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,
208 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
209 /* Self-export gets no proc entry */
212 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
214 /* Mask error for already created
222 struct nid_stat *tmp = exp->exp_nid_stats;
223 LASSERT(tmp != NULL);
225 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
226 if (tmp->nid_brw_stats == NULL)
229 init_brw_stats(tmp->nid_brw_stats);
230 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
231 0644, &filter_per_nid_stats_fops,
234 CWARN("Error adding the brw_stats file\n");
236 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
240 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
244 /* Always add in ldlm_stats */
245 tmp->nid_ldlm_stats = lprocfs_alloc_stats(LDLM_LAST_OPC -
247 if (tmp->nid_ldlm_stats == NULL)
250 lprocfs_init_ldlm_stats(tmp->nid_ldlm_stats);
251 rc = lprocfs_register_stats(tmp->nid_proc, "ldlm_stats",
252 tmp->nid_ldlm_stats);
260 /* Add client data to the FILTER. We use a bitmap to locate a free space
261 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
262 * Otherwise, we have just read the data from the last_rcvd file and
263 * we know its offset. */
264 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
267 struct filter_obd *filter = &obd->u.filter;
268 struct filter_export_data *fed = &exp->exp_filter_data;
269 unsigned long *bitmap = filter->fo_last_rcvd_slots;
270 int new_client = (cl_idx == -1);
274 LASSERT(bitmap != NULL);
275 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
278 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
281 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
282 * there's no need for extra complication here
285 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
287 if (cl_idx >= LR_MAX_CLIENTS) {
288 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
292 if (test_and_set_bit(cl_idx, bitmap)) {
293 cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
298 if (test_and_set_bit(cl_idx, bitmap)) {
299 CERROR("FILTER client %d: bit already set in bitmap!\n",
305 fed->fed_lr_idx = cl_idx;
306 fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
307 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
308 LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
310 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
311 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
314 struct lvfs_run_ctxt saved;
315 loff_t off = fed->fed_lr_off;
319 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
320 fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_lcd));
322 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
323 /* Transaction needed to fix bug 1403 */
324 handle = fsfilt_start(obd,
325 filter->fo_rcvd_filp->f_dentry->d_inode,
326 FSFILT_OP_SETATTR, NULL);
327 if (IS_ERR(handle)) {
328 rc = PTR_ERR(handle);
329 CERROR("unable to start transaction: rc %d\n", rc);
331 rc = fsfilt_add_journal_cb(obd, 0, handle,
332 target_client_add_cb, exp);
334 spin_lock(&exp->exp_lock);
335 exp->exp_need_sync = 1;
336 spin_unlock(&exp->exp_lock);
338 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
340 sizeof(*fed->fed_lcd),
341 &off, rc /* sync if no cb */);
343 filter->fo_rcvd_filp->f_dentry->d_inode,
346 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
349 CERROR("error writing %s client idx %u: rc %d\n",
350 LAST_RCVD, fed->fed_lr_idx, rc);
357 struct lsd_client_data zero_lcd; /* globals are implicitly zeroed */
359 static int filter_client_free(struct obd_export *exp)
361 struct filter_export_data *fed = &exp->exp_filter_data;
362 struct filter_obd *filter = &exp->exp_obd->u.filter;
363 struct obd_device *obd = exp->exp_obd;
364 struct lvfs_run_ctxt saved;
369 if (fed->fed_lcd == NULL)
372 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
373 if (strcmp(fed->fed_lcd->lcd_uuid, obd->obd_uuid.uuid ) == 0)
376 LASSERT(filter->fo_last_rcvd_slots != NULL);
378 off = fed->fed_lr_off;
380 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
381 fed->fed_lr_idx, fed->fed_lr_off, fed->fed_lcd->lcd_uuid);
383 /* Don't clear fed_lr_idx here as it is likely also unset. At worst
384 * we leak a client slot that will be cleaned on the next recovery. */
386 CERROR("%s: client idx %d has med_off %lld\n",
387 obd->obd_name, fed->fed_lr_idx, off);
388 GOTO(free, rc = -EINVAL);
391 /* Clear the bit _after_ zeroing out the client so we don't
392 race with filter_client_add and zero out new clients.*/
393 if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
394 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
399 if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
400 /* Don't force sync on disconnect if aborting recovery,
401 * or it does num_clients * num_osts. b=17194 */
402 int need_sync = (!exp->exp_libclient || exp->exp_need_sync) &&
403 !(exp->exp_flags&OBD_OPT_ABORT_RECOV);
404 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
405 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_lcd,
406 sizeof(zero_lcd), &off, 0);
408 /* Make sure the server's last_transno is up to date. Do this
409 * after the client is freed so we know all the client's
410 * transactions have been committed. */
412 filter_update_server_data(obd, filter->fo_rcvd_filp,
413 filter->fo_fsd, need_sync);
414 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
416 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
417 "zero out client %s at idx %u/%llu in %s %ssync rc %d\n",
418 fed->fed_lcd->lcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
419 LAST_RCVD, need_sync ? "" : "a", rc);
422 if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
423 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
430 OBD_FREE_PTR(fed->fed_lcd);
436 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
437 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
438 struct filter_mod_data *fmd)
440 LASSERT_SPIN_LOCKED(&fed->fed_lock);
441 if (--fmd->fmd_refcount == 0) {
442 /* XXX when we have persistent reservations and the handle
443 * is stored herein we need to drop it here. */
444 fed->fed_mod_count--;
445 list_del(&fmd->fmd_list);
446 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
450 /* drop fmd reference, free it if last ref */
451 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
453 struct filter_export_data *fed;
458 fed = &exp->exp_filter_data;
459 spin_lock(&fed->fed_lock);
460 filter_fmd_put_nolock(fed, fmd); /* caller reference */
461 spin_unlock(&fed->fed_lock);
464 /* expire entries from the end of the list if there are too many
465 * or they are too old */
466 static void filter_fmd_expire_nolock(struct filter_obd *filter,
467 struct filter_export_data *fed,
468 struct filter_mod_data *keep)
470 struct filter_mod_data *fmd, *tmp;
472 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
476 if (time_before(jiffies, fmd->fmd_expire) &&
477 fed->fed_mod_count < filter->fo_fmd_max_num)
480 list_del_init(&fmd->fmd_list);
481 filter_fmd_put_nolock(fed, fmd); /* list reference */
485 void filter_fmd_expire(struct obd_export *exp)
487 spin_lock(&exp->exp_filter_data.fed_lock);
488 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
489 &exp->exp_filter_data, NULL);
490 spin_unlock(&exp->exp_filter_data.fed_lock);
493 /* find specified objid, group in export fmd list.
494 * caller must hold fed_lock and take fmd reference itself */
495 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
496 struct filter_export_data *fed,
497 obd_id objid, obd_gr group)
499 struct filter_mod_data *found = NULL, *fmd;
501 LASSERT_SPIN_LOCKED(&fed->fed_lock);
503 list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
504 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
506 list_del(&fmd->fmd_list);
507 list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
508 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
513 filter_fmd_expire_nolock(filter, fed, found);
518 /* Find fmd based on objid and group, or return NULL if not found. */
519 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
520 obd_id objid, obd_gr group)
522 struct filter_mod_data *fmd;
524 spin_lock(&exp->exp_filter_data.fed_lock);
525 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
526 &exp->exp_filter_data, objid, group);
528 fmd->fmd_refcount++; /* caller reference */
529 spin_unlock(&exp->exp_filter_data.fed_lock);
534 /* Find fmd based on objid and group, or create a new one if none is found.
535 * It is possible for this function to return NULL under memory pressure,
536 * or if objid = 0 is passed (which will only cause old entries to expire).
537 * Currently this is not fatal because any fmd state is transient and
538 * may also be freed when it gets sufficiently old. */
539 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
540 obd_id objid, obd_gr group)
542 struct filter_export_data *fed = &exp->exp_filter_data;
543 struct filter_mod_data *found = NULL, *fmd_new = NULL;
545 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
547 spin_lock(&fed->fed_lock);
548 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
551 list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
552 fmd_new->fmd_id = objid;
553 fmd_new->fmd_gr = group;
554 fmd_new->fmd_refcount++; /* list reference */
556 fed->fed_mod_count++;
558 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
562 found->fmd_refcount++; /* caller reference */
563 found->fmd_expire = jiffies +
564 exp->exp_obd->u.filter.fo_fmd_max_age;
567 spin_unlock(&fed->fed_lock);
573 /* drop fmd list reference so it will disappear when last reference is put.
574 * This isn't so critical because it would in fact only affect the one client
575 * that is doing the unlink and at worst we have an stale entry referencing
576 * an object that should never be used again. */
577 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
579 struct filter_mod_data *found = NULL;
581 spin_lock(&exp->exp_filter_data.fed_lock);
582 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
584 list_del_init(&found->fmd_list);
585 filter_fmd_put_nolock(&exp->exp_filter_data, found);
587 spin_unlock(&exp->exp_filter_data.fed_lock);
590 #define filter_fmd_drop(exp, objid, group)
593 /* remove all entries from fmd list */
594 static void filter_fmd_cleanup(struct obd_export *exp)
596 struct filter_export_data *fed = &exp->exp_filter_data;
597 struct filter_mod_data *fmd = NULL, *tmp;
599 spin_lock(&fed->fed_lock);
600 list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
601 list_del_init(&fmd->fmd_list);
602 filter_fmd_put_nolock(fed, fmd);
604 spin_unlock(&fed->fed_lock);
607 static int filter_init_export(struct obd_export *exp)
609 spin_lock_init(&exp->exp_filter_data.fed_lock);
610 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
612 spin_lock(&exp->exp_lock);
613 exp->exp_connecting = 1;
614 spin_unlock(&exp->exp_lock);
616 return ldlm_init_export(exp);
619 static int filter_free_server_data(struct filter_obd *filter)
621 OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
622 filter->fo_fsd = NULL;
623 OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
624 filter->fo_last_rcvd_slots = NULL;
628 /* assumes caller is already in kernel ctxt */
629 int filter_update_server_data(struct obd_device *obd, struct file *filp,
630 struct lr_server_data *fsd, int force_sync)
636 CDEBUG(D_INODE, "server uuid : %s\n", fsd->lsd_uuid);
637 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
638 le64_to_cpu(fsd->lsd_last_transno));
639 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
640 le64_to_cpu(fsd->lsd_mount_count));
642 fsd->lsd_compat14 = fsd->lsd_last_transno;
643 rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
645 CERROR("error writing lr_server_data: rc = %d\n", rc);
650 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
653 struct filter_obd *filter = &obd->u.filter;
659 if (filter->fo_last_objid_files[group] == NULL) {
660 CERROR("Object group "LPU64" not fully setup; not updating "
661 "last_objid\n", group);
665 CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
666 obd->obd_name, group, filter->fo_last_objids[group]);
668 tmp = cpu_to_le64(filter->fo_last_objids[group]);
669 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
670 &tmp, sizeof(tmp), &off, force_sync);
672 CERROR("error writing group "LPU64" last objid: rc = %d\n",
676 extern int ost_handle(struct ptlrpc_request *req);
677 /* assumes caller has already in kernel ctxt */
678 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
680 struct filter_obd *filter = &obd->u.filter;
681 struct lr_server_data *fsd;
682 struct lsd_client_data *lcd = NULL;
683 struct inode *inode = filp->f_dentry->d_inode;
684 unsigned long last_rcvd_size = i_size_read(inode);
690 /* ensure padding in the struct is the correct size */
691 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
692 sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
693 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
694 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
696 OBD_ALLOC(fsd, sizeof(*fsd));
699 filter->fo_fsd = fsd;
701 OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
702 if (filter->fo_last_rcvd_slots == NULL) {
703 OBD_FREE(fsd, sizeof(*fsd));
707 if (last_rcvd_size == 0) {
708 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
710 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
711 fsd->lsd_last_transno = 0;
712 mount_count = fsd->lsd_mount_count = 0;
713 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
714 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
715 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
716 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
717 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
718 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
720 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
722 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
726 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
727 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
728 "using the wrong disk %s. Were the "
729 "/dev/ assignments rearranged?\n",
730 obd->obd_uuid.uuid, fsd->lsd_uuid);
731 GOTO(err_fsd, rc = -EINVAL);
733 mount_count = le64_to_cpu(fsd->lsd_mount_count);
734 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
736 /* Assume old last_rcvd format unless I_C_LR is set */
737 if (!(fsd->lsd_feature_incompat &
738 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
739 fsd->lsd_last_transno = fsd->lsd_compat14;
743 if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
744 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
745 obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
746 ~FILTER_INCOMPAT_SUPP);
747 GOTO(err_fsd, rc = -EINVAL);
749 if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
750 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
751 obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
752 ~FILTER_ROCOMPAT_SUPP);
753 /* Do something like remount filesystem read-only */
754 GOTO(err_fsd, rc = -EINVAL);
757 CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
758 obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
759 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
760 obd->obd_name, mount_count + 1);
761 CDEBUG(D_INODE, "%s: server data size: %u\n",
762 obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
763 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
764 obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
765 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
766 obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
767 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
768 obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
769 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
770 last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
771 (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
772 le16_to_cpu(fsd->lsd_client_size));
774 if (!obd->obd_replayable) {
775 CWARN("%s: recovery support OFF\n", obd->obd_name);
779 for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
780 off < last_rcvd_size; cl_idx++) {
782 struct obd_export *exp;
783 struct filter_export_data *fed;
788 GOTO(err_client, rc = -ENOMEM);
791 /* Don't assume off is incremented properly by
792 * fsfilt_read_record(), in case sizeof(*lcd)
793 * isn't the same as fsd->lsd_client_size. */
794 off = le32_to_cpu(fsd->lsd_client_start) +
795 cl_idx * le16_to_cpu(fsd->lsd_client_size);
796 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
798 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
799 LAST_RCVD, cl_idx, off, rc);
800 break; /* read error shouldn't cause startup to fail */
803 if (lcd->lcd_uuid[0] == '\0') {
804 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
809 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
811 /* These exports are cleaned up by filter_disconnect(), so they
812 * need to be set up like real exports as filter_connect() does.
