1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * lustre/obdfilter/filter.c
41 * Author: Peter Braam <braam@clusterfs.com>
42 * Author: Andreas Dilger <adilger@clusterfs.com>
46 * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
47 * (which need to get journal_lock, may block if journal full).
49 * Invariant: Call filter_start_transno() before any journal ops to avoid the
50 * same deadlock problem. We can (and want) to get rid of the
51 * transno sem in favour of the dir/inode i_mutex to avoid single
52 * threaded operation on the OST.
55 #define DEBUG_SUBSYSTEM S_FILTER
57 #ifndef AUTOCONF_INCLUDED
58 #include <linux/config.h>
60 #include <linux/module.h>
62 #include <linux/dcache.h>
63 #include <linux/init.h>
64 #include <linux/version.h>
65 #include <linux/sched.h>
66 #include <linux/mount.h>
67 #include <linux/buffer_head.h>
69 #include <obd_cksum.h>
70 #include <obd_class.h>
72 #include <lustre_dlm.h>
73 #include <lustre_fsfilt.h>
74 #include <lprocfs_status.h>
75 #include <lustre_log.h>
76 #include <libcfs/list.h>
77 #include <lustre_disk.h>
78 #include <lustre_quota.h>
79 #include <linux/slab.h>
80 #include <lustre_param.h>
81 #include <lustre/ll_fiemap.h>
83 #include "filter_internal.h"
85 static struct lvfs_callback_ops filter_lvfs_ops;
86 cfs_mem_cache_t *ll_fmd_cachep;
88 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
89 void *cb_data, int error)
91 struct obd_export *exp = cb_data;
92 LASSERT(exp->exp_obd == obd);
93 obd_transno_commit_cb(obd, transno, exp, error);
94 class_export_cb_put(exp);
97 int filter_version_get_check(struct obd_export *exp,
98 struct obd_trans_info *oti, struct inode *inode)
102 if (inode == NULL || oti == NULL)
105 curr_version = fsfilt_get_version(exp->exp_obd, inode);
106 if ((__s64)curr_version == -EOPNOTSUPP)
108 /* VBR: version is checked always because costs nothing */
109 if (oti->oti_pre_version != 0 &&
110 oti->oti_pre_version != curr_version) {
111 CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
112 oti->oti_pre_version, curr_version);
113 cfs_spin_lock(&exp->exp_lock);
114 exp->exp_vbr_failed = 1;
115 cfs_spin_unlock(&exp->exp_lock);
118 oti->oti_pre_version = curr_version;
122 /* Assumes caller has already pushed us into the kernel context. */
123 int filter_finish_transno(struct obd_export *exp, struct inode *inode,
124 struct obd_trans_info *oti, int rc, int force_sync)
126 struct obd_device_target *obt = &exp->exp_obd->u.obt;
127 struct tg_export_data *ted = &exp->exp_target_data;
128 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
129 struct lsd_client_data *lcd;
132 int err, log_pri = D_RPCTRACE;
134 /* Propagate error code. */
138 if (!exp->exp_obd->obd_replayable || oti == NULL)
141 cfs_mutex_down(&ted->ted_lcd_lock);
143 /* if the export has already been disconnected, we have no last_rcvd slot,
144 * update server data with latest transno then */
146 cfs_mutex_up(&ted->ted_lcd_lock);
147 CWARN("commit transaction for disconnected client %s: rc %d\n",
148 exp->exp_client_uuid.uuid, rc);
149 err = filter_update_server_data(exp->exp_obd);
153 /* we don't allocate new transnos for replayed requests */
154 cfs_spin_lock(&obt->obt_lut->lut_translock);
155 if (oti->oti_transno == 0) {
156 last_rcvd = le64_to_cpu(lsd->lsd_last_transno) + 1;
157 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
158 LASSERT(last_rcvd >= le64_to_cpu(lcd->lcd_last_transno));
160 last_rcvd = oti->oti_transno;
161 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
162 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
163 if (unlikely(last_rcvd < le64_to_cpu(lcd->lcd_last_transno))) {
164 CERROR("Trying to overwrite bigger transno, on-disk: "
165 LPU64", new: "LPU64"\n",
166 le64_to_cpu(lcd->lcd_last_transno), last_rcvd);
167 cfs_spin_lock(&exp->exp_lock);
168 exp->exp_vbr_failed = 1;
169 cfs_spin_unlock(&exp->exp_lock);
170 cfs_spin_unlock(&obt->obt_lut->lut_translock);
171 cfs_mutex_up(&ted->ted_lcd_lock);
175 oti->oti_transno = last_rcvd;
177 lcd->lcd_last_transno = cpu_to_le64(last_rcvd);
178 lcd->lcd_pre_versions[0] = cpu_to_le64(oti->oti_pre_version);
179 lcd->lcd_last_xid = cpu_to_le64(oti->oti_xid);
180 cfs_spin_unlock(&obt->obt_lut->lut_translock);
183 fsfilt_set_version(exp->exp_obd, inode, last_rcvd);
185 off = ted->ted_lr_off;
187 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
188 ted->ted_lr_idx, ted->ted_lr_off);
191 class_export_cb_get(exp); /* released when the cb is called */
193 force_sync = fsfilt_add_journal_cb(exp->exp_obd,
199 err = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
200 lcd, sizeof(*lcd), &off,
201 force_sync | exp->exp_need_sync);
203 filter_commit_cb(exp->exp_obd, last_rcvd, exp, err);
211 CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
212 last_rcvd, lcd->lcd_uuid, ted->ted_lr_idx, err);
213 cfs_mutex_up(&ted->ted_lcd_lock);
217 void f_dput(struct dentry *dentry)
219 /* Can't go inside filter_ddelete because it can block */
220 CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
221 dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
222 LASSERT(atomic_read(&dentry->d_count) > 0);
227 static void init_brw_stats(struct brw_stats *brw_stats)
230 for (i = 0; i < BRW_LAST; i++)
231 cfs_spin_lock_init(&brw_stats->hist[i].oh_lock);
234 static int lprocfs_init_rw_stats(struct obd_device *obd,
235 struct lprocfs_stats **stats)
239 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
240 LPROC_FILTER_LAST - 1;
241 *stats = lprocfs_alloc_stats(num_stats, LPROCFS_STATS_FLAG_NOPERCPU);
245 lprocfs_init_ops_stats(LPROC_FILTER_LAST, *stats);
246 lprocfs_counter_init(*stats, LPROC_FILTER_READ_BYTES,
247 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
248 lprocfs_counter_init(*stats, LPROC_FILTER_WRITE_BYTES,
249 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
254 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
255 plus the procfs overhead :( */
256 static int filter_export_stats_init(struct obd_device *obd,
257 struct obd_export *exp,
263 if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
264 /* Self-export gets no proc entry */
267 rc = lprocfs_exp_setup(exp, client_nid, &newnid);
269 /* Mask error for already created
277 struct nid_stat *tmp = exp->exp_nid_stats;
278 LASSERT(tmp != NULL);
280 OBD_ALLOC(tmp->nid_brw_stats, sizeof(struct brw_stats));
281 if (tmp->nid_brw_stats == NULL)
282 GOTO(clean, rc = -ENOMEM);
284 init_brw_stats(tmp->nid_brw_stats);
285 rc = lprocfs_seq_create(exp->exp_nid_stats->nid_proc, "brw_stats",
286 0644, &filter_per_nid_stats_fops,
289 CWARN("Error adding the brw_stats file\n");
291 rc = lprocfs_init_rw_stats(obd, &exp->exp_nid_stats->nid_stats);
295 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
299 rc = lprocfs_nid_ldlm_stats_init(tmp);
309 /* Add client data to the FILTER. We use a bitmap to locate a free space
310 * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
311 * Otherwise, we have just read the data from the last_rcvd file and
312 * we know its offset. */
313 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
316 struct obd_device_target *obt = &obd->u.obt;
317 struct tg_export_data *ted = &exp->exp_target_data;
318 struct lr_server_data *lsd = class_server_data(obd);
319 unsigned long *bitmap = obt->obt_lut->lut_client_bitmap;
320 int new_client = (cl_idx == -1);
324 LASSERT(bitmap != NULL);
325 LASSERTF(cl_idx > -2, "%d\n", cl_idx);
328 if (strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid) == 0)
331 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
332 * there's no need for extra complication here
335 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
337 if (cl_idx >= LR_MAX_CLIENTS) {
338 CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
342 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
343 cl_idx = cfs_find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
348 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
349 CERROR("FILTER client %d: bit already set in bitmap!\n",
355 ted->ted_lr_idx = cl_idx;
356 ted->ted_lr_off = le32_to_cpu(lsd->lsd_client_start) +
357 cl_idx * le16_to_cpu(lsd->lsd_client_size);
358 cfs_init_mutex(&ted->ted_lcd_lock);
359 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
361 CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
362 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
365 struct lvfs_run_ctxt saved;
366 loff_t off = ted->ted_lr_off;
370 CDEBUG(D_INFO, "writing client lcd at idx %u (%llu) (len %u)\n",
371 ted->ted_lr_idx,off,(unsigned int)sizeof(*ted->ted_lcd));
373 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
376 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
377 /* Transaction needed to fix bug 1403 */
378 handle = fsfilt_start(obd,
379 obt->obt_rcvd_filp->f_dentry->d_inode,
380 FSFILT_OP_SETATTR, NULL);
381 if (IS_ERR(handle)) {
382 rc = PTR_ERR(handle);
383 CERROR("unable to start transaction: rc %d\n", rc);
385 ted->ted_lcd->lcd_last_epoch = lsd->lsd_start_epoch;
386 exp->exp_last_request_time = cfs_time_current_sec();
387 rc = fsfilt_add_journal_cb(obd, 0, handle,
388 target_client_add_cb,
389 class_export_cb_get(exp));
391 cfs_spin_lock(&exp->exp_lock);
392 exp->exp_need_sync = 1;
393 cfs_spin_unlock(&exp->exp_lock);
395 rc = fsfilt_write_record(obd, obt->obt_rcvd_filp,
397 sizeof(*ted->ted_lcd),
398 &off, rc /* sync if no cb */);
400 obt->obt_rcvd_filp->f_dentry->d_inode,
403 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
406 CERROR("error writing %s client idx %u: rc %d\n",
407 LAST_RCVD, ted->ted_lr_idx, rc);
414 static int filter_client_del(struct obd_export *exp)
416 struct tg_export_data *ted = &exp->exp_target_data;
417 struct obd_device_target *obt = &exp->exp_obd->u.obt;
418 struct lvfs_run_ctxt saved;
423 if (ted->ted_lcd == NULL)
426 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
427 if (strcmp(ted->ted_lcd->lcd_uuid, exp->exp_obd->obd_uuid.uuid ) == 0)
430 LASSERT(obt->obt_lut->lut_client_bitmap != NULL);
432 off = ted->ted_lr_off;
434 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
435 ted->ted_lr_idx, ted->ted_lr_off, ted->ted_lcd->lcd_uuid);
437 /* Don't clear ted_lr_idx here as it is likely also unset. At worst
438 * we leak a client slot that will be cleaned on the next recovery. */
440 CERROR("%s: client idx %d has med_off %lld\n",
441 exp->exp_obd->obd_name, ted->ted_lr_idx, off);
442 GOTO(free, rc = -EINVAL);
445 /* Clear the bit _after_ zeroing out the client so we don't
446 race with filter_client_add and zero out new clients.*/
447 if (!cfs_test_bit(ted->ted_lr_idx, obt->obt_lut->lut_client_bitmap)) {
448 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
453 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
454 /* Make sure the server's last_transno is up to date.
455 * This should be done before zeroing client slot so last_transno will
456 * be in server data or in client data in case of failure */
457 filter_update_server_data(exp->exp_obd);
459 cfs_mutex_down(&ted->ted_lcd_lock);
460 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
461 rc = fsfilt_write_record(exp->exp_obd, obt->obt_rcvd_filp,
463 sizeof(*ted->ted_lcd), &off, 0);
464 cfs_mutex_up(&ted->ted_lcd_lock);
465 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
467 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
468 "zero out client %s at idx %u/%llu in %s, rc %d\n",
469 ted->ted_lcd->lcd_uuid, ted->ted_lr_idx, ted->ted_lr_off,
476 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
477 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
478 struct filter_mod_data *fmd)
480 LASSERT_SPIN_LOCKED(&fed->fed_lock);
481 if (--fmd->fmd_refcount == 0) {
482 /* XXX when we have persistent reservations and the handle
483 * is stored herein we need to drop it here. */
484 fed->fed_mod_count--;
485 cfs_list_del(&fmd->fmd_list);
486 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
490 /* drop fmd reference, free it if last ref */
491 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
493 struct filter_export_data *fed;
498 fed = &exp->exp_filter_data;
499 cfs_spin_lock(&fed->fed_lock);
500 filter_fmd_put_nolock(fed, fmd); /* caller reference */
501 cfs_spin_unlock(&fed->fed_lock);
504 /* expire entries from the end of the list if there are too many
505 * or they are too old */
506 static void filter_fmd_expire_nolock(struct filter_obd *filter,
507 struct filter_export_data *fed,
508 struct filter_mod_data *keep)
510 struct filter_mod_data *fmd, *tmp;
512 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
516 if (cfs_time_before(jiffies, fmd->fmd_expire) &&
517 fed->fed_mod_count < filter->fo_fmd_max_num)
520 cfs_list_del_init(&fmd->fmd_list);
521 filter_fmd_put_nolock(fed, fmd); /* list reference */
525 void filter_fmd_expire(struct obd_export *exp)
527 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
528 filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
529 &exp->exp_filter_data, NULL);
530 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
533 /* find specified objid, group in export fmd list.
534 * caller must hold fed_lock and take fmd reference itself */
535 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
536 struct filter_export_data *fed,
537 obd_id objid, obd_seq group)
539 struct filter_mod_data *found = NULL, *fmd;
541 LASSERT_SPIN_LOCKED(&fed->fed_lock);
543 cfs_list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
544 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
546 cfs_list_del(&fmd->fmd_list);
547 cfs_list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
548 fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
553 filter_fmd_expire_nolock(filter, fed, found);
558 /* Find fmd based on objid and group, or return NULL if not found. */
559 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
560 obd_id objid, obd_seq group)
562 struct filter_mod_data *fmd;
564 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
565 fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
566 &exp->exp_filter_data, objid, group);
568 fmd->fmd_refcount++; /* caller reference */
569 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
574 /* Find fmd based on objid and group, or create a new one if none is found.
575 * It is possible for this function to return NULL under memory pressure,
576 * or if objid = 0 is passed (which will only cause old entries to expire).
