4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
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 #include <linux/module.h>
56 #include <linux/dcache.h>
57 #include <linux/init.h>
58 #include <linux/version.h>
59 #include <linux/sched.h>
60 #include <linux/mount.h>
61 #include <linux/buffer_head.h>
63 #include <obd_cksum.h>
64 #include <obd_class.h>
66 #include <lustre_dlm.h>
67 #include <lustre_fsfilt.h>
68 #include <lprocfs_status.h>
69 #include <lustre_log.h>
70 #include <libcfs/list.h>
71 #include <lustre_disk.h>
72 #include <lustre_quota.h>
73 #include <linux/slab.h>
74 #include <lustre_param.h>
75 #include <lustre/ll_fiemap.h>
77 #include "filter_internal.h"
79 static struct lvfs_callback_ops filter_lvfs_ops;
80 cfs_mem_cache_t *ll_fmd_cachep;
82 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
83 void *cb_data, int error)
85 struct obd_export *exp = cb_data;
86 LASSERT(exp->exp_obd == obd);
87 obd_transno_commit_cb(obd, transno, exp, error);
88 class_export_cb_put(exp);
91 int filter_version_get_check(struct obd_export *exp,
92 struct obd_trans_info *oti, struct inode *inode)
96 if (inode == NULL || oti == NULL)
99 curr_version = fsfilt_get_version(exp->exp_obd, inode);
100 if ((__s64)curr_version == -EOPNOTSUPP)
102 /* VBR: version is checked always because costs nothing */
103 if (oti->oti_pre_version != 0 &&
104 oti->oti_pre_version != curr_version) {
105 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
106 oti->oti_pre_version, curr_version);
107 cfs_spin_lock(&exp->exp_lock);
108 exp->exp_vbr_failed = 1;
109 cfs_spin_unlock(&exp->exp_lock);
112 oti->oti_pre_version = curr_version;
116 /* Assumes caller has already pushed us into the kernel context. */
117 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
118 struct obd_trans_info *oti, int rc, int force_sync)
120 struct obd_device_target *obt = &exp->exp_obd->u.obt;
121 struct tg_export_data *ted = &exp->exp_target_data;
122 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
123 struct lsd_client_data *lcd;
126 int err, log_pri = D_RPCTRACE;
128 /* Propagate error code. */
132 if (!exp->exp_obd->obd_replayable || oti == NULL)
135 cfs_mutex_lock(&ted->ted_lcd_lock);
137 /* if the export has already been disconnected, we have no last_rcvd slot,
138 * update server data with latest transno then */
140 cfs_mutex_unlock(&ted->ted_lcd_lock);
141 CWARN("commit transaction for disconnected client %s: rc %d\n",
142 exp->exp_client_uuid.uuid, rc);
143 err = filter_update_server_data(exp->exp_obd);
147 /* we don't allocate new transnos for replayed requests */
148 cfs_spin_lock(&obt->obt_lut->lut_translock);
149 if (oti->oti_transno == 0) {
150 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
151 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
152 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
154 last_rcvd = oti->oti_transno;
155 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
156 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
157 if (unlikely(last_rcvd < le64_to_cpu(lcd->lcd_last_transno))) {
158 CERROR("Trying to overwrite bigger transno, on-disk: "
159 LPU64", new: "LPU64"\n",
160 le64_to_cpu(lcd->lcd_last_transno), last_rcvd);
161 cfs_spin_lock(&exp->exp_lock);
162 exp->exp_vbr_failed = 1;
163 cfs_spin_unlock(&exp->exp_lock);
164 cfs_spin_unlock(&obt->obt_lut->lut_translock);
165 cfs_mutex_unlock(&ted->ted_lcd_lock);
169 oti->oti_transno = last_rcvd;
171 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
172 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
173 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
174 cfs_spin_unlock(&obt->obt_lut->lut_translock);
177 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
179 off = ted->ted_lr_off;
181 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
182 ted->ted_lr_idx, ted->ted_lr_off);
185 class_export_cb_get(exp); /* released when the cb is called */
187 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
193 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
194 lcd, sizeof(*lcd), &off,
195 force_sync | exp->exp_need_sync);
197 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
205 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
206 last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
207 cfs_mutex_unlock(&ted->ted_lcd_lock);
211 void f_dput(struct dentry *dentry)
213 /* Can't go inside filter_ddelete because it can block */
214 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
215 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
216 LASSERT(atomic_read(&dentry->d_count) > 0);
221 static void init_brw_stats(struct brw_stats *brw_stats)
224 for (i = 0; i < BRW_LAST; i++)
225 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
228 static int lprocfs_init_rw_stats(struct obd_device *obd,
229 struct lprocfs_stats **stats)
233 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
234 LPROC_FILTER_LAST - 1;
235 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
239 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
240 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
241 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
242 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
243 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
248 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
249 plus the procfs overhead :( */
250 static int filter_export_stats_init(struct obd_device *obd,
251 struct obd_export *exp,
257 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
258 /* Self-export gets no proc entry */
261 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
263 /* Mask error for already created
271 struct nid_stat *tmp = exp->exp_nid_stats;
272 LASSERT(tmp != NULL);
274 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
275 if (tmp->nid_brw_stats == NULL)
276 GOTO(clean, rc = -ENOMEM);
278 init_brw_stats(tmp->nid_brw_stats);
279 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
280 0644, &filter_per_nid_stats_fops,
283 CWARN("Error adding the brw_stats file\n");
285 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
289 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
293 rc = lprocfs_nid_ldlm_stats_init(tmp);
303 /* Add client data to the FILTER. We use a bitmap to locate a free space
304 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
305 * Otherwise, we have just read the data from the last_rcvd file and
306 * we know its offset. */
307 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
310 struct obd_device_target *obt = &obd->u.obt;
311 struct tg_export_data *ted = &exp->exp_target_data;
312 struct lr_server_data *lsd = class_server_data(obd);
313 unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
314 int new_client = (cl_idx == -1);
318 LASSERT(bitmap != NULL);
319 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
322 if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
325 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
326 * there's no need for extra complication here
329 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
331 if (cl_idx >= LR_MAX_CLIENTS) {
332 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
336 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
337 cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
342 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
343 CERROR("FILTER client %d: bit already set in bitmap!\n",
349 ted->ted_lr_idx = cl_idx;
350 ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
351 cl_idx * le16_to_cpu(lsd->lsd_client_size);
352 cfs_mutex_init(&ted->ted_lcd_lock);
353 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
355 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
356 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
359 struct lvfs_run_ctxt saved;
360 loff_t off = ted->ted_lr_off;
364 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
365 ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
367 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
370 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
371 /* Transaction needed to fix bug 1403 */
372 handle = fsfilt_start(obd,
373 obt->obt_rcvd_filp->f_dentry->d_inode,
374 FSFILT_OP_SETATTR, NULL);
375 if (IS_ERR(handle)) {
376 rc = PTR_ERR(handle);
377 CERROR("unable to start transaction: rc %d\n", rc);
379 ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
380 exp->exp_last_request_time = cfs_time_current_sec();
381 rc = fsfilt_add_journal_cb(obd, 0, handle,
382 target_client_add_cb,
383 class_export_cb_get(exp));
385 cfs_spin_lock(&exp->exp_lock);
386 exp->exp_need_sync = 1;
387 cfs_spin_unlock(&exp->exp_lock);
389 rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
391 sizeof(*ted->ted_lcd),
392 &off, rc /* sync if no cb */);
394 obt->obt_rcvd_filp->f_dentry->d_inode,
397 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
400 CERROR("error writing %s client idx %u: rc %d\n",
401 LAST_RCVD, ted->ted_lr_idx, rc);
408 static int filter_client_del(struct obd_export *exp)
410 struct tg_export_data *ted = &exp->exp_target_data;
411 struct obd_device_target *obt = &exp->exp_obd->u.obt;
412 struct lvfs_run_ctxt saved;
417 if (ted->ted_lcd == NULL)
420 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
421 if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
424 LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
426 off = ted->ted_lr_off;
428 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
429 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
431 /* Don't clear ted_lr_idx here as it is likely also unset. At worst
432 * we leak a client slot that will be cleaned on the next recovery. */
434 CERROR("%s: client idx %d has med_off %lld\n",
435 exp->exp_obd->obd_name, ted->ted_lr_idx, off);
436 GOTO(free, rc = -EINVAL);
439 /* Clear the bit _after_ zeroing out the client so we don't
440 race with filter_client_add and zero out new clients.*/
441 if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
442 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
447 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
448 /* Make sure the server's last_transno is up to date.
449 * This should be done before zeroing client slot so last_transno will
450 * be in server data or in client data in case of failure */
451 filter_update_server_data(exp->exp_obd);
453 cfs_mutex_lock(&ted->ted_lcd_lock);
454 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
455 rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
457 sizeof(*ted->ted_lcd), &off, 0);
458 cfs_mutex_unlock(&ted->ted_lcd_lock);
459 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
461 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
462 "zero out client %s at idx %u/%llu in %s, rc %d\n",
463 ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
470 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
471 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
472 struct filter_mod_data *fmd)
474 LASSERT_SPIN_LOCKED(&fed->fed_lock);
475 if (--fmd->fmd_refcount == 0) {
476 /* XXX when we have persistent reservations and the handle
477 * is stored herein we need to drop it here. */
478 fed->fed_mod_count--;
479 cfs_list_del(&fmd->fmd_list);
480 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
484 /* drop fmd reference, free it if last ref */
485 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
487 struct filter_export_data *fed;
492 fed = &exp->exp_filter_data;
493 cfs_spin_lock(&fed->fed_lock);
494 filter_fmd_put_nolock(fed, fmd); /* caller reference */
495 cfs_spin_unlock(&fed->fed_lock);
498 /* expire entries from the end of the list if there are too many
499 * or they are too old */
500 static void filter_fmd_expire_nolock(struct filter_obd *filter,
501 struct filter_export_data *fed,
502 struct filter_mod_data *keep)
504 struct filter_mod_data *fmd, *tmp;
506 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
510 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
511 fed->fed_mod_count < filter->fo_fmd_max_num)
514 cfs_list_del_init(&fmd->fmd_list);
515 filter_fmd_put_nolock(fed, fmd); /* list reference */
519 void filter_fmd_expire(struct obd_export *exp)
521 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
522 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
523 &exp->exp_filter_data, NULL);
524 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
527 /* find specified objid, group in export fmd list.
528 * caller must hold fed_lock and take fmd reference itself */
529 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
530 struct filter_export_data *fed,
531 obd_id objid, obd_seq group)
533 struct filter_mod_data *found = NULL, *fmd;
535 LASSERT_SPIN_LOCKED(&fed->fed_lock);
537 cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
538 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
540 cfs_list_del(&fmd->fmd_list);
541 cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
542 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
547 filter_fmd_expire_nolock(filter, fed, found);
552 /* Find fmd based on objid and group, or return NULL if not found. */
553 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
554 obd_id objid, obd_seq group)
556 struct filter_mod_data *fmd;
558 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
559 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
560 &exp->exp_filter_data, objid, group);
562 fmd->fmd_refcount++; /* caller reference */
563 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
568 /* Find fmd based on objid and group, or create a new one if none is found.
569 * It is possible for this function to return NULL under memory pressure,
570 * or if objid = 0 is passed (which will only cause old entries to expire).
571 * Currently this is not fatal because any fmd state is transient and
572 * may also be freed when it gets sufficiently old. */
573 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
574 obd_id objid, obd_seq group)
576 struct filter_export_data *fed = &exp->exp_filter_data;
577 struct filter_mod_data *found = NULL, *fmd_new = NULL;
579 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
581 cfs_spin_lock(&fed->fed_lock);
582 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
585 cfs_list_add_tail(&fmd_new->fmd_list,
587 fmd_new->fmd_id = objid;
588 fmd_new->fmd_gr = group;
589 fmd_new->fmd_refcount++; /* list reference */
591 fed->fed_mod_count++;
593 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
597 found->fmd_refcount++; /* caller reference */
598 found->fmd_expire = jiffies +
599 exp->exp_obd->u.filter.fo_fmd_max_age;
602 cfs_spin_unlock(&fed->fed_lock);
608 /* drop fmd list reference so it will disappear when last reference is put.
609 * This isn't so critical because it would in fact only affect the one client
610 * that is doing the unlink and at worst we have an stale entry referencing
611 * an object that should never be used again. */
612 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
614 struct filter_mod_data *found = NULL;
616 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
617 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
619 cfs_list_del_init(&found->fmd_list);
620 filter_fmd_put_nolock(&exp->exp_filter_data, found);
622 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
625 #define filter_fmd_drop(exp, objid, group)
628 /* remove all entries from fmd list */
629 static void filter_fmd_cleanup(struct obd_export *exp)
631 struct filter_export_data *fed = &exp->exp_filter_data;
632 struct filter_mod_data *fmd = NULL, *tmp;
634 cfs_spin_lock(&fed->fed_lock);
635 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
636 cfs_list_del_init(&fmd->fmd_list);
637 filter_fmd_put_nolock(fed, fmd);
639 cfs_spin_unlock(&fed->fed_lock);
642 static int filter_init_export(struct obd_export *exp)
647 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
648 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
650 cfs_spin_lock(&exp->exp_lock);
651 exp->exp_connecting = 1;
652 cfs_spin_unlock(&exp->exp_lock);
654 /* self-export doesn't need client data and ldlm initialization */
655 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
656 &exp->exp_client_uuid)))
659 rc = lut_client_alloc(exp);
661 rc = ldlm_init_export(exp);
663 CERROR("%s: Can't initialize export: rc %d\n",
664 exp->exp_obd->obd_name, rc);
669 static int filter_free_server_data(struct obd_device_target *obt)
671 lut_fini(NULL, obt->obt_lut);
672 OBD_FREE_PTR(obt->obt_lut);
676 /* assumes caller is already in kernel ctxt */
677 int filter_update_server_data(struct obd_device *obd)
679 struct file *filp = obd->u.obt.obt_rcvd_filp;
680 struct lr_server_data *lsd = class_server_data(obd);
685 CDEBUG(D_INODE, "server uuid : %s\n", lsd->lsd_uuid);
686 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
687 le64_to_cpu(lsd->lsd_last_transno));
688 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
689 le64_to_cpu(lsd->lsd_mount_count));
691 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
693 CERROR("error writing lr_server_data: rc = %d\n", rc);
698 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
701 struct filter_obd *filter = &obd->u.filter;
707 if (filter->fo_last_objid_files[group] == NULL) {
708 CERROR("Object seq "LPU64" not fully setup; not updating "
709 "last_objid\n", group);
713 CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
714 obd->obd_name, filter->fo_last_objids[group], group);
716 tmp = cpu_to_le64(filter->fo_last_objids[group]);
717 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
718 &tmp, sizeof(tmp), &off, force_sync);
720 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
724 extern int ost_handle(struct ptlrpc_request *req);
725 /* assumes caller has already in kernel ctxt */
726 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
728 struct filter_obd *filter = &obd->u.filter;
729 struct lr_server_data *lsd;
730 struct lsd_client_data *lcd = NULL;
731 struct inode *inode = filp->f_dentry->d_inode;
732 unsigned long last_rcvd_size = i_size_read(inode);
733 struct lu_target *lut;
740 /* ensure padding in the struct is the correct size */
741 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
742 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
743 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
744 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
746 /* allocate and initialize lu_target */
750 rc = lut_init(NULL, lut, obd, NULL);
753 lsd = class_server_data(obd);
754 if (last_rcvd_size == 0) {
755 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
757 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
758 lsd->lsd_last_transno = 0;
759 mount_count = lsd->lsd_mount_count = 0;
760 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
761 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
762 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
763 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
764 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
765 /* OBD_COMPAT_OST is set in filter_connect_internal when the
766 * MDS first connects and assigns the OST index number. */
767 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
770 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
772 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
776 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
777 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
778 "using the wrong disk %s. Were the "
779 "/dev/ assignments rearranged?\n",
780 obd->obd_uuid.uuid, lsd->lsd_uuid);
781 GOTO(err_lut, rc = -EINVAL);
783 mount_count = le64_to_cpu(lsd->lsd_mount_count);
784 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
786 /* Assume old last_rcvd format unless I_C_LR is set */
787 if (!(lsd->lsd_feature_incompat &
788 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
789 lsd->lsd_last_transno = lsd->lsd_compat14;
791 /* OBD_COMPAT_OST is set in filter_connect_internal when the
792 * MDS first connects and assigns the OST index number. */
793 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
797 if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
798 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
799 obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
800 ~FILTER_INCOMPAT_SUPP);
801 GOTO(err_lut, rc = -EINVAL);
803 if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
804 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
805 obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
806 ~FILTER_ROCOMPAT_SUPP);
807 /* Do something like remount filesystem read-only */
808 GOTO(err_lut, rc = -EINVAL);
811 start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
813 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
814 obd->obd_name, start_epoch);
815 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
816 obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
817 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
818 obd->obd_name, mount_count + 1);
819 CDEBUG(D_INODE, "%s: server data size: %u\n",
820 obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
821 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
822 obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
823 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
824 obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
825 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
826 obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
827 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
828 last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
829 (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
830 le16_to_cpu(lsd->lsd_client_size));
832 if (!obd->obd_replayable) {
833 CWARN("%s: recovery support OFF\n", obd->obd_name);
839 GOTO(err_client, rc = -ENOMEM);
841 for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
842 off < last_rcvd_size; cl_idx++) {
844 struct obd_export *exp;
845 struct filter_export_data *fed;
847 /* Don't assume off is incremented properly by
848 * fsfilt_read_record(), in case sizeof(*lcd)
849 * isn't the same as lsd->lsd_client_size. */
850 off = le32_to_cpu(lsd->lsd_client_start) +
851 cl_idx * le16_to_cpu(lsd->lsd_client_size);
852 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
854 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
855 LAST_RCVD, cl_idx, off, rc);
856 break; /* read error shouldn't cause startup to fail */
859 if (lcd->lcd_uuid[0] == '\0') {
860 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
865 check_lcd(obd->obd_name, cl_idx, lcd);
867 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
869 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
870 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
871 last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
873 /* These exports are cleaned up by filter_disconnect(), so they
874 * need to be set up like real exports as filter_connect() does.
