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 lsd->lsd_mount_count = cpu_to_le64(obd->u.obt.obt_mount_count);
909 /* save it, so mount count and last_transno is current */
910 rc = filter_update_server_data(obd);
912 GOTO(err_client, rc);
917 class_disconnect_exports(obd);
919 filter_free_server_data(&obd->u.obt);
923 static int filter_cleanup_groups(struct obd_device *obd)
925 struct filter_obd *filter = &obd->u.filter;
927 struct dentry *dentry;
931 if (filter->fo_dentry_O_groups != NULL) {
932 for (i = 0; i < filter->fo_group_count; i++) {
933 dentry = filter->fo_dentry_O_groups[i];
937 OBD_FREE(filter->fo_dentry_O_groups,
938 filter->fo_group_count *
939 sizeof(*filter->fo_dentry_O_groups));
940 filter->fo_dentry_O_groups = NULL;
942 if (filter->fo_last_objid_files != NULL) {
943 for (i = 0; i < filter->fo_group_count; i++) {
944 filp = filter->fo_last_objid_files[i];
948 OBD_FREE(filter->fo_last_objid_files,
949 filter->fo_group_count *
950 sizeof(*filter->fo_last_objid_files));
951 filter->fo_last_objid_files = NULL;
953 if (filter->fo_dentry_O_sub != NULL) {
954 for (i = 0; i < filter->fo_group_count; i++) {
955 for (j = 0; j < filter->fo_subdir_count; j++) {
956 dentry = filter->fo_dentry_O_sub[i].dentry[j];
961 OBD_FREE(filter->fo_dentry_O_sub,
962 filter->fo_group_count *
963 sizeof(*filter->fo_dentry_O_sub));
964 filter->fo_dentry_O_sub = NULL;
966 if (filter->fo_last_objids != NULL) {
967 OBD_FREE(filter->fo_last_objids,
968 filter->fo_group_count *
969 sizeof(*filter->fo_last_objids));
970 filter->fo_last_objids = NULL;
972 if (filter->fo_dentry_O != NULL) {
973 f_dput(filter->fo_dentry_O);
974 filter->fo_dentry_O = NULL;
979 static int filter_update_last_group(struct obd_device *obd, int group)
981 struct filter_obd *filter = &obd->u.filter;
982 struct file *filp = NULL;
983 int last_group = 0, rc;
987 if (group <= filter->fo_committed_group)
990 filp = filp_open("LAST_GROUP", O_RDWR, 0700);
994 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
998 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1000 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
1004 CDEBUG(D_INODE, "%s: previous %d, new %d\n",
1005 obd->obd_name, last_group, group);
1009 /* must be sync: bXXXX */
1010 rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
1012 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
1016 filter->fo_committed_group = group;
1019 filp_close(filp, 0);
1023 static int filter_read_group_internal(struct obd_device *obd, int group,
1026 struct filter_obd *filter = &obd->u.filter;
1027 __u64 *new_objids = NULL;
1028 struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
1029 struct dentry **new_groups = NULL;
1030 struct file **new_files = NULL;
1031 struct dentry *dentry;
1033 int old_count = filter->fo_group_count, rc, stage = 0, i;
1037 int len = group + 1;
1039 snprintf(name, 24, "%d", group);
1043 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
1045 if (IS_ERR(dentry)) {
1046 CERROR("Cannot lookup expected object group %d: %ld\n",
1047 group, PTR_ERR(dentry));
1048 RETURN(PTR_ERR(dentry));
1051 dentry = simple_mkdir(filter->fo_dentry_O,
1052 obd->u.obt.obt_vfsmnt, name, 0700, 1);
1053 if (IS_ERR(dentry)) {
1054 CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
1056 RETURN(PTR_ERR(dentry));
1061 snprintf(name, 24, "O/%d/LAST_ID", group);
1063 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
1065 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
1066 GOTO(cleanup, rc = PTR_ERR(filp));
1070 rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
1072 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
1076 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1077 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
1078 if (tmp_subdirs == NULL)
1079 GOTO(cleanup, rc = -ENOMEM);
1082 for (i = 0; i < filter->fo_subdir_count; i++) {
1084 snprintf(dir, sizeof(dir), "d%u", i);
1086 tmp_subdirs->dentry[i] = simple_mkdir(dentry,
1087 obd->u.obt.obt_vfsmnt,
1089 if (IS_ERR(tmp_subdirs->dentry[i])) {
1090 rc = PTR_ERR(tmp_subdirs->dentry[i]);
1091 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
1096 CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
1097 tmp_subdirs->dentry[i]);
1101 /* 'group' is an index; we need an array of length 'group + 1' */
1102 if (group + 1 > old_count) {
1103 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
1104 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1105 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1106 OBD_ALLOC(new_files, len * sizeof(*new_files));
1108 if (new_objids == NULL || new_subdirs == NULL ||
1109 new_groups == NULL || new_files == NULL)
1110 GOTO(cleanup, rc = -ENOMEM);
1113 memcpy(new_objids, filter->fo_last_objids,
1114 old_count * sizeof(*new_objids));
1115 memcpy(new_subdirs, filter->fo_dentry_O_sub,
1116 old_count * sizeof(*new_subdirs));
1117 memcpy(new_groups, filter->fo_dentry_O_groups,
1118 old_count * sizeof(*new_groups));
1119 memcpy(new_files, filter->fo_last_objid_files,
1120 old_count * sizeof(*new_files));
1122 OBD_FREE(filter->fo_last_objids,
1123 old_count * sizeof(*new_objids));
1124 OBD_FREE(filter->fo_dentry_O_sub,
1125 old_count * sizeof(*new_subdirs));
1126 OBD_FREE(filter->fo_dentry_O_groups,
1127 old_count * sizeof(*new_groups));
1128 OBD_FREE(filter->fo_last_objid_files,
1129 old_count * sizeof(*new_files));
1131 filter->fo_last_objids = new_objids;
1132 filter->fo_dentry_O_sub = new_subdirs;
1133 filter->fo_dentry_O_groups = new_groups;
1134 filter->fo_last_objid_files = new_files;
1135 filter->fo_group_count = len;
1138 filter->fo_dentry_O_groups[group] = dentry;
1139 filter->fo_last_objid_files[group] = filp;
1140 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1141 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1142 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1145 filter_update_last_group(obd, group);
1147 if (i_size_read(filp->f_dentry->d_inode) == 0) {
1148 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1149 rc = filter_update_last_objid(obd, group, 1);
1153 filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1154 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1155 obd->obd_name, group, last_objid);
1160 if (new_objids != NULL)
1161 OBD_FREE(new_objids, len * sizeof(*new_objids));
1162 if (new_subdirs != NULL)
1163 OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1164 if (new_groups != NULL)
1165 OBD_FREE(new_groups, len * sizeof(*new_groups));
1166 if (new_files != NULL)
1167 OBD_FREE(new_files, len * sizeof(*new_files));
1169 if (filter->fo_subdir_count && fid_seq_is_mdt(group)) {
1170 for (i = 0; i < filter->fo_subdir_count; i++) {
1171 if (tmp_subdirs->dentry[i] != NULL)
1172 dput(tmp_subdirs->dentry[i]);
1174 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1177 filp_close(filp, 0);
1184 static int filter_read_groups(struct obd_device *obd, int last_group,
1187 struct filter_obd *filter = &obd->u.filter;
1188 int old_count, group, rc = 0;
1190 cfs_down(&filter->fo_init_lock);
1191 old_count = filter->fo_group_count;
1192 for (group = old_count; group <= last_group; group++) {
1193 rc = filter_read_group_internal(obd, group, create);
1197 cfs_up(&filter->fo_init_lock);
1201 /* FIXME: object groups */
1202 static int filter_prep_groups(struct obd_device *obd)
1204 struct filter_obd *filter = &obd->u.filter;
1205 struct dentry *O_dentry;
1207 int last_group, rc = 0, cleanup_phase = 0;
1211 O_dentry = simple_mkdir(cfs_fs_pwd(current->fs), obd->u.obt.obt_vfsmnt,
1213 CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1214 if (IS_ERR(O_dentry)) {
1215 rc = PTR_ERR(O_dentry);
1216 CERROR("cannot open/create O: rc = %d\n", rc);
1219 filter->fo_dentry_O = O_dentry;
1220 cleanup_phase = 1; /* O_dentry */
1222 /* we have to initialize all groups before first connections from
1223 * clients because they may send create/destroy for any group -bzzz */
1224 filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1226 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1227 GOTO(cleanup, rc = PTR_ERR(filp));
1229 cleanup_phase = 2; /* filp */
1231 rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1233 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1238 last_group = FID_SEQ_OST_MDT0;
1240 CWARN("%s: initialize groups [%u,%u]\n", obd->obd_name,
1241 FID_SEQ_OST_MDT0, last_group);
1242 filter->fo_committed_group = last_group;
1243 rc = filter_read_groups(obd, last_group, 1);
1247 filp_close(filp, 0);
1251 switch (cleanup_phase) {
1253 filp_close(filp, 0);
1255 filter_cleanup_groups(obd);
1256 f_dput(filter->fo_dentry_O);
1257 filter->fo_dentry_O = NULL;
1265 /* setup the object store with correct subdirectories */
1266 static int filter_prep(struct obd_device *obd)
1268 struct lvfs_run_ctxt saved;
1269 struct filter_obd *filter = &obd->u.filter;
1271 struct inode *inode;
1275 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1276 file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1277 if (!file || IS_ERR(file)) {
1279 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1283 obd->u.obt.obt_rcvd_filp = file;
1284 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1285 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1286 file->f_dentry->d_inode->i_mode);
1287 GOTO(err_filp, rc = -ENOENT);
1290 inode = file->f_dentry->d_parent->d_inode;
1291 /* We use i_op->unlink directly in filter_vfs_unlink() */
1292 if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1293 CERROR("%s: filesystem does not support create/unlink ops\n",
1295 GOTO(err_filp, rc = -EOPNOTSUPP);
1298 rc = filter_init_server_data(obd, file);
1300 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1303 LASSERT(obd->u.obt.obt_lut);
1304 target_recovery_init(obd->u.obt.obt_lut, ost_handle);
1306 /* open and create health check io file*/
1307 file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1310 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1312 GOTO(err_server_data, rc);
1314 filter->fo_obt.obt_health_check_filp = file;
1315 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1316 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1317 file->f_dentry->d_inode->i_mode);
1318 GOTO(err_health_check, rc = -ENOENT);
1320 rc = lvfs_check_io_health(obd, file);
1322 GOTO(err_health_check, rc);
1324 rc = filter_prep_groups(obd);
1326 GOTO(err_health_check, rc);
1328 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1333 if (filp_close(filter->fo_obt.obt_health_check_filp, 0))
1334 CERROR("can't close %s after error\n", HEALTH_CHECK);
1335 filter->fo_obt.obt_health_check_filp = NULL;
1337 target_recovery_fini(obd);
1338 filter_free_server_data(&obd->u.obt);
1340 if (filp_close(obd->u.obt.obt_rcvd_filp, 0))
1341 CERROR("can't close %s after error\n", LAST_RCVD);
1342 obd->u.obt.obt_rcvd_filp = NULL;
1346 /* cleanup the filter: write last used object id to status file */
1347 static void filter_post(struct obd_device *obd)
1349 struct lvfs_run_ctxt saved;
1350 struct filter_obd *filter = &obd->u.filter;
1353 /* XXX: filter_update_lastobjid used to call fsync_dev. It might be
1354 * best to start a transaction with h_sync, because we removed this
1357 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1358 rc = filter_update_server_data(obd);
1360 CERROR("error writing server data: rc = %d\n", rc);
1362 for (i = 0; i < filter->fo_group_count; i++) {
1363 rc = filter_update_last_objid(obd, i,
1364 (i == filter->fo_group_count - 1));
1366 CERROR("error writing group %d lastobjid: rc = %d\n",
1370 rc = filp_close(obd->u.obt.obt_rcvd_filp, 0);
1371 obd->u.obt.obt_rcvd_filp = NULL;
1373 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1375 rc = filp_close(filter->fo_obt.obt_health_check_filp, 0);
1376 filter->fo_obt.obt_health_check_filp = NULL;
1378 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1380 filter_cleanup_groups(obd);
1381 filter_free_server_data(&obd->u.obt);
1382 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1384 filter_free_capa_keys(filter);
1385 cleanup_capa_hash(filter->fo_capa_hash);
1388 static void filter_set_last_id(struct filter_obd *filter,
1389 obd_id id, obd_seq group)
1391 LASSERT(group <= filter->fo_group_count);
1393 cfs_spin_lock(&filter->fo_objidlock);
1394 filter->fo_last_objids[group] = id;
1395 cfs_spin_unlock(&filter->fo_objidlock);
1398 obd_id filter_last_id(struct filter_obd *filter, obd_seq group)
1401 LASSERT(group <= filter->fo_group_count);
1402 LASSERT(filter->fo_last_objids != NULL);
1404 /* FIXME: object groups */
1405 cfs_spin_lock(&filter->fo_objidlock);
1406 id = filter->fo_last_objids[group];
1407 cfs_spin_unlock(&filter->fo_objidlock);
1411 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1413 LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
1417 /* We never dget the object parent, so DON'T dput it either */
1418 struct dentry *filter_parent(struct obd_device *obd, obd_seq group, obd_id objid)
1420 struct filter_obd *filter = &obd->u.filter;
1421 struct filter_subdirs *subdirs;
1423 if (group >= filter->fo_group_count) /* FIXME: object groups */
1424 return ERR_PTR(-EBADF);
1426 if (!fid_seq_is_mdt(group) || filter->fo_subdir_count == 0)
1427 return filter->fo_dentry_O_groups[group];
1429 subdirs = &filter->fo_dentry_O_sub[group];
1430 return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1433 /* We never dget the object parent, so DON'T dput it either */
1434 struct dentry *filter_parent_lock(struct obd_device *obd, obd_seq group,
1437 unsigned long now = jiffies;
1438 struct dentry *dparent = filter_parent(obd, group, objid);
1441 if (IS_ERR(dparent))
1443 if (dparent == NULL)
1444 return ERR_PTR(-ENOENT);
1446 rc = filter_lock_dentry(obd, dparent);
1447 fsfilt_check_slow(obd, now, "parent lock");
1448 return rc ? ERR_PTR(rc) : dparent;
1451 /* We never dget the object parent, so DON'T dput it either */
1452 static void filter_parent_unlock(struct dentry *dparent)
1454 UNLOCK_INODE_MUTEX(dparent->d_inode);
1457 /* How to get files, dentries, inodes from object id's.
