stats);
2065 filter->fo_read_cache = 1; /* enable read-only cache by default */
2066 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2067 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2068 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2069 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2071 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2072 spin_lock_init(&filter->fo_llog_list_lock);
2074 filter->fo_fl_oss_capa = 1;
2075 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2076 filter->fo_capa_hash = init_capa_hash();
2077 if (filter->fo_capa_hash == NULL)
2078 GOTO(err_ops, rc = -ENOMEM);
2080 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2081 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2082 LDLM_NAMESPACE_GREEDY);
2083 if (obd->obd_namespace == NULL)
2084 GOTO(err_post, rc = -ENOMEM);
2085 obd->obd_namespace->ns_lvbp = obd;
2086 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2087 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2089 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2090 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2092 rc = obd_llog_init(obd, &obd->obd_olg, obd, 1, NULL, NULL);
2094 CERROR("failed to setup llogging subsystems\n");
2098 rwlock_init(&filter->fo_sptlrpc_lock);
2099 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2100 /* do this after llog being initialized */
2101 filter_adapt_sptlrpc_conf(obd, 1);
2103 rc = lquota_setup(filter_quota_interface_ref, obd);
2107 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2108 if (q->max_sectors < q->max_hw_sectors &&
2109 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2110 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2111 "for larger I/O requests: max_sectors = %u "
2112 "could be up to max_hw_sectors=%u\n",
2113 obd->obd_name, mnt->mnt_sb->s_id,
2114 q->max_sectors, q->max_hw_sectors);
2116 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2117 if (uuid_ptr != NULL) {
2118 class_uuid_unparse(uuid_ptr, &uuid);
2124 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2126 if (obd->obd_recovering) {
2127 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2128 "recovery for at least %d:%.02d, or until %d "
2129 "client%s reconnect. During this time new clients"
2130 " will not be allowed to connect. "
2131 "Recovery progress can be monitored by watching "
2132 "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
2133 obd->obd_name, lustre_cfg_string(lcfg, 1),
2134 label ?: "", label ? "/" : "", str,
2135 obd->obd_recovery_timeout / 60,
2136 obd->obd_recovery_timeout % 60,
2137 obd->obd_max_recoverable_clients,
2138 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2141 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2142 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2143 label ?: "", label ? "/" : "", str,
2144 obd->obd_replayable ? "enabled" : "disabled");
2152 fsfilt_put_ops(obd->obd_fsops);
2153 filter_iobuf_pool_done(filter);
2155 server_put_mount(obd->obd_name, mnt);
2156 obd->u.obt.obt_sb = 0;
2160 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2162 struct lprocfs_static_vars lvars;
2167 CLASSERT(offsetof(struct obd_device, u.obt) ==
2168 offsetof(struct obd_device, u.filter.fo_obt));
2170 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2173 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2174 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2177 addr = (unsigned long)cfs_page_address(page);
2178 clear_page((void *)addr);
2180 /* lprocfs must be setup before the filter so state can be safely added
2181 * to /proc incrementally as the filter is setup */
2182 lprocfs_filter_init_vars(&lvars);
2183 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2184 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2185 /* Init obdfilter private stats here */
2186 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2187 LPROCFS_CNTR_AVGMINMAX,
2188 "read_bytes", "bytes");
2189 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2190 LPROCFS_CNTR_AVGMINMAX,
2191 "write_bytes", "bytes");
2192 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2193 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2194 "get_page", "usec");
2195 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2196 LPROCFS_CNTR_AVGMINMAX,
2197 "get_page_failures", "num");
2198 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2199 LPROCFS_CNTR_AVGMINMAX,
2200 "cache_access", "pages");
2201 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2202 LPROCFS_CNTR_AVGMINMAX,
2203 "cache_hit", "pages");
2204 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2205 LPROCFS_CNTR_AVGMINMAX,
2206 "cache_miss", "pages");
2208 lproc_filter_attach_seqstat(obd);
2209 obd->obd_proc_exports_entry = lprocfs_register("exports",
2210 obd->obd_proc_entry,
2212 if (IS_ERR(obd->obd_proc_exports_entry)) {
2213 rc = PTR_ERR(obd->obd_proc_exports_entry);
2214 CERROR("error %d setting up lprocfs for %s\n",
2216 obd->obd_proc_exports_entry = NULL;
2219 if (obd->obd_proc_exports_entry)
2220 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2221 lprocfs_nid_stats_clear_read,
2222 lprocfs_nid_stats_clear_write, obd, NULL);
2224 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2225 LUSTRE_CFG_BUFLEN(lcfg, 4));
2226 rc = filter_common_setup(obd, lcfg, (void *)addr);
2227 OBD_PAGE_FREE(page);
2230 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2231 lprocfs_free_per_client_stats(obd);
2232 lprocfs_free_obd_stats(obd);
2233 lprocfs_obd_cleanup(obd);
2239 static struct llog_operations filter_mds_ost_repl_logops;
2241 static struct llog_operations filter_size_orig_logops = {
2242 .lop_setup = llog_obd_origin_setup,
2243 .lop_cleanup = llog_obd_origin_cleanup,
2244 .lop_add = llog_obd_origin_add
2247 static int filter_olg_fini(struct obd_llog_group *olg)
2249 struct llog_ctxt *ctxt;
2250 int rc = 0, rc2 = 0;
2253 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2255 rc = llog_cleanup(ctxt);
2257 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2259 rc2 = llog_cleanup(ctxt);
2264 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2266 rc2 = llog_cleanup(ctxt);
2275 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2276 struct obd_device *tgt)
2281 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2282 &filter_mds_ost_repl_logops);
2286 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2287 &filter_size_orig_logops);
2293 filter_olg_fini(olg);
2298 * Init the default olg, which is embeded in the obd_device, for filter.
2301 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2302 struct obd_device *tgt)
2304 struct filter_obd *filter = &obd->u.filter;
2305 struct llog_ctxt *ctxt;
2309 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2310 if (!filter->fo_lcm)
2313 filter_mds_ost_repl_logops = llog_client_ops;
2314 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2315 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2316 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2318 rc = filter_olg_init(obd, olg, tgt);
2320 GOTO(cleanup_lcm, rc);
2322 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2325 GOTO(cleanup_olg, rc);
2327 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2329 CERROR("Can't get ctxt for %p:%x\n", olg,
2330 LLOG_MDS_OST_REPL_CTXT);
2331 GOTO(cleanup_olg, rc = -ENODEV);
2333 ctxt->loc_lcm = filter->fo_lcm;
2334 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2335 llog_ctxt_put(ctxt);
2339 filter_olg_fini(olg);
2341 llog_recov_thread_fini(filter->fo_lcm, 1);
2342 filter->fo_lcm = NULL;
2347 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2348 struct obd_device *tgt, int count, struct llog_catid *catid,
2349 struct obd_uuid *uuid)
2351 struct filter_obd *filter = &obd->u.filter;
2352 struct llog_ctxt *ctxt;
2356 LASSERT(olg != NULL);
2357 if (olg == &obd->obd_olg)
2358 return filter_default_olg_init(obd, olg, tgt);
2360 LASSERT(filter->fo_lcm != NULL);
2361 rc = filter_olg_init(obd, olg, tgt);
2364 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2366 CERROR("Can't get ctxt for %p:%x\n", olg,
2367 LLOG_MDS_OST_REPL_CTXT);
2368 filter_olg_fini(olg);
2371 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2372 ctxt->loc_lcm = filter->fo_lcm;
2373 llog_ctxt_put(ctxt);
2377 static int filter_llog_finish(struct obd_device *obd, int count)
2379 struct filter_obd *filter = &obd->u.filter;
2380 struct llog_ctxt *ctxt;
2383 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2386 * Make sure that no cached llcds left in recov_thread.