814 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
816 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
817 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
818 last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
820 if (PTR_ERR(exp) == -EALREADY) {
821 /* export already exists, zero out this one */
822 CERROR("Zeroing out duplicate export due to "
824 lcd->lcd_uuid[0] = '\0';
826 GOTO(err_client, rc = PTR_ERR(exp));
829 fed = &exp->exp_filter_data;
831 fed->fed_group = 0; /* will be assigned at connect */
832 filter_export_stats_init(obd, exp, NULL);
833 rc = filter_client_add(obd, exp, cl_idx);
834 /* can't fail for existing client */
835 LASSERTF(rc == 0, "rc = %d\n", rc);
838 spin_lock(&exp->exp_lock);
839 exp->exp_connecting = 0;
840 exp->exp_in_recovery = 0;
841 spin_unlock(&exp->exp_lock);
842 obd->obd_max_recoverable_clients++;
843 class_export_put(exp);
846 /* Need to check last_rcvd even for duplicated exports. */
847 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
850 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
851 fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
857 obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
859 target_recovery_init(obd, ost_handle);
862 filter->fo_mount_count = mount_count + 1;
863 fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
865 /* save it, so mount count and last_transno is current */
866 rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
868 GOTO(err_client, rc);
873 target_recovery_fini(obd);
875 filter_free_server_data(filter);
879 static int filter_cleanup_groups(struct obd_device *obd)
881 struct filter_obd *filter = &obd->u.filter;
883 struct dentry *dentry;
887 if (filter->fo_dentry_O_groups != NULL) {
888 for (i = 0; i < filter->fo_group_count; i++) {
889 dentry = filter->fo_dentry_O_groups[i];
893 OBD_FREE(filter->fo_dentry_O_groups,
894 filter->fo_group_count *
895 sizeof(*filter->fo_dentry_O_groups));
896 filter->fo_dentry_O_groups = NULL;
898 if (filter->fo_last_objid_files != NULL) {
899 for (i = 0; i < filter->fo_group_count; i++) {
900 filp = filter->fo_last_objid_files[i];
904 OBD_FREE(filter->fo_last_objid_files,
905 filter->fo_group_count *
906 sizeof(*filter->fo_last_objid_files));
907 filter->fo_last_objid_files = NULL;
909 if (filter->fo_dentry_O_sub != NULL) {
910 for (i = 0; i < filter->fo_group_count; i++) {
911 for (j = 0; j < filter->fo_subdir_count; j++) {
912 dentry = filter->fo_dentry_O_sub[i].dentry[j];
917 OBD_FREE(filter->fo_dentry_O_sub,
918 filter->fo_group_count *
919 sizeof(*filter->fo_dentry_O_sub));
920 filter->fo_dentry_O_sub = NULL;
922 if (filter->fo_last_objids != NULL) {
923 OBD_FREE(filter->fo_last_objids,
924 filter->fo_group_count *
925 sizeof(*filter->fo_last_objids));
926 filter->fo_last_objids = NULL;
928 if (filter->fo_dentry_O != NULL) {
929 f_dput(filter->fo_dentry_O);
930 filter->fo_dentry_O = NULL;
935 static int filter_update_last_group(struct obd_device *obd, int group)
937 struct filter_obd *filter = &obd->u.filter;
938 struct file *filp = NULL;
939 int last_group = 0, rc;
943 if (group <= filter->fo_committed_group)
946 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
950 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
954 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
956 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
959 LASSERTF(off == 0 || CHECK_MDS_GROUP(last_group),
960 "off = %llu and last_group = %d\n", off, last_group);
962 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
963 obd->obd_name, last_group, group);
967 /* must be sync: bXXXX */
968 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
970 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
974 filter->fo_committed_group = group;
981 static int filter_read_group_internal(struct obd_device *obd, int group,
984 struct filter_obd *filter = &obd->u.filter;
985 __u64 *new_objids = NULL;
986 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
987 struct dentry **new_groups = NULL;
988 struct file **new_files = NULL;
989 struct dentry *dentry;
991 int old_count = filter->fo_group_count, rc, stage = 0, i;
997 snprintf(name, 24, "%d", group);
1001 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1003 if (IS_ERR(dentry)) {
1004 CERROR("Cannot lookup expected object group %d: %ld\n",
1005 group, PTR_ERR(dentry));
1006 RETURN(PTR_ERR(dentry));
1009 dentry = simple_mkdir(filter->fo_dentry_O, filter->fo_vfsmnt,
1011 if (IS_ERR(dentry)) {
1012 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1014 RETURN(PTR_ERR(dentry));
1019 snprintf(name, 24, "O/%d/LAST_ID", group);
1021 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1023 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1024 GOTO(cleanup, rc = PTR_ERR(filp));
1028 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1030 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1034 if (filter->fo_subdir_count) {
1035 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1036 if (tmp_subdirs == NULL)
1037 GOTO(cleanup, rc = -ENOMEM);
1040 for (i = 0; i < filter->fo_subdir_count; i++) {
1042 snprintf(dir, sizeof(dir), "d%u", i);
1044 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1047 if (IS_ERR(tmp_subdirs->dentry[i])) {
1048 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1049 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1054 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1055 tmp_subdirs->dentry[i]);
1059 /* 'group' is an index; we need an array of length 'group + 1' */
1060 if (group + 1 > old_count) {
1061 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1062 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1063 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1064 OBD_ALLOC(new_files, len * sizeof(*new_files));
1066 if (new_objids == NULL || new_subdirs == NULL ||
1067 new_groups == NULL || new_files == NULL)
1068 GOTO(cleanup, rc = -ENOMEM);
1071 memcpy(new_objids, filter->fo_last_objids,
1072 old_count * sizeof(*new_objids));
1073 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1074 old_count * sizeof(*new_subdirs));
1075 memcpy(new_groups, filter->fo_dentry_O_groups,
1076 old_count * sizeof(*new_groups));
1077 memcpy(new_files, filter->fo_last_objid_files,
1078 old_count * sizeof(*new_files));
1080 OBD_FREE(filter->fo_last_objids,
1081 old_count * sizeof(*new_objids));
1082 OBD_FREE(filter->fo_dentry_O_sub,
1083 old_count * sizeof(*new_subdirs));
1084 OBD_FREE(filter->fo_dentry_O_groups,
1085 old_count * sizeof(*new_groups));
1086 OBD_FREE(filter->fo_last_objid_files,
1087 old_count * sizeof(*new_files));
1089 filter->fo_last_objids = new_objids;
1090 filter->fo_dentry_O_sub = new_subdirs;
1091 filter->fo_dentry_O_groups = new_groups;
1092 filter->fo_last_objid_files = new_files;
1093 filter->fo_group_count = len;
1096 filter->fo_dentry_O_groups[group] = dentry;
1097 filter->fo_last_objid_files[group] = filp;
1098 if (filter->fo_subdir_count) {
1099 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1100 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1103 filter_update_last_group(obd, group);
1105 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1106 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1107 rc = filter_update_last_objid(obd, group, 1);
1111 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1112 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1113 obd->obd_name, group, last_objid);
1118 if (new_objids != NULL)
1119 OBD_FREE(new_objids, len * sizeof(*new_objids));
1120 if (new_subdirs != NULL)
1121 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1122 if (new_groups != NULL)
1123 OBD_FREE(new_groups, len * sizeof(*new_groups));
1124 if (new_files != NULL)
1125 OBD_FREE(new_files, len * sizeof(*new_files));
1127 if (filter->fo_subdir_count) {
1128 for (i = 0; i < filter->fo_subdir_count; i++) {
1129 if (tmp_subdirs->dentry[i] != NULL)
1130 dput(tmp_subdirs->dentry[i]);
1132 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1135 filp_close(filp, 0);
1142 static int filter_read_groups(struct obd_device *obd, int last_group,
1145 struct filter_obd *filter = &obd->u.filter;
1146 int old_count, group, rc = 0;
1148 down(&filter->fo_init_lock);
1149 old_count = filter->fo_group_count;
1150 for (group = old_count; group <= last_group; group++) {
1152 rc = filter_read_group_internal(obd, group, create);
1156 up(&filter->fo_init_lock);
1160 /* FIXME: object groups */
1161 static int filter_prep_groups(struct obd_device *obd)
1163 struct filter_obd *filter = &obd->u.filter;
1164 struct dentry *dentry, *O_dentry;
1166 int last_group, rc = 0, cleanup_phase = 0;
1170 O_dentry = simple_mkdir(current->fs->pwd, filter->fo_vfsmnt,
1172 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1173 if (IS_ERR(O_dentry)) {
1174 rc = PTR_ERR(O_dentry);
1175 CERROR("cannot open/create O: rc = %d\n", rc);
1178 filter->fo_dentry_O = O_dentry;
1179 cleanup_phase = 1; /* O_dentry */
1181 /* Lookup "R" to tell if we're on an old OST FS and need to convert
1182 * from O/R/<dir>/<objid> to O/0/<dir>/<objid>. This can be removed
1183 * some time post 1.0 when all old-style OSTs have converted along
1184 * with the init_objid hack. */
1185 dentry = ll_lookup_one_len("R", O_dentry, 1);
1187 GOTO(cleanup, rc = PTR_ERR(dentry));
1188 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1189 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1192 CWARN("converting OST to new object layout\n");
1193 if (IS_ERR(O0_dentry)) {
1194 rc = PTR_ERR(O0_dentry);
1195 CERROR("error looking up O/0: rc %d\n", rc);
1196 GOTO(cleanup_R, rc);
1199 if (O0_dentry->d_inode) {
1200 CERROR("Both O/R and O/0 exist. Fix manually.\n");
1201 GOTO(cleanup_O0, rc = -EEXIST);
1204 LOCK_INODE_MUTEX(O_dentry->d_inode);
1205 rc = ll_vfs_rename(O_dentry->d_inode, dentry, filter->fo_vfsmnt,
1206 O_dentry->d_inode, O0_dentry,
1208 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1211 CERROR("error renaming O/R to O/0: rc %d\n", rc);
1212 GOTO(cleanup_O0, rc);
1214 filter->fo_fsd->lsd_feature_incompat |=
1215 cpu_to_le32(OBD_INCOMPAT_GROUPS);
1216 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1218 GOTO(cleanup_O0, rc);
1230 cleanup_phase = 2; /* groups */
1232 /* we have to initialize all groups before first connections from
1233 * clients because they may send create/destroy for any group -bzzz */
1234 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1236 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1237 GOTO(cleanup, rc = PTR_ERR(filp));
1239 cleanup_phase = 3; /* filp */
1241 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1243 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1247 last_group = FILTER_MIN_GROUPS;
1249 LASSERT_MDS_GROUP(last_group);
1252 CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1253 FILTER_MIN_GROUPS, last_group);
1254 filter->fo_committed_group = last_group;
1255 rc = filter_read_groups(obd, last_group, 1);
1259 filp_close(filp, 0);
1263 switch (cleanup_phase) {
1265 filp_close(filp, 0);
1267 filter_cleanup_groups(obd);
1269 f_dput(filter->fo_dentry_O);
1270 filter->fo_dentry_O = NULL;
1278 /* setup the object store with correct subdirectories */
1279 static int filter_prep(struct obd_device *obd)
1281 struct lvfs_run_ctxt saved;
1282 struct filter_obd *filter = &obd->u.filter;
1284 struct inode *inode;
1288 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1289 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1290 if (!file || IS_ERR(file)) {
1292 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1296 filter->fo_rcvd_filp = file;
1297 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1298 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1299 file->f_dentry->d_inode->i_mode);
1300 GOTO(err_filp, rc = -ENOENT);
1303 inode = file->f_dentry->d_parent->d_inode;
1304 /* We use i_op->unlink directly in filter_vfs_unlink() */
1305 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1306 CERROR("%s: filesystem does not support create/unlink ops\n",
1308 GOTO(err_filp, rc = -EOPNOTSUPP);
1311 rc = filter_init_server_data(obd, file);
1313 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1316 /* open and create health check io file*/
1317 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1320 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1322 GOTO(err_server_data, rc);
1324 filter->fo_health_check_filp = file;
1325 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1326 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1327 file->f_dentry->d_inode->i_mode);
1328 GOTO(err_health_check, rc = -ENOENT);
1330 rc = lvfs_check_io_health(obd, file);
1332 GOTO(err_health_check, rc);
1334 rc = filter_prep_groups(obd);
1336 GOTO(err_health_check, rc);
1338 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1343 if (filp_close(filter->fo_health_check_filp, 0))
1344 CERROR("can't close %s after error\n", HEALTH_CHECK);
1345 filter->fo_health_check_filp = NULL;
1347 target_recovery_fini(obd);
1348 filter_free_server_data(filter);
1350 if (filp_close(filter->fo_rcvd_filp, 0))
1351 CERROR("can't close %s after error\n", LAST_RCVD);
1352 filter->fo_rcvd_filp = NULL;
1356 /* cleanup the filter: write last used object id to status file */
1357 static void filter_post(struct obd_device *obd)
1359 struct lvfs_run_ctxt saved;
1360 struct filter_obd *filter = &obd->u.filter;
1363 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1364 * best to start a transaction with h_sync, because we removed this
1367 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1368 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1371 CERROR("error writing server data: rc = %d\n", rc);
1373 for (i = 0; i < filter->fo_group_count; i++) {
1374 rc = filter_update_last_objid(obd, i,
1375 (i == filter->fo_group_count - 1));
1377 CERROR("error writing group %d lastobjid: rc = %d\n",
1381 rc = filp_close(filter->fo_rcvd_filp, 0);
1382 filter->fo_rcvd_filp = NULL;
1384 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1386 rc = filp_close(filter->fo_health_check_filp, 0);
1387 filter->fo_health_check_filp = NULL;
1389 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1391 filter_cleanup_groups(obd);
1392 filter_free_server_data(filter);
1393 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1395 filter_free_capa_keys(filter);
1396 cleanup_capa_hash(filter->fo_capa_hash);
1399 static void filter_set_last_id(struct filter_obd *filter,
1400 obd_id id, obd_gr group)
1402 LASSERT(filter->fo_fsd != NULL);
1403 LASSERT(group <= filter->fo_group_count);
1405 spin_lock(&filter->fo_objidlock);
1406 filter->fo_last_objids[group] = id;
1407 spin_unlock(&filter->fo_objidlock);
1410 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1413 LASSERT(filter->fo_fsd != NULL);
1414 LASSERT(group <= filter->fo_group_count);
1416 /* FIXME: object groups */
1417 spin_lock(&filter->fo_objidlock);
1418 id = filter->fo_last_objids[group];
1419 spin_unlock(&filter->fo_objidlock);
1423 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1425 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1429 /* We never dget the object parent, so DON'T dput it either */
1430 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1432 struct filter_obd *filter = &obd->u.filter;
1433 struct filter_subdirs *subdirs;
1434 LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1436 if ((group > FILTER_GROUP_MDS0 && group < FILTER_GROUP_MDS1_N_BASE) ||
1437 filter->fo_subdir_count == 0)
1438 return filter->fo_dentry_O_groups[group];
1440 subdirs = &filter->fo_dentry_O_sub[group];
1441 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1444 /* We never dget the object parent, so DON'T dput it either */
1445 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1448 unsigned long now = jiffies;
1449 struct dentry *dparent = filter_parent(obd, group, objid);
1452 if (IS_ERR(dparent))
1454 if (dparent == NULL)
1455 return ERR_PTR(-ENOENT);
1457 rc = filter_lock_dentry(obd, dparent);
1458 fsfilt_check_slow(obd, now, "parent lock");
1459 return rc ? ERR_PTR(rc) : dparent;
1462 /* We never dget the object parent, so DON'T dput it either */
1463 static void filter_parent_unlock(struct dentry *dparent)
1465 UNLOCK_INODE_MUTEX(dparent->d_inode);
1468 /* How to get files, dentries, inodes from object id's.