1459 * If dir_dentry is passed, the caller has already locked the parent
1460 * appropriately for this operation (normally a write lock). If
1461 * dir_dentry is NULL, we do a read lock while we do the lookup to
1462 * avoid races with create/destroy and such changing the directory
1463 * internal to the filesystem code. */
1464 struct dentry *filter_fid2dentry(struct obd_device *obd,
1465 struct dentry *dir_dentry,
1466 obd_seq group, obd_id id)
1468 struct dentry *dparent = dir_dentry;
1469 struct dentry *dchild;
1474 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1475 obd->u.filter.fo_destroys_in_progress == 0) {
1476 /* don't fail lookups for orphan recovery, it causes
1477 * later LBUGs when objects still exist during precreate */
1478 CDEBUG(D_INFO, "*** cfs_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1479 RETURN(ERR_PTR(-ENOENT));
1482 CERROR("fatal: invalid object id 0\n");
1483 RETURN(ERR_PTR(-ESTALE));
1486 len = sprintf(name, LPU64, id);
1487 if (dir_dentry == NULL) {
1488 dparent = filter_parent_lock(obd, group, id);
1489 if (IS_ERR(dparent)) {
1490 CERROR("%s: error getting object "POSTID
1491 " parent: rc %ld\n", obd->obd_name,
1492 id, group, PTR_ERR(dparent));
1496 CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1497 dparent->d_name.len, dparent->d_name.name, name);
1498 /* dparent is already locked here, so we cannot use ll_lookup_one_len() */
1499 dchild = lookup_one_len(name, dparent, len);
1500 if (dir_dentry == NULL)
1501 filter_parent_unlock(dparent);
1502 if (IS_ERR(dchild)) {
1503 CERROR("%s: object "LPU64":"LPU64" lookup error: rc %ld\n",
1504 obd->obd_name, id, group, PTR_ERR(dchild));
1508 if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1509 CERROR("%s: got bad object "LPU64" inode %lu\n",
1510 obd->obd_name, id, dchild->d_inode->i_ino);
1512 RETURN(ERR_PTR(-ENOENT));
1515 CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1516 name, dchild, atomic_read(&dchild->d_count));
1518 LASSERT(atomic_read(&dchild->d_count) > 0);
1523 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1524 obd_id group, struct lustre_handle *lockh)
1526 int flags = LDLM_AST_DISCARD_DATA, rc;
1527 struct ldlm_res_id res_id;
1528 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1531 osc_build_res_name(objid, group, &res_id);
1532 /* Tell the clients that the object is gone now and that they should
1533 * throw away any cached pages. */
1534 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1535 &policy, LCK_PW, &flags, ldlm_blocking_ast,
1536 ldlm_completion_ast, NULL, NULL, 0, NULL,
1543 static void filter_fini_destroy(struct obd_device *obd,
1544 struct lustre_handle *lockh)
1546 if (lustre_handle_is_used(lockh))
1547 ldlm_lock_decref(lockh, LCK_PW);
1550 /* This is vfs_unlink() without down(i_sem). If we call regular vfs_unlink()
1551 * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1552 * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1553 * reverse. Caller must take i_sem before starting the transaction and we
1554 * drop it here before the inode is removed from the dentry. bug 4180/6984 */
1555 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry,
1556 struct vfsmount *mnt)
1561 /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1562 * itself we already hold dir->i_mutex for child create/unlink ops */
1563 LASSERT(dentry->d_inode != NULL);
1564 LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1565 LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1569 if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1570 GOTO(out, rc = -ENOENT);
1572 rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1577 GOTO(out, rc = -EPERM);
1579 /* check_sticky() */
1580 if ((dentry->d_inode->i_uid != cfs_curproc_fsuid() &&
1581 !cfs_capable(CFS_CAP_FOWNER)) || IS_APPEND(dentry->d_inode) ||
1582 IS_IMMUTABLE(dentry->d_inode))
1583 GOTO(out, rc = -EPERM);
1585 /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1586 * that was done because of journal_start (which is already done
1587 * here) or some other ordering issue. */
1588 ll_vfs_dq_init(dir);
1590 rc = ll_security_inode_unlink(dir, dentry, mnt);
1594 rc = dir->i_op->unlink(dir, dentry);
1596 /* need to drop i_mutex before we lose inode reference */
1597 UNLOCK_INODE_MUTEX(dentry->d_inode);
1604 /* Caller must hold LCK_PW on parent and push us into kernel context.
1605 * Caller must hold child i_mutex, we drop it always.
1606 * Caller is also required to ensure that dchild->d_inode exists. */
1607 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1608 obd_seq group, struct dentry *dparent,
1609 struct dentry *dchild)
1611 struct inode *inode = dchild->d_inode;
1614 /* There should be 2 references to the inode:
1615 * 1) taken by filter_prepare_destroy
1616 * 2) taken by filter_destroy */
1617 if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 2) {
1618 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1619 dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1620 (unsigned long)inode->i_nlink,
1621 atomic_read(&inode->i_count));
1624 rc = filter_vfs_unlink(dparent->d_inode, dchild, obd->u.obt.obt_vfsmnt);
1626 CERROR("error unlinking objid %.*s: rc %d\n",
1627 dchild->d_name.len, dchild->d_name.name, rc);
1631 struct filter_intent_args {
1632 struct ldlm_lock **victim;
1637 static enum interval_iter filter_intent_cb(struct interval_node *n,
1640 struct ldlm_interval *node = (struct ldlm_interval *)n;
1641 struct filter_intent_args *arg = (struct filter_intent_args*)args;
1642 __u64 size = arg->size;
1643 struct ldlm_lock **v = arg->victim;
1644 struct ldlm_lock *lck;
1646 /* If the interval is lower than the current file size,
1648 if (interval_high(n) <= size)
1649 return INTERVAL_ITER_STOP;
1651 cfs_list_for_each_entry(lck, &node->li_group, l_sl_policy) {
1652 /* Don't send glimpse ASTs to liblustre clients.
1653 * They aren't listening for them, and they do
1654 * entirely synchronous I/O anyways. */
1655 if (lck->l_export == NULL ||
1656 lck->l_export->exp_libclient == 1)
1659 if (*arg->liblustre)
1660 *arg->liblustre = 0;
1663 *v = LDLM_LOCK_GET(lck);
1664 } else if ((*v)->l_policy_data.l_extent.start <
1665 lck->l_policy_data.l_extent.start) {
1666 LDLM_LOCK_RELEASE(*v);
1667 *v = LDLM_LOCK_GET(lck);
1670 /* the same policy group - every lock has the
1671 * same extent, so needn't do it any more */
1675 return INTERVAL_ITER_CONT;
1678 static int filter_intent_policy(struct ldlm_namespace *ns,
1679 struct ldlm_lock **lockp, void *req_cookie,
1680 ldlm_mode_t mode, int flags, void *data)
1682 CFS_LIST_HEAD(rpc_list);
1683 struct ptlrpc_request *req = req_cookie;
1684 struct ldlm_lock *lock = *lockp, *l = NULL;
1685 struct ldlm_resource *res = lock->l_resource;
1686 ldlm_processing_policy policy;
1687 struct ost_lvb *res_lvb, *reply_lvb;
1688 struct ldlm_reply *rep;
1690 int idx, rc, tmpflags = 0, only_liblustre = 1;
1691 struct ldlm_interval_tree *tree;
1692 struct filter_intent_args arg;
1693 __u32 repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1694 [DLM_LOCKREPLY_OFF] = sizeof(*rep),
1695 [DLM_REPLY_REC_OFF] = sizeof(*reply_lvb) };
1698 policy = ldlm_get_processing_policy(res);
1699 LASSERT(policy != NULL);
1700 LASSERT(req != NULL);
1702 rc = lustre_pack_reply(req, 3, repsize, NULL);
1704 RETURN(req->rq_status = rc);
1706 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1707 LASSERT(rep != NULL);
1709 reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1710 sizeof(*reply_lvb));
1711 LASSERT(reply_lvb != NULL);
1713 //fixup_handle_for_resent_req(req, lock, &lockh);
1715 /* Call the extent policy function to see if our request can be
1716 * granted, or is blocked.
1717 * If the OST lock has LDLM_FL_HAS_INTENT set, it means a glimpse
1718 * lock, and should not be granted if the lock will be blocked.
1721 LASSERT(ns == ldlm_res_to_ns(res));
1723 rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1724 check_res_locked(res);
1726 /* FIXME: we should change the policy function slightly, to not make
1727 * this list at all, since we just turn around and free it */
1728 while (!cfs_list_empty(&rpc_list)) {
1729 struct ldlm_lock *wlock =
1730 cfs_list_entry(rpc_list.next, struct ldlm_lock,
1732 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1733 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1734 lock->l_flags &= ~LDLM_FL_CP_REQD;
1735 cfs_list_del_init(&wlock->l_cp_ast);
1736 LDLM_LOCK_RELEASE(wlock);
1739 /* The lock met with no resistance; we're finished. */
1740 if (rc == LDLM_ITER_CONTINUE) {
1741 /* do not grant locks to the liblustre clients: they cannot
1742 * handle ASTs robustly. We need to do this while still
1743 * holding lr_lock to avoid the lock remaining on the res_link
1744 * list (and potentially being added to l_pending_list by an
1745 * AST) when we are going to drop this lock ASAP. */
1746 if (lock->l_export->exp_libclient ||
1747 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1748 ldlm_resource_unlink_lock(lock);
1749 err = ELDLM_LOCK_ABORTED;
1751 err = ELDLM_LOCK_REPLACED;
1757 /* Do not grant any lock, but instead send GL callbacks. The extent
1758 * policy nicely created a list of all PW locks for us. We will choose
1759 * the highest of those which are larger than the size in the LVB, if
1760 * any, and perform a glimpse callback. */
1761 res_lvb = res->lr_lvb_data;
1762 LASSERT(res_lvb != NULL);
1763 *reply_lvb = *res_lvb;
1766 * lr_lock guarantees that no new locks are granted, and,
1767 * therefore, that res->lr_lvb_data cannot increase beyond the
1768 * end of already granted lock. As a result, it is safe to
1769 * check against "stale" reply_lvb->lvb_size value without
1772 arg.size = reply_lvb->lvb_size;
1774 arg.liblustre = &only_liblustre;
1775 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1776 tree = &res->lr_itree[idx];
1777 if (tree->lit_mode == LCK_PR)
1780 interval_iterate_reverse(tree->lit_root,
1781 filter_intent_cb, &arg);
1785 /* There were no PW locks beyond the size in the LVB; finished. */
1787 if (only_liblustre) {
1788 /* If we discovered a liblustre client with a PW lock,
1789 * however, the LVB may be out of date! The LVB is
1790 * updated only on glimpse (which we don't do for
1791 * liblustre clients) and cancel (which the client
1792 * obviously has not yet done). So if it has written
1793 * data but kept the lock, the LVB is stale and needs
1794 * to be updated from disk.
1796 * Of course, this will all disappear when we switch to
1797 * taking liblustre locks on the OST. */
1798 ldlm_res_lvbo_update(res, NULL, 1);
1800 RETURN(ELDLM_LOCK_ABORTED);
1804 * This check is for lock taken in filter_prepare_destroy() that does
1805 * not have l_glimpse_ast set. So the logic is: if there is a lock
1806 * with no l_glimpse_ast set, this object is being destroyed already.
1808 * Hence, if you are grabbing DLM locks on the server, always set
1809 * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1811 if (l->l_glimpse_ast == NULL) {
1812 /* We are racing with unlink(); just return -ENOENT */
1813 rep->lock_policy_res1 = -ENOENT;
1817 LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1818 rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1821 *reply_lvb = *res_lvb;
1825 LDLM_LOCK_RELEASE(l);
1827 RETURN(ELDLM_LOCK_ABORTED);
1831 * per-obd_device iobuf pool.
1833 * To avoid memory deadlocks in low-memory setups, amount of dynamic
1834 * allocations in write-path has to be minimized (see bug 5137).
1836 * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1837 * OST threads (see ost_thread_{init,done}()).
1839 * "iobuf's" used by filter cannot be attached to OST thread, however, because
1840 * at the OST layer there are only (potentially) multiple obd_device of type
1841 * unknown at the time of OST thread creation.
1843 * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1844 * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1845 * it's very own iobuf.