2387 * We actually do sync in disconnect time, but disconnect
2388 * may not come being marked rq_no_resend = 1.
2390 llog_sync(ctxt, NULL);
2393 * Balance class_import_get() in llog_receptor_accept().
2394 * This is safe to do, as llog is already synchronized
2395 * and its import may go.
2397 mutex_down(&ctxt->loc_sem);
2398 if (ctxt->loc_imp) {
2399 class_import_put(ctxt->loc_imp);
2400 ctxt->loc_imp = NULL;
2402 mutex_up(&ctxt->loc_sem);
2403 llog_ctxt_put(ctxt);
2406 if (filter->fo_lcm) {
2407 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2408 filter->fo_lcm = NULL;
2410 RETURN(filter_olg_fini(&obd->obd_olg));
2414 * Find the group llog according to group index in the llog group list.
2416 static struct obd_llog_group *
2417 filter_find_olg_internal(struct filter_obd *filter, int group)
2419 struct obd_llog_group *olg;
2421 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2422 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2423 if (olg->olg_group == group)
2430 * Find the group llog according to group index on the filter
2432 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2434 struct obd_llog_group *olg = NULL;
2435 struct filter_obd *filter;
2437 filter = &obd->u.filter;
2439 if (group == FILTER_GROUP_LLOG)
2440 RETURN(&obd->obd_olg);
2442 spin_lock(&filter->fo_llog_list_lock);
2443 olg = filter_find_olg_internal(filter, group);
2444 spin_unlock(&filter->fo_llog_list_lock);
2449 * Find the llog_group of the filter according to the group. If it can not
2450 * find, create the llog_group, which only happens when mds is being synced
2453 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2455 struct obd_llog_group *olg = NULL;
2456 struct filter_obd *filter;
2459 filter = &obd->u.filter;
2461 if (group == FILTER_GROUP_LLOG)
2462 RETURN(&obd->obd_olg);
2464 spin_lock(&filter->fo_llog_list_lock);
2465 olg = filter_find_olg_internal(filter, group);
2467 if (olg->olg_initializing) {
2468 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2470 GOTO(out_unlock, olg);
2475 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2477 llog_group_init(olg, group);
2478 list_add(&olg->olg_list, &filter->fo_llog_list);
2479 olg->olg_initializing = 1;
2480 spin_unlock(&filter->fo_llog_list_lock);
2482 rc = llog_cat_initialize(obd, olg, 1, NULL);
2484 spin_lock(&filter->fo_llog_list_lock);
2485 list_del(&olg->olg_list);
2486 spin_unlock(&filter->fo_llog_list_lock);
2488 GOTO(out, olg = ERR_PTR(-ENOMEM));
2490 spin_lock(&filter->fo_llog_list_lock);
2491 olg->olg_initializing = 0;
2492 spin_unlock(&filter->fo_llog_list_lock);
2493 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2494 obd->obd_name, group, olg);
2499 spin_unlock(&filter->fo_llog_list_lock);
2503 static int filter_llog_connect(struct obd_export *exp,
2504 struct llogd_conn_body *body)
2506 struct obd_device *obd = exp->exp_obd;
2507 struct llog_ctxt *ctxt;
2508 struct obd_llog_group *olg;
2512 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2513 obd->obd_name, body->lgdc_logid.lgl_oid,
2514 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2516 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2518 CERROR(" %s: can not find olg of group %d\n",
2519 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2522 llog_group_set_export(olg, exp);
2524 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2525 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2526 body->lgdc_ctxt_idx);
2528 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2529 obd->obd_name, body->lgdc_logid.lgl_oid,
2530 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2532 rc = llog_connect(ctxt, &body->lgdc_logid,
2533 &body->lgdc_gen, NULL);
2534 llog_ctxt_put(ctxt);
2536 CERROR("failed to connect rc %d idx %d\n", rc,
2537 body->lgdc_ctxt_idx);
2542 static int filter_llog_preclean(struct obd_device *obd)
2544 struct obd_llog_group *olg, *tmp;
2545 struct filter_obd *filter;
2546 struct list_head remove_list;
2550 rc = obd_llog_finish(obd, 0);
2552 CERROR("failed to cleanup llogging subsystem\n");
2554 filter = &obd->u.filter;
2555 CFS_INIT_LIST_HEAD(&remove_list);
2557 spin_lock(&filter->fo_llog_list_lock);
2558 while (!list_empty(&filter->fo_llog_list)) {
2559 olg = list_entry(filter->fo_llog_list.next,
2560 struct obd_llog_group, olg_list);
2561 list_del(&olg->olg_list);
2562 list_add(&olg->olg_list, &remove_list);
2564 spin_unlock(&filter->fo_llog_list_lock);
2566 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2567 list_del_init(&olg->olg_list);
2568 rc = filter_olg_fini(olg);
2570 CERROR("failed to cleanup llogging subsystem for %u\n",
2578 static int filter_precleanup(struct obd_device *obd,
2579 enum obd_cleanup_stage stage)
2585 case OBD_CLEANUP_EARLY:
2587 case OBD_CLEANUP_EXPORTS:
2588 target_cleanup_recovery(obd);
2589 rc = filter_llog_preclean(obd);
2595 static int filter_cleanup(struct obd_device *obd)
2597 struct filter_obd *filter = &obd->u.filter;
2601 LCONSOLE_WARN("%s: shutting down for failover; client state "
2602 "will be preserved.\n", obd->obd_name);
2604 if (!list_empty(&obd->obd_exports)) {
2605 CERROR("%s: still has clients!\n", obd->obd_name);
2606 class_disconnect_exports(obd);
2607 if (!list_empty(&obd->obd_exports)) {
2608 CERROR("still has exports after forced cleanup?\n");
2613 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2614 lprocfs_free_per_client_stats(obd);
2615 lprocfs_free_obd_stats(obd);
2616 lprocfs_obd_cleanup(obd);
2617 lquota_cleanup(filter_quota_interface_ref, obd);
2619 /* Stop recovery before namespace cleanup. */
2620 target_stop_recovery_thread(obd);
2621 target_cleanup_recovery(obd);
2623 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2624 obd->obd_namespace = NULL;
2626 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2628 if (obd->u.obt.obt_sb == NULL)
2633 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2634 shrink_dcache_sb(obd->u.obt.obt_sb);
2636 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2637 obd->u.obt.obt_sb = NULL;
2639 fsfilt_put_ops(obd->obd_fsops);
2641 filter_iobuf_pool_done(filter);
2643 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2648 static int filter_connect_internal(struct obd_export *exp,
2649 struct obd_connect_data *data)
2651 struct filter_export_data *fed = &exp->exp_filter_data;
2656 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2657 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2658 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2659 data->ocd_connect_flags, data->ocd_version,
2660 data->ocd_grant, data->ocd_index);
2662 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2663 CWARN("!!! This export (nid %s) used object group %d "
2664 "earlier; now it's trying to use group %d! This could "
2665 "be a bug in the MDS. Please report to "
2666 "http://bugzilla.lustre.org/\n",
2667 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2670 fed->fed_group = data->ocd_group;
2672 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2673 exp->exp_connect_flags = data->ocd_connect_flags;
2674 data->ocd_version = LUSTRE_VERSION_CODE;
2676 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2677 obd_size left, want;
2679 spin_lock(&exp->exp_obd->obd_osfs_lock);
2680 left = filter_grant_space_left(exp);
2681 want = data->ocd_grant;
