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)
651 cfs_spin_lock_init(&exp->exp_filter_data.fed_lock);
652 CFS_INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
654 cfs_spin_lock(&exp->exp_lock);
655 exp->exp_connecting = 1;
656 cfs_spin_unlock(&exp->exp_lock);
657 rc = lut_client_alloc(exp);
659 rc = ldlm_init_export(exp);
664 static int filter_free_server_data(struct obd_device_target *obt)
666 lut_fini(NULL, obt->obt_lut);
667 OBD_FREE_PTR(obt->obt_lut);
671 /* assumes caller is already in kernel ctxt */
672 int filter_update_server_data(struct obd_device *obd)
674 struct file *filp = obd->u.obt.obt_rcvd_filp;
675 struct lr_server_data *lsd = class_server_data(obd);
680 CDEBUG(D_INODE, "server uuid : %s\n", lsd->lsd_uuid);
681 CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
682 le64_to_cpu(lsd->lsd_last_transno));
683 CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
684 le64_to_cpu(lsd->lsd_mount_count));
686 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off, 0);
688 CERROR("error writing lr_server_data: rc = %d\n", rc);
693 int filter_update_last_objid(struct obd_device *obd, obd_seq group,
696 struct filter_obd *filter = &obd->u.filter;
702 if (filter->fo_last_objid_files[group] == NULL) {
703 CERROR("Object seq "LPU64" not fully setup; not updating "
704 "last_objid\n", group);
708 CDEBUG(D_INODE, "%s: server last_objid for "POSTID"\n",
709 obd->obd_name, filter->fo_last_objids[group], group);
711 tmp = cpu_to_le64(filter->fo_last_objids[group]);
712 rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
713 &tmp, sizeof(tmp), &off, force_sync);
715 CERROR("error writing seq "LPU64" last objid: rc = %d\n",
719 extern int ost_handle(struct ptlrpc_request *req);
720 /* assumes caller has already in kernel ctxt */
721 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
723 struct filter_obd *filter = &obd->u.filter;
724 struct lr_server_data *lsd;
725 struct lsd_client_data *lcd = NULL;
726 struct inode *inode = filp->f_dentry->d_inode;
727 unsigned long last_rcvd_size = i_size_read(inode);
728 struct lu_target *lut;
735 /* ensure padding in the struct is the correct size */
736 CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
737 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
738 CLASSERT (offsetof(struct lsd_client_data, lcd_padding) +
739 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
741 /* allocate and initialize lu_target */
745 rc = lut_init(NULL, lut, obd, NULL);
748 lsd = class_server_data(obd);
749 if (last_rcvd_size == 0) {
750 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
752 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
753 lsd->lsd_last_transno = 0;
754 mount_count = lsd->lsd_mount_count = 0;
755 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
756 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
757 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
758 lsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
759 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
760 /* OBD_COMPAT_OST is set in filter_connect_internal when the
761 * MDS first connects and assigns the OST index number. */
762 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
765 rc = fsfilt_read_record(obd, filp, lsd, sizeof(*lsd), &off);
767 CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
771 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
772 LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
773 "using the wrong disk %s. Were the "
774 "/dev/ assignments rearranged?\n",
775 obd->obd_uuid.uuid, lsd->lsd_uuid);
776 GOTO(err_lut, rc = -EINVAL);
778 mount_count = le64_to_cpu(lsd->lsd_mount_count);
779 filter->fo_subdir_count = le16_to_cpu(lsd->lsd_subdir_count);
781 /* Assume old last_rcvd format unless I_C_LR is set */
782 if (!(lsd->lsd_feature_incompat &
783 cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
784 lsd->lsd_last_transno = lsd->lsd_compat14;
786 /* OBD_COMPAT_OST is set in filter_connect_internal when the
787 * MDS first connects and assigns the OST index number. */
788 lsd->lsd_feature_incompat |= cpu_to_le32(OBD_INCOMPAT_COMMON_LR|
792 if (lsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
793 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
794 obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
795 ~FILTER_INCOMPAT_SUPP);
796 GOTO(err_lut, rc = -EINVAL);
798 if (lsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
799 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
800 obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
801 ~FILTER_ROCOMPAT_SUPP);
802 /* Do something like remount filesystem read-only */
803 GOTO(err_lut, rc = -EINVAL);
806 start_epoch = le32_to_cpu(lsd->lsd_start_epoch);
808 CDEBUG(D_INODE, "%s: server start_epoch : %#x\n",
809 obd->obd_name, start_epoch);
810 CDEBUG(D_INODE, "%s: server last_transno : "LPX64"\n",
811 obd->obd_name, le64_to_cpu(lsd->lsd_last_transno));
812 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
813 obd->obd_name, mount_count + 1);
814 CDEBUG(D_INODE, "%s: server data size: %u\n",
815 obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
816 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
817 obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
818 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
819 obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
820 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
821 obd->obd_name, le16_to_cpu(lsd->lsd_subdir_count));
822 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
823 last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
824 (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
825 le16_to_cpu(lsd->lsd_client_size));
827 if (!obd->obd_replayable) {
828 CWARN("%s: recovery support OFF\n", obd->obd_name);
834 GOTO(err_client, rc = -ENOMEM);
836 for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
837 off < last_rcvd_size; cl_idx++) {
839 struct obd_export *exp;
840 struct filter_export_data *fed;
842 /* Don't assume off is incremented properly by
843 * fsfilt_read_record(), in case sizeof(*lcd)
844 * isn't the same as lsd->lsd_client_size. */
845 off = le32_to_cpu(lsd->lsd_client_start) +
846 cl_idx * le16_to_cpu(lsd->lsd_client_size);
847 rc = fsfilt_read_record(obd, filp, lcd, sizeof(*lcd), &off);
849 CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
850 LAST_RCVD, cl_idx, off, rc);
851 break; /* read error shouldn't cause startup to fail */
854 if (lcd->lcd_uuid[0] == '\0') {
855 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
860 check_lcd(obd->obd_name, cl_idx, lcd);
862 last_rcvd = le64_to_cpu(lcd->lcd_last_transno);
864 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
865 " srv lr: "LPU64"\n", lcd->lcd_uuid, cl_idx,
866 last_rcvd, le64_to_cpu(lsd->lsd_last_transno));
868 /* These exports are cleaned up by filter_disconnect(), so they
869 * need to be set up like real exports as filter_connect() does.
871 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
873 if (PTR_ERR(exp) == -EALREADY) {
874 /* export already exists, zero out this one */
875 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
879 GOTO(err_client, rc = PTR_ERR(exp));
882 fed = &exp->exp_filter_data;
883 *fed->fed_ted.ted_lcd = *lcd;
884 fed->fed_group = 0; /* will be assigned at connect */
885 filter_export_stats_init(obd, exp, NULL);
886 rc = filter_client_add(obd, exp, cl_idx);
887 /* can't fail for existing client */
888 LASSERTF(rc == 0, "rc = %d\n", rc);
890 /* VBR: set export last committed */
891 exp->exp_last_committed = last_rcvd;
892 cfs_spin_lock(&exp->exp_lock);
893 exp->exp_connecting = 0;
894 exp->exp_in_recovery = 0;
895 cfs_spin_unlock(&exp->exp_lock);
896 obd->obd_max_recoverable_clients++;
897 class_export_put(exp);
899 if (last_rcvd > le64_to_cpu(lsd->lsd_last_transno))
900 lsd->lsd_last_transno = cpu_to_le64(last_rcvd);
904 obd->obd_last_committed = le64_to_cpu(lsd->lsd_last_transno);
906 obd->u.obt.obt_mount_count = mount_count + 1;
907 obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
908 lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
910 /* save it, so mount count and last_transno is current */
911 rc = filter_update_server_data(obd);
913 GOTO(err_client, rc);
918 class_disconnect_exports(obd);
920 filter_free_server_data(&obd->u.obt);
924 static int filter_cleanup_groups(struct obd_device *obd)
926 struct filter_obd *filter = &obd->u.filter;
928 struct dentry *dentry;
932 if (filter->fo_dentry_O_groups != NULL) {
933 for (i = 0; i < filter->fo_group_count; i++) {
934 dentry = filter->fo_dentry_O_groups[i];
938 OBD_FREE(filter->fo_dentry_O_groups,
939 filter->fo_group_count *
940 sizeof(*filter->fo_dentry_O_groups));
941 filter->fo_dentry_O_groups = NULL;
943 if (filter->fo_last_objid_files != NULL) {
944 for (i = 0; i < filter->fo_group_count; i++) {
945 filp = filter->fo_last_objid_files[i];
949 OBD_FREE(filter->fo_last_objid_files,
950 filter->fo_group_count *
951 sizeof(*filter->fo_last_objid_files));
952 filter->fo_last_objid_files = NULL;
954 if (filter->fo_dentry_O_sub != NULL) {
955 for (i = 0; i < filter->fo_group_count; i++) {
956 for (j = 0; j < filter->fo_subdir_count; j++) {
957 dentry = filter->fo_dentry_O_sub[i].dentry[j];
962 OBD_FREE(filter->fo_dentry_O_sub,
963 filter->fo_group_count *
964 sizeof(*filter->fo_dentry_O_sub));
965 filter->fo_dentry_O_sub = NULL;
967 if (filter->fo_last_objids != NULL) {
968 OBD_FREE(filter->fo_last_objids,
969 filter->fo_group_count *
970 sizeof(*filter->fo_last_objids));
971 filter->fo_last_objids = NULL;
973 if (filter->fo_dentry_O != NULL) {
974 f_dput(filter->fo_dentry_O);
975 filter->fo_dentry_O = NULL;
980 static int filter_update_last_group(struct obd_device *obd, int group)
982 struct filter_obd *filter = &obd->u.filter;
983 struct file *filp = NULL;
984 int last_group = 0, rc;
988 if (group <= filter->fo_committed_group)
991 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
995 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
999 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1001 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1005 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1006 obd->obd_name, last_group, group);
1010 /* must be sync: bXXXX */
1011 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1013 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1017 filter->fo_committed_group = group;
1020 filp_close(filp, 0);
1024 static int filter_read_group_internal(struct obd_device *obd, int group,
1027 struct filter_obd *filter = &obd->u.filter;
1028 __u64 *new_objids = NULL;
1029 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1030 struct dentry **new_groups = NULL;
1031 struct file **new_files = NULL;
1032 struct dentry *dentry;
1034 int old_count = filter->fo_group_count, rc, stage = 0, i;
1038 int len = group + 1;
1040 snprintf(name, 24, "%d", group);
1044 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1046 if (IS_ERR(dentry)) {
1047 CERROR("Cannot lookup expected object group %d: %ld\n",
1048 group, PTR_ERR(dentry));
1049 RETURN(PTR_ERR(dentry));
1052 dentry = simple_mkdir(filter->fo_dentry_O,
1053 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1054 if (IS_ERR(dentry)) {
1055 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1057 RETURN(PTR_ERR(dentry));
1062 snprintf(name, 24, "O/%d/LAST_ID", group);
1064 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1066 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1067 GOTO(cleanup, rc = PTR_ERR(filp));
1071 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1073 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1077 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1078 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1079 if (tmp_subdirs == NULL)
1080 GOTO(cleanup, rc = -ENOMEM);
1083 for (i = 0; i < filter->fo_subdir_count; i++) {
1085 snprintf(dir, sizeof(dir), "d%u", i);
1087 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1088 obd->u.obt.obt_vfsmnt,
1090 if (IS_ERR(tmp_subdirs->dentry[i])) {
1091 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1092 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1097 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1098 tmp_subdirs->dentry[i]);
1102 /* 'group' is an index; we need an array of length 'group + 1' */
1103 if (group + 1 > old_count) {
1104 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1105 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1106 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1107 OBD_ALLOC(new_files, len * sizeof(*new_files));
1109 if (new_objids == NULL || new_subdirs == NULL ||
1110 new_groups == NULL || new_files == NULL)
1111 GOTO(cleanup, rc = -ENOMEM);
1114 memcpy(new_objids, filter->fo_last_objids,
1115 old_count * sizeof(*new_objids));
1116 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1117 old_count * sizeof(*new_subdirs));
1118 memcpy(new_groups, filter->fo_dentry_O_groups,
1119 old_count * sizeof(*new_groups));
1120 memcpy(new_files, filter->fo_last_objid_files,
1121 old_count * sizeof(*new_files));
1123 OBD_FREE(filter->fo_last_objids,
1124 old_count * sizeof(*new_objids));
1125 OBD_FREE(filter->fo_dentry_O_sub,
1126 old_count * sizeof(*new_subdirs));
1127 OBD_FREE(filter->fo_dentry_O_groups,
1128 old_count * sizeof(*new_groups));
1129 OBD_FREE(filter->fo_last_objid_files,
1130 old_count * sizeof(*new_files));
1132 filter->fo_last_objids = new_objids;
1133 filter->fo_dentry_O_sub = new_subdirs;
1134 filter->fo_dentry_O_groups = new_groups;
1135 filter->fo_last_objid_files = new_files;
1136 filter->fo_group_count = len;
1139 filter->fo_dentry_O_groups[group] = dentry;
1140 filter->fo_last_objid_files[group] = filp;
1141 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1142 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1143 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1146 filter_update_last_group(obd, group);