577 * Currently this is not fatal because any fmd state is transient and
578 * may also be freed when it gets sufficiently old. */
579 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
580 obd_id objid, obd_seq group)
582 struct filter_export_data *fed = &exp->exp_filter_data;
583 struct filter_mod_data *found = NULL, *fmd_new = NULL;
585 OBD_SLAB_ALLOC_PTR_GFP(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO);
587 cfs_spin_lock(&fed->fed_lock);
588 found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
591 cfs_list_add_tail(&fmd_new->fmd_list,
593 fmd_new->fmd_id = objid;
594 fmd_new->fmd_gr = group;
595 fmd_new->fmd_refcount++; /* list reference */
597 fed->fed_mod_count++;
599 OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
603 found->fmd_refcount++; /* caller reference */
604 found->fmd_expire = jiffies +
605 exp->exp_obd->u.filter.fo_fmd_max_age;
608 cfs_spin_unlock(&fed->fed_lock);
614 /* drop fmd list reference so it will disappear when last reference is put.
615 * This isn't so critical because it would in fact only affect the one client
616 * that is doing the unlink and at worst we have an stale entry referencing
617 * an object that should never be used again. */
618 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_seq group)
620 struct filter_mod_data *found = NULL;
622 cfs_spin_lock(&exp->exp_filter_data.fed_lock);
623 found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
625 cfs_list_del_init(&found->fmd_list);
626 filter_fmd_put_nolock(&exp->exp_filter_data, found);
628 cfs_spin_unlock(&exp->exp_filter_data.fed_lock);
631 #define filter_fmd_drop(exp, objid, group)
634 /* remove all entries from fmd list */
635 static void filter_fmd_cleanup(struct obd_export *exp)
637 struct filter_export_data *fed = &exp->exp_filter_data;
638 struct filter_mod_data *fmd = NULL, *tmp;
640 cfs_spin_lock(&fed->fed_lock);
641 cfs_list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
642 cfs_list_del_init(&fmd->fmd_list);
643 filter_fmd_put_nolock(fed, fmd);
645 cfs_spin_unlock(&fed->fed_lock);
648 static int filter_init_export(struct obd_export *exp)
653 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
654 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
656 cfs_spin_lock(&exp->exp_lock);
657 exp->exp_connecting = 1;
658 cfs_spin_unlock(&exp->exp_lock);
660 /* self-export doesn't need client data and ldlm initialization */
661 if (unlikely(exp == exp->exp_obd->obd_self_export))
664 rc = lut_client_alloc(exp);
666 rc = ldlm_init_export(exp);
668 CERROR("%s: Can't initialize export: rc %d\n",
669 exp->exp_obd->obd_name, rc);
674 static int filter_free_server_data(struct obd_device_target *obt)
676 lut_fini(NULL, obt->obt_lut);
677 OBD_FREE_PTR(obt->obt_lut);
681 /* assumes caller is already in kernel ctxt */
682 int filter_update_server_data(struct obd_device *obd)
684 struct file *filp = obd->u.obt.obt_rcvd_filp;
685 struct lr_server_data *lsd = class_server_data(obd);
690 CDEBUG(D_INODE, "server uuid : %s\n", lsd->lsd_uuid);
691 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
692 le64_to_cpu(lsd->lsd_last_transno));
693 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
694 le64_to_cpu(lsd->lsd_mount_count));
696 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
698 CERROR("error writing lr_server_data: rc = %d\n", rc);
703 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
706 struct filter_obd *filter = &obd->u.filter;
712 if (filter->fo_last_objid_files[group] == NULL) {
713 CERROR("Object seq "LPU64" not fully setup; not updating "
714 "last_objid\n", group);
718 CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
719 obd->obd_name, filter->fo_last_objids[group], group);
721 tmp = cpu_to_le64(filter->fo_last_objids[group]);
722 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
723 &tmp, sizeof(tmp), &off, force_sync);
725 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
729 extern int ost_handle(struct ptlrpc_request *req);
730 /* assumes caller has already in kernel ctxt */
731 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
733 struct filter_obd *filter = &obd->u.filter;
734 struct lr_server_data *lsd;
735 struct lsd_client_data *lcd = NULL;
736 struct inode *inode = filp->f_dentry->d_inode;
737 unsigned long last_rcvd_size = i_size_read(inode);
738 struct lu_target *lut;
745 /* ensure padding in the struct is the correct size */
746 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
747 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
748 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
749 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
751 /* allocate and initialize lu_target */
755 rc = lut_init(NULL, lut, obd, NULL);
758 lsd = class_server_data(obd);
759 if (last_rcvd_size == 0) {
760 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
762 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
763 lsd->lsd_last_transno = 0;
764 mount_count = lsd->lsd_mount_count = 0;
765 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
766 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
767 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
768 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
769 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
770 /* OBD_COMPAT_OST is set in filter_connect_internal when the
771 * MDS first connects and assigns the OST index number. */
772 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
775 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
777 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
781 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
782 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
783 "using the wrong disk %s. Were the "
784 "/dev/ assignments rearranged?\n",
785 obd->obd_uuid.uuid, lsd->lsd_uuid);
786 GOTO(err_lut, rc = -EINVAL);
788 mount_count = le64_to_cpu(lsd->lsd_mount_count);
789 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
791 /* Assume old last_rcvd format unless I_C_LR is set */
792 if (!(lsd->lsd_feature_incompat &
793 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
794 lsd->lsd_last_transno = lsd->lsd_compat14;
796 /* OBD_COMPAT_OST is set in filter_connect_internal when the
797 * MDS first connects and assigns the OST index number. */
798 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
802 if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
803 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
804 obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
805 ~FILTER_INCOMPAT_SUPP);
806 GOTO(err_lut, rc = -EINVAL);
808 if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
809 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
810 obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
811 ~FILTER_ROCOMPAT_SUPP);
812 /* Do something like remount filesystem read-only */
813 GOTO(err_lut, rc = -EINVAL);
816 start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
818 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
819 obd->obd_name, start_epoch);
820 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
821 obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
822 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
823 obd->obd_name, mount_count + 1);
824 CDEBUG(D_INODE, "%s: server data size: %u\n",
825 obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
826 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
827 obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
828 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
829 obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
830 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
831 obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
832 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
833 last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
834 (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
835 le16_to_cpu(lsd->lsd_client_size));
837 if (!obd->obd_replayable) {
838 CWARN("%s: recovery support OFF\n", obd->obd_name);
844 GOTO(err_client, rc = -ENOMEM);
846 for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
847 off < last_rcvd_size; cl_idx++) {
849 struct obd_export *exp;
850 struct filter_export_data *fed;
852 /* Don't assume off is incremented properly by
853 * fsfilt_read_record(), in case sizeof(*lcd)
854 * isn't the same as lsd->lsd_client_size. */
855 off = le32_to_cpu(lsd->lsd_client_start) +
856 cl_idx * le16_to_cpu(lsd->lsd_client_size);
857 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
859 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
860 LAST_RCVD, cl_idx, off, rc);
861 break; /* read error shouldn't cause startup to fail */
864 if (lcd->lcd_uuid[0] == '\0') {
865 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
870 check_lcd(obd->obd_name, cl_idx, lcd);
872 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
874 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
875 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
876 last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
878 /* These exports are cleaned up by filter_disconnect(), so they
879 * need to be set up like real exports as filter_connect() does.
881 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
883 if (PTR_ERR(exp) == -EALREADY) {
884 /* export already exists, zero out this one */
885 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
889 GOTO(err_client, rc = PTR_ERR(exp));
892 fed = &exp->exp_filter_data;
893 *fed->fed_ted.ted_lcd = *lcd;
894 fed->fed_group = 0; /* will be assigned at connect */
895 filter_export_stats_init(obd, exp, NULL);
896 rc = filter_client_add(obd, exp, cl_idx);
897 /* can't fail for existing client */
898 LASSERTF(rc == 0, "rc = %d\n", rc);
900 /* VBR: set export last committed */
901 exp->exp_last_committed = last_rcvd;
902 cfs_spin_lock(&exp->exp_lock);
903 exp->exp_connecting = 0;
904 exp->exp_in_recovery = 0;
905 cfs_spin_unlock(&exp->exp_lock);
906 obd->obd_max_recoverable_clients++;
907 class_export_put(exp);
909 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
910 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
914 obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
916 obd->u.obt.obt_mount_count = mount_count + 1;
917 lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
919 /* save it, so mount count and last_transno is current */
920 rc = filter_update_server_data(obd);
922 GOTO(err_client, rc);
927 class_disconnect_exports(obd);
929 filter_free_server_data(&obd->u.obt);
933 static int filter_cleanup_groups(struct obd_device *obd)
935 struct filter_obd *filter = &obd->u.filter;
937 struct dentry *dentry;
941 if (filter->fo_dentry_O_groups != NULL) {
942 for (i = 0; i < filter->fo_group_count; i++) {
943 dentry = filter->fo_dentry_O_groups[i];
947 OBD_FREE(filter->fo_dentry_O_groups,
948 filter->fo_group_count *
949 sizeof(*filter->fo_dentry_O_groups));
950 filter->fo_dentry_O_groups = NULL;
952 if (filter->fo_last_objid_files != NULL) {
953 for (i = 0; i < filter->fo_group_count; i++) {
954 filp = filter->fo_last_objid_files[i];
958 OBD_FREE(filter->fo_last_objid_files,
959 filter->fo_group_count *
960 sizeof(*filter->fo_last_objid_files));
961 filter->fo_last_objid_files = NULL;
963 if (filter->fo_dentry_O_sub != NULL) {
964 for (i = 0; i < filter->fo_group_count; i++) {
965 for (j = 0; j < filter->fo_subdir_count; j++) {
966 dentry = filter->fo_dentry_O_sub[i].dentry[j];
971 OBD_FREE(filter->fo_dentry_O_sub,
972 filter->fo_group_count *
973 sizeof(*filter->fo_dentry_O_sub));
974 filter->fo_dentry_O_sub = NULL;
976 if (filter->fo_last_objids != NULL) {
977 OBD_FREE(filter->fo_last_objids,
978 filter->fo_group_count *
979 sizeof(*filter->fo_last_objids));
980 filter->fo_last_objids = NULL;
982 if (filter->fo_dentry_O != NULL) {
983 f_dput(filter->fo_dentry_O);
984 filter->fo_dentry_O = NULL;
989 static int filter_update_last_group(struct obd_device *obd, int group)
991 struct filter_obd *filter = &obd->u.filter;
992 struct file *filp = NULL;
993 int last_group = 0, rc;
997 if (group <= filter->fo_committed_group)
1000 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
1004 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
1008 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1010 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1014 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1015 obd->obd_name, last_group, group);
1019 /* must be sync: bXXXX */
1020 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1022 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1026 filter->fo_committed_group = group;
1029 filp_close(filp, 0);
1033 static int filter_read_group_internal(struct obd_device *obd, int group,
1036 struct filter_obd *filter = &obd->u.filter;
1037 __u64 *new_objids = NULL;
1038 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1039 struct dentry **new_groups = NULL;
1040 struct file **new_files = NULL;
1041 struct dentry *dentry;
1043 int old_count = filter->fo_group_count, rc, stage = 0, i;
1047 int len = group + 1;
1049 snprintf(name, 24, "%d", group);
1053 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1055 if (IS_ERR(dentry)) {
1056 CERROR("Cannot lookup expected object group %d: %ld\n",
1057 group, PTR_ERR(dentry));
1058 RETURN(PTR_ERR(dentry));
1061 dentry = simple_mkdir(filter->fo_dentry_O,
1062 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1063 if (IS_ERR(dentry)) {
1064 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1066 RETURN(PTR_ERR(dentry));
1071 snprintf(name, 24, "O/%d/LAST_ID", group);
1073 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1075 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1076 GOTO(cleanup, rc = PTR_ERR(filp));
1080 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1082 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1086 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1087 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1088 if (tmp_subdirs == NULL)
1089 GOTO(cleanup, rc = -ENOMEM);
1092 for (i = 0; i < filter->fo_subdir_count; i++) {
1094 snprintf(dir, sizeof(dir), "d%u", i);
1096 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1097 obd->u.obt.obt_vfsmnt,
1099 if (IS_ERR(tmp_subdirs->dentry[i])) {
1100 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1101 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1106 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1107 tmp_subdirs->dentry[i]);
1111 /* 'group' is an index; we need an array of length 'group + 1' */
1112 if (group + 1 > old_count) {
1113 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1114 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1115 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1116 OBD_ALLOC(new_files, len * sizeof(*new_files));
1118 if (new_objids == NULL || new_subdirs == NULL ||
1119 new_groups == NULL || new_files == NULL)
1120 GOTO(cleanup, rc = -ENOMEM);
1123 memcpy(new_objids, filter->fo_last_objids,
1124 old_count * sizeof(*new_objids));
1125 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1126 old_count * sizeof(*new_subdirs));
1127 memcpy(new_groups, filter->fo_dentry_O_groups,
1128 old_count * sizeof(*new_groups));
1129 memcpy(new_files, filter->fo_last_objid_files,
1130 old_count * sizeof(*new_files));
1132 OBD_FREE(filter->fo_last_objids,
1133 old_count * sizeof(*new_objids));
1134 OBD_FREE(filter->fo_dentry_O_sub,
1135 old_count * sizeof(*new_subdirs));
1136 OBD_FREE(filter->fo_dentry_O_groups,
1137 old_count * sizeof(*new_groups));
1138 OBD_FREE(filter->fo_last_objid_files,
1139 old_count * sizeof(*new_files));
1141 filter->fo_last_objids = new_objids;
1142 filter->fo_dentry_O_sub = new_subdirs;
1143 filter->fo_dentry_O_groups = new_groups;
1144 filter->fo_last_objid_files = new_files;
1145 filter->fo_group_count = len;
1148 filter->fo_dentry_O_groups[group] = dentry;
1149 filter->fo_last_objid_files[group] = filp;
1150 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1151 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1152 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1155 filter_update_last_group(obd, group);
1157 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1158 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1159 rc = filter_update_last_objid(obd, group, 1);
1163 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1164 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1165 obd->obd_name, group, last_objid);
1170 if (new_objids != NULL)
1171 OBD_FREE(new_objids, len * sizeof(*new_objids));
1172 if (new_subdirs != NULL)
1173 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1174 if (new_groups != NULL)
1175 OBD_FREE(new_groups, len * sizeof(*new_groups));
1176 if (new_files != NULL)
1177 OBD_FREE(new_files, len * sizeof(*new_files));
1179 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1180 for (i = 0; i < filter->fo_subdir_count; i++) {
1181 if (tmp_subdirs->dentry[i] != NULL)
1182 dput(tmp_subdirs->dentry[i]);
1184 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1187 filp_close(filp, 0);
1194 static int filter_read_groups(struct obd_device *obd, int last_group,
1197 struct filter_obd *filter = &obd->u.filter;
1198 int old_count, group, rc = 0;