1470 * If dir_dentry is passed, the caller has already locked the parent
1471 * appropriately for this operation (normally a write lock). If
1472 * dir_dentry is NULL, we do a read lock while we do the lookup to
1473 * avoid races with create/destroy and such changing the directory
1474 * internal to the filesystem code. */
1475 struct dentry *filter_fid2dentry(struct obd_device *obd,
1476 struct dentry *dir_dentry,
1477 obd_gr group, obd_id id)
1479 struct dentry *dparent = dir_dentry;
1480 struct dentry *dchild;
1485 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1486 obd->u.filter.fo_destroys_in_progress == 0) {
1487 /* don't fail lookups for orphan recovery, it causes
1488 * later LBUGs when objects still exist during precreate */
1489 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1490 RETURN(ERR_PTR(-ENOENT));
1493 CERROR("fatal: invalid object id 0\n");
1494 RETURN(ERR_PTR(-ESTALE));
1497 len = sprintf(name, LPU64, id);
1498 if (dir_dentry == NULL) {
1499 dparent = filter_parent_lock(obd, group, id);
1500 if (IS_ERR(dparent)) {
1501 CERROR("%s: error getting object "LPU64":"LPU64
1502 " parent: rc %ld\n", obd->obd_name,
1503 id, group, PTR_ERR(dparent));
1507 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1508 dparent->d_name.len, dparent->d_name.name, name);
1509 dchild = /*ll_*/lookup_one_len(name, dparent, len);
1510 if (dir_dentry == NULL)
1511 filter_parent_unlock(dparent);
1512 if (IS_ERR(dchild)) {
1513 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1518 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1519 CERROR("%s: got bad object "LPU64" inode %lu\n",
1520 obd->obd_name, id, dchild->d_inode->i_ino);
1522 RETURN(ERR_PTR(-ENOENT));
1525 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1526 name, dchild, atomic_read(&dchild->d_count));
1528 LASSERT(atomic_read(&dchild->d_count) > 0);
1533 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1536 struct lustre_handle lockh;
1537 int flags = LDLM_AST_DISCARD_DATA, rc;
1538 struct ldlm_res_id res_id;
1539 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1542 osc_build_res_name(objid, group, &res_id);
1543 /* Tell the clients that the object is gone now and that they should
1544 * throw away any cached pages. */
1545 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1546 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1547 ldlm_completion_ast, NULL, NULL, 0, NULL,
1550 /* We only care about the side-effects, just drop the lock. */
1552 ldlm_lock_decref(&lockh, LCK_PW);
1557 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1558 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1559 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1560 * reverse. Caller must take i_sem before starting the transaction and we
1561 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1562 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1563 struct vfsmount *mnt)
1568 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1569 * itself we already hold dir->i_mutex for child create/unlink ops */
1570 LASSERT(dentry->d_inode != NULL);
1571 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1572 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1576 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1577 GOTO(out, rc = -ENOENT);
1579 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1584 GOTO(out, rc = -EPERM);
1586 /* check_sticky() */
1587 if ((dentry->d_inode->i_uid != current->fsuid &&
1588 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1589 IS_IMMUTABLE(dentry->d_inode))
1590 GOTO(out, rc = -EPERM);
1592 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1593 * that was done because of journal_start (which is already done
1594 * here) or some other ordering issue. */
1597 rc = ll_security_inode_unlink(dir, dentry, mnt);
1601 rc = dir->i_op->unlink(dir, dentry);
1603 /* need to drop i_mutex before we lose inode reference */
1604 UNLOCK_INODE_MUTEX(dentry->d_inode);
1611 /* Caller must hold LCK_PW on parent and push us into kernel context.
1612 * Caller must hold child i_mutex, we drop it always.
1613 * Caller is also required to ensure that dchild->d_inode exists. */
1614 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1615 obd_gr group, struct dentry *dparent,
1616 struct dentry *dchild)
1618 struct inode *inode = dchild->d_inode;
1619 struct filter_obd *filter = &obd->u.filter;
1622 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1623 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1624 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1625 (unsigned long)inode->i_nlink,
1626 atomic_read(&inode->i_count));
1629 rc = filter_vfs_unlink(dparent->d_inode, dchild, filter->fo_vfsmnt);
1631 CERROR("error unlinking objid %.*s: rc %d\n",
1632 dchild->d_name.len, dchild->d_name.name, rc);
1636 struct filter_intent_args {
1637 struct ldlm_lock **victim;
1642 static enum interval_iter filter_intent_cb(struct interval_node *n,
1645 struct ldlm_interval *node = (struct ldlm_interval *)n;
1646 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1647 __u64 size = arg->size;
1648 struct ldlm_lock **v = arg->victim;
1649 struct ldlm_lock *lck;
1651 /* If the interval is lower than the current file size,
1653 if (interval_high(n) <= size)
1654 return INTERVAL_ITER_STOP;
1656 list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1657 /* Don't send glimpse ASTs to liblustre clients.
1658 * They aren't listening for them, and they do
1659 * entirely synchronous I/O anyways. */
1660 if (lck->l_export == NULL ||
1661 lck->l_export->exp_libclient == 1)
1664 if (*arg->liblustre)
1665 *arg->liblustre = 0;
1668 *v = LDLM_LOCK_GET(lck);
1669 } else if ((*v)->l_policy_data.l_extent.start <
1670 lck->l_policy_data.l_extent.start) {
1671 LDLM_LOCK_RELEASE(*v);
1672 *v = LDLM_LOCK_GET(lck);
1675 /* the same policy group - every lock has the
1676 * same extent, so needn't do it any more */
1680 return INTERVAL_ITER_CONT;
1683 static int filter_intent_policy(struct ldlm_namespace *ns,
1684 struct ldlm_lock **lockp, void *req_cookie,
1685 ldlm_mode_t mode, int flags, void *data)
1687 CFS_LIST_HEAD(rpc_list);
1688 struct ptlrpc_request *req = req_cookie;
1689 struct ldlm_lock *lock = *lockp, *l = NULL;
1690 struct ldlm_resource *res = lock->l_resource;
1691 ldlm_processing_policy policy;
1692 struct ost_lvb *res_lvb, *reply_lvb;
1693 struct ldlm_reply *rep;
1695 int idx, rc, tmpflags = 0, only_liblustre = 1;
1696 struct ldlm_interval_tree *tree;
1697 struct filter_intent_args arg;
1698 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1699 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1700 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1703 policy = ldlm_get_processing_policy(res);
1704 LASSERT(policy != NULL);
1705 LASSERT(req != NULL);
1707 rc = lustre_pack_reply(req, 3, repsize, NULL);
1709 RETURN(req->rq_status = rc);
1711 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1712 LASSERT(rep != NULL);
1714 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1715 sizeof(*reply_lvb));
1716 LASSERT(reply_lvb != NULL);
1718 //fixup_handle_for_resent_req(req, lock, &lockh);
1720 /* Call the extent policy function to see if our request can be
1721 * granted, or is blocked.
1722 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1723 * lock, and should not be granted if the lock will be blocked.
1726 LASSERT(ns == res->lr_namespace);
1728 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1729 check_res_locked(res);
1731 /* FIXME: we should change the policy function slightly, to not make
1732 * this list at all, since we just turn around and free it */
1733 while (!list_empty(&rpc_list)) {
1734 struct ldlm_lock *wlock =
1735 list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1736 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1737 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1738 lock->l_flags &= ~LDLM_FL_CP_REQD;
1739 list_del_init(&wlock->l_cp_ast);
1740 LDLM_LOCK_RELEASE(wlock);
1743 /* The lock met with no resistance; we're finished. */
1744 if (rc == LDLM_ITER_CONTINUE) {
1745 /* do not grant locks to the liblustre clients: they cannot
1746 * handle ASTs robustly. We need to do this while still
1747 * holding ns_lock to avoid the lock remaining on the res_link
1748 * list (and potentially being added to l_pending_list by an
1749 * AST) when we are going to drop this lock ASAP. */
1750 if (lock->l_export->exp_libclient ||
1751 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1752 ldlm_resource_unlink_lock(lock);
1753 err = ELDLM_LOCK_ABORTED;
1755 err = ELDLM_LOCK_REPLACED;
1761 /* Do not grant any lock, but instead send GL callbacks. The extent
1762 * policy nicely created a list of all PW locks for us. We will choose
1763 * the highest of those which are larger than the size in the LVB, if
1764 * any, and perform a glimpse callback. */
1765 res_lvb = res->lr_lvb_data;
1766 LASSERT(res_lvb != NULL);
1767 *reply_lvb = *res_lvb;
1770 * ->ns_lock guarantees that no new locks are granted, and,
1771 * therefore, that res->lr_lvb_data cannot increase beyond the
1772 * end of already granted lock. As a result, it is safe to
1773 * check against "stale" reply_lvb->lvb_size value without
1776 arg.size = reply_lvb->lvb_size;
1778 arg.liblustre = &only_liblustre;
1779 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1780 tree = &res->lr_itree[idx];
1781 if (tree->lit_mode == LCK_PR)
1784 interval_iterate_reverse(tree->lit_root,
1785 filter_intent_cb, &arg);
1789 /* There were no PW locks beyond the size in the LVB; finished. */
1791 if (only_liblustre) {
1792 /* If we discovered a liblustre client with a PW lock,
1793 * however, the LVB may be out of date! The LVB is
1794 * updated only on glimpse (which we don't do for
1795 * liblustre clients) and cancel (which the client
1796 * obviously has not yet done). So if it has written
1797 * data but kept the lock, the LVB is stale and needs
1798 * to be updated from disk.
1800 * Of course, this will all disappear when we switch to
1801 * taking liblustre locks on the OST. */
1802 ldlm_res_lvbo_update(res, NULL, 0, 1);
1804 RETURN(ELDLM_LOCK_ABORTED);
1808 * This check is for lock taken in filter_prepare_destroy() that does
1809 * not have l_glimpse_ast set. So the logic is: if there is a lock
1810 * with no l_glimpse_ast set, this object is being destroyed already.