1148 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1149 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1150 rc = filter_update_last_objid(obd, group, 1);
1154 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1155 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1156 obd->obd_name, group, last_objid);
1161 if (new_objids != NULL)
1162 OBD_FREE(new_objids, len * sizeof(*new_objids));
1163 if (new_subdirs != NULL)
1164 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1165 if (new_groups != NULL)
1166 OBD_FREE(new_groups, len * sizeof(*new_groups));
1167 if (new_files != NULL)
1168 OBD_FREE(new_files, len * sizeof(*new_files));
1170 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1171 for (i = 0; i < filter->fo_subdir_count; i++) {
1172 if (tmp_subdirs->dentry[i] != NULL)
1173 dput(tmp_subdirs->dentry[i]);
1175 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1178 filp_close(filp, 0);
1185 static int filter_read_groups(struct obd_device *obd, int last_group,
1188 struct filter_obd *filter = &obd->u.filter;
1189 int old_count, group, rc = 0;
1191 cfs_down(&filter->fo_init_lock);
1192 old_count = filter->fo_group_count;
1193 for (group = old_count; group <= last_group; group++) {
1194 rc = filter_read_group_internal(obd, group, create);
1198 cfs_up(&filter->fo_init_lock);
1202 /* FIXME: object groups */
1203 static int filter_prep_groups(struct obd_device *obd)
1205 struct filter_obd *filter = &obd->u.filter;
1206 struct dentry *O_dentry;
1208 int last_group, rc = 0, cleanup_phase = 0;
1212 O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1214 CDEBUG(D_INODE, "%s: got/created O: %p\n", obd->obd_name, O_dentry);
1215 if (IS_ERR(O_dentry)) {
1216 rc = PTR_ERR(O_dentry);
1217 CERROR("%s: cannot open/create O: rc = %d\n", obd->obd_name,rc);
1220 filter->fo_dentry_O = O_dentry;
1221 cleanup_phase = 1; /* O_dentry */
1223 /* we have to initialize all groups before first connections from
1224 * clients because they may send create/destroy for any group -bzzz */
1225 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1227 CERROR("%s: cannot create LAST_GROUP: rc = %ld\n",
1228 obd->obd_name, PTR_ERR(filp));
1229 GOTO(cleanup, rc = PTR_ERR(filp));
1231 cleanup_phase = 2; /* filp */
1233 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1235 CERROR("%s: error reading LAST_GROUP: rc %d\n",
1241 last_group = FID_SEQ_OST_MDT0;
1243 CDEBUG(D_INODE, "%s: initialize group %u (max %u)\n", obd->obd_name,
1244 FID_SEQ_OST_MDT0, last_group);
1245 filter->fo_committed_group = last_group;
1246 rc = filter_read_groups(obd, last_group, 1);
1250 filp_close(filp, 0);
1254 switch (cleanup_phase) {
1256 filp_close(filp, 0);
1258 filter_cleanup_groups(obd);
1259 f_dput(filter->fo_dentry_O);
1260 filter->fo_dentry_O = NULL;
1268 /* setup the object store with correct subdirectories */
1269 static int filter_prep(struct obd_device *obd)
1271 struct lvfs_run_ctxt saved;
1272 struct filter_obd *filter = &obd->u.filter;
1274 struct inode *inode;
1278 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1279 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1280 if (!file || IS_ERR(file)) {
1282 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1286 obd->u.obt.obt_rcvd_filp = file;
1287 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1288 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1289 file->f_dentry->d_inode->i_mode);
1290 GOTO(err_filp, rc = -ENOENT);
1293 inode = file->f_dentry->d_parent->d_inode;
1294 /* We use i_op->unlink directly in filter_vfs_unlink() */
1295 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1296 CERROR("%s: filesystem does not support create/unlink ops\n",
1298 GOTO(err_filp, rc = -EOPNOTSUPP);
1301 rc = filter_init_server_data(obd, file);
1303 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1306 LASSERT(obd->u.obt.obt_lut);
1307 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1309 /* open and create health check io file*/
1310 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1313 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1315 GOTO(err_server_data, rc);
1317 filter->fo_obt.obt_health_check_filp = file;
1318 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1319 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1320 file->f_dentry->d_inode->i_mode);
1321 GOTO(err_health_check, rc = -ENOENT);
1323 rc = lvfs_check_io_health(obd, file);
1325 GOTO(err_health_check, rc);
1327 rc = filter_prep_groups(obd);
1329 GOTO(err_health_check, rc);
1331 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1336 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1337 CERROR("can't close %s after error\n", HEALTH_CHECK);
1338 filter->fo_obt.obt_health_check_filp = NULL;
1340 target_recovery_fini(obd);
1341 filter_free_server_data(&obd->u.obt);
1343 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1344 CERROR("can't close %s after error\n", LAST_RCVD);
1345 obd->u.obt.obt_rcvd_filp = NULL;
1349 /* cleanup the filter: write last used object id to status file */
1350 static void filter_post(struct obd_device *obd)
1352 struct lvfs_run_ctxt saved;
1353 struct filter_obd *filter = &obd->u.filter;
1356 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1357 * best to start a transaction with h_sync, because we removed this
1360 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1361 rc = filter_update_server_data(obd);
1363 CERROR("error writing server data: rc = %d\n", rc);
1365 for (i = 0; i < filter->fo_group_count; i++) {
1366 rc = filter_update_last_objid(obd, i,
1367 (i == filter->fo_group_count - 1));
1369 CERROR("error writing group %d lastobjid: rc = %d\n",
1373 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1374 obd->u.obt.obt_rcvd_filp = NULL;
1376 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1378 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1379 filter->fo_obt.obt_health_check_filp = NULL;
1381 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1383 filter_cleanup_groups(obd);
1384 filter_free_server_data(&obd->u.obt);
1385 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1387 filter_free_capa_keys(filter);
1388 cleanup_capa_hash(filter->fo_capa_hash);
1391 static void filter_set_last_id(struct filter_obd *filter,
1392 obd_id id, obd_seq group)
1394 LASSERT(group <= filter->fo_group_count);
1396 cfs_spin_lock(&filter->fo_objidlock);
1397 filter->fo_last_objids[group] = id;
1398 cfs_spin_unlock(&filter->fo_objidlock);
1401 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1404 LASSERT(group <= filter->fo_group_count);
1405 LASSERT(filter->fo_last_objids != NULL);
1407 /* FIXME: object groups */
1408 cfs_spin_lock(&filter->fo_objidlock);
1409 id = filter->fo_last_objids[group];
1410 cfs_spin_unlock(&filter->fo_objidlock);
1414 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1416 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1420 /* We never dget the object parent, so DON'T dput it either */
1421 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1423 struct filter_obd *filter = &obd->u.filter;
1424 struct filter_subdirs *subdirs;
1426 if (group >= filter->fo_group_count) /* FIXME: object groups */
1427 return ERR_PTR(-EBADF);
1429 if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1430 return filter->fo_dentry_O_groups[group];
1432 subdirs = &filter->fo_dentry_O_sub[group];
1433 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1436 /* We never dget the object parent, so DON'T dput it either */
1437 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1440 unsigned long now = jiffies;
1441 struct dentry *dparent = filter_parent(obd, group, objid);
1444 if (IS_ERR(dparent))
1446 if (dparent == NULL)
1447 return ERR_PTR(-ENOENT);
1449 rc = filter_lock_dentry(obd, dparent);
1450 fsfilt_check_slow(obd, now, "parent lock");
1451 return rc ? ERR_PTR(rc) : dparent;
1454 /* We never dget the object parent, so DON'T dput it either */
1455 static void filter_parent_unlock(struct dentry *dparent)
1457 UNLOCK_INODE_MUTEX(dparent->d_inode);
1460 /* How to get files, dentries, inodes from object id's.
1462 * If dir_dentry is passed, the caller has already locked the parent
1463 * appropriately for this operation (normally a write lock). If
1464 * dir_dentry is NULL, we do a read lock while we do the lookup to
1465 * avoid races with create/destroy and such changing the directory
1466 * internal to the filesystem code. */
1467 struct dentry *filter_fid2dentry(struct obd_device *obd,
1468 struct dentry *dir_dentry,
1469 obd_seq group, obd_id id)
1471 struct dentry *dparent = dir_dentry;
1472 struct dentry *dchild;
1477 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1478 obd->u.filter.fo_destroys_in_progress == 0) {
1479 /* don't fail lookups for orphan recovery, it causes
1480 * later LBUGs when objects still exist during precreate */
1481 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1482 RETURN(ERR_PTR(-ENOENT));
1485 CERROR("fatal: invalid object id 0\n");
1486 RETURN(ERR_PTR(-ESTALE));
1489 len = sprintf(name, LPU64, id);
1490 if (dir_dentry == NULL) {
1491 dparent = filter_parent_lock(obd, group, id);
1492 if (IS_ERR(dparent)) {
1493 CERROR("%s: error getting object "POSTID
1494 " parent: rc %ld\n", obd->obd_name,
1495 id, group, PTR_ERR(dparent));
1499 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1500 dparent->d_name.len, dparent->d_name.name, name);
1501 /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1502 dchild = lookup_one_len(name, dparent, len);
1503 if (dir_dentry == NULL)
1504 filter_parent_unlock(dparent);
1505 if (IS_ERR(dchild)) {
1506 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1507 obd->obd_name, id, group, PTR_ERR(dchild));
1511 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1512 CERROR("%s: got bad object "LPU64" inode %lu\n",
1513 obd->obd_name, id, dchild->d_inode->i_ino);
1515 RETURN(ERR_PTR(-ENOENT));
1518 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1519 name, dchild, atomic_read(&dchild->d_count));
1521 LASSERT(atomic_read(&dchild->d_count) > 0);
1526 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1527 obd_id group, struct lustre_handle *lockh)
1529 int flags = LDLM_AST_DISCARD_DATA, rc;
1530 struct ldlm_res_id res_id;
1531 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1534 osc_build_res_name(objid, group, &res_id);
1535 /* Tell the clients that the object is gone now and that they should
1536 * throw away any cached pages. */
1537 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1538 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1539 ldlm_completion_ast, NULL, NULL, 0, NULL,
1546 static void filter_fini_destroy(struct obd_device *obd,
1547 struct lustre_handle *lockh)
1549 if (lustre_handle_is_used(lockh))
1550 ldlm_lock_decref(lockh, LCK_PW);
1553 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1554 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1555 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1556 * reverse. Caller must take i_sem before starting the transaction and we
1557 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1558 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1559 struct vfsmount *mnt)
1564 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1565 * itself we already hold dir->i_mutex for child create/unlink ops */
1566 LASSERT(dentry->d_inode != NULL);
1567 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1568 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1572 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1573 GOTO(out, rc = -ENOENT);
1575 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1580 GOTO(out, rc = -EPERM);
1582 /* check_sticky() */
1583 if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1584 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1585 IS_IMMUTABLE(dentry->d_inode))
1586 GOTO(out, rc = -EPERM);
1588 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1589 * that was done because of journal_start (which is already done
1590 * here) or some other ordering issue. */
1591 ll_vfs_dq_init(dir);
1593 rc = ll_security_inode_unlink(dir, dentry, mnt);
1597 rc = dir->i_op->unlink(dir, dentry);
1599 /* need to drop i_mutex before we lose inode reference */
1600 UNLOCK_INODE_MUTEX(dentry->d_inode);
1607 /* Caller must hold LCK_PW on parent and push us into kernel context.
1608 * Caller must hold child i_mutex, we drop it always.
1609 * Caller is also required to ensure that dchild->d_inode exists. */
1610 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1611 obd_seq group, struct dentry *dparent,
1612 struct dentry *dchild)
1614 struct inode *inode = dchild->d_inode;
1617 /* There should be 2 references to the inode:
1618 * 1) taken by filter_prepare_destroy
1619 * 2) taken by filter_destroy */
1620 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1621 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1622 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1623 (unsigned long)inode->i_nlink,
1624 atomic_read(&inode->i_count));
1627 rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1629 CERROR("error unlinking objid %.*s: rc %d\n",
1630 dchild->d_name.len, dchild->d_name.name, rc);
1634 struct filter_intent_args {
1635 struct ldlm_lock **victim;
1640 static enum interval_iter filter_intent_cb(struct interval_node *n,
1643 struct ldlm_interval *node = (struct ldlm_interval *)n;
1644 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1645 __u64 size = arg->size;
1646 struct ldlm_lock **v = arg->victim;
1647 struct ldlm_lock *lck;
1649 /* If the interval is lower than the current file size,
1651 if (interval_high(n) <= size)
1652 return INTERVAL_ITER_STOP;
1654 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1655 /* Don't send glimpse ASTs to liblustre clients.