876 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
878 if (PTR_ERR(exp) == -EALREADY) {
879 /* export already exists, zero out this one */
880 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
884 GOTO(err_client, rc = PTR_ERR(exp));
887 fed = &exp->exp_filter_data;
888 *fed->fed_ted.ted_lcd = *lcd;
889 fed->fed_group = 0; /* will be assigned at connect */
890 filter_export_stats_init(obd, exp, NULL);
891 rc = filter_client_add(obd, exp, cl_idx);
892 /* can't fail for existing client */
893 LASSERTF(rc == 0, "rc = %d\n", rc);
895 /* VBR: set export last committed */
896 exp->exp_last_committed = last_rcvd;
897 cfs_spin_lock(&exp->exp_lock);
898 exp->exp_connecting = 0;
899 exp->exp_in_recovery = 0;
900 cfs_spin_unlock(&exp->exp_lock);
901 obd->obd_max_recoverable_clients++;
902 class_export_put(exp);
904 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
905 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
909 obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
911 obd->u.obt.obt_mount_count = mount_count + 1;
912 obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
913 lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
915 /* save it, so mount count and last_transno is current */
916 rc = filter_update_server_data(obd);
918 GOTO(err_client, rc);
923 class_disconnect_exports(obd);
925 filter_free_server_data(&obd->u.obt);
929 static int filter_cleanup_groups(struct obd_device *obd)
931 struct filter_obd *filter = &obd->u.filter;
933 struct dentry *dentry;
937 if (filter->fo_dentry_O_groups != NULL) {
938 for (i = 0; i < filter->fo_group_count; i++) {
939 dentry = filter->fo_dentry_O_groups[i];
943 OBD_FREE(filter->fo_dentry_O_groups,
944 filter->fo_group_count *
945 sizeof(*filter->fo_dentry_O_groups));
946 filter->fo_dentry_O_groups = NULL;
948 if (filter->fo_last_objid_files != NULL) {
949 for (i = 0; i < filter->fo_group_count; i++) {
950 filp = filter->fo_last_objid_files[i];
954 OBD_FREE(filter->fo_last_objid_files,
955 filter->fo_group_count *
956 sizeof(*filter->fo_last_objid_files));
957 filter->fo_last_objid_files = NULL;
959 if (filter->fo_dentry_O_sub != NULL) {
960 for (i = 0; i < filter->fo_group_count; i++) {
961 for (j = 0; j < filter->fo_subdir_count; j++) {
962 dentry = filter->fo_dentry_O_sub[i].dentry[j];
967 OBD_FREE(filter->fo_dentry_O_sub,
968 filter->fo_group_count *
969 sizeof(*filter->fo_dentry_O_sub));
970 filter->fo_dentry_O_sub = NULL;
972 if (filter->fo_last_objids != NULL) {
973 OBD_FREE(filter->fo_last_objids,
974 filter->fo_group_count *
975 sizeof(*filter->fo_last_objids));
976 filter->fo_last_objids = NULL;
978 if (filter->fo_dentry_O != NULL) {
979 f_dput(filter->fo_dentry_O);
980 filter->fo_dentry_O = NULL;
985 static int filter_update_last_group(struct obd_device *obd, int group)
987 struct filter_obd *filter = &obd->u.filter;
988 struct file *filp = NULL;
989 int last_group = 0, rc;
993 if (group <= filter->fo_committed_group)
996 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1000 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1004 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1006 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1010 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1011 obd->obd_name, last_group, group);
1015 /* must be sync: bXXXX */
1016 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1018 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1022 filter->fo_committed_group = group;
1025 filp_close(filp, 0);
1029 static int filter_read_group_internal(struct obd_device *obd, int group,
1032 struct filter_obd *filter = &obd->u.filter;
1033 __u64 *new_objids = NULL;
1034 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1035 struct dentry **new_groups = NULL;
1036 struct file **new_files = NULL;
1037 struct dentry *dentry;
1039 int old_count = filter->fo_group_count, rc, stage = 0, i;
1043 int len = group + 1;
1045 snprintf(name, 24, "%d", group);
1049 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1051 if (IS_ERR(dentry)) {
1052 CERROR("Cannot lookup expected object group %d: %ld\n",
1053 group, PTR_ERR(dentry));
1054 RETURN(PTR_ERR(dentry));
1057 dentry = simple_mkdir(filter->fo_dentry_O,
1058 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1059 if (IS_ERR(dentry)) {
1060 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1062 RETURN(PTR_ERR(dentry));
1067 snprintf(name, 24, "O/%d/LAST_ID", group);
1069 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1071 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1072 GOTO(cleanup, rc = PTR_ERR(filp));
1076 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1078 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1082 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1083 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1084 if (tmp_subdirs == NULL)
1085 GOTO(cleanup, rc = -ENOMEM);
1088 for (i = 0; i < filter->fo_subdir_count; i++) {
1090 snprintf(dir, sizeof(dir), "d%u", i);
1092 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1093 obd->u.obt.obt_vfsmnt,
1095 if (IS_ERR(tmp_subdirs->dentry[i])) {
1096 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1097 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1102 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1103 tmp_subdirs->dentry[i]);
1107 /* 'group' is an index; we need an array of length 'group + 1' */
1108 if (group + 1 > old_count) {
1109 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1110 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1111 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1112 OBD_ALLOC(new_files, len * sizeof(*new_files));
1114 if (new_objids == NULL || new_subdirs == NULL ||
1115 new_groups == NULL || new_files == NULL)
1116 GOTO(cleanup, rc = -ENOMEM);
1119 memcpy(new_objids, filter->fo_last_objids,
1120 old_count * sizeof(*new_objids));
1121 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1122 old_count * sizeof(*new_subdirs));
1123 memcpy(new_groups, filter->fo_dentry_O_groups,
1124 old_count * sizeof(*new_groups));
1125 memcpy(new_files, filter->fo_last_objid_files,
1126 old_count * sizeof(*new_files));
1128 OBD_FREE(filter->fo_last_objids,
1129 old_count * sizeof(*new_objids));
1130 OBD_FREE(filter->fo_dentry_O_sub,
1131 old_count * sizeof(*new_subdirs));
1132 OBD_FREE(filter->fo_dentry_O_groups,
1133 old_count * sizeof(*new_groups));
1134 OBD_FREE(filter->fo_last_objid_files,
1135 old_count * sizeof(*new_files));
1137 filter->fo_last_objids = new_objids;
1138 filter->fo_dentry_O_sub = new_subdirs;
1139 filter->fo_dentry_O_groups = new_groups;
1140 filter->fo_last_objid_files = new_files;
1141 filter->fo_group_count = len;
1144 filter->fo_dentry_O_groups[group] = dentry;
1145 filter->fo_last_objid_files[group] = filp;
1146 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1147 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1148 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1151 filter_update_last_group(obd, group);
1153 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1154 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1155 rc = filter_update_last_objid(obd, group, 1);
1159 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1160 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1161 obd->obd_name, group, last_objid);
1166 if (new_objids != NULL)
1167 OBD_FREE(new_objids, len * sizeof(*new_objids));
1168 if (new_subdirs != NULL)
1169 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1170 if (new_groups != NULL)
1171 OBD_FREE(new_groups, len * sizeof(*new_groups));
1172 if (new_files != NULL)
1173 OBD_FREE(new_files, len * sizeof(*new_files));
1175 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1176 for (i = 0; i < filter->fo_subdir_count; i++) {
1177 if (tmp_subdirs->dentry[i] != NULL)
1178 dput(tmp_subdirs->dentry[i]);
1180 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1183 filp_close(filp, 0);
1190 static int filter_read_groups(struct obd_device *obd, int last_group,
1193 struct filter_obd *filter = &obd->u.filter;
1194 int old_count, group, rc = 0;
1196 cfs_mutex_lock(&filter->fo_init_lock);
1197 old_count = filter->fo_group_count;
1198 for (group = old_count; group <= last_group; group++) {
1199 rc = filter_read_group_internal(obd, group, create);
1203 cfs_mutex_unlock(&filter->fo_init_lock);
1207 /* FIXME: object groups */
1208 static int filter_prep_groups(struct obd_device *obd)
1210 struct filter_obd *filter = &obd->u.filter;
1211 struct dentry *O_dentry;
1213 int last_group, rc = 0, cleanup_phase = 0;
1217 O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1219 CDEBUG(D_INODE, "%s: got/created O: %p\n", obd->obd_name, O_dentry);
1220 if (IS_ERR(O_dentry)) {
1221 rc = PTR_ERR(O_dentry);
1222 CERROR("%s: cannot open/create O: rc = %d\n", obd->obd_name,rc);
1225 filter->fo_dentry_O = O_dentry;
1226 cleanup_phase = 1; /* O_dentry */
1228 /* we have to initialize all groups before first connections from
1229 * clients because they may send create/destroy for any group -bzzz */
1230 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1232 CERROR("%s: cannot create LAST_GROUP: rc = %ld\n",
1233 obd->obd_name, PTR_ERR(filp));
1234 GOTO(cleanup, rc = PTR_ERR(filp));
1236 cleanup_phase = 2; /* filp */
1238 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1240 CERROR("%s: error reading LAST_GROUP: rc %d\n",
1246 last_group = FID_SEQ_OST_MDT0;
1248 CDEBUG(D_INODE, "%s: initialize group %u (max %u)\n", obd->obd_name,
1249 FID_SEQ_OST_MDT0, last_group);
1250 filter->fo_committed_group = last_group;
1251 rc = filter_read_groups(obd, last_group, 1);
1255 filp_close(filp, 0);
1259 switch (cleanup_phase) {
1261 filp_close(filp, 0);
1263 filter_cleanup_groups(obd);
1264 f_dput(filter->fo_dentry_O);
1265 filter->fo_dentry_O = NULL;
1273 /* setup the object store with correct subdirectories */
1274 static int filter_prep(struct obd_device *obd)
1276 struct lvfs_run_ctxt saved;
1277 struct filter_obd *filter = &obd->u.filter;
1279 struct inode *inode;
1283 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1284 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1285 if (!file || IS_ERR(file)) {
1287 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1291 obd->u.obt.obt_rcvd_filp = file;
1292 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1293 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1294 file->f_dentry->d_inode->i_mode);
1295 GOTO(err_filp, rc = -ENOENT);
1298 inode = file->f_dentry->d_parent->d_inode;
1299 /* We use i_op->unlink directly in filter_vfs_unlink() */
1300 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1301 CERROR("%s: filesystem does not support create/unlink ops\n",
1303 GOTO(err_filp, rc = -EOPNOTSUPP);
1306 rc = filter_init_server_data(obd, file);
1308 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1311 LASSERT(obd->u.obt.obt_lut);
1312 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1314 /* open and create health check io file*/
1315 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1318 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1320 GOTO(err_server_data, rc);
1322 filter->fo_obt.obt_health_check_filp = file;
1323 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1324 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1325 file->f_dentry->d_inode->i_mode);
1326 GOTO(err_health_check, rc = -ENOENT);
1328 rc = lvfs_check_io_health(obd, file);
1330 GOTO(err_health_check, rc);
1332 rc = filter_prep_groups(obd);
1334 GOTO(err_health_check, rc);
1336 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1341 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1342 CERROR("can't close %s after error\n", HEALTH_CHECK);
1343 filter->fo_obt.obt_health_check_filp = NULL;
1345 target_recovery_fini(obd);
1346 filter_free_server_data(&obd->u.obt);
1348 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1349 CERROR("can't close %s after error\n", LAST_RCVD);
1350 obd->u.obt.obt_rcvd_filp = NULL;
1354 /* cleanup the filter: write last used object id to status file */
1355 static void filter_post(struct obd_device *obd)
1357 struct lvfs_run_ctxt saved;
1358 struct filter_obd *filter = &obd->u.filter;
1361 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1362 * best to start a transaction with h_sync, because we removed this
1365 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1366 rc = filter_update_server_data(obd);
1368 CERROR("error writing server data: rc = %d\n", rc);
1370 for (i = 0; i < filter->fo_group_count; i++) {
1371 rc = filter_update_last_objid(obd, i,
1372 (i == filter->fo_group_count - 1));
1374 CERROR("error writing group %d lastobjid: rc = %d\n",
1378 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1379 obd->u.obt.obt_rcvd_filp = NULL;
1381 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1383 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1384 filter->fo_obt.obt_health_check_filp = NULL;
1386 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1388 filter_cleanup_groups(obd);
1389 filter_free_server_data(&obd->u.obt);
1390 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1392 filter_free_capa_keys(filter);
1393 cleanup_capa_hash(filter->fo_capa_hash);
1396 static void filter_set_last_id(struct filter_obd *filter,
1397 obd_id id, obd_seq group)
1399 LASSERT(group <= filter->fo_group_count);
1401 cfs_spin_lock(&filter->fo_objidlock);
1402 filter->fo_last_objids[group] = id;
1403 cfs_spin_unlock(&filter->fo_objidlock);
1406 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1409 LASSERT(group <= filter->fo_group_count);
1410 LASSERT(filter->fo_last_objids != NULL);
1412 /* FIXME: object groups */
1413 cfs_spin_lock(&filter->fo_objidlock);
1414 id = filter->fo_last_objids[group];
1415 cfs_spin_unlock(&filter->fo_objidlock);
1419 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1421 mutex_lock_nested(&dparent->d_inode->i_mutex, I_MUTEX_PARENT);
1425 /* We never dget the object parent, so DON'T dput it either */
1426 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1428 struct filter_obd *filter = &obd->u.filter;
1429 struct filter_subdirs *subdirs;
1431 if (group >= filter->fo_group_count) /* FIXME: object groups */
1432 return ERR_PTR(-EBADF);
1434 if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1435 return filter->fo_dentry_O_groups[group];
1437 subdirs = &filter->fo_dentry_O_sub[group];
1438 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1441 /* We never dget the object parent, so DON'T dput it either */
1442 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1445 unsigned long now = jiffies;
1446 struct dentry *dparent = filter_parent(obd, group, objid);
1449 if (IS_ERR(dparent))
1451 if (dparent == NULL)
1452 return ERR_PTR(-ENOENT);
1454 rc = filter_lock_dentry(obd, dparent);
1455 fsfilt_check_slow(obd, now, "parent lock");
1456 return rc ? ERR_PTR(rc) : dparent;
1459 /* We never dget the object parent, so DON'T dput it either */
1460 static void filter_parent_unlock(struct dentry *dparent)
1462 mutex_unlock(&dparent->d_inode->i_mutex);
1465 /* How to get files, dentries, inodes from object id's.