1812 * Hence, if you are grabbing DLM locks on the server, always set
1813 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1815 if (l->l_glimpse_ast == NULL) {
1816 /* We are racing with unlink(); just return -ENOENT */
1817 rep->lock_policy_res1 = -ENOENT;
1821 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1822 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1823 /* Update the LVB from disk if the AST failed (this is a legal race) */
1825 * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1826 * sending ast is not handled. This can result in lost client writes.
1829 ldlm_res_lvbo_update(res, NULL, 0, 1);
1832 *reply_lvb = *res_lvb;
1836 LDLM_LOCK_RELEASE(l);
1838 RETURN(ELDLM_LOCK_ABORTED);
1842 * per-obd_device iobuf pool.
1844 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1845 * allocations in write-path has to be minimized (see bug 5137).
1847 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1848 * OST threads (see ost_thread_{init,done}()).
1850 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1851 * at the OST layer there are only (potentially) multiple obd_device of type
1852 * unknown at the time of OST thread creation.
1854 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1855 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1856 * it's very own iobuf.
1860 * filter_kiobuf_pool_init()
1862 * filter_kiobuf_pool_done()
1864 * filter_iobuf_get()
1866 * operate on this array. They are "generic" in a sense that they don't depend
1867 * on actual type of iobuf's (the latter depending on Linux kernel version).
1871 * destroy pool created by filter_iobuf_pool_init
1873 static void filter_iobuf_pool_done(struct filter_obd *filter)
1875 struct filter_iobuf **pool;
1880 pool = filter->fo_iobuf_pool;
1882 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1883 if (pool[i] != NULL)
1884 filter_free_iobuf(pool[i]);
1886 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1887 filter->fo_iobuf_pool = NULL;
1892 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1894 struct filter_obd *filter = &obd->u.filter;
1895 struct sptlrpc_rule_set tmp_rset;
1898 sptlrpc_rule_set_init(&tmp_rset);
1899 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1901 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1906 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1908 write_lock(&filter->fo_sptlrpc_lock);
1909 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1910 filter->fo_sptlrpc_rset = tmp_rset;
1911 write_unlock(&filter->fo_sptlrpc_lock);
1917 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1919 static int filter_iobuf_pool_init(struct filter_obd *filter)
1926 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1928 if (filter->fo_iobuf_pool == NULL)
1931 filter->fo_iobuf_count = OSS_THREADS_MAX;
1936 /* Return iobuf allocated for @thread_id. We don't know in advance how
1937 * many threads there will be so we allocate a large empty array and only
1938 * fill in those slots that are actually in use.
1939 * If we haven't allocated a pool entry for this thread before, do so now. */
1940 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1942 int thread_id = (oti && oti->oti_thread) ?
1943 oti->oti_thread->t_id : -1;
1944 struct filter_iobuf *pool = NULL;
1945 struct filter_iobuf **pool_place = NULL;
1947 if (thread_id >= 0) {
1948 LASSERT(thread_id < filter->fo_iobuf_count);
1949 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1952 if (unlikely(pool == NULL)) {
1953 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1954 PTLRPC_MAX_BRW_PAGES);
1955 if (pool_place != NULL)
1962 /* mount the file system (secretly). lustre_cfg parameters are:
1965 * 3 = flags: failover=f, failout=n
1968 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1971 struct filter_obd *filter = &obd->u.filter;
1972 struct vfsmount *mnt;
1973 struct lustre_mount_info *lmi;
1974 struct obd_uuid uuid;
1982 if (lcfg->lcfg_bufcount < 3 ||
1983 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1984 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1987 lmi = server_get_mount(obd->obd_name);
1989 /* We already mounted in lustre_fill_super.
1990 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1991 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1993 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1995 /* old path - used by lctl */
1996 CERROR("Using old MDS mount method\n");
1997 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1998 MS_NOATIME|MS_NODIRATIME,
1999 lustre_cfg_string(lcfg, 1), option);
2002 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2003 lustre_cfg_string(lcfg, 1), rc);
2007 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2009 if (IS_ERR(obd->obd_fsops))
2010 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2012 rc = filter_iobuf_pool_init(filter);
2016 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2017 CERROR("%s: Underlying device is marked as read-only. "
2018 "Setup failed\n", obd->obd_name);
2019 GOTO(err_ops, rc = -EROFS);
2022 /* failover is the default */
2023 obd->obd_replayable = 1;
2025 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2026 str = lustre_cfg_string(lcfg, 3);
2027 if (strchr(str, 'n')) {
2028 CWARN("%s: recovery disabled\n", obd->obd_name);
2029 obd->obd_replayable = 0;
2033 filter->fo_vfsmnt = mnt;
2034 obd->u.obt.obt_sb = mnt->mnt_sb;
2035 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2036 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2038 fsfilt_setup(obd, obd->u.obt.obt_sb);
2040 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2041 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2042 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2043 obd->obd_lvfs_ctxt.fs = get_ds();
2044 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2046 init_mutex(&filter->fo_init_lock);
2047 filter->fo_committed_group = 0;
2049 rc = filter_prep(obd);
2053 filter->fo_destroys_in_progress = 0;
2054 for (i = 0; i < 32; i++)
2055 sema_init(&filter->fo_create_locks[i], 1);
2057 spin_lock_init(&filter->fo_translock);
2058 spin_lock_init(&filter->fo_objidlock);
2059 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2060 sema_init(&filter->fo_alloc_lock, 1);
2061 init_brw_stats(&filter->fo_filter_stats);
2062 filter->fo_read_cache = 1; /* enable read-only cache by default */
2063 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2064 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2065 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2066 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2068 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2069 spin_lock_init(&filter->fo_llog_list_lock);
2071 filter->fo_fl_oss_capa = 1;
2072 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2073 filter->fo_capa_hash = init_capa_hash();
2074 if (filter->fo_capa_hash == NULL)
2075 GOTO(err_ops, rc = -ENOMEM);
2077 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2078 obd->obd_namespace = ldlm_namespace_new(obd, ns_name, LDLM_NAMESPACE_SERVER,
2079 LDLM_NAMESPACE_GREEDY);
2080 if (obd->obd_namespace == NULL)
2081 GOTO(err_post, rc = -ENOMEM);
2082 obd->obd_namespace->ns_lvbp = obd;
2083 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2084 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2086 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2087 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2089 rc = obd_llog_init(obd, &obd->obd_olg, obd, 1, NULL, NULL);
2091 CERROR("failed to setup llogging subsystems\n");
2095 rwlock_init(&filter->fo_sptlrpc_lock);
2096 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2097 /* do this after llog being initialized */
2098 filter_adapt_sptlrpc_conf(obd, 1);
2100 rc = lquota_setup(filter_quota_interface_ref, obd);
2104 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2105 if (q->max_sectors < q->max_hw_sectors &&
2106 q->max_sectors < PTLRPC_MAX_BRW_SIZE >> 9)
2107 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2108 "for larger I/O requests: max_sectors = %u "
2109 "could be up to max_hw_sectors=%u\n",
2110 obd->obd_name, mnt->mnt_sb->s_id,
2111 q->max_sectors, q->max_hw_sectors);
2113 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2114 if (uuid_ptr != NULL) {
2115 class_uuid_unparse(uuid_ptr, &uuid);
2121 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2123 if (obd->obd_recovering) {
2124 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
2125 "recovery for at least %d:%.02d, or until %d "
2126 "client%s reconnect. During this time new clients"
2127 " will not be allowed to connect. "
2128 "Recovery progress can be monitored by watching "
2129 "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
2130 obd->obd_name, lustre_cfg_string(lcfg, 1),
2131 label ?: "", label ? "/" : "", str,
2132 obd->obd_recovery_timeout / 60,
2133 obd->obd_recovery_timeout % 60,
2134 obd->obd_max_recoverable_clients,
2135 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2138 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2139 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2140 label ?: "", label ? "/" : "", str,
2141 obd->obd_replayable ? "enabled" : "disabled");
2149 fsfilt_put_ops(obd->obd_fsops);
2150 filter_iobuf_pool_done(filter);
2152 server_put_mount(obd->obd_name, mnt);
2153 obd->u.obt.obt_sb = 0;
2157 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2159 struct lprocfs_static_vars lvars;
2164 CLASSERT(offsetof(struct obd_device, u.obt) ==
2165 offsetof(struct obd_device, u.filter.fo_obt));
2167 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2170 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2171 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2174 addr = (unsigned long)cfs_page_address(page);
2175 clear_page((void *)addr);
2177 /* lprocfs must be setup before the filter so state can be safely added
2178 * to /proc incrementally as the filter is setup */
2179 lprocfs_filter_init_vars(&lvars);
2180 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2181 lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2182 /* Init obdfilter private stats here */
2183 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2184 LPROCFS_CNTR_AVGMINMAX,
2185 "read_bytes", "bytes");
2186 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2187 LPROCFS_CNTR_AVGMINMAX,
2188 "write_bytes", "bytes");
2189 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2190 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2191 "get_page", "usec");
2192 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2193 LPROCFS_CNTR_AVGMINMAX,
2194 "get_page_failures", "num");
2195 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2196 LPROCFS_CNTR_AVGMINMAX,
2197 "cache_access", "pages");
2198 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2199 LPROCFS_CNTR_AVGMINMAX,
2200 "cache_hit", "pages");
2201 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2202 LPROCFS_CNTR_AVGMINMAX,
2203 "cache_miss", "pages");
2205 lproc_filter_attach_seqstat(obd);
2206 obd->obd_proc_exports_entry = lprocfs_register("exports",
2207 obd->obd_proc_entry,
2209 if (IS_ERR(obd->obd_proc_exports_entry)) {
2210 rc = PTR_ERR(obd->obd_proc_exports_entry);
2211 CERROR("error %d setting up lprocfs for %s\n",
2213 obd->obd_proc_exports_entry = NULL;
2216 if (obd->obd_proc_exports_entry)
2217 lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2218 lprocfs_nid_stats_clear_read,
2219 lprocfs_nid_stats_clear_write, obd, NULL);
2221 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2222 LUSTRE_CFG_BUFLEN(lcfg, 4));
2223 rc = filter_common_setup(obd, lcfg, (void *)addr);
2224 OBD_PAGE_FREE(page);
2227 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2228 lprocfs_free_per_client_stats(obd);
2229 lprocfs_free_obd_stats(obd);
2230 lprocfs_obd_cleanup(obd);
2236 static struct llog_operations filter_mds_ost_repl_logops;
2238 static struct llog_operations filter_size_orig_logops = {
2239 .lop_setup = llog_obd_origin_setup,
2240 .lop_cleanup = llog_obd_origin_cleanup,
2241 .lop_add = llog_obd_origin_add
2244 static int filter_olg_fini(struct obd_llog_group *olg)
2246 struct llog_ctxt *ctxt;
2247 int rc = 0, rc2 = 0;
2250 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2252 rc = llog_cleanup(ctxt);
2254 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2256 rc2 = llog_cleanup(ctxt);
2261 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2263 rc2 = llog_cleanup(ctxt);
2272 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2273 struct obd_device *tgt)
2278 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2279 &filter_mds_ost_repl_logops);
2283 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2284 &filter_size_orig_logops);
2290 filter_olg_fini(olg);
2295 * Init the default olg, which is embeded in the obd_device, for filter.
2298 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2299 struct obd_device *tgt)
2301 struct filter_obd *filter = &obd->u.filter;
2302 struct llog_ctxt *ctxt;
2306 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2307 if (!filter->fo_lcm)
2310 filter_mds_ost_repl_logops = llog_client_ops;
2311 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2312 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2313 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2315 rc = filter_olg_init(obd, olg, tgt);
2317 GOTO(cleanup_lcm, rc);
2319 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2322 GOTO(cleanup_olg, rc);
2324 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2326 CERROR("Can't get ctxt for %p:%x\n", olg,
2327 LLOG_MDS_OST_REPL_CTXT);
2328 GOTO(cleanup_olg, rc = -ENODEV);
2330 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2331 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2332 llog_ctxt_put(ctxt);
2336 filter_olg_fini(olg);
2338 llog_recov_thread_fini(filter->fo_lcm, 1);
2339 filter->fo_lcm = NULL;
2344 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2345 struct obd_device *tgt, int count, struct llog_catid *catid,
2346 struct obd_uuid *uuid)
2348 struct filter_obd *filter = &obd->u.filter;
2349 struct llog_ctxt *ctxt;
2353 LASSERT(olg != NULL);
2354 if (olg == &obd->obd_olg)
2355 return filter_default_olg_init(obd, olg, tgt);
2357 LASSERT(filter->fo_lcm != NULL);
2358 rc = filter_olg_init(obd, olg, tgt);
2361 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2363 CERROR("Can't get ctxt for %p:%x\n", olg,
2364 LLOG_MDS_OST_REPL_CTXT);
2365 filter_olg_fini(olg);
2368 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2369 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2370 llog_ctxt_put(ctxt);
2374 static int filter_llog_finish(struct obd_device *obd, int count)
2376 struct filter_obd *filter = &obd->u.filter;
2377 struct llog_ctxt *ctxt;
2380 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2383 * Make sure that no cached llcds left in recov_thread.
2384 * We actually do sync in disconnect time, but disconnect
2385 * may not come being marked rq_no_resend = 1.