1656 * They aren't listening for them, and they do
1657 * entirely synchronous I/O anyways. */
1658 if (lck->l_export == NULL ||
1659 lck->l_export->exp_libclient == 1)
1662 if (*arg->liblustre)
1663 *arg->liblustre = 0;
1666 *v = LDLM_LOCK_GET(lck);
1667 } else if ((*v)->l_policy_data.l_extent.start <
1668 lck->l_policy_data.l_extent.start) {
1669 LDLM_LOCK_RELEASE(*v);
1670 *v = LDLM_LOCK_GET(lck);
1673 /* the same policy group - every lock has the
1674 * same extent, so needn't do it any more */
1678 return INTERVAL_ITER_CONT;
1681 static int filter_intent_policy(struct ldlm_namespace *ns,
1682 struct ldlm_lock **lockp, void *req_cookie,
1683 ldlm_mode_t mode, int flags, void *data)
1685 CFS_LIST_HEAD(rpc_list);
1686 struct ptlrpc_request *req = req_cookie;
1687 struct ldlm_lock *lock = *lockp, *l = NULL;
1688 struct ldlm_resource *res = lock->l_resource;
1689 ldlm_processing_policy policy;
1690 struct ost_lvb *res_lvb, *reply_lvb;
1691 struct ldlm_reply *rep;
1693 int idx, rc, tmpflags = 0, only_liblustre = 1;
1694 struct ldlm_interval_tree *tree;
1695 struct filter_intent_args arg;
1696 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1697 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1698 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1701 policy = ldlm_get_processing_policy(res);
1702 LASSERT(policy != NULL);
1703 LASSERT(req != NULL);
1705 rc = lustre_pack_reply(req, 3, repsize, NULL);
1707 RETURN(req->rq_status = rc);
1709 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1710 LASSERT(rep != NULL);
1712 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1713 sizeof(*reply_lvb));
1714 LASSERT(reply_lvb != NULL);
1716 //fixup_handle_for_resent_req(req, lock, &lockh);
1718 /* Call the extent policy function to see if our request can be
1719 * granted, or is blocked.
1720 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1721 * lock, and should not be granted if the lock will be blocked.
1724 LASSERT(ns == ldlm_res_to_ns(res));
1726 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1727 check_res_locked(res);
1729 /* FIXME: we should change the policy function slightly, to not make
1730 * this list at all, since we just turn around and free it */
1731 while (!cfs_list_empty(&rpc_list)) {
1732 struct ldlm_lock *wlock =
1733 cfs_list_entry(rpc_list.next, struct ldlm_lock,
1735 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1736 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1737 lock->l_flags &= ~LDLM_FL_CP_REQD;
1738 cfs_list_del_init(&wlock->l_cp_ast);
1739 LDLM_LOCK_RELEASE(wlock);
1742 /* The lock met with no resistance; we're finished. */
1743 if (rc == LDLM_ITER_CONTINUE) {
1744 /* do not grant locks to the liblustre clients: they cannot
1745 * handle ASTs robustly. We need to do this while still
1746 * holding lr_lock to avoid the lock remaining on the res_link
1747 * list (and potentially being added to l_pending_list by an
1748 * AST) when we are going to drop this lock ASAP. */
1749 if (lock->l_export->exp_libclient ||
1750 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1751 ldlm_resource_unlink_lock(lock);
1752 err = ELDLM_LOCK_ABORTED;
1754 err = ELDLM_LOCK_REPLACED;
1760 /* Do not grant any lock, but instead send GL callbacks. The extent
1761 * policy nicely created a list of all PW locks for us. We will choose
1762 * the highest of those which are larger than the size in the LVB, if
1763 * any, and perform a glimpse callback. */
1764 res_lvb = res->lr_lvb_data;
1765 LASSERT(res_lvb != NULL);
1766 *reply_lvb = *res_lvb;
1769 * lr_lock guarantees that no new locks are granted, and,
1770 * therefore, that res->lr_lvb_data cannot increase beyond the
1771 * end of already granted lock. As a result, it is safe to
1772 * check against "stale" reply_lvb->lvb_size value without
1775 arg.size = reply_lvb->lvb_size;
1777 arg.liblustre = &only_liblustre;
1778 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1779 tree = &res->lr_itree[idx];
1780 if (tree->lit_mode == LCK_PR)
1783 interval_iterate_reverse(tree->lit_root,
1784 filter_intent_cb, &arg);
1788 /* There were no PW locks beyond the size in the LVB; finished. */
1790 if (only_liblustre) {
1791 /* If we discovered a liblustre client with a PW lock,
1792 * however, the LVB may be out of date! The LVB is
1793 * updated only on glimpse (which we don't do for
1794 * liblustre clients) and cancel (which the client
1795 * obviously has not yet done). So if it has written
1796 * data but kept the lock, the LVB is stale and needs
1797 * to be updated from disk.
1799 * Of course, this will all disappear when we switch to
1800 * taking liblustre locks on the OST. */
1801 ldlm_res_lvbo_update(res, NULL, 1);
1803 RETURN(ELDLM_LOCK_ABORTED);
1807 * This check is for lock taken in filter_prepare_destroy() that does
1808 * not have l_glimpse_ast set. So the logic is: if there is a lock
1809 * with no l_glimpse_ast set, this object is being destroyed already.
1811 * Hence, if you are grabbing DLM locks on the server, always set
1812 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1814 if (l->l_glimpse_ast == NULL) {
1815 /* We are racing with unlink(); just return -ENOENT */
1816 rep->lock_policy_res1 = -ENOENT;
1820 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1821 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1824 *reply_lvb = *res_lvb;
1828 LDLM_LOCK_RELEASE(l);
1830 RETURN(ELDLM_LOCK_ABORTED);
1834 * per-obd_device iobuf pool.
1836 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1837 * allocations in write-path has to be minimized (see bug 5137).
1839 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1840 * OST threads (see ost_thread_{init,done}()).
1842 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1843 * at the OST layer there are only (potentially) multiple obd_device of type
1844 * unknown at the time of OST thread creation.
1846 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1847 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1848 * it's very own iobuf.
1852 * filter_kiobuf_pool_init()
1854 * filter_kiobuf_pool_done()
1856 * filter_iobuf_get()
1858 * operate on this array. They are "generic" in a sense that they don't depend
1859 * on actual type of iobuf's (the latter depending on Linux kernel version).
1863 * destroy pool created by filter_iobuf_pool_init
1865 static void filter_iobuf_pool_done(struct filter_obd *filter)
1867 struct filter_iobuf **pool;
1872 pool = filter->fo_iobuf_pool;
1874 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1875 if (pool[i] != NULL)
1876 filter_free_iobuf(pool[i]);
1878 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1879 filter->fo_iobuf_pool = NULL;
1884 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1886 struct filter_obd *filter = &obd->u.filter;
1887 struct sptlrpc_rule_set tmp_rset;
1890 sptlrpc_rule_set_init(&tmp_rset);
1891 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1893 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1898 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1900 cfs_write_lock(&filter->fo_sptlrpc_lock);
1901 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1902 filter->fo_sptlrpc_rset = tmp_rset;
1903 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1909 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1911 static int filter_iobuf_pool_init(struct filter_obd *filter)
1918 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1920 if (filter->fo_iobuf_pool == NULL)
1923 filter->fo_iobuf_count = OSS_THREADS_MAX;
1928 /* Return iobuf allocated for @thread_id. We don't know in advance how
1929 * many threads there will be so we allocate a large empty array and only
1930 * fill in those slots that are actually in use.
1931 * If we haven't allocated a pool entry for this thread before, do so now. */
1932 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1934 int thread_id = (oti && oti->oti_thread) ?
1935 oti->oti_thread->t_id : -1;
1936 struct filter_iobuf *pool = NULL;
1937 struct filter_iobuf **pool_place = NULL;
1939 if (thread_id >= 0) {
1940 LASSERT(thread_id < filter->fo_iobuf_count);
1941 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1944 if (unlikely(pool == NULL)) {
1945 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1946 PTLRPC_MAX_BRW_PAGES);
1947 if (pool_place != NULL)
1954 /* mount the file system (secretly). lustre_cfg parameters are:
1957 * 3 = flags: failover=f, failout=n
1960 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1963 struct filter_obd *filter = &obd->u.filter;
1964 struct vfsmount *mnt;
1965 struct lustre_mount_info *lmi;
1966 struct obd_uuid uuid;
1970 struct request_queue *q;
1974 if (lcfg->lcfg_bufcount < 3 ||
1975 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1976 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1979 lmi = server_get_mount(obd->obd_name);
1981 /* We already mounted in lustre_fill_super.
1982 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1983 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1985 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1987 /* old path - used by lctl */
1988 CERROR("Using old MDS mount method\n");
1989 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1990 MS_NOATIME|MS_NODIRATIME,
1991 lustre_cfg_string(lcfg, 1), option);
1994 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1995 lustre_cfg_string(lcfg, 1), rc);
1999 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2001 if (IS_ERR(obd->obd_fsops))
2002 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2004 rc = filter_iobuf_pool_init(filter);
2008 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2009 CERROR("%s: Underlying device is marked as read-only. "
2010 "Setup failed\n", obd->obd_name);
2011 GOTO(err_ops, rc = -EROFS);
2014 /* failover is the default */
2015 obd->obd_replayable = 1;
2017 /* disable connection until configuration finishes */
2018 obd->obd_no_conn = 1;
2020 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2021 str = lustre_cfg_string(lcfg, 3);
2022 if (strchr(str, 'n')) {
2023 CWARN("%s: recovery disabled\n", obd->obd_name);
2024 obd->obd_replayable = 0;
2028 obd->u.obt.obt_magic = OBT_MAGIC;
2029 obd->u.obt.obt_vfsmnt = mnt;
2030 obd->u.obt.obt_sb = mnt->mnt_sb;
2031 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2032 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2034 fsfilt_setup(obd, obd->u.obt.obt_sb);
2036 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2037 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2038 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2039 obd->obd_lvfs_ctxt.fs = get_ds();
2040 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2042 cfs_init_mutex(&filter->fo_init_lock);
2043 filter->fo_committed_group = 0;
2044 filter->fo_destroys_in_progress = 0;
2045 for (i = 0; i < 32; i++)
2046 cfs_sema_init(&filter->fo_create_locks[i], 1);
2048 cfs_spin_lock_init(&filter->fo_objidlock);
2049 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2050 cfs_sema_init(&filter->fo_alloc_lock, 1);
2051 init_brw_stats(&filter->fo_filter_stats);
2052 cfs_spin_lock_init(&filter->fo_flags_lock);
2053 filter->fo_read_cache = 1; /* enable read-only cache by default */
2054 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2055 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2056 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2057 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2058 filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2059 filter_slc_set(filter); /* initialize sync on lock cancel */
2061 rc = filter_prep(obd);
2065 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2066 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2068 filter->fo_fl_oss_capa = 1;
2070 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2071 filter->fo_capa_hash = init_capa_hash();
2072 if (filter->fo_capa_hash == NULL)
2073 GOTO(err_post, rc = -ENOMEM);
2075 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2076 obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2077 LDLM_NAMESPACE_SERVER,
2078 LDLM_NAMESPACE_GREEDY,
2080 if (obd->obd_namespace == NULL)
2081 GOTO(err_post, rc = -ENOMEM);
2082 obd->obd_namespace->ns_lvbp = obd;
2083 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2084 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2086 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2087 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2089 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2091 CERROR("failed to setup llogging subsystems\n");
2095 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2096 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2097 /* do this after llog being initialized */
2098 filter_adapt_sptlrpc_conf(obd, 1);
2100 rc = lquota_setup(filter_quota_interface_ref, obd);
2104 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2105 if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2106 queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2107 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2108 "for larger I/O requests: max_sectors = %u "
2109 "could be up to max_hw_sectors=%u\n",
2110 obd->obd_name, mnt->mnt_sb->s_id,
2111 queue_max_sectors(q), queue_max_hw_sectors(q));
2113 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2114 if (uuid_ptr != NULL) {
2115 class_uuid_unparse(uuid_ptr, &uuid);
2121 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2122 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2123 obd->obd_name, label ?: str, lmi ? "on " : "",
2124 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2125 obd->obd_replayable ? "enabled" : "disabled");
2132 fsfilt_put_ops(obd->obd_fsops);
2133 filter_iobuf_pool_done(filter);
2135 server_put_mount(obd->obd_name, mnt);
2136 obd->u.obt.obt_sb = 0;
2140 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2142 struct lprocfs_static_vars lvars;
2143 cfs_proc_dir_entry_t *entry;
2149 CLASSERT(offsetof(struct obd_device, u.obt) ==
2150 offsetof(struct obd_device, u.filter.fo_obt));
2152 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2155 /* lprocfs must be setup before the filter so state can be safely added
2156 * to /proc incrementally as the filter is setup */
2157 lprocfs_filter_init_vars(&lvars);
2158 rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2160 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2165 rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2167 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2169 GOTO(obd_cleanup, rc);
2172 /* Init obdfilter private stats here */
2173 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2174 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2175 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2176 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2177 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2178 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2179 "get_page", "usec");
2180 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2181 LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2182 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2183 LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2184 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2185 LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2186 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2187 LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2189 rc = lproc_filter_attach_seqstat(obd);
2191 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2192 GOTO(free_obd_stats, rc);
2195 entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2196 if (IS_ERR(entry)) {
2197 rc = PTR_ERR(entry);
2198 CERROR("%s: error %d setting up lprocfs for %s\n",
2199 obd->obd_name, rc, "exports");
2200 GOTO(free_obd_stats, rc);
2202 obd->obd_proc_exports_entry = entry;
2204 entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2205 lprocfs_nid_stats_clear_read,
2206 lprocfs_nid_stats_clear_write, obd, NULL);
2207 if (IS_ERR(entry)) {
2208 rc = PTR_ERR(entry);
2209 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2211 GOTO(free_obd_stats, rc);
2214 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2215 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2217 GOTO(remove_entry_clear, rc = -ENOMEM);
2218 addr = (unsigned long)cfs_page_address(page);
2219 clear_page((void *)addr);
2220 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2221 LUSTRE_CFG_BUFLEN(lcfg, 4));
2222 rc = filter_common_setup(obd, lcfg, (void *)addr);
2223 OBD_PAGE_FREE(page);
2225 CERROR("%s: filter_common_setup failed: %d.\n",
2227 GOTO(remove_entry_clear, rc);
2233 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2235 lprocfs_free_obd_stats(obd);
2237 lprocfs_obd_cleanup(obd);
2241 static struct llog_operations filter_mds_ost_repl_logops;
2243 static struct llog_operations filter_size_orig_logops = {
2244 .lop_setup = llog_obd_origin_setup,
2245 .lop_cleanup = llog_obd_origin_cleanup,
2246 .lop_add = llog_obd_origin_add
2249 static int filter_olg_fini(struct obd_llog_group *olg)
2251 struct llog_ctxt *ctxt;
2252 int rc = 0, rc2 = 0;
2255 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2257 rc = llog_cleanup(ctxt);
2259 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2261 rc2 = llog_cleanup(ctxt);
2266 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2268 rc2 = llog_cleanup(ctxt);
2277 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2278 struct obd_device *tgt)
2283 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2284 &filter_mds_ost_repl_logops);
2288 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2289 &filter_size_orig_logops);
2295 filter_olg_fini(olg);
2300 * Init the default olg, which is embeded in the obd_device, for filter.