2682 filter_grant(exp, fed->fed_grant, want, left);
2683 data->ocd_grant = fed->fed_grant;
2684 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2686 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2687 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2688 exp->exp_client_uuid.uuid, exp,
2689 data->ocd_grant, want, left);
2692 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2693 struct filter_obd *filter = &exp->exp_obd->u.filter;
2694 struct lr_server_data *lsd = filter->fo_fsd;
2695 int index = le32_to_cpu(lsd->lsd_ost_index);
2697 if (!(lsd->lsd_feature_compat &
2698 cpu_to_le32(OBD_COMPAT_OST))) {
2699 /* this will only happen on the first connect */
2700 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2701 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2702 filter_update_server_data(exp->exp_obd,
2703 filter->fo_rcvd_filp, lsd, 1);
2704 } else if (index != data->ocd_index) {
2705 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2706 " %u doesn't match actual OST index"
2707 " %u in last_rcvd file, bad "
2709 obd_export_nid2str(exp), index,
2715 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2716 data->ocd_brw_size = 65536;
2717 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2718 data->ocd_brw_size = min(data->ocd_brw_size,
2719 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2720 LASSERT(data->ocd_brw_size);
2723 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2724 __u32 cksum_types = data->ocd_cksum_types;
2726 /* The client set in ocd_cksum_types the checksum types it
2727 * supports. We have to mask off the algorithms that we don't
2729 if (cksum_types & OBD_CKSUM_ALL)
2730 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2732 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2734 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2735 "%x\n", exp->exp_obd->obd_name,
2736 obd_export_nid2str(exp), cksum_types,
2737 data->ocd_cksum_types);
2739 /* This client does not support OBD_CONNECT_CKSUM
2740 * fall back to CRC32 */
2741 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2742 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2743 exp->exp_obd->obd_name,
2744 obd_export_nid2str(exp));
2747 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2748 * FIXME: We don't strictly need the COMPAT flag for that,
2749 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2750 * FIXME: We needed it for the index, as index 0 is valid. */
2755 static int filter_reconnect(const struct lu_env *env,
2756 struct obd_export *exp, struct obd_device *obd,
2757 struct obd_uuid *cluuid,
2758 struct obd_connect_data *data,
2764 if (exp == NULL || obd == NULL || cluuid == NULL)
2767 rc = filter_connect_internal(exp, data);
2769 filter_export_stats_init(obd, exp, localdata);
2774 /* nearly identical to mds_connect */
2775 static int filter_connect(const struct lu_env *env,
2776 struct lustre_handle *conn, struct obd_device *obd,
2777 struct obd_uuid *cluuid,
2778 struct obd_connect_data *data, void *localdata)
2780 struct lvfs_run_ctxt saved;
2781 struct obd_export *exp;
2782 struct filter_export_data *fed;
2783 struct lsd_client_data *lcd = NULL;
2788 if (conn == NULL || obd == NULL || cluuid == NULL)
2791 rc = class_connect(conn, obd, cluuid);
2794 exp = class_conn2export(conn);
2795 LASSERT(exp != NULL);
2797 fed = &exp->exp_filter_data;
2799 rc = filter_connect_internal(exp, data);
2803 filter_export_stats_init(obd, exp, localdata);
2804 if (obd->obd_replayable) {
2805 OBD_ALLOC(lcd, sizeof(*lcd));
2807 CERROR("filter: out of memory for client data\n");
2808 GOTO(cleanup, rc = -ENOMEM);
2811 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2813 rc = filter_client_add(obd, exp, -1);
2818 group = data->ocd_group;
2820 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2821 obd->obd_name, exp->exp_handle.h_cookie, group);
2823 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2824 rc = filter_read_groups(obd, group, 1);
2825 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2827 CERROR("can't read group %u\n", group);
2837 fed->fed_lcd = NULL;
2839 class_disconnect(exp);
2841 class_export_put(exp);
2847 /* Do extra sanity checks for grant accounting. We do this at connect,
2848 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2849 * always get rid of it or turn it off when we know accounting is good. */
2850 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2852 struct filter_export_data *fed;
2853 struct obd_export *exp;
2854 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2855 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2856 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2858 if (list_empty(&obd->obd_exports))
2861 /* We don't want to do this for large machines that do lots of
2862 mounts or unmounts. It burns... */
2863 if (obd->obd_num_exports > 100)
2866 spin_lock(&obd->obd_osfs_lock);
2867 spin_lock(&obd->obd_dev_lock);
2868 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2870 fed = &exp->exp_filter_data;
2871 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2874 if (maxsize > 0) { /* we may not have done a statfs yet */
2875 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2876 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2877 exp->exp_client_uuid.uuid, exp,
2878 fed->fed_grant, fed->fed_pending, maxsize);
2879 LASSERTF(fed->fed_dirty <= maxsize,
2880 "%s: cli %s/%p %ld > "LPU64"\n", func,
2881 exp->exp_client_uuid.uuid, exp,
2882 fed->fed_dirty, maxsize);
2885 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2886 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2887 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2889 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2890 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2891 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2892 tot_granted += fed->fed_grant + fed->fed_pending;
2893 tot_pending += fed->fed_pending;
2894 tot_dirty += fed->fed_dirty;
2896 fo_tot_granted = obd->u.filter.fo_tot_granted;
2897 fo_tot_pending = obd->u.filter.fo_tot_pending;
2898 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2899 spin_unlock(&obd->obd_dev_lock);
2900 spin_unlock(&obd->obd_osfs_lock);
2902 /* Do these assertions outside the spinlocks so we don't kill system */
2903 if (tot_granted != fo_tot_granted)
2904 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2905 func, tot_granted, fo_tot_granted);
2906 if (tot_pending != fo_tot_pending)
2907 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2908 func, tot_pending, fo_tot_pending);
2909 if (tot_dirty != fo_tot_dirty)
2910 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2911 func, tot_dirty, fo_tot_dirty);
2912 if (tot_pending > tot_granted)
2913 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2914 func, tot_pending, tot_granted);
2915 if (tot_granted > maxsize)
2916 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2917 func, tot_granted, maxsize);
2918 if (tot_dirty > maxsize)
2919 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2920 func, tot_dirty, maxsize);
2923 /* Remove this client from the grant accounting totals. We also remove
2924 * the export from the obd device under the osfs and dev locks to ensure
2925 * that the filter_grant_sanity_check() calculations are always valid.