1849 * filter_kiobuf_pool_init()
1851 * filter_kiobuf_pool_done()
1853 * filter_iobuf_get()
1855 * operate on this array. They are "generic" in a sense that they don't depend
1856 * on actual type of iobuf's (the latter depending on Linux kernel version).
1860 * destroy pool created by filter_iobuf_pool_init
1862 static void filter_iobuf_pool_done(struct filter_obd *filter)
1864 struct filter_iobuf **pool;
1869 pool = filter->fo_iobuf_pool;
1871 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1872 if (pool[i] != NULL)
1873 filter_free_iobuf(pool[i]);
1875 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1876 filter->fo_iobuf_pool = NULL;
1881 static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
1883 struct filter_obd *filter = &obd->u.filter;
1884 struct sptlrpc_rule_set tmp_rset;
1887 sptlrpc_rule_set_init(&tmp_rset);
1888 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
1890 CERROR("obd %s: failed get sptlrpc rules: %d\n",
1895 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
1897 cfs_write_lock(&filter->fo_sptlrpc_lock);
1898 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
1899 filter->fo_sptlrpc_rset = tmp_rset;
1900 cfs_write_unlock(&filter->fo_sptlrpc_lock);
1906 * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1908 static int filter_iobuf_pool_init(struct filter_obd *filter)
1915 OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1917 if (filter->fo_iobuf_pool == NULL)
1920 filter->fo_iobuf_count = OSS_THREADS_MAX;
1925 /* Return iobuf allocated for @thread_id. We don't know in advance how
1926 * many threads there will be so we allocate a large empty array and only
1927 * fill in those slots that are actually in use.
1928 * If we haven't allocated a pool entry for this thread before, do so now. */
1929 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1931 int thread_id = (oti && oti->oti_thread) ?
1932 oti->oti_thread->t_id : -1;
1933 struct filter_iobuf *pool = NULL;
1934 struct filter_iobuf **pool_place = NULL;
1936 if (thread_id >= 0) {
1937 LASSERT(thread_id < filter->fo_iobuf_count);
1938 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1941 if (unlikely(pool == NULL)) {
1942 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1943 PTLRPC_MAX_BRW_PAGES);
1944 if (pool_place != NULL)
1951 /* mount the file system (secretly). lustre_cfg parameters are:
1954 * 3 = flags: failover=f, failout=n
1957 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1960 struct filter_obd *filter = &obd->u.filter;
1961 struct vfsmount *mnt;
1962 struct lustre_mount_info *lmi;
1963 struct obd_uuid uuid;
1967 struct request_queue *q;
1971 if (lcfg->lcfg_bufcount < 3 ||
1972 LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1973 LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1976 lmi = server_get_mount(obd->obd_name);
1978 /* We already mounted in lustre_fill_super.
1979 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1980 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1982 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1984 /* gets recovery timeouts from mount data */
1985 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_soft)
1986 obd->obd_recovery_timeout =
1987 lsi->lsi_lmd->lmd_recovery_time_soft;
1988 if (lsi->lsi_lmd && lsi->lsi_lmd->lmd_recovery_time_hard)
1989 obd->obd_recovery_time_hard =
1990 lsi->lsi_lmd->lmd_recovery_time_hard;
1992 /* old path - used by lctl */
1993 CERROR("Using old MDS mount method\n");
1994 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1995 MS_NOATIME|MS_NODIRATIME,
1996 lustre_cfg_string(lcfg, 1), option);
1999 LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
2000 lustre_cfg_string(lcfg, 1), rc);
2004 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
2006 if (IS_ERR(obd->obd_fsops))
2007 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
2009 rc = filter_iobuf_pool_init(filter);
2013 if (lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb))) {
2014 CERROR("%s: Underlying device is marked as read-only. "
2015 "Setup failed\n", obd->obd_name);
2016 GOTO(err_ops, rc = -EROFS);
2019 /* failover is the default */
2020 obd->obd_replayable = 1;
2022 if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2023 str = lustre_cfg_string(lcfg, 3);
2024 if (strchr(str, 'n')) {
2025 CWARN("%s: recovery disabled\n", obd->obd_name);
2026 obd->obd_replayable = 0;
2030 obd->u.obt.obt_vfsmnt = mnt;
2031 obd->u.obt.obt_sb = mnt->mnt_sb;
2032 obd->u.obt.obt_magic = OBT_MAGIC;
2033 filter->fo_fstype = mnt->mnt_sb->s_type->name;
2034 CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
2036 fsfilt_setup(obd, obd->u.obt.obt_sb);
2038 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
2039 obd->obd_lvfs_ctxt.pwdmnt = mnt;
2040 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
2041 obd->obd_lvfs_ctxt.fs = get_ds();
2042 obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
2044 cfs_init_mutex(&filter->fo_init_lock);
2045 filter->fo_committed_group = 0;
2046 filter->fo_destroys_in_progress = 0;
2047 for (i = 0; i < 32; i++)
2048 cfs_sema_init(&filter->fo_create_locks[i], 1);
2050 cfs_spin_lock_init(&filter->fo_objidlock);
2051 CFS_INIT_LIST_HEAD(&filter->fo_export_list);
2052 cfs_sema_init(&filter->fo_alloc_lock, 1);
2053 init_brw_stats(&filter->fo_filter_stats);
2054 cfs_spin_lock_init(&filter->fo_flags_lock);
2055 filter->fo_read_cache = 1; /* enable read-only cache by default */
2056 filter->fo_writethrough_cache = 1; /* enable writethrough cache */
2057 filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
2058 filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
2059 filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
2060 filter->fo_syncjournal = 0; /* Don't sync journals on i/o by default */
2061 filter_slc_set(filter); /* initialize sync on lock cancel */
2063 rc = filter_prep(obd);
2067 CFS_INIT_LIST_HEAD(&filter->fo_llog_list);
2068 cfs_spin_lock_init(&filter->fo_llog_list_lock);
2070 filter->fo_fl_oss_capa = 1;
2072 CFS_INIT_LIST_HEAD(&filter->fo_capa_keys);
2073 filter->fo_capa_hash = init_capa_hash();
2074 if (filter->fo_capa_hash == NULL)
2075 GOTO(err_post, rc = -ENOMEM);
2077 sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
2078 obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
2079 LDLM_NAMESPACE_SERVER,
2080 LDLM_NAMESPACE_GREEDY,
2082 if (obd->obd_namespace == NULL)
2083 GOTO(err_post, rc = -ENOMEM);
2084 obd->obd_namespace->ns_lvbp = obd;
2085 obd->obd_namespace->ns_lvbo = &filter_lvbo;
2086 ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
2088 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2089 "filter_ldlm_cb_client", &obd->obd_ldlm_client);
2091 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
2093 CERROR("failed to setup llogging subsystems\n");
2097 cfs_rwlock_init(&filter->fo_sptlrpc_lock);
2098 sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
2099 /* do this after llog being initialized */
2100 filter_adapt_sptlrpc_conf(obd, 1);
2102 rc = lquota_setup(filter_quota_interface_ref, obd);
2106 q = bdev_get_queue(mnt->mnt_sb->s_bdev);
2107 if (queue_max_sectors(q) < queue_max_hw_sectors(q) &&
2108 queue_max_sectors(q) < PTLRPC_MAX_BRW_SIZE >> 9)
2109 LCONSOLE_INFO("%s: underlying device %s should be tuned "
2110 "for larger I/O requests: max_sectors = %u "
2111 "could be up to max_hw_sectors=%u\n",
2112 obd->obd_name, mnt->mnt_sb->s_id,
2113 queue_max_sectors(q), queue_max_hw_sectors(q));
2115 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2116 if (uuid_ptr != NULL) {
2117 class_uuid_unparse(uuid_ptr, &uuid);
2123 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2124 LCONSOLE_INFO("%s: Now serving %s %s%s with recovery %s\n",
2125 obd->obd_name, label ?: str, lmi ? "on " : "",
2126 lmi ? s2lsi(lmi->lmi_sb)->lsi_lmd->lmd_dev : "",
2127 obd->obd_replayable ? "enabled" : "disabled");
2129 if (obd->obd_recovering)
2130 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
2131 "or until %d client%s reconnect%s\n",
2133 obd->obd_recovery_timeout / 60,
2134 obd->obd_recovery_timeout % 60,
2135 obd->obd_max_recoverable_clients,
2136 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
2137 (obd->obd_max_recoverable_clients == 1) ? "s": "");
2145 fsfilt_put_ops(obd->obd_fsops);
2146 filter_iobuf_pool_done(filter);
2148 server_put_mount(obd->obd_name, mnt);
2149 obd->u.obt.obt_sb = 0;
2153 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2155 struct lprocfs_static_vars lvars;
2156 cfs_proc_dir_entry_t *entry;
2162 CLASSERT(offsetof(struct obd_device, u.obt) ==
2163 offsetof(struct obd_device, u.filter.fo_obt));
2165 if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2168 /* lprocfs must be setup before the filter so state can be safely added
2169 * to /proc incrementally as the filter is setup */
2170 lprocfs_filter_init_vars(&lvars);
2171 rc = lprocfs_obd_setup(obd, lvars.obd_vars);
2173 CERROR("%s: lprocfs_obd_setup failed: %d.\n",
2178 rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
2180 CERROR("%s: lprocfs_alloc_obd_stats failed: %d.\n",
2182 GOTO(obd_cleanup, rc);
2185 /* Init obdfilter private stats here */
2186 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2187 LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
2188 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2189 LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
2190 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_GET_PAGE,
2191 LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
2192 "get_page", "usec");
2193 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_NO_PAGE,
2194 LPROCFS_CNTR_AVGMINMAX, "get_page_failures", "num");
2195 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_ACCESS,
2196 LPROCFS_CNTR_AVGMINMAX, "cache_access", "pages");
2197 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_HIT,
2198 LPROCFS_CNTR_AVGMINMAX, "cache_hit", "pages");
2199 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_CACHE_MISS,
2200 LPROCFS_CNTR_AVGMINMAX, "cache_miss", "pages");
2202 rc = lproc_filter_attach_seqstat(obd);
2204 CERROR("%s: create seqstat failed: %d.\n", obd->obd_name, rc);
2205 GOTO(free_obd_stats, rc);
2208 entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL);
2209 if (IS_ERR(entry)) {
2210 rc = PTR_ERR(entry);
2211 CERROR("%s: error %d setting up lprocfs for %s\n",
2212 obd->obd_name, rc, "exports");
2213 GOTO(free_obd_stats, rc);
2215 obd->obd_proc_exports_entry = entry;
2217 entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear",
2218 lprocfs_nid_stats_clear_read,
2219 lprocfs_nid_stats_clear_write, obd, NULL);
2220 if (IS_ERR(entry)) {
2221 rc = PTR_ERR(entry);
2222 CERROR("%s: add proc entry 'clear' failed: %d.\n",
2224 GOTO(free_obd_stats, rc);
2227 /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2228 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2230 GOTO(remove_entry_clear, rc = -ENOMEM);
2231 addr = (unsigned long)cfs_page_address(page);
2232 clear_page((void *)addr);
2233 memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2234 LUSTRE_CFG_BUFLEN(lcfg, 4));
2235 rc = filter_common_setup(obd, lcfg, (void *)addr);
2236 OBD_PAGE_FREE(page);
2238 CERROR("%s: filter_common_setup failed: %d.\n",
2240 GOTO(remove_entry_clear, rc);
2246 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2248 lprocfs_free_obd_stats(obd);
2250 lprocfs_obd_cleanup(obd);
2254 static struct llog_operations filter_mds_ost_repl_logops;
2256 static struct llog_operations filter_size_orig_logops = {
2257 .lop_setup = llog_obd_origin_setup,
2258 .lop_cleanup = llog_obd_origin_cleanup,
2259 .lop_add = llog_obd_origin_add
2262 static int filter_olg_fini(struct obd_llog_group *olg)
2264 struct llog_ctxt *ctxt;
2265 int rc = 0, rc2 = 0;
2268 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2270 rc = llog_cleanup(ctxt);
2272 ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2274 rc2 = llog_cleanup(ctxt);
2279 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_ORIG_CTXT);
2281 rc2 = llog_cleanup(ctxt);
2290 filter_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2291 struct obd_device *tgt)
2296 rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2297 &filter_mds_ost_repl_logops);
2301 rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2302 &filter_size_orig_logops);
2308 filter_olg_fini(olg);
2313 * Init the default olg, which is embeded in the obd_device, for filter.
2316 filter_default_olg_init(struct obd_device *obd, struct obd_llog_group *olg,
2317 struct obd_device *tgt)
2319 struct filter_obd *filter = &obd->u.filter;
2320 struct llog_ctxt *ctxt;
2324 filter->fo_lcm = llog_recov_thread_init(obd->obd_name);
2325 if (!filter->fo_lcm)
2328 filter_mds_ost_repl_logops = llog_client_ops;
2329 filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2330 filter_mds_ost_repl_logops.lop_connect = llog_obd_repl_connect;
2331 filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2333 rc = filter_olg_init(obd, olg, tgt);
2335 GOTO(cleanup_lcm, rc);
2337 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
2340 GOTO(cleanup_olg, rc);
2342 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2344 CERROR("Can't get ctxt for %p:%x\n", olg,
2345 LLOG_MDS_OST_REPL_CTXT);
2346 GOTO(cleanup_olg, rc = -ENODEV);
2348 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2349 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2350 llog_ctxt_put(ctxt);
2354 filter_olg_fini(olg);
2356 llog_recov_thread_fini(filter->fo_lcm, 1);
2357 filter->fo_lcm = NULL;
2362 filter_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2363 struct obd_device *tgt, int *index)
2365 struct filter_obd *filter = &obd->u.filter;
2366 struct llog_ctxt *ctxt;
2370 LASSERT(olg != NULL);
2371 if (olg == &obd->obd_olg)
2372 return filter_default_olg_init(obd, olg, tgt);
2374 LASSERT(filter->fo_lcm != NULL);
2375 rc = filter_olg_init(obd, olg, tgt);
2378 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2380 CERROR("Can't get ctxt for %p:%x\n", olg,
2381 LLOG_MDS_OST_REPL_CTXT);
2382 filter_olg_fini(olg);
2385 ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2386 ctxt->loc_lcm = lcm_get(filter->fo_lcm);
2387 llog_ctxt_put(ctxt);
2391 static int filter_llog_finish(struct obd_device *obd, int count)
2393 struct filter_obd *filter = &obd->u.filter;
2394 struct llog_ctxt *ctxt;
2397 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_MDS_OST_REPL_CTXT);
2400 * Make sure that no cached llcds left in recov_thread.