2303 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2304 struct obd_device *tgt)
2306 struct filter_obd *filter = &obd->u.filter;
2307 struct llog_ctxt *ctxt;
2311 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2312 if (!filter->fo_lcm)
2315 filter_mds_ost_repl_logops = llog_client_ops;
2316 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2317 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2318 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2320 rc = filter_olg_init(obd, olg, tgt);
2322 GOTO(cleanup_lcm, rc);
2324 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2327 GOTO(cleanup_olg, rc);
2329 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2331 CERROR("Can't get ctxt for %p:%x\n", olg,
2332 LLOG_MDS_OST_REPL_CTXT);
2333 GOTO(cleanup_olg, rc = -ENODEV);
2335 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2336 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2337 llog_ctxt_put(ctxt);
2341 filter_olg_fini(olg);
2343 llog_recov_thread_fini(filter->fo_lcm, 1);
2344 filter->fo_lcm = NULL;
2349 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2350 struct obd_device *tgt, int *index)
2352 struct filter_obd *filter = &obd->u.filter;
2353 struct llog_ctxt *ctxt;
2357 LASSERT(olg != NULL);
2358 if (olg == &obd->obd_olg)
2359 return filter_default_olg_init(obd, olg, tgt);
2361 LASSERT(filter->fo_lcm != NULL);
2362 rc = filter_olg_init(obd, olg, tgt);
2365 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2367 CERROR("Can't get ctxt for %p:%x\n", olg,
2368 LLOG_MDS_OST_REPL_CTXT);
2369 filter_olg_fini(olg);
2372 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2373 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2374 llog_ctxt_put(ctxt);
2378 static int filter_llog_finish(struct obd_device *obd, int count)
2380 struct filter_obd *filter = &obd->u.filter;
2381 struct llog_ctxt *ctxt;
2384 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2387 * Make sure that no cached llcds left in recov_thread.
2388 * We actually do sync in disconnect time, but disconnect
2389 * may not come being marked rq_no_resend = 1.
2391 llog_sync(ctxt, NULL);
2394 * Balance class_import_get() in llog_receptor_accept().
2395 * This is safe to do, as llog is already synchronized
2396 * and its import may go.
2398 cfs_mutex_down(&ctxt->loc_sem);
2399 if (ctxt->loc_imp) {
2400 class_import_put(ctxt->loc_imp);
2401 ctxt->loc_imp = NULL;
2403 cfs_mutex_up(&ctxt->loc_sem);
2404 llog_ctxt_put(ctxt);
2407 if (filter->fo_lcm) {
2408 cfs_mutex_down(&ctxt->loc_sem);
2409 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2410 filter->fo_lcm = NULL;
2411 cfs_mutex_up(&ctxt->loc_sem);
2413 RETURN(filter_olg_fini(&obd->obd_olg));
2417 * Find the group llog according to group index in the llog group list.
2419 static struct obd_llog_group *
2420 filter_find_olg_internal(struct filter_obd *filter, int group)
2422 struct obd_llog_group *olg;
2424 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2425 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2426 if (olg->olg_seq == group)
2433 * Find the group llog according to group index on the filter
2435 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2437 struct obd_llog_group *olg = NULL;
2438 struct filter_obd *filter;
2440 filter = &obd->u.filter;
2442 if (group == FID_SEQ_LLOG)
2443 RETURN(&obd->obd_olg);
2445 cfs_spin_lock(&filter->fo_llog_list_lock);
2446 olg = filter_find_olg_internal(filter, group);
2447 cfs_spin_unlock(&filter->fo_llog_list_lock);
2452 * Find the llog_group of the filter according to the group. If it can not
2453 * find, create the llog_group, which only happens when mds is being synced
2456 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2458 struct obd_llog_group *olg = NULL, *olg_new = NULL;
2459 struct filter_obd *filter;
2462 filter = &obd->u.filter;
2464 if (group == FID_SEQ_LLOG)
2465 RETURN(&obd->obd_olg);
2467 OBD_ALLOC_PTR(olg_new);
2468 if (olg_new == NULL)
2469 RETURN(ERR_PTR(-ENOMEM));
2471 cfs_spin_lock(&filter->fo_llog_list_lock);
2472 olg = filter_find_olg_internal(filter, group);
2474 if (olg->olg_initializing) {
2475 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2477 GOTO(out_unlock, olg);
2480 /* set as the newly allocated one */
2485 llog_group_init(olg, group);
2486 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2487 olg->olg_initializing = 1;
2488 cfs_spin_unlock(&filter->fo_llog_list_lock);
2490 rc = obd_llog_init(obd, olg, obd, NULL);
2492 cfs_spin_lock(&filter->fo_llog_list_lock);
2493 cfs_list_del(&olg->olg_list);
2494 cfs_spin_unlock(&filter->fo_llog_list_lock);
2496 GOTO(out, olg = ERR_PTR(-ENOMEM));
2498 cfs_spin_lock(&filter->fo_llog_list_lock);
2499 olg->olg_initializing = 0;
2500 cfs_spin_unlock(&filter->fo_llog_list_lock);
2501 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2502 obd->obd_name, group, olg);
2507 cfs_spin_unlock(&filter->fo_llog_list_lock);
2509 OBD_FREE_PTR(olg_new);
2513 static int filter_llog_connect(struct obd_export *exp,
2514 struct llogd_conn_body *body)
2516 struct obd_device *obd = exp->exp_obd;
2517 struct llog_ctxt *ctxt;
2518 struct obd_llog_group *olg;
2522 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2523 obd->obd_name, body->lgdc_logid.lgl_oid,
2524 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2526 olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2528 CERROR(" %s: can not find olg of group %d\n",
2529 obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2532 llog_group_set_export(olg, exp);
2534 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2535 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2536 body->lgdc_ctxt_idx);
2538 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2539 obd->obd_name, body->lgdc_logid.lgl_oid,
2540 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2542 cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2543 obd->u.filter.fo_mds_ost_sync = 1;
2544 cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2545 rc = llog_connect(ctxt, &body->lgdc_logid,
2546 &body->lgdc_gen, NULL);
2547 llog_ctxt_put(ctxt);
2549 CERROR("failed to connect rc %d idx %d\n", rc,
2550 body->lgdc_ctxt_idx);
2555 static int filter_llog_preclean(struct obd_device *obd)
2557 struct obd_llog_group *olg, *tmp;
2558 struct filter_obd *filter;
2559 cfs_list_t remove_list;
2563 rc = obd_llog_finish(obd, 0);
2565 CERROR("failed to cleanup llogging subsystem\n");
2567 filter = &obd->u.filter;
2568 CFS_INIT_LIST_HEAD(&remove_list);
2570 cfs_spin_lock(&filter->fo_llog_list_lock);
2571 while (!cfs_list_empty(&filter->fo_llog_list)) {
2572 olg = cfs_list_entry(filter->fo_llog_list.next,
2573 struct obd_llog_group, olg_list);
2574 cfs_list_del(&olg->olg_list);
2575 cfs_list_add(&olg->olg_list, &remove_list);
2577 cfs_spin_unlock(&filter->fo_llog_list_lock);
2579 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2580 cfs_list_del_init(&olg->olg_list);
2581 rc = filter_olg_fini(olg);
2583 CERROR("failed to cleanup llogging subsystem for %u\n",
2591 static int filter_precleanup(struct obd_device *obd,
2592 enum obd_cleanup_stage stage)
2598 case OBD_CLEANUP_EARLY:
2600 case OBD_CLEANUP_EXPORTS:
2601 /* Stop recovery before namespace cleanup. */
2602 target_recovery_fini(obd);
2603 rc = filter_llog_preclean(obd);
2609 static int filter_cleanup(struct obd_device *obd)
2611 struct filter_obd *filter = &obd->u.filter;
2615 LCONSOLE_WARN("%s: shutting down for failover; client state "
2616 "will be preserved.\n", obd->obd_name);
2618 obd_exports_barrier(obd);
2619 obd_zombie_barrier();
2621 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2622 lprocfs_free_per_client_stats(obd);
2623 lprocfs_free_obd_stats(obd);
2624 lprocfs_obd_cleanup(obd);
2625 lquota_cleanup(filter_quota_interface_ref, obd);
2627 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2628 obd->obd_namespace = NULL;
2630 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2632 if (obd->u.obt.obt_sb == NULL)
2637 ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2638 shrink_dcache_sb(obd->u.obt.obt_sb);
2640 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2641 obd->u.obt.obt_sb = NULL;
2643 fsfilt_put_ops(obd->obd_fsops);
2645 filter_iobuf_pool_done(filter);
2647 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2652 static int filter_connect_internal(struct obd_export *exp,
2653 struct obd_connect_data *data,
2656 struct filter_export_data *fed = &exp->exp_filter_data;
2661 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2662 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2663 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2664 data->ocd_connect_flags, data->ocd_version,
2665 data->ocd_grant, data->ocd_index);
2667 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2668 CWARN("!!! This export (nid %s) used object group %d "
2669 "earlier; now it's trying to use group %d! This could "
2670 "be a bug in the MDS. Please report to "
2671 "http://bugs.whamcloud.com/\n",
2672 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2675 fed->fed_group = data->ocd_group;
2677 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2678 exp->exp_connect_flags = data->ocd_connect_flags;
2679 data->ocd_version = LUSTRE_VERSION_CODE;
2681 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2682 if (data->ocd_connect_flags & OBD_CONNECT_MDS)
2683 CWARN("%s: Received MDS connection for group %u\n",
2684 exp->exp_obd->obd_name, data->ocd_group);
2685 else if (data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN)
2688 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2689 struct filter_obd *filter = &exp->exp_obd->u.filter;
2690 obd_size left, want;
2692 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2693 left = filter_grant_space_left(exp);
2694 want = data->ocd_grant;
2695 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2696 data->ocd_grant = fed->fed_grant;
2697 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2699 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2700 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2701 exp->exp_client_uuid.uuid, exp,
2702 data->ocd_grant, want, left);
2704 filter->fo_tot_granted_clients ++;
2707 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2708 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2709 int index = le32_to_cpu(lsd->lsd_ost_index);
2711 if (!(lsd->lsd_feature_compat &
2712 cpu_to_le32(OBD_COMPAT_OST))) {
2713 /* this will only happen on the first connect */
2714 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2715 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2716 /* sync is not needed here as filter_client_add will
2717 * set exp_need_sync flag */
2718 filter_update_server_data(exp->exp_obd);
2719 } else if (index != data->ocd_index) {
2720 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2721 " %u doesn't match actual OST index"
2722 " %u in last_rcvd file, bad "
2724 obd_export_nid2str(exp), index,
2728 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2729 * FIXME: We don't strictly need the COMPAT flag for that,
2730 * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2731 * FIXME: We needed it for the index, as index 0 is valid. */
2734 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2735 data->ocd_brw_size = 65536;
2736 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2737 data->ocd_brw_size = min(data->ocd_brw_size,
2738 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2739 if (data->ocd_brw_size == 0) {
2740 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2741 " ocd_version: %x ocd_grant: %d ocd_index: %u "
2742 "ocd_brw_size is unexpectedly zero, "
2743 "network data corruption?"