1200 cfs_down(&filter->fo_init_lock);
1201 old_count = filter->fo_group_count;
1202 for (group = old_count; group <= last_group; group++) {
1203 rc = filter_read_group_internal(obd, group, create);
1207 cfs_up(&filter->fo_init_lock);
1211 /* FIXME: object groups */
1212 static int filter_prep_groups(struct obd_device *obd)
1214 struct filter_obd *filter = &obd->u.filter;
1215 struct dentry *O_dentry;
1217 int last_group, rc = 0, cleanup_phase = 0;
1221 O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1223 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1224 if (IS_ERR(O_dentry)) {
1225 rc = PTR_ERR(O_dentry);
1226 CERROR("cannot open/create O: rc = %d\n", rc);
1229 filter->fo_dentry_O = O_dentry;
1230 cleanup_phase = 1; /* O_dentry */
1232 /* we have to initialize all groups before first connections from
1233 * clients because they may send create/destroy for any group -bzzz */
1234 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1236 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1237 GOTO(cleanup, rc = PTR_ERR(filp));
1239 cleanup_phase = 2; /* filp */
1241 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1243 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1248 last_group = FID_SEQ_OST_MDT0;
1250 CWARN("%s: initialize groups [%u,%u]\n", obd->obd_name,
1251 FID_SEQ_OST_MDT0, last_group);
1252 filter->fo_committed_group = last_group;
1253 rc = filter_read_groups(obd, last_group, 1);
1257 filp_close(filp, 0);
1261 switch (cleanup_phase) {
1263 filp_close(filp, 0);
1265 filter_cleanup_groups(obd);
1266 f_dput(filter->fo_dentry_O);
1267 filter->fo_dentry_O = NULL;
1275 /* setup the object store with correct subdirectories */
1276 static int filter_prep(struct obd_device *obd)
1278 struct lvfs_run_ctxt saved;
1279 struct filter_obd *filter = &obd->u.filter;
1281 struct inode *inode;
1285 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1286 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1287 if (!file || IS_ERR(file)) {
1289 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1293 obd->u.obt.obt_rcvd_filp = file;
1294 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1295 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1296 file->f_dentry->d_inode->i_mode);
1297 GOTO(err_filp, rc = -ENOENT);
1300 inode = file->f_dentry->d_parent->d_inode;
1301 /* We use i_op->unlink directly in filter_vfs_unlink() */
1302 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1303 CERROR("%s: filesystem does not support create/unlink ops\n",
1305 GOTO(err_filp, rc = -EOPNOTSUPP);
1308 rc = filter_init_server_data(obd, file);
1310 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1313 LASSERT(obd->u.obt.obt_lut);
1314 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1316 /* open and create health check io file*/
1317 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1320 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1322 GOTO(err_server_data, rc);
1324 filter->fo_obt.obt_health_check_filp = file;
1325 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1326 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1327 file->f_dentry->d_inode->i_mode);
1328 GOTO(err_health_check, rc = -ENOENT);
1330 rc = lvfs_check_io_health(obd, file);
1332 GOTO(err_health_check, rc);
1334 rc = filter_prep_groups(obd);
1336 GOTO(err_health_check, rc);
1338 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1343 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1344 CERROR("can't close %s after error\n", HEALTH_CHECK);
1345 filter->fo_obt.obt_health_check_filp = NULL;
1347 target_recovery_fini(obd);
1348 filter_free_server_data(&obd->u.obt);
1350 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1351 CERROR("can't close %s after error\n", LAST_RCVD);
1352 obd->u.obt.obt_rcvd_filp = NULL;
1356 /* cleanup the filter: write last used object id to status file */
1357 static void filter_post(struct obd_device *obd)
1359 struct lvfs_run_ctxt saved;
1360 struct filter_obd *filter = &obd->u.filter;
1363 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1364 * best to start a transaction with h_sync, because we removed this
1367 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1368 rc = filter_update_server_data(obd);
1370 CERROR("error writing server data: rc = %d\n", rc);
1372 for (i = 0; i < filter->fo_group_count; i++) {
1373 rc = filter_update_last_objid(obd, i,
1374 (i == filter->fo_group_count - 1));
1376 CERROR("error writing group %d lastobjid: rc = %d\n",
1380 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1381 obd->u.obt.obt_rcvd_filp = NULL;
1383 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1385 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1386 filter->fo_obt.obt_health_check_filp = NULL;
1388 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1390 filter_cleanup_groups(obd);
1391 filter_free_server_data(&obd->u.obt);
1392 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1394 filter_free_capa_keys(filter);
1395 cleanup_capa_hash(filter->fo_capa_hash);
1398 static void filter_set_last_id(struct filter_obd *filter,
1399 obd_id id, obd_seq group)
1401 LASSERT(group <= filter->fo_group_count);
1403 cfs_spin_lock(&filter->fo_objidlock);
1404 filter->fo_last_objids[group] = id;
1405 cfs_spin_unlock(&filter->fo_objidlock);
1408 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1411 LASSERT(group <= filter->fo_group_count);
1412 LASSERT(filter->fo_last_objids != NULL);
1414 /* FIXME: object groups */
1415 cfs_spin_lock(&filter->fo_objidlock);
1416 id = filter->fo_last_objids[group];
1417 cfs_spin_unlock(&filter->fo_objidlock);
1421 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1423 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1427 /* We never dget the object parent, so DON'T dput it either */
1428 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1430 struct filter_obd *filter = &obd->u.filter;
1431 struct filter_subdirs *subdirs;
1433 if (group >= filter->fo_group_count) /* FIXME: object groups */
1434 return ERR_PTR(-EBADF);
1436 if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1437 return filter->fo_dentry_O_groups[group];
1439 subdirs = &filter->fo_dentry_O_sub[group];
1440 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1443 /* We never dget the object parent, so DON'T dput it either */
1444 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1447 unsigned long now = jiffies;
1448 struct dentry *dparent = filter_parent(obd, group, objid);
1451 if (IS_ERR(dparent))
1453 if (dparent == NULL)
1454 return ERR_PTR(-ENOENT);
1456 rc = filter_lock_dentry(obd, dparent);
1457 fsfilt_check_slow(obd, now, "parent lock");
1458 return rc ? ERR_PTR(rc) : dparent;
1461 /* We never dget the object parent, so DON'T dput it either */
1462 static void filter_parent_unlock(struct dentry *dparent)
1464 UNLOCK_INODE_MUTEX(dparent->d_inode);
1467 /* How to get files, dentries, inodes from object id's.
1469 * If dir_dentry is passed, the caller has already locked the parent
1470 * appropriately for this operation (normally a write lock). If
1471 * dir_dentry is NULL, we do a read lock while we do the lookup to
1472 * avoid races with create/destroy and such changing the directory
1473 * internal to the filesystem code. */
1474 struct dentry *filter_fid2dentry(struct obd_device *obd,
1475 struct dentry *dir_dentry,
1476 obd_seq group, obd_id id)
1478 struct dentry *dparent = dir_dentry;
1479 struct dentry *dchild;
1484 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1485 obd->u.filter.fo_destroys_in_progress == 0) {
1486 /* don't fail lookups for orphan recovery, it causes
1487 * later LBUGs when objects still exist during precreate */
1488 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1489 RETURN(ERR_PTR(-ENOENT));
1492 CERROR("fatal: invalid object id 0\n");
1493 RETURN(ERR_PTR(-ESTALE));
1496 len = sprintf(name, LPU64, id);
1497 if (dir_dentry == NULL) {
1498 dparent = filter_parent_lock(obd, group, id);
1499 if (IS_ERR(dparent)) {
1500 CERROR("%s: error getting object "POSTID
1501 " parent: rc %ld\n", obd->obd_name,
1502 id, group, PTR_ERR(dparent));
1506 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1507 dparent->d_name.len, dparent->d_name.name, name);
1508 /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1509 dchild = lookup_one_len(name, dparent, len);
1510 if (dir_dentry == NULL)
1511 filter_parent_unlock(dparent);
1512 if (IS_ERR(dchild)) {
1513 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1514 obd->obd_name, id, group, PTR_ERR(dchild));
1518 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1519 CERROR("%s: got bad object "LPU64" inode %lu\n",
1520 obd->obd_name, id, dchild->d_inode->i_ino);
1522 RETURN(ERR_PTR(-ENOENT));
1525 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1526 name, dchild, atomic_read(&dchild->d_count));
1528 LASSERT(atomic_read(&dchild->d_count) > 0);
1533 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1534 obd_id group, struct lustre_handle *lockh)
1536 int flags = LDLM_AST_DISCARD_DATA, rc;
1537 struct ldlm_res_id res_id;
1538 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1541 osc_build_res_name(objid, group, &res_id);
1542 /* Tell the clients that the object is gone now and that they should
1543 * throw away any cached pages. */
1544 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1545 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1546 ldlm_completion_ast, NULL, NULL, 0, NULL,
1553 static void filter_fini_destroy(struct obd_device *obd,
1554 struct lustre_handle *lockh)
1556 if (lustre_handle_is_used(lockh))
1557 ldlm_lock_decref(lockh, LCK_PW);
1560 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1561 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1562 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1563 * reverse. Caller must take i_sem before starting the transaction and we
1564 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1565 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1566 struct vfsmount *mnt)
1571 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1572 * itself we already hold dir->i_mutex for child create/unlink ops */
1573 LASSERT(dentry->d_inode != NULL);
1574 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1575 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1579 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1580 GOTO(out, rc = -ENOENT);
1582 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1587 GOTO(out, rc = -EPERM);
1589 /* check_sticky() */
1590 if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1591 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1592 IS_IMMUTABLE(dentry->d_inode))
1593 GOTO(out, rc = -EPERM);
1595 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1596 * that was done because of journal_start (which is already done
1597 * here) or some other ordering issue. */
1598 ll_vfs_dq_init(dir);
1600 rc = ll_security_inode_unlink(dir, dentry, mnt);
1604 rc = dir->i_op->unlink(dir, dentry);
1606 /* need to drop i_mutex before we lose inode reference */
1607 UNLOCK_INODE_MUTEX(dentry->d_inode);
1614 /* Caller must hold LCK_PW on parent and push us into kernel context.
1615 * Caller must hold child i_mutex, we drop it always.
1616 * Caller is also required to ensure that dchild->d_inode exists. */
1617 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1618 obd_seq group, struct dentry *dparent,
1619 struct dentry *dchild)
1621 struct inode *inode = dchild->d_inode;
1624 /* There should be 2 references to the inode:
1625 * 1) taken by filter_prepare_destroy
1626 * 2) taken by filter_destroy */
1627 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1628 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1629 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1630 (unsigned long)inode->i_nlink,
1631 atomic_read(&inode->i_count));
1634 rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1636 CERROR("error unlinking objid %.*s: rc %d\n",
1637 dchild->d_name.len, dchild->d_name.name, rc);
1641 struct filter_intent_args {
1642 struct ldlm_lock **victim;
1647 static enum interval_iter filter_intent_cb(struct interval_node *n,
1650 struct ldlm_interval *node = (struct ldlm_interval *)n;
1651 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1652 __u64 size = arg->size;
1653 struct ldlm_lock **v = arg->victim;
1654 struct ldlm_lock *lck;
1656 /* If the interval is lower than the current file size,
1658 if (interval_high(n) <= size)
1659 return INTERVAL_ITER_STOP;
1661 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1662 /* Don't send glimpse ASTs to liblustre clients.
1663 * They aren't listening for them, and they do
1664 * entirely synchronous I/O anyways. */
1665 if (lck->l_export == NULL ||
1666 lck->l_export->exp_libclient == 1)
1669 if (*arg->liblustre)
1670 *arg->liblustre = 0;
1673 *v = LDLM_LOCK_GET(lck);
1674 } else if ((*v)->l_policy_data.l_extent.start <
1675 lck->l_policy_data.l_extent.start) {
1676 LDLM_LOCK_RELEASE(*v);
1677 *v = LDLM_LOCK_GET(lck);
1680 /* the same policy group - every lock has the
1681 * same extent, so needn't do it any more */
1685 return INTERVAL_ITER_CONT;
1688 static int filter_intent_policy(struct ldlm_namespace *ns,
1689 struct ldlm_lock **lockp, void *req_cookie,
1690 ldlm_mode_t mode, int flags, void *data)
1692 CFS_LIST_HEAD(rpc_list);
1693 struct ptlrpc_request *req = req_cookie;
1694 struct ldlm_lock *lock = *lockp, *l = NULL;
1695 struct ldlm_resource *res = lock->l_resource;
1696 ldlm_processing_policy policy;
1697 struct ost_lvb *res_lvb, *reply_lvb;
1698 struct ldlm_reply *rep;
1700 int idx, rc, tmpflags = 0, only_liblustre = 1;
1701 struct ldlm_interval_tree *tree;
1702 struct filter_intent_args arg;
1703 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1704 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1705 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1708 policy = ldlm_get_processing_policy(res);
1709 LASSERT(policy != NULL);
1710 LASSERT(req != NULL);
1712 rc = lustre_pack_reply(req, 3, repsize, NULL);
1714 RETURN(req->rq_status = rc);
1716 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1717 LASSERT(rep != NULL);
1719 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1720 sizeof(*reply_lvb));
1721 LASSERT(reply_lvb != NULL);
1723 //fixup_handle_for_resent_req(req, lock, &lockh);
1725 /* Call the extent policy function to see if our request can be
1726 * granted, or is blocked.
1727 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1728 * lock, and should not be granted if the lock will be blocked.
1731 LASSERT(ns == ldlm_res_to_ns(res));
1733 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1734 check_res_locked(res);
1736 /* FIXME: we should change the policy function slightly, to not make
1737 * this list at all, since we just turn around and free it */
1738 while (!cfs_list_empty(&rpc_list)) {
1739 struct ldlm_lock *wlock =
1740 cfs_list_entry(rpc_list.next, struct ldlm_lock,
1742 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1743 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1744 lock->l_flags &= ~LDLM_FL_CP_REQD;
1745 cfs_list_del_init(&wlock->l_cp_ast);
1746 LDLM_LOCK_RELEASE(wlock);
1749 /* The lock met with no resistance; we're finished. */
1750 if (rc == LDLM_ITER_CONTINUE) {
1751 /* do not grant locks to the liblustre clients: they cannot
1752 * handle ASTs robustly. We need to do this while still
1753 * holding lr_lock to avoid the lock remaining on the res_link
1754 * list (and potentially being added to l_pending_list by an
1755 * AST) when we are going to drop this lock ASAP. */
1756 if (lock->l_export->exp_libclient ||
1757 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1758 ldlm_resource_unlink_lock(lock);
1759 err = ELDLM_LOCK_ABORTED;
1761 err = ELDLM_LOCK_REPLACED;
1767 /* Do not grant any lock, but instead send GL callbacks. The extent
1768 * policy nicely created a list of all PW locks for us. We will choose
1769 * the highest of those which are larger than the size in the LVB, if
1770 * any, and perform a glimpse callback. */
1771 res_lvb = res->lr_lvb_data;
1772 LASSERT(res_lvb != NULL);
1773 *reply_lvb = *res_lvb;
1776 * lr_lock guarantees that no new locks are granted, and,
1777 * therefore, that res->lr_lvb_data cannot increase beyond the
1778 * end of already granted lock. As a result, it is safe to
1779 * check against "stale" reply_lvb->lvb_size value without
1782 arg.size = reply_lvb->lvb_size;
1784 arg.liblustre = &only_liblustre;
1785 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1786 tree = &res->lr_itree[idx];
1787 if (tree->lit_mode == LCK_PR)
1790 interval_iterate_reverse(tree->lit_root,
1791 filter_intent_cb, &arg);
1795 /* There were no PW locks beyond the size in the LVB; finished. */
1797 if (only_liblustre) {
1798 /* If we discovered a liblustre client with a PW lock,
1799 * however, the LVB may be out of date! The LVB is
1800 * updated only on glimpse (which we don't do for
1801 * liblustre clients) and cancel (which the client
1802 * obviously has not yet done). So if it has written
1803 * data but kept the lock, the LVB is stale and needs
1804 * to be updated from disk.