2401 * We actually do sync in disconnect time, but disconnect
2402 * may not come being marked rq_no_resend = 1.
2404 llog_sync(ctxt, NULL);
2407 * Balance class_import_get() in llog_receptor_accept().
2408 * This is safe to do, as llog is already synchronized
2409 * and its import may go.
2411 cfs_mutex_down(&ctxt->loc_sem);
2412 if (ctxt->loc_imp) {
2413 class_import_put(ctxt->loc_imp);
2414 ctxt->loc_imp = NULL;
2416 cfs_mutex_up(&ctxt->loc_sem);
2417 llog_ctxt_put(ctxt);
2420 if (filter->fo_lcm) {
2421 cfs_mutex_down(&ctxt->loc_sem);
2422 llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
2423 filter->fo_lcm = NULL;
2424 cfs_mutex_up(&ctxt->loc_sem);
2426 RETURN(filter_olg_fini(&obd->obd_olg));
2430 * Find the group llog according to group index in the llog group list.
2432 static struct obd_llog_group *
2433 filter_find_olg_internal(struct filter_obd *filter, int group)
2435 struct obd_llog_group *olg;
2437 LASSERT_SPIN_LOCKED(&filter->fo_llog_list_lock);
2438 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2439 if (olg->olg_seq == group)
2446 * Find the group llog according to group index on the filter
2448 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2450 struct obd_llog_group *olg = NULL;
2451 struct filter_obd *filter;
2453 filter = &obd->u.filter;
2455 if (group == FID_SEQ_LLOG)
2456 RETURN(&obd->obd_olg);
2458 cfs_spin_lock(&filter->fo_llog_list_lock);
2459 olg = filter_find_olg_internal(filter, group);
2460 cfs_spin_unlock(&filter->fo_llog_list_lock);
2465 * Find the llog_group of the filter according to the group. If it can not
2466 * find, create the llog_group, which only happens when mds is being synced
2469 struct obd_llog_group *filter_find_create_olg(struct obd_device *obd, int group)
2471 struct obd_llog_group *olg = NULL, *olg_new = NULL;
2472 struct filter_obd *filter;
2475 filter = &obd->u.filter;
2477 if (group == FID_SEQ_LLOG)
2478 RETURN(&obd->obd_olg);
2480 OBD_ALLOC_PTR(olg_new);
2481 if (olg_new == NULL)
2482 RETURN(ERR_PTR(-ENOMEM));
2484 cfs_spin_lock(&filter->fo_llog_list_lock);
2485 olg = filter_find_olg_internal(filter, group);
2487 if (olg->olg_initializing) {
2488 GOTO(out_unlock, olg = ERR_PTR(-EBUSY));
2490 GOTO(out_unlock, olg);
2493 /* set as the newly allocated one */
2498 llog_group_init(olg, group);
2499 cfs_list_add(&olg->olg_list, &filter->fo_llog_list);
2500 olg->olg_initializing = 1;
2501 cfs_spin_unlock(&filter->fo_llog_list_lock);
2503 rc = obd_llog_init(obd, olg, obd, NULL);
2505 cfs_spin_lock(&filter->fo_llog_list_lock);
2506 cfs_list_del(&olg->olg_list);
2507 cfs_spin_unlock(&filter->fo_llog_list_lock);
2509 GOTO(out, olg = ERR_PTR(-ENOMEM));
2511 cfs_spin_lock(&filter->fo_llog_list_lock);
2512 olg->olg_initializing = 0;
2513 cfs_spin_unlock(&filter->fo_llog_list_lock);
2514 CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2515 obd->obd_name, group, olg);
2520 cfs_spin_unlock(&filter->fo_llog_list_lock);
2522 OBD_FREE_PTR(olg_new);
2526 static int filter_llog_connect(struct obd_export *exp,
2527 struct llogd_conn_body *body)
2529 struct obd_device *obd = exp->exp_obd;
2530 struct llog_ctxt *ctxt;
2531 struct obd_llog_group *olg;
2535 CDEBUG(D_OTHER, "%s: LLog connect for: "LPX64"/"LPX64":%x\n",
2536 obd->obd_name, body->lgdc_logid.lgl_oid,
2537 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2539 olg = filter_find_olg(obd, body->lgdc_logid.lgl_oseq);
2541 CERROR(" %s: can not find olg of group %d\n",
2542 obd->obd_name, (int)body->lgdc_logid.lgl_oseq);
2545 llog_group_set_export(olg, exp);
2547 ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2548 LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2549 body->lgdc_ctxt_idx);
2551 CWARN("%s: Recovery from log "LPX64"/"LPX64":%x\n",
2552 obd->obd_name, body->lgdc_logid.lgl_oid,
2553 body->lgdc_logid.lgl_oseq, body->lgdc_logid.lgl_ogen);
2555 cfs_spin_lock(&obd->u.filter.fo_flags_lock);
2556 obd->u.filter.fo_mds_ost_sync = 1;
2557 cfs_spin_unlock(&obd->u.filter.fo_flags_lock);
2558 rc = llog_connect(ctxt, &body->lgdc_logid,
2559 &body->lgdc_gen, NULL);
2560 llog_ctxt_put(ctxt);
2562 CERROR("failed to connect rc %d idx %d\n", rc,
2563 body->lgdc_ctxt_idx);
2568 static int filter_llog_preclean(struct obd_device *obd)
2570 struct obd_llog_group *olg, *tmp;
2571 struct filter_obd *filter;
2572 cfs_list_t remove_list;
2576 rc = obd_llog_finish(obd, 0);
2578 CERROR("failed to cleanup llogging subsystem\n");
2580 filter = &obd->u.filter;
2581 CFS_INIT_LIST_HEAD(&remove_list);
2583 cfs_spin_lock(&filter->fo_llog_list_lock);
2584 while (!cfs_list_empty(&filter->fo_llog_list)) {
2585 olg = cfs_list_entry(filter->fo_llog_list.next,
2586 struct obd_llog_group, olg_list);
2587 cfs_list_del(&olg->olg_list);
2588 cfs_list_add(&olg->olg_list, &remove_list);
2590 cfs_spin_unlock(&filter->fo_llog_list_lock);
2592 cfs_list_for_each_entry_safe(olg, tmp, &remove_list, olg_list) {
2593 cfs_list_del_init(&olg->olg_list);
2594 rc = filter_olg_fini(olg);
2596 CERROR("failed to cleanup llogging subsystem for %u\n",
2604 static int filter_precleanup(struct obd_device *obd,
2605 enum obd_cleanup_stage stage)
2611 case OBD_CLEANUP_EARLY:
2613 case OBD_CLEANUP_EXPORTS:
2614 /* Stop recovery before namespace cleanup. */
2615 target_recovery_fini(obd);
2616 rc = filter_llog_preclean(obd);
2622 static int filter_cleanup(struct obd_device *obd)
2624 struct filter_obd *filter = &obd->u.filter;
2628 LCONSOLE_WARN("%s: shutting down for failover; client state "
2629 "will be preserved.\n", obd->obd_name);
2631 obd_exports_barrier(obd);
2632 obd_zombie_barrier();
2634 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
2635 lprocfs_free_per_client_stats(obd);
2636 lprocfs_free_obd_stats(obd);
2637 lprocfs_obd_cleanup(obd);
2638 lquota_cleanup(filter_quota_interface_ref, obd);
2640 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
2641 obd->obd_namespace = NULL;
2643 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2645 if (obd->u.obt.obt_sb == NULL)
2650 ll_vfs_dq_off(obd->u.obt.obt_sb, 0);
2651 shrink_dcache_sb(obd->u.obt.obt_sb);
2653 server_put_mount(obd->obd_name, obd->u.obt.obt_vfsmnt);
2654 obd->u.obt.obt_sb = NULL;
2656 fsfilt_put_ops(obd->obd_fsops);
2658 filter_iobuf_pool_done(filter);
2660 LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2665 static int filter_connect_internal(struct obd_export *exp,
2666 struct obd_connect_data *data,
2669 struct filter_export_data *fed = &exp->exp_filter_data;
2674 CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2675 " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2676 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2677 data->ocd_connect_flags, data->ocd_version,
2678 data->ocd_grant, data->ocd_index);
2680 if (fed->fed_group != 0 && fed->fed_group != data->ocd_group) {
2681 CWARN("!!! This export (nid %s) used object group %d "
2682 "earlier; now it's trying to use group %d! This could "
2683 "be a bug in the MDS. Please report to "
2684 "http://bugs.whamcloud.com/\n",
2685 obd_export_nid2str(exp), fed->fed_group,data->ocd_group);
2688 fed->fed_group = data->ocd_group;
2690 data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2691 exp->exp_connect_flags = data->ocd_connect_flags;
2692 data->ocd_version = LUSTRE_VERSION_CODE;
2694 /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
2695 if (!ergo(data->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN,
2696 data->ocd_connect_flags & OBD_CONNECT_MDS))
2699 if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2700 struct filter_obd *filter = &exp->exp_obd->u.filter;
2701 obd_size left, want;
2703 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
2704 left = filter_grant_space_left(exp);
2705 want = data->ocd_grant;
2706 filter_grant(exp, fed->fed_grant, want, left, (reconnect == 0));
2707 data->ocd_grant = fed->fed_grant;
2708 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
2710 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2711 LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2712 exp->exp_client_uuid.uuid, exp,
2713 data->ocd_grant, want, left);
2715 filter->fo_tot_granted_clients ++;
2718 if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2719 struct lr_server_data *lsd = class_server_data(exp->exp_obd);
2720 int index = le32_to_cpu(lsd->lsd_ost_index);
2722 if (!(lsd->lsd_feature_compat &
2723 cpu_to_le32(OBD_COMPAT_OST))) {
2724 /* this will only happen on the first connect */
2725 lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2726 lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2727 /* sync is not needed here as filter_client_add will
2728 * set exp_need_sync flag */
2729 filter_update_server_data(exp->exp_obd);
2730 } else if (index != data->ocd_index) {
2731 LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2732 " %u doesn't match actual OST index"
2733 " %u in last_rcvd file, bad "
2735 obd_export_nid2str(exp), index,
2739 /* FIXME: Do the same with the MDS UUID and lsd_peeruuid.
2740 * FIXME: We don't strictly need the COMPAT flag for that,
2741 * FIXME: as lsd_peeruuid[0] will tell us if that is set.
2742 * FIXME: We needed it for the index, as index 0 is valid. */
2745 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2746 data->ocd_brw_size = 65536;
2747 } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2748 data->ocd_brw_size = min(data->ocd_brw_size,
2749 (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2750 if (data->ocd_brw_size == 0) {
2751 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
2752 " ocd_version: %x ocd_grant: %d ocd_index: %u "
2753 "ocd_brw_size is unexpectedly zero, "
2754 "network data corruption?"