2387 llog_sync(ctxt, NULL);
2390 * Balance class_import_get() in llog_receptor_accept().
2391 * This is safe to do, as llog is already synchronized
2392 * and its import may go.
2394 mutex_down(&ctxt->loc_sem);
2395 if (ctxt->loc_imp) {
2396 class_import_put(ctxt->loc_imp);
2397 ctxt->loc_imp = NULL;
2399 mutex_up(&ctxt->loc_sem);
2400 llog_ctxt_put(ctxt);
2403 if (filter->fo_lcm) {
2404 mutex_down(&ctxt->loc_sem);
2405 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2406 filter->fo_lcm = NULL;
2407 mutex_up(&ctxt->loc_sem);
2409 RETURN(filter_olg_fini(&obd->obd_olg));
2413 * Find the group llog according to group index in the llog group list.
2415 static struct obd_llog_group *
2416 filter_find_olg_internal(struct filter_obd *filter, int group)
2418 struct obd_llog_group *olg;
2420 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2421 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2422 if (olg->olg_group == group)
2429 * Find the group llog according to group index on the filter
2431 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2433 struct obd_llog_group *olg = NULL;
2434 struct filter_obd *filter;
2436 filter = &obd->u.filter;
2438 if (group == FILTER_GROUP_LLOG)
2439 RETURN(&obd->obd_olg);
2441 spin_lock(&filter->fo_llog_list_lock);
2442 olg = filter_find_olg_internal(filter, group);
2443 spin_unlock(&filter->fo_llog_list_lock);
2448 * Find the llog_group of the filter according to the group. If it can not
2449 * find, create the llog_group, which only happens when mds is being synced
2452 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2454 struct obd_llog_group *olg = NULL;
2455 struct filter_obd *filter;
2458 filter = &obd->u.filter;
2460 if (group == FILTER_GROUP_LLOG)
2461 RETURN(&obd->obd_olg);
2463 spin_lock(&filter->fo_llog_list_lock);
2464 olg = filter_find_olg_internal(filter, group);
2466 if (olg->olg_initializing) {
2467 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2469 GOTO(out_unlock, olg);
2474 GOTO(out_unlock, olg = ERR_PTR(-ENOMEM));
2476 llog_group_init(olg, group);
2477 list_add(&olg->olg_list, &filter->fo_llog_list);
2478 olg->olg_initializing = 1;
2479 spin_unlock(&filter->fo_llog_list_lock);
2481 rc = llog_cat_initialize(obd, olg, 1, NULL);
2483 spin_lock(&filter->fo_llog_list_lock);
2484 list_del(&olg->olg_list);
2485 spin_unlock(&filter->fo_llog_list_lock);
2487 GOTO(out, olg = ERR_PTR(-ENOMEM));
2489 spin_lock(&filter->fo_llog_list_lock);
2490 olg->olg_initializing = 0;
2491 spin_unlock(&filter->fo_llog_list_lock);
2492 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2493 obd->obd_name, group, olg);
2498 spin_unlock(&filter->fo_llog_list_lock);
2502 static int filter_llog_connect(struct obd_export *exp,
2503 struct llogd_conn_body *body)
2505 struct obd_device *obd = exp->exp_obd;
2506 struct llog_ctxt *ctxt;
2507 struct obd_llog_group *olg;
2511 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2512 obd->obd_name, body->lgdc_logid.lgl_oid,
2513 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2515 olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2517 CERROR(" %s: can not find olg of group %d\n",
2518 obd->obd_name, (int)body->lgdc_logid.lgl_ogr);
2521 llog_group_set_export(olg, exp);
2523 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2524 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2525 body->lgdc_ctxt_idx);
2527 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2528 obd->obd_name, body->lgdc_logid.lgl_oid,
2529 body->lgdc_logid.lgl_ogr, body->lgdc_logid.lgl_ogen);
2531 rc = llog_connect(ctxt, &body->lgdc_logid,
2532 &body->lgdc_gen, NULL);
2533 llog_ctxt_put(ctxt);
2535 CERROR("failed to connect rc %d idx %d\n", rc,
2536 body->lgdc_ctxt_idx);
2541 static int filter_llog_preclean(struct obd_device *obd)
2543 struct obd_llog_group *olg, *tmp;
2544 struct filter_obd *filter;
2545 struct list_head remove_list;
2549 rc = obd_llog_finish(obd, 0);
2551 CERROR("failed to cleanup llogging subsystem\n");
2553 filter = &obd->u.filter;
2554 CFS_INIT_LIST_HEAD(&remove_list);
2556 spin_lock(&filter->fo_llog_list_lock);
2557 while (!list_empty(&filter->fo_llog_list)) {
2558 olg = list_entry(filter->fo_llog_list.next,
2559 struct obd_llog_group, olg_list);
2560 list_del(&olg->olg_list);
2561 list_add(&olg->olg_list, &remove_list);
2563 spin_unlock(&filter->fo_llog_list_lock);
2565 list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2566 list_del_init(&olg->olg_list);
2567 rc = filter_olg_fini(olg);
2569 CERROR("failed to cleanup llogging subsystem for %u\n",
2577 static int filter_precleanup(struct obd_device *obd,
2578 enum obd_cleanup_stage stage)
2584 case OBD_CLEANUP_EARLY:
2586 case OBD_CLEANUP_EXPORTS:
2587 target_cleanup_recovery(obd);
2588 rc = filter_llog_preclean(obd);
2594 static int filter_cleanup(struct obd_device *obd)
2596 struct filter_obd *filter = &obd->u.filter;
2600 LCONSOLE_WARN("%s: shutting down for failover; client state "
2601 "will be preserved.\n", obd->obd_name);
2603 if (!list_empty(&obd->obd_exports)) {
2604 CERROR("%s: still has clients!\n", obd->obd_name);
2605 class_disconnect_exports(obd);
2606 if (!list_empty(&obd->obd_exports)) {
2607 CERROR("still has exports after forced cleanup?\n");
2612 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2613 lprocfs_free_per_client_stats(obd);
2614 lprocfs_free_obd_stats(obd);
2615 lprocfs_obd_cleanup(obd);
2616 lquota_cleanup(filter_quota_interface_ref, obd);
2618 /* Stop recovery before namespace cleanup. */
2619 target_stop_recovery_thread(obd);
2620 target_cleanup_recovery(obd);
2622 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2623 obd->obd_namespace = NULL;
2625 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2627 if (obd->u.obt.obt_sb == NULL)
2632 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2633 shrink_dcache_sb(obd->u.obt.obt_sb);
2635 server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2636 obd->u.obt.obt_sb = NULL;
2638 fsfilt_put_ops(obd->obd_fsops);
2640 filter_iobuf_pool_done(filter);
2642 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2647 static int filter_connect_internal(struct obd_export *exp,
2648 struct obd_connect_data *data)
2650 struct filter_export_data *fed = &exp->exp_filter_data;
2655 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2656 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2657 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2658 data->ocd_connect_flags, data->ocd_version,
2659 data->ocd_grant, data->ocd_index);
2661 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2662 CWARN("!!! This export (nid %s) used object group %d "
2663 "earlier; now it's trying to use group %d! This could "
2664 "be a bug in the MDS. Please report to "
2665 "http://bugzilla.lustre.org/\n",
2666 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2669 fed->fed_group = data->ocd_group;
2671 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2672 exp->exp_connect_flags = data->ocd_connect_flags;
2673 data->ocd_version = LUSTRE_VERSION_CODE;
2675 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2676 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2677 data->ocd_connect_flags & OBD_CONNECT_MDS))
2680 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2681 struct filter_obd *filter = &exp->exp_obd->u.filter;
2682 obd_size left, want;
2684 spin_lock(&exp->exp_obd->obd_osfs_lock);
2685 left = filter_grant_space_left(exp);
2686 want = data->ocd_grant;
2687 filter_grant(exp, fed->fed_grant, want, left);
2688 data->ocd_grant = fed->fed_grant;
2689 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2691 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2692 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2693 exp->exp_client_uuid.uuid, exp,
2694 data->ocd_grant, want, left);
2696 filter->fo_tot_granted_clients ++;
2699 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2700 struct filter_obd *filter = &exp->exp_obd->u.filter;
2701 struct lr_server_data *lsd = filter->fo_fsd;
2702 int index = le32_to_cpu(lsd->lsd_ost_index);
2704 if (!(lsd->lsd_feature_compat &
2705 cpu_to_le32(OBD_COMPAT_OST))) {
2706 /* this will only happen on the first connect */
2707 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2708 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2709 filter_update_server_data(exp->exp_obd,
2710 filter->fo_rcvd_filp, lsd, 1);
2711 } else if (index != data->ocd_index) {
2712 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2713 " %u doesn't match actual OST index"
2714 " %u in last_rcvd file, bad "
2716 obd_export_nid2str(exp), index,
2722 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2723 data->ocd_brw_size = 65536;
2724 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2725 data->ocd_brw_size = min(data->ocd_brw_size,
2726 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2727 LASSERT(data->ocd_brw_size);
2730 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2731 __u32 cksum_types = data->ocd_cksum_types;
2733 /* The client set in ocd_cksum_types the checksum types it
2734 * supports. We have to mask off the algorithms that we don't
2736 if (cksum_types & OBD_CKSUM_ALL)
2737 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2739 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2741 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2742 "%x\n", exp->exp_obd->obd_name,
2743 obd_export_nid2str(exp), cksum_types,
2744 data->ocd_cksum_types);
2746 /* This client does not support OBD_CONNECT_CKSUM
2747 * fall back to CRC32 */
2748 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2749 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2750 exp->exp_obd->obd_name,
2751 obd_export_nid2str(exp));
2754 /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2755 * FIXME: We don't strictly need the COMPAT flag for that,
2756 * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2757 * FIXME: We needed it for the index, as index 0 is valid. */
2762 static int filter_reconnect(const struct lu_env *env,
2763 struct obd_export *exp, struct obd_device *obd,
2764 struct obd_uuid *cluuid,
2765 struct obd_connect_data *data,
2771 if (exp == NULL || obd == NULL || cluuid == NULL)
2774 rc = filter_connect_internal(exp, data);
2776 filter_export_stats_init(obd, exp, localdata);
2781 /* nearly identical to mds_connect */
2782 static int filter_connect(const struct lu_env *env,
2783 struct obd_export **exp, struct obd_device *obd,
2784 struct obd_uuid *cluuid,
2785 struct obd_connect_data *data, void *localdata)
2787 struct lvfs_run_ctxt saved;
2788 struct lustre_handle conn = { 0 };
2789 struct obd_export *lexp;
2790 struct filter_export_data *fed;
2791 struct lsd_client_data *lcd = NULL;
2796 if (exp == NULL || obd == NULL || cluuid == NULL)
2799 rc = class_connect(&conn, obd, cluuid);
2802 lexp = class_conn2export(&conn);
2803 LASSERT(lexp != NULL);
2805 fed = &lexp->exp_filter_data;
2807 rc = filter_connect_internal(lexp, data);
2811 filter_export_stats_init(obd, lexp, localdata);
2812 if (obd->obd_replayable) {
2813 OBD_ALLOC(lcd, sizeof(*lcd));
2815 CERROR("filter: out of memory for client data\n");
2816 GOTO(cleanup, rc = -ENOMEM);
2819 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2821 rc = filter_client_add(obd, lexp, -1);
2826 group = data->ocd_group;
2828 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2829 obd->obd_name, lexp->exp_handle.h_cookie, group);
2831 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2832 rc = filter_read_groups(obd, group, 1);
2833 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2835 CERROR("can't read group %u\n", group);
2845 fed->fed_lcd = NULL;
2847 class_disconnect(lexp);
2856 /* Do extra sanity checks for grant accounting. We do this at connect,
2857 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2858 * always get rid of it or turn it off when we know accounting is good. */
2859 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2861 struct filter_export_data *fed;
2862 struct obd_export *exp;
2863 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2864 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2865 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2867 if (list_empty(&obd->obd_exports))
2870 /* We don't want to do this for large machines that do lots of
2871 mounts or unmounts. It burns... */
2872 if (obd->obd_num_exports > 100)
2875 spin_lock(&obd->obd_osfs_lock);
2876 spin_lock(&obd->obd_dev_lock);
2877 list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2879 fed = &exp->exp_filter_data;
2880 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2883 if (maxsize > 0) { /* we may not have done a statfs yet */
2884 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2885 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2886 exp->exp_client_uuid.uuid, exp,
2887 fed->fed_grant, fed->fed_pending, maxsize);
2888 LASSERTF(fed->fed_dirty <= maxsize,
2889 "%s: cli %s/%p %ld > "LPU64"\n", func,
2890 exp->exp_client_uuid.uuid, exp,
2891 fed->fed_dirty, maxsize);
2894 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2895 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2896 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2898 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2899 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2900 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2901 tot_granted += fed->fed_grant + fed->fed_pending;
2902 tot_pending += fed->fed_pending;
2903 tot_dirty += fed->fed_dirty;
2905 fo_tot_granted = obd->u.filter.fo_tot_granted;
2906 fo_tot_pending = obd->u.filter.fo_tot_pending;
2907 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2908 spin_unlock(&obd->obd_dev_lock);
2909 spin_unlock(&obd->obd_osfs_lock);
2911 /* Do these assertions outside the spinlocks so we don't kill system */
2912 if (tot_granted != fo_tot_granted)
2913 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2914 func, tot_granted, fo_tot_granted);
2915 if (tot_pending != fo_tot_pending)
2916 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2917 func, tot_pending, fo_tot_pending);
2918 if (tot_dirty != fo_tot_dirty)
2919 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2920 func, tot_dirty, fo_tot_dirty);
2921 if (tot_pending > tot_granted)
2922 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2923 func, tot_pending, tot_granted);
2924 if (tot_granted > maxsize)
2925 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2926 func, tot_granted, maxsize);
2927 if (tot_dirty > maxsize)
2928 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2929 func, tot_dirty, maxsize);
2932 /* Remove this client from the grant accounting totals. We also remove
2933 * the export from the obd device under the osfs and dev locks to ensure
2934 * that the filter_grant_sanity_check() calculations are always valid.