1806 * Of course, this will all disappear when we switch to
1807 * taking liblustre locks on the OST. */
1808 ldlm_res_lvbo_update(res, NULL, 1);
1810 RETURN(ELDLM_LOCK_ABORTED);
1814 * This check is for lock taken in filter_prepare_destroy() that does
1815 * not have l_glimpse_ast set. So the logic is: if there is a lock
1816 * with no l_glimpse_ast set, this object is being destroyed already.
1818 * Hence, if you are grabbing DLM locks on the server, always set
1819 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1821 if (l->l_glimpse_ast == NULL) {
1822 /* We are racing with unlink(); just return -ENOENT */
1823 rep->lock_policy_res1 = -ENOENT;
1827 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1828 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1831 *reply_lvb = *res_lvb;
1835 LDLM_LOCK_RELEASE(l);
1837 RETURN(ELDLM_LOCK_ABORTED);
1841 * per-obd_device iobuf pool.
1843 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1844 * allocations in write-path has to be minimized (see bug 5137).
1846 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1847 * OST threads (see ost_thread_{init,done}()).
1849 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1850 * at the OST layer there are only (potentially) multiple obd_device of type
1851 * unknown at the time of OST thread creation.
1853 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1854 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1855 * it's very own iobuf.
1859 * filter_kiobuf_pool_init()
1861 * filter_kiobuf_pool_done()
1863 * filter_iobuf_get()
1865 * operate on this array. They are "generic" in a sense that they don't depend
1866 * on actual type of iobuf's (the latter depending on Linux kernel version).
1870 * destroy pool created by filter_iobuf_pool_init
1872 static void filter_iobuf_pool_done(struct filter_obd *filter)
1874 struct filter_iobuf **pool;
1879 pool = filter->fo_iobuf_pool;
1881 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1882 if (pool[i] != NULL)
1883 filter_free_iobuf(pool[i]);
1885 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1886 filter->fo_iobuf_pool = NULL;
1891 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1893 struct filter_obd *filter = &obd->u.filter;
1894 struct sptlrpc_rule_set tmp_rset;
1897 sptlrpc_rule_set_init(&tmp_rset);
1898 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1900 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1905 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1907 cfs_write_lock(&filter->fo_sptlrpc_lock);
1908 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1909 filter->fo_sptlrpc_rset = tmp_rset;
1910 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1916 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1918 static int filter_iobuf_pool_init(struct filter_obd *filter)
1925 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1927 if (filter->fo_iobuf_pool == NULL)
1930 filter->fo_iobuf_count = OSS_THREADS_MAX;
1935 /* Return iobuf allocated for @thread_id. We don't know in advance how
1936 * many threads there will be so we allocate a large empty array and only
1937 * fill in those slots that are actually in use.
1938 * If we haven't allocated a pool entry for this thread before, do so now. */
1939 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1941 int thread_id = (oti && oti->oti_thread) ?
1942 oti->oti_thread->t_id : -1;
1943 struct filter_iobuf *pool = NULL;
1944 struct filter_iobuf **pool_place = NULL;
1946 if (thread_id >= 0) {
1947 LASSERT(thread_id < filter->fo_iobuf_count);
1948 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1951 if (unlikely(pool == NULL)) {
1952 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1953 PTLRPC_MAX_BRW_PAGES);
1954 if (pool_place != NULL)
1961 /* mount the file system (secretly). lustre_cfg parameters are:
1964 * 3 = flags: failover=f, failout=n
1967 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1970 struct filter_obd *filter = &obd->u.filter;
1971 struct vfsmount *mnt;
1972 struct lustre_mount_info *lmi;
1973 struct obd_uuid uuid;
1977 struct request_queue *q;
1981 if (lcfg->lcfg_bufcount < 3 ||
1982 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1983 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1986 lmi = server_get_mount(obd->obd_name);
1988 /* We already mounted in lustre_fill_super.
1989 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1990 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1992 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1994 /* gets recovery timeouts from mount data */
1995 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1996 obd->obd_recovery_timeout =
1997 lsi->lsi_lmd->lmd_recovery_time_soft;
1998 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1999 obd->obd_recovery_time_hard =
2000 lsi->lsi_lmd->lmd_recovery_time_hard;
2002 /* old path - used by lctl */
2003 CERROR("Using old MDS mount method\n");
2004 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
2005 MS_NOATIME|MS_NODIRATIME,
2006 lustre_cfg_string(lcfg, 1), option);
2009 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2010 lustre_cfg_string(lcfg, 1), rc);
2014 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2016 if (IS_ERR(obd->obd_fsops))
2017 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2019 rc = filter_iobuf_pool_init(filter);
2023 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2024 CERROR("%s: Underlying device is marked as read-only. "
2025 "Setup failed\n", obd->obd_name);
2026 GOTO(err_ops, rc = -EROFS);
2029 /* failover is the default */
2030 obd->obd_replayable = 1;
2032 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2033 str = lustre_cfg_string(lcfg, 3);
2034 if (strchr(str, 'n')) {
2035 CWARN("%s: recovery disabled\n", obd->obd_name);
2036 obd->obd_replayable = 0;
2040 obd->u.obt.obt_vfsmnt = mnt;
2041 obd->u.obt.obt_sb = mnt->mnt_sb;
2042 obd->u.obt.obt_magic = OBT_MAGIC;
2043 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2044 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2046 fsfilt_setup(obd, obd->u.obt.obt_sb);
2048 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2049 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2050 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2051 obd->obd_lvfs_ctxt.fs = get_ds();
2052 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2054 cfs_init_mutex(&filter->fo_init_lock);
2055 filter->fo_committed_group = 0;
2056 filter->fo_destroys_in_progress = 0;
2057 for (i = 0; i < 32; i++)
2058 cfs_sema_init(&filter->fo_create_locks[i], 1);
2060 cfs_spin_lock_init(&filter->fo_objidlock);
2061 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2062 cfs_sema_init(&filter->fo_alloc_lock, 1);
2063 init_brw_stats(&filter->fo_filter_stats);
2064 cfs_spin_lock_init(&filter->fo_flags_lock);
2065 filter->fo_read_cache = 1; /* enable read-only cache by default */
2066 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2067 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2068 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2069 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2070 filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2071 filter_slc_set(filter); /* initialize sync on lock cancel */
2073 rc = filter_prep(obd);
2077 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2078 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2080 filter->fo_fl_oss_capa = 1;
2082 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2083 filter->fo_capa_hash = init_capa_hash();
2084 if (filter->fo_capa_hash == NULL)
2085 GOTO(err_post, rc = -ENOMEM);
2087 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2088 obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2089 LDLM_NAMESPACE_SERVER,
2090 LDLM_NAMESPACE_GREEDY,
2092 if (obd->obd_namespace == NULL)
2093 GOTO(err_post, rc = -ENOMEM);
2094 obd->obd_namespace->ns_lvbp = obd;
2095 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2096 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2098 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2099 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2101 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2103 CERROR("failed to setup llogging subsystems\n");
2107 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2108 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2109 /* do this after llog being initialized */
2110 filter_adapt_sptlrpc_conf(obd, 1);
2112 rc = lquota_setup(filter_quota_interface_ref, obd);
2116 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2117 if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2118 queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2119 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2120 "for larger I/O requests: max_sectors = %u "
2121 "could be up to max_hw_sectors=%u\n",
2122 obd->obd_name, mnt->mnt_sb->s_id,
2123 queue_max_sectors(q), queue_max_hw_sectors(q));
2125 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2126 if (uuid_ptr != NULL) {
2127 class_uuid_unparse(uuid_ptr, &uuid);
2133 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2134 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2135 obd->obd_name, label ?: str, lmi ? "on " : "",
2136 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2137 obd->obd_replayable ? "enabled" : "disabled");
2139 if (obd->obd_recovering)
2140 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2141 "or until %d client%s reconnect%s\n",
2143 obd->obd_recovery_timeout / 60,
2144 obd->obd_recovery_timeout % 60,
2145 obd->obd_max_recoverable_clients,
2146 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2147 (obd->obd_max_recoverable_clients == 1) ? "s": "");
2155 fsfilt_put_ops(obd->obd_fsops);
2156 filter_iobuf_pool_done(filter);
2158 server_put_mount(obd->obd_name, mnt);
2159 obd->u.obt.obt_sb = 0;
2163 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2165 struct lprocfs_static_vars lvars;
2166 cfs_proc_dir_entry_t *entry;
2172 CLASSERT(offsetof(struct obd_device, u.obt) ==
2173 offsetof(struct obd_device, u.filter.fo_obt));
2175 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2178 /* lprocfs must be setup before the filter so state can be safely added
2179 * to /proc incrementally as the filter is setup */
2180 lprocfs_filter_init_vars(&lvars);
2181 rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2183 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2188 rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2190 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2192 GOTO(obd_cleanup, rc);
2195 /* Init obdfilter private stats here */
2196 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2197 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2198 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2199 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2200 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2201 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2202 "get_page", "usec");
2203 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2204 LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2205 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2206 LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2207 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2208 LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2209 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2210 LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2212 rc = lproc_filter_attach_seqstat(obd);
2214 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2215 GOTO(free_obd_stats, rc);
2218 entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2219 if (IS_ERR(entry)) {
2220 rc = PTR_ERR(entry);
2221 CERROR("%s: error %d setting up lprocfs for %s\n",
2222 obd->obd_name, rc, "exports");
2223 GOTO(free_obd_stats, rc);
2225 obd->obd_proc_exports_entry = entry;
2227 entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2228 lprocfs_nid_stats_clear_read,
2229 lprocfs_nid_stats_clear_write, obd, NULL);
2230 if (IS_ERR(entry)) {
2231 rc = PTR_ERR(entry);
2232 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2234 GOTO(free_obd_stats, rc);
2237 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2238 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2240 GOTO(remove_entry_clear, rc = -ENOMEM);
2241 addr = (unsigned long)cfs_page_address(page);
2242 clear_page((void *)addr);
2243 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2244 LUSTRE_CFG_BUFLEN(lcfg, 4));
2245 rc = filter_common_setup(obd, lcfg, (void *)addr);
2246 OBD_PAGE_FREE(page);
2248 CERROR("%s: filter_common_setup failed: %d.\n",
2250 GOTO(remove_entry_clear, rc);
2256 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2258 lprocfs_free_obd_stats(obd);
2260 lprocfs_obd_cleanup(obd);
2264 static struct llog_operations filter_mds_ost_repl_logops;
2266 static struct llog_operations filter_size_orig_logops = {
2267 .lop_setup = llog_obd_origin_setup,
2268 .lop_cleanup = llog_obd_origin_cleanup,
2269 .lop_add = llog_obd_origin_add
2272 static int filter_olg_fini(struct obd_llog_group *olg)
2274 struct llog_ctxt *ctxt;
2275 int rc = 0, rc2 = 0;
2278 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2280 rc = llog_cleanup(ctxt);
2282 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2284 rc2 = llog_cleanup(ctxt);
2289 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2291 rc2 = llog_cleanup(ctxt);
2300 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2301 struct obd_device *tgt)
2306 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2307 &filter_mds_ost_repl_logops);
2311 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2312 &filter_size_orig_logops);
2318 filter_olg_fini(olg);
2323 * Init the default olg, which is embeded in the obd_device, for filter.
2326 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2327 struct obd_device *tgt)
2329 struct filter_obd *filter = &obd->u.filter;
2330 struct llog_ctxt *ctxt;
2334 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2335 if (!filter->fo_lcm)
2338 filter_mds_ost_repl_logops = llog_client_ops;
2339 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2340 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2341 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2343 rc = filter_olg_init(obd, olg, tgt);
2345 GOTO(cleanup_lcm, rc);
2347 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2350 GOTO(cleanup_olg, rc);
2352 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2354 CERROR("Can't get ctxt for %p:%x\n", olg,
2355 LLOG_MDS_OST_REPL_CTXT);
2356 GOTO(cleanup_olg, rc = -ENODEV);
2358 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2359 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2360 llog_ctxt_put(ctxt);
2364 filter_olg_fini(olg);
2366 llog_recov_thread_fini(filter->fo_lcm, 1);
2367 filter->fo_lcm = NULL;
2372 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2373 struct obd_device *tgt, int *index)
2375 struct filter_obd *filter = &obd->u.filter;
2376 struct llog_ctxt *ctxt;
2380 LASSERT(olg != NULL);
2381 if (olg == &obd->obd_olg)
2382 return filter_default_olg_init(obd, olg, tgt);
2384 LASSERT(filter->fo_lcm != NULL);
2385 rc = filter_olg_init(obd, olg, tgt);
2388 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2390 CERROR("Can't get ctxt for %p:%x\n", olg,
2391 LLOG_MDS_OST_REPL_CTXT);
2392 filter_olg_fini(olg);
2395 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2396 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2397 llog_ctxt_put(ctxt);
2401 static int filter_llog_finish(struct obd_device *obd, int count)
2403 struct filter_obd *filter = &obd->u.filter;
2404 struct llog_ctxt *ctxt;
2407 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2410 * Make sure that no cached llcds left in recov_thread.
2411 * We actually do sync in disconnect time, but disconnect
2412 * may not come being marked rq_no_resend = 1.
2414 llog_sync(ctxt, NULL);
2417 * Balance class_import_get() in llog_receptor_accept().
2418 * This is safe to do, as llog is already synchronized
2419 * and its import may go.