2926 * The client should do something similar when it invalidates its import. */
2927 static void filter_grant_discard(struct obd_export *exp)
2929 struct obd_device *obd = exp->exp_obd;
2930 struct filter_obd *filter = &obd->u.filter;
2931 struct filter_export_data *fed = &exp->exp_filter_data;
2933 spin_lock(&obd->obd_osfs_lock);
2934 spin_lock(&obd->obd_dev_lock);
2935 list_del_init(&exp->exp_obd_chain);
2936 spin_unlock(&obd->obd_dev_lock);
2938 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2939 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2940 obd->obd_name, filter->fo_tot_granted,
2941 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2942 filter->fo_tot_granted -= fed->fed_grant;
2943 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2944 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2945 obd->obd_name, filter->fo_tot_pending,
2946 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2947 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2948 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2949 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2950 obd->obd_name, filter->fo_tot_dirty,
2951 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2952 filter->fo_tot_dirty -= fed->fed_dirty;
2956 spin_unlock(&obd->obd_osfs_lock);
2959 static int filter_destroy_export(struct obd_export *exp)
2963 if (exp->exp_filter_data.fed_pending)
2964 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2965 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2966 exp, exp->exp_filter_data.fed_pending);
2968 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2970 target_destroy_export(exp);
2971 ldlm_destroy_export(exp);
2973 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2977 if (exp->exp_obd->obd_replayable)
2978 filter_client_free(exp);
2980 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2982 filter_grant_discard(exp);
2983 filter_fmd_cleanup(exp);
2985 if (!(exp->exp_flags & OBD_OPT_FORCE))
2986 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2991 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2993 struct obd_llog_group *olg_min, *olg;
2994 struct filter_obd *filter;
2995 int worked = -1, group;
2996 struct llog_ctxt *ctxt;
2999 filter = &obd->u.filter;
3001 /* we can't sync log holding spinlock. also, we do not want to get
3002 * into livelock. so we do following: loop over MDS's exports in
3003 * group order and skip already synced llogs -bzzz */
3005 /* look for group with min. number, but > worked */
3008 spin_lock(&filter->fo_llog_list_lock);
3009 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3010 if (olg->olg_group <= worked) {
3011 /* this group is already synced */
3014 if (group < olg->olg_group) {
3015 /* we have group with smaller number to sync */
3018 /* store current minimal group */
3020 group = olg->olg_group;
3022 spin_unlock(&filter->fo_llog_list_lock);
3024 if (olg_min == NULL)
3027 worked = olg_min->olg_group;
3028 if (olg_min->olg_exp &&
3029 (dexp == olg_min->olg_exp || dexp == NULL)) {
3031 ctxt = llog_group_get_ctxt(olg_min,
3032 LLOG_MDS_OST_REPL_CTXT);
3034 err = llog_sync(ctxt, olg_min->olg_exp);
3035 llog_ctxt_put(ctxt);
3037 CERROR("error flushing logs to MDS: "
3042 } while (olg_min != NULL);
3045 /* Also incredibly similar to mds_disconnect */
3046 static int filter_disconnect(struct obd_export *exp)
3048 struct obd_device *obd = exp->exp_obd;
3053 class_export_get(exp);
3055 if (!(exp->exp_flags & OBD_OPT_FORCE))
3056 filter_grant_sanity_check(obd, __FUNCTION__);
3057 filter_grant_discard(exp);
3059 /* Flush any remaining cancel messages out to the target */
3060 filter_sync_llogs(obd, exp);
3062 /* Disconnect early so that clients can't keep using export */
3063 rc = class_disconnect(exp);
3064 if (exp->exp_obd->obd_namespace != NULL)
3065 ldlm_cancel_locks_for_export(exp);
3067 fsfilt_sync(obd, obd->u.obt.obt_sb);
3069 lprocfs_exp_cleanup(exp);
3071 class_export_put(exp);
3075 /* reverse import is changed, sync all cancels */
3076 static void filter_revimp_update(struct obd_export *exp)
3081 class_export_get(exp);
3083 /* flush any remaining cancel messages out to the target */
3084 filter_sync_llogs(exp->exp_obd, exp);
3085 class_export_put(exp);
3089 static int filter_ping(struct obd_export *exp)
3091 filter_fmd_expire(exp);
3096 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3097 const char *what, int quiet)
3099 struct dentry *dchild = NULL;
3102 if (oa->o_valid & OBD_MD_FLGROUP)
3105 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3107 if (IS_ERR(dchild)) {
3108 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3109 what, group, oa->o_id);
3113 if (dchild->d_inode == NULL) {
3115 CERROR("%s: %s on non-existent object: "LPU64"\n",
3116 obd->obd_name, what, oa->o_id);
3118 RETURN(ERR_PTR(-ENOENT));
3124 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3126 struct dentry *dentry = NULL;
3127 struct obd_device *obd;
3131 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3132 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3136 obd = class_exp2obd(exp);
3138 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3142 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3144 RETURN(PTR_ERR(dentry));
3146 /* Limit the valid bits in the return data to what we actually use */
3147 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3148 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3154 /* this should be enabled/disabled in condition to enabled/disabled large
3155 * inodes (fast EAs) in backing store FS. */
3156 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3157 void *handle, struct obdo *oa)
3159 struct obd_device *obd = exp->exp_obd;
3163 if (oa->o_valid & OBD_MD_FLFID) {
3164 struct filter_fid ff;
3166 if (!(oa->o_valid & OBD_MD_FLGROUP))
3168 /* packing fid and converting it to LE for storing into EA.
3169 * Here ->o_stripe_idx should be filled by LOV and rest of
3170 * fields - by client. */
3171 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3172 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3173 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3174 ff.ff_objid = cpu_to_le64(oa->o_id);
3175 ff.ff_group = cpu_to_le64(oa->o_gr);
3177 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3178 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3179 oa->o_generation, oa->o_id, oa->o_gr);
3181 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3183 CERROR("store fid in object failed! rc: %d\n", rc);
3185 CDEBUG(D_HA, "OSS object without fid info!\n");
3191 /* this is called from filter_truncate() until we have filter_punch() */
3192 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3193 struct obdo *oa, struct obd_trans_info *oti)
3195 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3196 struct llog_cookie *fcc = NULL;
3197 struct filter_obd *filter;
3198 int rc, err, locked = 0, sync = 0;
3199 loff_t old_size = 0;
3200 unsigned int ia_valid;
3201 struct inode *inode;
3206 LASSERT(dentry != NULL);
3207 LASSERT(!IS_ERR(dentry));
3209 inode = dentry->d_inode;
3210 LASSERT(inode != NULL);
3212 filter = &exp->exp_obd->u.filter;
3213 iattr_from_obdo(&iattr, oa, oa->o_valid);
3214 ia_valid = iattr.ia_valid;
3216 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3217 OBD_ALLOC(fcc, sizeof(*fcc));
3219 *fcc = oa->o_lcookie;
3222 if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
3224 LOCK_INODE_MUTEX(inode);
3225 old_size = i_size_read(inode);
3229 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3230 * then we will accept the UID+GID sent by the client during write for
3231 * initializing the ownership of this inode. We only allow this to
3232 * happen once so clear these bits in setattr. In 2.6 kernels it is
3233 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3234 * the flags that are actually being set. */
3235 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3236 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3237 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3239 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3240 if (!(ia_valid & ATTR_MODE)) {
3241 iattr.ia_mode = inode->i_mode;
3242 iattr.ia_valid |= ATTR_MODE;
3244 iattr.ia_mode &= ~S_ISUID;