2935 * The client should do something similar when it invalidates its import. */
2936 static void filter_grant_discard(struct obd_export *exp)
2938 struct obd_device *obd = exp->exp_obd;
2939 struct filter_obd *filter = &obd->u.filter;
2940 struct filter_export_data *fed = &exp->exp_filter_data;
2942 spin_lock(&obd->obd_osfs_lock);
2943 spin_lock(&obd->obd_dev_lock);
2944 list_del_init(&exp->exp_obd_chain);
2945 spin_unlock(&obd->obd_dev_lock);
2947 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2948 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2949 obd->obd_name, filter->fo_tot_granted,
2950 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2951 filter->fo_tot_granted -= fed->fed_grant;
2952 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2953 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2954 obd->obd_name, filter->fo_tot_pending,
2955 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2956 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2957 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2958 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2959 obd->obd_name, filter->fo_tot_dirty,
2960 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2961 filter->fo_tot_dirty -= fed->fed_dirty;
2965 spin_unlock(&obd->obd_osfs_lock);
2968 static int filter_destroy_export(struct obd_export *exp)
2972 if (exp->exp_filter_data.fed_pending)
2973 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2974 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2975 exp, exp->exp_filter_data.fed_pending);
2977 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2979 target_destroy_export(exp);
2980 ldlm_destroy_export(exp);
2982 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2986 if (exp->exp_obd->obd_replayable)
2987 filter_client_free(exp);
2989 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2991 filter_grant_discard(exp);
2992 filter_fmd_cleanup(exp);
2994 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2995 struct filter_obd *filter = &exp->exp_obd->u.filter;
2996 if (filter->fo_tot_granted_clients > 0)
2997 filter->fo_tot_granted_clients --;
3000 if (!(exp->exp_flags & OBD_OPT_FORCE))
3001 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
3006 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3008 struct obd_llog_group *olg_min, *olg;
3009 struct filter_obd *filter;
3010 int worked = -1, group;
3011 struct llog_ctxt *ctxt;
3014 filter = &obd->u.filter;
3016 /* we can't sync log holding spinlock. also, we do not want to get
3017 * into livelock. so we do following: loop over MDS's exports in
3018 * group order and skip already synced llogs -bzzz */
3020 /* look for group with min. number, but > worked */
3023 spin_lock(&filter->fo_llog_list_lock);
3024 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3025 if (olg->olg_group <= worked) {
3026 /* this group is already synced */
3029 if (group < olg->olg_group) {
3030 /* we have group with smaller number to sync */
3033 /* store current minimal group */
3035 group = olg->olg_group;
3037 spin_unlock(&filter->fo_llog_list_lock);
3039 if (olg_min == NULL)
3042 worked = olg_min->olg_group;
3043 if (olg_min->olg_exp &&
3044 (dexp == olg_min->olg_exp || dexp == NULL)) {
3046 ctxt = llog_group_get_ctxt(olg_min,
3047 LLOG_MDS_OST_REPL_CTXT);
3049 err = llog_sync(ctxt, olg_min->olg_exp);
3050 llog_ctxt_put(ctxt);
3052 CERROR("error flushing logs to MDS: "
3057 } while (olg_min != NULL);
3060 /* Also incredibly similar to mds_disconnect */
3061 static int filter_disconnect(struct obd_export *exp)
3063 struct obd_device *obd = exp->exp_obd;
3068 class_export_get(exp);
3070 if (!(exp->exp_flags & OBD_OPT_FORCE))
3071 filter_grant_sanity_check(obd, __FUNCTION__);
3072 filter_grant_discard(exp);
3074 /* Flush any remaining cancel messages out to the target */
3075 filter_sync_llogs(obd, exp);
3077 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3079 /* Disconnect early so that clients can't keep using export */
3080 rc = class_disconnect(exp);
3081 if (exp->exp_obd->obd_namespace != NULL)
3082 ldlm_cancel_locks_for_export(exp);
3084 fsfilt_sync(obd, obd->u.obt.obt_sb);
3086 lprocfs_exp_cleanup(exp);
3088 class_export_put(exp);
3092 /* reverse import is changed, sync all cancels */
3093 static void filter_revimp_update(struct obd_export *exp)
3098 class_export_get(exp);
3100 /* flush any remaining cancel messages out to the target */
3101 filter_sync_llogs(exp->exp_obd, exp);
3102 class_export_put(exp);
3106 static int filter_ping(struct obd_export *exp)
3108 filter_fmd_expire(exp);
3113 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
3114 const char *what, int quiet)
3116 struct dentry *dchild = NULL;
3119 if (oa->o_valid & OBD_MD_FLGROUP)
3122 dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
3124 if (IS_ERR(dchild)) {
3125 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
3126 what, group, oa->o_id);
3130 if (dchild->d_inode == NULL) {
3132 CERROR("%s: %s on non-existent object: "LPU64"\n",
3133 obd->obd_name, what, oa->o_id);
3135 RETURN(ERR_PTR(-ENOENT));
3141 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3143 struct dentry *dentry = NULL;
3144 struct obd_device *obd;
3148 rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3149 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3153 obd = class_exp2obd(exp);
3155 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3159 dentry = filter_oa2dentry(obd, oinfo->oi_oa);
3161 RETURN(PTR_ERR(dentry));
3163 /* Limit the valid bits in the return data to what we actually use */
3164 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3165 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3171 /* this should be enabled/disabled in condition to enabled/disabled large
3172 * inodes (fast EAs) in backing store FS. */
3173 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3174 void *handle, struct obdo *oa)
3176 struct obd_device *obd = exp->exp_obd;
3180 if (oa->o_valid & OBD_MD_FLFID) {
3181 struct filter_fid ff;
3183 if (!(oa->o_valid & OBD_MD_FLGROUP))
3185 /* packing fid and converting it to LE for storing into EA.
3186 * Here ->o_stripe_idx should be filled by LOV and rest of
3187 * fields - by client. */
3188 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
3189 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
3190 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
3191 ff.ff_objid = cpu_to_le64(oa->o_id);
3192 ff.ff_group = cpu_to_le64(oa->o_gr);
3194 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
3195 LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
3196 oa->o_generation, oa->o_id, oa->o_gr);
3198 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3200 CERROR("store fid in object failed! rc: %d\n", rc);
3202 CDEBUG(D_HA, "OSS object without fid info!\n");
3208 /* this is called from filter_truncate() until we have filter_punch() */
3209 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3210 struct obdo *oa, struct obd_trans_info *oti)
3212 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3213 struct llog_cookie *fcc = NULL;
3214 struct filter_obd *filter;
3215 int rc, err, sync = 0;
3216 loff_t old_size = 0;
3217 unsigned int ia_valid;
3218 struct inode *inode;
3223 LASSERT(dentry != NULL);
3224 LASSERT(!IS_ERR(dentry));
3226 inode = dentry->d_inode;
3227 LASSERT(inode != NULL);
3229 filter = &exp->exp_obd->u.filter;
3230 iattr_from_obdo(&iattr, oa, oa->o_valid);
3231 ia_valid = iattr.ia_valid;
3233 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3234 OBD_ALLOC(fcc, sizeof(*fcc));
3236 *fcc = oa->o_lcookie;
3238 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3240 /* Filter truncates and writes are serialized by
3241 * i_alloc_sem, see the comment in
3242 * filter_preprw_write.*/
3243 if (ia_valid & ATTR_SIZE)
3244 down_write(&inode->i_alloc_sem);
3245 LOCK_INODE_MUTEX(inode);
3246 old_size = i_size_read(inode);
3249 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3250 * then we will accept the UID+GID sent by the client during write for
3251 * initializing the ownership of this inode. We only allow this to
3252 * happen once so clear these bits in setattr. In 2.6 kernels it is
3253 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3254 * the flags that are actually being set. */
3255 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3256 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3257 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3259 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3260 if (!(ia_valid & ATTR_MODE)) {
3261 iattr.ia_mode = inode->i_mode;
3262 iattr.ia_valid |= ATTR_MODE;
3264 iattr.ia_mode &= ~S_ISUID;
3266 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3267 if (!(iattr.ia_valid & ATTR_MODE)) {
3268 iattr.ia_mode = inode->i_mode;
3269 iattr.ia_valid |= ATTR_MODE;
3271 iattr.ia_mode &= ~S_ISGID;
3274 orig_ids[USRQUOTA] = inode->i_uid;
3275 orig_ids[GRPQUOTA] = inode->i_gid;
3276 handle = fsfilt_start_log(exp->exp_obd, inode,
3277 FSFILT_OP_SETATTR, oti, 1);
3279 GOTO(out_unlock, rc = PTR_ERR(handle));
3281 /* update inode EA only once when inode is suid bit marked. As
3282 * on 2.6.x UID and GID may be set separately, we check here
3283 * only one of them to avoid double setting. */
3284 if (inode->i_mode & S_ISUID)
3285 filter_update_fidea(exp, inode, handle, oa);
3287 handle = fsfilt_start(exp->exp_obd, inode,
3288 FSFILT_OP_SETATTR, oti);
3290 GOTO(out_unlock, rc = PTR_ERR(handle));
3292 if (oa->o_valid & OBD_MD_FLFLAGS) {
3293 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
3294 EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
3296 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3298 /* set cancel cookie callback function */
3299 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3300 filter_cancel_cookies_cb,
3304 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3305 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3307 /* The truncate might have used up our transaction credits. Make
3308 * sure we have one left for the last_rcvd update. */
3309 err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
3311 rc = filter_finish_transno(exp, oti, rc, sync);
3313 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3317 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3319 CERROR("error on commit, err = %d\n", err);
3326 /* For a partial-page truncate flush the page to disk immediately
3327 * to avoid data corruption during direct disk write. b=17397 */
3328 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3329 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3330 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3338 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3339 UNLOCK_INODE_MUTEX(inode);
3340 if (ia_valid & ATTR_SIZE)
3341 up_write(&inode->i_alloc_sem);
3343 OBD_FREE(fcc, sizeof(*fcc));
3345 /* trigger quota release */
3346 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3347 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3348 int rc2 = lquota_adjust(filter_quota_interface_ref,
3349 exp->exp_obd, cur_ids,
3350 orig_ids, rc, FSFILT_OP_SETATTR);
3351 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3352 "filter adjust qunit. (rc:%d)\n", rc2);
3357 /* this is called from filter_truncate() until we have filter_punch() */
3358 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3359 struct obd_trans_info *oti)
3361 struct obdo *oa = oinfo->oi_oa;
3362 struct lustre_capa *capa = oinfo_capa(oinfo);
3363 struct ldlm_res_id res_id;
3364 struct filter_mod_data *fmd;
3365 struct lvfs_run_ctxt saved;
3366 struct filter_obd *filter;
3367 struct ldlm_resource *res;
3368 struct dentry *dentry;
3369 __u64 opc = CAPA_OPC_META_WRITE;
3373 if (oa->o_valid & OBD_FL_TRUNC)
3374 opc |= CAPA_OPC_OSS_TRUNC;
3375 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa), capa, opc);
3379 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3380 rc = filter_capa_fixoa(exp, oa, obdo_mdsno(oa), capa);
3385 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
3386 /* This would be very bad - accidentally truncating a file when
3387 * changing the time or similar - bug 12203. */
3388 if (oa->o_valid & OBD_MD_FLSIZE &&
3389 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3390 static char mdsinum[48];
3392 if (oa->o_valid & OBD_MD_FLFID)
3393 snprintf(mdsinum, sizeof(mdsinum) - 1,
3394 " of inode "LPU64"/%u", oa->o_fid,
3399 CERROR("%s: setattr from %s trying to truncate objid "LPU64
3401 exp->exp_obd->obd_name, obd_export_nid2str(exp),
3406 dentry = __filter_oa2dentry(exp->exp_obd, oa, __FUNCTION__, 1);
3408 RETURN(PTR_ERR(dentry));
3410 filter = &exp->exp_obd->u.filter;
3411 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3415 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3416 fmd = filter_fmd_get(exp, oa->o_id, oa->o_gr);
3417 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3418 fmd->fmd_mactime_xid = oti->oti_xid;
3419 filter_fmd_put(exp, fmd);
3422 /* setting objects attributes (including owner/group) */
3423 rc = filter_setattr_internal(exp, dentry, oa, oti);
3425 GOTO(out_unlock, rc);
3427 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3428 &res_id, LDLM_EXTENT, 0);
3431 LDLM_RESOURCE_ADDREF(res);
3432 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3433 LDLM_RESOURCE_DELREF(res);
3434 ldlm_resource_putref(res);
3437 oa->o_valid = OBD_MD_FLID;
3439 /* Quota release need uid/gid info */
3440 obdo_from_inode(oa, dentry->d_inode,
3441 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3447 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3451 /* XXX identical to osc_unpackmd */
3452 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3453 struct lov_mds_md *lmm, int lmm_bytes)
3459 if (lmm_bytes < sizeof (*lmm)) {
3460 CERROR("lov_mds_md too small: %d, need %d\n",
3461 lmm_bytes, (int)sizeof(*lmm));
3464 /* XXX LOV_MAGIC etc check? */
3466 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3467 CERROR("lov_mds_md: zero lmm_object_id\n");
3472 lsm_size = lov_stripe_md_size(1);