2421 cfs_mutex_down(&ctxt->loc_sem);
2422 if (ctxt->loc_imp) {
2423 class_import_put(ctxt->loc_imp);
2424 ctxt->loc_imp = NULL;
2426 cfs_mutex_up(&ctxt->loc_sem);
2427 llog_ctxt_put(ctxt);
2430 if (filter->fo_lcm) {
2431 cfs_mutex_down(&ctxt->loc_sem);
2432 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2433 filter->fo_lcm = NULL;
2434 cfs_mutex_up(&ctxt->loc_sem);
2436 RETURN(filter_olg_fini(&obd->obd_olg));
2440 * Find the group llog according to group index in the llog group list.
2442 static struct obd_llog_group *
2443 filter_find_olg_internal(struct filter_obd *filter, int group)
2445 struct obd_llog_group *olg;
2447 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2448 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2449 if (olg->olg_seq == group)
2456 * Find the group llog according to group index on the filter
2458 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2460 struct obd_llog_group *olg = NULL;
2461 struct filter_obd *filter;
2463 filter = &obd->u.filter;
2465 if (group == FID_SEQ_LLOG)
2466 RETURN(&obd->obd_olg);
2468 cfs_spin_lock(&filter->fo_llog_list_lock);
2469 olg = filter_find_olg_internal(filter, group);
2470 cfs_spin_unlock(&filter->fo_llog_list_lock);
2475 * Find the llog_group of the filter according to the group. If it can not
2476 * find, create the llog_group, which only happens when mds is being synced
2479 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2481 struct obd_llog_group *olg = NULL, *olg_new = NULL;
2482 struct filter_obd *filter;
2485 filter = &obd->u.filter;
2487 if (group == FID_SEQ_LLOG)
2488 RETURN(&obd->obd_olg);
2490 OBD_ALLOC_PTR(olg_new);
2491 if (olg_new == NULL)
2492 RETURN(ERR_PTR(-ENOMEM));
2494 cfs_spin_lock(&filter->fo_llog_list_lock);
2495 olg = filter_find_olg_internal(filter, group);
2497 if (olg->olg_initializing) {
2498 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2500 GOTO(out_unlock, olg);
2503 /* set as the newly allocated one */
2508 llog_group_init(olg, group);
2509 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2510 olg->olg_initializing = 1;
2511 cfs_spin_unlock(&filter->fo_llog_list_lock);
2513 rc = obd_llog_init(obd, olg, obd, NULL);
2515 cfs_spin_lock(&filter->fo_llog_list_lock);
2516 cfs_list_del(&olg->olg_list);
2517 cfs_spin_unlock(&filter->fo_llog_list_lock);
2519 GOTO(out, olg = ERR_PTR(-ENOMEM));
2521 cfs_spin_lock(&filter->fo_llog_list_lock);
2522 olg->olg_initializing = 0;
2523 cfs_spin_unlock(&filter->fo_llog_list_lock);
2524 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2525 obd->obd_name, group, olg);
2530 cfs_spin_unlock(&filter->fo_llog_list_lock);
2532 OBD_FREE_PTR(olg_new);
2536 static int filter_llog_connect(struct obd_export *exp,
2537 struct llogd_conn_body *body)
2539 struct obd_device *obd = exp->exp_obd;
2540 struct llog_ctxt *ctxt;
2541 struct obd_llog_group *olg;
2545 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2546 obd->obd_name, body->lgdc_logid.lgl_oid,
2547 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2549 olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2551 CERROR(" %s: can not find olg of group %d\n",
2552 obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2555 llog_group_set_export(olg, exp);
2557 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2558 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2559 body->lgdc_ctxt_idx);
2561 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2562 obd->obd_name, body->lgdc_logid.lgl_oid,
2563 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2565 cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2566 obd->u.filter.fo_mds_ost_sync = 1;
2567 cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2568 rc = llog_connect(ctxt, &body->lgdc_logid,
2569 &body->lgdc_gen, NULL);
2570 llog_ctxt_put(ctxt);
2572 CERROR("failed to connect rc %d idx %d\n", rc,
2573 body->lgdc_ctxt_idx);
2578 static int filter_llog_preclean(struct obd_device *obd)
2580 struct obd_llog_group *olg, *tmp;
2581 struct filter_obd *filter;
2582 cfs_list_t remove_list;
2586 rc = obd_llog_finish(obd, 0);
2588 CERROR("failed to cleanup llogging subsystem\n");
2590 filter = &obd->u.filter;
2591 CFS_INIT_LIST_HEAD(&remove_list);
2593 cfs_spin_lock(&filter->fo_llog_list_lock);
2594 while (!cfs_list_empty(&filter->fo_llog_list)) {
2595 olg = cfs_list_entry(filter->fo_llog_list.next,
2596 struct obd_llog_group, olg_list);
2597 cfs_list_del(&olg->olg_list);
2598 cfs_list_add(&olg->olg_list, &remove_list);
2600 cfs_spin_unlock(&filter->fo_llog_list_lock);
2602 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2603 cfs_list_del_init(&olg->olg_list);
2604 rc = filter_olg_fini(olg);
2606 CERROR("failed to cleanup llogging subsystem for %u\n",
2614 static int filter_precleanup(struct obd_device *obd,
2615 enum obd_cleanup_stage stage)
2621 case OBD_CLEANUP_EARLY:
2623 case OBD_CLEANUP_EXPORTS:
2624 /* Stop recovery before namespace cleanup. */
2625 target_recovery_fini(obd);
2626 rc = filter_llog_preclean(obd);
2632 static int filter_cleanup(struct obd_device *obd)
2634 struct filter_obd *filter = &obd->u.filter;
2638 LCONSOLE_WARN("%s: shutting down for failover; client state "
2639 "will be preserved.\n", obd->obd_name);
2641 obd_exports_barrier(obd);
2642 obd_zombie_barrier();
2644 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2645 lprocfs_free_per_client_stats(obd);
2646 lprocfs_free_obd_stats(obd);
2647 lprocfs_obd_cleanup(obd);
2648 lquota_cleanup(filter_quota_interface_ref, obd);
2650 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2651 obd->obd_namespace = NULL;
2653 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2655 if (obd->u.obt.obt_sb == NULL)
2660 ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2661 shrink_dcache_sb(obd->u.obt.obt_sb);
2663 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2664 obd->u.obt.obt_sb = NULL;
2666 fsfilt_put_ops(obd->obd_fsops);
2668 filter_iobuf_pool_done(filter);
2670 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2675 static int filter_connect_internal(struct obd_export *exp,
2676 struct obd_connect_data *data,
2679 struct filter_export_data *fed = &exp->exp_filter_data;
2684 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2685 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2686 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2687 data->ocd_connect_flags, data->ocd_version,
2688 data->ocd_grant, data->ocd_index);
2690 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2691 CWARN("!!! This export (nid %s) used object group %d "
2692 "earlier; now it's trying to use group %d! This could "
2693 "be a bug in the MDS. Please report to "
2694 "http://bugs.whamcloud.com/\n",
2695 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2698 fed->fed_group = data->ocd_group;
2700 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2701 exp->exp_connect_flags = data->ocd_connect_flags;
2702 data->ocd_version = LUSTRE_VERSION_CODE;
2704 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2705 if (data->ocd_connect_flags & OBD_CONNECT_MDS)
2706 CWARN("%s: Received MDS connection for group %u\n",
2707 exp->exp_obd->obd_name, data->ocd_group);
2708 else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
2711 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2712 struct filter_obd *filter = &exp->exp_obd->u.filter;
2713 obd_size left, want;
2715 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2716 left = filter_grant_space_left(exp);
2717 want = data->ocd_grant;
2718 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2719 data->ocd_grant = fed->fed_grant;
2720 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2722 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2723 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2724 exp->exp_client_uuid.uuid, exp,
2725 data->ocd_grant, want, left);
2727 filter->fo_tot_granted_clients ++;
2730 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2731 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2732 int index = le32_to_cpu(lsd->lsd_ost_index);
2734 if (!(lsd->lsd_feature_compat &
2735 cpu_to_le32(OBD_COMPAT_OST))) {
2736 /* this will only happen on the first connect */
2737 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2738 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2739 /* sync is not needed here as filter_client_add will
2740 * set exp_need_sync flag */
2741 filter_update_server_data(exp->exp_obd);
2742 } else if (index != data->ocd_index) {
2743 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2744 " %u doesn't match actual OST index"
2745 " %u in last_rcvd file, bad "
2747 obd_export_nid2str(exp), index,
2751 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2752 * FIXME: We don't strictly need the COMPAT flag for that,
2753 * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2754 * FIXME: We needed it for the index, as index 0 is valid. */
2757 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2758 data->ocd_brw_size = 65536;
2759 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2760 data->ocd_brw_size = min(data->ocd_brw_size,
2761 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2762 if (data->ocd_brw_size == 0) {
2763 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2764 " ocd_version: %x ocd_grant: %d ocd_index: %u "
2765 "ocd_brw_size is unexpectedly zero, "
2766 "network data corruption?"
2767 "Refusing connection of this client\n",
2768 exp->exp_obd->obd_name,
2769 exp->exp_client_uuid.uuid,
2770 exp, data->ocd_connect_flags, data->ocd_version,
2771 data->ocd_grant, data->ocd_index);
2776 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2777 __u32 cksum_types = data->ocd_cksum_types;
2779 /* The client set in ocd_cksum_types the checksum types it
2780 * supports. We have to mask off the algorithms that we don't
2782 if (cksum_types & OBD_CKSUM_ALL)
2783 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2785 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2787 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2788 "%x\n", exp->exp_obd->obd_name,
2789 obd_export_nid2str(exp), cksum_types,
2790 data->ocd_cksum_types);
2792 /* This client does not support OBD_CONNECT_CKSUM
2793 * fall back to CRC32 */
2794 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2795 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2796 exp->exp_obd->obd_name,
2797 obd_export_nid2str(exp));
2800 if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2801 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2806 static int filter_reconnect(const struct lu_env *env,
2807 struct obd_export *exp, struct obd_device *obd,
2808 struct obd_uuid *cluuid,
2809 struct obd_connect_data *data,
2815 if (exp == NULL || obd == NULL || cluuid == NULL)
2818 rc = filter_connect_internal(exp, data, 1);
2820 filter_export_stats_init(obd, exp, localdata);
2825 static int filter_connect(const struct lu_env *env,
2826 struct obd_export **exp, struct obd_device *obd,
2827 struct obd_uuid *cluuid,
2828 struct obd_connect_data *data, void *localdata)
2830 struct lvfs_run_ctxt saved;
2831 struct lustre_handle conn = { 0 };
2832 struct obd_export *lexp;
2836 if (exp == NULL || obd == NULL || cluuid == NULL)
2839 rc = class_connect(&conn, obd, cluuid);
2842 lexp = class_conn2export(&conn);
2843 LASSERT(lexp != NULL);
2845 rc = filter_connect_internal(lexp, data, 0);
2849 filter_export_stats_init(obd, lexp, localdata);
2850 if (obd->obd_replayable) {
2851 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2853 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2854 rc = filter_client_add(obd, lexp, -1);
2859 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2860 rc = filter_read_groups(obd, data->ocd_group, 1);
2861 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2863 CERROR("can't read group %u\n", data->ocd_group);
2871 class_disconnect(lexp);
2880 /* Do extra sanity checks for grant accounting. We do this at connect,
2881 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2882 * always get rid of it or turn it off when we know accounting is good. */
2883 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2885 struct filter_export_data *fed;
2886 struct obd_export *exp;
2887 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2888 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2889 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2891 if (cfs_list_empty(&obd->obd_exports))
2894 /* We don't want to do this for large machines that do lots of
2895 mounts or unmounts. It burns... */
2896 if (obd->obd_num_exports > 100)
2899 cfs_spin_lock(&obd->obd_osfs_lock);
2900 cfs_spin_lock(&obd->obd_dev_lock);
2901 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2903 fed = &exp->exp_filter_data;
2904 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2907 if (maxsize > 0) { /* we may not have done a statfs yet */
2908 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2909 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2910 exp->exp_client_uuid.uuid, exp,
2911 fed->fed_grant, fed->fed_pending, maxsize);
2912 LASSERTF(fed->fed_dirty <= maxsize,
2913 "%s: cli %s/%p %ld > "LPU64"\n", func,
2914 exp->exp_client_uuid.uuid, exp,
2915 fed->fed_dirty, maxsize);
2918 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2919 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2920 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2922 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2923 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2924 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2925 tot_granted += fed->fed_grant + fed->fed_pending;
2926 tot_pending += fed->fed_pending;
2927 tot_dirty += fed->fed_dirty;
2929 fo_tot_granted = obd->u.filter.fo_tot_granted;
2930 fo_tot_pending = obd->u.filter.fo_tot_pending;
2931 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2932 cfs_spin_unlock(&obd->obd_dev_lock);
2933 cfs_spin_unlock(&obd->obd_osfs_lock);
2935 /* Do these assertions outside the spinlocks so we don't kill system */
2936 if (tot_granted != fo_tot_granted)
2937 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2938 func, tot_granted, fo_tot_granted);
2939 if (tot_pending != fo_tot_pending)
2940 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2941 func, tot_pending, fo_tot_pending);
2942 if (tot_dirty != fo_tot_dirty)
2943 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2944 func, tot_dirty, fo_tot_dirty);
2945 if (tot_pending > tot_granted)
2946 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2947 func, tot_pending, tot_granted);
2948 if (tot_granted > maxsize)
2949 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2950 func, tot_granted, maxsize);
2951 if (tot_dirty > maxsize)
2952 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2953 func, tot_dirty, maxsize);
2956 /* Remove this client from the grant accounting totals. We also remove
2957 * the export from the obd device under the osfs and dev locks to ensure
2958 * that the filter_grant_sanity_check() calculations are always valid.