1467 * If dir_dentry is passed, the caller has already locked the parent
1468 * appropriately for this operation (normally a write lock). If
1469 * dir_dentry is NULL, we do a read lock while we do the lookup to
1470 * avoid races with create/destroy and such changing the directory
1471 * internal to the filesystem code. */
1472 struct dentry *filter_fid2dentry(struct obd_device *obd,
1473 struct dentry *dir_dentry,
1474 obd_seq group, obd_id id)
1476 struct dentry *dparent = dir_dentry;
1477 struct dentry *dchild;
1482 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1483 obd->u.filter.fo_destroys_in_progress == 0) {
1484 /* don't fail lookups for orphan recovery, it causes
1485 * later LBUGs when objects still exist during precreate */
1486 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1487 RETURN(ERR_PTR(-ENOENT));
1490 CERROR("fatal: invalid object id 0\n");
1491 RETURN(ERR_PTR(-ESTALE));
1494 len = sprintf(name, LPU64, id);
1495 if (dir_dentry == NULL) {
1496 dparent = filter_parent_lock(obd, group, id);
1497 if (IS_ERR(dparent)) {
1498 CERROR("%s: error getting object "POSTID
1499 " parent: rc %ld\n", obd->obd_name,
1500 id, group, PTR_ERR(dparent));
1504 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1505 dparent->d_name.len, dparent->d_name.name, name);
1506 /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1507 dchild = lookup_one_len(name, dparent, len);
1508 if (dir_dentry == NULL)
1509 filter_parent_unlock(dparent);
1510 if (IS_ERR(dchild)) {
1511 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1512 obd->obd_name, id, group, PTR_ERR(dchild));
1516 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1517 CERROR("%s: got bad object "LPU64" inode %lu\n",
1518 obd->obd_name, id, dchild->d_inode->i_ino);
1520 RETURN(ERR_PTR(-ENOENT));
1523 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1524 name, dchild, atomic_read(&dchild->d_count));
1526 LASSERT(atomic_read(&dchild->d_count) > 0);
1531 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1532 obd_id group, struct lustre_handle *lockh)
1534 int flags = LDLM_AST_DISCARD_DATA, rc;
1535 struct ldlm_res_id res_id;
1536 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1539 osc_build_res_name(objid, group, &res_id);
1540 /* Tell the clients that the object is gone now and that they should
1541 * throw away any cached pages. */
1542 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1543 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1544 ldlm_completion_ast, NULL, NULL, 0, NULL,
1551 static void filter_fini_destroy(struct obd_device *obd,
1552 struct lustre_handle *lockh)
1554 if (lustre_handle_is_used(lockh))
1555 ldlm_lock_decref(lockh, LCK_PW);
1558 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1559 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1560 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1561 * reverse. Caller must take i_sem before starting the transaction and we
1562 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1563 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1564 struct vfsmount *mnt)
1569 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1570 * itself we already hold dir->i_mutex for child create/unlink ops */
1571 LASSERT(dentry->d_inode != NULL);
1572 LASSERT(mutex_trylock(&dir->i_mutex) == 0);
1573 LASSERT(mutex_trylock(&dentry->d_inode->i_mutex) == 0);
1577 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1578 GOTO(out, rc = -ENOENT);
1580 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1585 GOTO(out, rc = -EPERM);
1587 /* check_sticky() */
1588 if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1589 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1590 IS_IMMUTABLE(dentry->d_inode))
1591 GOTO(out, rc = -EPERM);
1593 /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
1594 ll_vfs_dq_init(dir);
1596 rc = ll_security_inode_unlink(dir, dentry, mnt);
1600 rc = dir->i_op->unlink(dir, dentry);
1602 /* need to drop i_mutex before we lose inode reference */
1603 mutex_unlock(&dentry->d_inode->i_mutex);
1610 /* Caller must hold LCK_PW on parent and push us into kernel context.
1611 * Caller must hold child i_mutex, we drop it always.
1612 * Caller is also required to ensure that dchild->d_inode exists. */
1613 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1614 obd_seq group, struct dentry *dparent,
1615 struct dentry *dchild)
1617 struct inode *inode = dchild->d_inode;
1620 /* There should be 2 references to the inode:
1621 * 1) taken by filter_prepare_destroy
1622 * 2) taken by filter_destroy */
1623 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1624 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1625 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1626 (unsigned long)inode->i_nlink,
1627 atomic_read(&inode->i_count));
1630 rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1632 CERROR("error unlinking objid %.*s: rc %d\n",
1633 dchild->d_name.len, dchild->d_name.name, rc);
1637 struct filter_intent_args {
1638 struct ldlm_lock **victim;
1643 static enum interval_iter filter_intent_cb(struct interval_node *n,
1646 struct ldlm_interval *node = (struct ldlm_interval *)n;
1647 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1648 __u64 size = arg->size;
1649 struct ldlm_lock **v = arg->victim;
1650 struct ldlm_lock *lck;
1652 /* If the interval is lower than the current file size,
1654 if (interval_high(n) <= size)
1655 return INTERVAL_ITER_STOP;
1657 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1658 /* Don't send glimpse ASTs to liblustre clients.
1659 * They aren't listening for them, and they do
1660 * entirely synchronous I/O anyways. */
1661 if (lck->l_export == NULL ||
1662 lck->l_export->exp_libclient == 1)
1665 if (*arg->liblustre)
1666 *arg->liblustre = 0;
1669 *v = LDLM_LOCK_GET(lck);
1670 } else if ((*v)->l_policy_data.l_extent.start <
1671 lck->l_policy_data.l_extent.start) {
1672 LDLM_LOCK_RELEASE(*v);
1673 *v = LDLM_LOCK_GET(lck);
1676 /* the same policy group - every lock has the
1677 * same extent, so needn't do it any more */
1681 return INTERVAL_ITER_CONT;
1684 static int filter_intent_policy(struct ldlm_namespace *ns,
1685 struct ldlm_lock **lockp, void *req_cookie,
1686 ldlm_mode_t mode, int flags, void *data)
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) };
1701 struct ldlm_glimpse_work gl_work;
1702 CFS_LIST_HEAD(gl_list);
1705 policy = ldlm_get_processing_policy(res);
1706 LASSERT(policy != NULL);
1707 LASSERT(req != NULL);
1709 rc = lustre_pack_reply(req, 3, repsize, NULL);
1711 RETURN(req->rq_status = rc);
1713 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1714 LASSERT(rep != NULL);
1716 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1717 sizeof(*reply_lvb));
1718 LASSERT(reply_lvb != NULL);
1720 //fixup_handle_for_resent_req(req, lock, &lockh);
1722 /* Call the extent policy function to see if our request can be
1723 * granted, or is blocked.
1724 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1725 * lock, and should not be granted if the lock will be blocked.
1728 if (flags & LDLM_FL_BLOCK_NOWAIT) {
1729 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_AGL_DELAY, 5);
1731 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_AGL_NOLOCK))
1732 RETURN(ELDLM_LOCK_ABORTED);
1735 LASSERT(ns == ldlm_res_to_ns(res));
1737 rc = policy(lock, &tmpflags, 0, &err, NULL);
1738 check_res_locked(res);
1740 /* The lock met with no resistance; we're finished. */
1741 if (rc == LDLM_ITER_CONTINUE) {
1742 /* do not grant locks to the liblustre clients: they cannot
1743 * handle ASTs robustly. We need to do this while still
1744 * holding lr_lock to avoid the lock remaining on the res_link
1745 * list (and potentially being added to l_pending_list by an
1746 * AST) when we are going to drop this lock ASAP. */
1747 if (lock->l_export->exp_libclient ||
1748 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1749 ldlm_resource_unlink_lock(lock);
1750 err = ELDLM_LOCK_ABORTED;
1752 err = ELDLM_LOCK_REPLACED;
1756 } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1757 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1758 * callback for glimpse size. The real size user will trigger
1759 * the glimpse callback when necessary. */
1761 RETURN(ELDLM_LOCK_ABORTED);
1764 /* Do not grant any lock, but instead send GL callbacks. The extent
1765 * policy nicely created a list of all PW locks for us. We will choose
1766 * the highest of those which are larger than the size in the LVB, if
1767 * any, and perform a glimpse callback. */
1768 res_lvb = res->lr_lvb_data;
1769 LASSERT(res_lvb != NULL);
1770 *reply_lvb = *res_lvb;
1773 * lr_lock guarantees that no new locks are granted, and,
1774 * therefore, that res->lr_lvb_data cannot increase beyond the
1775 * end of already granted lock. As a result, it is safe to
1776 * check against "stale" reply_lvb->lvb_size value without
1777 * res->lr_lvb_mutex.
1779 arg.size = reply_lvb->lvb_size;
1781 arg.liblustre = &only_liblustre;
1782 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1783 tree = &res->lr_itree[idx];
1784 if (tree->lit_mode == LCK_PR)
1787 interval_iterate_reverse(tree->lit_root,
1788 filter_intent_cb, &arg);
1792 /* There were no PW locks beyond the size in the LVB; finished. */
1794 if (only_liblustre) {
1795 /* If we discovered a liblustre client with a PW lock,
1796 * however, the LVB may be out of date! The LVB is
1797 * updated only on glimpse (which we don't do for
1798 * liblustre clients) and cancel (which the client
1799 * obviously has not yet done). So if it has written
1800 * data but kept the lock, the LVB is stale and needs
1801 * to be updated from disk.
1803 * Of course, this will all disappear when we switch to
1804 * taking liblustre locks on the OST. */
1805 ldlm_res_lvbo_update(res, NULL, 1);
1807 RETURN(ELDLM_LOCK_ABORTED);
1811 * This check is for lock taken in filter_prepare_destroy() that does
1812 * not have l_glimpse_ast set. So the logic is: if there is a lock
1813 * with no l_glimpse_ast set, this object is being destroyed already.
1815 * Hence, if you are grabbing DLM locks on the server, always set
1816 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1818 if (l->l_glimpse_ast == NULL) {
1819 /* We are racing with unlink(); just return -ENOENT */
1820 rep->lock_policy_res1 = -ENOENT;
1824 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1826 /* Populate the gl_work structure.
1827 * Grab additional reference on the lock which will be released in
1828 * ldlm_work_gl_ast_lock() */
1829 gl_work.gl_lock = LDLM_LOCK_GET(l);
1830 /* The glimpse callback is sent to one single extent lock. As a result,
1831 * the gl_work list is just composed of one element */
1832 cfs_list_add_tail(&gl_work.gl_list, &gl_list);
1833 /* the ldlm_glimpse_work structure is allocated on the stack */
1834 gl_work.gl_flags = LDLM_GL_WORK_NOFREE;
1836 rc = ldlm_glimpse_locks(res, &gl_list); /* this will update the LVB */
1838 if (!cfs_list_empty(&gl_list))
1839 LDLM_LOCK_RELEASE(l);
1842 *reply_lvb = *res_lvb;
1846 LDLM_LOCK_RELEASE(l);
1848 RETURN(ELDLM_LOCK_ABORTED);
1852 * per-obd_device iobuf pool.
1854 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1855 * allocations in write-path has to be minimized (see bug 5137).
1857 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1858 * OST threads (see ost_thread_{init,done}()).
1860 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1861 * at the OST layer there are only (potentially) multiple obd_device of type
1862 * unknown at the time of OST thread creation.
1864 * We create a cfs_hash for struct filter_obd (->fo_iobuf_hash field) on
1865 * initializing, each OST thread will create it's own iobuf on the first
1866 * access and insert it into ->fo_iobuf_hash with thread ID as key,
1867 * so the iobuf can be found again by thread ID.
1871 * filter_iobuf_pool_init()
1873 * filter_iobuf_pool_done()
1875 * filter_iobuf_get()
1877 * operate on this array. They are "generic" in a sense that they don't depend
1878 * on actual type of iobuf's (the latter depending on Linux kernel version).