2755 "Refusing connection of this client\n",
2756 exp->exp_obd->obd_name,
2757 exp->exp_client_uuid.uuid,
2758 exp, data->ocd_connect_flags, data->ocd_version,
2759 data->ocd_grant, data->ocd_index);
2764 if (data->ocd_connect_flags & OBD_CONNECT_CKSUM) {
2765 __u32 cksum_types = data->ocd_cksum_types;
2767 /* The client set in ocd_cksum_types the checksum types it
2768 * supports. We have to mask off the algorithms that we don't
2770 if (cksum_types & OBD_CKSUM_ALL)
2771 data->ocd_cksum_types &= OBD_CKSUM_ALL;
2773 data->ocd_cksum_types = OBD_CKSUM_CRC32;
2775 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
2776 "%x\n", exp->exp_obd->obd_name,
2777 obd_export_nid2str(exp), cksum_types,
2778 data->ocd_cksum_types);
2780 /* This client does not support OBD_CONNECT_CKSUM
2781 * fall back to CRC32 */
2782 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
2783 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
2784 exp->exp_obd->obd_name,
2785 obd_export_nid2str(exp));
2788 if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
2789 data->ocd_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
2794 static int filter_reconnect(const struct lu_env *env,
2795 struct obd_export *exp, struct obd_device *obd,
2796 struct obd_uuid *cluuid,
2797 struct obd_connect_data *data,
2803 if (exp == NULL || obd == NULL || cluuid == NULL)
2806 rc = filter_connect_internal(exp, data, 1);
2808 filter_export_stats_init(obd, exp, localdata);
2813 /* nearly identical to mds_connect */
2814 static int filter_connect(const struct lu_env *env,
2815 struct obd_export **exp, struct obd_device *obd,
2816 struct obd_uuid *cluuid,
2817 struct obd_connect_data *data, void *localdata)
2819 struct lvfs_run_ctxt saved;
2820 struct lustre_handle conn = { 0 };
2821 struct obd_export *lexp;
2826 if (exp == NULL || obd == NULL || cluuid == NULL)
2829 rc = class_connect(&conn, obd, cluuid);
2832 lexp = class_conn2export(&conn);
2833 LASSERT(lexp != NULL);
2835 rc = filter_connect_internal(lexp, data, 0);
2839 filter_export_stats_init(obd, lexp, localdata);
2840 if (obd->obd_replayable) {
2841 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
2843 memcpy(lcd->lcd_uuid, cluuid, sizeof(lcd->lcd_uuid));
2844 rc = filter_client_add(obd, lexp, -1);
2849 group = data->ocd_group;
2851 CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2852 obd->obd_name, lexp->exp_handle.h_cookie, group);
2854 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2855 rc = filter_read_groups(obd, group, 1);
2856 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2858 CERROR("can't read group %u\n", group);
2866 class_disconnect(lexp);
2875 /* Do extra sanity checks for grant accounting. We do this at connect,
2876 * disconnect, and statfs RPC time, so it shouldn't be too bad. We can
2877 * always get rid of it or turn it off when we know accounting is good. */
2878 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2880 struct filter_export_data *fed;
2881 struct obd_export *exp;
2882 obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2883 obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2884 obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2886 if (cfs_list_empty(&obd->obd_exports))
2889 /* We don't want to do this for large machines that do lots of
2890 mounts or unmounts. It burns... */
2891 if (obd->obd_num_exports > 100)
2894 cfs_spin_lock(&obd->obd_osfs_lock);
2895 cfs_spin_lock(&obd->obd_dev_lock);
2896 cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2898 fed = &exp->exp_filter_data;
2899 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2902 if (maxsize > 0) { /* we may not have done a statfs yet */
2903 LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2904 "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2905 exp->exp_client_uuid.uuid, exp,
2906 fed->fed_grant, fed->fed_pending, maxsize);
2907 LASSERTF(fed->fed_dirty <= maxsize,
2908 "%s: cli %s/%p %ld > "LPU64"\n", func,
2909 exp->exp_client_uuid.uuid, exp,
2910 fed->fed_dirty, maxsize);
2913 CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2914 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2915 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2917 CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2918 obd->obd_name, exp->exp_client_uuid.uuid, exp,
2919 fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2920 tot_granted += fed->fed_grant + fed->fed_pending;
2921 tot_pending += fed->fed_pending;
2922 tot_dirty += fed->fed_dirty;
2924 fo_tot_granted = obd->u.filter.fo_tot_granted;
2925 fo_tot_pending = obd->u.filter.fo_tot_pending;
2926 fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2927 cfs_spin_unlock(&obd->obd_dev_lock);
2928 cfs_spin_unlock(&obd->obd_osfs_lock);
2930 /* Do these assertions outside the spinlocks so we don't kill system */
2931 if (tot_granted != fo_tot_granted)
2932 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2933 func, tot_granted, fo_tot_granted);
2934 if (tot_pending != fo_tot_pending)
2935 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2936 func, tot_pending, fo_tot_pending);
2937 if (tot_dirty != fo_tot_dirty)
2938 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2939 func, tot_dirty, fo_tot_dirty);
2940 if (tot_pending > tot_granted)
2941 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2942 func, tot_pending, tot_granted);
2943 if (tot_granted > maxsize)
2944 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2945 func, tot_granted, maxsize);
2946 if (tot_dirty > maxsize)
2947 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2948 func, tot_dirty, maxsize);
2951 /* Remove this client from the grant accounting totals. We also remove
2952 * the export from the obd device under the osfs and dev locks to ensure
2953 * that the filter_grant_sanity_check() calculations are always valid.
2954 * The client should do something similar when it invalidates its import. */
2955 static void filter_grant_discard(struct obd_export *exp)
2957 struct obd_device *obd = exp->exp_obd;
2958 struct filter_obd *filter = &obd->u.filter;
2959 struct filter_export_data *fed = &exp->exp_filter_data;
2961 cfs_spin_lock(&obd->obd_osfs_lock);
2962 LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2963 "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2964 obd->obd_name, filter->fo_tot_granted,
2965 exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2966 filter->fo_tot_granted -= fed->fed_grant;
2967 LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2968 "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2969 obd->obd_name, filter->fo_tot_pending,
2970 exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2971 /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2972 LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2973 "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2974 obd->obd_name, filter->fo_tot_dirty,
2975 exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2976 filter->fo_tot_dirty -= fed->fed_dirty;
2980 cfs_spin_unlock(&obd->obd_osfs_lock);
2983 static int filter_destroy_export(struct obd_export *exp)
2985 struct filter_export_data *fed = &exp->exp_filter_data;
2988 if (fed->fed_pending)
2989 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2990 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2991 exp, fed->fed_pending);
2993 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2995 target_destroy_export(exp);
2996 ldlm_destroy_export(exp);
2997 lut_client_free(exp);
2999 if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
3002 if (!exp->exp_obd->obd_replayable)
3003 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
3005 filter_grant_discard(exp);
3006 filter_fmd_cleanup(exp);
3008 if (exp->exp_connect_flags & OBD_CONNECT_GRANT_SHRINK) {
3009 struct filter_obd *filter = &exp->exp_obd->u.filter;
3010 if (filter->fo_tot_granted_clients > 0)
3011 filter->fo_tot_granted_clients --;
3014 if (!(exp->exp_flags & OBD_OPT_FORCE))
3015 filter_grant_sanity_check(exp->exp_obd, __func__);
3020 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
3022 struct obd_llog_group *olg_min, *olg;
3023 struct filter_obd *filter;
3024 int worked = -1, group;
3025 struct llog_ctxt *ctxt;
3028 filter = &obd->u.filter;
3030 /* we can't sync log holding spinlock. also, we do not want to get
3031 * into livelock. so we do following: loop over MDS's exports in
3032 * group order and skip already synced llogs -bzzz */
3034 /* look for group with min. number, but > worked */
3037 cfs_spin_lock(&filter->fo_llog_list_lock);
3038 cfs_list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
3039 if (olg->olg_seq <= worked) {
3040 /* this group is already synced */
3043 if (group < olg->olg_seq) {
3044 /* we have group with smaller number to sync */
3047 /* store current minimal group */
3049 group = olg->olg_seq;
3051 cfs_spin_unlock(&filter->fo_llog_list_lock);
3053 if (olg_min == NULL)
3056 worked = olg_min->olg_seq;
3057 if (olg_min->olg_exp &&
3058 (dexp == olg_min->olg_exp || dexp == NULL)) {
3060 ctxt = llog_group_get_ctxt(olg_min,
3061 LLOG_MDS_OST_REPL_CTXT);
3063 err = llog_sync(ctxt, olg_min->olg_exp);
3064 llog_ctxt_put(ctxt);
3066 CERROR("error flushing logs to MDS: "
3071 } while (olg_min != NULL);
3074 /* Also incredibly similar to mds_disconnect */
3075 static int filter_disconnect(struct obd_export *exp)
3077 struct obd_device *obd = exp->exp_obd;
3082 class_export_get(exp);
3084 if (!(exp->exp_flags & OBD_OPT_FORCE))
3085 filter_grant_sanity_check(obd, __func__);
3086 filter_grant_discard(exp);
3088 /* Flush any remaining cancel messages out to the target */
3089 filter_sync_llogs(obd, exp);
3091 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
3093 rc = server_disconnect_export(exp);
3095 /* Do not erase record for recoverable client. */
3096 if (obd->obd_replayable && (!obd->obd_fail || exp->exp_failed))
3097 filter_client_del(exp);
3099 fsfilt_sync(obd, obd->u.obt.obt_sb);
3101 class_export_put(exp);
3105 /* reverse import is changed, sync all cancels */
3106 static void filter_revimp_update(struct obd_export *exp)
3111 class_export_get(exp);
3113 /* flush any remaining cancel messages out to the target */
3114 filter_sync_llogs(exp->exp_obd, exp);
3115 class_export_put(exp);
3119 static int filter_ping(struct obd_export *exp)
3121 filter_fmd_expire(exp);
3125 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct ost_id *ostid,
3126 const char *what, int quiet)
3128 struct dentry *dchild = NULL;
3130 dchild = filter_fid2dentry(obd, NULL, ostid->oi_seq, ostid->oi_id);
3132 if (IS_ERR(dchild)) {
3133 CERROR("%s error looking up object: "POSTID"\n",
3134 what, ostid->oi_id, ostid->oi_seq);
3138 if (dchild->d_inode == NULL) {
3140 CERROR("%s: %s on non-existent object: "POSTID" \n",
3141 obd->obd_name, what, ostid->oi_id,ostid->oi_seq);
3143 RETURN(ERR_PTR(-ENOENT));
3149 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
3151 struct dentry *dentry = NULL;
3152 struct obd_device *obd;
3156 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
3157 oinfo_capa(oinfo), CAPA_OPC_META_READ);
3161 obd = class_exp2obd(exp);
3163 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3167 dentry = filter_oa2dentry(obd, &oinfo->oi_oa->o_oi);
3169 RETURN(PTR_ERR(dentry));
3171 /* Limit the valid bits in the return data to what we actually use */
3172 oinfo->oi_oa->o_valid = OBD_MD_FLID;
3173 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL, FILTER_VALID_FLAGS);
3179 /* this should be enabled/disabled in condition to enabled/disabled large
3180 * inodes (fast EAs) in backing store FS. */
3181 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
3182 void *handle, struct obdo *oa)
3184 struct obd_device *obd = exp->exp_obd;
3188 if (oa->o_valid & OBD_MD_FLFID) {
3189 struct filter_fid ff;
3191 if (!(oa->o_valid & OBD_MD_FLGROUP))
3193 /* packing fid and converting it to LE for storing into EA.