3246 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3247 if (!(iattr.ia_valid & ATTR_MODE)) {
3248 iattr.ia_mode = inode->i_mode;
3249 iattr.ia_valid |= ATTR_MODE;
3251 iattr.ia_mode &= ~S_ISGID;
3254 orig_ids[USRQUOTA] = inode->i_uid;
3255 orig_ids[GRPQUOTA] = inode->i_gid;
3256 handle = fsfilt_start_log(exp->exp_obd, inode,
3257 FSFILT_OP_SETATTR, oti, 1);
3259 GOTO(out_unlock, rc = PTR_ERR(handle));
3261 /* update inode EA only once when inode is suid bit marked. As
3262 * on 2.6.x UID and GID may be set separately, we check here
3263 * only one of them to avoid double setting. */
3264 if (inode->i_mode & S_ISUID)
3265 filter_update_fidea(exp, inode, handle, oa);
3267 handle = fsfilt_start(exp->exp_obd, inode,
3268 FSFILT_OP_SETATTR, oti);
3270 GOTO(out_unlock, rc = PTR_ERR(handle));
3272 if (oa->o_valid & OBD_MD_FLFLAGS) {
3273 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3274 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3276 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3278 /* set cancel cookie callback function */
3279 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3280 filter_cancel_cookies_cb,
3284 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3285 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3287 /* The truncate might have used up our transaction credits. Make
3288 * sure we have one left for the last_rcvd update. */
3289 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3291 rc = filter_finish_transno(exp, oti, rc, sync);
3293 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3297 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3299 CERROR("error on commit, err = %d\n", err);
3306 /* For a partial-page truncate flush the page to disk immediately
3307 * to avoid data corruption during direct disk write. b=17397 */
3308 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3309 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3310 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3317 UNLOCK_INODE_MUTEX(inode);
3324 UNLOCK_INODE_MUTEX(inode);
3327 OBD_FREE(fcc, sizeof(*fcc));
3329 /* trigger quota release */
3330 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3331 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3332 int rc2 = lquota_adjust(filter_quota_interface_ref,
3333 exp->exp_obd, cur_ids,
3334 orig_ids, rc, FSFILT_OP_SETATTR);
3335 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3336 "filter adjust qunit. (rc:%d)\n", rc2);
3341 /* this is called from filter_truncate() until we have filter_punch() */
3342 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3343 struct obd_trans_info *oti)
3345 struct obdo *oa = oinfo->oi_oa;
3346 struct lustre_capa *capa = oinfo_capa(oinfo);
3347 struct ldlm_res_id res_id;
3348 struct filter_mod_data *fmd;
3349 struct lvfs_run_ctxt saved;
3350 struct filter_obd *filter;
3351 struct ldlm_resource *res;
3352 struct dentry *dentry;
3353 __u64 opc = CAPA_OPC_META_WRITE;
3357 if (oa->o_valid & OBD_FL_TRUNC)
3358 opc |= CAPA_OPC_OSS_TRUNC;
3359 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa, opc);
3363 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3364 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
3369 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3370 /* This would be very bad - accidentally truncating a file when
3371 * changing the time or similar - bug 12203. */
3372 if (oa->o_valid & OBD_MD_FLSIZE &&
3373 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3374 static char mdsinum[48];
3376 if (oa->o_valid & OBD_MD_FLFID)
3377 snprintf(mdsinum, sizeof(mdsinum) - 1,
3378 " of inode "LPU64"/%u", oa->o_fid,
3383 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3385 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3390 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3392 RETURN(PTR_ERR(dentry));
3394 filter = &exp->exp_obd->u.filter;
3395 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3399 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3400 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3401 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3402 fmd->fmd_mactime_xid = oti->oti_xid;
3403 filter_fmd_put(exp, fmd);
3406 /* setting objects attributes (including owner/group) */
3407 rc = filter_setattr_internal(exp, dentry, oa, oti);
3409 GOTO(out_unlock, rc);
3411 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3412 &res_id, LDLM_EXTENT, 0);
3415 LDLM_RESOURCE_ADDREF(res);
3416 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3417 LDLM_RESOURCE_DELREF(res);
3418 ldlm_resource_putref(res);
3421 oa->o_valid = OBD_MD_FLID;
3423 /* Quota release need uid/gid info */
3424 obdo_from_inode(oa, dentry->d_inode,
3425 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3431 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3435 /* XXX identical to osc_unpackmd */
3436 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3437 struct lov_mds_md *lmm, int lmm_bytes)
3443 if (lmm_bytes < sizeof (*lmm)) {
3444 CERROR("lov_mds_md too small: %d, need %d\n",
3445 lmm_bytes, (int)sizeof(*lmm));
3448 /* XXX LOV_MAGIC etc check? */
3450 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3451 CERROR("lov_mds_md: zero lmm_object_id\n");
3456 lsm_size = lov_stripe_md_size(1);
3460 if (*lsmp != NULL && lmm == NULL) {
3461 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3462 OBD_FREE(*lsmp, lsm_size);
3467 if (*lsmp == NULL) {
3468 OBD_ALLOC(*lsmp, lsm_size);
3472 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3473 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3474 OBD_FREE(*lsmp, lsm_size);
3477 loi_init((*lsmp)->lsm_oinfo[0]);
3481 /* XXX zero *lsmp? */
3482 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3483 LASSERT((*lsmp)->lsm_object_id);
3486 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3491 /* caller must hold fo_create_locks[oa->o_gr] */
3492 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3493 struct filter_obd *filter)
3495 struct obdo doa; /* XXX obdo on stack */
3501 LASSERT_MDS_GROUP(oa->o_gr);
3502 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3503 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3505 memset(&doa, 0, sizeof(doa));
3507 doa.o_valid |= OBD_MD_FLGROUP;
3508 doa.o_gr = oa->o_gr;
3509 doa.o_mode = S_IFREG;
3511 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3512 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3513 exp->exp_obd->obd_name, doa.o_gr);
3517 last = filter_last_id(filter, doa.o_gr);
3519 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3520 exp->exp_obd->obd_name, oa->o_id + 1, last);
3522 for (id = last; id > oa->o_id; id--) {
3524 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3525 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3526 CEMERG("error destroying precreate objid "LPU64": %d\n",
3528 filter_set_last_id(filter, id - 1, doa.o_gr);
3529 /* update last_id on disk periodically so that if we restart
3530 * we don't need to re-scan all of the just-deleted objects. */
3531 if ((id & 511) == 0)
3532 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3535 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3536 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3538 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3539 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3544 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3545 obd_gr group, int *num);
3546 /* returns a negative error or a nonnegative number of files to create */
3547 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3548 obd_gr group, struct obd_trans_info *oti)
3550 struct obd_device *obd = exp->exp_obd;
3551 struct filter_obd *filter = &obd->u.filter;
3555 /* delete orphans request */
3556 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3557 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3558 CERROR("%s: dropping old orphan cleanup request\n",
3562 /* This causes inflight precreates to abort and drop lock */
3563 set_bit(group, &filter->fo_destroys_in_progress);
3564 down(&filter->fo_create_locks[group]);
3565 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3566 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3567 exp->exp_obd->obd_name, group);