1882 * destroy pool created by filter_iobuf_pool_init
1884 static void filter_iobuf_pool_done(struct filter_obd *filter)
1888 if (filter->fo_iobuf_hash != NULL) {
1889 cfs_hash_putref(filter->fo_iobuf_hash);
1890 filter->fo_iobuf_hash = NULL;
1895 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1897 struct filter_obd *filter = &obd->u.filter;
1898 struct sptlrpc_rule_set tmp_rset;
1901 sptlrpc_rule_set_init(&tmp_rset);
1902 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1904 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1909 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1911 cfs_write_lock(&filter->fo_sptlrpc_lock);
1912 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1913 filter->fo_sptlrpc_rset = tmp_rset;
1914 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1920 filter_iobuf_hop_hash(cfs_hash_t *hs, const void *key, unsigned mask)
1922 __u64 val = *((__u64 *)key);
1924 return cfs_hash_long(val, hs->hs_cur_bits);
1928 filter_iobuf_hop_key(cfs_hlist_node_t *hnode)
1930 struct filter_iobuf *pool;
1932 pool = cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist);
1933 return &pool->dr_hkey;
1937 filter_iobuf_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
1939 struct filter_iobuf *pool;
1941 pool = cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist);
1942 return pool->dr_hkey == *((__u64 *)key);
1946 filter_iobuf_hop_object(cfs_hlist_node_t *hnode)
1948 return cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist);
1952 filter_iobuf_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1954 /* dummy, required by cfs_hash */
1958 filter_iobuf_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1960 /* dummy, required by cfs_hash */
1964 filter_iobuf_hop_exit(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1966 struct filter_iobuf *pool;
1968 pool = cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist);
1969 filter_free_iobuf(pool);
1972 static struct cfs_hash_ops filter_iobuf_hops = {
1973 .hs_hash = filter_iobuf_hop_hash,
1974 .hs_key = filter_iobuf_hop_key,
1975 .hs_keycmp = filter_iobuf_hop_keycmp,
1976 .hs_object = filter_iobuf_hop_object,
1977 .hs_get = filter_iobuf_hop_get,
1978 .hs_put_locked = filter_iobuf_hop_put_locked,
1979 .hs_exit = filter_iobuf_hop_exit
1982 #define FILTER_IOBUF_HASH_BITS 9
1983 #define FILTER_IOBUF_HBKT_BITS 4
1986 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1988 static int filter_iobuf_pool_init(struct filter_obd *filter)
1990 filter->fo_iobuf_hash = cfs_hash_create("filter_iobuf",
1991 FILTER_IOBUF_HASH_BITS,
1992 FILTER_IOBUF_HASH_BITS,
1993 FILTER_IOBUF_HBKT_BITS, 0,
1997 CFS_HASH_RW_BKTLOCK |
1998 CFS_HASH_NO_ITEMREF);
2000 return filter->fo_iobuf_hash != NULL ? 0 : -ENOMEM;
2003 /* Return iobuf allocated for @thread_id.
2004 * If we haven't allocated a pool entry for this thread before, do so now and
2005 * insert it into fo_iobuf_hash, otherwise we can find it from fo_iobuf_hash */
2006 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
2008 struct filter_iobuf *pool = NULL;
2013 thread_id = (oti && oti->oti_thread) ? oti->oti_thread->t_id : -1;
2014 if (thread_id >= 0) {
2015 struct ptlrpc_service_part *svcpt;
2017 svcpt = oti->oti_thread->t_svcpt;
2018 LASSERT(svcpt != NULL);
2020 key = (__u64)(svcpt->scp_cpt) << 32 | thread_id;
2021 pool = cfs_hash_lookup(filter->fo_iobuf_hash, &key);
2026 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE, PTLRPC_MAX_BRW_PAGES);
2030 if (thread_id >= 0) {
2031 pool->dr_hkey = key;
2032 rc = cfs_hash_add_unique(filter->fo_iobuf_hash,
2033 &key, &pool->dr_hlist);
2034 /* ptlrpc service thould guarantee thread ID is unique */
2035 LASSERT(rc != -EALREADY);
2041 /* mount the file system (secretly). lustre_cfg parameters are:
2044 * 3 = flags: failover=f, failout=n
2047 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
2050 struct filter_obd *filter = &obd->u.filter;
2051 struct vfsmount *mnt;
2052 struct file_system_type *type;
2053 struct lustre_mount_info *lmi;
2054 struct obd_uuid uuid;
2058 struct request_queue *q;
2062 if (lcfg->lcfg_bufcount < 3 ||
2063 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
2064 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
2067 lmi = server_get_mount(obd->obd_name);
2069 /* We already mounted in lustre_fill_super.
2070 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
2071 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
2073 obd->obd_fsops = fsfilt_get_ops(lsi->lsi_fstype);
2075 /* old path - used by lctl */
2076 CERROR("Using old MDS mount method\n");
2077 type = get_fs_type(lustre_cfg_string(lcfg, 2));
2079 CERROR("get_fs_type failed\n");
2082 mnt = vfs_kern_mount(type, MS_NOATIME|MS_NODIRATIME,
2083 lustre_cfg_string(lcfg, 1), option);
2084 cfs_module_put(type->owner);
2087 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2088 lustre_cfg_string(lcfg, 1), rc);
2092 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2094 if (IS_ERR(obd->obd_fsops))
2095 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2097 rc = filter_iobuf_pool_init(filter);
2101 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2102 CERROR("%s: Underlying device is marked as read-only. "
2103 "Setup failed\n", obd->obd_name);
2104 GOTO(err_ops, rc = -EROFS);
2107 /* failover is the default */
2108 obd->obd_replayable = 1;
2110 /* disable connection until configuration finishes */
2111 obd->obd_no_conn = 1;
2113 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2114 str = lustre_cfg_string(lcfg, 3);
2115 if (strchr(str, 'n')) {
2116 CWARN("%s: recovery disabled\n", obd->obd_name);
2117 obd->obd_replayable = 0;
2121 obd->u.obt.obt_magic = OBT_MAGIC;
2122 obd->u.obt.obt_vfsmnt = mnt;
2123 obd->u.obt.obt_sb = mnt->mnt_sb;
2124 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2125 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2127 rc = fsfilt_setup(obd, obd->u.obt.obt_sb);
2131 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2132 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2133 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2134 obd->obd_lvfs_ctxt.fs = get_ds();
2135 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2137 cfs_mutex_init(&filter->fo_init_lock);
2138 filter->fo_committed_group = 0;
2139 filter->fo_destroys_in_progress = 0;
2140 for (i = 0; i < 32; i++)
2141 cfs_mutex_init(&filter->fo_create_locks[i]);
2143 cfs_spin_lock_init(&filter->fo_objidlock);
2144 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2145 cfs_mutex_init(&filter->fo_alloc_lock);
2146 init_brw_stats(&filter->fo_filter_stats);
2147 cfs_spin_lock_init(&filter->fo_flags_lock);
2148 filter->fo_read_cache = 1; /* enable read-only cache by default */
2149 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2150 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2151 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2152 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2153 filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2154 filter_slc_set(filter); /* initialize sync on lock cancel */
2156 rc = filter_prep(obd);
2160 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2161 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2163 filter->fo_fl_oss_capa = 1;
2165 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2166 filter->fo_capa_hash = init_capa_hash();
2167 if (filter->fo_capa_hash == NULL)
2168 GOTO(err_post, rc = -ENOMEM);
2170 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2171 obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2172 LDLM_NAMESPACE_SERVER,
2173 LDLM_NAMESPACE_GREEDY,
2175 if (obd->obd_namespace == NULL)
2176 GOTO(err_post, rc = -ENOMEM);
2177 obd->obd_namespace->ns_lvbp = obd;
2178 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2179 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2181 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2182 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2184 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2186 CERROR("failed to setup llogging subsystems\n");
2190 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2191 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2192 /* do this after llog being initialized */
2193 filter_adapt_sptlrpc_conf(obd, 1);
2195 rc = lquota_setup(filter_quota_interface_ref, obd);
2199 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2200 if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2201 queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2202 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2203 "for larger I/O requests: max_sectors = %u "
2204 "could be up to max_hw_sectors=%u\n",
2205 obd->obd_name, mnt->mnt_sb->s_id,
2206 queue_max_sectors(q), queue_max_hw_sectors(q));
2208 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2209 if (uuid_ptr != NULL) {
2210 class_uuid_unparse(uuid_ptr, &uuid);
2216 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2217 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2218 obd->obd_name, label ?: str, lmi ? "on " : "",
2219 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2220 obd->obd_replayable ? "enabled" : "disabled");
2227 fsfilt_put_ops(obd->obd_fsops);
2228 filter_iobuf_pool_done(filter);
2230 server_put_mount(obd->obd_name, mnt);
2231 obd->u.obt.obt_sb = 0;
2235 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2237 struct lprocfs_static_vars lvars;
2238 cfs_proc_dir_entry_t *entry;
2244 CLASSERT(offsetof(struct obd_device, u.obt) ==
2245 offsetof(struct obd_device, u.filter.fo_obt));
2247 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2250 /* lprocfs must be setup before the filter so state can be safely added
2251 * to /proc incrementally as the filter is setup */
2252 lprocfs_filter_init_vars(&lvars);
2253 rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2255 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2260 rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2262 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2264 GOTO(obd_cleanup, rc);
2267 /* Init obdfilter private stats here */
2268 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2269 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2270 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2271 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2272 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2273 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2274 "get_page", "usec");
2275 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2276 LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2277 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2278 LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2279 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2280 LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2281 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2282 LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2284 rc = lproc_filter_attach_seqstat(obd);
2286 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2287 GOTO(free_obd_stats, rc);
2290 entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2291 if (IS_ERR(entry)) {
2292 rc = PTR_ERR(entry);
2293 CERROR("%s: error %d setting up lprocfs for %s\n",
2294 obd->obd_name, rc, "exports");
2295 GOTO(free_obd_stats, rc);
2297 obd->obd_proc_exports_entry = entry;
2299 entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2300 lprocfs_nid_stats_clear_read,
2301 lprocfs_nid_stats_clear_write, obd, NULL);
2302 if (IS_ERR(entry)) {
2303 rc = PTR_ERR(entry);
2304 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2306 GOTO(free_obd_stats, rc);
2309 rc = lprocfs_job_stats_init(obd, LPROC_FILTER_STATS_LAST,
2310 filter_stats_counter_init);
2312 GOTO(remove_entry_clear, rc);
2314 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2315 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2317 GOTO(job_stats_fini, rc = -ENOMEM);
2318 addr = (unsigned long)cfs_page_address(page);
2319 clear_page((void *)addr);
2320 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2321 LUSTRE_CFG_BUFLEN(lcfg, 4));
2322 rc = filter_common_setup(obd, lcfg, (void *)addr);
2323 OBD_PAGE_FREE(page);
2325 CERROR("%s: filter_common_setup failed: %d.\n",
2327 GOTO(job_stats_fini, rc);
2333 lprocfs_job_stats_fini(obd);
2335 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2337 lprocfs_free_obd_stats(obd);
2339 lprocfs_obd_cleanup(obd);
2343 static struct llog_operations filter_mds_ost_repl_logops;
2345 static struct llog_operations filter_size_orig_logops = {
2346 .lop_setup = llog_obd_origin_setup,
2347 .lop_cleanup = llog_obd_origin_cleanup,
2348 .lop_add = llog_obd_origin_add
2351 static int filter_olg_fini(struct obd_llog_group *olg)
2353 struct llog_ctxt *ctxt;
2354 int rc = 0, rc2 = 0;
2357 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2359 rc = llog_cleanup(ctxt);
2361 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2363 rc2 = llog_cleanup(ctxt);
2368 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2370 rc2 = llog_cleanup(ctxt);
2379 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2380 struct obd_device *tgt)
2385 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2386 &filter_mds_ost_repl_logops);
2390 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2391 &filter_size_orig_logops);
2397 filter_olg_fini(olg);
2402 * Init the default olg, which is embeded in the obd_device, for filter.
2405 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2406 struct obd_device *tgt)
2408 struct filter_obd *filter = &obd->u.filter;
2409 struct llog_ctxt *ctxt;
2413 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2414 if (!filter->fo_lcm)
2417 filter_mds_ost_repl_logops = llog_client_ops;
2418 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2419 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2420 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2422 rc = filter_olg_init(obd, olg, tgt);
2424 GOTO(cleanup_lcm, rc);
2426 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2429 GOTO(cleanup_olg, rc);
2431 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2433 CERROR("Can't get ctxt for %p:%x\n", olg,
2434 LLOG_MDS_OST_REPL_CTXT);
2435 GOTO(cleanup_olg, rc = -ENODEV);
2437 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2438 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2439 llog_ctxt_put(ctxt);
2443 filter_olg_fini(olg);
2445 llog_recov_thread_fini(filter->fo_lcm, 1);
2446 filter->fo_lcm = NULL;
2451 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2452 struct obd_device *tgt, int *index)
2454 struct filter_obd *filter = &obd->u.filter;
2455 struct llog_ctxt *ctxt;
2459 LASSERT(olg != NULL);
2460 if (olg == &obd->obd_olg)
2461 return filter_default_olg_init(obd, olg, tgt);
2463 LASSERT(filter->fo_lcm != NULL);
2464 rc = filter_olg_init(obd, olg, tgt);
2467 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2469 CERROR("Can't get ctxt for %p:%x\n", olg,
2470 LLOG_MDS_OST_REPL_CTXT);
2471 filter_olg_fini(olg);
2474 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2475 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2476 llog_ctxt_put(ctxt);
2480 static int filter_llog_finish(struct obd_device *obd, int count)
2482 struct filter_obd *filter = &obd->u.filter;
2483 struct llog_ctxt *ctxt;
2486 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2489 * Make sure that no cached llcds left in recov_thread.
2490 * We actually do sync in disconnect time, but disconnect
2491 * may not come being marked rq_no_resend = 1.
2493 llog_sync(ctxt, NULL, OBD_LLOG_FL_EXIT);
2496 * Balance class_import_get() in llog_receptor_accept().
2497 * This is safe to do, as llog is already synchronized
2498 * and its import may go.
2500 cfs_mutex_lock(&ctxt->loc_mutex);
2501 if (ctxt->loc_imp) {
2502 class_import_put(ctxt->loc_imp);
2503 ctxt->loc_imp = NULL;
2506 if (filter->fo_lcm) {
2507 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2508 filter->fo_lcm = NULL;
2511 cfs_mutex_unlock(&ctxt->loc_mutex);
2512 llog_ctxt_put(ctxt);
2515 RETURN(filter_olg_fini(&obd->obd_olg));
2519 * Find the group llog according to group index in the llog group list.