3194 * Here ->o_stripe_idx should be filled by LOV and rest of
3195 * fields - by client. */
3196 ff.ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
3197 ff.ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
3198 /* XXX: we are ignoring o_parent_ver here, since this should
3199 * be the same for all objects in this fileset. */
3200 ff.ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
3201 ff.ff_objid = cpu_to_le64(oa->o_id);
3202 ff.ff_seq = cpu_to_le64(oa->o_seq);
3204 CDEBUG(D_INODE, "storing filter fid EA (parent "DFID" "
3205 LPU64"/"LPU64")\n", PFID(&ff.ff_parent), oa->o_id,
3208 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
3210 CERROR("store fid in object failed! rc: %d\n", rc);
3212 CDEBUG(D_HA, "OSS object without fid info!\n");
3218 /* this is called from filter_truncate() until we have filter_punch() */
3219 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
3220 struct obdo *oa, struct obd_trans_info *oti)
3222 unsigned int orig_ids[MAXQUOTAS] = {0, 0};
3223 struct llog_cookie *fcc = NULL;
3224 struct filter_obd *filter;
3225 int rc, err, sync = 0;
3226 loff_t old_size = 0;
3227 unsigned int ia_valid;
3228 struct inode *inode;
3229 struct page *page = NULL;
3234 LASSERT(dentry != NULL);
3235 LASSERT(!IS_ERR(dentry));
3237 inode = dentry->d_inode;
3238 LASSERT(inode != NULL);
3240 filter = &exp->exp_obd->u.filter;
3241 iattr_from_obdo(&iattr, oa, oa->o_valid);
3242 ia_valid = iattr.ia_valid;
3244 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3245 OBD_ALLOC(fcc, sizeof(*fcc));
3247 *fcc = oa->o_lcookie;
3249 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3250 unsigned long now = jiffies;
3251 ll_vfs_dq_init(inode);
3252 /* Filter truncates and writes are serialized by
3253 * i_alloc_sem, see the comment in
3254 * filter_preprw_write.*/
3255 if (ia_valid & ATTR_SIZE)
3256 down_write(&inode->i_alloc_sem);
3257 LOCK_INODE_MUTEX(inode);
3258 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
3259 old_size = i_size_read(inode);
3262 /* VBR: version recovery check */
3263 rc = filter_version_get_check(exp, oti, inode);
3265 GOTO(out_unlock, rc);
3267 /* Let's pin the last page so that ldiskfs_truncate
3268 * should not start GFP_FS allocation. */
3269 if (ia_valid & ATTR_SIZE) {
3270 page = grab_cache_page(inode->i_mapping,
3271 iattr.ia_size >> PAGE_CACHE_SHIFT);
3273 GOTO(out_unlock, rc = -ENOMEM);
3278 /* If the inode still has SUID+SGID bits set (see filter_precreate())
3279 * then we will accept the UID+GID sent by the client during write for
3280 * initializing the ownership of this inode. We only allow this to
3281 * happen once so clear these bits in setattr. In 2.6 kernels it is
3282 * possible to get ATTR_UID and ATTR_GID separately, so we only clear
3283 * the flags that are actually being set. */
3284 if (ia_valid & (ATTR_UID | ATTR_GID)) {
3285 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
3286 (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
3288 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
3289 if (!(ia_valid & ATTR_MODE)) {
3290 iattr.ia_mode = inode->i_mode;
3291 iattr.ia_valid |= ATTR_MODE;
3293 iattr.ia_mode &= ~S_ISUID;
3295 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
3296 if (!(iattr.ia_valid & ATTR_MODE)) {
3297 iattr.ia_mode = inode->i_mode;
3298 iattr.ia_valid |= ATTR_MODE;
3300 iattr.ia_mode &= ~S_ISGID;
3303 orig_ids[USRQUOTA] = inode->i_uid;
3304 orig_ids[GRPQUOTA] = inode->i_gid;
3305 handle = fsfilt_start_log(exp->exp_obd, inode,
3306 FSFILT_OP_SETATTR, oti, 1);
3308 GOTO(out_unlock, rc = PTR_ERR(handle));
3310 /* update inode EA only once when inode is suid bit marked. As
3311 * on 2.6.x UID and GID may be set separately, we check here
3312 * only one of them to avoid double setting. */
3313 if (inode->i_mode & S_ISUID)
3314 filter_update_fidea(exp, inode, handle, oa);
3316 handle = fsfilt_start(exp->exp_obd, inode,
3317 FSFILT_OP_SETATTR, oti);
3319 GOTO(out_unlock, rc = PTR_ERR(handle));
3321 if (oa->o_valid & OBD_MD_FLFLAGS) {
3322 rc = fsfilt_iocontrol(exp->exp_obd, dentry,
3323 FSFILT_IOC_SETFLAGS, (long)&oa->o_flags);
3325 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
3327 /* set cancel cookie callback function */
3328 sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
3329 filter_cancel_cookies_cb,
3333 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
3334 fsfilt_extend(exp->exp_obd, inode, 0, handle);
3336 /* The truncate might have used up our transaction credits. Make sure
3337 * we have two left for the last_rcvd and VBR inode version updates. */
3338 err = fsfilt_extend(exp->exp_obd, inode, 2, handle);
3340 /* Update inode version only if data has changed => size has changed */
3341 rc = filter_finish_transno(exp, ia_valid & ATTR_SIZE ? inode : NULL,
3345 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
3349 err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
3351 CERROR("error on commit, err = %d\n", err);
3358 /* For a partial-page truncate flush the page to disk immediately
3359 * to avoid data corruption during direct disk write. b=17397 */
3360 if (!sync && (iattr.ia_valid & ATTR_SIZE) &&
3361 old_size != iattr.ia_size && (iattr.ia_size & ~CFS_PAGE_MASK)) {
3362 err = filemap_fdatawrite_range(inode->i_mapping, iattr.ia_size,
3372 page_cache_release(page);
3374 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
3375 UNLOCK_INODE_MUTEX(inode);
3376 if (ia_valid & ATTR_SIZE)
3377 up_write(&inode->i_alloc_sem);
3379 OBD_FREE(fcc, sizeof(*fcc));
3381 /* trigger quota release */
3382 if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3383 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3384 int rc2 = lquota_adjust(filter_quota_interface_ref,
3385 exp->exp_obd, cur_ids,
3386 orig_ids, rc, FSFILT_OP_SETATTR);
3387 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3388 "filter adjust qunit. (rc:%d)\n", rc2);
3393 /* this is called from filter_truncate() until we have filter_punch() */
3394 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3395 struct obd_trans_info *oti)
3397 struct obdo *oa = oinfo->oi_oa;
3398 struct lustre_capa *capa = oinfo_capa(oinfo);
3399 struct ldlm_res_id res_id;
3400 struct filter_mod_data *fmd;
3401 struct lvfs_run_ctxt saved;
3402 struct filter_obd *filter;
3403 struct ldlm_resource *res;
3404 struct dentry *dentry;
3405 __u64 opc = CAPA_OPC_META_WRITE;
3409 if (oa->o_valid & OBD_FL_TRUNC)
3410 opc |= CAPA_OPC_OSS_TRUNC;
3412 rc = filter_auth_capa(exp, NULL, oa->o_seq, capa, opc);
3416 if (oa->o_valid & (OBD_MD_FLUID | OBD_MD_FLGID)) {
3417 rc = filter_capa_fixoa(exp, oa, oa->o_seq, capa);
3422 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
3423 /* This would be very bad - accidentally truncating a file when
3424 * changing the time or similar - bug 12203. */
3425 if (oa->o_valid & OBD_MD_FLSIZE &&
3426 oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3427 static char mdsinum[48];
3429 if (oa->o_valid & OBD_MD_FLFID)
3430 snprintf(mdsinum, sizeof(mdsinum) - 1, " of inode "DFID,
3431 oa->o_parent_seq, oa->o_parent_oid,
3436 CERROR("%s: setattr from %s trying to truncate objid "POSTID
3437 "%s\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
3438 oa->o_id, oa->o_seq, mdsinum);
3442 dentry = __filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi, __func__, 1);
3444 RETURN(PTR_ERR(dentry));
3446 filter = &exp->exp_obd->u.filter;
3447 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3450 * We need to be atomic against a concurrent write
3451 * (which takes the semaphore for reading). fmd_mactime_xid
3452 * checks will have no effect if a write request with lower
3453 * xid starts just before a setattr and finishes later than
3454 * the setattr (see bug 21489, comment 27).
3457 (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3458 unsigned long now = jiffies;
3459 down_write(&dentry->d_inode->i_alloc_sem);
3460 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem");
3461 fmd = filter_fmd_get(exp, oa->o_id, oa->o_seq);
3462 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3463 fmd->fmd_mactime_xid = oti->oti_xid;
3464 filter_fmd_put(exp, fmd);
3465 up_write(&dentry->d_inode->i_alloc_sem);
3468 /* setting objects attributes (including owner/group) */
3469 rc = filter_setattr_internal(exp, dentry, oa, oti);
3471 GOTO(out_unlock, rc);
3473 res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3474 &res_id, LDLM_EXTENT, 0);
3477 LDLM_RESOURCE_ADDREF(res);
3478 rc = ldlm_res_lvbo_update(res, NULL, 0);
3479 LDLM_RESOURCE_DELREF(res);
3480 ldlm_resource_putref(res);
3483 oa->o_valid = OBD_MD_FLID;
3485 /* Quota release need uid/gid info */
3486 obdo_from_inode(oa, dentry->d_inode, NULL,
3487 FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3492 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3496 /* XXX identical to osc_unpackmd */
3497 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3498 struct lov_mds_md *lmm, int lmm_bytes)
3504 if (lmm_bytes < sizeof (*lmm)) {
3505 CERROR("lov_mds_md too small: %d, need %d\n",
3506 lmm_bytes, (int)sizeof(*lmm));
3509 /* XXX LOV_MAGIC etc check? */
3511 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3512 CERROR("lov_mds_md: zero lmm_object_id\n");
3517 lsm_size = lov_stripe_md_size(1);
3521 if (*lsmp != NULL && lmm == NULL) {
3522 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3523 OBD_FREE(*lsmp, lsm_size);
3528 if (*lsmp == NULL) {
3529 OBD_ALLOC(*lsmp, lsm_size);
3533 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3534 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3535 OBD_FREE(*lsmp, lsm_size);
3538 loi_init((*lsmp)->lsm_oinfo[0]);
3542 /* XXX zero *lsmp? */
3543 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3544 LASSERT((*lsmp)->lsm_object_id);
3547 (*lsmp)->lsm_maxbytes = exp->exp_obd->u.obt.obt_sb->s_maxbytes;
3552 /* caller must hold fo_create_locks[oa->o_seq] */
3553 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3554 struct filter_obd *filter)
3556 struct obdo doa = { 0 }; /* XXX obdo on stack */
3562 LASSERT(down_trylock(&filter->fo_create_locks[oa->o_seq]) != 0);
3564 memset(&doa, 0, sizeof(doa));
3566 doa.o_valid |= OBD_MD_FLGROUP;
3567 doa.o_seq = oa->o_seq;
3568 doa.o_mode = S_IFREG;
3570 if (!cfs_test_bit(doa.o_seq, &filter->fo_destroys_in_progress)) {
3571 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3572 exp->exp_obd->obd_name, doa.o_seq);
3576 last = filter_last_id(filter, doa.o_seq);
3578 skip_orphan = !!(exp->exp_connect_flags & OBD_CONNECT_SKIP_ORPHAN);
3580 CDEBUG(D_HA, "%s: deleting orphan objects from "LPU64" to "LPU64"%s\n",
3581 exp->exp_obd->obd_name, oa->o_id + 1, last,
3582 skip_orphan ? ", orphan objids won't be reused any more." : ".");
3584 for (id = last; id > oa->o_id; id--) {
3586 rc = filter_destroy(exp, &doa, NULL, NULL, NULL, NULL);
3587 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3588 CEMERG("error destroying precreate objid "LPU64": %d\n",
3591 /* update last_id on disk periodically so that if we restart
3592 * we don't need to re-scan all of the just-deleted objects. */
3593 if ((id & 511) == 0 && !skip_orphan) {
3594 filter_set_last_id(filter, id - 1, doa.o_seq);
3595 filter_update_last_objid(exp->exp_obd, doa.o_seq, 0);
3599 CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3600 exp->exp_obd->obd_name, doa.o_seq, oa->o_id);
3603 filter_set_last_id(filter, id, doa.o_seq);
3604 rc = filter_update_last_objid(exp->exp_obd, doa.o_seq, 1);
3607 * We have destroyed orphan objects, but don't want to reuse
3608 * them. Therefore we don't reset last_id to the last created
3609 * objects. Instead, we report back to the MDS the object id
3610 * of the last orphan, so that the MDS can restart allocating
3611 * objects from this id + 1 and thus skip the whole orphan
3617 cfs_clear_bit(doa.o_seq, &filter->fo_destroys_in_progress);
3622 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3623 obd_seq group, int *num);
3624 /* returns a negative error or a nonnegative number of files to create */
3625 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3626 obd_seq group, struct obd_trans_info *oti)
3628 struct obd_device *obd = exp->exp_obd;
3629 struct filter_obd *filter = &obd->u.filter;
3633 /* delete orphans request */
3634 if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3635 obd_id last = filter_last_id(filter, group);
3637 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3638 CERROR("%s: dropping old orphan cleanup request\n",
3642 /* This causes inflight precreates to abort and drop lock */
3643 cfs_set_bit(group, &filter->fo_destroys_in_progress);
3644 cfs_down(&filter->fo_create_locks[group]);
3645 if (!cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3646 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3647 exp->exp_obd->obd_name, group);
3648 cfs_up(&filter->fo_create_locks[group]);
3651 diff = oa->o_id - last;
3652 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3655 if (-diff > OST_MAX_PRECREATE) {
3656 CERROR("%s: ignoring bogus orphan destroy request: "
3657 "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3659 /* FIXME: should reset precreate_next_id on MDS */
3660 GOTO(out, rc = -EINVAL);
3663 rc = filter_destroy_precreated(exp, oa, filter);
3665 CERROR("%s: unable to write lastobjid, but "
3666 "orphans were deleted\n", obd->obd_name);
3669 /* XXX: Used by MDS for the first time! */
3670 cfs_clear_bit(group, &filter->fo_destroys_in_progress);
3673 cfs_down(&filter->fo_create_locks[group]);
3674 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3675 CERROR("%s: dropping old precreate request\n",
3679 /* only precreate if group == 0 and o_id is specfied */
3680 if (!fid_seq_is_mdt(group) || oa->o_id == 0)
3683 diff = oa->o_id - filter_last_id(filter, group);
3684 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3685 filter_last_id(filter, group), diff);
3687 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3688 oa->o_id, filter_last_id(filter, group), diff);
3692 oa->o_id = filter_last_id(&obd->u.filter, group);
3693 rc = filter_precreate(obd, oa, group, &diff);
3694 oa->o_id = filter_last_id(&obd->u.filter, group);
3696 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3699 /* else diff == 0 */
3702 cfs_up(&filter->fo_create_locks[group]);
3706 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3707 __u64 max_age, __u32 flags)
3709 struct filter_obd *filter = &obd->u.filter;
3710 int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3714 /* at least try to account for cached pages. its still racey and
3715 * might be under-reporting if clients haven't announced their
3716 * caches with brw recently */
3717 cfs_spin_lock(&obd->obd_osfs_lock);
3718 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3719 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3720 cfs_spin_unlock(&obd->obd_osfs_lock);
3722 CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3723 " pending "LPU64" free "LPU64" avail "LPU64"\n",
3724 filter->fo_tot_dirty, filter->fo_tot_granted,
3725 filter->fo_tot_pending,
3726 osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3728 filter_grant_sanity_check(obd, __func__);
3730 osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3731 ((filter->fo_tot_dirty + filter->fo_tot_pending +
3732 osfs->os_bsize - 1) >> blockbits));
3734 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) {
3735 struct lr_server_data *lsd = class_server_data(obd);
3736 int index = le32_to_cpu(lsd->lsd_ost_index);
3738 if (cfs_fail_val == -1 || index == cfs_fail_val)
3739 osfs->os_bfree = osfs->os_bavail = 2;
3740 else if (cfs_fail_loc & OBD_FAIL_ONCE)
3741 cfs_fail_loc &= ~OBD_FAILED; /* reset flag */
3744 /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3745 * stop creating files on MDS if OST is not good shape to create
3749 if (filter->fo_obt.obt_sb->s_flags & MS_RDONLY)
3750 osfs->os_state = OS_STATE_READONLY;
3752 if (filter->fo_raid_degraded)
3753 osfs->os_state |= OS_STATE_DEGRADED;
3757 static int filter_use_existing_obj(struct obd_device *obd,
3758 struct dentry *dchild, void **handle,
3761 struct inode *inode = dchild->d_inode;
3765 if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3768 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3769 if (IS_ERR(*handle))
3770 return PTR_ERR(*handle);
3772 iattr.ia_valid = ATTR_MODE;
3773 iattr.ia_mode = S_ISUID | S_ISGID |0666;
3774 rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3781 static __u64 filter_calc_free_inodes(struct obd_device *obd)
3784 __u64 os_ffree = -1;
3786 cfs_spin_lock(&obd->obd_osfs_lock);
3787 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, cfs_time_shift_64(1));
3789 os_ffree = obd->obd_osfs.os_ffree;
3790 cfs_spin_unlock(&obd->obd_osfs_lock);
3795 /* We rely on the fact that only one thread will be creating files in a given
3796 * group at a time, which is why we don't need an atomic filter_get_new_id.