2959 * The client should do something similar when it invalidates its import. */
2960 static void filter_grant_discard(struct obd_export *exp)
2962 struct obd_device *obd = exp->exp_obd;
2963 struct filter_obd *filter = &obd->u.filter;
2964 struct filter_export_data *fed = &exp->exp_filter_data;
2966 cfs_spin_lock(&obd->obd_osfs_lock);
2967 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2968 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2969 obd->obd_name, filter->fo_tot_granted,
2970 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2971 filter->fo_tot_granted -= fed->fed_grant;
2972 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2973 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2974 obd->obd_name, filter->fo_tot_pending,
2975 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2976 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2977 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2978 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2979 obd->obd_name, filter->fo_tot_dirty,
2980 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2981 filter->fo_tot_dirty -= fed->fed_dirty;
2985 cfs_spin_unlock(&obd->obd_osfs_lock);
2988 static int filter_destroy_export(struct obd_export *exp)
2990 struct filter_export_data *fed = &exp->exp_filter_data;
2993 if (fed->fed_pending)
2994 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2995 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2996 exp, fed->fed_pending);
2998 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3000 target_destroy_export(exp);
3002 if (unlikely(exp == exp->exp_obd->obd_self_export))
3005 ldlm_destroy_export(exp);
3006 lut_client_free(exp);
3008 if (!exp->exp_obd->obd_replayable)
3009 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3011 filter_grant_discard(exp);
3012 filter_fmd_cleanup(exp);
3014 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3015 struct filter_obd *filter = &exp->exp_obd->u.filter;
3016 if (filter->fo_tot_granted_clients > 0)
3017 filter->fo_tot_granted_clients --;
3020 if (!(exp->exp_flags & OBD_OPT_FORCE))
3021 filter_grant_sanity_check(exp->exp_obd, __func__);
3026 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3028 struct obd_llog_group *olg_min, *olg;
3029 struct filter_obd *filter;
3030 int worked = -1, group;
3031 struct llog_ctxt *ctxt;
3034 filter = &obd->u.filter;
3036 /* we can't sync log holding spinlock. also, we do not want to get
3037 * into livelock. so we do following: loop over MDS's exports in
3038 * group order and skip already synced llogs -bzzz */
3040 /* look for group with min. number, but > worked */
3043 cfs_spin_lock(&filter->fo_llog_list_lock);
3044 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3045 if (olg->olg_seq <= worked) {
3046 /* this group is already synced */
3049 if (group < olg->olg_seq) {
3050 /* we have group with smaller number to sync */
3053 /* store current minimal group */
3055 group = olg->olg_seq;
3057 cfs_spin_unlock(&filter->fo_llog_list_lock);
3059 if (olg_min == NULL)
3062 worked = olg_min->olg_seq;
3063 if (olg_min->olg_exp &&
3064 (dexp == olg_min->olg_exp || dexp == NULL)) {
3066 ctxt = llog_group_get_ctxt(olg_min,
3067 LLOG_MDS_OST_REPL_CTXT);
3069 err = llog_sync(ctxt, olg_min->olg_exp);
3070 llog_ctxt_put(ctxt);
3072 CERROR("error flushing logs to MDS: "
3077 } while (olg_min != NULL);
3080 /* Also incredibly similar to mds_disconnect */
3081 static int filter_disconnect(struct obd_export *exp)
3083 struct obd_device *obd = exp->exp_obd;
3088 class_export_get(exp);
3090 if (!(exp->exp_flags & OBD_OPT_FORCE))
3091 filter_grant_sanity_check(obd, __func__);
3092 filter_grant_discard(exp);
3094 /* Flush any remaining cancel messages out to the target */
3095 filter_sync_llogs(obd, exp);
3097 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3099 rc = server_disconnect_export(exp);
3101 /* Do not erase record for recoverable client. */
3102 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3103 filter_client_del(exp);
3105 fsfilt_sync(obd, obd->u.obt.obt_sb);
3107 class_export_put(exp);
3111 /* reverse import is changed, sync all cancels */
3112 static void filter_revimp_update(struct obd_export *exp)
3117 class_export_get(exp);
3119 /* flush any remaining cancel messages out to the target */
3120 filter_sync_llogs(exp->exp_obd, exp);
3121 class_export_put(exp);
3125 static int filter_ping(struct obd_export *exp)
3127 filter_fmd_expire(exp);
3131 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3132 const char *what, int quiet)
3134 struct dentry *dchild = NULL;
3136 dchild = filter_fid2dentry(obd, NULL, ostid->oi_seq, ostid->oi_id);
3138 if (IS_ERR(dchild)) {
3139 CERROR("%s error looking up object: "POSTID"\n",
3140 what, ostid->oi_id, ostid->oi_seq);
3144 if (dchild->d_inode == NULL) {
3146 CERROR("%s: %s on non-existent object: "POSTID" \n",
3147 obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3149 RETURN(ERR_PTR(-ENOENT));
3155 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3157 struct dentry *dentry = NULL;
3158 struct obd_device *obd;
3162 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3163 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3167 obd = class_exp2obd(exp);
3169 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3173 dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3175 RETURN(PTR_ERR(dentry));
3177 /* Limit the valid bits in the return data to what we actually use */
3178 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3179 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3185 /* this should be enabled/disabled in condition to enabled/disabled large
3186 * inodes (fast EAs) in backing store FS. */
3187 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3188 void *handle, struct obdo *oa)
3190 struct obd_device *obd = exp->exp_obd;
3194 if (oa->o_valid & OBD_MD_FLFID) {
3195 struct filter_fid ff;
3197 if (!(oa->o_valid & OBD_MD_FLGROUP))
3199 /* packing fid and converting it to LE for storing into EA.
3200 * Here ->o_stripe_idx should be filled by LOV and rest of
3201 * fields - by client. */
3202 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3203 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3204 /* XXX: we are ignoring o_parent_ver here, since this should
3205 * be the same for all objects in this fileset. */
3206 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3207 ff.ff_objid = cpu_to_le64(oa->o_id);
3208 ff.ff_seq = cpu_to_le64(oa->o_seq);
3210 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3211 LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3214 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3216 CERROR("store fid in object failed! rc: %d\n", rc);
3218 CDEBUG(D_HA, "OSS object without fid info!\n");
3224 /* this is called from filter_truncate() until we have filter_punch() */
3225 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3226 struct obdo *oa, struct obd_trans_info *oti)
3228 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3229 struct llog_cookie *fcc = NULL;
3230 struct filter_obd *filter;
3231 int rc, err, sync = 0;
3232 loff_t old_size = 0;
3233 unsigned int ia_valid;
3234 struct inode *inode;
3235 struct page *page = NULL;
3240 LASSERT(dentry != NULL);
3241 LASSERT(!IS_ERR(dentry));
3243 inode = dentry->d_inode;
3244 LASSERT(inode != NULL);
3246 filter = &exp->exp_obd->u.filter;
3247 iattr_from_obdo(&iattr, oa, oa->o_valid);
3248 ia_valid = iattr.ia_valid;
3250 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3251 OBD_ALLOC(fcc, sizeof(*fcc));
3253 *fcc = oa->o_lcookie;
3255 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3256 unsigned long now = jiffies;
3257 ll_vfs_dq_init(inode);
3258 /* Filter truncates and writes are serialized by
3259 * i_alloc_sem, see the comment in
3260 * filter_preprw_write.*/
3261 if (ia_valid & ATTR_SIZE)
3262 down_write(&inode->i_alloc_sem);
3263 LOCK_INODE_MUTEX(inode);
3264 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
3265 old_size = i_size_read(inode);
3268 /* VBR: version recovery check */
3269 rc = filter_version_get_check(exp, oti, inode);
3271 GOTO(out_unlock, rc);
3273 /* Let's pin the last page so that ldiskfs_truncate
3274 * should not start GFP_FS allocation. */
3275 if (ia_valid & ATTR_SIZE) {
3276 page = grab_cache_page(inode->i_mapping,
3277 iattr.ia_size >> PAGE_CACHE_SHIFT);
3279 GOTO(out_unlock, rc = -ENOMEM);
3284 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3285 * then we will accept the UID+GID sent by the client during write for
3286 * initializing the ownership of this inode. We only allow this to
3287 * happen once so clear these bits in setattr. In 2.6 kernels it is
3288 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3289 * the flags that are actually being set. */
3290 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3291 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3292 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3294 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3295 if (!(ia_valid & ATTR_MODE)) {
3296 iattr.ia_mode = inode->i_mode;
3297 iattr.ia_valid |= ATTR_MODE;
3299 iattr.ia_mode &= ~S_ISUID;
3301 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3302 if (!(iattr.ia_valid & ATTR_MODE)) {
3303 iattr.ia_mode = inode->i_mode;
3304 iattr.ia_valid |= ATTR_MODE;
3306 iattr.ia_mode &= ~S_ISGID;
3309 orig_ids[USRQUOTA] = inode->i_uid;
3310 orig_ids[GRPQUOTA] = inode->i_gid;
3311 handle = fsfilt_start_log(exp->exp_obd, inode,
3312 FSFILT_OP_SETATTR, oti, 1);
3314 GOTO(out_unlock, rc = PTR_ERR(handle));
3316 /* update inode EA only once when inode is suid bit marked. As
3317 * on 2.6.x UID and GID may be set separately, we check here
3318 * only one of them to avoid double setting. */
3319 if (inode->i_mode & S_ISUID)
3320 filter_update_fidea(exp, inode, handle, oa);
3322 handle = fsfilt_start(exp->exp_obd, inode,
3323 FSFILT_OP_SETATTR, oti);
3325 GOTO(out_unlock, rc = PTR_ERR(handle));
3327 if (oa->o_valid & OBD_MD_FLFLAGS) {
3328 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3329 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3331 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3333 /* set cancel cookie callback function */
3334 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3335 filter_cancel_cookies_cb,
3339 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3340 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3342 /* The truncate might have used up our transaction credits. Make sure
3343 * we have two left for the last_rcvd and VBR inode version updates. */
3344 err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3346 /* Update inode version only if data has changed => size has changed */
3347 rc = filter_finish_transno(exp, ia_valid & ATTR_SIZE ? inode : NULL,
3351 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3355 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3357 CERROR("error on commit, err = %d\n", err);
3364 /* For a partial-page truncate flush the page to disk immediately
3365 * to avoid data corruption during direct disk write. b=17397 */
3366 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3367 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3368 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3378 page_cache_release(page);
3380 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3381 UNLOCK_INODE_MUTEX(inode);
3382 if (ia_valid & ATTR_SIZE)
3383 up_write(&inode->i_alloc_sem);
3385 OBD_FREE(fcc, sizeof(*fcc));
3387 /* trigger quota release */
3388 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3389 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3390 int rc2 = lquota_adjust(filter_quota_interface_ref,
3391 exp->exp_obd, cur_ids,
3392 orig_ids, rc, FSFILT_OP_SETATTR);
3393 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3394 "filter adjust qunit. (rc:%d)\n", rc2);
3399 /* this is called from filter_truncate() until we have filter_punch() */
3400 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3401 struct obd_trans_info *oti)
3403 struct obdo *oa = oinfo->oi_oa;
3404 struct lustre_capa *capa = oinfo_capa(oinfo);
3405 struct ldlm_res_id res_id;
3406 struct filter_mod_data *fmd;
3407 struct lvfs_run_ctxt saved;
3408 struct filter_obd *filter;
3409 struct ldlm_resource *res;
3410 struct dentry *dentry;
3411 __u64 opc = CAPA_OPC_META_WRITE;
3415 if (oa->o_valid & OBD_FL_TRUNC)
3416 opc |= CAPA_OPC_OSS_TRUNC;
3418 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3422 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3423 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3428 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3429 /* This would be very bad - accidentally truncating a file when
3430 * changing the time or similar - bug 12203. */
3431 if (oa->o_valid & OBD_MD_FLSIZE &&
3432 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3433 static char mdsinum[48];
3435 if (oa->o_valid & OBD_MD_FLFID)
3436 snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3437 oa->o_parent_seq, oa->o_parent_oid,
3442 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3443 "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3444 oa->o_id, oa->o_seq, mdsinum);
3448 dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3450 RETURN(PTR_ERR(dentry));
3452 filter = &exp->exp_obd->u.filter;
3453 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3456 * We need to be atomic against a concurrent write
3457 * (which takes the semaphore for reading). fmd_mactime_xid
3458 * checks will have no effect if a write request with lower
3459 * xid starts just before a setattr and finishes later than
3460 * the setattr (see bug 21489, comment 27).