2521 static struct obd_llog_group *
2522 filter_find_olg_internal(struct filter_obd *filter, int group)
2524 struct obd_llog_group *olg;
2526 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2527 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2528 if (olg->olg_seq == group)
2535 * Find the group llog according to group index on the filter
2537 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2539 struct obd_llog_group *olg = NULL;
2540 struct filter_obd *filter;
2542 filter = &obd->u.filter;
2544 if (group == FID_SEQ_LLOG)
2545 RETURN(&obd->obd_olg);
2547 cfs_spin_lock(&filter->fo_llog_list_lock);
2548 olg = filter_find_olg_internal(filter, group);
2549 cfs_spin_unlock(&filter->fo_llog_list_lock);
2554 * Find the llog_group of the filter according to the group. If it can not
2555 * find, create the llog_group, which only happens when mds is being synced
2558 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2560 struct obd_llog_group *olg = NULL, *olg_new = NULL;
2561 struct filter_obd *filter;
2564 filter = &obd->u.filter;
2566 if (group == FID_SEQ_LLOG)
2567 RETURN(&obd->obd_olg);
2569 OBD_ALLOC_PTR(olg_new);
2570 if (olg_new == NULL)
2571 RETURN(ERR_PTR(-ENOMEM));
2573 cfs_spin_lock(&filter->fo_llog_list_lock);
2574 olg = filter_find_olg_internal(filter, group);
2576 if (olg->olg_initializing) {
2577 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2579 GOTO(out_unlock, olg);
2582 /* set as the newly allocated one */
2587 llog_group_init(olg, group);
2588 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2589 olg->olg_initializing = 1;
2590 cfs_spin_unlock(&filter->fo_llog_list_lock);
2592 rc = obd_llog_init(obd, olg, obd, NULL);
2594 cfs_spin_lock(&filter->fo_llog_list_lock);
2595 cfs_list_del(&olg->olg_list);
2596 cfs_spin_unlock(&filter->fo_llog_list_lock);
2598 GOTO(out, olg = ERR_PTR(-ENOMEM));
2600 cfs_spin_lock(&filter->fo_llog_list_lock);
2601 olg->olg_initializing = 0;
2602 cfs_spin_unlock(&filter->fo_llog_list_lock);
2603 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2604 obd->obd_name, group, olg);
2609 cfs_spin_unlock(&filter->fo_llog_list_lock);
2611 OBD_FREE_PTR(olg_new);
2615 static int filter_llog_connect(struct obd_export *exp,
2616 struct llogd_conn_body *body)
2618 struct obd_device *obd = exp->exp_obd;
2619 struct llog_ctxt *ctxt;
2620 struct obd_llog_group *olg;
2624 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2625 obd->obd_name, body->lgdc_logid.lgl_oid,
2626 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2628 olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2630 CERROR(" %s: can not find olg of group %d\n",
2631 obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2634 llog_group_set_export(olg, exp);
2636 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2637 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2638 body->lgdc_ctxt_idx);
2640 CDEBUG(D_HA, "%s: Recovery from log "LPX64"/"LPX64":%x\n",
2641 obd->obd_name, body->lgdc_logid.lgl_oid,
2642 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2644 cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2645 obd->u.filter.fo_mds_ost_sync = 1;
2646 cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2647 rc = llog_connect(ctxt, &body->lgdc_logid,
2648 &body->lgdc_gen, NULL);
2649 llog_ctxt_put(ctxt);
2651 CERROR("failed to connect rc %d idx %d\n", rc,
2652 body->lgdc_ctxt_idx);
2657 static int filter_llog_preclean(struct obd_device *obd)
2659 struct obd_llog_group *olg, *tmp;
2660 struct filter_obd *filter;
2661 cfs_list_t remove_list;
2665 rc = obd_llog_finish(obd, 0);
2667 CERROR("failed to cleanup llogging subsystem\n");
2669 filter = &obd->u.filter;
2670 CFS_INIT_LIST_HEAD(&remove_list);
2672 cfs_spin_lock(&filter->fo_llog_list_lock);
2673 while (!cfs_list_empty(&filter->fo_llog_list)) {
2674 olg = cfs_list_entry(filter->fo_llog_list.next,
2675 struct obd_llog_group, olg_list);
2676 cfs_list_del(&olg->olg_list);
2677 cfs_list_add(&olg->olg_list, &remove_list);
2679 cfs_spin_unlock(&filter->fo_llog_list_lock);
2681 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2682 cfs_list_del_init(&olg->olg_list);
2683 rc = filter_olg_fini(olg);
2685 CERROR("failed to cleanup llogging subsystem for %u\n",
2693 static int filter_precleanup(struct obd_device *obd,
2694 enum obd_cleanup_stage stage)
2700 case OBD_CLEANUP_EARLY:
2702 case OBD_CLEANUP_EXPORTS:
2703 /* Stop recovery before namespace cleanup. */
2704 target_recovery_fini(obd);
2706 obd_exports_barrier(obd);
2707 obd_zombie_barrier();
2709 rc = filter_llog_preclean(obd);
2710 lprocfs_job_stats_fini(obd);
2711 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2712 lprocfs_free_per_client_stats(obd);
2713 lprocfs_obd_cleanup(obd);
2714 lprocfs_free_obd_stats(obd);
2715 lquota_cleanup(filter_quota_interface_ref, obd);
2721 static int filter_cleanup(struct obd_device *obd)
2723 struct filter_obd *filter = &obd->u.filter;
2727 LCONSOLE_WARN("%s: shutting down for failover; client state "
2728 "will be preserved.\n", obd->obd_name);
2730 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2731 obd->obd_namespace = NULL;
2733 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2735 if (obd->u.obt.obt_sb == NULL)
2740 ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2741 shrink_dcache_sb(obd->u.obt.obt_sb);
2743 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2744 obd->u.obt.obt_sb = NULL;
2746 fsfilt_put_ops(obd->obd_fsops);
2748 filter_iobuf_pool_done(filter);
2750 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2755 static int filter_connect_internal(struct obd_export *exp,
2756 struct obd_connect_data *data,
2759 struct filter_export_data *fed = &exp->exp_filter_data;
2764 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2765 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2766 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2767 data->ocd_connect_flags, data->ocd_version,
2768 data->ocd_grant, data->ocd_index);
2770 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2771 CWARN("!!! This export (nid %s) used object group %d "
2772 "earlier; now it's trying to use group %d! This could "
2773 "be a bug in the MDS. Please report to "
2774 "http://bugs.whamcloud.com/\n",
2775 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2778 fed->fed_group = data->ocd_group;
2780 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2781 exp->exp_connect_flags = data->ocd_connect_flags;
2782 data->ocd_version = LUSTRE_VERSION_CODE;
2784 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2785 if (data->ocd_connect_flags & OBD_CONNECT_MDS)
2786 CDEBUG(D_HA, "%s: Received MDS connection for group %u\n",
2787 exp->exp_obd->obd_name, data->ocd_group);
2788 else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
2791 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2792 struct filter_obd *filter = &exp->exp_obd->u.filter;
2793 obd_size left, want;
2795 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2796 left = filter_grant_space_left(exp);
2797 want = data->ocd_grant;
2798 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2799 data->ocd_grant = fed->fed_grant;
2800 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2802 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2803 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2804 exp->exp_client_uuid.uuid, exp,
2805 data->ocd_grant, want, left);
2807 filter->fo_tot_granted_clients ++;
2810 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2811 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2812 int index = le32_to_cpu(lsd->lsd_ost_index);
2814 if (!(lsd->lsd_feature_compat &
2815 cpu_to_le32(OBD_COMPAT_OST))) {
2816 /* this will only happen on the first connect */
2817 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2818 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2819 /* sync is not needed here as filter_client_add will
2820 * set exp_need_sync flag */
2821 filter_update_server_data(exp->exp_obd);
2822 } else if (index != data->ocd_index) {
2823 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2824 " %u doesn't match actual OST index"
2825 " %u in last_rcvd file, bad "
2827 obd_export_nid2str(exp), index,
2831 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2832 * FIXME: We don't strictly need the COMPAT flag for that,
2833 * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2834 * FIXME: We needed it for the index, as index 0 is valid. */
2837 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2838 data->ocd_brw_size = 65536;
2839 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2840 data->ocd_brw_size = min(data->ocd_brw_size,
2841 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2842 if (data->ocd_brw_size == 0) {
2843 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2844 " ocd_version: %x ocd_grant: %d ocd_index: %u "
2845 "ocd_brw_size is unexpectedly zero, "
2846 "network data corruption?"
2847 "Refusing connection of this client\n",
2848 exp->exp_obd->obd_name,
2849 exp->exp_client_uuid.uuid,
2850 exp, data->ocd_connect_flags, data->ocd_version,
2851 data->ocd_grant, data->ocd_index);
2856 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2857 __u32 cksum_types = data->ocd_cksum_types;
2859 /* The client set in ocd_cksum_types the checksum types it
2860 * supports. We have to mask off the algorithms that we don't
2862 data->ocd_cksum_types &= cksum_types_supported_server();
2864 /* 1.6.4 clients are not supported any more */
2866 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2867 "%x\n", exp->exp_obd->obd_name,
2868 obd_export_nid2str(exp), cksum_types,
2869 data->ocd_cksum_types);
2871 /* This client does not support OBD_CONNECT_CKSUM
2872 * fall back to CRC32 */
2873 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2874 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2875 exp->exp_obd->obd_name,
2876 obd_export_nid2str(exp));
2879 if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2880 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2885 static int filter_reconnect(const struct lu_env *env,
2886 struct obd_export *exp, struct obd_device *obd,
2887 struct obd_uuid *cluuid,
2888 struct obd_connect_data *data,
2894 if (exp == NULL || obd == NULL || cluuid == NULL)
2897 rc = filter_connect_internal(exp, data, 1);
2899 filter_export_stats_init(obd, exp, localdata);
2904 static int filter_connect(const struct lu_env *env,
2905 struct obd_export **exp, struct obd_device *obd,
2906 struct obd_uuid *cluuid,
2907 struct obd_connect_data *data, void *localdata)
2909 struct lvfs_run_ctxt saved;
2910 struct lustre_handle conn = { 0 };
2911 struct obd_export *lexp;
2915 if (exp == NULL || obd == NULL || cluuid == NULL)
2918 rc = class_connect(&conn, obd, cluuid);
2921 lexp = class_conn2export(&conn);
2922 LASSERT(lexp != NULL);
2924 rc = filter_connect_internal(lexp, data, 0);
2928 filter_export_stats_init(obd, lexp, localdata);
2929 if (obd->obd_replayable) {
2930 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2932 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2933 rc = filter_client_add(obd, lexp, -1);
2938 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2939 rc = filter_read_groups(obd, data->ocd_group, 1);
2940 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2942 CERROR("can't read group %u\n", data->ocd_group);
2950 class_disconnect(lexp);
2959 /* Do extra sanity checks for grant accounting. We do this at connect,
2960 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2961 * always get rid of it or turn it off when we know accounting is good. */
2962 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2964 struct filter_export_data *fed;
2965 struct obd_export *exp;
2966 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2967 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2968 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2970 if (cfs_list_empty(&obd->obd_exports))
2973 /* We don't want to do this for large machines that do lots of
2974 mounts or unmounts. It burns... */
2975 if (obd->obd_num_exports > 100)
2978 cfs_spin_lock(&obd->obd_osfs_lock);
2979 cfs_spin_lock(&obd->obd_dev_lock);
2980 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2982 fed = &exp->exp_filter_data;
2983 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2986 if (maxsize > 0) { /* we may not have done a statfs yet */
2987 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2988 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2989 exp->exp_client_uuid.uuid, exp,
2990 fed->fed_grant, fed->fed_pending, maxsize);
2991 LASSERTF(fed->fed_dirty <= maxsize,
2992 "%s: cli %s/%p %ld > "LPU64"\n", func,
2993 exp->exp_client_uuid.uuid, exp,
2994 fed->fed_dirty, maxsize);
2997 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2998 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2999 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
3001 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
3002 obd->obd_name, exp->exp_client_uuid.uuid, exp,
3003 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
3004 tot_granted += fed->fed_grant + fed->fed_pending;
3005 tot_pending += fed->fed_pending;
3006 tot_dirty += fed->fed_dirty;
3008 fo_tot_granted = obd->u.filter.fo_tot_granted;
3009 fo_tot_pending = obd->u.filter.fo_tot_pending;
3010 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
3011 cfs_spin_unlock(&obd->obd_dev_lock);
3012 cfs_spin_unlock(&obd->obd_osfs_lock);
3014 /* Do these assertions outside the spinlocks so we don't kill system */
3015 if (tot_granted != fo_tot_granted)
3016 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
3017 func, tot_granted, fo_tot_granted);
3018 if (tot_pending != fo_tot_pending)
3019 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
3020 func, tot_pending, fo_tot_pending);
3021 if (tot_dirty != fo_tot_dirty)
3022 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
3023 func, tot_dirty, fo_tot_dirty);
3024 if (tot_pending > tot_granted)
3025 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
3026 func, tot_pending, tot_granted);
3027 if (tot_granted > maxsize)
3028 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
3029 func, tot_granted, maxsize);
3030 if (tot_dirty > maxsize)
3031 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
3032 func, tot_dirty, maxsize);
3035 /* Remove this client from the grant accounting totals. We also remove
3036 * the export from the obd device under the osfs and dev locks to ensure
3037 * that the filter_grant_sanity_check() calculations are always valid.
3038 * The client should do something similar when it invalidates its import. */
3039 static void filter_grant_discard(struct obd_export *exp)
3041 struct obd_device *obd = exp->exp_obd;
3042 struct filter_obd *filter = &obd->u.filter;
3043 struct filter_export_data *fed = &exp->exp_filter_data;
3045 cfs_spin_lock(&obd->obd_osfs_lock);
3046 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
3047 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
3048 obd->obd_name, filter->fo_tot_granted,
3049 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
3050 filter->fo_tot_granted -= fed->fed_grant;
3051 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
3052 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
3053 obd->obd_name, filter->fo_tot_pending,
3054 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
3055 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
3056 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
3057 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
3058 obd->obd_name, filter->fo_tot_dirty,
3059 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
3060 filter->fo_tot_dirty -= fed->fed_dirty;
3064 cfs_spin_unlock(&obd->obd_osfs_lock);
3067 static int filter_destroy_export(struct obd_export *exp)
3069 struct filter_export_data *fed = &exp->exp_filter_data;
3072 if (fed->fed_pending)
3073 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
3074 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
3075 exp, fed->fed_pending);
3077 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3079 target_destroy_export(exp);
3081 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
3082 &exp->exp_client_uuid)))
3085 ldlm_destroy_export(exp);
3086 lut_client_free(exp);
3088 if (!exp->exp_obd->obd_replayable)
3089 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3091 filter_grant_discard(exp);
3092 filter_fmd_cleanup(exp);
3094 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3095 struct filter_obd *filter = &exp->exp_obd->u.filter;
3096 if (filter->fo_tot_granted_clients > 0)
3097 filter->fo_tot_granted_clients --;
3100 if (!(exp->exp_flags & OBD_OPT_FORCE))
3101 filter_grant_sanity_check(exp->exp_obd, __func__);
3106 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3108 struct obd_llog_group *olg_min, *olg;
3109 struct filter_obd *filter;
3110 int worked = -1, group;
3111 struct llog_ctxt *ctxt;
3114 filter = &obd->u.filter;
3116 /* we can't sync log holding spinlock. also, we do not want to get
3117 * into livelock. so we do following: loop over MDS's exports in
3118 * group order and skip already synced llogs -bzzz */
3120 /* look for group with min. number, but > worked */
3123 cfs_spin_lock(&filter->fo_llog_list_lock);
3124 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3125 if (olg->olg_seq <= worked) {
3126 /* this group is already synced */
3129 if (group < olg->olg_seq) {
3130 /* we have group with smaller number to sync */
3133 /* store current minimal group */
3135 group = olg->olg_seq;
3137 cfs_spin_unlock(&filter->fo_llog_list_lock);
3139 if (olg_min == NULL)
3142 worked = olg_min->olg_seq;
3143 if (olg_min->olg_exp &&
3144 (dexp == olg_min->olg_exp || dexp == NULL)) {
3146 ctxt = llog_group_get_ctxt(olg_min,
3147 LLOG_MDS_OST_REPL_CTXT);
3149 err = llog_sync(ctxt, olg_min->olg_exp, 0);
3150 llog_ctxt_put(ctxt);
3152 CERROR("error flushing logs to MDS: "
3157 } while (olg_min != NULL);
3160 /* Also incredibly similar to mds_disconnect */
3161 static int filter_disconnect(struct obd_export *exp)
3163 struct obd_device *obd = exp->exp_obd;
3168 class_export_get(exp);
3170 if (!(exp->exp_flags & OBD_OPT_FORCE))
3171 filter_grant_sanity_check(obd, __func__);
3172 filter_grant_discard(exp);
3174 /* Flush any remaining cancel messages out to the target */
3175 filter_sync_llogs(obd, exp);
3177 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3179 rc = server_disconnect_export(exp);
3181 /* Do not erase record for recoverable client. */
3182 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3183 filter_client_del(exp);
3185 fsfilt_sync(obd, obd->u.obt.obt_sb);
3187 class_export_put(exp);
3191 /* reverse import is changed, sync all cancels */
3192 static void filter_revimp_update(struct obd_export *exp)
3197 class_export_get(exp);
3199 /* flush any remaining cancel messages out to the target */
3200 filter_sync_llogs(exp->exp_obd, exp);
3201 class_export_put(exp);
3205 static int filter_ping(const struct lu_env *env, struct obd_export *exp)
3207 filter_fmd_expire(exp);
3211 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3212 const char *what, int quiet)
3214 struct dentry *dchild = NULL;
3216 dchild = filter_fid2dentry(obd, NULL, ostid->oi_seq, ostid->oi_id);
3218 if (IS_ERR(dchild)) {
3219 CERROR("%s error looking up object: "POSTID"\n",
3220 what, ostid->oi_id, ostid->oi_seq);
3224 if (dchild->d_inode == NULL) {
3226 CERROR("%s: %s on non-existent object: "POSTID" \n",
3227 obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3229 RETURN(ERR_PTR(-ENOENT));
3232 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
3233 /* Try to correct for a bug in 2.1.0 (LU-221) that caused negative
3234 * timestamps to appear to be in the far future, due old timestamp
3235 * being stored on disk as an unsigned value. This fixes up any
3236 * bad values stored on disk before returning them to the client,
3237 * and ensures any timestamp updates are correct. LU-1042 */
3238 if (unlikely(LTIME_S(dchild->d_inode->i_atime) == LU221_BAD_TIME))
3239 LTIME_S(dchild->d_inode->i_atime) = 0;
3240 if (unlikely(LTIME_S(dchild->d_inode->i_mtime) == LU221_BAD_TIME))
3241 LTIME_S(dchild->d_inode->i_mtime) = 0;
3242 if (unlikely(LTIME_S(dchild->d_inode->i_ctime) == LU221_BAD_TIME))
3243 LTIME_S(dchild->d_inode->i_ctime) = 0;
3245 #warning "remove old LU-221/LU-1042 workaround code"
3251 static int filter_getattr(const struct lu_env *env, struct obd_export *exp,
3252 struct obd_info *oinfo)
3254 struct dentry *dentry = NULL;
3255 struct obd_device *obd;
3260 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3261 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3265 obd = class_exp2obd(exp);
3267 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3271 dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3273 RETURN(PTR_ERR(dentry));
3275 /* Limit the valid bits in the return data to what we actually use */
3276 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3277 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
3279 /* Store inode version in reply */
3280 curr_version = fsfilt_get_version(exp->exp_obd, dentry->d_inode);
3281 if ((__s64)curr_version != -EOPNOTSUPP) {
3282 oinfo->oi_oa->o_valid |= OBD_MD_FLDATAVERSION;
3283 oinfo->oi_oa->o_data_version = curr_version;
3290 /* this should be enabled/disabled in condition to enabled/disabled large
3291 * inodes (fast EAs) in backing store FS. */
3292 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3293 void *handle, struct obdo *oa)
3295 struct obd_device *obd = exp->exp_obd;
3299 if (oa->o_valid & OBD_MD_FLFID) {
3300 struct filter_fid ff;
3302 if (!(oa->o_valid & OBD_MD_FLGROUP))
3304 /* packing fid and converting it to LE for storing into EA.