3476 if (*lsmp != NULL && lmm == NULL) {
3477 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3478 OBD_FREE(*lsmp, lsm_size);
3483 if (*lsmp == NULL) {
3484 OBD_ALLOC(*lsmp, lsm_size);
3488 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3489 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3490 OBD_FREE(*lsmp, lsm_size);
3493 loi_init((*lsmp)->lsm_oinfo[0]);
3497 /* XXX zero *lsmp? */
3498 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3499 LASSERT((*lsmp)->lsm_object_id);
3502 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3507 /* caller must hold fo_create_locks[oa->o_gr] */
3508 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3509 struct filter_obd *filter)
3511 struct obdo doa; /* XXX obdo on stack */
3518 LASSERT_MDS_GROUP(oa->o_gr);
3519 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3520 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3522 memset(&doa, 0, sizeof(doa));
3524 doa.o_valid |= OBD_MD_FLGROUP;
3525 doa.o_gr = oa->o_gr;
3526 doa.o_mode = S_IFREG;
3528 if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3529 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3530 exp->exp_obd->obd_name, doa.o_gr);
3534 last = filter_last_id(filter, doa.o_gr);
3536 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3538 CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3539 exp->exp_obd->obd_name, oa->o_id + 1, last,
3540 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3542 for (id = last; id > oa->o_id; id--) {
3544 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3545 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3546 CEMERG("error destroying precreate objid "LPU64": %d\n",
3549 /* update last_id on disk periodically so that if we restart
3550 * we don't need to re-scan all of the just-deleted objects. */
3551 if ((id & 511) == 0 && !skip_orphan) {
3552 filter_set_last_id(filter, id - 1, doa.o_gr);
3553 filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3557 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3558 exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3561 filter_set_last_id(filter, id, doa.o_gr);
3562 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3564 /* don't reuse orphan object, return last used objid */
3568 clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3573 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3574 obd_gr group, int *num);
3575 /* returns a negative error or a nonnegative number of files to create */
3576 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3577 obd_gr group, struct obd_trans_info *oti)
3579 struct obd_device *obd = exp->exp_obd;
3580 struct filter_obd *filter = &obd->u.filter;
3584 /* delete orphans request */
3585 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3586 obd_id last = filter_last_id(filter, group);
3588 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3589 CERROR("%s: dropping old orphan cleanup request\n",
3593 /* This causes inflight precreates to abort and drop lock */
3594 set_bit(group, &filter->fo_destroys_in_progress);
3595 down(&filter->fo_create_locks[group]);
3596 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3597 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3598 exp->exp_obd->obd_name, group);
3599 up(&filter->fo_create_locks[group]);
3602 diff = oa->o_id - last;
3603 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3606 if (-diff > OST_MAX_PRECREATE) {
3607 CERROR("%s: ignoring bogus orphan destroy request: "
3608 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3610 /* FIXME: should reset precreate_next_id on MDS */
3611 GOTO(out, rc = -EINVAL);
3614 rc = filter_destroy_precreated(exp, oa, filter);
3616 CERROR("%s: unable to write lastobjid, but "
3617 "orphans were deleted\n", obd->obd_name);
3620 /* XXX: Used by MDS for the first time! */
3621 clear_bit(group, &filter->fo_destroys_in_progress);
3624 down(&filter->fo_create_locks[group]);
3625 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3626 CERROR("%s: dropping old precreate request\n",
3630 /* only precreate if group == 0 and o_id is specfied */
3631 if (group == FILTER_GROUP_LLOG || oa->o_id == 0)
3634 diff = oa->o_id - filter_last_id(filter, group);
3635 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3636 filter_last_id(filter, group), diff);
3638 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3639 oa->o_id, filter_last_id(filter, group), diff);
3643 oa->o_id = filter_last_id(&obd->u.filter, group);
3644 rc = filter_precreate(obd, oa, group, &diff);
3645 oa->o_id = filter_last_id(&obd->u.filter, group);
3647 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3650 /* else diff == 0 */
3653 up(&filter->fo_create_locks[group]);
3657 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3658 __u64 max_age, __u32 flags)
3660 struct filter_obd *filter = &obd->u.filter;
3661 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3665 /* at least try to account for cached pages. its still racey and
3666 * might be under-reporting if clients haven't announced their
3667 * caches with brw recently */
3668 spin_lock(&obd->obd_osfs_lock);
3669 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3670 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3671 spin_unlock(&obd->obd_osfs_lock);
3673 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3674 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3675 filter->fo_tot_dirty, filter->fo_tot_granted,
3676 filter->fo_tot_pending,
3677 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3679 filter_grant_sanity_check(obd, __FUNCTION__);
3681 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3682 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3683 osfs->os_bsize - 1) >> blockbits));
3685 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3686 * stop creating files on MDS if OST is not good shape to create
3688 osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3693 static int filter_use_existing_obj(struct obd_device *obd,
3694 struct dentry *dchild, void **handle,
3697 struct inode *inode = dchild->d_inode;
3701 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3704 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3705 if (IS_ERR(*handle))
3706 return PTR_ERR(*handle);
3708 iattr.ia_valid = ATTR_MODE;
3709 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3710 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3718 /* We rely on the fact that only one thread will be creating files in a given
3719 * group at a time, which is why we don't need an atomic filter_get_new_id.
3720 * Even if we had that atomic function, the following race would exist:
3722 * thread 1: gets id x from filter_next_id
3723 * thread 2: gets id (x + 1) from filter_next_id
3724 * thread 2: creates object (x + 1)
3725 * thread 1: tries to create object x, gets -ENOSPC
3727 * Caller must hold fo_create_locks[group]
3729 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3730 obd_gr group, int *num)
3732 struct dentry *dchild = NULL, *dparent = NULL;
3733 struct filter_obd *filter;
3734 struct obd_statfs *osfs;
3735 int err = 0, rc = 0, recreate_obj = 0, i;
3736 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3738 void *handle = NULL;
3741 filter = &obd->u.filter;
3743 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3745 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3747 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3748 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3751 OBD_ALLOC(osfs, sizeof(*osfs));
3754 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ, 0);
3755 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3756 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3757 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3758 filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3762 OBD_FREE(osfs, sizeof(*osfs));
3767 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3768 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3770 for (i = 0; i < *num && err == 0; i++) {
3771 int cleanup_phase = 0;
3773 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3774 CWARN("%s: create aborted by destroy\n",
3783 last_id = filter_last_id(filter, group);
3784 if (next_id > last_id) {
3785 CERROR("Error: Trying to recreate obj greater"
3786 "than last id "LPD64" > "LPD64"\n",
3788 GOTO(cleanup, rc = -EINVAL);
3791 next_id = filter_last_id(filter, group) + 1;
3793 dparent = filter_parent_lock(obd, group, next_id);
3794 if (IS_ERR(dparent))
3795 GOTO(cleanup, rc = PTR_ERR(dparent));
3796 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3798 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3800 GOTO(cleanup, rc = PTR_ERR(dchild));
3801 cleanup_phase = 2; /* f_dput(dchild) */
3803 if (dchild->d_inode != NULL) {
3804 /* This would only happen if lastobjid was bad on disk*/
3805 /* Could also happen if recreating missing obj but it
3806 * already exists. */
3808 CERROR("%s: recreating existing object %.*s?\n",
3809 obd->obd_name, dchild->d_name.len,
3810 dchild->d_name.name);
3812 /* Use these existing objects if they are
3814 if (dchild->d_inode->i_size == 0) {
3815 rc = filter_use_existing_obj(obd,dchild,
3816 &handle, &cleanup_phase);
3823 CERROR("%s: Serious error: objid %.*s already "
3824 "exists; is this filesystem corrupt?\n",
3825 obd->obd_name, dchild->d_name.len,
3826 dchild->d_name.name);
3829 GOTO(cleanup, rc = -EEXIST);
3832 handle = fsfilt_start_log(obd, dparent->d_inode,
3833 FSFILT_OP_CREATE, NULL, 1);
3835 GOTO(cleanup, rc = PTR_ERR(handle));
3838 CDEBUG(D_INODE, "%s: filter_precreate(od->o_gr="LPU64
3839 ",od->o_id="LPU64")\n", obd->obd_name, group,
3842 /* We mark object SUID+SGID to flag it for accepting UID+GID
3843 * from client on first write. Currently the permission bits
3844 * on the OST are never used, so this is OK. */
3845 rc = ll_vfs_create(dparent->d_inode, dchild,
3846 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3848 CERROR("create failed rc = %d\n", rc);
3853 if (!recreate_obj) {
3854 filter_set_last_id(filter, next_id, group);
3855 err = filter_update_last_objid(obd, group, 0);
3857 CERROR("unable to write lastobjid "
3858 "but file created\n");
3862 switch(cleanup_phase) {
3864 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3866 CERROR("error on commit, err = %d\n", err);
3873 filter_parent_unlock(dparent);
3880 if (time_after(jiffies, enough_time)) {
3882 "%s: precreate slow - want %d got %d \n",
3883 obd->obd_name, *num, i);
3890 "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3891 obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3896 static int filter_create(struct obd_export *exp, struct obdo *oa,
3897 struct lov_stripe_md **ea, struct obd_trans_info *oti)
3899 struct obd_device *obd = exp->exp_obd;
3900 struct filter_export_data *fed;
3901 struct filter_obd *filter;
3902 struct lvfs_run_ctxt saved;
3903 struct lov_stripe_md *lsm = NULL;
3904 int rc = 0, diff, group = oa->o_gr;
3907 CDEBUG(D_INODE, "%s: filter_create(od->o_gr="LPU64",od->o_id="
3908 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
3910 if (!(oa->o_valid & OBD_MD_FLGROUP)) {
3911 CERROR("!!! nid %s sent invalid object group %d\n",
3912 obd_export_nid2str(exp), group);
3916 fed = &exp->exp_filter_data;
3917 filter = &obd->u.filter;
3919 if (fed->fed_group != group) {
3920 CERROR("!!! this export (nid %s) used object group %d "
3921 "earlier; now it's trying to use group %d! This could "
3922 "be a bug in the MDS. Please report to "
3923 "http://bugzilla.lustre.org/\n",
3924 obd_export_nid2str(exp), fed->fed_group, group);
3931 rc = obd_alloc_memmd(exp, &lsm);
3938 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3940 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3941 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3942 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3943 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3944 oa->o_id, filter_last_id(filter,
3949 down(&filter->fo_create_locks[oa->o_gr]);
3950 rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3951 up(&filter->fo_create_locks[oa->o_gr]);
3954 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3957 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3958 if (rc && ea != NULL && *ea != lsm) {
3959 obd_free_memmd(exp, &lsm);
3960 } else if (rc == 0 && ea != NULL) {
3961 /* XXX LOV STACKING: the lsm that is passed to us from
3962 * LOV does not have valid lsm_oinfo data structs, so
3963 * don't go touching that. This needs to be fixed in a
3965 lsm->lsm_object_id = oa->o_id;
3972 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3973 struct lov_stripe_md *md, struct obd_trans_info *oti,
3974 struct obd_export *md_exp, void *capa)
3976 unsigned int qcids[MAXQUOTAS] = {0, 0};
3977 struct obd_device *obd;
3978 struct filter_obd *filter;
3979 struct dentry *dchild = NULL, *dparent = NULL;
3980 struct lvfs_run_ctxt saved;
3981 void *handle = NULL;
3982 struct llog_cookie *fcc = NULL;
3983 int rc, rc2, cleanup_phase = 0, sync = 0;
3987 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3989 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
3990 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
3995 filter = &obd->u.filter;
3997 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4000 CDEBUG(D_INODE, "%s: filter_destroy(od->o_gr="LPU64",od->o_id="
4001 LPU64")\n", obd->obd_name, oa->o_gr, oa->o_id);
4003 dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
4005 GOTO(cleanup, rc = PTR_ERR(dchild));
4008 if (dchild->d_inode == NULL) {
4009 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
4011 /* If object already gone, cancel cookie right now */
4012 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4013 struct llog_ctxt *ctxt;
4014 struct obd_llog_group *olg;
4015 fcc = &oa->o_lcookie;
4016 olg = filter_find_olg(obd, oa->o_gr);
4018 CERROR(" %s: can not find olg of group %d\n",
4019 obd->obd_name, (int)oa->o_gr);
4020 GOTO(cleanup, rc = PTR_ERR(olg));
4022 llog_group_set_export(olg, exp);