3797 * Even if we had that atomic function, the following race would exist:
3799 * thread 1: gets id x from filter_next_id
3800 * thread 2: gets id (x + 1) from filter_next_id
3801 * thread 2: creates object (x + 1)
3802 * thread 1: tries to create object x, gets -ENOSPC
3804 * Caller must hold fo_create_locks[group]
3806 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3807 obd_seq group, int *num)
3809 struct dentry *dchild = NULL, *dparent = NULL;
3810 struct filter_obd *filter;
3811 struct obd_statfs *osfs;
3812 int err = 0, rc = 0, recreate_obj = 0, i;
3813 cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT/2);
3816 void *handle = NULL;
3819 filter = &obd->u.filter;
3821 LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3823 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3825 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3826 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3829 OBD_ALLOC(osfs, sizeof(*osfs));
3832 rc = filter_statfs(obd, osfs,
3833 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
3835 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3836 CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3837 LPU64"\n", obd->obd_name, osfs->os_bavail <<
3838 obd->u.obt.obt_vfsmnt->mnt_sb->s_blocksize_bits);
3842 OBD_FREE(osfs, sizeof(*osfs));
3847 CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3848 " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3850 for (i = 0; i < *num && err == 0; i++) {
3851 int cleanup_phase = 0;
3853 if (cfs_test_bit(group, &filter->fo_destroys_in_progress)) {
3854 CWARN("%s: create aborted by destroy\n",
3863 last_id = filter_last_id(filter, group);
3864 if (next_id > last_id) {
3865 CERROR("Error: Trying to recreate obj greater"
3866 "than last id "LPD64" > "LPD64"\n",
3868 GOTO(cleanup, rc = -EINVAL);
3871 next_id = filter_last_id(filter, group) + 1;
3873 /* Don't create objects beyond the valid range for this SEQ */
3874 if (unlikely(fid_seq_is_mdt0(group) &&
3875 next_id >= IDIF_MAX_OID)) {
3876 CERROR("%s:"POSTID" hit the IDIF_MAX_OID (1<<48)!\n",
3877 obd->obd_name, next_id, group);
3878 GOTO(cleanup, rc = -ENOSPC);
3879 } else if (unlikely(!fid_seq_is_mdt0(group) &&
3880 next_id >= OBIF_MAX_OID)) {
3881 CERROR("%s:"POSTID" hit the OBIF_MAX_OID (1<<32)!\n",
3882 obd->obd_name, next_id, group);
3883 GOTO(cleanup, rc = -ENOSPC);
3886 dparent = filter_parent_lock(obd, group, next_id);
3887 if (IS_ERR(dparent))
3888 GOTO(cleanup, rc = PTR_ERR(dparent));
3889 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3891 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3893 GOTO(cleanup, rc = PTR_ERR(dchild));
3894 cleanup_phase = 2; /* f_dput(dchild) */
3896 if (dchild->d_inode != NULL) {
3897 /* This would only happen if lastobjid was bad on disk*/
3898 /* Could also happen if recreating missing obj but it
3899 * already exists. */
3901 CERROR("%s: recreating existing object %.*s?\n",
3902 obd->obd_name, dchild->d_name.len,
3903 dchild->d_name.name);
3905 /* Use these existing objects if they are
3907 if (dchild->d_inode->i_size == 0) {
3908 rc = filter_use_existing_obj(obd,dchild,
3909 &handle, &cleanup_phase);
3916 CERROR("%s: Serious error: objid %.*s already "
3917 "exists; is this filesystem corrupt?\n",
3918 obd->obd_name, dchild->d_name.len,
3919 dchild->d_name.name);
3922 GOTO(cleanup, rc = -EEXIST);
3925 handle = fsfilt_start_log(obd, dparent->d_inode,
3926 FSFILT_OP_CREATE, NULL, 1);
3928 GOTO(cleanup, rc = PTR_ERR(handle));
3931 CDEBUG(D_INODE, "%s: filter_precreate(od->o_seq="LPU64
3932 ",od->o_id="LPU64")\n", obd->obd_name, group,
3935 /* We mark object SUID+SGID to flag it for accepting UID+GID
3936 * from client on first write. Currently the permission bits
3937 * on the OST are never used, so this is OK. */
3938 rc = ll_vfs_create(dparent->d_inode, dchild,
3939 S_IFREG | S_ISUID | S_ISGID | 0666, NULL);
3941 CERROR("create failed rc = %d\n", rc);
3942 if (rc == -ENOSPC) {
3943 os_ffree = filter_calc_free_inodes(obd);
3945 CERROR("%s: free inode "LPU64"\n",
3946 obd->obd_name, os_ffree);
3951 if (dchild->d_inode)
3952 CDEBUG(D_INFO, "objid "LPU64" got inum %lu\n", next_id,
3953 dchild->d_inode->i_ino);
3956 if (!recreate_obj) {
3957 filter_set_last_id(filter, next_id, group);
3958 err = filter_update_last_objid(obd, group, 0);
3960 CERROR("unable to write lastobjid "
3961 "but file created\n");
3965 switch(cleanup_phase) {
3967 err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3969 CERROR("error on commit, err = %d\n", err);
3976 filter_parent_unlock(dparent);
3983 if (cfs_time_after(jiffies, enough_time)) {
3986 "%s: precreate slow - want %d got %d \n",
3987 obd->obd_name, *num, i);
3994 "%s: created %d objects for group "POSTID" rc %d\n",
3995 obd->obd_name, i, filter->fo_last_objids[group], group, rc);
4000 int filter_create(struct obd_export *exp, struct obdo *oa,
4001 struct lov_stripe_md **ea, struct obd_trans_info *oti)
4003 struct obd_device *obd = exp->exp_obd;
4004 struct filter_export_data *fed;
4005 struct filter_obd *filter;
4006 struct lvfs_run_ctxt saved;
4007 struct lov_stripe_md *lsm = NULL;
4011 CDEBUG(D_INODE, "%s: filter_create(group="LPU64",id="
4012 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4014 fed = &exp->exp_filter_data;
4015 filter = &obd->u.filter;
4017 if (fed->fed_group != oa->o_seq) {
4018 CERROR("%s: this export (nid %s) used object group %d "
4019 "earlier; now it's trying to use group "LPU64"!"
4020 " This could be a bug in the MDS. Please report to "
4021 "http://bugzilla.lustre.org/\n", obd->obd_name,
4022 obd_export_nid2str(exp), fed->fed_group, oa->o_seq);
4029 rc = obd_alloc_memmd(exp, &lsm);
4036 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4038 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
4039 (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
4040 if (!obd->obd_recovering ||
4041 oa->o_id > filter_last_id(filter, oa->o_seq)) {
4042 CERROR("recreate objid "LPU64" > last id "LPU64"\n",
4043 oa->o_id, filter_last_id(filter, oa->o_seq));
4047 cfs_down(&filter->fo_create_locks[oa->o_seq]);
4048 rc = filter_precreate(obd, oa, oa->o_seq, &diff);
4049 cfs_up(&filter->fo_create_locks[oa->o_seq]);
4052 rc = filter_handle_precreate(exp, oa, oa->o_seq, oti);
4055 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4056 if (rc && ea != NULL && *ea != lsm) {
4057 obd_free_memmd(exp, &lsm);
4058 } else if (rc == 0 && ea != NULL) {
4059 /* XXX LOV STACKING: the lsm that is passed to us from
4060 * LOV does not have valid lsm_oinfo data structs, so
4061 * don't go touching that. This needs to be fixed in a
4063 lsm->lsm_object_id = oa->o_id;
4070 int filter_destroy(struct obd_export *exp, struct obdo *oa,
4071 struct lov_stripe_md *md, struct obd_trans_info *oti,
4072 struct obd_export *md_exp, void *capa)
4074 unsigned int qcids[MAXQUOTAS] = {0, 0};
4075 struct obd_device *obd;
4076 struct filter_obd *filter;
4077 struct dentry *dchild = NULL, *dparent = NULL;
4078 struct lustre_handle lockh = { 0 };
4079 struct lvfs_run_ctxt saved;
4080 void *handle = NULL;
4081 struct llog_cookie *fcc = NULL;
4082 int rc, rc2, cleanup_phase = 0, sync = 0;
4087 rc = filter_auth_capa(exp, NULL, oa->o_seq,
4088 (struct lustre_capa *)capa, CAPA_OPC_OSS_DESTROY);
4093 filter = &obd->u.filter;
4095 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4098 CDEBUG(D_INODE, "%s: filter_destroy(group="LPU64",oid="
4099 LPU64")\n", obd->obd_name, oa->o_seq, oa->o_id);
4101 dchild = filter_fid2dentry(obd, NULL, oa->o_seq, oa->o_id);
4103 GOTO(cleanup, rc = PTR_ERR(dchild));
4106 if (dchild->d_inode == NULL) {
4107 CDEBUG(D_INODE, "destroying non-existent object "POSTID"\n",
4108 oa->o_id, oa->o_seq);
4109 /* If object already gone, cancel cookie right now */
4110 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4111 struct llog_ctxt *ctxt;
4112 struct obd_llog_group *olg;
4114 olg = filter_find_olg(obd, oa->o_seq);
4116 CERROR(" %s: can not find olg of group %d\n",
4117 obd->obd_name, (int)oa->o_seq);
4118 GOTO(cleanup, rc = PTR_ERR(olg));
4120 fcc = &oa->o_lcookie;
4121 ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
4122 llog_cancel(ctxt, NULL, 1, fcc, 0);
4123 llog_ctxt_put(ctxt);
4124 fcc = NULL; /* we didn't allocate fcc, don't free it */
4126 GOTO(cleanup, rc = -ENOENT);
4129 rc = filter_prepare_destroy(obd, oa->o_id, oa->o_seq, &lockh);
4133 /* Our MDC connection is established by the MDS to us */
4134 if (oa->o_valid & OBD_MD_FLCOOKIE) {
4135 OBD_ALLOC(fcc, sizeof(*fcc));
4137 *fcc = oa->o_lcookie;
4139 ll_vfs_dq_init(dchild->d_inode);
4141 /* we're gonna truncate it first in order to avoid possible deadlock:
4143 * open trasaction open transaction
4144 * down(i_zombie) down(i_zombie)
4145 * restart transaction
4146 * (see BUG 4180) -bzzz
4148 * take i_alloc_sem too to prevent other threads from writing to the
4149 * file while we are truncating it. This can cause lock ordering issue
4150 * between page lock, i_mutex & starting new journal handle.
4151 * (see bug 20321) -johann
4154 down_write(&dchild->d_inode->i_alloc_sem);
4155 LOCK_INODE_MUTEX(dchild->d_inode);
4156 fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
4158 /* VBR: version recovery check */
4159 rc = filter_version_get_check(exp, oti, dchild->d_inode);
4161 UNLOCK_INODE_MUTEX(dchild->d_inode);
4162 up_write(&dchild->d_inode->i_alloc_sem);
4166 handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
4168 if (IS_ERR(handle)) {
4169 UNLOCK_INODE_MUTEX(dchild->d_inode);
4170 up_write(&dchild->d_inode->i_alloc_sem);
4171 GOTO(cleanup, rc = PTR_ERR(handle));
4174 iattr.ia_valid = ATTR_SIZE;
4176 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
4177 rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
4178 UNLOCK_INODE_MUTEX(dchild->d_inode);
4179 up_write(&dchild->d_inode->i_alloc_sem);
4183 GOTO(cleanup, rc = rc2);
4185 /* We don't actually need to lock the parent until we are unlinking
4186 * here, and not while truncating above. That avoids holding the
4187 * parent lock for a long time during truncate, which can block other
4188 * threads from doing anything to objects in that directory. bug 7171 */
4189 dparent = filter_parent_lock(obd, oa->o_seq, oa->o_id);
4190 if (IS_ERR(dparent))
4191 GOTO(cleanup, rc = PTR_ERR(dparent));
4192 cleanup_phase = 3; /* filter_parent_unlock */
4194 LOCK_INODE_MUTEX(dchild->d_inode);
4195 handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
4196 if (IS_ERR(handle)) {
4197 UNLOCK_INODE_MUTEX(dchild->d_inode);
4198 GOTO(cleanup, rc = PTR_ERR(handle));
4200 cleanup_phase = 4; /* fsfilt_commit */
4202 /* Quota release need uid/gid of inode */
4203 obdo_from_inode(oa, dchild->d_inode, NULL, OBD_MD_FLUID|OBD_MD_FLGID);
4205 filter_fmd_drop(exp, oa->o_id, oa->o_seq);
4207 /* this drops dchild->d_inode->i_mutex unconditionally */
4208 rc = filter_destroy_internal(obd, oa->o_id, oa->o_seq, dparent, dchild);
4212 switch(cleanup_phase) {
4215 sync = fsfilt_add_journal_cb(obd, 0, oti ?