3305 * Here ->o_stripe_idx should be filled by LOV and rest of
3306 * fields - by client. */
3307 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3308 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3309 /* XXX: we are ignoring o_parent_ver here, since this should
3310 * be the same for all objects in this fileset. */
3311 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3312 ff.ff_objid = cpu_to_le64(oa->o_id);
3313 ff.ff_seq = cpu_to_le64(oa->o_seq);
3315 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3316 LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3319 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3321 CERROR("store fid in object failed! rc: %d\n", rc);
3323 CDEBUG(D_HA, "OSS object without fid info!\n");
3329 /* this is called from filter_truncate() until we have filter_punch() */
3330 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3331 struct obdo *oa, struct obd_trans_info *oti)
3333 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3334 struct llog_cookie *fcc = NULL;
3335 struct filter_obd *filter;
3336 int rc, err, sync = 0;
3337 loff_t old_size = 0;
3338 unsigned int ia_valid;
3339 struct inode *inode;
3340 struct page *page = NULL;
3345 LASSERT(dentry != NULL);
3346 LASSERT(!IS_ERR(dentry));
3348 inode = dentry->d_inode;
3349 LASSERT(inode != NULL);
3351 filter = &exp->exp_obd->u.filter;
3352 iattr_from_obdo(&iattr, oa, oa->o_valid);
3353 ia_valid = iattr.ia_valid;
3355 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3356 OBD_ALLOC(fcc, sizeof(*fcc));
3358 *fcc = oa->o_lcookie;
3360 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3361 unsigned long now = jiffies;
3362 /* Filter truncates and writes are serialized by
3363 * i_alloc_sem, see the comment in
3364 * filter_preprw_write.*/
3365 if (ia_valid & ATTR_SIZE)
3366 down_write(&inode->i_alloc_sem);
3367 mutex_lock(&inode->i_mutex);
3368 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
3369 old_size = i_size_read(inode);
3372 /* VBR: version recovery check */
3373 rc = filter_version_get_check(exp, oti, inode);
3375 GOTO(out_unlock, rc);
3377 /* Let's pin the last page so that ldiskfs_truncate
3378 * should not start GFP_FS allocation. */
3379 if (ia_valid & ATTR_SIZE) {
3380 page = grab_cache_page(inode->i_mapping,
3381 iattr.ia_size >> PAGE_CACHE_SHIFT);
3383 GOTO(out_unlock, rc = -ENOMEM);
3388 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3389 * then we will accept the UID+GID sent by the client during write for
3390 * initializing the ownership of this inode. We only allow this to
3391 * happen once so clear these bits in setattr. In 2.6 kernels it is
3392 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3393 * the flags that are actually being set. */
3394 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3395 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3396 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3398 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3399 if (!(ia_valid & ATTR_MODE)) {
3400 iattr.ia_mode = inode->i_mode;
3401 iattr.ia_valid |= ATTR_MODE;
3403 iattr.ia_mode &= ~S_ISUID;
3405 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3406 if (!(iattr.ia_valid & ATTR_MODE)) {
3407 iattr.ia_mode = inode->i_mode;
3408 iattr.ia_valid |= ATTR_MODE;
3410 iattr.ia_mode &= ~S_ISGID;
3413 orig_ids[USRQUOTA] = inode->i_uid;
3414 orig_ids[GRPQUOTA] = inode->i_gid;
3415 handle = fsfilt_start_log(exp->exp_obd, inode,
3416 FSFILT_OP_SETATTR, oti, 1);
3418 GOTO(out_unlock, rc = PTR_ERR(handle));
3420 /* update inode EA only once when inode is suid bit marked. As
3421 * on 2.6.x UID and GID may be set separately, we check here
3422 * only one of them to avoid double setting. */
3423 if (inode->i_mode & S_ISUID)
3424 filter_update_fidea(exp, inode, handle, oa);
3426 handle = fsfilt_start(exp->exp_obd, inode,
3427 FSFILT_OP_SETATTR, oti);
3429 GOTO(out_unlock, rc = PTR_ERR(handle));
3432 /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
3433 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3434 ll_vfs_dq_init(inode);
3436 if (oa->o_valid & OBD_MD_FLFLAGS) {
3437 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3438 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3440 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3442 /* set cancel cookie callback function */
3443 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3444 filter_cancel_cookies_cb,
3448 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3449 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3451 /* The truncate might have used up our transaction credits. Make sure
3452 * we have two left for the last_rcvd and VBR inode version updates. */
3453 err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3455 /* Update inode version only if data has changed => size has changed */
3456 rc = filter_finish_transno(exp, ia_valid & ATTR_SIZE ? inode : NULL,
3460 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3464 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3466 CERROR("error on commit, err = %d\n", err);
3473 /* For a partial-page truncate flush the page to disk immediately
3474 * to avoid data corruption during direct disk write. b=17397 */
3475 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3476 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3477 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3487 page_cache_release(page);
3489 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3490 mutex_unlock(&inode->i_mutex);
3491 if (ia_valid & ATTR_SIZE)
3492 up_write(&inode->i_alloc_sem);
3494 OBD_FREE(fcc, sizeof(*fcc));
3496 /* trigger quota release */
3497 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3498 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3499 int rc2 = lquota_adjust(filter_quota_interface_ref,
3500 exp->exp_obd, cur_ids,
3501 orig_ids, rc, FSFILT_OP_SETATTR);
3502 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3503 "filter adjust qunit. (rc:%d)\n", rc2);
3508 /* this is called from filter_truncate() until we have filter_punch() */
3509 int filter_setattr(const struct lu_env *env, struct obd_export *exp,
3510 struct obd_info *oinfo, struct obd_trans_info *oti)
3512 struct obdo *oa = oinfo->oi_oa;
3513 struct lustre_capa *capa = oinfo_capa(oinfo);
3514 struct ldlm_res_id res_id;
3515 struct filter_mod_data *fmd;
3516 struct lvfs_run_ctxt saved;
3517 struct filter_obd *filter;
3518 struct ldlm_resource *res;
3519 struct dentry *dentry;
3520 __u64 opc = CAPA_OPC_META_WRITE;
3524 if (oinfo->oi_flags & OBD_FL_PUNCH)
3525 opc |= CAPA_OPC_OSS_TRUNC;
3527 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3531 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3532 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3537 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3538 /* This would be very bad - accidentally truncating a file when
3539 * changing the time or similar - bug 12203. */
3540 if (oa->o_valid & OBD_MD_FLSIZE &&
3541 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3542 static char mdsinum[48];
3544 if (oa->o_valid & OBD_MD_FLFID)
3545 snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3546 oa->o_parent_seq, oa->o_parent_oid,
3551 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3552 "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3553 oa->o_id, oa->o_seq, mdsinum);
3557 dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3559 RETURN(PTR_ERR(dentry));
3561 filter = &exp->exp_obd->u.filter;
3562 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3565 * We need to be atomic against a concurrent write
3566 * (which takes the semaphore for reading). fmd_mactime_xid
3567 * checks will have no effect if a write request with lower
3568 * xid starts just before a setattr and finishes later than
3569 * the setattr (see bug 21489, comment 27).
3572 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3573 unsigned long now = jiffies;
3574 down_write(&dentry->d_inode->i_alloc_sem);
3575 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem");
3576 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3577 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3578 fmd->fmd_mactime_xid = oti->oti_xid;
3579 filter_fmd_put(exp, fmd);
3580 up_write(&dentry->d_inode->i_alloc_sem);
3583 /* setting objects attributes (including owner/group) */
3584 rc = filter_setattr_internal(exp, dentry, oa, oti);
3586 GOTO(out_unlock, rc);
3588 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3589 &res_id, LDLM_EXTENT, 0);
3592 LDLM_RESOURCE_ADDREF(res);
3593 rc = ldlm_res_lvbo_update(res, NULL, 0);
3594 LDLM_RESOURCE_DELREF(res);
3595 ldlm_resource_putref(res);
3598 oa->o_valid = OBD_MD_FLID;
3600 /* Quota release need uid/gid info */
3601 obdo_from_inode(oa, dentry->d_inode,
3602 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3604 filter_counter_incr(exp, LPROC_FILTER_STATS_SETATTR,
3605 oti ? oti->oti_jobid : NULL, 1);
3609 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3613 /* XXX identical to osc_unpackmd */
3614 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3615 struct lov_mds_md *lmm, int lmm_bytes)
3621 if (lmm_bytes < sizeof (*lmm)) {
3622 CERROR("lov_mds_md too small: %d, need %d\n",
3623 lmm_bytes, (int)sizeof(*lmm));
3626 /* XXX LOV_MAGIC etc check? */
3628 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3629 CERROR("lov_mds_md: zero lmm_object_id\n");
3634 lsm_size = lov_stripe_md_size(1);
3638 if (*lsmp != NULL && lmm == NULL) {
3639 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3640 OBD_FREE(*lsmp, lsm_size);
3645 if (*lsmp == NULL) {
3646 OBD_ALLOC(*lsmp, lsm_size);
3650 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3651 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3652 OBD_FREE(*lsmp, lsm_size);
3655 loi_init((*lsmp)->lsm_oinfo[0]);
3659 /* XXX zero *lsmp? */
3660 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3661 LASSERT((*lsmp)->lsm_object_id);
3664 (*lsmp)->lsm_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
3669 /* caller must hold fo_create_locks[oa->o_seq] */
3670 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3671 struct filter_obd *filter)
3673 struct obdo doa = { 0 }; /* XXX obdo on stack */
3679 LASSERT_MUTEX_LOCKED(&filter->fo_create_locks[oa->o_seq]);
3681 memset(&doa, 0, sizeof(doa));
3683 doa.o_valid |= OBD_MD_FLGROUP;
3684 doa.o_seq = oa->o_seq;
3685 doa.o_mode = S_IFREG;
3687 if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3688 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3689 exp->exp_obd->obd_name, doa.o_seq);
3693 last = filter_last_id(filter, doa.o_seq);
3695 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3697 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3698 exp->exp_obd->obd_name, oa->o_id + 1, last,
3699 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3701 for (id = last; id > oa->o_id; id--) {
3703 rc = filter_destroy(NULL, exp, &doa, NULL, NULL, NULL, NULL);
3704 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3705 CEMERG("error destroying precreate objid "LPU64": %d\n",
3708 /* update last_id on disk periodically so that if we restart
3709 * we don't need to re-scan all of the just-deleted objects. */
3710 if ((id & 511) == 0 && !skip_orphan) {
3711 filter_set_last_id(filter, id - 1, doa.o_seq);
3712 filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3716 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3717 exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3720 filter_set_last_id(filter, id, doa.o_seq);
3721 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3724 * We have destroyed orphan objects, but don't want to reuse
3725 * them. Therefore we don't reset last_id to the last created
3726 * objects. Instead, we report back to the MDS the object id
3727 * of the last orphan, so that the MDS can restart allocating
3728 * objects from this id + 1 and thus skip the whole orphan
3734 cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3739 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3740 obd_seq group, int *num);
3741 /* returns a negative error or a nonnegative number of files to create */
3742 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3743 obd_seq group, struct obd_trans_info *oti)
3745 struct obd_device *obd = exp->exp_obd;
3746 struct filter_obd *filter = &obd->u.filter;
3750 /* delete orphans request */
3751 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3752 obd_id last = filter_last_id(filter, group);
3754 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3755 CERROR("%s: dropping old orphan cleanup request\n",
3759 /* This causes inflight precreates to abort and drop lock */
3760 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3761 cfs_mutex_lock(&filter->fo_create_locks[group]);
3762 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3763 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3764 exp->exp_obd->obd_name, group);
3765 cfs_mutex_unlock(&filter->fo_create_locks[group]);
3768 diff = oa->o_id - last;
3769 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3772 if (-diff > (OST_MAX_PRECREATE * 3) / 2) {
3773 CERROR("%s: ignoring bogus orphan destroy request: "
3774 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3776 /* FIXME: should reset precreate_next_id on MDS */
3777 GOTO(out, rc = -EINVAL);
3780 rc = filter_destroy_precreated(exp, oa, filter);
3782 CERROR("%s: unable to write lastobjid, but "
3783 "orphans were deleted\n", obd->obd_name);
3786 /* XXX: Used by MDS for the first time! */
3787 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3790 cfs_mutex_lock(&filter->fo_create_locks[group]);
3791 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3792 CERROR("%s: dropping old precreate request\n",
3796 /* only precreate if group == 0 and o_id is specfied */
3797 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3800 diff = oa->o_id - filter_last_id(filter, group);
3801 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3802 filter_last_id(filter, group), diff);
3805 * Check obd->obd_recovering to handle the race condition
3806 * while recreating missing precreated objects through
3807 * filter_preprw_write() and mds_lov_clear_orphans()
3810 LASSERTF(ergo(!obd->obd_recovering, diff >= 0),
3811 "%s: "LPU64" - "LPU64" = %d\n", obd->obd_name,
3812 oa->o_id, filter_last_id(filter, group), diff);
3816 oa->o_id = filter_last_id(&obd->u.filter, group);
3817 rc = filter_precreate(obd, oa, group, &diff);
3818 oa->o_id = filter_last_id(&obd->u.filter, group);
3820 oa->o_valid |= (OBD_MD_FLID | OBD_MD_FLGROUP);
3823 /* else diff == 0 */
3826 cfs_mutex_unlock(&filter->fo_create_locks[group]);
3830 static int filter_statfs(const struct lu_env *env, struct obd_export *exp,
3831 struct obd_statfs *osfs, __u64 max_age, __u32 flags)
3833 struct obd_device *obd = class_exp2obd(exp);
3834 struct filter_obd *filter = &obd->u.filter;
3835 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3836 struct lr_server_data *lsd = class_server_data(obd);
3840 /* at least try to account for cached pages. its still racey and
3841 * might be under-reporting if clients haven't announced their
3842 * caches with brw recently */
3843 cfs_spin_lock(&obd->obd_osfs_lock);
3844 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3845 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3846 cfs_spin_unlock(&obd->obd_osfs_lock);
3848 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3849 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3850 filter->fo_tot_dirty, filter->fo_tot_granted,
3851 filter->fo_tot_pending,
3852 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3854 filter_grant_sanity_check(obd, __func__);
3856 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3857 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3858 osfs->os_bsize - 1) >> blockbits));
3860 if (OBD_FAIL_CHECK_VALUE(OBD_FAIL_OST_ENOSPC,
3861 le32_to_cpu(lsd->lsd_ost_index)))
3862 osfs->os_bfree = osfs->os_bavail = 2;
3864 if (OBD_FAIL_CHECK_VALUE(OBD_FAIL_OST_ENOINO,
3865 le32_to_cpu(lsd->lsd_ost_index)))
3868 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3869 * stop creating files on MDS if OST is not good shape to create
3873 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3874 osfs->os_state = OS_STATE_READONLY;
3876 if (filter->fo_raid_degraded)
3877 osfs->os_state |= OS_STATE_DEGRADED;
3881 static int filter_use_existing_obj(struct obd_device *obd,
3882 struct dentry *dchild, void **handle,
3885 struct inode *inode = dchild->d_inode;
3889 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3892 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3893 if (IS_ERR(*handle))
3894 return PTR_ERR(*handle);
3896 iattr.ia_valid = ATTR_MODE;
3897 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3898 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3905 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3908 __u64 os_ffree = -1;
3910 cfs_spin_lock(&obd->obd_osfs_lock);
3911 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3913 os_ffree = obd->obd_osfs.os_ffree;
3914 cfs_spin_unlock(&obd->obd_osfs_lock);
3919 /* We rely on the fact that only one thread will be creating files in a given
3920 * group at a time, which is why we don't need an atomic filter_get_new_id.