3568 up(&filter->fo_create_locks[group]);
3571 diff = oa->o_id - filter_last_id(filter, group);
3572 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3573 filter_last_id(filter, group), diff);
3575 if (-diff > OST_MAX_PRECREATE) {
3576 CERROR("%s: ignoring bogus orphan destroy request: "
3577 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3578 oa->o_id, filter_last_id(filter, group));
3579 /* FIXME: should reset precreate_next_id on MDS */
3580 GOTO(out, rc = -EINVAL);
3583 rc = filter_destroy_precreated(exp, oa, filter);
3585 CERROR("%s: unable to write lastobjid, but "
3586 "orphans were deleted\n", obd->obd_name);
3589 /* XXX: Used by MDS for the first time! */
3590 clear_bit(group, &filter->fo_destroys_in_progress);
3593 down(&filter->fo_create_locks[group]);
3594 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3595 CERROR("%s: dropping old precreate request\n",
3599 /* only precreate if group == 0 and o_id is specfied */
3600 if (group == FILTER_GROUP_LLOG || oa->o_id == 0)
3603 diff = oa->o_id - filter_last_id(filter, group);
3604 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3605 filter_last_id(filter, group), diff);
3607 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3608 oa->o_id, filter_last_id(filter, group), diff);
3612 oa->o_id = filter_last_id(&obd->u.filter, group);
3613 rc = filter_precreate(obd, oa, group, &diff);
3614 oa->o_id = filter_last_id(&obd->u.filter, group);
3616 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3619 /* else diff == 0 */
3622 up(&filter->fo_create_locks[group]);
3626 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3627 __u64 max_age, __u32 flags)
3629 struct filter_obd *filter = &obd->u.filter;
3630 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3634 /* at least try to account for cached pages. its still racey and
3635 * might be under-reporting if clients haven't announced their
3636 * caches with brw recently */
3637 spin_lock(&obd->obd_osfs_lock);
3638 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3639 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3640 spin_unlock(&obd->obd_osfs_lock);
3642 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3643 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3644 filter->fo_tot_dirty, filter->fo_tot_granted,
3645 filter->fo_tot_pending,
3646 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3648 filter_grant_sanity_check(obd, __FUNCTION__);
3650 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3651 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3652 osfs->os_bsize - 1) >> blockbits));
3654 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3655 * stop creating files on MDS if OST is not good shape to create
3657 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3662 static int filter_use_existing_obj(struct obd_device *obd,
3663 struct dentry *dchild, void **handle,
3666 struct inode *inode = dchild->d_inode;
3670 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3673 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3674 if (IS_ERR(*handle))
3675 return PTR_ERR(*handle);
3677 iattr.ia_valid = ATTR_MODE;
3678 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3679 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3687 /* We rely on the fact that only one thread will be creating files in a given
3688 * group at a time, which is why we don't need an atomic filter_get_new_id.
3689 * Even if we had that atomic function, the following race would exist:
3691 * thread 1: gets id x from filter_next_id
3692 * thread 2: gets id (x + 1) from filter_next_id
3693 * thread 2: creates object (x + 1)
3694 * thread 1: tries to create object x, gets -ENOSPC
3696 * Caller must hold fo_create_locks[group]
3698 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3699 obd_gr group, int *num)
3701 struct dentry *dchild = NULL, *dparent = NULL;
3702 struct filter_obd *filter;
3703 struct obd_statfs *osfs;
3704 int err = 0, rc = 0, recreate_obj = 0, i;
3705 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3707 void *handle = NULL;
3710 filter = &obd->u.filter;
3712 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3714 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3716 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3717 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3720 OBD_ALLOC(osfs, sizeof(*osfs));
3723 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3724 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3725 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3726 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3727 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3731 OBD_FREE(osfs, sizeof(*osfs));
3736 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3737 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3739 for (i = 0; i < *num && err == 0; i++) {
3740 int cleanup_phase = 0;
3742 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3743 CWARN("%s: create aborted by destroy\n",
3752 last_id = filter_last_id(filter, group);
3753 if (next_id > last_id) {
3754 CERROR("Error: Trying to recreate obj greater"
3755 "than last id "LPD64" > "LPD64"\n",
3757 GOTO(cleanup, rc = -EINVAL);
3760 next_id = filter_last_id(filter, group) + 1;
3762 dparent = filter_parent_lock(obd, group, next_id);
3763 if (IS_ERR(dparent))
3764 GOTO(cleanup, rc = PTR_ERR(dparent));
3765 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3767 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3769 GOTO(cleanup, rc = PTR_ERR(dchild));
3770 cleanup_phase = 2; /* f_dput(dchild) */
3772 if (dchild->d_inode != NULL) {
3773 /* This would only happen if lastobjid was bad on disk*/
3774 /* Could also happen if recreating missing obj but it
3775 * already exists. */
3777 CERROR("%s: recreating existing object %.*s?\n",
3778 obd->obd_name, dchild->d_name.len,
3779 dchild->d_name.name);
3781 /* Use these existing objects if they are
3783 if (dchild->d_inode->i_size == 0) {
3784 rc = filter_use_existing_obj(obd,dchild,
3785 &handle, &cleanup_phase);
3792 CERROR("%s: Serious error: objid %.*s already "
3793 "exists; is this filesystem corrupt?\n",
3794 obd->obd_name, dchild->d_name.len,
3795 dchild->d_name.name);
3798 GOTO(cleanup, rc = -EEXIST);
3801 handle = fsfilt_start_log(obd, dparent->d_inode,
3802 FSFILT_OP_CREATE, NULL, 1);
3804 GOTO(cleanup, rc = PTR_ERR(handle));
3807 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3808 ",od->o_id="LPU64")\n", obd->obd_name, group,
3811 /* We mark object SUID+SGID to flag it for accepting UID+GID
3812 * from client on first write. Currently the permission bits
3813 * on the OST are never used, so this is OK. */
3814 rc = ll_vfs_create(dparent->d_inode, dchild,
3815 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3817 CERROR("create failed rc = %d\n", rc);
3822 if (!recreate_obj) {
3823 filter_set_last_id(filter, next_id, group);
3824 err = filter_update_last_objid(obd, group, 0);
3826 CERROR("unable to write lastobjid "
3827 "but file created\n");
3831 switch(cleanup_phase) {
3833 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3835 CERROR("error on commit, err = %d\n", err);
3842 filter_parent_unlock(dparent);
3849 if (time_after(jiffies, enough_time)) {
3851 "%s: precreate slow - want %d got %d \n",
3852 obd->obd_name, *num, i);
3859 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3860 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3865 static int filter_create(struct obd_export *exp, struct obdo *oa,
3866 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3868 struct obd_device *obd = exp->exp_obd;
3869 struct filter_export_data *fed;
3870 struct filter_obd *filter;
3871 struct lvfs_run_ctxt saved;
3872 struct lov_stripe_md *lsm = NULL;
3873 int rc = 0, diff, group = oa->o_gr;
3876 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3877 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3879 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3880 CERROR("!!! nid %s sent invalid object group %d\n",
3881 obd_export_nid2str(exp), group);
3885 fed = &exp->exp_filter_data;
3886 filter = &obd->u.filter;
3888 if (fed->fed_group != group) {
3889 CERROR("!!! this export (nid %s) used object group %d "