2744 "Refusing connection of this client\n",
2745 exp->exp_obd->obd_name,
2746 exp->exp_client_uuid.uuid,
2747 exp, data->ocd_connect_flags, data->ocd_version,
2748 data->ocd_grant, data->ocd_index);
2753 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2754 __u32 cksum_types = data->ocd_cksum_types;
2756 /* The client set in ocd_cksum_types the checksum types it
2757 * supports. We have to mask off the algorithms that we don't
2759 data->ocd_cksum_types &= cksum_types_supported();
2761 /* 1.6.4- only support CRC32 and didn't set ocd_cksum_types */
2762 if (unlikely(data->ocd_cksum_types == 0))
2763 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2765 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2766 "%x\n", exp->exp_obd->obd_name,
2767 obd_export_nid2str(exp), cksum_types,
2768 data->ocd_cksum_types);
2770 /* This client does not support OBD_CONNECT_CKSUM
2771 * fall back to CRC32 */
2772 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2773 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2774 exp->exp_obd->obd_name,
2775 obd_export_nid2str(exp));
2778 if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2779 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2784 static int filter_reconnect(const struct lu_env *env,
2785 struct obd_export *exp, struct obd_device *obd,
2786 struct obd_uuid *cluuid,
2787 struct obd_connect_data *data,
2793 if (exp == NULL || obd == NULL || cluuid == NULL)
2796 rc = filter_connect_internal(exp, data, 1);
2798 filter_export_stats_init(obd, exp, localdata);
2803 static int filter_connect(const struct lu_env *env,
2804 struct obd_export **exp, struct obd_device *obd,
2805 struct obd_uuid *cluuid,
2806 struct obd_connect_data *data, void *localdata)
2808 struct lvfs_run_ctxt saved;
2809 struct lustre_handle conn = { 0 };
2810 struct obd_export *lexp;
2814 if (exp == NULL || obd == NULL || cluuid == NULL)
2817 rc = class_connect(&conn, obd, cluuid);
2820 lexp = class_conn2export(&conn);
2821 LASSERT(lexp != NULL);
2823 rc = filter_connect_internal(lexp, data, 0);
2827 filter_export_stats_init(obd, lexp, localdata);
2828 if (obd->obd_replayable) {
2829 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2831 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2832 rc = filter_client_add(obd, lexp, -1);
2837 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2838 rc = filter_read_groups(obd, data->ocd_group, 1);
2839 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2841 CERROR("can't read group %u\n", data->ocd_group);
2849 class_disconnect(lexp);
2858 /* Do extra sanity checks for grant accounting. We do this at connect,
2859 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2860 * always get rid of it or turn it off when we know accounting is good. */
2861 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2863 struct filter_export_data *fed;
2864 struct obd_export *exp;
2865 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2866 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2867 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2869 if (cfs_list_empty(&obd->obd_exports))
2872 /* We don't want to do this for large machines that do lots of
2873 mounts or unmounts. It burns... */
2874 if (obd->obd_num_exports > 100)
2877 cfs_spin_lock(&obd->obd_osfs_lock);
2878 cfs_spin_lock(&obd->obd_dev_lock);
2879 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2881 fed = &exp->exp_filter_data;
2882 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2885 if (maxsize > 0) { /* we may not have done a statfs yet */
2886 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2887 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2888 exp->exp_client_uuid.uuid, exp,
2889 fed->fed_grant, fed->fed_pending, maxsize);
2890 LASSERTF(fed->fed_dirty <= maxsize,
2891 "%s: cli %s/%p %ld > "LPU64"\n", func,
2892 exp->exp_client_uuid.uuid, exp,
2893 fed->fed_dirty, maxsize);
2896 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2897 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2898 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2900 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2901 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2902 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2903 tot_granted += fed->fed_grant + fed->fed_pending;
2904 tot_pending += fed->fed_pending;
2905 tot_dirty += fed->fed_dirty;
2907 fo_tot_granted = obd->u.filter.fo_tot_granted;
2908 fo_tot_pending = obd->u.filter.fo_tot_pending;
2909 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2910 cfs_spin_unlock(&obd->obd_dev_lock);
2911 cfs_spin_unlock(&obd->obd_osfs_lock);
2913 /* Do these assertions outside the spinlocks so we don't kill system */
2914 if (tot_granted != fo_tot_granted)
2915 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2916 func, tot_granted, fo_tot_granted);
2917 if (tot_pending != fo_tot_pending)
2918 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2919 func, tot_pending, fo_tot_pending);
2920 if (tot_dirty != fo_tot_dirty)
2921 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2922 func, tot_dirty, fo_tot_dirty);
2923 if (tot_pending > tot_granted)
2924 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2925 func, tot_pending, tot_granted);
2926 if (tot_granted > maxsize)
2927 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2928 func, tot_granted, maxsize);
2929 if (tot_dirty > maxsize)
2930 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2931 func, tot_dirty, maxsize);
2934 /* Remove this client from the grant accounting totals. We also remove
2935 * the export from the obd device under the osfs and dev locks to ensure
2936 * that the filter_grant_sanity_check() calculations are always valid.
2937 * The client should do something similar when it invalidates its import. */
2938 static void filter_grant_discard(struct obd_export *exp)
2940 struct obd_device *obd = exp->exp_obd;
2941 struct filter_obd *filter = &obd->u.filter;
2942 struct filter_export_data *fed = &exp->exp_filter_data;
2944 cfs_spin_lock(&obd->obd_osfs_lock);
2945 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2946 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2947 obd->obd_name, filter->fo_tot_granted,
2948 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2949 filter->fo_tot_granted -= fed->fed_grant;
2950 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2951 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2952 obd->obd_name, filter->fo_tot_pending,
2953 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2954 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2955 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2956 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2957 obd->obd_name, filter->fo_tot_dirty,
2958 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2959 filter->fo_tot_dirty -= fed->fed_dirty;
2963 cfs_spin_unlock(&obd->obd_osfs_lock);
2966 static int filter_destroy_export(struct obd_export *exp)
2968 struct filter_export_data *fed = &exp->exp_filter_data;
2971 if (fed->fed_pending)
2972 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2973 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2974 exp, fed->fed_pending);
2976 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2978 target_destroy_export(exp);
2979 ldlm_destroy_export(exp);
2980 lut_client_free(exp);
2982 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2985 if (!exp->exp_obd->obd_replayable)
2986 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2988 filter_grant_discard(exp);
2989 filter_fmd_cleanup(exp);
2991 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
2992 struct filter_obd *filter = &exp->exp_obd->u.filter;
2993 if (filter->fo_tot_granted_clients > 0)
2994 filter->fo_tot_granted_clients --;
2997 if (!(exp->exp_flags & OBD_OPT_FORCE))
2998 filter_grant_sanity_check(exp->exp_obd, __func__);
3003 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3005 struct obd_llog_group *olg_min, *olg;
3006 struct filter_obd *filter;
3007 int worked = -1, group;
3008 struct llog_ctxt *ctxt;
3011 filter = &obd->u.filter;
3013 /* we can't sync log holding spinlock. also, we do not want to get
3014 * into livelock. so we do following: loop over MDS's exports in
3015 * group order and skip already synced llogs -bzzz */
3017 /* look for group with min. number, but > worked */
3020 cfs_spin_lock(&filter->fo_llog_list_lock);
3021 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3022 if (olg->olg_seq <= worked) {
3023 /* this group is already synced */
3026 if (group < olg->olg_seq) {
3027 /* we have group with smaller number to sync */
3030 /* store current minimal group */
3032 group = olg->olg_seq;
3034 cfs_spin_unlock(&filter->fo_llog_list_lock);
3036 if (olg_min == NULL)
3039 worked = olg_min->olg_seq;
3040 if (olg_min->olg_exp &&
3041 (dexp == olg_min->olg_exp || dexp == NULL)) {
3043 ctxt = llog_group_get_ctxt(olg_min,
3044 LLOG_MDS_OST_REPL_CTXT);
3046 err = llog_sync(ctxt, olg_min->olg_exp);
3047 llog_ctxt_put(ctxt);
3049 CERROR("error flushing logs to MDS: "
3054 } while (olg_min != NULL);
3057 /* Also incredibly similar to mds_disconnect */
3058 static int filter_disconnect(struct obd_export *exp)
3060 struct obd_device *obd = exp->exp_obd;
3065 class_export_get(exp);
3067 if (!(exp->exp_flags & OBD_OPT_FORCE))
3068 filter_grant_sanity_check(obd, __func__);
3069 filter_grant_discard(exp);
3071 /* Flush any remaining cancel messages out to the target */
3072 filter_sync_llogs(obd, exp);
3074 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3076 rc = server_disconnect_export(exp);
3078 /* Do not erase record for recoverable client. */
3079 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3080 filter_client_del(exp);
3082 fsfilt_sync(obd, obd->u.obt.obt_sb);
3084 class_export_put(exp);
3088 /* reverse import is changed, sync all cancels */
3089 static void filter_revimp_update(struct obd_export *exp)
3094 class_export_get(exp);
3096 /* flush any remaining cancel messages out to the target */
3097 filter_sync_llogs(exp->exp_obd, exp);
3098 class_export_put(exp);
3102 static int filter_ping(struct obd_export *exp)
3104 filter_fmd_expire(exp);
3108 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3109 const char *what, int quiet)
3111 struct dentry *dchild = NULL;
3113 dchild = filter_fid2dentry(obd, NULL, ostid->oi_seq, ostid->oi_id);
3115 if (IS_ERR(dchild)) {
3116 CERROR("%s error looking up object: "POSTID"\n",
3117 what, ostid->oi_id, ostid->oi_seq);
3121 if (dchild->d_inode == NULL) {
3123 CERROR("%s: %s on non-existent object: "POSTID" \n",
3124 obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3126 RETURN(ERR_PTR(-ENOENT));
3132 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3134 struct dentry *dentry = NULL;
3135 struct obd_device *obd;
3139 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3140 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3144 obd = class_exp2obd(exp);
3146 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3150 dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3152 RETURN(PTR_ERR(dentry));
3154 /* Limit the valid bits in the return data to what we actually use */
3155 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3156 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3162 /* this should be enabled/disabled in condition to enabled/disabled large
3163 * inodes (fast EAs) in backing store FS. */
3164 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3165 void *handle, struct obdo *oa)
3167 struct obd_device *obd = exp->exp_obd;
3171 if (oa->o_valid & OBD_MD_FLFID) {
3172 struct filter_fid ff;
3174 if (!(oa->o_valid & OBD_MD_FLGROUP))
3176 /* packing fid and converting it to LE for storing into EA.