3921 * Even if we had that atomic function, the following race would exist:
3923 * thread 1: gets id x from filter_next_id
3924 * thread 2: gets id (x + 1) from filter_next_id
3925 * thread 2: creates object (x + 1)
3926 * thread 1: tries to create object x, gets -ENOSPC
3928 * Caller must hold fo_create_locks[group]
3930 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3931 obd_seq group, int *num)
3933 struct dentry *dchild = NULL, *dparent = NULL;
3934 struct filter_obd *filter;
3935 struct obd_statfs *osfs;
3937 int err = 0, rc = 0, recreate_obj = 0, i;
3938 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3941 void *handle = NULL;
3944 filter = &obd->u.filter;
3946 LASSERT_MUTEX_LOCKED(&filter->fo_create_locks[group]);
3948 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3950 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3951 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3954 OBD_ALLOC(osfs, sizeof(*osfs));
3957 rc = filter_statfs(NULL, obd->obd_self_export, osfs,
3958 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
3960 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3961 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3962 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3963 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3965 if (oa->o_valid & OBD_MD_FLFLAGS)
3966 oa->o_flags |= OBD_FL_NOSPC_BLK;
3968 oa->o_valid |= OBD_MD_FLFLAGS;
3969 oa->o_flags = OBD_FL_NOSPC_BLK;
3974 OBD_FREE(osfs, sizeof(*osfs));
3979 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3980 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3982 for (i = 0; i < *num && err == 0; i++) {
3983 int cleanup_phase = 0;
3985 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3986 CWARN("%s: create aborted by destroy\n",
3995 last_id = filter_last_id(filter, group);
3996 if (next_id > last_id) {
3997 CERROR("Error: Trying to recreate obj greater"
3998 "than last id "LPD64" > "LPD64"\n",
4000 GOTO(cleanup, rc = -EINVAL);
4003 next_id = filter_last_id(filter, group) + 1;
4005 /* Don't create objects beyond the valid range for this SEQ */
4006 if (unlikely(fid_seq_is_mdt0(group) &&
4007 next_id >= IDIF_MAX_OID)) {
4008 CERROR("%s:"POSTID" hit the IDIF_MAX_OID (1<<48)!\n",
4009 obd->obd_name, next_id, group);
4010 GOTO(cleanup, rc = -ENOSPC);
4011 } else if (unlikely(!fid_seq_is_mdt0(group) &&
4012 next_id >= OBIF_MAX_OID)) {
4013 CERROR("%s:"POSTID" hit the OBIF_MAX_OID (1<<32)!\n",
4014 obd->obd_name, next_id, group);
4015 GOTO(cleanup, rc = -ENOSPC);
4018 dparent = filter_parent_lock(obd, group, next_id);
4019 if (IS_ERR(dparent))
4020 GOTO(cleanup, rc = PTR_ERR(dparent));
4021 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
4023 dchild = filter_fid2dentry(obd, dparent, group, next_id);
4025 GOTO(cleanup, rc = PTR_ERR(dchild));
4026 cleanup_phase = 2; /* f_dput(dchild) */
4028 if (dchild->d_inode != NULL) {
4029 /* This would only happen if lastobjid was bad on disk*/
4030 /* Could also happen if recreating missing obj but it
4031 * already exists. */
4033 CERROR("%s: recreating existing object %.*s?\n",
4034 obd->obd_name, dchild->d_name.len,
4035 dchild->d_name.name);
4037 /* Use these existing objects if they are
4039 if (i_size_read(dchild->d_inode) == 0) {
4040 rc = filter_use_existing_obj(obd,dchild,
4041 &handle, &cleanup_phase);
4048 CERROR("%s: Serious error: objid %.*s already "
4049 "exists; is this filesystem corrupt?\n",
4050 obd->obd_name, dchild->d_name.len,
4051 dchild->d_name.name);
4054 GOTO(cleanup, rc = -EEXIST);
4057 handle = fsfilt_start_log(obd, dparent->d_inode,
4058 FSFILT_OP_CREATE, NULL, 1);
4060 GOTO(cleanup, rc = PTR_ERR(handle));
4063 CDEBUG(D_INODE, "%s: filter_precreate(od->o_seq="LPU64
4064 ",od->o_id="LPU64")\n", obd->obd_name, group,
4067 /* We mark object SUID+SGID to flag it for accepting UID+GID
4068 * from client on first write. Currently the permission bits
4069 * on the OST are never used, so this is OK. */
4070 rc = ll_vfs_create(dparent->d_inode, dchild,
4071 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
4073 CWARN("%s: create failed: rc = %d\n", obd->obd_name,rc);
4074 if (rc == -ENOSPC) {
4075 os_ffree = filter_calc_free_inodes(obd);
4079 if (obd->obd_osfs.os_bavail <
4080 (obd->obd_osfs.os_blocks >> 10)) {
4081 if (oa->o_valid & OBD_MD_FLFLAGS) {
4082 oa->o_flags |= OBD_FL_NOSPC_BLK;
4084 oa->o_valid |= OBD_MD_FLFLAGS;
4085 oa->o_flags = OBD_FL_NOSPC_BLK;
4088 CWARN("%s: free inode "LPU64"\n",
4089 obd->obd_name, os_ffree);
4095 if (dchild->d_inode)
4096 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
4097 dchild->d_inode->i_ino);
4100 /* Initialize a/c/m time so any client timestamp will always
4101 * be newer and update the inode. ctime = 0 is also handled
4102 * specially in fsfilt_ext3_setattr(). See LU-221, LU-1042 */
4103 iattr.ia_valid = ATTR_ATIME | ATTR_MTIME | ATTR_CTIME;
4104 LTIME_S(iattr.ia_atime) = 0;
4105 LTIME_S(iattr.ia_mtime) = 0;
4106 LTIME_S(iattr.ia_ctime) = 0;
4107 err = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
4109 CWARN("%s: unable to initialize a/c/m time of newly "
4110 "created object %.*s: rc = %d\n",
4111 obd->obd_name, dchild->d_name.len,
4112 dchild->d_name.name, err);
4114 if (!recreate_obj) {
4115 filter_set_last_id(filter, next_id, group);
4116 err = filter_update_last_objid(obd, group, 0);
4118 CERROR("%s: unable to write lastobjid "
4119 "but file created: rc = %d\n",
4120 obd->obd_name, err);
4124 switch(cleanup_phase) {
4126 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4128 CERROR("error on commit, err = %d\n", err);
4135 filter_parent_unlock(dparent);
4142 if (cfs_time_after(jiffies, enough_time)) {
4145 "%s: precreate slow - want %d got %d \n",
4146 obd->obd_name, *num, i);
4153 "%s: created %d objects for group "POSTID" rc %d\n",
4154 obd->obd_name, i, filter->fo_last_objids[group], group, rc);
4159 int filter_create(const struct lu_env *env, struct obd_export *exp,
4160 struct obdo *oa, struct lov_stripe_md **ea,
4161 struct obd_trans_info *oti)
4163 struct obd_device *obd = exp->exp_obd;
4164 struct filter_export_data *fed;
4165 struct filter_obd *filter;
4166 struct lvfs_run_ctxt saved;
4167 struct lov_stripe_md *lsm = NULL;
4171 CDEBUG(D_INODE, "%s: filter_create(group="LPU64",id="
4172 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4174 fed = &exp->exp_filter_data;
4175 filter = &obd->u.filter;
4177 /* 1.8 client doesn't carry the ocd_group with connect request,
4178 * so the fed_group will always be zero for 1.8 client. */
4179 if (!(exp->exp_connect_flags & OBD_CONNECT_FULL20)) {
4180 if (oa->o_seq != FID_SEQ_OST_MDT0 &&
4181 oa->o_seq != FID_SEQ_LLOG &&
4182 oa->o_seq != FID_SEQ_ECHO) {
4183 CERROR("The request from older client has invalid"
4184 " group "LPU64"!\n", oa->o_seq);
4187 } else if (fed->fed_group != oa->o_seq) {
4188 CERROR("%s: this export (nid %s) used object group %d "
4189 "earlier; now it's trying to use group "LPU64"!"
4190 " This could be a bug in the MDS. Please report to "
4191 "http://bugzilla.lustre.org/\n", obd->obd_name,
4192 obd_export_nid2str(exp), fed->fed_group, oa->o_seq);
4199 rc = obd_alloc_memmd(exp, &lsm);
4206 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4208 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4209 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4210 if (!obd->obd_recovering ||
4211 oa->o_id > filter_last_id(filter, oa->o_seq)) {
4212 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4213 oa->o_id, filter_last_id(filter, oa->o_seq));
4217 cfs_mutex_lock(&filter->fo_create_locks[oa->o_seq]);
4218 rc = filter_precreate(obd, oa, oa->o_seq, &diff);
4219 cfs_mutex_unlock(&filter->fo_create_locks[oa->o_seq]);
4222 rc = filter_handle_precreate(exp, oa, oa->o_seq, oti);
4225 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4226 if (rc && ea != NULL && *ea != lsm) {
4227 obd_free_memmd(exp, &lsm);
4228 } else if (rc == 0 && ea != NULL) {
4229 /* XXX LOV STACKING: the lsm that is passed to us from
4230 * LOV does not have valid lsm_oinfo data structs, so
4231 * don't go touching that. This needs to be fixed in a
4233 lsm->lsm_object_id = oa->o_id;
4240 int filter_destroy(const struct lu_env *env, struct obd_export *exp,
4241 struct obdo *oa, struct lov_stripe_md *md,
4242 struct obd_trans_info *oti, struct obd_export *md_exp,
4245 unsigned int qcids[MAXQUOTAS] = {0, 0};
4246 struct obd_device *obd;
4247 struct filter_obd *filter;
4248 struct dentry *dchild = NULL, *dparent = NULL;
4249 struct lustre_handle lockh = { 0 };
4250 struct lvfs_run_ctxt saved;
4251 void *handle = NULL;
4252 struct llog_cookie *fcc = NULL;
4253 int rc, rc2, cleanup_phase = 0, sync = 0;
4258 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4259 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4264 filter = &obd->u.filter;
4266 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4269 CDEBUG(D_INODE, "%s: filter_destroy(group="LPU64",oid="
4270 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4272 dchild = filter_fid2dentry(obd, NULL, oa->o_seq, oa->o_id);
4274 GOTO(cleanup, rc = PTR_ERR(dchild));
4277 if (dchild->d_inode == NULL) {
4278 CDEBUG(D_INODE, "destroying non-existent object "POSTID"\n",
4279 oa->o_id, oa->o_seq);
4280 /* If object already gone, cancel cookie right now */
4281 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4282 struct llog_ctxt *ctxt;
4283 struct obd_llog_group *olg;
4285 olg = filter_find_olg(obd, oa->o_seq);
4287 CERROR(" %s: can not find olg of group %d\n",
4288 obd->obd_name, (int)oa->o_seq);
4289 GOTO(cleanup, rc = PTR_ERR(olg));
4291 fcc = &oa->o_lcookie;
4292 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4293 llog_cancel(NULL, ctxt, NULL, 1, fcc, 0);
4294 llog_ctxt_put(ctxt);
4295 fcc = NULL; /* we didn't allocate fcc, don't free it */
4297 GOTO(cleanup, rc = -ENOENT);
4300 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_seq, &lockh);
4304 /* Our MDC connection is established by the MDS to us */
4305 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4306 OBD_ALLOC(fcc, sizeof(*fcc));
4308 *fcc = oa->o_lcookie;
4311 /* we're gonna truncate it first in order to avoid possible deadlock:
4313 * open trasaction open transaction
4314 * down(i_zombie) down(i_zombie)
4315 * restart transaction
4316 * (see BUG 4180) -bzzz
4318 * take i_alloc_sem too to prevent other threads from writing to the
4319 * file while we are truncating it. This can cause lock ordering issue
4320 * between page lock, i_mutex & starting new journal handle.