3890 "earlier; now it's trying to use group %d! This could "
3891 "be a bug in the MDS. Please report to "
3892 "http://bugzilla.lustre.org/\n",
3893 obd_export_nid2str(exp), fed->fed_group, group);
3900 rc = obd_alloc_memmd(exp, &lsm);
3907 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3909 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3910 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3911 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3912 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3913 oa->o_id, filter_last_id(filter,
3918 down(&filter->fo_create_locks[oa->o_gr]);
3919 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3920 up(&filter->fo_create_locks[oa->o_gr]);
3923 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3926 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3927 if (rc && ea != NULL && *ea != lsm) {
3928 obd_free_memmd(exp, &lsm);
3929 } else if (rc == 0 && ea != NULL) {
3930 /* XXX LOV STACKING: the lsm that is passed to us from
3931 * LOV does not have valid lsm_oinfo data structs, so
3932 * don't go touching that. This needs to be fixed in a
3934 lsm->lsm_object_id = oa->o_id;
3941 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3942 struct lov_stripe_md *md, struct obd_trans_info *oti,
3943 struct obd_export *md_exp, void *capa)
3945 unsigned int qcids[MAXQUOTAS] = {0, 0};
3946 struct obd_device *obd;
3947 struct filter_obd *filter;
3948 struct dentry *dchild = NULL, *dparent = NULL;
3949 struct lvfs_run_ctxt saved;
3950 void *handle = NULL;
3951 struct llog_cookie *fcc = NULL;
3952 int rc, rc2, cleanup_phase = 0, sync = 0;
3956 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3958 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
3959 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
3964 filter = &obd->u.filter;
3966 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3969 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
3970 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3972 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3974 GOTO(cleanup, rc = PTR_ERR(dchild));
3977 if (dchild->d_inode == NULL) {
3978 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3980 /* If object already gone, cancel cookie right now */
3981 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3982 struct llog_ctxt *ctxt;
3983 struct obd_llog_group *olg;
3984 fcc = &oa->o_lcookie;
3985 olg = filter_find_olg(obd, oa->o_gr);
3987 CERROR(" %s: can not find olg of group %d\n",
3988 obd->obd_name, (int)oa->o_gr);
3989 GOTO(cleanup, rc = PTR_ERR(olg));
3991 llog_group_set_export(olg, exp);
3993 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
3994 llog_cancel(ctxt, NULL, 1, fcc, 0);
3995 llog_ctxt_put(ctxt);
3996 fcc = NULL; /* we didn't allocate fcc, don't free it */
3998 GOTO(cleanup, rc = -ENOENT);
4001 filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
4003 /* Our MDC connection is established by the MDS to us */
4004 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4005 OBD_ALLOC(fcc, sizeof(*fcc));
4007 *fcc = oa->o_lcookie;
4009 DQUOT_INIT(dchild->d_inode);
4011 /* we're gonna truncate it first in order to avoid possible deadlock:
4013 * open trasaction open transaction
4014 * down(i_zombie) down(i_zombie)
4015 * restart transaction
4016 * (see BUG 4180) -bzzz
4018 LOCK_INODE_MUTEX(dchild->d_inode);
4019 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4021 if (IS_ERR(handle)) {
4022 UNLOCK_INODE_MUTEX(dchild->d_inode);
4023 GOTO(cleanup, rc = PTR_ERR(handle));
4026 iattr.ia_valid = ATTR_SIZE;
4028 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4029 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4030 UNLOCK_INODE_MUTEX(dchild->d_inode);
4034 GOTO(cleanup, rc = rc2);
4036 /* We don't actually need to lock the parent until we are unlinking
4037 * here, and not while truncating above. That avoids holding the
4038 * parent lock for a long time during truncate, which can block other
4039 * threads from doing anything to objects in that directory. bug 7171 */
4040 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4041 if (IS_ERR(dparent))
4042 GOTO(cleanup, rc = PTR_ERR(dparent));
4043 cleanup_phase = 3; /* filter_parent_unlock */
4045 LOCK_INODE_MUTEX(dchild->d_inode);
4046 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4047 if (IS_ERR(handle)) {
4048 UNLOCK_INODE_MUTEX(dchild->d_inode);
4049 GOTO(cleanup, rc = PTR_ERR(handle));
4051 cleanup_phase = 4; /* fsfilt_commit */
4053 /* Quota release need uid/gid of inode */
4054 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4056 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4058 /* this drops dchild->d_inode->i_mutex unconditionally */
4059 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4063 switch(cleanup_phase) {
4066 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4067 oti->oti_handle : handle,
4068 filter_cancel_cookies_cb,
4070 /* If add_journal_cb failed, then filter_finish_transno
4071 * will commit the handle and we will do a sync
4072 * on commit. then we call callback directly to free
4075 rc = filter_finish_transno(exp, oti, rc, sync);
4077 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4080 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4082 CERROR("error on commit, err = %d\n", rc2);
4089 filter_parent_unlock(dparent);
4093 OBD_FREE(fcc, sizeof(*fcc));
4095 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4098 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4102 /* trigger quota release */
4103 qcids[USRQUOTA] = oa->o_uid;
4104 qcids[GRPQUOTA] = oa->o_gid;
4105 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4108 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4112 /* NB start and end are used for punch, but not truncate */
4113 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4114 struct obd_trans_info *oti,
4115 struct ptlrpc_request_set *rqset)
4120 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4121 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4122 oinfo->oi_policy.l_extent.end);
4126 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4127 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4128 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4130 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4131 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4132 rc = filter_setattr(exp, oinfo, oti);
4133 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4137 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4138 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4141 struct lvfs_run_ctxt saved;
4142 struct filter_obd *filter;
4143 struct dentry *dentry;
4147 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
4148 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4152 filter = &exp->exp_obd->u.filter;
4154 /* An objid of zero is taken to mean "sync whole filesystem" */
4155 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4156 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4157 /* Flush any remaining cancel messages out to the target */
4158 filter_sync_llogs(exp->exp_obd, exp);
4162 dentry = filter_oa2dentry(exp->exp_obd, oa);
4164 RETURN(PTR_ERR(dentry));
4166 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4168 LOCK_INODE_MUTEX(dentry->d_inode);
4170 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4172 /* just any file to grab fsync method - "file" arg unused */
4173 struct file *file = filter->fo_rcvd_filp;
4175 if (file->f_op && file->f_op->fsync)
4176 rc = file->f_op->fsync(NULL, dentry, 1);
4178 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4182 UNLOCK_INODE_MUTEX(dentry->d_inode);
4184 oa->o_valid = OBD_MD_FLID;
4185 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4187 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4193 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4194 void *key, __u32 *vallen, void *val,
4195 struct lov_stripe_md *lsm)
4197 struct obd_device *obd;
4200 obd = class_exp2obd(exp);
4202 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4206 if (KEY_IS(KEY_BLOCKSIZE)) {
4207 __u32 *blocksize = val;
4209 if (*vallen < sizeof(*blocksize))
4211 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4213 *vallen = sizeof(*blocksize);