3177 * Here ->o_stripe_idx should be filled by LOV and rest of
3178 * fields - by client. */
3179 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3180 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3181 /* XXX: we are ignoring o_parent_ver here, since this should
3182 * be the same for all objects in this fileset. */
3183 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3184 ff.ff_objid = cpu_to_le64(oa->o_id);
3185 ff.ff_seq = cpu_to_le64(oa->o_seq);
3187 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3188 LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3191 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3193 CERROR("store fid in object failed! rc: %d\n", rc);
3195 CDEBUG(D_HA, "OSS object without fid info!\n");
3201 /* this is called from filter_truncate() until we have filter_punch() */
3202 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3203 struct obdo *oa, struct obd_trans_info *oti)
3205 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3206 struct llog_cookie *fcc = NULL;
3207 struct filter_obd *filter;
3208 int rc, err, sync = 0;
3209 loff_t old_size = 0;
3210 unsigned int ia_valid;
3211 struct inode *inode;
3212 struct page *page = NULL;
3217 LASSERT(dentry != NULL);
3218 LASSERT(!IS_ERR(dentry));
3220 inode = dentry->d_inode;
3221 LASSERT(inode != NULL);
3223 filter = &exp->exp_obd->u.filter;
3224 iattr_from_obdo(&iattr, oa, oa->o_valid);
3225 ia_valid = iattr.ia_valid;
3227 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3228 OBD_ALLOC(fcc, sizeof(*fcc));
3230 *fcc = oa->o_lcookie;
3232 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3233 unsigned long now = jiffies;
3234 ll_vfs_dq_init(inode);
3235 /* Filter truncates and writes are serialized by
3236 * i_alloc_sem, see the comment in
3237 * filter_preprw_write.*/
3238 if (ia_valid & ATTR_SIZE)
3239 down_write(&inode->i_alloc_sem);
3240 LOCK_INODE_MUTEX(inode);
3241 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
3242 old_size = i_size_read(inode);
3245 /* VBR: version recovery check */
3246 rc = filter_version_get_check(exp, oti, inode);
3248 GOTO(out_unlock, rc);
3250 /* Let's pin the last page so that ldiskfs_truncate
3251 * should not start GFP_FS allocation. */
3252 if (ia_valid & ATTR_SIZE) {
3253 page = grab_cache_page(inode->i_mapping,
3254 iattr.ia_size >> PAGE_CACHE_SHIFT);
3256 GOTO(out_unlock, rc = -ENOMEM);
3261 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3262 * then we will accept the UID+GID sent by the client during write for
3263 * initializing the ownership of this inode. We only allow this to
3264 * happen once so clear these bits in setattr. In 2.6 kernels it is
3265 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3266 * the flags that are actually being set. */
3267 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3268 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3269 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3271 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3272 if (!(ia_valid & ATTR_MODE)) {
3273 iattr.ia_mode = inode->i_mode;
3274 iattr.ia_valid |= ATTR_MODE;
3276 iattr.ia_mode &= ~S_ISUID;
3278 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3279 if (!(iattr.ia_valid & ATTR_MODE)) {
3280 iattr.ia_mode = inode->i_mode;
3281 iattr.ia_valid |= ATTR_MODE;
3283 iattr.ia_mode &= ~S_ISGID;
3286 orig_ids[USRQUOTA] = inode->i_uid;
3287 orig_ids[GRPQUOTA] = inode->i_gid;
3288 handle = fsfilt_start_log(exp->exp_obd, inode,
3289 FSFILT_OP_SETATTR, oti, 1);
3291 GOTO(out_unlock, rc = PTR_ERR(handle));
3293 /* update inode EA only once when inode is suid bit marked. As
3294 * on 2.6.x UID and GID may be set separately, we check here
3295 * only one of them to avoid double setting. */
3296 if (inode->i_mode & S_ISUID)
3297 filter_update_fidea(exp, inode, handle, oa);
3299 handle = fsfilt_start(exp->exp_obd, inode,
3300 FSFILT_OP_SETATTR, oti);
3302 GOTO(out_unlock, rc = PTR_ERR(handle));
3304 if (oa->o_valid & OBD_MD_FLFLAGS) {
3305 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3306 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3308 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3310 /* set cancel cookie callback function */
3311 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3312 filter_cancel_cookies_cb,
3316 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3317 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3319 /* The truncate might have used up our transaction credits. Make sure
3320 * we have two left for the last_rcvd and VBR inode version updates. */
3321 err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3323 /* Update inode version only if data has changed => size has changed */
3324 rc = filter_finish_transno(exp, ia_valid & ATTR_SIZE ? inode : NULL,
3328 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3332 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3334 CERROR("error on commit, err = %d\n", err);
3341 /* For a partial-page truncate flush the page to disk immediately
3342 * to avoid data corruption during direct disk write. b=17397 */
3343 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3344 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3345 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3355 page_cache_release(page);
3357 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3358 UNLOCK_INODE_MUTEX(inode);
3359 if (ia_valid & ATTR_SIZE)
3360 up_write(&inode->i_alloc_sem);
3362 OBD_FREE(fcc, sizeof(*fcc));
3364 /* trigger quota release */
3365 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3366 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3367 int rc2 = lquota_adjust(filter_quota_interface_ref,
3368 exp->exp_obd, cur_ids,
3369 orig_ids, rc, FSFILT_OP_SETATTR);
3370 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3371 "filter adjust qunit. (rc:%d)\n", rc2);
3376 /* this is called from filter_truncate() until we have filter_punch() */
3377 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3378 struct obd_trans_info *oti)
3380 struct obdo *oa = oinfo->oi_oa;
3381 struct lustre_capa *capa = oinfo_capa(oinfo);
3382 struct ldlm_res_id res_id;
3383 struct filter_mod_data *fmd;
3384 struct lvfs_run_ctxt saved;
3385 struct filter_obd *filter;
3386 struct ldlm_resource *res;
3387 struct dentry *dentry;
3388 __u64 opc = CAPA_OPC_META_WRITE;
3392 if (oa->o_valid & OBD_FL_TRUNC)
3393 opc |= CAPA_OPC_OSS_TRUNC;
3395 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3399 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3400 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3405 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3406 /* This would be very bad - accidentally truncating a file when
3407 * changing the time or similar - bug 12203. */
3408 if (oa->o_valid & OBD_MD_FLSIZE &&
3409 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3410 static char mdsinum[48];
3412 if (oa->o_valid & OBD_MD_FLFID)
3413 snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3414 oa->o_parent_seq, oa->o_parent_oid,
3419 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3420 "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3421 oa->o_id, oa->o_seq, mdsinum);
3425 dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3427 RETURN(PTR_ERR(dentry));
3429 filter = &exp->exp_obd->u.filter;
3430 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3433 * We need to be atomic against a concurrent write
3434 * (which takes the semaphore for reading). fmd_mactime_xid
3435 * checks will have no effect if a write request with lower
3436 * xid starts just before a setattr and finishes later than
3437 * the setattr (see bug 21489, comment 27).
3440 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3441 unsigned long now = jiffies;
3442 down_write(&dentry->d_inode->i_alloc_sem);
3443 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem");
3444 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3445 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3446 fmd->fmd_mactime_xid = oti->oti_xid;
3447 filter_fmd_put(exp, fmd);
3448 up_write(&dentry->d_inode->i_alloc_sem);
3451 /* setting objects attributes (including owner/group) */
3452 rc = filter_setattr_internal(exp, dentry, oa, oti);
3454 GOTO(out_unlock, rc);
3456 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3457 &res_id, LDLM_EXTENT, 0);
3460 LDLM_RESOURCE_ADDREF(res);
3461 rc = ldlm_res_lvbo_update(res, NULL, 0);
3462 LDLM_RESOURCE_DELREF(res);
3463 ldlm_resource_putref(res);
3466 oa->o_valid = OBD_MD_FLID;
3468 /* Quota release need uid/gid info */
3469 obdo_from_inode(oa, dentry->d_inode, NULL,
3470 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3475 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3479 /* XXX identical to osc_unpackmd */
3480 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3481 struct lov_mds_md *lmm, int lmm_bytes)
3487 if (lmm_bytes < sizeof (*lmm)) {
3488 CERROR("lov_mds_md too small: %d, need %d\n",
3489 lmm_bytes, (int)sizeof(*lmm));
3492 /* XXX LOV_MAGIC etc check? */
3494 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3495 CERROR("lov_mds_md: zero lmm_object_id\n");
3500 lsm_size = lov_stripe_md_size(1);
3504 if (*lsmp != NULL && lmm == NULL) {
3505 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3506 OBD_FREE(*lsmp, lsm_size);
3511 if (*lsmp == NULL) {
3512 OBD_ALLOC(*lsmp, lsm_size);
3516 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3517 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3518 OBD_FREE(*lsmp, lsm_size);
3521 loi_init((*lsmp)->lsm_oinfo[0]);
3525 /* XXX zero *lsmp? */
3526 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3527 LASSERT((*lsmp)->lsm_object_id);
3530 (*lsmp)->lsm_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
3535 /* caller must hold fo_create_locks[oa->o_seq] */
3536 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3537 struct filter_obd *filter)
3539 struct obdo doa = { 0 }; /* XXX obdo on stack */
3545 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_seq]) != 0);
3547 memset(&doa, 0, sizeof(doa));
3549 doa.o_valid |= OBD_MD_FLGROUP;
3550 doa.o_seq = oa->o_seq;
3551 doa.o_mode = S_IFREG;
3553 if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3554 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3555 exp->exp_obd->obd_name, doa.o_seq);
3559 last = filter_last_id(filter, doa.o_seq);
3561 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3563 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3564 exp->exp_obd->obd_name, oa->o_id + 1, last,
3565 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3567 for (id = last; id > oa->o_id; id--) {
3569 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3570 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3571 CEMERG("error destroying precreate objid "LPU64": %d\n",
3574 /* update last_id on disk periodically so that if we restart
3575 * we don't need to re-scan all of the just-deleted objects. */
3576 if ((id & 511) == 0 && !skip_orphan) {
3577 filter_set_last_id(filter, id - 1, doa.o_seq);
3578 filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3582 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3583 exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3586 filter_set_last_id(filter, id, doa.o_seq);
3587 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3590 * We have destroyed orphan objects, but don't want to reuse
3591 * them. Therefore we don't reset last_id to the last created
3592 * objects. Instead, we report back to the MDS the object id
3593 * of the last orphan, so that the MDS can restart allocating
3594 * objects from this id + 1 and thus skip the whole orphan
3600 cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3605 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3606 obd_seq group, int *num);
3607 /* returns a negative error or a nonnegative number of files to create */
3608 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3609 obd_seq group, struct obd_trans_info *oti)
3611 struct obd_device *obd = exp->exp_obd;
3612 struct filter_obd *filter = &obd->u.filter;
3616 /* delete orphans request */
3617 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3618 obd_id last = filter_last_id(filter, group);
3620 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3621 CERROR("%s: dropping old orphan cleanup request\n",
3625 /* This causes inflight precreates to abort and drop lock */
3626 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3627 cfs_down(&filter->fo_create_locks[group]);
3628 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3629 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3630 exp->exp_obd->obd_name, group);
3631 cfs_up(&filter->fo_create_locks[group]);
3634 diff = oa->o_id - last;
3635 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3638 if (-diff > OST_MAX_PRECREATE) {
3639 CERROR("%s: ignoring bogus orphan destroy request: "
3640 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3642 /* FIXME: should reset precreate_next_id on MDS */
3643 GOTO(out, rc = -EINVAL);
3646 rc = filter_destroy_precreated(exp, oa, filter);
3648 CERROR("%s: unable to write lastobjid, but "
3649 "orphans were deleted\n", obd->obd_name);
3652 /* XXX: Used by MDS for the first time! */
3653 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3656 cfs_down(&filter->fo_create_locks[group]);
3657 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3658 CERROR("%s: dropping old precreate request\n",
3662 /* only precreate if group == 0 and o_id is specfied */
3663 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3666 diff = oa->o_id - filter_last_id(filter, group);
3667 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3668 filter_last_id(filter, group), diff);
3670 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3671 oa->o_id, filter_last_id(filter, group), diff);
3675 oa->o_id = filter_last_id(&obd->u.filter, group);
3676 rc = filter_precreate(obd, oa, group, &diff);
3677 oa->o_id = filter_last_id(&obd->u.filter, group);
3679 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3682 /* else diff == 0 */
3685 cfs_up(&filter->fo_create_locks[group]);
3689 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3690 __u64 max_age, __u32 flags)
3692 struct filter_obd *filter = &obd->u.filter;
3693 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3697 /* at least try to account for cached pages. its still racey and
3698 * might be under-reporting if clients haven't announced their
3699 * caches with brw recently */
3700 cfs_spin_lock(&obd->obd_osfs_lock);
3701 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3702 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3703 cfs_spin_unlock(&obd->obd_osfs_lock);
3705 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3706 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3707 filter->fo_tot_dirty, filter->fo_tot_granted,
3708 filter->fo_tot_pending,
3709 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3711 filter_grant_sanity_check(obd, __func__);
3713 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3714 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3715 osfs->os_bsize - 1) >> blockbits));
3717 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3718 struct lr_server_data *lsd = class_server_data(obd);
3719 int index = le32_to_cpu(lsd->lsd_ost_index);
3721 if (cfs_fail_val == -1 || index == cfs_fail_val)
3722 osfs->os_bfree = osfs->os_bavail = 2;
3723 else if (cfs_fail_loc & OBD_FAIL_ONCE)
3724 cfs_fail_loc &= ~OBD_FAILED; /* reset flag */
3727 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3728 * stop creating files on MDS if OST is not good shape to create
3732 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3733 osfs->os_state = OS_STATE_READONLY;
3735 if (filter->fo_raid_degraded)
3736 osfs->os_state |= OS_STATE_DEGRADED;
3740 static int filter_use_existing_obj(struct obd_device *obd,
3741 struct dentry *dchild, void **handle,
3744 struct inode *inode = dchild->d_inode;
3748 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3751 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3752 if (IS_ERR(*handle))
3753 return PTR_ERR(*handle);
3755 iattr.ia_valid = ATTR_MODE;
3756 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3757 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3764 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3767 __u64 os_ffree = -1;
3769 cfs_spin_lock(&obd->obd_osfs_lock);
3770 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3772 os_ffree = obd->obd_osfs.os_ffree;
3773 cfs_spin_unlock(&obd->obd_osfs_lock);
3778 /* We rely on the fact that only one thread will be creating files in a given
3779 * group at a time, which is why we don't need an atomic filter_get_new_id.