4321 * (see bug 20321) -johann
4324 down_write(&dchild->d_inode->i_alloc_sem);
4325 mutex_lock(&dchild->d_inode->i_mutex);
4326 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
4328 /* VBR: version recovery check */
4329 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4331 mutex_unlock(&dchild->d_inode->i_mutex);
4332 up_write(&dchild->d_inode->i_alloc_sem);
4336 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4338 if (IS_ERR(handle)) {
4339 mutex_unlock(&dchild->d_inode->i_mutex);
4340 up_write(&dchild->d_inode->i_alloc_sem);
4341 GOTO(cleanup, rc = PTR_ERR(handle));
4344 /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
4345 ll_vfs_dq_init(dchild->d_inode);
4347 iattr.ia_valid = ATTR_SIZE;
4349 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4350 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4351 mutex_unlock(&dchild->d_inode->i_mutex);
4352 up_write(&dchild->d_inode->i_alloc_sem);
4356 GOTO(cleanup, rc = rc2);
4358 /* We don't actually need to lock the parent until we are unlinking
4359 * here, and not while truncating above. That avoids holding the
4360 * parent lock for a long time during truncate, which can block other
4361 * threads from doing anything to objects in that directory. bug 7171 */
4362 dparent = filter_parent_lock(obd, oa->o_seq, oa->o_id);
4363 if (IS_ERR(dparent))
4364 GOTO(cleanup, rc = PTR_ERR(dparent));
4365 cleanup_phase = 3; /* filter_parent_unlock */
4367 mutex_lock(&dchild->d_inode->i_mutex);
4368 handle = fsfilt_start_log(obd, dparent->d_inode,
4369 FSFILT_OP_UNLINK, oti, 1);
4370 if (IS_ERR(handle)) {
4371 mutex_unlock(&dchild->d_inode->i_mutex);
4372 GOTO(cleanup, rc = PTR_ERR(handle));
4374 cleanup_phase = 4; /* fsfilt_commit */
4376 /* Quota release need uid/gid of inode */
4377 obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID | OBD_MD_FLGID);
4379 filter_fmd_drop(exp, oa->o_id, oa->o_seq);
4381 /* this drops dchild->d_inode->i_mutex unconditionally */
4382 rc = filter_destroy_internal(obd, oa->o_id, oa->o_seq, dparent, dchild);
4386 switch(cleanup_phase) {
4389 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4390 oti->oti_handle : handle,
4391 filter_cancel_cookies_cb,
4393 /* If add_journal_cb failed, then filter_finish_transno
4394 * will commit the handle and we will do a sync
4395 * on commit. then we call callback directly to free
4398 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4400 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4403 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4405 CERROR("error on commit, err = %d\n", rc2);
4412 filter_parent_unlock(dparent);
4414 filter_fini_destroy(obd, &lockh);
4418 OBD_FREE(fcc, sizeof(*fcc));
4420 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4423 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4427 /* trigger quota release */
4428 qcids[USRQUOTA] = oa->o_uid;
4429 qcids[GRPQUOTA] = oa->o_gid;
4430 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4433 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4437 /* NB start and end are used for punch, but not truncate */
4438 static int filter_truncate(const struct lu_env *env, struct obd_export *exp,
4439 struct obd_info *oinfo, struct obd_trans_info *oti,
4440 struct ptlrpc_request_set *rqset)
4445 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4446 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4447 oinfo->oi_policy.l_extent.end);
4451 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4452 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,oinfo->oi_oa->o_valid,
4453 oinfo->oi_policy.l_extent.start);
4455 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4456 rc = filter_setattr(env, exp, oinfo, oti);
4460 static int filter_sync(const struct lu_env *env, struct obd_export *exp,
4461 struct obd_info *oinfo, obd_off start, obd_off end,
4462 struct ptlrpc_request_set *set)
4464 struct lvfs_run_ctxt saved;
4465 struct obd_device_target *obt;
4466 struct dentry *dentry;
4470 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
4471 (struct lustre_capa *)oinfo->oi_capa,
4472 CAPA_OPC_OSS_WRITE);
4476 obt = &exp->exp_obd->u.obt;
4478 /* An objid of zero is taken to mean "sync whole filesystem" */
4479 if (!oinfo->oi_oa || !(oinfo->oi_oa->o_valid & OBD_MD_FLID)) {
4480 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4481 /* Flush any remaining cancel messages out to the target */
4482 filter_sync_llogs(exp->exp_obd, exp);
4486 dentry = filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi);
4488 RETURN(PTR_ERR(dentry));
4490 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4492 mutex_lock(&dentry->d_inode->i_mutex);
4494 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4496 /* just any file to grab fsync method - "file" arg unused */
4497 struct file *file = obt->obt_rcvd_filp;
4499 if (file->f_op && file->f_op->fsync)
4500 rc = file->f_op->fsync(NULL, dentry, 1);
4502 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4506 mutex_unlock(&dentry->d_inode->i_mutex);
4508 oinfo->oi_oa->o_valid = OBD_MD_FLID;
4509 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
4511 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4513 filter_counter_incr(exp, LPROC_FILTER_STATS_SYNC, oinfo->oi_jobid, 1);
4518 static int filter_get_info(const struct lu_env *env, struct obd_export *exp,
4519 __u32 keylen, void *key, __u32 *vallen, void *val,
4520 struct lov_stripe_md *lsm)
4522 struct obd_device *obd;
4525 obd = class_exp2obd(exp);
4527 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4531 if (KEY_IS(KEY_BLOCKSIZE)) {
4532 __u32 *blocksize = val;
4534 if (*vallen < sizeof(*blocksize))
4536 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4538 *vallen = sizeof(*blocksize);
4542 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4543 __u32 *blocksize_bits = val;
4544 if (blocksize_bits) {
4545 if (*vallen < sizeof(*blocksize_bits))
4547 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4549 *vallen = sizeof(*blocksize_bits);
4553 if (KEY_IS(KEY_LAST_ID)) {
4554 obd_id *last_id = val;
4555 /* FIXME: object groups */
4557 if (*vallen < sizeof(*last_id))
4559 *last_id = filter_last_id(&obd->u.filter,
4560 exp->exp_filter_data.fed_group);
4562 *vallen = sizeof(*last_id);
4566 if (KEY_IS(KEY_FIEMAP)) {
4567 struct ll_fiemap_info_key *fm_key = key;
4568 struct dentry *dentry;
4569 struct ll_user_fiemap *fiemap = val;
4570 struct lvfs_run_ctxt saved;
4573 if (fiemap == NULL) {
4574 *vallen = fiemap_count_to_size(
4575 fm_key->fiemap.fm_extent_count);
4579 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa.o_oi,
4582 RETURN(PTR_ERR(dentry));
4584 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4585 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4586 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4588 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4594 if (KEY_IS(KEY_SYNC_LOCK_CANCEL)) {
4595 *((__u32 *) val) = obd->u.filter.fo_sync_lock_cancel;
4596 *vallen = sizeof(__u32);
4600 CDEBUG(D_IOCTL, "invalid key\n");
4604 static inline int filter_setup_llog_group(struct obd_export *exp,
4605 struct obd_device *obd,
4608 struct obd_llog_group *olg;
4609 struct llog_ctxt *ctxt;
4612 olg = filter_find_create_olg(obd, group);
4614 RETURN(PTR_ERR(olg));
4616 llog_group_set_export(olg, exp);
4618 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4619 LASSERTF(ctxt != NULL, "ctxt is null\n");
4621 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4622 llog_ctxt_put(ctxt);
4626 static int filter_set_grant_shrink(struct obd_export *exp,
4627 struct ost_body *body)
4629 /* handle shrink grant */
4630 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4631 filter_grant_incoming(exp, &body->oa);
4632 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4638 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4640 struct obd_device *obd;
4646 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4650 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4651 obd_export_nid2str(exp));
4652 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4654 /* setup llog imports */
4656 group = (int)(*(__u32 *)val);
4658 group = 0; /* default value */
4660 LASSERT_SEQ_IS_MDT(group);
4661 rc = filter_setup_llog_group(exp, obd, group);
4665 if (group == FID_SEQ_OST_MDT0) {
4666 /* setup llog group 1 for interop */
4667 filter_setup_llog_group(exp, obd, FID_SEQ_LLOG);
4670 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4675 static int filter_set_info_async(const struct lu_env *env,
4676 struct obd_export *exp, __u32 keylen,
4677 void *key, __u32 vallen, void *val,
4678 struct ptlrpc_request_set *set)
4680 struct obd_device *obd;
4685 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4689 if (KEY_IS(KEY_CAPA_KEY)) {
4691 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4693 CERROR("filter update capability key failed: %d\n", rc);
4697 if (KEY_IS(KEY_REVIMP_UPD)) {
4698 filter_revimp_update(exp);
4699 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4703 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4704 filter_adapt_sptlrpc_conf(obd, 0);
4708 if (KEY_IS(KEY_MDS_CONN))
4709 RETURN(filter_set_mds_conn(exp, val));
4711 if (KEY_IS(KEY_GRANT_SHRINK))
4712 RETURN(filter_set_grant_shrink(exp, val));
4717 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4718 int len, void *karg, void *uarg)
4720 struct obd_device *obd = exp->exp_obd;
4721 struct obd_ioctl_data *data = karg;
4725 case OBD_IOC_ABORT_RECOVERY: {
4726 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4727 target_stop_recovery_thread(obd);
4731 case OBD_IOC_SYNC: {
4732 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4733 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4737 case OBD_IOC_SET_READONLY: {
4739 struct super_block *sb = obd->u.obt.obt_sb;
4740 struct inode *inode = sb->s_root->d_inode;
4741 BDEVNAME_DECLARE_STORAGE(tmp);
4742 CERROR("*** setting device %s read-only ***\n",
4743 ll_bdevname(sb, tmp));
4745 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4746 if (!IS_ERR(handle))
4747 rc = fsfilt_commit(obd, inode, handle, 1);
4749 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4750 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4752 rc = lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4756 case OBD_IOC_CATLOGLIST: {
4757 rc = llog_catalog_list(obd, 1, data);
4761 case OBD_IOC_LLOG_CANCEL:
4762 case OBD_IOC_LLOG_REMOVE:
4763 case OBD_IOC_LLOG_INFO:
4764 case OBD_IOC_LLOG_PRINT: {
4765 /* FIXME to be finished */
4766 RETURN(-EOPNOTSUPP);
4768 struct llog_ctxt *ctxt = NULL;
4770 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4771 rc = llog_ioctl(ctxt, cmd, data);
4772 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4785 static int filter_health_check(const struct lu_env *env, struct obd_device *obd)
4787 #ifdef USE_HEALTH_CHECK_WRITE
4788 struct filter_obd *filter = &obd->u.filter;
4793 * health_check to return 0 on healthy
4794 * and 1 on unhealthy.
4796 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4799 #ifdef USE_HEALTH_CHECK_WRITE
4800 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4801 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4806 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4809 return filter_fid2dentry(data, NULL, gr, id);
4812 static int filter_process_config(struct obd_device *obd, obd_count len,
4815 struct lustre_cfg *lcfg = buf;
4816 struct lprocfs_static_vars lvars;
4819 switch (lcfg->lcfg_command) {
4821 lprocfs_filter_init_vars(&lvars);
4823 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4833 static int filter_notify(struct obd_device *obd,
4834 struct obd_device *unused,
4835 enum obd_notify_event ev, void *data)
4838 case OBD_NOTIFY_CONFIG:
4839 LASSERT(obd->obd_no_conn);
4840 cfs_spin_lock(&obd->obd_dev_lock);
4841 obd->obd_no_conn = 0;
4842 cfs_spin_unlock(&obd->obd_dev_lock);
4845 CDEBUG(D_INFO, "%s: Unhandled notification %#x\n",
4851 static struct lvfs_callback_ops filter_lvfs_ops = {
4852 l_fid2dentry: filter_lvfs_fid2dentry,
4855 static struct obd_ops filter_obd_ops = {
4856 .o_owner = THIS_MODULE,
4857 .o_get_info = filter_get_info,
4858 .o_set_info_async = filter_set_info_async,
4859 .o_setup = filter_setup,
4860 .o_precleanup = filter_precleanup,
4861 .o_cleanup = filter_cleanup,
4862 .o_connect = filter_connect,
4863 .o_reconnect = filter_reconnect,
4864 .o_disconnect = filter_disconnect,
4865 .o_ping = filter_ping,
4866 .o_init_export = filter_init_export,
4867 .o_destroy_export = filter_destroy_export,
4868 .o_statfs = filter_statfs,
4869 .o_getattr = filter_getattr,
4870 .o_unpackmd = filter_unpackmd,
4871 .o_create = filter_create,
4872 .o_setattr = filter_setattr,
4873 .o_destroy = filter_destroy,
4874 .o_punch = filter_truncate,
4875 .o_sync = filter_sync,
4876 .o_preprw = filter_preprw,
4877 .o_commitrw = filter_commitrw,
4878 .o_llog_init = filter_llog_init,
4879 .o_llog_connect = filter_llog_connect,
4880 .o_llog_finish = filter_llog_finish,
4881 .o_iocontrol = filter_iocontrol,
4882 .o_health_check = filter_health_check,
4883 .o_process_config = filter_process_config,
4884 .o_notify = filter_notify,
4887 quota_interface_t *filter_quota_interface_ref;
4888 extern quota_interface_t filter_quota_interface;
4890 static int __init obdfilter_init(void)
4892 struct lprocfs_static_vars lvars;
4895 /** sanity check for group<->mdsno conversion */
4896 for (i = 0; i < MAX_MDT_COUNT; i++)
4897 LASSERT(objseq_to_mdsno(mdt_to_obd_objseq(i)) == i);
4899 lprocfs_filter_init_vars(&lvars);
4901 cfs_request_module("%s", "lquota");
4902 OBD_ALLOC(obdfilter_created_scratchpad,
4903 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4904 sizeof(*obdfilter_created_scratchpad));
4905 if (obdfilter_created_scratchpad == NULL)
4908 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4909 sizeof(struct filter_mod_data),
4912 GOTO(out, rc = -ENOMEM);
4914 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4915 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4917 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4918 LUSTRE_OST_NAME, NULL);
4922 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4923 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4924 ll_fmd_cachep = NULL;
4926 if (filter_quota_interface_ref)
4927 PORTAL_SYMBOL_PUT(filter_quota_interface);
4929 OBD_FREE(obdfilter_created_scratchpad,
4930 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4931 sizeof(*obdfilter_created_scratchpad));
4937 static void __exit obdfilter_exit(void)
4939 if (filter_quota_interface_ref)
4940 PORTAL_SYMBOL_PUT(filter_quota_interface);
4942 if (ll_fmd_cachep) {
4943 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4944 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4945 ll_fmd_cachep = NULL;
4948 class_unregister_type(LUSTRE_OST_NAME);
4949 OBD_FREE(obdfilter_created_scratchpad,
4950 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4951 sizeof(*obdfilter_created_scratchpad));
4954 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4955 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4956 MODULE_LICENSE("GPL");
4958 module_init(obdfilter_init);
4959 module_exit(obdfilter_exit);