4024 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4025 llog_cancel(ctxt, NULL, 1, fcc, 0);
4026 llog_ctxt_put(ctxt);
4027 fcc = NULL; /* we didn't allocate fcc, don't free it */
4029 GOTO(cleanup, rc = -ENOENT);
4032 filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
4034 /* Our MDC connection is established by the MDS to us */
4035 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4036 OBD_ALLOC(fcc, sizeof(*fcc));
4038 *fcc = oa->o_lcookie;
4040 DQUOT_INIT(dchild->d_inode);
4042 /* we're gonna truncate it first in order to avoid possible deadlock:
4044 * open trasaction open transaction
4045 * down(i_zombie) down(i_zombie)
4046 * restart transaction
4047 * (see BUG 4180) -bzzz
4049 LOCK_INODE_MUTEX(dchild->d_inode);
4050 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4052 if (IS_ERR(handle)) {
4053 UNLOCK_INODE_MUTEX(dchild->d_inode);
4054 GOTO(cleanup, rc = PTR_ERR(handle));
4057 iattr.ia_valid = ATTR_SIZE;
4059 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4060 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4061 UNLOCK_INODE_MUTEX(dchild->d_inode);
4065 GOTO(cleanup, rc = rc2);
4067 /* We don't actually need to lock the parent until we are unlinking
4068 * here, and not while truncating above. That avoids holding the
4069 * parent lock for a long time during truncate, which can block other
4070 * threads from doing anything to objects in that directory. bug 7171 */
4071 dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
4072 if (IS_ERR(dparent))
4073 GOTO(cleanup, rc = PTR_ERR(dparent));
4074 cleanup_phase = 3; /* filter_parent_unlock */
4076 LOCK_INODE_MUTEX(dchild->d_inode);
4077 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4078 if (IS_ERR(handle)) {
4079 UNLOCK_INODE_MUTEX(dchild->d_inode);
4080 GOTO(cleanup, rc = PTR_ERR(handle));
4082 cleanup_phase = 4; /* fsfilt_commit */
4084 /* Quota release need uid/gid of inode */
4085 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
4087 filter_fmd_drop(exp, oa->o_id, oa->o_gr);
4089 /* this drops dchild->d_inode->i_mutex unconditionally */
4090 rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
4094 switch(cleanup_phase) {
4097 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4098 oti->oti_handle : handle,
4099 filter_cancel_cookies_cb,
4101 /* If add_journal_cb failed, then filter_finish_transno
4102 * will commit the handle and we will do a sync
4103 * on commit. then we call callback directly to free
4106 rc = filter_finish_transno(exp, oti, rc, sync);
4108 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4111 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4113 CERROR("error on commit, err = %d\n", rc2);
4120 filter_parent_unlock(dparent);
4124 OBD_FREE(fcc, sizeof(*fcc));
4126 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4129 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4133 /* trigger quota release */
4134 qcids[USRQUOTA] = oa->o_uid;
4135 qcids[GRPQUOTA] = oa->o_gid;
4136 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4139 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4143 /* NB start and end are used for punch, but not truncate */
4144 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4145 struct obd_trans_info *oti,
4146 struct ptlrpc_request_set *rqset)
4151 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4152 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4153 oinfo->oi_policy.l_extent.end);
4157 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4158 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
4159 oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
4161 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4162 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4163 rc = filter_setattr(exp, oinfo, oti);
4164 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4168 static int filter_sync(struct obd_export *exp, struct obdo *oa,
4169 struct lov_stripe_md *lsm, obd_off start, obd_off end,
4172 struct lvfs_run_ctxt saved;
4173 struct filter_obd *filter;
4174 struct dentry *dentry;
4178 rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
4179 (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
4183 filter = &exp->exp_obd->u.filter;
4185 /* An objid of zero is taken to mean "sync whole filesystem" */
4186 if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
4187 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
4188 /* Flush any remaining cancel messages out to the target */
4189 filter_sync_llogs(exp->exp_obd, exp);
4193 dentry = filter_oa2dentry(exp->exp_obd, oa);
4195 RETURN(PTR_ERR(dentry));
4197 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4199 LOCK_INODE_MUTEX(dentry->d_inode);
4201 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4203 /* just any file to grab fsync method - "file" arg unused */
4204 struct file *file = filter->fo_rcvd_filp;
4206 if (file->f_op && file->f_op->fsync)
4207 rc = file->f_op->fsync(NULL, dentry, 1);
4209 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4213 UNLOCK_INODE_MUTEX(dentry->d_inode);
4215 oa->o_valid = OBD_MD_FLID;
4216 obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
4218 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4224 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4225 void *key, __u32 *vallen, void *val,
4226 struct lov_stripe_md *lsm)
4228 struct obd_device *obd;
4231 obd = class_exp2obd(exp);
4233 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4237 if (KEY_IS(KEY_BLOCKSIZE)) {
4238 __u32 *blocksize = val;
4240 if (*vallen < sizeof(*blocksize))
4242 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4244 *vallen = sizeof(*blocksize);
4248 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4249 __u32 *blocksize_bits = val;
4250 if (blocksize_bits) {
4251 if (*vallen < sizeof(*blocksize_bits))
4253 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4255 *vallen = sizeof(*blocksize_bits);
4259 if (KEY_IS(KEY_LAST_ID)) {
4260 obd_id *last_id = val;
4261 /* FIXME: object groups */
4263 if (*vallen < sizeof(*last_id))
4265 *last_id = filter_last_id(&obd->u.filter,
4266 exp->exp_filter_data.fed_group);
4268 *vallen = sizeof(*last_id);
4272 if (KEY_IS(KEY_FIEMAP)) {
4273 struct ll_fiemap_info_key *fm_key = key;
4274 struct dentry *dentry;
4275 struct ll_user_fiemap *fiemap = val;
4276 struct lvfs_run_ctxt saved;
4279 if (fiemap == NULL) {
4280 *vallen = fiemap_count_to_size(
4281 fm_key->fiemap.fm_extent_count);
4285 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa,
4288 RETURN(PTR_ERR(dentry));
4290 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4291 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4292 rc = fsfilt_iocontrol(obd, dentry->d_inode, NULL,
4293 EXT3_IOC_FIEMAP, (long)fiemap);
4294 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4300 CDEBUG(D_IOCTL, "invalid key\n");
4304 static inline int filter_setup_llog_group(struct obd_export *exp,
4305 struct obd_device *obd,
4308 struct obd_llog_group *olg;
4309 struct llog_ctxt *ctxt;
4312 olg = filter_find_create_olg(obd, group);
4314 RETURN(PTR_ERR(olg));
4316 llog_group_set_export(olg, exp);
4318 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4319 LASSERTF(ctxt != NULL, "ctxt is null\n");
4321 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4322 llog_ctxt_put(ctxt);
4325 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4326 void *key, __u32 vallen, void *val,
4327 struct ptlrpc_request_set *set)
4329 struct obd_device *obd;
4335 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4339 if (KEY_IS(KEY_CAPA_KEY)) {
4340 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4342 CERROR("filter update capability key failed: %d\n", rc);
4346 if (KEY_IS(KEY_REVIMP_UPD)) {
4347 filter_revimp_update(exp);
4348 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4352 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4353 filter_adapt_sptlrpc_conf(obd, 0);
4357 if (KEY_IS(KEY_GRANT_SHRINK)) {
4358 struct ost_body *body = (struct ost_body *)val;
4359 /* handle shrink grant */
4360 spin_lock(&exp->exp_obd->obd_osfs_lock);
4361 filter_grant_incoming(exp, &body->oa);
4362 spin_unlock(&exp->exp_obd->obd_osfs_lock);
4366 if (!KEY_IS(KEY_MDS_CONN))
4369 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4370 obd_export_nid2str(exp));
4371 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4373 /* setup llog imports */
4375 group = (int)(*(__u32 *)val);
4377 group = 0; /* default value */
4379 LASSERT_MDS_GROUP(group);
4380 rc = filter_setup_llog_group(exp, obd, group);
4384 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4386 if (group == FILTER_GROUP_MDS0) {
4387 /* setup llog group 1 for interop */
4388 filter_setup_llog_group(exp, obd, FILTER_GROUP_LLOG);
4394 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4395 int len, void *karg, void *uarg)
4397 struct obd_device *obd = exp->exp_obd;
4398 struct obd_ioctl_data *data = karg;
4402 case OBD_IOC_ABORT_RECOVERY: {
4403 CERROR("aborting recovery for device %s\n", obd->obd_name);
4404 target_stop_recovery_thread(obd);
4408 case OBD_IOC_SYNC: {
4409 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4410 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4414 case OBD_IOC_SET_READONLY: {
4416 struct super_block *sb = obd->u.obt.obt_sb;
4417 struct inode *inode = sb->s_root->d_inode;
4418 BDEVNAME_DECLARE_STORAGE(tmp);
4419 CERROR("*** setting device %s read-only ***\n",
4420 ll_bdevname(sb, tmp));
4422 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4423 if (!IS_ERR(handle))
4424 rc = fsfilt_commit(obd, inode, handle, 1);
4426 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4427 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4429 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4433 case OBD_IOC_CATLOGLIST: {
4434 rc = llog_catalog_list(obd, 1, data);
4438 case OBD_IOC_LLOG_CANCEL:
4439 case OBD_IOC_LLOG_REMOVE:
4440 case OBD_IOC_LLOG_INFO:
4441 case OBD_IOC_LLOG_PRINT: {
4442 /* FIXME to be finished */
4443 RETURN(-EOPNOTSUPP);
4445 struct llog_ctxt *ctxt = NULL;
4447 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4448 rc = llog_ioctl(ctxt, cmd, data);
4449 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4462 static int filter_health_check(struct obd_device *obd)
4464 #ifdef USE_HEALTH_CHECK_WRITE
4465 struct filter_obd *filter = &obd->u.filter;
4470 * health_check to return 0 on healthy
4471 * and 1 on unhealthy.
4473 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4476 #ifdef USE_HEALTH_CHECK_WRITE
4477 LASSERT(filter->fo_health_check_filp != NULL);
4478 rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4483 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4486 return filter_fid2dentry(data, NULL, gr, id);
4489 static int filter_process_config(struct obd_device *obd, obd_count len,
4492 struct lustre_cfg *lcfg = buf;
4493 struct lprocfs_static_vars lvars;
4496 switch (lcfg->lcfg_command) {
4498 lprocfs_filter_init_vars(&lvars);
4500 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4510 static struct lvfs_callback_ops filter_lvfs_ops = {
4511 l_fid2dentry: filter_lvfs_fid2dentry,
4514 static struct obd_ops filter_obd_ops = {
4515 .o_owner = THIS_MODULE,
4516 .o_get_info = filter_get_info,
4517 .o_set_info_async = filter_set_info_async,
4518 .o_setup = filter_setup,
4519 .o_precleanup = filter_precleanup,
4520 .o_cleanup = filter_cleanup,
4521 .o_connect = filter_connect,
4522 .o_reconnect = filter_reconnect,
4523 .o_disconnect = filter_disconnect,
4524 .o_ping = filter_ping,
4525 .o_init_export = filter_init_export,
4526 .o_destroy_export = filter_destroy_export,
4527 .o_statfs = filter_statfs,
4528 .o_getattr = filter_getattr,
4529 .o_unpackmd = filter_unpackmd,
4530 .o_create = filter_create,
4531 .o_setattr = filter_setattr,
4532 .o_destroy = filter_destroy,
4533 .o_brw = filter_brw,
4534 .o_punch = filter_truncate,
4535 .o_sync = filter_sync,
4536 .o_preprw = filter_preprw,
4537 .o_commitrw = filter_commitrw,
4538 .o_llog_init = filter_llog_init,
4539 .o_llog_connect = filter_llog_connect,
4540 .o_llog_finish = filter_llog_finish,
4541 .o_iocontrol = filter_iocontrol,
4542 .o_health_check = filter_health_check,
4543 .o_process_config = filter_process_config,
4546 quota_interface_t *filter_quota_interface_ref;
4547 extern quota_interface_t filter_quota_interface;
4549 static int __init obdfilter_init(void)
4551 struct lprocfs_static_vars lvars;
4554 lprocfs_filter_init_vars(&lvars);
4556 request_module("lquota");
4557 OBD_ALLOC(obdfilter_created_scratchpad,
4558 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4559 sizeof(*obdfilter_created_scratchpad));
4560 if (obdfilter_created_scratchpad == NULL)
4563 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4564 sizeof(struct filter_mod_data),
4567 GOTO(out, rc = -ENOMEM);
4569 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4570 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4572 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4573 LUSTRE_OST_NAME, NULL);
4577 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4578 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4579 ll_fmd_cachep = NULL;
4581 if (filter_quota_interface_ref)
4582 PORTAL_SYMBOL_PUT(filter_quota_interface);
4584 OBD_FREE(obdfilter_created_scratchpad,
4585 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4586 sizeof(*obdfilter_created_scratchpad));
4592 static void __exit obdfilter_exit(void)
4594 if (filter_quota_interface_ref)
4595 PORTAL_SYMBOL_PUT(filter_quota_interface);
4597 if (ll_fmd_cachep) {
4598 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4599 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4600 ll_fmd_cachep = NULL;
4603 class_unregister_type(LUSTRE_OST_NAME);
4604 OBD_FREE(obdfilter_created_scratchpad,
4605 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4606 sizeof(*obdfilter_created_scratchpad));
4609 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4610 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4611 MODULE_LICENSE("GPL");
4613 module_init(obdfilter_init);
4614 module_exit(obdfilter_exit);