4217 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4218 __u32 *blocksize_bits = val;
4219 if (blocksize_bits) {
4220 if (*vallen < sizeof(*blocksize_bits))
4222 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4224 *vallen = sizeof(*blocksize_bits);
4228 if (KEY_IS(KEY_LAST_ID)) {
4229 obd_id *last_id = val;
4230 /* FIXME: object groups */
4232 if (*vallen < sizeof(*last_id))
4234 *last_id = filter_last_id(&obd->u.filter,
4235 exp->exp_filter_data.fed_group);
4237 *vallen = sizeof(*last_id);
4241 if (KEY_IS(KEY_FIEMAP)) {
4242 struct ll_fiemap_info_key *fm_key = key;
4243 struct dentry *dentry;
4244 struct ll_user_fiemap *fiemap = val;
4245 struct lvfs_run_ctxt saved;
4248 if (fiemap == NULL) {
4249 *vallen = fiemap_count_to_size(
4250 fm_key->fiemap.fm_extent_count);
4254 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4257 RETURN(PTR_ERR(dentry));
4259 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4260 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4261 rc = fsfilt_iocontrol(obd, dentry->d_inode, NULL,
4262 EXT3_IOC_FIEMAP, (long)fiemap);
4267 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4273 CDEBUG(D_IOCTL, "invalid key\n");
4277 static inline int filter_setup_llog_group(struct obd_export *exp,
4278 struct obd_device *obd,
4281 struct obd_llog_group *olg;
4282 struct llog_ctxt *ctxt;
4285 olg = filter_find_create_olg(obd, group);
4287 RETURN(PTR_ERR(olg));
4289 llog_group_set_export(olg, exp);
4291 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4292 LASSERTF(ctxt != NULL, "ctxt is null\n");
4294 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4295 llog_ctxt_put(ctxt);
4298 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4299 void *key, __u32 vallen, void *val,
4300 struct ptlrpc_request_set *set)
4302 struct obd_device *obd;
4308 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4312 if (KEY_IS(KEY_CAPA_KEY)) {
4313 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4315 CERROR("filter update capability key failed: %d\n", rc);
4319 if (KEY_IS(KEY_REVIMP_UPD)) {
4320 filter_revimp_update(exp);
4321 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4325 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4326 filter_adapt_sptlrpc_conf(obd, 0);
4330 if (!KEY_IS(KEY_MDS_CONN))
4333 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4334 obd_export_nid2str(exp));
4335 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4337 /* setup llog imports */
4339 group = (int)(*(__u32 *)val);
4341 group = 0; /* default value */
4343 LASSERT_MDS_GROUP(group);
4344 rc = filter_setup_llog_group(exp, obd, group);
4348 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4350 if (group == FILTER_GROUP_MDS0) {
4351 /* setup llog group 1 for interop */
4352 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4358 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4359 int len, void *karg, void *uarg)
4361 struct obd_device *obd = exp->exp_obd;
4362 struct obd_ioctl_data *data = karg;
4366 case OBD_IOC_ABORT_RECOVERY: {
4367 CERROR("aborting recovery for device %s\n", obd->obd_name);
4368 target_stop_recovery_thread(obd);
4372 case OBD_IOC_SYNC: {
4373 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4374 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4378 case OBD_IOC_SET_READONLY: {
4380 struct super_block *sb = obd->u.obt.obt_sb;
4381 struct inode *inode = sb->s_root->d_inode;
4382 BDEVNAME_DECLARE_STORAGE(tmp);
4383 CERROR("*** setting device %s read-only ***\n",
4384 ll_bdevname(sb, tmp));
4386 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4387 if (!IS_ERR(handle))
4388 rc = fsfilt_commit(obd, inode, handle, 1);
4390 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4391 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4393 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4397 case OBD_IOC_CATLOGLIST: {
4398 rc = llog_catalog_list(obd, 1, data);
4402 case OBD_IOC_LLOG_CANCEL:
4403 case OBD_IOC_LLOG_REMOVE:
4404 case OBD_IOC_LLOG_INFO:
4405 case OBD_IOC_LLOG_PRINT: {
4406 /* FIXME to be finished */
4407 RETURN(-EOPNOTSUPP);
4409 struct llog_ctxt *ctxt = NULL;
4411 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4412 rc = llog_ioctl(ctxt, cmd, data);
4413 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4426 static int filter_health_check(struct obd_device *obd)
4428 #ifdef USE_HEALTH_CHECK_WRITE
4429 struct filter_obd *filter = &obd->u.filter;
4434 * health_check to return 0 on healthy
4435 * and 1 on unhealthy.
4437 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4440 #ifdef USE_HEALTH_CHECK_WRITE
4441 LASSERT(filter->fo_health_check_filp != NULL);
4442 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4447 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4450 return filter_fid2dentry(data, NULL, gr, id);
4453 static int filter_process_config(struct obd_device *obd, obd_count len,
4456 struct lustre_cfg *lcfg = buf;
4457 struct lprocfs_static_vars lvars;
4460 switch (lcfg->lcfg_command) {
4462 lprocfs_filter_init_vars(&lvars);
4464 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4474 static struct lvfs_callback_ops filter_lvfs_ops = {
4475 l_fid2dentry: filter_lvfs_fid2dentry,
4478 static struct obd_ops filter_obd_ops = {
4479 .o_owner = THIS_MODULE,
4480 .o_get_info = filter_get_info,
4481 .o_set_info_async = filter_set_info_async,
4482 .o_setup = filter_setup,
4483 .o_precleanup = filter_precleanup,
4484 .o_cleanup = filter_cleanup,
4485 .o_connect = filter_connect,
4486 .o_reconnect = filter_reconnect,
4487 .o_disconnect = filter_disconnect,
4488 .o_ping = filter_ping,
4489 .o_init_export = filter_init_export,
4490 .o_destroy_export = filter_destroy_export,
4491 .o_statfs = filter_statfs,
4492 .o_getattr = filter_getattr,
4493 .o_unpackmd = filter_unpackmd,
4494 .o_create = filter_create,
4495 .o_setattr = filter_setattr,
4496 .o_destroy = filter_destroy,
4497 .o_brw = filter_brw,
4498 .o_punch = filter_truncate,
4499 .o_sync = filter_sync,
4500 .o_preprw = filter_preprw,
4501 .o_commitrw = filter_commitrw,
4502 .o_llog_init = filter_llog_init,
4503 .o_llog_connect = filter_llog_connect,
4504 .o_llog_finish = filter_llog_finish,
4505 .o_iocontrol = filter_iocontrol,
4506 .o_health_check = filter_health_check,
4507 .o_process_config = filter_process_config,
4510 quota_interface_t *filter_quota_interface_ref;
4511 extern quota_interface_t filter_quota_interface;
4513 static int __init obdfilter_init(void)
4515 struct lprocfs_static_vars lvars;
4518 lprocfs_filter_init_vars(&lvars);
4520 request_module("lquota");
4521 OBD_ALLOC(obdfilter_created_scratchpad,
4522 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4523 sizeof(*obdfilter_created_scratchpad));
4524 if (obdfilter_created_scratchpad == NULL)
4527 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4528 sizeof(struct filter_mod_data),
4531 GOTO(out, rc = -ENOMEM);
4533 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4534 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4536 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4537 LUSTRE_OST_NAME, NULL);
4541 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4542 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4543 ll_fmd_cachep = NULL;
4545 if (filter_quota_interface_ref)
4546 PORTAL_SYMBOL_PUT(filter_quota_interface);
4548 OBD_FREE(obdfilter_created_scratchpad,
4549 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4550 sizeof(*obdfilter_created_scratchpad));
4556 static void __exit obdfilter_exit(void)
4558 if (filter_quota_interface_ref)
4559 PORTAL_SYMBOL_PUT(filter_quota_interface);
4561 if (ll_fmd_cachep) {
4562 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4563 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4564 ll_fmd_cachep = NULL;
4567 class_unregister_type(LUSTRE_OST_NAME);
4568 OBD_FREE(obdfilter_created_scratchpad,
4569 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4570 sizeof(*obdfilter_created_scratchpad));
4573 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4574 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4575 MODULE_LICENSE("GPL");
4577 module_init(obdfilter_init);
4578 module_exit(obdfilter_exit);