3463 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3464 unsigned long now = jiffies;
3465 down_write(&dentry->d_inode->i_alloc_sem);
3466 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem");
3467 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3468 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3469 fmd->fmd_mactime_xid = oti->oti_xid;
3470 filter_fmd_put(exp, fmd);
3471 up_write(&dentry->d_inode->i_alloc_sem);
3474 /* setting objects attributes (including owner/group) */
3475 rc = filter_setattr_internal(exp, dentry, oa, oti);
3477 GOTO(out_unlock, rc);
3479 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3480 &res_id, LDLM_EXTENT, 0);
3483 LDLM_RESOURCE_ADDREF(res);
3484 rc = ldlm_res_lvbo_update(res, NULL, 0);
3485 LDLM_RESOURCE_DELREF(res);
3486 ldlm_resource_putref(res);
3489 oa->o_valid = OBD_MD_FLID;
3491 /* Quota release need uid/gid info */
3492 obdo_from_inode(oa, dentry->d_inode, NULL,
3493 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3498 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3502 /* XXX identical to osc_unpackmd */
3503 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3504 struct lov_mds_md *lmm, int lmm_bytes)
3510 if (lmm_bytes < sizeof (*lmm)) {
3511 CERROR("lov_mds_md too small: %d, need %d\n",
3512 lmm_bytes, (int)sizeof(*lmm));
3515 /* XXX LOV_MAGIC etc check? */
3517 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3518 CERROR("lov_mds_md: zero lmm_object_id\n");
3523 lsm_size = lov_stripe_md_size(1);
3527 if (*lsmp != NULL && lmm == NULL) {
3528 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3529 OBD_FREE(*lsmp, lsm_size);
3534 if (*lsmp == NULL) {
3535 OBD_ALLOC(*lsmp, lsm_size);
3539 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3540 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3541 OBD_FREE(*lsmp, lsm_size);
3544 loi_init((*lsmp)->lsm_oinfo[0]);
3548 /* XXX zero *lsmp? */
3549 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3550 LASSERT((*lsmp)->lsm_object_id);
3553 (*lsmp)->lsm_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
3558 /* caller must hold fo_create_locks[oa->o_seq] */
3559 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3560 struct filter_obd *filter)
3562 struct obdo doa = { 0 }; /* XXX obdo on stack */
3568 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_seq]) != 0);
3570 memset(&doa, 0, sizeof(doa));
3572 doa.o_valid |= OBD_MD_FLGROUP;
3573 doa.o_seq = oa->o_seq;
3574 doa.o_mode = S_IFREG;
3576 if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3577 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3578 exp->exp_obd->obd_name, doa.o_seq);
3582 last = filter_last_id(filter, doa.o_seq);
3584 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3586 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3587 exp->exp_obd->obd_name, oa->o_id + 1, last,
3588 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3590 for (id = last; id > oa->o_id; id--) {
3592 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3593 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3594 CEMERG("error destroying precreate objid "LPU64": %d\n",
3597 /* update last_id on disk periodically so that if we restart
3598 * we don't need to re-scan all of the just-deleted objects. */
3599 if ((id & 511) == 0 && !skip_orphan) {
3600 filter_set_last_id(filter, id - 1, doa.o_seq);
3601 filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3605 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3606 exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3609 filter_set_last_id(filter, id, doa.o_seq);
3610 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3613 * We have destroyed orphan objects, but don't want to reuse
3614 * them. Therefore we don't reset last_id to the last created
3615 * objects. Instead, we report back to the MDS the object id
3616 * of the last orphan, so that the MDS can restart allocating
3617 * objects from this id + 1 and thus skip the whole orphan
3623 cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3628 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3629 obd_seq group, int *num);
3630 /* returns a negative error or a nonnegative number of files to create */
3631 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3632 obd_seq group, struct obd_trans_info *oti)
3634 struct obd_device *obd = exp->exp_obd;
3635 struct filter_obd *filter = &obd->u.filter;
3639 /* delete orphans request */
3640 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3641 obd_id last = filter_last_id(filter, group);
3643 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3644 CERROR("%s: dropping old orphan cleanup request\n",
3648 /* This causes inflight precreates to abort and drop lock */
3649 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3650 cfs_down(&filter->fo_create_locks[group]);
3651 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3652 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3653 exp->exp_obd->obd_name, group);
3654 cfs_up(&filter->fo_create_locks[group]);
3657 diff = oa->o_id - last;
3658 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3661 if (-diff > OST_MAX_PRECREATE) {
3662 CERROR("%s: ignoring bogus orphan destroy request: "
3663 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3665 /* FIXME: should reset precreate_next_id on MDS */
3666 GOTO(out, rc = -EINVAL);
3669 rc = filter_destroy_precreated(exp, oa, filter);
3671 CERROR("%s: unable to write lastobjid, but "
3672 "orphans were deleted\n", obd->obd_name);
3675 /* XXX: Used by MDS for the first time! */
3676 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3679 cfs_down(&filter->fo_create_locks[group]);
3680 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3681 CERROR("%s: dropping old precreate request\n",
3685 /* only precreate if group == 0 and o_id is specfied */
3686 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3689 diff = oa->o_id - filter_last_id(filter, group);
3690 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3691 filter_last_id(filter, group), diff);
3693 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3694 oa->o_id, filter_last_id(filter, group), diff);
3698 oa->o_id = filter_last_id(&obd->u.filter, group);
3699 rc = filter_precreate(obd, oa, group, &diff);
3700 oa->o_id = filter_last_id(&obd->u.filter, group);
3702 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3705 /* else diff == 0 */
3708 cfs_up(&filter->fo_create_locks[group]);
3712 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3713 __u64 max_age, __u32 flags)
3715 struct filter_obd *filter = &obd->u.filter;
3716 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3720 /* at least try to account for cached pages. its still racey and
3721 * might be under-reporting if clients haven't announced their
3722 * caches with brw recently */
3723 cfs_spin_lock(&obd->obd_osfs_lock);
3724 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3725 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3726 cfs_spin_unlock(&obd->obd_osfs_lock);
3728 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3729 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3730 filter->fo_tot_dirty, filter->fo_tot_granted,
3731 filter->fo_tot_pending,
3732 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3734 filter_grant_sanity_check(obd, __func__);
3736 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3737 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3738 osfs->os_bsize - 1) >> blockbits));
3740 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3741 struct lr_server_data *lsd = class_server_data(obd);
3742 int index = le32_to_cpu(lsd->lsd_ost_index);
3744 if (cfs_fail_val == -1 || index == cfs_fail_val)
3745 osfs->os_bfree = osfs->os_bavail = 2;
3746 else if (cfs_fail_loc & OBD_FAIL_ONCE)
3747 cfs_fail_loc &= ~OBD_FAILED; /* reset flag */
3750 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3751 * stop creating files on MDS if OST is not good shape to create
3755 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3756 osfs->os_state = OS_STATE_READONLY;
3758 if (filter->fo_raid_degraded)
3759 osfs->os_state |= OS_STATE_DEGRADED;
3763 static int filter_use_existing_obj(struct obd_device *obd,
3764 struct dentry *dchild, void **handle,
3767 struct inode *inode = dchild->d_inode;
3771 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3774 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3775 if (IS_ERR(*handle))
3776 return PTR_ERR(*handle);
3778 iattr.ia_valid = ATTR_MODE;
3779 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3780 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3787 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3790 __u64 os_ffree = -1;
3792 cfs_spin_lock(&obd->obd_osfs_lock);
3793 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3795 os_ffree = obd->obd_osfs.os_ffree;
3796 cfs_spin_unlock(&obd->obd_osfs_lock);
3801 /* We rely on the fact that only one thread will be creating files in a given
3802 * group at a time, which is why we don't need an atomic filter_get_new_id.
3803 * Even if we had that atomic function, the following race would exist:
3805 * thread 1: gets id x from filter_next_id
3806 * thread 2: gets id (x + 1) from filter_next_id
3807 * thread 2: creates object (x + 1)
3808 * thread 1: tries to create object x, gets -ENOSPC
3810 * Caller must hold fo_create_locks[group]
3812 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3813 obd_seq group, int *num)
3815 struct dentry *dchild = NULL, *dparent = NULL;
3816 struct filter_obd *filter;
3817 struct obd_statfs *osfs;
3819 int err = 0, rc = 0, recreate_obj = 0, i;
3820 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3823 void *handle = NULL;
3826 filter = &obd->u.filter;
3828 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3830 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3832 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3833 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3836 OBD_ALLOC(osfs, sizeof(*osfs));
3839 rc = filter_statfs(obd, osfs,
3840 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
3842 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3843 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3844 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3845 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3847 if (oa->o_valid & OBD_MD_FLFLAGS)
3848 oa->o_flags |= OBD_FL_NOSPC_BLK;
3850 oa->o_valid |= OBD_MD_FLFLAGS;
3851 oa->o_flags = OBD_FL_NOSPC_BLK;
3856 OBD_FREE(osfs, sizeof(*osfs));
3861 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3862 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3864 for (i = 0; i < *num && err == 0; i++) {
3865 int cleanup_phase = 0;
3867 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3868 CWARN("%s: create aborted by destroy\n",
3877 last_id = filter_last_id(filter, group);
3878 if (next_id > last_id) {
3879 CERROR("Error: Trying to recreate obj greater"
3880 "than last id "LPD64" > "LPD64"\n",
3882 GOTO(cleanup, rc = -EINVAL);
3885 next_id = filter_last_id(filter, group) + 1;
3887 /* Don't create objects beyond the valid range for this SEQ */
3888 if (unlikely(fid_seq_is_mdt0(group) &&
3889 next_id >= IDIF_MAX_OID)) {
3890 CERROR("%s:"POSTID" hit the IDIF_MAX_OID (1<<48)!\n",
3891 obd->obd_name, next_id, group);
3892 GOTO(cleanup, rc = -ENOSPC);
3893 } else if (unlikely(!fid_seq_is_mdt0(group) &&
3894 next_id >= OBIF_MAX_OID)) {
3895 CERROR("%s:"POSTID" hit the OBIF_MAX_OID (1<<32)!\n",
3896 obd->obd_name, next_id, group);
3897 GOTO(cleanup, rc = -ENOSPC);
3900 dparent = filter_parent_lock(obd, group, next_id);
3901 if (IS_ERR(dparent))
3902 GOTO(cleanup, rc = PTR_ERR(dparent));
3903 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3905 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3907 GOTO(cleanup, rc = PTR_ERR(dchild));
3908 cleanup_phase = 2; /* f_dput(dchild) */
3910 if (dchild->d_inode != NULL) {
3911 /* This would only happen if lastobjid was bad on disk*/
3912 /* Could also happen if recreating missing obj but it
3913 * already exists. */
3915 CERROR("%s: recreating existing object %.*s?\n",
3916 obd->obd_name, dchild->d_name.len,
3917 dchild->d_name.name);
3919 /* Use these existing objects if they are
3921 if (dchild->d_inode->i_size == 0) {
3922 rc = filter_use_existing_obj(obd,dchild,
3923 &handle, &cleanup_phase);
3930 CERROR("%s: Serious error: objid %.*s already "
3931 "exists; is this filesystem corrupt?\n",
3932 obd->obd_name, dchild->d_name.len,
3933 dchild->d_name.name);
3936 GOTO(cleanup, rc = -EEXIST);
3939 handle = fsfilt_start_log(obd, dparent->d_inode,
3940 FSFILT_OP_CREATE, NULL, 1);
3942 GOTO(cleanup, rc = PTR_ERR(handle));
3945 CDEBUG(D_INODE, "%s: filter_precreate(od->o_seq="LPU64
3946 ",od->o_id="LPU64")\n", obd->obd_name, group,
3949 /* We mark object SUID+SGID to flag it for accepting UID+GID
3950 * from client on first write. Currently the permission bits
3951 * on the OST are never used, so this is OK. */
3952 rc = ll_vfs_create(dparent->d_inode, dchild,
3953 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3955 CERROR("create failed rc = %d\n", rc);
3956 if (rc == -ENOSPC) {
3957 os_ffree = filter_calc_free_inodes(obd);
3961 if (obd->obd_osfs.os_bavail <
3962 (obd->obd_osfs.os_blocks >> 10)) {
3963 if (oa->o_valid & OBD_MD_FLFLAGS)
3964 oa->o_flags |= OBD_FL_NOSPC_BLK;
3966 oa->o_valid |= OBD_MD_FLFLAGS;
3967 oa->o_flags = OBD_FL_NOSPC_BLK;
3970 CERROR("%s: free inode "LPU64"\n",
3971 obd->obd_name, os_ffree);
3977 if (dchild->d_inode)
3978 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3979 dchild->d_inode->i_ino);
3982 /* Set a/c/m time to a insane large negative value at creation
3983 * time so that any timestamp arriving from the client will
3984 * always be newer and update the inode.
3985 * See LU-221 for details */
3986 iattr.ia_valid = ATTR_ATIME | ATTR_MTIME | ATTR_CTIME;
3987 LTIME_S(iattr.ia_atime) = INT_MIN + 24 * 3600;
3988 LTIME_S(iattr.ia_mtime) = INT_MIN + 24 * 3600;
3989 LTIME_S(iattr.ia_ctime) = INT_MIN + 24 * 3600;
3990 err = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
3992 CERROR("unable to initialize a/c/m time of newly"
3995 if (!recreate_obj) {
3996 filter_set_last_id(filter, next_id, group);
3997 err = filter_update_last_objid(obd, group, 0);
3999 CERROR("unable to write lastobjid "
4000 "but file created\n");
4004 switch(cleanup_phase) {
4006 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4008 CERROR("error on commit, err = %d\n", err);
4015 filter_parent_unlock(dparent);
4022 if (cfs_time_after(jiffies, enough_time)) {
4025 "%s: precreate slow - want %d got %d \n",
4026 obd->obd_name, *num, i);
4033 "%s: created %d objects for group "POSTID" rc %d\n",
4034 obd->obd_name, i, filter->fo_last_objids[group], group, rc);
4039 int filter_create(struct obd_export *exp, struct obdo *oa,
4040 struct lov_stripe_md **ea, struct obd_trans_info *oti)
4042 struct obd_device *obd = exp->exp_obd;
4043 struct filter_export_data *fed;
4044 struct filter_obd *filter;
4045 struct lvfs_run_ctxt saved;
4046 struct lov_stripe_md *lsm = NULL;
4050 CDEBUG(D_INODE, "%s: filter_create(group="LPU64",id="
4051 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4053 fed = &exp->exp_filter_data;
4054 filter = &obd->u.filter;
4056 /* 1.8 client doesn't carry the ocd_group with connect request,
4057 * so the fed_group will always be zero for 1.8 client. */
4058 if (!(exp->exp_connect_flags & OBD_CONNECT_FULL20)) {
4059 if (oa->o_seq != FID_SEQ_OST_MDT0 &&
4060 oa->o_seq != FID_SEQ_LLOG &&
4061 oa->o_seq != FID_SEQ_ECHO) {
4062 CERROR("The request from older client has invalid"
4063 " group "LPU64"!\n", oa->o_seq);
4066 } else if (fed->fed_group != oa->o_seq) {
4067 CERROR("%s: this export (nid %s) used object group %d "
4068 "earlier; now it's trying to use group "LPU64"!"
4069 " This could be a bug in the MDS. Please report to "
4070 "http://bugzilla.lustre.org/\n", obd->obd_name,
4071 obd_export_nid2str(exp), fed->fed_group, oa->o_seq);
4078 rc = obd_alloc_memmd(exp, &lsm);
4085 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4087 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4088 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4089 if (!obd->obd_recovering ||
4090 oa->o_id > filter_last_id(filter, oa->o_seq)) {
4091 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4092 oa->o_id, filter_last_id(filter, oa->o_seq));
4096 cfs_down(&filter->fo_create_locks[oa->o_seq]);
4097 rc = filter_precreate(obd, oa, oa->o_seq, &diff);
4098 cfs_up(&filter->fo_create_locks[oa->o_seq]);
4101 rc = filter_handle_precreate(exp, oa, oa->o_seq, oti);
4104 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4105 if (rc && ea != NULL && *ea != lsm) {
4106 obd_free_memmd(exp, &lsm);
4107 } else if (rc == 0 && ea != NULL) {
4108 /* XXX LOV STACKING: the lsm that is passed to us from
4109 * LOV does not have valid lsm_oinfo data structs, so
4110 * don't go touching that. This needs to be fixed in a
4112 lsm->lsm_object_id = oa->o_id;
4119 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4120 struct lov_stripe_md *md, struct obd_trans_info *oti,
4121 struct obd_export *md_exp, void *capa)
4123 unsigned int qcids[MAXQUOTAS] = {0, 0};
4124 struct obd_device *obd;
4125 struct filter_obd *filter;
4126 struct dentry *dchild = NULL, *dparent = NULL;
4127 struct lustre_handle lockh = { 0 };
4128 struct lvfs_run_ctxt saved;
4129 void *handle = NULL;
4130 struct llog_cookie *fcc = NULL;
4131 int rc, rc2, cleanup_phase = 0, sync = 0;
4136 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4137 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4142 filter = &obd->u.filter;
4144 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4147 CDEBUG(D_INODE, "%s: filter_destroy(group="LPU64",oid="
4148 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4150 dchild = filter_fid2dentry(obd, NULL, oa->o_seq, oa->o_id);
4152 GOTO(cleanup, rc = PTR_ERR(dchild));
4155 if (dchild->d_inode == NULL) {
4156 CDEBUG(D_INODE, "destroying non-existent object "POSTID"\n",
4157 oa->o_id, oa->o_seq);
4158 /* If object already gone, cancel cookie right now */
4159 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4160 struct llog_ctxt *ctxt;
4161 struct obd_llog_group *olg;
4163 olg = filter_find_olg(obd, oa->o_seq);
4165 CERROR(" %s: can not find olg of group %d\n",
4166 obd->obd_name, (int)oa->o_seq);
4167 GOTO(cleanup, rc = PTR_ERR(olg));
4169 fcc = &oa->o_lcookie;
4170 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4171 llog_cancel(ctxt, NULL, 1, fcc, 0);
4172 llog_ctxt_put(ctxt);
4173 fcc = NULL; /* we didn't allocate fcc, don't free it */
4175 GOTO(cleanup, rc = -ENOENT);
4178 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_seq, &lockh);
4182 /* Our MDC connection is established by the MDS to us */
4183 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4184 OBD_ALLOC(fcc, sizeof(*fcc));
4186 *fcc = oa->o_lcookie;
4188 ll_vfs_dq_init(dchild->d_inode);
4190 /* we're gonna truncate it first in order to avoid possible deadlock:
4192 * open trasaction open transaction
4193 * down(i_zombie) down(i_zombie)
4194 * restart transaction
4195 * (see BUG 4180) -bzzz
4197 * take i_alloc_sem too to prevent other threads from writing to the
4198 * file while we are truncating it. This can cause lock ordering issue
4199 * between page lock, i_mutex & starting new journal handle.