3780 * Even if we had that atomic function, the following race would exist:
3782 * thread 1: gets id x from filter_next_id
3783 * thread 2: gets id (x + 1) from filter_next_id
3784 * thread 2: creates object (x + 1)
3785 * thread 1: tries to create object x, gets -ENOSPC
3787 * Caller must hold fo_create_locks[group]
3789 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3790 obd_seq group, int *num)
3792 struct dentry *dchild = NULL, *dparent = NULL;
3793 struct filter_obd *filter;
3794 struct obd_statfs *osfs;
3796 int err = 0, rc = 0, recreate_obj = 0, i;
3797 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3800 void *handle = NULL;
3803 filter = &obd->u.filter;
3805 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3807 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3809 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3810 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3813 OBD_ALLOC(osfs, sizeof(*osfs));
3816 rc = filter_statfs(obd, osfs,
3817 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
3819 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3820 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3821 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3822 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3824 if (oa->o_valid & OBD_MD_FLFLAGS)
3825 oa->o_flags |= OBD_FL_NOSPC_BLK;
3827 oa->o_valid |= OBD_MD_FLFLAGS;
3828 oa->o_flags = OBD_FL_NOSPC_BLK;
3833 OBD_FREE(osfs, sizeof(*osfs));
3838 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3839 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3841 for (i = 0; i < *num && err == 0; i++) {
3842 int cleanup_phase = 0;
3844 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3845 CWARN("%s: create aborted by destroy\n",
3854 last_id = filter_last_id(filter, group);
3855 if (next_id > last_id) {
3856 CERROR("Error: Trying to recreate obj greater"
3857 "than last id "LPD64" > "LPD64"\n",
3859 GOTO(cleanup, rc = -EINVAL);
3862 next_id = filter_last_id(filter, group) + 1;
3864 /* Don't create objects beyond the valid range for this SEQ */
3865 if (unlikely(fid_seq_is_mdt0(group) &&
3866 next_id >= IDIF_MAX_OID)) {
3867 CERROR("%s:"POSTID" hit the IDIF_MAX_OID (1<<48)!\n",
3868 obd->obd_name, next_id, group);
3869 GOTO(cleanup, rc = -ENOSPC);
3870 } else if (unlikely(!fid_seq_is_mdt0(group) &&
3871 next_id >= OBIF_MAX_OID)) {
3872 CERROR("%s:"POSTID" hit the OBIF_MAX_OID (1<<32)!\n",
3873 obd->obd_name, next_id, group);
3874 GOTO(cleanup, rc = -ENOSPC);
3877 dparent = filter_parent_lock(obd, group, next_id);
3878 if (IS_ERR(dparent))
3879 GOTO(cleanup, rc = PTR_ERR(dparent));
3880 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3882 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3884 GOTO(cleanup, rc = PTR_ERR(dchild));
3885 cleanup_phase = 2; /* f_dput(dchild) */
3887 if (dchild->d_inode != NULL) {
3888 /* This would only happen if lastobjid was bad on disk*/
3889 /* Could also happen if recreating missing obj but it
3890 * already exists. */
3892 CERROR("%s: recreating existing object %.*s?\n",
3893 obd->obd_name, dchild->d_name.len,
3894 dchild->d_name.name);
3896 /* Use these existing objects if they are
3898 if (dchild->d_inode->i_size == 0) {
3899 rc = filter_use_existing_obj(obd,dchild,
3900 &handle, &cleanup_phase);
3907 CERROR("%s: Serious error: objid %.*s already "
3908 "exists; is this filesystem corrupt?\n",
3909 obd->obd_name, dchild->d_name.len,
3910 dchild->d_name.name);
3913 GOTO(cleanup, rc = -EEXIST);
3916 handle = fsfilt_start_log(obd, dparent->d_inode,
3917 FSFILT_OP_CREATE, NULL, 1);
3919 GOTO(cleanup, rc = PTR_ERR(handle));
3922 CDEBUG(D_INODE, "%s: filter_precreate(od->o_seq="LPU64
3923 ",od->o_id="LPU64")\n", obd->obd_name, group,
3926 /* We mark object SUID+SGID to flag it for accepting UID+GID
3927 * from client on first write. Currently the permission bits
3928 * on the OST are never used, so this is OK. */
3929 rc = ll_vfs_create(dparent->d_inode, dchild,
3930 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3932 CERROR("create failed rc = %d\n", rc);
3933 if (rc == -ENOSPC) {
3934 os_ffree = filter_calc_free_inodes(obd);
3938 if (obd->obd_osfs.os_bavail <
3939 (obd->obd_osfs.os_blocks >> 10)) {
3940 if (oa->o_valid & OBD_MD_FLFLAGS)
3941 oa->o_flags |= OBD_FL_NOSPC_BLK;
3943 oa->o_valid |= OBD_MD_FLFLAGS;
3944 oa->o_flags = OBD_FL_NOSPC_BLK;
3947 CERROR("%s: free inode "LPU64"\n",
3948 obd->obd_name, os_ffree);
3954 if (dchild->d_inode)
3955 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3956 dchild->d_inode->i_ino);
3959 /* Set a/c/m time to a insane large negative value at creation
3960 * time so that any timestamp arriving from the client will
3961 * always be newer and update the inode.
3962 * See LU-221 for details */
3963 iattr.ia_valid = ATTR_ATIME | ATTR_MTIME | ATTR_CTIME;
3964 LTIME_S(iattr.ia_atime) = INT_MIN + 24 * 3600;
3965 LTIME_S(iattr.ia_mtime) = INT_MIN + 24 * 3600;
3966 LTIME_S(iattr.ia_ctime) = INT_MIN + 24 * 3600;
3967 err = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
3969 CERROR("unable to initialize a/c/m time of newly"
3972 if (!recreate_obj) {
3973 filter_set_last_id(filter, next_id, group);
3974 err = filter_update_last_objid(obd, group, 0);
3976 CERROR("unable to write lastobjid "
3977 "but file created\n");
3981 switch(cleanup_phase) {
3983 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3985 CERROR("error on commit, err = %d\n", err);
3992 filter_parent_unlock(dparent);
3999 if (cfs_time_after(jiffies, enough_time)) {
4002 "%s: precreate slow - want %d got %d \n",
4003 obd->obd_name, *num, i);
4010 "%s: created %d objects for group "POSTID" rc %d\n",
4011 obd->obd_name, i, filter->fo_last_objids[group], group, rc);
4016 int filter_create(struct obd_export *exp, struct obdo *oa,
4017 struct lov_stripe_md **ea, struct obd_trans_info *oti)
4019 struct obd_device *obd = exp->exp_obd;
4020 struct filter_export_data *fed;
4021 struct filter_obd *filter;
4022 struct lvfs_run_ctxt saved;
4023 struct lov_stripe_md *lsm = NULL;
4027 CDEBUG(D_INODE, "%s: filter_create(group="LPU64",id="
4028 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4030 fed = &exp->exp_filter_data;
4031 filter = &obd->u.filter;
4033 /* 1.8 client doesn't carry the ocd_group with connect request,
4034 * so the fed_group will always be zero for 1.8 client. */
4035 if (!(exp->exp_connect_flags & OBD_CONNECT_FULL20)) {
4036 if (oa->o_seq != FID_SEQ_OST_MDT0 &&
4037 oa->o_seq != FID_SEQ_LLOG &&
4038 oa->o_seq != FID_SEQ_ECHO) {
4039 CERROR("The request from older client has invalid"
4040 " group "LPU64"!\n", oa->o_seq);
4043 } else if (fed->fed_group != oa->o_seq) {
4044 CERROR("%s: this export (nid %s) used object group %d "
4045 "earlier; now it's trying to use group "LPU64"!"
4046 " This could be a bug in the MDS. Please report to "
4047 "http://bugzilla.lustre.org/\n", obd->obd_name,
4048 obd_export_nid2str(exp), fed->fed_group, oa->o_seq);
4055 rc = obd_alloc_memmd(exp, &lsm);
4062 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4064 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4065 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4066 if (!obd->obd_recovering ||
4067 oa->o_id > filter_last_id(filter, oa->o_seq)) {
4068 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4069 oa->o_id, filter_last_id(filter, oa->o_seq));
4073 cfs_down(&filter->fo_create_locks[oa->o_seq]);
4074 rc = filter_precreate(obd, oa, oa->o_seq, &diff);
4075 cfs_up(&filter->fo_create_locks[oa->o_seq]);
4078 rc = filter_handle_precreate(exp, oa, oa->o_seq, oti);
4081 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4082 if (rc && ea != NULL && *ea != lsm) {
4083 obd_free_memmd(exp, &lsm);
4084 } else if (rc == 0 && ea != NULL) {
4085 /* XXX LOV STACKING: the lsm that is passed to us from
4086 * LOV does not have valid lsm_oinfo data structs, so
4087 * don't go touching that. This needs to be fixed in a
4089 lsm->lsm_object_id = oa->o_id;
4096 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4097 struct lov_stripe_md *md, struct obd_trans_info *oti,
4098 struct obd_export *md_exp, void *capa)
4100 unsigned int qcids[MAXQUOTAS] = {0, 0};
4101 struct obd_device *obd;
4102 struct filter_obd *filter;
4103 struct dentry *dchild = NULL, *dparent = NULL;
4104 struct lustre_handle lockh = { 0 };
4105 struct lvfs_run_ctxt saved;
4106 void *handle = NULL;
4107 struct llog_cookie *fcc = NULL;
4108 int rc, rc2, cleanup_phase = 0, sync = 0;
4113 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4114 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4119 filter = &obd->u.filter;
4121 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4124 CDEBUG(D_INODE, "%s: filter_destroy(group="LPU64",oid="
4125 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4127 dchild = filter_fid2dentry(obd, NULL, oa->o_seq, oa->o_id);
4129 GOTO(cleanup, rc = PTR_ERR(dchild));
4132 if (dchild->d_inode == NULL) {
4133 CDEBUG(D_INODE, "destroying non-existent object "POSTID"\n",
4134 oa->o_id, oa->o_seq);
4135 /* If object already gone, cancel cookie right now */
4136 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4137 struct llog_ctxt *ctxt;
4138 struct obd_llog_group *olg;
4140 olg = filter_find_olg(obd, oa->o_seq);
4142 CERROR(" %s: can not find olg of group %d\n",
4143 obd->obd_name, (int)oa->o_seq);
4144 GOTO(cleanup, rc = PTR_ERR(olg));
4146 fcc = &oa->o_lcookie;
4147 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4148 llog_cancel(ctxt, NULL, 1, fcc, 0);
4149 llog_ctxt_put(ctxt);
4150 fcc = NULL; /* we didn't allocate fcc, don't free it */
4152 GOTO(cleanup, rc = -ENOENT);
4155 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_seq, &lockh);
4159 /* Our MDC connection is established by the MDS to us */
4160 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4161 OBD_ALLOC(fcc, sizeof(*fcc));
4163 *fcc = oa->o_lcookie;
4165 ll_vfs_dq_init(dchild->d_inode);
4167 /* we're gonna truncate it first in order to avoid possible deadlock:
4169 * open trasaction open transaction
4170 * down(i_zombie) down(i_zombie)
4171 * restart transaction
4172 * (see BUG 4180) -bzzz
4174 * take i_alloc_sem too to prevent other threads from writing to the
4175 * file while we are truncating it. This can cause lock ordering issue
4176 * between page lock, i_mutex & starting new journal handle.