4216 oti->oti_handle : handle,
4217 filter_cancel_cookies_cb,
4219 /* If add_journal_cb failed, then filter_finish_transno
4220 * will commit the handle and we will do a sync
4221 * on commit. then we call callback directly to free
4224 rc = filter_finish_transno(exp, NULL, oti, rc, sync);
4226 filter_cancel_cookies_cb(obd, 0, fcc, rc);
4229 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
4231 CERROR("error on commit, err = %d\n", rc2);
4238 filter_parent_unlock(dparent);
4240 filter_fini_destroy(obd, &lockh);
4244 OBD_FREE(fcc, sizeof(*fcc));
4246 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4249 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
4253 /* trigger quota release */
4254 qcids[USRQUOTA] = oa->o_uid;
4255 qcids[GRPQUOTA] = oa->o_gid;
4256 rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
4259 CERROR("filter adjust qunit! (rc:%d)\n", rc2);
4263 /* NB start and end are used for punch, but not truncate */
4264 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
4265 struct obd_trans_info *oti,
4266 struct ptlrpc_request_set *rqset)
4271 if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
4272 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
4273 oinfo->oi_policy.l_extent.end);
4277 CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
4278 ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,oinfo->oi_oa->o_valid,
4279 oinfo->oi_policy.l_extent.start);
4281 oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
4282 oinfo->oi_oa->o_valid |= OBD_FL_TRUNC;
4283 rc = filter_setattr(exp, oinfo, oti);
4284 oinfo->oi_oa->o_valid &= ~OBD_FL_TRUNC;
4288 static int filter_sync(struct obd_export *exp, struct obd_info *oinfo,
4289 obd_off start, obd_off end,
4290 struct ptlrpc_request_set *set)
4292 struct lvfs_run_ctxt saved;
4293 struct obd_device_target *obt;
4294 struct dentry *dentry;
4298 rc = filter_auth_capa(exp, NULL, oinfo->oi_oa->o_seq,
4299 (struct lustre_capa *)oinfo->oi_capa,
4300 CAPA_OPC_OSS_WRITE);
4304 obt = &exp->exp_obd->u.obt;
4306 /* An objid of zero is taken to mean "sync whole filesystem" */
4307 if (!oinfo->oi_oa || !(oinfo->oi_oa->o_valid & OBD_MD_FLID)) {
4308 rc = fsfilt_sync(exp->exp_obd, obt->obt_sb);
4309 /* Flush any remaining cancel messages out to the target */
4310 filter_sync_llogs(exp->exp_obd, exp);
4314 dentry = filter_oa2dentry(exp->exp_obd, &oinfo->oi_oa->o_oi);
4316 RETURN(PTR_ERR(dentry));
4318 push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4320 LOCK_INODE_MUTEX(dentry->d_inode);
4322 rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
4324 /* just any file to grab fsync method - "file" arg unused */
4325 struct file *file = obt->obt_rcvd_filp;
4327 if (file->f_op && file->f_op->fsync)
4328 rc = file->f_op->fsync(NULL, dentry, 1);
4330 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
4334 UNLOCK_INODE_MUTEX(dentry->d_inode);
4336 oinfo->oi_oa->o_valid = OBD_MD_FLID;
4337 obdo_from_inode(oinfo->oi_oa, dentry->d_inode, NULL,
4338 FILTER_VALID_FLAGS);
4340 pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
4346 static int filter_get_info(struct obd_export *exp, __u32 keylen,
4347 void *key, __u32 *vallen, void *val,
4348 struct lov_stripe_md *lsm)
4350 struct obd_device *obd;
4353 obd = class_exp2obd(exp);
4355 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
4359 if (KEY_IS(KEY_BLOCKSIZE)) {
4360 __u32 *blocksize = val;
4362 if (*vallen < sizeof(*blocksize))
4364 *blocksize = obd->u.obt.obt_sb->s_blocksize;
4366 *vallen = sizeof(*blocksize);
4370 if (KEY_IS(KEY_BLOCKSIZE_BITS)) {
4371 __u32 *blocksize_bits = val;
4372 if (blocksize_bits) {
4373 if (*vallen < sizeof(*blocksize_bits))
4375 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
4377 *vallen = sizeof(*blocksize_bits);
4381 if (KEY_IS(KEY_LAST_ID)) {
4382 obd_id *last_id = val;
4383 /* FIXME: object groups */
4385 if (*vallen < sizeof(*last_id))
4387 *last_id = filter_last_id(&obd->u.filter,
4388 exp->exp_filter_data.fed_group);
4390 *vallen = sizeof(*last_id);
4394 if (KEY_IS(KEY_FIEMAP)) {
4395 struct ll_fiemap_info_key *fm_key = key;
4396 struct dentry *dentry;
4397 struct ll_user_fiemap *fiemap = val;
4398 struct lvfs_run_ctxt saved;
4401 if (fiemap == NULL) {
4402 *vallen = fiemap_count_to_size(
4403 fm_key->fiemap.fm_extent_count);
4407 dentry = __filter_oa2dentry(exp->exp_obd, &fm_key->oa.o_oi,
4410 RETURN(PTR_ERR(dentry));
4412 memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
4413 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4414 rc = fsfilt_iocontrol(obd, dentry, FSFILT_IOC_FIEMAP,
4416 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
4422 if (KEY_IS(KEY_SYNC_LOCK_CANCEL)) {
4423 *((__u32 *) val) = obd->u.filter.fo_sync_lock_cancel;
4424 *vallen = sizeof(__u32);
4428 CDEBUG(D_IOCTL, "invalid key\n");
4432 static inline int filter_setup_llog_group(struct obd_export *exp,
4433 struct obd_device *obd,
4436 struct obd_llog_group *olg;
4437 struct llog_ctxt *ctxt;
4440 olg = filter_find_create_olg(obd, group);
4442 RETURN(PTR_ERR(olg));
4444 llog_group_set_export(olg, exp);
4446 ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
4447 LASSERTF(ctxt != NULL, "ctxt is null\n");
4449 rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
4450 llog_ctxt_put(ctxt);
4454 static int filter_set_grant_shrink(struct obd_export *exp,
4455 struct ost_body *body)
4457 /* handle shrink grant */
4458 cfs_spin_lock(&exp->exp_obd->obd_osfs_lock);
4459 filter_grant_incoming(exp, &body->oa);
4460 cfs_spin_unlock(&exp->exp_obd->obd_osfs_lock);
4466 static int filter_set_mds_conn(struct obd_export *exp, void *val)
4468 struct obd_device *obd;
4474 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4478 LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
4479 obd_export_nid2str(exp));
4480 obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
4482 /* setup llog imports */
4484 group = (int)(*(__u32 *)val);
4486 group = 0; /* default value */
4488 LASSERT_SEQ_IS_MDT(group);
4489 rc = filter_setup_llog_group(exp, obd, group);
4493 if (group == FID_SEQ_OST_MDT0) {
4494 /* setup llog group 1 for interop */
4495 filter_setup_llog_group(exp, obd, FID_SEQ_LLOG);
4498 lquota_setinfo(filter_quota_interface_ref, obd, exp);
4503 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
4504 void *key, __u32 vallen, void *val,
4505 struct ptlrpc_request_set *set)
4507 struct obd_device *obd;
4512 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
4516 if (KEY_IS(KEY_CAPA_KEY)) {
4518 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
4520 CERROR("filter update capability key failed: %d\n", rc);
4524 if (KEY_IS(KEY_REVIMP_UPD)) {
4525 filter_revimp_update(exp);
4526 lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
4530 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4531 filter_adapt_sptlrpc_conf(obd, 0);
4535 if (KEY_IS(KEY_MDS_CONN))
4536 RETURN(filter_set_mds_conn(exp, val));
4538 if (KEY_IS(KEY_GRANT_SHRINK))
4539 RETURN(filter_set_grant_shrink(exp, val));
4544 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
4545 int len, void *karg, void *uarg)
4547 struct obd_device *obd = exp->exp_obd;
4548 struct obd_ioctl_data *data = karg;
4552 case OBD_IOC_ABORT_RECOVERY: {
4553 LCONSOLE_WARN("%s: Aborting recovery.\n", obd->obd_name);
4554 target_stop_recovery_thread(obd);
4558 case OBD_IOC_SYNC: {
4559 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
4560 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4564 case OBD_IOC_SET_READONLY: {
4566 struct super_block *sb = obd->u.obt.obt_sb;
4567 struct inode *inode = sb->s_root->d_inode;
4568 BDEVNAME_DECLARE_STORAGE(tmp);
4569 CERROR("*** setting device %s read-only ***\n",
4570 ll_bdevname(sb, tmp));
4572 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
4573 if (!IS_ERR(handle))
4574 rc = fsfilt_commit(obd, inode, handle, 1);
4576 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
4577 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
4579 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
4583 case OBD_IOC_CATLOGLIST: {
4584 rc = llog_catalog_list(obd, 1, data);
4588 case OBD_IOC_LLOG_CANCEL:
4589 case OBD_IOC_LLOG_REMOVE:
4590 case OBD_IOC_LLOG_INFO:
4591 case OBD_IOC_LLOG_PRINT: {
4592 /* FIXME to be finished */
4593 RETURN(-EOPNOTSUPP);
4595 struct llog_ctxt *ctxt = NULL;
4597 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4598 rc = llog_ioctl(ctxt, cmd, data);
4599 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
4612 static int filter_health_check(struct obd_device *obd)
4614 #ifdef USE_HEALTH_CHECK_WRITE
4615 struct filter_obd *filter = &obd->u.filter;
4620 * health_check to return 0 on healthy
4621 * and 1 on unhealthy.
4623 if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4626 #ifdef USE_HEALTH_CHECK_WRITE
4627 LASSERT(filter->fo_obt.obt_health_check_filp != NULL);
4628 rc |= !!lvfs_check_io_health(obd, filter->fo_obt.obt_health_check_filp);
4633 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4636 return filter_fid2dentry(data, NULL, gr, id);
4639 static int filter_process_config(struct obd_device *obd, obd_count len,
4642 struct lustre_cfg *lcfg = buf;
4643 struct lprocfs_static_vars lvars;
4646 switch (lcfg->lcfg_command) {
4648 lprocfs_filter_init_vars(&lvars);
4650 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4660 static struct lvfs_callback_ops filter_lvfs_ops = {
4661 l_fid2dentry: filter_lvfs_fid2dentry,
4664 static struct obd_ops filter_obd_ops = {
4665 .o_owner = THIS_MODULE,
4666 .o_get_info = filter_get_info,
4667 .o_set_info_async = filter_set_info_async,
4668 .o_setup = filter_setup,
4669 .o_precleanup = filter_precleanup,
4670 .o_cleanup = filter_cleanup,
4671 .o_connect = filter_connect,
4672 .o_reconnect = filter_reconnect,
4673 .o_disconnect = filter_disconnect,
4674 .o_ping = filter_ping,
4675 .o_init_export = filter_init_export,
4676 .o_destroy_export = filter_destroy_export,
4677 .o_statfs = filter_statfs,
4678 .o_getattr = filter_getattr,
4679 .o_unpackmd = filter_unpackmd,
4680 .o_create = filter_create,
4681 .o_setattr = filter_setattr,
4682 .o_destroy = filter_destroy,
4683 .o_brw = filter_brw,
4684 .o_punch = filter_truncate,
4685 .o_sync = filter_sync,
4686 .o_preprw = filter_preprw,
4687 .o_commitrw = filter_commitrw,
4688 .o_llog_init = filter_llog_init,
4689 .o_llog_connect = filter_llog_connect,
4690 .o_llog_finish = filter_llog_finish,
4691 .o_iocontrol = filter_iocontrol,
4692 .o_health_check = filter_health_check,
4693 .o_process_config = filter_process_config,
4696 quota_interface_t *filter_quota_interface_ref;
4697 extern quota_interface_t filter_quota_interface;
4699 static int __init obdfilter_init(void)
4701 struct lprocfs_static_vars lvars;
4704 /** sanity check for group<->mdsno conversion */
4705 for (i = 0; i < MAX_MDT_COUNT; i++)
4706 LASSERT(objseq_to_mdsno(mdt_to_obd_objseq(i)) == i);
4708 lprocfs_filter_init_vars(&lvars);
4710 cfs_request_module("%s", "lquota");
4711 OBD_ALLOC(obdfilter_created_scratchpad,
4712 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4713 sizeof(*obdfilter_created_scratchpad));
4714 if (obdfilter_created_scratchpad == NULL)
4717 ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4718 sizeof(struct filter_mod_data),
4721 GOTO(out, rc = -ENOMEM);
4723 filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4724 init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4726 rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4727 LUSTRE_OST_NAME, NULL);
4731 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4732 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4733 ll_fmd_cachep = NULL;
4735 if (filter_quota_interface_ref)
4736 PORTAL_SYMBOL_PUT(filter_quota_interface);
4738 OBD_FREE(obdfilter_created_scratchpad,
4739 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4740 sizeof(*obdfilter_created_scratchpad));
4746 static void __exit obdfilter_exit(void)
4748 if (filter_quota_interface_ref)
4749 PORTAL_SYMBOL_PUT(filter_quota_interface);
4751 if (ll_fmd_cachep) {
4752 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4753 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4754 ll_fmd_cachep = NULL;
4757 class_unregister_type(LUSTRE_OST_NAME);
4758 OBD_FREE(obdfilter_created_scratchpad,
4759 OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4760 sizeof(*obdfilter_created_scratchpad));
4763 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4764 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4765 MODULE_LICENSE("GPL");
4767 module_init(obdfilter_init);
4768 module_exit(obdfilter_exit);