4200 * (see bug 20321) -johann
4203 down_write(&dchild->d_inode->i_alloc_sem);
4204 LOCK_INODE_MUTEX(dchild->d_inode);
4205 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
4207 /* VBR: version recovery check */
4208 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4210 UNLOCK_INODE_MUTEX(dchild->d_inode);
4211 up_write(&dchild->d_inode->i_alloc_sem);
4215 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4217 if (IS_ERR(handle)) {
4218 UNLOCK_INODE_MUTEX(dchild->d_inode);
4219 up_write(&dchild->d_inode->i_alloc_sem);
4220 GOTO(cleanup, rc = PTR_ERR(handle));
4223 iattr.ia_valid = ATTR_SIZE;
4225 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4226 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4227 UNLOCK_INODE_MUTEX(dchild->d_inode);
4228 up_write(&dchild->d_inode->i_alloc_sem);
4232 GOTO(cleanup, rc = rc2);
4234 /* We don't actually need to lock the parent until we are unlinking
4235 * here, and not while truncating above. That avoids holding the
4236 * parent lock for a long time during truncate, which can block other
4237 * threads from doing anything to objects in that directory. bug 7171 */
4238 dparent = filter_parent_lock(obd, oa->o_seq, oa->o_id);
4239 if (IS_ERR(dparent))
4240 GOTO(cleanup, rc = PTR_ERR(dparent));
4241 cleanup_phase = 3; /* filter_parent_unlock */
4243 LOCK_INODE_MUTEX(dchild->d_inode);
4244 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4245 if (IS_ERR(handle)) {
4246 UNLOCK_INODE_MUTEX(dchild->d_inode);
4247 GOTO(cleanup, rc = PTR_ERR(handle));
4249 cleanup_phase = 4; /* fsfilt_commit */
4251 /* Quota release need uid/gid of inode */
4252 obdo_from_inode(oa, dchild->d_inode, NULL, OBD_MD_FLUID|OBD_MD_FLGID);
4254 filter_fmd_drop(exp, oa->o_id, oa->o_seq);
4256 /* this drops dchild->d_inode->i_mutex unconditionally */
4257 rc = filter_destroy_internal(obd, oa->o_id, oa->o_seq, dparent, dchild);
4261 switch(cleanup_phase) {
4264 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4265 oti->oti_handle : handle,
4266 filter_cancel_cookies_cb,
4268 /* If add_journal_cb failed, then filter_finish_transno
4269 * will commit the handle and we will do a sync
4270 * on commit. then we call callback directly to free
4273 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4275 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4278 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4280 CERROR("error on commit, err = %d\n", rc2);
4287 filter_parent_unlock(dparent);
4289 filter_fini_destroy(obd, &lockh);
4293 OBD_FREE(fcc, sizeof(*fcc));
4295 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4298 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4302 /* trigger quota release */
4303 qcids[USRQUOTA] = oa->o_uid;
4304 qcids[GRPQUOTA] = oa->o_gid;
4305 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4308 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4312 /* NB start and end are used for punch, but not truncate */
4313 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4314 struct obd_trans_info *oti,
4315 struct ptlrpc_request_set *rqset)
4320 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4321 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4322 oinfo->oi_policy.l_extent.end);
4326 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4327 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,oinfo->oi_oa->o_valid,
4328 oinfo->oi_policy.l_extent.start);
4330 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4331 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4332 rc = filter_setattr(exp, oinfo, oti);
4333 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4337 static int filter_sync(struct obd_export *exp, struct obd_info *oinfo,
4338 obd_off start, obd_off end,
4339 struct ptlrpc_request_set *set)
4341 struct lvfs_run_ctxt saved;
4342 struct obd_device_target *obt;
4343 struct dentry *dentry;
4347 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
4348 (struct lustre_capa *)oinfo->oi_capa,
4349 CAPA_OPC_OSS_WRITE);
4353 obt = &exp->exp_obd->u.obt;
4355 /* An objid of zero is taken to mean "sync whole filesystem" */
4356 if (!oinfo->oi_oa || !(oinfo->oi_oa->o_valid & OBD_MD_FLID)) {
4357 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4358 /* Flush any remaining cancel messages out to the target */
4359 filter_sync_llogs(exp->exp_obd, exp);
4363 dentry = filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi);
4365 RETURN(PTR_ERR(dentry));
4367 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4369 LOCK_INODE_MUTEX(dentry->d_inode);
4371 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4373 /* just any file to grab fsync method - "file" arg unused */
4374 struct file *file = obt->obt_rcvd_filp;
4376 if (file->f_op && file->f_op->fsync)
4377 rc = file->f_op->fsync(NULL, dentry, 1);
4379 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4383 UNLOCK_INODE_MUTEX(dentry->d_inode);
4385 oinfo->oi_oa->o_valid = OBD_MD_FLID;
4386 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL,
4387 FILTER_VALID_FLAGS);
4389 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4395 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4396 void *key, __u32 *vallen, void *val,
4397 struct lov_stripe_md *lsm)
4399 struct obd_device *obd;
4402 obd = class_exp2obd(exp);
4404 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4408 if (KEY_IS(KEY_BLOCKSIZE)) {
4409 __u32 *blocksize = val;
4411 if (*vallen < sizeof(*blocksize))
4413 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4415 *vallen = sizeof(*blocksize);
4419 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4420 __u32 *blocksize_bits = val;
4421 if (blocksize_bits) {
4422 if (*vallen < sizeof(*blocksize_bits))
4424 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4426 *vallen = sizeof(*blocksize_bits);
4430 if (KEY_IS(KEY_LAST_ID)) {
4431 obd_id *last_id = val;
4432 /* FIXME: object groups */
4434 if (*vallen < sizeof(*last_id))
4436 *last_id = filter_last_id(&obd->u.filter,
4437 exp->exp_filter_data.fed_group);
4439 *vallen = sizeof(*last_id);
4443 if (KEY_IS(KEY_FIEMAP)) {
4444 struct ll_fiemap_info_key *fm_key = key;
4445 struct dentry *dentry;
4446 struct ll_user_fiemap *fiemap = val;
4447 struct lvfs_run_ctxt saved;
4450 if (fiemap == NULL) {
4451 *vallen = fiemap_count_to_size(
4452 fm_key->fiemap.fm_extent_count);
4456 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa.o_oi,
4459 RETURN(PTR_ERR(dentry));
4461 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4462 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4463 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4465 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4471 if (KEY_IS(KEY_SYNC_LOCK_CANCEL)) {
4472 *((__u32 *) val) = obd->u.filter.fo_sync_lock_cancel;
4473 *vallen = sizeof(__u32);
4477 CDEBUG(D_IOCTL, "invalid key\n");
4481 static inline int filter_setup_llog_group(struct obd_export *exp,
4482 struct obd_device *obd,
4485 struct obd_llog_group *olg;
4486 struct llog_ctxt *ctxt;
4489 olg = filter_find_create_olg(obd, group);
4491 RETURN(PTR_ERR(olg));
4493 llog_group_set_export(olg, exp);
4495 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4496 LASSERTF(ctxt != NULL, "ctxt is null\n");
4498 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4499 llog_ctxt_put(ctxt);
4503 static int filter_set_grant_shrink(struct obd_export *exp,
4504 struct ost_body *body)
4506 /* handle shrink grant */
4507 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4508 filter_grant_incoming(exp, &body->oa);
4509 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4515 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4517 struct obd_device *obd;
4523 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4527 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4528 obd_export_nid2str(exp));
4529 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4531 /* setup llog imports */
4533 group = (int)(*(__u32 *)val);
4535 group = 0; /* default value */
4537 LASSERT_SEQ_IS_MDT(group);
4538 rc = filter_setup_llog_group(exp, obd, group);
4542 if (group == FID_SEQ_OST_MDT0) {
4543 /* setup llog group 1 for interop */
4544 filter_setup_llog_group(exp, obd, FID_SEQ_LLOG);
4547 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4552 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4553 void *key, __u32 vallen, void *val,
4554 struct ptlrpc_request_set *set)
4556 struct obd_device *obd;
4561 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4565 if (KEY_IS(KEY_CAPA_KEY)) {
4567 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4569 CERROR("filter update capability key failed: %d\n", rc);
4573 if (KEY_IS(KEY_REVIMP_UPD)) {
4574 filter_revimp_update(exp);
4575 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4579 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4580 filter_adapt_sptlrpc_conf(obd, 0);
4584 if (KEY_IS(KEY_MDS_CONN))
4585 RETURN(filter_set_mds_conn(exp, val));
4587 if (KEY_IS(KEY_GRANT_SHRINK))
4588 RETURN(filter_set_grant_shrink(exp, val));
4593 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4594 int len, void *karg, void *uarg)
4596 struct obd_device *obd = exp->exp_obd;
4597 struct obd_ioctl_data *data = karg;
4601 case OBD_IOC_ABORT_RECOVERY: {
4602 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4603 target_stop_recovery_thread(obd);
4607 case OBD_IOC_SYNC: {
4608 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4609 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4613 case OBD_IOC_SET_READONLY: {
4615 struct super_block *sb = obd->u.obt.obt_sb;
4616 struct inode *inode = sb->s_root->d_inode;
4617 BDEVNAME_DECLARE_STORAGE(tmp);
4618 CERROR("*** setting device %s read-only ***\n",
4619 ll_bdevname(sb, tmp));
4621 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4622 if (!IS_ERR(handle))
4623 rc = fsfilt_commit(obd, inode, handle, 1);
4625 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4626 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4628 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4632 case OBD_IOC_CATLOGLIST: {
4633 rc = llog_catalog_list(obd, 1, data);
4637 case OBD_IOC_LLOG_CANCEL:
4638 case OBD_IOC_LLOG_REMOVE:
4639 case OBD_IOC_LLOG_INFO:
4640 case OBD_IOC_LLOG_PRINT: {
4641 /* FIXME to be finished */
4642 RETURN(-EOPNOTSUPP);
4644 struct llog_ctxt *ctxt = NULL;
4646 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4647 rc = llog_ioctl(ctxt, cmd, data);
4648 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4661 static int filter_health_check(struct obd_device *obd)
4663 #ifdef USE_HEALTH_CHECK_WRITE
4664 struct filter_obd *filter = &obd->u.filter;
4669 * health_check to return 0 on healthy
4670 * and 1 on unhealthy.
4672 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4675 #ifdef USE_HEALTH_CHECK_WRITE
4676 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4677 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4682 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4685 return filter_fid2dentry(data, NULL, gr, id);
4688 static int filter_process_config(struct obd_device *obd, obd_count len,
4691 struct lustre_cfg *lcfg = buf;
4692 struct lprocfs_static_vars lvars;
4695 switch (lcfg->lcfg_command) {
4697 lprocfs_filter_init_vars(&lvars);
4699 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4709 static struct lvfs_callback_ops filter_lvfs_ops = {
4710 l_fid2dentry: filter_lvfs_fid2dentry,
4713 static struct obd_ops filter_obd_ops = {
4714 .o_owner = THIS_MODULE,
4715 .o_get_info = filter_get_info,
4716 .o_set_info_async = filter_set_info_async,
4717 .o_setup = filter_setup,
4718 .o_precleanup = filter_precleanup,
4719 .o_cleanup = filter_cleanup,
4720 .o_connect = filter_connect,
4721 .o_reconnect = filter_reconnect,
4722 .o_disconnect = filter_disconnect,
4723 .o_ping = filter_ping,
4724 .o_init_export = filter_init_export,
4725 .o_destroy_export = filter_destroy_export,
4726 .o_statfs = filter_statfs,
4727 .o_getattr = filter_getattr,
4728 .o_unpackmd = filter_unpackmd,
4729 .o_create = filter_create,
4730 .o_setattr = filter_setattr,
4731 .o_destroy = filter_destroy,
4732 .o_brw = filter_brw,
4733 .o_punch = filter_truncate,
4734 .o_sync = filter_sync,
4735 .o_preprw = filter_preprw,
4736 .o_commitrw = filter_commitrw,
4737 .o_llog_init = filter_llog_init,
4738 .o_llog_connect = filter_llog_connect,
4739 .o_llog_finish = filter_llog_finish,
4740 .o_iocontrol = filter_iocontrol,
4741 .o_health_check = filter_health_check,
4742 .o_process_config = filter_process_config,
4745 quota_interface_t *filter_quota_interface_ref;
4746 extern quota_interface_t filter_quota_interface;
4748 static int __init obdfilter_init(void)
4750 struct lprocfs_static_vars lvars;
4753 /** sanity check for group<->mdsno conversion */
4754 for (i = 0; i < MAX_MDT_COUNT; i++)
4755 LASSERT(objseq_to_mdsno(mdt_to_obd_objseq(i)) == i);
4757 lprocfs_filter_init_vars(&lvars);
4759 cfs_request_module("%s", "lquota");
4760 OBD_ALLOC(obdfilter_created_scratchpad,
4761 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4762 sizeof(*obdfilter_created_scratchpad));
4763 if (obdfilter_created_scratchpad == NULL)
4766 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4767 sizeof(struct filter_mod_data),
4770 GOTO(out, rc = -ENOMEM);
4772 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4773 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4775 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4776 LUSTRE_OST_NAME, NULL);
4780 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4781 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4782 ll_fmd_cachep = NULL;
4784 if (filter_quota_interface_ref)
4785 PORTAL_SYMBOL_PUT(filter_quota_interface);
4787 OBD_FREE(obdfilter_created_scratchpad,
4788 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4789 sizeof(*obdfilter_created_scratchpad));
4795 static void __exit obdfilter_exit(void)
4797 if (filter_quota_interface_ref)
4798 PORTAL_SYMBOL_PUT(filter_quota_interface);
4800 if (ll_fmd_cachep) {
4801 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4802 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4803 ll_fmd_cachep = NULL;
4806 class_unregister_type(LUSTRE_OST_NAME);
4807 OBD_FREE(obdfilter_created_scratchpad,
4808 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4809 sizeof(*obdfilter_created_scratchpad));
4812 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4813 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4814 MODULE_LICENSE("GPL");
4816 module_init(obdfilter_init);
4817 module_exit(obdfilter_exit);