4177 * (see bug 20321) -johann
4180 down_write(&dchild->d_inode->i_alloc_sem);
4181 LOCK_INODE_MUTEX(dchild->d_inode);
4182 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
4184 /* VBR: version recovery check */
4185 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4187 UNLOCK_INODE_MUTEX(dchild->d_inode);
4188 up_write(&dchild->d_inode->i_alloc_sem);
4192 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4194 if (IS_ERR(handle)) {
4195 UNLOCK_INODE_MUTEX(dchild->d_inode);
4196 up_write(&dchild->d_inode->i_alloc_sem);
4197 GOTO(cleanup, rc = PTR_ERR(handle));
4200 iattr.ia_valid = ATTR_SIZE;
4202 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4203 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4204 UNLOCK_INODE_MUTEX(dchild->d_inode);
4205 up_write(&dchild->d_inode->i_alloc_sem);
4209 GOTO(cleanup, rc = rc2);
4211 /* We don't actually need to lock the parent until we are unlinking
4212 * here, and not while truncating above. That avoids holding the
4213 * parent lock for a long time during truncate, which can block other
4214 * threads from doing anything to objects in that directory. bug 7171 */
4215 dparent = filter_parent_lock(obd, oa->o_seq, oa->o_id);
4216 if (IS_ERR(dparent))
4217 GOTO(cleanup, rc = PTR_ERR(dparent));
4218 cleanup_phase = 3; /* filter_parent_unlock */
4220 LOCK_INODE_MUTEX(dchild->d_inode);
4221 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4222 if (IS_ERR(handle)) {
4223 UNLOCK_INODE_MUTEX(dchild->d_inode);
4224 GOTO(cleanup, rc = PTR_ERR(handle));
4226 cleanup_phase = 4; /* fsfilt_commit */
4228 /* Quota release need uid/gid of inode */
4229 obdo_from_inode(oa, dchild->d_inode, NULL, OBD_MD_FLUID|OBD_MD_FLGID);
4231 filter_fmd_drop(exp, oa->o_id, oa->o_seq);
4233 /* this drops dchild->d_inode->i_mutex unconditionally */
4234 rc = filter_destroy_internal(obd, oa->o_id, oa->o_seq, dparent, dchild);
4238 switch(cleanup_phase) {
4241 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4242 oti->oti_handle : handle,
4243 filter_cancel_cookies_cb,
4245 /* If add_journal_cb failed, then filter_finish_transno
4246 * will commit the handle and we will do a sync
4247 * on commit. then we call callback directly to free
4250 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4252 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4255 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4257 CERROR("error on commit, err = %d\n", rc2);
4264 filter_parent_unlock(dparent);
4266 filter_fini_destroy(obd, &lockh);
4270 OBD_FREE(fcc, sizeof(*fcc));
4272 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4275 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4279 /* trigger quota release */
4280 qcids[USRQUOTA] = oa->o_uid;
4281 qcids[GRPQUOTA] = oa->o_gid;
4282 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4285 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4289 /* NB start and end are used for punch, but not truncate */
4290 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4291 struct obd_trans_info *oti,
4292 struct ptlrpc_request_set *rqset)
4297 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4298 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4299 oinfo->oi_policy.l_extent.end);
4303 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4304 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,oinfo->oi_oa->o_valid,
4305 oinfo->oi_policy.l_extent.start);
4307 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4308 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4309 rc = filter_setattr(exp, oinfo, oti);
4310 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4314 static int filter_sync(struct obd_export *exp, struct obd_info *oinfo,
4315 obd_off start, obd_off end,
4316 struct ptlrpc_request_set *set)
4318 struct lvfs_run_ctxt saved;
4319 struct obd_device_target *obt;
4320 struct dentry *dentry;
4324 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
4325 (struct lustre_capa *)oinfo->oi_capa,
4326 CAPA_OPC_OSS_WRITE);
4330 obt = &exp->exp_obd->u.obt;
4332 /* An objid of zero is taken to mean "sync whole filesystem" */
4333 if (!oinfo->oi_oa || !(oinfo->oi_oa->o_valid & OBD_MD_FLID)) {
4334 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4335 /* Flush any remaining cancel messages out to the target */
4336 filter_sync_llogs(exp->exp_obd, exp);
4340 dentry = filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi);
4342 RETURN(PTR_ERR(dentry));
4344 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4346 LOCK_INODE_MUTEX(dentry->d_inode);
4348 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4350 /* just any file to grab fsync method - "file" arg unused */
4351 struct file *file = obt->obt_rcvd_filp;
4353 if (file->f_op && file->f_op->fsync)
4354 rc = file->f_op->fsync(NULL, dentry, 1);
4356 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4360 UNLOCK_INODE_MUTEX(dentry->d_inode);
4362 oinfo->oi_oa->o_valid = OBD_MD_FLID;
4363 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL,
4364 FILTER_VALID_FLAGS);
4366 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4372 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4373 void *key, __u32 *vallen, void *val,
4374 struct lov_stripe_md *lsm)
4376 struct obd_device *obd;
4379 obd = class_exp2obd(exp);
4381 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4385 if (KEY_IS(KEY_BLOCKSIZE)) {
4386 __u32 *blocksize = val;
4388 if (*vallen < sizeof(*blocksize))
4390 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4392 *vallen = sizeof(*blocksize);
4396 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4397 __u32 *blocksize_bits = val;
4398 if (blocksize_bits) {
4399 if (*vallen < sizeof(*blocksize_bits))
4401 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4403 *vallen = sizeof(*blocksize_bits);
4407 if (KEY_IS(KEY_LAST_ID)) {
4408 obd_id *last_id = val;
4409 /* FIXME: object groups */
4411 if (*vallen < sizeof(*last_id))
4413 *last_id = filter_last_id(&obd->u.filter,
4414 exp->exp_filter_data.fed_group);
4416 *vallen = sizeof(*last_id);
4420 if (KEY_IS(KEY_FIEMAP)) {
4421 struct ll_fiemap_info_key *fm_key = key;
4422 struct dentry *dentry;
4423 struct ll_user_fiemap *fiemap = val;
4424 struct lvfs_run_ctxt saved;
4427 if (fiemap == NULL) {
4428 *vallen = fiemap_count_to_size(
4429 fm_key->fiemap.fm_extent_count);
4433 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa.o_oi,
4436 RETURN(PTR_ERR(dentry));
4438 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4439 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4440 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4442 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4448 if (KEY_IS(KEY_SYNC_LOCK_CANCEL)) {
4449 *((__u32 *) val) = obd->u.filter.fo_sync_lock_cancel;
4450 *vallen = sizeof(__u32);
4454 CDEBUG(D_IOCTL, "invalid key\n");
4458 static inline int filter_setup_llog_group(struct obd_export *exp,
4459 struct obd_device *obd,
4462 struct obd_llog_group *olg;
4463 struct llog_ctxt *ctxt;
4466 olg = filter_find_create_olg(obd, group);
4468 RETURN(PTR_ERR(olg));
4470 llog_group_set_export(olg, exp);
4472 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4473 LASSERTF(ctxt != NULL, "ctxt is null\n");
4475 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4476 llog_ctxt_put(ctxt);
4480 static int filter_set_grant_shrink(struct obd_export *exp,
4481 struct ost_body *body)
4483 /* handle shrink grant */
4484 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4485 filter_grant_incoming(exp, &body->oa);
4486 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4492 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4494 struct obd_device *obd;
4500 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4504 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4505 obd_export_nid2str(exp));
4506 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4508 /* setup llog imports */
4510 group = (int)(*(__u32 *)val);
4512 group = 0; /* default value */
4514 LASSERT_SEQ_IS_MDT(group);
4515 rc = filter_setup_llog_group(exp, obd, group);
4519 if (group == FID_SEQ_OST_MDT0) {
4520 /* setup llog group 1 for interop */
4521 filter_setup_llog_group(exp, obd, FID_SEQ_LLOG);
4524 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4529 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4530 void *key, __u32 vallen, void *val,
4531 struct ptlrpc_request_set *set)
4533 struct obd_device *obd;
4538 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4542 if (KEY_IS(KEY_CAPA_KEY)) {
4544 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4546 CERROR("filter update capability key failed: %d\n", rc);
4550 if (KEY_IS(KEY_REVIMP_UPD)) {
4551 filter_revimp_update(exp);
4552 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4556 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4557 filter_adapt_sptlrpc_conf(obd, 0);
4561 if (KEY_IS(KEY_MDS_CONN))
4562 RETURN(filter_set_mds_conn(exp, val));
4564 if (KEY_IS(KEY_GRANT_SHRINK))
4565 RETURN(filter_set_grant_shrink(exp, val));
4570 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4571 int len, void *karg, void *uarg)
4573 struct obd_device *obd = exp->exp_obd;
4574 struct obd_ioctl_data *data = karg;
4578 case OBD_IOC_ABORT_RECOVERY: {
4579 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4580 target_stop_recovery_thread(obd);
4584 case OBD_IOC_SYNC: {
4585 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4586 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4590 case OBD_IOC_SET_READONLY: {
4592 struct super_block *sb = obd->u.obt.obt_sb;
4593 struct inode *inode = sb->s_root->d_inode;
4594 BDEVNAME_DECLARE_STORAGE(tmp);
4595 CERROR("*** setting device %s read-only ***\n",
4596 ll_bdevname(sb, tmp));
4598 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4599 if (!IS_ERR(handle))
4600 rc = fsfilt_commit(obd, inode, handle, 1);
4602 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4603 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4605 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4609 case OBD_IOC_CATLOGLIST: {
4610 rc = llog_catalog_list(obd, 1, data);
4614 case OBD_IOC_LLOG_CANCEL:
4615 case OBD_IOC_LLOG_REMOVE:
4616 case OBD_IOC_LLOG_INFO:
4617 case OBD_IOC_LLOG_PRINT: {
4618 /* FIXME to be finished */
4619 RETURN(-EOPNOTSUPP);
4621 struct llog_ctxt *ctxt = NULL;
4623 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4624 rc = llog_ioctl(ctxt, cmd, data);
4625 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4638 static int filter_health_check(struct obd_device *obd)
4640 #ifdef USE_HEALTH_CHECK_WRITE
4641 struct filter_obd *filter = &obd->u.filter;
4646 * health_check to return 0 on healthy
4647 * and 1 on unhealthy.
4649 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4652 #ifdef USE_HEALTH_CHECK_WRITE
4653 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4654 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4659 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4662 return filter_fid2dentry(data, NULL, gr, id);
4665 static int filter_process_config(struct obd_device *obd, obd_count len,
4668 struct lustre_cfg *lcfg = buf;
4669 struct lprocfs_static_vars lvars;
4672 switch (lcfg->lcfg_command) {
4674 lprocfs_filter_init_vars(&lvars);
4676 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4686 static int filter_notify(struct obd_device *obd,
4687 struct obd_device *unused,
4688 enum obd_notify_event ev, void *data)
4691 case OBD_NOTIFY_CONFIG:
4692 LASSERT(obd->obd_no_conn);
4693 cfs_spin_lock(&obd->obd_dev_lock);
4694 obd->obd_no_conn = 0;
4695 cfs_spin_unlock(&obd->obd_dev_lock);
4698 CDEBUG(D_INFO, "%s: Unhandled notification %#x\n",
4704 static struct lvfs_callback_ops filter_lvfs_ops = {
4705 l_fid2dentry: filter_lvfs_fid2dentry,
4708 static struct obd_ops filter_obd_ops = {
4709 .o_owner = THIS_MODULE,
4710 .o_get_info = filter_get_info,
4711 .o_set_info_async = filter_set_info_async,
4712 .o_setup = filter_setup,
4713 .o_precleanup = filter_precleanup,
4714 .o_cleanup = filter_cleanup,
4715 .o_connect = filter_connect,
4716 .o_reconnect = filter_reconnect,
4717 .o_disconnect = filter_disconnect,
4718 .o_ping = filter_ping,
4719 .o_init_export = filter_init_export,
4720 .o_destroy_export = filter_destroy_export,
4721 .o_statfs = filter_statfs,
4722 .o_getattr = filter_getattr,
4723 .o_unpackmd = filter_unpackmd,
4724 .o_create = filter_create,
4725 .o_setattr = filter_setattr,
4726 .o_destroy = filter_destroy,
4727 .o_brw = filter_brw,
4728 .o_punch = filter_truncate,
4729 .o_sync = filter_sync,
4730 .o_preprw = filter_preprw,
4731 .o_commitrw = filter_commitrw,
4732 .o_llog_init = filter_llog_init,
4733 .o_llog_connect = filter_llog_connect,
4734 .o_llog_finish = filter_llog_finish,
4735 .o_iocontrol = filter_iocontrol,
4736 .o_health_check = filter_health_check,
4737 .o_process_config = filter_process_config,
4738 .o_notify = filter_notify,
4741 quota_interface_t *filter_quota_interface_ref;
4742 extern quota_interface_t filter_quota_interface;
4744 static int __init obdfilter_init(void)
4746 struct lprocfs_static_vars lvars;
4749 /** sanity check for group<->mdsno conversion */
4750 for (i = 0; i < MAX_MDT_COUNT; i++)
4751 LASSERT(objseq_to_mdsno(mdt_to_obd_objseq(i)) == i);
4753 lprocfs_filter_init_vars(&lvars);
4755 cfs_request_module("%s", "lquota");
4756 OBD_ALLOC(obdfilter_created_scratchpad,
4757 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4758 sizeof(*obdfilter_created_scratchpad));
4759 if (obdfilter_created_scratchpad == NULL)
4762 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4763 sizeof(struct filter_mod_data),
4766 GOTO(out, rc = -ENOMEM);
4768 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4769 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4771 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4772 LUSTRE_OST_NAME, NULL);
4776 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4777 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4778 ll_fmd_cachep = NULL;
4780 if (filter_quota_interface_ref)
4781 PORTAL_SYMBOL_PUT(filter_quota_interface);
4783 OBD_FREE(obdfilter_created_scratchpad,
4784 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4785 sizeof(*obdfilter_created_scratchpad));
4791 static void __exit obdfilter_exit(void)
4793 if (filter_quota_interface_ref)
4794 PORTAL_SYMBOL_PUT(filter_quota_interface);
4796 if (ll_fmd_cachep) {
4797 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4798 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4799 ll_fmd_cachep = NULL;
4802 class_unregister_type(LUSTRE_OST_NAME);
4803 OBD_FREE(obdfilter_created_scratchpad,
4804 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4805 sizeof(*obdfilter_created_scratchpad));
4808 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4809 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4810 MODULE_LICENSE("GPL");
4812 module_init(obdfilter_init);
4813 module_exit(obdfilter_exit);